content
stringlengths 1
15.9M
|
|---|
\section{Introduction}
Neutron stars are macroscopic objects formed by gravitating nuclear matter. Unfortunately, it is not possible to study neutron stars within the fundamental quantum theory of strong interactions coupled to gravity. The obvious reason is that neutron stars are made of neutrons (and a fraction of protons, electrons and perhaps some other particles) which are low energy non-perturbative excitations of QCD. Presently, it is not known how to describe non-perturbative phenomena of strongly interacting matter directly from QCD. To overcome this difficulty and study the low energy sector of QCD, one has to deal with effective field theories (EFTs) which are usually proposed (or motivated by some general arguments) rather than derived from the underlying fundamental theory. Even if one assumes an effective low energy action of QCD (for example quantum hadron dynamics, or the Nambu-Jona-Lasino model or some bag models), however, it is not possible to find neutron stars as solutions of that field theory coupled to Einstein gravity. This most obvious way to describe neutron stars is usually too complicated. A widely accepted solution of this problem is provided by the Tolman-Oppenheimer-Volkoff (TOV) approach \cite{OV}, \cite{Tol} where the Einstein equations are solved for a perfect fluid energy-momentum tensor. To complete the set of equations,
one usually assumes that the perfect fluid is of a barotropic nature with a barotropic equation of state (EoS),
i.e., an algebraic relation between the energy density $\rho$ and the pressure $p$. This is the place where a particular EFT finally enters.
To derive the equation of state from a given effective action (without gravity), however, is again a non-trivial task, since EFTs at the fundamental level (i.e., using their field degrees of freedom and the corresponding action) usually are not perfect fluids. Therefore, one has to go from a microscopic (field theoretic) description to a macroscopic description and a thermodynamic limit must be performed, which typically implies a mean-field approximation. In this manner, an average (spatially constant) value of the energy density is calculated and all deviations from spacial constancy are ignored. However, as gravity directly couples to derivative terms, any deviation from a constant value may influence some properties of neutron stars, like the maximal mass, maximal radius, or the mass-radius relation. Hence, studying this phenomenon is an important but, at the same time, complicated issue. The main difficulty comes from the fact that a full field theoretic calculation is required if we want to understand the error introduced by a mean field approximation. So we need a low energy effective field theory of QCD which is capable of describing nuclear matter and which has a perfect fluid energy-momentum tensor already at the microscopic (field theoretic) level. These two requirements are very strong and, in principle, one may even doubt whether such a model exists, at all.
Quite surprisingly, it has been shown recently that there exists an effective low energy action, the so-called {\em BPS Skyrme model}, for which the corresponding energy-momentum tensor possesses the {\it perfect fluid} form and, therefore, no mean-field approximation has to be made. Furthermore, one can solve this model even after its coupling to gravity and find solutions representing neutron stars. Hence, this theory can also serve as a laboratory for studying the validity and accuracy of the mean field approximation, which is one of the objectives of the present work.
Among EFTs for low energy QCD, Skyrme type models play a very prominent role \cite{skyrme}, \cite{sk}. It has been conjectured by Skyrme that baryons and nuclei may be described as a sort of {\it vorticity} in a mesonic fluid or, in a more precise mathematical language, as solitons in a mesonic effective field theory \cite{skyrme}.
This proposal received some further motivation from the large $N_c$ expansion. Indeed, QCD in the limit of an infinite number of colours becomes equivalent to a a weakly interacting theory with only mesonic degrees of freedom \cite{thooft}. Unfortunately, the derivation of the effective mesonic theory from QCD is still an open problem and its correct form is unknown. In the simplest form, as proposed by Skyrme, the Skyrme model lagrangian reads
\begin{equation}
\mathcal{L}=\mathcal{L}_2+\mathcal{L}_4+\mathcal{L}_0
\end{equation}
where the first two terms represent the sigma model (kinetic) part and the so-called Skyrme part
\begin{equation}
\mathcal{L}_2= \lambda_2 \mbox{Tr} \; \partial_\mu U \partial U^\dagger, \;\; \mathcal{L}_4=\lambda_4 \mbox{Tr} ([L_\mu, L_\nu])^2 ,
\end{equation}
and $\mathcal{L}_0=-\lambda_0 \mathcal{U} (\mbox{Tr} \; U)$ is a non-derivative term, i.e., a potential. Usually one chooses $\mathcal{U}=\mathcal{U}_\pi = 1-\mbox{Tr} \, U$, which provides a mass for the perturbative excitations (the pions). Here, all $\lambda_n$ are non-negative, dimensionful coupling constants. Moreover, $U$ is an $SU(2)$ valued matrix Skyrme field and $L_\mu = U^\dagger \partial_\mu U$ is the left-invariant Maurer-Cartan current. The baryon number is identified as a topological charge $B$ corresponding to the conserved topological current $\mathcal{B}^\mu$
\begin{equation}
\mathcal{B}^\mu = \frac{1}{24\pi^2} \epsilon^{\mu \nu \rho \sigma} \mbox{Tr} \; L_\nu L_\rho L_\sigma, \;\;\; B= \int d^3 x \mathcal{B}^0 .
\end{equation}
From a phenomenological perspective, the original Skyrme action has been quite successful in many respects. After a semiclassical rigid body quantization (and carefully taking into account the Finkelstein-Rubinstein constrains \cite{FR}, \cite{FR-Kr}), it provides a good description of nucleons (charge $B=1$ sector) \cite{anw} as well as the deuteron \cite{braaten} and some additional light nuclei \cite{carson}. Furthermore, it gives rotational excitation bands of some light nuclei with a good qualitative as well as quantitative agreement with experimental results \cite{light} (for a possible improvement related to deformations of spinning skyrmions see \cite{rot}). As a very recent example, the model allows to model the structure and properties (e.g., the so-called Hoyle state) of Carbon 12 \cite{carbon}.
This notable success of the original Skyrme model applied to excited states of the lowest nuclei has to be contrasted with its inability to describe the proper binding energies of atomic nuclei. Skyrmions with higher topological charges have binding energies approximately 10 times bigger than experimentally measured. Another important issue is the fact that skyrmions with large baryon number form crystals, which again is in contrast to the liquid behaviour of nuclear matter. In order to cure the binding energy problem, two possible generalizations of the original Lagrangian to a so-called {\it near BPS} Skyrme model have been proposed. Although they explore the same idea - to bring the model close to a BPS theory - they achieve it in rather different ways. One possibility is to flow the usual (static) Skyrme model into the conformal Yang-Mills $SU(2)$ theory in four dimensional Euclidean space by the inclusion of (infinitely many) higher vector mesons \cite{SutBPS} (see also \cite{rho}). Alternatively, one can add a dominating BPS submodel based entirely on the mesonic Skyrme field \cite{BPS}
\begin{equation}
\mathcal{L}_{BPS} \equiv \mathcal{L}_{06}=\mathcal{L}_6+\tilde{\mathcal{L}}_0,
\end{equation}
where the derivative dependent sextic term is just the baryon current squared
\begin{equation}
\mathcal{L}_6= -(24\pi^2)^2 \lambda_6 \mathcal{B}_\mu \mathcal{B}^\mu ,
\end{equation}
and $\tilde{\mathcal{L}}_0$ is again a potential.
Then, the full {\it near BPS Skyrme} model reads
\begin{equation} \label{near-BPS}
\mathcal{L}=\mathcal{L}_6+\tilde{\mathcal{L}}_0+ \epsilon( \mathcal{L}_2+\mathcal{L}_4+\mathcal{L}_0 )
\end{equation}
where $\epsilon$ is a small parameter (see also \cite{Marl}, \cite{Sp2}; for the influence of the sextic term on the usual
Skyrme model, see, e.g., \cite{sextic}).
It has been shown that this improved Skyrme action, already in the exact BPS limit $\epsilon \to 0$, leads to very accurate binding energies if one takes into account the semiclassical quantization of (iso)-rotational degrees of freedoms, the Coulomb interaction as well as the isospin symmetry breaking \cite{nearBPS}. Furthermore, the BPS Skyrme static energy functional is invariant under volume preserving diffeomorphisms (VPDs) on physical space, which are the symmetries of a perfect fluid. In addition, the energy-momentum tensor of the BPS Skyrme model is the energy-momentum tensor of a perfect fluid. This solves the second, before mentioned, long standing problem in the Skyrme framework. All these results support the conjecture that the near BPS Skyrme model might be the right low-energy EFT for the description of nuclear matter. One has to underline that already the BPS part of the full near BPS model gives a very good approximation to some static properties of nuclear matter (masses, binding energies). This is important since the BPS sector is {\it solvable} which renders computations possible. We remark that another variant of the Skyrme model leading to low binding energies, where a 'repulsive' potential is added to the standard Skyrme action, and whose existence is, again, based on topological energy bounds \cite{harland}, \cite{top-bounds}, has been proposed and investigated recently in \cite{har-gil-sp}.
The solvability of the BPS Skyrme model allows to find solitonic solutions after coupling to gravity \cite{star}. The important point is that the obtained gravitating skyrmions (neutron stars) are solutions of the field equations with the Einstein gravity fully taken into account. This means that no assumption on an EoS was made. Instead, the Skyrme matter acts as a source (energy-momentum tensor) for the Einstein equations (for previous applications of skyrmions in the context of neutron stars and as a source for GR equations see \cite{bizon} - \cite{gravSk}). The validity of this approach follows from the observation that, as mentioned above, the energy-momentum tensor is of the perfect fluid type and no thermodynamical limit has to be taken to reach a fluid regime. An (algebraic) on-shell equation of state may be computed after solving the equations, but it turns out that this on-shell EoS is not unique but depends on the total mass of the neutron star, i.e., on the specific solution. This is related to the fact that the original EoS in the BPS Skyrme model (without gravity) is not of an algebraic form $\rho = \rho (P)$ (i.e., the perfect fluid is not barotropic), but depends on spatial coordinates i.e., $\rho=\rho(r,P)$ since the energy density $\rho$ is typically not spatially constant.
$P$, on the other hand, is an integration constant in the field theory context, so it is constant by definition.
The objective of the present paper is twofold. We shall continue the investigation of neutron star properties within the BPS Skyrme model commenced in \cite{star}, and we want to compare the full field theory and gravity computations with the mean-field TOV approach. To accomplished this second aim, we have to derive an algebraic (global) EoS from the correct (coordinate dependent) EoS by a sort of mean-field approximation.
The outline of the paper is the following. In the next section, we briefly summarize some thermodynamical properties of the BPS Skyrme model. In section III it is shown how an average, mean-field equation of state (MF-EoS) may be derived from the exact EoS. In section IV we describe the results of our numerical neutron star calculations and compare the neutron star properties obtained in the full field theoretical and gravity set-up with the TOV approach, where the MF-EoS is used. In section V, we discuss our results, with emphasis on the mass-radius relation ($M(R)$ curve), and including the TOV inversion issue. We leave the comparison with other approaches to section VI where, in particular, we compare both to generic and to Skyrme-model related results. Finally, in section VII we summarize our conclusion and give some outlook.
\section{Equation of state in the BPS Skyrme model}
One of the most striking features of the BPS Skyrme model is the fact that, at the same time, it provides both a description in terms of microscopic degrees of freedom (mesonic fields in the BPS Skyrme action) and a description by macroscopic quantities, i.e., thermodynamical functions \cite{term}. Both descriptions agree completely in the sense that the field theoretical energy $E$ and pressure $P$ together with the geometrical volume $V$ obey the standard thermodynamical relation
\begin{equation}
P=-\frac{d E}{d V}.
\end{equation}
The main consequence of this result is that, in the BPS Skyrme model, we are all the time in a thermodynamical limit. In other words, no additional thermodynamical limit (for example a kind of mean field approximation) has to be performed to reach a macroscopic description in the language of the proper thermodynamical functions. This is a profound difference between the BPS Skyrme model and any other low energy effective theory of Quantum Chromodynamics.
Below we briefly summarize some thermodynamical properties of the BPS Skyrme model \cite{term}.
\\
First of all, the energy-momentum tensor of the BPS Skyrme model has the form of the energy-momentum tensor of a perfect fluid (for the moment we consider the case of flat Minkowski space, i.e., without gravity)
\begin{equation}
T^{\mu \nu} = (p+\rho) u^\mu u^\nu - p \eta^{\mu \nu}
\end{equation}
where the energy density and pressure for static configurations, which are the ones relevant for us (and where $u^\mu = (1,0,0,0)$), read
(for convenience, we introduce the new coupling constants $\lambda^2 = (24)^2 \lambda_6$ and $\mu^2 = \lambda_0$)
\begin{equation}
\rho = \lambda^2 \pi^4 \mathcal{B}_0^2+\mu^2\mathcal{U}, \;\;\; p=\lambda^2 \pi^4 \mathcal{B}_0^2-\mu^2\mathcal{U} .
\end{equation}
From the conservation equation $\partial_\mu T^{\mu \nu}=0$ we get $p=P=const.$ Hence, the pressure inside BPS skyrmions (skyrmionic matter) is constant.
\\
Moreover, the equation which defines the pressure is, in fact, a first integral of the full second order equation of motion. It provides solutions (field configurations with nontrivial baryonic charge) with non-zero (external) pressure. In particular, for the axially symmetric ansatz
\begin{equation}
U=\cos \xi + i \sin \xi \vec{n} \cdot \vec{\tau}
\end{equation}
where $\vec{\tau}$ are Pauli matrices and
\begin{equation}
\xi =\xi (r), \;\;\; \vec{n}=(\sin \theta \cos B \phi, \sin \theta \sin B \phi, \cos \theta)
\end{equation}
we get ($\xi_r \equiv \partial_r \xi$)
\begin{equation} \label{BPS-Pnot0}
\frac{|B|\lambda}{2r^2} \sin^2 \xi \xi_r= - \mu \sqrt{\mathcal{U}+\frac{P}{\mu^2}}
\end{equation}
(the usual BPS equation is obtained as a zero pressure condition
\cite{bazeia}).
Solving this equation with topologically nontrivial boundary conditions ($\xi(0)=\pi, \; \xi(R)=0$), we get a profile of the skyrmion. Here $R$ is the geometric radius of the soliton. Inserting it back into the energy density we get a spatially dependent function which also changes if the pressure is varied, i.e.,
\begin{equation}
\rho = \rho (r, P) .
\end{equation}
This formula can be regarded as an on-shell density-pressure equation of state in the BPS Skyrme model for spherically symmetric nuclear matter.
\\
Another important quantity considered in the context of nuclear matter is the particle density. Here, it is equivalent to the baryon topological charge density $\mathcal{B}_0$. The corresponding equation of state may be easily read-off from the components of the energy-momentum tensor. Namely,
\begin{equation}
\mathcal{B}_0 (r,P)= \frac{1}{\lambda \pi^2} \sqrt{\rho(r,P)+P}
\end{equation}
where the plus sign has been assumed as we have baryons (not anti-baryons). This, again, is the particle number - pressure equation of state which together with the above energy density - pressure EoS completely defines the system. Obviously, in our construction both equations of state
follow from the BPS Skyrme action.
\\
Furthermore, one can find the total energy and the geometrical volume
\begin{equation}
E(P)=2\pi \lambda \mu |B|\tilde{E}, \;\;\; V=2\pi |B| \frac{\lambda}{\mu} \tilde{V}
\end{equation}
where
\begin{equation} \label{tilde-E}
\tilde{E}=\int_0^\pi d\xi \sin^2 \xi \frac{2\mathcal{U}+\tilde{P}}{\sqrt{\mathcal{U}+\tilde{P}}}, \;\;\; \tilde{V}=\int_0^\pi d\xi \sin^2 \xi \frac{1}{\sqrt{\mathcal{U}+\tilde{P}}} .
\end{equation}
Here, the base space integrals defining $E$ and $V$ in the usual way may be converted into the (solution-independent) target space integrals (\ref{tilde-E}) with the help of the first-order (BPS) equation (\ref{BPS-Pnot0}).
Further,
$\tilde{P}=P/\mu^2$. In contrast to the density, the total energy and the volume are related to the pressure in a global (coordinate independent) manner. So, some well-defined energy-pressure $E=E(P)$ as well as volume-pressure $V=V(P)$ equations of state do exist. It should be noted that the energy and volume can be found without knowing a particular solution. They are given by integrals of some functions of the potential over the target space.
\section{Mean-field equation of state - a hadronic bag model}
As underlined already, it is an inherent property of the BPS Skyrme model (and undoubtedly its near BPS extensions) that the energy density of skyrmionic matter is not constant - except for the step function potential. On the other hand, the model allows to describe solutions with non-zero external pressure, and the pressure is always {\it constant} inside skyrmions (nuclear matter). Therefore, the energy density is a space (radial) dependent function which changes if a non-zero pressure is imposed,
\begin{equation}
\rho=\rho(r, P), \;\; P=const.
\end{equation}
It should be emphasized, again, that this is the proper equation of state of skyrmionic matter in the BPS Skyrme model. Hence, in general, the pressure and the energy density are not related by an algebraic equation (which does not depend on coordinates) i.e., there is no (barotropic) equation of state of the form
\begin{equation}
\rho = \rho (P).
\end{equation}
This may be a source of some difficulties if one wants to compare the thermodynamical properties of the BPS Skyrme model with the usually used density-pressure equations of state derived in other effective field theories.
However, it is possible to cast the equation of state obtained in the BPS Skyrme model into such an algebraic (global) form. This requires a simple averaging procedure, which may be interpreted as a mean field approximation. The obvious definition of the average energy density is
\begin{equation}
\bar{\rho}(P)=\frac{E(P)}{V(P)}
\end{equation}
and the resulting mean field equation of state (MF-EoS) is just a function of some target space integrals (averages)
\begin{equation}
\bar{\rho}(P)=\mu^2 \frac{\left<\frac{2\mathcal{U}}{\sqrt{\mathcal{U}+P/\mu^2}}\right>}{\left< \frac{1}{\sqrt{\mathcal{U}+P/\mu^2}} \right>} +P.
\end{equation}
Here
\begin{equation}
\left< F(\mathcal{U})\right> \equiv \frac{1}{{\rm Vol}_{{\mathcal S}^3}} \int {\rm vol}_{{\mathcal S}^3} F(\mathcal{U}) = \frac{2}{\pi} \int_0^\pi d\xi \sin^2 \xi F(\mathcal{U})
\end{equation}
where we use that, as a manifold, the target space SU(2) may be identified with the unit three-sphere ${\mathcal S}^3$. Further,
${\rm Vol}_{\mathcal{S}^3} = 2\pi^2 $ is the volume of the (target space) unit three-sphere, and ${\rm vol}_{\mathcal{S}^3}$ the corresponding volume form. The right hand side follows because the potential ${\mathcal U}$, by assumption, depends on the Skyrme field only via the field variable $\xi$.
For later convenience, we also want to define the average baryon density as baryon number divided by volume, i.e.,
\begin{equation} \label{av-b-dens}
\bar n_{B} = \frac{B}{V} = \frac{\mu}{\pi^2 \lambda \left< \frac{1}{\sqrt{\mathcal{U}+P/\mu^2}} \right>}.
\end{equation}
Notice that it is not necessary to know explicit soliton solutions of the model to find the MF-EoS. Instead, it is encoded in geometrical (target space) average values.
Obviously, a particular form of the MF-EoS follows from a particular potential. However, some general observations can be easily made.
\subsection{High pressure limit - a hadronic bag model.}
For large values of the pressure we find the following expansion (up to the first two terms; we use $\left< 1\right> = 1$)
\begin{equation}
\bar{\rho}(P)=\mu^2 \frac{\left< 2\mathcal{U}\right> }{\left< 1\right> }+P= P+ 2 \mu^2 \left< \mathcal{U}\right> .
\end{equation}
This is an EoS of a bag type matter with the asymptotical bag constant $B_\infty$
\begin{equation}
\bar{\rho}(P)=P+B_\infty, \;\;\; B_\infty=2\mu^2 \left< \mathcal{U}\right>
\end{equation}
(in the neutron star context, this EoS is known as the "maximally compact EoS"; see below).
Hence, at least for high pressure and from the average MF-EoS point of view, matter described by the BPS Skyrme model behaves as a matter in a bag type model, where the asymptotic bag constant is given again by a target space average of the potential. In comparison, the MIT bag model leads to the following linear equation of state \cite{MITbag}
\begin{equation}
P=\frac{1}{3} (\bar{\rho}-4B_{\rm MIT}) .
\end{equation}
As the Skyrme model is based entirely on a mesonic matrix field with all quark contributions assumed to be integrated out, we seem to have a kind of {\it hadronic bag model} instead of the typical quark bag model - the BPS Skyrme model describes nuclear matter, not quark matter. Moreover, in the MIT bag model the proportionality constant is $1/3$ instead of 1.
Nevertheless, in the high density (high pressure) limit, the EoS of the BPS Skyrme model coincides with the high-density EoS of several other important models of nuclear matter. First of all, let us observe that in the limit of very high pressure, even the exact (non-constant) energy and baryon number densities become approximately constant ($\rho \sim \bar \rho$ and ${\cal B}_0 \sim \bar n_B \equiv (B/V)$) and lead to the approximate EoS
\begin{equation}
\bar \rho = P = \pi^4 \lambda^2 \bar n_B^2.
\end{equation}
But this is exactly the high-density limit of the EoS of the Walecka model \cite{walecka} (for a good review see, e.g., \cite{schmitt}), where the constant $\pi^4 \lambda^2$ is replaced by $(1/2)(g_\omega^2/m_\omega^2)$ (here $g_\omega$ and $m_\omega$ are the coupling constant and mass of the vector meson of the Walecka model).
Secondly, the same high-density EoS $\bar\rho \sim P$ also occurs in a modification of the MIT bag EoS, where interactions with higher mesons are taken into account. Specifically, for high $P$, contributions from the vector meson interactions always win over the usual free fermion gas part. Therefore, asymptotically one gets again $P \sim \bar{\rho}$ \cite{thomas}, coinciding with the result obtained in the BPS Skyrme model. An explanation of this agreement may be related to the fact that the interaction used in \cite{thomas} in the derivation of the EoS is effectively described by a Nambu-Jona-Lasinio (NJL) interaction term, i.e., the square of the quark vector current
\begin{equation}
J^\mu_{V} = \bar{q} \gamma^\mu q.
\end{equation}
Further, the quark vector current is, up to a constant, just the baryon density current, and the NJL interaction term is, therefore, equivalent to the sextic (baryon density current squared) term in the BPS Skyrme model, although the field parametrization of the current is quite different in the two models - a Noether current based on quark fields in the NJL case, and a topological current based on the Skyrme field in the BPS Skyrme model. As the vector channel gives the leading contribution for the MIT bag EoS in the large pressure limit, the effective action is also dominated by this term. Thus, both models coincide in the high pressure regime, even though the baryonic current is defined by very different degrees of freedom: quark and mesonic, respectively. This observation may point towards the interesting possibility that the BPS Skyrme model is a solitonic realization of a vector MIT bag model, as conjectured by M. Nowak and M. Rho.
\subsection{Small pressure limit}
For small pressure $P$, the power series expansion in $P$ does not work. The reason is that one gets target space averages of arbitrary negative powers of the potential, $\left< \mathcal{U}^{-a}\right>$, where $a>0$. As the potential has at least one zero (the vacuum) we get singularities for sufficiently large $a$, and the expansion breaks down.
\\
In general the small $P$ behaviour is
\begin{equation}
\bar{\rho}(P)=2\mu^2 \frac{<\mathcal{U}^{1/2}>}{<\mathcal{U}^{-1/2}>}+ f(P)
\end{equation}
where $f(P)$ is a non-polynomial function of $P$, such that $f(P)=0$ as $P \rightarrow 0$. Its particular form strongly depends on the potential.
We may interpret the first constant as an equilibrium bag constant (i.e., the average energy density of nuclear matter at equilibrium or nuclear saturation density)
\begin{equation}
B_0=2\mu^2 \frac{<\mathcal{U}^{1/2}>}{<\mathcal{U}^{-1/2}>}.
\end{equation}
Let us notice that this expression is nonzero only if $<\mathcal{U}^{-1/2}>$ is not singular. This condition is equivalent to the fact that the geometrical volume $V$ (which, as we already know, is also the thermodynamical volume) is finite at $P=0$ , which as a consequence leads to the requirement that all skyrmions have to be compactons. However, compact solitons are obtained only if the approach to the vacuum of the potential is weaker than $\xi^6$, i.e., $\lim_{\xi \to 0}{\cal U} \sim \xi^\alpha$ for $\alpha <6$. In this paper, we shall consider only potentials which satisfy this condition.
\subsection{Causality}
There is one physically important condition which our MF-EoS should obey. Namely, the speed of sound
\begin{equation}
\frac{1}{v^2} = \frac{\partial \bar{\rho}}{\partial P}
\end{equation}
has to be smaller than or equal to the speed of light as required by causality,
\begin{equation}
v \leq 1.
\end{equation}
Therefore,
\begin{equation}
\frac{\partial \bar{\rho}}{\partial P} \geq 1
\end{equation}
which leads to the following requirement for the target space averages
\begin{equation}
\left<\mathcal{U} \left(\mathcal{U} +\frac{P}{\mu^2} \right)^{-1/2}\right>
\left<\left(\mathcal{U} +\frac{P}{\mu^2} \right)^{-3/2}\right> -\left<\mathcal{U} \left(\mathcal{U} +\frac{P}{\mu^2} \right)^{-3/2}\right>
\left<\left(\mathcal{U} +\frac{P}{\mu^2} \right)^{-1/2}\right>\geq 0
\end{equation}
This formula provides admissible potentials.
We will check this condition for all potentials considered below.
\\
Furthermore, using the Chebyshev integral inequality it is possible to prove that the asymptotic bag constant $B_\infty$ is always greater than or equal to the equilibrium bag constant $B_0$,
\begin{equation}
B_\infty \geq B_0.
\end{equation}
\subsection{Examples}
\subsubsection{The step function potential}
Let us now consider three particular examples. First of all, we analyze the step function potential
\begin{equation}
\mathcal{U}=\Theta ({\rm Tr}(1- \, U)).
\end{equation}
From a phenomenological point of view, this potential is not the proper choice. It leads to an unphysically large compression modulus \cite{term}. Nevertheless, it is still interesting to consider this case, as for the step function potential the exact equation of state precisely agrees with its mean-field average version. In more physical terms, this implies that the step function potential corresponds to the extreme case leading to completely flat (constant) energy and baryon number densities. Indeed, both exact and MF approaches give the following EoS
\begin{equation}
\rho = P+2\mu^2.
\end{equation}
Hence $B_\infty = B_0=2\mu^2$ and $\bar{\rho}=\rho$. The reason why these two approaches lead to the same answer is the fact that the energy density is a constant function. Thus, the mean field (average) energy density is equal to the energy density computed from the BPS equation. It is also clear why, asymptotically for large pressure, the average MF-EoS for any admissible potential looks the same as for the step function potential. From the BPS equation
\begin{equation}
\frac{|B|\lambda}{2r^2} \sin^2 \xi \xi_r = -\mu\sqrt{\mathcal{U}+\frac{P}{\mu^2}}
\end{equation}
we see that for $P >> \mu^2$, the right hand side is effectively equal to $-\mu \sqrt{P}$. This follows from the observation that the original potential takes values between the vacuum value $\mathcal{U}(\xi=0)=0$ and a maximum value $\mathcal{U}(\xi =\pi)$, which is negligible for sufficiently large pressure. Hence, in this limit, the nontrivial and field dependent right hand side effectively behaves as a field independent (constant) quantity. Therefore, also the left hand side, i.e., the baryon charge density as well as the corresponding energy density are constant, and the resulting EoS must correspond to the EoS of the step function potential.
\subsubsection{The Skyrme potential $\mathcal{U}=\mathcal{U}_\pi$}
Another obvious choice is the Skyrme potential originally used to provide masses for (pionic) field perturbations
\begin{equation}
\mathcal{U}=\mathcal{U}_\pi=2 \sin^2 \frac{\xi}{2} =1-\cos \xi \equiv 2h
\end{equation}
(where we defined the new field variable $h$ for later convenience).
In this case, the energy density obtained from the BPS equation is not constant and the mean-field averaging leads to a different MF-EoS
\begin{equation}
\bar{\rho}=\frac{\mu^2}{5} \left( 2-3 \frac{P}{\mu^2} +\frac{6}{1+\frac{P}{\mu^2} \left( 1- \frac{K \left[ \frac{2}{2+P/\mu^2} \right]}{E \left[ \frac{2}{2+P/\mu^2} \right]} \right)} \right)
\end{equation}
where $K$ and $E$ are the complete elliptic integrals of the first and second kind, respectively. Here, the bag constants are
\begin{equation}
B_\infty= 2 \mu^2, \;\;\; B_0=\frac{8}{5} \mu^2.
\end{equation}
The expansion at zero pressure gives
\begin{equation}
\bar{\rho}=\frac{\mu^2}{5} \left( 8 - \frac{P}{\mu^2} \ln \frac{P}{2 \mu^2} \right)
\end{equation}
where the subleading terms have been omitted. Then, $\frac{\partial \bar{\rho}}{\partial P} =\infty$ at $P=0$. As the MF-EoS is a monotonous function of the pressure and tends to $\bar{\rho}=P+B_\infty$ (see Fig. 1a) we can conclude that the speed of sound is always smaller than 1 and the model satisfies the causality condition.
\begin{figure}
\subfloat[]{
\label{fig:subfig:a}
\includegraphics[height=4.5cm]{h-1.pdf}}
\hspace{0.3in}
\subfloat[]{
\label{fig:subfig:b}
\includegraphics[height=4.5cm]{h-2.pdf}}
\caption{(Color online) The mean-field EoS for $\mathcal{U}_\pi$ and $\mathcal{U}_\pi^2$. Here $\mu^2=1$.}
\end{figure}
\subsubsection{The Skyrme potential squared $\mathcal{U}=\mathcal{U}_\pi^2$}
The last example we will explore further in the next section is the Skyrme potential squared
\begin{equation}
\mathcal{U}=\mathcal{U}_\pi^2.
\end{equation}
Then, we find the following exact expression for the MF-EoS
\begin{equation}
\bar{\rho}=\mu^2 \left( \frac{P}{\mu^2}+ \frac{5}{2}\, \frac{{}_3F_2 [\{ \frac{1}{2}, \frac{7}{4}, \frac{9}{4} \}, \{ \frac{5}{2}, 3 \}, -\frac{4 \mu^2}{P} ] }{{}_3F_2 [\{ \frac{1}{2}, \frac{3}{4}, \frac{5}{4} \}, \{ \frac{3}{2}, 2 \}, -\frac{4 \mu^2}{P} ] } \right)
\end{equation}
where ${}_pF_q[\{a_1,..,a_p\}, \{b_1,..,b_q\},z] $ is a generalized hypergeometric function. Here
\begin{equation}
B_\infty= \frac{5}{2} \mu^2,\;\;\;\; B_0=\mu^2.
\end{equation}
It can be shown that the model obeys the causality requirement. In Fig. 1b we plot this MF-EoS. As is clearly visible, it tends quite rapidly to a linear density-pressure function. Hence, the EoS approaches the EoS for the step function potential rather quickly.
\section{TOV vs. full gravitating field theory}
\subsection{Full field theoretical computations}
The BPS Skyrme model in curved space-time has the following form (for a more detailed analysis see \cite{star})
\begin{equation}
S_{06}=\int d^4 x |g|^{\frac{1}{2}} \left( -\lambda^2 \pi^4 |g|^{-1} g_{\mu \nu} \mathcal{B}^\mu \mathcal{B}^\nu - \mu^2 \mathcal{U} \right) .
\end{equation}
As in the flat space case, the corresponding energy-momentum tensor has a perfect fluid form, which for static solutions and for a diagonal metric reads
\begin{equation}
T^{00}=\rho g^{00}, \;\;\; T^{ij} = - p g^{ij}
\end{equation}
where now the pressure and energy density are metric dependent functions
\begin{equation}
\rho= \lambda^2\pi^4 |g|^{-1} g_{00} \mathcal{B}^0 \mathcal{B}^0 + \mu^2 \mathcal{U}
\end{equation}
\begin{equation}
p= \lambda^2\pi^4 |g|^{-1} g_{00} \mathcal{B}^0 \mathcal{B}^0 - \mu^2 \mathcal{U}.
\end{equation}
In order to find neutron stars in the BPS Skyrme model (in a full field theoretic calculation) one has to solve the Einstein equations
\begin{equation}
G_{\mu \nu} = \frac{\kappa^2}{2} T_{\mu \nu}
\end{equation}
(here $G_{\mu\nu}$ is the Einstein tensor and $\kappa^2 = 16 \pi G = 6.654 \cdot 10^{-41} \, {\rm fm} \, {\rm MeV}^{-1}$)
with the energy-momentum tensor provided by the BPS action and in a spherically symmetric metric
\begin{equation}
ds^2 = {\bf A} (r) dt^2-{\bf B} (r) dr^2 - r^2 (d\theta^2 +\sin^2 \theta d\phi^2).
\end{equation}
The key point is that the full field theoretical equations and the Einstein equations are compatible with the metric ansatz together with the previously introduced axially symmetric ansatz for the Skyrme field. This allows to reduce the equations to a set of three ordinary differential equations: two which couple the Skyrme field (profile) $h\equiv (1/2) (1-\cos \xi)$ and ${\bf B}$, while the third one determines ${\bf A}$ in terms of $h$ and ${\bf B}$
\begin{eqnarray}
\frac{1}{r} \frac{{\bf B}'}{{\bf B} } & =- & \frac{1}{r^2} ({\bf B} -1) +\frac{\kappa^2}{2} {\bf B} \rho \label{eq1} \\
r({\bf B} p)' &=& \frac{1}{2} (1-{\bf B} ) {\bf B} (\rho +3p) +\frac{\kappa^2}{4} r^2 {\bf B}^2 (\rho-p)p \label{eq2}\\
\frac{{\bf A}'}{{\bf A} } &=&\frac{1}{r} ({\bf B} -1) +\frac{\kappa^2}{2} r {\bf B} p \label{eq3}
\end{eqnarray}
where now
\begin{equation} \label{rho-p-r}
\rho = \frac{4B^2\lambda^2}{{\bf B} r^4} h(1-h) h_r^2+\mu^2 \mathcal{U}(h), \;\;\; p=\rho-2\mu^2 \mathcal{U}(h).
\end{equation}
These equations are solved for topologically non-trivial boundary conditions for the Skyrme profile
\begin{equation}
h(r=0)=1, \;\;\; h(r=R)=0
\end{equation}
together with natural conditions for the metric field and the pressure (which is equivalent to a condition for $h_r(r=R)$)
\begin{equation}
{\bf B}(r=0)=1, \;\;\; p(r=R)=0.
\end{equation}
A detailed discussion of these numerical computations is provided in \cite{star}. We want to emphasize that in this field theoretical approach with the gravitational interaction fully taken into account no fixed (unique) EoS has been assumed. On the contrary, the energy density $\rho$ and pressure $p$ are both metric dependent functions and, therefore, it is not expected that any {\it unique} EoS for {\it all} solution may exist.
We remark that a similar geometry-dependent (so-called quasi-local) EoS has been used in
an analysis of gravastars and neutron stars with
anisotropic matter, see, e.g., \cite{cat} - \cite{silva}.
However, for the assumed ansatz, which is natural for neutron stars, one can derive an on-shell EoS. Indeed, as $p$ and $\rho$ (after finding a particular solution) are functions of only one coordinate $r$, we may eliminate it and derive a relation $\rho=\rho(p)$. It should be stressed that this relation has to be found for each solution (each neutron star with a given mass) independently. One important result is that in this full field theoretical approach, there is no unique EoS (even for the spherically symmetric gravitating skyrmions). The obtained equation of state not only relates local quantities (pressure and energy density) but, in addition, depends on a global parameter i.e., the total mass $M$ (or equivalently the baryon charge) of the neutron star, i.e.,
\begin{equation}
\rho=\rho (p, M)
\end{equation}
Numerically, it has been found that the EoS is of a polytropic form $p=a\rho^b$, where the two parameters $a$ and $b$ depend on the total mass \cite{star}.
\subsection{Mean-field approximation computations}
The same set of Einstein equations (\ref{eq1}), (\ref{eq2}) can be used for the usual mean-field TOV computation. The only difference is that now the matter field is "averaged" in the sense that instead of the field dependent energy density and pressure we deal with their mean-field versions related by the mean-field equation of state (MF-EoS) $\bar{\rho}=\bar{\rho} (p)$ introduced above. Computationally it means that we solve (\ref{eq1}), (\ref{eq2}) where $p$ and $\bar \rho$ (which replaces $\rho$) are treated as independent functions, and a fixed (spatially independent) MF-EoS $\bar{\rho}=\bar{\rho} (p)$ is assumed, which closes the system. The only initial condition is ${\bf B}(r=0)=1$.
\\
Obviously, for the step-function potential both approaches are exactly the same, as the full field theoretical EoS is equal to its mean-field version.
\subsection{Parameter values and initial conditions}
Before performing the numerical calculations, we have to choose numerical values for the coupling constants $\lambda$ and $\mu$ of the BPS Skyrme model for the three potentials ${\cal U} = \Theta (h)$, ${\cal U} = {\cal U}_\pi = 2h$ and ${\cal U} = {\cal U}_\pi^2 = 4h^2$ we want to consider. We shall determine these values by fitting the BPS skyrmions to properties of nuclear matter. In \cite{star} we fitted to the mass of the helium nucleus and to the nucleon radius for simplicity, but here we prefer to fit to the binding energy per nucleon of infinite nuclear matter $E_{\rm b} = 16.3 \, {\rm MeV}$ and to the nuclear saturation density (baryon density of nuclear matter at equilibrium at zero pressure), $n_0 = 0.153 \; {\rm fm}^{-3}$ \cite{schmitt}, because infinite nuclear matter is a more appropriate choice for neutron stars. (In any case, the differences in the values of physical observables induced by the two fits are rather small and typically do not exceed 15\%). With the nucleon mass $E_{\rm n} = 939.6 \; {\rm MeV}$, the soliton energy per nucleon is $E_{B=1} = E_{\rm n} - E_{\rm b} = 923.3 \, {\rm MeV}$. Further, $V_{B=1} = (1/0.153)\,{\rm fm}^3$. Then, using the expressions for energy and volume at zero pressure (at nuclear saturation), the fit in the three cases leads to
\begin{equation} \label{param-pot-theta}
\Theta (h): \; \; E = 2\pi^2 B\lambda \mu , \; \; V = \pi^2 B\frac{\lambda}{\mu} \; \;
\Rightarrow \; \; \lambda^2 = 30.99 \; {\rm MeV} \,{\rm fm}^3 , \; \; \mu^2 = 70.61 \; {\rm MeV}\, {\rm fm}^{-3}
\end{equation}
\begin{equation} \label{param-pot-2}
{\cal U}_\pi: \; \; E = \frac{64 \sqrt{2}\pi}{15} B\lambda \mu , \; \; V = \frac{8}{3}\sqrt{2}\pi B\frac{\lambda}{\mu} \; \;
\Rightarrow \; \; \lambda^2 = 26.88 \; {\rm MeV} \,{\rm fm}^3 , \; \; \mu^2 = 88.26 \; {\rm MeV}\, {\rm fm}^{-3}
\end{equation}
\begin{equation} \label{param-pot-4}
{\cal U}_\pi^2: \; \; E = 2\pi^2 B \lambda \mu , \; \; V = 2\pi^2 B \frac{\lambda}{\mu} \; \;
\Rightarrow \; \; \lambda^2 = 15.493 \; {\rm MeV} \,{\rm fm}^3 , \; \; \mu^2 = 141.22 \; {\rm MeV}\, {\rm fm}^{-3}
\end{equation}
It is of some interest to consider the "initial conditions" at $r=0$ and the resulting free integration constants in the two cases (exact field theory, and TOV equations using the MF-EoS, respectively). At first sight, it seems that we have one free constant at $r=0$ in both cases, which may be chosen to be the value of $\rho$ ($\bar \rho$) at the center, i.e., $\rho (0)$ or $\bar \rho (0)$, respectively. Here, in the exact field theory case, $\rho (0)$ is related to the second Taylor coefficient $h_2$ (where $h = 1-h_2 r^2 + {\bf O}(r^3)$) via $\rho (0) = 16 B^2 \lambda^2 h_2^3 + \mu^2 {\cal U}(1)$, as may be checked easily. It seems that for each initial value of $\rho (0)$ or $\bar \rho (0)$ we just have to integrate up to a value $r=R$ where $p(R)=0$ which defines the surface of the resulting neutron star, such that there is one solution per initial value (at least as long as the initial values are not too big, such that the condition $p(R)=0$ can be satisfied for some $R$).
It turns out, however, that the initial value $\rho (0)$ may not be chosen arbitrarily in the exact field theory case. The reason is that, at the surface $r=R$, the Skyrme field $h$ must take its vacuum value $h=0$ where the potential is zero, ${\cal U}(0)=0$. But this immediately implies that, at the surface $r=R$, the energy density must be zero, too, i.e., $\rho (R)=0$. It then follows from Eq. (\ref{eq2}) that a metric function ${\bf B}$ which is nonsingular at the surface may exist only if $p$ satisfies the condition $p'(R)=0$, as well. In other words, we have to impose the condition $p'(R)=0$ in addition to $p(R)=0$, and, in general, both conditions may be satisfied simultaneously at most for a discrete set of initial values $\rho (0)$. We may, nevertheless, find different solutions with different neutron star masses, because in the exact field theory the baryon number $B$ enters as an additional free parameter, see Eq. (\ref{rho-p-r}). Concretely, we find one solution (i.e. one initial condition $\rho(0)$) for sufficiently small values of $B$, two solutions (i.e. two initial values $\rho_1(0)<\rho_2 (0)$) for intermediate values of $B$ (where only the smaller $\rho_1 (0)$ corresponds to a stable solution), and no solution for $B>B_{\rm max}$ (where the value of $B_{\rm max}$ depends on the potential).
In the MF-EoS case, instead, $\bar\rho (R)$ is nonzero, and no additional condition for $p'(R)$ follows. Different initial values $\bar \rho (0)$ will, therefore, lead to different solutions with different neutron star masses. As is usually done in the TOV approach, we shall assume that solutions are stable as long as increasing $\bar \rho (0)$ lead to increasing neutron star masses. For practical reasons, we shall, nevertheless, plot both stable and unstable branches in most of the figures, because the numerical integration does not distinguish stable from unstable solutions. On the other hand, the baryon number is no longer a free parameter in the MF-EoS (TOV equation) case. Instead, the baryon number must be determined a posteriori from a given solution via
\begin{equation}
B = 4\pi \int_0^R dr r^2 \sqrt{\bf B} \bar n_{B}
\end{equation}
where the average baryon density $\bar n_B$ is defined in (\ref{av-b-dens}), and $P$ must be replaced by the TOV solution $p(r)$ in Eq. (\ref{av-b-dens}).
\subsection{Results of numerical calculations}
\begin{figure}
\includegraphics[height=11.5cm]{mass.pdf}
\caption{(Color online) Neutron star masses (in units of the solar mass) and radii $R$ (in kilometers) for different potentials, both for exact field theory solutions (BPS) and for MF-EoS TOV solutions (Mean EoS). Details are explained in the main text.}
\end{figure}
Numerical solutions are found by a shooting from the center $r=0$. In the MF-EoS TOV case, we impose ${\bf B}(0)=1$ and choose a $\bar \rho (0)= \bar \rho_0$ (which determines $p(0)=p_0$ via the MF-EoS). Then we integrate until we reach a radius value $r=R$ such that $p(R)=0$. The corresponding neutron star mass is then determined from
\begin{equation} \label{NS-mass}
M=4\pi \int_0^R dr r^2 \bar \rho (r).
\end{equation}
Here, different initial values $\bar \rho_0$ lead to different solutions with different radii and masses. Further, formal solutions exist for arbitrary values of $\bar \rho_0$. We identify stable solutions by the condition that the neutron star mass should grow with growing $\bar \rho_0$, which holds only up to a certain maximum value of $\bar \rho_0$.
In the full field theory case, we impose ${\bf B}(0)=1$, $h(0)=1$. Then, for each given value of the baryon number $B$, we vary $\rho_0$ until we find a solution which obeys the two conditions $p(R)=0$ and $p'(R)=0$ for some radius $R$. For a fixed value of $B$, we find at most one stable solution. Different neutron star solutions are found for different values of $B$. The neutron star mass is again calculated from Eq. (\ref{NS-mass}) (replacing $\bar \rho (r)$ by $\rho (r)$).
In this case, it turns out that solutions obeying the two conditions $p(R)=0$ and $p'(R)=0$ only exist up to a certain maximum value of $B$.
The relations between the resulting neutron star radii $R$ and masses $M$ are shown in Fig. 2. We find that all radii increase with increasing masses (except very close to the maximum masses), both for the exact field theory and for the TOV calculations. This is most likely related to the rather stiff character of the EoS for nuclear matter described by the BPS Skyrme model. For small neutron stars (i.e., for solutions with a small $B$ or $\bar \rho_0$, respectively), we find that the $M(R)$ curve is well approximated by the curve for the EoS $\bar \rho =\mbox{const}$ (by a curve $M\propto R^3$). For the MF-EoS cases, we also show the unstable branches. For these unstable branches of the TOV solutions, we find that in the limit of very large $\bar \rho_0$ they approach a curve $M\propto R$. This is related to the fact that, for all potentials, the MF-EoS approaches $\bar \rho = p$ in the limit of very large $\bar \rho$.
When comparing exact field theory solutions with MF-EoS TOV solutions for the same potential, we find that the results depend quite significantly on the potential we choose. For the theta function potential, exact field theory calculations and MF-EoS TOV calculations lead to identical results. For the potential ${\cal U}_\pi = 2h$, the two curves are quite similar, the main difference being that the MF-EoS neutron star solution reaches slightly higher neutron star masses. In the case of the ${\cal U}_\pi^2 = 4h^2 $ potential, the difference is more pronounced. The MF-EoS solution not only reaches higher masses, but is also significantly more compact, i.e., the "compactness parameter" $2GM/R$ is significantly bigger than in the full field theory calculation. This difference is probably related to the fact that the potential $4h^2$ is quite peaked about the antivacuum $h=1$ and approaches the vacuum $h=0$ quite fast, i.e., its shape strongly deviates from the constant theta function potential. We plot the compactness of solutions (both full BPS and MF-EoS) for different potentials in Fig. 3.
In contrast to global properties (like the $M(R)$ curves), which are not too different between full field theoretic and MF TOV calculations, the results for local quantities (like the energy densities $\rho (r)$ and $\bar \rho (r)$) are completely different. We plot the energy densities in Fig. 4, the pressures in Fig. 5, and the metric functions ${\bf B}(r)$ in Fig. 6.
\begin{figure}
\includegraphics[height=11.5cm]{v2_vs_m.pdf}
\caption{(Color online) Compactness of neutron stars as a function of the neutron star masses (in units of the solar mass) for different solutions. For the MF-EoS solutions we also show the unstable branches. It may be seen that in the region of stable solutions all compactness curves are quite similar, with the exception of the MF-EoS solutions for the potential ${\cal U} =4h^2$. }
\end{figure}
\begin{figure}
\includegraphics[height=11.5cm]{rhodens_vs_r.pdf}
\caption{(Color online) Energy densities (in MeV fm$^{-3}$) as functions of the radius in units of the neutron star radius $R$. Full BPS energy densities are peaked about the center $r=0$ (especially for heavier neutron stars) and approach zero at the neutron star surface, $\rho (R)=0$. Here, $n$ is the baryon number in solar units, $n=B/B_\odot$. MF-EoS energy densities, on the other hand, vary only slowly and approach a nonzero value (nuclear saturation density) at the neutron star surface. Here, $M$ is the neutron star mass in solar units.}
\end{figure}
\begin{figure}
\includegraphics[height=11.5cm]{pdens_vs_r.pdf}
\caption{(Color online) Pressures (in MeV fm$^{-3}$) as functions of the radius in units of the neutron star radius $R$. Full BPS pressures are peaked about the center $r=0$ (especially for heavier neutron stars). Here, $n$ is the baryon number in solar units, $n=B/B_\odot$. MF-EoS pressures, on the other hand, vary only slowly. Here, $M$ is the neutron star mass in solar units.}
\end{figure}
\begin{figure}
\includegraphics[height=11.5cm]{B_vs_r.pdf}
\caption{(Color online) Metric functions ${\bf B} (r)$ as functions of the radius in units of the neutron star radius $R$. Full BPS solutions ${\bf B}(r)$ have their maxima in the interior of the neutron star, especially for the potential $4h^2$ which has a very peaked energy density. Here, $n$ is the baryon number in solar units, $n=B/B_\odot$. MF-EoS solutions, on the other hand, take their maxima at the neutron star surface $r=R_S$. Here, $M$ is the neutron star mass in solar units.}
\end{figure}
In Fig. 7 and Fig. 8 we show the central values of true and average energy densities and pressures for different neutron star solutions. Again, the difference between true densities (full field theory calculations) and average densities (MF-EoS TOV calculations) is huge. In Fig. 9, we plot the central value of $p$ as a function of the central value of $\rho$.
In Fig. 10, we plot the equations of state (EoS) $p(\rho)$ for different neutron star solutions. All shown EoS are on-shell in the sense that they are reconstructed from the numerical solutions $\rho (r)$ and $p(r)$ by eliminating the independent variable $r$. In the MF-EoS case, the reconstructed on-shell EoS must, of course, coincide with the original MF-EoS. In particular, the on-shell EoS for different solutions corresponding to the same MF-EoS (the same potential) must coincide, and the degree to which they coincide demonstrates the precision of our numerical calculations.
Finally, in Fig. 11 we show the gravitational mass loss of neutron stars for the full BPS model calculations for the two potentials ${\cal U}_\pi$ and
${\cal U}_\pi^2$. More precisely, the figure also includes the (small) mass loss due to the binding energy of infinite nuclear matter. In this paper, we define the solar baryon number $B_\odot$ as solar mass divided by proton mass,
\begin{equation} \label{B-sol}
B_\odot = (M_\odot /m_{\rm p}) = ((1.988 \cdot 10^{30} \; {\rm kg}) / (1,673 \cdot 10^{-27} \; {\rm kg})) = 1.188 \cdot 10^{57}
\end{equation}
(which, strictly speaking, does not exactly coincide with the number of baryon charges in the sun). On the other hand, the neutron star mass approaches $M\sim B E_{B=1}$ in the small mass limit (where gravity can be neglected), where $E_{B=1} = 923,3 \; {\rm MeV}$ is the mass per baryon number of infinite nuclear matter. The ratio $((M/M_\odot)/(B/B_\odot))$ in Fig. 11, therefore, approaches the limiting value $(E_{B=1}/m_{\rm p}) = 0.984$ in the limit of small mass (small $B$).
\begin{figure}
\includegraphics[height=11.5cm]{rho_vs_m.pdf}
\caption{(Color online) Central values of the energy density (and of the pressure for the full BPS solutions) for different neutron star solutions, as a function of the neutron star mass. There is a big difference between full BPS solutions, which lead to much larger central values, and MF-EoS TOV solutions. Besides, the central values for the potential $4h^2$ are much larger than for the potential $2h$. In the MF-EoS case, we also show the unstable branches.}
\end{figure}
\begin{figure}
\includegraphics[height=11.5cm]{p_vs_r.pdf}
\caption{(Color online) Central values of the pressure for different neutron star solutions, as a function of the neutron star radius. There is a big difference between full BPS solutions, which lead to much larger central values, and MF-EoS TOV solutions.
Besides, the central values for the potential $4h^2$ are much larger than for the potential $2h$. In the MF-EoS case, we also show the unstable branches.}
\end{figure}
\begin{figure}
\includegraphics[height=11.5cm]{p_vs_rho.pdf}
\caption{(Color online) The central value of the pressure $p(0)$ as a function of the central value of the energy density $\rho (0)$, for different solutions. A given central value of the pressure requires much higher central values of the energy density for full BPS models solutions than for MF-EoS TOV solutions. Observe that, as different solutions for the same potential in the MF-EoS TOV case correspond to the same MF-EoS, the curves for the TOV calculations are, at the same time, the MF-EoS graphs for the corresponding potentials. This is not true for the full BPS model calculations, where different solutions lead to different on-shell EoS even for the same potential, see Fig. 10. }
\end{figure}
\begin{figure}
\includegraphics[height=11.5cm]{pdens_vs_rhodens.pdf}
\caption{(Color online) Equations of state (EoS) for different solutions $\rho (r)$ and $p(r)$. In the case of full BPS model calculations, it is clearly visible that different solutions lead to different on-shell EoS, even for the same potential. Further, the on-shell EoS are approximated with high precision by a homogeneous EoS (polytrope) $p\sim a\rho^b$ where the values of $a$ and $b$ vary even for different solutions for the same potential. The behavior of the MF-EoS, on the other hand, is completely different. In addition, the calculated on-shell EoS for different solutions for the same potential must coincide with each other (and with the original MF-EoS). And indeed, the two different MF-EoS TOV solutions for the potential $2h$ lead to exactly identical numerical EoS in the region of overlapping values of $\rho$ and $p$. Here, $n=(B/B_\odot)$, and $M$ is the neutron star mass in solar units. }
\end{figure}
\begin{figure}
\includegraphics[height=11.5cm]{mn_vs_r.pdf}
\caption{(Color online) Gravitational mass loss: we plot the ratio of total mass (with the gravitational mass loss included) to non gravitational (pure neutron matter) mass, $((M/M_\odot)/(B/B_\odot))$, as a function of the neutron star radius, for the full BPS model calculations for the two potentials $2h$ and $4h^2$. The maximum mass loss is about 25\%. Further details are explained in the main text.}
\end{figure}
\section{Discussion of results}
The present work had the two-fold objective to continue, on the one hand, the application of the BPS Skyrme model - a field theory realization of a fluid model of nuclear matter - to the physics of neutron stars, and to investigate, on the other hand, the differences between full field theory calculations and approximate MF calculations of neutron star properties. Concerning the differences between exact and MF calculations, we find that they depend both on the chosen potential and on whether one considers global properties (like the neutron star masses and radii) or local quantities (like energy densities or the metric). There exists one limiting case - the potential $\Theta (h)$ - where MF and exact calculations coincide. For the other two potentials considered, we find that the potential $2h$ leads to rather similar $M(R)$ curves for both exact and MF calculations, where the MF curve reaches, however, slightly larger values for the neutron star mass $M$ (up to about $3.8 \, M_\odot $) than the exact field theory curve (about $3.3 \, M_\odot$), see Fig. 2. For the potential $4h^2$, the difference between the exact and the MF case is much bigger. The $M(R)$ curve for the exact calculation is, indeed, quite similar to the curves for $2h$, but terminates for a smaller mass (about $2.2 \, M_\odot $; the maximum mass values in \cite{star} are slightly bigger because of the different fit values used there). The curve for the MF calculation for the potential $4h^2$, on the other hand, leads to much more compact neutron stars (smaller radii for the same masses), see Figs. 2 and 3. From our results it is, in fact, easy to understand why the MF result for the potential $4h^2$ is quite different from the MF results for the other two potentials. The MF-EoS for all three potentials rather quickly approach the stiff "hadronic bag type" or maximally compact EoS
\begin{equation}
\bar \rho = P + B_\infty
\end{equation}
where, however, the numerical values for the "asymptotic bag constant" $B_\infty$ are different for the three potentials. Indeed, using the numerical fit values (\ref{param-pot-theta}), (\ref{param-pot-2}) and (\ref{param-pot-4}) we find
\begin{eqnarray}
\Theta (h): && B_\infty = 2\mu^2 \sim 141 \; {\rm MeV}\, {\rm fm}^{-3} \nonumber \\ 2h: && B_\infty = 2\mu^2 \sim 176 \; {\rm MeV} \, {\rm fm}^{-3} , \nonumber \\
4h^2 : && B_\infty = \frac{5}{2} \mu^2 \sim 353 \; {\rm MeV} \, {\rm fm}^{-3}.
\end{eqnarray}
So $B_\infty$ is much bigger for $4h^2$, which means that for a given value of the pressure $P$ the energy density is much bigger, and the corresponding nuclear matter more compressed, which leads to more compact neutron stars. In the full field theory calculation, it is still true that the matter is more compressed in the center for $4h^2$ but, on the other hand, the potential $4h^2$ approaches the vacuum value $h=0$ faster
(i.e. takes smaller values near $h=0$)
than the other potentials. This implies that the neutron star (or the soliton in the case without gravity) has a rather large "tail" of low density, which tends to increase its radius. This shows that the potential $4h^2$ has two properties which have opposite effects on the resulting neutron star radii. On the one hand, this potential is quite peaked about the anti-vacuum $h=1$, which leads to high densities close to the center. On the other hand, the potential $4h^2$ approaches the vacuum $h=0$ fast, which leads to a rather large tail of low density. It would be interesting to consider potentials where these two effects ("peakedness" near $h=1$, and approach to the vacuum near $h=0$) can be varied independently, and to study their influence on neutron star properties. In general, we can say that the differences in global properties (masses, radii) of neutron stars between exact and MF calculations depend on the potential, the most significant difference being that the exact calculations tend to give smaller maximum neutron star masses than the MF calculations.
For local quantities (like the energy densities, pressures, or metric coefficients as functions of the radial coordinate $r$), the differences between exact and MF calculations are much more pronounced. The true energy densities $\rho (r)$, for instance, take much larger values in the center and approach zero at the neutron star radius $R$, whereas the averaged, MF densities $\bar \rho (r)$ do not vary too much with $r$ (just by a factor of about 2 or 3) and take non-zero values (corresponding to the nuclear saturation density $\rho_s$ of nuclear matter) at the neutron star surface $r=R$. This does {\em not} mean that gravity causes a huge compression for the true energy densities. Instead, the energy densities have a rather similar profile already in the case without gravity (the BPS Skyrme solitons), and, at least for the potentials we considered, the compression induced by gravity is always smaller than by a factor of three in the center, even for the maximum mass cases, see \cite{star}. In other words, the compression at the center in the MF case is not too different from the exact case, although the absolute values of the densities are very different. For the metric functions ${\bf B}(r)$, the most interesting difference is that they take their maximum values inside the neutron star in the exact calculation, but on the surface in the MF calculation, see Fig. 6. These rather big differences for local quantities between exact and MF calculations lead to the plausible conjecture that certain global physical observables with a stronger dependence on the shapes of local quantities will be quite different, too, the most obvious candidate being the moment of inertia relevant for the description of (slow) rotations of neutron stars. Indeed, in the Newtonian case, the moment of inertia is just given by a volume integral of tensorial expressions of the type $x^i x^j \rho (\vec x)$, and the dependence both on the total energy and on the shape of $\rho$ is obvious. In the general relativistic case, however, the moment of inertia is {\em not} just the volume integral of certain moments of the energy density. Instead, its calculation requires to solve the Einstein equations for a more general metric depending on three independent metric functions \cite{mom-inert}, which is beyond the scope of the present article. Still, we expect the Newtonian arguments to be qualitatively valid and, therefore, rather pronounced differences between exact and MF moments of inertia. This issue shall be investigated in detail in a forthcoming publication.
In general, the differences between exact and MF calculations in the BPS Skyrme model will be the stronger the more the potential deviates from the step function potential which provides flat energy and particle densities. Beyond the BPS Skyrme model, this difference should become more important for theories which lead to appreciable inhomogeneities in the energy and particle distributions. In the standard Skyrme model, for instance, one may expect significant variations due to the considerable energy inhomogeneities of the skyrmion crystal \cite{piette3}, \cite{rho}.
Concerning the relevance and usefulness of the BPS Skyrme model (and its near-BPS generalizations) for the description of nuclear matter and neutron stars, we believe that in the present paper we simply added some further strong arguments to an already quite impressive body of evidence.
Let us briefly repeat some key properties of the model which support this claim. The model
\begin{enumerate}
\item has the energy-momentum tensor of a perfect fluid, so describes a perfect fluid state of nuclear matter.
\item has the SDiff symmetries on physical space among its symmetries, so locally volume-preserving deformations do not cost any energy.
\item has (infinitely many) static solutions (BPS solutions) saturating an energy bound linear in the baryon number (a BPS bound). This allows to easily accommodate the small binding energies of physical nuclei and nuclear matter.
\item incorporates the property of nuclear saturation in the sense that, for arbitrary baryon number $B$, there exist solutions with finite energy and volume (both proportional to $B$) and with zero pressure (the BPS solutions), that is, describing nuclear matter at equilibrium.
\item
allows for a derivation of its macroscopic, thermodynamical properties directly from the exact, microscopic description (BPS solutions), without the necessity of a thermodynamical limit.
\end{enumerate}
In the present paper we find, among other new results, that the EoS for a MF averaged energy density $\bar \rho$, for bigger values of $\bar \rho$ and pressure $P$, rather quickly approaches the asymptotic form $\bar\rho = P + B_\infty$. This EoS is called the "maximally compact" EoS in the neutron star literature and, indeed, leads to $M(R)$ curves which are rather similar to the ones we find, although slightly more compact \cite{latt} (the case of the step function potential is, of course, exactly equivalent to the case of the maximally compact EoS, with $B_\infty = B_0 = \rho_s$). For small $P$, on the other hand, our MF-EoS are much softer. Indeed, for potentials with an approach to the vacuum like $\lim_{\xi \to 0} {\cal U} \sim \xi^\alpha$ such that $\alpha \ge 2$ - to which both the potential $2h$ ($\alpha =2$) and the potential $4h^2$ ($\alpha =4$) belong - it follows easily from the results of \cite{term} that the speed of sound at nuclear saturation is zero, $v^2_{P=0} = \lim_{P\to 0} (\partial \bar \rho /\partial P)^{-1} =0$. The general picture we find is that of a maximally stiff (maximally compact) nuclear matter in the core of the neutron star and, more generally, for densities which are sufficiently above nuclear saturation density $\rho_s$. Near nuclear saturation and, therefore, near the neutron star surface, on the other hand, the nuclear matter gets much softer, where the details of this transition between "core" and "mantle" depend on the chosen potential. Very close to the neutron star surface ("neutron star crust") we believe, in any case, that the BPS Skyrme model is not sufficient, and more terms of the near-BPS model (\ref{near-BPS}) should be included for a reliable description \cite{star}.
Some bulk properties of neutron stars (like their maximum masses, or $M(R)$ curves), however, probably do not depend too much on the crust properties, and for these the BPS Skyrme model makes some rather robust predictions, like $M(R)$ curves which are quite similar to the $M(R)$ curves for nuclear matter with the "maximally compact" EoS where, e.g., $(dM/dR)>0$ for almost all neutron stars (except, probably, very close to the maximum mass). This differs from the $M(R)$ curves which result from a large class of nuclear physics EoS (see, e.g., \cite{latt} - \cite{piek}), but is completely compatible with the (still not very precise and not very abundant) observational data.
More precisely, modifications of the equation of state at and below nuclear saturation (e.g., via a generalization to the near-BPS Skyrme model, or by ad-hoc modifying the EoS at low densities using results from nuclear physics) will have almost no influence on $M$ or $R$ for sufficiently heavy neutron stars. The reason is that these regions of small density are, at the same time, regions of very small pressure, and regions of very small pressure may occupy only the very thin outermost shell of the star, due to the strong gravitational pull of a heavy neutron star. For light neutron stars, on the other hand, the influence of low-density regions on the mass will still be rather small, but their influence on the radius may be more appreciable, leading to larger radii. As of today, however, there is no firm observational evidence for the existence of neutron stars with masses significantly below one solar mass, so the discussion about the "correct" $M(R)$ curve for light neutron stars might well be purely academic.
Another issue where our results may be of some relevance is the so-called TOV inversion. Indeed, for a given EoS for barotropic nuclear matter,
$p=p(\rho)$, and for a given initial value $\rho (r=0)=\rho_c$ within a certain allowed range, a unique stable neutron star solution with its mass $M$ and radius $R$ follows from the TOV equations. Therefore, by varying $\rho_c$ over its allowed range, the whole $M(R)$ curve may be constructed for a given EoS. The TOV inversion now consists in the inverse operation, i.e., in the reconstruction of a barotropic EoS from a given curve $M(R)$. The formal reconstruction method was developed in \cite{lind}, and combined with statistical methods for an approximate reconstruction of $p(\rho)$ from a finite number of observational data, e.g., in \cite{stein}. Obviously, the TOV inversion hinges on the assumption of barotropic nuclear matter with a barotropic EoS $p=p(\rho)$. The BPS Skyrme model (and certainly also its near-BPS extensions), on the other hand, represents a well-motivated model of non-barotropic nuclear matter. A barotropic EoS may still be found by a mean-field limit (which is well-defined and straight-forward in this model), but the TOV inversion will, nevertheless, be problematic if the $M(R)$ curves of the exact theory differ from the ones for the MF theory. In this case, the EoS reconstructed from an observed $M(R)$ curve (which should correspond to the full field theory) via the TOV inversion will, in general, be different from the EoS resulting from the MF limit. We remark that this is not a specific problem of our model but will be present whenever nuclear matter is described by a field theory which leads to a non-barotropic fluid beyond mean field theory.
In our model, there exists a formal limit (the limit of very large density) where the difference between exact and MF EoS disappears. Indeed, the exact model has the non-barotropic EoS (off-shell) $p=\rho -2\mu^2 {\cal U}$ where (on-shell) ${\cal U}={\cal U}(r)$ depends on $r$, whereas the MF model EoS soon approaches $p = \bar \rho -B_\infty$, $B_\infty = $ const. Obviously, the difference between the two becomes immaterial in the limit of large $\rho$, or $(\rho -p)/\rho <<1$. This limit is, however, never attained for stable neutron star solutions, for which it is always true that $(\rho_c -p_c)/\rho_c >0.5$ even for the central values.
\section{Comparison with other results}
In our investigation, we presented an ample variety of both qualitative and quantitative results about neutron stars described by the BPS Skyrme model.
Still, precise quantitative predictions of neutron star properties within our model should be considered preliminary at the present moment. Firstly, as said, we think that a reliable description of nuclear matter requires the inclusion of further terms of the full near-BPS Skyrme model, although the contribution of these further terms to bulk quantities (like neutron star masses or radii) is probably small. Secondly, we fitted the two parameters of our model to the mass per baryon number of infinite nuclear matter, $E_{B=1} = 923.3 \; {\rm MeV}$, and to the nuclear saturation density $n_0 = 0.153 \; {\rm fm}^{-3}$. Here, the first value $E_{B=1} = 923.3 \; {\rm MeV}$ is quite precise, and almost all nuclear models lead to values $923 \le E_{B=1}/{\rm MeV} \le 926$ (see, e.g., Table 1 in \cite{weber}). The nuclear saturation density, on the other hand, is quite model dependent, where most models lead to a range of values $0.145 \le n_0 \cdot {\rm fm}^3 \le 0.175$ (see, again, Table 1 in \cite{weber}). Thirdly, at the moment, we do not know the "true" potential and so have the freedom to choose different potentials.
It is, nevertheless, interesting to compare our results with results of neutron star studies within other approaches. In particular, there exist some rather generic results which are either independent of a specific EoS or depend on it only weakly, which makes a comparison all the more relevant.
\subsection{Comparison with generic results}
\subsubsection{The Rhoades-Ruffini bound}
The Rhoades-Ruffini bound is a bound for the maximal neutron star mass originally derived by Rhoades and Ruffini in \cite{RR} which is based on the following observation (a transparent discussion may be found, e.g., in \cite{glend}). Let us restrict to "reasonable" EoS, i.e., EoS satisfying the constraints $0\le dp/d\rho \le 1$ (the pressure increases with the density, and the speed of sound is bound by the speed of light), and let us assume that one reference point $(\rho_f ,p_f)$ of the EoS is known. The result of Rhoades and Ruffini then says that among all EoS the contribution to the neutron star mass for $\rho \ge \rho_f$ is maximal for the maximally compact EoS
\begin{eqnarray}
\rho &=& \frac{1}{2}\left[ \rho_f - p_f + (\rho_f + p_f)\left( \frac{n}{n_f} \right)^2 \right] , \nonumber \\
p &=& \rho - (\rho_f - p_f) \, , \hspace*{1cm} n \ge n_f
\end{eqnarray}
(here $n$ is the baryon density and $n_f$ its value at the reference point $(\rho_f ,p_f)$). In a next step one has to choose values $\rho_f$ and $p_f$ for the reference point and an EoS for $\rho <\rho_f$ which smoothly joins the reference point, where both the values for $\rho_f$ and $p_f$ and the low-density EoS should follow from the properties of nuclear matter. Then a numerical value for the maximal mass may be calculated by numerically integrating the TOV equations. In the original paper \cite{RR}, $\rho_f = 4.6 \cdot 10^{17}\, \mbox{kg}\, \mbox{m}^{-3}$ together with the EoS for free degenerate neutrons for $\rho < \rho_f$ was chosen, leading to the bound $M_{\rm max}\le 3.2 M_\odot$. In \cite{glend} a bound is reported for the same value of $\rho_f$ but for an improved EoS for $\rho < \rho_f$, leading to $M_{\rm max}\le 3.14 M_\odot$. In \cite{glend} also the extreme case $\rho_f = \rho_s = 2,51 \cdot 10^{17} \, \mbox{kg}\, \mbox{m}^{-3}$ is considered, leading to the bound $M_{\rm max}\le 4.3 M_\odot$.
Here, $\rho_s$ is the nuclear saturation density
\begin{equation}
\rho_s = 923\, {\rm MeV} \, \cdot \, 0.153 \, {\rm fm}^{-3} = 141\, {\rm MeV} \, {\rm fm}^{-3} =
2,51 \cdot 10^{17}\, \mbox{kg}\, \mbox{m}^{-3} .
\end{equation}
If, instead, the maximally compact EoS is used all the way down to $p=0$, i.e., $\rho_f = \rho_s$, $p_f =0$ (so that there is no matching to a soft EoS for $\rho < \rho_f$), then the mass bound is $M_{\rm max} \le 4.09 M_\odot$, see \cite{latt}. But this case is exactly equivalent to the BPS Skyrme model for the step function potential, so it is a gratifying consistency check that the maximal masses in the two cases precisely agree, $M_{\rm max} = 4.1 M_\odot$ for ${\cal U} =\Theta (h)$. We display our maximum masses in table 1. We find that the maximum masses are slightly above the Rhoades-Ruffini bound of $3.2 M_\odot$ for the potential $2h$, but below the bound for $4h^2$.
\begin{table} \label{tab1}
\begin{center}
\begin{tabular}{|c|c|c|c|c|c|}
\hline
Potential & \; $\Theta (h) $\; & $2h$, exact & $2h$, MF & $4h^2$, exact & $4h^2$, MF \\
\hline
$M_{\rm max}/M_\odot$ & 4.1 & 3.34 & 3.79 & 2.15 &2.82 \\
\hline
\end{tabular}
\caption{Maximal neutron star masses for different potentials, both for the exact and the MF (mean field) calculations.}
\end{center} \label{table1}
\end{table}
\subsubsection{The compactness limit}
The compactness of a neutron star with mass $M$ and radius $R$ is defined as
\begin{equation}
\beta = \frac{2GM}{R}
\end{equation}
(remember that we use units where the speed of light $c=1$). Obviously, the radius of a neutron star must always be bigger than its Schwarzschild radius $R_S =2GM$, which implies the bound $\beta <1$. Just using relativity and the TOV equations, a tighter bound $\beta <(8/9)$ may be proved \cite{buch}. With additional assumptions, even tighter bounds may be proved. Using the same assumptions on the EoS as Rhoades and Ruffini in their mass bound ($0\le dp/d\rho \le 1$, and a smooth matching to calculable nuclear physics EoS near nuclear saturation), Glendenning derived the improved bound \cite{glend1}
\begin{equation} \label{glend-bound}
\beta \le \frac{1}{1.47} = 0.68 \equiv \beta_{\rm G}.
\end{equation}
We plot the compacntess values for the maximum mass solutions of our model in table 2.
\begin{table}
\begin{center}
\begin{tabular}{|c|c|c|c|c|c|}
\hline
Potential &\; $\Theta (h) $\; & $2h$, exact & $2h$, MF & $4h^2$, exact & $4h^2$, MF \\
\hline
$\beta$ for $M_{\rm max}$ & 0.7 & 0.58 & 0.68 & 0.42 &0.66 \\
\hline
\end{tabular}
\caption{Compactenss of neutron stars of maximal mass for different potentials, both for the exact and the MF (mean field) calculations.}
\end{center}
\end{table}
We find that the step function potential does not satisfy the bound. This is not surprising, because the step function potential leads to the maximally compact EoS for all densities and so cannot be matched to a softer EoS near nuclear saturation. All other potentials are compatible with the Glendenning bound for the compactness parameter. It can be seen that the exact solutions lead to significantly lower values for the compactness parameter than the MF solutions.
\subsubsection{The energy density limit}
In the TOV approach for a given EoS, a particular solution (together with its resulting neutron star mass $M$ and radius $R$) is determined by the energy density at the center $\rho_c \equiv \rho (r=0)$. For dimensional reasons, the three parameters are related by $\rho_c = \tilde\gamma MR^{-3}$ where, however, even for the same EoS different solutions may lead to different values for the dimensionless constant $\tilde\gamma$. We want to focus, however, on the maximal mass solution, where for a given EoS there exists just one physically acceptable solution leading to one unique relation
\begin{equation}
\rho_c = \gamma M R^{-3} .
\end{equation}
Here, the important point is that, for a given EoS, the maximal mass solution provides the maximum value for $\rho_c$ (remember that formal solutions for even higher $\rho_c$ and lower masses are unstable and, therefore, do not correspond to neutron stars). If this expression is now combined with the compactness limit of the previous section (which may be rewritten as a lower bound for the radius, $R\ge (2GM/\beta_{\rm G})$), an upper bound for the central density $\rho_c$ in terms or the maximum mass $M$ follows,
\begin{equation}
\rho_c \le \left( \frac{\beta_{\rm G}}{2G}\right)^3 \frac{\gamma}{M^2} .
\end{equation}
For a specific upper bound (a specific value for $\gamma$), one would still have to find the maximum mass solution for a particular EoS. Lattimer and Prakash \cite{latt1} considered, instead, the possibility to derive an universal, EoS-independent bound by using the exact solution of the TOV system known as Tolman VII solution \cite{Tol}. The Tolman VII solution has the energy density
\begin{equation}
\rho = \rho_c \left( 1-\frac{r^2}{R^2}\right)
\end{equation}
and has the following interesting properties. i) From the above expression, exact solutions may be determined for the metric functions and the pressure, using the TOV equations. ii) Using the scaling symmetry of the TOV equations, $r\to \lambda r$, $m \to \lambda m$, $\rho \to \lambda^{-2} \rho$, $p\to \lambda^{-2}p$, Tolman VII solutions with different values for the parameters $\rho_c$, $R$ and $M$ may be produced from a given solution. Here, $m=m(r)$ is the mass function related to the metric function via ${\bf B}=(1-(2Gm/r))^{-1}$ such that the first TOV equation simplifies to $m' = 4\pi r^2 \rho$. Further, $M=m(R)$. iii) As a consequence, there exist Tolman VII solutions with arbitrary values for the mass $M$ and radius $R$. This does not contradict the mass bounds derived above, because different Tolman VII solutions do not correspond to the same EoS. iv) There exists, however, a bound $\beta_{\rm TVII}$ on the compactness parameter $\beta$ such that for
$\beta \ge \beta_{\rm TVII}$ physically sensible Tolman VII solutions no longer exist. Concretely, $\beta_{\rm TVII}= 0.771 >\beta_{\rm G}$ ($\beta_{\rm G}$ is defined in (\ref{glend-bound})) such that Tolman VII solutions are still valid for $\beta = \beta_{\rm G}$. Consequently, they allow for a rather high compactness and so support high values of $\rho_c$. It is, therefore, quite natural to use the Tolman VII solution for $\beta = \beta_{\rm G}$ as a phenomenological upper limit for $\rho_c$. One easily calculates
\begin{equation}
\rho_c = \frac{15}{8\pi} \frac{M}{R^3} \quad \Rightarrow \quad \gamma = \frac{15}{8\pi}
\end{equation}
such that the energy density bound becomes \cite{latt1}
\begin{equation} \label{rho-c-bound}
\rho_c \le \left( \frac{\beta_{\rm G}}{2G}\right)^3 \frac{15}{8\pi } \frac{1}{M^2} \equiv \rho_{c,{\rm TVII}} \simeq 1.45 \cdot 10^{19} \left( \frac{M_\odot}{M} \right)^2 {\rm kg}\,
{\rm m}^{-3} .
\end{equation}
Lattimer and Prakash checked this bound for a large number of neutron star EoS and found that all values for $M_{\rm max}$ together with their corresponding $\rho_c$ values saturate the bound, and that some values are, in fact, quite close to the bound, i.e., (\ref{rho-c-bound}) is a rather tight bound.
In table 3 we show the values of $\rho_c$ for the maximum masses for our model for different potentials, together with the values of the bound (\ref{rho-c-bound}).
\begin{table}
\begin{center}
\begin{tabular}{|c|c|c|c|c|c|}
\hline
Potential & \; $\Theta (h) \; $ & $2h$, exact & $2h$, MF & $4h^2$, exact & $4h^2$, MF \\
\hline
$\rho_c/(10^{18} \, {\rm kg}\, {\rm m}^{-3})$ & 0.753 & 1.45 & 0.774 & 4.01 &1.34 \\
\hline
$ \rho_{c,{\rm TVII}}/(10^{18} \, {\rm kg}\, {\rm m}^{-3})$ & 0.863 & 1.30 & 1.01 & 3.14 &1.82 \\
\hline
\end{tabular}
\caption{Central energy densities of neutron stars of maximal mass, together with the Lattimer-Prakash bounds \ref{rho-c-bound}. }
\end{center}
\end{table}
We find that, both for the $2h$ and for the $4h^2$ potential, in the case of the exact calculation the central density is somewhat above the Lattimer-Prakash bound. We believe that these potentials behave very reasonable near the vacuum $h=(1/2)\sin^2 \xi =0$ (quadratic or quartic approach to the vacuum, respectively), e.g. from the point of view of pion physics. On the other hand, these potentials are quite spiky (non-flat) also near the anti-vacuum $h=1$, which leads to rather non-flat energy and baryon charge densities (and, consequently, to rather high central densities) already in the case without gravity.
The violation of the bound might, therefore, indicate that more realistic potentials should be flatter near the anti-vacuum, while maintaining the same approaches to the vacuum. In any case, we want to emphasize that the Lattimer-Prakash bound is of a phenomenological rather than absolute nature.
\subsection{Skyrme related models}
There have been previous attempts to couple the original Skyrme model to gravity and to use the resulting system to model neutron stars \cite{kli}-\cite{gravSk}. In \cite{kli}, \cite{bizon} spherically symmetric skyrmion configurations (hedgehogs) were coupled to gravity, but it was found that - as in the case without gravity - stable solutions do not exist. In \cite{piette1}, \cite{piette2} rational map ansaetze were considered which are known to provide reasonable approximations for not too high $B$ in the case without gravity. The true minimum energy skyrmions at large $B$, however, are known to be skyrmion crystals, and these crystals were used in the analysis of \cite{piette3}, \cite{nelmes}. More precisely, the EoS of the Skyrme crystal was calculated and the resulting energy-momentum tensor then coupled to gravity. As the Skyrme crystal is not a perfect fluid, there may, in fact, exist two pressures (the radial pressure $p_r (r)$ and the tangential pressure $p_t (r)$) even for the spherically symmetric ansatz.
Numerically, it is found that for sufficiently heavy neutron stars (above about $1.5 M_\odot$), self-gravitation, indeed, leads to an anisotropic deformation of the Skyrme crystal (e.g. to $p_r (r) \ne p_t (r)$). The maximum mass found in \cite{piette3} is about $M_{\rm max} \sim 1.9 M_\odot$. Further, the EoS is very stiff. The radial speed of sound, e.g., is about $v_r \sim 0.57 c$ already in the Skyrme crystal without gravity, and further grows (but always respecting $v_r \le c$) when self-gravitation is taken into account. A more detailed comparison of the results of \cite{piette3} with our BPS Skyrme neutron star results may be found in \cite{star}.
A different type of skyrmion fluid was developed by K\"albermann in \cite{kalb} and applied to the description of neutron stars in \cite{ouy1}, \cite{ouy2}. In \cite{kalb}, the Skyrme field was coupled to the dilaton (playing the role of the scalar $\sigma$ meson) and to the (vector) $\omega$ meson. Then a dilute fluid approximation was assumed, where only Skyrme fields which are superpositions of non-overlapping $B=1$ skyrmions (nucleons) are considered. The dilute fluid approximation was partly motivated by the non-existence of spherically symmetric higher $B$ skyrmions, and this motivation is to a certain point superseded by the existence of non-spherically symmetric skyrmions (concretely, the skyrmion crystal in the limit of large $B$). The dilute fluid approximation, however, implies that individual $B=1$ skyrmions only interact via mesons, which allows for a mean-field treatment very similar in spirit to the Walecka model, the main difference being that the nucleons are described by (extended) skyrmions instead of the (point-like) fermions of the Walecka model. In particular, a partition function and a related (barotropic) EoS may be derived in the MF limit. The MF theory is then fitted to the properties of nuclear matter near saturation by choosing the right parameter values for the dilaton potential. In \cite{ouy1}, the skyrmion fluid of K\"albermann in the MF limit was coupled to gravity, and the resulting solutions of the TOV equations were compared to neutron star properties. In addition, in \cite{ouy2}, also the isovector-vector $\rho$ meson was coupled to the Skyrme field, allowing for a different in-medium treatment of protons and neutrons (the in-vacuum properties are still the same for both in this model, because the same $B=1$ hedgehog skyrmion solution is used for both of them). The resulting neutron star $M(R)$ curves are, in fact, quite similar to the ones we find in our model, and the maximum masses are about $3M_\odot$, both for the case with and without the inclusion of the $\rho$ meson, see, e.g. Fig. 2 of Ref. \cite{ouy2}.
Finally, we want to emphasize once more that the $\omega$ meson is, in fact, implicitly present in the BPS Skyrme model. On the one hand, the sextic term
${\cal L}_6 = -\pi^4 \lambda^2 {\cal B}_\mu{\cal B}^\mu$ is induced by the coupling of the Skyrme model to the $\omega$ meson in the limit of large meson mass. On the other hand, the MF-EoS of the BPS Skyrme model exactly coincides with the EoS of the Walecka model in the large-density limit, where it is precisely the (repulsive force of the) $\omega$ meson which determines the EoS in that limit. This even allows to match the coupling constants of the $\omega$ meson with those of the BPS Skyrme model, as discussed in Section III.A.
\subsection{The double pulsar J0737-3093}
\begin{figure}
\includegraphics[height=11.5cm]{constraint.pdf}
\caption{(Color online) Comparison of the mass vs. baryon number curves (both in solar units) of the BPS model (exact calculations for $2h$ and $4h^2$, MF-EoS TOV calculation for $\Theta (h)$) with the values determined for the lighter neutron star of the double pulsar J0737-3039 (small yellow rectangle). Details are explained in the main text.}
\end{figure}
Even considering precise predictions preliminary,
it may still be of some interest to consider one particular case where some properties of a specific neutron star are known quite precisely, and to check what our model predicts for this case. Concretely, we refer to the double pulsar J0737-3039 \cite{pods}. The mass of the lighter pulsar ${\rm P}_2$ in this system is determined to be $M_{{\rm P}_2} = 1.249 \pm 0.001 \; M_\odot$ which makes it the lightest firmly established neutron star. If, in addition, it is assumed that this pulsar was formed from an ONeMg white dwarf via an electron-capture supernova - which is plausible given its small mass - then stellar structure calculations allow to determine its baryon number with a rather good precision. Concretely, it is found that the corresponding "baryonic mass" $\bar M_{{\rm P}_2}$ is in the interval $1.366 \le (\bar M_{{\rm P}_2}/M_\odot) \le 1.375$ \cite{pods}. In our model, the baryon number is a much more natural observable than the baryonic mass. Further, in the present paper we defined the solar baryon number $B_\odot$ as solar mass divided by proton mass (\ref{B-sol}). In \cite{pods}, on the other hand, the baryonic mass $\bar M_{{\rm P}_2}$ was determined from the baryon number by assuming a mass per baryon number equal to the atomic mass unit $u=931.5 \; {\rm MeV}$ (which is justified by the most abundant elements in a ONeMg white dwarf). This implies that in the transition from the interval for $ (\bar M_{{\rm P}_2}/M_\odot) $ to the interval for $(B_{{\rm P}_2}/B_\odot) $ we have to multiply by a factor of $(m_{\rm p}/u) = ((938.3 \; {\rm MeV})/(931.5 \; {\rm MeV})) = 1.0073$ leading to the baryon number interval
$1.376 \le (B_{{\rm P}_2}/B_\odot) \le 1.385$. The resulting small intervals for $M_{{\rm P}_2}$ and $B_{{\rm P}_2}$ may now be compared with the predictions of different models of nuclear matter (i.e., different EoS), see, e.g., \cite{blaschke}. We make this comparison with the $M(B)$ curves of our BPS model in Fig. 12. We find a reasonably good agreement, where especially the potential $\Theta (h)$ gets quite close, which probably means that quite flat potentials are preferred (but remember the caveats at the beginning of this section).
\section{Conclusions and Outlook}
To summarize, we found further evidence for the relevance of the (near-)BPS Skyrme model as a physical model for nuclear matter. The model incorporates in a completely natural fashion many nontrivial qualitative properties of nuclear matter, and, due to its simplicity (at least for the BPS submodel), allows for explicit and at least partly analytical calculations of many properties of nuclear matter, in general, and neutron stars, in particular. Specifically, it makes clear predictions for some neutron star properties, whereas other, more detailed predictions depend on the potential of the model and would, therefore, require a determination of this potential, e.g., from fits to nuclear data. If more precise data on neutron stars will be available in the future, it might even be conceivable to reconstruct the potential (as well as additional terms which are required for a further refinement and completion of the model) from fits to neutron stars, and to use it for further predictions of nuclear properties. We already found some evidence that, while a quadratic or quartic vacuum approach of the potential is expected on theoretical grounds, it should probably be flatter for larger field values (away from the vacuum), leading to flatter energy and baryon density profiles. This issue certainly deserves further investigation.
Further, we found strong indications that MF calculations and exact field theoretic calculations of neutron star properties can lead to different results. If these differences are appreciable, this casts some doubts on the general validity of the TOV inversion, i.e., on the reconstruction of a barotropic EoS $p=p(\rho)$ from an observed $M(R)$ curve. More generally, this implies that if a well-motivated model of nuclear matter does not lead to a satisfactory description of neutron star properties, it might not be the fault of the model but, rather, just indicate the inadequacy of the MF approximation, and the necessity to go beyond MF theory for a reliable description of neutron stars.
On the other hand, the study of rotating neutron stars within the BPS Skyrme model is a rather obvious next step. We believe that the difference between exact and MF calculations will be even more significant in that case. Another important issue is a different treatment of neutrons and protons, which are treated identically in our model, because they correspond to the same classical skyrmion solution. Here, a first possibility is a different in-medium treatment via the introduction of an isospin chemical potential or via a coupling to isovector mesons (like the $\rho$ meson in the dilute skyrmion fluid in \cite{ouy2}). The introduction of a difference between neutrons and protons in Skyrme models already in-vacuum requires the collective coordinate quantization of the isospin collective coordinates. A different treatment of neutrons and protons is certainly required for a realistic description of neutron stars, e.g., to maintain beta equilibrium, or for a consistent coupling to other (nuclear physics) EoS at lower densities. The generalizations of our model necessary to achieve this aim are, however, demanding, since some of the integrability and solvability properties simplifying the analysis of the model will be lost. These generalizations are, therefore, beyond the scope of the present article and will be investigated in forthcoming publications.
\section*{Acknowledgement}
The authors acknowledge financial support from the Ministry of Education, Culture, and Sports, Spain (Grant No. FPA2011-22776), the Xunta de Galicia (Grant No. INCITE09.296.035PR and Conselleria de Educacion), the Spanish Consolider-Ingenio 2010 Programme CPAN (CSD2007-00042), and FEDER.
CN thanks the Spanish Ministery of
Education, Culture and Sports for financial support (grant FPU AP2010-5772).
AW thanks D. Blaschke and M. Kutschera for comments. AW acknowledge also T. Kl\"ahn for inspiring discussions.
|
\section{Introduction}
Graphene is a one-atom thick crystal membrane with extraordinary
mechanical and electronic properties
\cite{sci04nov,PNA05nov,RMP09cas}. Experimental characterization of
suspended graphene shows that it is covered with ripples. These
ripples are several nanometers long waves of the sheet without a
preferred direction \cite{nature07mey,pss09ban}, modify the electronic
band structure \cite{PRB08gui}, and are expected to be relevant in the
understanding of electronic transport in graphene
\cite{PTRS08kat}. Also, more recently, buckling in which the unit
structures consist of only two-three unit cells of the graphene
honeycomb lattice has been experimentally observed \cite{ASC11mao}.
There have been many studies of ripples. The earliest studies using
Monte Carlo \cite{nat07fas} or molecular dynamics simulations
\cite{PRB07abe} have shown that ripples may be connected to variable
length $\sigma$ bonds of carbon atoms and may be caused by thermal
fluctuations. Other studies have explored the connection between
rippling and electron-phonon coupling \cite{EPL08Eun,PRB09gaz}. In
particular, it has been suggested that, at zero temperature, the
electron-phonon coupling may drive the graphene sheet into a quantum
critical point characterized by the vanishing of the bending rigidity
of the membrane \cite{PRL11san}. The continuation of this work by
J.~Gonzalez \cite{PRB14gon} discusses the nonzero expectation value of
the mean curvature (the Laplacian of the flexural phonon field) once
the bending rigidity of the membrane vanishes, and its role as order
parameter. Alternatively, assuming that the graphene membrane is
fluctuating in $2+d$ dimensions (with $d\gg 1$), Guinea \textit{et
al.\/} have calculated the dressed two-particle propagators of the
elastic and electron interactions. They have found a collective mode
which becomes unstable at a nonzero wave vector and causes the
appearance of Gaussian curvature \cite{PRB14gui}. Amorim et al.\/
\cite{PRB14amo} estimate the crossover temperature between quantum and
classical descriptions to be 70-90 K. Thus a quantum description of
ripples is not necessary at room temperature. All these studies
investigate and characterize rippling as equilibrium phenomena.
We are interested in rippling dynamics and stability of static
corrugations under disturbances. In experiments to visualize ripples
using a transmission electron microscope (TEM), the suspended graphene
sheet is hit by a low-intensity electron beam that may push atoms
out-of-plane upward or downward in a random fashion. An alternative
technique to visualize ripples is using a scanning tunneling
microscope (STM)\cite{zan12}. In this case, the graphene sheet is
pushed and locally heated in the region close to the tip. Depending on
the tunneling current and the voltage between tip and sheet, the
latter may undergo a phase transition from a flexible (rippled) to a
rigid (buckled) state \cite{sch15}.
In Ref.~\cite{nl12oha}, the authors simplify the distortion
of the $2$d crystal by modeling it with two-state spin-like variables
($+1$ upward, $-1$ downward). There are antiferromagnetic interactions
among these spins, because the out-of-plane shift of the atoms in
opposite directions stabilize the strictly $2$d system while keeping
the gapless band structure of graphene \cite{nl12oha}. A rich phase
diagram is found, including paramagnetic, ordered and glassy phases,
depending on the temperature and the values of the nearest-neighbor
and next-nearest-neighbor couplings. In this way, they describe the
formation and origin of the atomic scale rippling found in
Ref.~\cite{ASC11mao}. On the other hand, there are also models that
investigate rippling by considering at each lattice site a continuous
variable $u$ describing the out-of-plane displacement of the carbon
atom coupled to a spin variable ($\sigma=\pm 1$) representing an
internal degree of freedom \cite{pre12bon,jsm12bon}. This may be
understood as a mechanical system coupled to a spin system. The spin
at each lattice site represents the non-saturated fourth bond that, by
a physical mechanism similar to the one discussed in \cite{nl12oha},
tries to pull the corresponding atom upward ($u>0$) or downward
($u<0$) from the flat sheet configuration. Mathematically, this is
done by introducing a linear coupling term in the system energy,
proportional to $-u\sigma$ for each lattice site. In these simple
models, the mechanical system is either a chain of oscillators
\cite{pre12bon} or a discrete elasticity model of the hexagonal
graphene lattice \cite{jsm12bon}, while the spins are in contact with
a thermal bath and flip randomly according to Glauber dynamics at the
temperature $T$ of the thermal bath. In both models, the system forms
metastable but long-lived ripples assuming slow spin relaxation
\cite{pre12bon,jsm12bon}.
It is worth investigating a combination of the two approaches
described in the previous paragraph. Firstly, it seems sensible to
model the out-of-plane displacement at each lattice site by a
continuous variable as in Refs.~\cite{pre12bon,jsm12bon}, which is
driven by the internal degree of freedom represented by the
spin. Secondly, these spins certainly interact among themselves, by
the mechanism proposed in \cite{nl12oha}. Thus, in this paper we
discuss the formation and dynamics of ripples in graphene through a
system of atoms connected by harmonic springs and coupled to
interacting Ising spins. We start from the spin-oscillator chain model
\cite{pre12bon,jsm12bon,jstat10} and add interactions among spins that
produce stable rippling states. There appear different phases and
transitions between them depending on parameter values.
The plan of the paper is as follows. In Section \ref{model}, we
introduce the one-dimensional model and calculate analytically the
equilibrium state. We also present the equations that determine the
dynamics of the system. Section \ref{2d} is devoted to the expansion
of the model to $2$ dimensions on a hexagonal lattice, with first and
second-neighbors interactions between spins. Besides, we do a
systematic analysis of the system modifying the parameters and
studying the stationary configuration after the transitory, using a
phase diagram in Section \ref{pd}, and describing the different
phenomenology of each phase in Section \ref{rc}. The main conclusions
are presented in Sec. \ref{con}. Relevant information that is not
covered in the main text is presented in the Appendices: some
geometrical expressions for the hexagonal lattice are discussed in
\ref{AC}, while images of the different phases are collected in
\ref{img}.
\section{The one-dimensional model}\label{model}
To start with, we consider a one-dimensional chain of $N$ oscillators
with nearest-neighbor interactions, in which each oscillator is
linearly coupled to an Ising spin $\sigma_i=\pm 1$. A detailed
investigation of this model can be found in
Ref.~\cite{pre12bon}. Therein, it was shown that, for appropriate
temperatures, the stable equilibrium configuration has only one
ripple, although there appeared some more complex long-lived
metastable rippled states. To explore whether stable multi-rippled
equilibrium configurations are possible, we add an anti-ferromagnetic
term to the hamiltonian,
\begin{equation}
\mathcal{H}=\sum_{j=0}^{N} \biggl[ \frac{p_j^2}{2m}+ \frac{k}{2}(u_{j+1}-u_j)^2-fu_j\sigma_j
+ J \sigma_{j+1}\sigma_j\biggr].
\label{H1}
\end{equation}
Here $u$ and $p$ are the vertical displacement and momentum
respectively, and the extreme oscillators and spins are fixed
($u_0=p_0=\sigma_0=u_{N+1}=p_{N+1}=\sigma_{N+1}=0$) \cite{bc}.
The dynamics {of the model} is as follows: (i) the oscillators{'} equations of motion,
\begin{equation}
\label{eq:mot}
m\,\ddot{u}_j - k\, (u_{j+1}+u_{j-1}-2u_j)=f\sigma_j,
\end{equation}
{are the usual ones,} whereas the spins {evolve according to} Glauber dynamics \cite{Gl63}. The transition
rate from the configuration $(\bm{u},\bm{p},\bm{\sigma})$ to
$(\bm{u},\bm{p},R_j \bm{\sigma})$ (obtained from $\bm{\sigma}$ by
flipping the $j$-th spin) is
\begin{align}
W_j(\bm{\sigma}|\bm{u},\bm{p})=\frac{\alpha}{2}(1-\beta_j \sigma_j), \\
\beta_j= \tanh\! \left[\frac{f}{T} u_j -\frac{J}{T} (\sigma_{j-1}+\sigma_{j+1})\right]\!,
\end{align}
in which $\alpha$ is the characteristic attempt rate for the spin
flips. The Glauber transition rates ensure that the system
satisfies detailed balance, and therefore the system reaches
equilibrium for long enough times. In equilibrium, the probability of
a certain configuration $(\bm{u},\bm{p},\bm{\sigma})$ is proportional
to $e^{- \mathcal{H}/T}/Z$, where we measure the temperature $T$ in units of energy.
As explained in \cite{pre12bon}, for $J=0$ rippling appears
provided the temperature is less than
\begin{equation}
T_0=\frac{f^2K_{N}^{2}}{k}, \quad K_{N}\sim \frac{N}{\pi},
\end{equation}
as $N\to\infty$. To guarantee that the diffusive term in \eqref{eq:mot} remains finite in the continuum limit, it is convenient to nondimensionalize the equations of motion as follows:
\begin{eqnarray}
&&
u_j^*=\frac{k u_j}{f K_{N}^2},
\quad t^*=\frac{t}{K_{N}}\sqrt{\frac{k}{m}},\label{nondim_u_t} \label{par1}\\
&& \kappa=\frac{J}{ T_{0}},\quad\delta=\frac{\alpha K_N\sqrt{m}}{\sqrt{k}},
\quad
\theta=\frac{T}{T_{0}}=T \frac{k}{f^{2}K_{N}^{2}}.
\label{par2}
\end{eqnarray}
Then the transition rates and the equations of motion become
\begin{eqnarray}
&&W_j^*(\bm{\sigma}|\bm{u}^*,\bm{p})=\frac{\delta}{2}(1-\beta_j \sigma_j), \nonumber\\
&&\beta_j= \tanh\! \left[\frac{u^{*}_{j}}{\theta} -\frac{\kappa}{\theta} (\sigma_{j-1}+\sigma_{j+1})\right]\!,\nonumber\\
&&\frac{d^2u^*_j}{dt^{*\, 2}}-K_{N}^2(u^*_{j+1}+u^*_{j-1}-2u^*_j)= \sigma_j. \label{nondim}
\end{eqnarray}
We will omit the asterisks in nondimensional equations from now on. In
order to obtain the scaling of the critical temperature, we need to
know the scaling of the model parameters with the system size. In
principle, this could be done by deriving our model from a fundamental
microscopic one, but this is outside the scope of this
paper. Nevertheless, we discuss some possible scalings in the
following. If both the elastic constant $k$ and the antiferromagnetic
coupling constant $J$ are considered to be independent of the system
size, the only remaining parameter is $f$, the coupling between the
elastic and internal (spin) degrees of freedom. If $f$ is also
independent of the system size, the critical temperature $T_{0}$
diverges as $N^{2}$. In this case, rippling would be observed at all
temperatures. On the other hand, a finite value of $T_{0}$ in the
large system size limit is obtained when $f\propto N^{-1}$. Then,
rippling would be observed only for $T<T_{0}$. In principle, both
situations are compatible with current experiments, in which rippling
is observed over a wide temperature range.
Let us consider now the equilibrium situation.
Equation \eqref{nondim} can be averaged, with the result
\begin{equation}\label{curv}
\frac{1}{\pi^{2}} \frac{d^{2}}{dx^{2}} \langle u\rangle=-\langle
\sigma\rangle,
\end{equation}
in which $\langle u\rangle$ and $\langle\sigma\rangle$ are the
equilibrium average height and spin at position $x=i/N$;
the system has unit size in the continuous space variable
$x=i/N$, $0\leq x\leq 1$. Therefore, the average curvature of the
ripples is directly linked to the average spin. Very recently, this
idea has been used to develop a phenomenological Ising model to study
rippling in graphene in scanning tunneling microscopy experiments, in
which each spin represents a whole ripple and the spin sign gives its
corresponding convexity \cite{sch15}. Interestingly, {we can derive}
an effective free energy for the string, by integrating
the canonical distribution over the momenta $\bm{p}$ and the spins
$\bm{\sigma}$: the resulting probability $\mathcal{P}$ becomes a
functional of the string profile $u(x)$, which in dimensionless
variables reads \cite{unpub}
\begin{subequations}\label{F-cont}
\begin{equation}
\mathcal{P}[u] \propto \exp{\left(-\frac{\mathcal{F}}{\theta}\right)}, \quad
\mathcal{F}[u]=N \int_{0}^{1} dx \,
\underbrace{\left[\frac{1}{2\pi^{2}}\left(\frac{\partial u}{\partial
x}\right)^{2}-\theta \ln\zeta\left(
\frac{u}{\theta},\frac{\kappa}{\theta} \right)\right]}_{f(u,\frac{du}{dx})},
\end{equation}
\begin{equation}
\zeta\left(
\frac{u}{\theta},\frac{\kappa}{\theta} \right)=\exp\left(-\frac{\kappa}{\theta}\right)
\cosh\left(\frac{u}{\theta}\right)+\exp\left(\frac{\kappa}{\theta}
\right) \sqrt{1+\exp\left(-\frac{4\kappa}{\theta}\right)\sinh^{2}\left(\frac{u}{\theta}\right)}.
\end{equation}
\end{subequations}
The quantity $\ln\zeta$ is the logarithm of the spins
partition function per site, which depends on the ``field''
$f u_{j}/T \to u/\theta$ and the coupling $J/T\to
\kappa/\theta$.
The particularization of this free energy to the $J=0$ case was
obtained in Ref.~\cite{pre12bon}. For $J\neq 0$, in order to
calculate $\ln\zeta$, the system is divided into a set of nearly
independent subsystems with $N_{s}\gg 1$ sites each, but such that
$N_{s}\ll N$ and the ``field'' $u$ can be considered almost constant
within each subsystem. We may denote the subsystems by $S(x)$, in
which $x$ corresponds to the position of the subsystem in the
continuum limit. In each subsystem, the local average magnetization
$\mu=N_{s}^{-1}\sum_{j\in S(x)}\langle\sigma_{j}\rangle$ and
correlation
$C=N_{s}^{-1}\sum_{j\in S(x)}\langle\sigma_{j}\sigma_{j+1}\rangle$
are {given by} the usual formulas
\begin{equation}
\label{eq:magn_corr}
\mu=\theta\frac{\partial\ln\zeta}{\partial u}, \quad C=-\theta\frac{\partial\ln\zeta}{\partial\kappa}.
\end{equation}
{as} $u$ plays the role of the external field and {$\kappa$} is the
coupling constant. Global order parameters may be defined as
\begin{equation}
\label{eq-global-or-param}
M=\left|\,\int_{0}^{1} dx \, \mu \,\right|, \quad \mathcal{DL}=\frac{1}{2}\left(1-\int_{0}^{1} dx \, C\right),
\end{equation}
{in which $M$ is} the absolute value of the total magnetization
per site and the fraction of spin pairs in an antiferromagnetic
configuration {is $\mathcal{DL}=1$ } ($\mathcal{DL}=0$) for
perfect antiferromagnetic (ferromagnetic) ordering.
Equation~\eqref{F-cont} clearly shows that the
free energy is an extensive quantity if the dimensionless variables
are of the order of unity, which {is consistent with} our
scalings. Taking into account the expressions for the free energy,
Eq.~\eqref{F-cont}, and the average magnetization,
Eq.~\eqref{eq:magn_corr}, the equation giving the equilibrium
profile \eqref{curv} is nothing but the Euler-Lagrange equation for
the free energy functional, as it should be: the string profile
minimizes the free energy.
In what follows, we summarize the main physical implications
of the short-ranged antiferromagnetic interaction, as compared to
the $J=0$ case. The flat profile $u=0$ is always a solution of
Eq.~\eqref{curv}, but it becomes unstable for
$\exp(-2\kappa/\theta)/\theta>1$. For $J=0$, {there appear} rippled configurations
with non-zero magnetization {that are stable} for
$\theta<1$ \cite{pre12bon}{. F}or $J\neq 0$ the bifurcation
condition $\exp(-2\kappa/\theta)/\theta=1$ {produces two temperatures}
$\theta_{1}$ and $\theta_{2}$ for $\kappa<0.18$, as seen in the left
panel of Fig.~\ref{bif}. Specifically, these rippled configurations
become unstable for low (high) enough temperatures,
$\theta<\theta_{1}$ ($\theta>\theta_{2}$). The transitions at
$\theta_{1}$ and $\theta_{2}$ are of second order, the order
parameters $M$ and $\mathcal{DL}$ are continuous because $u$
bifurcates continuously from the solution $u=0$, similarly to the
behaviour found in the $J=0$ case. For $\kappa>0.18$, this rippled
ferromagnetic phase no longer exists because the antiferromagnetic
coupling is too strong.
\begin{figure}
\centering
\includegraphics[width=0.49\textwidth]{bifurcation_condition_2ndorder.pdf}
\includegraphics[width=0.49\textwidth]{bifurcation_condition_low_T.pdf}
\caption{\label{bif} (Left) Function controlling the bifurcation to the
ferromagnetic phase. The bifurcation condition
$\exp(-2\kappa/\theta)/\theta=1$ is equivalent to
$y(\theta)=-\frac{1}{2}\theta\ln\theta=\kappa$. It is clearly seen
that the bifurcation condition is fulfilled by two temperatures
$\theta_{1}$ and $\theta_{2}$ for $\kappa$ smaller than the maximum $y_{\text{max}}=(2e)^{-1}$
of $y$, that is, $\kappa<y_{\text{max}}$. (Right) Function controlling the
length $x_{0}$ of the antiferromagnetic regions near the borders of
the system at low temperatures. This length is given by
$\pi^{2}x_{0}(1-2x_{0})/4=\kappa$, which has two solutions $x_{01}$
and $x_{02}$, $x_{02}=1/2-x_{01}$, for $\kappa<\pi^{2}/32\simeq 0.3$
and no solutions for $\kappa>0.3$. At the limit value
$\kappa=0.3$, it is $x_{01}=x_{02}=1/4$. }
\end{figure}
In the limit $\theta\to 0^{+}$, the partition function {of the spins becomes}
\begin{subequations}
\label{eq:T=0}
\begin{equation}
\lim_{\theta\to 0}\theta\ln\zeta\left(
\frac{u}{\theta},\frac{\kappa}{\theta} \right)=\kappa+(|u|-2\kappa)\, \eta(|u|-2\kappa),
\end{equation}
\begin{equation}
\lim_{\theta\to 0}\mu= \text{sgn}(u)\,\eta(|u|-2\kappa), \quad \lim_{\theta\to
0}C=\text{sgn}(|u|-2\kappa).
\end{equation}
\end{subequations}
where $\eta(x)$ is the Heaviside step function, $\eta(x)=1$
for $x>0$ and $\eta(x)=0$ for $x<0$, and $\text{sgn}(x)$ is the sign
function, $\text{sgn}(x)=2\eta(x)-1$. In the flat configuration, $\mu=0$
and $C=-1$ everywhere. {There appears a new rippled low temperature phase,
that is antiferromagnetic near the boundaries because of} the clamped
boundary conditions{. I}nside an interval of length
$x_{0}$ close to the boundaries, $|u|<2\kappa$, and
$\mu=0$, $C=-1$ therein. Then, the simplest configuration is
composed of (i) two straight lines ($u''=0$) near the boundaries,
that is, in the intervals $(0,x_{0})$ and $(1-x_{0},1)$, and (ii) a
parabolic ripple with $|u|>2\kappa$ ($u''=\pm 1$) in between, for
$x\in(x_{0},1-x_{0})$, which corresponds to ferromagnetic ordering
because $\mu= \pm 1$ and $C= 1$ for $|u|>2\kappa$. The continuity of
$u$ and $u'$ at $x=x_{0}$ implies that
$\pi^{2}x_{0}(1-2x_{0})=4\kappa$, which determines two possible
values of $x_{0}$ for $\kappa<0.3$, as seen in the right panel of
Fig.~\ref{bif}. The configuration corresponding to the smallest
value $x_{01}$ is the absolute minimum of the free energy for
$0\leq x_{0}<1/8$, whereas both the flat string and the
configuration corresponding to $x_{02}$ are metastable. For
$1/4>x_{01}>1/8$, the absolute minimum of the free energy
corresponds to the flat string, whereas both configurations
corresponding to $x_{01}$ and $x_{02}$ are metastable. The
transition at $x_{c}=1/8$ (which corresponds to $25$ percent of the
spins being antiferromagnetic) is first order, because the order
parameters $M$ and $\mathcal{DL}$ change discontinuously: in the
zero temperature flat configuration, it is $M=0$ and
$\mathcal{DL}=1$, whereas in the rippled state we have $M=1-2x_{0}$
and $\mathcal{DL}=2x_{0}$. Moreover, as is usually the case in
first-order phase transitions, the string has many other metastable
configurations: they have $n$ internal nodes $x_{i}$,
$i=1,\ldots,n$, at which $u$ changes sign. The existence of the
different phases and their relative stability will be thoroughly
discussed elsewhere \cite{unpub}.
\section{The two-dimensional model} \label{2d}
Here, we extend the model to dimension $d=2$, in the
hope that this will make it possible to find more complex behaviors.
This extension is almost direct and, as we are
interested in applying the model to mimic a graphene sheet, we
consider a hexagonal lattice. Due to the symmetry, it is important to
write the hamiltonian carefully. First, each atom is indexed: $i$ will
be the row index and $j$ the column index, with the peculiarity that
each row comprises atoms with two different heights in a zigzag
distribution, see Figure \ref{esq}. It is important to note that the
form of the equations will be qualitatively different for atoms for
which $|i-j|$ is an even number \textit{(e-atoms)}, which have one
nearest neighbor above and two below, and those for which $|i-j|$ is
an odd number \textit{(o-atoms)}, which have one nearest neighbor
below and two above, that is, the opposite situation. It is quite
obvious that if the plane is rotated by an angle of $\pi$, the two
types of atoms are interchanged.
\begin{figure}
\centering
\includegraphics[width=0.6\textwidth]{esq.pdf}
\caption{Figure summarizing the atoms indexes and the
parameters of the unit cell of the hexagonal lattice.}
\label{esq}
\end{figure}
Taking into account the notation described above, we can write down
the extension of the 1$d$ Hamiltonian to $d=2$. Moreover, we introduce
next-nearest-neighbor interactions,
\begin{eqnarray}
\mathcal{H} &=&\sum_{ij} \left[\frac{p_{ij}^2}{2m}-fu_{ij}\sigma_{ij}
+ J'
\sigma_{ij}(\sigma_{i-1,j-1}+\sigma_{i,j-2}+\sigma_{i+1,j-1})\right]
\nonumber
\\ &+& \sum_{|i-j|=\text{even}} \Bigg\{\frac{k}{2}\left[
(u_{ij}-u_{i+1,j})^2+
(u_{ij}-u_{i,j-1})^2+(u_{ij}-u_{i,j+1})^2\right] \nonumber \\
&& \qquad\qquad+ J
\sigma_{ij}(\sigma_{i+1,j}+\sigma_{i,j-1}+\sigma_{i,j+1})\Bigg\},
\label{H}
\end{eqnarray}
where $i$ and $j$ take values $1 \to i_{\text{max}}$ and $1 \to
j_{\text{max}}$, respectively.
Following the same steps as in the previous section, the
nondimensional equation of motion for each atom and the expressions
for the transition rate become
\begin{equation}
\ddot{u}_{ij} - K_N^2 (u_{i+1,j}+u_{i,j-1}+u_{i,j+1}-3u_{ij})= \sigma_{ij}, \label{mov}
\end{equation}
\begin{equation}
\omega_{ij}(\bm{\sigma}|\bm{u})=\frac{\delta}{2}(1-\gamma_{ij}
\sigma_{ij}),
\end{equation}
\begin{eqnarray}
\gamma_{ij}&=&\tanh[\frac{u_{ij}}{\theta} -\frac{\kappa}{\theta}(\sigma_{i+1,j}+\sigma_{i,j+1}+\sigma_{i,j-1})\nonumber
\\ && \;\;
-\frac{\lambda}{\theta}(\sigma_{i,j-2}+\sigma_{i,j+2}+\sigma_{i-1,j-1}
+\sigma_{i-1,j+1}+\sigma_{i+1,j-1}+\sigma_{i+1,j+1})],
\label{gamma}
\end{eqnarray}
where $K_N$ is a large scale parameter to be calculated later. As we
said before, the difference between e-atoms with o-atoms follows from
the rotation by $\pi$ of the plane. For that reason, only equations
for e-atoms have been written. In the latter equations, the height
variable $u$ and time are dimensionless. In the
nondimensionalization, the same parameters as in equations
\eqref{par1} and \eqref{par2} appear, with the addition of
$\lambda=J'/T_{0}$, which corresponds to the new next-nearest-neighbor
interaction.
The length of each side of the finite hexagonal lattice is
$\tilde{L}=[3(n-1)+1]\tilde{l}/2$, where $\tilde{l}$ is the side of a
unit hexagonal cell and $n$ is the maximum value of the row index $i$
in Fig.\ \ref{esq}. Let us measure all lengths in units of
$\tilde{L}$, so that $l=\tilde{l}/\tilde{L}$ tends to zero as the
hexagonal lattice fills the plane. Then the expression within
parenthesis in \eqref{mov} has the limit
\cite{prb08car,prb12bon86,jsm12bon}
\begin{equation}
u_{i+1,j}+u_{i,j-1}+u_{i,j+1}-3u_{ij} \rightarrow \frac{a^2}{4}(\partial_x^2 u +\partial_y^2 u), \label{lap}
\end{equation}
as $a=\sqrt{3}l\to 0$.
Therefore, we take $K_{N}$ proportional to $a^{-1}$, namely
\begin{equation}
K_N=\frac{\sqrt{2}}{\pi} a^{-1}=\frac{3n-2}{\sqrt{6}\pi}\propto n,
\end{equation}
to guarantee that the diffusive term in \eqref{mov} remains finite as
$l\to 0$ (continuum limit). Note that the increments of the
continuous variables are $\Delta x(i\to i+1)=3l/2$ and
$\Delta y(j\to j+1)=a/2=\sqrt{3}l/2$, as seen in Fig.~\ref{esq}, so
that the hexagonal lattice goes to the unit square $0\leq x,y \leq
1$ in the continuum limit. For details, see \ref{AC}.
Once the system reaches the stationary state, equation \eqref{mov} can be averaged ignoring thermal fluctuations. Thus, using equation \eqref{lap} for $n \gg 1 $ we get
\begin{equation}
\label{cont_2d}
\frac{1}{2\pi^{2}}\nabla^2 \langle u\rangle=- \langle \sigma \rangle,
\end{equation}
where $\langle u\rangle$ and $\langle \sigma\rangle$ are the
average height and spin at the point $(x,y)$ of the
unit square. For $\kappa=\lambda=0$, we have that
$\langle\sigma\rangle=\tanh(u/\theta)$ and the flat configuration
$\langle u\rangle=0$ becomes unstable at $\theta=1$, similarly to
the situation in the 1d case. This kind of rigidly buckled
configurations have been observed in graphene in recent STM
experiments \cite{sch15}. Equation \eqref{cont_2d} tells us that there is a correspondence between lattice
patterns given by the average height profile and the spin
configuration. Specifically, the curvature of the rippling is
directly proportional to the average spin. Therefore, in the
following section we will mainly characterize the phases by the spin
configuration.
\subsection{Phase diagram} \label{pd}
{As it has already been said, the stable steady state is a rigidly buckled configuration below the critical temperature ($\theta<1$), provided there is no interaction between spins}. We expect that the introduction of the nearest neighbour interaction among the spins should introduce new phases{. By analogy with the 1d system, an antiferromagnetic nearest neighbour
interaction should make antiferromagnetic ordered phases
to appear} for low enough temperatures. Looking for a more
complex phenomenology, we introduce a next-nearest-neighbor
interaction, as in Ref.~\cite{nl12oha}. This term appears in \eqref{H}
through $J'$ and in \eqref{gamma} through its dimensionless
counterpart $\lambda$.
It is important to note that both $J$ ($\kappa$) and $J'$
($\lambda$) may take positive or negative values, corresponding to
antiferromagnetic and ferromagnetic interactions,
respectively. However, only positive values of $\lambda$ will be
considered, since a next-nearest-neighbor ferromagnetic interaction
just strengthens the nearest-neighbor one \cite{note}. For positive
values of $\kappa$ and $\lambda$ the qualitative behavior is quite
different. The nearest-neighbor interaction provides a defined minimum
energy distribution in which each spin and its nearest-neighbors point
in opposite directions. However, the next-nearest-neighbor interaction
does not yield a defined minimum energy distribution. In fact, the
second-neighbors of each atom are second-neighbors to each other, and
therefore the next-nearest-neighbor interaction causes the system to
be frustrated \cite{nl12oha}. An enlightening discussion about
frustration is given in the introduction of \cite{mezard}. In
principle, it is tempting to exclude negative values of $\kappa$
from the analysis. On intuitive grounds, one may conclude that the
nearest-neighbour ferromagnetic coupling with $\kappa<0$ should only
strengthen the already long-ranged ferromagnetic interaction among
the spins induced by the spin-lattice coupling term
$-f u_{ij}\sigma_{ij}$ \cite{pre12bon}. Nevertheless, the situation
is a little bit more complex, as discussed below.
\begin{figure}
\centering
\includegraphics[width=0.6\textwidth]{diag3_v2.pdf}
\caption{Phase diagram for a hexagonal lattice coupled to Ising
spins. Different regions have been delimited using the domain-wall
parameter, the magnetization and the specific heat. Once
the equilibrium state is reached, each region has a
different behavior, which is explained in the
text. Also plotted is the line $\kappa/\lambda=4$, which is
a good estimate for the transition line between zones $C$ and $D$.
This agrees with the line separating phases Ordered $1$ and $2$ in
Ref.~\cite{nl12oha}.}
\label{diag}
\end{figure}
We plot a phase diagram to show in only one graph all the different
behaviors, see Figure \ref{diag}. In our simulations, we have chosen
{a nondimensional temperature $\theta=0.01$,} which is far
below critical for $\kappa=\lambda=0$. A key parameter is
\begin{equation}
\mathcal{DL}=\frac{1}{N}\sum_{|i-j|=even} [3+\sigma_{ij}(\sigma_{i+1,j}+\sigma_{i,j-1}+\sigma_{i,j+1})],
\end{equation}
where $N$ denotes the number of atoms in the lattice.
This parameter estimates the domain-wall length
\cite{nl12oha}, and it is equal to $3$ (resp.~$0$) for
completely ferromagnetic (resp.~antiferromagnetic) behavior. In
addition, to delimit the regions on the diagram, we have used the
absolute value of the usual magnetization
\begin{equation}
M=\left|\frac{1}{N}\sum_{ij}\sigma_{ij}\right|\!,
\end{equation}
and energy fluctuations (proportional to the specific heat),
\begin{equation}
F=\sqrt{\langle(\Delta^{*}
e)^{2}\rangle}.
\end{equation}
Here $e$ is the system energy and $\Delta^{*}e=(e-\langle e\rangle)$,
where the angular brackets stands for the mean value that is calculated once the stationary state has been reached.
\subsection{Region characterization} \label{rc}
The different regions in the phase diagram have been characterized
using the three parameters $\mathcal{DL}$, $M$ and $F$. Figure
\ref{diag} is the superposition of the projections of $M$ and
$\mathcal{DL}$ on the plane $\lambda/\theta-\kappa/\theta$. Each region of the plane
correspond to different combinations of $M$ and $\mathcal{DL}$ values.
Once the regions have been delimited using the magnitudes described
above, the system is allowed to evolve with $\kappa$ and $\lambda$ in
one of the regions. Next, we verify that the system reaches the
equilibrium state and we obtain the basic structures in the spin
domains. Moreover, to check that we have actually reached the
equilibrium state, another simulation is carried out with this
distribution as the initial condition: if, aside from thermal
fluctuations, no evolution is found, equilibrium has been reached.
\begin{itemize}
\item Region $A$. $\mathcal{DL}\sim 3$, $M\sim 1$. The plane is
completely curved, and the spins are all pointing in the same
direction. This situation corresponds to small values of
$\kappa$ { and $\lambda$}, for which the interaction that dominates is the one
between the surface and the spins, in agreement with the simple
picture already present in the 1$d$ model, see Section \ref{model}.
\item Region $B$. It is the zone surrounding $A$, on which
$\mathcal{DL}$ and $M$ decrease from the $A$ values to those on the
other regions. Here, the system displays a behavior that is
analogous to the one described at the end of Section \ref{model} (in
1$d$). The interaction between the surface and the spins is an
effective ferromagnetic interaction with an intensity that decreases
from the center to the border. Thus, the plane is curved but the
spins close to the border are antiferromagnetically arranged.
\item Region $C$. $\mathcal{DL}\sim 0.5$, $M\sim 0$. The predominant
interaction is the antiferromagnetic first-neighbor one. The
equilibrium state (starting from a random initial spin distribution)
is composed of antiferromagnetic domains.
\item Region $D$. $\mathcal{DL}$ increases from $0.5$ to $1.2$, $M\sim
0$. The states in this region are metastable. Taking the
distribution corresponding to the equilibrium state of $C$ or $E$ as
the initial condition, the system does not evolve to the other
state, at least in a simulation time much greater than the
relaxation time from random initial conditions.
\item Region $E$. $\mathcal{DL}\sim 1.2$, $M\sim 0$. In this case, the
spins are distributed in rows of two atoms in the lowest energy
configuration. Beginning with random initial conditions, these
two-atoms domains were created, with the rows in any of the three
symmetrical directions.
\item Region $F$. $\mathcal{DL}\sim 1.5$, $M\sim 0.2$. The interaction
between the plane and the spins is relevant again, since the
antiferromagnetic next-nearest-neighbor interaction (with $\kappa
\sim 0$) has no defined minimum energy distribution. The plane is
curved, leading to a non-zero magnetization.
\item Region $G$. $\mathcal{DL}\sim 1.8$, $M\sim 0$. The typical
equilibrium configurations are long serpentine lines, with zero
magnetization. Taking as initial conditions the spins arranged in
rows, the system remains static.
\item Regions $H$ and $I$. In them, the system evolves from the
characteristic configurations of $G$ to the ferromagnetic
configurations of $J$, with the difference that in $H$ the
magnetization is different from zero whereas in $I$ it is not. $I$
is a ferromagnetic first-neighbor state, but with domains smaller
than in $J$ ($\mathcal{DL}$ smaller than in $J$).
\item Region $J$. $\mathcal{DL}\sim 2.5$, $M\sim 0.1$. In this case
the system behaves as a completely ferromagnetic first-neighbor
system. Starting from random initial conditions, ferromagnetic
domains grow until reaching a stationary state. In this case $M$ is
close to zero since spins are pointing to different directions in
adjoining domains.
\end{itemize}
The plots of the typical equilibrium configurations for each region
are in \ref{img}.
It should be noted that our phase diagram does not contain a
paramagnetic state because the chosen temperature, $\theta=0.01$, is
far below the critical temperature for $\kappa=0$ and
$\lambda=0$ (unity in our dimensionless variables). Each point of
the phase diagram corresponds to the energy minimum to which the
system evolves for the considered parameters. Once it is in the
neighborhood of this minimum, the energy barriers are so high that
ergodicity is no longer valid, and the system remains \textit{frozen}
\cite{hemmen}. This causes an Edwards-Anderson order parameter
\cite{edw},
\begin{equation}
q_{EA} \equiv \frac{1}{N} \sum_{ij} \mu_{ij}^{2}, \quad \mu_{ij} \equiv \langle \sigma_{ij} \rangle,
\label{q}
\end{equation}
to be different from zero at every point of the
plotted phase diagram. On the other hand, close to
$\kappa=\lambda=0$, the order parameter $q_{EA}$ will vanish as
$\theta\to \infty$, once ergodicity is recovered. In
Eq.~\eqref{q} the average should be understood as a time average or an extended Gibbs average in a phase space composed of disjoint ergodic components \cite{hemmen}.
\section{Conclusions}
\label{con}
We have studied a system of atoms connected by harmonic springs and
coupled to Glauber spins. The spins are in contact with a thermal
bath and interact with their neighbors. The $1d$ system forms one
ripple and becomes antiferromagnetic at the boundaries as $\rho$
increases, until it becomes completely antiferromagnetic. When the
system is on a $2d$ hexagonal lattice, each spin interacts with its
nearest-neighbors and next-nearest-neighbors, aside from the coupling
with the out-of-plane displacement. This situation generates
different phases which are included in a phase diagram.
The range of parameters in our phase diagram includes negative values
for the nearest neighbor coupling constant $\kappa$ and is thus wider
than the one used in \cite{nl12oha}, in which only antiferromagnetic
interactions were considered. The change in the sign of the spin-spin
interaction can be produced by the scattering of the conduction
electrons at the spins, see \cite{ruder,binder}. We are interested in
zero magnetization phases since they correspond to no overall
bending. Our model provides different phases obeying this constraint:
I and J are long wave length phases, similar to those observed in
\cite{nature07mey,pss09ban}, whereas C, E and G are phases with
atomic wave length. G is a stripy phase (see Figure \ref{FG}),
which could be associated with patterns seen in
\cite{ASC11mao}. The atomic wave length phases C and E correspond
with phases Ordered 1 and 2, respectively, from
Ref. \cite{nl12oha}. Therein, the line between these two phases is
(in our variables) $\kappa/\lambda=4$, which agrees with the limit of
true stability of C here. Interestingly, neither the metastable phase
$D$ or the other phases (including the long wavelength phases I and
J) were found in Ref.~\cite{nl12oha}. In that reference, (i) only
positive values of $\kappa$ were considered, and (ii) there was no
spin-atom coupling.
The buckling phase A is surrounded by rippled phases C, E, and G with
no overall bending. Starting from a point of the phase
diagram belonging to region C (rippled phase), if we
increase the temperature while keeping $\kappa$ and $\lambda$
(supposed temperature independent)
fixed, we move along a straight line of slope $\lambda/\kappa$ in
Figure \ref{diag} from phase C to the buckled phase A. In
experiments with STM at fixed current, the temperature is locally
increased at the tip region and this triggers a transition from a
rippled flexible phase to a rigid buckled phase \cite{sch15}. Thus our
model contains the ripples-to-buckling transition observed in
experiments although more work needs to be done to explain STM
observations in detail \cite{unpub}.
To conclude, our model is based in a few parameters controlling simple interactions
which generate complex collective behaviors. This allows us to
identify the interactions responsible for each pattern. In addition,
the elastic feature of the model makes it possible to visualize and
quantify the magnitude of the rippling, which could be compared with
experiments once height measurements had been improved.
\ack
This work has been supported by the Spanish Mi\-nisterio de Econom\'\i
a y Competitividad grants FIS2011-28838-C02-01 (MRG \& LLB), and
FIS2011-24460 (AP). MRG acknowledges support from Ministerio de
Educaci\'on, Cultura y Deporte through the FPU program grant FPU13/02971.
|
\section{Introduction}
The starburst event is important to understand evolution of galaxies.
Most galaxies experience starburst phases in their evolutionary histories, and present-day stars might have been formed mostly in the starbursts (e.g. \citealp{Elb03}).
Massive stars formed from the starbursts play an important role in stellar feedback such as chemical enrichment.
Therefore, the starburst is one of the main events for evolution of galaxies, and a study of nearby starburst galaxies can provide insights into the processes of the starburst.
One of the robust definitions for starburst galaxies is that galaxies have much higher star formation rate than that of their past average (e.g. \citealp{Meu95}). We need to derive star formation history and current star formation rate of galaxies to determine the starburst galaxies by this definition (e.g. \citealp{McQ10}). However, it is hard to derive star formation histories of galaxies, so the starburst galaxies can be categorised by finding other properties of the starburst events such as high star formation rates, existence of massive stars, and presence of super star clusters.
IC 10 is the nearest starburst galaxy \citep{Mas02,Kim09}.
An unusually large number of HII regions \citep{Hod90}, and far-IR luminosity \citep{Mel94} in this galaxy show that it has a much higher star formation rate than other dwarf galaxies in the Local Group.
IC 10 has a large number of Wolf-Rayet (W-R) stars, considering its total luminosity \citep{Mas02}.
The surface number density of W-R stars in IC 10 is about one hundred times higher than that of the Large Magellanic Cloud, and it is the highest among the Local Group galaxies, showing that the star formation in IC 10 is on the burst mode \citep{Mas95,Mas02}.
IC 10 is also known as the nearest blue compact dwarf galaxy (BCD) based on its surface brightness and structure. The $B-$band surface brightness of IC 10 is much higher than that of other dwarf galaxies, and it is similar to that of BCDs \citep{Ric01}. BCDs possess underlying old stellar populations with low surface brightness (e.g. \citealp{Dro01,Cro02}). IC 10 also has old stellar populations such as red giant branch (RGB) stars and C stars \citep{Dem04,Kim09,San10}, and they are much more extended and much fainter than the young stellar main body. BCDs mostly have very high rates of star formation as suggested by their strong H$\alpha$ emission \citep{Gil03}, and this property of BCDs is similar to that of the starburst galaxies. These properties of BCDs are also similar to those of IC 10.
To study formation and evolution of IC 10, several studies focused on the star clusters in IC 10.
\citet{Kar93} and \citet{Geo96} presented the first list of seven star cluster candidates in IC 10 detected in the ground-based images, including only their positions.
\citet{Hun01} found 13 stellar associations and clusters from $F336W$, $F555W$, $F814W$, and $F656N$ images of the eastern part of the main body of IC 10 taken with Wide Field Planetary Camera 2 (WFPC2) on the Hubble Space Telescope (HST). Two of them are old (with ages$>350$Myr) and the rest are young (ages with 4--30 Myr). They have half-light radii of
1.5--6.0 pc and absolute magnitudes of $-10.0<M_V<-6.6$.
\citet{Hun01} suggested that these young star clusters might have been formed during the recent starburst phase.
Later \citet{Tik09} searched for star clusters in IC 10, using WFPC2 and Advanced Camera for Survey (ACS)/Wide Field Channel (WFC) images of the main body of IC 10, and WFPC2 images of two outer fields in the HST archive, finding 57 star clusters.
They confirmed that four of the seven star cluster candidates in \citet{Kar93} covered by the HST images are indeed star clusters.
They found that
two of the star clusters in \citet{Hun01} are single stars, and two other clusters in \citet{Hun01} are part of an extended star complex.
They classified roughly the star clusters into three age groups based on their morphology: young (34), intermediate-age (5), and old (18).
However, no photometry data is available for these star clusters.
More recently, \citet{Sha10} provided a catalog of star clusters in IC 10 including three band photometric data based on the same HST images as used in \citet{Tik09}.
These previous studies covered mainly the main disk region of IC 10 so that little is known about the star clusters in the outer region of this galaxy.
The existence of the RGB stars in the outer region \citep{Dem04,San10} suggests that there may be significant population of old star clusters in the outer region.
In addition, photometric information is not enough to study physical properties for most of the previously known star clusters.
In this study we survey the outer region of IC 10 to find any star clusters in the halo using the wide field images available in the archive.
We also derive homogeneous photometric magnitudes of all star clusters in IC 10 and investigate their photometric properties.
This paper is organized of as follows.
Section 2 describes data,
the method of star cluster selection,
how to derive the integrated photometry
of the star clusters, and
how to estimate ages of the star clusters.
\S 3 presents a catalog of star clusters newly discovered as well as previously known in IC 10. Then we show color-magnitude diagrams, color-color diagrams, age distribution, and spatial distribution of the star clusters.
Implications of the results are discussed in \S 4,
and primary results are summarized in the final section.
We adopt a distance to IC 10, 715 kpc ($(m-M)_0=24.27\pm0.03$) and a foreground extinction values, $E(B-V)=0.52\pm0.04$ that were determined by the measurement of $K_S$-band luminosity of the tip of the red giant branch, and $UBV$ photometry of early-type stars, respectively \citep{Kim09}.
At this distance, one arcsecond corresponds to 3.5 pc, and one arcminute is 208 pc.
$D_{25}$ of IC 10 is $6\arcmin.3$ \citep{deV91} that is 1.3 kpc at the distance of IC 10.
The star formation rate of IC 10 is $0.6 M_{\odot}/yr$, derived from Br$\gamma$ IR imaging \citep{Bor00}, and the surface star formation rate of IC 10 is $0.03 M_{\odot} yr^{-1} kpc^{-2}$, based on the integrated H$\alpha$ luminosity of IC 10 \citep{Hun01}.
However, we adopted the surface star formation rate, $0.08 M_{\odot}yr^{-1}kpc^{-2}$ that is a logarithmic mean value of those in \citet{Bor00} and \citet{Hun01}.
\section{Data Reduction \& Analysis}
\subsection{Data}
We use $UBVRI$ images of IC 10 in the Local Group Survey (LGS, \citealp{Mas07}) and $R$ band Suprime-cam images of IC 10 in the Subaru archive (SMOKA).
The LGS images cover a $37\arcmin \times 37\arcmin$ field with a pixel scale of
$0\arcsec.27$/pixel, and the Subaru images have $29\arcmin \times 37\arcmin$ field of view
with a pixel scale of $0\arcsec.20$/pixel.
The seeing values of the images are about $0\arcsec.9$ and $0\arcsec.7$ for the LGS and Suprime-cam images, respectively.
The field of view of the overlapped region between these two fields is about $25\arcmin \times 34\arcmin$,
and we searched for star clusters in this overlapped region.
The positions of these fields are marked in Figure 1.
\begin{figure}
\epsscale{1.0} \plotone{fig1.eps}
\caption{A grey scale map of the digitized sky survey image of IC 10 showing the positions of the fields used in this study: Subaru field (dashed line), NOAO Local Group Survey field (dot-dashed line), and HST fields (small polygons). Solid line contours represent an HI map from \citet{van02}.
The solid line box represents the survey area in this study
\label{finder}}
\end{figure}
\subsection{Star Cluster Selection}
Typical star clusters appear to be slightly extended in the Subaru R-band images, but not in the LGS images. In addition, a small number of individual faint stars in the outskirts of star clusters in IC 10 are resolved in the Subaru images because of its depth and better seeing. Therefore we used Subaru $R$-band images for star cluster selection. However, Subaru images have only in $V$,$R$, and $I$ band data, so they are limited in estimating star cluster ages. We used the LGS images to extend the wavelength coverage of the SEDs for star clusters. Therefore we selected star clusters in IC 10 with following steps.
We ran Source Extractor \citep{Ber96} to the LGS $R$-band images of IC 10 and made a list of extended sources with small stellarity values ($<0.8$).
Then we inspected the images of the bright extended sources with $R<21.0$ mag (similar to the magnitude limit of the known star clusters in IC 10) in the
Subaru $R$-band images to select star clusters according to the selection criteria: (a) circular shape of the sources, and (b) the existence of some resolved stars in the outer region of each source.
We searched only the outer region of this galaxy, because the star clusters in the main body of the galaxy were already found in previous studies based on HST images \citep{Hun01,Tik09}.
Finally, we selected seven star clusters. Two of these were found in \cite{Kar93} and the rest are new ones.
Thumbnail images of the five new star clusters are shown in Figure \,\ref{thu}.
We also show a thumbnail image of one known star cluster (No. 1 in \citet{Tik09}) for comparison.
\begin{figure}
\epsscale{1.0} \plotone{fig2.eps
\caption{Gray scale maps of thumbnail $R$ band Suprime-cam images of five newly found star clusters in IC 10, and one known star cluster (ID 1). Field of view of each image is 10\arcsec by 10\arcsec. North is up and east to the left.
\label{thu}}
\end{figure}
\subsection{Photometry and Size Estimation}
We derived instrumental $UBVRI$ magnitudes of the five new star clusters as well as the 61 known star clusters \citep{Kar93,Tik09} using an aperture with $1\arcsec.6$ radius from the LGS images with PHOT/DAOPHOT task in IRAF\footnote{IRAF is distributed by the National Optical Astronomy Observatories, which are operated by the Association of Universities for Research in Astronomy, Inc., under cooperative agreement with the National Science Foundation.}.
We used annuli of $5\arcsec.4$--$8\arcsec.1$ for sky estimation.
Then we transformed these magnitudes on the standard system using the instrumental magnitudes of bright point sources in the same images and the $UBVRI$ photometry catalog of the point sources in IC 10 given by \citet{Mas07}.
Full Width at Half Maximum (FWHM) values of $UBVRI$ band images are similar ($0\arcsec.9$-$1\arcsec.0$), and we used $1\arcsec.6$ radius aperture to derive the color of star clusters. Therefore the missing flux due to the use of a fixed aperture photometry is negligible.
The star clusters in IC 10 appear as extended sources in the images so that we need to apply different aperture correction depending on the cluster size.
We adopted the aperture correction method that was applied by \citet{Hwa08} and \citet{Lim13}.
This method uses the correlation between the slope of light profiles and the calculated aperture correction values.
We calculated the slope of light profiles using magnitudes of 6 pixel and 10 pixel apertures.
The aperture correction values were calculated from the magnitude difference of 6 pixel and 20 pixel apertures derived for the isolated and bright star clusters.
We restrict the maximum correction values as 0.6 mag to avoid over correction.
The aperture correction can be applied to all $UBVRI$ magnitudes, but the aperture correction for each filter may increase color errors due to their uncertainties. In addition, the FWHM values for $UBVRI$ band images are similar ($0\arcsec.9$-$1\arcsec.0$). Therefore, we apply the aperture correction only for $V$ to derive the total $V$-band magnitudes, and using the same colors for SEDs.
We compared our photometry with that of \citet{Sha10}. There are differences in both $V$-band magnitudes and ($V-I$) colors between the two studies. The mean value of the difference in the $V$-band magnitudes (This study minus \citealp{Sha10}) is -0.25 with the standard deviation of 0.33, and the mean value of the difference in the ($V-I$) colors is 0.25 with the standard deviation of 0.17. These differences may be mainly due to different filter systems or aperture sizes. We used Johnson’s $UBVRI$ filter system, while they used the HST filter system. In addition, they used both $F555W$ and $F606W$ filters for $V$-band, but they did not distinguish $F555W$ and $F606W$ filters in their catalog. \citet{Sha10} did not provide any information on the aperture sizes so that we cannot discuss it more.
We estimated the sizes of the star clusters using 2-dimensional fitting tool ISHAPE \citep{Lar99} on the Suprime-cam $R$-band images.
The MOFFAT15 function is adopted to estimate the FWHM of the star clusters.
We derived effective radii of the star clusters using $r_{\rm eff} = 1.13 {\rm FWHM}$.
Several clusters were saturated on this image so that we could not obtain their sizes.
Finally, we derived the sizes of 58 star clusters among the cluster sample.
\subsection{Estimation of Star Cluster Age }
We derived ages, masses, and extinction values of the star clusters by comparing $UBVRI$ spectral energy distribution (SED) of the star clusters with simple stellar population (SSP) models given by \citep{Bru03}.
We assume a Salpeter initial stellar mass function (IMF) of $N(m)dm \varpropto m^{-2.35}$ with the mass range from $0.1 M_\odot$ to $100 M_\odot$.
The Salpeter IMF was often adopted in the previous studies (see \citealp{Lim13}) so that we adopt it for comparison with the previous studies.
The effect of assuming different IMFs is little for the optical colors of SSP models \citep{Lim13}, so that age distributions of the star clusters are not much affected by the different IMFs.
The metallicity of IC 10 is known to be similar to that of SMC ($Z=0.004$, \citet{Ski89,Gar90,Mag03}), but \citet{Hun01} suggested that the integrated $UVI$ colors of the star clusters in IC 10 follow the $Z=0.008$ evolutionary track better than the $Z=0.004$ track in the color-color diagrams.
Therefore, we assumed two values for the metallicity, $Z=0.004$ and $Z=0.008$ for SED fitting.
\citet{Kim09} suggested that the foreground reddening value of IC 10 is $E(B-V)=0.52\pm0.04$ and the total reddening value including the internal reddening of IC 10 is $E(B-V)=0.98\pm0.06$.
Therefore, we adopted the foreground reddening value, $E(B-V)=0.52$, and assigned the range of internal reddening value $E(B-V)=0.0$ to 0.6.
\section{Results}
\subsection{A Catalog of the Star Clusters in IC 10}
We made a catalog of 66 star clusters in IC 10 including 61 known star clusters \citep{Kar93,Hun01,Tik09} and 5 new star clusters found in this study, as listed in Table 1.
Table 1 includes positions, $UBVRI$ photometry, ages and reddening values of the star clusters.
We investigated optical properties of these star clusters using this catalog.
Figure \ref{spa} shows a spatial distribution of the star clusters in IC 10.
We divided the star cluster sample into two groups according to their position relative to R$_{25}$ of IC 10 ($R_{25} =3\arcmin.15$, $\approx 655$ pc):
the main body star clusters at $R<3\arcmin.15$
and the halo star clusters at $R \geq 3\arcmin.15$.
Several features are noted in this figure.
First, most star clusters are distributed in the main body, forming an elongated structure from south-east to north-west.
Second, four of the halo star clusters are located along the east direction,
and two are located at $6\arcmin$-$7\arcmin$ south east direction from the galaxy center.
Third, the most distant star cluster (ID No. 64) is at 9.8 arcmin ($\sim2$ kpc) from the galaxy center.
Number density profiles of the star clusters in IC 10 are shown in Figure \ref{np}.
The number density profile decreases slightly with the galactocentric distance
at $R=2\arcmin$, and drops but stays almost constant at $2\arcmin < R <4\arcmin$.
\begin{figure}
\epsscale{1.0} \plotone{fig3.eps}
\caption{Spatial distribution of the star clusters in IC 10. Crosses and filled circles show the known star clusters and the new star clusters, respectively. A large circle shows a main body region with radius of $R_{25}=3\arcmin.15$. Small circles represent the star clusters in the halo. The box with the grey scale image shows zoom-in of the main body of IC 10. The grey scale image is the LGS $R-$band image.
\label{spa}}
\end{figure}
\begin{figure}
\epsscale{1.0} \plotone{fig4.eps
\caption{Number density profiles of the star clusters in IC 10.
Open squares and error bars represent number density and its error of star clusters, respectively.
The vertical dashed line shows the location of R$_{25}$ for IC 10.
\label{np}}
\end{figure}
\subsection{Color-Magnitude Diagrams, Color-Color Diagrams and Sizes of the Star Clusters}
Figure \ref{cmd} displays color-magnitude diagrams (CMDs) and color histograms of the star clusters in IC 10.
The brightest star cluster in the main body is as bright as $V\sim17$ ($M_V\sim-9$ after correction of the foreground reddening), while that in the halo is as bright as $V\sim18.5$ ($M_V\sim-7.4$, after correction of the foreground reddening).
The color ranges of the star clusters are $-0.5\lesssim(U-B)\lesssim1.0$, $0.5\lesssim(B-V)\lesssim2.0$, and $0.5\lesssim(V-I)\lesssim3.0$.
The ($B-V$) color histogram of the main body star clusters shows two peaks at ($B-V$)$=1.1$ and ($B-V$)$=1.5$, while that of the halo star clusters has only one peak at ($B-V$)$=1.5$.
The star clusters in the halo are mostly redder and fainter than those in the main body.
Above result suggests that the halo star clusters are mostly older than those in the main body. However the reddening effect should be considered.
If we adopt the foreground extinction with $E(B-V)=0.52$ \citep{Kim09}, the reddening-corrected ($B-V$) colors of the halo star clusters are mostly $(B-V)_0 \leq 0.9$, which is similar to that of the metal-poor globular clusters in the Milky Way.
Figure \ref{ubvccd} shows ($U-B$)-($B-V$) color-color diagrams of the bright star clusters ($V<21$) in IC 10.
We compared the color-color diagrams with SSP models \citep{Bru03}.
Distinguishable features are as follows.
First, the color-color diagram of the main body star clusters follows the SSP model when the model was shifted with the total extinction value
($E(B-V)=0.98$), while the color-color diagram of the halo star clusters matches the SSP model when only the foreground extinction was corrected ($E(B-V)=0.52$).
Second, the star clusters in the main body follows the SSP model from young ages (1--10 Myr) to intermediate ages ($\sim1$ Gyr) when the foreground extinction and internal extinction are adjusted on the SSP model.
Third, the halo star clusters are concentrated on the old age area of the SSP model with the foreground extinction correction.
The color-color diagram of the star clusters gives a hint for their ages and extinctions. However, there is age-extinction degeneracy, which makes hard to derive ages and extinctions of the star clusters separately.
Therefore, we study ages of the star clusters quantitatively using the SED fit method in the next section.
Figure \,\ref{siz}(a) shows size distributions of the star clusters.
Effective radii of the star clusters are mostly ranging from $0.5$ pc to 8 pc, with a few larger than 8 pc.
The peak value of the size distribution is about 3.5 pc, which is similar to those of the star clusters in M51 and M82 \citep{Hwa08, Lim13}.
The sizes of the star clusters in the halo are similar to those of the star clusters in the main body.
There is little correlation between the magnitudes and sizes of the star clusters (Figure \,\ref{siz}(b)).
We also investigate the relation between the sizes and galactocentric distance (Figure \,\ref{siz}(c)), finding that the size of the star clusters changes little depending on the galactocentric distance.
\begin{figure}
\epsscale{1.0} \plotone{fig5.eps}
\caption{Color-magnitude diagrams of the star clusters in IC 10 (upper panel), and color histograms of the bright star clusters with $V<21$ mag (lower panel). Dashed and solid lines represent histograms of the star clusters in the main body and halo, respectively. Crosses and filled circles represent known star clusters and newly found star clusters, respectively. Open circles represent the halo star clusters. \label{cmd}}
\end{figure}
\begin{figure}
\epsscale{1.0} \plotone{fig6.eps}
\caption{($U-B$)-($B-V$) color-color diagrams of the bright star clusters with $V<21.0$ mag in IC 10.
Symbols are the same as in Figure \ref{cmd}.
Solid lines and dashed lines represent the simple stellar population (SSP) models for Z=0.004 and 0.008 \citep{Bru03}, respectively. They are shifted according to the reddening of $E(B-V)=0.52$ in (a), and $E(B-V)=0.98$ in (b). Arrows represent the reddening direction. Numbers along the SSP models represent log(age(yr)).
\label{ubvccd}}
\end{figure}
\begin{figure}
\epsscale{1.0} \plotone{fig7.eps
\caption{Sizes (effective radii) of the star clusters in IC 10. (a) Size distribution of the star clusters in IC 10. Solid, dashed, and dotted lines represent all, main body, and halo star clusters. (b) Size versus $V$ band magnitude of the star clusters. Crosses and open circles represent main body and halo star cluster, respectively. (c) Size of the star clusters versus galactocentric distance.
\label{siz}}
\end{figure}
\subsection{Ages of the Star Clusters}
The left panels in Figure \ref{age1} display age distributions of the star clusters in the main body and halo region. It shows two distinguishable features. First, the ages of the star clusters in the main body span a large range of ages from young ($\sim5$ Myr) to old ($>1$ Gyr), while the ages of the halo star clusters are mostly old ($>1$ Gyr). Second, the age distributions appear to be varying, depending on the metallicities, but their peak positions at 6 Myr, 100 Myr, and 4 Gyr are consistent. The youngest peak at 6 Myr is consistent with the epoch of the recent starburst in IC 10 \citep{Hun01,Vac07}. It is noted that this normal age distribution may have effects due to sample size and logarithmic binning. Therefore, we plotted the age distribution with the number of star clusters per unit time in the right panels of Figure \ref{age1}. It is noted that the two young age peaks are shown in both age distributions with logarithmic binning and normalized numbers. However, they also suffer from stochastic effects of low mass star clusters. We divide the star clusters into three groups according to their ages: young group ($\leq 10$ Myr), intermediate-age group ($10$ Myr --$1$ Gyr), and old group ($> 1$ Gyr).
We compared the ages of the star clusters with previous studies \citep{Hun01,Tik09}.
\citet{Hun01} obtained the ages of seven star clusters in the cluster sample, using the $(U-V)$--$(V-I)$ color-color diagram compared with theoretical SSP models from \citet{Lei99} for Z$=0.004$ and Z$=0.008$.
She derived ages of two resolved star clusters using CMDs of the resolved stars in each cluster. We compare the ages of these nine star clusters with our results in Figure \,\ref{comp1}(a).
The ages derived in this study are consistent with those of \citet{Hun01}.
\citet{Tik09} distinguished ages of the 56 star clusters by their morphology, and they categorized their cluster sample as young, intermediate-age, and old star clusters. We compare the ages of these star clusters with our results in Figure \,\ref{comp1}(b).
It shows that there is a weak correlation between the two results.
\citet{Sha10} provided the age information for four star clusters in their sample, and they are mostly old ($> \sim1$ Gyr). We compared the ages of these star clusters with those in this study, as plotted in Figure \,\ref{comp1}(a). Their age estimates for these star clusters are consistent with those in this study.
Figure \,\ref{mass} shows the masses versus the ages of the star clusters.
The masses of the star clusters in IC 10 are ranging from $\sim 10^{2.5} M_{\odot}$ to $\sim 10^{6.0} M_{\odot}$.
The lower limit of the masses of the star clusters is due to the survey limit of this study.
Young star clusters ($\leq10$ Myr) are mostly less massive than $10^4 M_{\odot}$, while the old star clusters ($>1$ Gyr) are as massive as $\sim 10^5 M_{\odot}$.
The masses of the old star clusters are similar to those of typical globular clusters in the Milky Way, and the mass range of the young star clusters is similar to that of the young open clusters in the Milky Way.
It is noted that young star clusters have low masses (M$\lesssim10^4$M$_\odot$), so stochastic effects can affect the age estimation of these clusters \citep{Mai09,And13}. This stochastic effect increases errors of age estimation for the range between 10 Myr and 100 Myr \citep{And13}. Therefore, we should be careful to study the star clusters with this range.
Figure \ref{spaage2} displays spatial distributions of the star clusters with different ages (young, $\leq10$ Myr; intermediate-age, 10 Myr -- 1 Gyr; and old, $>1$ Gyr).
We compared these spatial distributions with the HI map \citep{van02}, H$\alpha$ map \citep{Mas07}, and C star density map \citep{Dem04}
(contours in the figure).
The H$\alpha$ map represents recent star-forming regions, while the map of C star indicates the spatial distribution of relatively old stars ($\sim1$ Gyr).
The peak position of the young star clusters ($\leq10$ Myr) is consistent with that of the H$\alpha$ emission, which is close to the HI peak position.
This indicates that the recent starburst took place in this region.
The intermediate-age star clusters (10 Myr -- 1 Gyr) are distributed in two groups around the main body of IC 10. One is close to, but about one arcmin from the peak position of the H$\alpha$ emission. The other is about four arcmin from the peak position of the H$\alpha$ emission.
The old star clusters ($>1$ Gyr) are not only distributed in the main body of IC 10 but also in the halo region.
The old star clusters in the main body shows relatively loose concentration compared with the young star clusters, and their peak position is roughly
consistent with the galaxy center.
The spatial distribution of the old star clusters in the halo region is very asymmetric, while that of the C star is approximately circular.
It is noted the the spatial distribution of the C stars is more extended ($15\arcmin$) than that of the star clusters ($10\arcmin$).
It is expected that these may be more star clusters outside our survey area.
\begin{figure}
\epsscale{1.0} \plottwo{fig8a.eps}{fig8b.eps}
\caption{$left$: Age distributions of the main body (a,c) and halo (b,d) star clusters in IC 10 derived for Z$=0.004$ (a,b) and Z$=0.008$ (c,d), respectively. Solid line histograms and hatched histograms represent the age distributions of all star clusters and those with good fits, respectively. $right$: Age distributions of the main body star clusters with parameters of dN/dt derived for Z$=0.004$ (e) and Z$=0.008$ (f), respectively. Pluses and open squares represent all ages of star clusters and well fitted ages of star clusters. Error bars at the bottom of figures show the sizes of the bins.
\label{age1}}
\end{figure}
\begin{figure}
\epsscale{1.0} \plotone{fig9.eps}
\caption{Comparison of the star cluster ages with previous studies. The left column shows ages from this study versus ages from \citet{Hun01} and \citet{Sha10}, and the right column displays the ages from this study versus the ages from \citet{Tik09}. The top row and bottom row show the ages of the star clusters in this study derived for Z$=0.004$ and Z$=0.008$, respectively.
\label{comp1}}
\end{figure}
\begin{figure}
\epsscale{1.0} \plotone{fig10.eps}
\caption{Mass and age relations of the main body (pluses) and halo (circles) star clusters derived for Z$=0.004$ (a) and Z$=0.008$ (b), respectively. Solid lines represent the magnitude limit ($R=21.0$) of this study.
\label{mass}}
\end{figure}
\begin{figure}
\epsscale{1.0} \plotone{fig11.eps}
\caption{Spatial distribution of the star clusters with different ages in IC 10 derived for Z$=0.008$. Contours in left, middle, and right columns represent the HI map \citep{van02}, H$\alpha$ map \citep{Mas07}, and number density contour map of C stars \citep{Dem04}, respectively.
\label{spaage2}}
\end{figure}
\section{Discussion}
\subsection{Super Star Clusters in IC 10?}
\citet{Hun01} detected 13 star clusters and OB associations from the HST/WFPC2 images covering about a half of the main body of IC 10, finding that the brightest of which (ID 4-1 in her study, ID 54 in this study) has an absolute magnitude of $M_V \approx-10.0$. \citet{Hun01} pointed out that ID 54 is an OB association based on its low stellar number density. From this she pointed out that there are no Super Star Cluster (SSC) in her field. Although we increased the sample size of the star clusters in IC 10, covering a much larger field than that used in \citet{Hun01}, the cluster ID 54 remains to be the brightest in our sample with $M_V \approx -10.6$, which is only 0.6 mag brighter than the value given by \citet{Hun01}. However, this star cluster ID 54 may not be an OB association based on new stellar number density.
\citet{Hun01} derived the stellar number density of this star cluster from the number of resolved stars within the radius of $3\arcsec.09$ in the WF chips of HST/WFPC2 (see her Figure 14) is comparable with those of OB associations in the Milky Way and the Large Magellanic Cloud. However, this value for the stellar number density should be increased for the following reasons. First, the spatial resolution of WF chips \citet{Hun01} used is limited to deblend stars in the star clusters. We checked the F606W and F814W images obtained with HST/ACS WFC (ID: 9683, PI:F. Bauer) that has twice higher spatial resolution than WF chip in WFPC2 to test the effect of the spatial resolution. We derived photometry of resolved stars on the HST/ACS images using ALLSTAR/DAOPHOT packages in IRAF. Photometric zeropoints for each band is obtained from HST/ACS website\footnote{http://www.stsci.edu/hst/acs/analysis/zeropoints/zpt.py}, and the HST/ACS filter systems are transformed to the Johnson system using the relation in \citet{Sir05}. We found 23 stars with $M_V \leq -4$ in the cluster ID 54 using the distance and reddening as those in \citet{Hun01}, and 26 stars with $M_V \leq -4$ if we adopt the distance and reddening as adopted in this study. With this value we calculate the stellar number density, 0.07 stars per square parsec. This number of the resolved stars in the cluster is about twice larger than that from \citet{Hun01}, increasing the value of the star number density by a factor of two. This shows that the star number density of cluster ID 54 is much larger than those of typical OB associations in other galaxies (see Figure 14 in \citet{Hun01}), indicating that this cluster is not an OB association but an SSC.
We plotted this SSC on the relation between $V-$band absolute magnitude of the brightest cluster and star formation rate of the host galaxy in Figure \ref{sfrmv} including the results from previous studies \citep{Lar02,Bas08,Ada11,Coo12}. The dashed line in the figure is a linear fit of the samples from \citet{Lar02} given by \citet{Wei04}. If we include ID 54 in our sample as a star cluster, then IC 10 follows well this relation. This result suggests that the masses of the most massive star clusters in galaxies are correlated to the star formation rate of their host galaxies. However, \citet{Ada11} pointed out that the brightest star clusters in BCDs are located about 1.0 magnitude above this relation due to their high cluster formation efficiency (CFE, defined as cluster formation rate/star formation rate by \citealp{Bas08}). IC 10 is known as a BCD, but it follows well the relation between the brightest star clusters and the star formation rate of the host galaxies. Therefore, we need to check the CFE of IC 10.
Figure \ref{ssfr} shows CFEs versus surface star formation rates of several galaxies from previous studies \citep{God10,Ada11,Sil11,Ann11,Coo12}. CFE is defined as cluster formation rate/star formation rate, where the cluster formation rate is the total star cluster mass divided by the time interval of the experiment (10 Myr in this study). The total cluster mass is estimated as follows. First, we summed star cluster masses that are more massive than $10^3 M_{\odot}$. Second, we assumed the cluster mass function as a power-law function, $N(m)dm \varpropto m^{\beta}$ with $\beta=-2$, and integrated this cluster mass function from $10^3 M_{\odot}$ to $10^2 M_{\odot}$ to derive a value of the total mass for the low mass clusters. Finally, we summed these two values to get a value of the total cluster mass, which was used in this study. Dwarf galaxy samples from \citet{Coo12} are diverse in the relation, and IC 10 also increases the dispersion of this relation. It may be caused by the large error of CFE due to a small number of star clusters. \citet{Coo12} found that the scatters mostly come from the stochastic effect, but several galaxies show significant deviations exceeding the stochastic scatter. The result of IC 10 increases the scatter in the relation. It is noted that IC 10 has the highest surface star formation rate among the outliers in the relation between CFE and surface star formation rate. However, it is hard to constrain a reason for this result of IC 10 due to the limit of our data. Further studies are needed to understand why IC 10 has lower CFE compared with that of other star forming galaxies.
\begin{figure}
\epsscale{1.0} \plotone{fig12.eps}
\caption{Log star formation rate versus V-band absolute magnitude of the brightest star cluster for nearby galaxies from previous studies \citep{Lar02,Bas08,Ada11,Coo12}. A dashed line shows the fitting line from \citet{Wei04}. The filled circle represents IC 10. The internal and foreground extinctions are corrected for the all V-band absolute magnitudes of the brightest star clusters including that of IC 10 in this figure.
\label{sfrmv}}
\end{figure}
\begin{figure}
\epsscale{1.0} \plotone{fig13.eps}
\caption{Star cluster formation efficiency versus log surface star formation rate. The filled circle displays IC 10. The other samples from previous studies \citep{God10,Ada11,Sil11,Ann11,Coo12}. A dashed line shows the fitting result of \citet{God10}. The theoretical expectation of \citet{Kru12} is indicated with a dotted line.
\label{ssfr}}
\end{figure}
\subsection{The Halo of IC 10}
IC 10 has an extended structure outside the main body, including HI clouds, old stars (C stars and RGB stars), planetary nebulae, and star clusters (\citealp{Wil98,Dem04,San10,Gon12}, and this study).
This extended structure is order of magnitude larger ($r\approx30\arcmin$) than the optical main body ($R_{25} = 3\arcmin.15$).
The extended structure has mostly old stellar populations including old stars and star clusters, while the main body is dominated by young stellar populations.
This indicates that IC 10 has an old halo, embedding a disk of the main body.
This kind of old stellar halo
has been found in several other dwarf irregular galaxies (e.g. NGC 6822 \citep{Lee05, Hwa11}, Large Magellanic Cloud \citep{Min03}, Leo A \citep{Van04}, NGC 1569, and NGC 4449 \citep{Rys11}).
The formation of stellar halos in large and massive galaxies including the Milky Way has been explained in terms of hierarchical merging and accretion \citep{Whi78,Sea78}. For dwarf galaxies, two scenarios were suggested to explain the formation of halos: merging between gas-rich dwarf irregular galaxies \citep{Bek08a} or early star formation activity with supernova feedback \citep{Sti09}. These two scenarios can explain reasonably the existence of extended old populations and HI clouds, but which of the two is more effective is not yet clear.
The linear alignment of the halo star clusters in IC 10 may provide a hint for understanding the halo formation. It is well known that halo globular clusters in giant elliptical galaxies are distributed with a spheroidal shape (e.g \citealp{Lee03}). In addition, the Large Magellanic Cloud that is a dwarf galaxy similar to IC 10 does not show any sub structures of halo globular clusters (e.g. \citealp{Bic08}). Therefore, the spheroidal structure has been regarded as a normal spatial distribution of globular clusters in the halo.
However, the linear structures or substructures of the halo star clusters are found in several other galaxies. NGC 6822, a dwarf irregular galaxy similar to IC 10 has a linear structure of halo globular clusters \citep{Hwa11}. It is suggested that the halo globular clusters in NGC 6822 may have been accreted in the halo \citep{Hwa14}. The anisotropic spatial distributions of satellite systems were also found in the Milky Way \citep{Hol69,Zar97} and M31 \citep{Iba13}. It has been interpreted in terms of the accretion of satellite galaxies along filaments (e.g. \citealp{Kne04}). A recent study of halo globular clusters in M31 show that the globular cluster system is aligned along the tidal streams, suggesting that the globular clusters may be remnants of recently accreted dwarf galaxies into M31 \citep{Vel14}. Several studies suggested that this kind of accretion can happen in dwarf galaxies (e.g. NGC 4449 \citep{Str12,Mar12}, NGC 6822 \citep{Hwa14}). If we adopt this interpretation to IC 10, the halo star clusters in IC 10 might be metal-poor globular clusters accreted along filaments.
\subsection{The Origin of the Starbursts }
The starburst phenomenon in BCDs such as IC 10 has been a mystery because it is not easy to understand how low mass galaxies such as BCDs can keep gas long and show
strong starbursts later.
Several processes were proposed for explaining the origin of the starbursts in BCDs: in-spiraling gas clump \citep{Elm12}, gas cloud collision \citep{Gor81}, mergers \citep{Ost01,Bek08a,Bek08b}, galactic winds \citep{Ost03}, and tidal interactions \citep{van98}.
We will discuss the origin of the starbursts in IC 10 using the observational results
found in this study and previous studies.
To explain complex structure and kinematics of HI gas in IC 10,
\citet{Wil98} suggested that the recent starburst started a few $10^7$ yr ago, and is still going on, as the gas keeps infalling from the reservoir in the outer region.
Recently \citet{Nid13} found a new interesting structure from deep HI observation,
a long (5.2 kpc) HI extension in the north-west region.
They pointed out that this HI extension is not related with the recent starburst,
but may be with an interaction with another dwarf galaxy before this starburst.
However, they did not mention the epoch of the dwarf galaxy interaction.
The velocity of the HI extension with respect to IC 10 is about 65 km s$^{-1}$.
Assuming that the HI extension move with this velocity in the projected plane, we estimate that it would take roughly 300 Myr for this to move from IC 10 to the current position (18 kpc).
We found from the age distribution of the clusters that there were two starbursts,
one at $6$ Myr and another at $100$ Myr.
The first burst at $100$ Myr produced star clusters mostly in two regions, one in the central star-forming region and another in the north-west star-forming region.
The second burst at 6 Myr formed star clusters mostly in the south-east star-forming region.
It is noted that the separation of these two age peaks can be caused by the limit of SED fitting method. Theoretical SSP models have a degeneracy around log$(age)=7$, so the SED fitting method does not recover well around the age of log$(age)=7$ (e.g. \citealp{Mai09}). In addition, many star clusters in IC 10 are less massive than $10^4$ M$_{\odot}$ , so they also have the stochastic effect of IMF sampling. Because of the above reason, there is a possibility that these two peaks may be originally from one broad distribution. However, there is still a possibility that these two peaks can be genuine separated peaks. The spatial distributions of these two peaks are different, and these peaks are shown in both age distributions with normal histogram and dN/dt. The color-color diagram of star clusters in IC 10 also suggests the presence of a gap in star cluster formation around 10 Myr (Figure \ref{ubvccd}b). In addition, the ages of these peaks are consistent with the epochs of recent starburst and past interaction suggested by previous studies. Therefore, we regard these two age peaks as probable separated peaks.
Star cluster disruption can also affect the age distribution of star clusters. There are two models for the cluster disruption that are mass dependent disruption (MDD, \citealp{Lam05}) and mass independent disruption (MID, \citealp{Whi07}) models. It is hard to constrain with our sample which cluster disruption model works better for IC10, and how much star clusters were disrupted in IC10. However, both disruption models may not affect our interpretation for the age distribution because of the following reasons. If the time scale of cluster disruption is long ($\geq 10^8$yr), then the cluster disruption has been going gradually. Therefore, the age distribution may not change drastically. If the time scale of cluster disruption is short ($\leq 10^7$ yr), then star clusters older than 10 Myr remain to be a small fraction of them (e.g. \citealp{Cha10}). In this case, the starburst at 100 Myr may be stronger than the starburst at 6 Myr, but the age peaks may not disappear. Therefore, cluster disruption may not remove the peaks in the age distribution.
The studies of HI gas in IC 10 and this study have links to understand the recent starbursts.
The first starburst epoch found in this study corresponds to the encounter epoch of the HI extension estimated with its radial velocity and projected distance to IC 10.
In addition, the location of the star clusters belonging to the first starburst is closer to the HI extension than those of other star clusters.
Therefore, it is concluded that the HI extension was formed when IC 10 interacted with another gas-rich dwarf galaxy $\sim$100 Myr ago.
It is not clear how the first starburst propagates to the second starburst in this study.
\citet{Wil98} suggested that the stellar wind from massive stars formed at the first starburst in this study triggers the second starburst by compressing the gas.
However, the time scale that they mentioned is a little different with the time interval between the first and second starbursts in this study.
\citet{Bek08b} mentioned that the dwarf-dwarf interaction can make a BCD and there may be two starbursts in this case, though he did not explain the mechanism for two starbursts.
From above, we suggest a scenario as follows.
Two gas-rich dwarf galaxies tidally interacted at $\sim$100 Myr ago,
and the remnant of this encounter traveled to the northwest, leaving a long HI extension \citep{Bek08b,Nid13}.
At this time, a starburst occurred in the north-west star forming region, forming several star clusters.
Massive stars in these star clusters affected the interstellar medium through stellar winds or supernova explosion.
Then the second starburst happened at 6 Myr ago, and it is the recent starburst of IC 10.
\section{Summary}
We presented a photometric study of 66 star clusters in IC 10 including five new star clusters found in this study.
The integrated $UBVRI$ magnitudes of these star clusters were obtained, and ages of them were estimated by the SED fitting method.
Our primary results are summarized below.
\begin{enumerate}
\item The number density profiles of the star clusters show a break at $R_{25}$ and we divided the star clusters in IC 10 into two groups: main body clusters at $R<R_{25}$ and halo star clusters at $R\geq R_{25}$.
All five star clusters found in this study are halo star clusters.
The halo star clusters in IC 10 shows a linear structure, and this structure seems to be correlated with the extended HI disk.
The halo star clusters are mostly red ($(B-V)>1.2$) and old ($>1$ Gyr), while the disk star clusters have various color ranges ($0.5\lesssim (B-V) \lesssim 2.0$) and age ranges (5 Myr -- 10 Gyr).
\item The age distribution of the star clusters in IC 10 shows three peaks at $\sim$6 Myr, $\sim$100 Myr, and $\sim$4 Gyr. The youngest peak is consistent with the recent starburst, and the intermediate-age peak suggests the existence of another starburst.
\item The spatial distribution of the young star clusters ($<10$ Myr) are well consistent with H$\alpha$ emission and concentrated on the small region in the main body, while the old star clusters ($>1$ Gyr) are distributed in a wider area than the disk.
These young star clusters can be a result of the recent starburst event.
It seems that the spatial distribution of the intermediate-age star clusters is not correlated with the structures of HI gas, H$\alpha$ emission, and C stars.
However, they seem to be related with past merger or tidal interaction with HI cloud at the northwest.
\item One SSC is found in IC 10. The brightest star cluster (ID 54) in IC 10 is $M_V\approx-10.6$ after correction of foreground and internal reddenings.
It has relatively low stellar number density compared with those of typical SSCs.
\item The $(B-V)$ color-histograms and the ages of the halo star clusters suggests that most halo star clusters may be metal-poor old star clusters, and it similar to that of halo star clusters in other starburst galaxies (e.g. NGC 4449 and M82) and normal spiral galaxies (e.g. M31 and Milky Way).
The anisotropic spatial distribution of the halo star clusters and the existence of these old metal-poor halo star clusters suggest that the halo of IC 10 may be formed by accretions.
\item We found a group of intermediate-age star clusters with $\sim 100$ Myr, and it may be related with the HI extension at the northwest.
This HI extension might have interacted with IC 10 at $\sim$100 Myr ago, suggesting that the tidal interaction between two gas-rich dwarf galaxies occurred at $\sim$100 Myr ago.
This tidal interaction led to the starburst in IC 10 at $\sim$100 Myr ago, and the stellar feed back process might have caused the recent starburst in IC 10.
\end{enumerate}
\acknowledgments
We thank the anonymous referee for helpful comments which improved the original manuscript.
This work was supported by the National Research Foundation of Korea (NRF) grant
funded by the Korea Government (MSIP) (No.2013R1A2A2A05005120).
\makeatletter
|
\section{Introduction}
\label{sec:intro}
\resetsec
The revival of the debate in recent years in the broader community of
philosophers over the ontic status of spacetime can trace its roots,
in part, to its revival in the community of physicists.
\citeN{belot-gr-interp} and \citeN{belot-earman-presoc-qg}, for
instance, claim that philosophers ought to take the debate seriously
because many physicists do. I do not think that fact suffices as a
good reason for philosophers to take the debate as interesting, much
less even well posed, and so enter into it. The active work of
physicists on our best physical theories should provide the fodder for
the work of the philosopher of physics most of the time. Sometimes,
however, the physicists are confused or just mistaken, and it is then
our job to try to help set matters straight. I believe that is the
case here.\footnote{See
Curiel~\citeyear{curiel-modesty,curiel-gr-needs-no-interp} for
extensive arguments to this effect on closely related matters, and
for a defence of this claim as a fruitful philosophical attitude.}
Other philosophers in recent work have taken inspiration from the
traditional debates themselves. \citeN{maudlin-st-subs}, for
instance, after a \emph{pr\`ecis} of the debate in the 17th and 18th
centuries and Kant's attempt to sidestep it, concludes, ``[G]ranting
that the world is \emph{an sich} a spatiotemporal object, we must face
a fundamental problem: Are space and time entities in their own
right?'' In this paper, I dispute that ``must.''
A virtue of Maudlin's approach, which his work shares with that of
many other contemporary philosophers no matter their inspiration, is
the foundation of his arguments on the structures of our best physical
theories and the use of those structures to guide metaphysical
argument. I think the method falls short, however, in so far as it
treats those structures in abstraction from their uses in actual
scientific enterprises, both theoretical and experimental. This
lacuna leaves the debate merely formulaic, without real content, at
the mercy of clever sophistications without basis in real, empirically
grounded scientific knowledge in the fullest sense.
\citeN[p.~1]{stein94-struct-know} admirably sums up the situation as I
see it. I quote him at length, as he says it better than I
could\label{pg:stein-struc-know}:
\begin{quote}
[L]et me \ldots hazard a rough diagnosis of the reason why some
things that are (in my view) true, important, and obvious tend to
get lost sight of in our discussions. I think ``lost sight of'' is
the right phrase: it is a matter of perspective, of directions of
looking and lines of sight. As at an earlier time philosophy was
affected by a disease of system-building---the \emph{\'esprit de{}
syst\`eme} against which a revulsion set in toward the end of the
last century---so it has (I believe) in our own time been affected
by an excess of what might be called the \emph{\'esprit de{}
technique}\ldots: a tendency both to concentrate on such matters
of detail as allow of highly formal systematic treatment (which can
lead to the neglect of important matters on which sensible even if
vague things can be said), and (on the other hand), in treating
matters of the latter sort, to subject them to quasi-technical
elaboration beyond what, in the present state of knowledge, they can
profitably bear. [W]hat I have described can be characterized
rather precisely as a species of scholasticism\ldots. In so far as
the word ``scholasticism,'' in its application to medieval thought,
has a pejorative connotation, it refers to a tendency to develop
sterile technicalities---characterized by ingenuity out of relation
to fruitfulness; and to a tradition burdened by a large set of
standard counterposed doctrines, with stores of arguments and
counterarguments. In such a tradition, philosophical discussion
becomes something like a series of games of chess, in which moves
are largely drawn from a familiar repertoire, with occasional
strokes of originality---whose effect is to increase the repertoire
of known plays.
\end{quote}
In the spirit of Stein's diagnosis, rather than something formally
sophisticated I'm going to propose something crude and simple: in
order to try to avoid the sort of sterility that purely formal
technical elaboration can lead to, we should look at the way that
spacetime structures are used in practice to model real systems in
order to try to make progress on issues closely related to those
treated in the standard debate. For I do think that there are
important, deep questions that we can make progress on in the vicinity
of that debate, questions of the sort that Maxwell alludes to in the
passage I quoted as one of this paper's epigraphs. As Maxwell
intimates, however, in order for such questions to be investigated
profitably, they must be such as to support and stimulate ``the
investigation of nature.'' And that, I submit, can be accomplished
only when the questions bear on scientific knowledge in all its
guises, as theoretical comprehension and understanding, as evidential
warrant and interpretative tool in the attempt to assimilate novel
experimental results, as technical and practical expertise in the
design and performance of experiments, and as facility in the bringing
together of theory and experiment in such a way that each may
fruitfully inform the other.
To that end, in this paper I will argue that the way to find the
philosophically and scientifically fruitful gold in the metaphysical
dross is to formulate and address the questions in a way that
explicitly makes contact with both the theoretical and the
experimental aspects of our best current knowledge about the kinds of
physical system at issue. One way of trying to do that is to pose and
investigate the questions explicitly in the context of what I will
call an investigative framework: roughly speaking, a set of more or
less exactly articulated and fixed theoretical structures for the
modeling of physical systems, along with a family of experimental
practices and techniques suited to the investigation of the type of
systems the theoretical tools appropriately model, in the way the
theory actually models them. Different investigative frameworks, as I
show by constructive example, provide different natural criteria with
which to render determinate content to the question of the ontic
status of spacetime, with none privileged \emph{sub specie
{\ae}ternitatis} over any of the others. Those different criteria
yield different answers to the question, suitably formulated in the
given frameworks. This should not be surprising, I think. After all,
different sorts of scientific investigations naturally assume and rely
on different relations between individual spacetime points and
metrical (and other forms of spatiotemporal) structure, and it is
those relations that are supposed to serve as the criteria for
existence of individual spacetime points; the mathematical formalism
of the theory does not by itself fix a univocal relation with clear
\emph{physical} significance between points of the spacetime manifold
and geometrical structures, both local and global ones, that live on
the manifold. I therefore dispute not only the force of Maudlin's
``must,'' but even more the cogency of the demand itself, baldly
formulated.
I begin in \S\ref{sec:hole} with an examination of a popular argument,
the so-called Hole Argument, that seems to urge a form of
relationalism. I do this for two reasons. First, because advertence
to the argument has become something of a mannerism in the debate, it
must be confronted; I conclude that it has no bearing one way or
another on the issues the debate purports to address. Second, I
discuss it because it yields a useful schema for the production of
concrete criteria in the terms of which one can try to explicate the
difference between substantivalists and relationalists, such as it is.
I use that schema---whether the identification of spacetime points
must depend on the prior stipulation of metrical structure---to frame
the argument of the subsequent two sections of the paper. In each of
those two sections I make the schematic criterion concrete in the
context of a particular form of investigative framework so as to
construct two arguments with contrary conclusions, one in support of
something like relationalism and the other something like
substantivalism to show that one can make the debate concrete in any
of a number of precise, physically significant ways, none \emph{a
priori} privileged over the others, and that those ways will not in
general agree in their consequences.\footnote{I do not know of anyone
in the literature who adopts exactly the schematic criterion I
propose to found my two arguments. (Perhaps
Hoefer~\citeyearNP{hoefer-meta-st-subst,hoefer-abs-vs-rel-st} comes
the closest.) I use it because I think it captures the flavor of
the criteria that are often stipulated when one or the other
position is being argued for or against, \emph{viz}., schematically
speaking, that the question of the existence of spacetime points
boils down to the relation of those points to some fixed, underlying
geometrical structure, such as the metric. (See,
\emph{e}.\emph{g}., \citeNP{earman-world-enough},
\citeNP{butterfield-hole-truth},
Maudlin~\citeyearNP{maudlin-subs-st-ari-eins,maudlin-st-subs},
\citeNP{rynasiewicz-lessons-hole}, \citeNP{belot-rehab-rel},
\citeNP{dorato-subst-rel-struc-st-real},
\citeNP{huggett-reg-acct-rel-st},
Pooley~\citeyearNP{pooley-pts-parts-struc-real,pooley-subst-rel-approach-st},
\citeNP{belot-geom-poss}.) This is all I require for the overall
argument of the paper. I use this particular schema, moreover, as
only one example of the sort of criterion one could with some
justification rely on in this debate, not because I think it is
canonical or privileged in some way, but because it is popular and
has a lot to say for it \emph{prima facie}. My hope is that showing
how the debate breaks down when this particular criterion is used
will, at the least, strongly suggest that it would similarly break
down no matter what sort of purely formal criterion of that sort one
used.
DiSalle~\citeyear{disalle-dyn-indiscern-st,disalle-understand-st} is
a notable example of a contemporary philosopher who takes an
approach much more sympathetic to my own. (See
\citeNP{friedman-rvw-disalle} for a thoughtful discussion of
DiSalle's work.) Robert Geroch (in private conversation) is a
notable example of a contemporary physicist who does so.
\citeN{dorato-is-struc-real-rel-disguise} is an interesting case of
a philosopher who agrees with me that the contemporary debate is not
well posed, but thinks there is a best answer to a proper
reformulation of the debate. \citeN{rynasiewicz-abs-vs-rel-outmode}
agrees with me that the contemporary debate is not well posed, but
he uses arguments I would not endorse.}
The opposed arguments and contrary conclusions of
\S\S\ref{sec:limits}--\ref{sec:pointless}, in conjunction with the
dismissal of the Hole Argument, do not decisively refute the claim
that there is a single, canonical way to explicate the idea of a
spacetime point and so to enter into debate over the existence of such
a thing. As I urge in \S\ref{sec:subs-rel}, they strongly suggest it
is a question best settled in the context of a particular form of
investigation. The investigation itself in tandem with pragmatic
considerations and {\ae}sthetic predilections will guide the
investigator in settling the form of the question and so the search
for its answer. For a given spacetime theory, and even a given model
within the theory, depending on one's purposes and the tools one
allows oneself, either one can treat spacetime points as entities and
individuate and identify them \emph{a priori}, or one can in any of a
number of ways construct spacetime points as factitious, convenient
pseudo-entities, as it were. Nothing of intrinsic physical
significance hangs on the choice, and so \emph{a fortiori} science
cannot guide us if we attempt to choose \emph{sub specie
{\ae}ternitatis} between the alternatives---such a choice must
become, if anything, an exercise in scholastic metaphysics only.
In \S\ref{sec:embarass}, I extend the discussion to a host of other
types of spacetime structure, such as Killing fields and topological
invariants. The attempt to formulate criteria for the physicality of
such other structures adds weight to the conclusion that such
questions require concrete realization in the context of something
akin to real science in order to acquire substantive content. I
conclude in \S\ref{sec:valedict} with a brief attempt to show that the
arguments of this paper ramify into the debate between realists and
instrumentalists more generally, by dint, in part, of the picture of
science the arguments implicitly rely on. The overarching lesson I
draw is that metaphysical argumentation abstracted from the pragmatics
of the scientific enterprise as we know it---science as an actually
achieved state of knowledge and as an ongoing enterprise of
inquiry---is vain. Very little of real substance can be learned about
the nature of the physical world by studying only theoretical
structures in isolation from how they hook up to experimental
knowledge in real scientific practice, as is the endemic practice in
the current debate. In particular, tracking the alleged ontological
commitments of a theory based on an analysis of its formal structure
alone is not a viable approach to the issue, as we cannot know what
structures the theory provides have real physical significance, and
what sort of real physical significance they do have, unless we
understand how the theory is successfully applied in practice.
The constructions I found the arguments on require the use of advanced
mathematical machinery from the theory of general relativity. The
format of the paper does not allow for an introduction to most of it.
(For the interested reader, \citeNP{wald-gr} or
\citeNP{malament-fnds-gr-ngt}, for example, contains comprehensive
coverage of all material required.) I have tried to segregate it as
much as possible so that those who do not want to trudge through it
will not have to while still following the general argument. For
those who do want to skip most of the technical material, I recommend
the following: in \S\ref{sec:hole}, ignore the sketch of the Hole
Argument (the second and third paragraphs of the section), but read
the rest; in \S\ref{sec:limits}, read the first two paragraphs and the
last one; in \S\ref{sec:pointless}, read the first two paragraphs
(including definition~\ref{def:pointless-criterion}), and the final
two paragraphs. (The remainder of the paper should not pose strenuous
technical difficulties.) This course will convey almost the entirety
of my argument, bar supportive details the technical material purports
to provide.
\section{The Hole Argument}
\label{sec:hole}
\resetsec
In recent times, several physicists and philosophers have construed
Einstein's infamous Hole Argument so as to place it at the heart of
questions about the ontic status of spacetime points. Its lesson, so
claimed, is that one cannot identify spacetime points without reliance
on metrical structure, that there is no ``bare manifold of points'',
as it were, under the metric field.\footnote{See, \emph{e}.\emph{g}.,
\citeN{belot-gr-interp} and \citeN{gaul-rovelli-loopqg-diffeoinv}.
Einstein himself originally formulated the Hole Argument to
highlight what he regarded as problems of indeterminism for any
generally covariant theory. See
\citeN{einstein14-form-grund-allg-rel} and
\citeN{einstein-grossmann-14-kovar-feldgleich} for two versions of
the original argument, Norton
\citeyear{norton-how-found-efe,norton-gen-covar} for historical and
critical discussion, and \citeN{earman-norton-hole-arg} for the
introduction of the argument to the contemporary philosophical
debate.} In the contemporary literature, the debate is often posed
thus: should the manifold $\mathcal{M}$ by itself or the ordered pair
$(\mathcal{M}, \, g_{ab})$ be properly construed as the represention
of ``physical spacetime''?
This, in brief, is the argument. Fix a spacetime model $(\mathcal{M},
\, g_{ab})$.\footnote{I may seem to be biasing the argument already,
by demanding that one fix both a manifold and a metric to fix a
model of spacetime. In fact, though, by ``model of spacetime''
here, I explicitly mean ``manifold \emph{cum} metrical structure'',
irrespective of how the debate over what really represents physical
spacetime resolves itself, so there is no bias here.} For ease of
exposition, we stipulate that the spacetime be globally hyperbolic,
and so possesses a global Cauchy surface, $\Sigma$. (We could do
without this condition at the cost of unnecessary technical details.)
Say that we know the metric tensor on $\Sigma$ and on the entire
region of spacetime to its causal past, $J^- [\Sigma]$. (Note that
$J^- [\Sigma]$ contains $\Sigma$.) It is known that this forms a well
set Cauchy problem, and so there is a solution to the Einstein field
equation that uniquely extends $g_{ab}$ on $J^- [\Sigma]$ to a metric
tensor on all of $\mathcal{M}$, yielding the original spacetime we
fixed.\footnote{This is not, strictly speaking, accurate. If no
restrictions are placed on the form of the metric, then in general
the initial-value problem is not well set. Indeed, even a few known
``physical'' solutions to the Einstein field equation possess no
well set initial-value formulation, for example those representing
homogeneous dust and some types of perfect fluid. (See,
\emph{e}.\emph{g}., \citeNP{geroch-pde}.) We can ignore these
technicalities for our purposes, though it may raise a serious
problem for those who worry about indeterminism in the theory, one
which, to the best of my knowledge, has not been addressed in the
literature.} In particular, the solution to the Cauchy problem
fixes the metric on the region to the causal future of $\Sigma$, $J^+
[\Sigma]$. Now, let $\phi$ be a diffeomorphism that is the identity
on $J^- [\Sigma]$ and smoothly becomes non-trivial on $J^+ [\Sigma] -
\Sigma$. No matter what else one takes the significance of the
diffeomorphism invariance of general relativity to be, at a minimum it
must include the proposition that the application of a diffeomorphism
to a solution of the Einstein field equation yields another, possibly
distinct solution. Apply $\phi$ to $g_{ab}$ (but not to $M$ itself);
this yields a seemingly different metric---a different ``physical
state of the gravitational field''---on $J^+ [\Sigma] - \Sigma$, in
the sense that the same points of $J^+ [\Sigma] - \Sigma$ now carry
(in general) a different value for the metric. This is the crux of
the issue, that the diffeomorphism applied to the metric has yielded a
different tensor field in the sense that the same points of the
spacetime manifold now carry a different metric tensor than before.
We now face a dilemma, the argument continues
(\citeNP{earman-norton-hole-arg}): we can either hold that the
fixation of the metric on $J^- [\Sigma]$ does not determine the metric
on $J^+ [\Sigma] - \Sigma$, a radical form of seeming indeterminism,
or else we can conclude that spacetime points in some sense have no
identifiability or existence or what-have-you independent of the prior
fixation of the metric tensor. The argument concludes that the denial
of the independent existence of spacetime points is the lesser of the
evils (or, depending on one's viewpoint, the greater of the
goods).\footnote{Though it does not seem to be recognized in the
literature, there are two different versions of the argument used by
different investigators. The one I rehearse here can be thought of,
in a sense, as a generalization of the other. The more specialized
form, which Einstein himself formulated and used, assumes that
spacetime has a region of compact closure, the nominal hole, which
is devoid of ponderable matter (\emph{i}.\emph{e}., in which the
stress-energy tensor vanishes) though it itself is surrounded by a
region of non-trivial stress-energy; the diffeomorphism is then
stipulated to vanish everywhere except in the hole, and the argument
goes more or less as in the general case, with the emendation that
now it is the distribution of ponderable matter that does not
suffice to fix the physical state of the gravitational field.
(\citeNP{earman-world-enough}, for example, uses the more general
argument, whereas \citeNP{stachel-mean-gen-covar-hole} uses the more
specialized form.) I think the specialized form of the argument
introduces a dangerously misleading red herring, \emph{viz}.,
physical differences between regions of spacetime with non-vanishing
stress-energy and those without. There seems to me no principled
way within the context of the theory itself to distinguish between
such regions in a way that bears on metaphysical or ontological
issues. One of the regions, that with stress-energy, has
non-trivial Ricci curvature; the other does not, though it may have
non-trivial Weyl curvature. That difference by itself, the only one
formulable strictly based on the theory, can tell us nothing in the
abstract about the ontic status of the spacetime manifold. The
introduction of the difference seems rather to bespeak an old
prejudice that material sources should suffice to determine the
physical state of associated fields, but this is not true even in
classical Maxwell theory. Indeed, the issue seems much less of a
problem in general relativity, for in the case of the Maxwell field
we cannot determine a \emph{physically} unique solution without
imposing boundary conditions; otherwise, we are always free to add a
field with vanishing divergence and curl to a solution to yield
another that will have different physical effects on charged bodies.
In general relativity, one does not need to do anything of the sort
to determine a physically unique solution, so long as the initial
data is well behaved in the first place. (See, \emph{e}.\emph{g}.,
\citeNP[ch.~10, pp.~243--268]{wald-gr}.)}
I want to make a crude and simple proposal, for it seems to me that
the debate has lost sight of a crude and simple, and yet fundamentally
important, fact: just because the mathematical apparatus of a theory
appears to admit particular mathematical manipulations does not
\emph{eo ipso} mean that those manipulations admit of physically
significant interpretation, much less that those apparently
mathematical manipulations are even coherent in and of
themselves.\footnote{\citeN{weatherall-hole}, whose conclusions I
endorse, argues vigorously that the sort of manipulation employed in
the standard form of the Hole Argument does not make even
mathematical sense. For the sake of argument, however, I will
assume here that it does. (If one likes, one can take that
assumption as being in the service of a \emph{reductio}.)} One has
the mathematical structure of the theory; one is not free to do
whatever it is one wants with that formalism and then claim, with no
foundation in practice, that what one has done has physical
import.\footnote{\citeN[p.~149]{stachel-mean-gen-covar-hole} neatly
describes the current attitude in the literature towards mathematics
in physical theories:
\begin{quote}
A current trend among some philosophers of science is toward what I
will call ``the fetishism of mathematics.'' By this I mean the
tendency to assume that all the mathematical elements introduced in
the formalization of a physical theory must necessarily correspond
to something meaningful in the physical theory and, even more, in
the world that the physical theory purports to help us understand.
\end{quote}} Once one has the mathematical formalism in hand, one must
determine what one is allowed to do with it, ``allowed'' in the sense
that what one does respects the way that the formalism actually
represents physical systems. A simple example will help explain what
I mean: adding 3-vectors representing spatial points in Newtonian
mechanics. This shows the need for an investigative context for the
fixing of what counts as admissible manipulations of the mathematical
formalism, for as a physical operation adding spatial points makes no
sense---there is no sense to be had from the idea of linearly
superposing two different spatial points in Newtonian theory as a
representation of a physical state of affairs. For the purposes of
computing factitious quantities such as the center of mass, however,
it does make sense, though, again, not as an operation that has a
physical correlate in the world.
General relativity, in its usual incarnation, is formulated with the
use of differential manifolds with pseudo-Riemannian metrics. It does
not \emph{ipso facto} follow that every well formed mathematical
operation one can perform on a manifold with such a metric has
physical significance. It arguably makes mathematical sense to apply
a diffeomorphism of the manifold to the metric only, and not to the
underlying manifold at the same time. That fact by itself does not
imbue the operation with physical significance. It is exactly
considerations such as the Hole Argument highlights that show how
diffeomorphisms ought to be applied to solutions of the Einstein field
equation so as to have physical significance. When one applies a
diffeomorphism, one must apply it to both the manifold and the metric.
As I shall argue, no other procedure has physical content.\footnote{If
one adopts a certain definition of a differential manifold,
\emph{viz}., that it is an equivalence class of ``diffeomorphic
presentations'', then one will say that the proposed operation does
not make even purely mathematical sense. (\citeNP{weatherall-hole}
comes to the same conclusion, based on different, but related,
arguments.) $\mathbb{S}^2$, for example, can presented as a certain
submanifold of $\mathbb{R}^3$, or as a certain submanifold of a
17-dimensional hyperbolid, or simply as a manifold in its own right;
$\mathbb{S}^2 \times \mathbb{R}^2$ can be presented, as here, as a
direct product of manifolds, or as $\mathbb{R}^4$ with a line
removed; and so on. In this case, ``pushing tensors around on the
manifold by a diffeomorphism without also pushing the points
around'', as required by the Hole Argument, is not an unambiguous
notion, for strictly speaking manifold points are defined only up to
diffeomorphism in the first place. I do in fact accept the
definition of a differential manifold as an equivalence class, but I
am trying to be as charitable as possible to the proponents of the
debate and the arguments standardly deployed in its carrying out, so
I am willing to grant for the sake of argument that the required
manipulations make mathematical sense. In any event, it is not only
philosophers who explicitly attempt to manipulate manifolds and
objects in them, in the context of general relativity, in the way
the Hole Argument requires; see, \emph{e}.\emph{g}.,
\citeN{pons-salisbury-time-gencovar-theors-komar-bergmann} for
physicists explicitly doing so.}
The Hole Argument is obviated by the fact that the application of
$\phi$ to the manifold \emph{cum} metric results only in a different
presentation of the same intrinsic metrical structure. All observers,
no matter which diffeomorphic presentation of the manifold \emph{cum}
metric they use in their respective models, will agree on what is of
intrinsic physical significance in the possible interaction of
physical systems. (Are those two bodies in physical contact? Is heat
flowing from this one to that or vice-versa? Can a light-signal be
sent from this to that? Is gravitational radiation present? And so
on.) There is no logical or physical contradiction in taking
different diffeomorphic presentations of the manifold \emph{cum}
metric each as the representation of the same physical structure. One
must simply stipulate that, in the context of general relativity, the
application of a diffeomorphism to the metric is a \emph{physically}
well defined procedure only when one also applies it to the (given
presentation of the) manifold itself. The worry about determinism
thus evaporates, doing away with the dilemma. How one then goes on to
try to characterize the ontic nature of spacetime points, if that is
the sort of thing one is into, may be influenced by this restriction
on the applicability of diffeomorphisms to solutions of the Einstein
field equation, or it may not. The point of fundamental importance is
that this restriction results from \emph{both} pragmatic and semantic
considerations about the way that one may employ the formal apparatus
the theory provides so as to respect how solutions to the Einstein
field equation represent physically possible spacetimes in
practice---how it is that the formal structures of the theory acquire
real physical meaning.
In sum, I do not see why the Hole Argument drives one to conclude that
one should or should not attribute some form of existence to spacetime
points independent of the metrical structure. There is no logical or
physical contradiction, for example, in taking the image of a point
under the action of $\phi$ to be ``the same spacetime point'' as its
pre-image, as depicted in a different presentation of spacetime,
irrespective of metrical structure. In this case, a spacetime point
would be something like an equivalence class of ordinary mathematical
points under the relation of being related by a diffeomorphism. An
exact formulation that avoids having this idea collapse into
triviality---given any finite number of points on a manifold, there is
a diffeomorphism that maps those points onto any permutation of them,
which seems to leave one with a single equivalence class containing
all points---requires some refinement. One could do something like
the following: a spacetime point is a physical entity that one can
uniquely, or at least adequately and reliably, individuate and
identify by what is of intrinsic physical significance at the physical
event that occupies it, no matter the diffeomorphic presentation of
the manifold of events; it is an entity, in other words, individuated
and identified by the equivalence class of physical events under
diffeomorphic presentation.\footnote{Such a characterization would not
necessarily rely on metrical structure at a point since, in general,
one needs to fix the physical state on an open neighborhood of a
point in order to fix the metric structure at that point by way of
the Einstein field equation; one cannot solve the Einstein field
equation ``point by point'', as it were. The easiest way to see
this is to note the non-uniqueness of vacuum solutions. This is
intimately bound up with the fact that the value of the
stress-energy tensor at a point does not determine the value of the
Weyl tensor (conformal structure) at that point.} If one wants to
respond that bare spacetime points \emph{per se} even with what are
tantamount to unique labels attached (if the spacetime is not overly
symmetric) are dependent on physical phenomena under this definition
and inobservable to boot, and so unnecessary in the formulation of
physical theory, so as to conclude that they have no independent
metaphysical existence of one sort or another, I would not necessarily
disagree, but neither should I think that one requires the Hole
Argument to make the point, for the game of the Hole Argument is that
one cannot identify spacetime points in the absence of metrical
structure. One need not invoke or rely on metrical structure to make
the sort of identification I suggest, as I will show by construction
in \S\ref{sec:pointless}.
The basis for my rejection of the Hole Argument, that a proper
understanding of diffeomorphism invariance and the way to properly
implement it as a formal procedure vitiates it, rests on a deeper
point. I think the most unproblematic and uncontroversial claim one
can make about diffeomorphic freedom is that it embodies an
irremediable mathematical arbitrariness in the apparatus provided by
general relativity for the modeling of physical systems: the choice of
the presentation of the spacetime manifold and metric one uses to
model a physical system is fixed only up to
diffeomorphism.\footnote{\citeN{einstein-aether} makes a closely
related point himself: ``The fact that the general theory of
relativity has no preferred space-time coordinates which stand in a
determinate relation to the metric is more a characteristic of the
mathematical form of the theory than of its physical content.''}
There are restrictions on how one can apply diffeomorphisms to
solutions in practice in order for that application to be consistent
with the physical content of the theory, and those restrictions may
have philosophical significance, but they may not as well. By itself,
that there is arbitrariness tells us nothing of interest about the
theory.
A comparison is edifying. Classical mechanics as embodied in either
Lagrangian or Hamiltonian mechanics has a similar arbitrariness,
slightly different in each formulation of the theory. In Lagrangian
mechanics, one is free to choose the Lagrangian function itself on the
tangent bundle of configuration space up to the addition of a scalar
field derived from a closed 1-form on configuration space (or, in more
traditional terms, up to the addition of a total time-derivative of a
function of configuration coordinates) without changing the family of
solutions the Lagrangian determines.\footnote{See, \emph{e}.\emph{g}.,
\citeN{curiel-geom-ele}.} In Hamiltonian mechanics, one is free to
choose any symplectomorphism between the space of states and the
cotangent bundle of configuration space, \emph{i}.\emph{e}., one may
choose, up to symplectomorphism, any presentation of phase space (or,
in more traditional terms, any complete set of canonical coordinates),
without changing the family of solutions the Hamiltonian function
determines.\footnote{\emph{Op.\ cit}.} One feels no lack of
understanding of Lagrangian mechanics, no lacuna in its conceptual
resources, merely because one is free to choose the form of the
Lagrangian with wide latitude; just so, in Hamiltonian mechanics one
is not driven to investigate the ontic status of points in phase space
or of the physical quantities whose values one uses to label those
points, which ones get nominated `configuration' and which `momentum',
merely because one is free to choose whatever symplectomorphism one
likes in its presentation. Consider the fact that one can run an
argument analogous to the Hole Argument in the context of Hamiltonian
mechanics, substituting ``phase space'' for ``spacetime manifold'',
``symplectomorphism'' for ``diffeomorphism'' and ``symplectic
structure'' for ``metric''. Does that show anything of intrinsic
physical significance? No serious person would argue so. And in this
case, it would be manifestly absurd to ``apply a symplectomorphism
only to the symplectic structure and not the underlying manifold'': in
general the underlying manifold is a cotangent bundle and the
symplectic structure is the canonical one on it; pushing the
symplectic structure around on its own will yield a new symplectic
structure that is \emph{not} the canonical one, and so one manifestly
unphysical for the purpose of formulating Hamilton's equation.
The choice of Lagrangian or the choice of symplectomorphism rests on
nothing more than pragmatic considerations of the type adumbrated by
\citeN{carnap-emp-sem-ont} in his discussion of the choice of a
linguistic framework for the investigation of philosophical and
physical problems.\footnote{This is not to say that I consider the
choice of a Lagrangian or a symplectomorphic presentation of phase
space to be the choice of a Carnapian linguistic framework, only
that the sorts of considerations that go into each choice are
similar.} One chooses on the basis of nothing more than what puts
one at ease in any of a variety of ways, from pragmatic considerations
such as what will be simple or useful for a particular investigation,
to those based on historical custom and {\ae}sthetic predilection. It
is clear that the existence of inevitable, more or less arbitrary,
non-physical elements in the presentation of the models of a theory by
itself does not require of one a decision on the ontic status of any
entities putatively designated by the mathematical structures of
either Lagrangian or Hamiltonian mechanics. More to the point, it is
clear in these cases that the physical significance of the theory's
models is not masked or polluted by the unavoidable arbitrariness in
the details of their presentations.
In the same way, the diffeomorphic freedom in the presentation of
relativistic spacetimes does not \emph{ipso facto} require
philosophical elucidation, in so far as it in no way prevents us from
focusing on and investigating what is of true physical relevance in
systems that general relativity models, what one may think of as the
intrinsic physics of the systems, so long as one respects the
pragmatic conditions for the application of diffeomorphisms to
solutions. It is neither formal relations nor substantive entities
that remain invariant when one applies a diffeomorphism to a
relativistic spacetime; it is the family of physical facts the
spacetime represents. (This line of thought already strongly suggests
that the debate between substantivalists and relationalists is not
well posed.) One may represent those facts in a language some of
whose primitive terms designate ``spacetime points'' or not. Further,
one may want to restrict the attribution of existence to what has
intrinsic physical significance in the context of our best physical
theories. Then again, one may not. It is irrelevant to our capacity
to use them in profitable ways in science and, more important, to our
comprehension of those facts and our understanding of the role they
play in our broader attempts to comprehend the physical world.
In the end, however, the most serious problem I have with the Hole
Argument, and all other arguments analogous to it, comes to this:
nothing I can see militates in favor of taking the Hole Argument as
bearing on the ontic status of spacetime points, \emph{just because}
the Hole argument by itself provides no independent, clear and precise
criterion for what ``existence independent of metrical structure''
comes to. That idea has no substantive content on its own. In the
next two sections, I will show this by exhibiting two plausible,
precise criteria for what the idea may mean in the contexts of two
different types of investigation, which in the event lead respectively
to opposed conclusions.
\section{Limits of Spacetimes}
\label{sec:limits}
\resetsec
In this section, I propose an argument in favor of the view that one
cannot identify spacetime points in the absence of metrical structure,
and so, \emph{a fortiori}, that one cannot attribute to the spacetime
manifold any existence independent of that structure; the provision of
a precise criterion for the existence of spacetime structure, grounded
in both the structure and the application of physical theory, drives
the argument. In the event, two criteria natural to the investigative
context will suggest themselves, a weaker one based on the idea of the
identifiability of spacetime points and a stronger one based on their
existence (in a precise sense).
To treat a spacetime as the limit, in some sense, of an ancestral
family of continuously changing spacetimes is one of the ways of
embodying in the framework of general relativity two of the most
fundamental and indispensable tools in the physicist's workshop: the
idealization of a system by means of the suppression of complexity, so
as to render the system more tractable to investigation; and the
enrichment of a system's representation in a theory by the addition
(or reimposition) of complexity previously ignored (or ellided) in the
model the theory provides for the system. As a general rule, the
fewer degrees of freedom a system has, the easier it becomes to study.
Schwarzschild spacetime (figure~\ref{fig:schwarz}) is far easier to
work with than Reissner-Nordstr\"om (figure~\ref{fig:reiss-nord}) in
large part because one ignores electric charge, and there is a natural
sense in which one can think of Schwarzschild spacetime as the limit
of Reissner-Nordstr\"om as the electric charge of the central black
hole decreases in magnitude to zero.\footnote{Schwarzschild spacetime
is the unique spherically symmetric vacuum solution to the Einstein
field equation (other than Minkowski spacetime); it represents a
spacetime that is empty except for an electrically neutral,
spherically symmetric, static central body or black hole of a fixed
mass. Reissner-Nordstr\"om is the generalization of Schwarzschild
spacetime that allows the central structure to have an electric
charge. See, \emph{e}.\emph{g}., \citeN[ch.5,
\S5]{hawking-ellis-lrg-scl-struc-st} for an exposition.}
Contrarily, as a general rule the more degrees of freedom one includes
in a system's model, the more phenomena that the system manifests the
model can represent, and with greater accuracy (or at least fineness
of detail).
\begin{figure}
\centering
\includegraphics{schwarzschild}
\caption{\label{fig:schwarz}Carter-Penrose diagram of Schwarzschild
spacetime. Each point in the diagram represents a 2-sphere in the
spacetime manifold. (This diagram is taken from \protect
\citeNP{geroch-lim-sts}, with the author's permission.)}
\end{figure}
\begin{figure}
\centering
\includegraphics[height=4in]{reissner-nordstrom}
\caption{\label{fig:reiss-nord}Carter-Penrose diagram of
Reissner-Nordstr\"om spacetime. Each point in the diagram
represents a 2-sphere in the spacetime manifold. (This diagram is
taken from \protect \citeNP{geroch-lim-sts}, with the author's
permission.)}
\end{figure}
A generic representation of such a limiting process can provide a
schema of both of these theoretical tools respectively, depending on
whether one enlarges or shrinks the number of degrees of freedom in
the limiting process. As we will see, what in the idealized model one
may reasonably identify and attribute existence to may depend in
sensitive ways on the character of the more complex or simpler models
one starts from and the nature of the limiting process itself. This
fact drives the argument I propose. I will first discuss in some
detail two examples of such a limiting process in order to motivate
the two precise criteria I propose for the existence of spacetime
points independent of metrical structure.
Before diving into the examples, however, I first characterize in the
abstract the limiting process itself. I use the construction of
\citeN{geroch-lim-sts} (whose exposition I closely follow), which I
only sketch, to capture it. (I simplify his construction in
non-essential ways for our purposes, and gloss over unnecessary
technicalities.) Consider a 1-parameter family of relativistic
spacetimes, by which I mean a family $\{(\mathcal{M}_\lambda, \,
g^{ab} (\lambda))\}_{\lambda \in (0, 1]}$, where each
$(\mathcal{M}_\lambda, \, g^{ab} (\lambda))$ is a relativistic
spacetime with signature $(+, \, -, \, -, \, -)$ for $g^{ab}
(\lambda)$. (It will be clear in a moment why I work with the
contravariant form of the metric tensor.) In particular, I do not
assume that $\mathcal{M}_\lambda$ is diffeomorphic to
$\mathcal{M}_{\lambda'}$ for $\lambda \neq \lambda'$. The problem is
to find a limit of this family, in some suitable sense, as $\lambda
\rightarrow 0$. To solve the problem in full generality, we will use
a geometrical construction, gluing the manifolds $\mathcal{M}_\lambda$
of the family together to form a 5-dimensional manifold
$\mathfrak{M}$, so that each $\mathcal{M}_\lambda$ is itself a
4-dimensional submanifold of $\mathfrak{M}$ in such a way that the
collection of all of them foliate $\mathfrak{M}$.\footnote{In general
what will result is not a foliation in the strict sense of
differential topology, but will rather be a stratified space
(\citeNP{thom-stratifies}). It is close enough to a foliation,
however, to warrant using the more familiar term for simplicity of
exposition.} $\lambda$ becomes a scalar field on $\mathfrak{M}$,
and the metrics $g^{ab} (\lambda)$ on each submanifold fit together to
form a tensor field $g^{AB}$ on $\mathfrak{M}$, of signature $(0, \,
+, \, -, \, -, \, -)$. (I use majuscule indices for objects on
$\mathfrak{M}$.) The gradient of $\lambda$ on $\mathfrak{M}$
determines the singular part of $g^{AB}$: $g^{AN} \nabla_N \lambda =
0$. (This is why I work with the contravariant form of the metric;
otherwise, we could not contravect its five-dimensional parent in any
natural way with the gradient of $\lambda$.) Note that $g^{AB}$ by
itself already determines the submanifolds $\mathcal{M}_\lambda$
(\emph{viz}., as the surfaces defined by $g^{AN} \nabla_N \lambda =
0$), and that it does so in a way that does not fix any identification
of points among them. In other words, the structure I posit does not
allow one to say that a point in $\mathcal{M}_\lambda$ is ``the same
point in spacetime'' as a point in a different
$\mathcal{M}_{\lambda'}$ (as I shall discuss at some length below).
To define a limit of the family now reduces to the problem of the
attachment of a suitable boundary to $\mathfrak{M}$ ``at $\lambda =
0$''. A \emph{limiting envelopment} for $\mathfrak{M}$, then, is an
ordered quadruplet $(\hat{\mathfrak{M}}, \, \hat{g}^{AB}, \,
\hat{\lambda}, \, \Psi)$, where $\hat{\mathfrak{M}}$ is a
5-dimensional manifold with paracompact, Hausdorff, connected and
non-trivial boundary $\partial \hat{\mathfrak{M}}$, $\hat{g}^{AB}$ a
tensor field on $\hat{\mathfrak{M}}$, $\hat{\lambda}$ a scalar field
on $\hat{\mathfrak{M}}$ taking values in $[0, \, 1]$, and $\Psi$ a
diffeomorphism of $\mathfrak{M}$ to the interior of
$\hat{\mathfrak{M}}$, all such that
\begin{enumerate}
\item $\Psi$ takes $g^{AB}$ to $\hat{g}^{AB}$ (\emph{i}.\emph{e}.,
$\Psi$ is an isometry) and takes $\lambda$ to $\hat{\lambda}$
\item $\partial \hat{\mathfrak{M}}$ is the region defined by
$\hat{\lambda} = 0$
\item $\hat{g}^{AB}$ has signature $(0, \, +, \, -, \, -, \, -)$
on $\partial \hat{\mathfrak{M}}$
\end{enumerate}
This makes precise the sense in which $\hat{\mathfrak{M}}$ represents
$\mathfrak{M}$ with a boundary attached in such a way that the metric
on the boundary ($\hat{g}^{AB}$ restricted to $\partial
\hat{\mathfrak{M}}$) can be naturally identified as a limit of the
metrics on the $\mathcal{M}_\lambda$ ($g^{AB}$ on $\mathfrak{M}$). I
call $\{(\mathcal{M}_\lambda, \, g^{ab} (\lambda))\}_{\lambda \in (0,
1]}$ an \emph{ancestral family} of the spacetime represented by
$\partial \hat{\mathfrak{M}}$, and I call $\partial
\hat{\mathfrak{M}}$ the \emph{limit space} of the family with respect
to the given envelopment. In general, a given spacetime will have
many ancestral families, and an ancestral family will have many
different limit spaces. For the sake of convenience I will often not
distinguish between $\mathfrak{M}$ and the interior of
$\hat{\mathfrak{M}}$. (Although it is tempting also to abbreviate
`$\partial \hat{\mathfrak{M}}$' by `$\mathcal{M}_0$', I will not do
so, because part of the point of the construction is that different
spacetimes can have the same ancestral family.)
Before giving an example of the construction and putting it to work,
I discuss one of its features, that it parametrizes not only the
metrics but also the spacetime manifolds themselves.
\citeN[p.~181]{geroch-lim-sts} himself states in illuminating terms
the reason behind this.
\begin{quote}
It might be asked at this point why we do not simply take the
$g^{ab} (\lambda)$ as a 1-parameter family of metrics on a given
fixed manifold $\mathcal{M}$. Such a formulation would certainly
simplify the problem: it amounts to a specification of when two
points $p_\lambda \in \mathcal{M}_{\lambda'}$ and $p_{\lambda'} \in
\mathcal{M}_\lambda$ ($\lambda \neq \lambda'$) are to be considered
as representing ``the same point'' of $\mathcal{M}$. It is not
appropriate to provide this additional information, for it always
involves singling out a particular limit, while we are interested in
the general problem of finding all limits and studying their
properties.
\end{quote}
To make the force of these remarks clear, consider the attempt to take
the limit of Schwarzschild spacetime as the central mass goes to 0.
In Schwarzschild coordinates, using the parameter $\lambda \equiv
M^{-1/3}$ (the inverse-third root of the Schwarzschild mass), the
metric takes the form
\begin{equation}
\label{eq:schwarzschild}
\left( 1 - \frac{2}{\lambda^3 r} \right) dt^2 - \left( 1 -
\frac{2}{\lambda^3 r} \right)^{-1} dr^2 - r^2 (d\theta^2 + \sin^2
\theta \hspace*{.1em} d\phi^2)
\end{equation}
This clearly has no well defined limit as $\lambda \rightarrow 0$.
Now, apply the coordinate transformation
\[
\tilde{r} \equiv \lambda r, \quad \tilde{t} \equiv \lambda^{-1} t,
\quad \tilde{\rho} \equiv \lambda^{-1} \theta
\]
In these coordinates, the metric takes the form
\[
\left( \lambda^2 - \frac{2}{\tilde{r}} \right) d\tilde{t}^2 - \left(
\lambda^2 - \frac{2}{\tilde{r}} \right)^{-1} d\tilde{r}^2 -
\tilde{r}^2 (d\tilde{\rho}^2 + \lambda^{-2} \sin^2 (\lambda
\tilde{\rho}) \hspace*{.1em} d\phi^2)
\]
The limit $\lambda \rightarrow 0$ exists and yields
\[
-\frac{2}{\tilde{r}} d\tilde{t}^2 + \frac{\tilde{r}}{2} d\tilde{r}^2 -
\tilde{r}^2 (d\tilde{\rho}^2 + \tilde{\rho}^2 \hspace*{.1em} d\phi^2)
\]
a flat solution discovered by \citeN{kasner-geom-thms-efe}. If
instead of that coordinate transformation we apply the following to
the original Schwarzschild form~\eqref{eq:schwarzschild},
\[
x \equiv r + \lambda^{-4}, \quad \rho \equiv \lambda^{-4} \theta
\]
then the resulting form also has a well defined limit, which is the
Minkowski metric. The two limiting processes yield different
spacetimes because it happens behind the scenes that ``the same points
of the underlying manifold get pushed around relative to each other in
different ways''. Because the coordinate relations of initially
nearby points differ in different coordinate systems, those
differences get magnified in the limit, so that their final metrical
relations differ. Thus, the limits in the different coordinates yield
different metrics.\footnote{\citeN{paiva-et-lims-sts} discuss in some
detail an interesting class of different limiting spacetimes one can
induce from Schwarzschild spacetime by taking the limit as the mass
goes to zero and to infinity in different ways. See
\citeN{bengtsson-et-classics-lims-sts} for a similar discussion for
Reissner-Nordstr\"om spacetime, as the electric charge and the mass
respectively are taken to zero.}
In the language I introduced above, we should say that the difference
between the two limits consists in the different identifications each
makes among the points of different $\mathcal{M}_\lambda$. That is
why it is inappropriate to work with a fixed manifold from the start.
To do so determines a unique limit, but we want to allow ourselves
different ways to take the limit, so that our ideal scientist can
ignore different facets of the complex system under study, and so
produce different idealized models of it.\footnote{Of course,
sometimes is is appropriate for the scientist to take the limit of a
family of metrics on a fixed background manifold. An excellent
example is in the statement and proof of the geodesic theorem of
\citeN{ehlers-geroch-eom-small-bods-gr}. In fact, they give an
illuminating discussion of this very issue on p.~233.} For example,
she may want to take the limit of Reissner-Nordstr\"om spacetime as
the mass goes to zero while leaving the electric charge fixed rather
than taking the limit as the electric charge vanishes, or she may want
to take the limit in a way that does not respect the spherical
symmetry of the initial system in order, \emph{e}.\emph{g}., to study
small perturbations of the original system.
To characterize the metrical structure of the limit space using
structure of members of the ancestral family, I introduce one more
construction. An orthonormal tetrad $\xi (\lambda)$ at a point
$p_\lambda \in \mathcal{M}_\lambda$ is a collection of 4 tangent
vectors at the point mutually orthogonal with respect to $g_{ab}
(\lambda)$. Let $\gamma$ be a smooth curve on $\mathfrak{M}$ nowhere
tangent to any $\mathcal{M}_\lambda$ such that it intersects each
exactly once. $\gamma$ then is composed of a set of points $p_\lambda
\in \mathcal{M}_\lambda$, one for each $\lambda$. A \emph{family of
frames} along $\gamma$ is a family of orthonormal tetrads, one at
each point of the curve such that each vector in the tetrad is tangent
to its associated $\mathcal{M}_\lambda$, whose members vary smoothly
along it. In general, a family of frames will have no well defined
limit in $\hat{\mathfrak{M}}$ as $\lambda \rightarrow 0$,
\emph{i}.\emph{e}., there will be no tetrad $\xi (0)$ at a point of
$\partial \hat{\mathfrak{M}}$ that the family $\xi (\lambda)$
converges to; in this case, I say the family is \emph{degenerate}. It
is always possible, however, given a tetrad $\xi (0)$ at a point on
the boundary to find some family of frames that does converge to it.
Now, fix $\xi (0)$ at $p_0 \in \partial \hat{\mathfrak{M}}$ and a
family of frames $\xi (\lambda)$ that converges to it. We can
represent the metric tensor $g_{ab} (\lambda)$ in a normal
neighborhood of $p_\lambda$ in $\mathcal{M}_\lambda$ using the normal
coordinate system that $\xi (\lambda)$ defines in the neighborhood.
In a normal neighborhood of $p_0$, the components of the metric with
respect to these coordinates converge as $\lambda \rightarrow 0$, and
the limiting numbers are just the components of $g_{ab} (0)$ at $p_0$
with respect to the normal coordinates that $\xi (0)$ defines. In
this way, we can characterize all structure on the limit space based
on the behavior of the corresponding structures along the family of
frames in the ancestral family.
We are finally in a position to use this machinery to construct
concrete examples. Consider a family $\{(\mathcal{M}_\lambda, \,
g^{ab} (\lambda))\}$ of Reissner-Nordstr\"om spacetimes each element
of the family having the same fixed value $M$ for its mass and all
parametrized by their respective electric charges $\lambda$, which
converge smoothly to 0.\footnote{I ignore the fact that electric
charge is a discrete quantity in the real world, an appropriate
idealization in this context.} Construct their envelopment. One
can now impose a natural collection of families of frames on the
family, with the limit space being Schwarzschild
spacetime.\footnote{The frames are natural in the sense that they
conform to and respect the spherical and the timelike symmetries in
all the spacetimes. One could use this fact to explicate the claim
that Schwarzschild spacetime is the canonical limit of
Reissner-Nordstr\"om spacetime, in the sense that it is what one
expects on physical grounds, whatever exactly that may come to, in
the limit of vanishing charge while leaving all else about the
spacetime fixed.} Now, comparison of figures~\ref{fig:schwarz} and
\ref{fig:reiss-nord} suggests that something drastic happens in the
limit. All the points in the throat of the Reissner-Nordstr\"om
spacetimes (the shaded region in the diagram) seem to get swallowed by
the central singularity in Schwarzschild spacetime---in some way or
other, they vanish. Using our machinery we can make precise the
question of their behavior in the limit $\lambda \rightarrow 0$ in the
envelopment.
Consider the points in the shaded region in
figure~\ref{fig:reiss-nord}, between the lines $r = 0$ and $r = r^-$.
($r$ is the radial coordinate in a system that respects the
spacetime's spherical symmetry; the coordinate values $r^-$ and $r^+$
define boundaries of physical significance in the spacetime, which in
large part serve to characterize the central region of the spacetime
as a black hole.) Fix a natural family of frames along a curve in
$\mathfrak{M}$ composed of points $q_\lambda$ each of which lies in
the shaded region in its respective spacetime. It is straightforward
to verify that the family of frames along the curve does not have a
well defined limit: roughly speaking, the curve runs into the
Schwarzschild singularity at $r = 0$. In this sense, no point in
Reissner-Nordstr\"om spacetime to the future of the horizon $r = r^-$
has a corresponding point in the limit space. To sum up: one begins
with a family of Reissner-Nordstr\"om spacetimes continuously
parametrized by electric charge, which converges to 0, and constructs
the envelopment of the family; one constructs the limit space by a
choice of families of frames; the collection of families of frames
enforces an identification of points among different members of the
family of spacetimes, including a division of those points that have a
limit from those that do not; and that identification, in turn,
dictates the identification of spacetime points in the limit space
(which points in the ancestral family lie within the Schwarzschild
radius, \emph{e}.\emph{g}., and which do not). Thus one can identify
points within the limit Schwarzschild spacetime, one's idealized
model, only by reference to the metrical structure of members of the
ancestral family; one can, moreover, identify points in the limit
space with points in the more complex, initial models one is
idealizing only by reference to the metrical structure of the members
of the ancestral family as well. It is only by the latter
identification, however, that one can construe the limit space as an
idealized model of one's initial models, for the whole point is to
simplify the reckoning of the physical behavior of systems in
particular spatiotemporal regions of one's initial models, and most of
all at individual spacetime points of one's initial models.
One can, moreover, use different families of natural frames to
construct Schwarzschild spacetime from the same ancestral family, with
the result that in each case the same point of Schwarzschild spacetime
is identified with a different family of points in the ancestral
family. More generally, different families of frames will yield limit
spaces different from Schwarzschild spacetime, with no canonical way
to identify a point in one limit space (one idealized model the
theoretician constructs) with one in another. In other words, the
identification of points in the limit space depends sensitively on the
way the limit is taken, \emph{i}.\emph{e}., on the way the model is
constructed. In consequence, in so far as one conceives of
Schwarzschild spacetime as an idealized model of a richer, more
complete representation, one can identify points in it only by
reference to the metrical structure of one of its ancestral families,
and one can do that in a variety of ways.
Now, say one wants to treat slightly aspherical, almost
Schwarzschildian spacetimes as a complexification of Minkowski
spacetime, in order to study how asphericities affect metrical
behavior.\footnote{One ought not confuse the idea of complexification
I employ here---the making of a model more complex by the
introduction of new representational structure---with the idea
bandied about in other contexts in mathematical physics often also
called `complexification', in which one takes a mathematical
structure based on the real numbers and extends it to one based on
the complex numbers.} Because the limit spacetime will be almost
Schwarzschildian, its appropriate manifold is still $\mathbb{R}^2
\times \mathbb{S}^2$, the natural topology of Schwarzschild spacetime.
In this case, in one intuitive sense points will ``appear'', because
the topology of Minkowski spacetime is $\mathbb{R}^4$, so in some
sense one must ``compactify two topological dimensions'' to derive a
Schwarzschildian spacetime as a more complex limit. There are many
ways to effect such a compactification; all the simplest, such as
Alexandrov compactification, work by the addition of an extra point or
set of points to the topological manifold to represent, intuitively
speaking, the bringing in of points at infinity to a manageable
distance from everything else.\footnote{See, \emph{e}.\emph{g}.,
\citeN{kelley-gen-topo} for an account of methods of
compactification, including the Alexandrov type.} The difficulty of
these issues, however, is underscored by the fact that one can also
think of this as a case in which points rather \emph{disappear}:
$\mathbb{R}^2 \times \mathbb{S}^2$, after all, is homeomorphic to
$\mathbb{R}^4$ with a line removed! Thus one could use an ancestral
family every member of which is $\mathbb{R}^4$ but that has as limit
space the manifold of Schwarzschild spacetime presented as the
manifold $\mathbb{R}^4$ with a line removed.\footnote{This is a
concrete instance where thinking of two different diffeomorphic
presentations of the same manifold---in this case, $\mathbb{R}^2
\times \mathbb{S}^2$ and $\mathbb{R}^4$ with a line removed---as
different manifolds leads to obvious difficulties, if not downright
confusions.}
In this example, we will consider the attempt to introduce a central,
slightly aspherical body by physical construction in a Minkowskian
laboratory, as an experimentalist might do it. For the sake of
concreteness, let us say that our experimentalist will, in his
representation of the experiment, use an Alexandrov compactification
of $\mathbb{R}^4$ to yield $\mathbb{R}^2 \times \mathbb{S}^2$ as the
presentation of the manifold of the limit space. The physical
construction will proceed in infinitesimal stages, with a tiny portion
of matter introduced at each step distributed in a slightly aspherical
way (keeping, in an intuitive sense, the aspherical shape of the body
the same), and an allowance of a finite time to allow the ambient
metrical structure to settle down to an almost Schwarzschildian
character before the next step is initiated, until the central body's
mass reaches the desired amount. (Intuitively, the finite time period
allows the metrical perturbations introduced by the movement of the
matter in and its distribution around the central body to radiate off
to infinity.) One can represent this process with a limiting
ancestral family of Geroch's type in a more or less obvious way,
starting with Minkowski spacetime, \emph{viz}., the empty, flat
laboratory, and each member of the ancestral family representing the
laboratory at a particular stage of the construction, when a bit more
matter has been introduced and the perturbations have settled down.
Now, consider at the beginning of the process a small patch of space
in the laboratory not too far from the position where the central body
will be constructed. We want to try to track, as it were, the
spacetime points in that patch during the enlargement of the central
body because we plan to investigate, say, how the metrical structure
in regions at that spatiotemporal remove from a central aspherical
body differ from each other for different masses of the central body.
(Because the Einstein field equation is nonlinear, and there is no
exact symmetry, one cannot just assume that slightly aspherical
spacetimes will scale in any straightforward way with increases in the
central mass.) There are several ways one might go about trying to
track the region as the construction progresses. One obvious, simple
way is by the triangulation of distances from some ``fixed'' markers
in the laboratory. Because the metrical structure within the lab is
constantly changing, however, and doing so in very complex ways during
the periods when new matter is being introduced and distributed, and
the concomitant metrical perturbations are radiating away, there is no
canonical way of implementing the triangulation procedures; in fact,
the different ways of doing so are exactly captured by the different
families of frames one can fix to identify points among the members of
the ancestral family of spacetimes (which in this case, recall, now
respectively represent the spacetime region enclosed by the laboratory
at different stages of the construction of the central body).
According to some of the concrete implementations of the triangulation
procedure, \emph{i}.\emph{e}., according to different families of
frames one uses to identify points among the several members of the
ancestral family, the patch one tries to track will end up inside the
central body; according to other procedures, it will end up outside
the central body. In consequence, what one means by ``the set of
spacetime points composing a small region at a fixed spatiotemporal
position relative to the central body'' will depend sensitively on how
one fixes and tracks relative spatiotemporal positions, which is to
say, depends sensitively on one's knowledge of the spacetime's
metrical structure.\footnote{One might object that, in this example,
the experimentalist is really trying to track ``the same points
through space over time'', not ``the same spatiotemporal points in
different spacetimes''. In fact, though, since the goal of the
investigation is to determine how global metrical structure in
slightly aspherical spacetimes differ for different values of the
central mass, it is natural for the experimentalist to consider each
static phase of the laboratory---the period after the last bit of
mass has been added and the perturbations have settled down, but
before the next bit of mass is added---as a separate spacetime in
its own right, for the purposes of comparison. An appropriate
analogue is the so-called ``physical process'' version of the First
Law of black-hole mechanics
(\citeNP{wald-gao-proc-1st-genl-2nd-charged-rot-bhs},
\citeNP{wald-qft-cst}), where one must identify two separate
spacetimes (in the sense of two different solutions to the Einstein
field equation) that differ in that one conceives of the one as the
result of a dynamical evolution of the other, even though there is
no concrete representation of that evolution as occurring in a
single spacetime.}
We are finally in a position to offer a precise criterion for
``existence of spacetime points independent of metrical structure''
natural to the investigative contexts we have considered. There are
in fact two natural criteria that suggest themselves, one weaker than
the other. The first, suggested by the example of complexification
and stated somewhat loosely, is
\begin{defn}
\label{def:limiting-criterion-identify}
Points in a spacetime manifold \emph{have existence independent of
metrical structure} if there is a canonical method to identify
spacetime points during gradual modifications to the local spacetime
structure.
\end{defn}
My discussion of the example of complexification shows that, in this
context and using this criterion, spacetime points do not have
existence independent of metrical structure.
Now, based on the discussion of simplification, I propose a second
criterion, stronger than the first and formulated more precisely and
rigorously. Fix an envelopment of a limiting family with a definite
limit space. I say that a point in $\mathcal{M}_1$ with an associated
degenerate family of frames \emph{vanishes} (or that the point itself
is a \emph{vanishing point}) with respect to the given family of
frames. I say that a point in $\partial \hat{\mathfrak{M}}$
\emph{appears} if there is no family of frames that converges to it.
\begin{defn}
\label{def:limiting-criterion}
Points in a spacetime manifold \emph{have existence independent of
metrical structure} if no specification of a family of frames in
any ancestral family of the spacetime has vanishing or appearing
points.
\end{defn}
I do not demand that one be able to identify in a preferred way a
spacetime point in the limit with any point of any member of one of
its ancestral families, much less for all its ancestral families; this
allows us to hold on to diffeomorphic freedom in the presentation of
the limit space. I do not even demand that the criterion hold for
every possible spacetime model---perhaps in some spacetimes it makes
sense to attribute existence to spacetime points independent of
metrical structure, whereas in others (say, completely homogeneous
spacetimes) it does not. I demand only that, for a given spacetime,
one not be able to make points in any of its ancestral families vanish
and not be able to make points in it, as the limit space, appear.
This attempts to capture the idea that, when we construct a spacetime
model and treat it as an idealized representation of a more complex
system---as it always is---then we can reliably identify spacetime
points in our model with points in the more complex system, albeit up
to diffeomorphic presentation. If we cannot do this irrespective of
the more complex model we start from, then we cannot without
arbitrariness and artifice regard results of an investigation in the
context of the idealized model as relevant to the physics of the more
complex system, for we will be unable to identify the regions in the
more complex system that the results of the idealizing investigation
pertain to. The example of Schwarzschild spacetime as a limit of a
family of Reissner-Nordstr\"om spacetimes clearly does not satisfy the
criterion, for there are points that vanish in the limiting procedure
(\emph{e}.\emph{g}., those in the shaded region of
figure~\ref{fig:reiss-nord}). One may suspect that the existence of
singular structure in the two spacetimes fouls things up. The
following result, however, establishes that no spacetime satisfies the
criterion, \emph{i}.\emph{e}., that its failure is universal and
depends on no special properties of any spacetime model.
Every spacetime has at least one ancestral family, the trivial one
consisting of the continuous sequence of itself, so to speak.
Construct an envelopment $\mathfrak{M}$ for it, with it itself as the
limit space, and apply a slight twist, so to speak, to every metric in
every model in the family so as to render each model non-isometric to
any other, \emph{i}.\emph{e}., so as to render the family non-trivial.
(One can make this idea precise in any of a number of simple ways,
such as using a smoothly varying 1-parameter family of linear
perturbations.) On a curve in $\mathfrak{M}$, fix a family of frames
that has a well defined limit on $\partial \hat{\mathfrak{M}}$. Now,
define a family of Lorentz transformations along that curve, one
transformation at each point, such that the family varies smoothly
along the curve, and such that when one applies each transformation to
the tetrad at its point, the result is a family of frames that has no
well defined limit. (One can always do this; for example, the Lorentz
transformations can cause the tetrads to oscillate wildly as $\lambda
\rightarrow 0$.) The points of the ancestral family along that curve
have no corresponding point in the limit space defined by the
resulting family of frames. This proves
\begin{prop}
\label{prop:vanish}
Every spacetime has a non-trivial ancestral family with vanishing
points. Every non-trivial ancestral family has a limit space with
respect to which some of its points vanish.
\end{prop}
In consequence, in every relativistic spacetime we treat as an
idealized model in the context of this sort of scientific
investigation, we can attribute existence to individual spacetime
points (or not), only by reference to the metrical structure of the
ancestral family we use to construct the model, and the limiting
process we choose for the construction.
An obvious objection to the relevance of these arguments to the ontic
status of spacetime points is that I deal here only with idealizations
and approximations, not with ``a real model of real spacetime''. But
we never work with anything that is not an idealization---it's
idealizations all the way down, young man, as part of the human
condition. If you can't show me how to argue for the existence of
spacetime points independently of metrical structure using our best
scientific theories \emph{as they are actually used in successful
practice}, then you are not relying on real science to ground your
arguments. You are paying only lip-service to the idea that science
should ground these sorts of metaphysical issues.
\section{Pointless Constructions}
\label{sec:pointless}
\resetsec
The argument of \S\ref{sec:limits} yields a conclusion that holds only
in a limited sphere, \emph{viz}., those investigations based on the
idealization of models of spacetime by means of limits. One may
wonder whether it could be parlayed into a more general argument. I
do not think so. Indeed, I think there is \emph{no} sound argument to
the effect that no matter the context of the investigation one can
identify spacetime points or attribute existence to them only by
reference to prior metrical structure. Sometimes, in some contexts,
one can attribute existence to them and identify them without any such
reference. To show this, I will present an argument that all the
structure accruing to a spacetime, considered simply as a differential
manifold that represents the collection of all possible (or, depending
on one's modal predilections, actual) physical events, can be given
definition with clear physical content in the absence of metrical
structure. The argument takes the form of the construction of the
point-manifold of a spacetime, its topology, its differential
structure and all tensor bundles over it from a collection of
primitive objects that, when the construction is complete, acquires a
natural interpretation as a family of covering charts from the
manifold's atlas, along with the families of bounded, continuous
scalar fields on the domain of each chart. That idea yields the
following precise criterion the argument will rely on.
\begin{defn}
\label{def:pointless-criterion}
Points in a spacetime manifold \emph{have existence independent of
metrical structure} if the manifold can be constructed from a
family of scalar fields, the values of which can be empirically
determined without knowledge of metrical structure.
\end{defn}
The basic idea of the construction is simple. I posit a class of sets
of rational numbers to represent the possible values of physical
fields, with a bit of additional structure in the form of primitive
relations among them just strong enough to ground the definition of a
derived relation whose natural interpretation is ``lives at the same
point of spacetime as''. A point of spacetime, then, consists of an
equivalence class of the derived relation. The derived relation,
moreover, provides just enough rope to allow for the definition of a
topology and a differential structure on the family of all equivalence
classes, and from this the definition of all tensor bundles over the
resultant manifold, completing the construction. The posited
primitive and derived relations have a straightforward physical
interpretation, as the designators of instances of a schematic
representation of a fundamental type of procedure the experimental
physicist performs on physical fields when he attempts to ascertain
relations of physical proximity and superposition among their observed
values. An important example of such an experimental procedure is his
use of the observed values of physical quantities associated with
experimental apparatus to determine the values of quantities
associated with other systems, those he investigates by use of the
apparatus. This interpretation of the relations motivates the claim
that the constructed structure suffices, for our purposes, as a
representation of spacetime in the context of a particular type of
experimental investigation as modeled by mathematical physics, and is
not (only) an abstract mathematical toy.
I begin the construction by laying down some definitions. Let
$\mathbb{Q}$ be the set of rational numbers. A \emph{simple pointless
field} $\phi$ (or just \emph{simple field}) is a disjoint union
$\displaystyle \biguplus_{p \in \mathbb{Q}^4} f_p$, indexed by the set
$\mathbb{Q}^4$, such that
\begin{enumerate}
\item\label{item:field-rational_vals}every $f_p \in \mathbb{Q}$
\item\label{item:field-all-q4}there is an $f_p \in \phi$ for every
$p \in \mathbb{Q}^4$
\item\label{item:field-bounded}there are two strictly positive
numbers $B_\text{l}$ and $B_\text{u}$ such that $B_\text{l} < |f_p|
< B_\text{u}$ for all $p \in \mathbb{Q}^4$
\item\label{item:field-continuous}the function $\bar{\phi}:
\mathbb{Q}^4 \rightarrow \mathbb{Q}$ defined by $\bar{\phi}(p) =
f_p$ is continuous in the natural topologies on those spaces, except
perhaps across a finite number of compact three-dimensional
boundaries in $\mathbb{Q}^4$
\end{enumerate}
Our eventual interpretation of such a thing as a candidate result for
an experimentalist's determination of the values for a physical field
motivates the set of conditions. That we index $\phi$ over
$\mathbb{Q}^4$ means that we assume from the start that the
experimentalist by the use of actual measurements and observations
alone can impose on spacetime at most the structure of a countable
lattice indexed by quadruplets of rational numbers (and even this only
in a highly idealized sense); in other words, the spatiotemporal
precision of measurements is limited.
Condition~\ref{item:field-rational_vals} says that all measurements
have only a finite precision in the determination of the field's
value. Condition~\ref{item:field-all-q4} says that the field the
experimentalist measures has a definite value at every point of
spacetime. Condition~\ref{item:field-bounded} says that there is an
upper and a lower limit to the magnitude of values the experimentalist
can attribute to the field using the proposed experimental apparatus
and technique; for instance, any device for the measurement of the
energy of a system has only a finite precision, and thus can attribute
only absolute values greater than a certain magnitude, and the device
will be unable to cope with energies above a given magnitude.
Condition~\ref{item:field-continuous} tries to capture the ideas that
(local) experiments involve only a finite number of bounded physical
systems (apparatuses and objects of study), and that classical
physical systems bear physical quantities the magnitudes of which vary
continously (if not more smoothly), except perhaps across the
boundaries of the systems.
Fix a family $\Phi$ of simple pointless fields. The \emph{link at
$p$}, $\lambda_p$, is a set containing exactly one element from each
simple field in $\Phi$ such that all the elements are indexed by $p$,
the same quadruplet of rational numbers. One link, for example,
consists of the set of all values in the fields in $\Phi$ indexed by
$(3/17, \, 2, \, -3001\frac{90}{91}, \, 2)$. A \emph{linked family of
simple pointless fields} $\mathfrak{F}$ is an ordered pair $(\Phi,
\, \Lambda)$ where $\Phi$ is a countable collection of simple fields,
and $\Lambda$ is the family of links on $\Phi$, a \emph{linkage},
complete in the sense that it contains exactly one link for each $p
\in \mathbb{Q}^4$. The idea is that the values of the simple fields
in the same link all live ``at the same point of spacetime'', namely
that designated by $p$. One can think of the linkage as a coordinate
system on an underlying, abstract point set.
We are almost ready to define the point-structure of the spacetime
manifold. We require only two more constructions, which I give in an
abbreviated fashion so as to convey the main points without getting
bogged down in unnecessary technical detail. Let $\mathfrak{F} =
(\Phi, \, \Lambda)$ be a linked family containing all simple fields;
we call it a \emph{simple fundamental family}. Let
$\hat{\mathfrak{F}} = (\hat{\Phi}, \, \hat{\Lambda})$ be another. We
want a way to relate the linkages of the two, so as to be able to
represent the relation between the coordinate systems of two different
charts on the same neighborhood of the spacetime manifold, or on the
intersection of two neighborhoods. A \emph{cross-linkage} on a simple
fundamental family is an ordered triplet $(O, \, \hat{O}, \, \chi)$
where $O \subseteq \mathbb{Q}^4$ and $\hat{O} \subseteq \mathbb{Q}^4$
are open sets, such that either both are the null set or else both are
homeomorphic to $\mathbb{Q}^4$, and $\chi$ is a homeomorphism of $O$
to $\hat{O}$. The link $\lambda_p \in \Lambda$ for $p \in O$, then,
will designate the same point in the underlying manifold as
$\hat{\lambda}_{\chi(p)} \in \hat{\Lambda}$ for $\chi(p) \in \hat{O}$;
in this case, we say the links \emph{touch}. If $O$ and $\hat{O}$ are
the null set, then the represented neighborhoods do not intersect.
(We do not require that the values of the scalar fields in the two
different simple fundamental families be numerically equal at any
given point, as the two scalar fields may represent different physical
quantities, \emph{e}.\emph{g}., a component of the fluid velocity and
a component of the shear-stress tensor of a viscous fluid.) One can
extend the idea of a cross-linkage to an arbitrary number of simple
fundamental families in the obvious way. (To make the idea precise we
would need to index the collection of families, and so on, but I think
it is clear enough without going through the bother.) We would then
identify a point in an underlying abstract point-set as an equivalence
class of links under the equivalence relation ``touches''.
To finish the preparatory work, we must move from rationals to reals.
Fix a simple, fundamental family $\mathfrak{F}$. First, we attribute
to $\mathfrak{F}$ the algebraic structure of a module over
$\mathbb{Q}$. For example, the sum of two simple pointless fields
$\phi$ and $\psi$ in $\Phi$ is a simple pointless field $\xi$ such
that $x_p \equiv f_p + g_p$ is the value in $\xi$ labeled by the index
$p$, where $f_p \in \phi$ and $g_p \in \psi$. $\xi$ is clearly itself
a simple pointless field with a natural embedding in the linkage on
$\mathfrak{F}$, and so belongs to $\Phi$. Now, roughly speaking, we
take a double Cauchy-like completion of $\Phi$ over both the points $p
\in \mathbb{Q}^4$ and the values $f_{\hat{p}} \in \mathbb{Q}$,
yielding the family $\bar{\Phi}$ of all disjoint unions of real
numbers continuously indexed by quadruplets of real
numbers.\footnote{In order to get the completion we require, standard
Cauchy convergence does not in fact suffice. We must rather use a
more general method, such as Moore-Smith convergence based on
topological nets. The technical details are not important. See,
\emph{e}.\emph{g}., \citeN[ch.~2]{kelley-gen-topo} for details.}
This procedure makes sense, because every continuous real scalar field
on $\mathbb{R}^4$ is, again roughly speaking, the limit of some
sequence of bounded, continuous rational fields defined on
$\mathbb{Q}^4$. We thus obtain what is in effect the family
$\bar{\Phi}$ of all continuous real scalar fields on $\mathbb{R}^4$,
though I refer to them as \emph{pointless fields}, in so far as, at
this point, they are still only indexed disjoint unions. The limiting
procedure, moreover, induces on $\bar{\Phi}$ the structure of a module
over $\mathbb{R}$ from that on $\Phi$. Finally, in the obvious way,
we take the completion, as it were, of $\Lambda$ using the same
limiting procedure to obtain a linkage $\bar{\Lambda}$ on
$\bar{\Phi}$. I call $\bar{\mathfrak{F}} = (\bar{\Phi}, \,
\bar{\Lambda})$ a \emph{fundamental family}. A cross-linkage on a
pair of fundamental families is the same as for simple fundamental
families, except only that one uses homeomorphisms on subsets of
$\mathbb{R}$ rather than $\mathbb{Q}$. If we have two simple
fundamental families with a cross-linkage on them and take limits to
yield two fundamental families, then the nature of the limiting
process guarantees a unique cross-linkage on the two fundamental
families consistent with the original.
We can at last construct a real topological manifold from a collection
of simple fundamental families. The basic idea is that a fundamental
family represents the family of continuous real functions on the
interior of a bounded, normal neighborhood of what will be the
spacetime manifold. Because a spacetime manifold must be paracompact
(otherwise it could not bear a Lorentz metric), there is always a
countable collection of such bounded, normal neighborhoods that cover
it. This suggests
\begin{defn}
\label{def:pointless-mnfld}
A \emph{pointless topological manifold} is an ordered pair $(\{
\mathfrak{F}_i \}_{i \in \mathbb{N}}, \, \chi)$ consisting of a
countable set of simple fundamental families and a cross-linkage on
them.
\end{defn}
To justify the definition, I sketch the construction of the full
point-manifold and its topology. First, we take the joint limit of
all simple fundamental families to yield a countable collection of
fundamental families with the induced cross-linkage. A point in the
manifold, then, is an equivalence class of links, at most one link
from each family, under the equivalence relation ``touches''. The set
of links associated with one of the families, then, becomes a
representation, with respect to the equivalence relation, of the
interior of a compact, normal neighborhood in the manifold, and the
fields in that family represent the collection of continous real
functions on that neighborhood. The cross-linkage defines the
intersections among all these neighborhoods, yielding the entire
point-set of the manifold. By assumption, the collection of all such
neighborhoods forms a sub-basis for the topology of the manifold, and
so, by constructing the unique topological basis from the given
sub-basis, the point-set becomes a true topological manifold. It is
straightforward to verify, for example, that a real scalar field on
the constructed manifold is continuous if and only if its restriction
to any of the basic neighborhoods defines a field in the fundamental
family associated with that neighborhood.
Now, to complete the construction, we can define the manifold's
differential structure in a straightforward way using similar
techniques. First, demarcate the family of smooth scalar fields as a
sub-set of the continuous ones, which one can do in any of a number
straightforward ways with clear physical content based on the idea of
directional derivatives. The family of all smooth scalar fields on a
topological manifold, however, fixes its differential structure
\cite{chevalley46}. The directional derivatives themselves suffice
for the definition of the tangent bundle over the manifold, and from
that one obtains all tensor bundles.
After so much abstruse and, worse, tedious technical material, we can
now judge whether the construction supports the argument I want to
found on it. The use of $\mathbb{Q}^4$ to index a simple pointless
field represents the fact that all points in a laboratory have been
uniquely labeled by 4 rational numbers, say, by the use of rulers and
stop-watches. Such an operation neither measures nor relies on
knowledge of metrical structure, for it yields in effect only a chart
on that spacetime region. (No assumption need be made about the
``metrical goodness'' of the rulers and clocks.) Neither does any
other operation used in the construction pertain to metrical
structure. One determines the values of the simple fields, for
example, by use of physical observations, which do not themselves
necessarily depend on knowledge of the ambient metrical structure. To
illustrate the idea, consider the use of a gravity gradiometer to
measure the components of the Riemann tensor in a region of spacetime,
which exemplifies many of the ideas in the construction. The
gradiometer is essentially a sophisticated torsion balance for
measuring the quadrupole (and higher) moments of an acceleration
field.\footnote{See, \emph{e}.\emph{g}., \citeNP[\S16.5,
pp.~401--402]{misner-et73}, for a description of the device and its
use.} Its fixed center and the ends of its two rotatable axes
continuously occupy at any given moment 5 proximate points, the
attribution to which of values for linear and angular acceleration
yields direct measures of the components of the Riemann tensor in a
normal frame adapted to the position and motion of the instrument.
One then identifies the spacetime points the parts of the instrument
respectively occupy, and by extension those in the normal frame
adapted to it, by the values of the components of the Riemann tensor
and their derivatives in that frame, by the values of its scalar
invariants, and so on.\footnote{See, for example, Bergmann and
Komar~\citeyear{komar-bergmann-pois-brck-loc-obs-gr,bergmann-komar-obs-commut}
for a concrete, albeit purely formal, example of a procedure for
implementing this idea.} One does not have to postulate a prior
metric structure in order to perform the measurements and label the
points, nor need one have already determined the metrical structure by
experiment. Indeed, in the performance of the gradiometer
measurements one determines much of spacetime's metrical structure.
Because the facts of intrinsic physical significance that the values
of the fields and the relations among them embody (is this body in
contact with another?\ does heat flow from that body to this or
vice-versa?), moreover, remain invariant under the action of a
diffeomorphism it follows that the equivalence classes we used to
construct points does so as well. Thus, we can fix all the manifold
structure, including metrical, only up to diffeomorphism, as we
expect. This shows that the construction delivers everything we need
and nothing more.
There is an obvious response to the argument based on this
construction. One may object that, so far from the argument's having
shown that the construction pushes us to attribute independent
existence to spacetime points, it rather suggests that points are
defined only by reference to prior physical systems, and hence exist
in only a Pickwickian sense, dependent on the identifiability of those
physical systems. This objection can be answered by, as it were,
throwing away the ladder. Once one has the identification of
spacetime points with equivalence classes of values of scalar fields,
one can as easily say that the points are the objects with primitive
ontological significance, and the physical systems are defined by the
values of fields at those points, those values being attributes of
their associated points only \emph{per
accidens}.\footnote{\citeN{stachel-mean-gen-covar-hole} provides an
elegant tool for describing the result of such a construction as I
propose and in particular this rebuttal to the proposed objection
(though I should say his work is not related to a project such as
this). In his terms, I have sketched the construction of an
\emph{individuating field} independent of the stipulation of
metrical structure, \emph{viz}., a field or system of fields on
spacetime that suffices for the identification of individual
spacetime points.} I do not pretend to endorse such a move, but I
do not have to. My constructive argument is \emph{ad hominem}.
\section{The Debate between Substantivalists and Relationalists}
\label{sec:subs-rel}
\resetsec
I do not consider the idea of pointless manifolds deep or of great
interest in its own right.\footnote{There are a few questions of
potential interest that accrue to it. Is it possible to determine
the topology of a non-compact manifold by the postulation of a
finite number of simple fields? If so, does the minimum number
depend on a topological invariant? Is it in any case greater than
the number of fields we currently believe to have physical import?}
There are, I am sure, many other constructions in the same spirit. If
one were so inclined, I suppose one could try to take something like
it to give a precise way for a relationalist to characterize the
spacetime manifold.\footnote{See \citeN{butterfield-relat-poss-wrlds}
for a survey of some ways one might attempt such a project.} I am
not so inclined, because I do not think the contemporary debate
between the relationalist and the substantivalist has been well posed,
and I am inclined to think it never will be in any interesting sense.
That is what I take to be the force of the opposed constructions of
\S\ref{sec:limits} and \S\ref{sec:pointless}, taken in tandem. They
show that ``dependence on prior metrical structure'' is formal,
\emph{i}.\emph{e}., without substantive content until given
explication in the framework of an investigative enterprise, even if
that framework be given only in schematic form. Once one grants this,
however, the game is up. Different investigative frameworks can and
do yield natural criteria that lead to contrary
conclusions.\footnote{This line of argument bears fruitful comparison
to the ideas of \citeN{ruetsche-interp-quantth}, though it was
developed independently of her work.}
An amusingly poignant feature of the constructions shows this clearly:
each yields a conclusion contrary to what the traditional debates
would have led one to have expected based on the tools and techniques
it employs. In the second, one uses independent values of physical
quantities (a stock in trade of the relationalist) in order to
identify and attribute existence to spacetime points without a prior
assumption of metric structure; and in the first, one uses structures
in mathematical physics that seem to presuppose the independent
identifiability of spacetime points (a stock in trade of the
substantivalist) in order to argue that in fact they are not
identifiable without a prior postulation of metric structure. One may
think that these features of the arguments make them, in the end,
self-defeating, but I do not think that is so. In the first, one
operates under the implicit assumption that the more complex models
one idealizes are themselves only idealizations of yet more complex
models. In the second, one implicitly assumes that, say, the
gradiometer is small enough and the temporal interval of the
measurement itself short enough to justify the use of the Minkowski
metric in making the initial attributions of the magnitudes of
spatiotemporal intervals in the experiment; one then uses this to
bootstrap one's way to a more accurate representation of the metrical
structure of spacetime, which is what is done in practice. I think
that this facet of the arguments, perhaps more than anything else,
illustrates the vanity of the traditional debate: one can use the
characteristic resources and moves of each side to construct arguments
contrary to it, once one takes the trouble to make the question
precise.
Most damning in my eyes, the constructions show the futility of the
debate, for they make explicit how very little one gains in
comprehension or understanding by having taken the considerable
trouble to have made the questions precise. Indeed, one may feel with
justice that nothing has been gained, but rather something has been
lost in a pettifoggery of irrelevant technical detail.\footnote{Jeremy
Butterfield in particular has vigorously tried to convince me that I
dismiss too readily the possible philosophical value of the
technical constructions and arguments of \S\S\ref{sec:limits} and
\ref{sec:pointless}. I would like to think he is right.}
Although I conclude the traditional debate is without real content, I
think there is a related, interesting question one \emph{can} give
clear sense to: what in one's investigative framework is naturally
taken to, or must one take to, have intrinsic physical significance?
Even putting aside existence and ontology as emotive distractions,
however, I do not think one can give even this question substantive
sense in the abstract: the question is a formal template that one must
give substance to by fixing the significance of its terms in
presumably different ways in different particular contexts.
Consider one way to rephrase the question that may seem on its face to
give it concrete content in abstraction from any schematic framework:
what propositions would all observers agree on? One cannot answer
this question in the abstract, or even give it definite sense, because
one has not yet fixed the way that one will schematically represent
the observer (or experimental apparatus) and the process of
observation. In order to do so, one must settle many questions of a
more concrete nature. Will one use the same theory to model the
observation as one uses to model the system? Will one take the
observer to be a test system, in the sense that the values of its
associated physical quantities do not contribute to the initial-value
formulation of the equations of motion of one's theoretical or
experimental models? And so on. Until one settles such issues, one
cannot even say with precision what any single observer can or will
observe, much more what all will agree on. In this sense, even claims
such as ``in general relativity, only what is invariant under
diffeomorphisms has intrinsic physical significance'' have only
schematic content. One must give definite substance to the ``what''
in ``what is invariant''---substance that involves the forms of the
physical systems at issue and the methods available for their probing
and representation---before one can make the claim play any definite
role in our attempts to comprehend the world. I take this to be the
lesson of \citeN{stein-phil-prehist-gr}, \emph{viz}., that the way to
proceed in these matters is the one Newton and Riemann relied on: we
must infer what we can about the spatiotemporal structure of the world
from the roles it plays in characterizing physical interactions; and
on this basis, neither substantivalism nor relationalism can claim any
great victory.\footnote{\citeN[p.~274]{disalle-dyn-indiscern-st}
trenchantly makes a very closely related point, one, indeed, that in
large part may be viewed as foundational for my analysis:
\begin{quote}
Since the work of Riemann and Helmholtz, however (not to mention
Einstein), it should be clear that our claims about `objective'
spatiotemporal relations always involve assumptions about the
physical processes we use for measurement and stipulations about
how those processes are to indicate aspects of geometry.
\end{quote}}
In the end, why should we ever have expected there to have been a
single, canonical way to explicate the physical significance of the
idea of a spacetime point, on the basis of which we might then attempt
to determine whether such a thing exists or not in some lofty or
mundane sense? What, after all, is lost to our comprehension of the
physical world without such a unique, canonical explication? We
purport after all, in these debates, to attempt to better comprehend
the \emph{physical} world. Hadn't we better ensure, then, that the
terms of our arguments have the capacity to come in some important way
into contact with the physical world by way of experiment and theory?
Once we take that demand seriously, we find an orgiastic crowd of
possible candidates to serve as concrete realizations of the question,
some of which will be fruitful in some kinds of enterprises, others in
others, and, most likely, several in none at all. Indeed, I am far
from convinced that the question of the existence of spacetime points
has ever itself been well posed. I think a necessary (though not
sufficient) condition for the scientific cogency and relevance of that
question is a demonstration that an answer to it would contribute
fruitfully to the proper comprehension of the performance of an
experiment or the proper construction of a model of a physical system
in the context of general relativity. (Recall this paper's epigraph
by Maxwell.) But what possible difference could an answer to it make
one way or another to those scientific issues?
I think there is a better question at hand: what mathematical
structures ``best'' represent our experience of spatiotemporal
localization? Again, this question cannot be answered in the
abstract, for it depends sensitively on the answers to other, more or
less independent and yet inextricable questions, such as: what
mathematical structures best represent our experience of other
features of spatiotemporal phenomena, such as the lack of absolute
simultaneity, the orientability of space, \emph{etc}.? And also
questions such as: what structures for representation of various kinds
of derivatives do we need to formulate equations of motion? And what
structures for representation of Maxwell fields? And so on. One has
to attempt to address these questions in a dialectical fashion,
answering part of one here, seeing what adjustments that requires in
other parts of the manifold of possible structures, so to speak, and
so on. The answer to one of these questions in one context may be
individual points of a spacetime manifold, to another question in
another context it may be area and volume operators as in loop quantum
gravity, and so on. Instead of asking whether the manifold itself or
the manifold plus the metric is ``\emph{really} spacetime'', we should
rather be asking what sorts of structure with real physical
significance a manifold by itself and a manifold with a metric can
respectively support---anything requiring only differential topology
or geometry for the former, and anything requiring Lorentzian geometry
for the latter. It is to the investigation of such questions that I
now turn.
\section{An Embarassment of Spacetime Structures}
\label{sec:embarass}
\resetsec
The arguments of this paper extend themselves naturally beyond the
realm of the debate over the existence of spacetime points, and do so
in a way that sheds further light on the futility of that debate.
There are many different senses one can give to the question whether
some putative entity or structure of any type has real physical
significance in the context of general relativity, each more or less
natural in different contexts. For lack of a better term, I shall say
that an entity (which, as we shall see, can encompass several
different types of thing), purportedly represented by a theoretical
structure, has \emph{physicality} if one has a reason to take that
structure seriously in a physical sense, \emph{viz}., if one can show
that it plays an ineliminable, or at least fruitful and important,
role in the way that theory makes contact with experiment. Of course,
as I stressed in \S\ref{sec:hole}, such an abstract, purely formal
schema as ``plays an ineliminable, or at least fruitful and important,
role in the way that theory makes contact with experiment'' has no
real content until one explicates it in the context of a more or less
well delineated investigative framework. It is, in fact, one of the
``important matters on which sensible even if vague things can be
said,'' which Stein discussed in the passage I quoted on
page~\pageref{pg:stein-struc-know}. As such, it is the examples that
give the idea life.
\subsection{Manifest Physicality}
\label{sec:manifest-physicality}
A Maxwell field, represented by the Faraday tensor $F_{ab}$, is
manifestly physical. One important sense in which this is true turns
on the fact that it contributes to the stress-energy tensor on the
righthand side of the Einstein field equation. The Maxwell field
itself possesses stress-energy, and in general relativity nothing is
physical if not that.
Consider now a Killing field on spacetime, a vector field $\xi^a$ that
satisfies Killing's equation
\begin{equation}
\label{eq:kill-eqn}
\nabla_{(a} \hspace*{.1em} \xi_{b)} = 0
\end{equation}
and so generates an isometry, in the sense that $\pounds_\xi \hspace*{.1em}
g_{ab} = 0$. In this guise, it seems not to possess the
characteristics of a physical field, in so far as it enters the
equations of motion of no manifestly physical system, such as a
Maxwell field. In other words, it does not couple with phenomena we
consider physical, does not contribute to the stress-energy tensor.
Now, define the 2-index covariant tensor $P_{ab} \equiv \nabla_a
\hspace*{.1em} \xi_b$. Equation~\eqref{eq:kill-eqn} implies that it is
anti-symmetric. Let us say that it happens as well to have vanishing
divergence and curl, $\nabla_n P^{na} = 0$ and $\nabla_{[a} P_{bc]} =
0$, and so satisfies the source-free Maxwell equations. Is it
\emph{eo ipso} a true Maxwell field, and so physical? Not
necessarily. There are always an innumerable number of 2-forms on a
spacetime that satisfy the source-free Maxwell equations. At most,
one of them represents a physical Maxwell field. If, however, it just
so happened that $P_{ab}$ were to represent the physical Maxwell field
on spacetime---one known as a Papapetrou field in this case---the fact
that one natural way to represent the field happened to generate an
isometry would appear to be an accident, in the sense that no property
of the field accruing to it by dint of its physicality, which is to
say, by dint of its satisfaction of the Maxwell equations and
concomitant coupling with other manifestly physical phenomena (such as
spacetime curvature, by way of the Einstein field equation), depends
on the satisfaction of equation~\eqref{eq:kill-eqn} by $\xi^a$ (except
in the trivial sense that satisfaction of equation~\eqref{eq:kill-eqn}
is necessary for $\xi^a$ to be a 4-vector potential for a Maxwell
field). Still, $\xi^a$ is a naturally distinguished geometrical
structure in the physical description of spacetime, forms a part of
the description of spacetime independent of the particulars of the
physical constitution of any observed phenomena, in particular in so
far as it places non-trivial contraints on a manifestly physical
structure, the spacetime metric. In this sense, different from that
pertaining to the Maxwell field, $\xi^a$ is physical, for the Maxwell
field, by contrast, is not naturally distinguished in this sense, but
rather depends in an essential way on the peculiar, contingent
physical constitution of a particular family of phenomena.
In what sense, though, is the metric manifestly physical? The metric
does not itself contribute to the stress-energy content of spacetime,
for one cannot attribute a localized gravitational stress-energy to
it.\footnote{See, \emph{e}.\emph{g}.,
\citeN{curiel-geom-objs-nonexist-sab-uniq-efe}.} That is not to say
that the metric does not appear in the stress-energy tensor of a given
spacetime, for it is almost always required for the construction of
the stress-energy tensor.\footnote{Indeed, the only example I know of
a stress-energy tensor for which the metric is not needed for its
definition is the case of a null gas, for which only the conformal
structure of spacetime is required. See
\citeN{lehmkuhl-mass-energy-mom-st} for discussion of these issues.}
The stress-energy tensor of a Maxwell field, for example, is $F_{an}
F^n {}_b + \frac{1}{4} g_{ab} F_{rs} F^{rs}$. (The metric appears not
only explicitly in the second term, but also implicitly in both terms,
raising the contracted indices.) The metric, however, is necessary
both for posing the initial-value formulation of every possible kind
of field that may appear in a relativistic spacetime, in particular
all of those (such as the Maxwell field) that we regard as manifestly
physical, and for formulating the equations of motion of the fields.
In particular, the metric dynamically couples with other physical
systems, \emph{i}.\emph{e}., enters into interaction with them in the
strong sense that there always exist terms in the equations of motion
for any given field in which the metric appears as one factor and the
tensor representation of the field as another. For the Maxwell field,
the metric appears contravected with the Faraday tensor in the field
equation representing the fact that its covariant divergence equals
the charge-current density of matter.\footnote{That the other defining
equation for a Maxwell field, representing the fact that the Faraday
tensor is curl-free, does not require the metric at all for its
formulation---the exterior derivative is determined by the
differential structure of the underlying manifold, and does not
require any other structure at all for its definition---may push one
to say that it is not a dynamical equation of motion, but rather a
kinematical constraint.}
The metric, of course, can play other roles as well, just as a Killing
field. A vacuum spacetime with non-zero cosmological constant has a
stress-energy tensor equal to the metric times a constant. In this
case, one plausible way of reading the Einstein field equation is to
have the metric play simultaneously two distinct roles, one as the
necessary ground of all spatiotemporal structure (embodied in the
Einstein tensor) and the other as a component in the tensor
representing the stress-energy content of spacetime, depending on
contingent features of the ambient matter field, in this case,
whatever field gives rise to the cosmological constant. Again, in the
former sense, as ground of spatiotemporal structure, the metric is a
naturally distinguished structure in any physical description of
spacetime; in the latter sense, it rather depends on the peculiar,
contingent physical constitution of a particular family of phenomena.
Consider the Riemann tensor. Again, it manifests physicality in
several different ways, in different contexts. Perhaps the most
important is in the equation of geodesic deviation, where it directly
measures the rate at which infinitesimally neighboring geodesics tend
to converge towards or diverge away from each other. In this case,
the Riemann tensor's physicality consists in the fact that it encodes
all information needed to model manifestly observable phenomena,
\emph{viz}., the relative acceleration of nearby freely falling
particles and the tidal force exerted between different parts of a
freely falling extended body. Another important role it plays in
general relativity is as the measure of the failure of the ambient
covariant derivative operator associated with the spacetime metric to
commute with itself when acting on vectors or tensors. Here, the
physical interpretation is not clear, but one way of trying to
explicate it is by considering the way that a tangent vector changes
when parallel-propagated around an ``infinitesimally small''
loop.\footnote{See \citeN[ch.~2, \S3]{wald-gr} for a thorough
exposition.} The infinitesimal change in the vector when it returns
to the initial point is directly proportional to the Riemann tensor.
Still, it is difficult to say that this has real \emph{physical}
significance, in so far as one could implement such a mechanism and
measure the result only in a spacetime with closed causal curves. And
yet so much of the mathematical apparatus of general relativity
depends on the fact that the ambient derivative operator is, in
general, not flat (\emph{i}.\emph{e}., fails to commute with itself),
that it would be absurd to say that the Riemann tensor is not playing
a physical role here. What exactly that role is, however, is not easy
to pin down. This is an example of the kind of philosophically
important problem whose resolution would have manifest physical
significance that I take Maxwell to be referring to in the passage I
used as one of this paper's epigraphs.
The Einstein tensor itself presents an interesting case. It has no
straightforward geometrical interpretation.\footnote{See
\citeN[\S2.1]{curiel-primer-econds} for a discussion.} It seems,
moreover, to have no straightforward physical interpretation
either---it enters into the equations of motion of no known fields; it
measures no quantitative feature of any known physical phenomena; it
does not represent a field possessing stress-energy; it constrains the
behavior of no other manifestly physical structure; and so on. And
yet it is the structure that matter fields couple to (via the Einstein
field equation) in their role as source for spatiotemporal curvature.
In this role, it dynamically couples with no individual matter fields,
but rather only to the aggregate physical quantity ``stress-energy''
they all possess, and which, according to the fundamental principle of
the fungibility of all forms of energy,\footnote{See
\citeN[ch.~\textsc{v}, \S97]{maxwell-matt-mot} and
\citeN[chs.~\textsc{i, iii, iv, viii,
xii}]{maxwell-theory-heat-1888} for illuminating discussion of
this principle.} in no way differs qualitatively among all known
fields. Again, then, it seems manifestly physical in some sense, but
it is difficult to put one's finger clearly on that sense, and, again,
this is an example of a philosophically important problem whose
resolution would provide real physical insight.
Global structures of various sorts (causal, topological, projective,
conformal, affine, \emph{et al}.\@) present interesting cases as
well.\footnote{I take a structure to be global if it is not local in
the sense explicated by \citeN[p.~55]{manchak-know-glob-struc-st}:
\begin{quote}
[A] condition $C$ on a spacetime is \emph{local} if, given any two
locally isometric spacetimes $(M, \, g_{ab})$ and $(M', \,
g'_{ab})$, $(M, \, g_{ab})$ satisfies $C$ if and only if $(M', \,
g'_{ab})$ satisfies $C$.
\end{quote}
I think Manchak's definition of ``local'' is superior, as judged by
its physical significance in the context of general relativity, to
the one I proposed in \citeN[\S5]{curiel-sing}, though the latter
may still be of interest in purely mathematical contexts, or in
contexts of physical investigation that transcend the scope of a
single theory.} Consider the conformal structure of a spacetime.
It governs and is embodied in the relative behavior of the null cones
across all spacetime points. One natural interpretation of the null
cones is as determining a finite, unachievable upper-limit for the
velocities of material systems.\footnote{See, however,
\citeN{geroch-faster-light} and
\citeN{earman-no-suplum-prop-class-q-flds} for dissenting
arguments.} The fact that the null cones determine a topological
boundary for the chronological future and past of every spacetime
point also has a natural interpretation in the same vein: if the
chronological future or past were topologically closed, then there
would be a limiting upper velocity for massive bodies that would be
\emph{actually achievable} by a massive body using only a finite
amount of energy. If one accepts these interpretative glosses, then
the conformal structure has physicality in so far as it constrains the
behavior of manifestly physical systems.
So, to sum up, the notions of physicality mooted here are:
\begin{itemize}
\item contributes to $T_{ab}$ (\emph{e}.\emph{g}., Maxwell field)
\item required for initial-value formulation of manifestly
physical fields (\emph{e}.\emph{g}., Maxwell field, $g_{ab}$)
\item dynamically couples to manifestly physical entities
(\emph{e}.\emph{g}., Maxwell field, $g_{ab}$)
\item dynamically couples to manifestly physical quantities that
more than one type of physical system can bear (\emph{e}.\emph{g}.,
Einstein tensor)
\item acts as a measure of an observable aspect of manifestly
physical entities (\emph{e}.\emph{g}., Riemann tensor)
\item enters the field equation of a manifestly physical structure
(\emph{e}.\emph{g}., Einstein tensor)
\item constrains the behavior of a manifestly physical entity
(\emph{e}.\emph{g}., Killing field, conformal structure)
\item plays an ineliminable, though physically obscure, role in
the mathematical structure required to formulate the theory
(\emph{e}.\emph{g}., Riemann tensor, Einstein tensor)
\end{itemize}
I am confident there are yet more senses of physicality I have not
touched upon.
\subsection{Observability}
\label{sec:observ}
One does not have to be an instrumentalist or an empiricist to accept
that the possible observability of physical phenomena is one of the
most fundamental reasons we have to think such things are physical in
the first place. The question of the observability of various kinds
of global structure in general relativity, therefore, poses
particularly interesting problems for arguments about physicality.
Manchak~\citeyear{manchak-know-glob-struc-st,manchak-phys-reas-st}
shows that, in a precise sense, local observations can never suffice
to determine the complete global structure of spacetime in general,
and in particular cannot determine whether a spacetime is inextendible
or stably causal \cite[p.~418, proposition~3]{manchak-phys-reas-st}.
Nonetheless, there remain several things to say and ask about the
matter of physicality here.
Take, for example, the Euler number of the spacetime manifold, a
global topological structure.\footnote{See, \emph{e}.\emph{g}.,
\citeN[ch.~\textsc{viii}]{alexandrov-comb-topol-2}.} It is a
topological invariant that, in part, constrains the possible existence
of everywhere non-zero vector fields on a manifold. That an
even-dimensional sphere, for example, possesses no everywhere non-zero
vector field (and indeed no Lorentzian metric) follows directly from
the computation of its Euler number. If we were to live in a world
whose underlying manifold possessed a non-trivial Euler number, and so
could support no physical process that would manifest itself as an
everywhere non-zero vector field, this would constitute a physical
fact about the world in an indubitable sense. It is not clear to me,
however, whether in some precise sense the Euler number of the
spacetime manifold could ever be determined by direct observation.
The orientability of spacetime is an example of a global topological
structure that seems to be strictly inobservable in an intuitive
sense. This follows from the fact that one can construct an
orientable manifold from any non-orientable one by lifting the
structures on it to a suitable covering space, which is automatically
orientable. The lift of the spacetime metric to a covering manifold,
however, would yield a representation of exactly the same physical
spacetime as the original: every physical phenomena in the one has an
isometric analogue, as it were, in the other, and vice-versa. Whether
or not a spacetime manifold is simply connected, moreover, seems to be
in the same boat, for the universal covering manifold of a manifold is
guaranteed to be simply connected.\footnote{In order for a manifold to
possess a universal covering manifold, it must be semi-locally
simply connected. Intuitively, this means that it cannot contain
``arbitrarily small holes''. More precisely, it means that every
point in the space has a neighborhood such that every loop in the
neighborhood can be continuously contracted to a point. (The
contraction need not occur entirely with the given neighborhood.)
The so-called Hawaiian Ear-Ring is an example of a topological space
that is not semi-locally simply connected
\cite{biss-genl-approach-fund-grp}. Whether or not a spacetime
manifold is semi-locally simply connected presents us with yet
another type of question related to physicality: strictly speaking,
there is no physical need for a manifold to possess a universal
cover, and it is difficult, to say the least, to see what other
physical relevance being semi-locally simply connected could have;
and yet the construction of the universal cover is such an
extraordinarily useful theoretical device \cite{geroch-topol-gr}
that one wants to demand that a candidate spacetime manifold be
semi-locally simply connected. What status does such a demand have?
A purely pragmatic one?}
Nonetheless, I think those answers about the possible observability of
a manifold's orientability and simple connectedness may be too pat.
If one were to observe that any member of a certain family of closed,
physically distinguished spatiotemporal loops could not be
continuously deformed into any member of another family of closed,
physically distinguished spatiotemporal loops, one would have shown
that the spacetime manifold is not simply connected. Similarly, if
one could show that to parallely propagate a fixed orthonormal tetrad
around a given closed spatiotemporal loop would result in its
inversion, one would have demonstrated that spacetime is not
orientable. I personally have no idea what sorts of experiment could
show either of those things. The history of physics, however, if it
shows us nothing else, does show us never to underestimate the
ingenuity of experimentalists, no matter what the theoretician may
tell them is impossible to observe or measure.
The first Betti number of the spacetime manifold offers another
interesting example of this sort. The first Betti number of a
topological space is the number of distinct connected components it
has; any manifold with a first Betti number greater than one is
\emph{ipso facto} not connected. Say that we posited a non-connected
spacetime manifold. According to the principles of general
relativity, any phenomena in one component would be strictly
inobservable in any other.\footnote{Perhaps one could posit some form
of quantum entanglement among phenomena in the different components.
The ramblings of many theorists of quantum gravity notwithstanding,
such a possibility lies so far beyond the ambit of current well
entrenched experimental technique and well founded theoretical
knowledge as to render it incomprehensible as physics. By the
nature of the case, for instance, we could perform no direct
experiments on the putatively entangled phenomena in the postulated
other component to verify the entanglement beyond a shadow of a
doubt.} By this criterion, it makes no sense to attribute
physicality to regions of spacetime disconnected from our
own.
So, are these possibly inobservable global structures physical? Well,
it seems to me that in one sense they are, and in others they are not.
The only lesson I want to draw here is that questions of this sort
require in-depth investigation sensitive both to the technical details
of the mathematics and to the physical details of how such structures
may and may not bear on other phenomena we think of as manifestly
physical, even if they turn out to be indubitably
inobservable.\footnote{The family of phenomena in relativistic
spacetimes grouped under the rubric ``singular stucture'' (or
``singularities'') provides on its own a rich and diverse selection
of examples, which I do not have room even to sketch here. See
\citeN{curiel-sing} for an extended discussion.}
\subsection{Physicality and Existence}
\label{sec:phys-exist}
What I have discussed so far in this section, I submit, are
philophically rich, scientifically significant questions and
arguments, of the sort Maxwell mentions in the epigraph to this paper.
Insight into and progress on any of the questions would constitute
real progress in our attempts to understand the world in a scientific
sense. The sterility of the current debate between substantivalists
and relationalists is shown in the fact that no questions it addresses
has scientific value in the sense of Maxwell---it has spurred no work
with direct scientific, as opposed to purely metaphysical, import.
Still, No matter how convincing or interesting or philosophically rich
these examples and arguments may be, one might still want to respond
that they show nothing about the possible \emph{existence} of
spatiotemporal entities, and so in the end they do not bear on the
debate between substantivalism and relationalism. I do not think that
is the correct lesson to leave with, though. I take physicality to be
a necessary condition for the attribution of existence to a
theoretical entity. If there are many possible ways an entity can
manifest physicality, therefore, and one can show that different
entities manifest some but not others of them, then it follows that it
is meaningless to attribute existence \emph{simpliciter} to such
theoretical entities. If there are two entities each manifesting a
different type of physicality, then, in so far as each is a necessary
condition for existence, if one attributes existence to those
entities, it must be of a different sort for each. Thus, in so far as
one wants to make sense of the idea of ``existence'' in the context of
physical entities purportedly represented by theoretical structures
(if that is the sort of thing one likes to do), it cannot be univocal.
To paraphrase Aristotle, existence is said, if at all in physics, in
many ways.
What light, if any, does all this shed on the cogency of the
traditional debate about the ontic status of spacetime? I think quite
a bit. A spacetime point is not physical in any of the ways I have
explicated: there is no such thing as an initial-value problem for
them; there is no equation of motion for them; no property of theirs
dynamically couples to any physical field; and so on. How, then, is
one supposed to try to answer the question of whether or not they
exist in any way that purports to be grounded in physics?
\section{Valedictory Remarks on Realism and Instrumentalism, and the
Structure of Our Knowledge of Physics}
\label{sec:valedict}
\resetsec
I think my conclusions about the vanity of metaphysical argumentation
abstracted from the pragmatics of the scientific enterprise carry over
into the general debate over realism and instrumentalism. Indeed, I
consider the argument about relationalism and substantivalism to be an
instance of the more general form of argument one can give for
existence claims about entities and structures in science. I will
consider two examples to make the point, the first somewhat
sophisticated, the second quite simple.
Consider, first, the Unruh effect.\footnote{See \citeN{wald-qft-cst}
for a rigorous exposition of the phenomenon.} The effect, roughly
speaking, is as follows. (I discuss it only in the context of a
special case, but this does not affect the point.) Consider two
observers in Minkowski spacetime pervaded by a scalar quantum field in
its vacuum state. Each observer carries a simple particle-detector
coupled to the field, with two states: an excited state (``particle
observed''), and a ground state (``particle not observed''). Both
detectors are initially in the ground state. The first observer
follows a geodesic, and so does not accelerate; in this case, quantum
field theory predicts that the particle detector will remain in the
ground state, \emph{i}.\emph{e}., the probability that he will detect
any particles is zero, as one would expect on physical grounds, since
the background field is in the vacuum state. The second observer,
however, begins to accelerate. Now, there is a high probability that
her detector \emph{will} change from the ground to the excited state;
she will ``see particles''. That is the Unruh effect. Even though
the two observers disagree on whether there are particles or not, they
both agree that the state of the second particle detector changes, so
there is a physical fact of the matter in that sense.\footnote{Roughly
speaking, the resolution of the paradox turns on the fact that an
accelerating system in Minkowski spacetime occupies a negative
energy-state: the accelerating detector, in dropping to an energy
level beneath that of the ambient vacuum, registers the vacuum as
having positive energy, which the accelerating observer interprets
as having ``detected a particle''; the inertial observer, however,
accounts for the drop in the accelerating detector's energy by
concluding that it \emph{emitted} a particle, and so changed its
state. If one likes, one may take this as one way to make precise
the idea that ``particle'' is not a natural notion in quantum field
theory, and is indeed at times not only not useful but downright
obfuscatory.} Now, the bit of most interest to us is that the
fluctuations in the field that determine the change in the state of
the detector do not contribute to the definition of the stress-energy
tensor. All observers, both inertial and accelerating, will still
conclude that the ambient stress-energy tensor is that of the vacuum
state. Is Unruh radiance, then, physical or not? Is it ``real''
radiation? Well, in the sense that it is a phenomenon that all
observers will agree on, one that manifests itself in directly
observable effects, yes; in the sense that it does not contribute to
the stress-energy of spacetime, no.
Now, consider the question ``do electrons exist?'' On its face, it
seems immune to the sorts of problems I raise about the ontic status
of spatiotemporal structure. Surely one can attribute canonical
significance to the question ``do electrons exist?'' independent of
investigative framework? In fact, one cannot. Think of the different
contexts in which the concept of an electron may come into play, and
the natural ways in those contexts one may want to attribute
physicality (or not) to electrons. A small sample:
\begin{itemize}
\item as a component in a quantum, non-relativistic model of the
Hydrogen atom
\item as an element in the relativistic computation of the Lamb
shift
\item as a possible ``constituent'' of Hawking radiation in an
analysis of its spectrum
\item as a measuring device in the observation of parton structure
from deep inelastic scattering of electrons off protons, as modeled
by the Standard Model
\end{itemize}
In the first case, one may want to attribute physicality to the
electron in so far as its associated quantities enter into the
initial-value formulation of the system's equations of motion; in the
second, one may base the attribution on the fact that one identifies
the electron as the bearer of definite values for the kinematic
Casimir invariants of spin and mass; there is no good definition in
general of an electron in the third, because there is no unambiguous,
physically significant definition of ``particle'' in quantum field
theory on a curved spacetime, and so \emph{a fortiori} no way to
attribute physicality to such a thing;\footnote{In essence, this is
because one has no privileged group of timelike symmetries in a
generic spacetime, as one has in Minkowski spacetime, on which to
ground the notion of a particle. See \citeN{wald-qft-cst} for a
detailed explanation.} in the fourth and final case, one can
attribute physicality to the electron because one can associate
localized charge, spin and lepton number with the mass-energy
resonance that represents it. Now, one cannot even formulate in a
rigorous, precise way (and, indeed, often not even in a loose and
frowzy way) the criterion for physicality in any of these frameworks
in the terms of at least some of the others.
It follows that even in this case any formulation of the question in
abstract terms, such as ``what all observers agree on'' or ``what has
manifestly observable effects'' or ``what couples with other systems
we already think of as physical'' or ``what is essential to the
formulation of the theory'', remains empty until one renders content
to it by the fixation of a framework, even if only schematic. To be
clear, I do not claim that one must always make the investigative
framework of one's work explicit, only that one ought to recognize it
must be there in the background, specifiable when push comes to shove,
as it will from time to time.
In the picture I have implicitly relied on in the construction of my
arguments, the structure of physics may be thought of as something
like a differential manifold itself, with different techniques and
concepts that find appropriate application in different sorts of
investigation, and even in similar sorts of investigation of different
subject matters, all covering their own idiosyncratic patches of the
global manifold, consonant with each other when they overlap but with
none necessarily able to cover the entirety of the space. In that
vein, I am confident there are many other interesting senses one can
render to the idea of the physicality of putative entities and
structures represented by our best physical theories, variously useful
or at least illuminating in investigations of different sorts. In
some of those senses, one will rightly, or at least usefully or
suggestively, say those things are physical. In others, one will not.
The words we use to further all the sorts of scientific and
philosophical investigations we pursue do not matter, only the
concepts behind the words, some of which find natural application in
some investigations and some of which do not.
This is not instrumentalism. Among other things, I neither make nor
rely on any privileged claim about how one ought to understand the
structures of our best theories as formal systems, the terms and
relations with which we formulate them, and their broader or deeper
relation to the world itself, only about how we ought not understand
them. The greatest physicists have always, it seems to me, had the
capacity to to think in both realist and instrumentalist ways about
both the best contemporary theories and the most promising lines of
theoretical attack as they were being developed. Often, they held
both sorts of views in their minds at the same time, keeping many
avenues open, sometimes moving forward along one, sometimes switching
to another, sometimes straddling the line, as best befit the demands
of the investigation, with a concomitant gain in richness of
conception and depth of
thought.\footnote{Stein~\citeyear{stein-phil-prehist-gr,stein-locke-huy-newt,stein-poincare}
forcefully argues this line of thought.} In some contexts and for
some purposes it is most useful to conceive, think and speak in
realist terms, and in others to do so in instrumentalist terms. They
are both good in their place, and neither is correct \emph{sub specie
{\ae}ternitatis}. In any event, what I sketch here is
\emph{certainly} not anti-realism.
I am not against asking questions that, in traditional terms, seem to
bear on issues of realism and instrumentalism. I am against the focus
on the questions as meaningful and valuable in themselves, without
regard to the roles they may or may not play in the ongoing enterprise
of our scientific attempts to comprehend the physical world. That
focus, it seems to me, leads only to a sterile form of ideological
back-and-forth that has all but crowded out the possibility of
formulating and addressing questions of real scientific and
philosophical clarity and value. I take that to be the thrust of the
epigraph from Maxwell at the head of this paper.
\addcontentsline{toc}{section}{\hspace*{-1.3em}\numberline{}References}
|
\section{Introduction}
The observational evidence of multiple stellar populations in Galactic globular
clusters (GCs) was already available since the late '70s, in the large
star-to-star abundance variations of light elements (C, N, O, Na, Mg, Al)
following patterns of correlations and anticorrelations (see the reviews by
Smith 1987, Kraft 1994, Sneden 2000, Gratton, Sneden \& Carretta 2004).
However, only in the late '80s/early '90s Denisenkov \& Denisenkova (1989) and
Langer et al. (1993) were able to explain the observed abundance variations with
the simultaneous action of the Ne-Na and Mg-Al chains in the same temperature
layers where the ON part of the complete CNO cycle was at work.
The
discovery of Na-O and Mg-Al anticorrelation among turn-off and subgiant (SGB)
stars first in NGC~6752 (Gratton et al. 2001) and then in other GCs (M~71,
Ram\'irez \& Cohen 2002; M~5, Ram\'irez \& Cohen 2003; 47~Tuc, Carretta et al.
2004) provided the key spectroscopic evidence of multiple populations. These
unevolved stars cannot either synthesise or transport these elements to the
surface, thus they must be formed from gas polluted by ejecta of more massive
stars of an earlier stellar generation. Hence, ``observing an anticorrelation
among proton-capture elements" is simply a rewording for ``seeing multiple
stellar populations".
The advent of efficient, high-multiplexing spectrographs permitted, in the last
10 years, to collect precise abundances for $large$ samples of stars from high
resolution spectroscopy. We exploited FLAMES@VLT and designed a large, systematic
survey of RGB stars in Galactic GCs (for results, see e.g., Carretta et al.
2006, 2009a,b,c, 2014a,b,c, and references therein; D'Orazi et al. 2010a, 2014;
Gratton et al. 2006, 2007). This investigation was recently extended to other
evolutionary phases like horizontal branch (HB), SGB and dwarf stars (e.g.,
D'Orazi et al. 2010b; Gratton et al. 2011, 2012a, 2014). Other, independent studies
concentrated mostly on the closest GCs like 47~Tuc (e.g. Cordero et al. 2014,
Dobrovolskas et al. 2014), M~4 (e.g. D'Orazi \& Marino 2010, D'Orazi et al.
2013, Marino et al. 2008), NGC~6397 (e.g. Lind et al. 2008, 2009; Gonz\'alez
Hern\'andez et al. 2009; Lovisi et al. 2012), NGC~6752 (e.g. Yong et al. 2005,
2013; Gruyters et al. 2014, Shen et al. 2010), and M~22 (Marino et al. 2011,
2012, 2013). The literature is vast\footnote{We concentrate on spectroscopy in
the present paper. However, the improvements in precision photometry from ground
or space-based facilities produced colour-magnitude diagrams (CMDs) of GCs
revealing more and more split SGBs, RGBs, and even multiple main sequences
(e.g., Grundahl et al. 1998, Han et al. 2009, Lee et al. 2009, Monelli et al.
2013, Milone et al. 2012,2013, Nardiello et al. 2015 and references therein).
The synergy between spectroscopy and photometry allows to uncover how absorption
features due to the involved light elements affect the colour indexes of split
sequences (e.g. Carretta et al. 2011a, Sbordone et al. 2011, Milone et al. 2012,
Cassisi et al. 2013).} and the reader is directed to the review by Gratton et
al. (2012b) and to the references in the more recent papers.
Our homogenous analysis of a large sample of GCs and large numer of stars in
each GC permits a quantitative approach (e.g. Carretta et al. 2009a,b), showing
that the Na-O anticorrelation differs from cluster to cluster in shape and
extent (measured by the interquartile range -IQR- of the [O/Na] ratio,
introduced by Carretta 2006). Adding also literature data, we see that the Na-O
anticorrelation stands out as the chief chemical signature of multiple stellar
populations, so widespread that it may even be used to define a GC and that the
phenomenon is primarily driven by the cluster mass (Carretta et al. 2010a).
However, our ignorance about the still elusive nature of first
generation polluters hampers our vision. Massive stars, either single, and fast
rotating: (FRMS: Decressin et al. 2007) or in close binary systems (de Mink et
al. 2009), intermediate-mass asymptotic giant branch (IM-AGB) and super-AGB
stars (Ventura et al. 2001, D'Ercole et al. 2012), or none of them (Bastian et
al. 2013), have been proposed as polluters. It is even possible that $both$ the
two commonly favoured candidates (FRMS and IM-AGB) were at work. In two GCs
where RGB stars were recently found clustered into discrete groups according to
their distinct composition (NGC~6752: Carretta et al. 2012; NGC~2808: Carretta
2014) the population with intermediate chemistry cannot be reproduced by mixing
unprocessed and heavily polluted matter from the primordial and the extremely
processed components, calling for at least two kind of polluters.
Our approach to these problems has been to gather an unprecedented large
statistics, both in term of number of GCs and of stars for which homogeneous
abundances were derived in each GC. We studied
24 massive GCs with different global parameters (metallicity, age, Galactic
population, HB morphology, concentration, etc.) and we present here the
abundance analysis of the last of them, NGC~6093 (M~80).
M~80 is a metal-poor ([Fe/H]$=-1.73$ dex, Cavallo et al. 2004), moderately
massive ($M_V=-8.23$ mag, Harris 1996, 2010 edition) cluster of Oosterhoff type
II (Kopacki 2013), one of the 30 densest GCs in the Milky Way. In the central
regions of M~80 Ferraro et al. (1999) identified over 300 blue straggler stars,
despite the relatively low inferred collision rate. Their inference was that
M~80 could be in an unusual dynamical state, on the edge of core collapse where
the density has recently become large enough to increase the
encounters involving primordial binaries, triggering a burst of blue straggler
formation (but see Heinke et al. 2003 for a different view, based on the
comparison with 47~Tuc and its possibly uncomplete census of blue stragglers).
The HB morphology is characterised by an extended blue tail (Ferraro et al.
1998), with the concentration of blue tail stars increasing toward the cluster
centre (Brocato et a;. 1998). M~80 presents a typical halo-type orbit,
with a short period and small size, never leaving the inner 3-4 kpc of the
Galaxy (Dinescu et al. 1999).
The only chemical evidence of multiple populations comes from the study of ten
RGB stars by Cavallo et al. (2004) who found a mean [Al/Fe] ratio
of +0.37 dex with a large spread of 0.43 dex (1 sigma) in aluminum abundance.
Recently Monelli et al. (2013) used private photometry from their SUMO project
to study the colour spread of the RGB.
The paper is organized as follows: in \S2 we present the basic data for
this cluster, the selection of stars to be observed, and the observations
themselves.
In \S3 we describe the analysis method and the results for abundances are
presented in \S4, whereas they are discussed in \S5 in the context of multiple
populations in GCs. Finally, in \S6 we summarize our findings.
\section{Observations}\label{obs}
The photometric catalog for NGC~6093 is based on $BV$ data collected at the
Wide-Field Imager at the 2.2-m ESO-MPI telescope on 9 July 1999. These data were
reduced by YM in a standard way (see Carretta et al. 2014a, Momany et al. 2004
for details) and the photometry is unpublished.
Following our standard procedure, we selected a pool of stars lying near the RGB
ridge line in the CMD and without close neighbours.
The stars in our spectroscopic sample are indicated as large symbols in
Fig.~\ref{fig1}(a), with a zooming of the RGB region in
Fig.~\ref{fig1}(b), where we also indicate the membership status of our
targets (based on the analysis of their spectra). NGC~6093 is heavily contaminated by field stars. The excision of
non-members was generally done via radial velocity (RV); however, in some cases,
we discarded stars on the basis of their metallicity (see next sections).
\begin{figure}
\centering
\includegraphics[bb=20 150 330 715, clip, scale=0.825]{fig1.eps}
\caption{(a) CMD of the cluster, with FLAMES targets indicated by red, open
symbols. (b) Enlargmenent of the RGB region, with member stars indicated by
filled red dots, non members on the basis of RV by cyan open stars, non members
on the basis of abundance by open blue triangles.}
\label{fig1}
\end{figure}
\begin{figure}
\centering
\includegraphics[bb=20 150 288 706, clip, scale=0.825]{fig2anto.eps}
\caption{(a) Histogram of RV values for all stars observed. (b) Plot of RV as
function of distance from the cluster centre; in red we indicate the member
stars. The lines show the average RV and the $\pm3\sigma$ limits.}
\label{fig2}
\end{figure}
We used FLAMES@VLT and the log of the observations is given in
Table~\ref{t:logobs}. We obtained three exposures with the high resolution
GIRAFFE grating and the setup HR11 covering the Na~{\sc i} 5682-88~\AA\ doublet
and three exposures with the setup HR13 including the [O~{\sc i}] forbidden
lines at 6300-63~\AA \ and another Na~{\sc i} doublet at 6154-60~\AA. We
observed a total of 14 giants with the fibres feeding the UVES spectrograph (Red
Arm, with spectral range from 4800 to 6800~\AA\ and R=47,000). All the 14 stars
turned out to be members. We observed 78 member stars of M~80 with GIRAFFE; ten
are in common with UVES, so the grand total of our sample consists of 82 red
giant stars.
We used the 1-D, wavelength calibrated spectra as reduced by the ESO personnel
with the dedicated FLAMES pipelines.
RVs for stars observed with the GIRAFFE spectrograph were
obtained using the {\sc IRAF}\footnote{IRAF is distributed by the National
Optical Astronomical Observatory, which are operated by the Association of
Universities for Research in Astronomy, under contract with the National Science
Foundation } task {\sc FXCORR}, with appropriate templates, while those of the
stars observed with UVES were derived with the {\sc IRAF} task {\sc RVIDLINES}.
The multiple exposures were then combined.
The median values of the S/N ratio of the combined spectra are 125, 68, and
78, for observations with UVES and with the GIRAFFE HR11 and
HR13 setups, respectively.
Our optical $B, V$ photometric data were integrated with $K$ band magnitudes
from the Point Source Catalogue of 2MASS (Skrutskie et al. 2006) to derive
atmospheric parameters as described in Section 3.
Coordinates, magnitudes, and heliocentric RVs are shown in Table~\ref{t:coo60}
(the full table is only available in electronic form at CDS).
\begin{table}
\centering
\caption{Log of FLAMES observations for NGC 6093.}
\begin{tabular}{cccccc}
\hline\hline
Date & UT & exp. & setup & seeing & airmass\\
(yyyy-mm-dd) & (hh:mm:ss) & (s) & & (") & \\
\hline
2009-09-08 & 00:28:42.650 & 2940 & HR11 & 0.89 &1.237 \\
2009-09-09 & 00:50:31.604 & 2940 & HR11 & 0.91 &1.343 \\
2009-08-06 & 00:16:09.457 & 2940 & HR11 & 0.95 &1.002 \\
2009-09-10 & 00:02:22.285 & 2940 & HR13 & 1.44 &1.177 \\
2009-09-12 & 23:40:18.289 & 2940 & HR13 & 0.86 &1.149 \\
2009-08-06 & 00:16:09.457 & 2940 & HR13 & 0.94 &1.330 \\
\hline
\end{tabular}
\label{t:logobs}
\end{table}
According to Harris (1996, 2010 edition), the metallicity of NGC~6093 is $-1.75$
and the RV is $8.1\pm1.5$ km~s$^{-1}$ .
Our choice of member stars rests both on RV and, for marginal or dubious cases,
also on metallicity. Figure~\ref{fig2} shows the histogram of the heliocentric
RVs (in panel a) and the run of RVs with distance from the cluster centre (panel
b). We selected as candidate cluster members 82 stars for which both RV and
metallicity were acceptable (see Table~\ref{t:coo60}). Not surprisingly, they
are concentrated at small distances from the cluster centre. For them we find
an average velocity of $\langle RV \rangle=11.9\pm0.7$~km~s$^{-1}$ ($rms=$ 6.1
km~s$^{-1}$), in good agreement with the value reported by Harris (1996).
\subsection{Kinematics}
The spectroscopic dataset presented here constitutes the largest sample of
intermediate/high resolution spectra collected so far for NGC 6093, and can
therefore be used to study the kinematical properties of this cluster.
In Fig.~\ref{f:rot} the radial velocities of the member stars are plotted
against their position angle. To test the possible presence of systemic rotation
we calculated the difference between the average velocity of stars located
within and outside a $180^\circ$ interval centered around various position
angles, and compared it with a set of Monte Carlo extractions of the same number
of stars randomly distributed across the field of view. We found no significant
signature of rotation with a maximum difference lying at $\sim 1 \sigma$ from
the mean distribution of randomly extracted stars. The bestfit sinusoidal curve
overplotted to Fig.~\ref{f:rot} has a maximum amplitude of
$A_{rot}=1.0\pm0.6~$km~s$^{-1}$, corresponding to a ratio
$A_{rot}/\sigma_{0}=0.09\pm0.06$ (see below). This value places NGC~6093 in the
group of non/slowly rotating GCs of Bellazzini et al. (2012).
\begin{figure}
\centering
\includegraphics[scale=0.40]{rot.eps}
\caption{Radial velocities of member stars in M~80 as a function of their
position angle.}
\label{f:rot}
\end{figure}
We then fitted the radial distribution of velocities with a set of single-mass
King (1966) and multi-mass King-Michie (Gunn and Griffin 1979) isotropic models.
For this purpose, the density profile has been derived by counting stars
brighter than the turn-off point (well above the limiting magnitude of the WFI
photometry where severe incompleteness affects star counts) in
concentric annuli at different distance to the cluster center. The obtained
density profile has been then matched with that provided by Noyola and Gebhardt
(2006) from integrated photometry on Hubble Space Telescope images to sample the
innermost 1.7\arcmin\ where the WFI spatial resolution does not allow a complete
sampling of stars. The total cluster $V$ magnitude, obtained by integrating the
observed profile, turns out to be $V=7.48\pm0.05$ corresponding to a total
luminosity of $log L/L_{\odot}=5.18\pm0.09$.
The velocity dispersion profile has been derived by dividing the sample in four
radial bins containing $\sim20$ stars each.
\begin{figure*}
\centering
\includegraphics[bb=39 154 572 430, clip, scale=0.52]{prof.eps}
\caption{Density and velocity dispersion profiles (left and right panel,
respectively) in M~80. The best-fit models are supersimposed (solid red line).}
\label{f:prof}
\end{figure*}
Multi-mass models has been constructed by assuming nine
mass bins and a mass function derived iteratively to match the observed star
counts in the deep ACS photometry of NGC~6093 from the {\it ACS globular
clusters treasury project} (Sarajedini et al. 2007) corrected for completness
(see Anderson et al. 2008). We adopted the same procedure described
in Sollima et al. (2012) to account for dark-remnants and binary heating.
The dynamical masses of the best-fit models have been
calculated using the maximum likelihood algorithm described by Pryor and Meylan
(1993) applied to our sample of radial velocities. The estimated central line of
sight (LOS) velocity dispersions and the corresponding dynamical masses are
$\sigma_{v,0}=11.4\pm0.9$~km s$^{-1}$ and $\log{M/M_{\odot}}=5.50\pm0.07$ for the
single-mass model and $\sigma_{v,0}=10.1\pm0.8$~km s$^{-1}$
$\log{M/M_{\odot}}=5.56\pm0.07$ for the multi-mass model. The bestfit models are
overplotted to the density and velocity dispersion profiles in Fig.~\ref{f:prof}.
It is apparent that both models provide a good fit of both profiles. The
resulting M/L ratio is $2.1\pm0.6$, slightly larger than that predicted for a
mass-follow-light stellar system with the mass function slope estimated in the
literature ($\alpha=-1.36$; Paust et al. 2010), although the large associated
error prevents any firm conclusion. This is in agreement with what found
by Sollima et al. (2012) for a sample of six Galactic GCs and interpreted as a
consequence of a high retention efficency of dark remnants and/or to tidal
heating. In this regard, NGC~6093 follows an eccentric orbit (e=0.85) in a inner
region of the Milky Way (Dinescu et al. 1999) being therefore subject to an
intense tidal stress which can inflate its velocity dispersion even within its
half-mass (K\"upper et al. 2010).
\section{Atmospheric parameters, metallicity and abundance analysis}
In the abundance analysis we used mostly equivalent widths ($EW$s), measuread as
described in details in Bragaglia et al. (2001) using the package ROSA (Gratton
1988). The $EW$s measured on the GIRAFFE spectra were shifted to a system
defined by $EW$s from the high resolution UVES spectra. For this correction we
used 275 lines measured in the 10 stars observed with both instruments.
Abundances were derived from $EW$ analysis, using the Kurucz (1993) grid of
solar-scaled LTE model atmospheres with the overshooting option switched off.
Effective temperatures T$_{\rm eff}$ were derived following our usual
two-steps procedure: first pass values were obtained from $V-K$ colours and the
Alonso et al. (1999, 2001) calibration, and these values were used as inputs to
derive the finally adopted T$_{\rm eff}$'s from a relation between
T$_{\rm eff}(V-K)$ and the star magnitudes. Given the non negligible value of
the reddening for M~80 ($E(B-V)=0.18$ mag, Harris 1996), we employed near
infrared $K$ magnitudes for this relation.
Surface gravities $\log g$ were obtained from bolometric corrections (from
Alonso et al.), the adopted effective temperatures, reddening and distance
modulus from Harris (1996), and assuming masses of 0.85 M$_\odot$ and $M_{\rm
bol,\odot} = 4.75$ as bolometric magnitude for the Sun.
Finally, we obtained values of the microturbulent velocity $v_t$ by eliminating
trends of the abundances from Fe~{\sc i} lines with the expected line strength
(see Magain 1984). The final adopted
atmospheric parameters are listed with the derived Fe abundances of individual
stars in Tab.~\ref{t:atmpar60}.
Errors in atmospheric parameters and their impact on the derived abundances are
estimated as usual (see Carretta et al. 2009a,b and the Appendix to the present
paper). Internal (star to star) errors in temperature, gravity, model abundance
and $v_t$ are 4 K, 0.04 dex, 0.02 dex, and 0.10 km~s$^{-1}$,
respectively\footnote{The internal error in $v_t$
increases to 0.36 km~s$^{-1}$ for stars with GIRAFFE spectra.}.
As in the vast majority of GCs, the metallicity is very homogeneous in M~80. We
find on average [Fe/H]~{\sc i}$=-1.791\pm0.006\pm0.076$ dex ($\sigma=0.023$
dex) from the 14 stars observed with UVES and
[Fe/H]~{\sc i}$=-1.791\pm0.003\pm0.070$ dex ($\sigma=0.023$ dex) from 78 stars
with GIRAFFE spectra (where the first and second error bars refer to the
statistical and systematic errors, respectively). The agreement with average
abundances derived from singly ionized Fe lines is very good:
[Fe/H]~{\sc ii}$=-1.791$ dex $(\sigma=0.016$ dex, 14 stars with UVES) and
[Fe/H]~{\sc ii}$=-1.792$ dex $(\sigma=0.040$ dex, 56 stars with GIRAFFE). The
derived abundances show no
trend as a function of the temperature (Fig.~\ref{f:feteff60}) and are in good
agreement, within the uncertainties, with the average value found by Cavallo et
al. (2004) from 10 giants.
\begin{figure}
\centering
\includegraphics[bb=78 171 460 691, clip,scale=0.52]{feteff60.eps}
\caption{Abundance ratios [Fe/H] {\sc i} (upper panel) and [Fe/H] {\sc ii}
(lower panel) as a function of T$_{\rm eff}$ for all member stars analysed in
M~80 with UVES (blue squares) and GIRAFFE spectra (red circles: setups HR11 and
HR13; black crosses: HR11 only; green crosses: HR13 only). Error bars on the
right and on the left side are star-to-star errors for targets observed with
UVES and GIRAFFE, respectively.}
\label{f:feteff60}
\end{figure}
We measured the abundances of 12 species (beside Fe: O, Na, Mg, Si, Ca, Ti,
Sc, V, Cr, Ni, and Ba) from both UVES and GIRAFFE
spectra. Additionally, from the UVES spectra with large spectral coverage we
also obtained abundances of Al, Ti (from both Ti~{\sc i} and Ti~{\sc ii} lines),
Cr (both from Cr~{\sc i} and Cr~{\sc ii} lines), Mn, Co, Cu, Zn,
Zr (both from Zr~{\sc i} and Zr~{\sc ii} lines), La, Ce, Pr, Nd, Sm, and Eu.
Adopted line list and solar reference abundances are from Gratton et al. (2003).
We did apply NLTE corrections to the Na abundances following the prescriptions
by Gratton et al. (1999). We applied abundance corrections for Sc, V, Mn, and
Co to account for the hyperfine structure (references in Gratton et al. 2003).
\section{Results}
The average values of all measured elements with their $r.m.s.$
scatter are listed in Tab.~\ref{t:meanabu60}. Abundances of
proton-capture, $\alpha-$capture, Fe-group and neutron-capture elements are
given for individual stars in Tab.~\ref{t:proton60}, Tab.~\ref{t:alpha60},
Tab.~\ref{t:fegroup60}, Tab~\ref{t:neutron60} and Tab.~\ref{t:ba60}.
Abundance ratios as a function of temperature are shown in
Fig.~\ref{f:eleteff60}. In this figure we also plot the average [$\alpha$/Fe]
values, where the mean includes [Si/Fe], [Ca/Fe], and [Ti/Fe]~{\sc i}.
\setcounter{table}{3}
\begin{table}
\centering
\caption{Mean abundances from UVES and GIRAFFE }
\begin{tabular}{lcc}
\hline
& & \\
Element & UVES & GIRAFFE \\
&n~~ avg~~ $rms$ &n~~ avg~~ $rms$ \\
\hline
$[$O/Fe$]${\sc i} &14 +0.24 0.24 &54 +0.23 0.21 \\
$[$Na/Fe$]${\sc i} &14 +0.44 0.22 &63 +0.40 0.30 \\
$[$Mg/Fe$]${\sc i} &14 +0.45 0.10 &66 +0.46 0.05 \\
$[$Al/Fe$]${\sc i} &14 +0.48 0.34 & \\
$[$Si/Fe$]${\sc i} &14 +0.34 0.04 &75 +0.36 0.03 \\
$[$Ca/Fe$]${\sc i} &14 +0.36 0.02 &78 +0.36 0.02 \\
$[$Sc/Fe$]${\sc ii} &14 $-$0.03 0.03 &78 $-$0.01 0.03 \\
$[$Ti/Fe$]${\sc i} &14 +0.19 0.05 &64 +0.19 0.02 \\
$[$Ti/Fe$]${\sc ii} &14 +0.18 0.02 & \\
$[$V/Fe$]${\sc i} &14 $-$0.04 0.02 &45 $-$0.04 0.03 \\
$[$Cr/Fe$]${\sc i} &14 $-$0.05 0.03 &34 $-$0.03 0.04 \\
$[$Cr/Fe$]${\sc ii} &14 $-$0.01 0.03 & \\
$[$Mn/Fe$]${\sc i} &14 $-$0.48 0.02 & \\
$[$Fe/H$]${\sc i} &14 $-$1.79 0.02 &78 $-$1.79 0.02 \\
$[$Fe/H$]${\sc ii} &14 $-$1.79 0.02 &56 $-$1.79 0.04 \\
$[$Co/Fe$]${\sc i} &14 $-$0.22 0.03 & \\
$[$Ni/Fe$]${\sc i} &14 $-$0.13 0.02 &69 $-$0.13 0.02 \\
$[$Cu/Fe$]${\sc i} &14 $-$0.52 0.04 & \\
$[$Zn/Fe$]${\sc i} &14 $-$0.02 0.06 & \\
$[$Y/Fe$]${\sc ii} &14 $-$0.07 0.04 & \\
$[$Zr/Fe$]${\sc i} & 8 $-$0.03 0.05 & \\
$[$Zr/Fe$]${\sc ii} &14 +0.00 0.06 & \\
$[$Ba/Fe$]${\sc ii} &14 +0.16 0.16 &59 +0.12 0.28 \\
$[$La/Fe$]${\sc ii} &14 +0.28 0.11 & \\
$[$Ce/Fe$]${\sc ii} &14 +0.18 0.14 & \\
$[$Pr/Fe$]${\sc ii} &14 +0.24 0.09 & \\
$[$Nd/Fe$]${\sc ii} &14 +0.22 0.08 & \\
$[$Sm/Fe$]${\sc ii} &14 +0.40 0.08 & \\
$[$Eu/Fe$]${\sc ii} &14 +0.51 0.02 & \\
\hline
\end{tabular}
\label{t:meanabu60}
\end{table}
\begin{figure*}
\centering
\includegraphics[scale=0.52]{eleteff60.eps}
\caption{Abundance ratios [el/Fe] as a function of T$_{\rm eff}$ for all
stars in our sample. Filled squares and open circles indicate stars with UVES
and GIRAFFE spectra, respectively. Internal error bars are listed in
Tab.~\ref{t:sensitivityu60} and Tab.~\ref{t:sensitivitym60} in the Appendix.}
\label{f:eleteff60}
\end{figure*}
\subsection{Proton-capture elements}
The Na~{\sc i} lines at 6154-60~\AA\ are weak in metal-poor, relatively warm
stars, as discussed in Carretta et al. (2009a).
The possibility that they are
measured only when noise spuriously enhances their strength prompted us to adopt
an empirical parameter ($T_{\rm eff}/100$ K$-10 \times$[Fe/H]) to set rejection
criteria in these cases. We found that 12 stars in M~80 (mostly warmer than
$\sim4800$ K and with [Na/Fe]$\geq 0.5$ dex) with only one or both Na lines
from the weak doublet had the parameter exceeding the value of 65 fixed in
Carretta et al. (2009a). For sake of homogeneity, in these stars we dropped the
Na abundance although all these objects have a measured O abundance that would
put them reasonably well on the Na-O anticorrelation. After this culling of our
sample, we ended with O abundances for 63 RGB stars (44 actual detections and 19
upper limits) and Na abundances for 67 stars. The number of giants with
both O and Na abundances measured in M~80 is 50.
\begin{figure}
\centering
\includegraphics[scale=0.40]{m60antiu.eps}
\caption{The anticorrelation between Na and O abundances in M~80. Empty circles
indicate stars observed with GIRAFFE, filled squares with UVES. Upper limits in
O are indicated with arrows. The error bars represent internal errors (see
Appendix).}
\label{f:m60antiu}
\end{figure}
The spread in [O/Fe] and [Na/Fe] ratios in M~80 is significant, about 1 dex for
both elements. The resulting Na-O anticorrelation (Fig.~\ref{f:m60antiu})
is of moderate extension, reaching down [O/Fe]$\sim -0.5$ dex. This
chemical signature of multiple stellar populations presents the usual features:
a minority of stars with typical composition from core-collapse supernovae (high
O, low Na), separated more or less clearly at [Na/Fe]$\sim +0.3$ dex from the
bulk of stars with increasing Na and decreasing O abundances. Finally, a few
stars show signatures of extreme processing: very low O and high Na levels.
The different components are clearly apparent in the distribution of [O/Na]
ratios, shown in Fig.~\ref{f:histoona}: a dip between 0.0 and $-0.1$ dex separates
the two main populations, followed by a short tail at very low [O/Na] values.
\begin{figure}
\centering
\includegraphics[scale=0.40]{histoona.eps}
\caption{The distribution of the [O/Na] ratios in M~80.}
\label{f:histoona}
\end{figure}
Following the criteria defined in Carretta et al. (2009a) we can identify also
in M~80 a fraction of stars with primordial (P) composition, and two components
of second generation with intermediate (I) and extreme (E) chemistry processed
through hot H-burning. In M~80 the fractions we found are P$=36\pm8$\%
I$=56\pm11$\%, and E$=8\pm4$\%, respectively. We caution that due to the
limitations of the mechanical positioning of fibres our sample is composed of
stars located outside two half mass radii from cluster center. Since the global
population fraction can be best characterized from stars between 1 and 2 half
mass radii, first generation stars could be overrepresented in the observed
sample, as pointed out by the referee. However, we note that the P, I, and E
fractions in M~80 are not exceptional with respect to the values found in other
GCs with a large variety of concentration, mass and dynamical age (see Carretta
et al. 2009a, their Table~5).
\begin{figure*}
\centering
\includegraphics[scale=0.52]{proton60.eps}
\caption{Relations among the abundances of proton-capture elements in red giants
of M~80. Filled squares and empty circles indicate stars observed with UVES and
GIRAFFE, respectively. Internal error bars are also shown.}
\label{f:proton60}
\end{figure*}
The relations among all the abundances available for the light elements O, Na,
Mg, Al, Si in M~80 are shown in Fig.~\ref{f:proton60}. Al was only available for
the 14 stars observed with UVES (one upper limit, 13 detections). Our average
value (+0.48 dex, $\sigma=0.34$ dex) is in good agreement with that from 10
stars in Cavallo et al. (2004: +0.37 dex, $\sigma=0.48$ dex). The [Al/Fe] ratios
present well defined anticorrelations with species depleted in proton-capture
reactions (O, and Mg) and a neat correlation with Na abundances (middle panels
of Fig.~\ref{f:proton60}).
The other elements involved in hot H-burning (Mg, Si) do not seem to be
particularly touched by nuclear processing in now extinct first generation
stars. The only exception is star 13987, observed with UVES, which shows the
lowest Mg and O abundances and, accordingly, the highest Al, Si (but not Na)
abundances in our sample. It is this object which drives the correlation
between Mg and O observed in our sample of stars with UVES spectra (top-left
panel in Fig.~\ref{f:proton60}) as well as the correlation Si-Al and the Mg-Si
anticorrelation (top-right panel and bottom right-panel, respectively, in
Fig.~\ref{f:proton60}).
We conclude (with a cautionary word about the limited size of our UVES
sample) that in M~80 the nuclear processing at extremely hot temperatures
affected only a trace fraction of stars, at odds with what occurred in other GCs,
e.g. NGC~4833 (Carretta et al. 2014b), a cluster with a very similar (present-day)
total mass.
\subsection{$\alpha$ and Fe-group elements}
The run of $\alpha-$elements as a function of the temperature is shown in
Fig.~\ref{f:eleteff60}. Apart for a small increase in the scatter for Mg, the
other elements of this group (Si, Ca, Ti) are very homogeneous in M~80, with
dispersion totally explained by uncertainties in the analysis. In
Fig.~\ref{f:eleteff60} we also plot the mean [$\alpha$/Fe] ratio, obtained
from the average of [Si/Fe], [Ca/Fe], and [Ti/Fe]~{\sc i}: $+0.308\pm0.003$ dex
($\sigma=0.026$ dex, 82 stars). Including also Mg in the mean the value would
have been $+0.342\pm0.004$ ($\sigma=0.032$ dex, 82 stars), another evidence that
only a few stars in M~80 were polluted by material processed at the high
temperatures required for a significant depletion of Mg.
The elements of the Fe-group do not show any intrinsic scatter in M~80, and
usually track iron, apart from Mn and Cu (only measured in stars with UVES
spectra), which present the usual underabundance typical of metal-poor GCs (see
Gratton et al. 2004).
In cases where we measured both neutral and singly ionized transitions for a
species (Fe, Ti, Cr), there is an excellent agreement between the abundances.
\subsection{Neutron-capture elements}
We derived the abundances of a few neutron-capture elements, mainly for stars
with UVES spectra, apart from Ba, for which a transition was measurable also in
stars with GIRAFFE spectra and HR13 setup.
Concerning Ba, the lines available are strong and more sensitive to velocity
fields than to the abundances. To avoid the ensuing trends as a function of the
$v_t$ (whose values are derived from a plaethora of weaker Fe lines) we adopted
the procedure followed in Carretta et al. (2013, 2014a). Abundances of Ba were
obtained using a common [Fe/H] for all stars (equal to the mean value -1.79 dex)
and values of $v_t$ provided by the relation found by Worley et al. (2013).
The abundances from $EW$s of all three Ba lines available in the UVES spectra
were checked with synthetic spectra employing the line
lists by D'Orazi et al. (2012), finding good agreement between the two methods.
The results are shown in the last panel of Fig.~\ref{f:eleteff60} as a function
of temperature: most of the observed spread is accounted for by uncertainties
derived from the analysis (see error tables in the Appendix). The abundances of
Ba are not correlated with any of the proton-capture elements.
We note that star 16162 (with T$_{\rm eff}=4050$ K) stands out with its
[Ba/Fe]=1.001 dex, the highest in all our sample. This star has also the highest
Na abundance and it is characterized by the reddest $B-V$ colour in our dataset
of stars. In Fig.~\ref{f:m60barich55} the spectrum of this Ba-rich star is
compared in the region of the Ba~{\sc ii} 6141~\AA\ line to the spectra of the
three stars with most similar temperatures (and much lower Ba abundances, see
Tab.~\ref{t:ba60}). The spectral lines of star 16162 show a clear broadening
with respect to the other stars. We conclude that this object likely presents
hints of a phenomenon of mass transfer from a binary companion, although its RV
does not show particular differences with respect to other stars.
Anyway, by eliminating this star the average from GIRAFFE spectra changes only a
little, becoming [Ba/Fe]$=+0.11\pm 0.04$ dex ($\sigma=0.25$ dex, 58 stars).
\begin{figure}
\centering
\includegraphics[scale=0.40]{m60barich55.eps}
\caption{The GIRAFFE spectrum of star 16162 (T$_{\rm eff}=4050$ K,
[Ba/Fe]=+1.001 dex) in the region of the Ba~{\sc ii} 6141~\AA\ line (solid line)
compared to stars with similar temperatures: 14224 (dashed line), 17886 (dotted
line), and 17874 (dot dashed line).}
\label{f:m60barich55}
\end{figure}
The results for Y~{\sc ii} were validated through spectral synthesis of the
strong line at 4883~\AA, using a line list from D'Orazi et al. (2013). La and
Ce abundances were obtained from the lines at 6390~\AA\ and at 5274~\AA, using
spectrum synthesis and EWs, respectively, whereas Pr (4 lines) and Sm (2 lines)
were analyzed by using atomic parameters from Sneden et al. (2003) and Koch and
McWilliam (2014). Abundances of Nd were derived as in Carretta et al. (2011b).
\section{Discussion: M~80 in context}
\subsection{Multiple stellar populations in M~80}
The spectroscopic observations in M~80 complete the core sample of our FLAMES
survey of the Na-O anticorrelation in Galactic globular clusters. We have
studied 24 GCs (about 15\% of those listed in the Harris 1996 catalogue): the
range of main parameters covered by our sample is reported in
Tab.~\ref{t:tabrange}, together with the values for M~80.
\setcounter{table}{9}
\begin{table}
\caption[]{Summary of the global parameters sampled in GCs of our FLAMES survey}
\begin{tabular}{lrrr}
\hline
parameter & min & max &NGC 6093\\
\hline
$M_V$ (mag) & -6.64 & -9.98 & -8.23 \\
& NGC 6397 & NGC 6715 & \\
& & & \\
R$_{GC}$ (kpc) & 3.1 & 18.9 & 3.8 \\
& NGC 6388 & NGC 6715 & \\
& & & \\
$$[Fe/H]$$ (dex) & -2.34 & -0.43 & -1.79 \\
& NGC 7099 & NGC 6441 & \\
& & & \\
c & 0.93 & 2.29 & 1.68 \\
& NGC 6809 & NGC 7078 & \\
& & & \\
$\log t_c$ (years) & 4.94 & 8.99 & 7.78 \\
& NGC 6397 & NGC 288 & \\
& & & \\
$\rho_0$ (L$_\odot/pc^3$) & 1.78 & 5.76 & 4.79 \\
& NGC 0288 & NGC 6397 & \\
& & & \\
r$_h$ (pc) & 1.56 & 5.68 & 1.89 \\
& NGC 6397 & NGC 0288 & \\
& & & \\
r$_t$ (pc) & 10.43 & 90.02 & 38.63 \\
& NGC 6838 & NGC 4590 & \\
& & & \\
HB index & -1.00 & 1.00 & 0.93 \\
& NGC 6388 & NGC 6752 & \\
& & & \\
rel. age & 0.84 & 1.06 & 1.03 \\
& NGC 0362 & NGC 7099 & \\
& & & \\
IQR[O/Na] & 0.257 & 1.169 & 0.784 \\
& NGC 6838 & NGC 6715 & \\
& & & \\
\hline
\end{tabular}
\begin{list}{}{}
\item[] - $M_V$, R$_{GC}$, c, $\log t_c$, $\rho_0$ from Harris (1996, 2010 edition).
\item[] - [Fe/H] from Carretta et al. (2009c,2010b,2011b,2013,2014b)
\item[] - c: excluding post-core collapse GCs (NGC 6397,6752,7099), where c=2.5
\item[] - r$_h$, r$_t$, HB index: Mackey \& Van den Bergh (2005)
\item[] - relative age: Carretta et al. (2010a)
\item[] - IQR[O/Na]: our FLAMES survey
\end{list}
\label{t:tabrange}
\end{table}
M~80 is a typical inner halo GC (Carretta et al. 2010a), and also the chemistry
of its multiple stellar populations appears to be normal. The extension of the
Na-O anticorrelation, the most prominent signature of multiple stellar
generations, is quantified by a IQR[O/Na]=0.784. This value locates M~80 right
in the middle of the main relations defined by other GCs in our core sample,
like the correlation between extent of the anticorrelation and total cluster
mass (upper panel of Fig~\ref{f:mviqrres60}).
\begin{figure}
\centering
\includegraphics[bb=121 155 418 708, clip, scale=0.52]{mviqrres60.eps}
\caption{Upper panel: correlation between total absolute magnitude and the
interquartile range of the [O/Na] ratio for GC in our FLAMES survey. Lower
panel: relation between cluster concentration and the residual around the best
fit relation IQR[O/Na]-$M_V$(tot). M~80 is indicated by a filled star symbol. In
each panel the Spearman rank correlation coefficient $r_s$ and the
Pearson linear correlation coefficient $r_p$ are listed.}
\label{f:mviqrres60}
\end{figure}
In Carretta et al. (2014b) we discovered a tight relation (once the post-core
collapse GCs are excluded) linking the cluster concentration and the residuals
around the relation between $M_V$ and IQR[O/Na]. More concentrated clusters seem
to develop a less extended Na-O anticorrelation with respect to more loose GCs.
The anticorrelation of M~80 is in very good agreement with this trend
(Fig.~\ref{f:mviqrres60}, bottom panel).
In M~80 we were able to observe a number of stars below the bump level on the
luminosity function of the RGB ($V=16.12\pm0.03$ mag, Zoccali et al. 1999). As
predicted by Salaris et al. (2006), the luminosity of the RGB bump is sensitive
to the He abundances, being brighter for a population where a He-enhanced
component is present. This prediction has been since a long time observationally
confirmed both with spectroscopy (using the component with Na enhancement in GCs
as a proxy, see Carretta et al. 2007a, Bragaglia et al. 2010) and with
photometry (e.g. Nataf et al. 2011). In Fig.~\ref{f:bumpcumu} we show the
cumulative distribution of the magnitude differences $V-V_{bump}$ for all stars
in M~80 with a measured Na abundance, separating the first generation stars (P
component) from second generation stars (I+E components).
\begin{figure}
\centering
\includegraphics[scale=0.40]{bumpcumu.eps}
\caption{Cumulative distribution of the difference in magnitude between
individual $V$ and $V_{bump}=16.12$ mag for all stars of the first generation
(solid line) and of the second generation (dashed line) with measured Na
abundances.}
\label{f:bumpcumu}
\end{figure}
Although a Kolmogorov-Smirnov test formally does not allow us to reject the null
hypothesis that the two subsamples are extracted from the same parent
distribution, it is clear that regardless of adopted magnitude cut the second
generation stars are on average located at slightly brighter luminosities than
the first generation giants, as they should be according to their He-enhanced
abundances.
In our present sample of giants in M~80 we did not find a statistically
significant difference in radial distribution. However, taking into account also
the physical limitations imposed by the fibres positioning tool, a more suitable
approach to study the radial concentration of different stellar generations is
offered by photometry of large samples over an extended area on the cluster.
The combination $(U-B)-(B-I)$, which Monelli et al. (2013) call $c_{U,B,I}$,
is a good way to study the separation of stars according to different populations in
a GC, sice it does not (significantly) depend on reddening
and distance modulus.
There are 16 objects, including M~80, in common between the SUMO sample and our
FLAMES survey. Unfortunately the photometry of individual stars is still
unpublished, but Monelli et al. provide the width $W_{RGB}$ of the RGB in
$c_{U,B,I}$ that can be used to better investigate the relation between
spectroscopic and photometric properties of multiple stellar populations in M~80
and in other GCs.
For M~80 they derive a value $W_{RGB}=0.14$, which is the same width observed
for NGC~288 and M~4 (NGC~6121). This occurrence is clearly at odds with the
behaviour of these GCs concerning the Na-O anticorrelation, as shown in the
upper panels of Fig.~\ref{f:naohb}, where we plot the [Na/Fe] and [O/Fe] ratios
of each cluster from our strictly homogeneous FLAMES survey.
\begin{figure}
\centering
\includegraphics[scale=0.40]{sumo.eps}
\caption{The width of the RGB from the SUMO survey ($W_{RGB}$) as a function of
the interquartile range IQR[O/Na] from our FLAMES survey for the 16 GCs analyzed
in both samples. M~80 is indicated by a star symbol, under which NGC~288 (blue
filled square) is almost hidden. M~4 is indicated by a filled red circle.}
\label{f:sumo}
\end{figure}
In this figure we indicate with a vertical line the lowest border of the O
abundance measured in M~80, and superimposed this border to the other two plots.
The Na-O anticorrelation derived for M~4 (with IQR[O/Na]=0.373)\footnote{We used
the values derived from our homogeneous work, but other independent studies also
support the evidence that the Na-O anticorrelation is not very extended in M~4
(see e.g. Marino et al. 2008).} is clearly much shorter than those observed in
both M~80 and NGC~288 (IQR=0.784 and 0.776, respectively).
Therefore it is not a surprise that $W_{RGB}$ and IQR[O/Na] are not well
correlated (see Fig.~\ref{f:sumo}).
The Pearson's linear correlation coefficient is $r=0.46$ which with 16 GCs
indicates that the correlation is significant only at a level of confidence
between 90 and 95\%.
These three GCs are however different in other relevant global parameters: the
total metallicity decreases going from M~4 to NGC~288 and to M~80
([Fe/H$=-1.17, -1.30, -1.79$ dex, respectively, from high resolution UVES
spectra, Carretta et al. 2009c, and this work), and the HB morphology
significantly
differs among these clusters, as evident from the CMDs from the ACS survey
(Sarajedini et al. 2007) plotted in the lower panels of Fig.~\ref{f:naohb}.
The HB morphological index (defined as (B-R)/(B+V+R) from the number of stars
B,V,R bluer than, inside, and redder than the instability strip, Lee 1990) is
0.93, 0.98 and -0.06 for NGC~288, M~80, and M~4, respectively (Mackey and van
den Bergh 2005), corresponding to the differences clearly visible in the CMDs of
the lower panels in Fig.~\ref{f:naohb}. The HB stellar distribution in the first
two GCs is all to the blue of the RR Lyrae instability strip, starting almost at
the same colour, but reaching far more fainter (hotter) levels in M~80 than in
NGC~288. On the other hand, the blue part of the HB in M~4 barely ends at the
colour where the HB distribution starts in the two other GCs.
\begin{figure*}
\centering
\includegraphics[bb=15 15 570 500, clip, scale=0.52]{naohb.eps}
\caption{Upper panels, from left to right: Na-O anticorrelation for NGC~288
(Carretta et al. 2009a,b), M~80 (this work), and M~4 (NGC~6121, Carretta et al.
2009a,b). The vertical dotted line indicates the reach of the lowest [O/Fe] ratio
measured in M~80, and it is superimposed to the two other GCs for comparison.
Lower panels: $M_V$ vs $V-I$ CMDs for the three GCs from the ACS survey by
Sarajedini et al. (2007). Distance moduli and reddening values are from Harris
(1996).}
\label{f:naohb}
\end{figure*}
\begin{figure*}
\centering
\includegraphics[bb=19 146 587 500, clip, scale=0.52]{sumores.eps}
\caption{Residuals of the $W_{RGB}$ vs IQR[O/Na] relation as a function of the
HB index and [Fe/H] ratio for the 16 GCs in common with the SUMO survey. Symbols
are as in Fig.~\ref{f:sumo}.}
\label{f:sumores}
\end{figure*}
To better understand the missing piece of information from Fig.~\ref{f:sumo} we
then computed the residuals about the relation between $W_{RGB}$ and IQR[O/Na].
These value are shown in Fig.~\ref{f:sumores} as a function of the HB index and
metallicity, and allow to better understand the difference between photometric and
spectroscopic properties of multiple populations. The residuals are
correlated with HB index and [Fe/H] (which is the $first$ parameter controlling
the star distribution on the HB): both relations are found statistically
significant at a level of confidence of about 99\%.
We conclude that star-to-star abundance variations in the lightest
proton-capture elements, like C and N, are easily revealed as enlargements in
the photometric sequences in the CMDs, through the impact of variations in the
abundance of N, in particular, on the absorption molecular bands in the
ultraviolet or most bluer filters. These variations also include contributions
not due to multiple populations but to the standard evolution of population II
low mass stars: the first dredge-up and the second mixing episode after the RGB
bump both concur to enhance the N abundance, equally in first and second
generation stars. From the study of Monelli et al. (2013), at face value the
changes in the abundances of C and N are comparable in extent for GCs like M~80
and M~4 and NGC~288.
Species like O and Na cannot heavily affect the broad band filters, and this
explain the somewhat loose correlation in Fig.~\ref{f:sumo}. Second generation
stars with O-poor/Na-rich composition are also expected to show N enhancement
(including the contribution from mere evolutive mixing), and this gives the
correlation with the width of the RGB in index like $c_{U,B,I}$. On the other
hand, when O and Na are more changed with respect to the primordial composition
of stars, also He is more and more modified. Taking into account this occurrence
results into a tighter correlation between photometric and spectroscopic
index, and produces the significant correlation with HB morphology, for which He
is the third parameter (see Gratton et al. 2010 and references therein).
As a sanity check, if O and Na are more related to He and the HB morphology we
should expect that M~80 participates to the tight correlation between the
extension of the Na-O anticorrelation and the bluest and hottest point reached
along the HB, discovered by Carretta et al. (2007b). This is just the case, as
shown in Fig.~\ref{f:iqrteff60}, where we used for M~80 the value
$\log T_{\rm eff}^{max}=4.477$ from Recio-Blanco et al. (2006). M~80 nicely
falls on the relation defined by the strict correspondence between the
phenomenon of multiple stellar populations in GCs and their HB morphology.
\begin{figure}
\centering
\includegraphics[scale=0.40]{iqrteff60.eps}
\caption{The extension of the Na-O anti-correlation (measured using
IQR[O/Na]) as a function of the maximum temperature along the HB, taken from
Recio-Blanco et al. (2006). M~80 is represented by a filled star symbol. The
Spearman rank and the Pearson linear correlation coefficients are listed.}
\label{f:iqrteff60}
\end{figure}
\subsection{Other elements}
Elements produced with $\alpha-$capture in explosive nucleosynthesis of
core-collapse supernovae (Si, Ca, Ti) track each other in M~80. The observed
scatter in Mg is from 2 to 3 times larger than that of the other
$\alpha-$elements because Mg is also involved in the proton-capture reactions
which change the chemical composition of second generation stars with respect to
the level established by Type II SNe.
The pattern of Fe-group elements in M~80 is not unusual; these elements nicely
track Fe, apart from Mn and Cu, which are found to be underabundant in
excellent agreement with the measurements in field stars of the Galactic halo
(see Fig.~\ref{f:mncu60}).
We conclude that concerning the abundance pattern of species from proton-
and $\alpha-$capture, as well as those of the Fe-peak, M~80 is a typical halo
cluster.
\begin{figure}
\centering
\includegraphics[scale=0.45]{mncu60.eps}
\caption{[Mn/Fe] (upper panel) and [Cu/Fe] (lower panel) average ratios of M~80
from stars with UVES spectra compared to various samples of field stars as a
function of metallicity. The $rms$ scatter for Mn, Cu and Fe are plotted for
M~80 (they are inside the large symbol).}
\label{f:mncu60}
\end{figure}
Also the pattern of the abundances of neutron-capture elements looks
normal in M~80. In Fig.~\ref{f:baeuy60} we show the average abundance ratios of
a light neutron-capture element, Y, of the
heavier element Ba, created by the main $s-$process, and of the Eu forged almost
totally in the $r-$process (Arlandini et al. 1999). The mean ratios in M~80 are
found compatible to the run of these elements in field stars of the
Galactic halo as a function of metallicity (Venn et al. 2004). We do not confirm
the very high value [Eu/Fe]$=+0.80$ dex found by Cavallo et al. (2004), which
made M~80 outstanding in the plot showed by Venn et al.
\begin{figure}
\centering
\includegraphics[bb=19 146 418 709, clip, scale=0.52]{baeuy60.eps}
\caption{From top to bottom: [Ba/Fe], [Eu/Fe], [Eu/Ba], and [Eu/Y] abundance
ratios in M~80 (large filled circle) and in Galactic field stars (Venn et al.
2004, empty squares), as a function of metallicity.}
\label{f:baeuy60}
\end{figure}
To characterize the modality of enrichment in heavy elements from neutron
capture it is safer to use the La abundance rather than Ba: although derived for
the limited sample of stars with UVES spectra, La lines are weak and less
sensitive to uncertainty in the abundance analysis. The comparison between La (a
species with predominance of the $s-$process in the solar system) and Eu (almost
totally $r-$process element in solar system) allows to study the aforementioned
enrichment. We found in M~80 a mean value of
$\log \epsilon$(La/Eu)=0.44 dex ($\sigma=0.11$ dex, 14 stars) which would
indicate a predominance from $r-$process production in this cluster (pure $r$
level is 0.09 dex, Simmerer et al. 2004).
\begin{figure}
\centering
\includegraphics[bb=30 175 520 704, clip, scale=0.52]{ncapt60.eps}
\caption{Average abundances of neutron-capture elements Y, Zr, Ba, La, Ce, Pr,
Nd, Sm, and Eu in M~80. The associated $rms$ scatters are also shown. In the
upper panel the solar scaled abundance of pure $r-$process (dashed line) and of
pure $s-$process (dotted line) from Simmerer et al. (2004) are shown. In the
lower panel the attempt to best fit the observed abundances with solar scaled
pure $r-$process and $s-$process abundances (Simmerer et al. 2004) is shown
(light blue line): the scaling factor adopted are -1.27 dex and -1.81 dex for
the $r-$ and $s-$process, respectively. The black dotted line connects the
abundances in the $r+s$ group of the GC M~22 derived by Roederer et al. (2011).
}
\label{f:ncapt60}
\end{figure}
A more detailed comparison can be performed by comparing all the available
ratios of neutron-capture elements in M~80 with the pattern of solar scaled
contributions from $s-$ and $r-$process nucleosynthesis. In the upper panel of
Fig.~\ref{f:ncapt60} the comparison is made using the fractions estimated in
Simmerer et al. (2004). Several elements occupy a position intermediate between
the pure $s-$process level and the pure $r-$ fraction (obtained as usual from
$N_{SS,r} \equiv N_{SS,total} - N_{SS,s}$).
Therefore, we tried to obtain a best fit by trying to reproduce the composition
of Tab.~\ref{t:meanabu60} as the sum of two contributions (a pure solar-scaled
$r-$process and a pure solar scaled $s-$process: $s-$ and $r-$ fractions are
as in Simmerer et al. 2004) with suitable scaling factors. We found that the
observed pattern can be reproduced by
using scaling factors of -1.27 dex and -1.81 dex for the $r-$ and $s-$process,
respectively (see lower panel of Fig.~\ref{f:ncapt60}). Hence, if we consider
that [Fe/H]$=-1.79$ dex, this implies
[$r$/Fe]=+0.52 and [$s$/Fe]=-0.02 dex. The excess of elements produced by
the $r-$process is similar to that obtained for the $\alpha-$elements and
can be interpreted as an Fe deficiency due to the fact that there
was no significant contribution of SN Ia to the material from which M~80
formed. On the other hand, a significant contribution by the $s-$process
is required to explain observations. The abundance ratio first/second
peak of the $s-$process is more favorable to the second. This is not
unexpected, given the low metallicity of M80.
To get more insight with a direct, model independent approach we
compared the pattern found in M~80 with that of two other GCs, one showing a
dispersion in the $r-$process content and another one where a clear dispersion
in $s-$process elements has been detected, chosen from the extensive survey by
Roederer (2011).
In Fig.~\ref{f:eula60r} our ratios [Eu/Fe] and [La/Fe] for individual stars in
M~80 are compared to the sample of field stars with no detectable $s-$process
contribution analyzed by Roederer et al. (2010). Superimposed to these we also
plot stars in M~15, a GC well know to show a dispersion in the $r-$process
content (see Roederer 2011). For M~15 we used data from Otsuki et al. (2006)
and Sobeck et al. (2011).
\begin{figure}
\centering
\includegraphics[scale=0.40]{eula60r.eps}
\caption{[La/Fe] ratios as a function of [Eu/Fe] ratios for individual stars in
M~80 (large filled circles) compared to $r-$only field stars from Roederer et
al. (2010; small empty squares) and to a GC with a dispersion in $r-$process
from Roederer (2011): M~15 (filled triangles), with abundances from Otsuki et
al. (2006) and Sobeck et al. (2011). The internal error bars refer to our
analysis of M~80.}
\label{f:eula60r}
\end{figure}
The M~15 data show the same correlation between [La/Fe] and [Eu/Fe] present in
the sample of {\it bona fide} $r-$only metal-poor field stars from the
compilation of Roederer et al. (2010). A dispersion in [La/Fe] was found to be
intrinsic to this sample, indicating that these variations are intrinsically
associated to the $r-$process yields (Roederer et al. 2010). On the other hand,
it is immediately evident that no intrinsic dispersion in Eu is present among
giants in M~80, whereas we cannot exclude a possible small
spread of La abundances in these stars.
\begin{figure}
\centering
\includegraphics[scale=0.40]{eula6018.eps}
\caption{[La/Fe] ratios as a function of [Eu/Fe] ratios for individual stars in
M~80 (large filled circles) compared to $r-$only field stars from Roederer et
al. (2010; small empty squares) and to GCs with a dispersion in $s-$process
elements. Blue filled squares indicate the $r-$only component in M~22 from
Marino et al. (2011), orange asterisks are used for the $r+s$ component in M~22
(see Roederer et al. 2011, Marino et al. 2011), and green filled pentagons are
giants in NGC~1851 from Carretta et al. (2011b). The internal error bars refer
to our present analysis of M~80.}
\label{f:eula6018}
\end{figure}
In Fig.~\ref{f:eula6018} we compare abundances of La
and Eu in M~80 with those of GCs known to host stellar populations with an
intrinsic dispersion in $s-$process elements: M~22 (Marino et al. 2011, Roederer
et al. 2011) and NGC~1851 (Carretta et al. 2011b). The $r-$only component in
M~22 (in the terminology by Roederer et al. 2011) nicely falls on the locus of
galactic field metal-poor stars. This behaviour is shared also by the majority
of stars analyzed in NGC~1851 and by a small fraction of the present sample in
M~80.
However, most giants in M~80 have distinctly higher La abundances,
and populate a region in the [La/Fe] vs [Eu/Fe] plane where are located stars of
the so-called $r+s$ component in M~22 (Roderer et al. 2011; they correspond to
the $s-$rich group in Marino et al. 2011). This trend extends to higher
metallicities thanks to a minority component in NGC~1851.
The spread observed in La for RGB stars in M~80 seems comparable to the range
observed in the $r+s$ group in M~22. On the other hand, although similarities
between M~22 and M~80 include the total mass and metal abundance
([Fe/H]$\sim -1.79$ dex in both), M~80 does not show a spread in [Fe/H] as
instead observed in M~22 (Da Costa et al. 2009, Marino et al. 2011).
Anyway, if we plot the contribution from the $r+s$ group in M~22 from the
analysis by Roederer et al. (2011) over the distribution of neutron-capture
elements observed in M~80 (dotted line in the lower panel of
Fig.~\ref{f:ncapt60}, the agreement is not dramatically unsatisfactory.
We conclude that M~80 could be a less extreme case of a GC with a
significant contribution from the $s-$process, and may be more similar to
NGC~1851, although at lower metallicity.
Although most of neutron-capture elements are only available for the limited
sample of stars with UVES spectra, we found no statistically significant
difference in the level of abundances of heavy elements between different
stellar generations in M~80.
\section{Summary}
NGC~6093=M~80 is the last of the 24 globular clusters of our core program on the
Na-O anticorrelation in RGB stars and its relation with HB morphology and
cluster parameters (see e.g., Carretta et al. 2006, 2009a,b). In this paper we
present the analysis of this moderately massive, metal-poor, very dense GC, seen
in the direction of the Galactic centre. We used FLAMES spectra of 82 red
giant stars, confirmed members on the basis of their RV and metallicity and
derived atmospheric parameters, metallicity, and chemical
abundances (12 species for GIRAFFE spectra, more than 20 for UVES spectra).
The main goal is the determination of O and Na abundances, to study the shape
and extension of the Na-O anticorrelation, the main chemical signature of
multiple populations in GCs (e.g., Carretta et al. 2010a). The ratios between
first and second generation stars is normal, with fractions of P, I, and E
stars at 36, 56, and 8\%, respectively. The three groups
of stars appear to isolate naturally (see Fig. 8). The spread of both O and Na
is large, about 1 dex, and the anticorrelation has a moderate extension (as
measured by a IQR[O/Na]=0.784), which places NGC~6093 in the middle of the
correlation with cluster mass (via $M_V$, see Fig. 11) and maximum temperature
along the HB (see Fig. 16). The cluster also conforms to the relation between
concentration and the residual around the best fit IQR[O/Na] and $M_V$.
Using the UVES spectra of 14 member stars we can also study
other p-capture elements. We find that Al is enhanced in I, E stars and is
anticorrelated with Mg, as usual. Briefly, this inner-halo cluster shows the
classical chemical signatures of multiple stellar populations in the normal
ways.
All $\alpha$, proton-capture, and iron peak elements behave in NGC~6093 like in
a typical halo cluster. The neutron-capture elements (with the exception of Ba)
can be measured only in the UVES spectra. They conform to the run of field stars
as function of metallicity (see Fig. 18). Comparing the values of [La/Fe] and
[Eu/Fe] with those of two clusters showing some spread in $r-$process or $s-$process
elements (M 15 and M 22, respectively) we see that NGC~6093 may be a mild case
of intrinsic dispersion is $s-$process elements (see Fig. 21), without showing,
however, any indication of metallicity spread.
\begin{acknowledgements}
This publication makes use of data products from the Two Micron All Sky Survey,
which is a joint project of the University of Massachusetts and the Infrared
Processing and Analysis Center/California Institute of Technology, funded by the
National Aeronautics and Space Administration and the National Science
Foundation. This research has been funded by PRIN INAF 2011
"Multiple populations in globular clusters: their role in the Galaxy assembly"
(PI E. Carretta), and PRIN MIUR 2010-2011, project ``The Chemical and Dynamical
Evolution of the Milky Way and Local Group Galaxies'' (PI F. Matteucci) . This
research has made use of the SIMBAD database, operated at CDS, Strasbourg,
France and of NASA's Astrophysical Data System.
\end{acknowledgements}
|
\section{Introduction}
Bright squeezed vacuum (BSV) is a state of light emerging from the output of a high-gain unseeded parametric amplifier (OPA). Due to its nonclassical properties such as photon-number entanglement and quadrature squeezing, this state is useful for various quantum-information applications, among them quantum metrology~\cite{metrology}, quantum imaging~\cite{q-imaging}, and quantum lithography~\cite{lithography}. Besides containing a high number of photons in each mode, the state is essentially multimode, both in the frequency and in the angle. These features provide its large information capacity, as quantum information can be encoded in the number of photons in different modes. At the same time, the presence of a large number of modes can be a disadvantage in certain experiments, for instance achieving phase super-sensitivity~\cite{super} or (related) gravitational-wave detection~\cite{gravit}. A possible way to reduce the number of modes without losing nonclassical correlations is to use a nonlinear interferometer, in which only part of the spectrum is amplified. This has been already demonstrated for the angular spectrum in Ref.~\cite{separation}. The goal of this work is to show similar behavior in the frequency domain.
The paper is organized as follows. In the next two subsections, we briefly describe the mode structure of BSV (subsection~\ref{S1.1}) and the idea and operation of a nonlinear interferometer (subsection~\ref{S1.2}). Section~\ref{S2} explains the idea of reducing the number of modes in BSV in space/angle and time/frequency and demonstrates the narrowing of the BSV angular spectrum in a nonlinear interferometer with spatially separated crystals. The experiment on the narrowing of the BSV frequency spectrum is described in Section~\ref{S3}. Section~\ref{S4} contains the conclusions.
\subsection{BSV and its eigenmodes}\label{S1.1}
The most convenient way of generating BSV is high-gain parametric down-conversion in a nonlinear crystal, which can be considered as an unseeded traveling-wave OPA. The frequency-angular spectrum and photon-number correlations are well described by the Bloch-Messiah formalism~\cite{Wasilewski,Silberhorn,Bloch}, in which the Hamiltonian is diagonalized by passing to the eigenmodes of the OPA. For instance, in the case of spatially multimode PDC, the Hamiltonian can be written as~\cite{Bloch}
\begin{equation}
H=i\hbar\Gamma\iint d \mathbf{q}_s d \mathbf{q}_i F(\mathbf{q}_s,\mathbf{q}_i)a^\dagger_{\mathbf{q}_s}a^\dagger_{\mathbf{q}_i}+h.c.,
\label{eq:Ham}
\end{equation}
where $\Gamma$ characterizes the coupling strength, $\mathbf{q}_{s,i}$ are the transverse wavevectors of the signal and idler radiation, and $a^\dagger_{\mathbf{q}_{s,i}}$ are the photon creation operators in the corresponding plane-wave modes. The central part of the Hamiltonian is the two-photon amplitude (TPA), $F(\mathbf{q}_s,\mathbf{q}_i)$, whose meaning is the probability amplitude of a photon pair created with the wavevectors $\mathbf{q}_s,\mathbf{q}_i$. The Hamiltonian (\ref{eq:Ham}) is diagonalized by representing the TPA as a Schmidt decomposition,
\begin{equation}
F(\mathbf{q}_s,\mathbf{q}_i)=\sum_k \sqrt{\lambda_k}u_k(\mathbf{q}_s)v_k(\mathbf{q}_i),
\label{eq:TPA Schmidt}
\end{equation}
where $\lambda_k$ are the Schmidt eigenvalues, $u_k(\mathbf{q}_s),\,v_k(\mathbf{q}_i)$ the Schmidt modes, and $k$ is a two-dimensional index. By definition, the modes are ordered so that $\lambda_{k+1}\le\lambda_k$. The Hamiltonian (\ref{eq:Ham}) can be now written as a sum of two-mode Hamiltonians,
\begin{equation}
H=\sum_k \sqrt{\lambda_k}H_k,\,\,H_k=i\hbar\Gamma A^\dagger_kB^\dagger_k+h.c.,
\label{eq:BM_dec}
\end{equation}
with the photon creation operators $A^\dagger_k,\,B^\dagger_k$ relating to the Schmidt modes. Moreover, if the signal and idler beams are indistinguishable, their Schmidt modes are the same, and
\begin{equation}
H_k=i\hbar\Gamma (A^\dagger_k)^2+h.c.
\label{eq:Ham_deg}
\end{equation}
It is worth mentioning that the Schmidt decomposition can be alternatively performed in the space coordinates, which is equivalent to the wavevector decomposition~(\ref{eq:TPA Schmidt}). A similar decomposition is valid for the frequency/temporal domain. In different works, the Schmidt modes are also called nonmonochromatic modes~\cite{Opatrny}, squeezing (eigen)modes~\cite{Boyd}, broadband modes~\cite{Silberhorn}, or supermodes~\cite{Fabre}.
Clearly, in terms of these new modes, photon-number correlations are only pairwise. The total mean photon number can be represented as a sum of incoherent contributions from all Schmidt modes,
\begin{equation}
\langle N\rangle=\sum_k \langle N_k\rangle,\,\,\langle N_k\rangle=\sinh^2[\sqrt{\lambda_k}G],
\label{eq:ph_N0}
\end{equation}
where $G=\int\Gamma dt$ is the parametric gain. This means that while at low gain ($G<<1$), the Schmidt modes are populated with the weights given by the Schmidt eigenvalues $\lambda_k$, at high gain these weights are changed to become~\cite{Bloch}
\begin{equation}
\tilde{\lambda}_k=\frac{\sinh^2[\sqrt{\lambda_k}G]}{\sum_k\sinh^2[\sqrt{\lambda_k}G]}.
\label{eq:renorm}
\end{equation}
According to this, at high gain the lower-order Schmidt modes, initially having higher eigenvalues, become more pronounced.
\subsection{SU(1,1) interferometers}\label{S1.2}
At the very start of nonlinear optics, an idea emerged to realize two nonlinear effects at spatially separated points and to see interference between them. Such \textit{nonlinear interference} would enable the observation of the relative phase between two effects. It was first realized by Chang et al.~\cite{Bloembergen} who measured in this way the complex values of surface quadratic susceptibility for several semiconductors. Later, it became a common way to measure the phases of nonlinear susceptibilities.
After the discovery of parametric amplification via PDC and four-wave mixing (FWM), it was soon suggested to realize nonlinear interference based on these effects. Yurke et al.~\cite{SU11} proposed an interferometer in which the signal and idler beams emitted in the first parametric amplifier were directed into the second one and got amplified or deamplified depending on the phases introduced in the pump, signal, or idler beams, $\phi_{p,s,i}$ (Fig.~\ref{nonlinear_int}). Because the transformations performed by the interferometer on the fields at its two output modes relate to the SU(1,1) group, this type of interferometer is usually referred to as SU(1,1).
\begin{figure}[h]
\begin{center}
\includegraphics[width=0.5\textwidth]{nonlinear_int.eps}
\caption{(Color online) An SU(1,1) nonlinear interferometer based on two high-gain parametric amplifiers.}\label{nonlinear_int}
\end{center}
\end{figure}
Initially the SU(1,1) interferometer was proposed as a method to perform phase measurement below the shot-noise level, which is possible due to the strong dependence of its output intensity on the phases $\phi_{s,i}$ at high parametric gain~\cite{Ou}. Nevertheless, the first implementations of SU(1,1) interferometers were based on low-gain (spontaneous) PDC~\cite{Herzog,big}. In particular, in a clever modification of such interferometer the effect of `induced coherence without induced emission' was observed~\cite{induced}, which later was successfully implemented for the measurement of absorption~\cite{abs} and dispersion~\cite{disp}, as well as imaging~\cite{imaging} in the infrared or terahertz~\cite{THz} spectral ranges without the detection of infrared or terahertz radiation.
Only recently, the SU(1,1) interferometer using FWM in rubidium vapor has been implemented for overcoming the shot-noise level of phase measurement~\cite{Zhang}. More than $4$ dB improvement has been obtained compared to a conventional (SU(2)) interferometer populated with the same mean number of photons. The operation was at high gain, which provided $7.4$ dB amplification of the radiation from the first FWM source in the second one.
The same mechanism can be used for shaping the spectrum of high-gain PDC or FWM, both in the angle and in the frequency. Because of the nonlinear amplification of the incident radiation, the modes that are not amplified in the second nonlinear crystal will be much weaker at the output than those amplified. Moreover, one can take advantage of the de-amplification of certain modes, which, however, is not accompanied by a noise increase. Such selective amplification of different modes can enable tailoring the structure of the spectrum.
\section{Diffractive and dispersive spreading, and reduction of the mode number}\label{S2}
\subsection{Angular spectrum tailoring}\label{S2.1}
In a nonlinear interferometer formed by two spatially separated traveling-wave high-gain parametric amplifiers (Fig.~\ref{separated})~\cite{separation}, broadband radiation emitted by the first crystal is amplified in the second one. If the distance between the crystals is considerable, only part of the radiation is amplified in the second one - namely, the part that passes through the pump beam in the second crystal. In accordance with this, it was shown~\cite{separation} that at a certain distance between the crystals, the angular spectrum becomes nearly single-mode.
At a sufficiently large distance between the crystals, the angular width of the spectrum amplified in the second crystal should be roughly given by the ratio between the pump diameter $a$ and the distance $L$ between the crystals,
\begin{equation}
\Delta\theta\approx\frac{a}{L}.
\label{eq:angular width}
\end{equation}
\begin{figure}[h]
\begin{center}
\includegraphics[width=0.5\textwidth]{separated2.eps}
\caption{(Color online) A nonlinear interferometer formed by two spatially separated crystals.}\label{separated}
\end{center}
\end{figure}
This effect of spatial spectrum narrowing has a simple explanation in terms of the Schmidt modes of high-gain PDC. Indeed, each of the spatial Schmidt modes of BSV emitted by the first crystal is diffracting (spreading) in the course of propagation to the second crystal. To a good approximation, the spatial Schmidt modes are given by the Hermite-Gauss or Laguerre-Gauss set, the lowest-order mode being simply a Gaussian beam. Lower-order modes spread slowly and in the second crystal they overlap with the pump beam. Therefore, they get amplified provided that the phase acquired in the course of propagation is appropriate. However, higher-order modes (no matter if they are given by Laguerre-Gauss or Hermite-Gauss beams) spread faster in the space between the crystals and do not get amplified. As a result, the spatial spectrum at the output of the second crystal gets narrower. This continues until only the first Schmidt mode gets amplified, after which the angular width remains constant; increasing the distance $L$ only reduces the total intensity.
The dependence of the angular width $\Delta\theta$ on the distance $L$ between the crystals can be derived from this picture as follows. The zeroth-order Schmidt mode is a Gaussian beam of waist radius $w_{0}$. As it propagates from the crystal, the waist radius at a distance $z$ is~\cite{Kogelnik}
\begin{equation}
w_0(z)=\sqrt{w_{0}^{2}+\left(\frac{\lambda z}{\pi w_{0}}\right)^2},\,\, w_0(0)=w_0,
\label{eq:waist_1st}
\end{equation}
with $\lambda$ being the wavelength. The parameter $\theta_0\equiv\lambda/(\pi w_0)$ is the half-angle divergence of the Gaussian beam. Higher-order modes have larger spatial sizes, $w_{m}=Mw_0$; for instance, for Hermite-Gauss beams, $M=\sqrt{2m+1}$. At the same time, they have larger divergences $\theta_m=M\theta_0$, so that as the distance $z$ increases, they spread as
\begin{equation}
w_m(z)=\sqrt{w_{m}^{2}+\left(M^2\frac{\lambda z}{\pi w_{m}}\right)^2}.
\label{eq:waist_m}
\end{equation}
Assuming that for $z=L$, only modes of orders from $0$ to $m$ are amplified in the second crystal, we find the corresponding $M$ from the condition $w_m(L)=a/2$. We obtain
\begin{equation}
M=\frac{a}{2}\left[w_0^{2}+\left(\frac{\lambda L}{\pi w_0}\right)^2\right]^{-1/2}.
\label{eq:M}
\end{equation}
Then, the divergence of the beam will be equal to twice the half-angle divergence of mode $m$, $\Delta\theta=2\theta_m=M\frac{\lambda}{\pi w_{0}}$:
\begin{equation}
\Delta\theta=\left[\frac{1}{\Delta\theta_0^{2}}+\left(\frac{L}{a}\right)^2\right]^{-1/2},
\label{eq:angular width_full}
\end{equation}
where $\Delta\theta_0=\frac{a\lambda}{\pi w_{0}^2}$ is the initial angular width.
We have measured the dependence of the angular width on the distance between the crystals under the same experimental conditions as in Ref.~\cite{separation}: two $3$ mm crystals were placed into a single pump beam of full width at half maximum (FWHM) waist $200$ $\mu$m, and the distance between them was changed from $10$ to $130$ mm. The results are shown in Fig.~\ref{angular}.
\begin{figure}[h]
\begin{center}
\includegraphics[width=0.4\textwidth]{angular.eps}
\caption{(Color online) The angular width of the spectrum measured at the output of the second crystal, versus the distance between the crystals. The solid line is calculated with Eq.~(\ref{eq:angular width_full}), without fitting parameters. The dashed line is a guide to an eye. The arrow shows the position at which $m=1.1$ spatial modes were reported in Ref.~\cite{separation}.}\label{angular}
\end{center}
\end{figure}
Equation (\ref{eq:angular width_full}) was used for fitting the dependence at $L\le60$ mm. At larger distances, the angular spectrum shows no dependence on $L$; the dashed line is just a guide to the eye. The position at which nearly single-mode was observed (the number of modes was measured to be $m=1.1$)~\cite{separation} is shown by an arrow.
\subsection{Frequency spectrum tailoring}\label{S2.2}
This effect has an analogue in the frequency/time domain. In this case, the role of the diffractive spreading of beams is played by the dispersive spreading of pulses. Indeed, let a dispersive material of length $d$ be placed between the two crystals. Each frequency Schmidt mode of the BSV from the first crystal will spread in time in the course of propagation through the material, and the spreading will be determined by the group-velocity dispersion (GVD) $k''=d^2k/d\omega^2$. Lower-order modes are narrower in time than higher-order ones, but they will still spread in the GVD material slower than higher-order ones. This is illustrated by Fig.~\ref{modes} where several temporal Schmidt modes are plotted for BSV emitted from a 3 mm crystal pumped by $6$ ps pulses at wavelength $355$ nm (a). The emission is at the degenerate wavelength $710$ nm. The modes are assumed to be the same as for spontaneous parametric down-conversion~\cite{Law}. To a good approximation, they are given by Hermite functions~\cite{Wasilewski}.
After propagation through a dispersive material with the GVD $k''$ and length $d$, the temporal modes will spread in time but maintain their shapes. The latter follows from the fact that, similar to diffractive spreading of beams, dispersive spreading of pulses acts as the Fourier transformation, so that after a sufficiently long GVD material the shape of a pulse becomes similar to its spectral amplitude. At the same time, Hermite functions are eigenfunctions of the Fourier transformation. Therefore, the whole set of Schmidt modes will be simply rescaled after the propagation through the dispersive material. For instance, the zeroth-order Schmidt mode (dotted line in Fig.~\ref{modes}a), initially a Gaussian pulse of duration $\tau_{0}$ will remain a Gaussian pulse with the duration depending on $d$~\cite{Yariv},
\begin{equation}
\tau_0(d)=\sqrt{\tau_{0}^{2}+\left(\frac{k'' d}{\tau_{0}}\right)^2}.
\label{eq:time_1st}
\end{equation}
Higher-order modes will also maintain their shapes but, similar to the case of the angular modes, will spread faster.
In the right panel of Fig.~\ref{modes}, the Schmidt modes are plotted after the propagation through various lengths $d$ of SF6 glass, whose GVD at the wavelength $710$ nm is $k''=238$ fs$^2$/mm~\cite{refractive index}. In the calculation, Eq.~(\ref{eq:time_1st}) was used for the Gaussian mode, and the higher-order modes were simply rescaled accordingly.) For the length $d=10$ cm (b), only higher-order modes (mode $50$ in the figure) get sufficiently spread so that they do not fully overlap with the pump pulse in the second crystal. Therefore, high-order modes will not be amplified.
However, at $d=20$ cm (c), even the tenth-order mode becomes considerably spread and will not get fully amplified in the second crystal. This should lead to the narrowing of the spectrum. As the length of the dispersive glass increases, the spectral width should reduce. After $d=60$ cm of glass (d), the zeroth-order Schmidt mode will overlap with the pump pulse. At high gain, this situation should lead to single-mode output emission.
\begin{figure}[h]
\begin{center}
\includegraphics[width=0.7\textwidth]{modes.eps}
\caption{(Color online) Temporal Schmidt modes of orders $0$ (red dotted line), $10$ (blue dashed line) and $50$ (green solid line) for a $3$ mm crystal pumped by $6$ ps pulses before (a) and after propagating through $10$ cm (b), $20$ cm (c) and $60$ cm (d) of SF6 glass. For comparison, the amplitude of the pump pulse is shown by magenta dash-dotted line.}\label{modes}
\end{center}
\end{figure}
By analogy with Eq.~\ref{eq:angular width_full}, one can estimate the frequency spectrum of BSV generated in the system of two crystals separated by a GVD material as
\begin{equation}
\Delta\omega=\left[\frac{1}{\Delta\omega_0^{2}}+\left(\frac{k''d}{T_p}\right)^2\right]^{-1/2},
\label{eq:frequency width_full}
\end{equation}
where $T_p$ is the pump pulse duration and $\Delta\omega_0$ the initial spectral width.
In the next section, we describe the experimental results confirming this behavior.
\section{Experiment}\label{S3}
The scheme of the experimental setup is shown in Fig.~\ref{setup}. Collinear high-gain PDC with the central wavelength at 709.3 nm was created in a type-I 3-mm-long BBO crystal by pumping it with a third harmonic of the pulsed Nd:YAG laser at 354.7 nm, with a pulse duration of 18 ps (the coherence time being $6$ ps) and a repetition rate of 1 kHz. The laser power was varied by a half wave plate $\lambda_p/2$ followed by a polarization beamsplitter $PBS_p$. A telescope, made of plano convex lenses $L_{p1}$ and $L_{p2}$ with the focal distances of 50 cm and 5 cm, respectively, compressed the beam size down to half-power beam width of 225 $\mu m$. A dichroic mirror $DM_1$ separated the pump beam and the PDC. The PDC pulses were passing through the group velocity dispersion (GVD) medium. We had three options of GVD media: SF-6 glass rods of length $9$ cm and $18.3$ cm, and SF-57 glass rod of length $19.4$ cm. The pump pulses were timed, by means of a delay line, to overlap with the time-stretched PDC pulses on the dichroic mirror $DM_2$ and amplify them in the second type-I 3-mm-long BBO crystal. After the crystal the pump was blocked by a pair of dichroic mirrors $DM_3$ and a long-pass filter OG580. The iris $A_1$ placed in the focal plane of the lens L with the focal distance of 10 cm was used to align the crystals for collinear geometry. The lens $L_i$ focused the PDC radiation onto the input slit of the spectrometer with a resolution of 0.15 nm.
\begin{figure}[h]
\begin{center}
\includegraphics[width=0.8\textwidth]{setup.eps}
\caption{(Color online) Schematic of the experimental setup.}\label{setup}
\end{center}
\end{figure}
Figure~\ref{results} (a) shows the measured PDC spectra with and without the GVD medium placed between the crystals. The PDC spectrum for the crystals separated by an air gap of $24.2$ cm is shown by a blue line. The measurement was performed at an average pump power of $73.4$ mW in the setup, published before in~\cite{separation}. The interference fringes due to different refractive indexes of the pump and the signal and idler beams in the air were avoided by averaging the spectra taken over different positions of the first crystal in the range from 22.7 cm to 25.7 cm with the step of 2 mm. The FWHM of the spectrum was found to be $45.6$ nm.
\begin{figure}[h]
\begin{center}
\includegraphics[width=0.7\textwidth]{results5.eps}
\caption{(Color online) (a) Measured PDC spectra with different GVD media inserted. (b) FWHM of the PDC spectra experimentally measured with and without GVD media placed between the crystals (blue triangles) and the theoretically calculated dependence according to Eq.~\ref{eq:frequency width_full} (red line) plotted versus
$k''d$. }
\label{results}
\end{center}
\end{figure}
Green and red lines in Fig. ~\ref{results} (a) show how the spectrum of the PDC evolves as the value of $k''d$ for the inserted GVD medium is increased. All the spectra are affected by the interference arising from the frequency-dependent phase of the broadband PDC generated in the first crystal after passing through the GVD media. As a result, constructive or destructive interference is observed for the different frequencies at the output of the second crystal. Despite the interference, one can clearly see that the spectra measured with the GVD media are narrower than the one obtained with the air gap between the crystals. The FWHM of the envelope for each measurement was considered as the width of the spectrum.
In Fig.~\ref{results} (b) we compare the measurement results for the PDC spectral width (blue triangles) and the calculation according to Eq.~\ref{eq:frequency width_full} (red line) using the GVD values for the Schott glass SF6 and SF57~\cite{refractive index}. Instead of the pulse duration $T_p$, the coherence time of the laser $T_c=6 $ ps was used since it is the coherence of the pump that matters for parametric amplification. One can see that the experimental FWHM values agree well with the calculations.
\section{Conclusion}\label{S4}
In conclusion, we have considered a nonlinear interferometer formed by two unseeded traveling-wave parametric amplifiers (based on parametric down-conversion in nonlinear crystals) and showed that its angular and frequency spectrum of emission can be modified by spatially separating the two crystals and/or placing a dispersive material between them. The effect has a simple interpretation in terms of Schmidt modes: higher-order modes spread in space and time faster than low-order ones and hence do not get amplified in the second crystal. Our experimental results show the narrowing of the spatial spectrum, leading ultimately to a single spatial mode. For frequency spectrum narrowing, preliminary results show 30\% narrowing for a glass rod with large group velocity dispersion inserted into the interferometer.
The research leading to these results has received funding from the EU FP7 under grant agreement No. 308803 (project BRISQ2). We thank Xin Jiang and Patricia Schrehardt for providing the samples of SF6 and SF57.
|
\section{Introduction}
\label{sec:introduction}
Excited charmed mesons are of great theoretical and experimental interest as they allow detailed studies of QCD in an interesting energy regime.
Good progress has been achieved in identifying and measuring the parameters of
the orbitally excited states, notably from Dalitz plot (DP) analyses of three-body $B$ decays.
Relevant examples include the studies of ${\ensuremath{\Bub}}\xspace \ensuremath{\rightarrow}\xspace {\ensuremath{\D^+}}\xspace{\ensuremath{\pion^-}}\xspace\pim$~\cite{Abe:2003zm,Aubert:2009wg} and ${\ensuremath{\Bbar{}^0}}\xspace\ensuremath{\rightarrow}\xspace{\ensuremath{\D^0}}\xspace{\ensuremath{\pion^+}}\xspace{\ensuremath{\pion^-}}\xspace$~\cite{Kuzmin:2006mw} decays, which provide information on excited neutral and charged charmed mesons (collectively referred to as $D^{**}$ states), respectively.
First results on excited charm-strange mesons have also recently been obtained
with the DP analysis technique~\cite{LHCb-PAPER-2014-035,LHCb-PAPER-2014-036,Lees:2014abp}.
Studies of prompt charm resonance production in {\ensuremath{\Pe^+\Pe^-}}\xspace and $pp$ collisions~\cite{delAmoSanchez:2010vq,LHCb-PAPER-2013-026} have revealed a number of additional high mass states.
Most of these higher mass states are not yet confirmed by independent analyses, and their spectroscopic identification is unclear.
Analyses of resonances produced directly from {\ensuremath{\Pe^+\Pe^-}}\xspace and $pp$ collisions do not allow determination of the quantum numbers of the produced states, but can distinguish whether or not they have natural spin parity (\mbox{\itshape i.e.}\xspace\ $J^P$ in the series $0^+, 1^-, 2^+, ...$).
The current experimental knowledge of the neutral $D^{**}$ states is summarised in Table~\ref{tab:PDG} (here and throughout the paper, natural units with $\hbar = c = 1$ are used).
The $D^*_0(2400)^0$, $D_1(2420)^0$, $D_1^\prime(2430)^0$ and $D^*_2(2460)^0$ mesons are generally understood to be the four orbitally excited (1P) states.
The experimental situation as well as the spectroscopic identification of the heavier states is less clear.
\begin{table}[!b]
\centering
\caption{\small
Measured properties of neutral $D^{**}$ states.
Where more than one uncertainty is given, the first is statistical and the others systematic.
}
\begin{tabular}{ccccc}
\hline
Resonance & Mass & Width & $J^P$ & Ref. \\
& $(\ensuremath{\mathrm{Me\kern -0.1em V}}\xspace)$ & $(\ensuremath{\mathrm{Me\kern -0.1em V}}\xspace)$ & & \\
\hline
$D^*_0(2400)^0$ & $2318 \pm 29$ & $267 \pm 40$ & $0^+$ & \cite{PDG2014} \\
$D_1(2420)^0$ & $2421.4 \pm 0.6$ & $27.4 \pm 2.5$ & $1^+$ & \cite{PDG2014} \\
$D_1^\prime(2430)^0$ & $2427 \pm 26 \pm 20 \pm 15$ & $384\,^{+107}_{-75} \pm 24 \pm 70$ & $1^+$ & \cite{Abe:2003zm} \\
$D^*_2(2460)^0$ & $2462.6 \pm 0.6$ & $49.0 \pm 1.3$ & $2^+$ & \cite{PDG2014} \\
\hline
$D^*(2600)$ & $2608.7 \pm 2.4 \pm 2.5$ & $93 \pm 6 \pm 13$ & natural & \cite{delAmoSanchez:2010vq} \\
$D^*(2650)$ & $2649.2 \pm 3.5 \pm 3.5$ & $140 \pm 17 \pm 19$ & natural & \cite{LHCb-PAPER-2013-026} \\
$D^*(2760)$ & $2763.3 \pm 2.3 \pm 2.3$ & $60.9 \pm 5.1 \pm 3.6$ & natural & \cite{delAmoSanchez:2010vq} \\
$D^*(2760)$ & $2760.1 \pm 1.1 \pm 3.7$ & $74.4 \pm 3.4 \pm 19.1$ & natural & \cite{LHCb-PAPER-2013-026} \\
\hline
\end{tabular}
\label{tab:PDG}
\end{table}
The ${\ensuremath{\Bub}}\xspace\ensuremath{\rightarrow}\xspace{\ensuremath{\D^+}}\xspace{\ensuremath{\kaon^-}}\xspace{\ensuremath{\pion^-}}\xspace$ decay can be used to study neutral $D^{**}$ states.
The ${\ensuremath{\D^+}}\xspace{\ensuremath{\kaon^-}}\xspace{\ensuremath{\pion^-}}\xspace$ final state is expected to exhibit resonant structure only in the ${\ensuremath{\D^+}}\xspace{\ensuremath{\pion^-}}\xspace$ channel, and unlike the Cabibbo-favoured ${\ensuremath{\D^+}}\xspace{\ensuremath{\pion^-}}\xspace\pim$ final state does not contain any pair of identical particles.
This simplifies the analysis of the contributing excited charm states, since partial wave analysis can be used to help determine the resonances that contribute.
One further motivation to study ${\ensuremath{\Bub}}\xspace \ensuremath{\rightarrow}\xspace {\ensuremath{\D^+}}\xspace{\ensuremath{\kaon^-}}\xspace{\ensuremath{\pion^-}}\xspace$ decays is related to the measurement of the angle $\gamma$ of the unitarity triangle defined as $\gamma \equiv \arg\left[ - V^{}_{ud}V_{ub}^*/(V^{}_{cd}V_{cb}^*) \right]$, where $V^{}_{xy}$ are elements of the Cabibbo-Kobayashi-Maskawa (CKM) quark mixing matrix~\cite{PhysRevLett.10.531,PTP.49.652}.
One of the most powerful methods to determine $\gamma$ uses ${\ensuremath{\Bub}}\xspace \ensuremath{\rightarrow}\xspace D {\ensuremath{\kaon^-}}\xspace$ decays, with the neutral $D$ meson decaying to {\ensuremath{C\!P}}\xspace eigenstates~\cite{Gronau:1990ra,Gronau:1991dp}.
The sensitivity to $\gamma$ arises due to the interference of amplitudes proportional to the CKM matrix elements $V^{}_{ub}$ and $V^{}_{cb}$, associated with ${\ensuremath{\Dbar{}^0}}\xspace$ and ${\ensuremath{\D^0}}\xspace$ production respectively.
However, a challenge for such methods is to determine the ratio of magnitudes of the two amplitudes, $r_B$, that must be known to extract $\gamma$.
This is usually handled by including {\ensuremath{\PD}}\xspace meson decays to additional final states in the analysis.
By contrast, in ${\ensuremath{\Bub}}\xspace \ensuremath{\rightarrow}\xspace D^{**}{\ensuremath{\kaon^-}}\xspace$ decays the efficiency-corrected ratio of yields of ${\ensuremath{\Bub}}\xspace \ensuremath{\rightarrow}\xspace D^{**} {\ensuremath{\kaon^-}}\xspace \ensuremath{\rightarrow}\xspace {\ensuremath{\D^-}}\xspace {\ensuremath{\pion^+}}\xspace {\ensuremath{\kaon^-}}\xspace$ and ${\ensuremath{\Bub}}\xspace \ensuremath{\rightarrow}\xspace D^{**} {\ensuremath{\kaon^-}}\xspace \ensuremath{\rightarrow}\xspace {\ensuremath{\D^+}}\xspace {\ensuremath{\pion^-}}\xspace {\ensuremath{\kaon^-}}\xspace$ decays gives $r_B^2$ directly~\cite{Sinha:2004ct}.
The decay ${\ensuremath{\Bub}}\xspace \ensuremath{\rightarrow}\xspace D^{**} {\ensuremath{\kaon^-}}\xspace \ensuremath{\rightarrow}\xspace D \pi^0 {\ensuremath{\kaon^-}}\xspace$ where the {\ensuremath{\PD}}\xspace meson is reconstructed in {\ensuremath{C\!P}}\xspace eigenstates can be used to search for {\ensuremath{C\!P}}\xspace violation driven by $\gamma$.
Measurement of the first two of these processes would therefore provide knowledge of $r_B$ in ${\ensuremath{\Bub}}\xspace \ensuremath{\rightarrow}\xspace D^{**}{\ensuremath{\kaon^-}}\xspace$ decays, indicating whether or not a competitive measurement of $\gamma$ can be made with this approach.
In this paper, the ${\ensuremath{\Bub}}\xspace\ensuremath{\rightarrow}\xspace{\ensuremath{\D^+}}\xspace{\ensuremath{\kaon^-}}\xspace{\ensuremath{\pion^-}}\xspace$ decay is studied for the first time, with the ${\ensuremath{\D^+}}\xspace$ meson reconstructed through the ${\ensuremath{\kaon^-}}\xspace{\ensuremath{\pion^+}}\xspace\pip$ decay mode.
The inclusion of charge conjugate processes is implied.
The topologically similar ${\ensuremath{\Bub}}\xspace \ensuremath{\rightarrow}\xspace {\ensuremath{\D^+}}\xspace{\ensuremath{\pion^-}}\xspace\pim$ decay is used as a control channel and for normalisation of the branching fraction measurement.
A large ${\ensuremath{\Bub}}\xspace\ensuremath{\rightarrow}\xspace{\ensuremath{\D^+}}\xspace{\ensuremath{\kaon^-}}\xspace{\ensuremath{\pion^-}}\xspace$ signal yield is found, corresponding to a clear first observation of the decay, and allowing investigation of the DP structure of the decay.
The amplitude analysis allows studies of known resonances, searches for higher mass states and measurement of the properties, including the quantum numbers, of any resonances that are observed.
The analysis is based on a data sample corresponding to an integrated luminosity of $3.0 \,{\rm fb}^{-1}$ of $pp$ collision data collected with the LHCb detector, approximately one third of which was collected during 2011 when the collision centre-of-mass energy was $\sqrt{s} = 7 \ifthenelse{\boolean{inbibliography}}{\ensuremath{~T\kern -0.05em eV}\xspace}{\ensuremath{\mathrm{\,Te\kern -0.1em V}}}\xspace$ and the rest during 2012 with $\sqrt{s} = 8 \ifthenelse{\boolean{inbibliography}}{\ensuremath{~T\kern -0.05em eV}\xspace}{\ensuremath{\mathrm{\,Te\kern -0.1em V}}}\xspace$.
The paper is organised as follows.
A brief description of the LHCb detector as well as reconstruction and simulation software is given in Sec.~\ref{sec:detector}.
The selection of signal candidates is described in Sec.~\ref{sec:selection},
and the branching fraction measurement is presented in Sec.~\ref{sec:BF}.
Studies of the backgrounds and the fit to the {\ensuremath{\PB}}\xspace candidate invariant mass distribution are in Sec.~\ref{sec:mass-fit}, with studies of the signal efficiency and a definition of the square Dalitz plot (SDP) in Sec.~\ref{sec:efficiency}.
Systematic uncertainties on, and the results for, the branching fraction are discussed in Secs.~\ref{sec:BF-syst} and~\ref{sec:BF-results} respectively.
A study of the angular moments of ${\ensuremath{\Bub}}\xspace\ensuremath{\rightarrow}\xspace{\ensuremath{\D^+}}\xspace{\ensuremath{\kaon^-}}\xspace{\ensuremath{\pion^-}}\xspace$ decays is given in Sec.~\ref{sec:moments}, with results used to guide the Dalitz plot analysis that follows.
An overview of the Dalitz plot analysis formalism is given in Sec.~\ref{sec:dalitz-generalities}, and details of the implementation of the amplitude analysis are presented in Sec.~\ref{sec:dalitz}.
The evaluation of systematic uncertainties is described in Sec.~\ref{sec:systematics}.
The results and a summary are given in Sec.~\ref{sec:results}.
\section{LHCb detector}
\label{sec:detector}
The \mbox{LHCb}\xspace detector~\cite{Alves:2008zz,LHCb-DP-2014-002} is a single-arm forward
spectrometer covering the \mbox{pseudorapidity} range $2<\eta <5$,
designed for the study of particles containing {\ensuremath{\Pb}}\xspace or {\ensuremath{\Pc}}\xspace
quarks. The detector includes a high-precision tracking system
consisting of a silicon-strip vertex detector~\cite{LHCb-DP-2014-001}
surrounding the $pp$ interaction region, a large-area silicon-strip detector
located upstream of a dipole magnet with a bending power of about
$4{\rm\,Tm}$, and three stations of silicon-strip detectors and straw
drift tubes~\cite{LHCb-DP-2013-003} placed downstream of the magnet.
The polarity of the dipole magnet is reversed periodically throughout data-taking.
The tracking system provides a measurement of momentum, \mbox{$p$}\xspace, of charged particles with
a relative uncertainty that varies from 0.5\% at low momentum to 1.0\% at 200\ensuremath{\mathrm{\,Ge\kern -0.1em V}}\xspace.
The minimum distance of a track to a primary vertex, the impact parameter (IP), is measured with a resolution of $(15+29/\mbox{$p_{\rm T}$}\xspace)\ensuremath{{\,\upmu\rm m}}\xspace$,
where \mbox{$p_{\rm T}$}\xspace is the component of the momentum transverse to the beam, in \ensuremath{\mathrm{\,Ge\kern -0.1em V}}\xspace.
Different types of charged hadrons are distinguished using information
from two ring-imaging Cherenkov detectors~\cite{LHCb-DP-2012-003}.
Photon, electron and
hadron candidates are identified by a calorimeter system consisting of
scintillating-pad and preshower detectors, an electromagnetic
calorimeter and a hadronic calorimeter. Muons are identified by a
system composed of alternating layers of iron and multiwire
proportional chambers~\cite{LHCb-DP-2012-002}.
The trigger~\cite{LHCb-DP-2012-004} consists of a
hardware stage, based on information from the calorimeter and muon
systems, followed by a software stage, in which all tracks
with $\mbox{$p_{\rm T}$}\xspace>500~(300)\ensuremath{\mathrm{\,Me\kern -0.1em V}}\xspace$ are reconstructed for data collected in 2011 (2012).
The software trigger line used in the analysis reported in this paper requires
a two-, three- or four-track secondary vertex with significant displacement
from the primary $pp$ interaction vertices~(PVs). At least one charged particle
must have $\mbox{$p_{\rm T}$}\xspace > 1.7\ensuremath{\mathrm{\,Ge\kern -0.1em V}}\xspace$ and be inconsistent with originating from the PV.
A multivariate algorithm~\cite{BBDT} is used for
the identification of secondary vertices consistent with the decay
of a {\ensuremath{\Pb}}\xspace hadron.
In the offline selection, the objects that fired the trigger are associated with reconstructed particles.
Selection requirements can therefore be made not only on the trigger line that fired, but on whether the decision was due to the signal candidate, other particles produced in the $pp$ collision, or a combination of both.
Signal candidates are accepted offline if one of the final state particles created a cluster in the hadronic calorimeter with sufficient transverse energy to fire the hardware trigger.
These candidates are referred to as ``triggered on signal'' or TOS.
Events that are triggered at the hardware level by another particle in the event, referred to as ``triggered independent of signal'' or TIS, are also retained.
After all selection requirements are imposed, 57\,\% of events in the sample
were triggered by the decay products of the signal candidate (TOS), while the remainder were triggered only by another particle in the event (TIS-only).
Simulated events are used to characterise the detector response to signal and
certain types of background events.
In the simulation, $pp$ collisions are generated using
\mbox{\textsc{Pythia}}\xspace~\cite{Sjostrand:2006za,*Sjostrand:2007gs} with a specific \mbox{LHCb}\xspace
configuration~\cite{LHCb-PROC-2010-056}. Decays of hadronic particles
are described by \mbox{\textsc{EvtGen}}\xspace~\cite{Lange:2001uf}, in which final state
radiation is generated using \mbox{\textsc{Photos}}\xspace~\cite{Golonka:2005pn}. The
interaction of the generated particles with the detector and its
response are implemented using the \mbox{\textsc{Geant4}}\xspace
toolkit~\cite{Allison:2006ve, *Agostinelli:2002hh} as described in
Ref.~\cite{LHCb-PROC-2011-006}.
\section{Selection requirements}
\label{sec:selection}
Most selection requirements are optimised using the ${\ensuremath{\Bub}}\xspace \ensuremath{\rightarrow}\xspace {\ensuremath{\D^+}}\xspace {\ensuremath{\pion^-}}\xspace {\ensuremath{\pion^-}}\xspace$ control channel.
Loose initial selection requirements on the quality of the tracks combined to form the {\ensuremath{\PB}}\xspace candidate, as well as on their $p$, $\mbox{$p_{\rm T}$}\xspace$ and $\chi^2_{\rm{IP}}$, are applied to obtain a visible peak in the invariant mass distribution.
The $\chi^2_{\rm{IP}}$ is the difference between the $\chi^2$ of the PV reconstruction with and without the considered particle.
Only candidates with invariant mass in the range $1770 < m({\ensuremath{\kaon^-}}\xspace{\ensuremath{\pion^+}}\xspace\pip) < 1968 \ensuremath{\mathrm{\,Me\kern -0.1em V}}\xspace$ are retained.
Further requirements are imposed on the vertex quality ($\chi^2_{\rm{vtx}}$) and flight distance from the associated PV of the {\ensuremath{\PB}}\xspace and {\ensuremath{\PD}}\xspace candidates.
The {\ensuremath{\PB}}\xspace candidate must also satisfy requirements on its invariant mass and on the cosine of the angle between the momentum vector and the line joining the PV under consideration to the ${\ensuremath{\PB}}\xspace$ vertex ($\cos \theta_{\rm{dir}}$).
The initial selection requirements are found to be about $90\,\%$ efficient on simulated signal decays.
Two neural networks~\cite{Feindt2006190} are used to further separate signal from background.
The first is designed to separate candidates that contain real ${\ensuremath{\D^+}}\xspace \ensuremath{\rightarrow}\xspace {\ensuremath{\kaon^-}}\xspace {\ensuremath{\pion^+}}\xspace {\ensuremath{\pion^+}}\xspace$ decays from those that do not; the second separates ${\ensuremath{\Bub}}\xspace \ensuremath{\rightarrow}\xspace {\ensuremath{\D^+}}\xspace {\ensuremath{\pion^-}}\xspace {\ensuremath{\pion^-}}\xspace$ signal decays from background combinations.
Both networks are trained using the ${\ensuremath{\D^+}}\xspace {\ensuremath{\pion^-}}\xspace {\ensuremath{\pion^-}}\xspace$ control channel, where the {\it sPlot} technique~\cite{Pivk:2004ty} is used to statistically separate ${\ensuremath{\Bub}}\xspace \ensuremath{\rightarrow}\xspace {\ensuremath{\D^+}}\xspace {\ensuremath{\pion^-}}\xspace {\ensuremath{\pion^-}}\xspace$ signal decays from background combinations using the {\ensuremath{\PD}}\xspace ({\ensuremath{\PB}}\xspace) candidate mass as the discriminating variable for the first (second) network.
The first network takes as input properties of the {\ensuremath{\PD}}\xspace candidate and its daughter tracks, including information about kinematics, track and vertex quality.
The second uses a total of 27 input variables.
They include the $\chi^2_{\rm{IP}}$ of the two ``bachelor'' pions (\mbox{\itshape i.e.}\xspace\ pions that originate directly from the {\ensuremath{\PB}}\xspace decay) and properties of the {\ensuremath{\PD}}\xspace candidate including its $\chi^2_{\rm{IP}}$, $\chi^2_{\rm{vtx}}$, $\cos \theta_{\rm{dir}}$, the output of the {\ensuremath{\PD}}\xspace neural network and the square of the flight distance divided by its uncertainty squared ($\chi^2_{\rm{flight}}$).
Variables associated to the {\ensuremath{\PB}}\xspace candidate are also used, including $\mbox{$p_{\rm T}$}\xspace$, $\chi^2_{\rm{IP}}$, $\chi^2_{\rm{vtx}}$, $\chi^2_{\rm{flight}}$ and $\cos \theta_{\rm{dir}}$.
The $\mbox{$p_{\rm T}$}\xspace$ asymmetry and track multiplicity in a cone with half-angle of 1.5 units of the plane of pseudorapidity and azimuthal angle (measured in radians) around the {\ensuremath{\PB}}\xspace candidate flight direction~\cite{LHCb-PAPER-2012-001}, which contain information about the isolation of the {\ensuremath{\PB}}\xspace candidate from the rest of the event, are also used in the network.
The neural network input quantities depend only weakly on the kinematics of the ${\ensuremath{\PB}}\xspace$ decay.
A requirement is imposed on the second neural network output that reduces the combinatorial background by an order of magnitude while retaining about $75\,\%$ of the signal.
The selection criteria for the ${\ensuremath{\Bub}}\xspace \ensuremath{\rightarrow}\xspace {\ensuremath{\D^+}}\xspace {\ensuremath{\kaon^-}}\xspace {\ensuremath{\pion^-}}\xspace$ and ${\ensuremath{\Bub}}\xspace \ensuremath{\rightarrow}\xspace {\ensuremath{\D^+}}\xspace {\ensuremath{\pion^-}}\xspace {\ensuremath{\pion^-}}\xspace$ candidates are identical except for the particle identification (PID) requirement on the bachelor track that differs between the two modes.
All five final state particles for each decay mode have PID criteria applied to preferentially select either pions or kaons.
Tight requirements are placed on the higher momentum pion from the ${\ensuremath{\D^+}}\xspace$ decay and on the bachelor kaon in ${\ensuremath{\Bub}}\xspace \ensuremath{\rightarrow}\xspace {\ensuremath{\D^+}}\xspace {\ensuremath{\kaon^-}}\xspace {\ensuremath{\pion^-}}\xspace$ to suppress backgrounds from ${\ensuremath{\D^+_\squark}}\xspace \ensuremath{\rightarrow}\xspace {\ensuremath{\kaon^-}}\xspace {\ensuremath{\kaon^+}}\xspace {\ensuremath{\pion^+}}\xspace$ and ${\ensuremath{\Bub}}\xspace \ensuremath{\rightarrow}\xspace {\ensuremath{\D^+}}\xspace {\ensuremath{\pion^-}}\xspace {\ensuremath{\pion^-}}\xspace$ decays, respectively.
The combined efficiency of the PID requirements on the five final state tracks is around $70\,\%$ for ${\ensuremath{\Bub}}\xspace \ensuremath{\rightarrow}\xspace {\ensuremath{\D^+}}\xspace {\ensuremath{\pion^-}}\xspace {\ensuremath{\pion^-}}\xspace$ decays and around $40\,\%$ for ${\ensuremath{\Bub}}\xspace \ensuremath{\rightarrow}\xspace {\ensuremath{\D^+}}\xspace {\ensuremath{\kaon^-}}\xspace {\ensuremath{\pion^-}}\xspace$ decays.
The PID efficiency depends on the kinematics of the tracks, as described in detail in Sec.~\ref{sec:efficiency}, and is determined using samples of $D^{0} \ensuremath{\rightarrow}\xspace {\ensuremath{\kaon^-}}\xspace {\ensuremath{\pion^+}}\xspace$ decays selected in data by exploiting the kinematics of the $D^{*+} \ensuremath{\rightarrow}\xspace D^{0}{\ensuremath{\pion^+}}\xspace$ decay chain to obtain clean samples without using the PID information.
To improve the {\ensuremath{\PB}}\xspace candidate invariant mass resolution, track momenta are scaled~\cite{LHCb-PAPER-2012-048,LHCb-PAPER-2013-011} with calibration parameters determined by matching the measured peak of the $J/\psi \ensuremath{\rightarrow}\xspace \mu^{+} \mu^{-}$ decay to the known $J/\psi$ mass~\cite{PDG2014}.
Furthermore, a fit to the kinematics and topology of the decay chain~\cite{Hulsbergen:2005pu} is used to adjust the four-momenta of the tracks from the {\ensuremath{\PD}}\xspace candidate so that their combined invariant mass matches the world average value for the ${\ensuremath{\D^+}}\xspace$ meson~\cite{PDG2014}.
An additional {\ensuremath{\PB}}\xspace mass constraint is applied in the calculation of the variables that are used in the Dalitz plot fit.
To remove potential background from misreconstructed ${\ensuremath{\Lz^+_\cquark}}\xspace$ decays, candidates are rejected if the invariant mass of the {\ensuremath{\PD}}\xspace candidate lies in the range $2280$--$2300\ensuremath{\mathrm{\,Me\kern -0.1em V}}\xspace$ when the proton mass hypothesis is applied to the low momentum pion track.
Possible backgrounds from ${\ensuremath{\Bub}}\xspace$ meson decays without an intermediate charm meson are suppressed by the requirement on the output value from the first neural network, and any surviving background of this type is removed by requiring that the {\ensuremath{\PD}}\xspace candidate vertex is displaced by at least $1 \ensuremath{\rm \,mm}\xspace$ from the {\ensuremath{\PB}}\xspace decay vertex.
The efficiency of this requirement is about $85\,\%$.
Signal candidates are retained for further analysis if they have an invariant mass in the range $5100$--$5800 \ensuremath{\mathrm{\,Me\kern -0.1em V}}\xspace$.
After all selection requirements are applied, fewer than $1\,\%$ of events with one candidate also contain a second candidate.
Such multiple candidates are retained and treated in the same manner as other candidates; the associated systematic uncertainty is negligible.
\section{Branching fraction determination}
\label{sec:BF}
The ratio of branching fractions is calculated from the signal yields with event-by-event efficiency corrections applied as a function of square Dalitz plot position.
The calculation is
\begin{equation}
\label{eq:BF-weighted}
\frac{\mathcal{B}({\ensuremath{\Bub}}\xspace\ensuremath{\rightarrow}\xspace{\ensuremath{\D^+}}\xspace{\ensuremath{\kaon^-}}\xspace{\ensuremath{\pion^-}}\xspace)}{\mathcal{B}({\ensuremath{\Bub}}\xspace\ensuremath{\rightarrow}\xspace{\ensuremath{\D^+}}\xspace{\ensuremath{\pion^-}}\xspace\pim)} =
\frac{N^{\rm corr}({\ensuremath{\Bub}}\xspace\ensuremath{\rightarrow}\xspace{\ensuremath{\D^+}}\xspace{\ensuremath{\kaon^-}}\xspace{\ensuremath{\pion^-}}\xspace)}{N^{\rm corr}({\ensuremath{\Bub}}\xspace\ensuremath{\rightarrow}\xspace{\ensuremath{\D^+}}\xspace{\ensuremath{\pion^-}}\xspace\pim)}\,,
\end{equation}
where $N^{\rm corr} = \sum_{i}W_i/\epsilon_i$ is the efficiency-corrected yield.
The index $i$ sums over all candidates in the data sample and $W_i$ is the signal weight for each candidate, which is determined from the fits described in Sec.~\ref{sec:mass-fit} and shown in Figs.~\ref{fig:dpimi-fit} and~\ref{fig:dkpi-fit}, using the {\it sPlot} technique~\cite{Pivk:2004ty}.
Each fit is performed simultaneously to decays in the TOS and TIS-only categories.
The efficiency of candidate $i$, $\epsilon_i$, is obtained separately for each trigger subsample as described in Sec.~\ref{sec:efficiency}.
\subsection{Determination of signal and background yields}
\label{sec:mass-fit}
The candidates that survive the selection requirements are comprised of signal decays and various categories of background.
Combinatorial background arises from random combinations of tracks (possibly including a real ${\ensuremath{\D^+}}\xspace\ensuremath{\rightarrow}\xspace{\ensuremath{\kaon^-}}\xspace{\ensuremath{\pion^+}}\xspace\pip$ decay).
Partially reconstructed backgrounds originate from {\ensuremath{\Pb}}\xspace~hadron decays with additional particles that are not part of the reconstructed decay chain.
Misidentified decays also originate from {\ensuremath{\Pb}}\xspace~hadron decays, but where one of the final state particles has been incorrectly identified (\mbox{\itshape e.g.}\xspace\ a pion as a kaon).
The signal (normalisation channel) and background yields are obtained from unbinned maximum likelihood fits to the ${\ensuremath{\D^+}}\xspace{\ensuremath{\kaon^-}}\xspace{\ensuremath{\pion^-}}\xspace$ (${\ensuremath{\D^+}}\xspace{\ensuremath{\pion^-}}\xspace\pim$) invariant mass distributions.
Both the ${\ensuremath{\Bub}}\xspace \ensuremath{\rightarrow}\xspace {\ensuremath{\D^+}}\xspace{\ensuremath{\kaon^-}}\xspace{\ensuremath{\pion^-}}\xspace$ and ${\ensuremath{\Bub}}\xspace \ensuremath{\rightarrow}\xspace {\ensuremath{\D^+}}\xspace{\ensuremath{\pion^-}}\xspace\pim$ signal shapes are modelled by the sum of two Crystal Ball (CB) functions~\cite{Skwarnicki:1986xj} with a common mean and tails on opposite sides, where the high-mass tail accounts for non-Gaussian reconstruction effects.
The ratio of widths of the CB shapes and the relative normalisation of the narrower CB shape are constrained within their uncertainties to the values found in fits to simulated signal samples.
The tail parameters of the CB shapes are also fixed to those found in simulation.
The combinatorial backgrounds in both ${\ensuremath{\D^+}}\xspace{\ensuremath{\kaon^-}}\xspace{\ensuremath{\pion^-}}\xspace$ and ${\ensuremath{\D^+}}\xspace{\ensuremath{\pion^-}}\xspace\pim$ samples are modelled with linear functions; the slope of this function is allowed to differ between the two trigger subsamples.
The decay ${\ensuremath{\Bub}}\xspace\ensuremath{\rightarrow}\xspace{\ensuremath{\D^{*+}}}\xspace{\ensuremath{\kaon^-}}\xspace{\ensuremath{\pion^-}}\xspace$ is a partially reconstructed background for ${\ensuremath{\D^+}}\xspace{\ensuremath{\kaon^-}}\xspace{\ensuremath{\pion^-}}\xspace$ candidates, where the ${\ensuremath{\D^{*+}}}\xspace$ decays to either ${\ensuremath{\D^+}}\xspace\gamma$ or ${\ensuremath{\D^+}}\xspace{\ensuremath{\pion^0}}\xspace$ and the neutral particle is not reconstructed.
Similarly the decay ${\ensuremath{\Bub}}\xspace\ensuremath{\rightarrow}\xspace{\ensuremath{\D^{*+}}}\xspace{\ensuremath{\pion^-}}\xspace\pim$ forms a partially reconstructed background to the ${\ensuremath{\D^+}}\xspace{\ensuremath{\pion^-}}\xspace\pim$ final state.
These are modelled with non-parametric shapes determined from simulated samples.
The shapes are characterised by a sharp edge around $100 \ensuremath{\mathrm{\,Me\kern -0.1em V}}\xspace$ below the {\ensuremath{\PB}}\xspace peak, where the exact position of the edge depends on properties of the decay including the ${\ensuremath{\D^{*+}}}\xspace$ polarisation.
The fit quality improves when the shape is allowed to be offset by a small shift that is determined from the data.
Most potential sources of misidentified backgrounds have broad ${\ensuremath{\PB}}\xspace$ candidate invariant mass distributions, and hence are absorbed in the combinatorial background component in the fit.
The decays ${\ensuremath{\Bub}}\xspace\ensuremath{\rightarrow}\xspace D^{(*)+} {\ensuremath{\pion^-}}\xspace\pim$ and ${\ensuremath{\Bub}}\xspace\ensuremath{\rightarrow}\xspace{\ensuremath{\D^+_\squark}}\xspace{\ensuremath{\kaon^-}}\xspace{\ensuremath{\pion^-}}\xspace$, however, give distinctive shapes in the mass distribution of ${\ensuremath{\D^+}}\xspace{\ensuremath{\kaon^-}}\xspace{\ensuremath{\pion^-}}\xspace$ candidates.
For ${\ensuremath{\D^+}}\xspace{\ensuremath{\pion^-}}\xspace\pim$ candidates the only significant misidentified background contribution is from ${\ensuremath{\Bub}}\xspace\ensuremath{\rightarrow}\xspace D^{(*)+} {\ensuremath{\kaon^-}}\xspace{\ensuremath{\pion^-}}\xspace$ decays.
The misidentified background shapes are also modelled with non-parametric shapes determined from simulated samples.
The simulated samples used to obtain signal and background shapes are generated with flat distributions in the phase space of their SDPs.
For ${\ensuremath{\Bub}}\xspace\ensuremath{\rightarrow}\xspace {\ensuremath{\D^+}}\xspace {\ensuremath{\pion^-}}\xspace\pim$ and ${\ensuremath{\Bub}}\xspace\ensuremath{\rightarrow}\xspace {\ensuremath{\D^{*+}}}\xspace {\ensuremath{\pion^-}}\xspace\pim$ decays, accurate models of the distributions across the SDP are known~\cite{Abe:2003zm,Aubert:2009wg}, so the simulated samples are reweighted using the ${\ensuremath{\Bub}}\xspace\ensuremath{\rightarrow}\xspace {\ensuremath{\D^+}}\xspace {\ensuremath{\pion^-}}\xspace {\ensuremath{\pion^-}}\xspace$ data sample; this affects the shape of the misidentified background component in the fit to the ${\ensuremath{\D^+}}\xspace{\ensuremath{\kaon^-}}\xspace{\ensuremath{\pion^+}}\xspace$ sample.
Additionally, the ${\ensuremath{\D^+}}\xspace$ and ${\ensuremath{\D^{*+}}}\xspace$ portions of this background are combined according to their known branching fractions.
All of the shapes, except for that of the combinatorial background, are common between the two trigger subsamples in each fit, but the signal and background yields in the subsamples are independent.
In total there are 15 free parameters in the fit to the ${\ensuremath{\D^+}}\xspace{\ensuremath{\pion^-}}\xspace\pim$ sample:
yields in each subsample for signal, combinatorial, ${\ensuremath{\Bub}}\xspace\ensuremath{\rightarrow}\xspace D^{(*)+} {\ensuremath{\kaon^-}}\xspace{\ensuremath{\pion^-}}\xspace$ and ${\ensuremath{\Bub}}\xspace\ensuremath{\rightarrow}\xspace{\ensuremath{\D^{*+}}}\xspace{\ensuremath{\pion^-}}\xspace\pim$ backgrounds;
the combinatorial slope in each subsample;
the double CB peak position, the width of the narrower CB, the ratio of CB widths and the fraction of entries in the narrower CB shape;
and the shift parameter of the partially reconstructed background.
The result of the ${\ensuremath{\D^+}}\xspace{\ensuremath{\pion^-}}\xspace\pim$ fit is shown in Fig.~\ref{fig:dpimi-fit} for both trigger subsamples and gives a combined signal yield of approximately 49\,000 decays.
Component yields are given in Table~\ref{tab:DpipiFit_yields}.
\begin{figure}[!tb]
\centering
\includegraphics[scale=0.38]{Fig1a.pdf}
\includegraphics[scale=0.38]{Fig1b.pdf}
\caption{\small
Results of the fit to the ${\ensuremath{\Bub}}\xspace \ensuremath{\rightarrow}\xspace {\ensuremath{\D^+}}\xspace{\ensuremath{\pion^-}}\xspace\pim$ candidate invariant mass distribution for the (left) TOS and (right) TIS-only subsamples.
Data points are shown in black, the full fitted model as solid blue lines and the components as shown in the legend.}
\label{fig:dpimi-fit}
\end{figure}
\begin{table}[!tb]
\centering
\caption{\small
Yields of the various components in the fit to ${\ensuremath{\Bub}}\xspace\ensuremath{\rightarrow}\xspace{\ensuremath{\D^+}}\xspace{\ensuremath{\pion^-}}\xspace\pim$ candidate invariant mass distribution.}
\centering
\begin{tabular}{lcc}
\hline
Component & TOS & TIS-only \\
\hline
$N({\ensuremath{\Bub}}\xspace\ensuremath{\rightarrow}\xspace{\ensuremath{\D^+}}\xspace{\ensuremath{\pion^-}}\xspace\pim)$ & $29\,190 \pm 204$ & $19\,416 \pm 159$ \\
$N({\ensuremath{\Bub}}\xspace \ensuremath{\rightarrow}\xspace D^{(*)+}{\ensuremath{\kaon^-}}\xspace{\ensuremath{\pion^-}}\xspace)$ & $\phantom{29\,}807 \pm 123$ & $\phantom{19\,}401 \pm 84\phantom{9}$ \\
$N({\ensuremath{\Bub}}\xspace \ensuremath{\rightarrow}\xspace {\ensuremath{\D^{*+}}}\xspace{\ensuremath{\pion^-}}\xspace\pim)$ & $12\,120 \pm 115$ & $\phantom{1\,}8551 \pm 96\phantom{9}$ \\
$N(\rm{comb. \ bkgd.})$ & $\phantom{29\,}784 \pm 54\phantom{4}$ & $\phantom{9\,4}746 \pm 47\phantom{9}$ \\
\hline
\end{tabular}
\label{tab:DpipiFit_yields}
\end{table}
There are a total of 17 free parameters in the fit to the ${\ensuremath{\D^+}}\xspace{\ensuremath{\kaon^-}}\xspace{\ensuremath{\pion^-}}\xspace$ sample:
yields in each subsample for signal, combinatorial, ${\ensuremath{\Bub}}\xspace\ensuremath{\rightarrow}\xspace{\ensuremath{\D^{*+}}}\xspace{\ensuremath{\kaon^-}}\xspace{\ensuremath{\pion^-}}\xspace$, ${\ensuremath{\Bub}}\xspace\ensuremath{\rightarrow}\xspace{\ensuremath{\D^+_\squark}}\xspace{\ensuremath{\kaon^-}}\xspace{\ensuremath{\pion^-}}\xspace$ and ${\ensuremath{\Bub}}\xspace\ensuremath{\rightarrow}\xspace D^{(*)+} {\ensuremath{\pion^-}}\xspace\pim$ backgrounds;
the combinatorial slope in each subsample;
the same signal shape parameters as for the ${\ensuremath{\D^+}}\xspace{\ensuremath{\pion^-}}\xspace\pim$ fit;
and the shift parameter of the partially reconstructed background.
Figure~\ref{fig:dkpi-fit} shows the result of the ${\ensuremath{\D^+}}\xspace{\ensuremath{\kaon^-}}\xspace{\ensuremath{\pion^-}}\xspace$ fit for the two trigger subsamples that yield a total of approximately 2000 ${\ensuremath{\Bub}}\xspace\ensuremath{\rightarrow}\xspace{\ensuremath{\D^+}}\xspace{\ensuremath{\kaon^-}}\xspace{\ensuremath{\pion^-}}\xspace$ decays.
The yields of all fit components are shown in Table~\ref{tab:DKpiFit_yields}.
The statistical signal significance, estimated in the conventional way from the change in negative log-likelihood from the fit when the signal component is removed, is in excess of $60$ standard deviations ($\sigma$).
\begin{figure}[!tb]
\centering
\includegraphics[scale=0.38]{Fig2a.pdf}
\includegraphics[scale=0.38]{Fig2b.pdf}
\caption{\small
Results of the fit to the ${\ensuremath{\Bub}}\xspace \ensuremath{\rightarrow}\xspace {\ensuremath{\D^+}}\xspace{\ensuremath{\kaon^-}}\xspace{\ensuremath{\pion^-}}\xspace$ candidate invariant mass distribution for the (left) TOS and (right) TIS-only subsamples.
Data points are shown in black, the full fitted model as solid blue lines and the components as shown in the legend.}
\label{fig:dkpi-fit}
\end{figure}
\begin{table}[!tb]
\centering
\caption{\small
Yields of the various components in the fit to ${\ensuremath{\Bub}}\xspace\ensuremath{\rightarrow}\xspace{\ensuremath{\D^+}}\xspace{\ensuremath{\kaon^-}}\xspace{\ensuremath{\pion^-}}\xspace$ candidate invariant mass distribution.}
\centering
\begin{tabular}{lcc}
\hline
Component & TOS & TIS-only \\
\hline
$N({\ensuremath{\Bub}}\xspace\ensuremath{\rightarrow}\xspace{\ensuremath{\D^+}}\xspace{\ensuremath{\kaon^-}}\xspace{\ensuremath{\pion^-}}\xspace)$ & $1112 \pm 37$ & $891 \pm 32$ \\
$N({\ensuremath{\Bub}}\xspace \ensuremath{\rightarrow}\xspace D^{(*)+}{\ensuremath{\pion^-}}\xspace\pim)$ & $\phantom{1}114 \pm 34$ & $\phantom{8}23 \pm 27$ \\
$N({\ensuremath{\Bub}}\xspace \ensuremath{\rightarrow}\xspace {\ensuremath{\D^+_\squark}}\xspace{\ensuremath{\kaon^-}}\xspace{\ensuremath{\pion^-}}\xspace)$ & $\phantom{11}69 \pm 17$ & $\phantom{8}40 \pm 15$ \\
$N({\ensuremath{\Bub}}\xspace \ensuremath{\rightarrow}\xspace {\ensuremath{\D^{*+}}}\xspace{\ensuremath{\kaon^-}}\xspace{\ensuremath{\pion^-}}\xspace)$ & $\phantom{1}518 \pm 26$ & $361 \pm 21$ \\
$N(\rm{comb. \ bkgd.})$ & $\phantom{1}238 \pm 38$ & $253 \pm 36$ \\
\hline
\end{tabular}
\label{tab:DKpiFit_yields}
\end{table}
\subsection{Signal efficiency}
\label{sec:efficiency}
Since both ${\ensuremath{\Bub}}\xspace \ensuremath{\rightarrow}\xspace {\ensuremath{\D^+}}\xspace {\ensuremath{\kaon^-}}\xspace {\ensuremath{\pion^-}}\xspace$ and ${\ensuremath{\Bub}}\xspace \ensuremath{\rightarrow}\xspace {\ensuremath{\D^+}}\xspace {\ensuremath{\pion^-}}\xspace {\ensuremath{\pion^-}}\xspace$ decays have
non-trivial DP distributions, it is necessary to understand the variation of
the efficiency across the phase space.
Since, moreover, the efficiency variation tends to be strongest close to the kinematic boundaries of the conventional Dalitz plot, it is convenient to model these effects in terms of the SDP defined by variables \mpr\ and \thpr\ which are valid in the range 0 to~1 and are given for the ${\ensuremath{\D^+}}\xspace {\ensuremath{\kaon^-}}\xspace {\ensuremath{\pion^-}}\xspace$ case by
\begin{equation}
\label{eq:sqdp-vars}
\mpr \equiv \frac{1}{\pi} \arccos\left(2\frac{m({\ensuremath{\D^+}}\xspace{\ensuremath{\pion^-}}\xspace) - m^{\rm min}_{{\ensuremath{\D^+}}\xspace{\ensuremath{\pion^-}}\xspace}}{m^{\rm max}_{{\ensuremath{\D^+}}\xspace{\ensuremath{\pion^-}}\xspace} - m^{\rm min}_{{\ensuremath{\D^+}}\xspace{\ensuremath{\pion^-}}\xspace}} - 1 \right)
\hspace{10mm}{\rm and}\hspace{10mm}
\thpr \equiv \frac{1}{\pi}\theta({\ensuremath{\D^+}}\xspace{\ensuremath{\pion^-}}\xspace)\,,
\end{equation}
where $m^{\rm max}_{{\ensuremath{\D^+}}\xspace{\ensuremath{\pion^-}}\xspace} = m_{{\ensuremath{\Bub}}\xspace} - m_{{\ensuremath{\kaon^-}}\xspace}$ and $m^{\rm min}_{{\ensuremath{\D^+}}\xspace{\ensuremath{\pion^-}}\xspace} =
m_{{\ensuremath{\D^+}}\xspace} + m_{{\ensuremath{\pion^-}}\xspace}$ are the kinematic boundaries of $m({\ensuremath{\D^+}}\xspace{\ensuremath{\pion^-}}\xspace)$ allowed in the ${\ensuremath{\Bub}}\xspace \ensuremath{\rightarrow}\xspace {\ensuremath{\D^+}}\xspace{\ensuremath{\kaon^-}}\xspace{\ensuremath{\pion^-}}\xspace$ decay and $\theta({\ensuremath{\D^+}}\xspace{\ensuremath{\pion^-}}\xspace)$ is the helicity angle of the ${\ensuremath{\D^+}}\xspace{\ensuremath{\pion^-}}\xspace$ system (the angle between the ${\ensuremath{\kaon^-}}\xspace$ and the ${\ensuremath{\D^+}}\xspace$ meson momenta in the ${\ensuremath{\D^+}}\xspace{\ensuremath{\pion^-}}\xspace$ rest frame).
For the ${\ensuremath{\D^+}}\xspace {\ensuremath{\pion^-}}\xspace {\ensuremath{\pion^-}}\xspace$ case, $\mpr$ and $\thpr$ are defined in terms of the ${\ensuremath{\pion^-}}\xspace\pim$ mass and helicity angle, respectively, since with this choice only the region of the SDP with $\thpr({\ensuremath{\pion^-}}\xspace\pim) < 0.5$ is populated due to the symmetry of the two pions in the final state.
Efficiency variation across the SDP is caused by the detector acceptance and by trigger, selection and PID requirements.
The efficiency variation is evaluated for both ${\ensuremath{\D^+}}\xspace{\ensuremath{\kaon^-}}\xspace{\ensuremath{\pion^-}}\xspace$ and ${\ensuremath{\D^+}}\xspace{\ensuremath{\pion^-}}\xspace\pim$ final states with simulated samples generated uniformly over the SDP.
Data-driven corrections are applied to correct for known differences between data and simulation in the tracking, trigger and PID efficiencies, using identical methods to those described in Ref.~\cite{LHCb-PAPER-2014-036}.
The efficiency functions are fitted with two-dimensional cubic splines to smooth out statistical fluctuations due to limited sample size.
The efficiency is studied separately for the TOS and TIS-only categories.
The efficiency maps for each trigger subsample are shown for ${\ensuremath{\Bub}}\xspace \ensuremath{\rightarrow}\xspace {\ensuremath{\D^+}}\xspace {\ensuremath{\kaon^-}}\xspace {\ensuremath{\pion^-}}\xspace$ decays in Fig.~\ref{fig:eff}.
Regions of relatively high efficiency are seen where all decay products have comparable momentum in the {\ensuremath{\PB}}\xspace rest frame; the efficiency drops sharply in regions with a low momentum bachelor track due to geometrical effects.
The efficiency maps are used to calculate the ratio of branching fractions and also as inputs to the ${\ensuremath{\D^+}}\xspace{\ensuremath{\kaon^-}}\xspace{\ensuremath{\pion^-}}\xspace$ Dalitz plot fit.
\begin{figure}[!tb]
\centering
\includegraphics[scale=0.38]{Fig3a.pdf}
\includegraphics[scale=0.38]{Fig3b.pdf}
\caption{\small
Signal efficiency across the SDP for (left) TOS and (right) TIS-only ${\ensuremath{\Bub}}\xspace \ensuremath{\rightarrow}\xspace {\ensuremath{\D^+}}\xspace {\ensuremath{\kaon^-}}\xspace {\ensuremath{\pion^-}}\xspace$ decays.
The relative uncertainty at each point is typically $5\,\%$.
}
\label{fig:eff}
\end{figure}
\subsection{Systematic uncertainties}
\label{sec:BF-syst}
Table~\ref{tab:bf-syst} summarises the systematic uncertainties on the
measurement of the ratio of branching fractions.
Selection effects cancel in the ratio of branching fractions, except for inefficiency due to the ${\ensuremath{\Lz^+_\cquark}}\xspace$ veto.
The invariant mass fits are repeated both with a wider veto ($2270$--$2310\ensuremath{\mathrm{\,Me\kern -0.1em V}}\xspace$) and with no veto, and changes in the yields are used to assign a relative systematic uncertainty of $0.2\,\%$.
To estimate the uncertainty arising from the choice of invariant mass fit model, the ${\ensuremath{\D^+}}\xspace{\ensuremath{\kaon^-}}\xspace{\ensuremath{\pion^-}}\xspace$ mass fit is varied by replacing the signal shape with the sum of two bifurcated Gaussian functions, removing the smoothing of the non-parametric functions, using exponential and second-order polynomial functions to describe the combinatorial background, varying fixed parameters within their uncertainties and varying the binning of histograms used to reweight the simulated background samples.
For the ${\ensuremath{\D^+}}\xspace{\ensuremath{\pion^-}}\xspace\pim$ fit the same variations are made.
The relative changes in the yields are summed in quadrature to give a relative systematic uncertainty on the ratio of branching fractions of $2.0\,\%$.
The systematic uncertainty due to PID is estimated by accounting for three sources:
the intrinsic uncertainty of the calibration ($1.0\,\%$);
possible differences in the kinematics of tracks in simulated samples, used to reweight the calibration data samples, to those in the data ($1.7\,\%$);
the granularity of the binning in the reweighting procedure ($0.7\,\%$).
Combining these in quadrature, the total relative systematic uncertainty from PID is $2.1\,\%$.
The bins of the efficiency maps are varied within uncertainties to make 100 new efficiency maps, for both ${\ensuremath{\D^+}}\xspace{\ensuremath{\kaon^-}}\xspace{\ensuremath{\pion^-}}\xspace$ and ${\ensuremath{\D^+}}\xspace{\ensuremath{\pion^-}}\xspace\pim$ modes.
The efficiency-corrected yields are evaluated for each new map and their distributions are fitted with Gaussian functions.
The widths of these are used to assign a relative systematic uncertainty on the ratio of branching fractions of $0.8\,\%$.
\begin{table}[!tb]
\centering
\caption{\small
Relative systematic uncertainties on the measurement of the ratio of branching fractions for ${\ensuremath{\Bub}}\xspace \ensuremath{\rightarrow}\xspace {\ensuremath{\D^+}}\xspace{\ensuremath{\kaon^-}}\xspace{\ensuremath{\pion^-}}\xspace$ and ${\ensuremath{\Bub}}\xspace \ensuremath{\rightarrow}\xspace {\ensuremath{\D^+}}\xspace{\ensuremath{\pion^-}}\xspace\pim$ decays.
}
\label{tab:bf-syst}
\begin{tabular}{lc}
\hline
Source & Uncertainty (\%) \\
\hline
${\ensuremath{\Lz^+_\cquark}}\xspace$ veto & 0.2 \\
Fit model & 2.0 \\
Particle identification & 2.1 \\
Efficiency modelling & 0.8 \\
\hline
Total & 3.0 \\
\hline
\end{tabular}
\end{table}
A number of additional cross-checks are performed to test the branching fraction result. The neural network and PID requirements are both tightened and loosened. The data sample is divided by dipole magnet polarity and year of data taking. The branching fraction is also calculated separately for TOS and TIS-only events. All cross-checks give consistent results.
\subsection{Results}
\label{sec:BF-results}
The ratio of branching fractions is found to be
\begin{equation}
\nonumber \frac{\mathcal{B}({\ensuremath{\Bub}}\xspace\ensuremath{\rightarrow}\xspace{\ensuremath{\D^+}}\xspace{\ensuremath{\kaon^-}}\xspace{\ensuremath{\pion^-}}\xspace)}{\mathcal{B}({\ensuremath{\Bub}}\xspace\ensuremath{\rightarrow}\xspace{\ensuremath{\D^+}}\xspace{\ensuremath{\pion^-}}\xspace\pim)} = 0.0720 \pm 0.0019 \pm 0.0021\,,
\end{equation}
where the first uncertainty is statistical and the second systematic.
The statistical uncertainty includes contributions from the event weighting used in Eq.~(\ref{eq:BF-weighted}) and from the shape parameters that are allowed to vary in the fit~\cite{LHCb-PAPER-2012-018}.
The world average value of $\mathcal{B}({\ensuremath{\Bub}}\xspace\ensuremath{\rightarrow}\xspace{\ensuremath{\D^+}}\xspace{\ensuremath{\pion^-}}\xspace\pim) = (1.07 \pm 0.05) \times 10^{-3}$~\cite{PDG2014} assumes that ${\ensuremath{\Bu}}\xspace{\ensuremath{\Bub}}\xspace$ and ${\ensuremath{\B^0}}\xspace{\ensuremath{\Bbar{}^0}}\xspace$ are produced equally in the decay of the $\Upsilon(4S)$ resonance.
Using $\Gamma(\Upsilon(4S)\ensuremath{\rightarrow}\xspace{\ensuremath{\Bu}}\xspace{\ensuremath{\Bub}}\xspace)/\Gamma(\Upsilon(4S)\ensuremath{\rightarrow}\xspace{\ensuremath{\B^0}}\xspace{\ensuremath{\Bbar{}^0}}\xspace) = 1.055 \pm 0.025$~\cite{PDG2014} gives a corrected value of $\mathcal{B}({\ensuremath{\Bub}}\xspace\ensuremath{\rightarrow}\xspace{\ensuremath{\D^+}}\xspace{\ensuremath{\pion^-}}\xspace\pim) = (1.01 \pm 0.05) \times 10^{-3}$.
This allows the branching fraction of ${\ensuremath{\Bub}}\xspace\ensuremath{\rightarrow}\xspace{\ensuremath{\D^+}}\xspace{\ensuremath{\kaon^-}}\xspace{\ensuremath{\pion^-}}\xspace$ decays to be determined as
\begin{equation}
\nonumber \mathcal{B}({\ensuremath{\Bub}}\xspace\ensuremath{\rightarrow}\xspace{\ensuremath{\D^+}}\xspace{\ensuremath{\kaon^-}}\xspace{\ensuremath{\pion^-}}\xspace) = (7.31 \pm 0.19 \pm 0.22 \pm 0.39) \times 10^{-5}\,,
\end{equation}
where the third uncertainty is from $\mathcal{B}({\ensuremath{\Bub}}\xspace\ensuremath{\rightarrow}\xspace{\ensuremath{\D^+}}\xspace{\ensuremath{\pion^-}}\xspace\pim)$.
This measurement represents the first observation of the ${\ensuremath{\Bub}}\xspace\ensuremath{\rightarrow}\xspace{\ensuremath{\D^+}}\xspace{\ensuremath{\kaon^-}}\xspace{\ensuremath{\pion^-}}\xspace$ decay.
\section{Study of angular moments}
\label{sec:moments}
To investigate which amplitudes should be included in the DP analysis of ${\ensuremath{\Bub}}\xspace
\ensuremath{\rightarrow}\xspace {\ensuremath{\D^+}}\xspace{\ensuremath{\kaon^-}}\xspace{\ensuremath{\pion^-}}\xspace$ decays, a study of its angular moments is performed.
Such an analysis is particularly useful for ${\ensuremath{\Bub}}\xspace\ensuremath{\rightarrow}\xspace{\ensuremath{\D^+}}\xspace{\ensuremath{\kaon^-}}\xspace{\ensuremath{\pion^-}}\xspace$ decays because resonant contributions are only expected to appear in the ${\ensuremath{\D^+}}\xspace{\ensuremath{\pion^-}}\xspace$ combination, and therefore the distributions should be free of effects from reflections that make them more difficult to interpret.
The analysis is performed by calculating moments from the Legendre polynomials $P_L$ of order up to $2 J_{\rm max}$, where $J_{\rm max}$ is the maximum spin of the resonances considered.
Each candidate is weighted according to its value of $P_L\left(\cos\theta({\ensuremath{\D^+}}\xspace{\ensuremath{\pion^-}}\xspace)\right)$ with an efficiency correction applied, and background contributions subtracted.
The results for $J_{\rm max}=3$ are shown in Fig.~\ref{fig:moments2} for the ${\ensuremath{\D^+}}\xspace{\ensuremath{\pion^-}}\xspace$ invariant mass range $2.0$--$3.0\ensuremath{\mathrm{\,Ge\kern -0.1em V}}\xspace$.
The distributions of $\left\langle P_5 \right\rangle$ and $\left\langle P_6 \right\rangle$ are compatible with being flat, which implies that there are no significant spin~3 contributions.
Considering only contributions up to spin~2, the following expressions are used to interpret Fig.~\ref{fig:moments2}
\begin{align}
\label{eq:p0}
\left\langle P_{0}\right\rangle \propto\, &
\left|h_0\right|{}^2+\left|h_1\right|{}^2+\left|h_2\right|{}^2\,, \\
\label{eq:p1}
\left\langle P_1\right\rangle \propto\, &
\frac{2}{\sqrt{3}} \left|h_0\right| \left|h_1\right| \cos \left(\delta _0-\delta
_1\right)+\frac{4}{\sqrt{15}} \left|h_1\right| \left|h_2\right| \cos \left(\delta
_1-\delta _2\right)\,, \\
\label{eq:p2}
\left\langle P_{2}\right\rangle \propto\, &
\frac{2}{\sqrt{5}}\left|h_0\right| \left|h_2\right| \cos \left(\delta
_0-\delta _2\right)+
\frac{2}{5} \left|h_1\right|{}^2 + \frac{2}{7} \left|h_2\right|{}^2\,, \\
\label{eq:p3}
\left\langle P_{3}\right\rangle \propto\, &
\frac{6}{7} \sqrt{\frac{3}{5}} \left|h_1\right| \left|h_2\right| \cos \left(\delta
_1-\delta _2\right)\,, \\
\label{eq:p4}
\left\langle P_{4}\right\rangle \propto\, &
\frac{2}{7}\left|h_2\right|{}^2\,,
\end{align}
where S-, P- and D-wave contributions are denoted by amplitudes $h_j e^{i \delta_j}$ ($j=0,1,2$ respectively).
The $D^*_2(2460)^0$ resonance is clearly seen in the $\left\langle P_4 \right\rangle$ distribution of Fig.~\ref{fig:moments2}(e).
The distribution of $\left\langle P_3 \right\rangle$ shows interference between spin~1 and 2 contributions, indicating the presence of a broad, possibly nonresonant, spin~1 contribution at low $m({\ensuremath{\D^+}}\xspace{\ensuremath{\pion^-}}\xspace)$.
The difference in shape between $\left\langle P_1 \right\rangle$ and $\left\langle P_3 \right\rangle$ shows interference between spin~1 and 0 indicating that a broad spin~0 component is similarly needed.
\begin{figure}[!htp]
\centering
\includegraphics[scale=0.34]{Fig4a.pdf}
\includegraphics[scale=0.34]{Fig4b.pdf}
\includegraphics[scale=0.34]{Fig4c.pdf}
\includegraphics[scale=0.34]{Fig4d.pdf}
\includegraphics[scale=0.34]{Fig4e.pdf}
\includegraphics[scale=0.34]{Fig4f.pdf}
\includegraphics[scale=0.34]{Fig4g.pdf}
\caption{\small
The first seven Legendre-polynomial weighted moments for background-subtracted and efficiency-corrected ${\ensuremath{\Bub}}\xspace \ensuremath{\rightarrow}\xspace {\ensuremath{\D^+}}\xspace{\ensuremath{\kaon^-}}\xspace{\ensuremath{\pion^-}}\xspace$ data (black points) as a function of $m({\ensuremath{\D^+}}\xspace{\ensuremath{\pion^-}}\xspace)$ in the range $2.0$--$3.0\ensuremath{\mathrm{\,Ge\kern -0.1em V}}\xspace$.
Candidates from both TOS and TIS-only subsamples are included.
The blue line shows the result of the DP fit described in Sec.~\ref{sec:dalitz}.}
\label{fig:moments2}
\end{figure}
\section{Dalitz plot analysis formalism}
\label{sec:dalitz-generalities}
A Dalitz plot~\cite{Dalitz:1953cp} is a representation of the phase-space for a three-body decay in terms of two of the three possible two-body invariant mass squared combinations.
In ${\ensuremath{\Bub}}\xspace \ensuremath{\rightarrow}\xspace {\ensuremath{\D^+}}\xspace{\ensuremath{\kaon^-}}\xspace{\ensuremath{\pion^-}}\xspace$ decays, resonances are expected in the $m^2({\ensuremath{\D^+}}\xspace{\ensuremath{\pion^-}}\xspace)$ combination, therefore this and $m^2({\ensuremath{\D^+}}\xspace{\ensuremath{\kaon^-}}\xspace)$ are chosen to define the DP axes.
For a fixed ${\ensuremath{\Bub}}\xspace$ mass, all other relevant kinematic quantities can be calculated from these two invariant mass squared combinations.
The complex decay amplitude is described using the isobar approach~\cite{Fleming:1964zz,Morgan:1968zza,Herndon:1973yn}, where the total amplitude is calculated as a coherent sum of amplitudes from resonant and nonresonant intermediate processes.
The total amplitude is then given by
\begin{equation}\label{eqn:amp}
{\cal A}\left(m^2({\ensuremath{\D^+}}\xspace{\ensuremath{\pion^-}}\xspace), m^2({\ensuremath{\D^+}}\xspace{\ensuremath{\kaon^-}}\xspace)\right) = \sum_{j=1}^{N} c_j F_j\left(m^2({\ensuremath{\D^+}}\xspace{\ensuremath{\pion^-}}\xspace), m^2({\ensuremath{\D^+}}\xspace{\ensuremath{\kaon^-}}\xspace)\right) \,,
\end{equation}
where $c_j$ are complex coefficients giving the relative contribution of each intermediate process.
The $F_j\left(m^2({\ensuremath{\D^+}}\xspace{\ensuremath{\pion^-}}\xspace),m^2({\ensuremath{\D^+}}\xspace{\ensuremath{\kaon^-}}\xspace)\right)$ terms contain the resonance dynamics, which are
composed of several terms and
are normalised such that the integral of the squared magnitude over the DP is unity for each term.
For a ${\ensuremath{\D^+}}\xspace{\ensuremath{\pion^-}}\xspace$ resonance
\begin{equation}
\label{eq:ResDynEqn}
F\left(m^2({\ensuremath{\D^+}}\xspace{\ensuremath{\pion^-}}\xspace), m^2({\ensuremath{\D^+}}\xspace{\ensuremath{\kaon^-}}\xspace)\right) =
R\left(m({\ensuremath{\D^+}}\xspace{\ensuremath{\pion^-}}\xspace)\right) \times X(|\vec{p}\,|\,r_{\rm BW}) \times X(|\vec{q}\,|\,r_{\rm BW})
\times T(\vec{p},\vec{q}\,) \, ,
\end{equation}
where the functions $R$, $X$ and $T$ are described below, and $\vec{p}$ and $\vec{q}$ are the bachelor particle momentum and the momentum of one of the resonance daughters, respectively, both evaluated in the ${\ensuremath{\D^+}}\xspace{\ensuremath{\pion^-}}\xspace$ rest frame.
The $X(z)$ terms, where $z=|\vec{q}\,|\,r_{\rm BW}$ or $|\vec{p}\,|\,r_{\rm BW}$, are Blatt--Weisskopf barrier factors~\cite{blatt-weisskopf} with barrier radius $r_{\rm BW}$, and are given by
\begin{equation}\begin{array}{rcl}
L = 0 \ : \ X(z) & = & 1\,, \\
L = 1 \ : \ X(z) & = & \sqrt{\frac{1 + z_0^2}{1 + z^2}}\,, \\
L = 2 \ : \ X(z) & = & \sqrt{\frac{z_0^4 + 3z_0^2 + 9}{z^4 + 3z^2 + 9}}\,,\\
L = 3 \ : \ X(z) & = & \sqrt{\frac{z_0^6 + 6z_0^4 + 45z_0^2 + 225}{z^6 + 6z^4 + 45z^2 + 225}}\,,
\end{array}\label{eq:BWFormFactors}\end{equation}
where $z_0$ is the value of $z$ when the invariant mass is equal to the pole mass of the resonance and $L$ is the spin of the resonance. For a ${\ensuremath{\D^+}}\xspace{\ensuremath{\pion^-}}\xspace$ resonance, since the {\ensuremath{\Bub}}\xspace meson has zero spin, $L$ is also the orbital angular momentum between the resonance and the kaon.
The barrier radius, $r_{\rm BW}$, is taken to be $4.0\ensuremath{\mathrm{\,Ge\kern -0.1em V}}\xspace^{-1}\approx 0.8\ensuremath{\rm \,fm}\xspace$~\cite{Aubert:2005ce,LHCb-PAPER-2014-036} for all resonances.
The terms $T(\vec{p},\vec{q})$ describe the angular probability distribution and are
given in the Zemach tensor formalism~\cite{Zemach:1963bc,Zemach:1968zz} by
\begin{equation}\begin{array}{rcl}
L = 0 \ : \ T(\vec{p},\vec{q}) & = & 1\,,\\
L = 1 \ : \ T(\vec{p},\vec{q}) & = & -\,2\,\vec{p}\cdot\vec{q}\,,\\
L = 2 \ : \ T(\vec{p},\vec{q}) & = & \frac{4}{3} \left[3(\vec{p}\cdot\vec{q}\,)^2 - (|\vec{p}\,||\vec{q}\,|)^2\right]\,,\\
L = 3 \ : \ T(\vec{p},\vec{q}) & = & -\,\frac{24}{15} \left[5(\vec{p}\cdot\vec{q}\,)^3 - 3(\vec{p}\cdot\vec{q}\,)(|\vec{p}\,||\vec{q}\,|)^2\right]\,,
\end{array}\label{eq:ZTFactors}\end{equation}
which are proportional to the Legendre polynomials, $P_L(x)$, where $x$ is the cosine of the angle between $\vec{p}$ and $\vec{q}$ (referred to as the helicity angle).
The function $R\left(m({\ensuremath{\D^+}}\xspace{\ensuremath{\pion^-}}\xspace)\right)$ of Eq.~(\ref{eq:ResDynEqn}) is the mass lineshape.
The resonant contributions considered in the DP model are described by the relativistic Breit--Wigner (RBW) function
\begin{equation}
\label{eq:RelBWEqn}
R(m) = \frac{1}{(m_0^2 - m^2) - i\, m_0 \Gamma(m)} \,,
\end{equation}
where the mass-dependent decay width is
\begin{equation}
\label{eq:GammaEqn}
\Gamma(m) = \Gamma_0 \left(\frac{q}{q_0}\right)^{2L+1}
\left(\frac{m_0}{m}\right) X^2(q\,r_{\rm BW}) \,,
\end{equation}
where $q_0$ is the value of $q = |\vec{q}\,|$ for $m = m_0$.
Virtual contributions, from resonances
with pole masses outside the kinematically accessible region of the phase space, can also be modelled by this shape with one modification:
the pole mass $m_0$ is replaced with $m_0^{\rm{eff}}$, a mass in the kinematically allowed region, in the calculation of the parameter $q_{0}$.
This effective mass is defined by the {\it ad hoc} formula~\cite{LHCb-PAPER-2014-036}
\begin{equation}\label{eqn:effmass}
m_0^{\rm{eff}}(m_0) = m^{\rm{min}} + (m^{\rm{max}} - m^{\rm{min}}) \left( 1 + \tanh\left( \frac{m_0 - \frac{m^{\rm{min}}+m^{\rm{max}}}{2}}{m^{\rm{max}}-m^{\rm{min}}} \right) \right)\, ,
\end{equation}
where $m^{\rm{max}}$ and $m^{\rm{min}}$ are the upper and lower limits of the kinematically allowed range, respectively.
For virtual contributions, only the tail of the RBW function enters the Dalitz plot.
Given the large available phase-space in the {\ensuremath{\PB}}\xspace decay, it is possible to have nonresonant amplitudes (\mbox{\itshape i.e.}\xspace\ contributions that are not from any known resonance, including virtual states) that vary across the Dalitz plot.
A model that has been found to describe well nonresonant contributions in several {\ensuremath{\PB}}\xspace decay DP analyses is an exponential form factor (EFF)~\cite{Garmash:2004wa},
\begin{equation}
R(m) = e^{-\alpha m^2} \, ,
\label{eq:nonres}
\end{equation}
where $m$ is a two-body (in this case $D\pi$) invariant mass and $\alpha$ is a shape parameter that must be determined from the data.
Neglecting reconstruction effects, the DP probability density function would be
\begin{equation}
\label{eq:SigDPLike}
{\cal{P}}_{\rm phys}\left(m^2({\ensuremath{\D^+}}\xspace{\ensuremath{\pion^-}}\xspace), m^2({\ensuremath{\D^+}}\xspace{\ensuremath{\kaon^-}}\xspace)\right) =
\frac
{|{\cal A}\left(m^2({\ensuremath{\D^+}}\xspace{\ensuremath{\pion^-}}\xspace), m^2({\ensuremath{\D^+}}\xspace{\ensuremath{\kaon^-}}\xspace)\right)|^2}
{\int\!\!\int_{\rm DP}~{|{\cal A}|^2}~dm^2({\ensuremath{\D^+}}\xspace{\ensuremath{\pion^-}}\xspace)\,dm^2({\ensuremath{\D^+}}\xspace{\ensuremath{\kaon^-}}\xspace)} \, ,
\end{equation}
where the dependence of ${\cal A}$ on the DP position has been suppressed in the denominator for brevity.
The complex coefficients, given by $c_j$ in Eq.~(\ref{eqn:amp}), are the primary results of most Dalitz plot analyses.
However, these depend on the choice of normalisation, phase convention and amplitude formalism in each analysis.
Fit fractions and interference fit fractions are also reported as these provide a convention-independent method to allow meaningful comparisons of results.
The fit fraction is defined as the integral of the amplitude for a single component squared divided by that of the coherent matrix element squared for the complete Dalitz plot,
\begin{equation}
{\it FF}_j =
\frac
{\int\!\!\int_{\rm DP}\left|c_j F_j\left(m^2({\ensuremath{\D^+}}\xspace{\ensuremath{\pion^-}}\xspace), m^2({\ensuremath{\D^+}}\xspace{\ensuremath{\kaon^-}}\xspace)\right)\right|^2~dm^2({\ensuremath{\D^+}}\xspace{\ensuremath{\pion^-}}\xspace)\,dm^2({\ensuremath{\D^+}}\xspace{\ensuremath{\kaon^-}}\xspace)}
{\int\!\!\int_{\rm DP}\left|{\cal A}\right|^2~dm^2({\ensuremath{\D^+}}\xspace{\ensuremath{\pion^-}}\xspace)\,dm^2({\ensuremath{\D^+}}\xspace{\ensuremath{\kaon^-}}\xspace)} \, .
\label{eq:fitfraction}
\end{equation}
The fit fractions do not necessarily sum to unity due to the potential presence of net constructive or destructive interference, described by interference fit fractions defined for $i<j$ only by
\begin{equation}
{\it FF}_{ij} =
\frac
{\int\!\!\int_{\rm DP} 2 \, \ensuremath{\mathcal{R}e}\xspace\left[c_ic_j^* F_iF_j^*\right]~dm^2({\ensuremath{\D^+}}\xspace{\ensuremath{\pion^-}}\xspace)\,dm^2({\ensuremath{\D^+}}\xspace{\ensuremath{\kaon^-}}\xspace)}
{\int\!\!\int_{\rm DP}\left|{\cal A}\right|^2~dm^2({\ensuremath{\D^+}}\xspace{\ensuremath{\pion^-}}\xspace)\,dm^2({\ensuremath{\D^+}}\xspace{\ensuremath{\kaon^-}}\xspace)} \, ,
\label{eq:intfitfraction}
\end{equation}
where the dependence of $F_i^{(*)}$ and ${\cal A}$ on the DP position has been omitted.
\section{Dalitz plot fit}
\label{sec:dalitz}
The {\sc Laura++}~\cite{Laura++} package is used to perform the Dalitz plot fit, with the two trigger subsamples fitted simultaneously using the {\it J}{\sc fit} method~\cite{Ben-Haim:2014afa}.
The two subsamples have separate signal and background yields, efficiency maps and background SDP distributions, but all parameters of the signal model are common.
The likelihood function that is used is
\begin{equation}
{\cal L} =
\prod_i^{N_c}
\Bigg[
\sum_k N_k {\cal P}_k\left(m^2_i({\ensuremath{\D^+}}\xspace{\ensuremath{\pion^-}}\xspace),m^2_i({\ensuremath{\D^+}}\xspace{\ensuremath{\kaon^-}}\xspace)\right)
\Bigg] \,,
\end{equation}
where the index $i$ runs over $N_c$ candidates, while $k$ distinguishes the signal and background components with $N_k$ the yield in each component.
The probability density function for signal events, ${\cal P}_{\rm sig}$, is given by Eq.~(\ref{eq:SigDPLike}) where
the $|{\cal A}\left(m^2({\ensuremath{\D^+}}\xspace{\ensuremath{\pion^-}}\xspace), m^2({\ensuremath{\D^+}}\xspace{\ensuremath{\kaon^-}}\xspace)\right)|^2$ terms are multiplied by the efficiency function described in
Sec.~\ref{sec:efficiency}.
The mass resolution is approximately $2.4\ensuremath{\mathrm{\,Me\kern -0.1em V}}\xspace$, which is much lower than the width of the narrowest contribution to the Dalitz plot ($\sim 50 \ensuremath{\mathrm{\,Me\kern -0.1em V}}\xspace$); therefore, this has negligible effect on the likelihood and is not considered further.
The signal and background yields that enter the Dalitz plot fit are taken from the mass fit described in Sec.~\ref{sec:mass-fit}.
Only candidates in the signal region, defined as $\pm 2.5\sigma$ around the {\ensuremath{\PB}}\xspace signal peak, where $\sigma$ is the width of the peak, are used in the Dalitz plot fit.
Within this region, in the TOS subsample the result of the {\ensuremath{\PB}}\xspace candidate invariant mass fit corresponds to yields of $1060 \pm 35$, $37 \pm 6$, $26 \pm 8$ and $16 \pm 4$ in the signal, combinatorial background, $D^{(*)+}{\ensuremath{\pion^-}}\xspace\pim$ and ${\ensuremath{\D^+_\squark}}\xspace{\ensuremath{\kaon^-}}\xspace{\ensuremath{\pion^-}}\xspace$ components, respectively.
The equivalent yields in the TIS-only subsample are $849 \pm 30$, $39 \pm 6$, $5 \pm 5$ and $9 \pm 3$ candidates.
The contribution from ${\ensuremath{\D^{*+}}}\xspace{\ensuremath{\kaon^-}}\xspace{\ensuremath{\pion^-}}\xspace$ decays is negligible in the signal window.
The distributions of the candidates in the signal region over the DP and SDP are shown in Fig.~\ref{fig:canddp}.
\begin{figure}[!tb]
\centering
\includegraphics[scale=0.396]{Fig5a.pdf}
\includegraphics[scale=0.396]{Fig5b.pdf}
\caption{\small Distribution of ${\ensuremath{\Bub}}\xspace\ensuremath{\rightarrow}\xspace{\ensuremath{\D^+}}\xspace{\ensuremath{\kaon^-}}\xspace{\ensuremath{\pion^-}}\xspace$ candidates in the signal
region over (left) the DP and (right) the SDP.
Candidates from both TOS and TIS-only subsamples are included.}
\label{fig:canddp}
\end{figure}
The SDP distributions of the $D^{(*)+}{\ensuremath{\pion^-}}\xspace\pim$ and ${\ensuremath{\D^+_\squark}}\xspace{\ensuremath{\kaon^-}}\xspace{\ensuremath{\pion^-}}\xspace$ background sources are obtained from simulated samples using the same procedures as described for their invariant mass distributions in Sec.~\ref{sec:mass-fit}. The distribution of combinatorial background events is modelled by considering ${\ensuremath{\D^+}}\xspace{\ensuremath{\kaon^-}}\xspace{\ensuremath{\pion^-}}\xspace$ candidates in the sideband high-mass range $5500$--$5800\ensuremath{\mathrm{\,Me\kern -0.1em V}}\xspace$, with contributions from $D^{(*)+}{\ensuremath{\pion^-}}\xspace\pim$ in this region subtracted.
The dependence of the SDP distribution on {\ensuremath{\PB}}\xspace candidate mass was investigated and found to be negligible.
The SDP distributions of these backgrounds are shown in Fig.~\ref{fig:bkgs}.
These histograms are used to model the background contributions in the Dalitz plot fit.
\begin{figure}
\centering
\includegraphics[scale=0.37]{Fig6a.pdf}
\includegraphics[scale=0.37]{Fig6b.pdf}
\includegraphics[scale=0.37]{Fig6c.pdf}
\includegraphics[scale=0.37]{Fig6d.pdf}
\includegraphics[scale=0.37]{Fig6e.pdf}
\includegraphics[scale=0.37]{Fig6f.pdf}
\caption{\small Square Dalitz plot distributions used in the Dalitz plot fit for (top) combinatorial background, (middle) ${\ensuremath{\Bub}}\xspace \ensuremath{\rightarrow}\xspace D^{(*)+}{\ensuremath{\pion^-}}\xspace\pim$ decays and (bottom) ${\ensuremath{\Bub}}\xspace \ensuremath{\rightarrow}\xspace {\ensuremath{\D^+_\squark}}\xspace{\ensuremath{\kaon^-}}\xspace{\ensuremath{\pion^-}}\xspace$ decays.
Candidates the TOS (TIS-only) subsamples are shown in the left (right) column.}
\label{fig:bkgs}
\end{figure}
Using the results of the moments analysis of Sec.~\ref{sec:moments} as a guide, the nominal Dalitz plot fit model for ${\ensuremath{\Bub}}\xspace\ensuremath{\rightarrow}\xspace{\ensuremath{\D^+}}\xspace{\ensuremath{\kaon^-}}\xspace{\ensuremath{\pion^-}}\xspace$ decays is determined by considering several resonant, nonresonant and virtual amplitudes.
Those that do not contribute significantly and that do not aid the stability of the fit are removed.
Only natural spin-parity intermediate states are considered, as unnatural spin-parity states do not decay to two pseudoscalars.
The resulting signal model, referred to below as the nominal DP model, consists of the seven amplitudes shown in Table~\ref{tab:resonances}: three resonances, two virtual resonances and two nonresonant terms.
Parts of the model are known to be approximations.
In particular both S- and P-waves in the $D\pi$ system are modelled with overlapping broad structures.
The nominal model gives a better description of the data than any of the alternative models considered; alternative models are used to assign systematic uncertainties as discussed in Sec.~\ref{sec:systematics}.
\begin{table}[!tb]
\centering
\caption{\small
Signal contributions to the fit model, where parameters and uncertainties are taken from Ref.~\cite{PDG2014}.
States labelled with subscript $v$ are virtual contributions.
}
\label{tab:resonances}
\begin{tabular}{lcccc}
\hline
Resonance & Spin & DP axis & Model & Parameters \\
\hline \\ [-2.5ex]
$D^{*}_{0}(2400)^{0}$ &0& $m^{2}(D\pi)$ & RBW & $m = 2318 \pm 29\ensuremath{\mathrm{\,Me\kern -0.1em V}}\xspace$, $\Gamma = 267 \pm 40\ensuremath{\mathrm{\,Me\kern -0.1em V}}\xspace$ \\
$D^{*}_{2}(2460)^{0}$ &2& $m^{2}(D\pi)$ & RBW & \multirow{2}{*}{Determined from data (see Table~\ref{tab:masswidth})} \\
$D^{*}_{J}(2760)^{0}$ &1& $m^{2}(D\pi)$ & RBW \\
\hline
Nonresonant &0& $m^{2}(D\pi)$ & EFF & \multirow{2}{*}{Determined from data (see text)} \\
Nonresonant &1& $m^{2}(D\pi)$ & EFF \\
\hline \\ [-2.5ex]
$D^{*}_{v}(2007)^{0}$ &1& $m^{2}(D\pi)$ & RBW & $m = 2006.98 \pm 0.15 \ensuremath{\mathrm{\,Me\kern -0.1em V}}\xspace$, $\Gamma = 2.1 \ensuremath{\mathrm{\,Me\kern -0.1em V}}\xspace$ \\
$B^{*0}_{v}$ &1& $m^{2}(DK)$ & RBW & $m = 5325.2 \pm 0.4 \ensuremath{\mathrm{\,Me\kern -0.1em V}}\xspace$, $\Gamma = 0.0 \ensuremath{\mathrm{\,Me\kern -0.1em V}}\xspace$ \\
\hline
\end{tabular}
\end{table}
The free parameters in the fit are the $c_j$ terms introduced in Eq.~(\ref{eqn:amp}), with the real and imaginary parts of these complex coefficients determined for each amplitude in the fit model.
The $D^*_2(2460)^0$ component, as the reference amplitude, is the exception with real and imaginary parts fixed to 1 and 0, respectively.
Fit fractions and interference fit fractions are derived from these free parameters, as are the magnitudes and phases of the complex coefficients.
Statistical uncertainties for the derived parameters are calculated using large samples of simulated pseudoexperiments to ensure that non-trivial correlations are accounted for.
Several other parameters are also determined from the fit as described below.
In Dalitz plot fits it is common for the minimisation procedure to find local minima of the likelihood function.
To find the global minimum, the fit is performed many times using randomised starting values for the complex coefficients.
In addition to the global minimum of the likelihood, corresponding to the results reported below, several additional minima are found.
Two of these have negative log-likelihood (NLL) values close to that of global minimum.
The main differences between secondary minima and the global minimum are the interference patterns in the $D\pi$ S- and P-waves, as shown in App.~\ref{app:minima}.
The shape parameters, defined in Eq.~(\ref{eq:nonres}), for the nonresonant components are determined from the fit to data to be $0.36 \pm 0.03 \ensuremath{\mathrm{\,Ge\kern -0.1em V}}\xspace^{-2}$ and $0.36 \pm 0.04 \ensuremath{\mathrm{\,Ge\kern -0.1em V}}\xspace^{-2}$ for the S-wave and P-wave, respectively, where the uncertainties are statistical only.
The mass and width of the $D^{*}_{2}(2460)^{0}$ resonance are determined from the fit to improve the fit quality.
Since the mass and width of the $D^{*}_{J}(2760)^{0}$ state have not been precisely determined by previous experiments, these parameters are also allowed to vary in the fit.
The masses and widths of the $D^{*}_{2}(2460)^{0}$ and $D^{*}_{J}(2760)^{0}$ are reported in Table~\ref{tab:masswidth}.
The spin of the $D^{*}_{J}(2760)^{0}$ state has not been determined previously.
Fits are performed with all values up to 3, and spin~1 is found to be preferred with changes relative to the spin~0, 2 and 3 hypotheses of $2\Delta{\rm NLL} = 37.3, 49.5$ and $48.2$ units, respectively.
For comparison, the value of $2\Delta{\rm NLL}$ obtained from a fit with the $D^{*}_{1}(2760)^{0}$ state excluded is 75.0 units.
The alternative models discussed in Sec.~\ref{sec:systematics} give very similar values and therefore do not affect the conclusion that the $D^{*}_{J}(2760)^{0}$ state has spin~1.
\begin{table}[!tb]
\centering
\caption{\small Masses and widths determined in the fit to data, with statistical uncertainties only.}
\label{tab:masswidth}
\begin{tabular}{lcc}
\hline
Resonance & Mass $(\ensuremath{\mathrm{Me\kern -0.1em V}}\xspace)$ & Width $(\ensuremath{\mathrm{Me\kern -0.1em V}}\xspace)$ \\
\hline \\ [-2.5ex]
$D^{*}_{2}(2460)^{0}$ & $ 2464.0 \pm 1.4\phantom{1} $ & $ 43.8 \pm 2.9 $\\
$D^{*}_{J}(2760)^{0}$ & $ 2781 \pm 18 $ & $ 177 \pm 32 $\\
\hline
\end{tabular}
\end{table}
The values of the complex coefficients and fit fractions returned by the fit are shown in Table~\ref{tab:ffstat}.
Results for the interference fit fractions are given in App.~\ref{app:iffstat}. The total fit fraction exceeds unity
mostly due to interference between the $D^{*}_{0}(2400)^{0}$ and S-wave nonresonant contributions.
\begin{table}[!tb]
\centering
\caption{\small Complex coefficients and fit fractions determined from the
Dalitz plot fit. Uncertainties are statistical only.
}
\label{tab:ffstat}
\resizebox{\textwidth}{!}{
\begin{tabular}{lccccc}
\hline
& & \multicolumn{4}{c}{Isobar model coefficients} \\
Resonance & Fit fraction (\%) & Real part & Imaginary part & Magnitude & Phase \\
\hline \\ [-2.5ex]
$D^{*}_{0}(2400)^{0}$ & $\phantom{1}8.3 \pm 2.6$ & $-0.04 \pm 0.07$ & $-0.51 \pm 0.07$ & $0.51 \pm 0.09$ & $-1.65 \pm 0.16$\\
$D^{*}_{2}(2460)^{0}$ & $31.8 \pm 1.5$ & $\phantom{-}1.00$ & $\phantom{-}0.00$ & $1.00$ & $\phantom{-}0.00$\\
$D^{*}_{1}(2760)^{0}$ & $\phantom{1}4.9 \pm 1.2$ & $-0.32 \pm 0.06$ & $-0.23 \pm 0.07$ & $0.39 \pm 0.05$ & $-2.53 \pm 0.24$\\
\hline
S-wave nonresonant & $38.0 \pm 7.4$ & $\phantom{-}0.93 \pm 0.09$ & $-0.58 \pm 0.08$ & $1.09 \pm 0.09$ & $-0.56 \pm 0.09$\\
P-wave nonresonant & $23.8 \pm 5.6$ & $-0.43 \pm 0.09$ & $\phantom{-}0.75 \pm 0.09$ & $0.87 \pm 0.09$ & $\phantom{-}2.09 \pm 0.15$\\
\hline \\ [-2.5ex]
$D^{*}_{v}(2007)^{0}$ & $\phantom{1}7.6 \pm 2.3$ & $\phantom{-}0.16 \pm 0.08$ & $\phantom{-}0.46 \pm 0.09$ & $0.49 \pm 0.07$ & $\phantom{-}1.24 \pm 0.17$\\
$B^{*}_{v}$ & $\phantom{1}3.6 \pm 1.9$ & $-0.07 \pm 0.08$ & $\phantom{-}0.33 \pm 0.07$ & $0.34 \pm 0.06$ & $\phantom{-}1.78 \pm 0.23$\\
\hline
Total fit fraction & $118.1$ \\
\hline
\end{tabular}
}
\end{table}
The consistency of the fit model and the data is evaluated in several ways.
Numerous one-dimensional projections (including several shown below and those shown in Sec.~\ref{sec:moments}) show good agreement.
A two-dimensional $\chi^2$ value is determined by comparing the data and the fit model in $100$ equally populated bins across the SDP.
The pull, \mbox{\itshape i.e.}\xspace\ the difference between the data and fit model divided by the uncertainty, is shown with this SDP binning in Fig.~\ref{fig:sdppull}.
The $\chi^2$ value obtained is found to be within the bulk of the distribution expected from simulated pseudoexperiments.
Other unbinned fit quality tests~\cite{Williams:2010vh} also show acceptable agreement between the data and the fit model.
\begin{figure}[!tb]
\centering
\includegraphics[scale=0.5]{Fig7.pdf}
\caption{\small Differences between the data SDP distribution and the fit model across the SDP, in terms of the per-bin pull.}
\label{fig:sdppull}
\end{figure}
Figure~\ref{fig:fitproj} shows projections of the nominal fit model and the data onto $m(D\pi)$, $m(DK)$ and $m(K\pi)$.
Zooms are provided around the resonant structures on $m(D\pi)$ in Fig.~\ref{fig:fitprojzoom}.
Projections of the cosine of the helicity angle of the $D\pi$ system are shown in Fig.~\ref{fig:fitprojcos}.
Good agreement is seen between the data and the fit model.
\begin{figure}[!tb]
\centering
\includegraphics[scale=0.395]{Fig8a.pdf}
\includegraphics[scale=0.395]{Fig8b.pdf}
\includegraphics[scale=0.395]{Fig8c.pdf}
\includegraphics[scale=0.395]{Fig8d.pdf}
\includegraphics[scale=0.395]{Fig8e.pdf}
\includegraphics[scale=0.395]{Fig8f.pdf}
\includegraphics[scale=0.395]{Fig8g.pdf}
\caption{\small Projections of the data and amplitude fit onto (a)~$m(D\pi)$, (c)~$m(DK)$ and (e)~$m(K\pi)$, with the same
projections shown in (b), (d) and (f) with a logarithmic $y$-axis scale. Components are described in the legend.}
\label{fig:fitproj}
\end{figure}
\begin{figure}[!tb]
\centering
\includegraphics[scale=0.395]{Fig9a.pdf}
\includegraphics[scale=0.395]{Fig9b.pdf}
\includegraphics[scale=0.395]{Fig9c.pdf}
\caption{\small Projections of the data and amplitude fit onto $m(D\pi)$ in (a)~the threshold region, (b)~the $D^*_2(2460)^0$ region and (c)~the $D^*_1(2760)^0$ region. Components are as shown in Fig.~\ref{fig:fitproj}.}
\label{fig:fitprojzoom}
\end{figure}
\begin{figure}[!tb]
\centering
\includegraphics[scale=0.395]{Fig10a.pdf}
\includegraphics[scale=0.395]{Fig10b.pdf}
\includegraphics[scale=0.395]{Fig10c.pdf}
\caption{\small Projections of the data and amplitude fit onto the cosine of the helicity angle for the $D\pi$ system in (a)~the threshold region, (b)~the $D^*_2(2460)^0$ region and (c)~the $D^*_1(2760)^0$ region. Components are as shown in Fig.~\ref{fig:fitproj}.}
\label{fig:fitprojcos}
\end{figure}
\section{Systematic uncertainties}
\label{sec:systematics}
Sources of systematic uncertainty are divided into two categories: experimental and model uncertainties.
The sources of experimental systematic uncertainty are: the signal and background yields in the signal region; the SDP distributions of the background components; the efficiency variation across the SDP; possible fit bias.
The considered model uncertainties are: the fixed parameters in the amplitude model; the addition or removal of marginal amplitudes; the choice of models for the nonresonant contributions.
The systematic uncertainties from each source are combined in quadrature.
The signal and background yields in the signal region are determined from the fit to the $B$ candidate invariant mass distribution, as described in Sec.~\ref{sec:mass-fit}.
The uncertainty on each yield (including systematic uncertainty evaluated as in Sec.~\ref{sec:BF-syst}) is calculated, and the yields varied accordingly in the DP fit.
The deviations from the nominal DP fit result are assigned as systematic uncertainties.
The effect of imperfect knowledge of the background distributions over the SDP is tested by varying the histograms used to model the shapes within their statistical uncertainties.
For $D^{(*)+}{\ensuremath{\pion^-}}\xspace\pim$ decays the ratio of the ${\ensuremath{\D^{*+}}}\xspace$ and ${\ensuremath{\D^+}}\xspace$ contributions is varied.
Where applicable, the reweighting of the SDP distribution of the simulated samples is removed.
The uncertainty related to the knowledge of the variation of efficiency across the SDP is determined by varying the efficiency histograms before the spline fit is performed.
The central bin in each cell of $3 \times 3$ bins is varied by its statistical uncertainty and the surrounding bins in the cell are varied by interpolation.
This procedure accounts for possible correlations between the bins, since a systematic effect on a given bin is likely also to affect neighbouring bins.
The effects on the DP fit results are assigned as systematic uncertainties.
An additional systematic uncertainty is assigned by varying the binning scheme of the control sample used to determine the PID efficiencies.
Systematic uncertainties related to possible intrinsic fit bias are investigated using an ensemble of pseudoexperiments.
Differences between the input and fitted values from the ensemble for the fit parameters are found to be small.
Systematic uncertainties are assigned as the sum in quadrature of the difference between the input and output values and the uncertainty on the mean of the output value determined from a fit to the ensemble.
Systematic uncertainties due to fixed parameters in the fit model are determined by varying the parameters within their
uncertainties and repeating the fit. The fixed parameters considered are the mass and width of the $D^*_0(2400)^0$ resonance
and the \mbox{Blatt--Weisskopf} barrier radius, $r_{\rm BW}$. The mass and width are varied by the uncertainties shown in
Table~\ref{tab:resonances} and the barrier radius is varied between 3 and 5\,$\ensuremath{\mathrm{\,Ge\kern -0.1em V}}\xspace^{-1}$~\cite{LHCb-PAPER-2014-036}.
For each fit parameter, the difference compared to the nominal fit model is assigned as a systematic uncertainty for each source.
The marginal $B^{*0}_{v}$ component is removed from the model and the changes in the other parameters are assigned as the systematic uncertainties.
Dalitz plot analysis of ${\ensuremath{\Bbar{}^0_\squark}}\xspace \ensuremath{\rightarrow}\xspace {\ensuremath{\D^0}}\xspace {\ensuremath{\kaon^+}}\xspace{\ensuremath{\pion^-}}\xspace$ revealed that a structure at $m({\ensuremath{\D^0}}\xspace{\ensuremath{\kaon^+}}\xspace) \sim 2.86 \ensuremath{\mathrm{\,Ge\kern -0.1em V}}\xspace$ has both spin~1 and spin~3 components~\cite{LHCb-PAPER-2014-035,LHCb-PAPER-2014-036}.
Although there is no evidence for a spin~3 resonance in this analysis, the excess at $m({\ensuremath{\D^+}}\xspace{\ensuremath{\pion^-}}\xspace) \sim 2.76 \ensuremath{\mathrm{\,Ge\kern -0.1em V}}\xspace$ could have a similar composition.
A putative $D^*_3(2760)$ resonance is added to the fit model, and the effect on the other parameters is used to assign systematic uncertainties.
The EFF lineshapes used to model the nonresonant S- and P-wave contributions are replaced by a power-law model and the change in the fit parameters used as a systematic uncertainty.
The dependence of the results on the effective pole mass description of Eq.~(\ref{eqn:effmass}) that is used for the virtual resonance contributions is found by using a fixed width in Eq.~(\ref{eq:RelBWEqn}), removing the dependency on $m^{\rm eff}_0$.
The total experimental and model systematic uncertainties for fit fractions and complex coefficients are summarised in Tables~\ref{tab:exptsyst} and~\ref{tab:modsyst}, respectively.
The contributions for the fit fractions, masses and widths are broken down in Tables~\ref{tab:exptsystbreak} and~\ref{tab:modsystbreak}.
Similar tables summarising the systematic uncertainties on the interference fit fractions are given in App.~\ref{app:iffstat}.
The largest source of experimental systematic uncertainty on the fit fractions is due to the efficiency variation.
For the model uncertainty on the fit fractions, the addition and removal of marginal components and variation of fixed parameters dominate.
In general, the model uncertainties are larger than the experimental systematic uncertainties for the fit fractions and the masses and widths.
\begin{table}[!tb]
\centering
\caption{\small Experimental systematic uncertainties on the fit fractions and complex amplitudes.}
\label{tab:exptsyst}
\resizebox{\textwidth}{!}{
\begin{tabular}{lccccc}
\hline
& & \multicolumn{4}{c}{Isobar model coefficients} \\
Resonance & Fit fraction (\%) & Real part & Imaginary part & Magnitude & Phase \\
\hline \\ [-2.5ex]
$D^{*}_{0}(2400)^{0}$ & 0.6 & 0.03 & 0.02 & 0.02 & 0.06\\
$D^{*}_{2}(2460)^{0}$ & 0.9 & -- & -- & -- & -- \\
$D^{*}_{1}(2760)^{0}$ & 0.4 & 0.03 & 0.03 & 0.01 & 0.08\\
\hline
S-wave nonresonant & 1.5 & 0.03 & 0.03 & 0.02 & 0.04\\
P-wave nonresonant & 2.1 & 0.03 & 0.05 & 0.03 & 0.05\\
\hline \\ [-2.5ex]
$D^{*}_{v}(2007)^{0}$ & 1.3 & 0.03 & 0.04 & 0.04 & 0.07\\
$B^{*}_{v}$ & 0.9 & 0.22 & 0.02 & 0.03 & 0.11\\
\hline
\end{tabular}
}
\end{table}
\begin{table}[!tb]
\centering
\caption{\small Model uncertainties on the fit fractions and complex amplitudes.}
\label{tab:modsyst}
\resizebox{\textwidth}{!}{
\begin{tabular}{lccccc}
\hline
& & \multicolumn{4}{c}{Isobar model coefficients} \\
Resonance & Fit fraction (\%) & Real part & Imaginary part & Magnitude & Phase \\
\hline \\ [-2.5ex]
$D^{*}_{0}(2400)^{0}$ & 1.9 & 0.28 & 0.13 & 0.15 & 0.51\\
$D^{*}_{2}(2460)^{0}$ & 1.4 & -- & -- & -- & -- \\
$D^{*}_{1}(2760)^{0}$ & 0.9 & 0.03 & 0.03 & 0.03 & 0.08\\
\hline
S-wave nonresonant & 10.8 & 0.17 & 0.15 & 0.20 & 0.11\\
P-wave nonresonant & 3.7 & 0.34 & 0.68 & 0.12 & 0.95\\
\hline \\ [-2.5ex]
$D^{*}_{v}(2007)^{0}$ & 1.5 & 0.56 & 0.77 & 0.05 & 0.60\\
$B^{*}_{v}$ & 1.6 & 0.09 & 0.08 & 0.07 & 0.27\\
\hline
\end{tabular}
}
\end{table}
\begin{table}[!tb]
\centering
\caption{\small Breakdown of experimental systematic uncertainties on the fit fractions (\%) and masses $(\ensuremath{\mathrm{Me\kern -0.1em V}}\xspace)$ and widths $(\ensuremath{\mathrm{Me\kern -0.1em V}}\xspace)$.}
\label{tab:exptsystbreak}
\begin{tabular}{lrccccc}
\hline
& Nominal & S/B frac. & Eff. & Bkg. & Fit bias & Total \\
\hline \\ [-2.5ex]
$D^{*}_{0}(2400)^{0}$ & $8.3 \pm 2.6$ & 0.2 & 0.5 & 0.1 & 0.3 & 0.6 \\
$D^{*}_{2}(2460)^{0}$ & $31.8 \pm 1.5$ & 0.2 & 0.8 & 0.0 & 0.2 & 0.9 \\
$D^{*}_{1}(2760)^{0}$ & $4.9 \pm 1.2$ & 0.2 & 0.2 & 0.1 & 0.2 & 0.3 \\
\hline
S-wave nonresonant & $38.0 \pm 7.4$ & 0.7 & 0.5 & 0.4 & 1.2 & 1.5 \\
P-wave nonresonant & $23.8 \pm 5.6$ & 1.0 & 1.6 & 0.7 & 0.5 & 2.1 \\
\hline \\ [-2.5ex]
$D^{*}_{v}(2007)^{0}$ & $7.6 \pm 2.3$ & 0.7 & 1.0 & 0.3 & 0.3 & 1.3 \\
$B^{*}_{v}$ & $3.6 \pm 1.9$ & 0.3 & 0.3 & 0.2 & 0.8 & 0.9 \\
\hline
\hline \\ [-2.5ex]
$m\left( D^{*}_{2}(2460)^{0}\right)$ & $2464.0 \pm 1.4$ & 0.1 & 0.1 & 0.0 & 0.2 & 0.2 \\
$\Gamma\left( D^{*}_{2}(2460)^{0}\right)$ & $43.8 \pm 2.9$ & 0.3 & 0.3 & 0.0 & 0.4 & 0.6 \\
\hline \\ [-2.5ex]
$m\left( D^{*}_{1}(2760)^{0}\right)$ & $2781 \pm 18\phantom{.}$ &1 & 4 & 0 & 2 & 6 \\
$\Gamma\left( D^{*}_{1}(2760)^{0}\right)$ & $177 \pm 32\phantom{.}$ & 3 & 1 & 2 & 5 & 7 \\
\hline
\end{tabular}
\end{table}
\begin{table}[!tb]
\centering
\caption{\small Breakdown of model uncertainties on the fit fractions (\%) and masses $(\ensuremath{\mathrm{Me\kern -0.1em V}}\xspace)$ and widths $(\ensuremath{\mathrm{Me\kern -0.1em V}}\xspace)$.}
\label{tab:modsystbreak}
\begin{tabular}{lrcccc}
\hline
& Nominal & Add/rem & Alt. models & Fixed params & Total \\
\hline \\ [-2.5ex]
$D^{*}_{0}(2400)^{0}$ & $8.3 \pm 2.6$ & 2.0 & 0.1 & 0.2 & 2.0 \\
$D^{*}_{2}(2460)^{0}$ & $31.8 \pm 1.5$ & 1.3 & 0.2 & 0.4 & 1.4 \\
$D^{*}_{1}(2760)^{0}$ & $4.9 \pm 1.2$ & 0.8 & 0.1 & 0.3 & 0.9 \\
\hline
S-wave nonresonant & $38.0 \pm 7.4$ & 4.8 & 4.5 & 5.4 & 10.8 \\
P-wave nonresonant & $23.8 \pm 5.6$ & 2.6 & 2.1 & 3.0 & 3.7 \\
\hline \\ [-2.5ex]
$D^{*}_{v}(2007)^{0}$ & $7.6 \pm 2.3$ & 0.6 & 0.1 & 1.4 & 1.5 \\
$B^{*}_{v}$ & $3.6 \pm 1.9$ & 0.7 & 1.0 & 1.1 & 1.6 \\
\hline
\hline \\ [-2.5ex]
$m\left( D^{*}_{2}(2460)^{0}\right)$ & $2464.0 \pm 1.4$ & 0.5 & 0.1 & 0.1 & 0.5 \\
$\Gamma\left( D^{*}_{2}(2460)^{0}\right)$ & $43.8 \pm 2.9$ & 0.8 & 1.4 & 0.6 & 1.7 \\
\hline \\ [-2.5ex]
$m\left( D^{*}_{1}(2760)^{0}\right)$ & $2781 \pm 18\phantom{.}$ & 6 & 6 & 1 & 11 \\
$\Gamma\left( D^{*}_{1}(2760)^{0}\right)$ & $177 \pm 32\phantom{.}$ & 16 & 9 & 1 & 20 \\
\hline
\end{tabular}
\end{table}
Several cross-checks are performed to confirm the stability of the results. The data sample is divided into two parts depending
on the charge of the {\ensuremath{\PB}}\xspace candidate, the polarity of the magnet and the year of data taking.
Selection effects are also checked by varying the requirement on the neural network output variable and the PID criteria applied to the bachelor kaon.
A fit is performed for each of the subsamples individually and each is seen to be consistent with the default fit results, although in some cases one of the secondary minima described in App.~\ref{app:minima} becomes the preferred solution.
To cross-check the amplitude model, the fit is repeated many times with an extra resonance with fixed mass, width and spin included in the model.
All possible mass and width values, and spin up to 3, were considered.
None of the additional resonances are found to contribute significantly.
\section{Results and summary}
\label{sec:results}
The results for the complex coefficients are reported in Tables~\ref{tab:cf-results2} and~\ref{tab:cf-results} in terms of real and imaginary parts and of magnitudes and phases, respectively.
The results for the fit fractions are given in Table~\ref{tab:cfitfrac-results} and the results for the interference fit fractions are given in App.~\ref{app:iffstat}.
The fit fractions for resonant contributions are converted into quasi-two-body product branching fractions by multiplying by $\mathcal{B}({\ensuremath{\Bub}}\xspace\ensuremath{\rightarrow}\xspace{\ensuremath{\D^+}}\xspace{\ensuremath{\kaon^-}}\xspace{\ensuremath{\pion^-}}\xspace) = (7.31 \pm 0.19 \pm 0.22 \pm 0.39) \times 10^{-5}$, as determined in Sec.~\ref{sec:BF-results}.
These product branching fractions are shown in Table~\ref{tab:BFresults}; they cannot be converted into absolute branching fractions because the branching fractions for the resonance decays to ${\ensuremath{\D^+}}\xspace{\ensuremath{\pion^-}}\xspace$ are unknown.
\begin{table}[!tb]
\centering
\caption{\small Results for the complex amplitudes and their uncertainties. The three quoted errors are statistical,
experimental systematic and model uncertainties, respectively.}
\label{tab:cf-results2}
\begin{tabular}{lcc}
\hline
& \multicolumn{2}{c}{Isobar model coefficients} \\
Resonance & Real part & Imaginary part \\
\hline \\ [-2.5ex]
$D^{*}_{0}(2400)^{0}$ & $-0.04 \pm 0.07 \pm 0.03 \pm 0.28$ & $-0.51 \pm 0.07 \pm 0.02 \pm 0.13$ \\
$D^{*}_{2}(2460)^{0}$ & $1.00$ & $0.00$ \\
$D^{*}_{1}(2760)^{0}$ & $-0.32 \pm 0.06 \pm 0.03 \pm 0.03$ & $-0.23 \pm 0.07 \pm 0.03 \pm 0.03$ \\
\hline
S-wave nonresonant & $\phantom{-}0.93 \pm 0.09 \pm 0.03 \pm 0.17$ & $-0.58 \pm 0.08 \pm 0.03 \pm 0.15$ \\
P-wave nonresonant & $-0.43 \pm 0.09 \pm 0.03 \pm 0.34$ & $\phantom{-}0.75 \pm 0.09 \pm 0.05 \pm 0.68$ \\
\hline \\ [-2.5ex]
$D^{*}_{v}(2007)^{0}$ & $\phantom{-}0.16 \pm 0.08 \pm 0.03 \pm 0.56$ & $\phantom{-}0.46 \pm 0.09 \pm 0.04 \pm 0.77$ \\
$B^*_v$ & $-0.07 \pm 0.08 \pm 0.22 \pm 0.09$ & $\phantom{-}0.33 \pm 0.07 \pm 0.02 \pm 0.08$ \\
\hline
\end{tabular}
\end{table}
\begin{table}[!tb]
\centering
\caption{\small Results for the complex amplitudes and their uncertainties. The three quoted errors are statistical,
experimental systematic and model uncertainties, respectively.}
\label{tab:cf-results}
\begin{tabular}{lcc}
\hline
& \multicolumn{2}{c}{Isobar model coefficients} \\
Resonance & Magnitude & Phase \\
\hline \\ [-2.5ex]
$D^{*}_{0}(2400)^{0}$ & $0.51 \pm 0.09 \pm 0.02 \pm 0.15$ & $-1.65 \pm 0.16 \pm 0.06 \pm 0.50$ \\
$D^{*}_{2}(2460)^{0}$ & $1.00$ & $0.00$ \\
$D^{*}_{1}(2760)^{0}$ & $0.39 \pm 0.05 \pm 0.01 \pm 0.03$ & $-2.53 \pm 0.24 \pm 0.08 \pm 0.08$ \\
\hline
S-wave nonresonant & $1.09 \pm 0.09 \pm 0.02 \pm 0.20$ & $-0.56 \pm 0.09 \pm 0.04 \pm 0.11$ \\
P-wave nonresonant & $0.87 \pm 0.09 \pm 0.03 \pm 0.11$ & $\phantom{-}2.09 \pm 0.15 \pm 0.05 \pm 0.95$ \\
\hline \\ [-2.5ex]
$D^{*}_{v}(2007)^{0}$ & $0.49 \pm 0.07 \pm 0.04 \pm 0.05$ & $\phantom{-}1.24 \pm 0.17 \pm 0.07 \pm 0.60$ \\
$B^*_v$ & $0.34 \pm 0.06 \pm 0.03 \pm 0.07$ & $\phantom{-}1.78 \pm 0.23 \pm 0.11 \pm 0.27$ \\
\hline
\end{tabular}
\end{table}
\begin{table}[!tb]
\centering
\caption{\small Results for the fit fractions and their uncertainties (\%). The three quoted errors are statistical, experimental
systematic and model uncertainties, respectively.}
\label{tab:cfitfrac-results}
\begin{tabular}{lc}
\hline
Resonance & Fit fraction \\
\hline \\ [-2.5ex]
$D^{*}_{0}(2400)^{0}$ & $\phantom{2}8.3 \pm 2.6 \pm 0.6 \pm \phantom{2}1.9$ \\
$D^{*}_{2}(2460)^{0}$ & $31.8 \pm 1.5 \pm 0.9 \pm \phantom{2}1.4$ \\
$D^{*}_{1}(2760)^{0}$ & $\phantom{2}4.9 \pm 1.2 \pm 0.3 \pm \phantom{2}0.9$ \\
\hline
S-wave nonresonant & $38.0 \pm 7.4 \pm 1.5 \pm 10.8$ \\
P-wave nonresonant & $23.8 \pm 5.6 \pm 2.1 \pm \phantom{2}3.7$ \\
\hline \\ [-2.5ex]
$D^{*}_{v}(2007)^{0}$ & $\phantom{2}7.6 \pm 2.3 \pm 1.3 \pm \phantom{2}1.5$ \\
$B^*_v$ & $\phantom{2}3.6 \pm 1.9 \pm 0.9 \pm \phantom{2}1.6$ \\
\hline
\end{tabular}
\end{table}
\begin{table}[!tb]
\centering
\caption{\small Results for the product branching fractions ${\cal B}({\ensuremath{\Bub}}\xspace \ensuremath{\rightarrow}\xspace R{\ensuremath{\kaon^-}}\xspace) \times {\cal B}(R \ensuremath{\rightarrow}\xspace {\ensuremath{\D^+}}\xspace{\ensuremath{\pion^-}}\xspace)$ ($10^{-6}$). The four quoted errors are statistical, experimental systematic, model and inclusive branching fraction uncertainties, respectively.}
\label{tab:BFresults}
\begin{tabular}{lc}
\hline
Resonance & Branching fraction \\
\hline \\ [-2.5ex]
$D^{*}_{0}(2400)^{0}$ & $\phantom{2}6.1 \pm 1.9 \pm 0.5 \pm 1.4 \pm 0.4$ \\
$D^{*}_{2}(2460)^{0}$ & $23.2 \pm 1.1 \pm 0.6 \pm 1.0 \pm 1.6$ \\
$D^{*}_{1}(2760)^{0}$ & $\phantom{2}3.6 \pm 0.9 \pm 0.3 \pm 0.7 \pm 0.2$ \\
\hline
S-wave nonresonant & $27.8 \pm 5.4 \pm 1.1 \pm 7.9 \pm 1.9$ \\
P-wave nonresonant & $17.4 \pm 4.1 \pm 1.5 \pm 2.7 \pm 1.2$ \\
\hline \\ [-2.5ex]
$D^{*}_{v}(2007)^{0}$ & $\phantom{2}5.6 \pm 1.7 \pm 1.0 \pm 1.1 \pm 0.4$ \\
$B^*_v$ & $\phantom{2}2.6 \pm 1.4 \pm 0.6 \pm 1.2 \pm 0.2$ \\
\hline
\end{tabular}
\end{table}
The masses and widths of the $D^{*}_{2}(2460)^{0}$ and $D^{*}_{1}(2760)^{0}$ are determined to be
\begin{eqnarray*}
m(D^{*}_{2}(2460)^{0}) & = & (2464.0 \pm 1.4 \pm 0.5 \pm 0.2) \ensuremath{\mathrm{\,Me\kern -0.1em V}}\xspace \, ,\\
\Gamma(D^{*}_{2}(2460)^{0}) & = & \phantom{24}(43.8 \pm 2.9 \pm 1.7 \pm 0.6) \ensuremath{\mathrm{\,Me\kern -0.1em V}}\xspace \, ,\\
m(D^{*}_{1}(2760)^{0}) & = & \phantom{.0}(2781 \pm \phantom{.}18 \pm \phantom{.}11 \pm \phantom{.0}6) \ensuremath{\mathrm{\,Me\kern -0.1em V}}\xspace \, ,\\
\Gamma(D^{*}_{1}(2760)^{0}) & = & \phantom{.02}(177 \pm \phantom{.}32 \pm \phantom{.}20 \pm \phantom{.0}7) \ensuremath{\mathrm{\,Me\kern -0.1em V}}\xspace \, ,
\end{eqnarray*}
where the three quoted errors are statistical, experimental systematic and model uncertainties, respectively.
The results for the $D^{*}_{2}(2460)^{0}$ are within $2\,\sigma$ of the world average values~\cite{PDG2014}.
The mass of the $D^{*}_{1}(2760)^{0}$ resonance is similarly consistent with previous measurements.
The measured width of this state is larger than previous measurements by 2 to 3 times the uncertainties.
Future studies based on much larger data samples will be required to better understand these states.
The measurement of $\mathcal{B}({\ensuremath{\Bub}}\xspace\ensuremath{\rightarrow}\xspace{\ensuremath{\D^+}}\xspace{\ensuremath{\kaon^-}}\xspace{\ensuremath{\pion^-}}\xspace)$ corresponds to the first observation of this decay mode.
Therefore, the resonant contributions to the decay are also first observations.
The significance of the ${\ensuremath{\Bub}}\xspace \ensuremath{\rightarrow}\xspace D^{*}_{1}(2760)^{0}{\ensuremath{\kaon^-}}\xspace$ observation is investigated by removing the corresponding resonance from the DP model.
A fit without the $D^{*}_{1}(2760)^{0}$ component increases the value of $2\Delta{\rm NLL}$ by $75.0$ units, corresponding to a high statistical significance.
Only the systematic effects due to uncertainties in the DP model could in principle significantly change the conclusion regarding the need for this resonance.
However, in alternative DP models where a $D\pi$ resonance with spin~3 is added and where the $B^*_v$ contribution is removed, the shift in $2\Delta{\rm NLL}$ remains above 50 units.
The alternative models also do not significantly impact the level at which the $D^{*}_{1}(2760)^{0}$ state is preferred to be spin~1.
Therefore, these results represent the first observation of the ${\ensuremath{\Bub}}\xspace \ensuremath{\rightarrow}\xspace D^{*}_{1}(2760)^{0}{\ensuremath{\kaon^-}}\xspace$ and the measurement of the spin of the $D^{*}_{1}(2760)^{0}$ resonance.
In summary, the ${\ensuremath{\Bub}}\xspace\ensuremath{\rightarrow}\xspace{\ensuremath{\D^+}}\xspace{\ensuremath{\kaon^-}}\xspace{\ensuremath{\pion^-}}\xspace$ decay has been observed in a data sample corresponding to $3.0\ensuremath{\mbox{\,fb}^{-1}}\xspace$ of $pp$ collision data recorded by the LHCb experiment.
An amplitude analysis of its Dalitz plot distribution has been performed, in which a model containing resonant contributions from the $D^{*}_{0}(2400)^{0}$, $D^{*}_{2}(2460)^{0}$ and $D^{*}_{1}(2760)^{0}$ states in addition to both S-wave and P-wave nonresonant amplitudes and components due to virtual $D^{*}_{v}(2007)^{0}$ and $B^{*0}_{v}$ resonances was found to give a good description of the data.
The ${\ensuremath{\Bub}}\xspace \ensuremath{\rightarrow}\xspace D^{*}_{2}(2460)^{0}{\ensuremath{\kaon^-}}\xspace$ decay may in future be used to determine the angle $\gamma$ of the CKM unitarity triangle.
The results provide insight into the spectroscopy of charm mesons, and demonstrate that further progress may be obtained with Dalitz plot analyses of larger data samples.
\section*{Acknowledgements}
\noindent We express our gratitude to our colleagues in the CERN
accelerator departments for the excellent performance of the LHC. We
thank the technical and administrative staff at the LHCb
institutes. We acknowledge support from CERN and from the national
agencies: CAPES, CNPq, FAPERJ and FINEP (Brazil); NSFC (China);
CNRS/IN2P3 (France); BMBF, DFG, HGF and MPG (Germany); INFN (Italy);
FOM and NWO (The Netherlands); MNiSW and NCN (Poland); MEN/IFA (Romania);
MinES and FANO (Russia); MinECo (Spain); SNSF and SER (Switzerland);
NASU (Ukraine); STFC (United Kingdom); NSF (USA).
The Tier1 computing centres are supported by IN2P3 (France), KIT and BMBF
(Germany), INFN (Italy), NWO and SURF (The Netherlands), PIC (Spain), GridPP
(United Kingdom).
We are indebted to the communities behind the multiple open
source software packages on which we depend. We are also thankful for the
computing resources and the access to software R\&D tools provided by Yandex LLC (Russia).
Individual groups or members have received support from
EPLANET, Marie Sk\l{}odowska-Curie Actions and ERC (European Union),
Conseil g\'{e}n\'{e}ral de Haute-Savoie, Labex ENIGMASS and OCEVU,
R\'{e}gion Auvergne (France), RFBR (Russia), XuntaGal and GENCAT (Spain), Royal Society and Royal
Commission for the Exhibition of 1851 (United Kingdom).
\addcontentsline{toc}{section}{References}
\setboolean{inbibliography}{true}
\ifx\mcitethebibliography\mciteundefinedmacro
\PackageError{LHCb.bst}{mciteplus.sty has not been loaded}
{This bibstyle requires the use of the mciteplus package.}\fi
\providecommand{\href}[2]{#2}
\begin{mcitethebibliography}{10}
\mciteSetBstSublistMode{n}
\mciteSetBstMaxWidthForm{subitem}{\alph{mcitesubitemcount})}
\mciteSetBstSublistLabelBeginEnd{\mcitemaxwidthsubitemform\space}
{\relax}{\relax}
\bibitem{Abe:2003zm}
Belle collaboration, K.~Abe {\em et~al.},
\ifthenelse{\boolean{articletitles}}{\emph{{Study of ${\ensuremath{\Bub}}\xspace \ensuremath{\rightarrow}\xspace D^{**0} {\ensuremath{\pion^-}}\xspace \
(D^{**0} \ensuremath{\rightarrow}\xspace D^{(*)+} {\ensuremath{\pion^-}}\xspace)$ decays}},
}{}\href{http://dx.doi.org/10.1103/PhysRevD.69.112002}{Phys.\ Rev.\
\textbf{D69} (2004) 112002}, \href{http://arxiv.org/abs/hep-ex/0307021}{{\tt
arXiv:hep-ex/0307021}}\relax
\mciteBstWouldAddEndPuncttrue
\mciteSetBstMidEndSepPunct{\mcitedefaultmidpunct}
{\mcitedefaultendpunct}{\mcitedefaultseppunct}\relax
\EndOfBibitem
\bibitem{Aubert:2009wg}
\mbox{BaBar}\xspace collaboration, B.~Aubert {\em et~al.},
\ifthenelse{\boolean{articletitles}}{\emph{{Dalitz plot analysis of ${\ensuremath{\Bub}}\xspace \ensuremath{\rightarrow}\xspace
{\ensuremath{\D^+}}\xspace{\ensuremath{\pion^-}}\xspace\pim$}}, }{}\href{http://dx.doi.org/10.1103/PhysRevD.79.112004}{Phys.\
Rev.\ \textbf{D79} (2009) 112004},
\href{http://arxiv.org/abs/0901.1291}{{\tt arXiv:0901.1291}}\relax
\mciteBstWouldAddEndPuncttrue
\mciteSetBstMidEndSepPunct{\mcitedefaultmidpunct}
{\mcitedefaultendpunct}{\mcitedefaultseppunct}\relax
\EndOfBibitem
\bibitem{Kuzmin:2006mw}
Belle collaboration, A.~Kuzmin {\em et~al.},
\ifthenelse{\boolean{articletitles}}{\emph{{Study of $\bar{B}^0 \ensuremath{\rightarrow}\xspace D^0 \pi^+
\pi^-$ decays}},
}{}\href{http://dx.doi.org/10.1103/PhysRevD.76.012006}{Phys.\ Rev.\
\textbf{D76} (2007) 012006}, \href{http://arxiv.org/abs/hep-ex/0611054}{{\tt
arXiv:hep-ex/0611054}}\relax
\mciteBstWouldAddEndPuncttrue
\mciteSetBstMidEndSepPunct{\mcitedefaultmidpunct}
{\mcitedefaultendpunct}{\mcitedefaultseppunct}\relax
\EndOfBibitem
\bibitem{LHCb-PAPER-2014-035}
LHCb collaboration, R.~Aaij {\em et~al.},
\ifthenelse{\boolean{articletitles}}{\emph{{Observation of overlapping
spin-$1$ and spin-$3$ $\bar{D}^0 K^-$ resonances at mass $2.86$~GeV/$c^2$}},
}{}\href{http://dx.doi.org/10.1103/PhysRevLett.113.162001}{Phys.\ Rev.\
Lett.\ \textbf{113} (2014) 162001},
\href{http://arxiv.org/abs/1407.7574}{{\tt arXiv:1407.7574}}\relax
\mciteBstWouldAddEndPuncttrue
\mciteSetBstMidEndSepPunct{\mcitedefaultmidpunct}
{\mcitedefaultendpunct}{\mcitedefaultseppunct}\relax
\EndOfBibitem
\bibitem{LHCb-PAPER-2014-036}
LHCb collaboration, R.~Aaij {\em et~al.},
\ifthenelse{\boolean{articletitles}}{\emph{{Dalitz plot analysis of
$B^0_s\ensuremath{\rightarrow}\xspace\bar{D}^0K^-\pi^+$ decays}},
}{}\href{http://dx.doi.org/10.1103/PhysRevD.90.072003}{Phys.\ Rev.\
\textbf{D90} (2014) 072003}, \href{http://arxiv.org/abs/1407.7712}{{\tt
arXiv:1407.7712}}\relax
\mciteBstWouldAddEndPuncttrue
\mciteSetBstMidEndSepPunct{\mcitedefaultmidpunct}
{\mcitedefaultendpunct}{\mcitedefaultseppunct}\relax
\EndOfBibitem
\bibitem{Lees:2014abp}
\mbox{BaBar}\xspace collaboration, J.~P. Lees {\em et~al.},
\ifthenelse{\boolean{articletitles}}{\emph{{Dalitz plot analyses of $B^0 \ensuremath{\rightarrow}\xspace
D^- D^0 K^+$ and $B^+ \ensuremath{\rightarrow}\xspace \overline{D}^0 D^0 K^+$ decays}},
}{}\href{http://dx.doi.org/10.1103/PhysRevD.91.052002}{Phys.\ Rev.\
\textbf{D91} (2015) 052002}, \href{http://arxiv.org/abs/1412.6751}{{\tt
arXiv:1412.6751}}\relax
\mciteBstWouldAddEndPuncttrue
\mciteSetBstMidEndSepPunct{\mcitedefaultmidpunct}
{\mcitedefaultendpunct}{\mcitedefaultseppunct}\relax
\EndOfBibitem
\bibitem{delAmoSanchez:2010vq}
\mbox{BaBar}\xspace collaboration, P.~del Amo~Sanchez {\em et~al.},
\ifthenelse{\boolean{articletitles}}{\emph{{Observation of new resonances
decaying to $D\pi$ and $D^*\pi$ in inclusive $e^+e^-$ collisions near
$\sqrt{s}=$10.58 GeV}},
}{}\href{http://dx.doi.org/10.1103/PhysRevD.82.111101}{Phys.\ Rev.\
\textbf{D82} (2010) 111101}, \href{http://arxiv.org/abs/1009.2076}{{\tt
arXiv:1009.2076}}\relax
\mciteBstWouldAddEndPuncttrue
\mciteSetBstMidEndSepPunct{\mcitedefaultmidpunct}
{\mcitedefaultendpunct}{\mcitedefaultseppunct}\relax
\EndOfBibitem
\bibitem{LHCb-PAPER-2013-026}
LHCb collaboration, R.~Aaij {\em et~al.},
\ifthenelse{\boolean{articletitles}}{\emph{{Study of $D_J$ meson decays to
$D^+\pi^-$, $D^0\pi^+$ and $D^{*+}\pi^-$ final states in $pp$ collisions}},
}{}\href{http://dx.doi.org/10.1007/JHEP09(2013)145}{JHEP \textbf{09} (2013)
145}, \href{http://arxiv.org/abs/1307.4556}{{\tt arXiv:1307.4556}}\relax
\mciteBstWouldAddEndPuncttrue
\mciteSetBstMidEndSepPunct{\mcitedefaultmidpunct}
{\mcitedefaultendpunct}{\mcitedefaultseppunct}\relax
\EndOfBibitem
\bibitem{PDG2014}
Particle Data Group, K.~A. Olive {\em et~al.},
\ifthenelse{\boolean{articletitles}}{\emph{{\href{http://pdg.lbl.gov/}{Review
of particle physics}}},
}{}\href{http://dx.doi.org/10.1088/1674-1137/38/9/090001}{Chin.\ Phys.\
\textbf{C38} (2014) 090001}\relax
\mciteBstWouldAddEndPuncttrue
\mciteSetBstMidEndSepPunct{\mcitedefaultmidpunct}
{\mcitedefaultendpunct}{\mcitedefaultseppunct}\relax
\EndOfBibitem
\bibitem{PhysRevLett.10.531}
N.~Cabibbo, \ifthenelse{\boolean{articletitles}}{\emph{{Unitary symmetry and
leptonic decays}},
}{}\href{http://dx.doi.org/10.1103/PhysRevLett.10.531}{Phys.\ Rev.\ Lett.\
\textbf{10} (1963) 531}\relax
\mciteBstWouldAddEndPuncttrue
\mciteSetBstMidEndSepPunct{\mcitedefaultmidpunct}
{\mcitedefaultendpunct}{\mcitedefaultseppunct}\relax
\EndOfBibitem
\bibitem{PTP.49.652}
M.~Kobayashi and T.~Maskawa,
\ifthenelse{\boolean{articletitles}}{\emph{{{\ensuremath{C\!P}}\xspace-violation in the
renormalizable theory of weak interaction}},
}{}\href{http://dx.doi.org/10.1143/PTP.49.652}{Progress of Theoretical
Physics \textbf{49} (1973) 652}\relax
\mciteBstWouldAddEndPuncttrue
\mciteSetBstMidEndSepPunct{\mcitedefaultmidpunct}
{\mcitedefaultendpunct}{\mcitedefaultseppunct}\relax
\EndOfBibitem
\bibitem{Gronau:1990ra}
M.~Gronau and D.~London, \ifthenelse{\boolean{articletitles}}{\emph{{How to
determine all the angles of the unitarity triangle from ${\ensuremath{\B^0}}\xspace \ensuremath{\rightarrow}\xspace D {\ensuremath{\kaon^0_{\rm\scriptscriptstyle S}}}\xspace$ and
${\ensuremath{\B^0_\squark}}\xspace \ensuremath{\rightarrow}\xspace D\phi$ }},
}{}\href{http://dx.doi.org/10.1016/0370-2693(91)91756-L}{Phys.\ Lett.\
\textbf{B253} (1991) 483}\relax
\mciteBstWouldAddEndPuncttrue
\mciteSetBstMidEndSepPunct{\mcitedefaultmidpunct}
{\mcitedefaultendpunct}{\mcitedefaultseppunct}\relax
\EndOfBibitem
\bibitem{Gronau:1991dp}
M.~Gronau and D.~Wyler, \ifthenelse{\boolean{articletitles}}{\emph{{On
determining a weak phase from charged {\ensuremath{\PB}}\xspace decay asymmetries}},
}{}\href{http://dx.doi.org/10.1016/0370-2693(91)90034-N}{Phys.\ Lett.\
\textbf{B265} (1991) 172}\relax
\mciteBstWouldAddEndPuncttrue
\mciteSetBstMidEndSepPunct{\mcitedefaultmidpunct}
{\mcitedefaultendpunct}{\mcitedefaultseppunct}\relax
\EndOfBibitem
\bibitem{Sinha:2004ct}
N.~Sinha, \ifthenelse{\boolean{articletitles}}{\emph{{Determining $\gamma$
using $B \ensuremath{\rightarrow}\xspace D^{**}K$}},
}{}\href{http://dx.doi.org/10.1103/PhysRevD.70.097501}{Phys.\ Rev.\
\textbf{D70} (2004) 097501}, \href{http://arxiv.org/abs/hep-ph/0405061}{{\tt
arXiv:hep-ph/0405061}}\relax
\mciteBstWouldAddEndPuncttrue
\mciteSetBstMidEndSepPunct{\mcitedefaultmidpunct}
{\mcitedefaultendpunct}{\mcitedefaultseppunct}\relax
\EndOfBibitem
\bibitem{Alves:2008zz}
LHCb collaboration, A.~A. Alves~Jr.\ {\em et~al.},
\ifthenelse{\boolean{articletitles}}{\emph{{The \mbox{LHCb}\xspace detector at the LHC}},
}{}\href{http://dx.doi.org/10.1088/1748-0221/3/08/S08005}{JINST \textbf{3}
(2008) S08005}\relax
\mciteBstWouldAddEndPuncttrue
\mciteSetBstMidEndSepPunct{\mcitedefaultmidpunct}
{\mcitedefaultendpunct}{\mcitedefaultseppunct}\relax
\EndOfBibitem
\bibitem{LHCb-DP-2014-002}
LHCb collaboration, R.~Aaij {\em et~al.},
\ifthenelse{\boolean{articletitles}}{\emph{{LHCb detector performance}},
}{}\href{http://arxiv.org/abs/1412.6352}{{\tt arXiv:1412.6352}}, {to appear
in Int. J. Mod. Phys. A}\relax
\mciteBstWouldAddEndPuncttrue
\mciteSetBstMidEndSepPunct{\mcitedefaultmidpunct}
{\mcitedefaultendpunct}{\mcitedefaultseppunct}\relax
\EndOfBibitem
\bibitem{LHCb-DP-2014-001}
R.~Aaij {\em et~al.}, \ifthenelse{\boolean{articletitles}}{\emph{{Performance
of the LHCb Vertex Locator}},
}{}\href{http://dx.doi.org/10.1088/1748-0221/9/09/P09007}{JINST \textbf{9}
(2014) P09007}, \href{http://arxiv.org/abs/1405.7808}{{\tt
arXiv:1405.7808}}\relax
\mciteBstWouldAddEndPuncttrue
\mciteSetBstMidEndSepPunct{\mcitedefaultmidpunct}
{\mcitedefaultendpunct}{\mcitedefaultseppunct}\relax
\EndOfBibitem
\bibitem{LHCb-DP-2013-003}
R.~Arink {\em et~al.}, \ifthenelse{\boolean{articletitles}}{\emph{{Performance
of the LHCb Outer Tracker}},
}{}\href{http://dx.doi.org/10.1088/1748-0221/9/01/P01002}{JINST \textbf{9}
(2014) P01002}, \href{http://arxiv.org/abs/1311.3893}{{\tt
arXiv:1311.3893}}\relax
\mciteBstWouldAddEndPuncttrue
\mciteSetBstMidEndSepPunct{\mcitedefaultmidpunct}
{\mcitedefaultendpunct}{\mcitedefaultseppunct}\relax
\EndOfBibitem
\bibitem{LHCb-DP-2012-003}
M.~Adinolfi {\em et~al.},
\ifthenelse{\boolean{articletitles}}{\emph{{Performance of the \mbox{LHCb}\xspace RICH
detector at the LHC}},
}{}\href{http://dx.doi.org/10.1140/epjc/s10052-013-2431-9}{Eur.\ Phys.\ J.\
\textbf{C73} (2013) 2431}, \href{http://arxiv.org/abs/1211.6759}{{\tt
arXiv:1211.6759}}\relax
\mciteBstWouldAddEndPuncttrue
\mciteSetBstMidEndSepPunct{\mcitedefaultmidpunct}
{\mcitedefaultendpunct}{\mcitedefaultseppunct}\relax
\EndOfBibitem
\bibitem{LHCb-DP-2012-002}
A.~A. Alves~Jr.\ {\em et~al.},
\ifthenelse{\boolean{articletitles}}{\emph{{Performance of the LHCb muon
system}}, }{}\href{http://dx.doi.org/10.1088/1748-0221/8/02/P02022}{JINST
\textbf{8} (2013) P02022}, \href{http://arxiv.org/abs/1211.1346}{{\tt
arXiv:1211.1346}}\relax
\mciteBstWouldAddEndPuncttrue
\mciteSetBstMidEndSepPunct{\mcitedefaultmidpunct}
{\mcitedefaultendpunct}{\mcitedefaultseppunct}\relax
\EndOfBibitem
\bibitem{LHCb-DP-2012-004}
R.~Aaij {\em et~al.}, \ifthenelse{\boolean{articletitles}}{\emph{{The \mbox{LHCb}\xspace
trigger and its performance in 2011}},
}{}\href{http://dx.doi.org/10.1088/1748-0221/8/04/P04022}{JINST \textbf{8}
(2013) P04022}, \href{http://arxiv.org/abs/1211.3055}{{\tt
arXiv:1211.3055}}\relax
\mciteBstWouldAddEndPuncttrue
\mciteSetBstMidEndSepPunct{\mcitedefaultmidpunct}
{\mcitedefaultendpunct}{\mcitedefaultseppunct}\relax
\EndOfBibitem
\bibitem{BBDT}
V.~V. Gligorov and M.~Williams,
\ifthenelse{\boolean{articletitles}}{\emph{{Efficient, reliable and fast
high-level triggering using a bonsai boosted decision tree}},
}{}\href{http://dx.doi.org/10.1088/1748-0221/8/02/P02013}{JINST \textbf{8}
(2013) P02013}, \href{http://arxiv.org/abs/1210.6861}{{\tt
arXiv:1210.6861}}\relax
\mciteBstWouldAddEndPuncttrue
\mciteSetBstMidEndSepPunct{\mcitedefaultmidpunct}
{\mcitedefaultendpunct}{\mcitedefaultseppunct}\relax
\EndOfBibitem
\bibitem{Sjostrand:2006za}
T.~Sj\"{o}strand, S.~Mrenna, and P.~Skands,
\ifthenelse{\boolean{articletitles}}{\emph{{PYTHIA 6.4 physics and manual}},
}{}\href{http://dx.doi.org/10.1088/1126-6708/2006/05/026}{JHEP \textbf{05}
(2006) 026}, \href{http://arxiv.org/abs/hep-ph/0603175}{{\tt
arXiv:hep-ph/0603175}}\relax
\mciteBstWouldAddEndPuncttrue
\mciteSetBstMidEndSepPunct{\mcitedefaultmidpunct}
{\mcitedefaultendpunct}{\mcitedefaultseppunct}\relax
\EndOfBibitem
\bibitem{Sjostrand:2007gs}
T.~Sj\"{o}strand, S.~Mrenna, and P.~Skands,
\ifthenelse{\boolean{articletitles}}{\emph{{A brief introduction to PYTHIA
8.1}}, }{}\href{http://dx.doi.org/10.1016/j.cpc.2008.01.036}{Comput.\ Phys.\
Commun.\ \textbf{178} (2008) 852},
\href{http://arxiv.org/abs/0710.3820}{{\tt arXiv:0710.3820}}\relax
\mciteBstWouldAddEndPuncttrue
\mciteSetBstMidEndSepPunct{\mcitedefaultmidpunct}
{\mcitedefaultendpunct}{\mcitedefaultseppunct}\relax
\EndOfBibitem
\bibitem{LHCb-PROC-2010-056}
I.~Belyaev {\em et~al.}, \ifthenelse{\boolean{articletitles}}{\emph{{Handling
of the generation of primary events in Gauss, the LHCb simulation
framework}}, }{}\href{http://dx.doi.org/10.1088/1742-6596/331/3/032047}{{J.\
Phys.\ Conf.\ Ser.\ } \textbf{331} (2011) 032047}\relax
\mciteBstWouldAddEndPuncttrue
\mciteSetBstMidEndSepPunct{\mcitedefaultmidpunct}
{\mcitedefaultendpunct}{\mcitedefaultseppunct}\relax
\EndOfBibitem
\bibitem{Lange:2001uf}
D.~J. Lange, \ifthenelse{\boolean{articletitles}}{\emph{{The EvtGen particle
decay simulation package}},
}{}\href{http://dx.doi.org/10.1016/S0168-9002(01)00089-4}{Nucl.\ Instrum.\
Meth.\ \textbf{A462} (2001) 152}\relax
\mciteBstWouldAddEndPuncttrue
\mciteSetBstMidEndSepPunct{\mcitedefaultmidpunct}
{\mcitedefaultendpunct}{\mcitedefaultseppunct}\relax
\EndOfBibitem
\bibitem{Golonka:2005pn}
P.~Golonka and Z.~Was, \ifthenelse{\boolean{articletitles}}{\emph{{PHOTOS Monte
Carlo: A precision tool for QED corrections in $Z$ and $W$ decays}},
}{}\href{http://dx.doi.org/10.1140/epjc/s2005-02396-4}{Eur.\ Phys.\ J.\
\textbf{C45} (2006) 97}, \href{http://arxiv.org/abs/hep-ph/0506026}{{\tt
arXiv:hep-ph/0506026}}\relax
\mciteBstWouldAddEndPuncttrue
\mciteSetBstMidEndSepPunct{\mcitedefaultmidpunct}
{\mcitedefaultendpunct}{\mcitedefaultseppunct}\relax
\EndOfBibitem
\bibitem{Allison:2006ve}
Geant4 collaboration, J.~Allison {\em et~al.},
\ifthenelse{\boolean{articletitles}}{\emph{{Geant4 developments and
applications}}, }{}\href{http://dx.doi.org/10.1109/TNS.2006.869826}{IEEE
Trans.\ Nucl.\ Sci.\ \textbf{53} (2006) 270}\relax
\mciteBstWouldAddEndPuncttrue
\mciteSetBstMidEndSepPunct{\mcitedefaultmidpunct}
{\mcitedefaultendpunct}{\mcitedefaultseppunct}\relax
\EndOfBibitem
\bibitem{Agostinelli:2002hh}
Geant4 collaboration, S.~Agostinelli {\em et~al.},
\ifthenelse{\boolean{articletitles}}{\emph{{Geant4: a simulation toolkit}},
}{}\href{http://dx.doi.org/10.1016/S0168-9002(03)01368-8}{Nucl.\ Instrum.\
Meth.\ \textbf{A506} (2003) 250}\relax
\mciteBstWouldAddEndPuncttrue
\mciteSetBstMidEndSepPunct{\mcitedefaultmidpunct}
{\mcitedefaultendpunct}{\mcitedefaultseppunct}\relax
\EndOfBibitem
\bibitem{LHCb-PROC-2011-006}
M.~Clemencic {\em et~al.}, \ifthenelse{\boolean{articletitles}}{\emph{{The
\mbox{LHCb}\xspace simulation application, Gauss: Design, evolution and experience}},
}{}\href{http://dx.doi.org/10.1088/1742-6596/331/3/032023}{{J.\ Phys.\ Conf.\
Ser.\ } \textbf{331} (2011) 032023}\relax
\mciteBstWouldAddEndPuncttrue
\mciteSetBstMidEndSepPunct{\mcitedefaultmidpunct}
{\mcitedefaultendpunct}{\mcitedefaultseppunct}\relax
\EndOfBibitem
\bibitem{Feindt2006190}
M.~Feindt and U.~Kerzel, \ifthenelse{\boolean{articletitles}}{\emph{{The
NeuroBayes neural network package}},
}{}\href{http://dx.doi.org/10.1016/j.nima.2005.11.166}{Nucl.\ Instrum.\
Meth.\ A \textbf{559} (2006) 190}\relax
\mciteBstWouldAddEndPuncttrue
\mciteSetBstMidEndSepPunct{\mcitedefaultmidpunct}
{\mcitedefaultendpunct}{\mcitedefaultseppunct}\relax
\EndOfBibitem
\bibitem{Pivk:2004ty}
M.~Pivk and F.~R. Le~Diberder,
\ifthenelse{\boolean{articletitles}}{\emph{{sPlot: A statistical tool to
unfold data distributions}},
}{}\href{http://dx.doi.org/10.1016/j.nima.2005.08.106}{Nucl.\ Instrum.\
Meth.\ \textbf{A555} (2005) 356},
\href{http://arxiv.org/abs/physics/0402083}{{\tt
arXiv:physics/0402083}}\relax
\mciteBstWouldAddEndPuncttrue
\mciteSetBstMidEndSepPunct{\mcitedefaultmidpunct}
{\mcitedefaultendpunct}{\mcitedefaultseppunct}\relax
\EndOfBibitem
\bibitem{LHCb-PAPER-2012-001}
LHCb collaboration, R.~Aaij {\em et~al.},
\ifthenelse{\boolean{articletitles}}{\emph{{Observation of $CP$ violation in
$B^\pm \ensuremath{\rightarrow}\xspace D K^\pm$ decays}},
}{}\href{http://dx.doi.org/10.1016/j.physletb.2012.04.060}{Phys.\ Lett.\
\textbf{B712} (2012) 203}, Erratum
\href{http://dx.doi.org/10.1016/j.physletb.2012.05.060}{ibid.\
\textbf{B713} (2012) 351}, \href{http://arxiv.org/abs/1203.3662}{{\tt
arXiv:1203.3662}}\relax
\mciteBstWouldAddEndPuncttrue
\mciteSetBstMidEndSepPunct{\mcitedefaultmidpunct}
{\mcitedefaultendpunct}{\mcitedefaultseppunct}\relax
\EndOfBibitem
\bibitem{LHCb-PAPER-2012-048}
LHCb collaboration, R.~Aaij {\em et~al.},
\ifthenelse{\boolean{articletitles}}{\emph{{Measurements of the
$\Lambda_b^0$, $\Xi_b^-$, and $\Omega_b^-$ baryon masses}},
}{}\href{http://dx.doi.org/10.1103/PhysRevLett.110.182001}{Phys.\ Rev.\
Lett.\ \textbf{110} (2013) 182001},
\href{http://arxiv.org/abs/1302.1072}{{\tt arXiv:1302.1072}}\relax
\mciteBstWouldAddEndPuncttrue
\mciteSetBstMidEndSepPunct{\mcitedefaultmidpunct}
{\mcitedefaultendpunct}{\mcitedefaultseppunct}\relax
\EndOfBibitem
\bibitem{LHCb-PAPER-2013-011}
LHCb collaboration, R.~Aaij {\em et~al.},
\ifthenelse{\boolean{articletitles}}{\emph{{Precision measurement of $D$
meson mass differences}},
}{}\href{http://dx.doi.org/10.1007/JHEP06(2013)065}{JHEP \textbf{06} (2013)
065}, \href{http://arxiv.org/abs/1304.6865}{{\tt arXiv:1304.6865}}\relax
\mciteBstWouldAddEndPuncttrue
\mciteSetBstMidEndSepPunct{\mcitedefaultmidpunct}
{\mcitedefaultendpunct}{\mcitedefaultseppunct}\relax
\EndOfBibitem
\bibitem{Hulsbergen:2005pu}
W.~D. Hulsbergen, \ifthenelse{\boolean{articletitles}}{\emph{{Decay chain
fitting with a Kalman filter}},
}{}\href{http://dx.doi.org/10.1016/j.nima.2005.06.078}{Nucl.\ Instrum.\
Meth.\ \textbf{A552} (2005) 566},
\href{http://arxiv.org/abs/physics/0503191}{{\tt
arXiv:physics/0503191}}\relax
\mciteBstWouldAddEndPuncttrue
\mciteSetBstMidEndSepPunct{\mcitedefaultmidpunct}
{\mcitedefaultendpunct}{\mcitedefaultseppunct}\relax
\EndOfBibitem
\bibitem{Skwarnicki:1986xj}
T.~Skwarnicki, {\em {A study of the radiative cascade transitions between the
Upsilon-prime and Upsilon resonances}}, PhD thesis, Institute of Nuclear
Physics, Krakow, 1986,
{\href{http://inspirehep.net/record/230779/}{DESY-F31-86-02}}\relax
\mciteBstWouldAddEndPuncttrue
\mciteSetBstMidEndSepPunct{\mcitedefaultmidpunct}
{\mcitedefaultendpunct}{\mcitedefaultseppunct}\relax
\EndOfBibitem
\bibitem{LHCb-PAPER-2012-018}
LHCb collaboration, R.~Aaij {\em et~al.},
\ifthenelse{\boolean{articletitles}}{\emph{{Observation of $B^0 \ensuremath{\rightarrow}\xspace \bar{D}^0
K^+ K^-$ and evidence for $B^0_s \ensuremath{\rightarrow}\xspace \bar{D}^0 K^+ K^-$}},
}{}\href{http://dx.doi.org/10.1103/PhysRevLett.109.131801}{Phys.\ Rev.\
Lett.\ \textbf{109} (2012) 131801},
\href{http://arxiv.org/abs/1207.5991}{{\tt arXiv:1207.5991}}\relax
\mciteBstWouldAddEndPuncttrue
\mciteSetBstMidEndSepPunct{\mcitedefaultmidpunct}
{\mcitedefaultendpunct}{\mcitedefaultseppunct}\relax
\EndOfBibitem
\bibitem{Dalitz:1953cp}
R.~H. Dalitz, \ifthenelse{\boolean{articletitles}}{\emph{{On the analysis of
tau-meson data and the nature of the tau-meson}},
}{}\href{http://dx.doi.org/10.1080/14786441008520365}{Phil.\ Mag.\
\textbf{44} (1953) 1068}\relax
\mciteBstWouldAddEndPuncttrue
\mciteSetBstMidEndSepPunct{\mcitedefaultmidpunct}
{\mcitedefaultendpunct}{\mcitedefaultseppunct}\relax
\EndOfBibitem
\bibitem{Fleming:1964zz}
G.~N. Fleming, \ifthenelse{\boolean{articletitles}}{\emph{{Recoupling effects
in the isobar model. 1. General formalism for three-pion scattering}},
}{}\href{http://dx.doi.org/10.1103/PhysRev.135.B551}{Phys.\ Rev.\
\textbf{135} (1964) B551}\relax
\mciteBstWouldAddEndPuncttrue
\mciteSetBstMidEndSepPunct{\mcitedefaultmidpunct}
{\mcitedefaultendpunct}{\mcitedefaultseppunct}\relax
\EndOfBibitem
\bibitem{Morgan:1968zza}
D.~Morgan, \ifthenelse{\boolean{articletitles}}{\emph{{Phenomenological
analysis of $I=\frac{1}{2}$ single-pion production processes in the energy
range 500 to 700 MeV}},
}{}\href{http://dx.doi.org/10.1103/PhysRev.166.1731}{Phys.\ Rev.\
\textbf{166} (1968) 1731}\relax
\mciteBstWouldAddEndPuncttrue
\mciteSetBstMidEndSepPunct{\mcitedefaultmidpunct}
{\mcitedefaultendpunct}{\mcitedefaultseppunct}\relax
\EndOfBibitem
\bibitem{Herndon:1973yn}
D.~Herndon, P.~Soding, and R.~J. Cashmore,
\ifthenelse{\boolean{articletitles}}{\emph{{A generalised isobar model
formalism}}, }{}\href{http://dx.doi.org/10.1103/PhysRevD.11.3165}{Phys.\
Rev.\ \textbf{D11} (1975) 3165}\relax
\mciteBstWouldAddEndPuncttrue
\mciteSetBstMidEndSepPunct{\mcitedefaultmidpunct}
{\mcitedefaultendpunct}{\mcitedefaultseppunct}\relax
\EndOfBibitem
\bibitem{blatt-weisskopf}
J.~Blatt and V.~E. Weisskopf, {\em Theoretical nuclear physics}, J. Wiley (New
York), 1952\relax
\mciteBstWouldAddEndPuncttrue
\mciteSetBstMidEndSepPunct{\mcitedefaultmidpunct}
{\mcitedefaultendpunct}{\mcitedefaultseppunct}\relax
\EndOfBibitem
\bibitem{Aubert:2005ce}
\mbox{BaBar}\xspace collaboration, B.~Aubert {\em et~al.},
\ifthenelse{\boolean{articletitles}}{\emph{{Dalitz-plot analysis of the
decays $B^\pm \ensuremath{\rightarrow}\xspace K^\pm \pi^\mp \pi^\pm$}},
}{}\href{http://dx.doi.org/10.1103/PhysRevD.72.072003}{Phys.\ Rev.\
\textbf{D72} (2005) 072003}, Erratum
\href{http://dx.doi.org/10.1103/PhysRevD.74.099903}{ibid.\ \textbf{D74}
(2006) 099903}, \href{http://arxiv.org/abs/hep-ex/0507004}{{\tt
arXiv:hep-ex/0507004}}\relax
\mciteBstWouldAddEndPuncttrue
\mciteSetBstMidEndSepPunct{\mcitedefaultmidpunct}
{\mcitedefaultendpunct}{\mcitedefaultseppunct}\relax
\EndOfBibitem
\bibitem{Zemach:1963bc}
C.~Zemach, \ifthenelse{\boolean{articletitles}}{\emph{{Three pion decays of
unstable particles}},
}{}\href{http://dx.doi.org/10.1103/PhysRev.133.B1201}{Phys.\ Rev.\
\textbf{133} (1964) B1201}\relax
\mciteBstWouldAddEndPuncttrue
\mciteSetBstMidEndSepPunct{\mcitedefaultmidpunct}
{\mcitedefaultendpunct}{\mcitedefaultseppunct}\relax
\EndOfBibitem
\bibitem{Zemach:1968zz}
C.~Zemach, \ifthenelse{\boolean{articletitles}}{\emph{{Use of angular-momentum
tensors}}, }{}\href{http://dx.doi.org/10.1103/PhysRev.140.B97}{Phys.\ Rev.\
\textbf{140} (1965) B97}\relax
\mciteBstWouldAddEndPuncttrue
\mciteSetBstMidEndSepPunct{\mcitedefaultmidpunct}
{\mcitedefaultendpunct}{\mcitedefaultseppunct}\relax
\EndOfBibitem
\bibitem{Garmash:2004wa}
Belle collaboration, A.~Garmash {\em et~al.},
\ifthenelse{\boolean{articletitles}}{\emph{{Dalitz analysis of the three-body
charmless decays ${\ensuremath{\Bu}}\xspace\ensuremath{\rightarrow}\xspace {\ensuremath{\kaon^+}}\xspace{\ensuremath{\pion^+}}\xspace{\ensuremath{\pion^-}}\xspace$ and ${\ensuremath{\Bu}}\xspace \ensuremath{\rightarrow}\xspace {\ensuremath{\kaon^+}}\xspace\Kp{\ensuremath{\kaon^-}}\xspace$}},
}{}\href{http://dx.doi.org/10.1103/PhysRevD.71.092003}{Phys.\ Rev.\
\textbf{D71} (2005) 092003}, \href{http://arxiv.org/abs/hep-ex/0412066}{{\tt
arXiv:hep-ex/0412066}}\relax
\mciteBstWouldAddEndPuncttrue
\mciteSetBstMidEndSepPunct{\mcitedefaultmidpunct}
{\mcitedefaultendpunct}{\mcitedefaultseppunct}\relax
\EndOfBibitem
\bibitem{Laura++}
{{\tt Laura++} Dalitz plot fitting package, \url{http://laura.hepfo
rge.org}}\relax
\mciteBstWouldAddEndPuncttrue
\mciteSetBstMidEndSepPunct{\mcitedefaultmidpunct}
{\mcitedefaultendpunct}{\mcitedefaultseppunct}\relax
\EndOfBibitem
\bibitem{Ben-Haim:2014afa}
E.~Ben-Haim, R.~Brun, B.~Echenard, and T.~E. Latham,
\ifthenelse{\boolean{articletitles}}{\emph{{JFIT: a framework to obtain
combined experimental results through joint fits}},
}{}\href{http://arxiv.org/abs/1409.5080}{{\tt arXiv:1409.5080}}\relax
\mciteBstWouldAddEndPuncttrue
\mciteSetBstMidEndSepPunct{\mcitedefaultmidpunct}
{\mcitedefaultendpunct}{\mcitedefaultseppunct}\relax
\EndOfBibitem
\bibitem{Williams:2010vh}
M.~Williams, \ifthenelse{\boolean{articletitles}}{\emph{{How good are your
fits? Unbinned multivariate goodness-of-fit tests in high energy physics}},
}{}\href{http://dx.doi.org/10.1088/1748-0221/5/09/P09004}{JINST \textbf{5}
(2010) P09004}, \href{http://arxiv.org/abs/1006.3019}{{\tt
arXiv:1006.3019}}\relax
\mciteBstWouldAddEndPuncttrue
\mciteSetBstMidEndSepPunct{\mcitedefaultmidpunct}
{\mcitedefaultendpunct}{\mcitedefaultseppunct}\relax
\EndOfBibitem
\end{mcitethebibliography}
\clearpage
|
\section{Introduction}
We begin by considering an immersed curve $\Gamma: S^1 \to \mathbb R^2$. The \emph{rotation number} $\mathrm{rot}(\Gamma)$ is defined to be the (signed) number of turns made by the tangent vector as we travel along $\Gamma$, or more formally the degree of $\Gamma$'s Gauss map. The \emph{index} $\mathrm{ind}_{\Gamma}(p)$ of a point $p \in \mathbb R^2 \setminus \Gamma$ with respect to $\Gamma$ is the (signed) number of times $\Gamma$ revolves around $p$, or more formally the degree of the map $S^1 \to S^1$ given by
\[ t \mapsto \frac{\Gamma(t) - p}{\|\Gamma(t) - p\|} \]
\begin{rem}The reader should beware that in some sources the rotation number is referred to as the index of the curve, and that the term ``winding number'' may be used to refer either to the index or the rotation number.
\end{rem}
It is a basic fact of differential geometry that $\Gamma$'s rotation number can be calculated by integrating its curvature and dividing by $2\pi$, i.e.
\begin{equation}\label{HU-multiplicities}
\frac{1}{2\pi} \int_{S^1} k(t) \, dt = \mathrm{rot}(\Gamma)
\end{equation}
This formula is referred to as Hopf's Umlaufsatz by Lanzat and Polyak in \cite{lp}, although other sources, including the paper in which Hopf introduced the term, \cite{hopf}, use ``Umlaufsatz'' to refer to the more specific case where $\Gamma$ is a simple curve (i.e. $\Gamma$ has no self-intersections), in which case
\begin{equation}\label{HU-original}
\frac{1}{2\pi}\int_{S^1} k(t) \, dt = \pm 1
\end{equation}
In this sense Equation (\ref{HU-multiplicities}) can be seen as an ``Umlaufsatz with multiplicities''. Later, when we consider curves in more general surfaces, we will see that we need a ``Gauss-Bonnet theorem with multiplicities'' in the same vein.
While the rotation number can be calculated for any immersed curve in the plane and is invariant under all regular homotopies, Arnold's invariants $J^\pm$ and $St$ are defined only for \emph{generic} curves in the plane---curves which have a finite number of transverse double points as their only self-intersections---and are invariant under regular homotopies within this narrower class.
It is well-known that if two generic curves are regular homotopic in the space of immersed curves, then one can be transformed into the other by a series of homotopies in the space of generic curves together with a finite number of \emph{self-tangency} and \emph{triple-point} moves, shown in Figure \ref{moves}. Self-tangency moves can be further classified into direct and opposite self-tangencies, according to whether the two involved arcs of the curve are pointing in the same direction or in opposite directions (see Figure \ref{st-classification}). Arnold's $J^+$ invariant can be understood as counting direct self-tangency moves: It increases by 2 under direct self-tangency moves and is unchanged under opposite self-tangency moves and triple-point moves. The $J^-$ and $St$ invariants play analogous roles for opposite self-tangency moves and triple-point moves, respectively \cite{j+}.
\begin{figure}
\centering
\begin{subfigure}{0.4\textwidth}
\includegraphics[width=\textwidth, trim = 1cm 18cm 4cm 3cm, clip]{self-tangency}
\caption{A self-tangency move}
\end{subfigure}
\hspace{2cm}
\begin{subfigure}{0.4\textwidth}
\includegraphics[width=\textwidth, trim = 0cm 19cm 1cm 0, clip]{triple-point}
\caption{A triple-point move}
\end{subfigure}
\caption{The self-tangency and triple-point moves}
\label{moves}
\end{figure}
\begin{figure}
\centering
\begin{subfigure}{0.4\textwidth}
\includegraphics[width=\textwidth, trim = 1cm 18cm 3cm 3cm, clip]{direct-self-tangency}
\caption{A direct self-tangency move}
\end{subfigure}
\hspace{2cm}
\begin{subfigure}{0.4\textwidth}
\includegraphics[width=\textwidth, trim = 1cm 18cm 3cm 3cm, clip]{opposite-self-tangency}
\caption{An opposite self-tangency move}
\end{subfigure}
\caption{Direct and opposite self-tangency moves}
\label{st-classification}
\end{figure}
This characterization of $J^+(\Gamma)$ determines it uniquely up to addition of a term depending only on the regular homotopy class of $\Gamma$ (in the class of immersed curves). Arnold specifies a value of $J^+$ on a standard representative of each homotopy class of $\Gamma$, thereby specifying it exactly \cite{j+}.
Lanzat and Polyak's polynomial invariant in \cite{lp} is likewise defined for generic curves in the plane. For such a $\Gamma$, they extend $\mathrm{ind}_{\Gamma}$ to a total function on $\mathbb R^2$ by first noting that $\mathrm{ind}_{\Gamma}$ is locally constant on $\mathbb R^2 \setminus \Gamma$ and then defining $\mathrm{ind}_\Gamma(p)$ for $p \in \Gamma$ to be the average of $\mathrm{ind}_\Gamma$ over the connected components of $\mathbb R^2 \setminus \Gamma$ in a neighborhood of $p$ (See Figure \ref{index}). Then Lanzat and Polyak's invariant $I_q(\Gamma) \in \mathbb R[q^{\frac12}, q^{-\frac12}]$ is given by
\begin{equation}\label{lp}
I_q(\Gamma) = \frac1{2\pi} \left(\int_{S^1} k(t) \cdot q^{\mathrm{ind}_\Gamma(\Gamma(t))} \, dt
- \sum_{d\in X} \theta_d \cdot q^{\mathrm{ind}_\Gamma(d)}(q^{\frac12} - q^{-\frac12})\right)
\end{equation}
where $X$ is the set of double points of $\Gamma$, and for each $X \ni d = \Gamma(t_1) = \Gamma(t_2)$, one defines $\theta_d \in (0,\pi)$ as the unsigned angle between $\Gamma'(t_1)$ and $-\Gamma'(t_2)$.
Substituting $q=1$ into the polynomial immediately gives
\[ I_1(\Gamma) = \frac{1}{2\pi} \int_{S^1} k(t) \, dt = \mathrm{rot}(\Gamma) \]
It is in this sense that Lanzat and Polyak's polynomial is considered to be a quantization or quantum deformation of the rotation number. Lanzat and Polyak show that their polynomial is invariant under regular homotopies in the space of generic curves. They calculate its value on representatives of the regular homotopy classes (with respect to regular homotopies in the space of \emph{all} immersed curves) and how it changes under the different kinds of self-tangency and triple-point moves. From these results, they easily show the relation
\[ I_1'(\Gamma) = \frac12 (1 - J^+(\Gamma)) \]
where $I_1'(\Gamma)$ is the linear term in the Taylor expansion of $I_q(\Gamma)$ at $q=1$. From this relation and Equation (\ref{lp}) they obtain an integral expression for $J^+(\Gamma)$.
\begin{figure}
\centering
\includegraphics[width = .33\textwidth, trim = 5.5cm 17cm 10cm 5cm, clip]{index}
\caption{A double point of index $i$}
\label{index}
\end{figure}
Both the rotation number and the $J^+$ invariant have been extended to curves in more general surfaces. For the rotation number, first note that the rotation number on the plane can be defined as the unique homomorphism from the group of regular homotopy classes of curves immersed in the plane to the group $\mathbb Z$, subject to the constraint that a small counterclockwise loop should map to 1. To determine a rotation number for a curve in a surface $S$, we first need to specify a set of smooth generators of the fundamental group of $S$; then there is a unique homomorphism from the group of regular homotopy classes of curves in $S$ to $\mathbb Z/|\chi(S)|\mathbb Z$, subject to the constraints that the representatives of the fundamental group map to 0 and that a small counterclockwise contractible loop maps to 1 \cite{mc}. Here $\chi(S)$ is the Euler characteristic of $S$, and $\mathbb Z/|\chi(S)|\mathbb Z$ is simply $\mathbb Z$ if $\chi(S)$ is 0. The choice of generators does not affect the rotation number of homologically trivial curves, which will be the focus of most of the remainder of the paper. For our purposes, then, we can take the following as the definition of the rotation number: The \emph{rotation number} for homologically trivial curves in a closed oriented surface $S$ is the unique homomorphism from the group of regular homotopy classes of homologically trivial curves in $S$ to $\mathbb Z/|\chi(S)|\mathbb Z$, subject to the constraint that a small counterclockwise contractible loop maps to 1.
Several generalizations of $J^+$ are given by Viro in \cite{giro}. Most significantly for our purposes, he defines an invariant $J^+(\Gamma)$ for generic homologically trivial curves immersed in oriented closed surfaces, which is unchanged by opposite self-tangency moves and triple-point moves, and which under direct self-tangency moves increases by 2.
\begin{rem}
Actually Viro generalizes the related invariant $J^-$, which is unchanged by direct self-tangency moves and triple-point moves, and which decreases by 2 at opposite self-tangencies. As he points out, though, given such an invariant $J^-$ we can immediately define an invariant $J^+$ by $J^+ = J^- + n$, where $n$ is the number of double points of the curve. It is easily checked that $J^+$ defined in this way is unchanged by opposite self-tangency moves and triple-point moves, and increases by 2 under direct self-tangency moves.
\end{rem}
Viro's $J^+$ generalizes Arnold's earlier construction of a $J^+$ invariant for curves on the sphere.
To calculate $SJ^+(\Gamma)$, as Arnold called it, stereographically project the sphere onto the plane, choosing some point in $S^2 \setminus \Gamma$ to be the point at infinity. Then letting $\Gamma'$ be the resulting curve in the plane,
\begin{equation}\label{sj+}
SJ^+(\Gamma) = J^+(\Gamma') + \frac{\mathrm{rot}(\Gamma')^2}{2}
\end{equation}
Arnold showed that $SJ^+$ is independent of the point we choose as the point at infinity in the stereographic projection, and it clearly changes in the way that a $J^+$ invariant is required to under self-tangency and triple-point moves.
In this paper we attempt to generalize Lanzat and Polyak's polynomial and related results to curves in surfaces, as far as is possible. We introduce a more general index function which makes sense in a general connected oriented surface, but the price for this generality is twofold: The new index function $\mathrm{ind}_{\Gamma,b}$ is defined not only in terms of $\Gamma$ but also a base point $b$ in the surface, and $\Gamma$ must be homologically trivial. Using the $\mathrm{ind}_{\Gamma,b}$ we construct a polynomial $I_q(\Gamma,b)$, which is defined by integrating local geometric data much as Lanzat and Polyak's polynomial is. We show that $I_q(\Gamma,b)$ is independent of the Riemannian metric on the surface. Evaluating the polynomial at $q=1$, we obtain a new integral formula for the rotation number, which is also a generalization of one form of the Gauss-Bonnet theorem. For surfaces with nonzero Euler characteristic, we show that $J^+(\Gamma)$ can be calculated from the value and first derivative of $I_q(\Gamma,b)$ at $q=1$, thereby obtaining an integral formula for $J^+(\Gamma)$. Lastly, we use this formula to give an explicit expression for $SJ^+$.
\section{Generalizing the Index Function}
In a general surface, the concept of how many times a curve goes around a point is meaningless. Note, however, that before extending its domain from $\mathbb R^2 \setminus \Gamma$ to $\mathbb R^2$, $\mathrm{ind}_\Gamma$ is the unique function $\mathbb R^2 \setminus \Gamma \to \mathbb Z$ satisfying the following conditions:
\begin{enumerate}[(i)]
\item $\mathrm{ind}_\Gamma$ is locally constant.
\item $\mathrm{ind}_\Gamma$ increases by 1 when we make a positive crossing over $\Gamma$.
\item $\mathrm{ind}_\Gamma(p) = 0$ for a point $p$ on the outside of $\Gamma$.
\end{enumerate}
Although the third condition has no meaning in a general surface, the first two can be considered in any oriented surface. For an oriented closed surface $S$ and a generic smooth curve $\Gamma:S^1 \to S$, constructing an integer-valued function on $\mathbb R^2 \setminus \Gamma$ satisfying the conditions (i) and (ii) is equivalent to constructing a singular 2-chain with boundary $\Gamma$. Thus such a function can be constructed if and only if $\Gamma$ is homologically trivial. In this case, if $S$ is connected then the function is uniquely determined up to addition of a constant. This motivates the following definition: For an oriented connected surface $S$, a homologically trivial generic curve $\Gamma:S^1 \to S$, and a base point $b \in S \setminus \Gamma$, let $\mathrm{ind}_{\Gamma,b}$ be the unique function $S \setminus \Gamma \to \mathbb Z$ satisfying the following conditions:
\begin{enumerate}[(i)]
\item $\mathrm{ind}_{\Gamma,b}$ is locally constant.
\item $\mathrm{ind}_{\Gamma,b}$ increases by 1 when we make a positive crossing over $\Gamma$.
\item $\mathrm{ind}_{\Gamma,b}(b) = 0$.
\end{enumerate}
\begin{rem}
It can be easily checked that $\mathrm{ind}_{\Gamma,b}(p)$ is the intersection index of any path from $b$ to $p$ with $\Gamma$, which can be taken as an alternative definition of $\mathrm{ind}_{\Gamma,b}$. In this case the requirement that $S$ is connected ensures that such a path exists, and the requirement that $\Gamma$ is homologically trivial ensures that the intersection index does not depend on the path chosen.
\end{rem}
As before, we extend $\mathrm{ind}_{\Gamma, b}$ to all of $S$ by saying that for $p$ in the image of $\Gamma$, $\mathrm{ind}_{\Gamma, b}(p)$ is found by averaging $\mathrm{ind}_{\Gamma, b}$ over the connected components of $S \setminus \Gamma$ in a neighborhood of $S$.
\section{Main Result}
Let $S$ be an oriented connected closed surface with Riemannian metric, let $\Gamma: S^1 \to S$ be a homologically trivial generic smooth curve on $S$, and let $b \in S \setminus \Gamma$. Let $X \subseteq S$ be the set of double points of $\Gamma$. For each double point $d \in X$, let $\theta_d$ be the unsigned angle between the tangent vectors $\Gamma'(t_1)$ and $-\Gamma'(t_2)$, where $\{t_1, t_2\} = \Gamma^{-1}(\{d\})$. Let $k_g(t)$ be the geodesic curvature of $\Gamma$ at $\Gamma(t)$, and let $K: S \to \mathbb R$ be the Gaussian curvature. Define $I_q(\Gamma, b)$ as
\[ \frac1{2\pi} \left(
\int_{S_1} k_g(t) \cdot q^{\mathrm{ind}_{\Gamma, b}(\Gamma(t))} \, dt
- \sum_{d \in X} \theta_d \cdot q^{\mathrm{ind}_{\Gamma, b}(d)}(q^{\frac12}-q^{-\frac12})
+ \iint_S K \cdot \frac{q^{\mathrm{ind}_{\Gamma, b}}-1}{q^{\frac12}-q^{-\frac12}} \, dA
\right) \]
\begin{thm} $I_q(\Gamma, b)$ is preserved under orientation-preserving diffeomorphisms. That is, $I_q(\Gamma, b)$ does not depend on the Riemannian metric on $S$.
\end{thm}
\begin{proof}
For $j \in \frac12 \mathbb Z \setminus \mathbb Z$, let $S_j$ be the subsurface of $S$ where the index is greater than $j$. Observe that $\partial S_j$ is a piecewise smooth curve whose pieces are the arcs of $\Gamma$ with index $j$. The orientation which $\partial S_j$ inherits from $\Gamma$ agrees with its orientation as the boundary of $S_j$. For each $i \in \mathbb Z$ let $X_i \subseteq S$ be the set of double points with degree $i$. At each $d \in X_{j-\frac12}$, $\partial S_j$ changes direction by $\pi-\theta_d$. At each $d \in X_{j-\frac12}$, $\partial S_j$ changes direction by $-(\pi-\theta_d)$. By the Gauss-Bonnet Theorem,
\[ 2\pi\chi(S_j) =
\iint_{S_j} K\,dA
+\int_{S^1} k_g(t) \cdot \mathbf1_{\mathrm{ind}_{\Gamma, b}} \, dt
+ \sum_{d \in X_{j - \frac12}}(\pi - \theta_d)
- \sum_{d \in X_{j + \frac12}}(\pi - \theta_d)
\]
and so
\[\sum_{j \in \frac12 \mathbb Z \setminus \mathbb Z} 2\pi \chi(S_j) q^j =\]
\[\iint_{S} K\sum_{i=0}^\infty q^{\mathrm{ind}_{\Gamma, b} - i - \frac12}\,dA
+\int_{S^1} k_g(t) \cdot q^{\mathrm{ind}_{\Gamma, b}(\Gamma(t))} \, dt
+\sum_{d \in X}(\pi - \theta_d)(q^{\mathrm{ind}_{\Gamma, b}(d) + \frac12} - q^{\mathrm{ind}_{\Gamma, b}(d) - \frac12})\]
\[ = \iint_{S} K\frac{q^{\mathrm{ind}_{\Gamma,b}}}{q^{\frac12} - q^{-\frac12}}\,dA
+\int_{S^1} k_g(t) \cdot q^{\mathrm{ind}_{\Gamma, b}(\Gamma(t))} \, dt
+\sum_{d \in X}(\pi - \theta_d)(q^{\mathrm{ind}_{\Gamma, b}(d) + \frac12} - q^{\mathrm{ind}_{\Gamma, b}(d) - \frac12}) \]
in $C^\infty((1,+\infty))$. Applying the Gauss-Bonnet Theorem to all of $S$ gives $2\pi \chi(S) = \iint_S K \,dA$, so
\[\sum_{j \in \frac12 \mathbb Z \setminus \mathbb Z} 2\pi \chi(S_j) q^j - \frac{2\pi\chi(S)}{q^{\frac12} - q^{-\frac12}}
- \sum_{d \in X}\pi(q^{\mathrm{ind}_{\Gamma, b}(d) + \frac12} - q^{\mathrm{ind}_{\Gamma, b}(d) - \frac12}) \]
\[ = \iint_S K \cdot \frac{q^{\mathrm{ind}_{\Gamma, b}}-1}{q^{\frac12}-q^{-\frac12}} \, dA
+\int_{S_1} k_g(t) \cdot q^{\mathrm{ind}_{\Gamma, b}(\Gamma(t))} \, dt
- \sum_{d \in X} \theta_d \cdot q^{\mathrm{ind}_{\Gamma, b}(d)}(q^{\frac12}-q^{-\frac12})
\]
\[ = 2\pi I_q(\Gamma,b) \]
Thus
\begin{equation}\label{topological-formula}
I_q(\Gamma,b) =
\sum_{j \in \frac12 \mathbb Z \setminus \mathbb Z} \chi(S_j) q^j
- \frac{\chi(S)}{q^{\frac12} - q^{-\frac12}}
- \frac12\sum_{d \in X}(q^{\mathrm{ind}_{\Gamma, b}(d) + \frac12} - q^{\mathrm{ind}_{\Gamma, b}(d) - \frac12})
\end{equation}
None of the terms on the right of (\ref{topological-formula}) depend on the Riemannian metric, so $I_q(\Gamma,b)$ is preserved under orientation-preserving diffeomorphisms.
\end{proof}
\begin{rem}
Although two of the terms on the right side of Equation (\ref{topological-formula}) diverge as $q \to 1$, $I_q(\Gamma,b)$ is a polynomial in $q^{\frac12}$ and $q^{-\frac12}$ (as is clear from its integral definition) and is thus defined for all $q>0$.
\end{rem}
\begin{rem}
It is not quite true that $I_q(\Gamma,b)$ is invariant under regular homotopies of $\Gamma$ in the space of generic curves; such a homotopy might change the position of $\Gamma$ relative to the base point $b$. However, if we view $\Gamma$ together with the choice of base point as a generic immersion $S^1 \sqcup \{\bullet\} \to S$, it follows immediately from the theorem that $I_q$ is invariant under regular homotopies in the space of such generic immersions.
\end{rem}
\section{Behavior under change of base point}
For a new base point $b'$, $\mathrm{ind}_{\Gamma, b'}$ and $\mathrm{ind}_{\Gamma, b}$ differ by a constant. Suppose $\mathrm{ind}_{\Gamma, b'}=\mathrm{ind}_{\Gamma, b} + C$. Then
\begin{multline}\label{basepoint-change}
I_q(\Gamma, b')
= \frac1{2\pi}\left(
\int_{S_1} k_g(t) \cdot q^{\mathrm{ind}_{\Gamma, b}(\Gamma(t))+C} \, dt
- \sum_{d \in X} \theta_d \cdot q^{\mathrm{ind}_{\Gamma, b}(d)+C}(q^{\frac12}-q^{-\frac12})\right.\\
\left.+ \iint_S K \cdot \frac{q^{\mathrm{ind}_{\Gamma, b}+C}-1}{q^{\frac12}-q^{-\frac12}} \, dA
\right)
\end{multline}
\[ = q^C I_q(\Gamma, b)
+ \frac1{2\pi} \iint_S K \cdot \frac{q^C - 1}{q^{\frac12}-q^{-\frac12}} \, dA\]
\[= q^C I_q(\Gamma, b) + \frac{q^C - 1}{q^{\frac12}-q^{-\frac12}} \chi(S) \]
\section{Relation to rotation number}
Since Lanzat and Polyak's polynomial $I_q(\Gamma)$ is a quantum deformation of the rotation number, a natural question is what $I_q(\Gamma,b)$ is a quantum deformation of.
Calculating $I_1(\Gamma,b)$ explicitly by substituting $q=1$ into the integral expression of $I_q(\Gamma,b)$, we get
\begin{equation}\label{GB-multiplicities}
I_1(\Gamma,b) = \frac1{2\pi} \left( \int_{S^1} k_g(t)\,dt + \iint_S K \cdot \mathrm{ind}_{\Gamma,b} \,dA \right)
\end{equation}
In the special case where $\Gamma$ is the boundary of a disk $D \subset S$ and the base point $b$ is on the outside of $D$, we have
\[ \frac1{2\pi} \left( \int_{S^1} k_g(t)\,dt + \iint_S K \cdot \mathrm{ind}_{\Gamma,b} \,dA \right)
= \frac1{2\pi} \left( \int_{\partial D} k_g \, ds + \iint_D K \, dA\right) = 1,\]
one form of the Gauss-Bonnet theorem. Thus Equation (\ref{GB-multiplicities}) is a ``Gauss-Bonnet theorem with multiplicities'' in which $\Gamma$ may wrap around multiple times. But what exactly does $I_1(\Gamma,b)$ measure? By the above intuitive reasoning and by analogy with Lanzat and Polyak's results, we should hope that $I_1(\Gamma, b)$ will coincide with the generalized definition of the rotation number. This indeed turns out to be the case, as we will demonstrate shortly.
McIntyre and Cairns give a formula for the rotation number for generic immersed curves (which they call \emph{normal}) in \cite{mc}. In the case where the curve is homologically trivial, their formula reduces to the following:
\begin{thm}[\cite{mc}]
For each positive (respectively negative) integer $i$, let $S'_i \subseteq S$ be the region where $\mathrm{ind}_{\Gamma, b}$ is greater than or equal to (respectively less than or equal to) $i$. Then the rotation number of $\Gamma$ is given by
\[ \sum_{i > 0} \chi(S'_i) - \sum_{i < 0} \chi(S'_i) \]
if $S$ is a torus, and
\[ \sum_{i > 0} \chi(S'_i) - \sum_{i < 0} \chi(S'_i) \mod |\chi(S)| \]
otherwise.
\end{thm}
Now observe that
\[ S'_i = \left\{ \begin{array}{ll} S \setminus S_{i + \frac12} & i<0 \\ S_{i - \frac12} & i>0 \end{array} \right. \]
so
\[ \sum_{i > 0} \chi(S'_i) - \sum_{i < 0} \chi(S'_i)
= \sum_{i > 0} \chi(S_{i - \frac12}) - \sum_{i < 0} \chi(S \setminus S_{i + \frac12}) \]
\[= \sum_{i > 0} \chi(S_{i - \frac12}) - \sum_{i < 0} (\chi(S) - \chi(S_{i + \frac12}))
= \sum_{j \in \frac12 \mathbb Z \setminus \mathbb Z} (\chi(S_j) - \mathbf 1_{j<0} \cdot \chi(S))\]
By Equation (\ref{topological-formula}),
\[ I_q(\Gamma, b)
= \sum_{j \in \frac12 \mathbb Z \setminus \mathbb Z} \chi(S_j) q^j
- \frac{\chi(S)}{q^{\frac12} - q^{-\frac12}}
- \frac12\sum_{d \in X}(q^{\mathrm{ind}_{\Gamma, b}(d) + \frac12} - q^{\mathrm{ind}_{\Gamma, b}(d) - \frac12}) \]
\[ = \sum_{j \in \frac12 \mathbb Z \setminus \mathbb Z}(\chi(S_j)- \chi(S))q^j - \frac12\sum_{d \in X}(q^{\mathrm{ind}_{\Gamma, b}(d) + \frac12} - q^{\mathrm{ind}_{\Gamma, b}(d) - \frac12}) \]
so evaluating $I_q(\Gamma, b)$ at $q=1$ gives
\[ I_1(\Gamma,b) = \sum_{j \in \frac12 \mathbb Z \setminus \mathbb Z} (\chi(S_j) - \mathbf 1_{j<0} \cdot \chi(S))
= \sum_{i > 0} \chi(S'_i) - \sum_{i < 0} \chi(S'_i)\]
Thus,
\begin{prop}
The rotation number of $\Gamma$ is given by
$ I_1(\Gamma,b) $
if $S$ is a torus, and
$ I_1(\Gamma,b) \mod |\chi(S)| $
otherwise.
\end{prop}
\begin{rem}
Using the formula for the change of base point, Equation (\ref{basepoint-change}), and plugging in $q=1$, we see that if $\mathrm{ind}_{\Gamma,b'} = \mathrm{ind}_{\Gamma,b} + C$,
\[ I_1(\Gamma, b') = I_1(\Gamma, b) + C\chi(S). \]
Thus $I_1(\Gamma, b)$ may depend on the choice of $b$ but $I_1(\Gamma, b) \mod |\chi(S)|$ does not.
\end{rem}
\section{Relation to the integral with respect to Euler characteristic}
Let $\tilde \Gamma$ be the smoothing of $\Gamma$, and let $\tilde S_j \subseteq S$ be the region where $\mathrm{ind}_{\tilde\Gamma,b} > 0$ for all $j \in \frac12 \mathbb Z \setminus \mathbb Z$. The Euler characteristic of the region where $\mathrm{ind}_{\tilde\Gamma,b}=i$ is
\[ \chi(\tilde S_{i -\frac12} \setminus \tilde S_{i+\frac12})
= \chi(\tilde S_{i -\frac12})- \chi(\tilde S_{i+\frac12})
= \chi(S_{i -\frac12})- \chi(S_{i+\frac12})
= a_{i -\frac12}- a_{i+\frac12} + \delta_{i,0} \cdot \chi(S)\]
so
\[ \int_{S \setminus \tilde \Gamma} q^{\mathrm{ind}_{\tilde\Gamma,b}} \, d\chi
= \chi(S) + (q^{\frac12} - q^{-\frac12}) \sum_{j \in \frac12 \mathbb Z \setminus \mathbb Z} a_j\]
\[= \chi(S) + (q^{\frac12} - q^{-\frac12}) \left( I_q(\Gamma, b) + \frac12 \sum_{d\in X} (q^\frac12 - q^{-\frac12})q^{\mathrm{ind}_{\Gamma,b}(d)} \right) \]
Solving for $I_q(\Gamma, b)$ yields
\begin{equation}\label{euler} I_q(\Gamma, b) \\
= -\frac12 \sum_{d\in X} (q^\frac12 - q^{-\frac12})q^{\mathrm{ind}_{\Gamma,b}(d)}
+ \int_{S \setminus \tilde \Gamma} \frac{q^{\mathrm{ind}_{\tilde\Gamma,b}}-1}{q^{\frac12} - q^{-\frac12}} \, d\chi
\end{equation}
\section{Relation to $J^+$ invariants}
In order to connect our results to Viro's generalization of $J^+$, it is necessary to further generalize our definition of an index function. Let a \emph{rational index function} for $\Gamma$ be any function $S \setminus \Gamma \to \mathbb Q$ which is locally constant and which increases by 1 when we make a positive crossing over $\Gamma$. Clearly any two rational index functions for $\Gamma$ differ by a rational constant. For any rational index function $\iota$ for $\Gamma$ we can extend $\iota$ to a function on $S$ by averaging over adjacent regions as before, and then define $I_q(\iota)$ by replacing $\mathrm{ind}_{\Gamma, b}$ with $\iota$ in the definition of $I_q(\Gamma,b)$. (Note that $I_q(\iota)$ no longer need belong to $\mathbb R[q^{\frac12}, q^{-\frac12}]$, but is a member of $\mathbb R[q^t: t\in \mathbb Q]$.) For any rational index function $\iota$ for $\Gamma$ there is a unique rational index function $\tilde\iota$ for $\tilde\Gamma$ such that $\iota$ and $\tilde\iota$ agree away from the curve. The formula for $I_q$ as an integral with respect to the Euler characteristic and the formula for the change of base point generalize in the obvious way. Explicitly,
\begin{equation}\label{eulerQ} I_q(\iota) \\
= -\frac12 \sum_{d\in X} (q^\frac12 - q^{-\frac12})q^{\iota(d)}
+ \int_{S \setminus \tilde \Gamma} \frac{q^{\tilde\iota}-1}{q^{\frac12} - q^{-\frac12}} \, d\chi
\end{equation}
and for any two rational index functions $\iota, \iota'$ with $\iota'-\iota = C$,
\[ I_q(\iota') = q^C I_q(\iota) + \frac{q^C - 1}{q^{\frac12}-q^{-\frac12}} \chi(S) \]
\begin{rem}
For any $\zeta$ in the relative homology group $H_2(S, \Gamma; \mathbb Q)$ such that $\partial(\zeta)$ is the fundamental class of $\Gamma$, one can define $\mathrm{ind}_\zeta(x)$ for $x \in S \setminus \Gamma$ as follows: $\mathrm{ind}_\zeta(x)$ is the image of $\zeta$ under the map
\[ H_2(S, \Gamma; \mathbb Q) \to H_2(S, S \setminus x; \mathbb Q) \to \mathbb Q \]
where the map $H_2(S, \Gamma; \mathbb Q) \to H_2(S, S \setminus x; \mathbb Q)$ is the relativization homomorphism and the map $H_2(S, S \setminus x; \mathbb Q) \to \mathbb Q$ is the canonical isomorphism. It is easily checked that $\zeta \mapsto \mathrm{ind}_\zeta$ gives an isomorphism between elements of $H_2(S, \Gamma; \mathbb Q)$ with boundary $\Gamma$ and rational index functions for $\Gamma$, so results about general rational index functions may be rephrased as results about functions of the form $\mathrm{ind}_\zeta$; indeed, it is the latter formulation that Viro uses to define $J^-$. Here we present his results using the language of rational index functions for greater congruence with the rest of the paper.
\end{rem}
For $\chi(S) \neq 0$, Viro defines $J^-(\Gamma)$ as follows: Find the unique rational index function $\tilde\iota_0$ for $\tilde \Gamma$ such that
\[ \int_{S \setminus \tilde\Gamma} \tilde\iota_0 \, d \chi = 0\]
(Note that the condition $\chi(S) \neq 0$ is necessary here to ensure the existence and uniqueness of $\tilde\iota_0$.)
Then
\[ J^-(\Gamma) = 1 - \int_{S \setminus \tilde\Gamma} \tilde\iota_0^2 \, d \chi \]
We can now state the relationship between $I_q$ and $J^+$.
\begin{prop}
Assuming that $\chi(S) \neq 0$,
\[ J^+(\Gamma) = \frac{I_1(\Gamma,b)^2}{\chi(S)} - 2 I_1'(\Gamma,b) + 1 \]
\end{prop}
\begin{proof}
From Equation (\ref{euler}) it immediately follows that
\[ I_1(\Gamma,b) = \int_{S \setminus \tilde \Gamma} (\mathrm{ind}_{\tilde\Gamma,b}) \, d\chi \]
and
\[ I'_1(\Gamma,b) = -\frac{|X|}{2} + \frac12 \int_{S \setminus \tilde \Gamma} (\mathrm{ind}_{\tilde\Gamma,b})^2 \, d\chi \]
Then $\mathrm{ind}_{\tilde\Gamma,b} - I_1(\Gamma,b)/\chi(S)$ is a rational index function for $\tilde\Gamma$ and
\[ \int_{S \setminus \tilde \Gamma} \left(\mathrm{ind}_{\tilde\Gamma,b} - \frac{I_1(\Gamma,b)}{\chi(S)}\right) \, d\chi
= I_1(\Gamma,b) - I_1(\Gamma,b) = 0\]
so $\tilde\iota_0 = \mathrm{ind}_{\tilde\Gamma,b} - I_1(\Gamma,b)/\chi(S)$ is the unique rational index function for $\tilde\Gamma$ satisfying $\int_{S \setminus \tilde\Gamma} \tilde\iota_0 \, d \chi = 0$.
Now
\[ J^-(\Gamma)
= 1 - \int_{S \setminus \tilde\Gamma} \left(\mathrm{ind}_{\tilde\Gamma,b} - \frac{I_1(\iota)}{\chi(S)}\right)^2 \, d \chi\]
\[= 1 - \int_{S \setminus \tilde\Gamma} (\mathrm{ind}_{\tilde\Gamma,b})^2 \, d \chi
+ \frac{2 I_1(\Gamma,b)}{\chi(S)}\int_{S \setminus \tilde\Gamma} (\mathrm{ind}_{\tilde\Gamma,b}) \, d \chi
- \frac{I_1(\Gamma,b)^2}{\chi(S)^2} \int_{S \setminus \tilde\Gamma} d \chi \]
\[ = 1 - (2I'_1(\Gamma,b) + |X|) + \frac{2I_1(\Gamma,b)^2}{\chi(S)} - \frac{I_1(\Gamma,b)^2}{\chi(S)} \]
\[= \frac{I_1(\Gamma,b)^2}{\chi(S)} - 2I_1'(\Gamma,b) + 1 - |X|\]
and so
\[J^+(\Gamma) = J^-(\Gamma) + |X| = \frac{I_1(\Gamma,b)^2}{\chi(S)} - 2I_1'(\Gamma,b) + 1\]
\end{proof}
Plugging in the integral definition of $I_q(\Gamma,b)$, we get
\begin{multline}\label{integral-j+}
J^+(\Gamma) =
\frac{1}{4\pi^2\chi(S)}\left( \int_{S^1} k_g(t)\,dt + \iint_S \mathrm{ind}_{\Gamma,b} \,dA \right)^2 \\
-\frac1\pi\left(\int_{S^1}k_g(t)\cdot\mathrm{ind}_{\Gamma, b}(\Gamma(t))\,dt
- \sum_{d \in X} \theta_d
+ \frac12\iint_S K \cdot (\mathrm{ind}_{\Gamma, b})^2 \, dA\right)
+ 1.
\end{multline}
\section{A formula for the $SJ^+$ invariant}
Let $\Gamma$ be any curve on the unit sphere. (Note that $\Gamma$ is automatically homologically trivial.)
Using the facts that $K=1$ on $S^2$ and $\chi(S^2) = 2$ to simplify the expression in Equation (\ref{integral-j+}) gives the following expression for the $SJ^+$ invariant:
\begin{multline*}
SJ^+(\Gamma) =
\frac{1}{8\pi^2}\left( \int_{S^1} k_g(t)\,dt + \iint_S \mathrm{ind}_{\Gamma,b} \,dA \right)^2 \\
-\frac1\pi\left(\int_{S^1}k_g(t)\cdot\mathrm{ind}_{\Gamma, b}(\Gamma(t))\,dt
- \sum_{d \in X} \theta_d
+ \frac12\iint_S(\mathrm{ind}_{\Gamma, b})^2 \, dA\right)
+ 1
\end{multline*}
\section{Acknowledgments}
The main results of this paper were obtained in the 2014 Knots \& Graphs summer undergraduate research group at the Ohio State University. I am grateful to the Ohio State University for funding the program, to fellow members of the research group for discussions, and especially to Dr. Sergei Chmutov for invaluable guidance throughout the research process.
|
\section{Introduction}
Let \((\xi_t)_{t\geq 0}\) be a L\'evy process with a L{\'e}vy triplet \(\left( \mu, \sigma^{2}, \nu \right)\). The main object of our study is the so-called generalized Ornstein-Uhlenbeck (GOU) process defined as
\begin{eqnarray}
\label{GOU}
X_{t} = e^{-\xi_{t}} \left(
X_{0} + \int_{0}^{t} e^{\xi_{u-}} d u
\right),\quad t\geq 0.
\end{eqnarray}
The GOU processes have recently got much attention in the literature. A comprehensive study of the GOU processes and an extended list of references can be found in the theses of Behme \cite{Behme}, where, in particular, it is shown that \(X_{t}\) satisfies the following SDE:
\begin{eqnarray*}
d X_{t}= X_{t-} dU_{t} +dt, \quad \mbox{where} \quad
U_{t}:= -\xi_{t} +\sum_{0<s\leq t} \left(
e^{-\Delta \xi_{s}} - 1 + \Delta \xi_{s}
\right) +\frac{1}{2} \sigma^{2} t.
\end{eqnarray*}
The popularity of GOU processes is related to the fact they appear to be useful in several applications.
For instance, the process \eqref{GOU} determines the volatility process in the COGARCH (COntinious Generalized AutoRegressive Conditionally Heteroscedastic) model introduced in Kl{\"u}ppelberg et al. \cite{cogarch}.
One important result from the theory of GOU processes is that, under some conditions, the process \eqref{GOU} is
stationary with invariant stationary distribution given by the distribution of the following exponential functional of \(\xi:\)
\begin{eqnarray}
\label{Ainfty}
A_{\infty}:= \int_{0}^{\infty} e^{-\xi_{t}} \; dt.
\end{eqnarray}
In fact, the properties of the functional \eqref{Ainfty} have been widely studied in the literature and we refer to the survey by Bertoin and Yor \cite{BertoinYor} for a theoretical background of the exponential functionals. In particular, it is known that the Mellin transform of the density \(\pi\) of exponential functional,
\[\mathcal{M}(z):= {\mathbb E} \left[A_{\infty}^{z-1}\right]=\int_{0}^{\infty}x^{z-1}\pi(x)\, dx,\]
satisfies the following recursive formula
\begin{eqnarray}
\label{momenty}
\mathcal{M}(z)
=
\frac{
\phi(z)
}{
z
}\;
\mathcal{M}(z+1),
\end{eqnarray}
where \(\phi(z)\) is a Laplace exponent of the process \(\xi\), i.e.,
\(
\phi(z) := - \log {\mathbb E} \left[
e^{ - z \xi_{1}}
\right],
\)
and complex \(z\) is taken from the strip
\begin{eqnarray}
\label{Ups}
\Upsilon:= \Bigl\{z \in {\mathbb C}: \; 0<\Re(z) < \theta \Bigr\} \quad \mbox{with}\quad \theta:=\sup\left\{x \geq 0: {\mathbb E}[ e^{-x \xi_{1}}] \leq 1\right\}.
\end{eqnarray}
The recursive formula \eqref{momenty} first appeared for real \(z\) in the paper by Maulik and Zwart \cite{MZ}. The validity of \eqref{momenty} for complex \(z\) was recently shown by Kuznetsov, Pardo and Savov \cite{Kuznets}. If \(\xi_{t}\) is a subordinator, the parameter \(\theta\) is equal to infinity. Let us note that
the functional \(A_{\infty}\) appeared in such application areas as finance (see, e.g. the monograph by Yor \cite{Yor}), carousel systems (see Litvak and Adan \cite{LitvakAdan}, Litvak and van Zwet \cite{LitvakZwet}), self-similar fragmentations (see Bertoin and Yor \cite{BertoinYor}), and information transmission problems (especially TCP/IP protocol, see Guillemin, Robert and Zwart \cite{Guill}). For the detailed discussion on the physical interpretations, we refer to Comtet, Monthus and Yor \cite{CMY} and the dissertation by Monthus \cite{Monthus}.
\par
In this paper, we mainly focus on the case when \(\xi\) is a subordinator with finite L{\'e}vy measure. In terms of the L{\'e}vy triplet \((\mu,\sigma^2,\nu)\), this means that \(\mu>0\), \(\sigma=0\), \(\nu(I\!\!R_{-})=0\) and moreover \(\a:=\nu(I\!\!R_{+})<\infty\). Suppose that the process \eqref{GOU} is observed at equidistant time points \(0=t_{0} < t_{1} < \ldots < t_{n}\). Since under some mild assumptions the process is stationary and the invariant distribution is given by the distribution of the exponential functional \(A_{\infty}\) (see Fasen, \cite{Fasen}), we assume that \(X_{t_{0}}, \ldots , X_{t_{n}}\) are also distributed as \(A_{\infty}\). Our main goal is statistical inference on the L{\'e}vy triplet \((\mu,\sigma^2,\nu)\) based on the observations \(X_{t_{0}}, \ldots , X_{t_{n}}\). More precisely, we will pursue the following two aims: (1) estimation of the drift term \(\c\) and the intensity parameter \(\a\); (2) estimation of the L{\'e}vy measure \(\nu\).
\par
To the best of our knowledge, the statistical inference for GOU processes of the form \eqref{GOU} from their low-frequency observations has not been yet studied in the literature. In fact the resulting statistical problem is quite challenging and needs careful treatment. Indeed, the only connection between the stationary distribution of a GOU process, which can be estimated from the data, and the parameters of the underlying L\'evy process is given by the recurrent relation \eqref{momenty} which is rather implicit. The main idea of our procedure for estimating the parameters of the process \(\xi\) can be described as follows. First, by making use of \eqref{momenty}, we estimate the Laplace exponent \(\phi(z)\) at the points \(z=u^{\circ}+\i v \in \Upsilon\), where \(u^{\circ}>0\) is fixed and \(v\) varies on the equidistant grid between \(\varepsilon V_{n}\) and \(V_{n}\) (with \(\varepsilon>0\) and \(V_{n} \to \infty\) as \(n \to \infty\)) Afterwards, we use the representation
\begin{eqnarray}
\label{estproc1}
\phi(u^{\circ}+\i v) &=& \a+ \c \left( u^{\circ}+\i v \right) - \mathcal{F}[\bar\nu](-v) , \qquad v \in I\!\!R,
\end{eqnarray}
where \(\bar\nu(dx):= e^{-u^{\circ} x} \nu(dx)\), and \(\mathcal{F}[\bar\nu] (v)\) stands for the Fourier transform of the measure \(\bar\nu\), i.e., \( \mathcal{F}[\bar\nu](v) := \int_{I\!\!R_{+}} e^{\i v x} \bar\nu(dx).\) Since \(\mathcal{F}[\bar\nu] (v) \to 0\) as \(v \to \infty\) by the Riemann-Lebesgue lemma, upon taking real and imaginary parts of the left and right hand sides of \eqref{estproc1}, we are able to consequently estimate the parameters \(\c\) and \(\a\). With no doubt, the second aim, a complete recovering of the L{\'e}vy measure \(\nu,\) is the most difficult task. Since the estimates of the parameters \(\c\) and \(\a\) are already obtained, we can estimate by \eqref{estproc1} the Fourier transform \(\mathcal{F}[\bar\nu] (v)\) of \(\bar\nu \) for \(v\) from \([-V_{n}, V_{n}]\). The last step of this procedure, the estimation of the L{\'e}vy measure \(\nu\), is based on the regularised inverse Fourier transform formula.
The above estimation algorithm bears some similarity to the spectral estimation algorithm introduced by Belomestny and Reiss \cite{DBReiss}, \cite{BelReiss}. Let us also mention that the problem of statistical inference for L\'evy processes (or some their generalizations) observed at low frequency was the subject of many studies, see, e.g. Neumann and Rei{\ss}~\cite{NeuReiss},
Rei{\ss}~\cite{TestingReiss}, Kappus~\cite{AdaptiveKappus}, Trabs~\cite{CalibrationTrabs} and Jongbloed et al. ~\cite{JMV}.
Note that the last reference deals with the L\'evy-driven Ornstein-Uhlenbeck processes, which are not of the form \eqref{GOU}.
\par
The paper is organized as follows. In the next section, we formulate our main assumptions and give some examples. In Section~\ref{estimation}, the main estimation algorithm is presented and discussed in details. Next, we analyze the convergence rates of the proposed algorithms in Section~\ref{sec4} and provide some numerical examples in Section~\ref{secsim}. The proofs of our theoretical results are collected in Section~\ref{theory}.
\section{Main setup}
\label{main_setup}
In this article, we study the class of subordinators with finite L{\'e}vy measures as possible choice for the L\'evy process $(\xi_{t})$. In terms of the L{\'e}vy triplet \((\c,\a,\nu)\), this means that
\begin{eqnarray}
\label{cond1}
\left\{
\begin{aligned}
\c\geq 0, &\qquad \sigma=0, \\
\nu(I\!\!R_{-}) = 0, &\qquad \a:= \nu(I\!\!R_{+}) < \infty.\\
\end{aligned}
\right.
\end{eqnarray}
A detailed discussion of the subordination theory as well as various examples of such processes (Gamma, Poisson, tempered stable, inverse Gaussian, Meixner processes, etc.), are given in \cite{BNS}, \cite{Bertoin}, \cite{ContTankov}, \cite{Sato}, \cite{Schoutens}. Note that in the case of subordinators, the truncation function in the L{\'e}vy-Khinchine formula can be omitted, that is, the characteristic exponent of \(\xi\) is equal to
\begin{eqnarray}
\label{phis}
\psi(z) = \log {\mathbb E} \left[
e^{\i z \xi_{1}}
\right]
=
\i \c z
+ \int_{0}^{\infty}
\left(
e^{\i z x} - 1
\right) \nu(dx).
\end{eqnarray}
Later on, we also need the Laplace exponent of \(\xi\), which is defined as \[\phi(z) := - \log {\mathbb E} \left[
e^{-z \xi_{1}}
\right] = -\psi\left(\i z \right).\] Under the assumption \eqref{cond1} the Laplace exponent \(\phi(\cdot)\) is given by
\begin{eqnarray}
\label{phis2}
\phi(z) &=&
\c z
+ \int_{0}^{\infty}
\left(
1 - e^{- z u}
\right) \nu(du) \\
\label{phis3}
&=&
\c z
+ z \int_{0}^{\infty}
e^{- z u} \nu\left( (u, +\infty) \right) du.
\end{eqnarray}
Let us summarise the main properties of the functional $A_{\infty}=\int_{0}^{\infty}e^{-\xi_{t}}dt$
in this case.
\begin{prop}
The random variable $A_{\infty}$ admits a bounded density $\pi$ and fulfills
$\mathrm{E}\left[A_{\infty}^{s-1}\right]<\infty$ for all $s>0.$
If $\mu>0$, then $0<A_{\infty}\leq1/\mu$ a.s. Moreover, the following
relation holds for $\mathrm{Re}[z]>0,$
\begin{eqnarray}
\nonumber
\phi(z)=
z\frac{\mathcal{M}(z)}{\mathcal{M}(z+1)} & = & \mu z+\int_{0}^{\infty}(1-e^{-zx})\nu(dx),\\
\label{Mz}
& = & \lambda+\mu z-\int_{0}^{\infty}e^{-zx}\nu(dx),
\end{eqnarray}
where \(\mathcal{M}(z)\) is the Mellin transform of \(\pi.\)
\end{prop}
\section{Estimation of the L{\'e}vy triplet}
\label{estimation}
In the sequel, we suppose that we are given by the observations \(X_{t_0},X_{t_1},\ldots, X_{t_n}\) of the process \eqref{GOU} at the equidistant time points \( 0=t_0<t_{1} < \ldots < t_{n}\) with \(t_j=j\cdot \Delta\) for some \(\Delta>0\). Assuming that the process \(X_{t}\) is stationary (see \cite{Behme}, \cite{Fasen}), we get that the random variables \(X_{k}:=X_{t_k}, \; k=1,\ldots, n,\) have all the same distribution, which coincides with the distribution of \(A_{\infty}\).
The first step of our estimation procedure consists in the estimation the Laplace exponent \(\phi(z)\) for \(z=u^{\circ}+\i v\), where \(u^{\circ}>0\) is fixed and \(v\) varies. An estimator of \(\phi(z)\) can be obtained from the recursive formula \eqref{momenty} for the Mellin transform of \(\pi\). In \cite{CPY}, this formula is proved for real positive \(z\) such that \(\phi(z)>0\) and \(\mathcal{M}(z+1)<\infty\). The case of complex \(z\) is considered in \cite{Kuznets}, where one can also find some generalizations of the formula \eqref{momenty} to the integrals with respect to Brownian motion with drift. In particular, applying Theorem 2 from \cite{Kuznets}, we get that \eqref{momenty} holds for any \(z \in \Upsilon\). In the situation when \((\xi_{t})\) is a subordinator, the set \(\Upsilon\) coincides with the positive half-plane (equivalently, the parameter \(\theta\) is equal to infinity) due to
\[
{\mathbb E}\left[
e^{-x \xi_{1}}
\right]
= - \phi (x) = - \c x - x \int_{I\!\!R_{+}} e^{-x u} \nu\left( (u, +\infty) \right) du < 0, \quad \forall \; x>0.
\]
Motivated by \eqref{Mz}, we first estimate the Mellin transform \(\mathcal{M}(z) \) via its empirical counterpart
\begin{eqnarray}
\label{step1}
\mathcal{M}_{n}(z) :
=
\frac{1}{n}
\sum_{k=1}^{n} X_{k}^{z-1}
\end{eqnarray}
and then define an estimate of the Laplace exponent \(\phi(z)\) by
\begin{eqnarray}
\label{step2
Y_{n} (z)
=
z \frac{\mathcal{M}_{n}(z)}{\mathcal{M}_{n}(z+1)}.
\end{eqnarray}
If the sequence \(X_1,\ldots,X_n\) has some mixing properties, then we can expect that \(Y_n(z)\to \phi(z)\) in probability.
\subsection{Estimation of $\a$ and $\c$}
Under our assumptions, the Laplace exponent of the L\'evy process \((\xi_t)\) can be represented in the form:
\begin{eqnarray}
\label{estproc}
\phi(u^{\circ} + \i v)
= \lambda+\mu \cdot \left( u^{\circ}+\i v \right) -\mathcal{F}[\bar\nu](-v)
, \qquad v \in I\!\!R,
\end{eqnarray}
where \(\lambda:=\int_{I\!\!R_{+}} \nu(dx)\) and \(\bar\nu(dx):= e^{-u^{\circ} x} \nu(dx)\). The general idea of the procedure described below is to estimate the Laplace exponent \(\phi(\cdot)\) at the points \(z=u^{\circ}+\i v\) with \(v\in \mathbb{R}\) and then use the relation \eqref{estproc} for the estimation of parameters. By the Riemann-Lebesque lemma, \(\mathcal{F}[\bar\nu] (-v) \to 0\) as \(v \to +\infty\) (see, e.g., \cite{Kawata}) and we conclude from \eqref{estproc} that \( \phi(u^{\circ}+\i v)\) is approximately (at least for large \(v\)) a linear function in \(v\) with the slope \(\c\) and the intercept term \(\a\). This observation suggests that a properly weighted least-squares approach can be applied to estimate \(\c\) and \(\a\). Let $V_{n}$ be a sequence of positive real numbers and $w(\cdot)$ be a
nonnegative weight function supported on $[0,1].$ Define a scaled weight function
$w_{n}(v)=V_{n}^{-1}w(v/V_{n})$ and introduce the estimators of the parameters \(\lambda\) and \(\mu\) as the solution of the following optimization problem:
\begin{eqnarray*}
(\lambda_{n},\mu_{n}): & = & \operatornamewithlimits{arg\,min}_{(\lambda,\mu)}\int_{0}^{\infty}w_{n}(v)\left|Y_{n}(u^{\circ}+\i v)-\mu\cdot (u^{\circ}+\i v)-\lambda\right|^{2}\, dv\\
& = & \operatornamewithlimits{arg\,min}_{(\lambda,\mu)}\int_{0}^{\infty}w_{n}(v)\left\{ \left|\mathrm{Im}\left[Y_{n}(u^{\circ}+\i v)\right]-\mu v\right|^{2}
\right. \\
&& \hspace{4cm} \left.
+ \left|\mathrm{Re}\left[Y_{n}(u^{\circ}+\i v)\right]-\lambda-\mu u^{\circ}\right|^{2}\right\} \, dv
\end{eqnarray*}
with $Y_{n}(z)$ defined in \eqref{step2}. The above optimisation problem admits an explicit solution given by
\begin{eqnarray*}
\mu_{n} & = & \frac{\int_{0}^{\infty}w_{n}(v)\mathrm{Im}\left[Y_{n}(u^{\circ}+\i v)\right]\, dv}{\int_{0}^{\infty}vw_{n}(v)\, dv}=\int_{0}^{\infty}w_{\mu,n}(v)\mathrm{Im}\left[Y_{n}(u^{\circ}+\i v)\right]\, dv\\
\lambda_{n} & = & \frac{\int_{0}^{\infty}w_{n}(v)\mathrm{Re}\left[Y_{n}(u^{\circ}+\i v)\right]\, dv}{\int_{0}^{\infty}w_{n}(v)\, dv}-\mu_{n}u^{\circ}\\\ && \hspace{4cm} =
\int_{0}^{\infty}w_{\lambda,n}(v)\mathrm{Re}\left[Y_{n}(u^{\circ}+\i v)\right]\, dv-\mu_{n}u^{\circ}
\end{eqnarray*}
with $w_{\mu,n}(v):=V_{n}^{-2} w_{\mu}(v/V_{n}) $ and $w_{\lambda,n}(v):=V_{n}^{-1}w_{\lambda} (v/V_{n})$,
where
\[
w_{\mu}(\cdot)= c_{1,w}^{-1} w(\cdot), \qquad
w_{\lambda}(\cdot) = c_{0,w}^{-1} w(\cdot), \qquad
c_{i,w}=\int_{0}^{1}v^{i}w(v)\, dv,\quad i=0,1.
\]
Taking into account the definition of the weight function \(w_{n}(\cdot)\), we get also some equivalent representations of the estimators \(\mu_{n}\) and \(\lambda_{n}\)
\begin{eqnarray*}
\c_{n} &=&
\operatornamewithlimits{arg\,min}_{\c} \int_{\varepsilon}^{1} w(\alpha)
\Bigl(
\Im [Y_{n} (u+\i \alpha V_{n})]
-
\c \alpha V_{n}
\Bigr)^{2} d\alpha \\
\a_{n} &:=&
\operatornamewithlimits{arg\,min}_{\a}\int_{\varepsilon}^{1} w(\alpha)
\Bigl(
\Re [Y_{n} (u+\i \alpha V_{n})]
-
\c_{n} u
-
\a
\Bigr)^{2}
d\alpha.
\end{eqnarray*}
In practice, we need to replace the above integrals by sums. To this end, let the numbers \(\alpha_{1}, \ldots, \alpha_{M}\) constitute an equidistant grid on the set \([\varepsilon, 1]\) for some \(\varepsilon>0.\) We
estimate the Mellin transform \(\mathcal{M}(z)\) for all \(z\in \{u^{\circ}+\i \alpha_{m} V_{n}, \; m=1,\ldots, M\}\) and \(z\in \{u^{\circ}-1+\i \alpha_{m} V_{n}, \; m=1,\ldots, M\}\)
and so get the estimates of the Laplace exponent at the discrete points \(z=u^{\circ}+\i \alpha_{m} V_{n}\) (see above). Now we define an estimate of the parameter \(\c\) via
\begin{eqnarray}
\label{hatcopt}
\hat{\c}_{n} &:=& \operatornamewithlimits{arg\,min}_{\c} \sum_{m=1}^{M} w(\alpha_{m})
\Bigl(
\Im [Y_{n} (u^{\circ}+\i \alpha_{m} V_{n})]
-
\c \alpha_{m} V_{n}
\Bigr)^{2}\\
\label{hatc}
&=&
\frac{
\sum_{m=1}^{M}
w(\alpha_{m}) \alpha_{m} \: \Im [Y_{n} (u^{\circ}+\i \alpha_{m} V_{n})]
}
{
V_{n} \: \cdot \: \sum_{m=1}^{M}
w(\alpha_{m}) \alpha_{m}^{2}
}.
\end{eqnarray}
Afterwards, we estimate the parameter \(\a\) by
\begin{eqnarray}
\label{hataopt}
\hat{\a}_{n} &:=&
\operatornamewithlimits{arg\,min}_{\a} \sum_{m=1}^{M} w(\alpha_{m})
\Bigl(
\Re [Y_{n} (u^{\circ}+\i \alpha_{m} V_{n})]
-
\hat{\c}_{n} u - \a
\Bigr)^{2}\\
\label{hata}
&=&
\frac{
\sum_{m=1}^{M} w (\alpha_{m}) \Re [Y_{n}(u^{\circ}+\i \alpha_{m} V_{n})]
}
{
\sum_{m=1}^{M}
w (\alpha_{m})
}
-\hat{\c}_{n} u^{\circ}.
\end{eqnarray}
The whole algorithm is described below.\vspace{0.5cm}
\begin{bclogo}[couleur=blue!15, logo=\bccrayon]
{Algorithm 1: Estimation of \(\a\) and \(\c\)}
\begin{algorithm}[H]
\SetAlgoLined
\textbf{Data:}
\(n\) observations \(X_{1},\ldots ,X_{n}\) of the GOU process \((X_t)\) observed at equidistant grid \(j\cdot \Delta,\) \(j=1,\ldots, n.\)
\vspace{0.3cm}
\textbf{Initiate:}
Fix \(V_{n} \to \infty\), \(\varepsilon \in (0,1)\) and \(u^{\circ} > -1\).
\\
Set \(\alpha_{j}=\varepsilon+j\cdot\left( 1- \varepsilon\right)/M,\) \(j=1,\ldots,M.\)
\\
Fix a function \(w (\cdot) \geq 0\) supported on \([\varepsilon,1]\).
\\
Denote \(v_{m,n}:= \alpha_{m} V_{n}\).
\vspace{0.3cm}
\textbf{Algorithm:}
\begin{enumerate}
\item
Estimate the Mellin transform \(\mathcal{M}(z) := {\mathbb E} \left[ A_{\infty}^{z-1} \right]\) \\
for \(z\in\{u^{\circ}+\i v_{m,n},1+u^{\circ}+\i v_{m,n},\, m=1,\ldots,M\}\) via
\begin{eqnarray*}
\mathcal{M}_{n}(z) =
\frac{1}{n}
\sum_{k=1}^{n} X_{k}^{z-1}.
\end{eqnarray*}
\item Estimate the Laplace exponent
\(\phi(z) := - \log {\mathbb E} \left[
e^{-z \xi_{1}}
\right]\) \\
at the points \(z\in\{u^{\circ}+\i v_{m,n},\, m=1,\ldots,M\}\) by
\begin{eqnarray*}
Y_{n} (z)
=
z \frac{\mathcal{M}_{n}(z)}{\mathcal{M}_{n}(z+1)}.
\end{eqnarray*}
\item[3.] Estimate \(\c\) by
\begin{eqnarray*}
\label{opt}
\c_{n} :=
\frac{
\sum_{m=1}^{M}
w(\alpha_{m}) \alpha_{m} \: \Im [Y_{n} (u^{\circ}+\i v_{m,n})]
}
{
V_{n} \: \cdot \: \sum_{m=1}^{M}
w(\alpha_{m})\, \alpha_{m}^{2}
}.
\end{eqnarray*}
\item[4.] Estimate \(\a\) by
\begin{eqnarray*}
\a_{n} :=
\frac{
\sum_{m=1}^{M} w (\alpha_{m}) \Re [Y_{n}(u^{\circ}+\i v_{m,n})]
}
{
\sum_{m=1}^{M}
w (\alpha_{m})
}
-\c_{n} u^{\circ}.
\end{eqnarray*}
\end{enumerate}
\end{algorithm}
\end{bclogo}
\subsection{Estimation of the L{\'e}vy measure $\nu$}
As a result of Algorithm~1, we obtain the estimates \(\c_{n}\) and \(\a_{n}\) of the parameters \(\c\) and \(\a,\) respectively. Based on \eqref{estproc}, we first define an estimate for the Fourier transform of \(\bar\nu\) via
\begin{eqnarray}
\label{step5}
\hat\mathcal{F}[\bar\nu](-v) = - Y_{n}(u^{\circ}+\i v) + \c_{n} \cdot (u^{\circ}+\i v) + \a_{n}.
\end{eqnarray}
Next we estimate the measure \(\nu\) by a regularised Fourier inversion formula
\begin{eqnarray}
\label{step6}
\nu_n(x) &=& \frac{e^{u^{\circ} x}}{2 \pi} \int_{I\!\!R} e^{\i v x } \hat\mathcal{F}[\bar\nu](-v) \mathcal{K} (-v / V_{n})\, dv,
\end{eqnarray}
where \(\mathcal{K}\) is a regularizing symmetric kernel supported on \([-1,1]\). Note that with a slight abuse of notation, we use \(\nu\) also for the density of the L{\'e}vy measure, and \(\nu_{n}\) for an estimate of this density. In what follows, we also use the notation \(\bar{\nu}_{n} = e^{-u^{\circ} x} \nu_{n}\). The formal description of the algorithm is given below.
\begin{bclogo}[couleur=blue!15, logo=\bccrayon]
{Algorithm 2: Estimation of \(\nu\)}
\begin{algorithm}[H]
\SetAlgoLined
\textbf{Data:}
\(n\) observations \(X_{1},\ldots ,X_{n}\) of the GOU process \((X_t)\) observed at equidistant grid points \(j\cdot \Delta,\) \(j=1,\ldots, n.\)
\vspace{0.3cm}
\textbf{Initiate:}
Fix \(V_{n} \to \infty\) and \(u^{\circ} > -1\).
\\
Set \(\alpha_{m}=-1+2\cdot j/M,\) \(m=0,\ldots,M.\)
\\
Fix a regularizing kernel \(\mathcal{K}\) supported on \([-1,1]\).\\
Denote \(v_{m,n}:= \alpha_{m} V_{n}\).
\vspace{0.3cm}
\textbf{Algorithm:}
\begin{enumerate}
\item[1-2] The first two steps coincide with ones of Algorithm 1.
\item[3.] Estimate \(\mathcal{F}[\bar\nu](-v_{m,n}) \) for \(\bar{\nu} (dx) =e^{-u^{\circ} x} \nu (dx)\) by
\begin{eqnarray*}
\hat\mathcal{F}[\bar\nu](-v_{m,n}) = - Y_{n}(u+\i v_{m,n}) + \c_{n} \cdot (u+\i v_{m,n}) + \a_{n}
\end{eqnarray*}
for \(m=0,\ldots,M.\)
\item[4.]
Estimate \(\nu\) by
\begin{eqnarray*}
\bar{\nu}_n(x) &=& e^{u^{\circ} x} \frac{1}{2\pi\cdot (1+M) } \sum_{m=0}^{M}{e^{\i v_{m,n} x }} \hat\mathcal{F}[\bar\nu](-v_{m,n}) \mathcal{K} (\alpha_{m}).
\end{eqnarray*}
\end{enumerate}
\end{algorithm}
\end{bclogo}
\begin{rem}
It is a worth mentioning that the estimation Algorithms 1 and 2 can be applied to a more general situation when
\begin{eqnarray}
\label{xi}
\xi_{t} =
\c t + \tau_{t},
\end{eqnarray}
where the process \(\tau_{t}\) is a difference between two subordinators, i.e., \(\tau_{t} = \tau^{+}_{t} + \tau^{-}_{t}\), and \(\tau^{+}\) and \(\tau^{-}\) are the processes of finite variation with L{\'e}vy measures \(\nu^{+}\) and \(\nu^{-}\) concentrated on \(I\!\!R_{+}\) and \(I\!\!R_{-},\) respectively. In fact, in this case, the formula \eqref{estproc} still holds with \[\nu(dx) = I\!\!I \{x>0\} \nu^{+} (dx) + I\!\!I \{x<0\} \nu^{-}(dx).\]
Therefore, the consequent estimation of \(\c\), \(\a\) and the Fourier transform of the measure \(e^{-u^{\circ} x} \nu(dx)\), as well as the estimation of \(\nu\) are still possible.
\end{rem}
\section{Convergence}
\label{sec4}
In order to analyse the convergence properties of the estimates \(\mu_n,\) \(\lambda_n\) and \(\nu_n\) we need to further specify the class of L\'evy processes \((\xi_t).\)
\label{further}
\begin{defi}
For $s\in\mathbb{N}\cup \{0\}$ and $R>0,$ let ${\cal G}(s,R)$
denote the set of all L\'evy triplets $(\mu,0,\nu)$, such
that $\nu$ is supported on $\mathbb{R}_{+}$
and
\begin{eqnarray}
\label{cond2}
\max\left\{\nu(\mathbb{R}_{+}), \int_{I\!\!R} |v|^{2 s} \left|
\mathcal{F}[\bar\nu] (v)
\right|^{2}\, dv\right\} \leq R,
\end{eqnarray}
where \(\bar{\nu}(dx)=e^{-u^{\circ}x}\nu(dx).\)
\end{defi}
Note that if \eqref{cond2} holds, then \(\bar\nu\) is $s$-times (weakly) differentiable
with
\begin{eqnarray}
\label{cond3}
\bigl\Vert \bar{\nu}^{(s)} \bigr\Vert _{\infty}\leq \frac{1}{2\pi}\int_{I\!\!R} |v|^s \left| \mathcal{F}[\bar\nu] (-v) \right|\, dv<\infty.
\end{eqnarray}
\par
It turns out that the convergence rates of the estimates \(\mu_n,\) \(\lambda_n\) and \(\nu_n\) crucially depend on the asymptotic behaviour of the Mellin transform of \(A_\infty.\) In order to specify this behaviour, let us fix some $u^{0}>0$ and introduce two classes of probability densities:
\begin{eqnarray}
\label{P}
\mathcal{P}(\beta,L) & := & \left\{ p:\,\liminf_{|v|\to\infty}\left[|v|^{\beta}\left|\mathcal{M}[p](u^{\circ}+\i v)\right|\right] \geq L\right\} ,\\
\label{E}
\mathcal{E}(\alpha,L) & := & \left\{ p:\,\liminf_{|v|\to\infty}\left[e^{\alpha|v|}\left|\mathcal{M}[p](u^{\circ}+\i v)\right|\right] \geq L\right\},
\end{eqnarray}
where \(\alpha, \beta \in I\!\!R,\) \(L>0\) and for any density \(p,\) \(\mathcal{M}[p]\) stands for the Mellin transform of \(p.\)
Before we formulate the main convergence results, let us look at some examples.
\begin{ex} Consider the class of L{\'e}vy processes with \(\c=0,\) \(\sigma=0\) and the L{\'e}vy density \(\nu\) of the form
\[
\nu(x) = \sum_{j=1}^{N} \left[\sum_{k=1}^{{m_{j}}} g_{jk} x^{k-1}\right] e^{-\rho_{j}x} \cdot I\!\!I\{x>0\}
\]
with \(N, m_{j} \in {\mathbb N}\), \(\rho_{j}>0\), \(g_{jk}>0\). First note that the assumption \eqref{cond1} obviously holds. Let us now check \eqref{cond2}. We can apply the well-known Erd{\'e}lyi lemma to derive
\[
\int_{I\!\!R_{+}} x^{k-1} f(x) e^{i v x} dx \asymp c_{1} v^{-k}, \qquad v \to \infty
\]
for any exponentially decaying and smooth function \(f\) on \(I\!\!R_+\), and some complex \(c_{1}\) depending on \(f\). Therefore, we conclude that
\begin{multline*}
\left| \mathcal{F}[\bar\nu] (-v) \right|
=
\left|
\sum_{j=1}^{N} \sum_{k=1}^{{m_{j}}} \alpha_{jk}
\int_{I\!\!R_{+}} x^{k-1} f_{j} (x)
e^{i v x}
dx
\right|
\asymp c_{2} v^{-k^{*}}, \\
\mbox{where} \quad
f_{j}(x)= e^{- (\rho_{j}+u^{\circ}) x},
\quad k^{*}:=\operatornamewithlimits{arg\,min}_{k}
\left\{
\exists \: j: \; \alpha_{jk} \ne 0
\right\},
\end{multline*}
where \(c_{2}>0\) depends on \(u^{\circ}\). Hence for any \(s<k^{*}-1\), the condition \eqref{cond2} holds for some \(R>0\).
Furthermore, taking into account the asymptotic behaviour of the Gamma function (see, e.g., formula~8.328 from \cite{Jeffrey}):
\begin{eqnarray}
\label{GR}
\left| \Gamma(u+\i v) \right|= \exp\left\{
-\frac{\pi}{2} v +\left(u - \frac{1}{2} \right) \ln v
\right\} \cdot \sqrt{2 \pi} \left( 1 + o(1) \right), \qquad v \to \infty,
\end{eqnarray}
we derive
\begin{eqnarray*}
\left|
\mathcal{M}(u^{\circ} + \i v )
\right|
\asymp \sqrt{2 \pi} A^{1- u^{\circ}} \exp \left\{
- \frac{\pi}{2} v +
\left(
u^{\circ} - \frac{1}{2} + \sum_{j=1}^{N} \rho_{j}m_{j} + \sum_{j=1}^{K} \Re(\zeta_{j})
\right) \ln v
\right\},
\end{eqnarray*}
where \(\zeta_{1},\ldots, \zeta_{K}\) are the roots of the equation
\begin{eqnarray*}
\sum_{j=1}^{N} \sum_{k=1}^{m_{j}} \frac{
g_{jk} (k-1)!
}{
\left( \rho_{j} +z \right)^{k}
}= \lambda - \mu z,
\end{eqnarray*}
see \cite{Kuznets2}.
Therefore, for any \(u^{\circ}>1/2\), we conclude that \( \pi \in \mathcal{E}(\pi /2 , L)\) with any \(L>0\).
\end{ex}
\begin{ex}\label{exx} Next, we provide an example of a L{\'e}vy process \(\xi_t\) with \(A_\infty=\int_0^\infty e^{-\xi_t}\,dt\) having a density from \(\mathcal{P}(\beta,L)\). Consider a subordinator \(\mathcal{T}\) with drift \(\c>0\) and the L{\'e}vy density
\begin{eqnarray*}
\nu(x)= ab \exp\{-bx\}\: I\{x>0\}, \quad a, b >0.
\end{eqnarray*}
The exponential functional \(A_\infty\) of the process \((\xi_t)\) has a density of the form
\begin{eqnarray*}
\pi(x) = C_{1} x^{b} (1- \c x)^{(a/\c) -1 } \: I\{0< x<1/\c\}
\end{eqnarray*}
with some \(C_{1}>0\), see \cite{CPY}. In other words, \(A_{\infty}\) has the same distribution as \(\xi/\c\), where the r.v. \(\xi\) has the Beta distribution with parameters \(\alpha=b+1\) and \(\beta = a/ \c = \lambda/ \c\). The Mellin transform of the function \(\pi(x)\) in the half-plane \(\Re(s)>-\alpha\) is hence given by
\begin{eqnarray*}
\mathcal{M}(z) = \frac{{\mathbb E} \left[
\xi^{z-1}
\right]}{\c^{z-1}}
&=&
\frac{1}{\c^{z-1}}
\frac{
B (z+\alpha - 1, \beta)
}{
B (\alpha, \beta)
}\\
&=&
\frac{
\Gamma (\alpha + \beta)
}{
\Gamma (\alpha)
}
\cdot
\frac{1}{\c^{z-1}}
\frac{
\Gamma (z+\alpha - 1)
}{
\Gamma (z+\alpha+\beta - 1)
}.
\end{eqnarray*}
Using \eqref{GR}, we conclude that the Mellin transform of \(A_\infty\) has a polynomial decay in this case. More precisely,
\[ |\mathcal{M}(u^{\circ}+ \i v)| \asymp L\cdot |v|^{-\lambda/\c} \quad \mbox{ with } \quad L=\c^{-u^{\circ}+1} \frac{\Gamma(\lambda/\mu+b+1)}{\Gamma(b+1)},\]
as \(|v|\to\infty\) and therefore \(\pi \in \mathcal{P}(\lambda/ \mu, L)\).
\end{ex}
Let us now formulate the main result concerning the convergence of the estimates \(\mu_{n}\) and \(\lambda_{n}\).
\begin{thm}[upper bounds for $\c_{n}$ and $\a_{n}$]
\label{corub}
Let \((\xi_{t})\) be a L{\'e}vy process with a triplet from $\mathcal{G} (s, R)$. Suppose that the sequence $X_{0},X_{1},\ldots,X_{n}$ is $\alpha$-mixing and strictly stationary. Denote the \(\alpha\)-mixing coefficients of the sequence $X_{0},X_{1},\ldots,X_{n}$ by \(\alpha(s)\).
\begin{enumerate}[(i)]
\item Assume that the density \(\pi\) of \(A_\infty\) belongs to \(\mathcal{P}(\beta,L)\) with some \(\beta \in I\!\!R\) and \(L>0\), and moreover
\begin{eqnarray}
\label{expcond}
\alpha(j)\lesssim e^{-j\alpha^{*}}, \quad j\in\mathbb{N}, \quad\mbox{for some} \quad \alpha^{*} \geq 0.
\end{eqnarray}
Then the quadratic risks of the estimates \(\mu_n\) and \(\lambda_n,\) under the choice \(V_{n}=n^{1/(2\beta+2s+3)},\) satisfy the following asymptotic relations
\[
\mathrm{E}\left[\left|\mu_{n}-\mu\right|^{2}\right]\lesssim n^{-2(s+2)/(2\beta+2s+3)}\log(n)
\]
and
\[
\mathrm{E}\left[\left|\lambda_{n}-\lambda\right|^{2}\right]\lesssim n^{-2(s+1)/(2\beta+2s+3)}\log(n),
\]
as \(n\to\infty.\)
\item
If \(\pi\in \mathcal{E}(\alpha, L)\) and
\begin{eqnarray}
\label{expcond_pol}
\alpha(j)\lesssim j^{-\alpha^{*}},\quad j\in\mathbb{N}, \quad \mbox{for some} \quad \alpha^{*} \geq 2,
\end{eqnarray}
then the choice
\[V_{n}=\frac{1}{2\alpha}\log(n)-\frac{s+2}{\alpha}\log(\log(n)),\] leads to the rates
\[
\mathrm{E}\left[\left|\mu_{n}-\mu\right|^{2}\right]\lesssim\log^{-2(s+2)}(n),
\]
\[
\mathrm{E}\left[\left|\lambda_{n}-\lambda\right|^{2}\right]\lesssim\log^{-2(s+1)}(n).
\]
\end{enumerate}
\end{thm}
\begin{proof}
Proof is given in Section~\ref{upplmu}.
\end{proof}
In a similar way, we can establish the upper bounds for the risk of \(\bar\nu_{n}\). In the theorem formulated below, the quality of the estimate \(\bar\nu_{n}\) is measured in terms of the mean integrated squared error (MISE)
\begin{eqnarray*}
\textrm{MISE} (\bar{\nu}_{n}) &:=&
{\mathbb E} \left[
\int_{I\!\!R} \left|
\bar{\nu}_{n}(x) - \bar\nu(x)
\right|^{2} dx
\right].
\end{eqnarray*}
\begin{thm}[upper bounds for $\bar\nu_{n}$]
\label{nuub}
Let the assumptions of Theorem~\ref{corub} be fulfilled and let \(\mathcal{K}(\cdot)\) be a kernel satisfying
\begin{eqnarray}
\label{assk}
|1 - \mathcal{K}(x)| \leq A |x|^{s}, \qquad \forall x \in I\!\!R \setminus \{0\}
\end{eqnarray}
with some \(A>0.\)
\begin{enumerate}[(i)]
\item Assume that the density of \(A_\infty\) belongs to \(\mathcal{P}(\beta,L)\) with some \(\beta \in I\!\!R\) and \(L>0\), and moreover
\begin{eqnarray*}
\alpha(j)\lesssim e^{-j\alpha^{*}},\quad j\in\mathbb{N}, \quad \mbox{for some} \quad \alpha^{*} > 0.
\end{eqnarray*}
Then under the choice \(V_{n}=n^{1/(2\beta+2s+3)}\), the MISE of the estimator \(\bar\nu_n\) is bounded as follows:
\[
\textrm{MISE} (\bar{\nu}_{n}) \lesssim n^{-2s/(2\beta+2s+3)},\quad n\to\infty.
\]
\item
If the density of \(A_\infty\) belongs to the class \(\mathcal{E}(\alpha, L)\) and
\begin{eqnarray*}
\label{expcond_pol}
\alpha(j)\lesssim j^{-\alpha^{*}},\quad j\in\mathbb{N}, \quad \mbox{for some} \quad \alpha^{*} \geq 2,
\end{eqnarray*}
then under the choice
\[V_{n}=\frac{1}{2\alpha}\log(n)-\frac{s+2}{\alpha}\log(\log(n))\]
we have
\[
\textrm{MISE} (\bar{\nu}_{n}) \lesssim\log^{-2 s}(n),\quad n\to\infty.
\]
\end{enumerate}
\end{thm}
\begin{proof}
Proof is given in Section~\ref{uppmise}.
\end{proof}
The next theorem shows that the rates obtained in the previous theorem are optimal up to a logarithmic factor.
\begin{thm}[lower bounds for $\bar\nu_{n}$]
\label{thm3}
Fix some \(s \in {\mathbb N}\cup\{0\},\) \(R>0,\) \(\alpha>0,\) \(\beta>0,\) \(L>0\) and define
\begin{eqnarray*}
\varphi_{n}(\pi):=\varphi_{n}(\pi,\rho)=
\begin{cases}
n^{s/(2\beta+2s+3)}\log^{-\rho}(n) , &\text{if $\pi \in \mathcal{P}(\beta,L)$,}\\
\log^{ s}(n), &\text{if $\pi \in \mathcal{E}(\alpha,L)$,}
\end{cases}
\end{eqnarray*}
for any \(\rho>0\) and any probability density \(\pi \in \mathcal{P}(\beta,L)\cup \mathcal{E}(\alpha,L), \)
Then for some \(\rho^{*}>0\), it holds
\begin{eqnarray}
\label{lowerbound}
\inf_{\bar{\nu}_{n}} \sup_{\substack{\mathcal{T}\in\mathcal{G}(s,R)\\ \pi_{\mathcal{T}} \in \mathcal{P}(\beta,L)\cup \mathcal{E}(\alpha,L)} }
\left\{
\varphi_{n}^{2}(\pi_{\mathcal{T}}, \rho^{*})\cdot
{\mathbb E}_{\pi^{\otimes n}_{\mathcal{T}}} \left[
\int_{I\!\!R} \left|
\bar{\nu}_{n}(x) - \bar\nu(x)
\right|^{2} dx
\right]
\right\} > 0,
\end{eqnarray}
where the infimum is taken over all possible estimates \(\bar{\nu}_{n}\) of the function \(\bar\nu\) based on i.i.d. sample \(X_1,\ldots,X_n\) from the distribution \(\pi_{\mathcal{T}}\) of \(A_{\infty}:= \int_{0}^{\infty} e^{-\xi_{t}} \; dt\) such that the L\'evy triplet \(\mathcal{T}\) of \((\xi_t)\) belongs to \(\mathcal{G}(s,R).\)
\end{thm}
\begin{proof}
Proof is given in Section~\ref{prooflb}.
\end{proof}
An important condition of Theorems ~\ref{corub} and \ref{nuub} is \eqref{expcond}, which means that the sequence \(X_{0}, X_{1},\ldots, X_{n}\) is exponentially \(\alpha\)-mixing. Since \(\beta\)-mixing coefficient between two sigma-algebras is larger than or equal to the corresponding \(\alpha\)-mixing coefficient, it is sufficient to show that \(X_{0}, X_{1},\ldots, X_{n}\) is an exponentially \(\beta\)-mixing sequence (see Section~1.1 from \cite{Bosq}). For the case of the GOU processes \eqref{GOU}, the latter question was addressed in \cite{Fasen}. The sufficient conditions for exponential \(\beta\)-mixing given in \cite{Fasen} are:
\begin{enumerate}
\item the distribution of \(A_{\infty}\) has a Pareto-like asymptotic behaviour, that is,
\begin{eqnarray*}
\P \left\{
A_{\infty}>x
\right\}
\asymp C x^{-\alpha}
\quad \mbox{as}\quad x \to \infty
\end{eqnarray*}
with some \(\alpha>0\) and \(C>0;\)
\item there exist \(A>0\), \(B>A\) and \(h>0\) such that \(\psi(A)=0, \psi(B)<\infty\) with \(\psi\) given in \eqref{phis}, and
\begin{eqnarray*}
{\mathbb E} \left|
e^{-\xi_{h}}\int_{0}^{h}e^{\xi_{u-}} du
\right|^{B} <\infty.
\end{eqnarray*}
\end{enumerate}
As it is proved in \cite{LindnerMaller}, both conditions are guaranteed by the positiveness of \(\mu\) and the existence of a positive zero of the function \(\psi(\cdot)\). We refer also to \cite{Lee} for some further results in this direction.
\section{Simulation study}
\label{secsim}
\sec{Example 1.}
Consider the subordinator \(\tau_{t}\) with the L{\'e}vy density
\begin{eqnarray}
\label{nu}
\nu(x)= ab \exp\{-bx\}\: I\!\!I\{x>0\}, \quad a, b >0.
\end{eqnarray}
Note that in this case, \(\lambda=\int_{I\!\!R_{+}} \nu (u) du = a\).
Define a L\'evy process
\begin{eqnarray}
\label{xiex}
\xi_{t}=\mu t + \sigma W_{t} + \tau_{t},
\end{eqnarray}
where \(W_{t}\) is a Brownian motion. The Laplace exponent of \(\xi_{t}\) is given by
\begin{eqnarray}
\label{phiss2}
\phi(z) = z \left(
\c - \frac{1}{2} \sigma^{2} z + \frac{a}{b+z}
\right).
\end{eqnarray}
In \cite{CPY}, it is shown that the exponential functional \(A_{\infty}=\int_{0}^{\infty} e^{-\xi_{t}} \; dt\) is finite for any \(\mu\) and \(\sigma\), and moreover the density function \(\pi\) of \(A_{\infty}\) satisfies the following differential equation
\begin{multline}
-\frac{\sigma^{2}}{2} x^{2} \pi''(x) +
\left[
\left(
\frac{\sigma^{2}}{2} (3-b) +\c
\right)
x
-1
\right] \pi'(x) \\
+
\left[
\left(
1-b
\right)
\left(
\frac{\sigma^{2}}{2} + \c
\right)
-a+\frac{b}{x}
\right] \pi(x) = 0.
\label{eq}
\end{multline}
Some special cases are considered below:
\begin{enumerate}
\item In the case \(\c=0, \: \sigma=0\) (pure jump process), this equation has a solution
\begin{eqnarray}
\label{kx1}
\pi_{1}(x) = C x^{b} e^{-a x} \: I\{x>0\},
\end{eqnarray}
and therefore \(A_{\infty} \stackrel{d}{=} G(b+1, a)\), where \(G(\alpha, \beta)\) is a Gamma distribution with shape parameter \(\alpha\) and rate \(\beta\).
\item If \(\c>0, \: \sigma=0\) (pure jump process with drift), then
\begin{eqnarray}
\label{kx2}
\pi_{2}(x) = C x^{b} (1- \c x)^{(a/\c) -1 } \: I\{0< x<1/\c\}.
\end{eqnarray}
In this situation \(A_{\infty} \stackrel{d}{=} B(b+1, a/\c) / \c\), where \(B(\alpha, \beta)\) is a Beta - distribution.
\item In the case \(\c \ne 0, \: \sigma \ne 0\), the equation \eqref{eq} also allows for the closed form solution. Assuming for simplicity \(\sigma^{2}/2 =1\), \(\c=-(b+1)\), we get the solution of \eqref{eq} in the following form:
\begin{eqnarray}
\label{kx}
\pi_{3}(x)=C\: x^{b-1/2} \exp\left\{\frac{1}{2x}\right\} I_{\mu}\left(\frac{1}{2x}\right),
\end{eqnarray}
where we denote by \(I_{\mu}\) the modified Bessel function of the first kind, \(\mu=\sqrt{a+1/4}\), and the constant \(C\) is later chosen to guarantee the condition \(\int_{0}^{\infty}\pi_{3}(x) dx =1\). \vspace{0.5cm}
\end{enumerate}
For our numerical study, we assume that the data are generated from the distribution of \eqref{Ainfty}, where the process \((\xi_{t})\) is defined by \eqref{xiex} with
\(\c=1.8, \sigma=0,\) and the subordinator \(\tau_{t}\) in the form \eqref{nu} with \(a=0.7\), \(b=0.2\). A sample from the distribution of the integral \(A_{\infty}\) can be simulated from the corresponding Beta-distribution, see \eqref{kx2}.
In the first step, we estimate the Mellin transform \(\mathcal{M}(z)\) for \(z=u+\i v\) with \(u=u^{\circ}=29\) and \(u=u^{\circ}+1=30\) and \(v\) lying on the equidistant grid between \(-30\) and \(30\). Next, we estimate the Laplace exponent of \(\xi\) by the formula \eqref{step2}. Figure~\ref{plot1} graphically compares the proposed estimator of the Laplace exponent \(\phi(u^{\circ}+\i v)\) with its theoretical values \((\mu+a/(b+u^{\circ}+\i v))\cdot \left(u^{\circ}+\i v\right)\) .
\begin{figure}
\begin{center}
\includegraphics[width=0.6\linewidth ]{phi.pdf}\caption{Plots of theoretical (blue dashed) and empirical (red solid) Laplace exponents in Example 1. Real, imaginary parts and absolute values are presented. \label{plot1}
\end{center}
\end{figure}
\begin{figure}
\begin{center}
\includegraphics[width=0.6\linewidth ]{boxplot.pdf}\caption{Boxplots for the estimates of \(\c=c\) and \(\a=a\) for different sample sizes \(n\) based on 25 simulation runs.\label{plot2}
\end{center}
\end{figure}
Estimates for the parameters \(\c\) and \(\a=a\) are given in \eqref{hatc} and \eqref{hata}, respectively. The boxplots of this estimates based on 25 simulation runs are presented on Figure \ref{plot2}. \vspace{0.5cm}
\sec{Example 2.} Consider the compound Poisson process
\[
\xi_{t} = -\log q \left(\sum_{k=1}^{N_{t}} \eta_{k} \right),
\]
where \(q \in (0,1)\) is fixed, \(N_{t}\) is a Poisson process with intensity \(\lambda\) and \(\eta_{k}\) are i.i.d. random variables with a distribution \(\L\). The integral \(A_{\infty}\) admits the representation
\begin{eqnarray*}
A_{\infty} = \int_{0}^{\infty} q^{-\xi_{t}} dt = \sum_{n=0}^{\infty} q^{S_{n}} \left(T_{n+1} - T_{n}\right),
\end{eqnarray*}
where \(T_{n}\) is the jump time of \(N,\) i.e., \(T_{n} = \inf\left\{ t: N_{t} =n \right\}\), and \(S_{n}=\sum_{k=1}^{n}\eta_{k}.\) Note that if \(\eta_{k}\) take only positive values, then \(\xi_{t}\) is a subordinator. For the overview of the properties of the integral \(A_{\infty}\) in the particular case \(\eta_{k} \equiv 1\) (that is, \(\xi_{t}\) is a Poisson process up to a constant), we refer to \cite{BertoinYor}.
Fix some positive \(\alpha\) and consider the case when \(\L\) is the standard normal distribution truncated on the interval \((\alpha, +\infty)\). The density function of \(\L\) is given by \[p_{\L} (x)= p(x) / (1-F(\alpha)),\] where \(p(\cdot)\) and \(F(\cdot)\) are the density and the distribution functions of the standard Normal distribution. In this case, the Laplace exponent of \(\xi_{t}\) is equal to
\begin{eqnarray*}
\phi(z) = \lambda \left[
1 - \frac{
1-F\left(\alpha + (\log q) z\right)
}
{
1 - F\left(\alpha\right)
}\;
\exp \left\{
-
\frac{
\left(
\log q
\right)^{2}
z^{2}
}{2}
\right\}
\right],
\end{eqnarray*}
where the function \(F(\cdot)\) can be calculated for complex arguments from the error function:
\[
F(z) := \frac{1}{2} \left( \erf\left( \frac{z}{\sqrt{2}} \right) + 1 \right), \quad \mbox{where} \quad
\erf(z) = \frac{2}{\sqrt{\pi}}\int_{0}^{z} e^{-s^{2}} ds.
\]
In this example, we aim to estimate the L{\'e}vy measure of the process \(\-(\xi_{t})\), which is given by
\[
\nu(dx) = \frac{\lambda}{1-F(\alpha)} \: p(x) I\!\!I\{x>\alpha\} dx.
\]
For our numerical study, we take \(q=0.5, \alpha=0.1\), and \(\lambda=1\). First, we estimate the Laplace exponent by \eqref{step2}. The quality of the corresponding estimate at the complex points \(z=u^{\circ}+\i v\) with \(u^{\circ}=1\) and \(v \in [-5,5]\) can be visually seen in Figure~\ref{fig3}.
\begin{figure}
\begin{center}
\includegraphics[width=0.6\linewidth ]{phi2.pdf}\caption{Plots of theoretical (blue dashed) and empirical (red solid) Laplace exponents for Example 2. Graphs present real, imaginary and absolute values. For \(v \in[-3,3]\) the curves are visually indistinguishable. \label{fig3}
\end{center}
\end{figure}
\begin{figure}
\begin{center}
\includegraphics[width=0.6\linewidth ]{Levy.pdf}\caption{Left: plots of the L{\'e}vy density (blue dashed line) and its estimate \(\bar{\nu}_n(z)\) (red solid). Right: the imaginary part of the estimate \(\bar{\nu}_n(z)\) (red solid) and the line \(Y=0\) (blue dashed line). \label{fig4}
\end{center}
\end{figure}
Next, we proceed with the estimation of the Fourier transform of the measure \(\bar{\nu}(x):=e^{-u^{\circ} x} \nu(x)\)
by applying \eqref{step5}. For the last step of the Algorithm~2, i.e. the reconstruction of the L{\'e}vy measure by \eqref{step6}, we follow \cite{Belomest2011} and use the so-called flat-top kernel, which is defined as follows:
\begin{eqnarray*}
\mathcal{K}(x)=
\begin{cases}
1, & |x|\leq 0.05, \\
\exp\left( -\frac{e^{-1/(|x|-0.05)}}{1-|x|} \right), & 0.05<|x|<1,\\
0, & |x|\geq 1.
\end{cases}
\end{eqnarray*}
The quality of the resulting estimate \(\bar{\nu}_n\) is shown in Figure~\ref{fig4}.
\section{Proofs}\label{theory}
\subsection{Upper bounds for the quadratic risks of $\mu_{n}$ and $\lambda_{n}$} \label{upp}
\label{upplmu}
The next proposition is the main technical result for this section.
\begin{prop}
\label{propub}
Let \(\xi_{t}\) be a L{\'e}vy triplet from ${\cal G} (s,R)$. Suppose that the sequence $X_{0},X_{1},\ldots,X_{n}$ of observations of the exponential functional \(A_{\infty}:=\lim_{T\to\infty} A_{T}= \int_{0}^{\infty} e^{-\xi_{t}} \; dt\) is $\alpha$-mixing and strictly stationary. Denote the mixing coefficients of the sequence $X_{0},X_{1},\ldots,X_{n}$ by \(\alpha(s)\).
Then for
any $p\in\{0,1,\ldots,n\}$ we have
\begin{multline*}
\mathrm{E}\left[\left|\mu_{n}-\mu\right|^{2}\right] \lesssim
\frac{p}{n}\int_{0}^{\infty}\frac{|u^{\circ}+\i v|^{2}}{\left|\mathcal{M}(u^{\circ}+\i v+1)\right|^{2}}\left|w_{\mu,n}(v)\right|^{2}\, dv\\
+\sum_{j=p+1}^{n}\alpha(j)\left[\int_{0}^{\infty}\frac{\left|u^{\circ}+\i v\right|\left|w_{\mu,n}(v)\right|}{\left|\mathcal{M}_{n}(u^{\circ}+\i v+1)\right|}\, dv\right]^{2}
\\
+ \|\bar{\nu}^{(s)}\|_{\infty}^{2}\|\mathcal{F}^{-1}[w_{\mu,n}(\cdot)/(-i\cdot)^{s}]\|_{L^{1}}^{2},
\end{multline*}
\begin{multline*}
\mathrm{E}\left[\left|\lambda_{n}-\lambda\right|^{2}\right] \lesssim \frac{p}{n}\int_{0}^{\infty}\frac{|u^{\circ}+\i v|^{2}}{\left|\mathcal{M}(u^{\circ}+\i v+1)\right|^{2}}\left|w_{\lambda,n}(v)\right|^{2}\, dv\\+\sum_{j=p+1}^{n}\alpha(j)\left[\int_{0}^{\infty}\frac{\left|u^{\circ}+\i v\right|\left|w_{\lambda,n}(v)\right|}{\left|\mathcal{M}_{n}(u^{\circ}+\i v+1)\right|}\, dv\right]^{2}\\
+ \|\bar{\nu}^{(s)}\|_{\infty}^{2}\|\mathcal{F}^{-1}[w_{\lambda,n}(\cdot)/(-i\cdot)^{s}]\|_{L^{1}}^{2},
\end{multline*}
provided $\sum_{j=1}^{\infty}\alpha^{1-\epsilon}(j)<\infty$ for some
$\epsilon>0$ and the sequence $V_{n}$ satisfies
\begin{equation}
\sup_{v\in[0,V_{n}]}\frac{1}{\left|\mathcal{M}(u^{\circ}+\i v+1)\right|}=o(n^{1/2}).\label{eq:Vn_cond}
\end{equation}
\end{prop}
\begin{proof}
\textbf{1.} Denote $Y(z) :=\phi(z) = z \cdot \mathcal{M}(z) / \mathcal{M}(z+1),$ then
\[\mu=\int_{0}^{\infty}w_{\mu,n}(v)\Im\left[Y(u^{\circ}+\i v)\right]\, dv + \int_{0}^{\infty}w_{\mu,n}(v)\mathrm{Im}[\mathcal{F}[\bar{\nu}](-v)]\, dv\]
and we have
\begin{eqnarray*}
\mu_{n}-\mu & = & \int_{0}^{\infty}w_{\mu,n}(v)\mathrm{Im}\left[Y_{n}(u^{\circ}+\i v)-Y(u^{\circ}+\i v)\right]\, dv\\
&& \hspace{4cm} - \int_{0}^{\infty}w_{\mu,n}(v)\mathrm{Im}[\mathcal{F}[\bar{\nu}](-v)]\, dv\\
& = & \mathrm{Im}\left[\int_{0}^{\infty}w_{\mu,n}(v)S_{n}(u^{\circ}+\i v)\, dv\right] - \mathrm{Im}[D_{n}(u^{\circ})]
\end{eqnarray*}
with
\[
S_{n}(u^{\circ}+\i v)=Y_{n}(u^{\circ}+\i v)-Y(u^{\circ}+\i v),\quad D_{n}(u^{\circ})=
\int_{0}^{\infty}w_{\mu,n}(v) \mathcal{F}[\bar{\nu}](-v) dv.
\]
Note that
\begin{eqnarray*}
{\mathbb E} \left[ \left(
\mu_{n}-\mu
\right)^{2}
\right] \leq 2 \cdot
{\mathbb E}\left[
\left(
\mathrm{Im}\left[\int_{0}^{\infty}w_{\mu,n}(v)S_{n}(u^{\circ}+\i v)\, dv\right]
\right)^{2}
\right]
+
2
\left|
D_{n}(u^{\circ})
\right|^{2}.
\end{eqnarray*}
\textbf{2.} Since
\begin{eqnarray*}
\frac{S_{n}(z)}{z} & = & \frac{\mathcal{M}_{n}(z)}{\mathcal{M}_{n}(z+1)}-\frac{\mathcal{M}(z)}{\mathcal{M}(z+1)}\\
& = & \frac{\mathcal{M}_{n}(z)\mathcal{M}(z+1)-\mathcal{M}(z)\mathcal{M}_{n}(z+1)}{\mathcal{M}_{n}(z+1)\mathcal{M}(z+1)}\\
& = & \frac{\left[\mathcal{M}_{n}(z)-\mathcal{M}(z)\right]\mathcal{M}_{n}(z+1)-\left[\mathcal{M}_{n}(z+1)-\mathcal{M}(z+1)\right]\mathcal{M}_{n}(z)}{\mathcal{M}_{n}(z+1)\mathcal{M}(z+1)}\\
& = & \frac{\left[\mathcal{M}_{n}(z)-\mathcal{M}(z)\right]}{\mathcal{M}(z+1)}-\frac{Y_{n}(z)}{z}\,\frac{\left[\mathcal{M}_{n}(z+1)-\mathcal{M}(z+1)\right]}{\mathcal{M}(z+1)}\\
& = & \frac{\left[\mathcal{M}_{n}(z)-\mathcal{M}(z)\right]}{\mathcal{M}(z+1)}-\frac{S_{n}(z)}{z}\,\frac{\left[\mathcal{M}_{n}(z+1)-\mathcal{M}(z+1)\right]}{\mathcal{M}(z+1)}\\
& & -\frac{Y(z)}{z}\,\frac{\left[\mathcal{M}_{n}(z+1)-\mathcal{M}(z+1)\right]}{\mathcal{M}(z+1)},
\end{eqnarray*}
we get
\[
S_{n} \cdot (1+R_{2,n} )=-Y\cdot R_{2,n} +R_{1,n}
\]
with
\[
R_{1,n}(z)=z\,\frac{\left[\mathcal{M}_{n}(z)-\mathcal{M}(z)\right]}{\mathcal{M}(z+1)},\quad R_{2,n}(z)=\frac{\left[\mathcal{M}_{n}(z+1)-\mathcal{M}(z+1)\right]}{\mathcal{M}(z+1)}.
\]
Following the lines of the proof of Theorem~1.5 from \cite{Bosq}, we get
\begin{multline}
\label{bosq}
\mathrm{E}\left[\left|\mathcal{M}_{n}(z)-\mathcal{M}(z)\right|^{2}\right] = \frac{1}{n^{2}}
\sum_{0\leq k,j\leq n-1}\mathrm{Cov}\left(X_{k}^{z-1},X_{j}^{z-1}\right)\\
=
\frac{1}{n} \operatorname{Var}\left(X_{0}^{z-1} \right) +
\frac{2}{n}\sum_{k=1}^{n-1} \left( 1 - \frac{k}{n} \right)
\mathrm{Cov}\left(X_{0}^{z-1},X_{k}^{z-1}\right).
\end{multline}
Note that the sum in the last representation converges as \(n \to \infty\), because by Davydov's inequality
\begin{eqnarray}
\label{davydov}
\left|
\mathrm{Cov}\left(X_{0}^{z-1},X_{k}^{z-1}\right)
\right|
\leq
\frac{2r}{r-2}
\left( 2 \alpha(k) \right)^{(r-2)/r}
\left(
{\mathbb E} \left[
X_{0}^{\left( u^{\circ}-1 \right) r}
\right]
\right)^{2/ r},
\end{eqnarray}
and therefore the series \(\sum \left(X_{0}^{z-1},X_{k}^{z-1}\right)\) is convergent if \(r=2 / \varepsilon\).
We have $\mathrm{E}\left[\left|\mathcal{M}_{n}(u^{\circ}+\i v)-\mathcal{M}(u^{\circ}+\i v)\right|^{2}\right]\lesssim n^{-1}$
uniformly in $v\in\mathbb{R}.$ As a result
\[
\mathrm{E}\left[\left|R_{2,n}(u^{\circ}+\i v)\right|^{2}\right]\lesssim\frac{1}{n \cdot |\mathcal{M}(u^{\circ}+\i v+1)|^{2}}.
\]
The condition (\ref{eq:Vn_cond}) implies now that $\sup_{v\in[0,V_{n}]}\left|R_{2,n}(u^{\circ}+\i v)\right|^{2}=o_{P}(1).$
Furthermore, we have
\begin{multline*}
\mathrm{Var}\left[\int_{0}^{\infty}R_{1,n}(u^{\circ}+\i v)w_{\mu,n}(v)\, dv\right] \\= \int_{0}^{\infty}\int_{0}^{\infty}\frac{\mathrm{Cov}(\mathcal{M}_{n}(u^{\circ}+\i v_{1}),\mathcal{M}_{n}(u^{\circ}+\i v_{2}))}{\mathcal{M}(u^{\circ}+\i v_{1}+1)\overline{\mathcal{M}(u^{\circ}+\i v_{2}+1)}}\\ \cdot (u^{\circ}+\i v_{1})(u^{\circ}-\i v_{2})w_{\mu,n}(v_{1})\, w_{\mu,n}(v_{2})\, dv_{1} dv_{2}.
\end{multline*}
Similar to \eqref{bosq}, we consider a representation
\begin{multline*}
\mathrm{Cov}(\mathcal{M}_{n}(u^{\circ}+\i v_{1}),\mathcal{M}_{n}(u^{\circ}+\i v_{2})) \\
=\frac{1}{n}\left[g_{0}(v_{1},v_{2}) +
2\sum_{j=1}^{p}g_{j}(v_{1},v_{2})+2\sum_{j=p+1}^{n-1}g_{j}(v_{1},v_{2})\right],
\end{multline*}
where \(g_{j}(v_{1}, v_{2}) := (1 - j/n) \cdot \mathrm{Cov}\left(X_{0}^{u^{\circ}+\i v_{1}-1},X_{j}^{u^{\circ}+\i v_{2}-1}\right), \; j=0..(n-1)\). Applying once more Davydov's inequality, we get
\begin{eqnarray}
\label{davydov}
\left|
g_{j}(v_{1}, v_{2})
\right|
\leq
\frac{2r}{r-2}
\left( 2 \alpha(j) \right)^{(r-2)/r}
\left(
{\mathbb E} \left[
X_{0}^{\left( u^{\circ}-1 \right) r}
\right]
\right)^{2/ r},
\end{eqnarray}
Now using the Cauchy-Schwarz inequality we get
\begin{eqnarray*}
\mathrm{Var}\left[\int_{0}^{\infty}R_{1,n}(u^{\circ}+\i v)w_{\mu,n}(v)\, dv\right] & \lesssim & p\int_{0}^{\infty}\frac{\left|u^{\circ}+\i v\right|^{2}\left|w_{\mu,n}(v)\right|^{2}}{\left|\mathcal{M}(u^{\circ}+\i v+1)\right|^{2}}\, dv\\
& &+\sum_{j=p+1}^{n}\alpha(j)\left[\int_{0}^{\infty}\frac{\left|u^{\circ}+\i v\right|\left|w_{\mu,n}(v)\right|}{\left|\mathcal{M}(u^{\circ}+\i v+1)\right|}\, dv\right]^{2}.
\end{eqnarray*}
Finally using the fact
$\sup_{v\in[0,V_{n}]}\left|R_{2,n}(u^{\circ}+\i v)\right|^{2}=o_{P}(1),$
we derive
\begin{eqnarray*}
\mathrm{Var}\left[\int_{0}^{\infty}S_{n}(u^{\circ}+\i v)w_{\mu,n}(v)\, dv\right] & \lesssim & p\int_{0}^{\infty}\frac{\left|u^{\circ}+\i v\right|^{2}\left|w_{\mu,n}(v)\right|^{2}}{\left|\mathcal{M}(u^{\circ}+\i v+1)\right|^{2}}\, dv\\
& & +\sum_{j=p+1}^{n}\alpha(j)\left[\int_{0}^{\infty}\frac{\left|u^{\circ}+\i v\right|\left|w_{\mu,n}(v)\right|}{\left|\mathcal{M}(u^{\circ}+\i v+1)\right|}\, dv\right]^{2}.
\end{eqnarray*}
\textbf{3.} Turn now to the term $D_{n}.$ By the Plancherel's identity
\begin{eqnarray*}
\left|\int_{0}^{\infty}w_{\mu,n}(v)\mathcal{F}[\bar{\nu}](-v)dv\right| & = & \left|\int_{0}^{\infty}\frac{w_{\mu,n}(v)}{(-\i v)^{s}}\left[(-\i v)^{s}\mathcal{F}[\bar{\nu}](-v)\right]dv\right|\\
& = & \left|\int_{0}^{\infty}\frac{w_{\mu,n}(v)}{(-\i v)^{s}}\left[\mathcal{F}[\bar{\nu}^{(s)}](-v)\right]dv\right|\\
& = & 2\pi\,\left|\int_{-\infty}^{\infty}\bar{\nu}^{(s)}(x)\overline{{\cal F}^{-1}[w_{\mu,n}(\cdot)/(-i\cdot)^{s}](x)}dx\right|\\
& \leq & 2\pi\|\bar{\nu}^{(s)}\|_{\infty}\|\mathcal{F}^{-1}[w_{\mu,n}(\cdot)/(-i\cdot)^{s}]\|_{L^{1}}.
\end{eqnarray*}
\end{proof}
\textbf{\underline{Proof of Theorem~\ref{corub}}}
\begin{enumerate}[(i)]
\item
Suppose that $\pi\in\mathcal{P}(\beta,L)$ and $\alpha(j)\lesssim e^{-j\alpha^{*}},$
then by taking $p=c\log(n)$ for $c$ large enough, we arrive at
\begin{eqnarray*}
\mathrm{E}\left[\left|\mu_{n}-\mu\right|^{2}\right] & \lesssim & \frac{V_{n}^{-4}\log(n)}{n}\int_{0}^{V_{n}}|v|^{2\beta+2}\left|w_{\mu}(v/V_{n})\right|^{2}\, dv+V_{n}^{-2(s+2)}\\
&\lesssim& n^{-1}\log(n)V_{n}{}^{2\beta-1}+V_{n}^{-2(s+2)},\\
\mathrm{E}\left[\left|\lambda_{n}-\lambda\right|^{2}\right] & \lesssim & \frac{V_{n}^{-2}\log(n)}{n}\int_{0}^{V_{n}}|v|^{2\beta+2}\left|w_{\lambda}(v/V_{n})\right|^{2}\, dv+V_{n}^{-2(s+1)}\\&\lesssim& n^{-1}\log(n)V_{n}{}^{2\beta+1}+V_{n}^{-2(s+1)}
\end{eqnarray*}
By taking $V_{n}=n^{1/(2\beta+2s+3)},$ we get
\[
\mathrm{E}\left[\left|\mu_{n}-\mu\right|^{2}\right]\lesssim n^{-2(s+2)/(2\beta+2s+3)}\log(n)
\]
and
\[
\mathrm{E}\left[\left|\lambda_{n}-\lambda\right|^{2}\right]\lesssim n^{-2(s+1)/(2\beta+2s+3)}\log(n).
\]
\item
Suppose that \textup{$\pi\in\mathcal{E}(\alpha,L),$ then by taking
$p=0,$ we get
\begin{eqnarray*}
\mathrm{E}\left[\left|\mu_{n}-\mu\right|^{2}\right] & \lesssim & \frac{V_{n}^{-4}}{n}\left[\int_{0}^{V_{n}}\frac{\left|u^{\circ}+\i v\right|\left|w_{\mu}(v/V_{n})\right|}{\exp(-\alpha|v|)}\, dv\right]^{2}+V_{n}^{-2(s+2)}\\&\lesssim&\frac{1}{n}\exp(2\alpha V_{n})+V_{n}^{-2(s+2)},\\
\mathrm{E}\left[\left|\lambda_{n}-\lambda\right|^{2}\right] & \lesssim & \frac{V_{n}^{-2}}{n}\left[\int_{0}^{V_{n}}\frac{\left|u^{\circ}+\i v\right|\left|w_{\lambda}(v/V_{n})\right|}{\exp(-\alpha|v|)}\, dv\right]^{2}+V_{n}^{-2(s+1)}\\&\lesssim&\frac{V_{n}^{2}}{n}\exp(2\alpha V_{n})+V_{n}^{-2(s+1)}.
\end{eqnarray*}
Under the choice $V_{n}=\frac{1}{2\alpha}\log(n)-\frac{s+2}{\alpha}\log(\log(n)),$
one derives
\[
\mathrm{E}\left[\left|\mu_{n}-\mu\right|^{2}\right]\lesssim\log^{-2(s+2)}(n)
\]
and
\[
\mathrm{E}\left[\left|\lambda_{n}-\lambda\right|^{2}\right]\lesssim\log^{-2(s+1)}(n).
\]
}
\end{enumerate}
\subsection{Upper bounds for $MISE(\bar\nu_{n})$}
\label{uppmise}
\begin{prop}
Let the assumptions of the Proposition~\ref{propub} be fulfilled and let the kernel \(\mathcal{K}(\cdot)\) satisfy the assumption \eqref{assk}. Then the mean integrated squared error of the estimator \(\bar\nu_{n}(x)\) satisfies the following asymptotic relation
\begin{eqnarray*}
\textrm{MISE} (\bar\nu_{n}) & \lesssim&
\frac{1}{n}
\int_{I\!\!R}
\frac{\left|u^{\circ}+\i v\right|^{2}\left[\mathcal{K}(v / V_{n})\right]^{2}}{\left|\mathcal{M}(u^{\circ}+\i v+1)\right|^{2}}\, dv\\
&& \hspace{1cm}
+C_{1 } V_{n}^{3} \cdot {\mathbb E} \left[
\left( \mu_{n} - \mu \right)^{2}
\right]
+
C_{2} V_{n} \cdot {\mathbb E} \left[
\left( \lambda_{n} - \lambda \right)^{2}
\right] + C_{3} \frac{AL}{V_{n}^{2 s}}
\end{eqnarray*}
with some \(C_{1}, C_{2}, C_{3} > 0.\)
\end{prop}
\begin{proof}
Recall that
\begin{eqnarray*}
\bar\nu_{n}(x) &=& \frac{1}{2 \pi} \int_{I\!\!R} e^{\i v x } \hat\mathcal{F}[\bar\nu] (-v) \mathcal{K}(-v / V_{n}) dv = \mathcal{F}^{-1}[\hat\mathcal{F}_{\bar\nu} (\cdot) \mathcal{K}(\cdot / V_{n})](x),
\end{eqnarray*}
and
\begin{eqnarray*}
\hat\mathcal{F}[\bar\nu](-v) &=& - Y_{n}(u^{\circ}+\i v) + \c_{n}\cdot (u^{\circ}+\i v) + \a_{n},\\
\mathcal{F}[\bar\nu](-v) &=& - Y(u^{\circ}+\i v) + \c\cdot (u^{\circ}+\i v) + \a.
\end{eqnarray*}
By the Parsenval's identity,
\begin{eqnarray*}
\textrm{MISE} &=& \frac{1}{2 \pi} {\mathbb E} \left[
\int_{I\!\!R} \left|
\mathcal{F}[\bar{\nu}_{n}] (v)
-
\mathcal{F}[\bar\nu] (v)
\right|^{2} dv
\right] \\
&=&
\frac{1}{2 \pi} {\mathbb E} \left[
\int_{I\!\!R} \left|
\hat\mathcal{F}[\bar\nu] (v) \mathcal{K}(v / V_{n})
-
\mathcal{F}[\bar\nu] (v)
\right|^{2} dv
\right]
\\
&=&
\frac{1}{2 \pi} {\mathbb E} \left[
\int_{I\!\!R} \left|
\left(
\hat\mathcal{F}[\bar\nu] (v)
-
\mathcal{F}[\bar\nu] (v)
\right)
\mathcal{K}(v / V_{n})
+
\left(
\mathcal{K}(v / V_{n}) -1
\right)
\mathcal{F}[\bar\nu] (v)
\right|^{2} dv
\right] \\
& \leq &
\frac{1}{\pi} {\mathbb E} \left[
\int_{I\!\!R} \left|
\left(
\hat\mathcal{F}[\bar\nu] (v)
-
\mathcal{F}[\bar\nu] (v)
\right)
\mathcal{K}(v / V_{n})
\right|^{2} dv
\right]\\
&& \hspace{4cm} +
\frac{1}{\pi} {\mathbb E} \left[
\int_{I\!\!R} \left|
\left(
\mathcal{K}(v / V_{n}) -1
\right)
\mathcal{F}[\bar\nu] (v)
\right|^{2} dv
\right]\\
& \leq &
\frac{3}{\pi} \left( J_{1} +J_{2} +J_{3} \right) +\frac{1}{\pi}J_{4},
\end{eqnarray*}
where
\begin{eqnarray*}
J_{1} &:=& {\mathbb E} \left[
\int_{I\!\!R} \left|
Y_{n}(u^{\circ}+\i v) -
Y(u^{\circ}+\i v)
\right|^{2}
\left [
\mathcal{K}(v / V_{n})
\right]^{2} dv
\right], \\
J_{2}&:=&
A_{n} \cdot
{\mathbb E} \left[
\left( \mu_{n} - \mu \right)^{2}
\right] \qquad \mbox{with} \;\; A_{n}:= \int_{I\!\!R}
|u^{\circ}+\i v|^{2} \cdot \left[ \mathcal{K}(v / V_{n}) \right]^{2}
dv,\\
J_{3}&:=&
B_{n} \cdot
{\mathbb E} \left[
\left( \lambda_{n} - \lambda \right)^{2}
\right] \qquad \mbox{with} \;\;
B_{n}:=\int_{I\!\!R}
\left[ \mathcal{K}(v / V_{n}) \right]^{2}
dv,\\
J_{4}&:=&
\int_{I\!\!R} \left|
\left(
\mathcal{K}(v / V_{n}) -1
\right)
\mathcal{F}[\bar\nu] (v)
\right|^{2} dv.
\end{eqnarray*}
The treatment of \(J_{1}\) is based on the observation that
\begin{eqnarray*}
Y_{n}(z) - Y(z) \asymp R_{1,n} =z\,\frac{\left[\mathcal{M}_{n}(z)-\mathcal{M}(z)\right]}{\mathcal{M}(z+1)}.
\end{eqnarray*}
We get that
\begin{eqnarray*}
J_{1} \asymp
\int_{I\!\!R}{\mathbb E} \left[
\left|\mathcal{M}_{n}(u^{\circ}+\i v)-\mathcal{M}(u^{\circ}+\i v)\right|^{2}
\right]
\frac{\left|u^{\circ}+\i v\right|^{2}\left[\mathcal{K}(v / V_{n})\right]^{2}}{\left|\mathcal{M}(u^{\circ}+\i v+1)\right|^{2}}\, dv.
\end{eqnarray*}
As it was shown before,
\(
{\mathbb E} \left[
\left|\mathcal{M}_{n}(u^{\circ}+\i v)-\mathcal{M}(u^{\circ}+\i v)\right|^{2}
\right] \lesssim n^{-1},
\)
see \eqref{bosq}-\eqref{davydov}. Therefore,
\begin{eqnarray*}
J_{1} \lesssim
\frac{1}{n} \cdot
\int_{I\!\!R}
\frac{\left|u^{\circ}+\i v\right|^{2}\left[\mathcal{K}(v / V_{n})\right]^{2}}{\left|\mathcal{M}(u^{\circ}+\i v+1)\right|^{2}}\, dv.
\end{eqnarray*}
To complete the proof, it is sufficient to note that
\begin{eqnarray*}
A_{n} \asymp V_{n}^{3} \cdot \int_{I\!\!R} y^{2} \left[ \mathcal{K}(y) \right]^{2} dy, \qquad
B_{n} = V_{n} \cdot \int_{I\!\!R} \left[ \mathcal{K}(y) \right]^{2} dy,
\end{eqnarray*}
and
\begin{eqnarray*}
J_{4} \leq
A
\int_{I\!\!R} \left|
\frac{v}{V_{n}}
\right|^{2 s}
\left|
\mathcal{F}[\bar\nu] (v)
\right|^{2} dv
\leq \frac{A L}{V_{n}^{2 s}}.
\end{eqnarray*}
\end{proof}
\textbf{\underline{Proof of Theorem~\ref{nuub}}}
\begin{enumerate}[(i)]
\item
Recall that if \(\pi \in \mathcal{P}(\beta,L), \) then
\begin{eqnarray*}
\mathrm{E}\left[\left|\mu_{n}-\mu\right|^{2}\right]
&\lesssim& n^{-1}\log(n)V_{n}{}^{2\beta-1}+V_{n}^{-2(s+2)},\\
\mathrm{E}\left[\left|\lambda_{n}-\lambda\right|^{2}\right] &\lesssim& n^{-1}\log(n)V_{n}^{2\beta+1}+V_{n}^{-2(s+1)},
\end{eqnarray*}
see the proof of Theorem~\ref{corub}. Taking into account that
\(
J_{1} \lesssim
n^{-1} V_{n}^{2 \beta +3},
\)
we arrive at
\begin{eqnarray*}
\textrm{MISE} (\bar\nu_{n}) \lesssim n^{-1} V_{n}^{3+2 \beta} +n^{-1}\log(n)V_{n}^{2\beta+1}+V_{n}^{-2(s+1)} + V_{n}^{-2s}.
\end{eqnarray*}
Choosing $V_{n}=n^{1/(2\beta+2s+3)},$ we get
\begin{eqnarray*}
n^{-1} V_{n}^{3+2 \beta} = V_{n}^{-2s} \gtrsim V_{n}^{-2(s+1)},
\end{eqnarray*}
and therefore
\begin{eqnarray*}
\textrm{MISE} (\bar\nu_{n}) \lesssim n^{-1}\log(n)V_{n}^{2\beta+1}+ V_{n}^{-2s} \lesssim n^{-2s/(2\beta+2s+3)}.
\end{eqnarray*}
\item Similarly, we derive the upper bound for the class \(\mathcal{E}(\alpha,L).\)
Recall that
\begin{eqnarray*}
\mathrm{E}\left[\left|\mu_{n}-\mu\right|^{2}\right] &\lesssim&n^{-1}\exp(2\alpha V_{n})+V_{n}^{-2(s+2)},\\
\mathrm{E}\left[\left|\lambda_{n}-\lambda\right|^{2}\right] &\lesssim&n^{-1} V_{n}^{2} \exp(2\alpha V_{n})+V_{n}^{-2(s+1)}
\end{eqnarray*}
and therefore
\begin{eqnarray*}
\textrm{MISE} (\bar\nu_{n}) \lesssim n^{-1}\log(n)V_{n}^{3} e^{2 \alpha V_{n}}+ V_{n}^{-2s} \lesssim \left( \log n \right)^{- 2 s}.
\end{eqnarray*}
\end{enumerate}
\subsection{Lower bounds for $MISE$}
\label{prooflb}
\textbf{\underline{Proof of Theorem~\ref{thm3}}.} The general idea of the proof is to apply Theorem~2.7 from \cite{Tsyb}. This theorem yields that \eqref{lowerbound} holds, if there exists a parameterized set of L{\'e}vy triplets
\[
\mathcal{T}_{\theta}=(1,0,\nu_{\theta}) \subset \mathcal{G}(s,R),\quad\theta\in\{0,1\}^{L}
\] for some \(s\in {\mathbb N}\cup 0, R>0\), \(L>0\) and a set of parameters $\{\theta^{(j)}, \; j=0,\ldots, M\}$ such that the following two properties hold.
\begin{enumerate}[(i)]
\item For any \(0 \leq j < k \leq M\),
\begin{eqnarray}
\label{1}
\int_{\mathbb{R}}\left|\nu_{\theta^{(j)}}(x)-\nu_{\theta^{(k)}}(x)\right| ^{2} dx\geq 2\varphi_{n}.
\end{eqnarray}
\item Denote by \(\pi_{\theta_{j}}, j=0,\ldots, M,\) the probability distribution of the exponential L{\'e}vy model $A_{j,\infty}=\int_{0}^{\infty}e^{-\xi_{j,s}}ds,$ where $\xi_{j,s}$ is a L{\'e}vy subordinator with triplet $\mathcal{T}_{\theta_{j}}.$ Then
\begin{eqnarray}
\label{2}
\frac{n}{M} \sum_{j=1}^{M} K \left(\pi_{\theta^{(j)}}, \pi_{\theta^{(0)}}\right) \leq \varkappa\log(M),
\end{eqnarray}
for \(n\) large enough, where \(K\) stands for the Kullback-Leibler divergence between models, and \(\varkappa\in (0,1/8)\).
\end{enumerate}
Below we present a detailed proof for the polynomial case.
\textbf{1. Presentation of the models.} Consider an exponential L{\'e}vy model $A_{0,\infty}=\int_{0}^{\infty}e^{-\xi_{0,s}}ds,$
where $\xi_{0,s}$ is a L{\'e}vy subordinator with a triplet $(1,0,\nu_{0})$
and $\nu_{0}(x)=abe^{-bx}$ for some $0<a\leq1,$ $0<b<1.$ It is clear that \((1,0,\nu_{0})\in \mathcal{G}(0,R)\) for some \(R>0\) and the Laplace
exponent of $\xi_{0,s}$ is given by
\begin{eqnarray*}
\phi_{0}(z) = z+\int_{0}^{\infty}(1-e^{-xz})\,\nu_{0}(x)\, dx= z\left[1+\frac{a}{z+b}\right],\qquad\Re(z)>-b,
\end{eqnarray*}
see Example~1 from Section~\ref{secsim}. For the case of general classes \(\mathcal{G}(s,R)\) with \(s>0,\) we could take a L\'evy density of the form \(\nu_0(x)=b^{1+s}x^se^{-bx} / \Gamma(s+1).\)
Fix some \(L>0\) and let us construct now a parameterized set of L{\'e}vy triplets
\(
\mathcal{T}_{\theta}=(1,0,\nu_{\theta}),\;\theta\in\{0,1\}^{L},
\)
with L{\'e}vy measure \(\nu_{\theta}\) defined by
\begin{eqnarray*}
\nu_{\theta}(x):=\nu_{0}(x)+\delta\cdot\Delta_{\theta}(x),
\end{eqnarray*}
where
$\delta>0$ small enough,
\begin{eqnarray*}
\Delta_{\theta}(x)&:=&\bigl(g_{\theta}(x)+a(g_{\theta}\star\exp(-b\cdot)(x))\bigr)' ,\\
g_{\theta}(x) &:=&\sum_{k=L+1}^{2L} \theta_{k-L}\cos(k\gamma_{L}x)g_{0}(x),
\end{eqnarray*}
\(\theta_{k-L}\) stands for the \((k-L)\)-th component of the vector \(\theta\), $\gamma_{L}\to\infty$ as $L\to\infty$, and
\[
g_{0}(x):= x^{-3/2}\exp(-1/x),\quad x>0.
\]
\textbf{2. Distributional properties of the models.}
In this step, we perform some technical calculations, which will be used later. It holds
\begin{multline*}
\mathcal{L}[\Delta_{\theta}](z) =
\int_{0}^{\infty}e^{-zx}\Delta_{\theta}(x)\, dx \\= z\left[1+\frac{a}{z+b}\right]\left[\int_{0}^{\infty}e^{-zx}g_{\theta}(x)\, dx\right]
= \phi_{0}(z)\cdot\mathcal{L}[g_{\theta}](z),
\end{multline*}
where \(\mathcal{L}[g_{\theta}](z)\) is the Laplace transform of the function \(g_{0}(\cdot)\), which is equal to
\[
\mathcal{L}[g_{\theta}](z)=\frac{1}{2}\sum_{k=L+1}^{2L}\theta_{k-L}\left[\mathcal{L}[g_{0}](z+\i \gamma_{L}k)+\mathcal{L}[g_{0}](z-\i \gamma_{L}k)\right].
\]
We see that $\int_{0}^{\infty}\Delta_{\theta}(x)\, dx=0$ and
\[
\phi_{\theta}(z)-\phi_{0}(z)=\delta\phi_{0}(z)\mathcal{L}[g_{\theta}](z),
\]
where \(\phi_{\theta}(\cdot)\) is the Laplace exponent of a L{\'e}vy process \(\xi_{\theta,s}\) with the L{\'e}vy triplet \(\mathcal{T}_{\theta}\). Furthermore, the Laplace transform of $g_{0}$ is given by
\[
\mathcal{L}[g_{0}](u+\i v)=\sqrt{\pi}e^{-2(z_{+}+iz_{-})}
\]
with $2z_{\pm}^{2}=\sqrt{u^{2}+v^{2}}\pm u.$
The Mellin transform of the density $\pi_{\theta}$ corresponding
to the L{\'e}vy model $\mathcal{T}_{\theta}$ satisfies the following
functional equation
\[
\frac{\mathcal{M}_{\theta}(z)}{\mathcal{M}_{0}(z)}=\frac{\phi_{\theta}(z)}{\phi_{0}(z)}\frac{\mathcal{M}_{\theta}(z+1)}{\mathcal{M}_{0}(z+1)}.
\]
Since
\[
\frac{\phi_{\theta}(z)}{\phi_{0}(z)}-1=\delta\mathcal{L}[g_{\theta}](z),
\]
and
\begin{eqnarray*}
\sum_{k=1}^{\infty}\left|\mathcal{L}[g_{\theta}](z+k)\right| & \leq & C\exp\left(-\sqrt{2\mathrm{Re}(z)}-\sqrt{2\left|\mathrm{Im}(z)\right|}\right)\\
&&\hspace{3cm}\cdot\sum_{k=1}^{\infty}\sum_{j=L+1}^{2L}\exp(-\sqrt{2\gamma_{L}j}-\sqrt{2k})\\
& \leq & C'\exp\left(-\sqrt{2\mathrm{Re}(z)}-\sqrt{2\left|\mathrm{Im}(z)\right|}\right),\quad\mathrm{Re}(z)\geq0,
\end{eqnarray*}
we derive the following infinite product representation for the ratio $\mathcal{M}_{\theta}(z)/\mathcal{M}_{0}(z)$
\[
\frac{\mathcal{M}_{\theta}(z)}{\mathcal{M}_{0}(z)}=\prod_{k=0}^{\infty}(1+\delta\mathcal{L}[g_{\theta}](z+k)).
\]
Furthermore, it can be proved that
\[
\left|\frac{\mathcal{M}_{\theta}(u+\i v)}{\mathcal{M}_{0}(u+\i v)}-1\right|\leq c\delta\left|\mathcal{L}[g_{\theta}](u+\i v)\right|
\]
for some absolute constant $c>0.$ Note that the random variables
$A_{\theta,\infty}=\int_{0}^{\infty}e^{-\xi_{\theta,s}}ds$ with $\xi_{\theta,s}$
being a L{\'e}vy process with the triplet $\mathcal{T}_{\theta},$ satisfies
$0<A_{\theta,\infty}<1$ a.s. Moreover the density $p_{0}$ of the
r.v. $A_{0,\infty}$ has the form
\[
\pi_{0}(x)=\frac{1}{B(b-1,a)}x^{b}(1-x)^{a-1}1_{\{0<x<1\}}
\]
and the Mellin transform $\mathcal{M}_{0}(z)$ of $A_{0,\infty}$ is given
by
\begin{equation}
\mathcal{M}_{0}(z)=\frac{B(z+b,a)}{B(b-1,a)},\label{eq:mellin_0}
\end{equation}
see Example~\ref{exx}.
\textbf{3. Class \(\mathcal{G}(s,R)\).} In this step, we check that constructed models \(\mathcal{T}_{\theta^{(j)}},\) \(j=1,\ldots, M\) belong to class \(\mathcal{G}(s,R)\) with \(s=0\) and some \(R>0\). We have for any $\theta\in\{0,1\}^{L},$
\begin{eqnarray*}
\int_{\mathbb{R}}\left|v\right|^{2s}\left|\mathcal{F}\left[\nu_{\theta}\right](v)\right|^{2}\, dx & \leq & \int_{\mathbb{R}}\left|v\right|^{2s}\left|\mathcal{F}\left[\nu_{0}\right](v)\right|^{2}\, dv\\
&&\hspace{2cm}+\int_{\mathbb{R}}\left|v\right|^{2s}\left|\mathcal{F}\left[\nu_{\theta}\right](v)-\mathcal{F}\left[\nu_{0}\right](v)\right|^{2}\, dv\\
& \leq & \int_{\mathbb{R}}\left|v\right|^{2s}\left|\mathcal{F}\left[\nu_{0}\right](v)\right|^{2}\, dv+\delta^{2}\int_{\mathbb{R}}\left|v\right|^{2s}\left|\mathcal{F}[\Delta_{\theta}](v)\right|^{2}\, dv.\\
\end{eqnarray*}
The inequality $\left|\phi_{0}(-\i v)\right|\leq c\cdot |v|$ for $v\in\mathbb{R}$, where \(c= 1 + a/b,\) implies
\begin{eqnarray*}
\int_{\mathbb{R}}\left|v\right|^{2s}\left|\mathcal{F}[\Delta_{\theta}](v)\right|^{2}\, dv & \leq & c \int_{\mathbb{R}}|v|^{2(s+1)}\left|\mathcal{L}[g_{\theta}](-\i v)\right|^{2}\, dv\\
& = & \frac{c}{2}\int_{\mathbb{R}}|v|^{2(s+1)}\cdot \left|\sum_{k=L+1}^{2L}\theta_{k-L}\left(\mathcal{L}[g_{0}](-\i v+\i \gamma_{L}k) \right.\right. \\
&& \hspace{3cm} \left. \left.
+\mathcal{L}[g_{0}](-\i v-\i \gamma_{L}k)\right)\right|^{2}\, dv\\
& \leq & \frac{c}{2}\sum_{k=L+1}^{2L}\int_{\mathbb{R}}|v|^{2(s+1)}\left|\mathcal{L}[g_{0}](-\i v+\i \gamma_{L}k)\right|^{2}\, dv\\
& & \hspace{0.2cm}+\frac{c}{2}\sum_{k=L+1}^{2L}\int_{\mathbb{R}}|v|^{2(s+1)}\left|\mathcal{L}[g_{0}](-\i v-\i \gamma_{L}k)\right|^{2}\, dv\\&&\hspace{6cm}+R_{L},\\
\end{eqnarray*}
where
\begin{eqnarray*}
R_{L} & = & 2\sum_{k\neq j}\int_{\mathbb{R}}|v|^{4}\mathcal{L}[g_{0}](-\i v-\i j\gamma_{L})\overline{\mathcal{L}[g_{0}](-\i v-\i k\gamma_{L})}\, dv\\
& & +2\sum_{k\neq j}\int_{\mathbb{R}}|v|^{4}\mathcal{L}[g_{0}](-\i v+\i j\gamma_{L})\overline{\mathcal{L}[g_{0}](-\i v+\i k\gamma_{L})}\, dv
\end{eqnarray*}
It holds
\begin{eqnarray*}
\left|R_{L}\right| & \leq & CL\sum_{j=1}^{2L}\left(j\gamma_{L}\right){}^{2(s+1)}\exp(-\sqrt{2\gamma_{L}j})\\
& \leq & CL^{2(s+1)+2}\gamma_{L}{}^{2(s+1)}\exp(-\sqrt{2\gamma_{L}})\\
& = & o\left(L^{2(s+1)+1}\right),
\end{eqnarray*}
provided $\gamma_{L}=c\log^{2}(L)$ for large enough $c>0.$ Hence
$\int_{\mathbb{R}}\left|v\right|^{2s}\left|\mathcal{F}[\Delta_{\theta}](v)\right|^{2}\, dv$
is bounded if $\delta^{2}\gamma_{L}^{2(s+1)}L^{2s+3}=O(1).$
\textbf{4. Upper bound for the $L^{2}$-distance between elements of \(\{\nu_{\theta}\}\).}
Fix two
vectors $\theta,\theta'\in\{0,1\}^{L}.$ We have
\begin{eqnarray*}
\int_{\mathbb{R}}\left|\nu_{\theta}(x)-\nu_{\theta'}(x)\right|^{2}\, dx & = & \frac{1}{2\pi}\delta^{2}\int_{\mathbb{R}}\left|\phi_{0}(-\i v)\mathcal{L}[g_{\theta}-g_{\theta'}](-\i v)\right|^{2}\, dv\\
& = & \frac{1}{2\pi}\delta^{2}\sum_{k=L+1}^{2L}\left(\theta_{k-L}-\theta'_{k-L}\right)^{2}\\
&& \hspace{2cm} \cdot\int_{\mathbb{R}}\left|\phi_{0}(-\i v)\mathcal{L}[g_{0}](-\i v+\i \gamma_{L}k)\right|^{2}\, dv\\
& & +\frac{1}{2\pi}\delta^{2}\sum_{k=L+1}^{2L}\left(\theta_{k-L}-\theta'_{k-L}\right)^{2}\\
&& \hspace{2cm} \cdot\int_{\mathbb{R}}\left|\phi_{0}(-\i v)\mathcal{L}[g_{0}](-\i v-\i \gamma_{L}k)\right|^{2}\, dv\\
& & +\frac{1}{2\pi}\delta^{2}R_{L},
\end{eqnarray*}
where
\begin{eqnarray*}
R_{L} & \leq & 2\sum_{k\neq j}\int_{\mathbb{R}}\left|\phi_{0}(-\i v)\right|^{2}\mathcal{L}[g_{0}](-\i v-\i j\gamma_{L})\overline{\mathcal{L}[g_{0}](-\i v-\i k\gamma_{L})}\, dv\\
& & +2\sum_{k\neq j}\int_{\mathbb{R}}\left|\phi_{0}(-\i v)\right|^{2}\mathcal{L}[g_{0}](-\i v+\i j\gamma_{L})\overline{\mathcal{L}[g_{0}](-\i v+\i k\gamma_{L})}\, dv.
\end{eqnarray*}
Consider, for example,
\begin{multline*}
\int_{\mathbb{R}}\left|\phi_{0}(-\i v)\mathcal{L}[g_{0}](-\i v+\i \gamma_{L}k)\right|^{2}\, dv = \int_{\mathbb{R}}\left|\phi_{0}(-i(v+\gamma_{L}k))\mathcal{L}[g_{0}](-\i v)\right|^{2}\, dv
\\= \int_{\mathbb{R}}\left|v+\gamma_{L}k\right|^{2}\left|1+\frac{a}{b-i(v+\gamma_{L}k)}\right|^{2} e^{-2\sqrt{2|v|}}\, dv\\=
\gamma_{L}^{2}k^{2}\int_{\mathbb{R}}\left|1+\frac{a}{b-i(v+\gamma_{L}k)}\right|^{2}e^{-2\sqrt{2|v|}}\, dv +O(\gamma_{L}k).
\end{multline*}
So we have
\begin{multline*}
\sum_{k=L+1}^{2L}\left(\theta_{k-L}-\theta'_{k-L}\right)^{2}\int_{\mathbb{R}}\left|\phi_{0}(-\i v)\mathcal{L}[g_{0}](-\i v+\i \gamma_{L}k)\right|^{2}\, dv \\= C\gamma_{L}^{2}\sum_{k=L+1}^{2L}\left(\theta_{k-L}-\theta'_{k-L}\right)^{2}k^{2}
+o\left(\gamma_{L}^{2}\sum_{k=L+1}^{2L}\left(\theta_{k-L}-\theta'_{k-L}\right)^{2}k^{2}\right)\\
\geq C'\gamma_{L}^{2}L^{2}\sum_{k=1}^{L}I(\theta_{k}\neq\theta'_{k}),
\end{multline*}
as $L\to\infty$ and $\rho(\theta,\theta')=\sum_{k=1}^{L}I(\theta_{k}\neq\theta'_{k})>0.$
Analogously,
\begin{eqnarray*}
\sum_{k=L+1}^{2L}\left(\theta_{k-L}-\theta'_{k-L}\right)^{2}\int_{\mathbb{R}}\left|\phi_{0}(-\i v)\mathcal{L}[g_{0}](-\i v-\i \gamma_{L}k)\right|^{2}\, dv & = & C''\gamma_{L}^{2}L^{2}\rho(\theta,\theta').\end{eqnarray*}
Furtheremore, one shows (see above) that
\[
\left|R_{L}\right|=o\left(L^{3}\right).
\]
\textbf{5. Choice of $\theta^{(0)},\ldots,\theta^{(M)}$.}
Our choice is based on the well-known Varshamov-Gilbert bound (see \cite{Tsyb}, Lemma~2.9), which implies that there are $M>2^{L/8}$
vectors $\theta^{(0)},\ldots,\theta^{(M)}\in\{0,1\}^{L}$ such that
\[
\rho(\theta^{(j)},\theta^{(k)})\geq L/8.
\]
\textbf{6. Upper bound for $K(\pi_{0},\pi_{\theta})$.}
By Parseval identity for Mellin transforms, we get
\begin{eqnarray*}
K(\pi_{0},\pi_{\theta}) & = & \int_{0}^{1}\frac{\left|\pi_{\theta}(x)-\pi_{0}(x)\right|^{2}}{\pi_{0}(x)}\, dx\\
& = & \int_{0}^{1}x^{-b}(1-x)^{1-a}\left|\pi_{\theta}(x)-\pi_{0}(x)\right|^{2}\, dx\\
& \leq & \int_{0}^{1}x^{-b}\left|\pi_{\theta}(x)-\pi_{0}(x)\right|^{2}\, dx\\
& = & \frac{1}{2\pi}\int_{-\infty}^{\infty}\left|\mathcal{M}_{\theta}((1-b)/2+\i v)-\mathcal{M}_{0}((1-b)/2+\i v)\right|^{2}\, dv\\
& \leq & \frac{c\delta^{2}}{2\pi}\int_{-\infty}^{\infty}\left|\mathcal{M}_{0}((1-b)/2+\i v)\right|^{2}\left|\mathcal{L}[g_{\theta}](u+\i v)\right|^{2}\, dv.
\end{eqnarray*}
So we get
\begin{eqnarray*}
K(\pi_{0},\pi_{\theta}) & \leq & \frac{c\delta^{2}}{2\pi}\sum_{k=L+1}^{2L}\int_{\mathbb{R}}\left|\mathcal{M}_{0}((1-b)/2+\i v)\right|^{2}\left|\mathcal{L}[g_{0}](-\i v+\i \gamma_{L}k)\right|^{2}\, dv,\\
& & +\frac{c\delta^{2}}{2\pi}\sum_{k=L+1}^{2L}\int_{\mathbb{R}}\left|\mathcal{M}_{0}((1-b)/2+\i v)\right|^{2}\left|\mathcal{L}[g_{0}](-\i v-\i \gamma_{L}k)\right|^{2}\, dv\\
& & +\frac{c\delta^{2}}{2\pi}R_{L},
\end{eqnarray*}
where
\begin{eqnarray*}
R_{L} & \leq & 2\sum_{k\neq j}\int_{\mathbb{R}}\left|\mathcal{M}_{0}((1-b)/2+\i v)\right|^{2}\mathcal{L}[g_{0}](-\i v-\i j\gamma_{L})\overline{\mathcal{L}[g_{0}](-\i v-\i k\gamma_{L})}\, dv\\
& & +2\sum_{k\neq j}\int_{\mathbb{R}}\left|\mathcal{M}_{0}((1-b)/2+\i v)\right|^{2}\mathcal{L}[g_{0}](-\i v+\i j\gamma_{L})\overline{\mathcal{L}[g_{0}](-\i v+\i k\gamma_{L})}\, dv.
\end{eqnarray*}
The equation (\ref{eq:mellin_0}) implies that $\mathcal{M}_{0}(z)$ is finite
for all $z$ with $\mathrm{Re}(z)\geq0$ and
\begin{eqnarray*}
\mathcal{M}_{0}(u+\i v) & = & C(a,b)\frac{\Gamma(u+\i v+b)}{\Gamma(u+\i v+b+a)}\\
& \asymp & C(a,b)e^{-a\log(u+\i v+b)}\\
& = & C(a,b)\left((u+b)^{2}+v^{2}\right)^{-a/2}e^{i\mathrm{Arg}(u+\i v+b)},\quad u^{2}+v^{2}\to\infty.
\end{eqnarray*}
Hence
\[
\left|\mathcal{M}_{0}(u+\i v)\right|\asymp C(a,b)\left((u+b)^{2}+v^{2}\right)^{-a/2},\quad u^{2}+v^{2}\to\infty
\]
and the density $\pi_{0}$ of $A_{0,\infty}$ belongs to the class
$\mathcal{P}(a,C(a,b))$ (see also Example~\ref{exx}). We have
\[
\sum_{k=L+1}^{2L}\int_{\mathbb{R}}\left|\mathcal{M}_{0}((1-b)/2+\i v)\right|^{2}\left|\mathcal{L}[g_{0}](-\i v+\i \gamma_{L}k)\right|^{2}\, dv=O(L^{-2a+1})
\]
and
\[
\sum_{k=L+1}^{2L}\int_{\mathbb{R}}\left|\mathcal{M}_{0}((1-b)/2+\i v)\right|^{2}\left|\mathcal{L}[g_{0}](-\i v-\i \gamma_{L}k)\right|^{2}\, dv=O(L^{-2a+1}).
\]
Hence
\begin{eqnarray}
\label{3}
\frac{n}{M}\sum_{m=1}^{M} K(\pi_{0},\pi_{\theta^{(m)}})\leq n\delta^{2}L^{-2a}\log(M),\quad L\to\infty
\end{eqnarray}
for large \(n\).
\textbf{7. Choice of $L$.} To complete the proof, we choose \(L\) such that the conditions \eqref{1} and \eqref{2} are fulfilled. First note that since our model belongs to the class \(\mathcal{G}(s,R)\), we can take
$\gamma_{L}=c\log^{2}(L)$ and $\delta^{2}= \gamma_{L}^{-2(s+1)}L^{-2s-3} \cdot O(1),$ see Step~3 of the proof for details. Second, comparing \eqref{3} with \eqref{2}, we fix
\(\varkappa =
n\delta^{2}L^{-2a}\). This leads to the choice of \(L\) as the solution of the equation
\[L^{2a+2s+3}
\log^{4(s+1)}(L) = n O(1)
\]
Combination of the results from Steps~4 and 5 yields the condition \eqref{1}, because
\begin{eqnarray*}
\int_{\mathbb{R}}\left|\nu_{\theta}(x)-\nu_{\theta'}(x)\right|^{2} dx &\geq& C_{1} \delta^{2}\gamma_{L}^{2} L^{3}\\
&=& C_{2}(\log L)^{-4s} L^{-2s } \\&=& C_{3}\left( \log L \right)^{4s \frac{-2a-1}{2a+2s+3}}
n^{-2s / (2a+2s+3)}
\end{eqnarray*}
for some \(C_{1}, C_{2}, C_{3}>0\) and \(L\) large enough. This observation completes the proof.
\bibliographystyle{plain}
|
\section{Introduction}
\label{sec:introduction}
Dark matter direct detection experiments have witnessed in the last years an impressive improvement in sensitivity, reflected in particular by the recent breaking of the zeptobarn-scale barrier in the scattering cross section by the LUX experiment~\cite{Akerib:2013tjd}. Moreover, the prospects for future improvements are bright: in the near future the XENON1T experiment~\cite{Aprile:2012zx} will extend the reach in cross section by about one order of magnitude and, in the longer term, a multi-ton xenon detector, such as DARWIN~\cite{Baudis:2012bc}, might even reach the yoctobarn scale.
The excellent sensitivity of experiments has allowed to rule out already some well motivated models, notably those where the dark matter is in the form of Weakly Interacting Massive Particles (WIMPs) which couple at tree level to the Standard Model via the exchange of the Z-boson, such as the sneutrino in the Minimal Supersymmetric Standard Model~\cite{Falk:1994es}. On the other hand, some other scenarios have suppressed tree-level couplings to the nucleons and can naturally evade detection at current experiments. This is the case when the dark matter interaction with the nucleus is mediated by the Higgs boson, and hence suppressed by the small Higgs coupling to the proton constituents (for example, when the dark matter is a singlet scalar~\cite{McDonald:1993ex, Cline:2013gha}, an inert $SU(2)_L$ doublet~\cite{Deshpande:1977rw,LopezHonorez:2006gr,Majumdar:2006nt} a hidden vector~\cite{Hambye:2008bq} or a singlet antisymmetric tensor~\cite{Cata:2014sta}). Alternatively, this is also the case when the dark matter only couples to the leptons or to the heavy quarks, where the tree-level couplings to the nucleons simply vanish~(as occurs, for instance, in the radiative seesaw model \cite{Ma:2006km}). Nevertheless, even if the dark matter-nucleon tree-level coupling vanishes, such a coupling is necessarily generated via quantum effects, due to the non-vanishing gauge and Higgs coupling interactions of all the Standard Model fermions. Given the excellent sensitivity of current instruments, such quantum effects could induce observable signals at experiments. The strength of the signal is, however, very model dependent and should be investigated case by case.
In this paper we focus on scenarios with a Dirac singlet fermion as dark matter candidate which couples to a Standard Model fermion via a scalar mediator. The model contains only three free parameters, namely the dark matter mass, the mediator mass and the Yukawa coupling strength, which can be fixed in terms of the other two parameters from requiring that the dark matter population in our Universe is generated via thermal freeze-out. In this paper we a present a comprehensive analysis of the scattering process of dark matter particles with nucleons for couplings to any Standard Model fermion. The scattering arises at tree level in the case of the first generation quarks and radiatively for other fermions, namely via the one-loop exchange of a Higgs, a Z-boson and a photon for any fermion~\cite{Agrawal:2011ze,Bai:2013iqa,Bai:2014osa,Chang:2014tea,Agrawal:2014ufa,Agrawal:2014una,Kopp:2014tsa,Yu:2014mfa,DiFranzo:2013vra}, or via a box-diagram with two external gluons for colored fermions \cite{Drees:1993bu,Hisano:2010ct,Gondolo:2013wwa}. We then calculate the scattering rate with the nucleons induced by this effective interaction, which we finally confront with the upper limit reported by LUX, as well as with the projected reach of the XENON1T and DARWIN experiments.
The paper is organized as follows. In section \ref{sec:model_introduction} we describe the model. Then, in section \ref{sec:L_eff} we discuss the effective dark matter-nucleon interaction induced in the model, for dark matter coupling to any Standard Model fermion. In section \ref{sec:event_rate} we calculate the corresponding scattering rate with nucleons for each of these possibilities and we confront the predicted rate to the present limits from the LUX experiment and to the projected sensitivities of future experiments. Finally, in section \ref{sec:conclusions} we present our conclusions and in Appendix~\ref{sec:appendix_lux_limits}, our method to calculate the event rate in LUX.
\section{Singlet Dirac dark matter with a charged mediator}
\label{sec:model_introduction}
We consider a simple extension of the Standard Model by a Dirac fermion $\chi$, singlet under the Standard Model gauge group and which constitutes our dark matter candidate, and a charged scalar particle $\eta$, which mediates the interaction between the dark matter and a Standard Model fermion $f$. Furthermore, in order to guarantee the stability of the dark matter particle, we impose a discrete $Z_2$ symmetry under which $\chi$ and $\eta$ are odd, while all the Standard Model particles are even. The most general from for the Lagrangian reads:
\begin{align}
\mathcal{L} = \mathcal{L}_\text{SM} + \mathcal{L}_\chi + \mathcal{L}_\eta + \mathcal{L}_\text{int}^\text{fermion} + \mathcal{L}_\text{int}^\text{scalar} \,.
\end{align}
Here, $\mathcal{L}_\text{SM}$ is the Standard Model Lagrangian, which includes a potential for the Higgs doublet $\Phi$, $V=m_1^2 \Phi^\dagger \Phi +\frac{1}{2}\lambda_1 (\Phi^\dagger \Phi)^2$, while $\mathcal{L}_\chi$ and $\mathcal{L}_\eta$ are the parts of the Lagrangian involving just the dark matter particle or the scalar mediator and which are given by:
\begin{align}
\mathcal{L}_\chi &= \bar{\chi} i \slashed{\partial} \chi - m_\chi \bar{\chi} \chi \quad \text { and}\\
\mathcal{L}_\eta &= (D_\mu \eta)^\dagger (D^\mu \eta) - m_2^2 \eta^\dagger \eta - \frac12 \lambda_2 (\eta^\dagger \eta)^2 \,,
\end{align}
where the covariant derivative $D_\mu$ depends on the gauge quantum numbers of $\eta$, which will be specified below.
On the other hand, $\mathcal{L}_\text{int}^\text{fermion}$ describes the Yukawa interaction between the dark matter particle, the scalar mediator and the Standard Model fermion $f$, which can be either a right-handed fermion singlet $f_R \in \left\{ u_R^i, d_R^i, e_R^i \right\}$, $i=1,2,3$ being a generation index, or a left-handed fermion doublet $f_L \in \left\{ Q_L^i, L_L^i \right\}$. Then, the interaction Lagrangian reads
\begin{align}
\label{eq:L_int_fermion_fR}
\mathcal{L}_\text{int}^\text{fermion} = -y \, \eta^\dagger \bar{\chi} f_R+ \text{h.c.} \,,
\end{align}
for coupling to a right-handed fermion and
\begin{align}
\label{eq:L_int_fermion_fL}
\mathcal{L}_\text{int}^\text{fermion} = -y \, \eta^\dagger \bar{\chi} f_L+ \text{h.c.} =
\begin{cases}
-y \, \eta_0^\dagger \bar{\chi} \nu_L -y \, \eta_-^\dagger \bar{\chi} e_L+ \text{h.c. } \text{ for coupling to }L_L \,, \text{ and} \\
-y \, \eta_u^\dagger \bar{\chi} u_L -y \, \eta_d^\dagger \bar{\chi} d_L+ \text{h.c. } \text{ for coupling to }Q_L
\end{cases}
\end{align}
for coupling to a left-handed fermion. Notice that specifying the fermion representation fixes the gauge quantum numbers of the scalar mediator.
Lastly $\mathcal{L}_\text{int}^\text{scalar}$ describes the interaction between the scalar mediator and the Standard Model Higgs boson and reads:
\begin{align}
\label{eq:L_int_scalar}
\mathcal{L}_\text{int}^\text{scalar} =
\begin{cases}
- \lambda_3 (\Phi^\dagger \Phi) (\eta^\dagger \eta) &\text{ for coupling to }f_R\,, \text{ and} \\
- \lambda_3 (\Phi^\dagger \Phi) (\eta^\dagger \eta) - \lambda_4 (\Phi^\dagger \eta) (\eta^\dagger \Phi) &\text{ for coupling to }f_L \,.
\end{cases}
\end{align}
These interaction terms only affect the phenomenology in a significant way when the couplings are ${\cal O}(1)$. Therefore, in what follows we will neglect these interactions and we will simply set these couplings to zero. This choice implies in particular that, for the scenario where the dark matter couples to left-handed fermions, the two components of the scalar doublet $\eta$ have a common mass, which we denote by $m_\eta$.
Provided that the coupling $y$ is large enough, the dark matter candidate $\chi$ is kept in thermal equilibrium in the Early Universe with the plasma of Standard Model particles. For large parts of the parameter space, the annihilation process which is most relevant for the equilibration is $\chi \bar{\chi} \rightarrow f \bar{f}$, where $f$ is the Standard Model fermion that couples to the dark matter particle via the interaction given in Eq.~(\ref{eq:L_int_fermion_fR}) or~(\ref{eq:L_int_fermion_fL}). The thermally averaged annihilation cross section at the time of dark matter decoupling is, for Dirac fermions, well approximated by the velocity independent part of the cross section, which reads:
\begin{align}
\label{eq:sigmav_ff}
\langle \sigma v \rangle \simeq \sigma v \big|_{v \rightarrow 0} =
\frac{y^4 N_c}{32 \pi}\, \frac{m_\chi^2 \sqrt{1-\left( m_f/m_\chi\right)^2}}{\left( m_\chi^2+m_\eta^2-m_f^2\right)^2} \,.
\end{align}
In this expression, $N_c=3$ for coupling to quarks and $N_c=1$ for coupling to leptons. Moreover, in scenarios where the dark matter particle couples to a fermion doublet $f_L$, the total annihilation cross section is given by the sum of the annihilation cross sections into the two states of the fermion doublet. As is well known, for Dirac dark matter the total annihilation cross section for the process $\chi \bar{\chi} \rightarrow f \bar{f}$ at the time of freeze-out should take the value $\langle \sigma v \rangle_\text{thermal} \simeq 4.4 \times 10^{-26} \, \text{cm}^3/\text{s}$ in order to match the dark matter density in our Universe, $\Omega_{\rm DM} h^2=0.1198\pm 0.0015$, as measured by the Planck satellite~\cite{Planck:2015xua}.
In this work we also consider the possibility that the scalar mediator $\eta$ is close in mass to the dark matter particle $\chi$, in which case the calculation of the dark matter abundance must be modified due to the presence of coannihilation processes. Namely, if $m_\eta \lesssim 1.2\, m_\chi$, the temperature at the time of dark matter decoupling can be large enough to keep $\eta$ also in thermal equilibrium with the plasma, which efficiently depletes the number of dark matter particles through the processes $\chi \eta \rightarrow q g$, $\eta \eta^\dagger \rightarrow g g$, and $\eta_u \eta_d^\dagger \rightarrow W^+ g$, if $\chi$ couples to quarks, and through $\chi \eta \rightarrow e^- \gamma$, $\eta \eta^\dagger \rightarrow \gamma \gamma, \gamma Z, ZZ, W^+ W^-$, or $\eta_0 \eta_-^\dagger \rightarrow W^+ \gamma$, if $\chi$ couples to leptons. In order to fully take into account all coannihilation channels in the calculation of the relic density, we use micrOMEGAs~\cite{Belanger:2013oya}, interfaced with FeynRules~\cite{Alloul:2013bka} and CalcHEP~\cite{Belyaev:2012qa} for the numerical solution of the corresponding Boltzmann equation.
In our numerical analysis we will focus in the well motivated scenario where the Dirac dark matter particle is produced via thermal freeze-out in the early Universe. This condition fixes one of the parameters of the model in terms of the remaining two, {\it e.g.} the Yukawa coupling $y = y_{\rm thermal}(m_\chi,m_\eta)$. Then, the parameter space of the model is spanned by only two parameters, for example, the dark matter mass $m_\chi$ and the relative difference between the dark matter mass and the mediator mass, $(m_\eta-m_\chi)/m_\chi$.
\section{Effective Lagrangian for dark matter-nucleon scattering}
\label{sec:L_eff}
The interactions of a Dirac dark matter particle with a nucleon can be described by the Feynman diagrams shown, up to one-loop in perturbation theory, in Fig.~\ref{fig:scattering_diagrams}, where we include for conciseness only one representative diagram of each class. Panel (a) corresponds to the tree-level interaction mediated by the exchange of the scalar $\eta$. Panels (b) and (c) show penguin diagrams mediated by the photon (or the $Z$-boson) and by the Higgs boson, respectively, which arise at the one loop level in all models of Dirac dark matter with a charged scalar mediator, regardless of the choice of the Standard Model fermion. Finally, panel (d) shows a box-diagram with two external gluons, which arises also at the one loop level when the dark matter couples to a colored scalar mediator. Let us discuss separately the effective Lagrangian that arises in each case.
If the dark matter couples at tree level to a first generation quark, the dominant contribution to the scattering cross section is the tree level process $\chi q\rightarrow \chi q$ with the $t$-channel exchange of the scalar mediator $\eta$, depicted in Fig.~\ref{fig:scattering_diagrams}, panel (a). Analogously, the dark matter antiparticle $\bar\chi$ scatters off quarks, with the scalar being in this case exchanged in the $s$-channel. After a Fierz rearrangement of the corresponding matrix elements, these diagrams give rise to a vector interaction of $\chi$ with the quark, which in turn translates into an effective vector interaction of the dark matter particle with the nucleon of the form:
\begin{align}
\label{eq:Leff_tree}
\mathcal{L}_\text{eff,tree} &= f_{V,\text{tree}}^{(N)}\,\bar{\chi} \gamma^\mu \chi \, \bar{N} \gamma_\mu N
\end{align}
and which leads, in the non-relativistic limit, to a spin-independent interaction. Here, the effective couplings for protons and neutrons, $f_{V,\text{tree}}^{(p)}$ and $f_{V,\text{tree}}^{(n)}$, read
\begin{align}
\label{eq:fV_tree}
f^{(p)}_{V,\text{tree}}&= \begin{cases} 2 \, \lambda_V &\text{for coupling to }u_R \\ \lambda_V &\text{for coupling to } d_R\\ 3 \, \lambda_V &\text{for coupling to }(u_L, d_L)\end{cases}
\quad \text{and} \quad
f^{(n)}_{V,\text{tree}}&= \begin{cases} \lambda_V &\text{for coupling to }u_R \\ 2 \, \lambda_V &\text{for coupling to } d_R\\ 3 \, \lambda_V &\text{for coupling to }(u_L, d_L)\end{cases}
\end{align}
with $\lambda_V = y^2/\left[ 8 \left( m_\eta^2- m_\chi^2\right) \right]$. Due to vector-current conservation, the spin-independent tree-level contribution is exactly zero for coupling to second- and third-generation quarks, and is of course also zero for dark matter coupling solely to a lepton. The same diagrams also give rise to an axial vector interaction of the form $f_{A,\text{tree}}^{(N)} \chi \gamma^\mu \gamma^5 \chi \bar{q} \gamma_\mu \gamma_5 q$, which leads to a spin-dependent interaction. The coupling constants of the axial vector and the vector interaction turn out to be comparable in size in this model, therefore, given that the experimental sensitivity to the spin-independent interaction is orders of magnitude better than to the spin-dependent one, this contribution can be in most instances safely neglected. Nevertheless, it will also be included in our numerical analysis for completeness.
\begin{figure}
\begin{center}
\includegraphics[scale=0.6]{./plots/tree_level.pdf}
\includegraphics[scale=0.6]{./plots/photon_Z_exchange.pdf}
\includegraphics[scale=0.6]{./plots/higgs_exchange.pdf}
\includegraphics[scale=0.6]{./plots/gluon_scattering.pdf}\\
(a) \hspace{3.5cm} (b) \hspace{3.5cm} (c) \hspace{3.5cm} (d)
\end{center}
\caption{\small Exemplary Feynman diagrams contributing to the dark matter-nucleon scattering: (a) tree-level scattering, (b) one-loop exchange of a photon or $Z$ boson, (c) one-loop exchange of a Higgs boson, and (d) scattering off gluons at one loop.}
\label{fig:scattering_diagrams}
\end{figure}
On the other hand, and regardless of the choice of the Standard Model fermion, the Yukawa interactions Eq.~(\ref{eq:L_int_fermion_fR}) and Eq.~(\ref{eq:L_int_fermion_fL}) necessarily induce at the quantum level the scattering of dark matter particles with nucleons, due to the penguin diagrams depicted in panels (b) and (c) of Fig.~\ref{fig:scattering_diagrams}. The first diagram induces electromagnetic moments for the dark matter particle, $\chi$, the most relevant ones for Dirac dark matter being the magnetic dipole moment $\mu_\chi$ and the charge radius $b_\chi$, defined by the effective dark matter-photon Lagrangian:\footnote{This diagram also leads in addition to an anapole moment, which corresponds to the effective operator $\bar{\chi} \gamma^\mu \gamma^5 \chi \partial^\nu F_{\mu \nu}$. This operator is suppressed in the non-relativistic limit by the square of the dark matter velocity and is always subdominant in this model. This contribution, however, is also included in our numerical analysis for completeness.}
\begin{align}
\label{eq:L_eff_gamma}
\mathcal{L}_{\text{eff},\gamma} = \frac{\mu_\chi}{2} \bar{\chi} \sigma^{\mu \nu} \chi F_{\mu \nu} + b_\chi \bar{\chi} \gamma^\mu \chi \partial^\nu F_{\mu \nu} \,.
\end{align}
The electromagnetic moments $\mu_\chi$ and $b_\chi$ are obtained by matching the coefficients of this effective Lagrangian to the results of an explicit calculation of the relevant loop diagrams. For dark matter coupling to a right-handed Standard Model fermion $f_R$, and taking the limit where the transferred momentum in the dark matter-nucleus interaction goes to zero, the result reads:\footnote{We have used FeynCalc~\cite{Mertig:1990an} for parts of the computation.}
\begin{align}
\label{eq:mu_chi}
\mu_\chi =& \frac{-Q_f e N_c y^2}{32 \pi^2 m_\chi}\, \bigg[ \frac{-\Delta+1-\mu-\epsilon}{\Delta^{1/2}} \arctanh \left( \frac{\Delta^{1/2}}{\mu+\epsilon-1}\right) + \frac12 (\epsilon-\mu) \log \left( \frac{\epsilon}{\mu} \right) -1 \bigg]\;,\\ \nonumber
b_\chi =& \frac{-Q_f e N_c y^2}{384 \pi^2 m_\chi^2}\, \bigg[ \frac{2}{\Delta^{3/2}} \left( 8 \Delta^2+ (9 \mu+7 \epsilon-5) \Delta - 4 \epsilon (3 \mu+\epsilon- 1) \right) \arctanh\left( \frac{\Delta^{1/2}}{\mu+\epsilon-1}\right) \\
\label{eq:b_chi}
& + (8 \mu-8 \epsilon + 1) \log \left( \frac{\epsilon}{\mu} \right) + 4 \left( 4+ \frac{\mu+3 \epsilon - 1}{\Delta} \right) \bigg]\;.
\end{align}
Here, $Q_f$ is the charge of $f_R$ in units of e (such that e.g. $Q_f=-1$ for coupling to $e_R$), $N_c=3 \, (1)$ for coupling to quarks (leptons) and
\begin{align}
\Delta \equiv \mu^2 + (\epsilon-1)^2-2 \mu (\epsilon+1)\;,
\label{eq:definition_Delta}
\end{align}
with $\mu \equiv m_\eta^2/m_\chi^2$ and $\epsilon \equiv m_f^2/m_\chi^2$ (see also \cite{Chang:2014tea}). For dark matter coupling to a left-handed Standard Model fermion doublet, $\mu_\chi$ and $b_\chi$ are simply given by the sum of the corresponding expressions for both components in the doublet.
The expression for $b_\chi$ presents, when $m_f\ll m_\eta$, a infrared divergence $\propto \log \left( m_f^2/m_\eta^2\right)$, which enhances the scattering rate for light fermions. Note, however, that Eq.~(\ref{eq:b_chi}) was derived in the limit where the transferred momentum goes to zero and is therefore not valid in dark matter scenarios where the fermion mass constitutes the smallest energy scale in the scattering process. More specifically, when the dark matter scatters off xenon nuclei, the typical transferred momentum is $\sim 50$ MeV, therefore for couplings to electrons and first generation quarks Eq.~(\ref{eq:b_chi}) is not valid. For these cases, and in order to cut-off the divergence, we replace in our numerical analysis the fermion mass by 50 MeV. On the other hand, and in contrast to the charge radius, the magnetic dipole moment $\mu_\chi$ has a finite value for $m_f\ll m_\eta$.
Analogously, a dark matter interaction with the $Z$ boson arises at the one-loop level, leading to an effective vector interaction of the form:
\begin{align}
\label{eq:Leff_Z}
\mathcal{L}_{\text{eff},Z} = f_{V,Z}^{(N)} \,\bar{\chi} \gamma^\mu \chi \, \bar{N} \gamma_\mu N \,,
\end{align}
where
\begin{align}
f_{V,Z}^{(p)} = \left( 4 s_W^2-1 \right) \frac{G_F a_Z}{\sqrt{2}} \quad , \quad f_{V,Z}^{(n)} = \frac{G_F a_Z}{\sqrt{2}} \,.
\label{eq:f_VZ}
\end{align}
Here, $a_Z$ is an effective form factor which, for dark matter coupling to a right-handed fermion, is given by
\begin{align}
\label{eq:aZ_fR}
a_Z^{\left(f_R\right)} = \frac{T_3^f N_c y^2\epsilon}{16 \pi^2}\, \bigg[ \frac12 \log \left( \frac{\epsilon}{\mu}\right) +\frac{1+\mu-\epsilon}{\Delta^{1/2}} \arctanh \left( \frac{\Delta^{1/2}}{\mu+\epsilon-1} \right) \bigg]\;,
\end{align}
with $\Delta$ defined as in Eq.~(\ref{eq:definition_Delta}). For the case of dark matter coupling to a left-handed fermion doublet one has $a_Z^{\left(u_L, d_L\right)} = -a_Z^{\left(u_R\right)}-a_Z^{\left(d_R\right)}$ and $a_Z^{\left(\nu_L, e_L\right)} = -a_Z^{\left(e_R\right)}$. Notice that the dark matter effective coupling to the $Z$ boson scales as $(m_f/m_\chi)^2$, making this contribution subdominant in comparison to the photon exchange, except for the interesting scenario where the dark matter couples to $t_R$ or $(t_L, b_L)$.
The penguin diagram shown in panel (c) of Fig.~\ref{fig:scattering_diagrams} induces a coupling of $\chi$ to the Standard Model Higgs boson $h$, which in turn induces the effective dark matter-nucleon interaction
\begin{align}
\label{eq:Leff_higgs}
\mathcal{L}_\text{eff,Higgs} &= f_{\text{S,Higgs}}^{(N)}\,\bar{\chi} \chi \, \bar{N} N\;,
\end{align}
where
\begin{align}
\nonumber
f_{\text{S,Higgs}}^{(N)} =& \frac{-\sqrt{2} G_F m_\chi m_N f_N^{(0)}}{m_h^2} \, \frac{3 \epsilon}{32 \pi^2 \Delta^{1/2}} \\
&\left[ 2 \left( \epsilon^2 + \mu \left(\mu-1 \right)-\epsilon \left(1 + 2\mu \right) \right) \arctanh \left( \frac{\Delta^{1/2}}{\mu+\epsilon-1} \right) +\Delta^{1/2} \left( 2+ \left( \mu-\epsilon \right) \log\left( \frac{\epsilon}{\mu} \right) \right) \right]\;,
\end{align}
with the Higgs-nucleon coupling $f_N^{(0)} \simeq 0.345$~ \cite{Cline:2013gha}.\footnote{The effective dark matter interactions to the Higgs and to the Z-boson have also been presented in \cite{Kumar:2013hfa} in the limit $m_\chi \rightarrow 0$ for a model where the dark matter couples to $t_R$. Our results in that limit agree with the expressions given in that work.} As the coupling of the fermion $f$ to the Higgs is proportional to $m_f$, this contribution is clearly subdominant for dark matter coupling to any of the light Standard Model fermions. In the case of dark matter coupling to a third-generation quark, we find that the contribution from the $Z$ exchange is always much larger, rendering the Higgs exchange diagram subdominant for all possible cases, although it is also included in our numerical results. An additional contribution to the effective dark matter-Higgs interaction arises when the quartic couplings $\lambda_3$ and $\lambda_4$ in Eq.~(\ref{eq:L_int_scalar}) are non-vanishing. However, we find numerically that, even for quartic couplings $\mathcal{O}(1)$, the Higgs interaction gives a subdominant contribution to the dark matter scattering off nucleons.
Finally, if the dark matter couples to quarks, it can scatter off gluons via one-loop box diagrams, such as the one shown in panel (d) of Fig.~\ref{fig:scattering_diagrams}. This generates an effective scalar interaction of $\chi$ with nucleons, which reads
\begin{align}
\label{eq:Leff_gluon}
\mathcal{L}_{\text{eff,gluon}} = f_{S,\text{gluon}}^{(N)} \, \bar{\chi} \chi \bar{N} N \,.
\end{align}
The effective coupling $f_{S,\text{gluon}}^{(N)}$ has been calculated for the case of Majorana dark matter interacting with a colored mediator in~\cite{Drees:1993bu,Hisano:2010ct,Gondolo:2013wwa}. By explicitly writing down the amplitudes for all relevant diagrams, it is straightforward to check that $f_{S,\text{gluon}}^{(N)}$ is exactly the same for Majorana and for Dirac dark matter. Hence, we do not repeat here the rather cumbersome expression for the effective scalar coupling, and we refer instead to these works. Also, following the detailed discussion in~\cite{Gondolo:2013wwa}, we include the gluon contribution to the effective dark matter-nucleon Lagrangian only for the case where dark matter couples to a heavy quark, {\it i.e.} $c$, $b$ or $t$.
\section{Event rate for dark matter-nucleus scattering and experimental limits}
\label{sec:event_rate}
The effective Lagrangians given in Eqs.~(\ref{eq:Leff_tree},~\ref{eq:Leff_Z},~\ref{eq:Leff_higgs},~\ref{eq:Leff_gluon}) describing the interaction of dark matter with nucleons, as well as the effective dark matter-photon interaction defined by Eq.~(\ref{eq:L_eff_gamma}), induce the elastic scattering of dark matter off nuclei. Taking into account all these contributions, the differential scattering cross section of dark matter off a nucleus with mass $m_A$ and spin $J_A$ which recoils with energy $E_R$, reads, including only the leading terms in an expansion in $1/E_R$ and $1/v^2$:
\begin{align}
\label{eq:scattering_crosssection}
\frac{\text{d}\sigma}{\text{d}E_R} = \, &\alpha_{\text{em}} \mu_\chi^2 Z^2 \left( \frac{1}{E_R} -\frac{m_A}{2 \mu_\text{red}^2 v^2} \right) \left[F_{\text{SI}} \left(E_R\right)\right]^2 \\\nonumber
&+ \frac{\mu_A^2 \mu_\chi^2 m_A}{\pi v^2} \,\frac{J_A+1}{3J_A}\, \left[F_{\text{dipole}} \left(E_R\right)\right]^2 \\\nonumber
&+ \frac{m_A}{2 \pi v^2} \left( f^{(A)}\right)^2 \left[F_{\text{SI}} \left(E_R\right)\right]^2 \,,
\end{align}
where $f^{(A)}$ is the effective scalar dark matter-nucleus coupling, defined as
\begin{align}
f^{(A)} = Z \left( f_{S,\text{gluon}}^{(p)} + f_{V,\text{tree}}^{(p)} + f_{V,Z}^{(p)} - e b_\chi - \frac{e \mu_\chi}{2 m_\chi} \right) + (A-Z) \left( f_{S,\text{gluon}}^{(n)} + f_{V,\text{tree}}^{(n)} + f_{V,Z}^{(n)}\right) \,.
\end{align}
For $F_{\text{SI}} \left( E_R \right)$, we use the Helm form factor given in~\cite{Lewin:1995rx}, while for the form factor $F_{\text{dipole}} \left(E_R\right)$ associated to the dipole-dipole scattering we follow~\cite{Banks:2010eh}.
Finally, we translate the scattering cross section into an event rate in a xenon experiment and we compare the predicted rate with the current limit by the LUX experiment~\cite{Akerib:2013tjd}, which provides the strongest sensitivity for the model discussed in this paper. It is important to remark that one can not simply map the limits on $\sigma_\text{SI}^p$ as they are published by the corresponding collaborations onto limits on the parameter space of the model considered in this work, due to basically two reasons: first, the scattering diagrams involving the exchange of a photon only induce a dark matter coupling to protons, but not to neutrons, leading to a (potentially large) source of isospin violation in the scattering. Secondly, due to the long-range force induced by the magnetic dipole moment of dark matter, the differential scattering cross section given in Eq.~(\ref{eq:scattering_crosssection}) has a different dependence on the recoil energy $E_R$ compared to the case of standard spin-independent or spin-dependent interactions. In our numerical analysis, we derive the corresponding upper limits arising from the null-result of LUX, fully taking into account the isospin violation as well as the non-standard dependence of the differential scattering cross section with $E_R$. We provide the details of our method of calculating the event rate in LUX in Appendix~\ref{sec:appendix_lux_limits}.
We will also calculate the regions of the parameter space which can be probed by the future experiments XENON1T~\cite{Aprile:2012zx}, already under construction since 2013, and DARWIN, a design study for a next generation
multi-ton dark matter detector~\cite{Baudis:2012bc}. Following~\cite{Cushman:2013zza}, we assume that XENON1T will have a reach in $\sigma_p^{\text{SI}}$ a factor of 40 better than the present LUX limit for dark matter masses above $\simeq 100$ GeV. For the model discussed in this work, this translates into a rescaling of the upper limits on the Yukawa coupling $y$ arising from the LUX data by a factor of $40^{1/4} \simeq 2.5$. Here, we refrain from deriving prospects for XENON1T for small dark matter masses, i.e. $m_\chi \lesssim 30$ GeV, as these will depend on so far unknown details of the detector performance near threshold. For DARWIN, we consider the most optimistic scenario where the reach in cross section reaches $\simeq 10^{-48} \text{ cm}^2$~\cite{Baudis:2012bc}, thereby approaching the irreducible background of coherent neutrino-nucleus scattering, and we derive prospects by rescaling our limits on the Yukawa coupling obtained from the LUX data by a factor of $1800^{1/4} \simeq 6.5$. As for XENON1T, we do not consider prospects for low-mass dark matter. For simplicity we also assume that DARWIN will consist only of xenon, although currently both liquid argon and xenon are discussed as possible target materials.
\begin{figure}[t]
\hspace{-0.7cm}
\includegraphics[scale=1.05]{./plots/MostImportantContris_Dirac_uR.pdf}
\hspace{-1.0cm}
\includegraphics[scale=1.05]
{./plots/MostImportantContris_Dirac_cR.pdf}
\hspace{-1.0cm}
\includegraphics[scale=1.05]{./plots/MostImportantContris_Dirac_tR.pdf}
\hspace{-0.7cm}
\includegraphics[scale=1.05]{./plots/MostImportantContris_Dirac_dR.pdf}
\hspace{-1.0cm}
\includegraphics[scale=1.05]
{./plots/MostImportantContris_Dirac_sR.pdf}
\hspace{-1.0cm}
\includegraphics[scale=1.05]{./plots/MostImportantContris_Dirac_bR.pdf}
\hspace{-0.7cm}
\includegraphics[scale=1.05]{./plots/MostImportantContris_Dirac_uL.pdf}
\hspace{-1.0cm}
\includegraphics[scale=1.05]
{./plots/MostImportantContris_Dirac_cL.pdf}
\hspace{-1.0cm}
\includegraphics[scale=1.05]{./plots/MostImportantContris_Dirac_tL.pdf}
\caption{\small Dominant contributions to the dark matter scattering rate in LUX, for coupling to right-handed up-type quarks (upper panels), right-handed down-type quarks (middle panels), and left-handed quark doublets (lower panels).}
\label{fig:MostImportantContris_QuarkCoupling}
\end{figure}
\begin{figure}
\hspace{-0.7cm}
\includegraphics[scale=1.05]{./plots/MostImportantContris_Dirac_e.pdf}
\hspace{-1.0cm}
\includegraphics[scale=1.05]
{./plots/MostImportantContris_Dirac_m.pdf}
\hspace{-1.0cm}
\includegraphics[scale=1.05]{./plots/MostImportantContris_Dirac_l.pdf}
\caption{\small Dominant contributions to the dark matter scattering rate in LUX, for coupling to leptons.}
\label{fig:MostImportantContris_LeptonCoupling}
\end{figure}
Before presenting the actual limits on the model, let us first discuss the relative importance of the different contributions to the scattering rate, which were introduced in Section~\ref{sec:L_eff} and were schematically shown in Fig.~\ref{fig:scattering_diagrams}. Which of these processes gives the largest contribution to the event rate in a direct detection experiment depends on which Standard Model fermion the dark matter particle couples to, as well as on the dark matter and mediator masses. In Figs.~\ref{fig:MostImportantContris_QuarkCoupling} and~\ref{fig:MostImportantContris_LeptonCoupling} we show, separately for each coupling scheme, which processes dominate the scattering in LUX, for given dark matter and mediator masses, $m_\chi$ and $m_\eta$. The left panels of Fig.~\ref{fig:MostImportantContris_QuarkCoupling} correspond to a scenario where the dark matter particle couples to right-handed up-quarks (top left), right-handed down-quarks (central left), or to the first generation of left-handed quarks (bottom left). As expected, in these cases the tree-level scattering of dark matter off the nucleons dominate the event rate, rendering all one-loop processes irrelevant. On the other hand, as discussed in Section~\ref{sec:L_eff}, the tree-level contribution is negligible if dark matter couples to quarks of the second or third generation, or to leptons. In all these scenarios, the one-loop processes dominate the event rate in LUX (and other direct detection experiments), as shown in Figs.~\ref{fig:MostImportantContris_QuarkCoupling} and~\ref{fig:MostImportantContris_LeptonCoupling}. More precisely, the scattering via photon exchange, which leads to the charge radius and magnetic dipole moment operators, is the dominant contribution to the cross section for large parts of the parameter space of each scenario. In particular, the charge radius operator typically leads to the largest event rate for $m_\chi \lesssim 1$ TeV, while the magnetic dipole moment dominates for $m_\chi \gtrsim 1$ TeV, although the concrete values depend on the coupling scheme and on the mass splitting between the mediator $\eta$ and the dark matter particle $\chi$. This behavior can be qualitatively understood from the dimensionality of each effective operator describing the coupling of dark matter to photons (see Eq.~(\ref{eq:L_eff_gamma})), and which is dimension five for the case of magnetic dipole interactions, while it is dimension six for the charge radius. Hence, the corresponding coefficients of these operators scale differently with the dark matter mass, namely as $\mu_\chi \sim m_\chi^{-1}$ and $b_\chi \sim m_\chi^{-2}$, respectively.
On the other hand, if the dark matter particle couples to the top-quark, the scattering via the exchange of a $Z$ boson becomes relevant, and generically dominates over the photon exchange for masses below $\simeq 5 - 10$ TeV, as shown in the right upper and right bottom panels of Fig.~\ref{fig:MostImportantContris_QuarkCoupling}. As mentioned in Section~\ref{sec:L_eff}, this is due to the dependence of the effective coupling $a_Z$ with the mass of the fermion coupling to $\chi$: from Eq.~(\ref{eq:aZ_fR}) it follows that for $m_f \ll m_\chi$ this effective coupling scales as $a_Z \propto (m_f/m_\chi)^2$, making it relevant only for the case of $\chi$ coupling to $t_R$ or $(b_L, t_L)$. Finally, we find that the dark matter-nucleon scattering induced by the one-loop coupling of dark matter to gluons is only relevant for rather small dark matter masses and small mass splittings $m_\eta/m_\chi$, for dark matter coupling to a heavy quark, as illustrated in the corresponding panels in Fig.~\ref{fig:MostImportantContris_QuarkCoupling}.
We finally determine the regions of the parameter space which are excluded by the LUX experiment for each of these scenarios, as well as the regions that can be probed with the future XENON1T and DARWIN experiments, fixing the value of the Yukawa coupling at each point to the value $y_\text{thermal}$ required to reproduce the observed dark matter abundance via thermal freeze-out. The results for the case of dark matter coupling to quarks are shown in Fig.~\ref{fig:thermal_limits_quarks}, and for coupling to leptons in Fig.~\ref{fig:thermal_limits_leptons}. The dotted black contour lines show the value of the Yukawa coupling which leads, at that point of the parameter space, to the observed dark matter abundance $\Omega_{\rm DM} h^2\simeq 0.12$. Besides, for each of the possible scenarios, the region of the parameter space corresponding to low mass and small mass splitting is theoretically inaccessible since the coannihilations suppress the relic abundance of $\chi$ below the observed value. This region is shown in Figs.~\ref{fig:thermal_limits_quarks} and ~\ref{fig:thermal_limits_leptons} as dark grey. On the other hand, large masses and/or large mass splittings correspond to Yukawa couplings $y>\sqrt{4\pi}$, where our perturbative calculation loses validity.\footnote{This condition is motivated by the requirement that the expansion parameter $\alpha_y \equiv y^2/4 \pi$ for higher-order corrections should be smaller than one.} It is worthwhile noticing that this requirement sets an upper limit on the dark matter mass which approximately reads $m_{\rm DM}\lesssim 20$ TeV for coupling to quarks and $m_{\rm DM}\lesssim 15 $ TeV for coupling to leptons.
\begin{figure}[t!]
\hspace{-0.7cm}
\includegraphics[scale=1.05]{./plots/ThermalLimits_Dirac_uR.pdf}
\hspace{-1.0cm}
\includegraphics[scale=1.05]
{./plots/ThermalLimits_Dirac_cR.pdf}
\hspace{-1.0cm}
\includegraphics[scale=1.05]{./plots/ThermalLimits_Dirac_tR.pdf}
\hspace{-0.7cm}
\includegraphics[scale=1.05]{./plots/ThermalLimits_Dirac_dR.pdf}
\hspace{-1.0cm}
\includegraphics[scale=1.05]
{./plots/ThermalLimits_Dirac_sR.pdf}
\hspace{-1.0cm}
\includegraphics[scale=1.05]{./plots/ThermalLimits_Dirac_bR.pdf}
\hspace{-0.7cm}
\includegraphics[scale=1.05]{./plots/ThermalLimits_Dirac_uL.pdf}
\hspace{-1.0cm}
\includegraphics[scale=1.05]
{./plots/ThermalLimits_Dirac_cL.pdf}
\hspace{-1.0cm}
\includegraphics[scale=1.05]{./plots/ThermalLimits_Dirac_tL.pdf}
\caption{\small Regions of the parameter space for Dirac dark matter particles with scalar mediator excluded by the LUX experiment, assuming coupling to right-handed up-quarks (upper panels), right-handed down-quarks (middle panels), and left-handed quark doublets (lower panels), and assuming that the dark matter density of the Universe is generated by thermal freeze-out (the necessary value of the Yukawa coupling is indicated by the dotted black contour lines). The dark gray shaded regions in the lower left corners of each figure are theoretically inaccessible, due to efficient coannihilations which deplete the dark matter density below the observed value, while the light gray shaded regions violate our perturbativity condition $y_\text{thermal} < \sqrt{4 \pi}$. }
\label{fig:thermal_limits_quarks}
\end{figure}
The left panels in Fig.~\ref{fig:thermal_limits_quarks} correspond to scenarios where $\chi$ couples to first generation quarks, namely to right-handed up-quarks (top panel), right-handed down-quarks (central panel), and to the first generation left-handed quark doublet (lower panel). As apparent from the plots, practically the whole parameter space for a thermal relic compatible with our perturbativity condition is already excluded by the LUX experiment, with XENON1T being able to constrain dark matter masses up to $\simeq 10$ TeV, thereby probing the whole parameter space of the model. These strong direct detection bounds result from the unsuppressed tree-level coupling of the dark matter particle $\chi$ to light quarks, \textit{cf.} Eq.~(\ref{eq:Leff_tree}).
On the other hand, for coupling to quarks of the second generation, the tree-level coupling is negligible and the limits dominantly arise from the dark matter electromagnetic moments, \textit{cf.} Fig.~\ref{fig:MostImportantContris_QuarkCoupling}. Consequently, the constraints from LUX are considerably weaker, as apparent from the second column of Fig.~\ref{fig:thermal_limits_quarks} for dark matter coupling to $c_R$ (top panel), $s_R$ (central panel), and $(c_L, s_L)$ (lower panel). Nevertheless, it is worth pointing out that despite the suppression of the one-loop diagrams, LUX already excludes dark matter masses up to $\simeq 100-500$ GeV, depending on the mass splitting of $\chi$ and $\eta$ as well as on the precise coupling scheme of dark matter. Most interestingly, in these scenarios there are exciting prospects for the upcoming direct detection experiments to probe large parts of the viable parameter space. For example, a DARWIN-like experiment would be able to cover practically the whole parameter space of thermally produced dark matter, compatible with our perturbativity condition, despite the fact that all direct detection processes occur only at the one-loop level. Furthermore it is worthwhile remarking that direct detection experiments have an enhanced sensitivity in the case of mass-degenerate scenarios, i.e. $m_\eta/m_\chi \simeq 1$, as apparent from the figures. This larger sensitivity follows from the parametric dependence of the charge radius operator $ b_\chi \propto 1/( (m_\eta/m_\chi)^2 - 1)$, which leads to an enhancement of the scattering cross section as the mass spectrum becomes more and more degenerate.
Lastly, for dark matter coupling to third-generation quarks, the constraints on the parameter space are presented in the right panels of Fig.~\ref{fig:thermal_limits_quarks}. In the case of coupling to $t_R$, shown in the upper right panel, we restrict our analysis to $m_\chi > m_t \simeq 175$ GeV, which ensures that the process $\chi \bar{\chi} \rightarrow t \bar{t}$ is kinematically allowed during freeze-out in all of the considered parameter space. In principle, a thermal relic can also exist for $m_\chi < m_t$, however in that case the relic density would instead be set by one-loop annihilations or two-to-three processes, changing the phenomenology of the model quite dramatically. For an extensive discussion of the model in that part of the parameter space we refer to~\cite{Kumar:2013hfa}. As argued in section \ref{sec:L_eff}, if the dark matter particle couples to either $t_R$ or $(t_L, b_L)$, the scattering rate in direct detection experiments is typically dominated by the $Z$ exchange contribution, while it is dominated by the electromagnetic moments for coupling to the right-handed bottom quark. Then, for coupling to the right-handed top quark, LUX excludes masses up to $200 - 400$ GeV, depending on the mass splitting of $\eta$ and $\chi$, while the constraints for coupling to $b_R$ (central panel) are somewhat weaker due to the absence of a large effective dark matter coupling to the $Z$ boson. Again, it is important to point out that XENON1T and eventually DARWIN will probe a significantly larger region of the parameter space than LUX, and will be able to rule out, or hopefully discover, dark matter particles with masses up to several TeV. Similar conclusions also hold for the case of $\chi$ coupling to the third generation quark-doublet, shown in the lower right panel of Fig.~\ref{fig:thermal_limits_quarks}.
Direct search experiments are subject to astrophysical and nuclear uncertainties which might modify the conclusions presented above. Hence, and in order to robustly exclude a region of the parameter space or to confirm a possible signal, it is important to exploit the complementarity with indirect and collider searches, which rely on very different physical processes, and which are affected by different systematic uncertainties. A thorough investigation of the complementarity among the different search strategies for Dirac dark matter is beyond the scope of this paper and here we will limit ourselves to briefly comment on the qualitative features of the indirect and collider searches. If the colored mediator is sufficiently light, it might be directly produced at the LHC via strong interactions. If, moreover, the mass difference between the scalar mediator and the dark matter particle is sizable, the subsequent decay produces a characteristic signal consisting in two jets plus missing transverse momentum, which has been searched for in the framework of simplified supersymmetric models \cite{Chatrchyan:2014lfa,Aad:2014wea,Aad:2014vma,Khachatryan:2014qwa}. In contrast, when the mass difference is small, the jets might be too soft to be detected and a different search strategy should be pursued, for instance the search for monophoton~\cite{Aad:2014tda} or monojet events~\cite{Aad:2014nra} plus missing transverse momentum. Finally, for very heavy mediators, only the dark matter particle can be pair-produced at the collider. In this regime, dark matter signals could be detected in searches for monojet, monophoton or mono-W/Z boson plus missing transverse momentum~\cite{Aad:2015zva,atlasMono,Khachatryan:2014rra}. Present collider limits constrain some regions of the parameter space with $m_\chi\lesssim 200$ GeV, although the precise reach of the limits crucially depends on the details of the model. Additional constraints from colliders follow from electroweak precision observables, notably the $Z$-boson decay width~\cite{ALEPH:2005ab}, and which exclude the existence of scalar mediators with mass $m_\eta\lesssim 40$ GeV.
On the other hand, the strongest indirect detection limits on this class of scenarios stem from the non-observation of an excess of gamma-rays correlated to the direction of dwarf galaxies. A recent analysis by the Fermi-LAT collaboration from a stacked analysis of 15 dwarf spheroidal galaxies excludes $m_{\rm DM}\lesssim 100$ GeV for dark matter coupling to the $b$-quark~\cite{Ackermann:2015zua}, which is competitive, and for some choices of parameters better, than the LUX limits shown in Fig.~\ref{fig:thermal_limits_quarks}. Additional constraints on the model can be derived from gamma-ray observations of the Milky Way center, although this search is affected by large, and still not fully understood, backgrounds. Interestingly, some authors have claimed evidence for a gamma-ray excess in this region~\cite{Goodenough:2009gk,Hooper:2010mq,Hooper:2011ti,Abazajian:2012pn,Daylan:2014rsa,Calore:2014xka}, which can be interpreted in terms of dark matter annihilations. Dirac dark matter with a charged mediator constitutes an excellent particle physics framework to explain the excess (in contrast to Majorana or scalar dark matter, which have annihilation cross sections at the Galactic center which are helicity- and velocity-suppressed). The mass ranges favored by the dark matter interpretation of the galactic center excess (GCE) for the annihilation final states selected in~\cite{Calore:2014nla}, and which point to an annihilation cross section of the order of the thermal value, are shown in Fig.~\ref{fig:thermal_limits_quarks}, as vertical hatched bands. It follows from the figure that direct detection experiment can test, for some channels, this interpretation of the excess, especially for couplings to first generation quarks, to $s_R$ or to $c_R$. One should bear in mind that the astrophysical uncertainties in the extraction of the dark matter parameters from the GeV excess are still large. Therefore, the value of the annihilation cross section can be significantly lower than the thermal value and accordingly the impact of the LUX limits.
Finally, we turn to scenarios in which the dark matter particle only couples to one of the right- or left-handed leptons of the Standard Model. As discussed in Section~\ref{sec:L_eff}, in this scenario, the tree-level interactions vanish, however the one-loop diagrams induce significant scattering rates in direct detection experiments. In fact, many of the qualitative features of the scenario where $\chi$ couples to heavy quarks also apply to this case. The parameter space for the various Dirac dark matter scenarios with coupling to leptons are shown in Fig.~\ref{fig:thermal_limits_leptons}, where again the region of the parameter space incompatible with thermal production and perturbativity are shown, respectively, as dark and light gray shaded regions. For coupling to leptons, the region of parameter space for which no thermal relic exists due to strong coannihilations is smaller compared to the case of dark matter coupling to quarks, since for leptophilic dark matter the relevant coannihilation channels (such as $\eta \eta^\dagger \rightarrow \gamma \gamma$) are suppressed by powers of the electromagnetic coupling, while for coupling to quarks, the corresponding processes are induced by the strong coupling between the scalar $\eta$ and the gluons. Besides, in the case of dark matter coupling to leptons, for $m_\chi \simeq m_\eta \simeq 45$ GeV the coannihilation process $\eta \eta^\dagger \rightarrow f \bar{f}$ is resonantly enhanced due to the diagram where the $Z$ boson is exchanged in the $s$-channel. Hence, around that dark matter mass, the region in parameter space where no thermal relic can exist due to coannihilations is extended to higher mass splittings.
The region of parameter space excluded by LUX, as well as the range of masses which will be accessible to XENON1T and DARWIN are very similar for the six different scenarios shown in Fig.~\ref{fig:thermal_limits_leptons}. For all mass splittings between the mediator $\eta$ and the dark matter particle $\chi$, LUX excludes dark matter masses between $8 \text{ GeV} \lesssim m_\chi \lesssim 100$ GeV, and XENON1T will be sensitive up to $m_\chi \simeq 500$ GeV. Moreover, as noted already before, the charge radius operator $b_\chi$ is enhanced for small mass splittings, leading to an increased sensitivity of the direct detection experiments for $m_\eta/m_\chi \rightarrow 1$. Moreover, from these figures it follows that a multi-ton detection experiment like DARWIN will practically cover the whole parameter space of these scenarios, again despite the fact that all direct detection signals are one-loop suppressed.
Scenarios with Dirac dark matter coupling to leptons can also be probed in collider or indirect search experiments, which give complementary limits on the parameter space of the model. The charged scalar mediator can be pair-produced via the Drell-Yan process in colliders and subsequently decay producing a characteristic signature of opposite-sign, same-flavor leptons, plus missing transverse momentum. This search has been undertaken in the context of slepton searches at the LHC by the ATLAS~\cite{Aad:2014vma} and CMS collaborations~\cite{Khachatryan:2014qwa}, and at LEP II by the ALEPH, DELPHI, L3 and OPAL collaborations~\cite{LEP-sleptons}, and exclude some regions of the parameter space with $m_\chi\lesssim 100$ GeV (see \cite{Garny:2015wea} for a compilation of the limits of the parameter space of a Majorana dark matter particle that couples to a lepton and a scalar mediator, and which can also be applied to Dirac dark matter, since the production mechanism of scalar mediators and their decay modes are identical in both cases). Finally, the precise measurement of the $Z$-boson decay width at LEP~\cite{ALEPH:2005ab} excludes mediators with mass $m_\eta\lesssim 40$ GeV.
Stringent limits on this scenario also follow from indirect searches. The annihilation into leptons produces an additional component in the cosmic positron flux which is severely constrained by the AMS-02 data on the positron fraction~\cite{Aguilar:2013qda,Accardo:2014lma}, defined as the flux of positrons divided by the total flux of electrons plus positrons, as well as the data on the positron flux~\cite{Aguilar:2014mma}. More specifically, the non-observation of the sharp feature in the positron fraction produced by the hard leptons produced in these annihilation channels translates into strong limits on the parameters of the model. From the limits on the cross section derived in~\cite{Bergstrom:2013jra} one can exclude $m_{\chi}\lesssim 120$ GeV for coupling to $e_R$, $m_{\chi}\lesssim 70$ GeV for coupling to $\mu_R$ and $m_{\chi}\lesssim 30$ GeV for coupling to $\tau_R$. The more conservative limits from~\cite{Ibarra:2013zia} exclude $m_{\chi}\lesssim 100$ GeV and $m_{\chi}\lesssim 40$ GeV for couplings to $e_R$ and $\mu_R$, respectively. Strong constraints on these channels can also be derived from dwarf galaxy observations, which exclude $m_\chi\lesssim 100$ GeV~\cite{Ackermann:2015zua}.
\begin{figure}[t!]
\hspace{-0.7cm}
\includegraphics[scale=1.05]{./plots/ThermalLimits_Dirac_eR.pdf}
\hspace{-1.0cm}
\includegraphics[scale=1.05]
{./plots/ThermalLimits_Dirac_mR.pdf}
\hspace{-1.0cm}
\includegraphics[scale=1.05]{./plots/ThermalLimits_Dirac_lR.pdf}
\hspace{-0.7cm}
\includegraphics[scale=1.05]{./plots/ThermalLimits_Dirac_eL.pdf}
\hspace{-1.0cm}
\includegraphics[scale=1.05]{./plots/ThermalLimits_Dirac_mL.pdf}
\hspace{-1.0cm}
\includegraphics[scale=1.05]{./plots/ThermalLimits_Dirac_lL.pdf}
\caption{\small Same as Fig.~\ref{fig:thermal_limits_quarks}, but for dark matter coupling to right-handed charged leptons (upper panels) and to left-handed lepton-doublets (lower panels).}
\label{fig:thermal_limits_leptons}
\end{figure}
\section{Conclusions}
\label{sec:conclusions}
In some scenarios, the dark matter particle only couples to the nucleons via quantum effects, with a scattering rate which is expected to be very small. On the other hand, the excellent sensitivity of current direct search experiments might allow to identify such rare events over the environmental and instrumental backgrounds, thus opening the possibility of probing a class of models which were not accessible to previous experiments, such as those where the dark matter couples only to leptons or to heavy quarks.
In this paper, we have investigated this exciting possibility in the framework of a toy model where the dark matter is a Dirac fermion, singlet under the Standard Model gauge group, which couples to one of the Standard Model fermions via a Yukawa coupling with a scalar mediator. For reasonable values of the parameters of the model, the observed dark matter abundance $\Omega_{\rm DM} h^2\simeq 0.12$ can be explained via thermal freeze-out. In this model, the scattering rate with nucleons receives various contributions, which depend on the choice of the Standard Model fermion. For couplings to first generation quarks, the dark matter interacts with the nucleons at tree level resulting in a comparatively large scattering rate. On the other hand, for couplings to any Standard Model fermion, the dark matter interacts with the nucleon via one-loop penguin diagrams mediated by the photon, the $Z$ boson or the Higgs boson. Lastly, for couplings to any colored fermion, the dark matter interacts with the nucleon also via one-loop box diagrams with two external gluons.
Remarkably, current experiments have reached the necessary sensitivity to probe these quantum-induced dark matter interactions with nucleons. More specifically, assuming thermal production for the dark matter population in our Universe, the current LUX data probe the parameter space of {\it all} scenarios considered in this paper, posing in some cases very stringent constraints on the parameter space of the model. For instance, the LUX data exclude dark matter masses below $\sim 3$ TeV for coupling to first generation quarks, masses between $\sim 10-100$ GeV for coupling to the right-handed charm or strange quarks, and masses also between $\sim 10-100$ GeV for coupling to leptons (of any generation and chirality). The future XENON1T experiment will continue closing in on the parameter space of the model and will be able to probe dark matter masses up to several hundreds of GeV, while in the longer term, a multi-ton xenon detector such as DARWIN will be able to exclude masses as large as a few TeV. Direct search experiments will then have a high sensitivity to this class of scenarios, even for multi-TeV dark matter masses, which are difficult to probe at the LHC.
\vspace{0.5cm}
\section*{Acknowledgements}
We are grateful to Riccardo Catena for useful discussions. This work was partially supported by the DFG cluster of excellence ``Origin and Structure of the Universe,'' the TUM Graduate School and the Studienstiftung des Deutschen Volkes.
\vspace{0.5cm}
|
\section{Introduction}
Supersymmetric (SUSY) extensions of the standard model (SM) at the weak scale have the advantages of solving the hierarchy problem, providing a dark matter candidate, and leading to gauge coupling unification. Even though weak scale SUSY is by no mean ruled out, lack of results at the LHC forces one to reconsider whether the Higgs mass might be fine-tuned to a certain degree. This is the case in Split-SUSY models \cite{Wells:2003tf, ArkaniHamed:2004fb, Giudice:2004tc} where fermion superpartners can still be close to the electroweak scale while the scalar superpartners are much heavier. These theories are no longer solutions to the hierarchy problem (which could be explained by an environmental selection principle for example), but maintain a dark matter candidate and can keep intact gauge coupling unification \cite{Wells:2003tf, ArkaniHamed:2004fb, Giudice:2004tc}. However, it was shown \cite{Arvanitaki:2012ps} that scalars heavier than $10^5$ TeV would make it difficult to reconcile Split-SUSY with the known mass of the Higgs boson \cite{Chatrchyan:2013mxa, Aad:2014aba}, therefore putting an upper limit on this splitting. These Split-SUSY theories with only a small gap are referred to as Mini-Split \cite{Arvanitaki:2012ps}.
One of the main phenomenological characteristics of Mini-Split models is the presence of a small hierarchy between the gauginos and the scalars. The conventional gaugino mass spectra associated to well-known mediation mechanisms like anomaly mediation \cite{Randall:1998uk, Giudice:1998xp} and gauge mediation \cite{Dine:1981gu, Nappi:1982hm, AlvarezGaume:1981wy, Dine:1993yw, Dine:1994vc, Dine:1995ag} are then modified, as the heavy superpartners deflect the gaugino masses from their standard renormalization group (RG) trajectory when they are integrated out. The resulting spectra are referred to as deflected anomaly mediation \cite{ Rattazzi:1999qg, Bagnaschi:2014rsa} or deflected gauge mediation. The precise phenomenology of Mini-Split models depends on the value of $\mu$ which could either be at the electroweak scale or at the same scale as the scalars. In this work, we focus on the case of large $\mu$. The case of small $\mu$ was considered in \cite{Jung:2013zya} which provides future prospects for anomaly mediation in Mini-Split theories at a 100 TeV collider with light Higgsinos (which minimizes the amount of deflection) and applies these results to gauge and mirror mediation. Reference \cite{Jung:2013zya} also studied cases with a large $\mu$ (50 TeV) but still somewhat smaller than what is considered in most of the parameter space we consider. Dark matter predictions for such models are presented in \cite{Cesarini:2006jp, Yokozaki:2009fu}. Other variants of deflected mediation are studied in \cite{Setzer:2010bz, Okada:2002mv, Okada:2012nr, deBlas:2011cr, Ibe:2006de, Ibe:2011aa, Ibe:2012hu, Bhattacherjee:2012ed, Kahn:2013pfa}.
The purpose of this paper is to constrain the parameter space of Mini-Split models with deflected anomaly and gauge mediation using LHC data and to predict future exclusion and discovery prospects at LHC 14 and a future 100 TeV collider. Current constraints are extracted from ATLAS \cite{TheATLAScollaboration:2013zia, Aad:2014vma, Aad:2014nua, Aad:2014pda, Aad:2014lra} and CMS \cite{Khachatryan:2014mma, Chatrchyan:2014lfa, CMS:2013ija} SUSY searches (mainly gluino pair production), the known mass of the Higgs boson \cite{Chatrchyan:2013mxa, Aad:2014aba}, and the absence of a color-breaking vacuum \cite{Bagnaschi:2014rsa}. Future prospects for LHC 14 and a 100 TeV collider are obtained by using the same theoretical tools in conjunction with background estimates. In the cases studied here, the deflection comes mainly from the Higgsino sector \cite{Bagnaschi:2014rsa, ArkaniHamed:2012gw}, which is assumed to be around the scalar scale and the light neutralinos/charginos are almost pure gauginos. As one generally expects the third generation of squarks to be lighter because of renormalization group effects for example, this paper makes the simplifying assumption of a slightly lighter third generation.
This paper is organized as follows. The necessary theoretical elements are presented first. This includes an explanation of how Mini-Split theories can arise in both anomaly and gauge mediation, as well as pole mass expressions and branching fractions. The procedure necessary to calculate the Higgs mass is also presented. The methodology used in obtaining both current limits and future prospects is then explained. This includes the LHC searches used to determine current limits. Finally, we present current LHC constraints and prospects at LHC 14 and a future 100 TeV collider.
\section{Theory}
\subsection{Mini-Split models}\label{SectionMiniSplitModels}
In this section, we review how Mini-Split spectra can be realized in both anomaly and gauge mediation (see for example \cite{Bagnaschi:2014rsa}). Quite generally, sfermions masses can be generated via terms of the form
\begin{equation}\label{Squarkmassesgeneration}
\int d^4 \theta \frac{X^{\dagger}X}{M_{\ast}^2}Q^{\dagger}Q,
\end{equation}
where $M_{\ast}$ is the mediation scale, $X=\theta^2 F_{X}$ is a SUSY breaking spurion, and $Q$ is a chiral superfield. This term is always allowed by symmetries, irrespective of the R-charge of $X$ or its gauge quantum numbers. On the other hand gaugino masses are generated via terms of the form
\begin{equation}
\int d^2 \theta \frac{X}{M_{\ast}}W_{i\alpha}W^\alpha_i,
\end{equation}
where $W_{i\alpha} (i = 1,2,3)$ are the gauge-strength superfields. Contrary to the sfermion masses of (\ref{Squarkmassesgeneration}), here $X$ is required to be a singlet under all gauge and global charges in order for this term to be allowed. It is therefore easier to forbid, and in the models that we consider we assume that it is absent. There is however an unavoidable contribution to gaugino masses coming from anomaly mediation
\begin{equation}
M_i=\frac{\beta_i}{g_i}m_{3/2}.
\end{equation}
The A-terms are also generated by anomaly mediation and are given by
\begin{equation}\label{eqAterm}
A_y=-\frac{\beta_y}{y}m_{3/2},
\end{equation}
where $y$ is the corresponding Yukawa and $\beta_y$ is its beta function. A $B_\mu$ term can be generated by a term of the form
\begin{equation}
\int d^4 \theta \frac{X^{\dagger}X}{M_{\ast}^2}H_u H_d.
\end{equation}
In Mini-Split scenarios, the $\mu$ term can either be large (at the scale of the scalars) or small (at the scale of the gauginos) depending on how it is generated. In this work we concentrate on the case where it is large, which could be generated through the Giudice-Masiero mechanism \cite{Giudice:1988yz} where a term of the following form is introduced
\begin{equation}
\int d^4\theta \Phi^{\dagger}\Phi\left[\hat{H}_{u,d}^{\dagger}\hat{H}_{u,d}+\left(c \hat{H}_u\hat{H}_d+\text{h.c}\right)\right].
\end{equation}
Here $c$ is an arbitrary dimensionless constant and $\Phi$ is the conformal compensator which gets a non-zero $F$-term once SUSY is broken: $\Phi = 1 - m_{3/2} \theta^2$ . Upon rescaling of the fields, this becomes
\begin{equation}
\int d^4\theta \left[H_{u,d}^{\dagger}H_{u,d}+\left(c \frac{\Phi^{\dagger}}{\Phi}H_u H_d+\text{h.c}\right)\right]
\end{equation}
and leads to a $\mu$ term, in addition to an additional contribution to $B_{\mu}$. These terms are of order $m_{3/2}$ and $m_{3/2}^2$ respectively. If gravity is the sole mediator of supersymmetry breaking, then $M_\ast$ is the Planck mass and this leads to the scalars and Higgsinos all having masses of roughly $m_{3/2}$ while the masses of the gauginos are a loop factor smaller, leading to a Mini-Split spectrum. The fact that the $\mu$ term is taken to be large will change the running of the gauge couplings compared to the more conventional split-spectrum with light Higgsinos. The prediction for $\alpha_s(M_Z)$ was found in \cite{ArkaniHamed:2012gw} to be smaller than with light Higgsinos, but still consistent with the measured value.
Gauge mediation can also lead to Mini-Split spectra. This can be done in a multitude of ways. We give an example taken from \cite{Arvanitaki:2012ps}. Assume a superpotential of the form
\begin{equation}
W=M_R\left(\Phi_1\overline{\Phi}_1+\Phi_2\overline{\Phi}_2\right)+X\Phi_1\overline{\Phi}_2,
\end{equation}
where the $\Phi_i$ and the $\overline{\Phi}_i$ are messengers and $X=M+F\theta^2$ is a spurion that breaks SUSY and R-symmetry. This leads to gauginos masses of
\begin{equation}
\label{eq:gaugemediation}
M_i=\frac{\alpha_i}{6\pi}\frac{M}{M_R}\frac{F^3}{M_R^5}+\mathcal{O}\left(\frac{M^3}{M_R^3}\frac{F^3}{M_R^5},\frac{F^5}{M_R^9}\right).
\end{equation}
On the other hand, the scalars masses are $\mathcal{O}(\alpha F/M_R)$. If R-symmetry is weakly broken $(M < M_R)$, a Mini-Split spectrum is again generated.
\subsection{Gaugino mass spectrum}
The main effect of the small mass hierarchy between the gauginos and scalars/Higgsinos is that radiative corrections to the pole masses of gauginos coming from integrating out the scalars and Higgsinos can be comparable to, if not larger than, the contributions coming from anomaly mediation or gauge mediation directly. In the case of anomaly mediation, the expressions are well known and can be read from different sources \cite{Bagnaschi:2014rsa, Gupta:2012gu}. In the limit of degenerate sfermion masses, the pole masses of the gauginos are
\begin{equation}\label{AnomalyPoleMasses}
\begin{aligned}
M_{\tilde{B}} = M_1(Q)&\left[1+\frac{C_\mu}{11}+\frac{8g_1^2}{80\pi^2}\left(-\frac{41}{2}\ln\frac{Q^2}{M_1^2}-\frac{1}{2}\ln\frac{\mu^2}{M_1^2}+\ln\frac{m_A^2}{M_1^2}\right.\right.\\
&\left.\left. +11\ln\frac{m_{\tilde{q}}^2}{M_1^2}+9\ln\frac{m_{\tilde{l}}^2}{M_1^2}\right)+\frac{g_3^2}{6\pi^2}-\frac{13g_t^2}{264\pi^2\sin^2\beta}\right]\\
M_{\tilde{W}} = M_2(Q)&\left[1+C_\mu+\frac{g_2^2}{16\pi^2}\left(\frac{19}{6}\ln\frac{Q^2}{M_2^2}-\frac{1}{6}\ln\frac{\mu^2}{M_2^2}+\frac{1}{3}\ln\frac{m_A^2}{M_2^2}\right.\right.\\
&\left.\left. +3\ln\frac{m_{\tilde{q}}^2}{M_2^2}+\ln\frac{m_{\tilde{l}}^2}{M_2^2}\right)+\frac{3g_3^2}{2\pi^2}-\frac{3g_t^2}{8\pi^2\sin^2\beta}\right]\\
M_{\tilde{G}} = M_3(Q)&\left[1+\frac{g_3^2}{16\pi^2}\left(7\ln\frac{Q^2}{M_3^2}+4\ln\frac{m_{\tilde{q}}^2}{M_3^2}+13-2F\left(\frac{M_3^2}{m_{\tilde{q}}^2} \right) \right)-\frac{7g_3^2}{24\pi^2}+\frac{g_t^2}{12\pi^2\sin^2\beta}\right]
\end{aligned}
\end{equation}
where
\begin{equation}\label{AnomalyBareMasses}
M_1(Q) = \frac{33 g_1^2(Q)}{80\pi^2}m_{3/2} \;\;\;\; M_2(Q) = \frac{g_2^2(Q)}{16\pi^2}m_{3/2} \;\;\;\; M_3(Q) = -\frac{3g_3^2(Q)}{16\pi^2}m_{3/2},
\end{equation}
$g_i(Q)$ are the gauge couplings of the SM in $\overline{MS}$ and SU(5) convention at scale $Q$, $g_t$ is the top Yukawa coupling in the SM, and
\begin{equation}
\begin{aligned}
C_\mu &= \frac{\mu}{m_{3/2}}\frac{m_A^2\sin^2\beta}{m_A^2-\mu^2}\ln\frac{m_A^2}{\mu^2}\\
F(x) &= 3\left[\frac{3}{2} -\frac{1}{x}-\left(\frac{1}{x}-1\right)^2\ln|1-x|\right].
\end{aligned}
\end{equation}
The main point of interest is that the corrections due to $C_\mu$ can be comparable if not bigger than the usual expressions. A typical mass spectrum is shown in the left panel of figure \ref{PlotsMassSpectrum}. Similar expressions hold for gauge mediation
\begin{equation}\label{GaugePoleMasses}
\begin{aligned}
M_{\tilde{B}} = M'_1(Q)&\left[1+\frac{3C'_\mu}{5}+\frac{g_1^2}{80\pi^2}\left(-\frac{41}{2}\ln\frac{Q^2}{M_1^2}-\frac{1}{2}\ln\frac{\mu^2}{M_1^2}+\ln\frac{m_A^2}{M_1^2}+11\ln\frac{m_{\tilde{q}}^2}{M_1^2}+9\ln\frac{m_{\tilde{l}}^2}{M_1^2}\right)\right]\\
M_{\tilde{W}} = M'_2(Q)&\left[1+C'_\mu+\frac{g_2^2}{16\pi^2}\left(\frac{19}{6}\ln\frac{Q^2}{M_2^2}-\frac{1}{6}\ln\frac{\mu^2}{M_2^2}+\frac{1}{3}\ln\frac{m_A^2}{M_2^2}+3\ln\frac{m_{\tilde{q}}^2}{M_2^2}+\ln\frac{m_{\tilde{l}}^2}{M_2^2}\right)\right]\\
M_{\tilde{G}} = M'_3(Q)&\left[1+\frac{g_3^2}{16\pi^2}\left(7\ln\frac{Q^2}{M_3^2}+4\ln\frac{m_{\tilde{q}}^2}{M_3^2}+13-2F\left(\frac{M_3^2}{m_{\tilde{q}}^2} \right) \right)+\frac{6g_3^2}{16\pi^2}\right]
\end{aligned}
\end{equation}
where we have kept only the terms proportional to $g_t$, $g_3$, or log-enhanced \cite{Gherghetta:1999sw},
\begin{equation}
\begin{aligned}
&M'_i(Q) = \frac{g_i^2}{16\pi^2} \Lambda \\
&C'_\mu = \frac{\mu}{\Lambda} \frac{m_A^2\sin^2\beta}{m_A^2-\mu^2} \ln\frac{m_A^2}{\mu^2}
\end{aligned}
\end{equation}
where $\Lambda$, in a given gauge mediation model, can be expressed in term of the SUSY breaking scale and the messenger scales (see for example eq. (\ref{eq:gaugemediation})). The last term of $M_{\tilde{G}}$ in (\ref{GaugePoleMasses}) can be extracted from \cite{Picariello:1998dy}. A typical mass spectrum is shown in the right panel of figure \ref{PlotsMassSpectrum}.
\begin{figure}
\centering
\begin{subfigure}[b]{0.475\textwidth}
\centering
\includegraphics[width=\textwidth,bb=0 0 260 206]{MassAnomaly-eps-converted-to.pdf}
\caption{}
\label{Figure1a}
\end{subfigure}%
~\qquad
\begin{subfigure}[b]{0.475\textwidth}
\centering
\includegraphics[width=\textwidth,bb=0 0 260 206]{MassGauge-eps-converted-to.pdf}
\caption{}
\label{Figure1b}
\end{subfigure}
\caption{Typical mass spectrum for (a) anomaly mediation and (b) gauge mediation. In (a), the masses appearing on the right side of (\ref{AnomalyPoleMasses}) are taken to be $m_{\text{scalars}}=\mu=m_{3/2}=50$ TeV with $\tan\beta=2$. In (b), the masses appearing on the right side of (\ref{GaugePoleMasses}) are taken to be $m_{\text{scalars}}=\mu=\Lambda=200$ TeV with $\tan\beta=2$. }\label{PlotsMassSpectrum}
\end{figure}
The parameters $C_\mu$ and $C'_\mu$ can be rewritten by requiring the fine-tuning condition, which needs to be imposed to have the weak scale parametrically smaller than the scalars \cite{Bagnaschi:2014rsa},
\begin{equation}
\tan^2\beta = \frac{m^2_{H_d}+\mu^2}{m^2_{H_u}+\mu^2}
\end{equation}
and the usual relation $m^2_A = m_{H_u}^2 + m_{H_d}^2 + 2\mu^2$. $C_{\mu}$ can then be expressed as \cite{Bagnaschi:2014rsa}
\begin{equation}\label{FinetuningCmu}
C_\mu = \frac{2\mu\tan\beta}{m_{3/2}}\frac{m_{H_d}^2+\mu^2}{(\tan^2\beta+1)m_{H_d}^2+\mu^2}\ln\left[(1+\cot^2\beta)\left(1+\frac{m_{H_d}^2}{\mu^2}\right)\right].
\end{equation}
The same applies to $C'_\mu$ with $m_{3/2} \to \Lambda$.
In these models the gauginos are the lightest sparticles and, because $\mu$ is large, the light neutralinos and charginos are almost pure binos and winos. As such, there is a neutralino of mass very close to $M_{\tilde{B}}$ and a pair of nearly degenerate neutralino and chargino of mass $M_{\tilde{W}}$. There is a small mass difference between the neutral and charged wino dominated by a loop effect \cite{Feng:1999fu}
\begin{equation}\label{LoopSplitting}
\Delta M \equiv m_{\chi_{\tilde{W}}^+}-m_{\chi_{\tilde{W}}^0} = \frac{\alpha_2 M_2}{4\pi}\left[f(r_W)-c_W^2f(r_Z)-s_W^2f(r_{\gamma})\right]
\end{equation}
where $f(y)=\int_0^1(2+2x)\log(x^2+(1-x)y^2)dx$ and $r_i=m_i/M_2$. The mass splitting is typically of the order of 150 MeV.
\subsection{Gaugino decays}\label{SectionGauginoDecays}
In this work we concentrate on gluino decay via third generation squarks. These decay modes dominate if the third generation squarks are lighter than the others, which is expected from RG effects or could be imposed for other model building reasons.\footnote{For example, flavor physics might require the first and second generations of squarks to be in the 1000 TeV range, while the third generation could be kept somewhat lighter to obtain the appropriate Higgs mass \cite{ArkaniHamed:2012gw}.} The decays that we consider are then
\begin{equation}\label{eqGluinoDecay}
\begin{aligned}
&\tilde{g} \to t \overline{t} \chi_1^0\quad &&\tilde{g} \to b \overline{b} \chi_1^0\quad &&\tilde{g} \to b \overline{t} \chi_1^+ \\
&\tilde{g} \to t \overline{t} \chi_2^0\quad &&\tilde{g} \to b \overline{b} \chi_2^0\quad &&\tilde{g} \to \overline{b} t \chi_1^-.
\end{aligned}
\end{equation}
The gluino can also decay to a gluon and a neutralino; however, it is negligible for heavy enough Higgsinos \cite{Jung:2013zya} and we ignore it. To compute the branching ratios we use analytical results that can be found in \cite{Toharia:2005gm}. An example of branching fractions is shown in figure \ref{PlotBranching}. In practice, $\chi_2^0$ always decays to $\chi_1^0$ and a Higgs boson \cite{ArkaniHamed:2012gw}, irrespective of whether $M_{\tilde{B}}$ is larger than $M_{\tilde{W}}$ or the opposite. In our scenario, the decay $\chi_2^0$ to $\chi_1^0$ and a Z boson is extremely suppressed due to the neutralinos being almost pure gauginos. When $M_{\tilde{W}} < M_{\tilde{B}}$, $\chi_1^+$ can only decay to $\chi_1^0$ and either light leptons or a pion which can cause this chargino to be metastable because of lack of phase-space \cite{Feng:1999fu}. As the decay is always very soft, the decay products are generally unaccounted for and the chargino is practically indistinguishable from the stable neutralino. When $M_{\tilde{W}} > M_{\tilde{B}}$ , $\chi_1^+$ decays to $\chi_1^0$ and a $W$ boson (we verified that the decay to $\chi_2^0$ only becomes relevant for $\mu$ at a scale considerably higher than anything relevant to this paper). The branching ratios we compute assume equal masses for the stops and the sbottoms. If the stops were lighter, the decays to two $b$ quarks, which can only proceed via off-shell sbottoms, would be relatively suppressed. As can be seen in figure \ref{PlotBranching}, these decays are already suppressed. The only thing that would change is the branching fraction of $\tilde{g} \to t \overline{t} \chi_1^0$, $\tilde{g} \to t \overline{t} \chi_2^0$ and $\tilde{g} \to b \overline{t} \chi_1^+$, which all have similar efficiencies for the searches we consider. We therefore do not expect that this assumption will affect our results greatly.
\begin{figure}[t!]
\centering
\includegraphics[width=0.5\textwidth,bb=0 0 360 315]{BRplot.pdf}
\caption{Branching ratios of the gluino for $M_{\tilde{G}}=1500$ GeV and $M_{\tilde{B}}=0$ GeV. The third generation scalar masses are assumed to be degenerate and much heavier than the gauginos.}\label{PlotBranching}
\end{figure}
\subsection{Higgs mass}\label{SectionHiggsMass}
To set the mass of the Higgs to its experimentally measured value, we follow the procedure outlined in \cite{Bagnaschi:2014rsa} which we summarize here.
First, $\overline{MS}$ parameters are taken from reference \cite{Buttazzo:2013uya} for the top Yukawa and the gauge coupling constants and from \cite{Mihaila:2012pz} for the bottom and tau Yukawas. The quartic coupling of the Higgs boson is extracted from its pole mass \cite{Holthausen:2011aa, Giudice:2011cg, Sirlin:1985ux, Cabrera:2011bi} using a value of 125.15 GeV, which is the naive average of the ATLAS \cite{Aad:2014aba} and CMS \cite{Chatrchyan:2013mxa} values. These parameters are then evolved up to the scalars scale using three-loops beta functions \cite{Bednyakov:2012rb, Bednyakov:2012en, Bednyakov:2013eba}. Threshold corrections are taken from \cite{Bagnaschi:2014rsa}. These include one-loop corrections and two-loop QCD corrections. The Higgs quartic is then matched with its SUSY expression and the threshold corrections. This determines one of the parameters, therefore reducing the dimension of the parameter space by one. As explained in the next section, we vary $\tan \beta$ to obtain the correct value of the Higgs mass.
In some regions of the parameter space it is not possible to obtain the correct Higgs mass because the required parameters lead to a color-breaking minimum that is deeper than the electroweak minimum. The necessary condition to avoid this is \cite{Bagnaschi:2014rsa}
\begin{equation}\label{eqVacuumStability}
\frac{(A_t-\mu\cot\beta)^2}{m_{Q_3} m_{U_3}} < \left(4-\frac{1}{\sin^2\beta}\right)\left(\frac{m_{Q_3}^2}{m_{U_3}^2}+\frac{m_{U_3}^2}{m_{Q_3}^2}\right),
\end{equation}
where $m_{Q_3}$ is the third generation soft mass for the SU(2) quark doublet and $m_{U_3}$ the right-handed stop soft mass.
\section{Methodology and results}
\subsection{Parameter space}
We begin by discussing the parameter space we use to study the models of interest. It is very similar for both anomaly and gauge mediation. There are essentially four parameters that control the phenomenology of anomaly mediation \cite{Bagnaschi:2014rsa}. They are $m_{3/2}$, $\tan\beta$, $m_{\text{scalars}}$, and $\mu$. As explained in section \ref{SectionMiniSplitModels}, $m_{3/2}$ and $m_{\text{scalars}}$ are expected to be of the same order of magnitude so we set them equal to each other. An additional parameter can be fixed by requiring the theory to predict the correct mass of the Higgs boson with the help of the results of section \ref{SectionHiggsMass}. Generally speaking, $\tan\beta$ is the best parameter to do so as varying it even slightly can have a substantial effect on the Higgs mass. The parameter space is then reduced to $\mu$ and $m_{3/2}$. However, we trade $\mu$ for $C_\mu$. The main advantage of this parametrization is that the ratio of gaugino masses depends mostly on $C_\mu$. The exact details of the scalar sector are relegated to two-loops corrections in (\ref{AnomalyPoleMasses}) and our results can therefore be applied to models where the scalar sector does not differ too significantly. To translate this to something more familiar, we provide each parameter space plot with contours of constant $M_{\tilde{B}}$, $M_{\tilde{W}}$, $\mu$, and $\tan\beta$.
The relationship between $\mu$ and $C_\mu$ depends on $m_{H_d}$ which we take to be at $m_{\text{scalars}}$. A different choice would lead, for the same $C_\mu$, to a different value of $\mu$ which in turn would affect mostly the color-breaking bounds (see equation (\ref{eqVacuumStability})). Taking $m_{H_d}$ much bigger than $m_{\text{scalars}}$ would limit $C_{\mu}$ to a narrow band around 0 and taking $m_{H_d}$ much smaller would push the bounds to large values of $C_\mu$ such that the gluino would be the LSP for most of the parameter space. With $m_{H_d}$ being set to $m_{\text{scalars}}$, we have a benchmark that does not suffer from any of these drawbacks. We assume the third generation to be lighter than the others, so as a benchmark we set the first and second generation squarks masses to $4m_{\text{scalars}}$ and all third generation masses to $m_{\text{scalars}}$. This is small enough to prevent problems with large logs, while keeping branching fraction to the first two generations below the percent level which is well below some of the uncertainties (e.g. gluino pair production cross-section). Sleptons masses are also set to $4m_{\text{scalars}}$. Lowering the masses of the first two generations of squarks would increase the branching ratio of the gluino to light jets, possibly affecting the reach of our searches (however, the high jet-multiplicity would still provide strong bounds). It would have only a slight effect on the gaugino spectrum and on the Higgs mass. Finally, we set the third generation A-term $A_t$ by equation (\ref{eqAterm}). Overall, changing our choice of benchmark parameters (mainly the choice of setting $m_{H_d}$ to $m_{\text{scalars}}$ and of taking $m_{Q_3}=m_{U_3}=m_{\text{scalars}}$) will mostly affect the $\mu$ and $\tan \beta$ contours in our results. Also, as a result of a modified relationship between $C_{\mu}$, $\mu$, and $\tan \beta$, the region of parameter space where there is a color-breaking vacuum would also be modified.
In almost all of our parameter space the Higgsinos are heavy, except for a region near $C_\mu=0$ where a Higgsino can be the lightest superpartner (LSP). More precisely, outside of $|C_\mu|<0.3$, the Higgsinos are always an order of magnitude heavier than the gluino while only inside $|C_\mu|<0.1$ are the Higgsinos comparable in mass to the bino and winos. This represents only a very narrow band in the parameter space and the efficiencies of the signal regions are not expected to change much in it. In addition, this case has already been studied in \cite{Jung:2013zya, Barnard:2012au}. As such, we neglect this effect. When $M_{\tilde{W}} < M_{\tilde{B}}$, the mass difference between $\chi_1^+$ and $\chi_1^0$ is calculated using (\ref{LoopSplitting}).
The previous discussion applies almost directly to gauge mediation by trading $m_{3/2}$ for $\Lambda$. In this case, we fix $m_{\text{scalars}}$ to $\Lambda$ while $\tan \beta$ is again set by requesting the correct mass of the Higgs boson.\footnote{There is considerable freedom on the choice of the scalar masses. The choice we make is more to keep in tune with our procedure for anomaly mediation. As explained above, the exact details of the scalar sector are not very relevant in our parametrization.} The masses of the sleptons and the first two generations squarks are still set to $4m_{\text{scalars}}$. $A_t$ is set to zero, as one would expect it to be small \cite{Bagnaschi:2014rsa}, and is then completely overshadowed by $\mu$. The mass $m_{H_d}$ is once more set to $m_{\text{scalars}}$.
Two other constraints are of importance for the parameter space. First of all, for a given value of $m_{3/2}$ ($\Lambda$), a small value of $A_t$ will lead to an upper bound on $C_\mu$ ($C'_\mu$) beyond which it is impossible to obtain the correct Higgs mass. Indeed if $C_\mu$ ($C'_\mu$) becomes large, the threshold corrections also become large and the quartic matching condition does not accept any solutions for real $\tan\beta$. In fact, requiring $C_\mu$ ($C'_\mu$) close to its upper bound can make the Higgsinos heavy enough that large logs could become a problem and perturbation expansions could fail. Fixing the stop mixing parameter $A_t-\mu\cot\beta$ to a small value would solve this problem, but this would imply $A_t$ reaching values that are too high to be readily explained in our framework without large fine-tuning. The second issue arises from the presence of a color-breaking vacuum which is controlled by equation (\ref{eqVacuumStability}). For the values of $m_{3/2}$ ($\Lambda$) considered in this work, it turns out that this limit is always stronger than the upper bound on $C_\mu$ ($C'_\mu$) coming from the mass of the Higgs boson. This latter constraint can therefore be ignored. We limit ourselves to the regions of parameter space where equation (\ref{eqVacuumStability}) is satisfied.
\subsection{Current LHC constraints}
To obtain current limits on anomaly and gauge mediation, we recast searches for gluino pair production. In particular we concentrate on searches with either many b-jets, leptons, or large jet-multiplicity. Of course, all of these searches have stringent cuts on missing transverse energy (MET). The chosen searches are summarized in table \ref{TableSearches}. As a general rule, \cite{Aad:2014lra} dominates over the others. For each of these searches, we implemented codes simulating the cuts. To validate our codes, we generated events with MadGraph 5 \cite{Alwall:2014hca} intefaced with Pythia 6 \cite{Sjostrand:2006za} and Delphes 3 \cite{deFavereau:2013fsa, Cacciari:2011ma}. We were able to reproduce all four searches with good accuracy. There are also constraints coming from electrowino production for which the experimental bounds found in \cite{TheATLAScollaboration:2013zia, Aad:2014vma, Aad:2014nua, Khachatryan:2014mma} apply directly. This is because the branching ratios for the charginos and neutralinos that are relevant for our models are the same as the one used in the simplified models considered in those searches. The bounds are in general much weaker than the one from gluino production and become relevant only in a tiny region of parameter space where the electrowinos are very light.
\begin{table}[t!]
\begin{center}
\begin{tabular}{ |C{3cm}|C{4cm}|C{4cm}|C{3cm}| }
\hline
Collaboration & Search & Strategy & Reference \\ \hline
ATLAS & JHEP 06 (2014) 035 & 2 same sign / 3 leptons + 0-3 b-jets + MET & \cite{Aad:2014pda} \\ \hline
ATLAS & JHEP 10 (2014) 024 & 0-1 leptons + ≥ 3 b-jets + MET & \cite{Aad:2014lra} \\ \hline
CMS & CMS-SUS-13-012 & High jet-multiplicity + MET & \cite{Chatrchyan:2014lfa} \\ \hline
CMS & CMS-PAS-SUS-12-016 & 2 opposite sign leptons + high-jet multiplicity + ≥3 b-jets + MET & \cite{CMS:2013ija} \\ \hline
\end{tabular}%
\caption{Gluino pair production searches.}\label{TableSearches}
\end{center}
\end{table}
Our method to reinterpret the experimental constraints follows closely the procedure of \cite{Papucci:2014rja}. We look at every possible combination of decay chains (\ref{eqGluinoDecay}) and evaluate for each of them the efficiency of every signal region. The branching fractions are then calculated using the procedure of section \ref{SectionGauginoDecays}. The gluino pair production cross-sections are calculated at NLO+NLL with NLL-fast \cite{Beenakker:1996ch, Kulesza:2008jb, Kulesza:2009kq, Beenakker:2009ha, Beenakker:2011fu}, which we verified using Prospino \cite{Beenakker:1996ed}. The number of expected signals in a given signal region can then be calculated. The $95\%$ confidence level signal upper limit can either be read directly from these searches or calculated using the known background and confidence level (CL) techniques \cite{Read:2002hq}. The different signal regions are combined in a boolean fashion \cite{Martin:2014qra}. A more thorough approach would require the correlation between the backgrounds of the different signal regions, which is not readily available.
The events are generated with MadGraph 5 \cite{Alwall:2014hca} interfaced with Pythia 6 \cite{Sjostrand:2006za} and Delphes 3 \cite{deFavereau:2013fsa, Cacciari:2011ma}. 10000 events are generated for each grid point. MadGraph generally takes care of decay chains up to the production of the LSP. The only exception is when either $\chi_2^0$ or $\chi_1^+$ are very close in mass to $\chi_1^0$. These decays can then be forced to be off-shell and the decay chains become too long to be handled by MadGraph comfortably. In the worst case scenario, $\chi_2^0$ can decay to $\chi_1^0$ and an off-shell Higgs which then decays to a W and a off-shell W which in turn decays to other particles. To handle these difficult decays, we calculate branching ratios in advance using the decay functionalities of MadGraph to produce decay tables. $\chi_2^0$ and $\chi_1^+$ are then decayed by Pythia using these results. Delphes handles the detector simulation and is tuned to simulate the ATLAS and CMS detectors.
The results for the 95$\%$ CL limits from ATLAS and CMS are given in figures \ref{AnomalyLHC} and \ref{GaugeLHC} for anomaly and gauge mediation respectively. Each one is provided with contour plots of $M_{\tilde{B}}$, $M_{\tilde{W}}$, $\mu$, and $\tan\beta$ to relate it to more familiar parameters. The regions forbidden by color-breaking vacuum are shown in purple. Overall, gluinos of mass up to $1.3$ TeV can be excluded over significant regions of parameter space. The results for the anomaly mediation spectrum can be easily understood. Over the entire covered parameter space, the gluino decays mainly to charginos. For $C_\mu$ between -4 and 4, the neutral wino is the LSP. The most relevant parameter in this region is then the ratio of the mass of the LSP and of the gluino. Below $C_\mu$ equal to 2, this ratio is large and the exclusion limits are strong. Above that value, the mass spectrum becomes compressed and kinematics quantities like MET become much smaller. As such, the exclusion limits drop considerably.
The results for gauge mediation are similar but with a few additional subtleties. Near $C'_\mu$ equal to -5, the spectrum is fairly compressed and the wino is too heavy to be produced. The gluino decays softly to $\chi_1^0$ and quarks, which results in lower constraints. As $C'_\mu$ increases, the spectrum becomes less compressed and the limits are stronger. However, near $C'_\mu$ equal to -3, the winos become light enough to be produced and the gluino decay to chargino dominates. As these decay chains are longer, there is less MET and the constraints are less strong. In a very narrow band around $C'_\mu$ equal -1.5, the wino is the LSP. The chargino then decays softly to a neutral wino. This is similar to gluino decaying to $\chi_1^0$ and the exclusion reaches the same levels as at $C'_\mu$ equal to -3. As $C'_\mu$ continues to increase, the mass spectrum again becomes compressed to the point where gluinos can only decay to $\chi_1^0$ and a pair of soft bottom quarks and the limits drop considerably. In addition, direct electroweakinos production searches from \cite{TheATLAScollaboration:2013zia, Aad:2014vma, Aad:2014nua, Khachatryan:2014mma} impose limits in a very narrow band near $C'_\mu$ equal to -2. This corresponds to when both the wino and bino are light which only occurs around $C'_\mu$ equal to $-2$. This region is shown as a grey band in figure \ref{GaugeLHC}.
\begin{figure}
\centering
\begin{subfigure}[b]{0.5\textwidth}
\centering
\caption{$M_{\tilde{B}}$ [GeV]}
\includegraphics[width=\textwidth,bb=0 0 580 390]{AnomalyLHCMB.pdf}
\label{Plot1a}
\end{subfigure}%
~
\begin{subfigure}[b]{0.5\textwidth}
\centering
\caption{$M_{\tilde{W}}$ [GeV]}
\includegraphics[width=\textwidth,bb=0 0 580 390]{AnomalyLHCMW.pdf}
\label{Plot1b}
\end{subfigure}
\begin{subfigure}[b]{0.5\textwidth}
\centering
\caption{$\mu$ [TeV]}
\includegraphics[width=\textwidth,bb=0 0 580 390]{AnomalyLHCmu.pdf}
\label{Plot1c}
\end{subfigure}%
~
\begin{subfigure}[b]{0.5\textwidth}
\centering
\caption{$\tan\beta$}
\includegraphics[width=\textwidth,bb=0 0 580 390]{AnomalyLHCtanb.pdf}
\label{Plot1d}
\end{subfigure}
\caption{95$\%$ exclusion limits for anomaly mediation at the LHC. The yellow band corresponds to the $1\sigma$ uncertainty on the gluino pair production cross-section and the purple bands are the forbidden region of color-breaking vacuum. Contour lines of constant $M_{\tilde{B}}$, $M_{\tilde{W}}$, $\mu$, and $\tan\beta$ are shown respectively in (a), (b), (c), and (d).}\label{AnomalyLHC}
\end{figure}
\begin{figure}
\centering
\begin{subfigure}[b]{0.5\textwidth}
\centering
\caption{$M_{\tilde{B}}$ [GeV]}
\includegraphics[width=\textwidth,bb=0 0 580 390]{GaugeLHCMB.pdf}
\label{Plot2a}
\end{subfigure}%
~
\begin{subfigure}[b]{0.5\textwidth}
\centering
\caption{$M_{\tilde{W}}$ [GeV]}
\includegraphics[width=\textwidth,bb=0 0 580 390]{GaugeLHCMW.pdf}
\label{Plot2b}
\end{subfigure}
\begin{subfigure}[b]{0.5\textwidth}
\centering
\caption{$\mu$ [TeV]}
\includegraphics[width=\textwidth,bb=0 0 580 390]{GaugeLHCmu.pdf}
\label{Plot2c}
\end{subfigure}%
~
\begin{subfigure}[b]{0.5\textwidth}
\centering
\caption{$\tan\beta$}
\includegraphics[width=\textwidth,bb=0 0 580 390]{GaugeLHCtanb.pdf}
\label{Plot2d}
\end{subfigure}
\caption{95$\%$ exclusion limits for gauge mediation at the LHC. The yellow band corresponds to the $1\sigma$ uncertainty on the gluino pair production cross-section and the purple bands are the forbidden region of color-breaking vacuum. The grey band corresponds to limits from direct electroweak searches. Contour lines of constant $M_{\tilde{B}}$, $M_{\tilde{W}}$, $\mu$, and $\tan\beta$ are shown respectively in (a), (b), (c), and (d).}\label{GaugeLHC}
\end{figure}
\subsection{Prospects at LHC 14}\label{SectionProspects14}
The procedure of the previous section can be modified to predict the discovery and exclusion prospects at the next phase of the LHC. The only differences amount to the signal regions and background estimations.
Two different strategies are adopted to cover the possibilities of the spectrum being compressed or not. When the LSP is considerably lighter than the gluino, kinematic quantities like MET are large and strong kinematic cuts are sufficient to eliminate most of the background. We refer to these signal regions as high MET cuts. On the contrary, when the gluino has a mass close to the LSP, quantities like MET become small and the cuts remove most signals. Lowering the cuts does not improve the limits much as the background increases considerably. However, adding the requirement of a pair of same sign dilepton (SSDL) drastically cuts the background and allows the kinematic cuts to be made less stringent by exploiting the possible production of leptons during the top decay. The only drawback to SSDL is that a large part of the signal is cut and the resulting limits are less strong than pure high MET cuts in the non-compact case. The net result is that high MET signal regions usually dominate until the spectrum becomes near degenerate. The exclusion then drops until the signal regions with SSDL become relevant which prevents the exclusion limits from dropping too fast. However, the SSDL cuts eventually also fail when there is not enough phase space for the gluino to produce top quarks.
For the high MET signal regions, we adopt the cuts of \cite{CMS:2013ega} for gluino decaying to top quarks and a single lepton. The cuts for SSDL are taken directly from \cite{Cohen:2013xda} and correspond to their gluino-neutralino model with heavy flavor decay for 14 TeV. We verified that we could reproduce both sets of results.
The detector card for Delphes is the standard 14 TeV card from Snowmass \cite{Anderson:2013kxz}. The background estimates for the high MET regions are obtained from the Snowmass online backgrounds \cite{Avetisyan:2013onh}. We simply apply our cuts on their events while taking into consideration their relative weight. The Snowmass backgrounds also provide events files with different average number of pile-up. In general, pile-up has very little effect on the high MET regions, while, for SSDL, leptons can possibly get lost in the pile-up jets \cite{Cohen:2013xda}, reducing the efficiency of the signal. We however concentrate on the case of 0 pile-up as the effect is generally small on most of the parameter space. For high MET cuts, we obtain backgrounds of (23.0, 12.1, 2.6, 2.1) for the four signal regions of \cite{CMS:2013ega} and 3000 $\text{fb}^{-1}$ of integrated luminosity. This can be compared with their result at 140 pile-up of (17.5, 4.8, 0.9, 1.6) and the same integrated luminosity. The backgrounds for SSDL are taken directly from \cite{Cohen:2013xda}, as we follow very closely their procedure. A 20$\%$ systematic uncertainty on all backgrounds is assumed \cite{Cohen:2013xda}. The gluino pair production cross-section is calculated using NLL-fast \cite{Beenakker:1996ch, Kulesza:2008jb, Kulesza:2009kq, Beenakker:2009ha, Beenakker:2011fu} customized for a 14 TeV collider. The possibility of 300 and 3000 $\text{fb}^{-1}$ of integrated luminosity are considered.
The results can be seen for anomaly mediation in figures \ref{AnomalyLHC1495} and \ref{AnomalyLHC145s} for 95$\%$ exclusion and $5\sigma$ discovery respectively, as well as for gauge mediation in figure \ref{GaugeLHC1495} and \ref{GaugeLHC145s} for 95$\%$ exclusion and $5\sigma$ discovery respectively. The curves are essentially scaled up versions of the 8 TeV constraints. The anomaly mediation limits curves are flatter than those for the current LHC constraints. This can be explained by the fact that the branching ratio to the LSP and two tops decreases more slowly as $C_{\mu}$ increases because heavier gluinos are being probed.
\begin{figure}
\centering
\begin{subfigure}[b]{0.5\textwidth}
\centering
\caption{$M_{\tilde{B}}$ [TeV]}
\includegraphics[width=\textwidth,bb=0 0 400 275]{Anomaly1495MB.pdf}
\label{Plot2a}
\end{subfigure}%
~
\begin{subfigure}[b]{0.5\textwidth}
\centering
\caption{$M_{\tilde{W}}$ [TeV]}
\includegraphics[width=\textwidth,bb=0 0 400 275]{Anomaly1495MW.pdf}
\label{Plot2b}
\end{subfigure}
\begin{subfigure}[b]{0.5\textwidth}
\centering
\caption{$\mu$ [PeV]}
\includegraphics[width=\textwidth,bb=0 0 400 275]{Anomaly1495mu.pdf}
\label{Plot2c}
\end{subfigure}%
~
\begin{subfigure}[b]{0.5\textwidth}
\centering
\caption{$\tan\beta$}
\includegraphics[width=\textwidth,bb=0 0 400 275]{Anomaly1495tanb.pdf}
\label{Plot2d}
\end{subfigure}
\caption{95$\%$ exclusion limits for anomaly mediation at LHC 14 for (solid) 300 $\text{fb}^{-1}$ and (dashed) 3000 $\text{fb}^{-1}$ integrated luminosity. The green band corresponds to the $1\sigma$ uncertainty on the gluino pair production cross-section for 300 $\text{fb}^{-1}$, the yellow band corresponds to the $1\sigma$ uncertainty on the gluino pair production cross-section for 3000 $\text{fb}^{-1}$, and the purple bands are the forbidden region of color-breaking vacuum. Contour lines of constant $M_{\tilde{B}}$, $M_{\tilde{W}}$, $\mu$, and $\tan\beta$ are shown respectively in (a), (b), (c), and (d).}\label{AnomalyLHC1495}
\end{figure}
\begin{figure}
\centering
\begin{subfigure}[b]{0.5\textwidth}
\centering
\caption{$M_{\tilde{B}}$ [TeV]}
\includegraphics[width=\textwidth,bb=0 0 400 275]{Anomaly145sMB.pdf}
\label{Plot2a}
\end{subfigure}%
~
\begin{subfigure}[b]{0.5\textwidth}
\centering
\caption{$M_{\tilde{W}}$ [TeV]}
\includegraphics[width=\textwidth,bb=0 0 400 275]{Anomaly145sMW.pdf}
\label{Plot2b}
\end{subfigure}
\begin{subfigure}[b]{0.5\textwidth}
\centering
\caption{$\mu$ [PeV]}
\includegraphics[width=\textwidth,bb=0 0 400 275]{Anomaly145smu.pdf}
\label{Plot2c}
\end{subfigure}%
~
\begin{subfigure}[b]{0.5\textwidth}
\centering
\caption{$\tan\beta$}
\includegraphics[width=\textwidth,bb=0 0 400 275]{Anomaly145stanb.pdf}
\label{Plot2d}
\end{subfigure}
\caption{5$\sigma$ discovery limits for anomaly mediation at LHC 14 for (solid) 300 $\text{fb}^{-1}$ and (dashed) 3000 $\text{fb}^{-1}$ integrated luminosity. The green band corresponds to the $1\sigma$ uncertainty on the gluino pair production cross-section for 300 $\text{fb}^{-1}$, the yellow band corresponds to the $1\sigma$ uncertainty on the gluino pair production cross-section for 3000 $\text{fb}^{-1}$, and the purple bands are the forbidden region of color-breaking vacuum. Contour lines of constant $M_{\tilde{B}}$, $M_{\tilde{W}}$, $\mu$, and $\tan\beta$ are shown respectively in (a), (b), (c), and (d).}\label{AnomalyLHC145s}
\end{figure}
\begin{figure}
\centering
\begin{subfigure}[b]{0.5\textwidth}
\centering
\caption{$M_{\tilde{B}}$ [TeV]}
\includegraphics[width=\textwidth,bb=0 0 400 275]{Gauge1495MB.pdf}
\label{Plot2a}
\end{subfigure}%
~
\begin{subfigure}[b]{0.5\textwidth}
\centering
\caption{$M_{\tilde{W}}$ [TeV]}
\includegraphics[width=\textwidth,bb=0 0 400 275]{Gauge1495MW.pdf}
\label{Plot2b}
\end{subfigure}
\begin{subfigure}[b]{0.5\textwidth}
\centering
\caption{$\mu$ [PeV]}
\includegraphics[width=\textwidth,bb=0 0 400 275]{Gauge1495mu.pdf}
\label{Plot2c}
\end{subfigure}%
~
\begin{subfigure}[b]{0.5\textwidth}
\centering
\caption{$\tan\beta$}
\includegraphics[width=\textwidth,bb=0 0 400 275]{Gauge1495tanb.pdf}
\label{Plot2d}
\end{subfigure}
\caption{95$\%$ exclusion limits for gauge mediation at LHC 14 for (solid) 300 $\text{fb}^{-1}$ and (dashed) 3000 $\text{fb}^{-1}$ integrated luminosity. The green band corresponds to the $1\sigma$ uncertainty on the gluino pair production cross-section for 300 $\text{fb}^{-1}$, the yellow band corresponds to the $1\sigma$ uncertainty on the gluino pair production cross-section for 3000 $\text{fb}^{-1}$, and the purple bands are the forbidden region of color-breaking vacuum. Contour lines of constant $M_{\tilde{B}}$, $M_{\tilde{W}}$, $\mu$, and $\tan\beta$ are shown respectively in (a), (b), (c), and (d).}\label{GaugeLHC1495}
\end{figure}
\begin{figure}
\centering
\begin{subfigure}[b]{0.5\textwidth}
\centering
\caption{$M_{\tilde{B}}$ [TeV]}
\includegraphics[width=\textwidth,bb=0 0 400 275]{Gauge145sMB.pdf}
\label{Plot2a}
\end{subfigure}%
~
\begin{subfigure}[b]{0.5\textwidth}
\centering
\caption{$M_{\tilde{W}}$ [TeV]}
\includegraphics[width=\textwidth,bb=0 0 400 275]{Gauge145sMW.pdf}
\label{Plot2b}
\end{subfigure}
\begin{subfigure}[b]{0.5\textwidth}
\centering
\caption{$\mu$ [PeV]}
\includegraphics[width=\textwidth,bb=0 0 400 275]{Gauge145smu.pdf}
\label{Plot2c}
\end{subfigure}%
~
\begin{subfigure}[b]{0.5\textwidth}
\centering
\caption{$\tan\beta$}
\includegraphics[width=\textwidth,bb=0 0 400 275]{Gauge145stanb.pdf}
\label{Plot2d}
\end{subfigure}
\caption{5$\sigma$ discovery limits for gauge mediation at LHC 14 for (solid) 300 $\text{fb}^{-1}$ and (dashed) 3000 $\text{fb}^{-1}$ integrated luminosity. The green band corresponds to the $1\sigma$ uncertainty on the gluino pair production cross-section for 300 $\text{fb}^{-1}$, the yellow band corresponds to the $1\sigma$ uncertainty on the gluino pair production cross-section for 3000 $\text{fb}^{-1}$, and the purple bands are the forbidden region of color-breaking vacuum. Contour lines of constant $M_{\tilde{B}}$, $M_{\tilde{W}}$, $\mu$, and $\tan\beta$ are shown respectively in (a), (b), (c), and (d).}\label{GaugeLHC145s}
\end{figure}
\subsection{Prospects at a 100 TeV collider}\label{SectionProspects}
To fully explore the possibility of discovering Split supersymmetry at colliders, we study the prospect of a 100 TeV collider following the same procedure as in the previous two sections. Our high MET cuts are adapted from \cite{Jung:2013zya}, which are themselves based on \cite{Hinchliffe:2002mn}. These cuts rely on $M_{\text{eff}}$ which is defined as
\begin{equation}
M_{\text{eff}}=\sum_i p_T(i)+\text{MET}.
\end{equation}
The sum is on jets with $p_T > 50$ GeV and $|\eta| < 5$ and leptons with $p_T > 15$ GeV and $|\eta| < 2.5$. We push things further than \cite{Jung:2013zya} by requiring b-jets, implementing detector simulations, and using a set of signal regions optimized for different regions of parameter space. The preselection cuts are given by \cite{Jung:2013zya}
\begin{itemize}
\item{Lepton veto,}
\item{At least two jets with $p_T > 0.1 M_{\text{eff}}$,}
\item{$\text{MET} > 0.2 M_{\text{eff}}$,}
\item{$p_T(j_1) < 0.35 M_{\text{eff}}$,}
\item{$\Delta \phi(j_1,\text{MET}) < \pi -0.2$,}
\item{$\Delta \phi(j_1,j_2) < 2\pi/3$. }
\end{itemize}
The different signal regions correspond to different combinations of minimum b-jets requirements and $M_{\text{eff}}$ cuts and are given in table \ref{TableSRleptonveto}.
\begin{table}[t!]
\begin{center}
\begin{tabular}{ |C{3cm}|C{2cm}|C{2cm}|C{3cm}|}
\hline
SR & b-jets & $M_{\text{eff}}$ [TeV] & Background \\ \hline
hMETb3A & $\ge$ 3 & $>$ 15.0 & 23.4 \\ \hline
hMETb3B & $\ge$ 3 & $>$ 17.5 & 7.8 \\ \hline
hMETb3C & $\ge$ 3 & $>$ 20.0 & 2.3 \\ \hline
hMETb4A & $\ge$ 4 & $>$ 12.5 & 12.6 \\ \hline
hMETb4B & $\ge$ 4 & $>$ 15.0 & 3.8 \\ \hline
hMETb4C & $\ge$ 4 & $>$ 17.5 & 1.5 \\ \hline
hMETb4D & $\ge$ 4 & $>$ 20.0 & 0.5 \\ \hline
\end{tabular}%
\caption{Signal regions for high MET. The background for 3 $\text{ab}^{-1}$ is also included.}\label{TableSRleptonveto}
\end{center}
\end{table}
The SSDL cuts and the corresponding backgrounds are taken directly from \cite{Cohen:2013xda} and correspond to their search for gluino-neutralino model with heavy flavor decays. We verified that we could reproduce their results.
The detector card for Delphes is the standard 100 TeV card from Snowmass \cite{Anderson:2013kxz}. The background estimates for high MET are again obtained from the Snowmass online backgrounds \cite{Avetisyan:2013onh}. The backgrounds for the high MET signal regions are shown in table \ref{TableSRleptonveto} for 3 $\text{ab}^{-1}$ integrated luminosity. A 20$\%$ systematic uncertainty on all backgrounds is assumed \cite{Cohen:2013xda}. The discussion of pile-up for high MET or SSDL from the previous section still holds. We concentrate on the 0 pile-up case, as the average pile-up of a future 100 TeV is still unknown and as it only has a non-negligible effect on a small portion of our parameter space. The gluino pair production cross-section is calculated using NLL-fast \cite{Beenakker:1996ch, Kulesza:2008jb, Kulesza:2009kq, Beenakker:2009ha, Beenakker:2011fu} customized for a 100 TeV collider.
The results are again scaled up versions of LHC constraints with possible exclusion of up to a 14 TeV gluino in a large region of parameter space and discovery of up to 12 TeV. These numbers are similar to those obtained by \cite{Jung:2013zya} which seem somewhat more optimistic (with a possible discovery of up to $\sim 15$ TeV).\footnote{ This might be due, for example, to the fact that we have used a detector simulation, but we haven't directly checked that hypothesis.} For anomaly mediation, exclusion limits are governed by high MET signal regions and are thus very high until $C_\mu$ reaches 1. At this point, the spectrum becomes compact and the limits drop. The SSDL bins then dominate and the limits stabilize with a discovery reach of about $7$ TeV (this number is in fact quite close to the result of \cite{Cohen:2013xda}). The exact same thing happens in the case of gauge mediation, except that the limits drop at $C'_\mu$ equal to 0.
\begin{figure}
\centering
\begin{subfigure}[b]{0.5\textwidth}
\centering
\caption{$M_{\tilde{B}}$ [TeV]}
\includegraphics[width=\textwidth,bb=0 0 400 275]{Anomaly100MB.pdf}
\label{Plot3a}
\end{subfigure}%
~
\begin{subfigure}[b]{0.5\textwidth}
\centering
\caption{$M_{\tilde{W}}$ [TeV]}
\includegraphics[width=\textwidth,bb=0 0 400 275]{Anomaly100MW.pdf}
\label{Plot3b}
\end{subfigure}
\begin{subfigure}[b]{0.5\textwidth}
\centering
\caption{$\mu$ [PeV]}
\includegraphics[width=\textwidth,bb=0 0 400 275]{Anomaly100mu.pdf}
\label{Plot3c}
\end{subfigure}%
~
\begin{subfigure}[b]{0.5\textwidth}
\centering
\caption{$\tan\beta$}
\includegraphics[width=\textwidth,bb=0 0 400 275]{Anomaly100tanb.pdf}
\label{Plot3d}
\end{subfigure}
\caption{95$\%$ (dashed) exclusion and $5\sigma$ (solid) discovery limits for anomaly mediation at a 100 TeV $pp$ collider with 3 $\text{ab}^{-1}$ integrated luminosity. The yellow band corresponds to the $1\sigma$ uncertainty on the gluino pair production cross-section for 95$\%$ exclusion, the green band corresponds to the $1\sigma$ uncertainty on the gluino pair production cross-section for $5\sigma$ discovery, and the purple bands are the forbidden region of color-breaking vacuum. Contour lines of constant $M_{\tilde{B}}$, $M_{\tilde{W}}$, $\mu$, and $\tan\beta$ are shown respectively in (a), (b), (c), and (d).}\label{Anomaly100}
\end{figure}
\begin{figure}
\centering
\begin{subfigure}[b]{0.5\textwidth}
\centering
\caption{$M_{\tilde{B}}$ [TeV]}
\includegraphics[width=\textwidth,bb=0 0 400 275]{Gauge100MB.pdf}
\label{Plot4a}
\end{subfigure}%
~
\begin{subfigure}[b]{0.5\textwidth}
\centering
\caption{$M_{\tilde{W}}$ [TeV]}
\includegraphics[width=\textwidth,bb=0 0 400 275]{Gauge100MW.pdf}
\label{Plot4b}
\end{subfigure}
\begin{subfigure}[b]{0.5\textwidth}
\centering
\caption{$\mu$ [PeV]}
\includegraphics[width=\textwidth,bb=0 0 400 275]{Gauge100mu.pdf}
\label{Plot4c}
\end{subfigure}%
~
\begin{subfigure}[b]{0.5\textwidth}
\centering
\caption{$\tan\beta$}
\includegraphics[width=\textwidth,bb=0 0 400 275]{Gauge100tanb.pdf}
\label{Plot4d}
\end{subfigure}
\caption{95$\%$ (dashed) exclusion and $5\sigma$ (solid) discovery limits for gauge mediation at a 100 TeV $pp$ collider with 3 $\text{ab}^{-1}$ integrated luminosity. The yellow band corresponds to the $1\sigma$ uncertainty on the gluino pair production cross-section for 95$\%$ exclusion, the green band corresponds to the $1\sigma$ uncertainty on the gluino pair production cross-section for $5\sigma$ discovery, and the purple bands are the forbidden region of color-breaking vacuum. Contour lines of constant $M_{\tilde{B}}$, $M_{\tilde{W}}$, $\mu$, and $\tan\beta$ are shown respectively in (a), (b), (c), and (d).}\label{Gauge100}
\end{figure}
\section{Conclusions}
In light of ever stronger constraints from collider physics, Mini-Split scenarios become more and more appealing. In these models, a small hierarchy exists between the sfermions and gauginos, with the gauginos being near the electroweak scale. This kind of spectrum could easily arise from anomaly mediation and also from gauge mediation. In these models the electroweak scale is tuned, but gauge couplings could still unify at a high scale, and the models have possible dark matter candidates. The hierarchy between the scalars and the gauginos leads to large radiative corrections which can greatly modify the standard mass spectra of anomaly and gauge mediation.
In this paper we studied hadron collider constraints and prospects on these deflected anomaly mediation and deflected gauge mediation models. By using a simple parametrization of the models and assuming a lighter third generation and a heavy Higgsino, we recast SUSY searches from ATLAS and CMS to obtain exclusions on the parameter space of the models. The known mass of the Higgs boson and the absence of color-vacuum were also taken into account. Results for anomaly and gauge mediation can be seen respectively in figures \ref{AnomalyLHC} and \ref{GaugeLHC}. We also obtained future prospects for deflected anomaly mediation and deflected gauge mediation for LHC 14 and a 100 TeV collider. For LHC 14, the 95$\%$ projected exclusion limits are shown in figure \ref{AnomalyLHC1495} and \ref{GaugeLHC1495} for anomaly and gauge mediation respectively and the 5$\sigma$ discovery prospects are shown in \ref{AnomalyLHC145s} and \ref{GaugeLHC145s}. The prospects at a 100 TeV collider for anomaly and gauge mediation are found in figures \ref{Anomaly100} and \ref{Gauge100} respectively.
While the goal of this work was to explore the collider phenomenology of Mini-Split models, dark matter properties could also be used to further restrict the parameter space. The thermal abundance of the dark matter candidate is strongly dependent on the identity of the LSP. For a Wino LSP, the correct thermal relic abundance can be obtained for a wino mass of 2.7 TeV \cite{Bagnaschi:2014rsa,ArkaniHamed:2012gw}. This region of parameter space is not constrained by the LHC, but is within reach of a 100 TeV collider. Wino LSP with lighter mass could be accommodated by invoking non thermal production \cite{ArkaniHamed:2012gw}. Similarly, Bino LSP, which tend to overclose the universe, could be accommodated if there was late entropy production or a low reheating temperature.
\acknowledgments
This work was supported in part by the Natural Sciences and Engineering Research Council of Canada (NSERC). HB acknowledges support from the Ontario Graduate Scholarship (OGS).
\bibliographystyle{JHEP}
|
\section{Introduction}
The recent impressive advances in the field of quantum simulators
allowed to probe the many-body physics of strongly correlated systems
at the level of the single quantum object. At present cold atoms trapped in optical lattices
can be considered among the most promising examples of quantum simulators.
By means of ultracold atomic and molecular gases, it is nowadays possible
to reach a degree of control and accuracy in engineering the dynamics of many-body
systems that were unimaginable in previously.
As a consequence, the coherent quantum dynamics emerging from carefully tailored
microscopic Hamiltonians can now be tested experimentally~\cite{Bloch2008}.
It has been possible, just to recall one example, to implement the Bose-Hubbard (BH)
model~\cite{Fisher1989, Jaksch1998} and to detect its zero-temperature superfluid (SF)
to Mott insulator (MI) quantum phase transition~\cite{Greiner2002}.
Other models involving spinor gases, Fermi systems, Bose-Fermi mixtures,
or dipolar gases have been also devised and realized, providing an even
richer phase diagram (see for example the review~\cite{Lewenstein2007}).
We mention the stabilization of density-wave (DW) phases for bosons,
as well as more peculiar topological or supersolid orderings, which can
arise in the presence of finite-range interactions~\cite{Pfau2009}.
More recently a novel kind of many-body quantum simulator has been introduced,
based on the idea to use single photons as quantum objects.
Since photons hardly interact in open space, the most natural way to significantly
increase their interactions is to trap them into an optical QED cavity, and couple
the field with atoms/molecules inside it in order to create an optical nonlinearity.
If the nonlinearity is sufficiently large, the so called photon blockade
sets in~\cite{Schmidt1996, Imamoglu1997}, namely, the presence of a single photon inside
a cavity prevents a second one to enter it.
In the rotating-wave approximation, the simplest light-matter interaction scheme
of this type can be accurately described by the Jaynes-Cummings (JC) model.
By arranging an array of cavities coupled through the photon hopping,
such to generate a competition between the hopping and the on-site nonlinearities,
one can devise a setup that is well described by the so called
Jaynes-Cummings-Hubbard (JCH) model~\cite{Hartmann2006, Greentree2006, Angelakis2007}.
In many respects, if one ignores dissipation, the physics emerging from the JCH Hamiltonian
resembles, at low-energies, that of an effective BH model.
Probably the main difference between the two systems is
that, instead of having neutral bosons as building blocks of the model, in the JCH Hamiltonian
one has to think in terms of polaritons, {\it i.e.}, combined photonic/atomic excitations.
Many different works already addressed the JCH equilibrium
phase diagram with analytical, as well as numerical methods, leading to a fairly complete theoretical
understanding of the nature and the location of the emerging
phases and phase transitions in terms of the parameters governing the system
(the field has been recently reviewed in, {\it e.g.}, Refs.~\cite{Review1, Review2, Review3, Review4}).
Additional interest in cavity arrays comes from the fact that these systems can be naturally
considered as open-system quantum simulators. Some related features have been recently
explored~\cite{NonEq1,NonEq2,NonEq3,NonEq4,NonEq5,NonEq6,NonEq7,Jin2013,Jin2014}.
In the following we will not touch on this and consider only the ``equilibrium'' phase diagram.
This intense activity has been very recently boosted by the first experiments on
QED cavity arrays~\cite{Underwood2012, Abbarchi2013, Toyoda2013}. As of today,
the most concrete possibility to realize controllable and scalable quantum simulators with
cavity arrays involves circuit-QED cavities~\cite{Lucero2012, Steffen2013, Chen2014}.
So far the coupling between cavities
has been mostly considered through photon hopping.
Only few works started addressing more general schemes,
where the cavity coupling can be induced also by means of non-linear
elements~\cite{Zueco2012, Peropadre2013, Jin2013, Jin2014}.
Such configurations include cross-Kerr interactions and/or
correlated hopping terms, which lead to generalizations of the JCH model
in a way similar to the extended BH (EBH) Hamiltonian
for atoms with large dipole momentum loaded in optical lattices~\cite{Sowinski2012}.
The underlying physical model is believed to possess a much richer structure,
with the emergence of exotic phases of correlated polaritons.
It is particularly interesting to address these schemes in one-dimensional (1D) systems,
where interactions become crucial to stabilize exotic phases
of matter~\cite{DallaTorre2006A, DallaTorre2006B, Sowinski2012, Rossini2012, Deng2013}.
These notably include a series of nontrivial density-wave (DW) states,
which can arise in the strong coupling regime~\cite{Wikberg2012},
as well as supersolidity and phase-separation effects~\cite{Batrouni2013, Batrouni2014}.
Extension to consider also counter-rotating terms in the ultrastrong coupling
regime, thus leading to the so called Rabi-Hubbard model~\cite{Schiro2012},
have been investigated~\cite{Kumar2013}.
However we are not aware of numerical investigations of coupled cavity models
beyond the JCH and Rabi-Hubbard model.
In all such situations, non-perturbative, either numerical or analytical calculations are necessary.
Here the density-matrix renormalization group (DMRG) algorithm~\cite{dmrgA, dmrgB}
has been employed to work out the quantitative zero-temperature phase diagram
of the JCH model~\cite{Rossini2007, Rossini2008, Souza2013}.
This is a particularly efficient method for the statics of 1D many-body problems.
Its key strategy consists in constructing a portion of the system (called block) and then
recursively enlarge it. At each step, the basis of the corresponding Hamiltonian is truncated
to a given value $m$, so that one can manage the Hamiltonian in an effective Hilbert space
of fixed dimensions, as the physical system grows. This truncation is performed
by retaining the eigenstates corresponding to the $m$ highest eigenvalues
of the reduced density matrix of the block.
The aim of this paper is to quantitatively study a generalization
of the JCH Hamiltonian, aimed at taking into account an effective
nearest-neighbor nonlinearity between cavities mediated
by an $N$-type four-level system as discussed for two cavities in Ref.~\cite{Hu2011}.
The presence of this coupling leads to an effective cross-Kerr non-linearity.
An analysis at the mean-field level of a dissipative open EBH as an effective model for nonlinearly
coupled cavities has been performed, unveiling the emergence of novel
photon crystal and supersolid phases~\cite{Jin2013, Jin2014}.
Here we do not resort to the effective EBH model and analyze the full model as introduced
in~\cite{Hu2011}. Using the DMRG algorithm, we work out the 1D ground-state phase diagram.
We show that a physics similar to the EBH model appears,
with a rich phase diagram including gapless SF, as well as MI
and DW phases of polaritons.
We postpone the analysis of the interplay of driving and dissipation to a future work.
The paper is organized as follows.
In the next two sections we introduce the model of coupled cavities
of our interest (Sec.~\ref{sec:model}) and the quantities we are going
to address, namely the energy gaps, and the staggered
number-number correlations (Sec.~\ref{sec:quantities}).
In Sec.~\ref{sec:pd} we discuss the zero-temperature equilibrium phase diagram,
focusing on the MI/SF boundary and on the boundary separating the DW from the other phases.
Finally, in Sec.~\ref{sec:summary} we draw our conclusions.
\section{The model}
\label{sec:model}
\begin{figure}[!b]
\includegraphics[width=0.9\hsize]{fig1.pdf}
\caption{{\bf Scheme for nonlinearly coupled QED cavities.}
(a) An array of QED cavities nonlinearly coupled by $N$-type atoms.
The photon hopping between nearest-neighbor cavities has a strength $t$.
Each effective site is composed of a cavity and an atom (dashed box).
(b) Level structure of the $N$-type atoms.
The transition $\ket{1} \leftrightarrow \ket{3}$ is resonantly coupled
to the cavity mode of its own site with strength $g_1$,
while $\ket{2} \leftrightarrow \ket{4}$ is coupled to the cavity mode of
its right nearest-neighbor site with strength $g_2$, and has a detuning $\Delta$.
The transition $\ket{2} \leftrightarrow \ket{3}$ is in resonance with
an external laser field of strength $\Omega$.}
\label{fig1}
\end{figure}
Let us consider a 1D array of QED cavities,
where photons can hop between neighboring cavities.
Moreover two adjacent resonators are also nonlinearly coupled to each other
via a $N$-type four-level system, as shown in Fig.~\ref{fig1}(a).
For the sake of clarity in our description, we shall divide the 1D array
into coupled effective sites composed of a cavity and an atom.
The four levels are denoted by
$\{ \ket{i} \}_{i = 1 \ldots 4}$, and are depicted in Fig.~\ref{fig1}(b).
An external laser with frequency $\Omega$ resonantly drives
the transition $\ket{3} \leftrightarrow \ket{2}$.
The transition $\ket{1} \leftrightarrow \ket{3}$
is resonantly coupled to the cavity mode of the same site with strength $g_1$,
while the transition $\ket{2} \leftrightarrow \ket{4}$ couples to the cavity mode
of its right nearest-neighbor site with strength $g_2$, and a detuning $\Delta$.
The use of such $N$-type atom for generating large Kerr nonlinearity
has been extensively studied in the literature~\cite{Schmidt1996,Imamoglu1997,Rebic2009},
however the vast majority of the scenarios only focused on a single-mode cavity.
Our work is inspired by the idea of Ref.~\cite{Hu2011}, where the cross-Kerr nonlinearity
is generated between two different and neighboring cavities, in circuit-QED systems.
In practice, we use the unbalanced couplings of atomic transition $|1\rangle\leftrightarrow|3\rangle$
with left cavity mode, and $|2\rangle\leftrightarrow|4\rangle$ with right cavity mode respectively,
in order to generate the local ($g_1$) and nonlocal ($g_2$) nonlinearities of our many-body system.
This kind of four-level artificial molecule can be realized using two Josephson
transmon qubits coupled by a superconducting quantum interference device.
Using the interaction picture and in the rotating-wave approximation, the system Hamiltonian reads
\begin{equation}
\mathcal H = \sum_i \bigg[ \Delta \sigma^{44}_{i} +
\left( \Omega \sigma^{23}_i + g_1 \sigma_i^{13} a_i^\dagger + \mathrm{H.c.} \right)
+ \left(-t a_{i} \, a_{i+1}^\dagger + g_2 \sigma_i^{24} a_{i+1}^\dagger
+ \mathrm{H.c.} \right) \bigg] \,,
\label{hamiltonian}
\end{equation}
where $\sigma^{mn} = \ket{m} \bra{n}, (m,n = 1,2,3,4)$,
and $a$ $(a^\dagger)$ is the annihilation (creation) operator of the cavity mode.
The subscripts denote the site position along the 1D chain.
The first three terms in the r.h.s. of Eq.~\eqref{hamiltonian} describe the local
terms and the nonlinearities on each site.
Inside the latter brackets, the first term is the photon hopping, while the second term
describes the coupling of the atom to its right neighboring cavity, which generates
an effective nonlocal cross-Kerr nonlinearity between the two cavities.
Hereafter we concentrate on the 1D model in Eq.~\eqref{hamiltonian}
at zero temperature, specifically addressing the case without dissipation with DMRG.
Let us also fix the Hamiltonian quantities in units of $\Omega$, set $\hbar = 1$,
and work with open boundary conditions.
We recall that, in the presence of dissipation, the problem becomes much more difficult
to be handled numerically~\footnote{It is however possible to address the effect
of dissipation with a DMRG approach in a 1D chain, when this is described
by a master equation within the Lindblad formalism.
In the language of tensor networks, one has to generalize the matrix-product-state
ansatz to a matrix-product-density-operator ansatz for mixed states,
as originally proposed in Refs.~\cite{Verstraete2004, Zwolak2004}.
The computational complexity is greater than for static computations, and
is eventually related to the amount of entanglement in the steady state.}.
For the system we are considering here, in the strong coupling
regime atoms and photons cannot be considered as two separate entities.
It is thus natural to investigate the phase diagram in terms of combined
atomic/photonic modes, named polaritons.
The polaritonic number operator on each site $i$, representing the number
of local excitations, is defined as
$n^{\mathrm{pol}}_i=2\sigma_i^{44}+\sigma^{33}_i+\sigma^{22}_i+a^\dagger_ia_i$.
For the closed system described by the Hamiltonian~\eqref{hamiltonian},
the total number $N^{\rm pol} = \sum_i n^{\rm pol}_i$ of such polaritons
is a conserved quantity.
In the following we work in the canonical ensemble for polaritons,
and focus on the integer filling situation.
\section{Energy gaps and correlation functions}
\label{sec:quantities}
The different nature of the various phases is sensitive to a number
of properties which we are going to focus on.
Here we are going to study quantities that resemble those characterizing
the various phases of the EBH model~\cite{Rossini2012}.
First of all, the ground-state energy gap is an important indicator
which characterizes the presence or absence of criticality in the model.
In particular, in the critical SF phase, the {\it charge gap} vanishes
in the thermodynamic limit. On the other side, in the insulating
MI and DW phases, such gap remains finite. In order to make connection with a similar
notation in the EBH model, below we introduce the so called charge and neutral gaps
referring respectively to the gaps corresponding to adding one extra particle (``charge" sector) or
remaining with the same number of particles (``neutral" sector). We stress however that in the
present model the excitation carry no real charge. This has to be understood only as a
convention.
The charge gap is defined as
\begin{equation}
\Delta E_c = \Delta E^+ - \Delta E^- \, ,
\end{equation}
where, in the canonical ensemble, $\Delta E^+$ ($\Delta E^-$) denotes
the extra energy needed to add (remove) one particle,
{\it i.e.} one polariton, in the system.
In the specific, focusing on the unit filling,
$\Delta E^+ = E_{L+1}-E_{L}$ and $\Delta E^- = E_{L}-E_{L-1}$,
where $E_L$ is the ground-state energy per site of an $L$-sites
cavity-array with exactly $L$ excitations,
and $E_{L+1}$ ($E_{L-1}$) is the corresponding energy per site
with one excitation more (less).
It is therefore possible to extrapolate $\Delta E_c$ by running
three different DMRG simulations with fixed number of polaritons
$N^{\rm pol} = L-1, \, L, \, L+1$~\cite{KuhnerA, KuhnerB}.
While the charge gap is able to detect particle-hole excitations,
in some circumstances it is possible that the dominant
low-energy excitations are of a different type.
Their presence can be detected only by the so called {\it neutral gap}
at a fixed number of particles,
\begin{equation}
\Delta E_n = E_L^{1} - E_L \, ,
\end{equation}
where, again working in the canonical ensemble, $E_{L}^{1}$
denotes the first excited energy per site
of an $L$-site system with $L$ excitations.
In the following, we also focus on the analysis of the staggered diagonal order
for the polaritons, in order to distinguish the DW from the other phases.
We do this by investigating the two-point correlation function
\begin{equation}
C_{\mathrm{DW}}(r)=(-1)^r \langle \delta n_i^{\mathrm{pol}} \delta n_{i+r}^{\mathrm{pol}} \rangle \, ,
\label{eq:DWcorr}
\end{equation}
where $\delta n_i^{\mathrm{pol}} = n^{\mathrm{pol}}_i-\bar{n}$ denotes the polariton fluctuation
from the average filling $\bar{n}$.
The order parameter identifying the DW phase is thus given by:
$\mathcal{O}_{\mathrm{DW}}\equiv\lim_{r\rightarrow\infty}{C_{\mathrm{DW}}(r)}$.
A finite value of $\mathcal{O}_{\mathrm{DW}}$ indicates a tendency to establish,
in the thermodynamic limit, a staggered occupation of the polaritons.
On the other side, in the MI as well as the SF phases,
$C_{\mathrm{DW}}(r)$ vanishes exponentially with increasing distance $r$.
\section{Phase diagram}
\label{sec:pd}
\begin{figure}[!t]
\includegraphics[width = 1\hsize]{fig2_N.pdf}
\caption{{\bf Phase diagram.}
Ground-state phase diagram for a 1D system of coupled cavities
described by Hamiltonian~\eqref{hamiltonian} in the $g_2-t$ parameter space.
Here and in the subsequent figures
we choose $g_1 / \Omega = 0.8$, $\Delta / \Omega = -2$.
The various symbols correspond to points belonging to different phases
(triangles = MI, circles = DW, squares = SF).
Filled points lie close to a phase transition, and are used to
draw the interpolating curves (dashed lines).
The DW-to-MI/SF boundary (filled circles) has been obtained
by analyzing the density-wave order parameter,
while the MI-SF boundary (filled squares) through the charge gap.
The two vertical dotted lines denoted two cuts in the phase diagram which
will be analyzed in details below.
The error bars in the points characterizing the DW-to-MI/SF
boundary take into account the discretization of the $g_2$ values
that we adopted in our numerical simulations (see the text).}
\label{fig2}
\end{figure}
The zero-temperature phase diagram of model~\eqref{hamiltonian}, at unit polariton
filling $\bar{n} = 1$ and in the $g_2-t$ plane, is summarized in Fig.~\ref{fig2}.
We observe that three different phases can be stabilized.
Their boundaries have been obtained by means of a finite-size scaling
of the numerical data, for systems up to $L=300$ sites.
In our simulations we imposed a cutoff photon number in each cavity,
such that $n^{\rm phot}_i \leq 3$.
We also truncated the effective Hilbert space dimension to a value
$m=80$ in all the simulations, except for those shown in Fig.~\ref{fig5}
for the neutral energy gap (see the discussion in Sec.~\ref{sec:neutral_gap}).
We checked that, by increasing $m$ and the local fock-space truncation
over the photon number, the results concerning the charge gap
and the DW order parameter do not change on the scales shown here.
For small photon hopping ($t / \Omega \lessapprox 0.2$),
by increasing the nonlocal nonlinearity $g_2$ the system exhibits
a direct transition from the MI to the DW phase.
On the other hand, for $t / \Omega \gtrapprox 0.2$, the MI-to-DW transition
is mediated by an extended region appearing at intermediate $g_2$ values,
where the system stabilizes into a gapless SF.
In the following we are going to elucidate our finite-size scaling procedure
and how we were able to distinguish between the different phases.
\subsection{Boundary between MI and SF phases}
In the limit of small $g_2$ and $t$ values, the dominant presence
of the on-site interactions stabilize the system into a MI phase
with exactly one polariton per cavity ($\bar{n} = 1$),
and where the charge energy gap has a finite value.
As long as the hopping strength $t$ is progressively increased
(and for fixed $g_1, \, g_2$), the system eventually enters a SF phase,
with a vanishing gap.
The filled squares of Fig.~\ref{fig2} denoting the MI/SF boundaries
have been obtained by means of a finite-size scaling of the charge gap.
We performed simulations up to $L=100$ sites and analyzed whether the gap
closes or remains finite in the thermodynamic limit $L \to \infty$.
In Fig.~\ref{fig3}, left panel, we highlight the size-dependence of $\Delta E_c$
as a function of $1/L$ for two points in the phase space
close to the MI/SF transition (see points along the dotted line in Fig.~\ref{fig2}).
We expect to see a quadratic dependence $\Delta E_c \sim L^{-2}$
(dashed line) at large $L$~\cite{KuhnerA, KuhnerB}, however a linear extrapolation
(solid line) is already a good approximation to the scaling,
and we can use it to determine $\Delta E_c$ in the thermodynamics limit.
Indeed, we observe that the difference between quadratic
and linear extrapolation is tiny ($\lesssim 10^{-3}$) and does not produce any
distinguishable modification on the scale of Fig.~\ref{fig2}.
In the specific case of Fig.~\ref{fig3}, we fixed $t / \Omega = 0.25$
and chose two different values of $g_2 / \Omega$ corresponding to configurations
in the gapped MI ($g_2 / \Omega = 1.35$, triangles) and in the gapless SF phase
($g_2 / \Omega = 1.5$, squares).
The MI is signaled by an extrapolated finite value of $\lim_{L \to \infty} \Delta E_c > 0$,
while in the SF this is zero.
\begin{figure}[!t]
\includegraphics[width = 0.495\hsize]{fig3a.pdf}
\includegraphics[width = 0.495\hsize]{fig3b.pdf}
\caption{{\bf Analysis of the MI-SF Boundary.}
Left panel: system-size dependence of the charge gap $\Delta E_c$ per site,
for fixed $t/\Omega = 0.25$ and two values of $g_2$ in the MI ($g_2/\Omega = 1.35$)
and in the SF ($g_2/\Omega = 1.5$) phase. Symbols denote the DMRG results.
Solid and dashed lines are linear and quadratic fitting curves, respectively.
The difference between the extrapolated values in the two fits
$\Delta E_c^{\infty} = \lim_{L \to \infty} \Delta E_c^{(L)}$
is negligible.
Right panel: determination of the critical $g_2^*$ value
for the quantum phase transition.
The triangles denote the charge gap per site $\Delta E_c^{\infty}$
at the thermodynamic limit, as extrapolated in the left panel.
The dashed line is a best linear fit of the data {\it vs.} $g_2$.
The critical point is obtained when $\Delta E_c$ vanishes.
For $t / \Omega =0.25$, we get $g_2^* / \Omega \approx 1.379$.}
\label{fig3}
\end{figure}
In order to locate the critical $g_2$ for a given value of $t$
(filled squares in Fig.~\ref{fig2}) we perform a linear extrapolation
of the charge gaps in the vicinity of the critical value of $g_2$.
An example of such procedure is shown in the right panel of Fig.~\ref{fig3},
where we plot $\Delta E_c$ as a function of $g_2$, when this is close
to the phase transition (in the specific, here we set $t / \Omega = 0.25$).
After a linear extrapolation, we get a critical $g_2$ value
corresponding to $g_2^* / \Omega \approx 1.379$.
An analogous procedure is repeated for all the filled squares
shown in Fig.~\ref{fig2}, thus identifying the MI/SF boundary.
\subsection{Boundary of the DW phase}
\begin{figure}[!t]
\includegraphics[width = 0.495\hsize]{fig4b.pdf}
\includegraphics[width = 0.495\hsize]{fig4a.pdf}
\caption{{\bf Determination of the DW boundaries.}
The two-point correlation function $C_{\mathrm{DW}}(r)$ for the polariton
number and its asymptotic value near the MI-DW quantum phase transition.
Here we fix $t/\Omega = 0.05$ and vary $g_2/\Omega$.
Left panel: behavior at fixed system size $L=300$, as a function
of the distance $r$ and for different values of
$g_2/\Omega = 1.3$, $1.35$, $1.4$, and $1.6$.
To minimize boundary effects, we chose the two points $(i, i+r)$
symmetrically with respect to the center of the array.
Right panel: finite-size scaling close to the transition.
Empty circles, filled circles, and triangles respectively are for
$g_2 / \Omega = 1.4$ (DW), $1.35$ (near the critical point), and $1.3$ (MI phase).
In the MI phase, $C_{\mathrm{DW}}$ vanishes exponentially with $L$,
according to the fits: $C_{\mathrm{DW}}^{g_2=1.35} \approx 0.241 \times e^{-0.032 L}$
and $C_{\mathrm{DW}}^{g_2=1.3} \approx 0.253 \times e^{-0.065 L}$.
In the DW phase, $C_{\mathrm{DW}}$ converges to a finite value.
The inset shows such obtained asymptotic value $\mathcal{O}_{\mathrm{DW}}$,
as a function of $g_2$.}
\label{fig4}
\end{figure}
The DW phase is characterized by a finite order parameter ${\mathcal O}_{\mathrm{DW}}$.
Let us therefore look at the two-point staggered correlator in Eq.~\eqref{eq:DWcorr}.
Since in DMRG simulations we are employing open boundary conditions,
to minimize the border effects we analyze the correlations in such a way that
the two points are taken symmetrically with respect to the center
of the system~\footnote{The two points of
$\langle \delta n_i^{\mathrm{pol}} \delta n_j^{\mathrm{pol}} \rangle$,
with $\vert i - j \vert = r$, have been chosen such that
$i = (L - r + 1)/2, j = (L + r + 1)/2$ for odd r,
and $i = (L - r)/2, j = (L + r)/2$ for even r
(e.g. for $L = 100$ sites, $r = 1$ corresponds to $i = 50, j = 51$;
$r = 2$ corresponds to $i = 49, j = 51$; $r = 3$ to $i = 49, j = 52$, and so on)}.
The left panel of Fig.~\ref{fig4} shows how differently such polariton correlations
behave when the system goes from the MI to DW phase, for a fixed system size.
To be more accurate, in the right panel we performed a finite-size scaling
and showed that the staggered correlation $C_{\rm DW}(r)$
approaches the zero value exponentially with $L$,
in the MI phase (a similar behavior occurs in the SF region).
On the other hand, in the DW such correlator asymptotically converges to a finite value.
In the specific, here we fix $t/\Omega=0.05$ and show that for $g_2/\Omega = 1.3, \, 1.35$
the DW order is exponentially suppressed with $L$, while for $g_2/\Omega = 1.4$
it remains finite.
The ${\mathcal O}_{\mathrm{DW}}$ order parameter reached for $L \to \infty$
is displayed in the inset as a function of $g_2$.
In order to determine the DW boundary in the phase diagram of Fig.~\ref{fig2},
we adopted the following protocol.
For a fixed value of $t/\Omega$, we start increasing $g_2$ from zero up to a finite value,
with a fixed increment $\delta g_2 = 0.05 \Omega$, and to compute the DW order parameter
for all such values of $g_2$.
The boundary of DW phase in the $g_2-t$ plane (filled circles in Fig.~\ref{fig2}),
for any fixed $t$, is located by the $g_2^*(t)$ that gives the first non-vanishing
order parameter $\mathcal{O}_{\mathrm{DW}}$.
Here we stress that, because of the arrangement
of our 1D array [see Fig.~\ref{fig1}(a)]
and of the asymmetric coupling between the atom and its right/left cavity,
the antiferromagnetic symmetry of the system is spontaneously broken.
In particular, the state $\ket{4}$ of the $L$-th atom
will be never occupied, since the transition $\ket{2} \leftrightarrow \ket{4}$
does not couple to any cavity mode [see Eq.~\eqref{hamiltonian}].
As a consequence, in our simulations we do not need any symmetry-breaking potential.
We can observe that the expectation value for the onsite number of polaritons
explicitly exhibits a staggered behavior, in that the occupation of the $(2n-1)$-th site
is always higher than that of the $(2n)$-th site (for any integer value of $n$).
Finally we notice that such staggering persists at finite size,
also for the set of parameters corresponding to the MI phase,
although it is extremely tiny and decreases with $L$.
This effect eventually disappears in the thermodynamic limit.
The extension of the DW phase depends on the cavity detuning $\Delta$.
In particular, the robustness of the order parameter increases with increasing the modulus
of the detuning (see Fig.~\ref{funDelta}). Quite remarkably, we note that a positive $\Delta$
will never stabilize an antiferromagnetic DW ordering.
\begin{figure}[!t]
\includegraphics[width = 0.5\hsize]{funDelta.pdf}
\caption{{\bf DW phase order parameter as a function of detuning.}
For a positive detuning, the system will never present as a DW phase,
meaning that $C_{DW}(L/2)=0$. Taking a negative value of the detuning,
we observe that $C_{DW}(L/2)$ increases with $\vert \Delta \vert$.
The parameters in this figure are $t/\Omega=0.05$, $g_1/\Omega=0.8$,
$g_2/\Omega=1.6$ and the system size is $L=100$.}
\label{funDelta}
\end{figure}
\subsection{Neutral gap}
\label{sec:neutral_gap}
The analysis leading to the phase diagram in Fig.~\ref{fig2} has been corroborated
by a study of the neutral gap, which vanishes both in proximity of
the phase transitions and in the entire superfluid region.
Differently for the charge gap, it is able to detect the presence of excitations
other than particle-hole, and thus locates the boundaries of insulating regions
(as the DW) beyond the MI.
\begin{figure}[!b]
\includegraphics[width = 0.495\hsize]{fig5a_N.pdf}
\includegraphics[width = 0.495\hsize]{fig5b_N.pdf}
\caption{{\bf Analysis of the neutral gap.}
Neutral energy gap as a function of $g_2$ for $t/\Omega=0.25$, {\it i.e.}, along
the vertical cut depicted in Fig.~\ref{fig2}.
In the left panel, the different curves are for various system sizes
according to the legend, and for a fixed number of kept states $m=80$
in the DMRG algorithm.
The right panel evidences the convergence of the data, at fixed $L=80$, by increasing $m$
(see also the inset, where we show the behavior of $\Delta E_n$ as
a function of $m$, for three different values of $g_2$.}
\label{fig5}
\end{figure}
The data displayed in Fig.~\ref{fig5} show the behavior of $\Delta E_n$ as a function
of $g_2$, for a fixed value of $t / \Omega$.
In particular we analyzed a vertical cut in the phase diagram of Fig.~\ref{fig2}
(see the rightmost vertical dotted line in that figure), where the system can be
in three different phases according to the value of $g_2$.
With increasing $g_2$, it goes from the MI phase (nonzero $\Delta E_n$, for $t/\Omega \lesssim 1.45$)
to the SF phase (zero $\Delta E_n$, for $1.45 \lesssim t/\Omega \lesssim 1.8$),
and then to the DW phase (nonzero $\Delta E_n$, for $t/\Omega \gtrsim 1.8$).
While we cannot see a clear signature of a finite gap for $g_2 = 1.8 \Omega$,
the scaling with the size displayed in the left panel of Fig.~\ref{fig5}
seems to suggest the scenario depicted above.
It is however important to stress that the DMRG simulations needed to compute the neutral gap
have to target the two lowest-lying eigenstates in a single run.
Thus they generally require a larger dimension $m$ of the effective Hilbert space,
as compared to all the other ground-state calculations discussed before.
The analysis of the neutral gap requires a careful convergence test of the
results with $m$, which we provide in the right panel of Fig.~\ref{fig5}.
We observe that the non monotonic features that are visible in the region
$1.45 \lesssim t/\Omega \lesssim 1.8$ have to be probably ascribed
to the inaccuracy of the method at small $m$ values.
This signals the presence of the gapless SF phase there, in agreement
with the results provided by the charge gap (MI/SF boundary)
and for the DW order parameter (SF/DW boundary).
\section{Summary}
\label{sec:summary}
Using the density-matrix renormalization group with open boundary conditions,
we studied the equilibrium phase diagram of a system of coupled QED cavities in one dimension.
We provided results beyond the standard model of couplings through photon hopping,
and also considered nearest-neighbor cross-Kerr nonlinearities.
Our analysis is based on a finite-size scaling of the ground-state charge
and neutral gaps, as well as of the density-wave order parameter,
for systems up to 300 sites.
We showed that, beyond the conventional Mott insulator and superfluid phases,
the presence of a nearest-neighbor nonlinear coupling can also
stabilize a density-wave ordering of polaritons.
\begin{backmatter}
\section*{Competing interests}
The authors declare that they have no competing interests.
\section*{Authors contributions}
All the authors participated in the design of the research, analysis of the results, and writing of the paper.
The DMRG code used to run all the simulations of this research has been
developed and written by D. Rossini and coworkers~\cite{dmrgB} (see also \texttt{www.dmrg.it}).
The DMRG simulations were performed by J. Jin.
\section*{Acknowledgements}
We would like to acknowledge our previous collaboration with M. Hartmann and M. Leib
which was inspiring for the present work.
This work was supported by Italian MIUR via FIRB Project RBFR12NLNA and PRIN Project 2010LLKJBX,
by EU through IP-SIQS, and by National Natural Science Foundation of China under Grant No. 11175033 and No. 11305021.
|
\section{Introduction}
\label{sect:intro}
Let $n, k_1, \ldots, k_n$ be positive integers. \emph{Multiple polylogarithms} (MPLs) in a single variable are defined by
\begin{align*}
\Li_{k_1,\ldots, k_n}(z) := \sum_{m_1>\cdots>m_n>0} \frac{z^{m_1}}{m_1^{k_1} \cdots m_n^{k_n}}
\end{align*}
when $z$ is a complex number. The function $\Li_{k_1,\ldots, k_n}(z)$ is of depth $\operatorname{dpt}(\mathbf{k}):=n \ge 1$ and weight $\operatorname{wt}(\mathbf k):=k_1+\cdots+k_n$, for $\mathbf{k}:=(k_1,\ldots,k_n)$. It is analytic in the open unit disk and, in the case $k_1 > 1$, continuous on the closed unit disk. In this case we observe the connection of MPLs and \emph{multiple zeta values} (MZVs)
\begin{align}
\label{MZVs-sum}
\zeta(k_1,\ldots,k_n):= \sum_{m_1>\cdots>m_n>0}\frac{1}{m_1^{k_1}\cdots m_n^{k_n}} =\Li_{k_1,\ldots,k_n}(1).
\end{align}
\noindent Equivalently we can define MPLs by induction on the weight $\operatorname{wt}(\mathbf k)$ as follows:
\begin{align}\label{eq:diff1}
z\frac{d}{dz}\Li_{k_1,\ldots,k_n}(z)=\Li_{k_1-1,k_2, \ldots, k_n}(z) \hspace{0.3cm} \text{if}\hspace{0.3cm} k_1>1,
\end{align}
\begin{align}\label{eq:diff2}
(1-z)\frac{d}{dz}\Li_{1,k_2,\ldots,k_n}(z)=\Li_{k_2,\ldots,k_n}(z) \hspace{0.3cm} \text{if}\hspace{0.3cm} n>1,
\end{align}
\begin{align}\label{eq:initial}
\Li_{k_1,\ldots,k_n}(0)=0.
\end{align}
Therefore one can observe an integral formula for MPLs using iterated Chen integrals. Indeed, let $\varphi_1, \ldots, \varphi_p$ be complex-valued differential $1$-forms defined on a compact interval. Then we define inductively for real numbers $x$ and $y$
\begin{align*}
\int_x^y \varphi_1\cdots \varphi_p:=\int_x^y\varphi_1(t)\int_x^t\varphi_2\cdots \varphi_p.
\end{align*}
Now we set
\begin{align*}
\omega_{k_1,\ldots,k_n}:=\omega_0^{k_1-1}\omega_1 \cdots \omega_0^{k_n-1}\omega_1,
\end{align*}
where
\begin{align*}
\omega_0(t):=\frac{dt}{t} \hspace{0.4cm} \text{and} \hspace{0.4cm} \omega_1(t):=\frac{dt}{1-t}.
\end{align*}
Using the differential equations \eqref{eq:diff1}, \eqref{eq:diff2} and the initial conditions \eqref{eq:initial} we obtain
\begin{align}
\label{eq:integralrep}
\Li_{k_1,\ldots,k_n}(z) = \int_0^z \omega_{k_1,\ldots,k_n}
\end{align}
using the convention $\Li_{\emptyset}(z)=1$.
This representation gives rise to the well known shuffle products of MPLs and MZVs (see e.g. \cite{Waldschmidt02,Waldschmidt11}).\\
Recall that the $\mathbb Q$-vector space spanned by MZVs forms an algebra equipped with two products. The quasi-shuffle product is obtained when one multiplies series (\ref{MZVs-sum}) directly, which yields a linear combination of MZVs due to the product rule for sums. The aforementioned shuffle product between MZVs derives from integration by parts for iterated integrals. The resulting so-called double shuffle relations among MZVs arise from the interplay between these two products. An alternative characterization of MPLs can be given by the following formula
\begin{align}
\label{eq:Jcharact}
\Li_{k_1,\ldots,k_n}(z) = J^{k_1}[yJ^{k_2}[y \cdots J^{k_n}[y]\cdots ]](z),
\end{align}
where $y(z):=\frac{z}{1-z}$ and $J[f](z):=\int_0^z \frac{f(t)}{t}\,dt$ (see Lemma \ref{lem:Jiteration}). Since $J$ is a Rota--Baxter operator of weight zero, iterations of the operator $J$ induce a product that coincides with the usual shuffle product for MPLs (see Lemma \ref{lem:jash}). For $|z|<1$ Equation \eqref{eq:Jcharact} is valid for all $k_1,\ldots,k_n\in \bZ$. The inverse of $J$ is given by $J^{-1}[f](t) = \delta[f](t):=t\frac{\partial f}{\partial t}(t)$ (Proposition \ref{prop:RBOJ}).
We study the $\bQ$-vector space
\begin{align*}
\mathcal{MP}:=\langle z\mapsto \Li_{-k_1,\ldots,-k_n}(z) \colon k_1,\ldots,k_n\in \bN_0, n\in \bN \rangle_{\bQ},
\end{align*}
which is indeed an algebra, where the product is induced by Equation \eqref{eq:Jcharact} (see Lemma \ref{lem:Jalgebra}).
The algebra $\mathcal{MP}$ admits also an interpretation for MZVs at non-positive integers. Indeed, let $k_1,\ldots,k_n \in \bN_0$. It is easily seen that $\Li_{-k_1,\ldots,-k_n}(z)$ is convergent for $|z|<1$ and divergent for $z=1$. Nevertheless we can perceive the product induced for $|z|<1$ as an analogue for the shuffle product of MZVs at non-positive integers. In order to make this connection more precise we have to establish a renormalization procedure. This permits us to extract explicit numbers for MZVs with non-positive arguments in a consistent way, such that they satisfy the shuffle product relations induced by the algebra structure of $\mathcal{MP}$.\\
We should keep in mind that a characterization of the shuffle product at non-positive integers -- in contrast to the quasi-shuffle product -- is a crucial point. Since the quasi-shuffle product is induced by the series representation of MZVs, the combinatorics is essentially the same as for positive arguments. On the other hand the shuffle product for positive indices is induced by the integral representation \eqref{eq:integralrep}. The combinatorics behind this product comes from shuffling of integration variables. It could be illustrated by the shuffling of two decks of cards, say a deck of red and blue cards, each consecutively numbered such that the internal numbering of red and blue cards is preserved. In this approach, however, it is not clear how to handle non-positive arguments, which corresponds to a non-positive number of cards.\\
Extracting finite numbers for MZVs at non-positive integers is accomplished by the process of renormalization, which involves two steps:
\begin{enumerate}[(I)]
\item introduction of a regularization scheme,
\item applying a subtraction method.
\end{enumerate}
\noindent In step (I) we consider divergent MZVs $\zeta(-k_1,\ldots,-k_n)$ with $k_1,\ldots,k_n\in \bN_0$, and introduce a so-called regularization parameter $z$, which systematically deforms the divergent MZV in order to obtain a meromorphic function in $z$ with the only singularity in $z=0$.
Step (II) involves a systematic procedure to eliminate singularities in terms of recursively defined subtractions. A rather natural way to achieve such eliminations is widely known as minimal subtraction scheme. The renormalization process is an integral part of perturbative quantum field theory (QFT). See e.g.~\cite{Collins}. The ``right" choice of the regularization scheme in QFT is essential in the light of constraints coming from physics. In our context those constraints are of mathematical nature: The deformation of divergent MZVs has to be established in such a way that the regularized MZVs coincide with the meromorphic continuation of (M)ZVs. The recursively defined subtractions in step (II) involve combinatorial structures, which are concisely captured by the Connes--Kreimer Hopf algebraic approach to renormalization \cite{Connes00, Connes01, Manchon08}.\\
One of the key points in our approach is based on providing an adequate Hopf algebra together with an algebra morphism from that Hopf algebra into the space $\mathcal{MP}$ (Theorem \ref{theo:HopfQuot}), which permits to define a consistent renormalization process. Regarding regularization schemes, we will use the fact, that MPLs may be considered as regularized classical MZVs. However, we also consider a specific $q$-analogue of MZVs \cite{Ohno12}, where the variable $q$ takes the role of a natural regulator.
\begin{remark} {\rm{Renormalization of MZVs at non-positive integers appeared already in \cite{Guo08} and \cite{Manchon10}. The authors applied regularization schemes together with well-chosen subtraction methods suitable for preserving the quasi-shuffle product for renormalized MZVs -- at non-positive arguments. Using Ecalle's Mould calculus, Bouillot proposes in his work \cite{Bouillot13} a unifying picture of MZVs at non-positive arguments respecting the quasi-shuffle product. The common point of our approach with those presented in the aforementioned references is the use of the Connes--Kreimer Hopf algebraic approach to renormalization, and the corresponding algebraic Birkhoff decomposition, which encodes the subtraction procedure for singularities. However, we should emphasize that in our work it is the shuffle product, in a naturally extended sense, which is satisfied by renormalized MZVs at non-positive arguments.
We remark that Zhao's approach in \cite{Zhao08} is rather different from the point of view we present in this paper, since it extended the approach of \cite{Guo08} to a particular $q$-analog of MZVs preserving the quasi-shuffle product, while not attributing any regularization properties to the $q$-parameter itself.}}
\end{remark}
In \cite{Castillo13b} the authors indicated that the $q$-parameter appearing in a specific $q$-analogue of MZVs \cite{Ohno12} (see Equation \eqref{eq:qMPLs} below) may be considered as a regularization parameter for MZVs at non-positive arguments. The approach presented in our paper can be consistently extended to this $q$-analogue of MZVs ($q$MZVs). Indeed, we would like to demonstrate that under the $q$-parameter regularization \cite{Castillo13a,Castillo13b,Ohno12} a proper renormalization of MZVs can be defined. The \emph{$q$-multiple polylogarithm} ($q$MPL) in one variable is defined as
\begin{align}
\label{qMPL}
\Li^q_{k_1,\ldots,k_n}(z):=\sum_{m_1>\cdots >m_n>0}\frac{z^{m_1}}{[m_1]_q^{k_1}\cdots [m_n]_q^{k_n}}
\end{align}
with $[m]_q:=\frac{1-q^m}{1-q}$. It turns out that for $|z|<1$ the series in (\ref{qMPL}) is convergent for $k_1,\ldots,k_n\in \bZ$, and especially for $|q|<1$ we obtain the formal power series
\begin{align}
\label{eq:qMPLs}
\mathfrak{z}_q(k_1,\ldots,k_n):= \Li^q_{k_1,\ldots,k_n}(q)
= \sum_{m_1>\cdots>m_n>0} \frac{q^{m_1}}{[m_1]_q^{k_1}\cdots [m_n]_q^{k_n}} \in \bZ[[q]].
\end{align}
These $q$MZVs were introduced by Ohno, Okuda and Zudilin in \cite{Ohno12}, and further studied in \cite{Castillo13a,Castillo13b}, see also \cite{Singer15,Singer14,Zhao14}. For $k_1 > 1$ and $k_2,\ldots,k_n\geq 1$ we see that
\begin{align}\label{eq:lim}
\lim_{q\nearrow 1}\mathfrak{z}_q(k_1,\ldots,k_n) = \z(k_1,\ldots,k_n),
\end{align}
where $q \nearrow 1$ means $q\to 1$ inside an angular sector $-\frac{\pi}{2} +\varepsilon \leq \operatorname{Arg}(1-q) \leq \frac{\pi}{2} - \varepsilon$ with $\varepsilon>0$ sufficiently small. It will be convenient for technical reasons to consider the \emph{modified $q$MZVs} introduced in \cite{Ohno12}
\begin{align}\label{eq:modify}
\overline{\mathfrak{z}}_q(k_1,\ldots,k_n) := (1-q)^{-(k_1+\cdots+k_n)}\mathfrak{z}(k_1,\ldots,k_n).
\end{align}
They are used to establish a Hopf algebra structure on the space of modified $q$MPLs. The modification has to be reversed after renormalization, in order to relate the renormalized $q$MZVs via \eqref{eq:lim} to renormalized MZVs.
\medskip
The paper is organized as follows. In Section \ref{sect:mero} we recall the basic results on the meromorphic continuation of MZVs. Section \ref{sect:algebra} contains the main result, i.e., the detailed construction of a Hopf algebra for MPLs at non-positive integers. A generalization of this finding to the $q$-analogue of MZVs defined by Ohno, Okuda and Zudilin is presented as well. In Section \ref{sect:Renorm} we recall the Hopf algebra approach to perturbative renormalization by Connes and Kreimer, and apply one of its main theorems to the renormalization of MPLs at non-positive integer arguments. The $q$-analogue of this result is as well presented, and compared to the classical case.
\medskip
\noindent {\bf{Acknowledgement}}: The first author is supported by a Ram\'on y Cajal research grant from the Spanish government. The second and third authors gratefully acknowledge support by ICMAT and the Severo Ochoa Excellence Program. The second author is supported by Agence Nationale de la Recherche (projet CARMA).\\
The authors gratefully acknowledge that most of the discussions and work was carried out during ICMAT Fall School Multiple Zeta Values, Multiple Polylogarithms and Quantum Field Theory (Oct. 7-11, 2013, ICMAT, Madrid), Clay Mathematics Institute Summer School 2014 (June 30 - July 25, 2014, Madrid) and ICMAT Research Trimester on Multiple Zeta Values, Multiple Polylogarithms, and Quantum Field Theory (Sept. 15 - Dec. 19, 2014, ICMAT, Madrid).
\section{Meromorphic Continuation of MZVs}
\label{sect:mero}
In this section we review some well-known facts about the meromorphic continuation of MZVs. For $n \in \bN$ we consider the function
\begin{align}
\label{eq:MRiemann}
\z_n \colon \bC^n \to \bC,
\hspace{1cm}
\z_n(s_1,\ldots,s_n):= \sum_{m_1 > \cdots > m_n >0}\frac{1}{m_1^{s_1}\cdots m_n^{s_n}}.
\end{align}
\begin{proposition}[\cite{Krattenthaler07}]
The domain of absolute convergence of the function \eqref{eq:MRiemann} is given by
\begin{align*}
\left\{(s_1,\ldots,s_n)\in \bC^n\colon \sum_{j=1}^k\RE(s_j)>k, k=1,\ldots,n \right\}.
\end{align*}
In this domain $\zeta_n$ defines an analytic function in $n$ variables.
\end{proposition}
\begin{theorem}[\cite{Akiyama01a,Akiyama01b,Manchon10}]
\label{theo:meroz}
The function $\z_n(s_1,\ldots,s_n)$ admits a meromorphic extension to $\bC^n$. The subvariety $\mathcal S_n$ of singularities is given by
\begin{align*}
\mathcal S_n= \left\{(s_1,\ldots,s_n)\in \bC^n\colon s_1=1; s_1+s_2=2,1,0,-2,-4,\ldots; \sum_{i=1}^js_i\in \bZ_{\leq j} ~(j=3,4,\ldots,n)\right\}.
\end{align*}
\end{theorem}
\noindent In the subsequent sections $\z_n$ always denotes the meromorphic continuation of MZVs.
\begin{remark}\label{rem:mero}
{\rm{In this paper we discuss $\z_n$ restricted to the set $(\bZ_{\leq0})^n$. The Bernoulli numbers are defined by the following generating series
\begin{align*}
\frac{te^t}{e^t-1} = \sum_{m\geq0} \frac{B_m}{m!}t^m.
\end{align*}
The first few values are $B_0=1, B_1=\frac{1}{2}, B_2=\frac{1}{6},B_3=0, B_4=-\frac{1}{30}, B_5=0,$ etc., especially $B_{2l+1}=0$ for $l\in \bN$.
Therefore we have the following cases for $\z_n$ restricted to non-positive arguments:
\allowdisplaybreaks{
\begin{itemize}
\item Case $n=1$: For $l \in \bN_0$ we have the well known formula
\begin{align*}
\z_1(-l) = -\frac{B_{l+1}}{l+1}.
\end{align*}
\item Case $n=2$: In the light of Theorem \ref{theo:meroz} we assume the sum $k_1+k_2$ to be odd. Therefore we obtain from \cite{Akiyama01a} that
\begin{align*}
\z_2(-k_1,-k_2) = \frac{1}{2} \left(1+\delta_0(k_2) \right) \frac{B_{k_1+k_2+1}}{k_1+k_2+1}.
\end{align*}
\item Case $n\geq 3$: From Theorem \ref{theo:meroz} we deduce that
\begin{align*}
(\bZ_{\leq 0})^n \subseteq \mathcal S_n.
\end{align*}
Therefore we obtain no information from the meromorphic continuation.
\end{itemize}
}}}
\end{remark}
\section{Algebraic Framework}
\label{sect:algebra}
We briefly introduce Rota--Baxter algebras, since they conveniently relate to shuffle-type products on word algebras. Two such shuffle products are presented, which encode products of MPLs and $q$MPLs at integer arguments. The main result of this section is the construction of a graded connected commutative and cocommutative shuffle Hopf algebra for ($q$)MPLs at non-positive integer arguments.
\subsection{Rota--Baxter Algebra and multiple zeta values}
\label{ssect:RBA}
Let $k$ be a ring, $\lambda \in k$ and $\mathcal{A}$ a $k$-algebra. A \emph{Rota--Baxter operator (RBO) of weight $\lambda$ on $\mathcal{A}$ over $k$} is a $k$-module endomorphism $L$ of $\mathcal{A}$ such that
\begin{align*}
L(x)L(y) = L(xL(y)) + L(L(x)y) + \lambda L(xy)
\end{align*}
for any $x,y \in \mathcal{A}$. A \emph{Rota--Baxter $k$-algebra (RBA) of weight $\lambda$} is a pair $(\mathcal{A},L)$ with a $k$-algebra $\mathcal{A}$ and a Rota--Baxter operator $L$ of weight $\lambda$ on $\mathcal{A}$ over $k$. On the algebra of continuous functions $C(\bR)$ the integration operator
\begin{align*}
R\colon C(\bR) \to C(\bR), \hspace{0.5cm} R[f](z):=\int_{0}^zf(x)\,dx
\end{align*}
is a RBO of weight zero, which is an immediate consequence of the integration by parts formula. We consider the $\bC$-algebra of power series
\begin{align*}
\mathcal{P}_{\geq 1}:=\bigg\{ f(t):=\sum_{k\geq 1}a_k t^k\colon R_f\geq 1 \bigg\}\subseteq t\bC[[t]]
\end{align*}
without a term of degree zero in $t$, and radius of convergence, $R_f$, of at least $1$. We define the operator
\begin{align*}
J\colon \mathcal{P}_{\geq 1} \to\mathcal{P}_{\geq 1}, \hspace{0.5cm} J[f](t):=\int_{0}^t f(z)\frac{dz}{z}.
\end{align*}
Further the \emph{Euler derivation} $\delta$ is given by
\begin{align*}
\delta\colon \mathcal{P}_{\geq 1} \to \mathcal{P}_{\geq 1}, \hspace{0.5cm} \delta[f](t):=t\frac{\partial f}{\partial t}(t).
\end{align*}
\begin{proposition}\label{prop:RBOJ} ~
\begin{enumerate}[(i)]
\item The pair $(\mathcal{P}_{\geq 1},J)$ is a RBA of weight $\lambda=0$.
\item The operator $\delta$ is a derivation, i.e., $\delta[fg] = \delta[f]g + f\delta[g]$, for any $f,g\in \mathcal{P}_{\geq 1}$.
\item The operators $J$ and $\delta$ are mutually inverse, i.e., $J\circ \delta = \delta \circ J = \Id$.
\end{enumerate}
\end{proposition}
\begin{proof}
Statement (i) follows form integration by part. The second claim is straightforward to show. Finally, item (iii) is an immediate consequence of the fundamental theorem of calculus together with the fact that $f(0)=0$ for any $f\in \mathcal{P}_{\geq 1}$.
\end{proof}
\begin{lemma}
\label{lem:Jiteration}
Let $k_1,\ldots,k_n$ be integers. Then $\Li_{k_1,\ldots,k_n}(t) \in \mathcal{P}_{\geq 1}$, explicitly
\begin{align*}
\Li_{k_1,\ldots,k_n}(t)=J^{k_1}[y J^{k_2}[y \cdots J^{k_n}[y]\cdots]](t),
\end{align*}
where $y(t):=\frac{t}{1- t}\in \mathcal{P}_{\geq 1}$.
\end{lemma}
\begin{proof}
Using the fact that $J^{-1}=\delta$ we prove the claim for $\bk:=(k_1,\ldots,k_n)\in \bZ^n$ by induction on its depth, $\operatorname{dpt}(\bk)=n$. For $\operatorname{dpt}(\bk)=1$ we easily compute
\begin{align*}
J^k[y](t)& = \left.\begin{cases}
\sum_{m\geq 1}\frac{t^m}{m^k}, & \text{for~} k\geq 0 \\
\sum_{m\geq 1}m^{|k|}{t^m} = \sum_{m\geq 1}\frac{t^m}{m^k}, & \text{for~} k< 0
\end{cases} \right\}
= \Li_k(t)
\end{align*}
for any $k \in \bZ$. In the inductive step we get
\allowdisplaybreaks{
\begin{align*}
J^{k_1}[y J^{k_2}[y \cdots J^{k_n}[y]\cdots]](t)
& = J^{k_1}\left[\sum_{m>0}t^m \sum_{m_2 > \cdots > m_n >0}\frac{t^{m_2}}{m_2^{k_2}\cdots m_n^{k_n}} \right] \\
& = J^{k_1}\left[ \sum_{m_1 >m_2 > \cdots >m_n >0}\frac{t^{m_1}}{m_2^{k_2}\cdots m_n^{k_n}} \right] \\
& = \sum_{m_1 >m_2 > \cdots >m_n >0}\frac{t^{m_1}}{m_1^{k_1}m_2^{k_2}\cdots m_n^{k_n}} \\
& = \Li_{k_1,\ldots,k_n}(t)
\end{align*}}
using the induction hypothesis.
\end{proof}
This lemma gives rise to the following algebraic formalism. Let $X_0:=\{j,d,y\}$, and $W_0$ denotes the set of words on the alphabet $X_0$, subject to the rule $jd=dj=\be$, where $\be$ denotes the empty word. Therefore any word $w\in W_0$ can be uniquely written in the canonical form
\begin{align*}
w=j^{k_1}yj^{k_2}y\cdots j^{k_{n-1}}y j^{k_n}
\end{align*}
for $k_1,\ldots,k_n\in \bZ$ using the notation $j^{-1}=d$ and $j^0=\be$. The length of the word $w$ above is $|w|=k_1+\cdots+k_n+n-1$.
Further, $\mathcal{A}_0$ denotes the vector space $\mathcal{A}_0:=\langle W_0 \rangle_\bQ$ spanned by the words in $W_0$.
Next we define the product $\sho \colon \mathcal{A}_0\otimes \mathcal{A}_0 \to \mathcal{A}_0$ by $\be\sho w := w \sho \be := w$ for any word $w\in W_0$, and recursively with respect to the sum of the length of two words in $W_0$:
\allowdisplaybreaks{
\begin{enumerate}[(i)]
\item $yu\sho v := u \sho yv := y(u\sho v)$,
\item $ju \sho jv := j(u\sho jv) + j(ju\sho v)$,
\item $du\sho dv := d(u \sho dv) - u \sho d^2v $,
\item $du\sho jv := d(u\sho jv)-u \sho v$,
\item $ju\sho dv := d(ju\sho v)-u\sho v$.
\end{enumerate}}
\begin{remark}{\rm{
\textbullet~ Note that (iv) can be deduced from (iii) by replacing $v$ by $j^2v$. \\
\textbullet~ (iii) does not really define $du\sho dv$ by induction on the sum of lengths of two words, because $|du|+|dv|=|u|+|d^2v|$. Using (i) and writing $u'=du=d^kyw$ for some $k\ge 1$, we can however get a recursive definition by iterating (iii) as follows:
\begin{align*}
d^kyw\sho dv =&~d\big(d^{k-1}yw\sho dv-d^{k-2}yw\sho d^2v+\cdots\\
& +(-1)^{k-1}yw\sho d^kv\big)+(-1)^ky(w\sho d^{k+1}v).
\end{align*}
}}
\end{remark}
\begin{lemma}\label{lem:idealT}
The $\bQ$-vector space
\begin{align*}
\mathcal{T}:=\langle j^{k_1}yj^{k_2}y\cdots j^{k_{n-1}}y j^{k_n}\in W_0 \colon k_n\neq 0, n\in \bN \rangle_{\bQ}
\end{align*}
is a two sided ideal of $(\mathcal{A}_0,\sho)$.
\end{lemma}
\begin{proof}
Let $a\in \{j,d\}$ and $u:=u'a\in W_0$ and $v\in W_0$. We prove $u\sho v \in \mathcal{T}$ by induction on $r:=|u|+|v|$. The base cases are true because we observe for $r=1$
that $d\sho \be = d, j\sho \be = j$ and for $r=2$ that
\begin{align*}
d\sho y = yd, \hskip 8mm j\sho y = yj,\\
d\sho d = 0,\hskip 4mm j\sho j = 2j^2,\hskip 4mm j\sho d=d\sho j=0.
\end{align*}
For the inductive step we have several cases: \\
\noindent\textbullet~ 1st case: $u=y\tilde{u}a$ or $v=y\tilde{v}$. This is an immediate consequence of (i) and the induction hypothesis. \\
\noindent\textbullet~ 2nd case: $u=j\tilde{u}a$ and $v=j\tilde{v}$. We observe using (ii) and the induction hypothesis that
\begin{align*}
j\tilde{u}a \sho j\tilde{v} = j(\tilde{u}a \sho j\tilde{v} + j\tilde{u}a \sho\tilde{v}) \in \mathcal{T}.
\end{align*}
\noindent\textbullet~ 3rd case: $u=d\tilde{u}a$ and $v=j\tilde{v}$. We observe using (iv) and the induction hypothesis that
\begin{align*}
d\tilde{u}a \sho j\tilde{v} = d(\tilde{u}a \sho j\tilde{v}) - \tilde{u}a \sho \tilde{v} \in \mathcal{T}.
\end{align*}
\noindent\textbullet~ 4th case: $u=j\tilde{u}a$ and $v=d\tilde{v}$. We observe using (v) and the induction hypothesis that
\begin{align*}
j\tilde{u}a \sho d\tilde{v} = d(j\tilde{u}a \sho \tilde{v}) - \tilde{u}a \sho \tilde{v} \in \mathcal{T}.
\end{align*}
\noindent\textbullet~ 5th case: $u=d\tilde{u}a$ and $v=d\tilde{v}$. We observe using (iii) that
\begin{align*}
d\tilde{u}a \sho d\tilde{v} = d(\tilde{u}a \sho d\tilde{v}) - \tilde{u}a \sho d^2\tilde{v}.
\end{align*}
By induction hypothesis the first term is an element of $\mathcal{T}$. If $\tilde{u}a$ is not a word consisting purely of $d$ we apply rule (iii) until we hit a letter not equal to $d$ and we are in one of the above cases. Therefore we only consider the case, where $\tilde{u}a$ is a word consisting purely of $d$, i.e., $\tilde{u}a=d^n$ for $n\in \bN$. Now we prove $d^n\sho d^mw=0$ for any $w\in W_0$ and $m\in \bN$. For $n=1$ we have $d\sho d^m w= d(\be \sho d^m w)- \be \sho d^{m+1}w = 0$. Therefore we obtain by induction hypothesis that
\begin{align*}
d^{n+1} \sho d^m w = d(d^{n}\sho d^{m}w) - d^n\sho d^{m+1}w =0.
\end{align*}
All in all we have shown that $\sho(\mathcal{T} \otimes \mathcal{A}_0 )\subseteq \mathcal{T} $.
Since $\sho$ is not commutative we also have to prove $v\sho u \in \mathcal{T}$ by induction on $r:=|v|+|u|$. The base cases are true. The first four cases are completely analogous to the first four cases above. We only discuss the following case: $v=d\tilde{v}$ and $u=d\tilde{u}a$. \\
We observe using (iii) that
\begin{align*}
d\tilde{v} \sho d\tilde{u}a = d(\tilde{v} \sho d \tilde{u}a ) - \tilde{v} \sho d^2\tilde{u}a.
\end{align*}
By induction hypothesis the first term is an element of $\mathcal{T}$. If $\tilde{v}$ starts with $j$ or $y$ we are in one of the above cases. Only the case $\tilde{v}=d^n$ for $n\in \bN$ has to be considered. By the same induction as in the previous case we obtain that the last term is zero. This proves $\sho(\mathcal{A}_0 \otimes \mathcal{T} )\subseteq \mathcal{T}$. The proof is now complete.
\end{proof}
Let $Y_0:=\{\be\}\cup W_0y$ be the set of \textsl{admissible words}, i.e., words which do not end up with a $j$ or a $d$. It is easily seen that $\mathcal{A}'_0:=\langle Y_0 \rangle_\bQ$ is a subalgebra of $(\mathcal{A}_0,\sho)$ isomorphic to $\mathcal{A}_0/\mathcal{T}$. A priori, the product $\sho$ on $\mathcal{A}_0$ is neither commutative nor associative. Now let $\mathcal{L}$ (resp. $\mathcal{L}'$) be the ideal of $\mathcal{A}_0$ (resp. $\mathcal{A}'_0$) generated by
$$\{j^k\big(d(u\sho v)-du\sho v-u\sho dv\big),\,u,v\in W_0y, \,k\in\bZ\}.$$
Let $\mathcal{B}_0$ (resp. $\mathcal{B}'_0$) be the quotient algebra $\mathcal{A}_0/\mathcal{L}$ (resp. $\mathcal{A}'_0/\mathcal{L}'$). We obviously have the isomorphism:
$$\mathcal{B}'_0\sim \mathcal{A}_0/(\mathcal{T}+\mathcal{L}).$$
\begin{proposition}\label{lem:Jalgebra}
The pair $(\mathcal{B}_0,\sho)$ is a commutative, associative and unital algebra.
\end{proposition}
\begin{proof}
We first prove commutativity $u'\sho v'-v'\sho u'\in \mathcal{L}$ by induction on $r=|u'|+|v'|$. The cases $r=0$ and $r=1$ are immediate. Several cases must be considered:
\noindent\textbullet~1st case: $u'=yu$ or $v'=yv$. The induction hypothesis immediately applies, using (i).\\
\noindent\textbullet~2nd case: $u'=ju$ and $v'=jv$. Then we have by induction hypothesis:
\begin{align*}
ju\sho jv-jv\sho ju=j(ju\sho v+u\sho jv-jv\sho u-v\sho ju)\in\mathcal{L}.
\end{align*}
\noindent\textbullet~3rd case: $u'=du$ and $v'=jv$ or vice-versa. We have then:
\begin{align*}
du\sho jv-jv\sho du&=d(u\sho jv)-u\sho v-d(jv\sho u)+v\sho u\\
&=d(u\sho jv-jv\sho u)-(u\sho v-v\sho u),
\end{align*}
which belongs to $\mathcal{L}$ by induction hypothesis.\\
\noindent\textbullet~4th case: $u'=du$ and $v'=dv$. Then $dv\sho du-d(dv\sho u)+d^2v\sho u\in\mathcal{L}$, hence:
\begin{align*}
du\sho dv-dv\sho du&=d(u\sho dv)-u\sho d^2v-d(dv\sho u)+d^2v\sho u \mod \mathcal{L}\\
&= d(u\sho dv-dv\sho u)-(u\sho d^2v-d^2v\sho u)\mod \mathcal{L}.
\end{align*}
The first term belongs to $\mathcal{L}$ by induction hypothesis. We further suppose that $u'$ is written $d^kyw$ for some $k\ge 1$ and $w\in Y_0$. Iterating the process using (iii) we finally get $du\sho dv-dv\sho du=(-1)^k(yw\sho d^{k+1}v-d^{k+1}v\sho yw)\mod\mathcal{L}$. We are then back to the first case.
\vskip 3mm
Associativity follows by showing $u'\sho (v'\sho w') = (u'\sho v')\sho w'$ via induction on the sum $|u'|+|v'|+|w'|$. If one of the words is the empty one nothing is to show. Now let $u'=au$, $v'=bv$ and $w'=cw$ with $a,b,c\in \{d,j,y\}$. \\
\noindent\textbullet~1st case: one of the letters is $y$, for example $u'=yu$. Using the induction hypothesis, we obtain
\allowdisplaybreaks{
\begin{align*}
(yu\sho v')\sho w'
= (y(u\sho v'))\sho w' &= y((u\sho v')\sho w') \\
&= y(u\sho (v'\sho w'))\mod \mathcal{L}\\
&= yu \sho (v'\sho w')\mod\mathcal{L}.
\end{align*}}
Note that the other cases $v'=yv$ or $w'=yw$ are similar, and the arguments are completely analogous.\\
\noindent\textbullet~2nd case: $a=b=c=j$. On the one hand we have
\allowdisplaybreaks{
\begin{align*}
(ju\sho jv)\sho jw
=&~ j((u\sho jv)\sho jw) + j((ju\sho v)\sho jw) \\
& + j(j(u\sho jv)\sho w) + j (j(ju \sho v)\sho w)\\
=&~ j((u\sho jv)\sho jw) + j((ju\sho v)\sho jw) + j((ju\sho jv)\sho w),
\end{align*}}
on the other hand
\allowdisplaybreaks{
\begin{align*}
ju\sho (jv\sho jw)
=&~ j(u\sho j(v\sho jw)) + j(u\sho j(jv\sho w)) \\
& + j(ju\sho (v\sho jw)) + j (ju \sho (jv\sho w))\\
=&~ j(u\sho (jv\sho jw)) + j(ju\sho (v\sho jw)) + j(ju\sho (jv\sho w)).
\end{align*}}
Hence $(ju\sho jv)\sho jw =ju\sho (jv\sho jw)\mod \mathcal{L}$.\\
\noindent\textbullet~3rd case: two $j$'s and one $d$. On the one hand we have
\allowdisplaybreaks{
\begin{align*}
(ju \sho jv) \sho dw =&~ d(j(u\sho jv)\sho w) - (u\sho jv)\sho w \\
&+ d(j(ju\sho v)\sho w) - (ju \sho v)\sho w \\
=&~d((ju\sho jv)\sho w) - (u\sho jv)\sho w - (ju\sho v)\sho w,
\end{align*}}
on the other hand
\begin{align*}
ju \sho (jv \sho dw) &= ju\sho d(jv\sho w)-ju\sho (v\sho w)\\
&= d(ju\sho (jv\sho w)) - u\sho (jv \sho w) - ju \sho (v\sho w).
\end{align*}
then $(ju \sho jv) \sho dw=ju \sho (jv \sho dw)\mod \mathcal{L}$.\\
\noindent\textbullet~4th case: two $d$'s and one $j$. We have to prove
\begin{align*}
(du \sho dv) \sho jw = du \sho (dv \sho jw)\mod\mathcal{L}.
\end{align*}
It suffices to show that
\begin{align*}
(d^kyu \sho dv) \sho jw = d^kyu \sho (dv \sho jw)\mod\mathcal{L}
\end{align*}
with $u\in W_0$ and $k\in \bN$. Using (iii) we observe
\allowdisplaybreaks{
\begin{align*}
& (d^kyu \sho dv) \sho jw =d(d^{k-1}yu\sho dv)\sho jw-(d^{k-1}yu\sho d^2v)\sho jw\\
&=d\big((d^{k-1}yu\sho dv)\sho jw\big) - (d^{k-1}yu\sho dv)\sho w-(d^{k-1}yu\sho d^2v)\sho jw \\
&=d\big(d^{k-1}yu\sho (dv\sho jw)\big) - d^{k-1}yu\sho (dv\sho w) -(d^{k-1}yu\sho d^2v)\sho jw \mod\mathcal{L}\\
&=d^kyu\sho (dv\sho jw) + d^{k-1} yu\sho (d^2v\sho jw) - (d^{k-1} yu\sho d^2v)\sho jw \mod\mathcal{L} .
\end{align*}}
For the difference of the two terms in the previous line to belong to $\mathcal{L}$, it suffices to prove
$$
d^{k-1} yu\sho (d^2v\sho jw) =(d^{k-1} yu\sho d^2v)\sho jw\mod\mathcal{L}.
$$
Applying the above procedure iteratively this could be reduced to
\begin{align*}
yu\sho (d^kv\sho jw) = (yu\sho d^kv)\sho jw \mod\mathcal{L},
\end{align*}
which is true by using (i) and the induction hypothesis.\\
\noindent\textbullet~5th case: $a=b=c=d$. We have to prove
\begin{align*}
(du \sho dv) \sho dw = du \sho (dv \sho dw)\mod\mathcal{L}.
\end{align*}
It suffices to show that
\begin{align*}
(d^kyu \sho dv) \sho dw = d^kyu \sho (dv \sho dw)\mod\mathcal{L},
\end{align*}
with $u\in W_0$ and $k\in \bN$. Using (iii) we observe
\allowdisplaybreaks{
\begin{align*}
& (d^kyu\sho dv) \sho dw = d(d^{k-1}yu\sho dv)\sho dw-(d^{k-1}yu\sho d^2v)\sho dw\\
& = d\big((d^{k-1}yu\sho dv)\sho dw\big) - (d^{k-1}yu\sho dv)\sho d^2w -(d^{k-1}yu\sho d^2v)\sho dw \\
& = d^kyu\sho (dv\sho dw) + d^{k-1}yu\sho (d^2v\sho dw)+ d^{k-1}yu\sho (dv\sho d^2w) \\
& ~-(d^{k-1}yu\sho d^2v)\sho dw - (d^{k-1}yu\sho dv)\sho d^2w\mod\mathcal{L}.
\end{align*}}
Iteratively applying this procedure leads -- as in the 4th case -- to the claim using (i) and the induction hypothesis. Proposition \ref{lem:Jalgebra} is thus proven.
\end{proof}
\noindent Now we define the map $\z_t^\shuffle\colon \mathcal{B}'_0 \to \bQ[[t]]$ by $\z_t^\shuffle(\be):=1$, and for any $k_1,\ldots,k_n\in \bZ$,
\begin{align*}
j^{k_1}y\cdots j^{k_n}y \mapsto \z_t^\shuffle(j^{k_1}y\cdots j^{k_n}y) := \Li_{k_1,\ldots,k_n}(t).
\end{align*}
\begin{lemma}\label{lem:characterJ}
The map $\z_t^\shuffle$ is multiplicative, i.e., is an algebra morphism.
\end{lemma}
\begin{proof}
From Proposition \ref{prop:RBOJ} (ii) and (iii) we obtain for any $f,g\in \mathcal{P}_{\geq 1}$
\begin{align}
\label{eq:mixed}
\delta[J[f]g] = J[f]\delta[g] + fg.
\end{align}
Therefore the definition of $\sho $ and Proposition \ref{prop:RBOJ} (i), (ii), (iii) and \eqref{eq:mixed} imply that
\begin{align*}
\z_t^{\,\shuffle\,} \colon \mathcal{B}'_0\to \bQ[[t]],
\hspace{0.5cm} j^{k_1}y \cdots j^{k_n}y \mapsto J^{k_1}[y \cdots J^{k_n}[y]\cdots](t)
\end{align*}
with $k_1,\ldots, k_n\in \bZ$ is an algebra morphism.
\end{proof}
Next we show that if we restrict the shuffle product $\sho $ to admissible words corresponding to positive arguments we obtain the ordinary shuffle product. Let $\mathcal{C}:=\bQ\be\oplus j \bQ \langle j,y \rangle y$ and $\mathcal{D}:=\bQ\be\oplus x_0\bQ\langle x_0,x_1\rangle x_1$.
\begin{lemma}\label{lem:jash}
The algebras $(\mathcal{C},\sho )$ and $(\mathcal{D},\,\shuffle\,)$ are isomorphic, where $\,\shuffle\,$ denotes the ordinary shuffle product.
\end{lemma}
\begin{proof}
It is easily seen that $ \Phi \colon (\mathcal{D},\,\shuffle\,) \to (\mathcal{C},\sho )$ given by $\be \mapsto \be$ and
\begin{align*}
x_0^{k_1-1}x_1 x_0^{k_2-1}x_1 \cdots x_0^{k_n-1}x_1 \mapsto j^{k_1}y j^{k_2}y \cdots j^{k_n}y
\end{align*}
is an algebra morphism, for $k_1,\ldots,k_n\in \bN$, with $k_1>1$, $n\in \bN$. Since $\Phi$ is bijective the proof is complete.
\end{proof}
\subsection{$q$-multiple zeta values}
\label{ssect:qRBA}
For a formal power series $f\in \bQ[[t]]$ we define the $q$-dilation operator as $E_q[f](t):=f(qt).$ Let $\mathcal{A}:=t\bQ[[t,q]]$ be the space of formal power series in the variables $t$ and $q$, without a term of degree zero in $t$. We can interpret $\mathcal{A}$ as the $\bQ[[q]]$-algebra $t\bQ[[t]]$. Then the $\bQ[[q]]$-linear map $P_q\colon \mathcal{A}\to \mathcal{A}$ is defined by
\begin{align}
\label{eq:opP}
P_q[f](t):=\sum_{n\geq 0}E_q^n[f](t).
\end{align}
Furthermore, the \emph{$q$-difference operator} $D_q\colon \mathcal{A} \to \mathcal{A}$ is defined as $D_q:=\Id - E_q.$ We have the following known result:
\begin{proposition}[\cite{Castillo13b}]
\label{prop:MRBO}~
\begin{enumerate}[(i)]
\item The pair $(\mathcal{A},P_q)$ is a RBA of weight $\lambda=-1$.
\item For any $f,g\in \mathcal{P}_{\geq 1}$ the operator $D_q$ satisfies the generalized Leibniz rule, i.e.,
\begin{align*}
D_q[fg] = D_q[f]g + fD_q[g] - D_q[f]D_q[g].
\end{align*}
\item The operators $P_q$ and $D_q$ are mutually inverse, i.e., $D_q\circ P_q = P_q \circ D_q = \Id$.
\end{enumerate}
\end{proposition}
\begin{remark}{\rm{
Recall that the Jackson integral
$$
\mathcal{J}[f](x) := \int_{0}^{x} f(y) d_qy =(1-q)\:\sum_{n \ge 0} f(q^nx) q^nx
$$
is the $q$-analogue of the classical indefinite Riemann integral $R$. For functions $\frac{f(x)}{x}$ -- where the Jackson integral is well defined -- it reduces to
$$
(1-q)\:\sum_{n \ge 0} f(q^nx) = \int_{0}^{x} \frac{f(y)}{y} d_qy = (1-q) P_q[f](x),
$$
which is the $q$-analogue of the integral operator $J$. Correspondingly, the $q$-analogue of the Euler derivation $\delta$ reduces to $(\Id - E_q)$.
}}
\end{remark}
\begin{lemma}[\cite{Castillo13b}]\label{lem:characterq}
Let $k_1,\ldots,k_n\in \bZ$. Then we have
\begin{align*}
\overline{\mathfrak{z}}_q(k_1,\ldots,k_n)= P_q^{k_1}[yP_q^{k_2}[y\cdots P_q^{k_n}[y]\cdots]](q).
\end{align*}
\end{lemma}
Surprisingly enough, the algebraic formalism for $q$MZVs is simpler than in the classical case. Let $X_{-1}:=\{p,d,y\}$. By $W_{-1}$ we denote the set of words on the alphabet $X_{-1}$, subject to the rule $pd=dp=\be$, where $\be$ denotes the empty word. Again, $\mathcal{A}_{-1}$ denotes the algebra spanned by the words in $W_{-1}$, i.e., $\mathcal{A}_{-1}:=\langle W_{-1} \rangle_\bQ$.
Then we define the product $\shm \colon \mathcal{A}_{-1}\otimes \mathcal{A}_{-1} \to \mathcal{A}_{-1}$ by $\be\shm w := w \shm \be := w$ for any $w\in W_{-1}$, and for any words $u,v\in W_{-1}$
\allowdisplaybreaks{
\begin{enumerate}[(i)]
\item $yu\shm v := u \shm yv := y(u\shm v)$,
\item $pu \shm pv := p(u\shm pv) + p(pu\shm v) -p(u \shm v)$,
\item $du \shm dv := u \shm dv +du \shm v -d(u \shm v)$,
\item $du\shm pv = pv\shm du := d(u\shm pv)+du \shm v - u\shm v$.
\end{enumerate}}
\begin{remark}{\rm{
We can deduce (iv) from (iii).}}
\end{remark}
\begin{lemma}\label{lem:qalgebra}
The pair $(\mathcal{A}_{-1},\shm )$ is a commutative, associative and unital algebra.
\end{lemma}
\begin{proof}
The proof is similar to that of \cite[Theorem~7]{Castillo13b}, and left to the reader.
\end{proof}
\noindent Next we introduce the set of words ending in the letter $y$ and containing the empty word
\begin{align*}
Y_{-1}:=W_{-1}y \cup\{\be\}\subseteq W_{-1},
\end{align*}
subject to the rule $pd=dp=\be$. Note that $(\langle Y_{-1} \rangle_\bQ,\shm )$ is a subalgebra of $(\mathcal{A}_{-1},\shm )$. Moreover we introduce the map $\overline{\mathfrak{z}}_q^\shuffle\colon \langle Y_{-1} \rangle_\bQ \to \bQ[[q]]$ by
\begin{align*}
p^{k_1}y\cdots p^{k_n}y \mapsto \overline{\mathfrak{z}}_q^\shuffle(p^{k_1}y\cdots p^{k_n}y) := \overline{\mathfrak{z}}_q(k_1,\ldots,k_n)
\end{align*}
for any integers $k_1,\ldots,k_n$.
\begin{lemma}[\cite{Castillo13b}]
The map $\overline{\mathfrak{z}}_q^\shuffle$ is an algebra morphism.
\end{lemma}
\subsection{General Word Algebraic Part}
\label{ssect:genalg}
In this section we explore the algebraic structure that is related to non-positive arguments for MZVs and $q$MZVs. For this reason we introduce a parameter $\lambda \in \bQ$. The case $\lambda=0$ corresponds to MZVs and the case $\lambda=-1$ to (modified) $q$MZVs.\\
Let $L:=\{d,y\}$ be an alphabet of two letters. The free monoid of $L$ with empty word $\be$ is denoted by $L^\ast$. We denote the free algebra of $L$ by $\bQ\langle L\rangle$ and define the subspace of words ending in $d$ by
\begin{align*}
\mathcal{T}_-:=\mathcal{T}\cap \bQ\langle L\rangle=\langle \left\{wd\colon w\in L^\ast \right\} \rangle_\bQ\subseteq \bQ\langle L\rangle,
\end{align*}
with $\mathcal{T}$ defined in Lemma \ref{lem:idealT}.
The set of \emph{admissible words} is defined as
\begin{align*}
Y:=L^\ast y\cup \{\be\},
\end{align*}
and the $\bQ$-vector space spanned by $Y$ is notated as $\mathcal{H}:=\langle Y \rangle_\bQ$. It is isomorphic to the quotient $\bQ\langle L \rangle / \mathcal{T}_-.$ The \emph{weight} $\operatorname{wt}(w)$ of a word $w \in Y$ is given by the number of letters of $w$, and we use the convention $\operatorname{wt}(\be):=0$.
Furthermore, the \emph{depth} $\operatorname{dpt}(w)$ of a word $w \in Y$ is given by the number of $y$ in $w$. The $\bQ$-vector space $\mathcal{H}$ is graded by depth, i.e.,
\begin{align*}
\mathcal{H} = \bigoplus_{n\geq 0} \mathcal{H}_{(n)}
\end{align*}
with $\mathcal{H}_{(n)}:=\langle w\in Y\colon \operatorname{dpt}(w)=n \rangle_\bQ$.
\subsubsection{The algebra $\mathcal{H}_\lambda,\,\lambda \neq 0$}
\label{ssect:wordbialgebra}
Let $\lambda \in \bQ\setminus \{0\}$. We define the product
\begin{align*}
\,\shuffle_\lambda\, \colon \bQ\langle L \rangle \otimes \bQ\langle L \rangle \to \bQ\langle L \rangle
\end{align*}
iteratively by
\allowdisplaybreaks{
\begin{enumerate}[(P1)]
\item $\be \,\shuffle_\lambda\, w := w \,\shuffle_\lambda\, \be := w$ for any $w\in L^\ast$;
\item $yu\,\shuffle_\lambda\, v := u \,\shuffle_\lambda\, yv := y(u\,\shuffle_\lambda\, v)$ for any $u,v\in L^\ast$;
\item $du\,\shuffle_\lambda\, dv:=\frac 1\lambda \big[d(u\,\shuffle_\lambda\, v)-du\,\shuffle_\lambda\, v-u \,\shuffle_\lambda\, dv\big]$ for any $u,v\in L^\ast$.
\end{enumerate}}
\noindent Furthermore, we define the unit map $\eta \colon \bQ \to \bQ\langle L \rangle$, $1 \mapsto \be$.
\begin{proposition}
\label{prop:algebraH}
For $\lambda \in \bQ$, the triple $(\bQ\langle L \rangle, \,\shuffle_\lambda\,,\eta)$ is a commutative, associative, and unital $\bQ$-algebra. The subspace $\mathcal{T}_-$ is a two-sided ideal of $\bQ\langle L\rangle$.
\end{proposition}
\begin{proof}
In the case $\lambda=-1$ the proof is a consequence of Lemma \ref{lem:qalgebra}. We give a proof for any $\lambda\neq 0$ for completeness, although it could be derived from the case $\lambda=-1$ by appropriate rescaling. Commutativity is clear from the definition. We only have to verify associativity if all words begin with a letter $d$. We apply induction on the sum of the lengths of the words. The base case is trivial. For the inductive step we observe for $a,b,c \in L^\ast$ that
\allowdisplaybreaks{
\begin{align*}
(da\,\shuffle_\lambda\, db)\,\shuffle_\lambda\, dc
=&~\frac{1}{\lambda} \left[d(a\,\shuffle_\lambda\, b)-da\,\shuffle_\lambda\, b - a\,\shuffle_\lambda\, db \right]\,\shuffle_\lambda\, dc\\
=&~\frac{1}{\lambda^2} \left[ d((a\,\shuffle_\lambda\, b)\,\shuffle_\lambda\, c) - d(a\,\shuffle_\lambda\, b)\,\shuffle_\lambda\, c - (a\,\shuffle_\lambda\, b)\,\shuffle_\lambda\, dc \right] \\
& - \frac{1}{\lambda} \left[ (da\,\shuffle_\lambda\, b)\,\shuffle_\lambda\, dc + (a\,\shuffle_\lambda\, db)\,\shuffle_\lambda\, dc\right] \\
=&~\frac{1}{\lambda^2} \left[d(a\,\shuffle_\lambda\, b\,\shuffle_\lambda\, c)-a\,\shuffle_\lambda\, b\,\shuffle_\lambda\, dc -da\,\shuffle_\lambda\, b\,\shuffle_\lambda\, c - a\,\shuffle_\lambda\, db \,\shuffle_\lambda\, c \right]\\
& -\frac{1}{\lambda} \left[ da\,\shuffle_\lambda\, db\,\shuffle_\lambda\, c + da\,\shuffle_\lambda\, b\,\shuffle_\lambda\, dc + a\,\shuffle_\lambda\, d b \,\shuffle_\lambda\, dc \right] \\
\end{align*}}
and
\allowdisplaybreaks{
\begin{align*}
da\,\shuffle_\lambda\, (db\,\shuffle_\lambda\, dc)
=&~\frac{1}{\lambda} da \,\shuffle_\lambda\, \left[ d(b\,\shuffle_\lambda\, c) -db\,\shuffle_\lambda\, c-b\,\shuffle_\lambda\, dc \right] \\
=&~\frac{1}{\lambda^2} \left[ d(a\,\shuffle_\lambda\,(b\,\shuffle_\lambda\, c)) - da\,\shuffle_\lambda\,(b\,\shuffle_\lambda\, c) - a\,\shuffle_\lambda\, d(b\,\shuffle_\lambda\, c) \right] \\
& - \frac{1}{\lambda}\left[ da\,\shuffle_\lambda\,(db\,\shuffle_\lambda\, c) + da\,\shuffle_\lambda\,(b\,\shuffle_\lambda\, dc) \right] \\
=&~\frac{1}{\lambda^2} \left[ d(a\,\shuffle_\lambda\, b\,\shuffle_\lambda\, c)-da\,\shuffle_\lambda\, b\,\shuffle_\lambda\, c - a\,\shuffle_\lambda\, db\,\shuffle_\lambda\, c - a\,\shuffle_\lambda\, b\,\shuffle_\lambda\, dc \right] \\
&-\frac{1}{\lambda}\left[ a\,\shuffle_\lambda\, db\,\shuffle_\lambda\, dc + da\,\shuffle_\lambda\, db\,\shuffle_\lambda\, c + da\,\shuffle_\lambda\, b\,\shuffle_\lambda\, dc \right],
\end{align*}}
which shows associativity. From definition we obtain that $\,\shuffle_\lambda\, (\mathcal{T}_- \otimes \bQ\langle L \rangle) = \,\shuffle_\lambda\, (\bQ\langle L \rangle\ \otimes \mathcal{T}_-) \subseteq \mathcal{T}_-$ and therefore $\mathcal{T}_-$ is a two-sided ideal of $\bQ\langle L \rangle$.
\end{proof}
\subsubsection{The coproduct $\overline\Delta_\lambda,\,\lambda\in\bQ$}
\noindent Now we define the coproduct
\begin{align*}
\overline{\Delta}_\lambda \colon \bQ\langle L \rangle \to \bQ\langle L \rangle \otimes \bQ\langle L \rangle
\end{align*}
by
\allowdisplaybreaks{
\begin{enumerate}[(C1)]
\item $\overline{\Delta}_\lambda(y):= \be\otimes y + y\otimes \be$,
\item $\overline{\Delta}_\lambda(d):= \be\otimes d + d\otimes \be+ \lambda d\otimes d$,
\end{enumerate}}
\noindent which extends uniquely to an algebra morphism (with respect to~concatenation) on the free algebra $\bQ\langle L \rangle$. The counit map $\varepsilon\colon \bQ\langle L\rangle \to \bQ$ is given by $\varepsilon(\be) = 1$ and $\varepsilon(w)=0$ for any word $w\in L^*\setminus \{\be\}$.
\begin{example}\label{ex:ocoproduct} We have
\allowdisplaybreaks{
\begin{align*}
\overline\Delta_\lambda(dy) = &~ \be \otimes dy + dy \otimes \be +d\otimes y +y\otimes d + \lambda dy\otimes d + \lambda d \otimes dy.
\end{align*}}
\end{example}
\begin{proposition}\label{prop:coalgebraH}
For $\lambda \in \bQ$ the triple $(\bQ\langle L \rangle,\overline{\Delta}_\lambda,\varepsilon)$ is a cocommmutative and counital coalgebra, and $\mathcal{T}_-$ is a coideal of $\bQ\langle L \rangle$.
\end{proposition}
\begin{proof}
Cocommutativity is clear by definitions (C1) and (C2). The counit axiom is not hard to verify. Finally we have to check coassociativity. We have
\allowdisplaybreaks{
\begin{align*}
\lefteqn{(\Id \otimes \overline{\Delta}_\lambda) \overline{\Delta}_\lambda(d)
= (\Id \otimes \overline{\Delta}_\lambda) (\be \otimes d + d\otimes \be +\lambda d\otimes d)} \\
= &~\be \otimes \be \otimes d + \be \otimes d \otimes \be + \lambda \be \otimes d \otimes d + d\otimes \be \otimes \be
+\lambda d \otimes \be \otimes d + \lambda d \otimes d \otimes \be + \lambda^2 d \otimes d \otimes d
\end{align*}}
and
\allowdisplaybreaks{
\begin{align*}
\lefteqn{ (\overline{\Delta}_\lambda \otimes \Id) \overline{\Delta}_\lambda(d)
= (\overline{\Delta}_\lambda \otimes \Id) (\be \otimes d + d\otimes \be +\lambda d\otimes d)} \\
= &~\be \otimes \be \otimes d + \be \otimes d \otimes \be + d \otimes \be \otimes \be +\lambda d\otimes d \otimes \be
+\lambda \be \otimes d \otimes d + \lambda d \otimes \be \otimes d + \lambda^2 d \otimes d \otimes d.
\end{align*}}
The case $ (\Id \otimes \overline{\Delta}_\lambda) \overline{\Delta}_\lambda(y) =(\overline{\Delta}_\lambda \otimes \Id) \overline{\Delta}_\lambda(y)$ is easy to see. We immediately obtain
\begin{align*}
\overline{\Delta}_\lambda(\mathcal{T}_-) \subseteq \mathcal{T}_- \otimes \bQ\langle L \rangle + \bQ\langle L \rangle \otimes \mathcal{T}_-,
\end{align*}
which concludes the proof.
\end{proof}
\subsubsection{Compatibility properties of the coproduct ($\lambda\neq 0$ case)}
\begin{lemma}
\label{lem:proco}
For words $u,v\in L^\ast$ we have
\begin{align}\label{eq:deltay}
\overline{\Delta}_\lambda(y)[\overline{\Delta}_\lambda(u)\,\shuffle_\lambda\,\overline{\Delta}_\lambda(v)]= \overline{\Delta}_\lambda(yu) \,\shuffle_\lambda\, \overline{\Delta}_\lambda(v) = \overline{\Delta}_\lambda(u) \,\shuffle_\lambda\, \overline{\Delta}_\lambda(yv)
\end{align}
and
\allowdisplaybreaks{
\begin{align}
\label{eq:deltad}
\begin{split}
\overline{\Delta}_\lambda(d)[\overline{\Delta}_\lambda(u)\,\shuffle_\lambda\, \overline{\Delta}_\lambda(v)] =~
& \overline{\Delta}_\lambda(du) \,\shuffle_\lambda\, \overline{\Delta}_\lambda(v) + \overline{\Delta}_\lambda(u)\,\shuffle_\lambda\,\overline{\Delta}_\lambda(dv)\\
&+\lambda [\overline{\Delta}_\lambda(du)\,\shuffle_\lambda\, \overline{\Delta}_\lambda(dv)].
\end{split}
\end{align}}
\end{lemma}
\begin{proof}
Using (P2) and Sweedler's notation, $ \overline{\Delta}_\lambda(u)=\sum_{(u)} u_1 \otimes u_2$, we obtain for the first equality of \eqref{eq:deltay}
\allowdisplaybreaks{
\begin{align*}
\overline{\Delta}_\lambda(y)\left[\overline{\Delta}_\lambda(u)\,\shuffle_\lambda\,\overline{\Delta}_\lambda(v)\right]
& = \overline{\Delta}_\lambda(y)\left[\sum_{(u),(v)}(u_1\,\shuffle_\lambda\, v_1)\otimes (u_2 \,\shuffle_\lambda\, v_2) \right] \\
& = \sum_{(u),(v)}\left[(u_1\,\shuffle_\lambda\, v_1)\otimes y(u_2 \,\shuffle_\lambda\, v_2)+y(u_1\,\shuffle_\lambda\, v_1)\otimes (u_2 \,\shuffle_\lambda\, v_2) \right] \\
& = \sum_{(u),(v)}\left[(u_1\,\shuffle_\lambda\, v_1)\otimes (yu_2 \,\shuffle_\lambda\, v_2)+(yu_1\,\shuffle_\lambda\, v_1)\otimes (u_2 \,\shuffle_\lambda\, v_2) \right] \\
& = \overline{\Delta}_\lambda(yu)\,\shuffle_\lambda\, \overline{\Delta}_\lambda(v).
\end{align*}}
The second equality follows completly analogously. For \eqref{eq:deltad} we observe
{\allowdisplaybreaks
\begin{align*}
& \overline{\Delta}_\lambda(d)[\overline{\Delta}_\lambda(u)\,\shuffle_\lambda\, \overline{\Delta}_\lambda(v)] \\
=~& \overline{\Delta}_\lambda(d)\left[ \sum_{(u),(v)}(u_1\,\shuffle_\lambda\, v_1)\otimes (u_2 \,\shuffle_\lambda\, v_2)\right]\\
=~& \sum_{(u),(v)}\left[ d(u_1\,\shuffle_\lambda\, v_1)\otimes (u_2 \,\shuffle_\lambda\, v_2) + (u_1\,\shuffle_\lambda\, v_1)\otimes d(u_2 \,\shuffle_\lambda\, v_2)+\lambda d(u_1\,\shuffle_\lambda\, v_1)\otimes d(u_2 \,\shuffle_\lambda\, v_2)\right]\\
=~& \sum_{(u),(v)}\left[ (\be\otimes d + d\otimes \be +\lambda d\otimes d)(u_1\otimes u_2)\,\shuffle_\lambda\,(v_1\otimes v_2) \right]\\
&+\sum_{(u),(v)}\left[ (u_1\otimes u_2)\,\shuffle_\lambda\,(\be\otimes d + d\otimes \be +\lambda d\otimes d)(v_1\otimes v_2) \right]\\
&+\lambda \sum_{(u),(v)}\left[ (\be\otimes d + d\otimes \be +\lambda d\otimes d)(u_1\otimes u_2)\,\shuffle_\lambda\,(\be\otimes d + d\otimes \be +\lambda d\otimes d)(v_1\otimes v_2) \right]\\
=~& \overline{\Delta}_\lambda(du) \,\shuffle_\lambda\, \overline{\Delta}_\lambda(v) + \overline{\Delta}_\lambda(u)\,\shuffle_\lambda\, \overline{\Delta}_\lambda(dv) +\lambda \left[\overline{\Delta}_\lambda(du)\,\shuffle_\lambda\, \overline{\Delta}_\lambda(dv)\right],
\end{align*}}
which yields the claim.
\end{proof}
\subsubsection{The Hopf algebra $\mathcal{H}_\lambda,\,\lambda\neq 0$}
\begin{theorem}\label{theo:HopfQuot}
The quintuple $\mathcal{H}_\lambda=(\mathcal{H},\,\shuffle_\lambda\,,\eta,\Delta_\lambda,\varepsilon)$ is a Hopf algebra with
\begin{align*}
\Delta_\lambda(w)&:= \overline{\Delta}_\lambda(w) \mod ( \mathcal{T}_-\otimes \bQ\langle L \rangle + \bQ\langle L \rangle \otimes \mathcal{T}_-)
\end{align*}
for any word $w\in Y$, where $\mathcal{H}=\bQ\langle L \rangle/\mathcal{T}_-$ is always identified with $\langle Y \rangle_{\bQ}$.
\end{theorem}
\begin{proof}
On the one hand $\mathcal{H}$ is the quotient by a two-sided ideal $\mathcal{T}_-$ and therefore $(\mathcal{H},\,\shuffle_\lambda\,,\eta)$ is an algebra. On the other hand $\mathcal{T}_-$ is a coideal. Hence, $(\mathcal{H},\Delta_\lambda,\varepsilon)$ is a coalgebra. Since $\mathcal{H}$ is connected it suffices to prove that $(\mathcal{H},\,\shuffle_\lambda\,,\eta,\Delta_\lambda,\varepsilon)$ is a bialgebra.
We show that
$$
\overline{\Delta}_\lambda(u'\,\shuffle_\lambda\, v') = \overline{\Delta}_\lambda(u')\,\shuffle_\lambda\, \overline{\Delta}_\lambda(v') \mod ( \mathcal{T}_-\otimes \bQ\langle L \rangle + \bQ\langle L \rangle \otimes \mathcal{T}_-)
$$
by induction on the sum of weights $\operatorname{wt}(u)+\operatorname{wt}(v)$, of the words $u,v\in Y$. The base cases are straightforward. \\
\textbullet~1st case: $u'=yu$ or $v'=yv$. We have with Lemma \ref{lem:proco}
\begin{align*}
\overline{\Delta}_\lambda(yu\,\shuffle_\lambda\, v') & = \overline{\Delta}_\lambda(y(u\,\shuffle_\lambda\, v')) = \overline{\Delta}_\lambda(y) \overline{\Delta}_\lambda(u\,\shuffle_\lambda\, v') \\
& = \overline{\Delta}_\lambda(y) (\overline{\Delta}_\lambda(u)\,\shuffle_\lambda\, \overline{\Delta}_\lambda(v)) \mod ( \mathcal{T}_-\otimes \bQ\langle L \rangle + \bQ\langle L \rangle \otimes \mathcal{T}_-) \\
& = \overline{\Delta}_\lambda(yu)\,\shuffle_\lambda\,\overline{\Delta}_\lambda(v) \mod ( \mathcal{T}_-\otimes \bQ\langle L \rangle + \bQ\langle L \rangle \otimes \mathcal{T}_-).
\end{align*}
\textbullet~2nd case: $u'=du$ and $v'=dv$. We have with Lemma \ref{lem:proco} and the induction hypothesis
\allowdisplaybreaks{
\begin{align*}
& \overline{\Delta}_\lambda(du\,\shuffle_\lambda\, dv) = \frac{1}{\lambda}\left[\overline{\Delta}_\lambda\left(d(u\,\shuffle_\lambda\, v)- du\,\shuffle_\lambda\, v - u \,\shuffle_\lambda\, dv \right)\right] \\
& = \frac{1}{\lambda}\left[\overline{\Delta}_\lambda(d)(\overline{\Delta}_\lambda(u)\,\shuffle_\lambda\, \overline{\Delta}_\lambda(v)) -\overline{\Delta}_\lambda(du)\,\shuffle_\lambda\, \overline{\Delta}_\lambda(v)\right. \\
& \left.-\overline{\Delta}_\lambda(u)\,\shuffle_\lambda\, \overline{\Delta}_\lambda(dv) \right] \mod ( \mathcal{T}_-\otimes \bQ\langle L \rangle + \bQ\langle L \rangle \otimes \mathcal{T}_-)\\
& = \frac{1}{\lambda}\left[\overline{\Delta}_\lambda(du)\,\shuffle_\lambda\, \overline{\Delta}_\lambda(v) + \overline{\Delta}_\lambda(u)\,\shuffle_\lambda\,\overline{\Delta}_\lambda(dv) + \lambda (\overline{\Delta}_\lambda(du)\,\shuffle_\lambda\, \overline{\Delta}_\lambda(dv))\right. \\
&\left. -\overline{\Delta}_\lambda(du)\,\shuffle_\lambda\, \overline{\Delta}_\lambda(v)-\overline{\Delta}_\lambda(u)\,\shuffle_\lambda\, \overline{\Delta}_\lambda(dv) \right] \mod ( \mathcal{T}_-\otimes \bQ\langle L \rangle + \bQ\langle L \rangle \otimes \mathcal{T}_-)\\
& = \overline{\Delta}_\lambda(du)\,\shuffle_\lambda\, \overline{\Delta}_\lambda(dv) \mod ( \mathcal{T}_-\otimes \bQ\langle L \rangle + \bQ\langle L \rangle \otimes \mathcal{T}_-),
\end{align*}}
\ignore{
$\diamond$ For $\lambda=0$ let $u'=d^nu$ and $v'=d^mv$ with $u,v\in Y$ beginning in $y$. Then we have with Lemma \ref{lem:productexp} and Lemma \ref{lem:proco}
\allowdisplaybreaks{
\begin{align*}
& \overline{\Delta}_\lambda(d^n u\sho d^m v) \\
=~& \frac{n}{n+m}\left\{ \sum_{l=n}^{n+m-1}(-1)^{n+l}\overline{\Delta}_0(d)\overline{\Delta}_0(d^lu\sho d^{n+m-l-1}v) + (-1)^m \overline{\Delta}_0(d^{n+m}u\sho v ) \right\}\\
&+ \frac{m}{n+m}\left\{ \sum_{l=m}^{n+m-1}(-1)^{m+l}\overline{\Delta}_0(d)\overline{\Delta}_0(d^{n+m-l-1}u\sho d^l v) + (-1)^n \overline{\Delta}_0(u\sho d^{n+m}v ) \right\} \\
=~& \frac{n}{n+m}\left\{ \sum_{l=n}^{n+m-1}(-1)^{n+l}\overline{\Delta}_0(d)(\overline{\Delta}_0(d^lu)\sho \overline{\Delta}_0(d^{n+m-l-1}v)) + (-1)^m \overline{\Delta}_0(d^{n+m}u\sho v ) \right\}\\
&+ \frac{m}{n+m}\left\{ \sum_{l=m}^{n+m-1}(-1)^{m+l}\overline{\Delta}_0(d)(\overline{\Delta}_0(d^{n+m-l-1}u)\sho \overline{\Delta}_0(d^l v)) + (-1)^n \overline{\Delta}_0(u\sho d^{n+m}v ) \right\} \\
=~& \frac{n}{n+m}\left\{ \sum_{l=n}^{n+m-1}(-1)^{n+l}(\overline{\Delta}_0(d^{l+1}u)\sho \overline{\Delta}_0(d^{n+m-l-1}v) + \overline{\Delta}_0(d^lu)\sho \overline{\Delta}_0(d^{n+m-l}v))\right.\\
& \left.+ (-1)^m \overline{\Delta}_0(d^{n+m}u\sho v ) \right\}\\
&+ \frac{m}{n+m}\left\{ \sum_{l=m}^{n+m-1}(-1)^{m+l}(\overline{\Delta}_0(d^{n+m-l}u)\sho \overline{\Delta}_0(d^l v)+\overline{\Delta}_0(d^{n+m-l-1}u)\sho \overline{\Delta}_0(d^{l+1} v))\right.\\
&+ \left.(-1)^n \overline{\Delta}_0(u\sho d^{n+m}v ) \right\} \\
=~& \overline{\Delta}_0(d^n u)\sho \overline{\Delta}_0(d^m v)\mod ( \mathcal{T}_-\otimes \bQ\langle L \rangle + \bQ\langle L \rangle \otimes \mathcal{T}_-),
\end{align*}}
}
which concludes the proof.
\end{proof}
\begin{example}\label{ex:coproduct} For $n\in \bN$ we have $\Delta_\lambda(y^n) = \sum_{l=0}^n \binom{n}{l}y^l\otimes y^{n-l};$, and
\allowdisplaybreaks{
\begin{align*}
\Delta_\lambda(d^ny) = &~ \be \otimes d^ny + d^ny \otimes \be; \\
\Delta_\lambda(yd^ny)= &~ \be \otimes yd^ny + y\otimes d^ny + d^ny\otimes y + yd^ny \otimes \be; \\
\Delta_\lambda(dyd^ny) = &~ \be \otimes dyd^ny + y\otimes d^{n+1}y + d^ny\otimes dy + dy\otimes d^ny + d^{n+1}y\otimes y+ dyd^ny \otimes \be \\
& + \lambda dy\otimes d^{n+1}y +\lambda d^{n+1}y \otimes dy.\\
\end{align*}}
\end{example}
\subsubsection{Compatibility between the product and the coproduct ($\lambda= 0$ case)}
Let us now focus on the case $\lambda=0$. Recall from Paragraph \ref{ssect:RBA} that $\mathcal{L}$ is the two-sided ideal of the (noncommutative and nonassociative) algebra $(\mathcal{A}_0,\sho)$ generated by the elements
\begin{align*}
j^k\big(d(u\sho v)-du\sho v-u\sho dv\big),\,k\in\bZ,\, u,v\in W_0y.
\end{align*}
Now let $\mathcal{L}_-$ be the two-sided ideal of the (noncommutative and nonassociative) subalgebra $(\bQ\langle L \rangle,\sho)$ generated by the elements
\begin{align*}
d^k\big(d(u\sho v)-du\sho v-u\sho dv\big),\,k\in\bN_0,\, u,v\in L^\ast.
\end{align*}
Further let {$\mathcal{L}_-^{(2)}:=\mathcal{L}_-\otimes\bQ\langle L\rangle+\bQ\langle L\rangle\otimes\mathcal{L}_-$.
\begin{proposition}\label{prop:comp-delta-zero}
For any $u',v'\in L^\ast$ we have:
\begin{equation}
\overline{\Delta}_0(u'\sho v')=\overline{\Delta}_0(u')\sho \overline{\Delta}_0(v')\mod\mathcal{L}_-^{(2)}.
\end{equation}
\end{proposition}
\begin{proof}
We use induction on $r:=|u'|+|v'|$. The cases $r=0$ and $r=1$ being immediate. The case $u'=yu$ or $v'=yv$ is easy and left to the reader. In the case $u'=du$ and $v'=dv$ we compute:
\begin{align*}
& \overline{\Delta}_0(du\sho dv) =\overline{\Delta}_0\big(d(u\sho dv)-u\sho d^2v)\big)\\
=&~\overline{\Delta}_0(d)\overline{\Delta}_0(u\sho dv)-\overline{\Delta}_0(u\sho d^2v)\\
=&~\overline{\Delta}_0(d)\big(\overline{\Delta}_0(u)\sho \overline{\Delta}_0(dv)\big)-\overline{\Delta}_0(u\sho d^2v)\mod \mathcal{L}_-^{(2)}\\
=&~\overline{\Delta}_0(d)\overline{\Delta}_0(u)\sho \overline{\Delta}_0(dv)+\overline{\Delta}_0(u)\sho \overline{\Delta}_0(d)\overline{\Delta}_0(d) \overline{\Delta}_0(v)-\overline{\Delta}_0(u\sho d^2v)\mod \mathcal{L}_-^{(2)},
\end{align*}
hence we get:
\begin{equation}
\overline{\Delta}_0(du\sho dv)-\overline{\Delta}_0(du)\sho\overline{\Delta}_0(dv)=-\big(\overline{\Delta}_0(u\sho d^2v)-\overline{\Delta}_0(u)\sho \overline{\Delta}_0(d^2v)\big)\mod \mathcal{L}_-^{(2)}.
\end{equation}
Iterating this process we return to the case when one of the arguments starts with a $y$.
\end{proof}
\subsubsection{The Hopf algebra $\mathcal{H}_0$}
Let $\mathcal{H}_0:=\bQ\langle L\rangle /(\mathcal{L}_-+\mathcal{T}_-)$.
\begin{proposition}\label{prop:coideal-L}
The ideal $\mathcal{L}_-$ is a coideal of $(\bQ\langle L\rangle,\overline{\Delta}_0)$, where $\overline{\Delta}_0$ is defined by (C1) and (C2) with $\lambda =0$.
\end{proposition}
\begin{proof}
Using Proposition \ref{prop:comp-delta-zero} we compute
\begin{align*}
&\overline\Delta_0 \big(d(u\sho v)-du\sho v-u\sho dv\big)\\
=&~\overline\Delta_0(d)\big(\overline\Delta_0(u)\sho \overline\Delta_0(v)\big)-\overline\Delta_0(du)\sho\overline\Delta_0(v)-\overline\Delta_0(u)\sho \overline\Delta_0(dv)\mod\mathcal{L}_-^{(2)}\\
=&~\overline\Delta_0(d)\big(\overline\Delta_0(u)\sho \overline\Delta_0(v)\big)-\overline\Delta_0(d)\overline\Delta_0(u)\sho\overline\Delta_0(v)-\overline\Delta_0(u)\sho \overline\Delta_0(d)\overline\Delta_0(v)\mod\mathcal{L}_-^{(2)}.
\end{align*}
Hence, $\overline\Delta_0\big(d(u\sho v)-du\sho v-u\sho dv\big)\in\mathcal{L}_-^{(2)}$.
\end{proof}
\begin{corollary}
$\mathcal{H}_0$ is a commutative Hopf algebra.
\end{corollary}
\begin{proof}
From Proposition \ref{prop:coalgebraH} and Proposition \ref{prop:coideal-L} we get that the ideal $\mathcal{T}_-+\mathcal{L}_-$ is also a coideal of $\bQ\langle L\rangle$. Hence the quotient $\mathcal{H}_0$ is a bialgebra. It is graded by the depth (and weight), hence connected, thus $\mathcal{H}_0$ is a Hopf algebra.
\end{proof}
We will denote by $\Delta_0$ the coproduct on $\mathcal{H}_0$. By a slight abuse of notation the coproduct on $\bQ\langle L\rangle/\mathcal{T}$, which was deduced from $\overline\Delta_0$, will also be denoted $\Delta_0$. Note that the ideal $\mathcal{T}_-+\mathcal{L}_-$ is also graded by depth. One then gets a grading on the quotient $\mathcal{H}_0$, which we still denote by $\operatorname{dpt}$.
\subsubsection{Shuffle factorization}
\noindent Let $\lambda\in\bQ$, including the case $\lambda=0$. From connectedness we can always write
\begin{align*}
\Delta_\lambda([w]) = \be \otimes [w] + [w] \otimes \be + \tilde{\Delta}_\lambda([w])\hspace{0.5cm}
\text{with}\hspace{0.5cm} \tilde{\Delta}_\lambda([w])\in \bigoplus_{p+q=n \atop p\neq 0,q\neq 0}\mathcal{H}_{(p)}\otimes \mathcal{H}_{(q)}.
\end{align*}
Therefore in the following we use two variants of Sweedler's notation
\begin{align*}
\Delta_\lambda([w])=\sum_{([w])}[w]_1\otimes [w]_2\hspace{0.5cm}
\text{and}\hspace{0.5cm}
\tilde{\Delta}_\lambda([w])=\sum_{([w])}[w]'\otimes [w]''.
\end{align*}
\noindent The following theorem, valid for any $\lambda$, including $\lambda=0$, provides a nice example of the theory outlined in \cite{Patras93}.
\begin{theorem}\label{theo:trivialrenorm}
Let $\lambda \in \bQ$. Then for all $w\in L^*$ we have
\begin{align*}
\,\shuffle_\lambda\, \circ \Delta_\lambda ([w]) = 2^{\operatorname{dpt}([w])}[w],
\end{align*}
where $[w]$ stands for the class of $w$ modulo $\mathcal{T}_-$ in the case $\lambda\neq 0$ (resp. modulo $\mathcal{T}_-+\mathcal{L}_-$ in the case $\lambda=0$).
\end{theorem}
\begin{proof}
We prove this by induction on the weight of $[w]$. For $\operatorname{wt}([w])=0$ we have $[w]=\be$ and obtain$\,\shuffle_\lambda\, \circ \Delta_\lambda(\be) = \,\shuffle_\lambda\,(\be \otimes \be) = \be.$ For the inductive step we consider two cases: \\
\noindent\textbullet~1st case: $w=yv$ with $v\in L^*$\\
We have $\operatorname{dpt}([w]) = \operatorname{dpt}([v])+1$ and obtain
\allowdisplaybreaks{
\begin{align*}
\,\shuffle_\lambda\, \circ \Delta_\lambda([yv])
& = \,\shuffle_\lambda\, \circ (\Delta_\lambda([y])\Delta_\lambda([v])) = \,\shuffle_\lambda\, \left(\sum_{([v])}[v]_1\otimes [yv]_2 +\sum_{([v])}[yv]_1\otimes [v]_2\right) \\
& = [y] \left(\sum_{([v])}\,\shuffle_\lambda\,([v]_1\otimes [v]_2) +\sum_{([v])}\,\shuffle_\lambda\,([v]_1\otimes [v]_2)\right) = 2[y] (\,\shuffle_\lambda\,\circ \Delta_\lambda([v])) \\
& = 2^{\operatorname{dpt}([v])+1}[yv] = 2^{\operatorname{dpt}([w])}[w].
\end{align*}}
\noindent\textbullet~2nd case: $w=dv$ with $v\in L^*$\\
Since $\operatorname{dpt}([w])=\operatorname{dpt}([v])$ we observe
\allowdisplaybreaks{
\begin{align*}
\,\shuffle_\lambda\, \circ \Delta_\lambda([dv])
& = \,\shuffle_\lambda\, \circ (\Delta_\lambda([d])\Delta_\lambda([v]))\\
&= \,\shuffle_\lambda\, \left(\sum_{([v])}[dv]_1\otimes [v]_2 + [v]_1\otimes [dv]_2 + \lambda ([dv]_1\otimes [dv]_2) \right)\\
& = [d] \left( \sum_{([v])} \,\shuffle_\lambda\, ( [v]_1 \otimes [v]_2) \right) = [d](\,\shuffle_\lambda\, \circ \Delta_\lambda([v]))\\
& = 2^{\operatorname{dpt}([v])}[dv]= 2^{\operatorname{dpt}([w])}[w].
\end{align*}}
\end{proof}
\begin{corollary}
\label{cor:decomposition}
Let $\lambda \in \bQ$.
Then for all $w \in L^*$, we have
\begin{align*}
(2^{\operatorname{dpt}([w])}-2) [w] &= \,\shuffle_\lambda\, \circ \tilde{\Delta}_\lambda([w])\\
&= \sum_{([w])} [w]' \,\shuffle_\lambda\, [w]'' = K \star K([w]).
\end{align*}
The linear map $K:=\Id - \eta \circ \varepsilon \in {\mathrm{End}}_\bQ(\mathcal{H}_\lambda)$ is a projector to the augmentation ideal $\mathcal{H}':=\bigoplus_{n > 0} \mathcal{H}_n$, and $f \star g := \,\shuffle_\lambda\, \circ (f \otimes g) \circ \Delta_\lambda$, $f,g\in {\mathrm{End}}_\bQ(\mathcal{H}_\lambda)$.
\end{corollary}
\subsubsection{A combinatorial description of the coproduct $\Delta_\lambda$}
\label{ssect:combCoprod}
In the following we give a combinatorial description of the coproduct $\Delta_\lambda$. However, note that we consider the construction only on an admissible representative $w \in Y$ of a given equivalence class in $\mathbb{Q}\langle L \rangle/ \mathcal{T}_-$ for $\lambda\neq 0$ (resp. in $\mathbb{Q}\langle L \rangle/ (\mathcal{L}_-+\mathcal{T}_-)$ for $\lambda= 0$).
Let $w:= d^{n_1-1}y \cdots d^{n_{k-1}-1}yd^{n_k-1}y \in Y$ be a word, with $n:=\sum_{i=1}^kn_i$, and define for $1 \leq m \leq k$, $N^m_w:=\{n_1,n_1+n_2,\ldots,n_1+ \cdots + n_{m}\}$. The coproduct $\Delta_\lambda(w)$ can be calculated as follows. Let $S:=\{s_1 < \cdots < s_l\} \subseteq [n]:=\{1,\ldots,n\}$ and $\bar S:=[n]\backslash S=\{\bar s_1 < \cdots < \bar s_{n-l}\} $. Define the words $w_S:= w_{s_1} \cdots w_{s_l}$ and $w_{\bar S}:= w_{\bar s_1} \cdots w_{\bar s_{n-l}}$. The set $S$ is called \emph{admissible} if both $w_S$ and $w_{\bar S}$ are in $Y$, i.e., if $s_l,\bar s_{n-l} \in N^k_w$ with $w\in Y$. The coproduct is then given by
\begin{align}
\label{lambdaCoproduct}
\Delta_\lambda(w) = \sum_{S \subseteq [n] \atop S\ {\rm{adm}}} w_S \otimes w_{\bar S}
+ \sum_{S \subseteq [n] \atop S\ {\rm{adm}}} \sum_{J=\{j_1 < \cdots < j_p\} \subset S \atop {J \neq \emptyset,\ J \cap N^k_w
= \emptyset \atop j_p < n_1+ \cdots + n_{k-1}}} \lambda^{|J|} w_S \otimes w_{[n] \backslash (S \backslash J)}.
\end{align}
For $\lambda=0$ this reduces to the coproduct corresponding to MPLs at non-positive integer arguments
$$
\Delta_0(w) := \sum_{S \subseteq [n] \atop S\ {\rm{adm}}} w_S \otimes w_{\bar S}.
$$
We introduce now a graphical notation, which should make the above more transparent. The set of vertices $V:=\{\d ,\y\}$ is used to define a polygon. The black vertex $\d \sim d$ and the white one $\y \sim y$. To each admissible word $w = d^{n_1-1}y \cdots d^{n_{k-1}-1}yd^{n_k-1}y \in Y$ corresponds an polygon with clockwise oriented edges, and the vertices colored clockwise according to the word $w$. For instance, the word $w=ddydy$ corresponds to
$$
\scalebox{0.7}{\ddydy}
$$
An admissible subset $S \subseteq [n]$ corresponds to a sub polygon. The admissible subsets for the above example are as follows:
$\{1,3\}, \{2,3\}, \{4,5\}$ correspond respectively to
$$
\scalebox{0.7}{\PdyA \quad
\PdyB \quad
\PdyC}
$$
$\{1,2,3\}, \{2,4,5\}$, $\{1,4,5\}$ correspond respectively to
$$
\scalebox{0.7}{\PddyA \quad
\PddyB \quad
\PddyC}
$$
$\{1,2,4,5\}$ and $\{3\}$ correspond respectively to
$$
\scalebox{0.7}{\Pdddy}
\quad
\scalebox{0.7}{\Py}
$$
In the last polygon we have marked the single vertex subpolygon by a circle around the white vertex.
The coproduct
\begin{align*}
\Delta_0(ddydy) &=\sum_{S \subseteq [n] \atop S\ {\rm{adm}}} w_S \otimes w_{\bar S} \\
&= ddydy \otimes \be + \be \otimes ddydy
+ 3 dy \otimes ddy + 3 ddy \otimes dy + dddy \otimes y + y \otimes dddy.
\end{align*}
The first two terms on the right-hand side correspond to $S=[n]$ and $S=\emptyset$, respectively.
The pictorial description of the weight-$\lambda$ coproduct (\ref{lambdaCoproduct}) is captured as follows. The second term on the right-hand side of the coproduct (\ref{lambdaCoproduct}) reflects the term $ \lambda d\otimes d$ in the coproduct
$$
\overline{\Delta}_\lambda(d):= \be\otimes d + d\otimes \be+ \lambda d\otimes d
$$
defined further above. It amounts to a certain doubling of those black vertices in an admissible word $w = d^{n_1-1}y \cdots d^{n_{k-1}-1}yd^{n_k-1}y \in Y$, which appear before the $y$ at position $n_1 + \cdots + n_{k-1}$. Algebraically this means that extracting a subpolygon corresponding to the admissible subset $S \subseteq [n]$ leads to a splitting of the word $w$ into $w_S$ and $w_{\bar S '}$, where the augmented complement sets ${\bar S '}$ contain $\bar S$, i.e., $\bar S \subset {\bar S '}$. This is due to not eliminating several black vertices, i.e., $d$'s that appear before the $y$ at position $n_1 + \cdots + n_{k-1}$. Pictorially we denoted this by doubling black vertices. Returning to the example above. We find for the admissible word $w=ddydy$
$$
\scalebox{0.7}{\PdyAx \quad
\PdyBx }
$$
which correspond to the admissible subsets and related augmented complement sets: $\{1,3\}$, ${\bar S '}=\{1,2,4,5\}$, and $\{2,3\}$, ${\bar S '}=\{1,2,4,5\}$, respectively. Next we consider
$$
\scalebox{0.7}{\PddyBx \quad
\PddyCx}
$$
where we have $\{2,4,5\}$, ${\bar S '}=\{1,2,3\}$, and $\{1,4,5\}$, ${\bar S '}=\{1,2,3\}$, respectively. For
$$
\scalebox{0.7}{\PddyAxa \quad
\PddyAxb \quad
\PddyAxx}
$$
we have $\{1,2,3\}$, ${\bar S '}=\{1,4,5\}$, and $\{1,2,3\}$, ${\bar S '}=\{2,4,5\}$, and $\{1,2,3\}$, ${\bar S '}=\{1,2,4,5\}$, respectively. Finally
$$
\scalebox{0.7}{\Pdddyxa \quad
\Pdddyxb \quad
\Pdddyxx}
$$
where we have $\{1,2,4,5\}$, ${\bar S '}=\{1,3\}$, and $\{1,2,4,5\}$, ${\bar S '}=\{2,3\}$, and $\{1,2,4,5\}$, ${\bar S '}=\{1,2,3\}$, respectively.
The coproduct
\allowdisplaybreaks{
\begin{align*}
\lefteqn{\Delta_\lambda(ddydy) =\sum_{S \subseteq [5] \atop S\ {\rm{adm}}} w_S \otimes w_{\bar S}
+ \sum_{S \subseteq [5] \atop S\ {\rm{adm}}} \sum_{J=\{j_1 < \cdots < j_p\} \subset S \atop {J \neq \emptyset,\ J \cap N^2_{ddydy} = \emptyset \atop j_p < 3}} \lambda^{|J|} w_S \otimes w_{[5] \backslash (S \backslash J)}}
\\
&= ddydy \otimes \be + \be \otimes ddydy
+ 3 dy \otimes ddy + 3 ddy \otimes dy + dddy \otimes y + y \otimes dddy\\
& \quad + 2 \lambda dy \otimes dddy
+ 4 \lambda ddy \otimes ddy
+ 2 \lambda dddy \otimes dy + \lambda^2 ddy \otimes dddy
+ \lambda^2 dddy \otimes ddy.
\end{align*}}
Again, the first two terms on the right-hand side correspond to $S=[n]$ and $S=\emptyset$, respectively.
\section{Renormalization of regularized MZVs}
\label{sect:Renorm}
Alain Connes and Dirk Kreimer discovered a Hopf algebraic approach to the BPHZ renormalization method in perturbative quantum field theory \cite{Connes00,Connes01}. See \cite{Manchon08} for a review. One of the fundamental results of these seminal works is the formulation of the process of perturbative renormalization in terms of a factorization theorem for regularized Hopf algebra characters. We briefly recall this theorem, and apply it in the context of the Hopf algebra introduced on $t$- and $q$-regularized MPLs, when considering them at non-positive arguments.
\subsection{Connes--Kreimer Renormalization in a Nutshell}
\label{ssect:CKrenorm}
We regard the commutative algebra $\mathcal{A}:=\bQ[z^{-1},z]]$ with the renormalization scheme
$\mathcal{A} = \mathcal{A}_{-}\oplus \mathcal{A}_{+},$ where $\mathcal{A}_{-}:=z^{-1}\bQ[z^{-1}]$ and $\mathcal{A}_{+}:=\bQ[[z]]$.
On $\mathcal{A}$ we define the corresponding projector $\pi: \mathcal{A} \to \mathcal{A}_{-}$ by
\begin{align*}
\pi\left(\sum_{n=-k}^\infty a_n z^n \right):= \sum_{n=-k}^{-1}a_n z^n
\end{align*}
with the common convention that the sum over the empty set is zero. Then $\pi$ and $\Id-\pi: \mathcal{A} \to \mathcal{A}_{+}$ are Rota--Baxter operators of weight $-1$. See e.\,g. \cite{Connes00,Ebrahimi02,Ebrahimi07}.
Let $(\mathcal{H},m_{\mathcal{H}},\Delta)$ be a bialgebra and $(\mathcal{A},m_{\mathcal{A}})$ an algebra. Then we define the convolution product $\star\colon \operatorname{Hom}(\mathcal{H},\mathcal{A}) \otimes \operatorname{Hom}(\mathcal{H},\mathcal{A}) \to \operatorname{Hom}(\mathcal{H},\mathcal{A})$ by the composition
\begin{align*}
\mathcal{H} \stackrel{\Delta}{\longrightarrow}\mathcal{H}\otimes \mathcal{H} \stackrel{\varphi\otimes \psi}{\longrightarrow} \mathcal{A}\otimes \mathcal{A} \stackrel{m_\mathcal{A}}{\longrightarrow}\mathcal{A}
\end{align*}
for $\varphi,\psi \in \operatorname{Hom}(\mathcal{H},\mathcal{A})$, or in Sweedler's notation
\begin{align*}
(\varphi \star \psi) (x):= m_{\mathcal{A}}(\varphi\otimes \psi)\Delta(x)=\sum_{(x)}\varphi(x_1)\psi(x_2).
\end{align*}
The Connes--Kreimer Hopf algebra approach unveiled a beautiful encoding of one of the key concepts of the renormalization process, i.e., Bogoliubov's counter term recursion, in terms of an algebraic Birkhoff decomposition:
\begin{theorem}[\cite{Connes00},\cite{Connes01},\cite{Manchon08},\cite{Ebrahimi07a}]
\label{theo:ConKre}
Let $(\mathcal{H},m_{\mathcal{H}},\Delta)$ be a connected filtered Hopf algebra and $\mathcal{A}$ a commutative unital algebra equipped with a renormalization scheme $\mathcal{A}=\mathcal{A}_{-}\oplus \mathcal{A}_{+}$ and corresponding idempotent Rota--Baxter operator $\pi$, where $\mathcal{A}_{-}=\pi(\mathcal{A})$ and $\mathcal{A}_{+}=(\Id-\pi)(\mathcal{A})$. Further let $\phi\colon \mathcal{H} \to \mathcal{A}$ be a Hopf algebra character. Then:
\allowdisplaybreaks{
\begin{enumerate}[a)]
\item The character $\phi$ admits a unique decomposition
\begin{align}
\label{Birkhoff}
\phi=\phi_{-}^{\star{(-1)}}\star \phi_{+}
\end{align}
called \emph{algebraic Birkhoff decomposition}, in which $\phi_{-}\colon \mathcal{H} \to \bQ \oplus \mathcal{A}_{-}$ and $\phi_{+}\colon \mathcal{H} \to \mathcal{A}_{+}$ are characters.
\item The maps $\phi_{-}$ and $\phi_{+}$ are recursively given fixed point equations
\begin{align}
\phi_{-} & = e -\pi\left(\phi_{-} \star (\phi - e) \right), \label{Bogo1}\\
\phi_{+} & = e + (\Id-\pi)\left(\phi_{-}\star (\phi - e\right)), \label{Bogo2}
\end{align}
where the unit for the convolution algebra product is $e=\eta_{\mathcal{A}} \circ \varepsilon$, and $\eta_{\mathcal{A}}: \bQ \to \mathcal{A}$ is the unit map of the algebra $\mathcal{A}$.
\end{enumerate}}
\end{theorem}
\subsection{Renormalization of MZVs}
\label{ssect:renormMZV}
An important remark is in order. To improve readability we skip brackets in the notation of classes of words, that is, in the following a word $w$ stands for the class $[w]$.
\smallskip
Let $k_1,\ldots,k_n\in \bN_0$. Then we define a map $\phi\colon \mathcal{H}_0 \to \bQ[z^{-1},z]]$ by
\begin{align}\label{eq:defphi}
d^{k_1}y\cdots d^{k_n}y \mapsto \phi(d^{k_1}y\cdots d^{k_n}y)(z) := \partial_z ^{k_1}[x \partial_z ^{k_2}[x \cdots \partial_z^{k_n}[ x ]] \cdots ](z),
\end{align}
where $x(z):=\frac{e^z}{1-e^z}$.
\begin{lemma}\label{lem:characterphi}
The map $\phi\colon (\mathcal{H}_0,\sho )\to (\bQ[z^{-1},z]],\cdot)$ is a Hopf algebra character. Furthermore, the following diagram commutes:
\begin{center}
\makebox[0pt]{
\begin{xy}
(0,20)*+{(\mathcal{H}_0,\sho )}="a"; (40,20)*+{(\bQ[[t]],\cdot)}="b";
(40,0)*+{(\bQ[z^{-1},z]],\cdot)}="d";
{\ar "a";"b"}?*!/_4mm/{\z_t^\shuffle};
{\ar "b";"d"}?*!/_8mm/{t\mapsto e^z};
{\ar"a";"d"};?*!/_2mm/{\phi};
\end{xy}}
\end{center}
\end{lemma}
\begin{proof}
From the chain and product rule of differentiation we easily obtain that $\phi = \z^{\,\shuffle\,}_{e^z}$. Furthermore, the evaluation maps $t\mapsto e^z$ and $\z_t^\shuffle$ are both algebra morphisms (see Lemma \ref{lem:characterJ}). Therefore $\phi$ is -- as a composition of multiplicative maps -- itself a character.
\end{proof}
Next we apply Theorem \ref{theo:ConKre} to the character $\phi$ (Lemma \ref{lem:characterphi}). Then we define \emph{renormalized MZVs} $\zeta_+$ -- using the character $\phi_+$ with image in $\bQ[[z]]$ in the Birkhoff decomposition (\ref{Birkhoff}) of $\phi$ -- by
\begin{align*}
\zeta_+(-k_1,\ldots,-k_n):=\lim_{z\to 0} \phi_+(d^{k_1}y \cdots d^{k_n}y)(z)
\end{align*}
for $k_1,\ldots,k_n\in \bN_0, n\in \bN$. The first values of $\zeta_+$ in depth two are given in Table \ref{table1} (for an explicit calculation example see Example \ref{ex:reno}).
Note that $\zeta_+$ respects the shuffle product $\sho $ as $\phi_+$ is a character with respect to the algebra $(\mathcal{H}_0,\sho )$. However, note that the quasi-shuffle relations are not verified because it would require $\zeta_+(0,0)=\frac{3}{8}$.
\begin{table}
\begin{center}
\renewcommand{\arraystretch}{2}
\begin{tabular}{r|c|c|c|c}
\backslashbox{$k_1$}{$k_2$}& $0$ & $-1$ & $-2$ & $-3$ \\
\hline
$0$ & $\frac{1}{4}$ & $\frac{1}{24}$ & $0$ & $-\frac{1}{240} $ \\
\hline
$-1$ & $\frac{1}{12}$ & $\frac{1}{144}$ & $-\frac{1}{240}$ & $-\frac{1}{1440} $ \\
\hline
$-2$ & $ \frac{1}{72} $ & $-\frac{1}{240}$ & $-\frac{1}{720} $ & $\frac{1}{504} $ \\
\hline
$-3$ & $-\frac{1}{120}$ & $-\frac{1}{360} $ & $\frac{1}{504}$ & $\frac{107}{100800} $ \\
\end{tabular}
\end{center}
\caption{The renormalized MZVs $\zeta_+(k_1,k_2)$.}\label{table1}
\end{table}
Next we show that the renormalized MZVs coincide with the meromorphic continuation of MZVs discussed in Section \ref{sect:mero}:
\begin{theorem}\label{theo:mero}
The renormalization procedure is compatible with the meromorphic continuation of MZVs, i.e., for $k\in \bN_0$
\allowdisplaybreaks{
\begin{align}
\label{eq:mero1}
\zeta_+(-k) = \z_1(-k)
\end{align}}
and for $a,b\geq 0$ with $a+b$ odd
\begin{align}\label{eq:mero2}
\zeta_+(-a,-b) = \z_2(-a,-b).
\end{align}
\end{theorem}
Note that for $\operatorname{dpt}(w)>2$ there is no information form the meromorphic continuation (see Remark \ref{rem:mero}).
\begin{proof}
We begin with the proof of \eqref{eq:mero1}. From Equation \eqref{eq:defphi} we obtain
\allowdisplaybreaks{
\begin{align*}
\phi(d^ky)(z)
& = \partial_z^k\left(\frac{e^z}{1-e^z}\right) \\
& = - \partial_z^k\left(\frac{1}{z}\frac{ze^z}{e^z-1} \right) \\
& = - \partial_z^k\left( \frac{B_0}{z} + \sum_{n \geq 0}\frac{B_{n+1}}{(n+1)!}z^n \right)\\
& = -\left(\frac{(-1)^{k}k! B_0}{z^{k+1}} + \sum_{n\geq 0} \frac{B_{n+k+1}}{n+k+1}\frac{1}{n!}z^{n}\right).
\end{align*}}
Since $d^ky\in Y$ is a primitive element for the coproduct $\Delta_0$ we obtain with Remark \ref{rem:mero} that
\begin{align*}
\phi_{+}(d^ky)(z) =(\Id-\pi)\phi(z) = -\frac{B_{k+1}}{k+1} + O(z) = \z_1(-k) + O(z).
\end{align*}
For Equation \eqref{eq:mero2} we calculate for $a+b$ odd with $a,b\geq 0$ (see Remark \ref{rem:mero})
{\allowdisplaybreaks
\begin{align*}
\phi(d^ayd^by)(z)
=&~\partial_z^a\left[x(z)\partial_z^b\left[x(z)\right] \right] \\
=&~\partial_z^a\left[\left(\frac{B_0}{z}+\sum_{m\geq 0}\frac{B_{m+1}}{(m+1)!}z^{m}\right) \left( \frac{(-1)^{b}b! B_0}{z^{b+1}} + \sum_{n\geq 0} \frac{B_{n+b+1}}{n+b+1}\frac{1}{n!}z^{n} \right) \right] \\
=&~\partial_z^a\left[ \text{pole part~}+\sum_{n\geq 0}\frac{B_0B_{n+b+2}}{n+b+2}\frac{z^n}{(n+1)!} + \sum_{m\geq 0}\frac{(-1)^bb!B_0B_{m+b+2}}{(m+b+2)!}z^m \right. \\
&~\left. +\sum_{l\geq 0} \sum_{\begin{smallmatrix} n+m=l \\ n,m\geq 0 \end{smallmatrix}} \frac{B_{m+1}B_{n+b+1}}{n+b+1}\frac{z^l}{n!(m+1)!} \right] \\
=&~\text{pole part~} +\frac{B_0B_{a+b+2}}{(a+b+2)(a+1)} + (-1)^b\frac{a!b!B_0B_{a+b+2}}{(a+b+2)!} \\
&~+\sum_{\begin{smallmatrix} n+m=a \\ n,m\geq 0 \end{smallmatrix}} \frac{B_{m+1}B_{n+b+1}}{(m+1)!(n+b+1)}\frac{a!}{n!} + O(z).
\end{align*}}
The second and third summand are zero since $a+b+2\geq 3$ is an odd number and therefore $B_{a+b+2}=0$. We have three possibilities for the last sum to be different from zero:
\allowdisplaybreaks{
\begin{itemize}
\item Case 1: $m+1$ and $n+b+1$ are even numbers. Then we have $m+n+b+2=a+b+2$ even, which contradicts that $a+b$ is odd.
\item Case 2: $m=0$. The last summand is equal to
\begin{align*}
\frac{B_1B_{a+b+1}}{a+b+1}.
\end{align*}
\item Case 3: $n+b=0$. Then $n=b=0$ and we have for the last summand
\begin{align*}
\frac{B_1B_{a+1}}{a+1}.
\end{align*}
\end{itemize}}
Therefore we obtain together with Remark \ref{rem:mero} that
\allowdisplaybreaks{
\begin{align*}
\phi(d^ayd^by)(z) & = \text{pole part~} +\frac{1}{2}\left(1+\delta_0(b) \right)\frac{B_{a+b+1}}{a+b+1}+ O(z)\\
& = \text{pole part~} +\z_2(-a,-b)+ O(z).
\end{align*}}
Let $s:=a+b$ be an odd number and $c,d\geq 0$ with $c+d=s$. Then we observe using the above calculation that
\allowdisplaybreaks{
\begin{align*}
\phi_{-}(p^cy)(z)\phi(p^dy)(z) & = \frac{(-1)^{c}c!B_0}{z^{c+1}} \left( \frac{(-1)^{d}d! B_0}{z^{d+1}} + \sum_{n\geq 0}\frac{B_{n+d+1}}{n+d+1}\frac{1}{n!}z^n \right)\\
&= \text{pole part~} +(-1)^c\frac{ B_0B_{c+d+2}}{(c+1)(c+d+2)}+ O(z)
\end{align*}}
Since $c+d+2 = a+b+2\geq 3$ is an odd number $B_{c+d+2}=0$ and the constant term is zero.
Hence, we obtain $\phi_{+}(d^ayd^by)(z) = \z_2(-a,-b) + O(z).$ Here we have used that $\tilde{\Delta}_0$ respects the weight graduation.
\end{proof}
In the light of Theorem \ref{theo:trivialrenorm} and Corollary \ref{cor:decomposition} we deduce a simple way to calculate the renormalized MPL, which is presented in
\begin{corollary}
\label{cor:primitive-renormalization}
For $w \in Y$, $\operatorname{dpt}(w)>1$
\begin{align}
\label{simple-renormal}
\phi_{+}(w)= \frac{1}{2^{\operatorname{dpt}(w)}-2} \sum_{(w)} \phi_{+}(w') \phi_{+}(w'').
\end{align}
\end{corollary}
Note that both $\operatorname{dpt}(w')$ and $\operatorname{dpt}(w'')$ are strictly smaller than $\operatorname{dpt}(w)$. On the right-hand side of (\ref{simple-renormal}) one can continue to apply Theorem \ref{theo:trivialrenorm} to the words $w',w''$, until $w$ has been fully decomposed into primitive elements. The renormalization of $w$ respectively the corresponding MZV reduce to the simple renormalization of single MPLs at non-positive arguments corresponding to primitive words in $\mathcal{H}$. For example
$$
\phi_{+}(dyd^ny) = \phi_{+}(y)\phi_{+}(d^{n+1}y) + \phi_{+}(dy)\phi_{+}(d^{n}y).
$$
\begin{proof}[Proof of Corollary \ref{cor:primitive-renormalization}]
Statement (\ref{simple-renormal}) follows directly from Theorem \ref{theo:trivialrenorm}, since $\phi_{+}$ is a character by construction. See (\ref{Bogo2}) in Theorem \ref{theo:ConKre}. Observe that (\ref{simple-renormal}) it is compatible with (\ref{Bogo2}), since ${(\Id-\pi)\phi_{+}=\phi_{+}}$.
\end{proof}
\begin{example}\label{ex:reno}{\rm{
Let us calculate the renormalized MZVs $\zeta_+(0,-2)$ and $\zeta_+(-1,-1)$.
From Example \ref{ex:coproduct} we find
\allowdisplaybreaks{
\begin{align*}
\tilde{\Delta}_0(yd^2y) = y\otimes d^2y + d^2y \otimes y \hspace{0.5cm}\text{and}\hspace{0.5cm} \tilde{\Delta}_0(dydy) = y\otimes d^2y + d^2y\otimes y + 2dy \otimes dy.
\end{align*}}
Therefore we obtain form the iterative formulas \eqref{Bogo1} and \eqref{Bogo2} of Theorem \ref{theo:ConKre}
\allowdisplaybreaks{
\begin{align*}
\phi_+(yd^2y) &= (\Id-\pi)\left[\phi(yd^2y) - \left(\phi_{-}(y)\phi(d^2y) + \phi_{-}(d^2y) \phi(y) \right) \right], \\
\phi_+(dydy) &=(\Id-\pi)\left[ \phi(dydy) - \left(\phi_{-}(y)\phi(d^2y) + \phi_{-}(d^2y) \phi(y) + 2 \phi_{-}(dy)\phi(dy) \right)\right].
\end{align*}}
Using
\allowdisplaybreaks{
\begin{align*}
\phi(y)(z) & =-{z}^{-1}-{\frac {1}{2}}-{\frac {1}{12}}z+{\frac {1}{720}}{z}^{3}+ O\!\left( {z}^{4} \right), \\
\phi(dy)(z)& ={z}^{-2}-{\frac {1}{12}}+{\frac {1}{240}}{z}^{2}+O\!\left( {z}^{4} \right), \\
\phi(d^2y)(z)& =-2\,{z}^{-3}+{\frac {1}{120}}z-{\frac {1}{1512}}{z}^{3}+ O\!\left( {z}^{4} \right) \\
\end{align*}}
and
\allowdisplaybreaks{
\begin{align*}
\phi(yd^2y)(z)&= 2\,{z}^{-4}+{z}^{-3}+\frac{1}{6}\,{z}^{-2}-{\frac {1}{90}}
-{\frac {1}{240}}\,z+{\frac {1}{30240}}\,{z}^{2}+{\frac {1}{3024}}\,{z
}^{3}+O\left( {z}^{4} \right), \\
\phi(dydy)(z)&=3\,{z}^{-4}+{z}^{-3}+{\frac {1}{240}}-{\frac {1}{240}}z-{\frac {1}{1008}}{z}^{2}+{\frac {1}{3024}}{z}^{3}+O\!\left( {z}^{4} \right),
\end{align*}}
we observe that
\begin{align*}
(\Id-\pi)\left[\phi_{-}(y)\phi(d^2y) + \phi_{-}(d^2y) \phi(y)\right](z) &= -\frac{1}{90} + O(z)
\end{align*}
and
\begin{align*}
(\Id-\pi)\left[\phi_{-}(y)\phi(d^2y) + \phi_{-}(d^2y) \phi(y) + 2 \phi_{-}(dy)\phi(dy) \right](z) &= -\frac{1}{360} + O(z).
\end{align*}
Hence,
\begin{align*}
\phi_+(yd^2y)(z) = O(z) \hspace{1cm} \text{and}\hspace{1cm} \phi_+(dydy)(z) = \frac{1}{240}+\frac{1}{360}+O(z)=\frac{1}{144} + O(z),
\end{align*}
which results in $\zeta_+(0,-2)=0$ and $\zeta_+(-1,-1)=\frac{1}{144}$.
Alternatively, we can use the shuffle product to calculate $\zeta_+(0,-2)$ and $\zeta_+(-1,-1)$. Note that since $y\sho d^2y = yd^2y$ we have $\zeta_+(0,-2)=\zeta_+(0)\zeta_+(-2)=0$.
Because of $dy\sho dy = dydy - yd^2y$ we see that $\zeta_+(-1,-1)=\zeta_+(0,-2) + \zeta_+(-1)^2= \frac{1}{144}$.
A third way to calculate, say, $\zeta_+(-1,-1)$, is based on Corollary \ref{cor:decomposition}, and described in Corollary \ref{cor:primitive-renormalization}:
\begin{align*}
\zeta_{+}(-1,-1) = \frac{1}{2}\left( \zeta_+(0)\zeta_+(-2)+ \zeta_+(-2)\zeta_+(0) + 2\zeta_+(-1)\zeta_+(-1) \right) =\frac{1}{144}.
\end{align*}}}
\end{example}
\subsection{Renormalization of $q$MZVs}
\label{ssect:renormqMZV}
The Hopf algebra $(\mathcal{H},\shm ,\Delta_{-1})$ is related to the modified $q$-analogue $\overline{\mathfrak{z}}_q$ whereas the relation between MZVs and $q$MZVs (see Equation \eqref{eq:lim}) relies on a limit process involving the non-modified $q$MZVs $\mathfrak{z}_q$. Therefore the renormalization related to the $q$MZV deformation is more involved than the renormalization described in the previous section. First of all we apply Theorem \ref{theo:ConKre} in the framework of modified $q$MZVs. We define the map $\psi\colon (\mathcal{H},\shm )\to \bQ[z^{-1},z]]$ by
\begin{align*}
d^{k_1}y\cdots d^{k_n}y \mapsto \psi(d^{k_1}y\cdots d^{k_n}y)(z):=\sum_{m_1,\ldots,m_n\geq 0}\frac{B_{m_1}}{m_1!}\cdots \frac{B_{m_n}}{m_n!} \cdot C^{k_1,\ldots,k_n}_{m_1,\ldots,m_n} z^{m_1+\cdots+m_n-n}
\end{align*}
for $k_1,\ldots,k_n\in \bN_0$, where
\begin{align*}
C^{k_1,\ldots,k_n}_{m_1,\ldots,m_n}:= \sum_{l_i=0 \atop i=1,\ldots,n}^{k_i}\left(\prod_{i=1}^n\binom{k_i}{l_i}(-1)^{l_i+1}(l_1+\cdots+l_i+1)^{m_i-1} \right).
\end{align*}
\begin{lemma}\label{lem:qcharacter}
The map $\psi\colon (\mathcal{H},\shm )\to (\bQ[z^{-1},z]],\cdot)$ is a character. Furthermore, the following diagram commutes:
\begin{center}
\makebox[0pt]{
\begin{xy}
(0,20)*+{(\mathcal{H},\shm )}="a"; (40,20)*+{(\bQ[[q]],\cdot)}="b";
(40,0)*+{(\bQ[z^{-1},z]],\cdot)}="d";
{\ar "a";"b"}?*!/_4mm/{\overline{\mathfrak{z}}_q^\shuffle};
{\ar "b";"d"}?*!/_8mm/{q\mapsto e^z};
{\ar"a";"d"};?*!/_2mm/{\psi};
\end{xy}}
\end{center}
\end{lemma}
\begin{proof}
Let $k_1,\ldots,k_n\in \bN_0$. First we observe that
\allowdisplaybreaks{
\begin{align*}
~ & \overline{\mathfrak{z}}_q^{\,\shuffle\,}(d^{k_1}y\cdots d^{k_n}y) \\
= & \sum_{m_1>\cdots > m_n >0} q^{m_1}(1-q^{m_1})^{k_1}(1-q^{m_2})^{k_2}\cdots(1-q^{m_n})^{k_n} \\
= & \sum_{m_1,\ldots, m_n >0} q^{m_1+\cdots+m_n}(1-q^{m_1+\cdots+ m_n})^{k_1}(1-q^{m_2+\cdots+m_n})^{k_2}\cdots(1-q^{m_n})^{k_n} \\
= & \sum_{l_1=0}^{k_1} \cdots \sum_{l_n=0}^{k_n}(-1)^{l_1+\cdots+l_n} \binom{k_1}{l_1} \cdots \binom{k_n}{l_n} \sum_{m_1,\ldots,m_n>0} q^{m_1(l_1+1)}q^{m_2(l_1+l_2+1)} \cdots q^{m_n(l_1+\cdots+l_n+1)} \\
= & \sum_{l_1=0}^{k_1} \cdots \sum_{l_n=0}^{k_n}(-1)^{l_1+\cdots+l_n+n} \binom{k_1}{l_1} \cdots \binom{k_n}{l_n} \frac{q^{l_1+1}}{q^{l_1+1}-1} \cdots \frac{q^{l_1+\cdots+l_n+1}}{q^{l_1+\cdots+l_n+1}-1}.
\end{align*}}
This leads to
\begin{align*}
\overline{\mathfrak{z}}_q^{\,\shuffle\,}(d^{k_1}y\cdots d^{k_n}y) \stackrel{q\mapsto e^z}{\longmapsto} & \sum_{l_i=0 \atop i=1,\ldots,n}^{k_i} \left(\prod_{j=1}^{n}(-1)^{l_j+1} \binom{k_j}{l_j} \frac{e^{z(l_1+\cdots+l_j+1)}}{e^{z(l_1+\cdots+l_j+1)}-1} \right) \\
= & \sum_{l_i=0 \atop i=1,\ldots,n}^{k_i} \left(\prod_{j=1}^{n}(-1)^{l_j+1} \binom{k_j}{l_j} \sum_{m_j\geq 0} \frac{B_{m_j}}{m_j!} (z(l_1+\cdots+l_j+1))^{m_j-1} \right)\\
= & \sum_{m_1,\ldots, m_n\geq 0} \frac{B_{m_1}}{m_1!} \cdots \frac{B_{m_n}}{m_n!}\cdot C^{k_1,\ldots,k_n}_{m_1,\ldots,m_n} z^{m_1+\cdots +m_n-n}.
\end{align*}
The map $\psi$ is a character since it is a composition of algebra morphisms.
\end{proof}
Next we reverse the modification process applied in Equation \eqref{eq:modify}. Therefore we apply Theorem \ref{theo:ConKre} to the character $\psi$ and define the \emph{renormalized $q$MZVs} $\mathfrak{z}_{+}$ by
\begin{align}\label{eq:limmod}
\mathfrak{z}_{+}(-k_1,\ldots,-k_n):=\lim_{z \to 0} \frac{(-1)^{k_1+\cdots+k_n}}{z^{k_1+\cdots+k_n}}\psi_{+}(d^{k_1}y\cdots d^{k_n}y)(z)
\end{align}
for $k_1,\ldots,k_n\in \bN_0$.
\begin{theorem}
\label{theo:renoqMZV}
Let $k_1,\ldots,k_n\in \bN_0$. Then $\mathfrak{z}_{+}(-k_1,\ldots,-k_n)$ is well defined, and we have
\begin{align*}
\mathfrak{z}_{+}(-k_1,\ldots,-k_n) = \zeta_+(-k_1,\ldots,-k_n).
\end{align*}
Especially, the renormalized $q$MZVs $\mathfrak{z}_{+}$ respect the shuffle product $\sho $.
\end{theorem}
For the proof of this theorem we need an auxiliary result:
\begin{lemma}\label{lem:power}
We have $\psi_{+}(d^ky)(z) = (-1)^{k}\zeta_+(-k)z^k + O(z^{k+1})$ for all $k\in \bN_0$.
\end{lemma}
\begin{proof}
Since $d^ky$ is primitive with respect to the coproduct $\Delta_{-1}$, we obtain from Lemma \ref{lem:qcharacter} that
\begin{align*}
\psi_{+}(d^ky)(z) =(\Id-\pi)\psi(d^ky)(z)
= (\Id-\pi)\left(\sum_{m\geq 0} \frac{B_m}{m!}\cdot C_m^k z^{m-1} \right)
= \sum_{m>0} \frac{B_m}{m!}\cdot C_m^k z^{m-1}
\end{align*}
with $C_m^k = \sum_{l=0}^k\binom{k}{l}(-1)^{l+1}(l+1)^{m-1}$. We have
\begin{align*}
C_m^k & = \sum_{l=0}^k\binom{k}{l}(-1)^{l+1}(l+1)^{m-1}
= \frac{1}{k+1} \sum_{l=0}^{k+1}\binom{k+1}{l}(-1)^l l^m \\
& = \left.\frac{1}{k+1} \delta_z^m\left( \sum_{l=0}^{k+1}\binom{k+1}{l} (-z)^l \right)\right|_{z=1}
= \left.\frac{1}{k+1} \delta_z^m (1-z)^{k+1} \right|_{z=1}
\end{align*}
This shows that $C_m^k=0$ for $m=1,\ldots,k$. Furthermore, we observe that $C_{k+1}^k = (-1)^{k+1}\frac{(k+1)!}{k+1}$, which completes the proof.
\end{proof}
\begin{proof}[Proof of Theorem \ref{theo:renoqMZV}]
Let $w:=d^{k_1}y\cdots d^{k_n}y$ with $k_1,\ldots,k_n\in \bN_0$. In order to prove that $\mathfrak{z}_{+}$ is well defined we show $\psi_+(w)\in O(z^{\operatorname{wt}(w)-n})$. We split up the coproduct $\Delta_{-1}(w)$ into two parts
\begin{align}
\label{eq:decom}
\Delta_{-1}(w)=\Delta_0(w) + \left( \Delta_{-1}(w)-\Delta_0(w) \right).
\end{align}
From Corollary \ref{cor:primitive-renormalization} we deduce that $\Delta_{-1}(w)$ induces a $\bQ$-linear combination of products with $n$ factors of $\psi_+$ in primitive elements of $\mathcal{H}$. The part of $\psi_+$ corresponding to $\Delta_0(w)$ in the coproduct factorization is homogeneous in weight $\operatorname{wt}(w)$ and the one related to $\Delta_{-1}(w)-\Delta_0(w)$ has weight greater than $\operatorname{wt}(w)$. Therefore Lemma \ref{lem:power} implies that $\psi_+(w)\in O(z^{\operatorname{wt}(w)-n})$. Hence, the limit in \eqref{eq:limmod} exists. On the one hand we can apply Corollary \ref{cor:primitive-renormalization} to $\phi_+(w)$, defined in the previous section, which corresponds to the factorization induced by $\Delta_0(w)$. On the other hand we can do the same with $\psi_+(w)$. However, this factorization is related to $\Delta_{-1}(w)$. After dividing by $z^{\operatorname{wt}(w)-n}$ and taking the limit $z\to 0$ only the first part in the decomposition \eqref{eq:decom} of $\Delta_{-1}(w)$ makes a contribution in the factorization of $\psi_+(w)$. Using the fact that the leading factor of $\psi_+(d^ky)(z)$ equals $(-1)^k\zeta_+(-k)z^k$ concludes the proof.
\end{proof}
\begin{example}\label{ex:renoq}
Let us calculate the renormalized $q$MZV $\mathfrak{z}_{+}(-1,-1)$. From Example \ref{ex:coproduct} we obtain
\begin{align*}
\tilde{\Delta}_0(dydy) = y\otimes d^2y + d^2y\otimes y +2 dy\otimes dy - dy\otimes d^2y - d^2y\otimes dy,
\end{align*}
which gives
\begin{align*}
\psi_+(dydy) =(\Id-\pi)& \left[ \psi(dydy) - \left(\psi_{-}(y)\psi(d^2y) + \psi_{-}(d^2y) \psi(y) + 2 \psi_{-}(dy)\psi(dy) \right.\right.\\
& \left.\left.- \psi_{-}(dy)\psi(d^2y) - \psi_{-}(d^2y)\psi(dy)\right)\right]
\end{align*}
Using
\allowdisplaybreaks{
\begin{align*}
\psi(y)(z) &= -{z}^{-1}-\frac{1}{2}-\frac{1}{12}\,z+{\frac {1}{720}}\,{z}^{3}-{\frac {1}{30240}}\,{z}^{5}+O\!\left( {z}^{7} \right),\\
\psi(dy)(z) &= -\frac{1}{2}\,{z}^{-1}+\frac{1}{12}\,z-{\frac {7}{720}}\,{z}^{3}+{\frac {31}{30240}}\,{z}^{5}+O\!\left( {z}^{7} \right),\\
\psi(d^2y)(z) & = -\frac{1}{3} \,{z}^{-1}+{\frac {1}{60}}\,{z}^{3}-{\frac {1}{168}}\,{z}^{5}+O\!\left( {z}^{7} \right)
\end{align*}}
and
\allowdisplaybreaks{
\begin{align*}
\psi(dydy)(z)={\frac {5}{12}}\,{z}^{-2}+\frac{1}{6}\,{z}^{-1}-\frac{1}{36}+{\frac {1}{216}}\,{z}^{2}-{\frac {1}{120}}\,{z}^{3}-{\frac {19}{9072}}\,{z}^{4}+O\!\left( {z}^{5} \right),
\end{align*}}
we observe that
\begin{align*}
(\Id-\pi) \left[\psi_{-}(y)\psi(d^2y) + \psi_{-}(d^2y) \psi(y) + 2 \psi_{-}(dy)\psi(dy) \right](z) = -\frac{1}{18}-\frac{1}{135}z^2+O(z^3)
\end{align*}
and
\begin{align*}
(\Id-\pi) \left[-\psi_{-}(dy)\psi(d^2y) - \psi_{-}(d^2y)\psi(dy)\right](z) = \frac{1}{36} +\frac{11}{2160}z^2 + O(z^3).
\end{align*}
Therefore we have $\psi_{+}(dydy)(z) = \frac{1}{144} z^2 + O(z^3)$ and consequently $\mathfrak{z}_{+}(-1,-1)=\frac{1}{144}$, which coincides with $\zeta_+(-1,-1)$.
\end{example}
\begin{remark}
{\rm{The crucial point in the renormalization of $q$MZVs is the fact that $\psi_+(w)(z)\in O(z^{\operatorname{wt}(w)-n})$. We used a corollary of Theorem \ref{theo:trivialrenorm} to prove this. However, in the light of Theorem \ref{theo:ConKre} the previous example shows that this is obtained by non-trivial cancellations.}}
\end{remark}
\bibliographystyle{alpha}
|
\section*{Acknowledgment}
\ifCLASSOPTIONcaptionsoff
\newpage
\fi
\bibliographystyle{unsrt}
\section{Introduction}
\newcommand{\alert}[1]{#1}
\newcommand{\SPM}{\text{SPM}}
\newcommand{\XPM}{\text{XPM}}
\newcommand{\FWM}{\text{FWM}}
\newcommand{\ASE}{\text{ASE}}
\newcommand{\text{eff}}{\text{eff}}
\newcommand{\scripttwo}[2]{ \scriptsize{\begin{array}{c} #1 \\ #2 \end{array}} }
\newcommand{\NLI}{\text{NLI}}
\newcommand{\tnr}[1]{{\textnormal{#1}}}
\newcommand{i.e.,~}{i.e.,~}
\newcommand{\mathbb{E}}{\mathbb{E}}
\newcommand{P_\text{ASE}}{P_\text{ASE}}
\newcommand{Z_{k}}\newcommand{\zn}{z_{k}}{Z_{k}}\newcommand{\zn}{z_{k}}
\newcommand{\tilde{Z}_{k}}{\tilde{Z}_{k}}
\newcommand{\Ex}{\mathbb{E}}
\IEEEPARstart{T}{he} majority of traffic in core networks is carried by optical fiber.
Understanding the ultimate limits of communication over optical fiber is thus of great importance
and would help to provide guidelines for designing networks to meet the growing traffic demand.
An appealing property of optical fiber is that it has low attenuation over a large range of frequencies
which allows the transmission of broadband signals over long distances.
Optical amplifiers compensate the power loss
but unfortunately they also add noise.
Moreover, a signal propagating in optical fiber experiences distortions due to chromatic dispersion and Kerr non-linearity.
The optical fiber channel thus suffers from three main impairments of different nature: noise, dispersion, and Kerr non-linearity.
The interaction between these three phenomena makes the problem of estimating the capacity challenging \cite{JLT2010}.
\subsection{Capacity Estimates}
There are several attempts to obtain capacity estimates
for optical fiber channels.
Narimanov and Mitra \cite{Narimanov2002}
studied a single-channel transmission over a multi-span dispersive fiber link.
A perturbation technique is used to approximate the solution to the nonlinear Schr\"{o}dinger (NLS) equation assuming that the non-linear term is small.
The capacity expression is found for small nonlinearity.
Xiang and Zhang \cite{Xiang2011Perturb} extended some of the results of \cite{Narimanov2002}.
Mecozzi \cite{Mecozzi94} models the propagation of a single signal in a dispersionless fiber link,
in which the fiber loss is compensated by equally spaced (lumped) amplifiers
with a stochastic differential equation.
Mecozzi derived an expression for the conditional distribution of the output field given the input field
by computing arbitrary (conditional) moments.
Turitsyn et al \cite{TuritsynPRL2003} also
studied single-channel transmission
over zero-dispersion fiber links.
They obtained the conditional distribution
using techniques from quantum mechanics.
For Gaussian inputs and direct-detection,
a lower bound on capacity was derived
that grows with SNR (prelog = 1/2).
In \cite{Yousefi2010SumProd,Yousefi2010FokkerPlank,PerSampleIT2011},
Yousefi and Kschischang derived the conditional probability using two different approaches:
a sum-product approach and a Fokker-Planck differential equation approach.
Wei and Plant \cite{WeiPlantArxiv2006} made some useful comments on the results of \cite{TuritsynPRL2003},\cite{Mitra2001} and \cite{Tang2001}.
Ivakovic et al \cite{IvakovicJLT2007}
studied a single-channel long-haul transmission system
that uses on-off keying (OOK) and return-to-zero (RZ) pulses.
They estimate numerically the achievable information rate for independent uniformly distributed inputs
when the effects of the Kerr non-linearity, chromatic dispersion and amplified spontaneous emission
are taken into account.
The information rate is upper bounded by 1 bit/symbol because OOK is used
and therefore the computed rate is not a good estimate for the capacity.
Mitra and Stark \cite{Mitra2001}
studied a wavelength division multiplexing (WDM) system in which cross-phase modulation (XPM) is the only non-linear effect considered (ignoring Four-wave mixing (FWM) and assuming that self-phase modulation (SPM) can be fully corrected).
A key simplification in \cite{Mitra2001} is approximating the sum of intensities of the interfering channels in the XPM term of the propagation equation by a Gaussian random process.
A lower bound on capacity (per WDM channel) was derived for Gaussian inputs
using only the input-output covariance matrix.
The conclusion of \cite{Mitra2001} was that capacity has a peak and does not increase indefinitely with the input power.
Tang \cite{Tang2001} studied WDM transmission over a single-span dispersion-free fiber link.
In this case, the propagation equation can be solved analytically in closed-form.
A lower bound on capacity was obtained for Gaussian inputs by computing the power spectral density of the input (the sum all WDM channels), the power spectral density of the output (the overall WDM signal after propagation) and the cross-spectral density of the input and output.
Tang extended the results of \cite{Tang2001} to a multi-span dispersion-free fiber link in \cite{TangMultispan2001}
and then to a multi-span dispersive fiber link in \cite{Tang2002}.
In \cite{Tang2002}, a truncated Volterra series \cite{Ped1997Volterra}
is used to approximate the solution to the NLS equation
assuming that the effect of non-linearity is small.
The lower bounds in \cite{Tang2001}, \cite{TangMultispan2001} and \cite{Tang2002} have a peak value as a function of input power.
We \alert{remark} that it is implicit in the analysis of \cite{Tang2001,TangMultispan2001,Tang2002} that
the receiver of each user has access to all the WDM channels and not just its own channel.
Ho and Khan \cite{HoKahnOFC2002}
studied WDM transmission over a multi-span dispersive fiber link.
It is argued that under constant-envelope (also called constant-intensity) modulation
with uniform phase\footnote{We refer to constant-envelope modulation with uniform phase as \emph{ring} modulation.},
SPM and XPM cause only non-time-variant phase shifts
and hence the phase distortion is eliminated.
By modeling FWM as additive Gaussian noise,
they obtain an estimate of the information rate achieved by constant-envelope modulation.
It is assumed
that the FWM components from individual fiber spans combine incoherently.
Wegener et al \cite{Wegener2004}
studied WDM transmission over a multi-span dispersive fiber link.
To simplify the solution of the coupled propagation equations analytically,
the same technique in \cite{Mitra2001} is used as well as replacing the FWM with a Gaussian random process.
A lower bound on capacity is evaluated using only the input-output covariance matrix.
Taghavi et al \cite{Taghavi2006} studied WDM transmission over a single-span dispersive fiber link.
They use a (truncated) Volterra series solution to
the propagation equation.
Each receiver uses a linear filter to compensate dispersion followed by a matched filter (matched to the transmitted pulse) whose output is sampled at the symbol rate.
Assuming that dispersion is weak (so that inter-symbol interference can be neglected),
a discrete-time memoryless model is obtained.
The authors consider the case in which
each receiver has access to the received signal of all channels
and thus allows joint decoding with high reliability.
This case was treated as a multiple-access channel.
It is found that non-linearity does not affect the capacity of a channel to the first-order approximation (in the nonlinear coefficient)
and is achieved by performing an interference cancellation step
before decoding.
Moreover, single-channel detection (i.e., the decoder for a given user has access to the received signal at its own wavelength only)
was considered in two regimes: XPM-dominated and FWM-dominated regimes.
It is concluded that the capacity for single-channel detection is significantly reduced compared to the multiple access channel capacity.
Essiambre et al \cite{JLT2010}
reviewed fundamental concepts of digital communications, information theory
and the physical phenomena present in transmission over optical fiber networks.
They also evaluated by numerical simulations a capacity limit estimate for the optical fiber channel using multi-ring constellations
in various scenarios, e.g., different constellation shapings and different fiber dispersion maps.
For WDM transmission, non-linear compensation through backpropagation (of individual channels) was used.
The trend in the various scenarios is that the capacity has a peak value as a function of the launch power.
Bosco et al \cite{Bosco2011GN,Bosco2012GN:erratum}
studied (ultra-dense) WDM transmission over uncompensated optical fiber links
with both distributed and lumped amplification.
They argue that, after digital signal processing (DSP) at the receiver,
the distribution of each of the received constellation points is approximately Gaussian
with independent components, even in the absence of additive ASE noise.
Hence, they adopt a model, called the Gaussian noise (GN) model,
in which the impact of non-linear propagation is approximated by excess additive Gaussian noise.
Using the GN model, capacity estimates are derived.
In \cite{PoggioliniJLT2014GN}, Poggiolini discusses the GN model in depth.
Mecozzi and Essiambre \cite{MecozziJLT2012}
studied multi-channel transmission over a dispersive fiber link with distributed amplification.
They developed a general first-order perturbation theory of the signal propagation
and simplified the expression for highly dispersive, or pseudolinear, transmission.
The signal is linearly-modulated\footnote{The signal is the sum of modulated pulses.} at the transmitter
and the detection apparatus at the receiver is made of an optical filter (to separate the channel), mixing with a local oscillator
and subsequent sampling at the symbol rate.
By concentrating on inter-channel non-linearity (in particular XPM),
they derive a capacity estimate per channel.
An important observation is that
(the kurtosis of)
the constellation of the interfering channels is important in determining the system impairments.
Secondini et al \cite{SecondiniJLT2013}
studied WDM transmission over a dispersive fiber link.
FWM is neglected.
The key simplification is
replacing the unknown intensities appearing in the propagation equation with those corresponding to linear propagation.
They then derive a first-order approximation to the solution
based on frequency-resolved logarithmic perturbations.
The approximate solution is used to develop a discrete-time channel model for the channel which is composed of the
optical fiber link followed by a back-propagation block (and thus it is assumed that SPM is fully compensated), a matched filter,
and sampling at the symbol rate.
The discrete-time model obtained is a linear time-varying model.
By using the theory of mismatched decoding, they compute
the information rate achieved by independent and identically distributed (i.i.d.) Gaussian input symbols and a maximum likelihood \emph{symbol-by-symbol} detector
designed for a memoryless additive white Gaussian noise (AWGN) auxiliary channel with the same covariance matrix as the true channel.
They also evaluate the information rate achieved by a maximum likelihood \emph{sequence} decoder
designed for an auxiliary AWGN channel with inter-symbol interference, with same input-output covariance matrix as the real channel.
Dar et al \cite{DarAllerton2013} proposed a block-memoryless discrete-time channel model
for WDM transmission in the pseudo-linear regime in which XPM is the dominant non-linear effect.
The model is a discrete-time phase noise channel
in which the phase noise process
models XPM and is assumed to be a block-independent process, i.e.,
it remains unchanged within a block but changes independently between blocks.
It is assumed that the phase noise is (real) Gaussian with zero mean and a variance
that depends on the type of modulation.
For the proposed model, two lower bounds were developed:
the first is tight in the low power regime
while the second is better at high power.
In \cite{DarArxiv2013, DarECOC2013}, Dar et al added an extra term to the previous model
to capture non-linear effects that do not manifest themselves as phase noise.
Agrell et al \cite{AgrellJLT2014}
proposed a discrete-time model called the \emph{finite-memory GN model}
for coherent long-haul fiber-optical links without dispersion compensation.
Using the finite-memory GN model, they derive semi-analytic lower bounds for non i.i.d. inputs.
It was shown through numerical simulations that the information rates
of the finite-memory GN model
are higher than the rates of the regular GN model.
We \alert{remark} that there is little justification for the proposed discrete-time model
as it was not derived from a continuous-time description of the system.
Yousefi and Kschischang
\cite{YousefiISIT2013,YousefiISIT2013Fiber,YousefiNFT1,YousefiNFT2,YousefiNFT3}
discuss the non-linear Fourier transform (NFT), a method for
solving a broad class of non-linear differential equations,
and in particular for solving the NLS equation
for noiseless propagation.
The NFT for such systems plays a similar diagonalization role as the ordinary Fourier transform plays for linear systems.
They propose a scheme,
called non-linear frequency-division multiplexing (NFDM),
which can be viewed as a non-linear analogue of orthogonal frequency-division multiplexing.
In NFDM, information is encoded in the
NFT of the signal consisting of two components: a discrete and a continuous spectral function.
By modulating non-interacting degrees of freedom of a signal,
deterministic crosstalk between signal components due to dispersion and non-linearity is eliminated,
i.e., inter-symbol and inter-channel interference are zero
if there is no noise.
\subsection{Contributions and Organization}
We develop discrete-time interference channel models for WDM transmission over a single-span of both dispersionless and dispersive fiber.
The models are based on coupled differential equations that capture SPM, XPM and group velocity mismatch (GVM).
Transmitters send linearly-modulated pulses while receivers use
matched filters with symbol rate sampling (for dispersionless transmission) or banks of filters (for dispersive transmission).
Rather than using Gaussian codebooks,
we design codebooks based on a new technique called interference focusing.
We show that all users achieve a pre-log of 1 simultaneously by using interference focusing.
This paper extends the results in \cite{GhozlanISIT2010} and \cite{GhozlanGVM2011}.
The paper is organized as follows.
We review in Sec. \ref{sec:info-theory} basic definitions and relations in information theory.
In Sec. \ref{sec:fiber_channel}, we describe the wave propagation equation in optical fiber and the impairments that arise in transmission.
We study the case of zero group velocity mismatch (zero dispersion) in Sec. \ref{sec:zero-gvm}.
We extend this model to non-zero group velocity mismatch in Sec. \ref{sec:nonzero-gvm}.
We develop discrete-time interference channel models for both cases
and show that a pre-log of 1 is achievable for all users,
despite XPM that arises due to the fiber non-linearity.
\section{Information Theory}
\label{sec:info-theory}
We review basic information-theoretic quantities and inequalities.
For more details, the reader is referred to \cite{CoverTomIT2006}.
\paragraph{Discrete Random Variables.}
Let $X$,$Y$ and $Z$ be discrete random variables.
Let $p_{X,Y}$ be the joint probability distribution\footnote
{We use the term "probability distribution" to mean the \emph{non-cumulative} probability distribution of a random variable.}
of $X$ and $Y$.
The marginal probability distribution $p_X$ of $X$ is
\begin{align}
p_{X}(x) = \sum_{y} p_{X,Y}(x,y).
\end{align}
The conditional probability distribution of $Y$ given $X$ is
\begin{align}
p_{Y|X}(y|x) = \frac{p_{X,Y}(x,y)}{p_X(x)}
\end{align}
for $p_X(x)>0$.
The entropy of $X$ is given by
\begin{align}
H(X)
= \mathbb{E}[ -\log{p_X(X)} ].
\end{align}
The joint entropy of $(X,Y)$ is
\begin{align}
H(X,Y)
= \mathbb{E}[ -\log{p_{X,Y}(X,Y)} ].
\end{align}
The conditional entropy of $Y$ given $X$ is
\begin{align}
H(Y|X)
&= \mathbb{E}[ -\log{p_{Y|X}(Y|X)} ] \\
&= \sum_{x:p_X(x)>0} p_X(x) H(Y|X=x)
\end{align}
where for $p_X(x)>0$ we define
\begin{align}
H(Y|X=x)
= \mathbb{E}[ -\log{p_{Y|X}(Y|X)} | X=x].
\end{align}
The mutual information between $X$ and $Y$ is
\begin{align}
I(X;Y)
&= H(Y) - H(Y|X) \nonumber\\
&= H(X) - H(X|Y)
= I(Y;X)
\end{align}
and
the conditional mutual information between $X$ and $Y$ given $Z$ is
\begin{align}
I(X;Y|Z)
&= H(X|Z) - H(Y|X,Z) \nonumber\\
&= \sum_{z:p_Z(z)>0} p_Z(z) I(X;Y|Z=z)
\end{align}
where we define
\begin{align}
I(X;Y|Z=z)
= H(X|Z=z) - H(Y|X,Z=z)
\end{align}
for $p_Z(z)>0$.
The following are basic results for the information theoretic quantities defined above.
\begin{itemize}
\item
Non-negativity of mutual information
\begin{align}
I(X;Y) \geq 0.
\end{align}
\item
Conditioning does not increase entropy
\begin{align}
H(Y|X) \leq H(Y).
\end{align}
\item
Chain rule for entropy
\begin{align}
H(X,Y)
&= H(X) + H(Y|Y) \\
&= H(Y) + H(Y|X).
\end{align}
A more general form is
\begin{align}
H(X_1,\ldots,X_n)
= \sum_{k=1}^n H(X_k|X_1,\ldots,X_{k-1}).
\end{align}
\item
Chain rule for mutual information
\begin{align}
I(X;Y,Z)
= I(X;Y) + I(X;Z|Y).
\end{align}
A more general form is
\begin{align}
I(X;Y_1,\ldots,Y_n)
= \sum_{k=1}^n I(X;Y_k|Y_1,\ldots,Y_{k-1}).
\end{align}
\item
Data processing inequality\\
\begin{align}
I(X;Y) \geq I(X;Z)
\end{align}
if $X$---$Y$---$Z$ forms a Markov chain, i.e., $p_{Y,Z|X} = p_{Y|X} p_{Z|Y}$.
As a corollary, we have
\begin{align}
I(X;Y) \geq I(X;g(Y))
\end{align}
with equality if the function $g(\cdot)$ is a one-to-one mapping.
\end{itemize}
\paragraph{Continuous Random Variables.}
Suppose that $X$ and $Y$ are real continuous random variables
with joint probability density function (pdf) $p_{X,Y}$. Let $p_X$ and $p_{Y|X}$ be the marginal pdf of $X$
and the conditional pdf of $Y$ given $X$, respectively.
The differential entropy of $X$ is
\begin{align}
h(X)
= \mathbb{E}[ -\log{p_X(X)} ]
\end{align}
while
the joint differential entropy and
the conditional differential entropy are the respective
\begin{align}
h(X,Y)
= \mathbb{E}[ -\log{p_{X,Y}(X,Y)} ]
\end{align}
\begin{align}
h(Y|X)
= \mathbb{E}[ -\log{p_{Y|X}(Y|X)} ] .
\end{align}
\paragraph{Complex Random Variables}
Suppose that $X$ is a complex continuous random variable.
The notation $h(X)$ is to be understood as
\begin{align}
h(X)
= h(\Re\{X\},\Im\{X\})
\end{align}
where $\Re\{\cdot\}$ and $\Im\{\cdot\}$ are the real and imaginary parts of a complex number.
In other words, $h(X)$ denotes the (joint) entropy of the Cartesian coordinates of $X$.
A complex number $x$ can also be represented in polar coordinates, i.e., by using the amplitude $|x|$ and the phase $\arg{x}$.
The relation between $h(\Re\{X\},\Im\{X\})$ and $h(|X|,\arg{X})$ is
\cite[Lemma 6.16]{Lapidoth2003}
\begin{align}
h(|X|,\arg{X})
= h(\Re\{X\},\Im\{X\}) - \mathbb{E}[\log{|X|}].
\end{align}
It is worth noting that
\begin{align}
I(\Re\{X\},\Im\{X\};Y)
= I(|X|,\arg{X};Y)
\end{align}
for any $X$ and $Y$.
\paragraph{Channel Capacity}
The channel between the inputs $(X_1,\ldots,X_n)$ and the outputs $(Y_1,\ldots,Y_n)$
is defined by a conditional probability of the outputs given the input.
For any input distribution, the information rate
\begin{align}
R = \frac{1}{n} I(X_1,\ldots,X_n;Y_1,\ldots,Y_n)
\end{align}
is achievable, i.e., there exist codes of this rate such that
the probability of decoding error approaches zero as $n$ tends to infinity.
The (information) capacity of a channel is the maximum achievable rate
and is given by
\begin{align}
C = \lim_{n \rightarrow \infty} \sup \frac{1}{n} I(X_1,\ldots,X_n;Y_1,\ldots,Y_n)
\end{align}
where the supremum is over all of possible joint input distributions
satisfying desired constraints, e.g., a power constraint.
For a memoryless channel, the unconstrained expressions simplify to
\begin{align}
R = I(X;Y)
\end{align}
and
\begin{align}
C = \max_{p_X} I(X;Y).
\end{align}
\paragraph{Gaussian Channel}
The discrete-time memoryless complex Gaussian channel is
\begin{align}
Y = X + Z
\end{align}
where the output $Y$ and the input $X$ are complex, and $Z$ is a circularly-symmetric complex Gaussian random variable with mean 0 and variance $\mathbb{E}[|Z|^2]=N$.
For the second moment constraint $\mathbb{E}[|X|^2] \leq P$, the capacity is
\cite{Shannon1948}\cite{CoverTomIT2006}
\begin{align}
C = \log\left(1+\frac{P}{N}\right)
\end{align}
and is achieved by a circularly-symmetric Gaussian input distribution.
\section{Fiber Models}
\label{sec:fiber_channel}
We next discuss noise, chromatic dispersion and Kerr non-linearity in optical fiber.
Lumped optical amplifiers (also referred to as discrete amplifiers), such as erbium-doped optical amplifiers (EDFA), or
distributed amplifiers, such as Raman amplifiers, add noise to the signal
due to amplified spontaneous emission (ASE). The noise is typically modeled as a white Gaussian process.
In lumped amplification, $N_s$ amplifiers are inserted periodically over a fiber link of total length $L$
which creates $N_s$ spans\footnote{In practice, a span is often composed of standard single-mode fiber (SSMF)
followed by dispersion-compensated fiber (DCF).
However, we consider fiber links made of one type of fiber for simplicity.}, each of length $L_s = L/N_s$.
In distributed amplification, the signal is amplified continuously as it propagates through the fiber.
The power spectral density (PSD) of the ASE noise is
\begin{align}
N_{\text{ASE}}^{\text{lumped}} = N_s (e^{\alpha L_s}-1) h\nu n_\text{sp}
\label{eq:ase_lumped}
\end{align}
for lumped amplification
and
\begin{align}
N_{\text{ASE}}^{\text{distrib}} = L \alpha h\nu n_\text{sp}
\label{eq:ase_distrib}
\end{align}
for (ideal) distributed amplification.
In the above expressions,
$h\nu$ is the photon energy,
$n_\text{sp}$ is the spontaneous emission factor
and $\alpha$ is the power attenuation coefficient
\cite{JLT2010}.
Note that
\begin{align}
N_{\text{ASE}}^{\text{lumped}} \longrightarrow N_{\text{ASE}}^{\text{distrib}}
\quad \text{as} \quad L_s \longrightarrow 0
\end{align}
because
\begin{align}
N_s (e^{\alpha L_s}-1) = L \frac{e^{\alpha L_s}-1}{L_s} \longrightarrow L \alpha
\quad \text{as} \quad L_s \longrightarrow 0.
\end{align}
Dispersion arises because the medium absorbs energy through the oscillations of bound electrons, causing a {\it frequency} dependence of the material refractive index~\cite[p.~7]{Agrawal}. The Kerr effect is caused by anharmonic motion of bound electrons in the presence of an intense electromagnetic field, causing an {\it intensity} dependence of the material refractive index~\cite[p.~17, 165]{Agrawal}.
Let $A(z,t)$ be a complex number representing the slowly-varying component (or envelope) of a linearly-polarized, electric field at position $z$ and time $t$ in single-mode fiber.
The equation governing the evolution of $A(z,t)$ as the wave propagates through the fiber is \cite[p. 44]{Agrawal}
\begin{align}
\frac{\partial A}{\partial z} + \beta_1 \frac{\partial A}{\partial t} + i \frac{\beta_2}{2} \frac{\partial^2 A}{\partial t^2}
= - \frac{\alpha}{2} A + i \gamma |A|^2 A
\label{eq:propagation_eq}
\end{align}
where $i = \sqrt{-1}$, $\alpha$ is the power attenuation coefficient,
$\beta_1$ is the reciprocal of the group velocity,
$\beta_2$ is the group velocity dispersion (GVD) coefficient,
$\gamma = n_2 \omega_0 / (c A_{\text{eff}})$,
$n_2$ is the non-linear refractive index,
$\omega_0$ is the carrier frequency,
$c$ is the speed of light in free space,
and $A_{\text{eff}}$ is the effective cross-section area of the fiber.
It is common to specify GVD through
the dispersion parameter $D$ which is related to $\beta_2$ by \cite[p. 11]{Agrawal}
\begin{align}
D = - \frac{2 \pi c \beta_2}{\lambda^2}
\label{eq:d_coeff_def}
\end{align}
where $\lambda$ is the wavelength of the wave in free-space, i.e., $\lambda = 2\pi c/\omega_0$.
By defining a retarded-time reference frame with $T=t-\beta_1 z$, we have \cite[p.~50]{Agrawal}
\begin{align}
\frac{\partial A}{\partial z} + i \frac{\beta_2}{2} \frac{\partial^2 A}{\partial T^2}
= - \frac{\alpha}{2} A + i \gamma |A|^2 A.
\label{eq:gnls_eq}
\end{align}
Moreover, by defining $U = A e^{\alpha z/2}$, the differential equation becomes
\begin{align}
\frac{\partial U}{\partial z} + i \frac{\beta_2}{2} \frac{\partial^2 U}{\partial t^2}
= i \gamma e^{-\alpha z} |U|^2 U
\end{align}
and thus the term $(\alpha/2) A$ is eliminated from the equation.
We will set $\alpha=0$ in the rest of this section for simplicity.
Therefore, we have
\begin{align}
i \frac{\partial A}{\partial z} - \frac{\beta_2}{2} \frac{\partial^2 A}{\partial T^2} + \gamma |A|^2 A = 0
\label{eq:nls_eq}
\end{align}
which is referred to as the non-linear Schr\"{o}dinger (NLS) equation \cite[p. 50]{Agrawal}
because of its similarity to the Schr\"{o}dinger equation with a non-linear potential term
when the roles of time and distance are exchanged.
When $\beta_2 = 0$, Equation (\ref{eq:nls_eq}) has the exact solution \cite[p. 98]{Agrawal}
\begin{align} \label{eq:solution}
A(L,T) = A(0,T) e^{i \gamma L |A(0,T)|^2}
\end{align}
where $L$ is the fiber length.
In other words, Kerr non-linearity leaves the pulse shape unchanged but causes an intensity-dependent phase shift. The phase shift phenomenon is called SPM.
Suppose two optical fields at different carrier frequencies $\omega_1$ and $\omega_2$ are launched at the same location and propagate simultaneously inside the fiber. The fields interact with each other through the Kerr effect \cite[Ch. 7]{Agrawal}.
Specifically, neglecting fiber losses, the propagation is governed by the coupled non-linear Schr\"{o}dinger (NLS) equations \cite[p. 264, 274]{Agrawal}:
\begin{align}
& i \frac{\partial A_1}{\partial z} - \frac{\beta_{21}}{2} \frac{\partial^2 A_1}{\partial T^2}
+ \gamma_1 (|A_1|^2 + 2 |A_2|^2) A_1 = 0
\label{eq:nls2_eq1} \\
& i \frac{\partial A_2}{\partial z} - \frac{\beta_{22}}{2} \frac{\partial^2 A_2}{\partial T^2}
+ \gamma_2 (|A_2|^2 + 2 |A_1|^2) A_2 + i d \frac{\partial A_2}{\partial T} = 0
\label{eq:nls2_eq2}
\end{align}
where
$A_k(z,T)$ is the time-retarded, slowly varying component of field $k$, $k=1,2$, the $\beta_{2k}$ are GVD coefficients, the $\gamma_k$ are non-linear parameters, and $d=\beta_{12}-\beta_{11}$ where the $\beta_{1k}$ are reciprocals of group velocities.
Observe that now the group velocities $1/\beta_{11}$ and $1/\beta_{12}$ enter the equations via $d$ since there is no common reference frame to remove both.
The parameter $d$ is a measure of GVM.
To simplify the model, we assume that the second order dispersion effects are negligible,
\textit{i.e.}, $\beta_{21}=\beta_{22}=0$.
If $d=0$, the coupled NLS equations \eqref{eq:nls2_eq1}-\eqref{eq:nls2_eq2} have the exact solutions \cite[p. 275]{Agrawal}
\begin{align}
& A_1(L,T) = A_1(0,T) e^{i \gamma_1 L ( |A_1(0,T)|^2 + 2 |A_2(0,T)|^2)}
\label{eq:xpm_exact1} \\
& A_2(L,T) = A_2(0,T) e^{i \gamma_2 L ( |A_2(0,T)|^2 + 2 |A_1(0,T)|^2)}
\label{eq:xpm_exact2}
\end{align}
where $z=0$ is the point at which both fields are launched. Kerr non-linearity again leaves the pulse shapes unchanged but causes interference through intensity-dependent phase shifts. The interference phenomenon is called XPM. XPM is an important impairment in optical networks using WDM, see~\cite{JLT2010}.
Kerr non-linearity leads to a third effect called FWM when more than two carriers co-propagate in the fiber.
In FWM, four fields propagating at frequencies $\omega_1$, $\omega_2$, $\omega_3$ and $\omega_4=\omega_1-\omega_2+\omega_3$ interact with each other.
The propagation is governed by the following set of coupled equations
\cite[p. 393]{Agrawal}:
\begin{align}
\frac{\partial A_1}{\partial z} & = i\gamma_1
\left[ (|A_1|^2 + 2 \sum_{k\neq1} |A_k|^2) A_1 + 2 A_2^* A_3 A_4 e^{+i \Delta \kappa z} \right] \nonumber\\
\frac{\partial A_2}{\partial z} & = i\gamma_2
\left[ (|A_2|^2 + 2 \sum_{k\neq2} |A_k|^2) A_2 + 2 A_1^* A_3 A_4 e^{+i \Delta \kappa z} \right] \nonumber\\
\frac{\partial A_3}{\partial z} & = i\gamma_3
\left[ (|A_3|^2 + 2 \sum_{k\neq3} |A_k|^2) A_3 + 2 A_1 A_2 A_4^* e^{-i \Delta \kappa z} \right] \nonumber\\
\frac{\partial A_4}{\partial z} & = i\gamma_4
\left[ (|A_4|^2 + 2 \sum_{k\neq4} |A_k|^2) A_4 + 2 A_1 A_2 A_3^* e^{-i \Delta \kappa z} \right]
\label{eq:clns_fwm}
\end{align}
where
\begin{align}
\gamma_k = \frac{n_2 \omega_k}{c A_{\text{eff}}}
\end{align}
and the wave-vector mismatch $\Delta \kappa$ is given by
\begin{align}
\Delta \kappa = \kappa_4 + \kappa_3 - \kappa_2 - \kappa_1
\end{align}
where $\kappa_1$, $\kappa_2$, $\kappa_3$ and $\kappa_4$ are the propagation constants of the four waves.
We ignore the FWM effect in our models.
\section{Zero Group Velocity Mismatch}
\label{sec:zero-gvm}
We first investigate the case of zero GVM $d=0$, i.e., there is no dispersion.
This will help to prepare for the more general model treated in Sec. \ref{sec:nonzero-gvm}.
We develop a discrete-time two-user channel model based on sampling of electric fields in Sec. \ref{sec:dt-model-2user},
and we show that a pre-log 1/2 is achievable using phase modulation only in Sec. \ref{sec:onering}.
In Sec. \ref{sec:multiring}, we introduce interference focusing
and show that it achieves a pre-log 1 for all users, and therefore no degrees of freedom are lost.
An extension of the discrete-time model to $K$ users is presented in Sec. \ref{sec:dt-model-kuser}.
\subsection{Discrete-Time Two-User Model}
\label{sec:dt-model-2user}
Transmitter $k$ sends a string of symbols $X^n_k = (X_{k}[1], X_{k}[2], \cdots, X_{k}[n])$ using rectangular pulses in the time domain.
The transmitted signals $A_1(0,T)$ and $A_2(0,T)$ propagate through a dispersionless optical fiber of length $L$.
At the end of the fiber ($z=L$), the signals $A_1(L,T)$ and $A_2(L,T)$ are amplified by discrete amplifiers that
introduce ASE noise modeled as additive, white and Gaussian.
Receiver $k$ obtains $Y^n_k = (Y_{k}[1], Y_{k}[2], \cdots, Y_{k}[n])$ by matched filtering the noisy amplified signal and sampling the filter output at the symbol rate.
In this case, \eqref{eq:xpm_exact1}-\eqref{eq:xpm_exact2} imply that the channel is memoryless. Hence, we drop the time indices and write the input-output relationships as
\begin{align}
Y_1 & = X_1 \exp\left({i h_{11} |X_1|^2+ i h_{12} |X_2|^2}\right) + Z_1 \label{eq:channel2a} \\
Y_2 & = X_2 \exp\left({i h_{21} |X_1|^2+ i h_{22} |X_2|^2}\right) + Z_2 \label{eq:channel2b}
\end{align}
where $Z_{k}$ is circularly-symmetric complex Gaussian noise with variance $N$.
We assume that the noise random variables at the receivers are independent.
The term $\exp(i h_{kk} |X_{k}|^2)$ models SPM and the term $\exp(i h_{k\ell} |X_{\ell}|^2)$, $k \neq \ell$, models XPM. We regard the $h_{k\ell}$ as \textit{channel coefficients} that are time invariant. These coefficients are known at the transmitters as well as the receivers.
We use the power constraints
\begin{align} \label{eq:power-constraint-kuser}
\frac{1}{n} \sum_{j=1}^{n} \mathbb{E}\left[ |X_{k}[j]|^2 \right] \le P, \quad k=1,2.
\end{align}
A \textit{scheme} is a collection $\{ (\mathcal{C}_1(P,N),\mathcal{C}_2(P,N) )\}$ of pairs of codes
such that at $(P,N)$, user $k$ uses the code $\mathcal{C}_k(P,N)$
that satisfies the power constraint and achieves an information rate $R_k(P,N)$ where $k=1,2$.
We distinguish between two limiting cases:
1) fixed noise with growing powers and
2) fixed powers with vanishing noise.
\begin{definition} \label{def:highp-prelog}
The \textit{high-power} pre-log pair $(\overline{r}_1, \overline{r}_2)$ is achieved by a scheme
if the rates satisfy
\begin{equation}
\overline{r}_k(N) = \lim_{P \rightarrow \infty} \frac{R_k(P,N)}{\log(P/N)}
\text{ for } k=1,2.
\label{eq:high-power_prelog_userk_def}
\end{equation}
\end{definition}
\begin{definition} \label{def:lowp-prelog}
The \textit{low-noise} pre-log pair $(\underline{r}_1, \underline{r}_2)$ is achieved by a scheme
if the rates satisfy
\begin{equation}
\underline{r}_k(P) = \lim_{N \rightarrow 0} \frac{R_k(P,N)}{\log(P/N)}
\text{ for } k=1,2.
\label{eq:low-noise_prelog_userk_def}
\end{equation}
\end{definition}
The (high-power or low-noise) pre-log pair $(1/2,1/2)$ can be achieved if both users use amplitude modulation only or phase modulation only,
as shown in Sec. \ref{sec:zero-gvm-am} and Sec. \ref{sec:onering}, respectively.
We show in Sec. \ref{sec:multiring} that the high-power pre-log pair $(1,1)$ can be achieved
through \textit{interference focusing}.
\subsection{Amplitude Modulation}
\label{sec:zero-gvm-am}
First, we introduce a result by Lapidoth \cite[Sec. IV]{LapidothPhaseNoise2002}.
\begin{lemma}
Let $Y=X+Z$ where $Z$ is a circularly-symmetric complex Gaussian random variable with mean $0$ and variance $N$.
Define $S \equiv |X|^2/P$. Suppose $S$ is distributed as
\begin{align}
p_S(s) = \frac{e^{-s/2}}{\sqrt{2 \pi s}}, \quad s \geq 0.
\label{eq:gamma_dist_1dof}
\end{align}
In other words, $|X|^2$ follows a Gamma distribution (or a Chi-squared distribution) with \emph{one} degree of freedom and has mean $P$.
Then we have
\begin{align}
I(|X|^2;|Y|^2) \geq \frac{1}{2} \log\left(\frac{P}{2 N} \right) + o(1)
\end{align}
where $o(1)$ tends to zero as $P/N$ tends to infinity.
$\blacksquare$
\end{lemma}
If $|X_1|^2/P$ and $|X_2|^2/P$ are distributed according to $p_S$ in (\ref{eq:gamma_dist_1dof}),
then we have for $k=1,2$
\begin{align}
I(X_k;Y_k)
&\geq I(|X_k|;|Y_k|) \nonumber \\
&= I(|X_k|^2;|Y_k|^2) \nonumber \\
&\ge \frac{1}{2} \log\left(\frac{P}{2 N} \right) + o(1).
\end{align}
It follows that the high-power and low-noise pre-log pair $(1/2,1/2)$ can be achieved when both users use amplitude modulation.
\subsection{Phase Modulation}
\label{sec:onering}
Suppose the transmitters use phase modulation with $|X_1| = \sqrt{P}$ and $|X_2| = \sqrt{P}$.
The input-output equations (\ref{eq:channel2a})--(\ref{eq:channel2b}) become
\begin{align}
Y_1 & = X_1 ~ e^{i h_{11} P + i h_{12} P } + Z_1 \label{eq:Y1_ring}\\
Y_2 & = X_2 ~ e^{i h_{21} P + i h_{22} P } + Z_2. \label{eq:Y2_ring}
\end{align}
Therefore, each receiver sees a constant phase shift which allows us to treat each transmitter-receiver pair separately as an AWGN channel.
We next show that the pre-log pair $(r_1,r_2)=(1/2,1/2)$ can be achieved by using phase modulation only.
\begin{theorem}[One-Ring Modulation]
\label{theorem:one-ring}
Fix $P>0$. Let $Y=X+Z$ where $Z$ is a circularly-symmetric complex Gaussian random variable with mean $0$ and variance $N$,
and $X = \sqrt{P} e^{i \Phi_X}$ where $\Phi_X$ is a real random variable uniformly distributed on $[0,2\pi)$.
Then we have
\begin{align}
I(X;Y)
&\ge \frac{1}{2} \log\left(\frac{2 P}{N} \right) - 1 ~ (\text{nats}).
\label{eq:one_ring_lb}
\end{align}
\end{theorem}
\begin{proof}
We have
\begin{align}
I(X;Y)
&= h(Y) - h(Y|X) \nonumber \\
&= h(Y) - h(Z) \nonumber \\
&= \mathbb{E}[-\log p_Y(Y)] - \log(\pi e N) \label{eq:I_PhiX_Y}
\end{align}
The pdf $p_Y$ of $Y$ can be shown to be \cite[p. 688]{JLT2010}
\begin{align}
p_Y(y)
&= \frac{1}{\pi N} e^{-(y_A^2+P)/N} I_0\left(\frac{2 y_A \sqrt{P} }{N}\right)
\end{align}
where $I_0(\cdot)$ is the modified Bessel function of the first kind of order zero
and $Y_A = |Y|$.
Therefore, we have
\begin{align}
h(Y)
&= \mathbb{E}\left[
-\log\left(\frac{1}{\pi N} e^{-(Y_A^2+P)/N} I_0\left(\frac{2 Y_A \sqrt{P} }{N}\right) \right)
\right] \nonumber \\
&\stackrel{(a)}{\geq}
\mathbb{E}\left[
-\log\left(\frac{1}{\pi N} \frac{e^{-(Y_A - \sqrt{P})^2 /N}}{\sqrt{2 Y_A \sqrt{P} /N}} \right)
\right] \nonumber \\
&\stackrel{(b)}{\geq}
\mathbb{E}\left[ \log\left({\pi N} {\sqrt{2 Y_A \sqrt{P}/N}} \right) \right] \nonumber \\
& =
\frac{1}{4} \log\left(\frac{2P}{N}\right)
+
\log\left({\pi N}\right)
+
\frac{1}{2} \mathbb{E} \left[ \log\left(Y_A \sqrt{\frac{2}{N}}\right) \right]
\label{eq:entropy_y_lowerbound}
\end{align}
where $(a)$ follows by using Lemma \ref{lemma:besseli_lb} in Appendix A
and ($b$) holds because $(Y_A-\sqrt{P})^2 \geq 0$.
The pdf of $Y_A$ is given by
\begin{align}
p_{Y_A}(y_A)
&= \int_{-\pi}^{\pi} p_{Y}(y) y_A d\phi_y \nonumber\\
&= \frac{2 y_A}{N} e^{-(y_A^2+P)/N}
I_0\left(\frac{2 y_A \sqrt{P}}{N}\right) .
\label{eq:pya}
\end{align}
The last expectation in \eqref{eq:entropy_y_lowerbound} is
\begin{align}
\mathbb{E} \left[ \log\left(Y_A \sqrt{\frac{2}{N}}\right) \right]
&= \int_{y_A=0}^{\infty} p_{Y_A}(y_A)
\log\left(y_A \sqrt{\frac{2}{N}}\right) \, dy_A
\nonumber \\
&\stackrel{(a)}{=} \int_{z=0}^{\infty} z e^{-(z^2 + \nu^2)/2}
I_0\left( z \nu \right) \log(z) \, dz \nonumber \\
&\stackrel{(b)}{=} \frac{1}{2} \left[ \Gamma\left(0,\frac{P}{N}\right) + \log\left(\frac{2P}{N}\right) \right]
\nonumber \\
&\stackrel{(c)}{\geq} \frac{1}{2} \log\left(\frac{2P}{N}\right)
\label{eq:elogya_lowerbound}
\end{align}
where $\Gamma(a,x)$ is the upper incomplete Gamma function defined in (\ref{eq:upper-incomplete-gamma}).
Step ($a$) follows by setting $\nu^2=2P/N$ and $z = y_A \sqrt{2/N}$,
($b$) follows from Lemma \ref{lemma:integral-lemma3} (see Appendix B) and
($c$) holds because $\Gamma\left(0,x\right) \ge 0$ for $x\ge0$.\footnote{Note that $\lim_{x \rightarrow \infty} \Gamma\left(0,x\right) = 0$.}
Combining (\ref{eq:I_PhiX_Y}), (\ref{eq:entropy_y_lowerbound}) and (\ref{eq:elogya_lowerbound}) gives
\begin{align}
I(X;Y)
&\ge \frac{1}{2} \log\left(\frac{2 P}{N} \right) - \log(e)
\end{align}
which concludes the proof.
\end{proof}
Now, suppose that $X_k = \sqrt{P} e^{i \Phi_{X,k}}$ for $k=1,2$ where
$\Phi_{X,1}$ and $\Phi_{X,2}$ are statistically independent and uniformly distributed on $[0,2\pi)$.
It follows from (\ref{eq:Y1_ring}), (\ref{eq:Y2_ring}) and Theorem \ref{theorem:one-ring} that
the high-power and low-noise pre-log
pair $(1/2,1/2)$ can be achieved when both users use phase modulation.
\subsection{Interference Focusing}
\label{sec:multiring}
We propose an \textit{interference focusing} technique in which the transmitters \textit{focus} their phase interference on one point by constraining their transmitted signals to satisfy
\begin{align}
h_{21} |X_1|^2 & = 2 \pi \tilde{n}_1, ~ \tilde{n}_1=1,2,3,\ldots \label{eq:ring-constraint-1} \\
h_{12} |X_2|^2 & = 2 \pi \tilde{n}_2, ~ \tilde{n}_2=1,2,3,\ldots \label{eq:ring-constraint-2}
\end{align}
In other words, the transmitters use \textit{multi-ring modulation} with specified spacings between the rings.\footnote{Multi-ring modulation was used in~\cite{JLT2010,Essiambre2008PRL,Essiambre2009} for symmetry and computational reasons. We here find that it is useful for improving rate.} We thereby remove XPM interference and \eqref{eq:channel2a}-\eqref{eq:channel2b} reduce to
\begin{align}
Y_k = X_k e^{i h_{kk} |X_k|^2} + Z_k, \quad k=1,2.
\end{align}
This channel is effectively an AWGN channel since $h_{kk}$ is known by receiver $k$ and the SPM phase shift is determined by the desired signal $X_k$.
We will show that the high-power pre-log pair $(1,1)$ is achieved under the constraints \eqref{eq:ring-constraint-1}-\eqref{eq:ring-constraint-2}.
\begin{theorem}[Multi-Ring Modulation]
\label{theorem:multi-ring}
Let $Y=X+Z$ where $Z$ is a circularly-symmetric complex Gaussian random variable with mean $0$ and variance $N$.
Suppose $\mathbb{E}[|X|^2] \leq P$ and $|X|^2$ is allowed to take on values that are multiples of a fixed real number $\hat{p}>0$,
i.e., $|X|^2 = m \hat{p}$ where $m \in \mathbb{N}$. Then there exists a probability distribution $p_X$ of $X$ such that
\begin{align}
\lim_{P \rightarrow \infty} \frac{I(X;Y)}{\log({P/N})}
&\ge 1.
\label{eq:multi_ring_lb}
\end{align}
\end{theorem}
\begin{proof}
Define $X_A=|X|$ and $\Phi_X = \arg{X}$.
Consider multi-ring modulation, i.e.,
$X_A$ and $\Phi_X$ are statistically independent,
$\Phi_X$ is uniformly distributed on the interval $[0,2\pi)$
and $X_A \in \{\sqrt{\textsf{P}_j}:j=1,\ldots,J\}$
where $J$ is the number of rings.
We choose the rings to be spaced uniformly in amplitude as
\begin{align} \label{eq:power-levels}
\textsf{P}_j = a j^2 \, \hat{p}
\end{align}
where $a$ is a positive integer.
We further use a uniform frequency of occupation of rings with $P_{X_A}(\sqrt{\textsf{P}_j}) = 1/J$, $j=1,2,\ldots,J$.
The power constraint is therefore
\begin{align} \label{eq:pwr_constraint_uniform}
\frac{1}{J} \sum_{j=1}^J a j^2 \, \hat{p} \le P .
\end{align}
For (\ref{eq:pwr_constraint_uniform}), we compute
\begin{align}
\frac{1}{J} \sum_{j=1}^{J} a \hat{p} \, j^2 = a \hat{p} \frac{(J+1)(2J+1)}{6}
\end{align}
and to satisfy the power constraint we choose\footnote{The solution for $J$ should be positive and rounded down to the nearest integer but we ignore these issues for notational simplicity.}
\begin{align}
J = \frac{-3 + \sqrt{1 + 48 P/(a \hat{p})}}{4}.
\end{align}
Moreover, we choose $a$ to scale as $N\log({P/N})$ so $J$ scales as $\sqrt{({P/N})/\log({P/N})}$.
We have
\begin{align}
I(X;Y)
&= I(X_A , \Phi_X;Y) \nonumber \\
& = I(X_A;Y) + I(\Phi_X;Y|X_A) \label{eq:mult-ring-rate}
\end{align}
The term $I(X_A ; Y)$ can be viewed as the amplitude contribution while the term $I(\Phi_X ; Y | X_A)$ is the phase contribution.
\subsubsection{Phase Contribution}
We show that the phase modulation contributes 1/2 to the pre-log
when using multi-ring modulation.
\begin{lemma} \label{lemma:sum-bounds-integral}
A non-decreasing function $f(x)$ in $x$ satisfies, for integers $a$ and $b$ with $a \leq b$,
\begin{align}
\int_{x=a-1}^b f(x) dx \leq \sum_{i=a}^b f(i).
\end{align}
\end{lemma}
We thus have
\begin{align}
I(\Phi_X;Y|X_A)
& \stackrel{(a)}{=} \sum_{j=1}^J \frac{1}{J} \, I(\Phi_X;Y|X_A=\sqrt{\textsf{P}_j}) \nonumber \\
& \stackrel{(b)}{\geq} \sum_{j=1}^J \frac{1}{J} \, \frac{1}{2} \log\left( \frac{\textsf{P}_j}{N} \right) - 1 \nonumber \\
& \stackrel{(c)}{=} \frac{1}{2 J} \sum_{j=1}^J \log\left(\frac{a j^2 \hat{p}}{N}\right) - 1 \nonumber \\
& \stackrel{(d)}{\geq} \frac{1}{2 J} \int_{x=0}^J \log\left(\frac{a x^2 \hat{p}}{N}\right) dx - 1 \nonumber \\
& \stackrel{(e)}{=} \frac{1}{2} \log\left( \frac{a J^2 \hat{p}}{N e^2} \right) - 1
\end{align}
where
($a$) follows from the uniform occupation of rings,
($b$) follows from Theorem \ref{theorem:one-ring},
($c$) holds by choosing the rings according to \eqref{eq:power-levels},
($d$) follows from Lemma \ref{lemma:sum-bounds-integral} since the logarithm is an increasing function and
($e$) follows by using $\log(a x^2 \hat{p}/N) = \log(a \hat{p}/N) + 2 \log(x)$ and
\begin{align}
\int \log(x) dx = x \log\left({x/e}\right).
\end{align}
We can therefore write
\begin{align}
\lim_{P \rightarrow \infty} \frac{I(\Phi_X;Y|X_A)}{\log({P/N})}
&\ge \lim_{P \rightarrow \infty} \frac{\frac{1}{2} \log(a J^2 \hat{p}/N)}{\log({P/N})}
= \frac{1}{2} \label{eq:phase-prelog}
\end{align}
where \eqref{eq:phase-prelog} follows because $a$ scales as $N\log({P/N})$, $J^2$ scales as $({P/N})/\log({P/N})$, and $\hat{p}$ is independent of $P$ and $N$.
The pre-log of the phase contribution is therefore at least $1/2$.
\subsubsection{Amplitude Contribution}
We show that amplitude modulation contributes $1/2$ to the pre-log.
We have
\begin{align}
I(X_A;Y) = H(X_A) - H(X_A|Y)
\end{align}
where $H(X_A) = \log(J)$. We showed previously that $J$ scales as $\sqrt{({P/N})/\log({P/N})}$ if $a$ scales as $N\log({P/N})$. We bound $H(X_A|Y)$ using Fano's inequality as
\begin{align}
H(X_A|Y) &\le H(X_A|\hat{X}_A) \nonumber \\
&\le H(P_e) + P_e \log(J-1)
\label{eq:fano}
\end{align}
where $\hat{X}_A$ is any estimate of $X_A$ given $Y$, $P_e=\Pr[\hat{X}_A \ne X_A]$ and $H(P_e)$ is the binary entropy function with a general logarithm base. Suppose we use the minimum distance estimator
\begin{align}
\hat{X}_A = \arg \min_{x_A \in \mathcal{X}_A} |Y_A-x_A|
\label{eq:min_distance_dec}
\end{align}
where $Y_A=|Y|$ and $\mathcal{X}_A = \{\sqrt{\textsf{P}_j}:j=1,\ldots,J\}$.
The probability of error $P_e$ is upper bounded by
(see Lemma \ref{lemma:min-distance-estimator})
\begin{align}
P_e \leq \frac{2}{J} \sum_{j=2}^{J} \exp\left(-\frac{\Delta_j^2}{4}\right)
\end{align}
where $\Delta_j = (\sqrt{\textsf{P}_j} - \sqrt{\textsf{P}_{j-1}})/\sqrt{N}$.
For the power levels (\ref{eq:power-levels}), we have $\Delta_j = \sqrt{a \hat{p}/N}$ for all $j$, and hence
\begin{align} \label{eq:Pe-bound-2}
P_e
\le \frac{2(J-1)}{J} \exp\left(-\frac{a \hat{p}}{4 N}\right)
\le 2 \exp\left(-\frac{a \hat{p}}{4 N}\right) .
\end{align}
We see from \eqref{eq:Pe-bound-2} that $\lim_{P \rightarrow \infty} P_e = 0$ if $a$ scales as $N\log({P/N})$
(recall that $J$ scales as $\sqrt{({P/N})/\log({P/N})}\,$).
We thus have $\lim_{P \rightarrow \infty} H(X_A|Y) = 0$ by using (\ref{eq:fano}).
Consequently, we have
\begin{align} \label{eq:amplitude-prelog}
\lim_{P \rightarrow \infty} \frac{I(X_A;Y)}{\log({P/N})}
= \lim_{P \rightarrow \infty} \frac{\log(J)}{\log({P/N})}
= \frac{1}{2}.
\end{align}
Finally, combining \eqref{eq:mult-ring-rate}, \eqref{eq:phase-prelog}, and \eqref{eq:amplitude-prelog} gives (\ref{eq:multi_ring_lb}).
\end{proof}
We conclude that interference focusing achieves the largest-possible high-power pre-log of 1. Each user can therefore exploit all the phase and amplitude degrees of freedom simultaneously.
\subsection{Discrete-Time K-User Model}
\label{sec:dt-model-kuser}
Equations \eqref{eq:xpm_exact1}-\eqref{eq:xpm_exact2} generalize to $K$ frequencies.
This motivates the following memoryless interference network model based on sampling the fields $A_k(z,T)$, $k=1,2,\ldots,K$, at $z=0$ and $z=L$. Transmitter $k$ sends a string of symbols $X^n_k = (X_{k}[1], X_{k}[2], \cdots, X_{k}[n])$ while receiver $k$ sees $Y^n_k = (Y_{k}[1], Y_{k}[2], \cdots, Y_{k}[n])$.
We model the input-output relationship at each time instant $j$ as
\begin{align}
Y_{k}[j] = X_{k}[j] \exp\left( i \sum_{\ell=1}^K h_{k\ell} |X_{\ell}[j]|^2 \right) + Z_{k}[j]
\label{eq:channel}
\end{align}
for $k=1,2,\ldots,K$ where $Z_{k}[j]$ is circularly-symmetric complex Gaussian noise with variance $N$.
All noise random variables at different receivers and different times are taken to be independent. The terms $\exp(i h_{kk} |X_{k}[j]|^2)$ model SPM and the terms $\exp(i h_{k\ell} |X_{\ell}[j]|^2)$, $k \neq \ell$, model XPM. The $h_{k\ell}$ are again \textit{channel coefficients} that are time invariant and are known at the transmitters as well as the receivers.
The power constraints are
\begin{align} \label{eq:power-constraint}
\frac{1}{n} \sum_{j=1}^{n} \mathbb{E}\left[ |X_{k}[j]|^2 \right] \le P, \quad k=1,2,\ldots,K.
\end{align}
\subsubsection{Interference Focusing}
We outline how to apply interference focusing to problems with $K>2$. Define the interference phase vector
\begin{align}
\underline{\Psi} \stackrel{\Delta}{=} [\Psi_1,\Psi_2,\ldots,\Psi_K]^T
\end{align}
where $\Psi_k = \sum_{\ell=1}^K h_{k\ell} |X_{\ell}|^2$ and the instantaneous power vector
\begin{align}
\underline{\Pi} \stackrel{\Delta}{=} \left[ |X_1|^2,\ldots,|X_K|^2 \right]^T.
\end{align}
The relation between the $\underline{\Psi}$ and $\underline{\Pi}$ in matrix form is
\begin{align}
\underline{\Psi} = H_{SP} \, \underline{\Pi} + H_{XP} \, \underline{\Pi}
\end{align}
where $H_{SP}$ is a diagonal matrix that accounts for SPM and $H_{XP}$ is a zero-diagonal matrix that accounts for XPM.
We outline the argument for $K=3$.
Suppose the XPM matrix for a 3-user interference network is
\begin{align}
H_{XP} = \left[
\begin{array}{ccc}
0 & 1/2 & 3/5 \\
3/4 & 0 & 2/3 \\
5/6 & 1/5 & 0
\end{array}
\right].
\end{align}
Suppose that each transmitter knows the channel coefficients between itself and all the receiving nodes. The transmitters can thus use power levels of the form
\begin{align}
\underline{\Pi} & = 2 \pi \cdot \left[ \, \text{lcm}(4,6) m_1, \text{lcm}(2,5) m_2, \text{lcm}(5,3) m_3 \,\right] \nonumber \\
& = 2 \pi \cdot \left[\, 12 m_1, 10 m_2, 15 m_3 \,\right]
\end{align}
where $\text{lcm}(a,b)$ is the least common multiple of $a$ and $b$, and $m_1,m_2,m_3$ are positive integers. We thus have
\begin{align}
H_{XP} \, \underline{\Pi}
=
2 \pi
\left[
\begin{array}{ccc}
0 & 5 & 9 \\
9 & 0 & 10 \\
10 & 2 & 0
\end{array}
\right]
\left[
\begin{array}{c}
m_1 \\
m_2 \\
m_3
\end{array}
\right]
\end{align}
which implies that the phase interference has been eliminated.
The above example combined with an analysis similar to Section \ref{sec:multiring} shows that interference focusing will give each user a pre-log of $1$ even for $K$-user interference networks. However, the XPM coefficients $h_{k\ell}$ must be \textit{rationals}. Modifying interference focusing for \textit{real-valued} XPM coefficients is an interesting problem. It is clear from the example that interference focusing does not require global channel state information.
\section{Non-Zero Group Velocity Mismatch}
\label{sec:nonzero-gvm}
For simplicity, we ignore polarization effects and fiber losses in this section.
The evolution of the slowly-varying components of three optical fields
at three different center frequencies inside the fiber is governed by
\ifdefined\twocolumnmode{
\begin{align}
& i \frac{\partial A_1}{\partial z} + i \beta_{11} \frac{\partial A_1}{\partial t}
- \frac{\beta_{21}}{2} \frac{\partial^2 A_1}{\partial t^2}
\nonumber\\&\qquad\qquad
+ \gamma_1 (|A_1|^2 + 2 |A_2|^2 + 2 |A_3|^2) A_1 = 0
\\
& i \frac{\partial A_2}{\partial z} + i \beta_{12} \frac{\partial A_2}{\partial t}
- \frac{\beta_{22}}{2} \frac{\partial^2 A_2}{\partial t^2}
\nonumber\\&\qquad\qquad
+ \gamma_2 (|A_2|^2 + 2 |A_1|^2 + 2 |A_3|^2) A_2 = 0
\\
& i \frac{\partial A_3}{\partial z} + i \beta_{13} \frac{\partial A_3}{\partial t}
- \frac{\beta_{23}}{2} \frac{\partial^2 A_3}{\partial t^2}
\nonumber\\&\qquad\qquad
+ \gamma_3 (|A_3|^2 + 2 |A_1|^2 + 2 |A_2|^2) A_3 = 0.
\end{align}
}\else{
\begin{align}
& i \frac{\partial A_1}{\partial z} + i \beta_{11} \frac{\partial A_1}{\partial t}
- \frac{\beta_{21}}{2} \frac{\partial^2 A_1}{\partial t^2}
+ \gamma_1 (|A_1|^2 + 2 |A_2|^2 + 2 |A_3|^2) A_1 = 0
\\
& i \frac{\partial A_2}{\partial z} + i \beta_{12} \frac{\partial A_2}{\partial t}
- \frac{\beta_{22}}{2} \frac{\partial^2 A_2}{\partial t^2}
+ \gamma_2 (|A_2|^2 + 2 |A_1|^2 + 2 |A_3|^2) A_2 = 0
\\
& i \frac{\partial A_3}{\partial z} + i \beta_{13} \frac{\partial A_3}{\partial t}
- \frac{\beta_{23}}{2} \frac{\partial^2 A_3}{\partial t^2}
+ \gamma_3 (|A_3|^2 + 2 |A_1|^2 + 2 |A_2|^2) A_3 = 0.
\end{align}
}\fi
Suppose $\beta_{21} = 0$, $\beta_{22} = 0$ and $\beta_{23} = 0$, \textit{i.e.},
we have no GVD.
The solution at $z=L$, where $L$ is the fiber length, is given by
\begin{align}
A_k(L,t) &= A_{k}(0,t-\beta_{k1} L) \exp\left( i \phi_{k}(L,t-\beta_{k1} L) \right)
\end{align}
where $k=1,2,3$ and the time-dependent nonlinear phase shifts $\phi_{k}(L,t)$ are
\ifdefined\twocolumnmode{
\begin{align}
\phi_{1}(L,t) &=
\int_0^L \gamma_1 \big( |A_{1}(0,t)|^2
+ 2 |A_{2}(0,t+(\beta_{11} -\beta_{12}) \zeta)|^2
\nonumber\\&\qquad\qquad
+ 2 |A_{3}(0,t+(\beta_{11} -\beta_{13}) \zeta)|^2 \big) d\zeta
\\
\phi_{2}(L,t) &=
\int_0^L \gamma_2 \big( |A_{2}(0,t)|^2
+ 2 |A_{1}(0,t+(\beta_{12} -\beta_{11}) \zeta)|^2
\nonumber\\&\qquad\qquad
+ 2 |A_{3}(0,t+(\beta_{12} -\beta_{13}) \zeta)|^2 \big) d\zeta
\\
\phi_{3}(L,t) &=
\int_0^L \gamma_3 \big( |A_{3}(0,t)|^2
+ 2 |A_{1}(0,t+(\beta_{13} -\beta_{11}) \zeta)|^2
\nonumber\\&\qquad\qquad
+ 2 |A_{2}(0,t+(\beta_{13} -\beta_{12}) \zeta)|^2 \big) d\zeta
\end{align}
}\else{
\begin{align}
\phi_{1}(L,t) &=
\int_0^L \gamma_1 \big( |A_{1}(0,t)|^2
+ 2 |A_{2}(0,t+(\beta_{11} -\beta_{12}) \zeta)|^2
+ 2 |A_{3}(0,t+(\beta_{11} -\beta_{13}) \zeta)|^2 \big) d\zeta
\\
\phi_{2}(L,t) &=
\int_0^L \gamma_2 \big( |A_{2}(0,t)|^2
+ 2 |A_{1}(0,t+(\beta_{12} -\beta_{11}) \zeta)|^2
+ 2 |A_{3}(0,t+(\beta_{12} -\beta_{13}) \zeta)|^2 \big) d\zeta
\\
\phi_{3}(L,t) &=
\int_0^L \gamma_3 \big( |A_{3}(0,t)|^2
+ 2 |A_{1}(0,t+(\beta_{13} -\beta_{11}) \zeta)|^2
+ 2 |A_{2}(0,t+(\beta_{13} -\beta_{12}) \zeta)|^2 \big) d\zeta
\end{align}
}\fi
This solution follows from steps
similar to the steps outlined in Appendix \ref{sec:coupled-pdes} for two coupled equations.
Define
\begin{align}
d_{kj} \stackrel{\Delta}{=} \beta_{1k}-\beta_{1j}.
\end{align}
\subsection{Continuous-Time Model}
\label{sec:ct_model}
Consider the case of $\beta_{13} \neq \beta_{12} \neq \beta_{11}$.
Without loss of generality, suppose that $\beta_{13} > \beta_{12} > \beta_{11}$.
Let $\{x_k[m]\}_{m=0}^{n-1}$ be the codeword sent by transmitter $k$.
Suppose the transmitters use a pulse of shape $p(t)$
where $p(t)=0$ for $t \notin [0,T_s]$ and
\begin{align}
\int_0^{T_s} |p(\lambda)|^2 d\lambda = E_s.
\label{eq:nonzerogvm-pulse-energy}
\end{align}
The signal sent by transmitter $k$ is
\begin{align}
A_k(0,t) = \sum_{m=0}^{n-1} x_k[m] \ p(t-m T_s).
\end{align}
The signal observed by receiver $k$ is
\begin{align}
r_k(\T) = A_k(L,\T) + z_k(\T)
\end{align}
where $z_k(\T)$ is circularly-symmetric white Gaussian noise with
$\mathbb{E}[z_k(\T)] = 0$, and
$\mathbb{E}[z_k(\T) z_k^*(\T+\tau)] = N \delta(\tau)$.
The processes $z_1(\T)$, $z_2(\T)$ and $z_3(\T)$ are statistically independent.
The signal seen by receiver $k$ is fed to a bank of linear time-invariant (LTI) filters with impulse responses
$\{h_{\idx}(\T)\}_{\idx \in \mathcal{F}_k}$,
where $\mathcal{F}_k \subset \mathbb{Z}=\{\ldots,-1,0,1,\ldots\}$
and
\begin{align}
h_{\idx}(t) = p^*(-t) \exp(-i 2\pi {\idx} K(-t))
\end{align}
where $K(t)$ is defined as
\begin{align}
K(t) = \frac{1}{E_s} \int_{0}^{t} |p(\lambda)|^2 d\lambda.
\label{eq:rise_func}
\end{align}
The choice of the set $\mathcal{F}_k$ is specified in Sec. \ref{sec:ifocus}.
We show in Appendix \ref{sec:ortho_impulses} that the impulse responses of the filters are orthogonal, i.e., if $\idx_1 \neq \idx_2$, then we have
\begin{align}
\int_{-\infty}^{\infty} h_{\idx_1}(\xi) h^*_{\idx_2}(\xi) d\xi
= 0.
\label{eq:ortho_diff}
\end{align}
The remaining analysis is similar for all receivers, hence we present the analysis for receiver 1 only.
The output of the filter with index ${\idx}$ is
\begin{align*}
y_{1,{\idx}}(\T) = r_1(\T) \star h_{\idx}(\T)
\end{align*}
where $\star$ denotes convolution.
The noiseless part $\tilde{y}_{1,{\idx}}(\T)$ of the output of this filter is
\begin{align}
\ifdefined\twocolumnmode & \fi
\tilde{y}_{1,{\idx}}(\T+\beta_{11} L)
\ifdefined\twocolumnmode \nonumber\\ \fi
&\stackrel{\Delta}{=}
A_1(L,\T+\beta_{11} L) \star h_{\idx}(\T) \nonumber\\
&= \left( A_1(0,\T) e^{i \phi_1(L,\T)} \right) \star h_{\idx}(\T) \nonumber \\
&= \left( \sum_{m=0}^{n-1} x_1[m] \ p(\T-m T_s) e^{i \phi_1(L,\T)} \right) \star h_{\idx}(\T) \nonumber \\
&= \sum_{m=0}^{n-1} x_1[m]
\int p(\tau-m T_s) p^*(\tau-\T)
e^{i \phi_1(L,\tau) - i 2\pi \idx K(\tau-\T)} d\tau
\end{align}
where the integral is over the whole real line.
Sampling the output signal $y_{1,{\idx}}(\T+\beta_{11} L)$ at the time instants $\T = j T_s$, for $j=1,2,\ldots,n$, yields
\begin{align}
\ifdefined\twocolumnmode & \fi
\tilde{y}_{1,{\idx}}(j T_s+\beta_{11} L)
\ifdefined\twocolumnmode \nonumber\\ \fi
&= x_1[j] \
\int_{j T_s}^{j T_s+T_s}
|p(\tau-j T_s)|^2 e^{i \phi_1(L,\tau) - i 2\pi \idx K(\tau-j T_s)} d\tau
\label{eq:y1nj_noiseless_integral}
\end{align}
where we used $p(t)=0$ for $t \notin [0,T_s]$.
We write $\phi_1(L,\tau)$ as
\begin{align}
\phi_1(L,\tau) = \phi_{11}(L,\tau) + \phi_{12}(L,\tau) + \phi_{13}(L,\tau)
\label{eq:phi1_ct}
\end{align}
where we have defined
\begin{align}
\phi_{11}(L,t) &\stackrel{\Delta}{=} \gamma_1 L \ |A_1(0,t)|^2
\label{eq:phi11_definition} \\
\phi_{12}(L,t) &\stackrel{\Delta}{=} 2 \gamma_1 L_{12} \
\frac{1}{T_s} \int_{t-L d_{21}}^{t} |A_2(0,\lambda)|^2 d\lambda
\label{eq:phi12_definition} \\
\phi_{13}(L,t) &\stackrel{\Delta}{=} 2 \gamma_1 L_{13} \
\frac{1}{T_s} \int_{t-L d_{31}}^{t} |A_3(0,\lambda)|^2 d\lambda
\label{eq:phi13_definition}
\end{align}
and where $L_{1k} \stackrel{\Delta}{=} T_s/|d_{1k}|$ for $k \neq 1$.
Since $p(t)=0$ for $t \notin [0,T_s]$, we have
\begin{align}
\phi_{12}(L,t)
&= \frac{2 \gamma_1 L_{12}}{T_s} \int_{t-L d_{21}}^{t} \sum_{m=0}^{n-1} |x_2[m]|^2 \ |p(\lambda-m T_s)|^2 d\lambda \nonumber\\
&= \frac{2 \gamma_1 L_{12}}{T_s} \sum_{m=0}^{n-1}
|x_2[m]|^2 \ \int_{t-L d_{21}}^{t} |p(\lambda-m T_s)|^2 d\lambda \nonumber\\
&= 2 \gamma_1 L_{12} \frac{E_s}{T_s} \sum_{m=0}^{n-1} |x_2[m]|^2 \ \psi(t-m T_s;d_{21})
\label{eq:psi12_expansion}
\end{align}
where $\psi(t;d)$ is defined as
\begin{align}
\psi(t;d) \stackrel{\Delta}{=} \frac{1}{E_s} \int_{t-L d}^{t} |p(\lambda)|^2 d\lambda.
\label{eq:psi12m_definition}
\end{align}
If $L d \geq T_s$, then
\begin{align}
\psi(t;d) =
\left\{
\begin{array}{ll}
K(t), & 0 \leq t < T_s \\
1, & T_s \leq t < L d \\
\tilde{K}(t;d), & L d \leq t < L d + T_s \\
0, & \text{otherwise}
\end{array}
\right.
\label{eq:psi12m_Ld_gt_Ts}
\end{align}
where $K(t)$ is defined by (\ref{eq:rise_func}) and $\tilde{K}(t;d)$ is given by
\begin{align}
\tilde{K}(t;d)
&= \frac{1}{E_s} \int_{t-L d}^{T_s} |p(\lambda)|^2 d\lambda \nonumber \\
&= \frac{1}{E_s} \int_{0}^{T_s} |p(\lambda)|^2 d\lambda - \frac{1}{E_s} \int_{0}^{t-L d} |p(\lambda)|^2 d\lambda \nonumber \\
&= 1 - K(t-L d).
\label{eq:fall_func}
\end{align}
One can express $\phi_{13}(L,t)$ in a similar manner.
Suppose that $L |d_{1k}| = M_{1k} T_s$ for some positive integer $M_{1k}$ for $k=2,3$.
Hence, for $ \tau \in [j T_s, j T_s+T_s]$,
we have\footnote{We use the convention of setting the quantities that involve a negative time index to zero.}
\ifdefined\twocolumnmode{
\begin{align}
\phi_{11}(L,\tau)
&= \gamma_1 L \frac{E_s}{T_s} |x_1[j]|^2 \label{eq:phi11}\\
\phi_{12}(L,\tau)
&= 2 \gamma_1 L_{12} \frac{E_s}{T_s} \Big(
\Big( \sum_{r=1}^{M_{12}} |x_2[j-r]|^2 \Big) + \nonumber \\& \qquad
\left( |x_2[j]|^2 - |x_2[j-M_{12}]|^2 \right) K(t-j T_s)
\Big) \label{eq:phi12}\\
\phi_{13}(L,\tau)
&= 2 \gamma_1 L_{13} \frac{E_s}{T_s} \Big(
\Big( \sum_{r=1}^{M_{13}} |x_3[j-r]|^2 \Big) + \nonumber \\& \qquad
\left( |x_3[j]|^2 - |x_3[j-M_{13}]|^2 \right) K(t-j T_s)
\Big) .
\label{eq:phi13}
\end{align}
}\else{
\begin{align}
\phi_{11}(L,\tau)
&= \gamma_1 L \frac{E_s}{T_s} |x_1[j]|^2 \label{eq:phi11}\\
\phi_{12}(L,\tau)
&= 2 \gamma_1 L_{12} \frac{E_s}{T_s} \Big(
\Big( \sum_{r=1}^{M_{12}} |x_2[j-r]|^2 \Big) +
\left( |x_2[j]|^2 - |x_2[j-M_{12}]|^2 \right) K(t-j T_s)
\Big) \label{eq:phi12}\\
\phi_{13}(L,\tau)
&= 2 \gamma_1 L_{13} \frac{E_s}{T_s} \Big(
\Big( \sum_{r=1}^{M_{13}} |x_3[j-r]|^2 \Big) +
\left( |x_3[j]|^2 - |x_3[j-M_{13}]|^2 \right) K(t-j T_s)
\Big) .
\label{eq:phi13}
\end{align}
}\fi
By substituting (\ref{eq:phi11})--(\ref{eq:phi13}) in (\ref{eq:phi1_ct}), we get
\begin{align}
\phi_1(L,\tau) = \phi_1[j] + 2 \pi v_1[j] K(\tau-j T_s)
\end{align}
where
\ifdefined\twocolumnmode{
\begin{align}
\phi_1[j]
&= h_{11} |x_1[j]|^2
+ h_{12} \sum_{r=1}^{M_{12}} |x_2[j-r]|^2
+ h_{13} \sum_{r=1}^{M_{13}} |x_3[j-r]|^2 \label{eq:phi1j} \\
v_1[j]
&= h_{12} \left( |x_2[j]|^2 - |x_2[j-M_{12}]|^2 \right)/2\pi
\nonumber\\&
+ h_{13} \left( |x_3[j]|^2 - |x_3[j-M_{13}]|^2 \right)/2\pi \label{eq:v1j}\\
h_{11} &= \gamma_1 L \frac{E_s}{T_s} , \quad
h_{12} = 2 \gamma_1 L_{12} \frac{E_s}{T_s}, \quad
h_{13} = 2 \gamma_1 L_{13} \frac{E_s}{T_s}.
\label{eq:hcoeffs}
\end{align}
}\else{
\begin{align}
\phi_1[j]
&= h_{11} |x_1[j]|^2
+ h_{12} \sum_{r=1}^{M_{12}} |x_2[j-r]|^2
+ h_{13} \sum_{r=1}^{M_{13}} |x_3[j-r]|^2 \label{eq:phi1j} \\
v_1[j]
&= h_{12} \left( |x_2[j]|^2 - |x_2[j-M_{12}]|^2 \right)/2\pi
+ h_{13} \left( |x_3[j]|^2 - |x_3[j-M_{13}]|^2 \right)/2\pi \label{eq:v1j}\\
h_{11} &= \gamma_1 L \frac{E_s}{T_s} , \quad
h_{12} = 2 \gamma_1 L_{12} \frac{E_s}{T_s}, \quad
h_{13} = 2 \gamma_1 L_{13} \frac{E_s}{T_s}.
\label{eq:hcoeffs}
\end{align}
}\fi
Then by substituting in (\ref{eq:y1nj_noiseless_integral}), we have
\begin{align}
\ifdefined\twocolumnmode & \fi
\tilde{y}_{1,{\idx}}(j T_s+\beta_{11} L)
\ifdefined\twocolumnmode \nonumber\\ \fi
&= x_1[j] \ E_s \ e^{i \phi_1[j]} \ \int_{0}^{T_s} \frac{|p(\tau)|^2}{E_s} e^{i 2\pi(v_1[j]-\idx) K(\tau)} d\tau.
\label{eq:y1n_integral}
\end{align}
By applying Lemma \ref{lemma:leibniz} in Appendix \ref{sec:ortho_impulses}
to evaluate the integral in (\ref{eq:y1n_integral}),
the noiseless part $\tilde{y}_{1,{\idx}}[j]$ of the output of the filter with index ${\idx}$ at time $j$
can be written as
\begin{align}
&\tilde{y}_{1,{\idx}}[j] =
x_1[j] E_s e^{i\phi_1[j]} \ u_{1,{\idx}}[j]
\label{eq:y1nj_without_noise}
\end{align}
where
\begin{align}
u_{1,{\idx}}[j]
& = \left\{
\begin{array}{ll}
\displaystyle
\frac{\exp\left(i 2\pi(v_1[j] - {\idx})\right) - 1 }
{ i 2\pi(v_1[j] - {\idx}) } ,&\text{ if } v_1[j] \neq {\idx} \\
1 ,&\text{ otherwise}.
\end{array}
\right.
\label{eq:u1nj}
\end{align}
The output of the filter with index ${\idx}$ at time $j$ is
\begin{align}
y_{1,{\idx}}[j] = y_{1,{\idx}}(j T_s+\beta_{11} L) = \tilde{y}_{1,{\idx}}[j] + z_{1,{\idx}}[j]
\label{eq:y1nj}
\end{align}
where
\begin{align}
z_{1,{\idx}}[j] = \left. z_1(\T) \star h_{\idx}(\T) \right|_{\T=j T_s+\beta_{11} L}.
\label{eq:z1nj}
\end{align}
The variable $z_{1,{\idx}}[j]$ is Gaussian with mean 0 and variance $N E_s$.
Moreover, due to the orthogonality of the filter bank impulse responses, we have
$\mathbb{E}(z_{1,{\idx_1}}[j] z_{1,\idx_2}^*[j]) = 0$ for all ${\idx_1} \neq \idx_2$,
which implies that
the random variables $\{z_{1,{\idx}}[j]\}_{\idx \in \mathcal{F}_1}$ are independent.
\subsection{Discrete-Time Model}
\label{sec:dt_model}
To compute mutual information, we now consider the codeword $X_k^{\blk} = (X_{k}[1], X_{k}[2], \cdots, X_{k}[{\blk}])$
and the receiver samples $\Yv{k}^{\blk} = (\Yv{k}[1], \Yv{k}[2], \cdots, \Yv{k}[{\blk}])$ as random variables.
The input $X_k[j]$ of transmitter $k$ to the channel at time $j$ is a scalar,
whereas the channel output $\Yv{k}[j]$ at receiver $k$ at time $j$ is a vector
whose components are $Y_{k,{\idx}}[j]$, $f \in \mathcal{F}_k$.
The input-output relations are
\begin{align}
&Y_{k,{\idx}}[j] = X_k[j] ~ e^{ i \Phi_k[j] } ~ U_{k,{\idx}}[j] + Z_{k,{\idx}}[j]
\label{eq:model_yknj}
\end{align}
with
\ifdefined\twocolumnmode{
\begin{align}
\Phi_1[j]
&= h_{11} |X_1[j]|^2
+ h_{12} \sum_{r=1}^{M_{12}} |X_2[j-r]|^2
\nonumber\\&\
+ h_{13} \sum_{r=1}^{M_{13}} |X_3[j-r]|^2
\\
\Phi_2[j]
&= h_{21} \sum_{r=1}^{M_{12}} |X_1[j+M_{12}-r]|^2
+ h_{22} |X_2[j]|^2
\nonumber\\&\
+ h_{23} \sum_{r=1}^{M_{23}} |X_3[j-r]|^2
\\
\Phi_3[j]
&= h_{31} \sum_{r=1}^{M_{13}} |X_1[j+M_{13}-r]|^2
\nonumber\\&\
+ h_{32} \sum_{r=1}^{M_{23}} |X_2[j+M_{23}-r]|^2
+ h_{33} |X_3[j]|^2
\label{eq:Phi_kj}
\end{align}
}\else{
\begin{align}
\Phi_1[j]
&= h_{11} |X_1[j]|^2
+ h_{12} \sum_{r=1}^{M_{12}} |X_2[j-r]|^2
+ h_{13} \sum_{r=1}^{M_{13}} |X_3[j-r]|^2
\\
\Phi_2[j]
&= h_{21} \sum_{r=1}^{M_{12}} |X_1[j+M_{12}-r]|^2
+ h_{22} |X_2[j]|^2
+ h_{23} \sum_{r=1}^{M_{23}} |X_3[j-r]|^2
\\
\Phi_3[j]
&= h_{31} \sum_{r=1}^{M_{13}} |X_1[j+M_{13}-r]|^2
+ h_{32} \sum_{r=1}^{M_{23}} |X_2[j+M_{23}-r]|^2
+ h_{33} |X_3[j]|^2
\label{eq:Phi_kj}
\end{align}
}\fi
where $M_{12}$, $M_{13}$ and $M_{23}$ are positive integers and
\begin{align}
U_{k,{\idx}}[j]
& = \left\{
\begin{array}{ll}
\displaystyle
\frac{\exp\left(i 2\pi(V_k[j] - {\idx})\right) - 1 }
{ i 2\pi(V_k[j] - {\idx}) } ,&\text{ if } V_k[j] \neq {\idx} \\
1 ,&\text{ otherwise}
\end{array}
\right.
\label{eq:model_uknj}
\end{align}
where we define
\ifdefined\twocolumnmode{
\begin{align}
V_1[j] &\stackrel{\Delta}{=} h_{12}(|X_2[j] |^2 - |X_2[j-M_{12}]|^2)/ 2\pi
\nonumber\\&
+ h_{13}(|X_3[j] |^2 - |X_3[j-M_{13}]|^2)/ 2\pi \\
V_2[j] &\stackrel{\Delta}{=} h_{21}(|X_1[j+M_{12}]|^2 - |X_1[j] |^2)/ 2\pi
\nonumber\\&
+ h_{23}(|X_3[j] |^2 - |X_3[j-M_{23}]|^2)/ 2\pi \\
V_3[j] &\stackrel{\Delta}{=} h_{31}(|X_1[j+M_{13}]|^2 - |X_1[j] |^2)/ 2\pi
\nonumber\\&
+ h_{32}(|X_2[j+M_{23}]|^2 - |X_2[j] |^2)/ 2\pi.
\label{eq:model_vk}
\end{align}
}\else{
\begin{align}
V_1[j] &\stackrel{\Delta}{=} h_{12}(|X_2[j] |^2 - |X_2[j-M_{12}]|^2)/ 2\pi
+ h_{13}(|X_3[j] |^2 - |X_3[j-M_{13}]|^2)/ 2\pi \\
V_2[j] &\stackrel{\Delta}{=} h_{21}(|X_1[j+M_{12}]|^2 - |X_1[j] |^2)/ 2\pi
+ h_{23}(|X_3[j] |^2 - |X_3[j-M_{23}]|^2)/ 2\pi \\
V_3[j] &\stackrel{\Delta}{=} h_{31}(|X_1[j+M_{13}]|^2 - |X_1[j] |^2)/ 2\pi
+ h_{32}(|X_2[j+M_{23}]|^2 - |X_2[j] |^2)/ 2\pi.
\label{eq:model_vk}
\end{align}
}\fi
$Z_{k,{\idx}}[j]$ models the noise at filter $\idx$ of receiver $k$ at time $j$,
the random variables $\{Z_{k,{\idx}}[j]\}_{k,\idx,j}$ are independent circularly-symmetric complex Gaussian random variables with mean 0 and variance $N$.
We regard the $h_{k\ell}$ as \textit{channel coefficients} that are time invariant and known globally.
The following power constraints are imposed:
\begin{align} \label{eq:power-constraint-gvm}
\frac{1}{{\blk}} \sum_{j=1}^{{\blk}} \mathbb{E}\left[ |X_{k}[j]|^2 \right] \le P, \quad k=1,2,3.
\end{align}
A \textit{scheme} is a collection
$\{ (\mathcal{C}_1(P,N),\mathcal{C}_2(P,N),\mathcal{C}_3(P,N) )\}$ of triples of codes
such that at $(P,N)$, user $k$ uses the code $\mathcal{C}_k(P,N)$
that satisfies the power constraint and achieves an information rate $R_k(P,N)$ for $k=1,2,3$
where
\begin{align}
R_k(P,N) = I(X_k;\Yv{k}) \equiv \lim_{n \rightarrow \infty} \frac{1}{n} I(X_k^n;\Yv{k}^n).
\end{align}
We extend the definitions of pre-logs made in Definitions \ref{def:highp-prelog} and \ref{def:lowp-prelog}.
\begin{definition}
The \textit{high-power} pre-log triple $(\overline{r}_1, \overline{r}_2, \overline{r}_3)$ is achieved by a scheme
if the rates satisfy
\begin{equation}
\overline{r}_k(N) = \lim_{P \rightarrow \infty} \frac{R_k(P,N)}{\log(P/N)}
\text{ for } k=1,2,3.
\end{equation}
\end{definition}
\begin{definition}
The \textit{low-noise} pre-log triple $(\underline{r}_1, \underline{r}_2, \underline{r}_3)$ is achieved by a scheme
if the rates satisfy
\begin{equation}
\underline{r}_k(P) = \lim_{N \rightarrow 0} \frac{R_k(P,N)}{\log(P/N)}
\text{ for } k=1,2,3.
\end{equation}
\end{definition}
The (high-power or low-noise) pre-log triple $(1/2,1/2,1/2)$ can be achieved if all users use phase modulation only (see Sec. \ref{sec:inner-bound}).
It is not obvious whether $(1/2,1/2,1/2)$ is achievable by using amplitude modulation only.
We show next that the high-power pre-log triple $(1,1,1)$ can be achieved
for any positive $N$ through \textit{interference focusing}.
\subsection{Inner Bound: Phase Modulation}
\label{sec:inner-bound}
Suppose we use only the filter with index $f=0$.
Suppose further that the inputs $X_k^n$ of user $k$ are i.i.d.
with a constant amplitude $\sqrt{P}$ and a uniformly random phase (a ring), i.e., we have
\begin{align}
X_k[j] = \sqrt{P} e^{i \Phi_{X,k}[j]}
\end{align}
where $\Phi_{X,k}[j]$ is uniform on $[-\pi,\pi)$ for $j=1,2,\ldots,n$.
Therefore, the outputs become
\begin{align}
&Y_{k,0}[j] = X_k[j] ~ e^{ i \Phi_k[j] } ~ U_{k,0}[j] + Z_{k,0}[j]
\end{align}
with
\begin{align}
\Phi_1[j] &= [ h_{11} + h_{12} M_{12} + h_{13} M_{13} ] P \nonumber
\\
\Phi_2[j] &= [ h_{21} M_{12} + h_{22} + h_{23} M_{23} ] P \nonumber
\\
\Phi_3[j] &= [ h_{31} M_{13} + h_{32} M_{23} + h_{33} ] P,
\end{align}
i.e., the phase $\Phi_k[j]$ is constant for all $j=1,\ldots,n$.
Moreover, we have
\begin{align}
U_{1,0}[j] &= 1, \qquad \max\{M_{12},M_{13}\} < j \leq n \nonumber \\
U_{2,0}[j] &= 1, \qquad M_{23} < j < n-M_{12} \nonumber \\
U_{3,0}[j] &= 1, \qquad 1 \leq j < n-\max\{M_{13},M_{23}\}.
\end{align}
Thus, the users are decoupled under constant amplitude modulation, except near the beginning and the end of transmission.
We have
\begin{align}
\frac{1}{n} I(X_1^n;\Yv{1}^n)
&\stackrel{(a)}{\geq} \frac{1}{n} I(X_1^n;Y_{1,0}^n) \nonumber \\
&\stackrel{(b)}{\geq} \frac{1}{n} \sum_{j=1}^{n} I(X_1[j];Y_{1,0}[j]) \nonumber \\
&\stackrel{(c)}{\geq} \frac{1}{n} \sum_{j=\max\{M_{12},M_{13}\}+1}^{n} I(X_1[j];Y_{1,0}[j]) \nonumber \\
&\stackrel{(d)}{\geq} \left( \frac{n-\max\{M_{12},M_{13}\}}{n} \right) \left[ \frac{1}{2} \log\left(\frac{P}{N}\right) - 1\right]
\end{align}
where
$(a)$ follows from the chain rule and the non-negativity of mutual information,
$(b)$ follows because $X_1,\ldots,X_n$ are i.i.d. and because conditioning does not increase entropy,
$(c)$ follows from the non-negativity of mutual information and
$(d)$ holds because (see Theorem \ref{theorem:one-ring})
\begin{align}
I(X_1[j];Y_{1,0}[j]) \geq \frac{1}{2} \log\left(\frac{2 P}{N}\right) - 1.
\end{align}
As $n \rightarrow \infty$, we have
\begin{align}
R_1^\textsf{ring}(P,N)
&\geq \frac{1}{2} \log\left(\frac{2 P}{N}\right) - 1.
\end{align}
By using similar steps for users 2 and 3, we have
\begin{align}
R_k^\textsf{ring}(P,N)
&\geq \frac{1}{2} \log\left(\frac{2 P}{N}\right) - 1
\end{align}
for $k=1,2,3$ which implies that the pre-log triple is $(1/2,1/2,1/2)$ is achieved simply by using one receiver filter and phase modulation.
\subsection{Interference Focusing}
\label{sec:ifocus}
We use \textit{interference focusing}, \textit{i.e.}, we focus the phase interference on \textit{one} point
by imposing the following constraints on the transmitted symbols:
\begin{align}
h_{21} |X_1[j]|^2 = 2\pi \tilde{n}_{21},\ h_{31} |X_1[j]|^2 = 2\pi \tilde{n}_{31}, \\
h_{12} |X_2[j]|^2 = 2\pi \tilde{n}_{12},\ h_{32} |X_2[j]|^2 = 2\pi \tilde{n}_{32}, \\
h_{13} |X_3[j]|^2 = 2\pi \tilde{n}_{13},\ h_{23} |X_3[j]|^2 = 2\pi \tilde{n}_{23},
\end{align}
where
$\tilde{n}_{21}$, $\tilde{n}_{31}$,
$\tilde{n}_{12}$, $\tilde{n}_{32}$,
$\tilde{n}_{13}$ and $\tilde{n}_{23} \in \mathbb{N}$,
which ensures that the XPM interference is eliminated.
Suppose that
$h_{21}$, $h_{31}$,
$h_{12}$, $h_{32}$,
$h_{13}$ and $h_{23}$
are rational.
Then the interference focusing constraints become
\begin{align}
|X_k[j]|^2 = 2\pi \hat{p}_k ~ \tilde{n}_{k}
\end{align}
where
\begin{align}
\hat{p}_1 \stackrel{\Delta}{=} \text{lcm}(\text{den}(h_{21}),\text{den}(h_{31})), \label{eq:p1_hat}\\
\hat{p}_2 \stackrel{\Delta}{=} \text{lcm}(\text{den}(h_{12}),\text{den}(h_{32})), \label{eq:p2_hat}\\
\hat{p}_3 \stackrel{\Delta}{=} \text{lcm}(\text{den}(h_{13}),\text{den}(h_{23})). \label{eq:p3_hat}
\end{align}
where $\text{den}(x)$ is the denominator of a rational number $x$.
Because of the power constraint,
only a subset $\mathcal{P}_k$ of the allowed rings is actually used.
In this case, $V_k[j] \in \mathcal{V}_k$, for $k=1,2,3$, where
\begin{align}
\mathcal{V}_k = \left\{ \sum_{j \neq k} d_j: d_j \in \mathcal{D}_j, j \in\{1,2,3\} \right\}
\end{align}
and
\begin{align}
\mathcal{D}_k = \left\{\frac{p-p^\prime}{\hat{p_k}}: p \in \mathcal{P}_k, p^\prime \in \mathcal{P}_k \right\}
\end{align}
which leads us to choose the sets of ``normalized frequencies'' $\mathcal{F}_k$
of the filter banks at the receivers as
$\mathcal{F}_k = \mathcal{V}_k$.
Thus, under interference focusing, the output at receiver $k$ at time $j$ is a vector $\Yv{k}[j]$,
whose components are $\{Y_{k,{\idx}}[j]\}_{{\idx} \in \mathcal{V}_k}$,
where
\begin{align}
Y_{k,{\idx}}[j] = X_k[j] \ \exp\left( i h_{kk} |X_k[j]|^2 \right) U_{k,{\idx}}[j] + Z_{k,{\idx}}[j],
\label{eq:ykn}
\end{align}
and where
\begin{align}
U_{k,{\idx}}[j]
= \left\{ \begin{array}{ll}
1 ,&\text{ if } V_k[j] = {\idx}, \\
0 ,&\text{ otherwise.}
\end{array} \right.
\label{eq:u_kn_ifocus}
\end{align}
This means that
exactly one filter (the filter with index $V_k[j]$) output among all the filters contains the signal corrupted by noise,
while all other filters put out noise.
Therefore, we have
\begin{align}
&\frac{1}{n} I(X_k^n;\Yv{k}^n) \nonumber\\
&\stackrel{(a)}{\geq}
\frac{1}{n} \sum_{j=1}^n I(X_k[j];\Yv{k}[j]) \nonumber\\
&\stackrel{(b)}{\geq}
\frac{1}{n} \sum_{j=1}^n I\left(X_k[j]; Y_{k,V_k[j]}[j] \right) \nonumber\\
&\stackrel{(c)}{=} I\left( X_k[1];X_k[1] e^{i h_{kk}|X_k[1]|^2}+Z_{k,V_k[1]}[1] \right)
\end{align}
where
$(a)$ follows because $X_1^n$ are i.i.d. and because conditioning does not increase entropy;
$(b)$ follows from the chain rule and the non-negativity of mutual information
(it can be shown that equality holds, see Appendix \ref{sec:sufficient})
; and
$(c)$ holds because $X_1^n$ are i.i.d. and the channel becomes a memoryless time-invariant channel under interference focusing.
It follows from Theorem \ref{theorem:multi-ring} that by using interference focusing,
we have
\begin{align}
\lim_{P \rightarrow \infty} \frac{I\left( X_k[1];X_k[1] e^{i h_{kk}|X_k[1]|^2}+Z_{k,V_k[1]}[1] \right)}{\log(P/N)}
\geq 1
\end{align}
which implies that $\overline{r}_k \geq 1$ for $k=1,2,3$.
Hence, the high-power pre-log triple $(1,1,1)$ is achievable.
We remark that the question of whether all users can simultaneously achieve a low-noise pre-log of 1 is open for both models with and without GVM.
The following example illustrates our receiver structure,
and the role that interference focusing plays in choosing its parameters.
\textit{Example:}
Consider 2 transmitters that use a rectangular pulse, i.e.,
\begin{align}
p(t) = \left\{
\begin{array}{ll}
\displaystyle \sqrt{E_s/T_s} , & 0 \leq t < T_s \\
0, & \text{otherwise}.
\end{array}
\right.
\label{eq:square_pulse_def}
\end{align}
Suppose that
$h_{12} = 5$,
$h_{21} = 4$,
$P_1 = 8$,
and
$P_2 = 7$.
Since this is a two-user system, we may use
(\ref{eq:ring-constraint-1}) and (\ref{eq:ring-constraint-2}) rather than (\ref{eq:p1_hat}) and (\ref{eq:p2_hat}), i.e.,
we use $\hat{p}_1 = 1/h_{21} = 0.25$ and $\hat{p}_2 = 1/h_{12} = 0.2$.
Suppose that the users choose
the power levels
$\mathcal{P}_1 = \{2\pi\hat{p}_{1} \tilde{n}_1: \tilde{n}_1 = 1,4,9\} = \{ 0.5\pi, 2\pi, 4.5\pi \}$
and
$\mathcal{P}_2 = \{2\pi\hat{p}_{2} \tilde{n}_2: \tilde{n}_2 = 2,8\} = \{ 0.8\pi, 3.2\pi \}$
(see Fig. \ref{fig:constell}).
These choices satisfy the power constraints
and eliminate the interference.
The parameters of the filter bank are
$\mathcal{F}_1 = \mathcal{V}_1 = \{-6,0,6\}$ and
$\mathcal{F}_2 = \mathcal{V}_2 = \{-8,-5,-3,0,3,5,8\}$.
In other words, receiver 1 has 3 filters
whose frequency responses are sinc functions
centered at $f_1-6/T_s$, $f_1$, and $f_1+6/T_s$,
whereas
receiver 2 has 7 filters
whose frequency responses are sinc functions
centered at 7 different frequencies
(see Fig. \ref{fig:filter_bank}).
This shows that, because of the non-linearity,
the receivers need to extract information from a ``bandwidth'' larger than
the ``bandwidth'' of the transmitted signal.
\begin{figure}
\centering
\includegraphics[width= {\ifdefined\twocolumnmode 0.49\columnwidth \else 0.36\textwidth \fi} ]{user1_constell}
\includegraphics[width= {\ifdefined\twocolumnmode 0.49\columnwidth \else 0.36\textwidth \fi} ]{user2_constell}
\caption{Ring modulation used by transmitter 1 (left) and transmitter 2 (right).
The thin lines are the rings allowed by interference focusing,
and the thick blue lines are the rings selected for transmission.}
\label{fig:constell}
\end{figure}
\begin{figure}
\centering
\includegraphics[width= {\ifdefined\twocolumnmode 1.0\columnwidth \else 0.75\textwidth \fi} ]{user1_bank} \\
\includegraphics[width= {\ifdefined\twocolumnmode 1.0\columnwidth \else 0.75\textwidth \fi} ]{user2_bank}
\caption{Frequency responses of the filters at receivers 1 (top) and 2 (bottom).}
\label{fig:filter_bank}
\end{figure}
\subsection{Outer Bound}
\subsubsection{Interference Focusing}
We show next that the maximal pre-log triple for the model of Sec. \ref{sec:dt_model} is $(1,1,1)$
when interference focusing is used.
\begin{align}
I( X_1^n ; \Yv{1}^n )
&\stackrel{(a)}{\leq}
I( X_1^n ; \Yv{1}^n , V_1^n) \nonumber\\
&\stackrel{(b)}{=}
I( X_1^n ; \Yv{1}^n | V_1^n) \nonumber\\
&\stackrel{(c)}{=}
I( X_1^n ; Y_{1,V_1[1]}, Y_{1,V_1[2]}, \ldots, Y_{1,V_1[n]} | V_1^n) \nonumber\\
&\stackrel{(d)}{=}
I( X_1^n ; Y_{1,V_1[1]}, Y_{1,V_1[2]}, \ldots, Y_{1,V_1[n]} ) \nonumber\\
&{\leq} n \log\left( 1 + \frac{P}{N} \right).
\end{align}
Step $(a)$ follows from the chain rule and the non-negativity of mutual information;
$(b)$ holds because $X_1^n$ is independent of $V_1^n$;
$(c)$ holds because $Y_{1,f}[j] = Z_{1,f}[j]$ for $f \neq V_1[j]$
and the random variables $\{ Z_{1,f}[j] : j=1,\ldots,n \text{ and } f \neq V_1[j] \}$ are independent of
$Y_{1,V_1[1]}, Y_{1,V_1[2]}, \ldots, Y_{1,V_1[n]}$ and $X_1^n$ and
$(d)$ holds because $X_1^n$ and $Y_{1,V_1[1]}, Y_{1,V_1[2]}, \ldots, Y_{1,V_1[n]}$ are independent of $V_1^n$
(which follows from $Z_{1,V_1[1]}, Z_{1,V_1[2]}, \ldots, Z_{1,V_1[n]}$ being i.i.d.).
By using a similar argument for receiver 2 and receiver 3, we eventually have
\begin{align}
R_k \leq \log\left( 1 + \frac{P}{N} \right)
\end{align}
for $k=1,2,3$ which implies that the maximal pre-log triple is $(1,1,1)$.
\subsubsection{General Modulation}
We show next that the maximal pre-log triple is $(1,1,1)$
for any modulation scheme.
We use a genie-aided strategy. Suppose a genie reveals the codewords $x_2^n$ and $x_3^n$ of users 2 and 3 to receiver 1 prior to transmission.
Receiver 1 generates $\phi_{12}(t)$ and $\phi_{13}(t)$ according to (\ref{eq:phi12_definition}) and (\ref{eq:phi13_definition}), respectively,
and use them to cancel XPM in the received signal, i.e., receiver 1 generates
\begin{align}
\tilde{r}_1(t) = r_1(t) ~ e^{-j \phi_{12}(t) -j \phi_{13}(t)}.
\end{align}
The XPM-free signal $\tilde{r}(t)$ is fed to a filter with an impulse response $p^*(-t)$ and the output of the filter is sampled at symbol rate.
Matched filtering with symbol rate sampling does not incur any information loss because XPM is canceled.
The $j$-th filter output is
\begin{align}
\tilde{y}_1[j]
&= \left. \tilde{r}_1(t) \star p^*(-t) \right|_{t=j \Ts} \\
&= x_1[j] e^{j h_{11} |x_1[j]|^2} + \tilde{z}_1[j]
\label{eq:ytilde_genie}
\end{align}
where $\tilde{z}_1[j]$ is a realization of a Gaussian random variable with mean zero and variance $N$.
The channel (\ref{eq:ytilde_genie}) is a memoryless AWGN channel and therefore we have
\begin{align}
I( X_1^n ; \tilde{Y}_1[1], \ldots, \tilde{Y}_1[n])
&{\leq} n \log\left( 1 + \frac{P}{N} \right).
\end{align}
Similarly, it can be shown that the maximum pre-log for users 2 and 3 is 1
implying that the maximal pre-log triple is $(1,1,1)$.
\section{Conclusion}
\label{sec:conclusion}
We introduced 2 discrete-time interference channel models
based on a simplified optical fiber model.
We used coupled NLS equations to develop our models.
In the first model there was no dispersion.
The non-linear nature of the fiber-optic medium causes the users to suffer from amplitude-dependent phase interference.
We introduced a new technique called interference focusing that lets the users take full advantage of all the available amplitude and phase degrees of freedom
at high transmission powers.
In the second model, the second-order dispersion is negligible.
However, we included non-zero group velocity mismatch
as well as non-linearity.
We showed that our discrete-time model is justified
by using square pulse shaping at the transmitters
and a bank of frequency-shifted matched filters at the receivers.
We proved that both users can achieve a high-power pre-log of 1 simultaneously
by using interference focusing,
thus exploiting all the available amplitude and phase degrees of freedom.
\appendices
\section{Upper Bound on the Modified Bessel Function of the First Kind}
\begin{lemma} \label{lemma:besseli_lb}
\begin{align}
I_0(z) \leq \frac{e^z}{\sqrt{z}}, \quad z > 0
\end{align}
\end{lemma}
\begin{proof}
We have
\begin{align}
&\cos{\theta} \leq 1 - 4 \theta^2/\pi^2, & 0 \leq \theta \leq \pi/2 \label{eq:cos_ub_domain1}\\
&\cos{\theta} \leq 0 , & \pi/2 \leq \theta \leq \pi \label{eq:cos_ub_domain2}
\end{align}
where (\ref{eq:cos_ub_domain1})
follows by using the infinite product form for cosine
\begin{align}
\cos x = \prod_{n=1}^\infty \left[ 1- \frac{4x^2}{\pi^2 (2 n -1)^2}\right].
\end{align}
We thus have
\begin{align}
I_0(z)
& = \frac{1}{\pi} \int_{0}^{\pi} e^{z \cos\theta} \, d\theta \nonumber \\
& \stackrel{(a)}{\le} \frac{1}{\pi} \left[ \int_{0}^{\pi/2} e^{z (1- 4 \theta^2/\pi^2)} \, d\theta + \int_{\pi/2}^{\pi} \, d\theta \right] \nonumber \\
& = \frac{\sqrt{\pi}}{4} \frac{e^z}{\sqrt{z}} \left( 1 - 2 Q(\sqrt{2z}) \right) + \frac{1}{2} \nonumber \\
& \stackrel{(b)}{\le} \frac{\sqrt{\pi}}{2} \frac{e^z}{\sqrt{z}} \nonumber \\
&\leq \frac{e^z}{\sqrt{z}}
\end{align}
where
\begin{align}
Q(z) = \int_{z}^{\infty} \frac{1}{\sqrt{2 \pi}} e^{-x^2/2} dx.
\end{align}
Step ($a$) follows because the exponential function is a monotonic increasing function and
(\ref{eq:cos_ub_domain1})--(\ref{eq:cos_ub_domain2})
while
($b$) holds because $Q(z) \ge 0$ and
\begin{align}
\frac{\sqrt{\pi}}{4} \frac{e^z}{\sqrt{z}}
\geq \frac{\sqrt{\pi}}{4} \min_{z \geq 0} \frac{e^z}{\sqrt{z}}
= \frac{\sqrt{2 e \pi}}{4}
\ge \frac{1}{2}.
\end{align}
\end{proof}
\section{Expected Value of the Logarithm of a Rician R.V.}
Consider the following functions.
\begin{itemize}
\item
Gamma function $\Gamma(z)$ \cite[6.1.1]{Abramowitz1972}
\begin{align}
\Gamma(z) = \int_{0}^{\infty} t^{z-1} e^{-t} dt
\label{eq:gamma-func-def}
\end{align}
\item
Psi (Digamma) function $\psi(z)$ \cite[6.3.1]{Abramowitz1972}
\begin{align}
\psi(z) =\frac{d}{dz} \ln\Gamma(z) = \frac{ \Gamma^\prime(z) }{\Gamma(z)}
\end{align}
\item
Upper incomplete Gamma function $\Gamma(a,x)$ \cite[6.5.3]{Abramowitz1972}
\begin{align}
\Gamma(a,x) = \int_{x}^{\infty} t^{a-1} e^{-t} dt, \qquad a>0
\label{eq:upper-incomplete-gamma}
\end{align}
\end{itemize}
We derive several useful lemmas concerning these functions.
\begin{lemma} \label{lemma:integral-lemma1}
\begin{align}
\int_{0}^\infty {e^{- \frac{x^2}{2} } x^{2k+1} \ln(x) dx}
= 2^{k-1} \left( \Gamma(k+1) \ln(2) + \Gamma^\prime(k+1) \right)
\end{align}
\end{lemma}
\begin{proof}
Consider
\begin{align}
\mathcal{I}_k
&\stackrel{\Delta}{=} \int_{x=0}^\infty {e^{- \frac{x^2}{2} } x^{2k+1} \ln(x) dx} \nonumber \\
&\stackrel{(a)}{=} \int_{u=0}^\infty {e^{- u }(\sqrt{2u})^{2k+1} \ln(\sqrt{2u}) \frac{1}{\sqrt{2u} } du} \nonumber \\
&= \int_{u=0}^\infty {e^{- u }(2u)^{k} \frac{1}{2} \left(\ln(2) + \ln(u)\right) du} \nonumber \\
&= 2^{k-1} \ln(2) \int_{u=0}^\infty {e^{-u} u^{k} du}
+ 2^{k-1} \int_{u=0}^\infty {e^{- u }u^{k} \ln(u) du} \nonumber \\
&\stackrel{(b)}{=} 2^{k-1} \left( \Gamma(k+1) \ln(2) + \Gamma^\prime(k+1) \right)
\end{align}
where ($a$) follows from the transformation of variables $u = x^2/2$
and $(b)$ follows by (\ref{eq:gamma-func-def}) and \cite[4.352 (4)]{Gradshtein2007}
\begin{align}
\int_{u=0}^\infty {e^{- u }u^{k} \ln(u) du} = \Gamma^\prime(k+1).
\end{align}
\end{proof}
\begin{lemma} \label{lemma:integral-lemma2}
\begin{align}
\sum_{k=0}^\infty \frac{t^k}{k!} \psi(k+1) = e^t ( \Gamma(0,t)+\ln(t) )
\end{align}
\end{lemma}
\begin{proof}
We use the following formula \cite[6.2.1 (60)]{Brychkov2008}
\begin{align}
\sum_{k=1}^\infty \frac{t^k}{k!} \psi(k+a)
&= \frac{t}{a} e^t \bigg[ a \psi(a) \frac{1-e^{-t}}{t} \nonumber\\&\qquad
+ ~ _2F_2(1,1;a+1,2;-t) \bigg]
\label{eq:psi_series_a}
\end{align}
where $_2F_2(a_1,a_2;b_1,b_2;x)$ is the generalized hypergeometric function defined as
\cite[9.14 (1)]{Gradshtein2007},\cite[p. 674]{Brychkov2008}
\begin{align}
_2F_2(a_1,a_2;b_1,b_2;x) = \sum_{k=0}^\infty \frac{(a_1)_k (a_2)_k}{(b_1)_k (b_2)_k} \frac{x^k}{k!}
\label{eq:hypergeometric_2F2}
\end{align}
where $(f)_k \stackrel{\Delta}{=} f(f+1)\ldots (f+k-1)$.
Setting $a=1$ in (\ref{eq:psi_series_a}) gives
\begin{align}
\sum_{k=1}^\infty \frac{t^k}{k!} \psi(k+1)
& = t e^t \left[ \psi(1) \frac{1-e^{-t}}{t} + ~ _2F_2(1,1;2,2;-t) \right] \nonumber \\
& = e^t \left[ F(t) + (1-e^{-t}) \psi(1) \right]
\label{eq:psi_series_1}
\end{align}
where we defined $F(t)$ as
\begin{align}
F(t)
&\stackrel{\Delta}{=} t \cdot {_2F_2}(1,1;2,2;-t) \nonumber \\
&= t ~ \sum_{k=0}^\infty \frac{1}{(k+1)^2} \frac{(-t)^k}{k!} \nonumber \\
&= \sum_{m=1}^\infty \frac{-1}{m} \frac{(-t)^{m}}{m!}.
\label{eq:Ft_def}
\end{align}
From the Fundamental Theorem of Calculus, we have
\begin{align}
\int_{0}^{z} F^\prime(t) ~ dt = F(z) - F(0)
\label{eq:theorem_of_calculus}
\end{align}
where
\begin{align}
F^\prime(t)
\stackrel{\Delta}{=} \frac{dF(t)}{dt}
= \sum_{m=1}^\infty \frac{(-t)^{m-1}}{m!}
= \frac{1-e^{-t}}{t}
\end{align}
and therefore the left-hand side of (\ref{eq:theorem_of_calculus}) is
\begin{align}
\int_{0}^{z} \frac{1-e^{-t}}{t} dt
&=
\int_{z}^{\infty} \frac{e^{-t}}{t} dt
+ \int_{1}^{z} \frac{1}{t} dt
\nonumber\\& \qquad
+
\left(
\int_{0}^{1} \frac{1-e^{-t}}{t} dt
- \int_{1}^{\infty} \frac{e^{-t}}{t} dt
\right) \nonumber \\
&= \Gamma(0,z) + \ln(z) + \gamma
\end{align}
where we used the definition of the upper incomplete Gamma function
and the following integral form for Euler's constant
\cite[8.367 (12)]{Gradshtein2007}
\begin{align}
\gamma = \int_0^1 \frac{1-e^{-t}}{t} dt - \int_1^\infty \frac{e^{-t}}{t} dt.
\end{align}
Since $F(0)=0$ and $\psi(1) = -\gamma$ \cite[8.367 (1)]{Gradshtein2007}, we have
\begin{align}
F(z) = \Gamma(0,z) + \ln(z) - \psi(1).
\label{eq:Fz_close_form}
\end{align}
The lemma follows from (\ref{eq:psi_series_1}) and (\ref{eq:Fz_close_form}).
\end{proof}
\begin{lemma} \label{lemma:integral-lemma3}
\begin{align}
\int_{0}^\infty {x e^{- \frac{x^2+\nu^2}{2} } I_0(x\nu) \ln(x) dx}
= \frac{1}{2} \left( \Gamma\left( 0,\frac{\nu^2}{2} \right) + \ln(\nu^2) \right)
\end{align}
\end{lemma}
\begin{proof}
We compute
\begin{align}
& \int_{0}^\infty {x e^{- \frac{x^2+\nu^2}{2} } I_0(x\nu) \ln(x) dx} \nonumber \\
& \stackrel{(a)}{=}
\int_{0}^\infty {x e^{- \frac{x^2+\nu^2}{2} } \left( \sum_{k=0}^{\infty} \frac{(x^2 \nu^2/4)^k}{(k!)^2} \right) \ln(x) dx} \nonumber\\
& = \sum_{k=0}^{\infty} \frac{1}{4^k (k!)^2}
\left( \int_{0}^\infty {x e^{- \frac{x^2+\nu^2}{2} }(x \nu)^{2k} \ln(x) dx} \right) \nonumber\\
& \stackrel{(b)}{=}
\sum_{k=0}^{\infty} \frac{1}{4^k (k!)^2}
2^{k-1} e^{- \frac{\nu^2}{2} } \nu^{2k} ( \Gamma(k+1) \ln(2) + \Gamma^\prime(k+1) ) \nonumber\\
& \stackrel{(c)}{=}
e^{ -\frac{\nu^2}{2} } \left[
\frac{\ln(2)}{2} \sum_{k=0}^{\infty} \frac{(\nu^2/2)^k}{k!}
+ \frac{1 }{2} \sum_{k=0}^{\infty} \frac{(\nu^2/2)^k}{k!} \psi(k+1)
\right] \nonumber \\
& \stackrel{(d)}{=}
e^{ -\frac{\nu^2}{2} } \left[
\frac{\ln(2)}{2} e^{ \frac{\nu^2}{2} }
+ \frac{1 }{2} e^{ \frac{\nu^2}{2} } \left( \Gamma\left( 0,\frac{\nu^2}{2} \right) + \ln\left( \frac{\nu^2}{2} \right) \right)
\right] \nonumber \\
& = \frac{1}{2} \left( \Gamma\left( 0,\frac{\nu^2}{2} \right) + \ln(\nu^2) \right)
\end{align}
where in ($a$) we used the series representation of $I_0(\cdot)$
\cite[9.6.10]{Abramowitz1972}
\begin{align}
I_0(x) = \sum_{k=0}^{\infty} \frac{(x^2/4)^k}{(k!)^2}.
\end{align}
Step
($b$) follows from Lemma \ref{lemma:integral-lemma1},
($c$) follows because \cite[6.1.6]{Abramowitz1972}
\begin{align}
\Gamma(k+1) = k!
\end{align}
and $\psi(k+1) = \Gamma^\prime(k+1)/\Gamma(k+1)$ and
($d$) follows by Lemma \ref{lemma:integral-lemma2}.
\end{proof}
\section{Minimum-Distance Estimator}
Let $Y=X+Z$ where $Z$ is a circularly-symmetric complex Gaussian random variable with mean $0$ and variance $N$.
Suppose $X_A \stackrel{\Delta}{=} |X| \in \mathcal{X}_A = \{\sqrt{\textsf{P}_j}:j=1,\ldots,J\}$
where $0 < \textsf{P}_1 < \textsf{P}_2 <\ldots <\textsf{P}_J$.
Define the minimum-distance estimator $\hat{X}_A$ as
\begin{align}
\hat{X}_A = \arg \min_{x_A \in \mathcal{X}_A} |Y_A-x_A|
\end{align}
where $Y_A=|Y|$.
\begin{lemma} \label{lemma:min-distance-estimator}
The probability of error for uniformly distributed $X_A$ satisfies
\begin{align}
P_e
\stackrel{\Delta}{=} \Pr[\hat{X}_A \ne X_A]
\leq \frac{2}{J} \sum_{j=2}^{J} \exp\left(-\frac{\Delta_j^2}{4}\right)
\end{align}
where $\Delta_j = (\sqrt{\textsf{P}_j} - \sqrt{\textsf{P}_{j-1}})/\sqrt{N}$.
\end{lemma}
\begin{proof}
Let $P_{e,j}$ be the error probability when $X_A=\sqrt{\textsf{P}_j}$. We have $P_e=\sum_{j=1}^J (1/J) P_{e,j}$ and
\begin{align}
P_{e,j} & = \left\{ \begin{array}{ll}
\Pr\left(Y_A \ge \frac{\sqrt{\textsf{P}_1}+\sqrt{\textsf{P}_{2}}}{2} \right), & j=1 \\
\Pr\left(Y_A \le \frac{\sqrt{\textsf{P}_{K-1}}+\sqrt{\textsf{P}_K}}{2} \right), & j=J \\
\Pr\left(Y_A \le \frac{\sqrt{\textsf{P}_{j-1}}+\sqrt{\textsf{P}_j}}{2} \right) & \\
\quad + \Pr\left(Y_A \ge \frac{\sqrt{\textsf{P}_j}+\sqrt{\textsf{P}_{j+1}}}{2} \right), & \text{otherwise}.
\end{array} \right.
\end{align}
Conditioned on $X_A=\sqrt{\textsf{P}_j}$, $Y_A$ is a Ricean random variable, and hence we compute \cite[p. 50]{ProakisSalehi2008}
\begin{align} \label{eq:interval-error}
\Pr\left(Y_A \ge \frac{\sqrt{\textsf{P}_j}+\sqrt{\textsf{P}_{j+1}}}{2} \right)
= Q\left(\frac{\sqrt{\textsf{P}_j}}{\sqrt{N/2}},\frac{\sqrt{\textsf{P}_j}+\sqrt{\textsf{P}_{j+1}}}{2 \sqrt{{N/2}} }\right)
\end{align}
where $Q(a,b)$ is the Marcum Q-function \cite{Corazza2002}.
Consider the following bounds.
\begin{itemize}
\item Upper bound for $b>a$ \cite[UB1MG]{Corazza2002}
\begin{align}
Q(a,b) \le \exp\left(-\frac{(b-a)^2}{2}\right) .
\label{eq:marcum_q_ub}
\end{align}
\item Lower bound for $b<a$ \cite[LB2aS]{Corazza2002}
\begin{align}
& Q(a,b) \nonumber \\
& \ge 1 - \frac{1}{2} \left[ \exp\left(-\frac{(a-b)^2}{2}\right) - \exp\left(-\frac{(a+b)^2}{2}\right) \right] .
\label{eq:marcum_q_lb}
\end{align}
\end{itemize}
The bound (\ref{eq:marcum_q_lb}) implies
\begin{align}
1 - Q(a,b) \le \exp\left(-\frac{(a-b)^2}{2}\right) .
\label{eq:marcum_q_lb_loosened}
\end{align}
We use \eqref{eq:interval-error} and \eqref{eq:marcum_q_ub} to write
\begin{align}
\Pr\left(Y_A \ge \frac{\sqrt{\textsf{P}_j}+\sqrt{\textsf{P}_{j+1}}}{2} \right)\
\le \exp\left(-\frac{\Delta_{j+1}^2}{4}\right) .
\end{align}
where $\Delta_j = (\sqrt{\textsf{P}_j} - \sqrt{\textsf{P}_{j-1}})/\sqrt{N}$. Similarly, we use inequality \eqref{eq:marcum_q_lb_loosened} to write
\begin{align}
\Pr\left(Y_A \le \frac{\sqrt{\textsf{P}_{j-1}}+\sqrt{\textsf{P}_j}}{2} \right)
& \le \exp\left(-\frac{\Delta_j^2}{4}\right) .
\end{align}
Collecting our results, we have
\begin{align}
P_e
& \leq \frac{1}{J} \left[
\exp\left(-\frac{\Delta_2^2}{4}\right)
+ \sum_{j=2}^{J-1} \exp\left(-\frac{\Delta_j^2}{4}\right)
\right. \nonumber \\ & \left.
\qquad + \sum_{j=2}^{J-1} \exp\left(-\frac{\Delta_{j+1}^2}{4}\right)
+ \exp\left(-\frac{\Delta_J^2}{4}\right)
\right] \nonumber \\
& = \frac{2}{J} \sum_{j=2}^{J} \exp\left(-\frac{\Delta_j^2}{4}\right).
\label{eq:Pe-bound}
\end{align}
\end{proof}
\section{Orthogonality of Impulse Responses}
\label{sec:ortho_impulses}
First, we introduce a useful lemma.
\begin{lemma}
\label{lemma:leibniz}
For any complex number $B$, we have
\begin{align}
\int_{0}^{T_s} \frac{|p(t)|^2}{E_s} e^{B K(\tau)} d\tau
= \left\{
\begin{array}{ll}
(e^B-1)/B, & \text{if } B \neq 0 ,\\
1, & \text{if } B = 0
\end{array}
\right.
\end{align}
where $K(\cdot)$ is defined in (\ref{eq:rise_func}).
\end{lemma}
\begin{proof}
\begin{itemize}
\item
For $B = 0$, we have
\begin{align}
\int_{0}^{T_s} \frac{|p(t)|^2}{E_s} e^{B K(\tau)} d\tau
= \int_{0}^{T_s} \frac{|p(t)|^2}{E_s} d\tau
= 1
\end{align}
where the last equality follows from (\ref{eq:nonzerogvm-pulse-energy}).
\item
For $B \neq 0$, we have
\begin{align}
\int_{0}^{T_s} \frac{|p(t)|^2}{E_s} e^{B K(\tau)} d\tau
&\stackrel{(a)}{=} \frac{1}{B} \int_{0}^{T_s} B K^\prime(\tau) e^{B K(\tau)} d\tau \nonumber\\
&\stackrel{(b)}{=} \frac{e^{B K(T_s)} - e^{B K(0)}}{B} \nonumber\\
&\stackrel{(c)}{=} \frac{e^{B} - 1}{B}
\end{align}
where
($a$) follows by applying Leibniz rule for differentiation under the integral sign \cite[3.3.7]{Abramowitz1972}:
\begin{align}
K^\prime(t) = \frac{1}{E_s} \frac{d}{d t} \int_0^t |p(\lambda)|^2 d\lambda = \frac{1}{E_s} |p(t)|^2;
\end{align}
($b$) is obtained through integration by substitution and
($c$) holds because $K(0)=0$ and $K(T_s)=1$.
\end{itemize}
\end{proof}
Next we show that the impulse responses
$\{h_{\idx}(\T)\}_{\idx \in \mathcal{F}_1}$ are orthogonal (cf. (\ref{eq:ortho_diff})).
For $\idx_1 \neq \idx_2$, we have
\begin{align}
\ifdefined\twocolumnmode & \fi
\int_{-\infty}^{\infty} h_{\idx_1}(\xi) h^*_{\idx_2}(\xi) d\xi
\ifdefined\twocolumnmode \nonumber\\ \fi
&\stackrel{(a)}{=} \int_{-\infty}^{\infty} |p(-\xi)|^2 \exp(-i 2\pi (\idx_1 - \idx_2) K(-\xi)) \ d\xi \nonumber \\
&\stackrel{(b)}{=} \int_{-\infty}^{\infty} |p(t)|^2 \exp(-i 2\pi (\idx_1 - \idx_2) K(t)) \ dt \nonumber \\
&\stackrel{(c)}{=} \int_{0}^{\Ts} |p(t)|^2 \exp(-i 2\pi (\idx_1 - \idx_2) K(t)) \ dt \nonumber \\
&\stackrel{(d)}{=} 0
\end{align}
where
($a$) follows from $K^*(\xi) = K(\xi)$,
($b$) follows from the transformation of variables $t = -\xi$,
($c$) holds because $p(t)=0$ for $t \notin [0,\Ts]$ and
($d$) follows from Lemma \ref{lemma:leibniz} with $B=-i 2 \pi (\idx_1 - \idx_2) \neq 0$.
\section{Solution to Coupled Partial Differential Equations}
\label{sec:coupled-pdes}
We want to solve the coupled equations
\begin{align}
& i \frac{\partial A_1}{\partial z} + i \beta_{11} \frac{\partial A_1}{\partial t}
+ \gamma_1 (|A_1|^2 + 2 |A_2|^2) A_1 = 0
\label{eq:coupled-pde1} \\
& i \frac{\partial A_2}{\partial z} + i \beta_{12} \frac{\partial A_2}{\partial t}
+ \gamma_2 (|A_2|^2 + 2 |A_1|^2) A_2 = 0.
\label{eq:coupled-pde2}
\end{align}
First, we introduce the following lemma.
\begin{lemma} \label{lemma:nonhomog-pde}
The solution to the non-homogeneous\footnote{The homogeneous (i.e., $f(z,t)=0$) version of the equation is sometimes called the transport equation.}
first-order partial differential equation (PDE)
\begin{align}
\frac{\partial y(z,t)}{\partial z} + c \frac{\partial y(z,t)}{\partial t} = f(z,t)
\label{eq:nonhomog-pde}
\end{align}
is
\begin{align}
y(z,t) = y(0,t-c z) + \int_0^z f(\zeta,t-c z+c \zeta) d\zeta
\end{align}
where $c$ is a real constant.
\end{lemma}
\begin{proof}
Let $\tau = t-c z$ and $y(z,t) = \psi(z,t-c z) = \psi(z,\tau)$. Then we have
\begin{align}
\frac{\partial y}{\partial z}
= \frac{\partial \psi}{\partial z} - c \frac{\partial \psi}{\partial \tau}
\text{ and }
\frac{\partial y}{\partial t}
= \frac{\partial \psi}{\partial \tau}.
\end{align}
Substituting in (\ref{eq:nonhomog-pde}) yields
\begin{align}
\frac{\partial \psi}{\partial z} = f(z,\tau+c z).
\end{align}
The solution to this differential equation is
\begin{align}
\psi(z,\tau) = \psi(0,\tau) + \int_0^z f(\zeta,\tau+c \zeta) d\zeta.
\end{align}
It follows that the solution to (\ref{eq:nonhomog-pde}) is
\begin{align}
y(z,t) = y(0,t-c z) + \int_0^z f(\zeta,t-c z+c \zeta) d\zeta.
\end{align}
\end{proof}
Next, we solve the coupled equations.
Let $A_j(z,t) = V_j(z,t) \exp\left(i \phi_j(z,t) \right)$, $j = 1,2$, where $V_j(z,t) = |A_j(z,t)|$.
Then, we have
\begin{align}
\frac{\partial A_j}{\partial z}
&= \frac{\partial V_j}{\partial z} e^{i \phi_j} + V_j e^{i \phi_j} i \frac{\partial \phi_j}{\partial z} \nonumber \\
&= \left( \frac{\partial V_j}{\partial z} + i V_j \frac{\partial \phi_j}{\partial z} \right) e^{i \phi_j}
\end{align}
and
\begin{align}
\frac{\partial A_j}{\partial t}
&= \left( \frac{\partial V_j}{\partial t} + i V_j \frac{\partial \phi_j}{\partial t} \right) e^{i \phi_j}
\end{align}
By substituting in the coupled equations
and separating the real and imaginary parts, we have
\begin{align}
& \frac{\partial V_1}{\partial z} + \beta_{11} \frac{\partial V_1}{\partial t} = 0 \label{eq:V1_PDE}\\
& \frac{\partial \phi_1}{\partial z} + \beta_{11} \frac{\partial \phi_1}{\partial t}
= \gamma_1 (V_1^2 + 2 V_2^2) \label{eq:Phi1_PDE}\\
& \frac{\partial V_2}{\partial z} + \beta_{12} \frac{\partial V_2}{\partial t} = 0 \label{eq:V2_PDE}\\
& \frac{\partial \phi_2}{\partial z} + \beta_{12} \frac{\partial \phi_2}{\partial t}
= \gamma_2 (V_2^2 + 2 V_1^2) \label{eq:Phi2_PDE}
\end{align}
It follows from Lemma \ref{lemma:nonhomog-pde} that the solutions of equations (\ref{eq:V1_PDE})--(\ref{eq:Phi2_PDE}) are
\ifdefined\twocolumnmode{
\begin{align}
V_1(z,t)
&= V_{1}(0,t-\beta_{11} z)
\\
V_2(z,t)
&= V_{2}(0,t-\beta_{12} z)
\\
\phi_1(z,t)
&= \phi_1(0,t-\beta_{11} z) +
\int_0^z \gamma_1 \Big( V_1^2(\zeta,t-\beta_{11} z+\beta_{11} \zeta) \nonumber\\& \qquad
+ 2 V_2^2(\zeta,t-\beta_{11} z+\beta_{11} \zeta) \Big) d\zeta \nonumber \\
&= \phi_1(0,t-\beta_{11} z) +
\int_0^z \gamma_1 \Big( V_{1}^2(0,t-\beta_{11} z) \nonumber\\& \qquad
+ 2 V_{2}^2(0,t-\beta_{11} z+(\beta_{11} -\beta_{12}) \zeta) \Big) d\zeta
\\
\phi_2(z,t)
&= \phi_2(0,t-\beta_{12} z) +
\int_0^z \gamma_2 \Big( V_2^2(\zeta,t-\beta_{12} z+\beta_{12} \zeta) \nonumber\\& \qquad
+ 2 V_1^2(\zeta,t-\beta_{12} z+\beta_{12} \zeta) \Big) d\zeta \nonumber \\
&= \phi_2(0,t-\beta_{12} z) +
\int_0^z \gamma_2 \Big( V_{2}^2(0,t-\beta_{12} z) \nonumber\\& \qquad
+ 2 V_{1}^2(0,t-\beta_{12} z+(\beta_{12} -\beta_{11}) \zeta) \Big) d\zeta
\end{align}
}\else{
\begin{align}
V_1(z,t)
&= V_{1}(0,t-\beta_{11} z)
\\
V_2(z,t)
&= V_{2}(0,t-\beta_{12} z)
\\
\phi_1(z,t)
&= \phi_1(0,t-\beta_{11} z) +
\int_0^z \gamma_1 \Big( V_1^2(\zeta,t-\beta_{11} z+\beta_{11} \zeta)
+ 2 V_2^2(\zeta,t-\beta_{11} z+\beta_{11} \zeta) \Big) d\zeta \nonumber \\
&= \phi_1(0,t-\beta_{11} z) +
\int_0^z \gamma_1 \Big( V_{1}^2(0,t-\beta_{11} z)
+ 2 V_{2}^2(0,t-\beta_{11} z+(\beta_{11} -\beta_{12}) \zeta) \Big) d\zeta
\\
\phi_2(z,t)
&= \phi_2(0,t-\beta_{12} z) +
\int_0^z \gamma_2 \Big( V_2^2(\zeta,t-\beta_{12} z+\beta_{12} \zeta)
+ 2 V_1^2(\zeta,t-\beta_{12} z+\beta_{12} \zeta) \Big) d\zeta \nonumber \\
&= \phi_2(0,t-\beta_{12} z) +
\int_0^z \gamma_2 \Big( V_{2}^2(0,t-\beta_{12} z)
+ 2 V_{1}^2(0,t-\beta_{12} z+(\beta_{12} -\beta_{11}) \zeta) \Big) d\zeta
\end{align}
}\fi
Therefore, the solution to the original coupled equations (\ref{eq:coupled-pde1}) and (\ref{eq:coupled-pde2}) is
\ifdefined\twocolumnmode{
\begin{align}
A_1(z,t+\beta_{11} z) &= A_{1}(0,t)
\exp\Big(
i \int_0^z \gamma_1 ( |A_{1}(0,t)|^2 \nonumber\\&
+ 2 |A_{2}(0,t+(\beta_{11} -\beta_{12}) \zeta)|^2 ) d\zeta
\Big) \\
A_2(z,t+\beta_{12} z) &= A_{2}(0,t)
\exp\Big(
i \int_0^z \gamma_2 ( |A_{2}(0,t)|^2 \nonumber\\&
+ 2 |A_{1}(0,t+(\beta_{12} -\beta_{11}) \zeta)|^2 ) d\zeta
\Big)
\end{align}
}\else{
\begin{align}
A_1(z,t+\beta_{11} z) &= A_{1}(0,t)
\exp\Big(
i \int_0^z \gamma_1 ( |A_{1}(0,t)|^2
+ 2 |A_{2}(0,t+(\beta_{11} -\beta_{12}) \zeta)|^2 ) d\zeta
\Big) \\
A_2(z,t+\beta_{12} z) &= A_{2}(0,t)
\exp\Big(
i \int_0^z \gamma_2 ( |A_{2}(0,t)|^2
+ 2 |A_{1}(0,t+(\beta_{12} -\beta_{11}) \zeta)|^2 ) d\zeta
\Big)
\end{align}
}\fi
\section{Independence of Filter Outputs}
\label{sec:sufficient}
We drop the user index and time index for notational simplicity.
We decompose $\textbf{Y}$ into $Y_V$ and
$\overline{\textbf{Y}}_{V}$ where $\overline{\textbf{Y}}_{V}$ = $(Y_q: q \in \mathcal{V} \backslash \{V\})$.
Let $\overline{\textbf{y}} = (y_q: q \in \mathcal{V} \backslash \{V\}) \in \mathbb{R}^{|\mathcal{V}|-1}$.
The joint pdf of $X$, $Y_V$ and $\overline{\textbf{Y}}_{V}$ is
\ifdefined\twocolumnmode{
\begin{align}
&p_{X,Y_V,\overline{\textbf{Y}}_{V}}(x,y_V,\bar{\textbf{y}}) \nonumber\\
&= p_{X}(x) \sum_{v \in \mathcal{V}} p_{Y_V,\overline{\textbf{Y}}_{V},V|X}(y_V,\bar{\textbf{y}},v|x) \nonumber \\
&= p_{X}(x) \sum_{v \in \mathcal{V}} p_{V}(v) p_{Z_v}(y_V-x) \prod_{q \in \mathcal{V} \backslash \{v\}} p_{Z_q}(y_q)
\end{align}
}\else{
\begin{align}
p_{X,Y_V,\overline{\textbf{Y}}_{V}}(x,y_V,\bar{\textbf{y}})
&= p_{X}(x) \sum_{v \in \mathcal{V}} p_{Y_V,\overline{\textbf{Y}}_{V},V|X}(y_V,\bar{\textbf{y}},v|x) \nonumber \\
&= p_{X}(x) \sum_{v \in \mathcal{V}} p_{V}(v) p_{Z_v}(y_V-x) \prod_{q \in \mathcal{V} \backslash \{v\}} p_{Z_q}(y_q)
\end{align}
}\fi
where $p_{Z_f}(\cdot)$ is the pdf of $Z_f$ for $f \in \mathcal{V}$.
By using
\begin{align}
p_{Z_f}(z) = p_Z(z) \stackrel{\Delta}{=} \frac{1}{\pi N} e^{-|z|^2/N}, \qquad \forall f \in \mathcal{V},
\end{align}
we have
\begin{align}
p_{X,Y_V,\overline{\textbf{Y}}_{V}}(x,y_V,\bar{\textbf{y}})
&= p_{X}(x) ~ p_{Z}(y_V-x) \prod_{q \in \mathcal{V} \backslash \{v\}} p_{Z}(y_q) \nonumber \\
&= p_{X}(x) ~ p_{Y_V|X}(y_V|x) ~ p_{\overline{\textbf{Y}}_{V}}(\bar{\textbf{y}}).
\end{align}
Therefore, we find that $X$ and $Y_V$ are statistically independent of $\overline{\textbf{Y}}_{V}$ and
\begin{align}
I(X ; \overline{\textbf{Y}}_{V}|Y_V) = 0.
\end{align}
Hence, we have
\begin{align}
I(X; \textbf{Y})
&= I(X ; Y_V) + I(X ; \overline{\textbf{Y}}_{V}|Y_V)
= I(X ; Y_V).
\end{align} |
\section{Introduction}
\noindent
Backward stochastic differential equations, BSDEs for short, is by now an established field of research. The solution to a classical BSDE, driven by a Wiener process $W$,
is a pair of processes $(Y,Z)$ such that
\begin{eqnarray*
Y_t= \xi + \int_t ^T f(s,Y_s,Z_s) ds - \int_t ^T Z_s dW_s,\ 0\leq t\leq T,
\end{eqnarray*}
where $\xi$ is a random variable that becomes known, with certainty, only at time $T$. In this setting $Y_t\in\mathbb R^d$, $d\geq 1$, and in the following we refer to the case $d=1$ as the one-dimensional case and to the case $d>1$ as the multi-dimensional case. Classical BSDEs have turned out important in many areas of mathematics including mathematical finance, see \cite{EPQ} and the long list of references therein, stochastic control theory and stochastic game
theory, see, e.g., \cite{CK} and \cite{HL}, as well as in the connection to partial differential equations, see, e.g., \cite{BBP} and \cite{PP}.
In \cite{EKPPQ} a notion of \textit{reflected} BSDE was introduced. A solution to a one-dimensional reflected BSDE is a triple of processes $(Y,Z,\Lambda)$ satisfying
\begin{eqnarray*
Y_t&=& \xi + \int_t ^T f(s,Y_s,Z_s) ds +\Lambda_T-\Lambda_t- \int_t ^T Z_s dW_s,\ 0\leq t\leq T,\notag\\
Y_t&\geq& S_t,
\end{eqnarray*}
where the barrier $S$ is a given (one-dimensional) stochastic process. $\Lambda$ is a continuous increasing process, with $\Lambda_0=0$, pushing the process $Y$ upwards
in order to keep it above the barrier. This is done with minimal energy in the sense that
\begin{eqnarray*
\int_0^T (S_t-Y_t)d\Lambda_t=0,
\end{eqnarray*}
and consequently $\Lambda$ increases only when $Y$ is at the boundary of the space-time domain $\{(t,s): s>S_t\}$. This type of reflected BSDE has important applications in the context of American options, optimal stopping and obstacle problem, see \cite{EKPPQ}, as well as in the context of stochastic game problems, see \cite{CK}.
In the multi-dimensional case there are at least two different types of reflected BSDEs studied in the literature.
\noindent The first type of
multi-dimensional reflected BSDE was first studied in \cite{GP} where the authors considered reflected BSDEs of the form
\begin{eqnarray}\label{Kn4}
Y_t&=& \xi + \int_t ^T f(s,Y_s,Z_s) ds +\Lambda_T-\Lambda_t- \int_t ^T Z_s dW_s,\ 0\leq t\leq T,\notag\\
Y_t&\in & \Omega, \ 0\leq t\leq T,
\end{eqnarray}
where $\Omega\subset\mathbb R^d$. In this case $\Lambda_0=0$ and
\begin{eqnarray}\label{Kn4+}
&&\Lambda_t=\int_{0}^{t}\gamma _{s}d\left\vert \Lambda \right\vert
_{s},\ \gamma _{s}\in N^{1}\left( Y_{s}\right),\notag\\
&& d\left\vert \Lambda \right\vert \left( \left\{ t\in \left[ 0,T\right]
:Y_{t} \in \Omega\right\} \right)=0,
\end{eqnarray}
where $N^{1}\left( Y_{s}\right)$ is the unit inner normal to $\Omega$ at $Y_{s}$.
In particular, the process $\Lambda_t$ is of bounded total variation $|\Lambda|$ and it increases only when $Y$ is at the boundary of $\Omega$. To be more precise, when $Y$ is at the boundary it is pushed into the domain along $\gamma\in N^1(Y)$. In \cite{GP} existence and uniqueness for this problem
is established and we note that this problem, and its analysis, is inspired by and resemble the corresponding theory for reflected stochastic differential equations, see \cite{T}, \cite{S}, and \cite{LS}. Naturally one can attempt, as in the case of reflected SDEs, to study this problem with oblique reflection instead of reflection in the direction of the inner unit normal. However, to the best of our knowledge the case of oblique reflection is a less developed area of research in the context of BSDEs and we are only aware of the work in \cite{R}, where the author studies an obliquely reflected BSDE in an orthant.
\noindent
The second type of multi-dimensional reflected BSDEs occurs in the study of optimal switching problems and stochastic games, see, e.g., \cite{AF}, \cite{AH}, \cite{DHP}, \cite{HT}, \cite{HZ}, and references therein. In the generic optimal switching problem a production facility is considered and it is assumed that the production can run in $d \geq 2$ different production modes. Furthermore, there is a stochastic process $X=(X_t)_{t\geq 0}$ which stands for the market price of the underlying commodities and other financial parameters that influence the production. When the facility is in mode $i$, the revenue per unit time is $f_i(t,X_t)$ and the cost of switching from mode $i$ to mode $j$, at time $t$, is $c_{ij}(t,X_t)$. Let $(Y_t^1,\dots,Y_t^d)$ be the value function associated with the optimal switching problem, on the time interval $[t,T]$, i.e., $Y_t^i$ stands for the optimal expected profit if, at time $t$,
the production is in mode $i$. In this case, one can prove, under various assumptions, see \cite{AF}, \cite{AH}, and \cite{DHP}, that
$(Y_t^1,\dots,Y_t^d)$ solves the reflected BSDE
\begin{align}\label{eq6}
&Y^i_t=\xi_i+\int_t^Tf_i(s, X_s)ds-\int_t^TZ_s^idW_s+\Lambda_T^i-\Lambda_t^i,\notag\\
&Y^i_t\geq \max_{j\in A_i} \left (Y^j_t - c_{ij}(t,X_t)\right), \notag \\
&\int_0^T \left ( Y^i_t-\max_{j\in A_i} \left (Y^j_t -c_{ij}(t,X_t) \right ) \right )d\Lambda_t^i=0,
\end{align}
where $i\in\{1,\dots,d\}$, $0\leq t\leq T,$ and $A_i=\{1,\dots,d\}\setminus\{i\}$. In this case the reflected BSDE evolves in the closure
of the time-dependent domain
\begin{eqnarray}\label{eq6+}
D&=&\{(t,y)=(t, y_1,\dots,y_d)\in\mathbb R^{d+1}:\ 0\leq t\leq T,\notag\\
&&y_i \geq \max_{j\in A_i} \left ( y_j - c_{ij}(t,X_t) \right ), \mbox{ for all }i\in\{1,\dots,d\}\}.
\end{eqnarray}
On the boundary of $D$ a reflection occurs and in \cite{HT} the authors refer to this as an oblique reflection. While this oblique reflection seems to have no clear relation to what is referred to as an oblique reflection in the context of \eqref{Kn4}, \eqref{Kn4+},
it is still fair to refer to the problem in \eqref{eq6} as an obliquely reflected BSDE. However, we emphasize that the problems in \eqref{Kn4}, \eqref{Kn4+} and \eqref{eq6} are significantly different.
In this paper we consider the problem in \eqref{Kn4}, \eqref{Kn4+} in time-dependent domains and with underlying stochastic processes beyond Brownian motion. In light of \eqref{eq6}, \eqref{eq6+}, and corresponding developments for
reflected SDEs, see \cite{C}, \cite{CGK}, \cite{LS}, \cite{NO}, \cite{S}, and \cite{T}, it is natural to allow for time-dependent domains and in many cases this extra feature calls for additional arguments in comparison with the case of time-independent domains. In particular, we here consider \eqref{Kn4}, \eqref{Kn4+} in the context of time-dependent domains, and along the lines of \cite{GP}. In addition, we allow the BSDE to be driven by a Wiener-Poisson type process and our main result is a generalization of \cite{GP} and \cite{O} to a time-dependent setting. In general, it seems difficult to generalize \cite{GP} and its proofs beyond the assumption of convexity of the time-slices of the domain. Indeed, the assumption on convexity is heavily explored in \cite{GP} and \cite{O}. Beyond ensuring the existence of projections, convexity establishes the positivity of certain terms appearing when applying the Ito formula. In this sense, one may say that the arguments are slightly rigid as the structural assumption of convexity seems crucial. In our analysis,
it turns out that we are only able to pull the arguments of \cite{GP} through in the context of time-dependent domain having a similar rigidity in time. More precisely, in our case the time slices must be non-increasing and hence the domain must be non-expanding as a function of time. Under such a structural assumption though, we are able to generalize \cite{GP} and \cite{O} to a time-dependent setting. Finally, we note that it is an interesting open problem to understand if, in analogy with the connection between optimal switching problems and the problem in \eqref{eq6}, the problem in \eqref{Kn4}, \eqref{Kn4+} can be naturally associated to some stochastic optimization problem.
\setcounter{equation}{0} \setcounter{theorem}{0}
\section{Statement of main result}
\noindent
In this section we state our main result. To do this properly we first briefly discuss the geometry and processes of Wiener-Poisson type and define the reflected BSDE studied in this paper.
\subsection{Geometry}
Given $d\geq 1$, we let $\left\langle \cdot ,\cdot
\right\rangle $ denote the standard inner product on
\mathbb{R}^{d}$ and $\left\vert z\right\vert =\left\langle z,z\right\rangle
^{1/2}$ be the Euclidean norm of $z\in\mathbb R^d.$ Whenever $z\in\mathbb{R}^{d}$ and $r>0$, we let $B_r(z)$ and $S_r(z)$ denote the ball and sphere of radius $r$, centered at $z$, respectively, i.e. $B_{r}\left( z\right) =\left\{ y\in\mathbb{R}^{d}:\left\vert z-y\right\vert <r\right\} $ and $S_{r}\left( z\right)=\left\{ y\in\mathbb{R}^{d}:\left\vert z-y\right\vert =r\right\}$. Moreover, given $F\subset\mathbb{R}^{d}$, $E\subset\mathbb{R}
^{d}$, we let $\bar{F}$, $\bar{E}$ be the closure of $F$ and $E$,
respectively, and we let $d\left( y,E\right) $ denote the Euclidean distance
from
$y\in\mathbb{R}^{d}$ to $E$. Given $d\geq 1$, $T>0$ and an open, connected set $D^{\prime}\subset\mathbb{R}
^{d+1}$
we will refer to
\begin{equation*
D=D^{\prime }\cap ([0,T]\times \mathbb{R}^{d}),
\end{equation*
as a time-dependent domain.
Given $D$ and $t\in \left[ 0,T\right] $, we
define the time sections of $D$ as
\begin{equation}\label{eq:timeslice}
D_{t}=\left\{ z:\left( t,z\right) \in D\right\}.
\end{equation} We assume that
\begin{equation}\label{timedep+}
D_{t}\neq \emptyset, D_t\text{ is open, bounded and connected for every }t\in \left[ 0,T\right],
\end{equation
and that
\begin{equation}
D_{t}\text{ is convex
for every }t\in \left[ 0,T\right]. \label{timedep+1}
\end{equation
Furthermore, following \cit
{CGK}, we let
\begin{equation*
l\left( r\right) =\sup_{\substack{ s,t\in \lbrack 0,T] \\ \left\vert s-t\right\vert \leq r}}\,\sup_{z\in \overline{D_{s}}}d\left( z,D_{t}\right) ,
\end{equation*}
be the modulus of continuity of the variation of $D$ in time and we assume that
\begin{equation}\label{limitzero}
\lim_{r \to 0^{+}}l\left( r\right) =0.
\end{equation}
We also assume that
\begin{equation}\label{timedep+2}
D_{t'}\subseteq D_t\text{ whenever $t'\geq t$, }t', t\in \left[ 0,T\right].
\end{equation
Note that \eqref{timedep+2} implies that
\begin{equation*
l\left( r\right) =\sup_{\substack{ t\in \lbrack 0,T], [t-r,t+r]\in [0,T]}}\sup_{z\in \overline{D_{t-r}}}d\left( z,D_{t+r}\right).
\end{equation*
We let $\partial D$ and $\partial D_{t}$, for $t\in \left[ 0,T\right] $,
denote the boundaries of $D$ and $D_{t}$, respectively, and we let $N_{t}\left(
z\right) $ denote the cone of inward normal vectors at $z\in \partial D_{t}
, $t\in \lbrack 0,T]$. Note that it follows from \eqref{timedep+1} that $N_{t}\left(
z\right)\neq\emptyset $ for every $z\in \partial D_{t}$, $t\in \lbrack 0,T]$. In general, the cone $N_{t}\left( z\right) $ of inward normal vectors at $z\in \partial
D_{t}$, $t\in \lbrack 0,T]$, is defined as being equal to the set consisting
of the union of the set $\left\{ 0\right\}$ and the set
\begin{equation*
\left\{ v\in\mathbb{R}^{d}:v\neq 0,\exists \rho >0\text{ such that }B_{\rho }\left( z-\rho
v/\left\vert v\right\vert \right) \subset \mathbb{R}^{d}\setminus D_t\right\} .
\end{equation*
Note that this definition does not rule out the possibility of several unit
inward normal vectors at the same boundary point. Given $N_{t}\left(
z\right) $, we let $N_{t}^{1}(z):=N_{t}(z)\cap S_{1}(0)$, so that
N_{t}^{1}(z)$ contains the set of vectors in $N_{t}(z)$ with unit length. In this paper we consider reflected BSDEs in the setting of time-dependent domains $D$ satisfying
\eqref{timedep+}-\eqref{timedep+2}. Furthermore, reflection at $z\in \partial D_{t}$, $t\in \lbrack 0,T]$, is considered in the direction of a
unit spatial inward normal in the cone $N_{t}\left(
z\right)$.
\subsection{Processes of Wiener-Poisson type}
Throughout the paper we let
$$\left(\Omega ,\mathcal{F} ,\{ \mathcal{F}_t\}_{t \in [0,T]}, \mathbb{P}\right)$$ be a complete Wiener-Poisson space in $\mathbb R^n\times \mathbb R^m\setminus\{0\}$ with Levy measure $\lambda$. In particular, $\left(\Omega ,\mathcal{F} ,\mathbb P \right) $ is a complete probability space and $\{\mathcal{F}_t\},_{t\in[0,T]}$ is an increasing, right continuous family of complete sub $\sigma$-algebras of $\mathcal{F}$. We let $(W_t,\{F_t\})_{t \in [0,T]}$ be a standard Wiener process in $\R^n$ and $(\mu_t, \{\F_t\})_{ t\in [0,T]}$ be a martingale measure in $\R^m \setminus\{0\}$, which is assumed to be independent of $W$, and which corresponds to a standard Poisson random measure $p(t,A)$. Indeed, for any Borel measurable subset $A$ of $\R^m \setminus \{0\}$ such that the Levy measure $\lambda$ satisfies $\lambda(A) <+ \infty $, we have
\begin{equation*}
\mu_t(A) = p(t,A) - t \lambda(A)
\end{equation*}
where $p(t,A)$ satisfies
\begin{equation*}
E[ p(t,A) ] = t \lambda(A).
\end{equation*}
We let $U:=\R^m \setminus \{0\}$ and we let $\mathcal{U}$ be its Borel $\sigma$-algebra. We assume that $\{\F_t\}_{t\in [0,T]}$ is the filtration generated by $W_t$ and the jump process corresponding to the Poission random measure $p$, augmented with the $\mathbb P$-null sets of $\F$, i.e.,
\begin{equation*
\F_t=\sigma\left( \int \int _{A \times [0,s]} p(ds,dx): s\leq t, A\in \mathcal{U} \right ) \vee \sigma \left (W_s, s\leq t \right) \vee \F_0,
\end{equation*}
where $\F_0$ denotes the $\P$-null sets of $\F$ and $\sigma_1 \vee \sigma_2 $ denotes, given two $\sigma$-algebras $\sigma_1$ and $\sigma_2$, the $\si$-algebra generated by $\si_1 \cup \si_2$.
\subsection{Reflected BSDEs} Given $T>0$, we let $\mathcal{D}\left( \left[ 0,T\right] ,
\mathbb{R}^{d}\right) $ denote the set of c\`{a}dl\`{a}g functions $v(t)=v_{t}:\left[ 0,T\right] \rightarrow
\mathbb{R}^{d}$,
i.e., functions which are right continuous and have left limits. We denote the set of functions $w(t)=w_{t}:\left[ 0,T\right]
\rightarrow \mathbb{R} ^{d}$ with bounded variation by $\mathcal{BV}\left( \left[ 0,T\right] , \mathbb{R}^{d}\right) $ and we let $\left\vert w \right\vert $ denote the total
variation of $w \in \mathcal{BV}\left( \left[ 0,T\right] ,\mathbb{R}^{d}\right) $.
Recall that the total variation process $\left\vert w \right\vert $ is defined as
$$
\left\vert w \right\vert_t=\sup\sum_{k=1}^{n}| w_{t_i}-w_{t_{i-1}}|,\ 0\leq t\leq T,
$$
where the supremum is taken over all finite partitions $0=t_0<t_1<\dots<t_n=t$. Furthermore, we have that
\begin{eqnarray}\label{bff}
w_t=\int_0^t\nu_sd\left\vert w \right\vert_s
\end{eqnarray}
where $\nu_s$ is a vector of length 1, i.e., $|\nu_s|=1$ for $\left\vert w \right\vert$-almost all $s$. Let $$\left(
\Omega ,\mathcal{F}, \{\mathcal{F}_t\}, \mathbb{P}, W_t,\mu_t, t\in[0,T] \right)$$ be the complete Wiener-Poisson space in $\mathbb R^n\times \mathbb R^m\setminus\{0\}$, with Levy measure $\lambda$, as outlined above. Let $L^2(\Omega, \mathcal{F}_T, \mathbb P)$ be the space of square integrable, $\mathcal{F}_T$-adapted random variables and let $L^2(U, \mathcal{U}, \lambda; \R^d)$ be the space of functions which are $\mathcal{U}$-measurable, maps values in $U$ to $\R^d$, and which are square integrable on $U$ with respect to the Levy-measure $\lambda$.
In the following we let
the norm
$$\|z\|:=(\sum_{i,j}|z_{ij}|^2)^{1/2}$$
be defined on real-valued $(d\times n)$-dimensional matrices and we define the norm
$$\| u(e) \| := \left (\int_V |u(e)|^2 \lambda(de) \right )^{1/2}$$
on $L^2(U, \mathcal{U}, \lambda; \R^d)$. Let $\xi=(\xi_1,\dots,\xi_d)$ be such that
\begin{eqnarray}\label{data}
\mbox{$\xi \in L^2(\Omega, \mathcal{F}_T, \mathbb P)$ and $\xi\in D_T$ a.s.}
\end{eqnarray}
Let $f: \Omega \times [0,T] \times \R^d \times\R^{d\times n}\times L^2(U, \mathcal{U}, \lambda; \R^d) \to \R^d$ be a function such that
\begin{eqnarray}\label{data++}
(i)&&\mbox{$(\omega,t)\to f(\omega,t,y,z,u) $ is $\mathcal{F}_t$ progressively measurable whenever}\notag\\
&&\mbox{$(y,z,u)\in \R^d \times\R^{d\times n}\times L^2(U, \mathcal{U}, \lambda; \R^d)$},\notag \\
(ii)&&\mbox{$E \bigl [\int_0 ^T |f(\omega,t,0,0,0)|^2 dt\bigr ] < \infty $},\notag\\
(iii)&&\mbox{$|f(\omega, t, y, z, u)-f(\omega, t, y', z', u')|\leq c(|y-y'|+\|z-z'\|+\|u-u'\|)$}\notag\\
&&\mbox{for some constant $c$ whenever}\notag\\
&&\mbox{$(y,z,u), (y',z',u') \in \R^d \times \R^{d\times n} \times L^2(U, \mathcal{U}, \lambda; \R^d)$, $(\omega,t)\in \Omega \times [0,T]$}.
\end{eqnarray}
In the context of BSDEs, $\xi$ and $f$ are usually referred to as terminal value and driver of the BSDE, respectively. We are now ready to formulate the notion of reflected BSDE considered in this paper.
\begin{definition}\label{rbsde} Let $d\geq 1$ and $T>0$. Let $D\subset\mathbb{R}^{d+1}$ be a time-dependent domain satisfying \eqref{timedep+}. Given $(\xi,f)$ as in \eqref{data}-\eqref{data++}, a quadruple $(Y_t, Z_t, U_t, \Lambda_t)$ of progressively measurable processes with values in $\R^d \times \R^{d \times m} \times L^2(U, \mathcal{U}, \lambda; \R ^d) \times \R^d$
is said to be a solution to a reflected BSDE, with reflection in the inward spatial normal direction, in $D$, and with data $(\xi,f),$ if the following holds. $Y\in \mathcal{D}\left( \left[ 0,T\right] ,
\mathbb{R}^{d}\right) $, $Z$ and $U$ are predictable processes, and
\begin{eqnarray*}
(i)&&E\left [\sup _{0 \leq t \leq T} |Y_t|^2\right ] <\infty, \\
(ii)&& E \left [\int_0 ^T \|Z_t\|^2 dt + \int _0^T \int _U|U_s(e)|^2 \lambda(de)ds \right ] < \infty,\notag\\
(iii)&&Y_t= \xi + \int_t ^T f(s,Y_s,Z_s,U_s) ds + \Lambda_T- \Lambda_t\notag\\
&&- \int_t ^T Z_s dW_s - \int _t^T \int _U U_s(e) \mu(de,ds) \quad \mbox{a.s.}, \\
(iv)&& Y_t \in\overline{D_t} \quad \mbox{a.s.},
\end{eqnarray*}
whenever $t \in [0, T]$. Furthermore, $\Lambda\in \mathcal{BV}\left( \left[ 0,T\right]
\mathbb{R}
^{d}\right) $ and
\begin{eqnarray*}
(v)&&\Lambda_t=\int_{0}^{t^{}}\gamma _{s}d\left\vert \Lambda \right\vert
_{s},\ \gamma _{s}\in N_{s}^{1}\left( Y_{s}\right) \mbox{ whenever } Y_s \in \partial D_{s},\\
(vi)&& d\left\vert \Lambda \right\vert \left( \left\{ t\in \left[ 0,T\right]
:\left(t,Y_{t}\right) \in D\right\} \right)=0.
\end{eqnarray*}
\end{definition}
\subsection{Statement of the main result} Concerning reflected BSDEs we establish the following result.
\begin{theorem}\label{maint1} Let $d\geq 1$ and $T>0$. Let $D\subset \mathbb{R}^{d+1}$ be a time-dependent domain satisfying \eqref{timedep+}-\eqref{timedep+2} and assume the terminal data $\xi$ and driver $f$ satisfy \eqref{data}-\eqref{data++}. Then there exists a unique solution $(Y_t, Z_t, U_t, \Lambda_t)$ to the reflected BSDE, with reflection in the inward spatial normal direction, in $D$, and with data $(\xi,f),$ in the sense of Definition \ref{rbsde}.
\end{theorem}
\subsection{Organization of the paper} The rest of the paper is organized as follows. Section \ref{sec:prel} is of preliminary nature and we here focus on the geometry of the time-dependent domain as well as smooth approximations of it. We also recall the Ito formula in the context of Wiener-Poisson processes. In section \ref{sec:lemmas} we introduce a sequence of non-reflected BSDEs, constructed by penalization techniques, and develop a number of technical lemmas for these. Finally, using the results of section \ref{sec:lemmas}, we prove the main result in section \ref{sec:proof}.
\section{Preliminaries} \label{sec:prel}
\noindent
Let $d\geq 1$ and $T>0$. Let $D\subset\mathbb{R}^{d+1}$ be a time-dependent domain satisfying \eqref{timedep+}~-~\eqref{timedep+2}. Let $N =N_{t}(z)=N (t,z)$ denote the cone of inward normal vectors given for all $z\in \partial D_{t}$, $t\in \lbrack 0,T]$. Note that by \eqref{timedep+1} there exists, for any $y\in\mathbb{R}^{d}\setminus \overline{D}_{t}$, $t\in \lbrack 0,T]$,
at least one projection of $y$ onto $\partial D_{t}$ along $N_{t}$,
denoted $\pi\left( t,y\right) $, which
satisfie
\begin{equation*
\left\vert y-\pi\left(t, y\right) \right\vert=d\left( y,D_{t}\right).
\end{equation*}
To have $\pi(t,\cdot)$ well-defined for all $y\in\mathbb R^d$ we also let $\pi\left(t, y\right)=y$ whenever $y\in\overline{D_t}$. The following lemma summarizes a few standard results from convex analysis
\begin{lemma}\label{lemmaa} Let $D\subset
\mathbb{R}^{d+1}$ be a time-dependent domain satisfying \eqref{timedep+} and assume \eqref{timedep+1} and \eqref{timedep+2}. Then the following holds whenever $t\in[0,T]$:
\begin{eqnarray*}
(i)&&\langle y^\prime-y,y-\pi\left(t, y\right)\rangle \leq 0,\mbox{ for } (y,y^\prime) \in \mathbb R^d\times\overline{D_t}, \mbox{ and }\notag \\
(ii)&&\langle y^\prime-y, y-\pi(t,y)\rangle \leq \langle y^\prime-\pi(t,y^\prime),y-\pi(t,y)\rangle,\notag \\
(iii)&& |\pi(t,y) - \pi(t,y')| \leq |y-y'|,
\end{eqnarray*}
whenever $(y,y^\prime) \in \mathbb R^d\times\mathbb R^d$. Furthermore, there exists $P_T\in D_T$ and $\gamma$, $1\leq\gamma<\infty$, depending on $d(P_T, \partial D_T)$, such that
\begin{eqnarray*
(iv)&& \langle y-P_T, y-\pi(t,y)\rangle \geq \gamma^{-1} |y-\pi(t,y)|, \mbox{ for any } y\in \mathbb R^d, t\in[0,T].
\end{eqnarray*}
\end{lemma}
\subsection{Geometry of time-dependent domains - smooth approximations}
Note that the assumptions in \eqref{timedep+}, \eqref{timedep+1}, and \eqref{timedep+2} contain no particular smoothness assumption. Instead, we will in the following construct smooth approximations of $D$ to enable the use of Ito's formula. In the following we let $h(t,y)=d(y,D_t)$ whenever $y\in\mathbb R^d$, $t\in [0,T]$, with the convention that $h(t,y)=0$ if $y\in\overline{D_t}$. Assuming that $D_T\neq\emptyset\neq D_0$ we
let $h(t,y)=h(T,y)$ whenever $y\in\mathbb R^d$, $t>T$, and $h(t,y)=h(0,y)$ whenever $y\in\mathbb R^d$, $t<0$. Using this notation we see that
$$\overline D=\{(t,x)\in [0,T]\times\mathbb R^d|\ h(t,x)=0\}.$$ Let
$\phi=\phi(s,y)$ be a smooth mollifier in $\mathbb R^{d+1}$, i.e., $\phi\in C_0^\infty(\mathbb R^{d+1})$, $0\leq\phi\leq 1$, the support
of $\phi$ is contained in the Euclidean unit ball in $\mathbb R^{d+1}$, centered at $0$, and $\int\phi dyds=1$. Let, for $\delta>0$ small,
$\phi_\delta(s,y)=\delta^{n+1}\phi(\delta^{-1}s,\delta^{-1}y)$. Based on $\phi_\delta$ we let, whenever $(t,y)\in\mathbb R\times\mathbb R^d$,
\begin{eqnarray*}
h_{\delta}(t,y)=(\phi_{\delta}\ast h)(t,y)=\int_{\mathbb R}\int_{\mathbb R^n}\phi_{\delta}(t-s,y-x)h(s,x)dxds
\end{eqnarray*}
be a smooth mollification of $h$. Furthermore, we let
\[
h ( F, G ) = \max ( \sup \{ d ( y, F ) : y \in G \}, \sup \{ d ( y, G ) : y \in F \} )
\]
denote the Hausdorff distance
between the sets $ F, G \subset \mathbb R^d$. Based on $h_{\delta}$ we introduce a smooth approximation of $D$ as follows. Given $\eta$ fixed and $\delta>0$, we let
\begin{equation*}
D^{\eta}_{\delta} = \{(t,x)\in [0,T]\times\mathbb R^d|h_\delta(t,x)< \eta\}.
\end{equation*}
Then $D^{\eta}_{\delta}$ converges to $D^{\eta}:= \{(t,x)\in [0,T]\times\mathbb R^d|h(t,x)< \eta\}$ in the Hausdorff distance sense as $\delta\to 0$. Note that $D^\eta_\delta$ is a $C^\infty$-smooth domain. Hence, letting $\delta, \eta \to 0$ we have the following lemma.
\begin{lemma}\label{lemmauu1-} Let $D\subset
\mathbb{R}
^{d+1}$ be a time-dependent domain satisfying \eqref{timedep+}~-~\eqref{timedep+2}. Then, for any
$\epsilon>0$ there exists a time-dependent domain $D_\epsilon\subset \mathbb{R} ^{d+1}$ satisfying \eqref{timedep+}~-~\eqref{timedep+2} such that $D_\epsilon$ is $C^\infty$-smooth and
\begin{eqnarray*
h(D_{t},D_{\epsilon,t})<\epsilon\mbox{ for all $t\in[0,T]$},
\end{eqnarray*}
where $D_t$ is as defined in \eqref{eq:timeslice}, and $D_{\e,t}=\{x: (x,t) \in D_\e\}$.
\end{lemma}
Let, for all $t\in \lbrack 0,T]$, $N_\epsilon =N_{\epsilon,t}(z)=N_\epsilon(t,z)$ denote the cone of inward normal vectors at $z \in \partial D_{\e,t}$. Due to the smoothness of $\partial D_{\epsilon,t}$, $N_\epsilon(t,z)$ consists of a single vector. For any $y\in
\mathbb{R}
^{d}\setminus \overline{D}_{\e,t}$, $t\in \lbrack 0,T]$, we let
$\pi_\epsilon\left( t,y\right) $ denote the projection of $y$ onto $\partial D_{\epsilon,t}$ along this unique direction. To have $\pi_\epsilon(t,\cdot)$ well-defined for all $y\in\mathbb R^d$ we also let $\pi_\epsilon\left(t, y\right)=y$ whenever $y\in\overline{D}_{\epsilon,t}$.
In this setting, the following lemma can be proven as Lemma 2.2 in \cite{GP} as we are only considering fixed time slices $D_t$ of $D$.
\begin{lemma}
\label{lemmauu1}
Let $D\subset\mathbb{R}^{d+1}$ be a time-dependent domain satisfying \eqref{timedep+}~-~\eqref{timedep+2} and let $D_{\epsilon}$
be constructed as above. There exists a constant $c$ such that, if $\epsilon\in (0,1)$, $y\in\mathbb R^d$ and $t\in[0,T]$, then
\begin{eqnarray*}
(i)&&|\pi(t,y)-\pi_\epsilon(t,y)| \leq c \sqrt{\epsilon^2+\epsilon d(y, D_{\epsilon,t})},\\
(ii)&&|\pi(t,y)-\pi_\epsilon(t,y)| \leq c \sqrt{\epsilon^2+\epsilon d(y, D_t)}.
\end{eqnarray*}
\end{lemma}
The following lemma is a corollary of Lemma \ref{lemmauu1}.
\begin{lemma} \label{lemmauu2} Let $D$ and $D_{\epsilon}$ be as in the statement of Lemma \ref{lemmauu1}. There exists a constant $c$ such that, if
$\epsilon\in (0,1)$ and $y\in\mathbb R^d$, $t\in[0,T]$, then
\begin{eqnarray*}
(i)&&|\pi(t,y)-\pi_\epsilon(t,y)| \leq c \sqrt{\epsilon}(1+d(y, D_{\epsilon,t})),\\
(ii)&&|\pi(t,y)-\pi_\epsilon(t,y)|\leq c \sqrt{\epsilon}\sqrt{d(y, D_{t,\epsilon})}\mbox{ whenever } d(y,D_{\epsilon,t})>\epsilon.
\end{eqnarray*}
\end{lemma}
We here also recall Ito's formula in the context of Wiener-Poisson processes, see \cite{OS}. Here and in the following, we denote by $\C^{1,2}([0,T]\times \R^d, \R)$ the space of functions $\varphi(t,y) : [0,T]\times \R^d \to \R$ which are once continuosly differentiable with respect to $t\in [0,T]$ and twice continuosly differentiable with respect to $y \in \R^d$ and we let $A^\ast$ denote the transpose of the matrix $A$.
\begin{lemma}\label{lemmauu2+}
Let $Y_t$ be a Levy process such that
\begin{equation*}
dY_t = f_t dt + \sigma _t dW_t + \int _U U_t(e) \mu (de,dt)
\end{equation*}
and let $\varphi(t,y) \in \C^{1,2}([0,T]\times \R^d, \R)$. Then
\begin{eqnarray*}
d\varphi(t, Y_t) &=& \partial_t\varphi(t, Y_{t}) dt+ (\nabla \varphi(t, Y_{t^-})) \left [ f_t dt + \sigma_t dW_t + \int _U U_s(e) \mu (de,ds) \right] \notag \\
&+& \sum_{i,j}\frac{1}{2} (\sigma_t\sigma_t^\ast)_{ij}\partial^2_{y_iy_j}\varphi_\epsilon(t, Y_{t}) dt \notag \\
&+& \int_U \left [\varphi(t,Y_{t^-} + U_t(e))-\varphi(t,Y_{t^-}) - \langle \nabla \varphi(t,Y_{t^-}), U_t(e) \rangle \right] p(de,dt).
\end{eqnarray*}
\end{lemma}
Based on the smooth domain $D_\e$ we define the function $\varphi_\epsilon(t,y):= (d(y,D_{\epsilon,t}))^2 = |y-\pi_\epsilon(t,y)|^2$ to which Ito's formula needs to be applied in the proof of Theorem \ref{maint1},. Note that although $D_\e$ is a smooth domain, the second (spatial) derivative of $\varphi_\e$ is not continuous at the boundary of $D_{\e,t}$ and thus Lemma \ref{lemmauu2+} is not directly applicable. To enable the use of Ito's formula we therefore proceed along the lines of \cite{LS}, see also \cite{GP}, and extend our distance function $\varphi_\e$ across the boundary and into the domain $D_{\e,t}$. Indeed, since $\partial D_{\e,t}$ is smooth there exists a neighbourhood $V_{\e,t}$ of $\partial D_{\e,t}$ such that, for $y \in D_{\e,t} \cap V_{\e,t}$, there exists a unique pair $(x, s) \in \partial D_{\e,t} \times \R^+$ such that $y=x+s\gamma$, where $\gamma \in N^1_{\e,t}(x)$. Recall that $N^1_{\e,t}(x)$, the cone of unit inward normal vectors to $D_{\e,t}$, at $x \in \partial D_{\e,t}$, contains only a single vector.
By the convexity of $D_{\epsilon,t}$ we also have
$$y=x+s \gamma, \ \mbox{ for } x=\pi_\epsilon(t,y),\ s=-d(y, D_{\epsilon,t}), \ \gamma \in N^1_{\e,t}(\pi_\e(t,x)),$$
whenever $y\in\mathbb R^d\setminus\overline{D_{\epsilon,t}}$. Hence, for $t \in [0,T]$ fixed, we can define a smooth map $\phi_\epsilon : \R^d \to\R$ such that
\begin{eqnarray*
\phi_\epsilon(t,y)&=&s\quad \mbox{ when } y\in (\mathbb R^d\setminus\overline{D_{\epsilon,t}})\cup (
V_{\epsilon,t}\cap \overline{D_{\epsilon,t}}),\notag\\
\phi_\epsilon(t,y)&>&0\quad \mbox{ otherwise.
\end{eqnarray*}
Using such a function $\phi_\epsilon$ we have
\noindent \begin{eqnarray*
D_\epsilon&=&\{(t,y):\ t\in [0,T], y\in\mathbb R^d,\ \phi_\epsilon(t,y)>0\},\notag\\
\partial D_{\epsilon,t}&=&\{y\in\mathbb R^d,\ \phi_\epsilon(t,y)=0\},\mbox{ for }t\in [0,T],\notag\\
(\mathbb R^d\times [0,T])\setminus\overline{D_\epsilon}&=&\{(t,y):\ t\in [0,T], y\in\mathbb R^d,\ \phi_\epsilon(t,y)<0\}.
\end{eqnarray*}
Note that $\phi_\e(t,y)$ is smooth also across the boundary of $D_{\e,t}$ and that $\varphi_\e(t,y)=(\phi_\e(t,y)^-)^2$. Following \cite{GP}, we can now take an approximating sequence of smooth functions $\{g_n\}_{n\geq0}$, tending to $g(x)=(x^-)^2$ as $n \to \infty$,
and construct a sequence of smooth functions $\{\varphi^n_\e(t,y)=g_n(\phi_\e(t,y))\}_{n\geq 0}$ such that Ito's formula can be applied to $\varphi^n_\e$ for every $n\geq 0$ and such that
$\varphi^n_\e(t,y)$, $\frac{\partial}{\partial t}\varphi^n_\e(t,y)$, $\frac{\partial}{\partial y_i}\varphi^n_\e(t,y)$, $\frac{\partial^2}{\partial y_iy_j}\varphi^n_\e(t,y)$ tend to $\varphi_\e(t,y)$, $\frac{\partial}{\partial t}\varphi_\e(t,y)$, $\frac{\partial}{\partial y_i}\varphi_\e(t,y)$, $\frac{\partial^2}{\partial y_iy_j}\varphi_\e(t,y)$, respectively, as $n \to \infty$. Having such an approximation in mind, we will from here on in slightly abuse notation and apply the Ito formula directly to $\varphi_\e(t,y)$.
Finally, the following lemma is the result of the geometric assumptions on $D$ that we will use in the context of Ito's formula.
\begin{lemma} \label{lemmauu2ny} Let $D$ and $D_{\epsilon}$ be as in the statement of Lemma \ref{lemmauu1} and let $\varphi_\e(t,y)$ be defined as
$$\varphi_\epsilon(t,y):= (d(y,D_{\epsilon,t}))^2 = |y-\pi_\epsilon(t,y)|^2,\ t\in [0,T],\ y\in\mathbb R^d.$$
Then
\begin{eqnarray} \label{lemmauu2+-}
&(i) &\partial_t\varphi_\epsilon(t,y)\geq 0, \notag \\
&(ii) &\partial^2_{y_iy_j}\varphi_\epsilon(t,y)\xi_i\xi_j\geq 0,
\end{eqnarray}
whenever $t\in [0,T],\ y, \xi \in\mathbb R^d$, and
\begin{equation} \label{lemmauu2+-+}
\varphi_{\e}(t,y+z) - \varphi_{\e}(t,y) - \langle \nabla \varphi_{\e}(t,y) ,z \rangle\geq 0
\end{equation}
whenever $t\in [0,T],\ y,z\in\mathbb R^d$.
\end{lemma}
\begin{proof}
\eqref{lemmauu2+-} $(i)$ follows from \eqref{timedep+2} and \eqref{lemmauu2+-} $(ii)$ follows from the convexity of $D_{\e,t}$. Finally, Taylors formula and \eqref{lemmauu2+-} $(ii)$ yields \eqref{lemmauu2+-+}.
\end{proof}
\section{Estimates for approximating problems: technical lemmas} \label{sec:lemmas}
\noindent
To prove the existence part of Theorem \ref{maint1} we use the method of penalization. Indeed, for each $n\geq 1$, we construct a quadruple $(Y_t^n, Z_t^n, U_t ^n, \Lambda_t^n)$ through penalization and we prove that $(Y_t^n, Z_t^n, U_t ^n, \Lambda_t^n)$ converges, as $n\to \infty$, to a solution $(Y_t, Z_t, U_ t, \Lambda_t)$ of the reflected backward stochastic differential equation, with reflection in the inward spatial normal direction, in $D$, and with data $(\xi,f),$ as defined in Definition \ref{rbsde}. Uniqueness is then proved by an argument based on Ito's formula. In the following we let $(\xi,f)$ be as in \eqref{data}-\eqref{data++}, and we let $D\subset\mathbb{R}^{d+1}$ be a time-dependent domain satisfying \eqref{timedep+}~-~\eqref{timedep+2}. Furthermore, we let $c$ denote a generic constant which may change value from line to line.
\subsection{Construction of $(Y_t^n, Z_t^n, U_t^n, \Lambda_t^n)$}
Let, for any $n\in\mathbb Z_+$,
\begin{eqnarray}\label{ffa1}
f_n(t,y,z,u):=f(t,y,z,u)-n(y-\pi(t,y)).
\end{eqnarray}
Then, for $n$ fixed, $f_n$ satisfies \eqref{data++} since $\pi$ has the Lipschitz property in space, see Lemma \ref{lemmaa} $(iii)$. Hence, using results concerning existence and uniqueness for (unconstrained) BSDEs driven by Wiener-Poisson type processes, see Lemma 2.4 in \cite{TL}, we can conclude that there exist, for each $n\in\mathbb Z_+$, a unique
triple $(Y_t^n, Z_t^n, U_t^n)$ and a constant $c_n$, $1\leq c_n<\infty$, such that
\begin{eqnarray}\label{ffa}(i)&&E \left [\sup_{0 \leq t \leq T} |Y_t^n|^2 \right] \leq c_n,\notag\\
(ii)&& E \left [\int_0 ^T \|Z^n_t
\|^2 dt + \int _0^T \int _U|U^n_s(e)|^2 \lambda(de)ds \right] < \infty,
\notag\\
(iii)&&Y^n_t= \xi + \int_t ^T f_n(s,Y^n_s,Z^n_s,U^n_s) ds\notag\\
&&- \int_t ^T Z^n_s dW_s- \int_t^T \int _U U^n_s(e) \mu(de,ds).
\end{eqnarray}
Note also that from \cite{TL} we have $Y^n \in \mathcal{D}\left( \left[ 0,T\right] , \R^{d}\right) $. Given $(Y_t^n, Z_t^n, U_t ^n)$ we define, for $n\in\mathbb Z_+$, the process $\Lambda_t^n$ through
\begin{equation}\label{ffa3}
\Lambda_t ^n =-n \int_0^t ( Y_s ^n-\pi(s,Y_s ^n)) ds.
\end{equation}
Note that
\begin{equation*
\Lambda_t ^n =\int_0^t \frac{-( Y_s ^n-\pi(s,Y_s ^n))}{| Y_s ^n-\pi(s,Y_s ^n)|}d|\Lambda^n|_s
\end{equation*}
and that $-( Y_s ^n-\pi(s,Y_s ^n))/{| Y_s ^n-\pi(s,Y_s ^n)|}$ is an element in the inwards directed normal cone to $D_s$ at
$\pi(s,Y_s ^n)\in\partial D_s$. Furthermore, using \eqref{ffa1} and \eqref{ffa3} we see that \eqref{ffa} $(iii)$ can be rewritten as
\begin{eqnarray*
Y_t ^n&=& \xi +\int_t ^T f(s, Y_s ^n, Z_s ^n, U_s ^n) ds +\Lambda^n_T-\Lambda_t^n\notag\\
&&-\int _t ^T Z_s^n dW_s - \int _t^T \int _U U_s^n(e) \mu(de,ds),
\end{eqnarray*}
for all $t\in [0,T]$. Recall that $Y^n, \Lambda^n, U^n$, $W_t$ and $Z^n$ are multi-dimensional processes. In particular, $Y_t^n=(Y_t ^{1,n}, \dots,Y_t ^{d,n})$, $\Lambda_t^n=(\Lambda_t ^{1,n},\dots,\Lambda_t ^{d,n})$, $U_t^n=(U_t ^{1,n},\dots, U_t ^{d,n})$, $W_t=(W_t^1,\dots,W_t^n)$ and $ Z_t^n$ is a $d\times n$-matrix with entries $Z_t^{i,j,n}$ and columns $Z_t^{j,n}$.
\subsection{A priori estimates for $(Y_t^n, Z_t^n, U_t^n, \Lambda_t^n)$}
\begin{lemma}\label{ll1} There exists a constant $c$, $1\leq c<\infty$, independent of $n$, such that
\begin{eqnarray*}
(i)&&E \left [\sup_{0 \leq t \leq T}|Y_t^n|^2 \right ] \leq c,\notag\\
(ii)&&E \left [\int _t ^T \|Z_s^n\|^2 + \int _t^T \int _U |U_s^n(e)|^2 \lambda(de)ds \right ]\leq c,\ t\in[0,T],\\
(iii)&& E \left [n\int _t ^T|Y_s^n-\pi(s,Y_s ^n)|ds \right ]\leq c, \ t\in[0,T].
\end{eqnarray*}
\end{lemma}
\begin{proof} Let $P_T\in D_T$ be as in Lemma \ref{lemmaa} $(iv)$. Applying Ito's formula to the process $|Y_t ^n-P_T|^2$ we deduce, for $t\in [0,T]$, that
\begin{eqnarray}\label{jaj}
&&|Y_t ^n-P_T|^2 + \int _t ^T
\|Z_s^n\|^2 ds + \int _t ^T \int _U |U_s^n(e)|^2 p(de, ds) \notag \\
&=&|\xi-P_T|^2 + 2 \int _t ^T \langle Y_s ^n-P_T, f(s,Y_s ^n,Z_s ^n,U_s^n)\rangle ds - 2 \int _t ^T \langle Y_s ^n-P_T, Z_s^n dW_s \rangle\notag \\
&&- 2 \int _t ^T \int _U \langle Y_s ^n-P_T, U^n_s(e) \rangle \mu(de,ds) - 2 n \int _t ^T \langle Y_s ^n-P_T,Y_s^n-\pi(s,Y_s ^n)\rangle ds.
\end{eqnarray}
Let
\begin{eqnarray*
A_n:=|Y_t ^n-P_T|^2 + \int _t ^T \|Z_s^n\|^2 ds + \int _t ^T \int _U |U_s^n(e)|^2 p(de, ds).
\end{eqnarray*}
Rearranging \eqref{jaj}, we find that
\begin{eqnarray*
&&A_n + 2 n \int _t ^T \langle Y_s ^n-P_T,Y_s^n-\pi(s,Y_s ^n)\rangle ds \notag\\
&=& |\xi-P_T|^2 + 2 \int _t ^T \langle Y_s ^n-P_T, f(s,Y_s ^n,Z_s ^n,U_s^n)\rangle ds\notag\\
&&- 2 \int _t ^T \langle Y_s ^n-P_T, Z_s^n dW_s\rangle - 2 \int _t ^T \int _U \langle Y_s^n-P_T, U^n_s(e) \rangle \mu(de,ds).
\end{eqnarray*}
Furthermore, using Lemma \ref{lemmaa} $(iv)$ we see that
\begin{eqnarray}\label{jaj++ny}
&&A_n + 2 \gamma^{-1} n \int _t ^T |Y_s^n-\pi(s,Y_s ^n)|ds \notag\\
&\leq& |\xi-P_T|^2 + 2 \int _t ^T \langle Y_s ^n-P_T, f(s,Y_s ^n,Z_s ^n,U_s^n)\rangle ds\notag\\
&&- 2 \int _t ^T \langle Y_s ^n-P_T, Z_s^n dW_s\rangle+ 2 \int _t ^T \int _U \langle Y_s^n-P_T, U^n_s(e) \rangle \mu(de,ds).
\end{eqnarray}
Next, taking expectations in \eqref{jaj++ny} and using the fact that $\mu$ is a martingale measure, we can conclude that
\begin{eqnarray}\label{eq:inequal1}
&&E \left [A_n+2 \gamma^{-1} n \int _t ^T |Y_s^n-\pi(s,Y_s ^n)|ds \right ] \leq I_{t,T}
\end{eqnarray}
where \begin{eqnarray}\label{jaj++1}
I_{t,T}=E \left [ |\xi-P_T|^2 \right] + 2 E\left [\int _t ^T \langle Y_s ^n-P_T, f(s,Y_s ^n,Z_s ^n, U_s^n)\rangle ds\right].
\end{eqnarray}
Using the Lipschitz character of $f$, \eqref{data++} $(iii)$, and the inequality $ab \leq \eta a^2 + \frac{b^2}{4\eta}$ it follows that we can estimate $I_{t,T}$ as,
\begin{eqnarray}\label{jaj++1+}
I_{t,T}&\leq& c\left( 1+ E \left[\int _t^T \left( |f(s,P_T,0,0)|^2+(1+2\eta)|Y^n_s-P_T|^2 ds \right ) \right] \right)\notag \\
&+& c E\left[ \int _t ^T \frac{1}{\eta} \left (\|Z_s^n\|^2 + \int _U |U_s^n(e)|^2 \lambda(de) \right) ds\right ]
\end{eqnarray}
where $\eta>0$ is a degree of freedom and $c$, $1\leq c<\infty$, is a constant depending on $\xi$, $f$ and $\mbox{diam($D_T$)}$. If we let $\eta$ be such that $c/\eta \leq1/2$, it follows from \eqref{eq:inequal1} and \eqref{jaj++1+}, after recalling the definition of $A_n$, that
\begin{eqnarray}\label{jaj++ha}
&&E[|Y_t ^n-P_T|^2]+ \frac 1 2 E \left [\int _t ^T \|Z_s^n\|^2 ds\right ] + \frac{1}{2}E \left [\int _t ^T \int _U |U_s^n(e)|^2 \lambda(de) ds \right ] \notag \\
&\leq& c \left( 1+ E \left [\int _t^T |Y^n_s-P_T|^2 ds \right ]\right).
\end{eqnarray}
Using \eqref{jaj++ha} and Gronwall's lemma we deduce that
\begin{equation}\label{ffa2--}
\sup_{0 \leq t \leq T}E[|Y_t^n-P_T|^2] \leq \tilde c\exp(\tilde c T)
\end{equation}
where $\tilde c$ is independent of $n$. In particular, we can conclude that there exists $\hat c$, $1\leq\hat c<\infty$, independent of $n$, such that
\begin{eqnarray}\label{ffa2--+}
(i)&&\sup_{0 \leq t \leq T}E[|Y_t^n|^2] \leq \hat c, \quad \mbox {and}\notag\\
(ii)&&\ E \left [\int _0 ^T \|Z_s^n\|^2 ds + \int _0 ^T\int _U |U_s^n(e)|^2 \lambda(de) ds \right]\leq\hat c.
\end{eqnarray}
We next prove that there exists $\check c$, $1\leq \check c<\infty$, independent of $n$, such that
\begin{eqnarray}\label{ffa2--++}
E[\sup_{0 \leq t \leq T}|Y_t^n|^2] \leq \check c.
\end{eqnarray}
To do this we first note, using \eqref{jaj++ny} and the Lipschitz character of
$f$, that, for a constant $c$, $1\leq c<\infty$,
\begin{eqnarray}\label{eq:BDGdoob}
&&c^{-1}|Y_t^n-P_T|^2\leq |\xi-P_T|^2+\biggl |\int _t ^T \langle Y_s ^n-P_T, f(s,P_T,0,0)\rangle ds\biggr |\notag\\
&&+\biggl |\int _t ^T |Y_s ^n-P_T|(|Y_s ^n-P_T|+\|Z_s ^n\|+\|U_s^n\|_2)ds\biggr |\notag\\
&&+\biggl |\int _t ^T \langle Y_s ^n-P_T, Z_s^n dW_s\rangle\biggr |+ \biggl | \int _t ^T \int _U \langle Y_s^n-P_T, U^n_s(e) \rangle \mu(de,ds)\biggr |.
\end{eqnarray}
We treat the last two terms on the right hand side of \eqref{eq:BDGdoob} using H\"{o}lders inequality and the Burkholder-Davis-Gundy inequality. Indeed, applying these yields
\begin{eqnarray}\label{jaj++nyhha}
&&E\left [\sup_{0 \leq t \leq T}\left |\int _t ^T \langle Y_s ^n-P_T, Z_s^n dW_s\rangle\right |\right]\notag\\
&&\leq \biggl (E \left [\sup_{0 \leq t \leq T}|Y_t ^n-P_T|^2\right ]\biggr )^{1/2}\biggl (E\biggl[\sup_{0 \leq t \leq T}\biggl |\int _t ^T Z_s^n dW_s \biggr |^2
\biggr ]\biggr )^{1/2}\notag\\
&&\leq c\biggl (E \left [\sup_{0 \leq t \leq T}|Y_t ^n-P_T|^2 \right ]\biggr )^{1/2}\biggl (E \left [\int _0 ^T \|Z_s^n\|^2 ds\right ]\biggr )^{1/2}.
\end{eqnarray}
Similarly,
\begin{eqnarray*
&&E\left [\sup_{0 \leq t \leq T}\biggl |\int _t ^T \int _U \langle Y_s^n-P_T, U^n_s(e) \rangle \mu(de,ds)\biggr |\right]\notag\\
&&\leq c\biggl (E \left [\sup_{0 \leq t \leq T}|Y_t ^n-P_T|^2 \right ]\biggr )^{1/2}\biggl (E \left [\int _0 ^T\int _U |U_s^n(e)|^2 \lambda(de) ds\right ]\biggr )^{1/2}.
\end{eqnarray*}
Using the above estimates as well as \eqref{ffa2--}, \eqref{ffa2--+} and assumption \eqref{timedep+} we get after taking expectation in \eqref{eq:BDGdoob} that
\begin{eqnarray*}
&&E \left [ \sup_{0\leq t \leq T}|Y_t^n-P_T|^2 \right ] \leq c \biggl (1+ \left ( E [\sup_{0 \leq t \leq T}|Y_t ^n-P_T|^2 ] \right )^{1/2} \biggr)
\end{eqnarray*}
from which we conclude that $E \left [ \sup_{0\leq t \leq T}|Y_t^n-P_T|^2 \right] \leq c$ for some constant $1 \leq c < \infty$ and, consequently, that \eqref{ffa2--++} holds.
Finally, starting from \eqref{jaj++ny} and repeating the arguments above we also deduce that
\begin{eqnarray*
E \left [n\int _t ^T|Y_s^n-\pi(s,Y_s ^n)|ds \right ]\leq\hat c
\end{eqnarray*}
for some constant $\hat c$ independent of $n$. This completes the proof of Lemma \ref{ll1}.\end{proof}
\subsection{Uniform control of $d(Y_t^n,D_t)$}
We here prove the following lemma.
\begin{lemma}\label{lemmbb} Let $D\subset\mathbb{R}^{d+1}$ be a time-dependent domain satisfying \eqref{timedep+}~-~\eqref{timedep+2}. Let, for $\epsilon>0$ small, $D_{\epsilon}$
be as in Lemma \ref{lemmauu1-}. Then there exist $\epsilon_0>0$ and $c$, $1\leq c<\infty$, both independent of $n$, such that
\begin{eqnarray*}
(i)&&E \left [\sup{_{0\leq t\leq T}d(Y_t^n,D_{\e,t})^2} \right ] \leq c\biggl ( \frac 1 n +\epsilon+n\epsilon^2\biggr ),\notag\\
(ii)&&E \left [\int_0^T(d(Y_t^n,D_{\e,t}))^2dt \right ] \leq c\biggl ( \frac 1 {n^2} +\frac \epsilon n+\epsilon^2\biggr ),
\end{eqnarray*}
whenever $0<\epsilon<\epsilon_0$ and $n\geq 1$.
\end{lemma}
\begin{proof} Let $\varphi_\epsilon(t,Y_t)=d(Y_t^n, D_{\e,t})^2=|Y_t^n-\pi_\epsilon(t,Y_t^n)|^2$. Then $\nabla_y\varphi_\epsilon(t,Y_t)=2(Y_t^n-\pi_\epsilon(t,Y_t^n))$. Using the
Ito formula of Lemma \ref {lemmauu2+} we see that
\begin{eqnarray}\label{aaa-li}
&&\varphi_\e(t,Y_t^n) + \int_t^T (\partial_{s}\varphi_\epsilon)(s,Y_{s}^n)ds+ \frac{1}{2} \int_t^T \sum_{i,j} (Z_s^{n} Z_s^{n,\ast})_{ij} \partial^2_{y_iy_j}\varphi_\epsilon(s, Y_{s}) ds \notag\\
&&+ \int _t ^T \int _U [\varphi_{\e}(s,Y_{s^-}^n+U^n_s(e)) - \varphi_{\e}(Y^n_{s^-}) - \langle \nabla \varphi_{\e}(s,Y^n_{s^-}) , U^n_s(e) \rangle ] p(de,ds) \notag \\
&&=\varphi_\epsilon(T,\xi)+I_1+I_2+I_3 + I_4,
\end{eqnarray}
where
\begin{eqnarray}\label{aaa+}
I_1&=&2\int _t ^T \langle Y_s^n-\pi_\epsilon(s,Y_s^n),f(s,Y_s,Z_s^n, U_s^n)\rangle ds,\notag\\
I_2&=& - 2\int _t ^T \langle Y_s^n-\pi_\epsilon(s,Y_s^n),Z^n_s
dW_s \rangle,\notag\\
I_3&=&- 2n \int_t ^T\langle Y_s^n-\pi_\epsilon(s,Y_s^n),Y_s^n-\pi(s,Y_s^n)\rangle ds, \notag \\
I_4&=&- 2 \int_t ^T \int _U \langle Y_s^n-\pi_\epsilon(s,Y_s^n), U^n_s(e) \rangle \mu(de,ds).
\end{eqnarray}
Using Lemma \ref{lemmauu2ny} we see that
\begin{eqnarray*
&&\int_t^T \partial_{s}\varphi_\epsilon(s,Y_s^n)ds+ \frac{1}{2} \int_t^T \sum_{i,j} (Z_s^{n} Z_s^{n,\ast})_{ij} \partial^2_{y_iy_j}\varphi_\epsilon(s, Y_s) ds \notag\\
&&+ \int _t ^T \int _U [\varphi_{\e}(s,Y_{s^-}^n+U^n_s(e)) - \varphi_{\e}(Y^n_{s^-} ) - \langle \nabla \varphi_{\e}(s,Y^n_{s^-}) , U^n_s(e) \rangle ] p(de,ds)\geq 0,
\end{eqnarray*}
and hence
\begin{eqnarray}\label{aaa}
\varphi_\e(t,Y_t^n)\leq \varphi_\epsilon(T,\xi)+I_1+I_2+I_3 + I_4.
\end{eqnarray}
Since $\xi \in D_T$ a.s. we see, using Lemma \ref{lemmauu1}, that $\varphi _\e(T,\xi)\leq c \e^2$ a.s. To simplify the notation in what follows, we define $\chi_\epsilon (t,y) : [0,T]\times\mathbb R^d \to \{0,1\}$ as
\begin{equation*}
\chi_\epsilon(t,y) =\begin{cases}1& \mbox{ if } d(y,D_{\epsilon,t})>\epsilon \\
0 & \mbox{otherwise}
\end{cases}.
\end{equation*}
We first focus on the term $I_1$ in \eqref{aaa+}. Then, using the above introduced notation we see that
\begin{eqnarray*
I_1&\leq &2\int _t ^T |\varphi_\epsilon(s,Y_s^n)|^{1/2}|f(s,Y_s^n,Z_s^n, U^n_s)|\chi_\epsilon(s,Y_s^n)ds\notag\\
&&+2\int _t ^T |\varphi_\epsilon(s,Y_s^n)|^{1/2}|f(s,Y_s^n,Z_s^n, U_s^n)|(1-\chi_\epsilon(s,Y_s^n))ds.
\end{eqnarray*}
Furthermore, by the inequality $ab \leq \eta a^2 + \frac{b^2}{4\eta}$ and $x \leq \max\{1,x^2\}$,
\begin{eqnarray*
(i)&& 2|\varphi_\epsilon(s,Y_s^n)|^{1/2}|f(s,Y_s^n,Z_s^n, U^n_s)|\chi_\epsilon(s,Y_s^n)\notag\\
&\leq&
\frac{n}{4} \varphi_\epsilon(s,Y_s^n) \chi_\epsilon(s,Y_s^n) + \frac{4}{n} |f(s,Y_s^n,Z_s^n, U^n_s)|^2 \chi_\epsilon(s,Y_s^n),\notag\\
(ii)&& 2|\varphi_\epsilon(s,Y_s^n)|^{1/2}|f(s,Y_s^n,Z_s^n, U^n_s)|(1-\chi_\epsilon(s,Y_s^n))\notag\\
&\leq&2({\e}+ {\e} |f(s,Y_s^n,Z_s^n,U^n_s)|^2)(1-\chi_\epsilon(s,Y_s^n)).
\end{eqnarray*}
Next, focusing on the term $I_3$, we have by the bilinearity of $\langle \cdot, \cdot\rangle$ that
\begin{eqnarray}\label{aaa++}
I_3&=&-2n \int _t ^T|Y_s^n-\pi_\epsilon(s,Y_s^n)|^2\chi_\epsilon(s,Y_s^n) ds\notag\\
&&- 2n \int _t ^T\langle Y_s^n-\pi_\epsilon(s,Y_s^n),\pi_\epsilon(s,Y_s^n)-\pi(s,Y_s^n)\rangle\chi_\epsilon(s,Y_s^n) ds\notag\\
&&- 2n \int _t ^T\langle Y_s^n-\pi_\epsilon(s,Y_s^n),Y_s^n-\pi(s,Y_s^n)\rangle(1-\chi_\epsilon(s,Y_s^n)) ds\notag\\
&:=&I_{31}+I_{32}+I_{33}.
\end{eqnarray}
By Lemma \ref{lemmauu1} $(i)$ we immediately see that $|I_{33}|\leq c n\epsilon^2$. Furthermore, using Lemma \ref{lemmauu2} $(ii)$ we see that
\begin{eqnarray}\label{aaa+++}
|I_{32}|\leq c\sqrt{\epsilon} n \int _t ^T(d(Y_s^n,D_{\epsilon,s}))^{3/2}\chi_\epsilon(s,Y_s^n) ds.
\end{eqnarray}
Using the inequality \( ab \leq \frac{3a^{\frac{4}{3}}}{4}+\frac{b^4}{4} \) with
$a= d(Y^n_s, D_{\e,s})^{\frac{3}{2}}$ and $b=c\sqrt{\epsilon}$ we deduce from
\eqref{aaa+++} that
\begin{eqnarray}\label{aaa+++jj}
|I_{32}|\leq cn{\epsilon^2}+n \int _t ^T(d(Y_s^n,D_{\epsilon,s}))^{2}\chi_\epsilon(s,Y_s^n) ds.
\end{eqnarray}
Putting the estimate \eqref{aaa+++jj} into \eqref{aaa++} together we can conclude that
\begin{eqnarray*}
I_3\leq cn\e^2-n \int _t ^T|Y_s^n-\pi_\epsilon(s,Y_s^n)|^2\chi_\epsilon(s,Y_s^n) ds.
\end{eqnarray*}
Hence, putting the estimate for $I_1$ and $I_3$ together we can conclude that
\begin{eqnarray}\label{aaad}
I_1+I_3&\leq & c\epsilon+cn\epsilon^2-\frac 3 4n \int _t ^T\varphi_\epsilon(s,Y_s^n)\chi_\epsilon(s,Y_s^n) ds\notag\\
&&+c\int _t ^T\biggl (\frac 1 n+\epsilon\biggr ) |f(s,Y_s^n,Z_s^n, U^n_s)|^2 ds.
\end{eqnarray}
Combining \eqref{aaa} and \eqref{aaad} we have proved that
\begin{eqnarray} \label{aaad+}
\varphi_\e(t,Y_t^n)&\leq& c( \e+n\e^2) -\frac{3}{4} n \int _t ^T \varphi_\e(s,Y^n_s) \chi_\e(s,Y^n_s) ds\notag\\
&&+ c\int _t^T(\frac{1}{n} + \e) |f(s,Y^n_s, Z^n_s, U^n_s)|^2 ds \notag \\
&&- 2\int _t ^T \langle Y_s^n-\pi_\epsilon(s,Y_s^n),Z^n_s
dW_s \rangle\notag\\
&&-2\int_t ^T \int _U \langle Y_s^n-\pi_\epsilon(s,Y_s^n), U^n_s(e) \rangle \mu(de,ds).
\end{eqnarray}
In particular,
\begin{eqnarray} \label{aaad+ia}
&&\varphi_\e(t,Y_t^n)+\frac{3}{4}n \int _t ^T \varphi_\e(s,Y^n_s) \chi_\e(s,Y^n_s) ds\notag\\
&\leq& c( \e+n\e^2)+ c\int _t^T(\frac{1}{n} + \e) |f(s,Y^n_s, Z^n_s, U^n_s)|^2 ds \notag \\
&&-2\int _t ^T \langle Y_s^n-\pi_\epsilon(s,Y_s^n),Z^n_s
dW_s \rangle-2\int_t ^T \int _U \langle Y_s^n-\pi_\epsilon(s,Y_s^n), U^n_s(e) \rangle \mu(de,ds).
\end{eqnarray}
where $c$ is independent of $\e$ and $n$. The estimate in Lemma \ref{lemmbb} $(ii)$ now follows from taking expectation in \eqref{aaad+ia} and using the Lipschitz property of $f$ and
Lemma \ref{ll1} $(i)$ and $(ii)$. Similarly, using \eqref{aaad+} we can conclude that
\begin{eqnarray*
\sup_{0\leq t\leq T}E[\varphi_\e(t,Y^n_t)] \leq c\bigl (\frac{1}{n} +\e + n\e^2\bigr),
\end{eqnarray*}
for a constant $c$, independent of $\e$ and $n$. Since $\varphi_\e(t, Y_t) \geq 0$ for all $t \in [0,T]$ we can, repeating the arguments above, also conclude from \eqref{aaa-li} that
\begin{eqnarray}\label{eq:allests}
&& E\left [ \int_t^T (\partial_{s}\varphi_\epsilon)(s,Y_{s}^n)ds \right ]+ E \left [\int_t^T \sum_{i,j} (Z_s^{n} Z_s^{n,\ast})_{ij} \partial^2_{y_iy_j}\varphi_\epsilon(s, Y_{s}) ds \right ] \notag\\
+&&E\left [ \int _t ^T \int _U [\varphi_{\e}(s,Y_{s^-}^n+U^n_s(e)) - \varphi_{\e}(Y^n_{s^-}) - \langle \nabla \varphi_{\e}(s,Y^n_{s^-}) , U^n_s(e) \rangle ] p(de,ds) \right ] \notag \\
&& \leq c\bigl (\frac{1}{n} +\e + n\e^2\bigr),
\end{eqnarray}
for some constant $1\leq c < \infty$, for all $t \in [0,T]$.
Once again using the Lipschitz property of $f$ and Lemma \ref{ll1} in \eqref{aaad+ia} we see that to complete the proof of Lemma \ref{lemmbb} $(i)$ it remains to control the terms
\begin{eqnarray}\label{eq:finests}
&(i)&E \left [\sup_{0\leq t \leq T} \left | \int _t ^T \langle Y_s^n-\pi_\epsilon(s,Y_s^n),Z^n_s dW_s \rangle \right | \right ]
\notag \\
&(ii)&E \left [ \sup_{0\leq t \leq T}\left |\int_t ^T \int _U \langle Y_s^n-\pi_\epsilon(s,Y_s^n), U^n_s(e) \rangle \mu(de,ds) \right | \right ]
\end{eqnarray}
We first treat \eqref{eq:finests} $(i)$. As in \eqref{jaj++nyhha} we use the Burkholder-Davis-Gundy inequality to see that
\begin{align}\label{eq:nuzest}
&E \left [\sup_{0\leq t \leq T} \left | \int _t ^T \langle Y_s^n-\pi_\epsilon(s,Y_s^n),Z^n_s dW_s \rangle \right | \right ] &\notag \\
\leq & E \left [ \left ( \int _0 ^T \left |(Y_s^n-\pi_\epsilon(s,Y_s^n))^\ast Z^n_s \right|^2 ds \right )^{1/2} \right ].
\end{align}
By the convexity of $\varphi_\e$ and equivalence of Euclidean norms (see \cite{GP}~p.~115) we have that
$$
\frac {|(Y_s^n-\pi_\e(s,Y_s^n))^\ast Z^n_s|^2}{|(Y_s^n-\pi_\e(s,Y_s^n))|^2} \I_{\{Y^n_s \not \in D_{e,s}\}}\leq c \left ( \sum_{i,j} (Z_s^{n} Z_s^{n,\ast})_{ij} \partial^2_{y_iy_j}\varphi_\epsilon(s, Y_{s}) \right),
$$
where $\I$ is the indicator function, i.e., $\I_{\{Y^n_s \not \in D_{\e,s}\}} = 1$ if $Y^n_s \not \in D_{\e,s}$ and $0$ otherwise. Hence, it follows from \eqref{eq:allests} that for $t\in [0,T]$
\begin{equation}\label{eq:nuzest1+}
\int_t ^T \frac{|(Y_s^n-\pi_\e(s,Y_s^n))^\ast Z^n_s|^2}{|(Y_s^n-\pi_\e(s,Y_s^n))|^2} \I_{\{Y^n_s \not \in D_{e,s}\}}ds\leq c\bigl (\frac{1}{n} +\e + n\e^2\bigr).
\end{equation}
Note also that $\left | Y_s^n-\pi_\epsilon(s,Y_s^n) \right |= \left |Y_s^n-\pi_\epsilon(s,Y_s^n) \right | \I_{\{Y^n_s \not \in D_{\e,s}\}}$ and that
\begin{align}\label{eq:nuzest2} \int _t ^T |(Y_s^n-\pi_\e(s,Y_s^n))^\ast Z^n_s|^2 ds & \notag \\
\leq \sup_{t\leq s \leq T}& \varphi_\e(s, Y^n_s) \int _t ^T \frac {| (Y_s^n-\pi_\e(s,Y_s^n))^\ast Z^n_s|^2}{|(Y_s^n-\pi_\e(s,Y_s^n))|^2} \I_{\{Y^n_s \not \in D_{e,s}\}} ds.
\end{align}
We again use $ab \leq \eta a^2 + \frac{b^2}{\eta}$ to conclude from \eqref{eq:nuzest}, \eqref{eq:nuzest1+} and \eqref{eq:nuzest2} that
\begin{align}\label{eq:wrty}
&E \left [\sup_{0\leq t \leq T} \left | \int _t ^T \langle Y_s^n-\pi_\epsilon(s,Y_s^n),Z^n_s dW_s \rangle \right | \right ] \notag \\
\leq &\eta E \left [\sup_{0 \leq t \leq T} \varphi_\e(t, Y^n_t) \right ] +c_\eta(\frac{1}{n} +\e + n\e^2\bigr)
\end{align}
for some constant $c_\eta <\infty$ depending on the degree of freedom $\eta >0$.
We now treat the term \eqref{eq:finests} $(ii)$. Using Taylor's theorem we see that
\begin{align*}E\left [ \int _t ^T \int _U [\varphi_{\e}(s,Y_{s^-}^n+U^n_s(e)) - \varphi_{\e}(Y^n_{s^-}) - \langle \nabla \varphi_{\e}(s,Y^n_{s^-}) , U^n_s(e) \rangle p(de,ds) \right ]& \notag \\
\geq \breve c E \left [ \int_t^T |U(e)|^2 \lambda(de) ds \right ],
\end{align*}
for some constant $\breve c\geq 0$. Furthermore, by the strong convexity of $\varphi_\e$, and this is a consequence of \eqref{timedep+}, there exists a constant $\kappa >0$ such that $\breve c\geq \kappa >0$. Therefore we can, in a way similar to the above, conclude that
\begin{align}\label{eq:wrty2}
&E \left [ \sup_{0\leq t \leq T}\left |\int_t ^T \int _U \langle Y_s^n-\pi_\epsilon(s,Y_s^n), U^n_s(e) \rangle \mu(de,ds) \right | \right ] \notag \\
\leq & E \left [\left (\int_t ^T \int _U |Y_s^n-\pi_\epsilon(s,Y_s^n)|^2 |U^n_s(e)|^2 \lambda(de) ds \right) ^{1/2} \right ] \notag \\
\leq & E \left [ \left ( \sup_{t\leq s \leq T}\varphi_\e(s,Y^n_s) \int_t ^T \int _U |U^n_s(e)|^2 \lambda(de) ds \right) ^{1/2}\right ] \notag \\
\leq &\eta E \left [\sup_{0 \leq t \leq T} \varphi_\e(t, Y^n_t) \right ] +c_\eta(\frac{1}{n} +\e + n\e^2\bigr).
\end{align}
Once again, $c_\eta <\infty$ is a constant depending on the degree of freedom $\eta>0$.
The proof of Lemma \ref{lemmbb} is now completed by choosing $\eta$ small enough (in analogy with \eqref{jaj++1+},\eqref{jaj++ha}) and combining \eqref{aaad+ia} with \eqref{eq:wrty}, \eqref{eq:wrty2}.
\end{proof}
\begin{lemma}\label{lemmbbaa} Let $D\subset \R^{d+1}$ be a time-dependent domain satisfying \eqref{timedep+}~-~\eqref{timedep+2}. Then there exists $c$, $1\leq c<\infty$, independent of $n$ such that
\begin{eqnarray*}
(i)&&E \left [\sup_{0\leq t\leq T}(d(Y_t^n,D_{t}))^2 \right] \leq \frac{ c} n,\notag\\
(ii)&&E \left [\int_0^T(d(Y_t^n,D_{t}))^2dt \right] \leq \frac {c} {n^2},
\end{eqnarray*}
whenever $n\geq 1$.
\end{lemma}
\begin{proof} Let, for $\epsilon>0$ small, $D_{\epsilon}$
be as in Lemma \ref{lemmauu1-}. Then, using Lemma \ref{lemmauu1-} we have
\begin{eqnarray*
h(D_{t},D_{\epsilon,t})<\epsilon\mbox{ for all $t\in[0,T]$.}
\end{eqnarray*}
Hence,
\begin{eqnarray*
d(Y_t^n,D_{t})\leq d(Y_t^n,D_{\epsilon,t})+\epsilon\mbox{ for all $t\in[0,T]$.}
\end{eqnarray*}
Applying Lemma \ref{lemmbb} and letting $\e \to 0$ completes the proof.
\end{proof}
\subsection{$(Y_t^n, Z_t^n, U_t^n)$ is a Cauchy sequence }
\begin{lemma}\label{anotherlemma} There exists a constant $c$ such that the following holds whenever $m, n \in \mathbb{Z}^+$:
\begin{eqnarray*}
E \left [\sup_{0 \leq t \leq T} |Y_t^n-Y_t^m|^2 + \int_0 ^t \|Z_t^n-Z_t^m\|^2 dt\right ] &\leq& c \left(\frac{1}{n} + \frac{1}{m}\right),\\
E\left [\int_0 ^T \int _U |U_s^n(e)-U^m_s(e)| ^2 \lambda(de) ds\right] &\leq& c \left(\frac{1}{n} + \frac{1}{m}\right).
\end{eqnarray*}
\end{lemma}
\begin{proof}Applying Ito's formula to $|Y^n_t-Y^m_t|^2$ we get that
\begin{eqnarray}\label{eq:itoexpmn}
&&|Y^n_t-Y^m_t|^2+\int _t ^T \|Z_s^n-Z_s^m\|^2 ds\notag\\
&& + \int _t ^T \int _U |U^n_s(e)- U^m_s(e)| ^2 p(de,ds) \notag\\
&=& 2 \int _t ^T \langle Y^n_s - Y^m_s , f(s,Y^n_s, Z^n_s, U^n_s) - f(s,Y^m_s, Z^m_s, U^m_s) \rangle ds \notag \\
&&- 2 \int _t ^T \langle Y^n_s - Y^m_s , (Z^n_s- Z^m_s) dW_s\rangle\notag\\
&&- 2 \int _t ^T \int _U \langle Y^n_s - Y^m_s , U^n_s(e)- U^m_s(e) \rangle \mu(de,ds) \notag \\
&&- 2n \int _t ^T \langle Y^n_s - Y^m_s , Y^n_s- \pi(s,Y^n_s) \rangle ds\notag\\
&& + 2m \int _t ^T \langle Y^n_s - Y^m_s , Y^m_s- \pi(s,Y^m_s) \rangle ds.
\end{eqnarray}
Hence, taking expectation and using the Lipschitz character of $f$ we deduce that
\begin{eqnarray}\label{apa1}
&&E \left [|Y^n_t-Y^m_t|^2 \right]+E\left [\int _t ^T \|Z_s^n-Z_s^m\|^2 ds\right ]\notag\\
&& + E\left [\int _t ^T \int _U |U^n_s(e)- U^m_s(e)| ^2 \lambda(de)ds\right] \notag\\
&\leq& c E\left [ \int _t ^T (|Y^n_s - Y^m _s |^2 + |Y^n_s - Y^m _s |\|Z^n_s - Z^m _s \|)ds\right ]\notag \\
&&+c E\left [ \int _t ^T |Y^n_s - Y^m _s | \int _U |U^n_s(e) - U^m _s(e)| \lambda(de) ds\right ] \notag \\
&&- 2E\left [ \int _t ^T \langle Y^n_s - Y^m_s , n(Y^n_s - \pi(s,Y^n_s)) - m(Y^m_s - \pi(s,Y^m_s)) \rangle ds\right ].
\end{eqnarray}
Note that
\begin{eqnarray*}
&&-\langle Y^n_s - Y^m_s , n(Y^n_s - \pi(s,Y^n_s)) - m(Y^m_s - \pi(s,Y^m_s)) \rangle \notag\\
&=&\langle Y^m_s - Y^n_s , n(Y^n_s - \pi(s,Y^n_s)) \rangle +\langle Y^n_s - Y^m_s , m(Y^m_s - \pi(s,Y^m_s)) \rangle.
\end{eqnarray*}
Using Lemma \ref{lemmaa} $(ii)$ we have that
\begin{eqnarray*}
&&\langle Y^m_s - Y^n_s , n(Y^n_s - \pi(s,Y^n_s)) \rangle\leq n \langle Y^m_s - \pi(s,Y^m_s) , Y^n_s - \pi(s,Y^n_s)\rangle \notag\\
&&\langle Y^n_s - Y^m_s , m(Y^m_s - \pi(s,Y^m_s)) \rangle\leq m \langle Y^n_s - \pi(s,Y^n_s) , Y^m_s - \pi(s,Y^m_s)\rangle.
\end{eqnarray*}
Furthermore,
\begin{eqnarray*}
&&2E\left [ \int _t ^T \langle Y^m_s - Y^n_s , n(Y^n_s - \pi(s,Y^n_s)) \rangle ds\right ]\notag\\
&\leq& 2nE\left [ \int _t ^T |Y^m_s - \pi(s,Y^m_s)|| Y^n_s - \pi(s,Y^n_s)|ds\right ]\notag\\
&\leq & nE\left [ \int _t ^T \beta(d(Y^m_s,D_s))^2+\beta^{-1}(d(Y^n_s,D_s))^2ds\right ]\notag\\
&\leq &c(n\beta m^{-2}+\beta^{-1} n^{-1})\leq cm^{-1}
\end{eqnarray*}
where we have used Lemma \ref{lemmbbaa} $(ii)$ and chosen the degree of freedom to equal $\beta=m/n$. This argument can be repeated with
$n\langle Y^m_s - Y^n_s , n(Y^n_s - \pi(s,Y^n_s)) \rangle$ replaced by $m \langle Y^n_s - \pi(s,Y^n_s) , Y^m_s - \pi(s,Y^m_s)\rangle $ resulting in the bound $ cn^{-1}$. Put together we can conclude that
\begin{eqnarray}\label{apa}
&&- 2E\left [ \int _t ^T \langle Y^n_s - Y^m_s , n(Y^n_s - \pi(s,Y^n_s)) - m(Y^m_s - \pi(s,Y^m_s)) \rangle ds \right ]\notag\\
&\leq &c(n^{-1}+m^{-1}).
\end{eqnarray}
Combining \eqref{apa1}, \eqref{apa} and using Cauchy's inequality as in \eqref{jaj++1+}, \eqref{jaj++ha} we can conclude that
\begin{eqnarray}\label{apa1+}
&&E \left [|Y^n_t-Y^m_t|^2 \right ]+ \frac{1}{2}E\left [\int _t ^T\|Z_s^n-Z_s^m\|^2 ds\right ]\notag\\
&& +\frac{1}{2} E\left [\int _t ^T \int _U |U^n_s(e)- U^m_s(e)| ^2 \lambda(de)ds\right ] \notag\\
&\leq& c E\left [ \int _t ^T |Y^n_s - Y^m _s |^2 ds\right ]+c(n^{-1}+m^{-1})
\end{eqnarray}
where $c$ is independent of $n$ and $m$. By Gronwall's inequality we then have, using \eqref{apa1+},
\begin{eqnarray}\label{apa1++}
E \left [|Y^n_t-Y^m_t|^2 \right ]\leq c(n^{-1}+m^{-1}).
\end{eqnarray}
Subsequently,
\begin{eqnarray}\label{apa1+a}
E\left [\int _t ^T\|Z_s^n-Z_s^m\|^2 ds\right ]&\leq& c(n^{-1}+m^{-1})\notag\\
E\left[\int _t ^T \int _U |U^n_s(e)- U^m_s(e)| ^2 \lambda(de)ds\right ]&\leq& c(n^{-1}+m^{-1}).
\end{eqnarray}
Using \eqref{apa1++}, \eqref{apa1+a}, Lemma \ref{lemmbbaa} $(i)$, and now familiar arguments based on the Burkholder-Davis-Gundy inequality, we can, starting from \eqref{eq:itoexpmn}, also deduce that
\begin{equation*}
E \left [ \sup_{0 \leq t \leq T} |Y^n_t -Y^m_t|^2 \right ] \leq c \left (\frac{1}{n}+\frac{1}{m}\right),
\end{equation*}
to complete the proof of Lemma \ref{anotherlemma}. We omit further details.
\end{proof}
\section{The final argument: proof of Theorem \ref{maint1}} \label{sec:proof}
\noindent
Using Lemma \ref{anotherlemma} we can conclude that $(Y_t^n,Z_t^n,U_t^n)$ is a Cauchy sequence in the space of progressively measurable processes $(Y_t,Z_t,U_t)$ satisfying
\begin{eqnarray*
E \left [\sup_{0 \leq t\leq T} |Y_t|^2 \right] +E \left [\int _0 ^T \|Z_t\|^2 ds \right] +E \left[\int _0 ^T \int _U |U_t(e)|^2 \lambda(de) ds \right]<\infty.
\end{eqnarray*}
Hence, taking a subsequence if necessary, we have a sequence $(Y_t^n,Z_t^n,U_t^n)_{n \geq 0}$ and a triple of processes $(Y_t,Z_t,U_t)$ such that
\begin{equation*}
Y_t=\lim _{n \to \infty} Y^n_t, \
Z_t=\lim _{n \to \infty} Z^n_t, \
U_t=\lim _{n \to \infty} U^n_t
\end{equation*}
in the sense that
\begin{eqnarray}\label{aadg}
(i)&&E \left [\sup_{0 \leq t\leq T} |Y_t^n-Y_t|^2 \right] \to 0,\notag\\
(ii)&&E\left [\int _0 ^T \|Z_t^n-Z_t\|^2 ds \right] \to 0, \notag \\
(iii)&& E \left [\int _0 ^T \int _U |U_t^n(e)-U_t(e)|^2 \lambda(de) ds \right] \to 0
\end{eqnarray}
as $n \to \infty$. Furthermore, by Lemma \ref{ll1} we have
\begin{eqnarray}\label{ffaagain}
(i)&&E \left[\sup_{0 \leq t \leq T} |Y_t|^2 \right] <\infty,\notag\\
(ii)&& E \left [\int_0 ^T \|Z_t\|^2 dt + \int _0^T \int _U|U_s(e)|^2 \lambda(de)ds \right ] < \infty.
\end{eqnarray}
Note that, as a uniform limit of c\`{a}dl\`{a}g functions $\{Y^n_t\}$, we immediately have that $Y_t\in \mathcal{D}\left( \left[ 0,T\right] ,\R^{d}\right)$ and, by Lemma \ref{lemmbbaa} $(i)$, we have $Y_t \in \overline D$. Recall that
\begin{equation*}
\Lambda_t^n= -n \int _0 ^t (Y^n_s- \pi(s,Y^n_s))ds=\int_0^t -\frac{( Y_s ^n-\pi(s,Y_s ^n))}{| Y_s ^n-\pi(s,Y_s ^n)|}d|\Lambda^n|_s
\end{equation*}
and that $-( Y_s ^n-\pi(s,Y_s ^n))/{| Y_s ^n-\pi(s,Y_s ^n)|}$ is an element in the inward directed normal cone to $D_s$ at
$\pi(s,Y_s ^n)\in\partial D_s$. Using \eqref{ffa} $(iii)$, \eqref{aadg} and \eqref{ffaagain} it follows that there exists $\Lambda_t$ such that
\begin{equation*}
\Lambda_t=\lim _{n \to \infty}\Lambda^n= \lim _{n \to \infty} -n \int _0 ^t (Y^n_s- \pi(s,Y^n_s))ds
\end{equation*}
in the sense that
\begin{equation*
E \left [\sup_{0 \leq t\leq T} |\Lambda_t^n-\Lambda_t|^2 \right ] \to 0.
\end{equation*}
Hence, as $(\Lambda^n_t(\omega))_{0\leq t \leq T}$ is continuous, $(\Lambda_t(\omega))_{0\leq t \leq T}$ is continuous in $t$ a.s.
\subsection{Existence: $(Y_t, Z_t, U_t, \Lambda_t)$ is a solution} We will now prove that the constructed quadruple $(Y_t,Z_t,U_t,\Lambda_t)$ is a solution to our original problem. We first note that, as a limit of $(Y^n_t, Z^n_t, U^n_t)$, $(Y_t,Z_t,U_t)$ are progressively measurable, $Y_t\in \mathcal{D}\left( \left[ 0,T\right] ,\R^{d}\right)$ and $Z$ and $U$ are predictable. Hence it remains to verify that $(Y_t, Z_t, U_t, \Lambda_t)$ satisfies $(i)$-$(vi)$ stated in Definition \ref{rbsde} and that $\Lambda\in \mathcal{BV}\left( \left[ 0,T\right]
\mathbb{R}
^{d}\right) $. That
$(Y_t, Z_t, U_t, \Lambda_t)$ satisfies $(i)$-$(iii)$ was proved above and $(iv)$ is a consequence of Lemma \ref{lemmbbaa} and
\eqref{aadg} $(i)$. Hence we in the following focus on properties $(v)$ and $(vi)$. As mentioned above, we have that $(\Lambda_t(\omega))_{0\leq t \leq T}$ is continuous in $t$ for almost all
$\omega$ by uniform convergence. Furthermore, using
that
$$\int _t ^T d|\Lambda^n|_s=n \int _t ^T |Y_s^n - \pi(s, Y^n _s)|ds$$ we see from Lemma \ref{ll1} $(iii)$
that
\begin{equation*
E \left[\int _0 ^T d|\Lambda^n|_s \right ] = E \left [ n \int _0 ^T |Y^n_s-\pi(s,Y^n_s)| ds \right ]\leq c\mbox{ for all $n\in\mathbb Z_+$},
\end{equation*}
\noindent for some constant $c$ which is independent of $n$. It follows that $\Lambda_t(\omega)$ is
of bounded total variation on $[0,T]$ for almost all
$\omega$. Hence, it only remains to verify that
\begin{eqnarray}\label{defo}
(v)&&\Lambda_t=\int_{0}^{t}\gamma _{s}d\left\vert \Lambda \right\vert
_{s},\ \gamma _{s}\in N_{s}^{1}\left( Y_{s}\right) \ \mbox{ whenever } Y_s \in \partial D_{s},\notag\\
(vi)&& d\left\vert \Lambda \right\vert \left( \left\{ t\in \left[ 0,T\right]
:\left(t,Y_{t}\right) \in D\right\} \right)=0.
\end{eqnarray}
To verify the statements in \eqref{defo} we will use the following lemma.
\begin{lemma} \label{lemma:adgda}
Let $\{\Lambda^n\}_{n \in \mathbb{Z}_+}$ be a sequence of continuous functions, $\Lambda^n:[0,T]\to \R^d$, which converges
uniformly to $\Lambda$ as $n\to \infty$. Assume $\Lambda^n\in \mathcal{BV}\left( \left[ 0,T\right]
\mathbb{R}
^{d}\right) $ and that $|\Lambda^n|_T \leq c$, for some $c<\infty$, hold for all $n$. Let
$\{f^n\}_{n \in \mathbb{Z}_+}$ be a sequence of c\`{a}dl\`{a}g functions, $f^n:[0,T]\to \R^d$, converging uniformly to $f$ as $n\to \infty$. Then,
\begin{equation*}
\lim _{n \to \infty}\int _0 ^t \langle f^n_s, d\Lambda^n_s \rangle = \int _0 ^t \langle f_s, d\Lambda_s \rangle
\end{equation*}
for all $t \in [0, T]$.
\end{lemma}
\begin{proof} This is essentially Lemma 5.8 in \cite{GP}, see also \cite{S}.
\end{proof}
Using Lemma \ref{lemmaa} $(i)$ we see that ,
$$\langle z_t- Y^n _t, Y^n_t-\pi(t,Y^n_t) \rangle \leq 0$$
for any c\`{a}dl\`{a}g process $z_t$ taking values in $\overline{D_t}$. Hence, for any such process $z_t$ we have that
\begin{eqnarray*}
\int_0 ^t -n \langle z_s- Y^n _s, Y^n_s-\pi(s,Y^n_s) \rangle ds =\int_0^t \langle z_s-Y^n_s , d\Lambda^n_s \rangle \geq 0.
\end{eqnarray*}
Passing to the limit we obtain, using Lemma \ref{lemma:adgda}, that
\begin{equation} \label{defo+}
\int_0 ^ t\langle Y_s- z_s, d\Lambda_s \rangle \leq 0
\end{equation}
for all $z \in \mathcal{D}\left( \left[ 0,T\right], \R^{d}\right)$ taking values in $\overline D$, and for all $t\in [0,T]$.
\noindent Next, let $\tau \in [0,T)$ be any time such that $Y_{\tau} \in D$ and let $\hat \gamma$ be a unit vector in $\R^d$. Since $Y_s$ is right-continuous, taking assumption \eqref{limitzero} into account, we see that there exists $\e >0$ and $\delta > 0$ such that $Y_s+\e \hat \gamma \in D$ and $Y_s-\e \hat \gamma \in D$ whenever $s \in [\tau, \tau+\delta]$. However, this in combination with \eqref{defo+} implies that
\begin{equation*}
0 \leq \int _{\tau} ^{\tau+\delta} \hat \gamma d\Lambda_t \leq 0,
\end{equation*}
which in turn implies $(vi)$ in \eqref{defo}. Hence
\begin{eqnarray*
\Lambda_t=\int_{0}^{t}\gamma _{s}d\left\vert \Lambda \right\vert_{s},
\end{eqnarray*}
for some vector field $\gamma_s\in \R^d$, with support on $\partial D$, and such that $|\gamma_{s}|=1$ (see \eqref{bff}). To conclude the existence part of Theorem \ref{maint1} it remains to show \eqref{defo}~$(v)$, i.e., that $\gamma_s\in N_{s}^{1}\left( Y_{s}\right)$ whenever $Y_s \in \partial D_{s}$. However, using the above and \eqref{bff} we see that to prove $\gamma_s\in N_{s}^{1}\left( Y_{s}\right)$ whenever $Y_s \in \partial D_{s}$, it is enough to prove that if $ Y_{s}\in\partial D_s$ and if
$\langle Y_{s}-z_s,\gamma_s\rangle\leq 0$
for all $z_s \in D_s$, then ${\gamma_s}\in N_{s}\left( Y_{s}\right)$. To do this, take $\beta=Y_s-\gamma_s \in \mathbb R^d$. Then,
\begin{equation*}
|\beta-z_s|^2=|\beta-Y_{s}|^2+|Y_{s}-z_s|^2 + 2\langle \beta -Y_{s},Y_{s}-z_s\rangle
\end{equation*}
for all $z_s\in D_s$. Hence, if $\langle \beta -Y_{s},Y_{s}-z_s\rangle = -\langle Y_{s}-z_s,\gamma_s\rangle\geq 0$, then we have that
\begin{equation*}
|\beta-z_s|^2\geq|\beta-Y_{s}|^2
\end{equation*}
for all $z_s \in D_s$, which implies $\gamma_s \in N_s(Y_s)$. This proves \eqref{defo} $(v)$ and thus the proof of the existence part of Theorem \ref{maint1} is complete. \hfill $\Box$
\subsection{Uniqueness: $(Y_t, Z_t, U_t, \Lambda_t)$ is the only solution}
We here prove the uniqueness part of Theorem \ref{maint1} using Ito's formula. Indeed, assume that $(Y^i,Z^i, U^i , \Lambda^i)$, $i=1,2$, are two solutions to the
the reflected BSDE under consideration and define
\begin{equation*}
\{\Delta Y_t, \Delta Z_t, \Delta U_t, \Delta \Lambda_t\}= \{Y^1_t - Y^2_t, Z^1_t - Z^2_t, U^1_t - U^2_t, \Lambda^1_t - \Lambda^2_t \}.
\end{equation*}
Then, applying Ito's formula to $|\Delta Y_t|^2$ and taking expectation we have that
\begin{align*}
& E \left [ |\Delta Y_t|^2+\int _t ^T |\Delta Z_s|^2 ds + \int _t ^T \int _ U |\Delta U_s(e)|^2 \lambda(de)ds \right ] \\
=& 2 E \left [ \int _ t ^T \langle \Delta Y_s, f(s,Y_s^1, Z_s^1, U_s^1)-f(s, Y_s^2,Z_s^2,U_s^2) \rangle ds \right ]\\
&+ 2 E \left [\int _t ^T \langle \Delta Y_s, d\Delta\Lambda_s \rangle \right].
\end{align*}
Using \eqref{defo+} we see that the last term on the right hand side in the above display is $\leq0$. Next, using the Lipschitz character of $f$ and standard manipulations, see \eqref{jaj++1}, \eqref{jaj++1+}, we can conclude that
\begin{align*}
& E \left [ |\Delta Y_t|^2+\int _t ^T |\Delta Z_s|^2 ds + \int _t ^T \int _ U |\Delta U_s(e)|^2 \lambda(de)ds \right ] \\
\leq& c E\left [ \int _ t ^T |\Delta Y_s|^2 ds + \frac{1}{2} \int _t ^T |Z_s| ^2 ds + \frac{1}{2} \int _t ^T |\Delta U_s(e)|^2 \lambda(de)ds \right ].
\end{align*}
Applying Gronwall's lemma we see that $\{\Delta Y_t, \Delta Z_t, \Delta U_t\}$ must be identically zero a.s. By $(iii)$ of Definition 1, the same applies to $\Delta \Lambda_t$ and the proof is hence complete. \hfill $\Box$
|
\section{Introduction\label{sec:Intro}}
The experimental discovery of charged resonances $Z_c(3900)^+$ \cite{Ablikim:2013mio} and $Z(4430)^{\pm}$ \cite{Choi:2007wga,Mizuk:2009da} gives signatures
for hadrons with minimal quark content $\bar cc\bar du$. The neutral $X(3872)$ and yet-unconfirmed $Y(4140)$ with charge parity
$C\!=\!+1$ also appear to have significant four-quark Fock components. Most of the observed exotic states have
$J^P=1^+$. The $J^P$ for some has not been settled experimentally and $J^P=1^+$ presents one possible option.
In this paper, we perform a lattice investigation of the charmonium spectrum, looking for charmonium-like
states with quantum numbers $J^{PC}=1^{++}$ and three quark contents: $\bar cc \bar du$, $\bar cc(\bar uu+\bar dd)$
and $\bar cc \bar ss$, where the later two channels have $I\!=\!0$ and can mix with $\bar cc$ ($C$ indicates $C$-parity of neutral isospin partners for charged states). Our main interest
in these channels is aimed at a first-principle study of $X(3872)$ and $Y(4140)$, which were observed in
$X(3872)\rightarrow J/\psi \rho,~J/\psi\, \omega,~ D\bar{D}^*$ and $Y(4140)\rightarrow J/\psi \phi$, for example.
From the experimental side, the long known exotic candidate $X(3872)$ \cite{Choi:2003ue}
is confirmed to have $J^{PC}=1^{++}$ \cite{Aaij:2013zoa}.
However, questions about its isospin remain unsettled. If it has
isospin $I=1$, one expects charged partners. Observation of a nearly equal branching fraction for
$X(3872) \rightarrow J/\psi\, \omega$ and $X(3872) \rightarrow J/\psi \,\rho$ decays \cite{Agashe:2014kda} and searches for
charged partner $X(3872)$ states decaying to $J/\psi \rho^{\pm}$ \cite{Aubert:2004zr} speak against
a pure $I=1$ state. There are a few other candidates with $C\!=\!+1$ that could possibly have
$J^{PC}=1^{++}$ like $X(3940)$ \cite{Abe:2004zs}, $Z(4050)^{\pm}$ \cite{Mizuk:2009da} and
$Z(4250)^{\pm}$ \cite{Mizuk:2009da}. A detailed review on these can be found in Ref. \cite{Olsen:2014qna}.
The growing evidence for the $Y(4140)$ resonance in the $J/\psi \phi$ invariant mass
\cite{Aaltonen:2009tz} serves as promising signature for exotic hadrons with
hidden strangeness. Similarities in the properties of $X(3930)$ and $Y(4140)$ led to an interpretation that
$X(3930)$ may be a $D^*\bar D^*$ molecule and $Y(4140)$ is its hidden strange counterpart $D_s^*D_s^*$ molecule \cite{Liu:2010hf}.
However, the upper limit for the production of $Y(4140)$ in $\gamma\gamma\rightarrow J/\psi\,\phi$ is observed to be much lower
than theoretical expectations for a $D_s^*D_s^*$ molecule with $J^{PC}=0^{++}$ and $2^{++}$\cite{Shen:2009vs}.
Hence the quantum numbers of $Y(4140)$ stay unsettled and it remains open for a $J^{PC}=1^{++}$ assignment.
From a theoretical perspective, the description of such resonances is not settled. Several
suggestions have been made interpreting them as mesonic molecules \cite{Swanson:2003tb}, as
diquark-antidiquark structures \cite{Maiani:2004vq}, as a cusp phenomena \cite{Bugg:2004sh} or as
a $|c\bar{c}g\rangle$ hybrid meson \cite{Close:2003sg}. A great deal of theoretical studies are based on
phenomenological approaches like quark model, (unitarized) effective field theory
and QCD sum rules (see reviews \cite{Olsen:2014qna}).
It is paramount to establish whether QCD supports the existence of resonances with exotic character using first principles
techniques such as lattice QCD. Simulations that considered only $\bar{c}c$ interpolators could not provide evidence for
$X(3872)$. The first evidence from a lattice simulation for $X(3872)$ with $I\!=\!0$ was reported in Ref. \cite{Prelovsek:2013cra}, where
a combination of $\bar cc$ as well as $D\bar D^*$ and $J/\psi \omega$ interpolators was used. Recently, another calculation using the Highly Improved Staggered Quark action also gave evidence for $X(3872)$, using $\bar{c}c$ and $D\bar{D}^*$ interpolating fields
\cite{Lee:2014uta}. The search for the $Y(4140)$ resonance was performed only in \cite{Ozaki:2012ce}, where a phase shift for $J/\psi \phi$ scattering in $s$-wave and $p$-wave was extracted from $N_f\!=\!2+1$ simulation using twisted boundary conditions, and neglecting strange-quark annihilation. The resulting phase shifts did not support existence of a resonance.
The novel feature of the present study is to add diquark-antidiquark $\dad{c}{q}{c}{q}$ operators to the basis of interpolating fields and to extend the
extraction of the charmonium spectrum with $J^{PC}=1^{++}$ to a higher energy range.
This is the first dynamical lattice calculation involving diquark-antidiquark operators
along with several two-meson and $\bar cc$ kind of interpolators to study $X(3872)$ and $Y(4140)$.
We consider the color structures ${\mathcal G}=\bar 3_c,6_c$ for diquarks, which
have been suggested already in the late seventies \cite{Jaffe:1976ih}. Recently many phenomenological studies
\cite{Maiani:2004uc,Maiani:2004vq} and a few lattice studies \cite{Chiu:2006hd,Chiu:2005ey}
used them to extract the light and heavy meson spectra.
In Ref. \cite{Chiu:2006hd} a calculation using two-meson and diquark-antidiquark interpolators was performed to investigate
mass spectrum of $1^{++}$ exotic mesons in quenched lattice QCD. However, only one energy level was extracted, which is not sufficient to provide evidence for $X(3872)$ or $Y(4140)$.
In this paper we address the following questions: Is the lattice candidate for $X(3872)$ reproduced in presence of
diquark-antidiquark operators? Which are the crucial operator structures for its emergence? How important are the
$\dad{c}{q}{c}{q}$ Fock components in the established $X(3872)$? Do we find a lattice candidate for charged or
neutral $X(3872)$ with $I\!=\!1$? Do operators with hidden strangeness render a candidate for $Y(4140)$? Do we
find candidate states for other possible exotic states in the channels being probed?
The paper is organized as follows. \scn{Interpolators} addresses the expected two-meson scattering channels below 4.2 GeV. The lattice methodology is discussed in Sect. \ref{sec:methodology}. In Sect. \ref{sec:Fierz} and the Appendix we discuss the relations
between our diquark-antidiquark and two-meson interpolators via Fierz transformations. \scn{Results} is dedicated to results
and we conclude in Sect. \ref{sec:Conc}.
\section{Two particle states in lattice QCD\label{sec:Interpolators}}
A major hurdle in excited-state spectroscopy is that most of the states lie above various
thresholds and decay strongly in experiments. All states carrying the same quantum numbers,
including the single-particle and multiparticle states, in principle contribute to the
eigenstates of the Hamiltonian. The determination of scattering properties relies
on precise identification of all the eigenstates below and close above the energy
of our interest. The continuous spectrum of scattering states in the continuum gets
reduced to a discrete set of eigenstates, because lattice momenta are discretized due to
the finite lattice size.
Considering two-meson states with total momentum zero and without interaction, their energies are just the sum of the
individual particle energies
\begin{equation}
\label{eni}
E^{n.i.}_{M_1(\mathbf{n})M_2(-\mathbf{n})}= E_1(p) + E_2(p),\ p=\frac{2\pi |\mathbf{n}|}{L},\ \mathbf{n}\in N^3.
\end{equation}
In the presence of interactions, the energies get shifted depending on the interaction strength. For our lattice setting the noninteracting two-meson levels
with $J^{PC}=1^{++}$, total momentum zero in the indicated energy range are
\beit
\item ~~$I = 0$; ~$\bar cc(\bar uu+\bar dd)$ and $\bar cc$; ~$E\lesssim4.2~$GeV
\begin{align*}
& D(0)\bar D^*(0), & & J/\psi(0)\omega(0), & & D(1)\bar D^*(-1),\\
&J/\psi(1)\omega(-1), & &\eta_c(1)\sigma(-1), & & \chi_{c1}(0)\sigma(0)\;.
\end{align*}
\item ~~$I = 1$; ~$\bar cc \bar du$ ; ~$E\lesssim4.2~$GeV
\begin{align*}
&D(0)\bar D^*(0),& &J/\psi(0)\rho(0),& &D(1)\bar D^*(-1),\\
&J/\psi(1)\rho(-1),& &\chi_{c1}(1)\pi(-1),& &\chi_{c0}(1)\pi(-1)\;.
\end{align*}
\item ~~$I = 0$; ~$\bar cc \bar ss$ and $\bar cc$; ~$E\lesssim4.3~$GeV
\begin{align*}
&D_s(0)\bar D_s^*(0),& &J/\psi(0)\phi(0),& &D_s(1)\bar D_s^*(-1),\\
&J/\psi(1)\phi(-1)\;.&&&&
\end{align*}
\eeit
The parentheses denote meson momenta in units of $2\pi/L$.
We consider the flavor sectors $\bar cc(\bar uu+\bar dd)$ and $\bar cc \bar ss$
separately. In nature these two $I=0$ sectors can mix and they could in principle mix
also in our simulation without dynamical strange quarks. However, if both flavor
sectors would be treated together, then $6+4=10$ two-particle $I=0$ states are expected
below $4.2~$ GeV. This would make the resulting spectrum denser and noisier, so the
identification of eigenstates and the search for exotics would be even more challenging.
We therefore consider these two sectors separately in this first search for possible
exotics in the extended energy region. The corresponding assumptions will be
discussed for each flavor channel along with the results.
The noninteracting energies will be shown by the horizontal lines in our plots, and
follow from the masses and the single meson energies determined on the same set of gauge
configurations \cite{Mohler:2012na,Lang:2011mn,Lang:2014tia}. The energies of the $\sigma$ meson
using single-hadron approximation are $a\,m_\sigma= 0.302(15)$ and $a\,E_{\sigma(1)}=0.534(22)$.
Including two-meson operators up to 4.2 GeV at $m_{\pi}=266$~MeV should be sufficient
in searching for narrow exotic candidates below 4.2 GeV. Details of all the interpolators
used, including the diquark-antidiquark interpolators, can be found in the next section.
The mesons $R=\rho,\sigma$ are resonances that decay to $\pi\pi$ or $\pi\eta$ in QCD with $N_f\!=\!2$.
A proper simulation which would consider the three-meson system \cite{Polejaeva:2012ut} has not been performed
in practice yet. In absence of this, a simplifying approximation for channels containing these resonances is
adopted. We determine the energy of $R(p)$ as the ground state energy obtained from the correlation matrix
with $\sum_x e^{ipx} \bar q(x)\Gamma q(x)$ interpolators. This energy is used for the horizontal lines in
the plots. This basis renders in all cases just one low-lying state. Within our approximation this low-lying
state corresponds to a resonance $R$ with momentum $p$, to a two-particle state $\pi\pi/\pi\eta$ with total
momentum $p$, or to some mixture of $R$ and the two-particle state. We also do
not consider nonresonant three-meson levels which could appear above $\eta_c \pi\pi$, $J/\psi \pi\pi$, $\eta_c K\bar K$, $J/\psi K\bar K$ thresholds.
Based on the experience with two-meson operators we do not expect that without
explicit incorporation of three-meson interpolating fields these three-meson states appear in the spectra.
\section{Lattice methodology\label{sec:methodology}}
These calculations are performed on $N_f\!=\!2$ dynamical gauge configurations with
$m_{\pi}\!\simeq \!266$ MeV \cite{Hasenfratz:2008ce} and with other parameters provided
in \tbn{latpar}. The mass-degenerate $u/d$ quarks are based on a tree-level improved
Wilson-clover action. The strange quark is present only in the valence sector and we
assume that the valence strange content could uncover hints on the possible existence of
the $\bar cc\bar ss$ exotics. The absence of dynamical strange quarks prevents
$\bar cc\bar ss$ intermediate states in the $\bar cc(\bar uu+\bar dd)$ and $\bar cc$
sector, in accordance with treating these two $I=0$ sectors separately in our study.
With a rather small box size of $L\!\simeq\! 2$ fm, one expects to have large finite
size effects. On the other hand this serves as a crucial practical advantage by reducing
the number of two-meson scattering states $M_1(\mathbf{n})M_2(-\mathbf{n})$ in the
energy range of our interest. This helps in easier identification of the possible
resonances that could exist along with the regular two-meson energy levels. It also
reduces computational cost as one needs to consider a smaller number of distillation
eigenvectors and two-meson interpolators with respect to a study in larger volume.
\bet[tbh]
\centering
\betb{ccccccccccccc}
\hline
Lattice size & $\kappa$ & $\beta$ & $N_{\mathrm{cfgs}}$ & $m_\pi$ [MeV] & $a$ [fm] & $L$ [fm] \\\hline
$16^3 \times 32$ & $0.1283$ & 7.1 & 280 & 266(3)(3) & $0.1239(13)$ & 1.98
\\\hline
\eetb
\caption{Details of the gauge field ensemble used.}
\eet{latpar}
We construct altogether 22 interpolators with $J^{PC}=1^{++}$ and total momentum zero for the three cases of our
interest ($T_1^{++}$ irreducible representation of the discrete lattice group $O_h$ is employed):
\begin{widetext}
\beqa
O_{1-8}^{\bar cc}&=& \bar c \hat{M} c(0) ~\tfrac{1}{2}(1+K_d), \qquad\mbox{see Table X of Ref. \cite{Mohler:2012na}}\ \label{operators0} \\
O_9^{MM}&=&\bar c \gamma_5 u(0)~\bar u\gamma_i c(0) - \bar c \gamma_i u(0)~\bar u\gamma_5 c(0) +K_d\{u\to d\}, \ \nonumber \\
O_{10}^{MM}&=&\epsilon_{ijk} ~\bar c \gamma_j c(0)~\{~\bar u\gamma_k u(0)+K_d\{u\to d\}~\}, \nonumber \\
O_{11}^{MM}&=&\!\!\!\!\sum_{e_p=\pm e_{x,y,z}}\{\bar c \gamma_5 u(e_p)~\bar u\gamma_i c(-e_p) - \bar c \gamma_i u(e_p)~\bar u\gamma_5 c(-e_p)\} +K_d\ \{u\to d\}, \nonumber\\
O_{12}^{MM}&=&\bar c \gamma_5 \gamma_4 u(0)~\bar u\gamma_i \gamma_4 c(0) -\bar c \gamma_i \gamma_4 u(0)~\bar u\gamma_5 \gamma_4 c(0)+K_d\{u\to d\}, \nonumber\\
O_{13}^{MM}&=&\epsilon_{ijk} ~\bar c \gamma_j \gamma_4 c(0)~\{\bar u\gamma_k \gamma_4 u(0)+K_d \{u\to d\}\}, \nonumber\\
O_{14}^{MM}&=&\sum_{e_p=\pm e_{x,y,z}}\epsilon_{ijl} ~\bar c \gamma_j c(e_p)~\{\bar u\gamma_l u(-e_p)+K_d \{u\to d\}\},\nonumber\\
O_{15}^{MM}&=&\{\bar c\gamma_5 c (e_p) ~\bar uu(-e_p)~-~\bar c\gamma_5 c (-e_p) ~\bar uu(e_p)\}_{p=i}+K_d \{u\to d\}, \nonumber\\
O_{16}^{MM}&=&\epsilon_{ijp} \{\bar c\gamma_j\gamma_5 c (-e_p) ~\bar u\gamma_5u(e_p)~-\bar c\gamma_j\gamma_5 c (e_p) ~\bar u\gamma_5u(-e_p)\} +K_d \{u\to d\},\nonumber\\
O_{17}^{MM}&=&\bar c\gamma_i\gamma_5 c (0) ~\bar uu(0) +K_d \{u\to d\},\nonumber\\
O_{18}^{MM}&=&\{\bar c c (e_p) ~\bar u\gamma_5u(-e_p)~-~\bar cc (-e_p) ~\bar u\gamma_5 u(e_p)\}_{p=i} +K_d \{u\to d\} ,\nonumber\\
O_{19}^{4q}&=&[ \bar c~ C\gamma_5 \bar u^T]_{3_c}[ c^T \gamma_i C u]_{\bar 3_c} + [ \bar c~ C\gamma_i \bar u^T]_{3_c}[ c^T \gamma_5 C u]_{\bar 3_c}+K_d \{u\to d\}, \nonumber\\
O_{20}^{4q}&=&[ \bar c~ C \bar u^T]_{3_c}[ c^T \gamma_i \gamma_5 C u]_{\bar 3_c} + [ \bar c~ C \gamma_i \gamma_5 \bar u^T]_{3_c}[ c^T C u]_{\bar 3_c}+K_d \{u\to d\}, \nonumber\\
O_{21}^{4q}&=&[ \bar c~ C\gamma_5 \bar u^T]_{\bar 6_c}[ c^T \gamma_i C u]_{6_c} + [ \bar c~ C\gamma_i \bar u^T]_{\bar 6_c}[ c^T \gamma_5 C u]_{6_c}+K_d \{u\to d\}, \nonumber\\
O_{22}^{4q}&=&[ \bar c~ C \bar u^T]_{\bar 6_c}[ c^T \gamma_i \gamma_5 C u]_{6_c} + [ \bar c~ C \gamma_i \gamma_5 \bar u^T]_{\bar 6_c}[ c^T C u]_{6_c}+K_d \{u\to d\}. \nn
\eeqa{operators1}
\end{widetext}
The indices $i, j, k$ and $l$ define the Euclidean Dirac gamma matrices, while the index $p$ indicates the momentum
direction. Einstein's summation convention is implied for repeated indices. The unsummed index $i$ in all the operators
defines the polarization. The $C=i\gamma_2\gamma_4$ is the charge conjugation matrix. The coefficient $K_d$ depends on the
quark content: $K_d\!=\!1$ is used for $\bar cc(\bar uu+\bar dd)$ and $K_d\!=\!0$
for $\bar cc \bar ss$ followed by using strange quark propagators instead of the light quark propagators. For $I=1$ channel
we apply $K_d\!=\!-1$ which gives the flavor content $\bar cc(\bar uu-\bar dd)$ and has the same spectrum as $\bar cc\bar du$
in the isospin limit.
We emphasize the use of four operators $O^{4q}$ with diquark-antidiquark structure and color antitriplet or sextet diquarks
\beqa
[ \bar c \Gamma_1 \bar q]_{\mathcal{G}}[ c \Gamma_2 q]_{\mathcal{\bar G}} &\equiv& \sum_{\mathbf{x_1}} \mathcal{G}_{ab_1c_1}\bar c_{b_1}^{\alpha_1}\Gamma_1^{\alpha_1\beta_1}\bar q^{\beta_1}_{c_1}(\mathbf{x_1},t_f) \nonumber \\
~&\cdot& \sum_{\mathbf{x_2}}\mathcal{G}_{ab_2c_2}c_{b_2}^{\alpha_2}\Gamma_2^{\alpha_2\beta_2}q^{\beta_2}_{c_2}(\mathbf{x_2},t_f).
\eeqa{4q}
Here $a=1,2,3$ for color triplet and $a=1,...,6$ for sextet, while $b,c=1,2,3$ for both:
\beqa
\mathcal{G}^{3}_{abc} = \mathcal{G}^{\bar 3}_{abc} &=&\epsilon_{abc}~ \\
\mathcal{G}^{6}_{abc}=\mathcal{G}^{\bar 6}_{abc}&=&1 ~\mbox{:}~ a=1,2,3 ~\mbox{and}~ a\ne b\ne c \nn \\
\mathcal{G}^{6}_{abc}=\mathcal{G}^{\bar 6}_{abc}&=&\sqrt{2}~\mbox{:}~ a=4,5,6 ~\mbox{and}~ a-3= b = c \nn
\eeqa{4qCG}
while the remaining $\mathcal{G}_{abc}$ are zero. The operator [\eqn{4q}] reduces to
$\sum_{\mathbf{x}} \bar c(\mathbf{x}) \bar q(\mathbf{x}) c(\mathbf{x}) q(\mathbf{x})$ on ensemble averaging,
where the gauge configurations are not gauge fixed.
The interpolators are related with the two-meson channels as listed in \tbn{operators2}. noninteracting levels corresponding
to some of these two-meson channels lie above our energy of interest, and the corresponding
interpolators are not considered.
\bet
\betb{c | c | c | c }
\hline
N & $\bar cc(\bar uu+\bar dd)$ & $\bar cc\bar ud$ & $\bar cc\bar ss$ \\\hline\hline
$O_{1-8}^{\bar cc}$ & $\bar c ~\hat{M}~c$ & Does not couple & $\bar c ~\hat{M}~c$ \\\hline
$O_{9}^{MM}$ & $D(0)\bar D^*(0)$ & $D(0)\bar D^*(0)$ & $D_s(0)\bar D_s^*(0)$ \\\hline
$O_{10}^{MM}$ & $J/\psi(0)\omega(0)$ & $J/\psi(0)\rho(0)$ & $J/\psi(0)\phi(0)$ \\\hline
$O_{11}^{MM}$ & $D(1)\bar D^*(-1)$ & $D(1)\bar D^*(-1)$ & $D_s(1)\bar D_s^*(-1)$ \\\hline
$O_{12}^{MM}$ & $D(0)\bar D^*(0)$ & $D(0)\bar D^*(0)$ & $D_s(0)\bar D_s^*(0)$ \\\hline
$O_{13}^{MM}$ & $J/\psi(0)\omega(0)$ & $J/\psi(0)\rho(0)$ & $J/\psi(0)\phi(0)$ \\\hline
$O_{14}^{MM}$ & $J/\psi(1)\omega(-1)$ & $J/\psi(1)\rho(-1)$ & $J/\psi(1)\phi(-1)$ \\\hline
$O_{15}^{MM}$ & $\eta_c(1)\sigma(-1)$ & $\eta_c(1)a_0(-1)$ & Not used \\\hline
$O_{16}^{MM}$ & $\chi_{c1}(1)\eta(-1)$ & $\chi_{c1}(1)\pi(-1)$ & Not used \\\hline
$O_{17}^{MM}$ & $\chi_{c1}(0)\sigma(0)$ & $\chi_{c1}(0)a_0(0)$ & Not used \\\hline
$O_{18}^{MM}$ & $\chi_{c0}(1)\eta(-1)$ & $\chi_{c0}(1)\pi(-1)$ & Not used \\\hline
$O_{19-20}^{4q}$ & $\dadt{c}{q}{c}{q}$ & $\dadt{c}{u}{c}{d}$ & $\dadt{c}{s}{c}{s}$ \\\hline
$O_{21-22}^{4q}$ & $\dads{c}{q}{c}{q}$ & $\dads{c}{u}{c}{d}$ & $\dads{c}{s}{c}{s}$ \\\hline
\eetb
\caption{List of interpolators ($J^{PC}=1^{++}$) and their correspondence with various two-meson
scattering channels.}
\eet{operators2}
The Wick contractions considered in the computation of the correlation functions are shown in \fgn{Wickcont}. There are two
other classes of diagrams, which are not considered: one in which no valence quark propagates from source to sink, and the
other class in which only the light/strange quarks propagate from source to sink and the $\bar{c} c$ pair annihilates. The
effects from these two classes of diagrams, with the charm quark not propagating from source to the sink, are known to be suppressed
due to the Okubo-Zweig-Iizuka rule. They correspond to mixing with a number of channels that contain only the $u/d$ and
$s$ quarks, which represents currently unsolved challenge in lattice QCD. Note that the annihilation of $u/d$ and $s$ quarks
as well as mixing with $\bar cc$ is taken into account, unlike in the simulation \cite{Ozaki:2012ce} aimed at $Y(4140)$, for example.
\begin{widetext}
\begin{figure*}[tbh]
\centering
\hspace{-1.5cm}
\parbox{.45\linewidth}{
\centering
\includegraphics[scale=0.35]{wick-squished_allcon.eps}\\
(a)}\hspace{0.25cm}
\hspace{0.25cm}
\parbox{.45\linewidth}{
\centering
\includegraphics[scale=0.3]{wick-squished_Ldiscon.eps}\\
(b)}
\caption{The Wick contractions considered in our calculations. (a) Connected contraction diagrams. (b) Diagrams,
in which the light/strange quarks do not propagate from source to sink. The correlation functions in the
$\bar cc(\bar uu+\bar dd)$ and $\bar cc\bar ss$ cases are linear combinations of the diagrams of kind (a) and (b),
while the correlation functions between the operators with quark content $\bar cc\bar ud$ are constructed purely
from diagrams of kind (a).}
\label{fg:Wickcont}
\end{figure*}
\end{widetext}
Using the interpolators listed in Eq. (\ref{operators0}) and \tbn{operators2}, we compute the full coupled
correlation functions
\beq
\mathcal C_{jk}(t)=\langle\Omega|O_j (t_{s}+t) O_k^\dagger (t_{s})|\Omega\rangle =\sum_{n}Z_k^{n*}Z_j^ne^{-E_n t}.
\eeq{Corr}
For an efficient computation of these correlation matrices, we utilize the ``distillation" method for
the quark sources as proposed in Ref. \cite{Peardon:2009gh}. In this method the quark sources are
build from the $N_v$ lowest eigenmodes of the gauge-covariant Laplacian on a given time slice, $t_{s}$.
We use $N_{v}\!=\!64$ for computation of correlators involving $u/d$ quarks, while for the correlators
with hidden strange content, we use $N_{v}\!=\!48$. The correlation functions with $u/d$ quarks are
computed only for polarization along the $x$-axis and averaged over all $t_{s}$, while correlation functions
involving hidden strange quarks are averaged over all polarizations and for all even values of $t_{s}$.
The energies $E_n$ and overlaps $Z_j^{(n)}=\langle \Omega|O_i|n\rangle$ for all eigenstates $n$ are extracted using
the well-established generalized eigenvalue problem \cite{Michael:1985ne}
\beq
\mathcal C(t)\,u^{(n)}(t) = \lambda^{(n)}(t,t_0)\,\mathcal C(t_0)\,u^{(n)}(t).
\eeq{gevp1}
The energies $E_n$ are extracted asymptotically from two-exponential fits to the eigenvalues
\beq \lambda^{(n)}(t,t_0)\propto A_n e^{-E_n t} + A'_n e^{-E'_n t}~,\quad E_n^\prime>E_n. \eeq{texpf}
We find consistent results for $t_0=2, 3$ and present the results for $t_0=2$.
The two-exponential fits were typically done in the range $3\leq t \leq 14$.
The eigenvectors $u^{(n)}$ determine the overlaps
\beq
Z_j^{(n)}(t)=e^{E_nt/2}\frac{|\mathcal C_{jk}(t)u_k^{(n)}(t)|}{|\mathcal C(t)u^{(n)}(t)|}\;.
\eeq{gevp2}
The statistical errors obtained using single-elimination jackknife analysis are quoted throughout.
The complete basis was used in the initial analysis, which was later reduced to an optimized basis,
separately in each of the three cases, based on a systematic operator pruning. This procedure is
aimed at getting better signals (in terms of the numbers of states and the quality of the effective
mass plateau and the overlap factors) in comparison with the spectrum extracted from the full set of operators. After
finalizing the optimized set of two-meson interpolators, we fixed the $\bar{c}c$ and $\dad{c}{q}{c}{q}$
operators that give good signals for a maximum number of extractable states below
4.2 GeV. The optimized basis that we used for the three cases of quark content are
\beqa
\bar cc(\bar uu+\bar dd) : & \quad &O_{1,3,5}^{\bar cc},~O_{9-12,14,15,17}^{MM},~O_{19,21}^{4q} \nonumber \\
\bar cc\bar ud :& \quad &O_{9-16,18}^{MM},~O_{19,21}^{4q} \nonumber \\
\bar cc\bar ss :& \quad &O_{1,5}^{\bar cc},~O_{9-11,14}^{MM},~O_{19,21}^{4q}\;.
\eeqa{optmbasis}
Our principal aim is to find out whether QCD supports exotic states in addition to the conventional charmonia and
the two-meson scattering levels, which inevitably appear in dynamical QCD. Analytic techniques
have been proposed for the determination of the scattering matrix for coupled two-hadron scattering channels based on
L\"uscher-type finite volume formalisms \cite{Luscher:1990ux}.
These would in principle allow extraction of the masses and decay widths for resonances of interest. A number of
lattice calculations have already dealt with resonances and shallow bound states in the elastic scattering (see
\cite{Dudek:2012xn} and \cite{Mohler:2013rwa} for an example of each). The first calculation of a scattering matrix
for two coupled channels also promises progress in this direction \cite{Dudek:2014qha}. However, such an analysis
is beyond the scope of current lattice simulations for more than two coupled channels and/or three-hadron scattering
channels, which applies to the case considered.
Therefore we take a simplified approach, where the existence of possible exotic states is investigated by analyzing
the number of energy levels, their positions and overlaps with the considered lattice operators
$\langle \Omega |O_j|n\rangle$.
The formalism does predict an appearance of a level in addition to the (shifted) two-particle levels if there is a
relatively narrow resonance in one channel. We have, for example, found additional levels related to the resonances
$\rho$ \cite{Lang:2011mn}, $K^*(892)$ \cite{Prelovsek:2013ela}, $D^*_0(2400)$ \cite{Mohler:2012na}, and the bound
state $D_{s0}^*(2317)$ \cite{Mohler:2013rwa}. Additional levels related to $K_0^*(1430)$ \cite{Dudek:2014qha} and
$X(3872)$ \cite{Prelovsek:2013cra} have been found in the simulations of two coupled channels. Based on this experience,
we expect an additional energy level if an exotic state is of similar origin, i.e. if it corresponds to a pole of
the scattering matrix near the physical axis.
Consider a noninteracting situation. Several two-meson operators considered in \tbn{operators2} contain the vector
meson $V(1)$ with one unit of momentum. This can reside in irreducible representations (irreps) $A_1$ or $E_2$ of
the corresponding symmetry group $Dic_4$ \cite{Moore:2006ng,Thomas:2011rh}. One expects two degenerate energy
levels for $P(1)V(-1)$ since there are two ways to combine the vector-meson irrep ($A_1$, $E_2$) with the
pseudoscalar-meson irrep ($A_2$) to obtain the rest frame irrep of interest $T_1^+$ (see Table III of
\cite{Moore:2006ng}). The underlying reason is that $PV$ state with $J^P=1^+$ can be in $s$-wave or in
$d$-wave (also in continuum) \cite{Briceno:2014oea,Briceno:2013lba,Briceno:2013bda}. In the limit of small coupling
between $s-$ and $d-$wave, one energy level is due solely to the $s$-wave and the other one to $d$-wave
\cite{Briceno:2013lba,Briceno:2013bda}.\footnote{The $PV$ spectrum resembles (in the noninteracting limit)
the spectrum in the deuterium channel $pn$, since $S=1$, $J^P=1^+$ and $l=0,2$ apply in both cases. Figure 2
of \cite{Briceno:2013lba} indicates that one level $n(1)p(-1)$ is related mostly to $s$-wave and the other to $d$-wave.
L\"uscher's quantisation condition \cite{Briceno:2014oea} does not depend on the spins of the individual particles,
but on their total spin $S$. } We implement only the $s$-wave interpolator $O^{P(1)V(-1)}$ [Eq. (\ref{operators0})]
and therefore expect to see only one energy level; this is verified in our observed spectra shown in Sect.
\ref{sec:Results}. One would need to employ two distinct interpolators in order to find two $P(1)V(-1)$ energy
levels, but the extraction of such eigenstates has not been attempted yet for two-meson systems in QCD to our
knowledge. Our two-meson operators contain also $V_1(1)V_2(-1)$, where three levels are expected based on analogous
arguments \cite{Moore:2006ng}; we expect to find only one level related to $s$-wave interpolators
[Eq. (\ref{operators0})], and indeed we do not find two other levels related to $d$-wave (for total spins
$S\!=\!1,2$). We emphasize that the omission of additional interpolator structures and avoidance of levels related
to $d$-waves makes the search for possible exotics within our approach less cumbersome and results more transparent.
\footnote{If one would find an extra state near $V(1)P(-1)$ or $V_1(1)V_2(-1)$, one would indeed have to identify
whether this extra state arises due to the presence of the $d$-wave or is related to exotics. We do not address
this question since we do not find such an extra state.}
Charm quarks being heavy are subject to large discretization errors. We treat the charm quarks using the Fermilab
formulation \cite{ElKhadra:1996mp}, according to which we tune the charm quark mass by equating the spin averaged
kinetic mass of the $1S$ charmonium to its physical value. With this formulation, the discretization errors are
highly suppressed in the energy splitting $E_n-m_{s.a.}$, $m_{s.a.} = \frac14(m_{\eta_c}+3m_{J/\psi})$, which will
be compared with the experiments. We utilized this method in our earlier calculations on this ensemble and found good
agreement with the experiments for conventional charmonium in Ref. \cite{Mohler:2012na} as well as, for masses and
widths of charmed mesons in Refs. \cite{Mohler:2012na,Lang:2014yfa,Mohler:2013rwa}.
\section{Fierz relations \label{sec:Fierz} }
The diquark-antidiquark operators $\dadt{c}{q}{c}{q}$ and $\dads{c}{q}{c}{q}$ can be expressed as linear
combinations of color singlet currents $(\bar cc)_{1_c}(\bar qq)_{1_c}$ and $(\bar cq)_{1_c}(\bar qc)_{1_c}$
\cite{Maiani:2004vq,Ali:2014dva}. These relations are obtained for local currents via Fierz rearrangement
\cite{Nieves:2003in} and are presented in the Appendix. Note that our quarks are smeared and each meson in $O^{MM}$
has definite momentum, but the Fierz relation suggests that $O^{4q}$ and $O^{MM}$ are still linearly dependent.
The Fierz rearrangement is the key idea behind Coleman's argument \cite{Coleman} that in the large $N_c$ limit
application of Fermion quadrilinears to the vacuum creates meson pairs and nothing else. In the physical world
with $N_c\!=\!3$, it is argued that tetraquarks could exist at subleading orders \cite{Weinberg:2013cfa}
of large $N_c$ QCD. However, in the presence of the leading order two-meson terms, one should
take caution in interpreting the nature of the levels purely based on their overlap factors onto various
four-quark interpolators.
\begin{figure}[tb]
\centering
\includegraphics[scale=0.33]{T1pp_cormat.eps}
\includegraphics[scale=0.07]{T1pp_cormat_leg.eps}\\
\caption{Time averaged normalized correlation matrix $\tilde{\mathcal C}$ [\eqn{TANCM}] for the operator basis
$O_{1-22}$ with quark content $\bar cc(\bar uu+\bar dd)$ and $\bar cc$. The axis ticks correspond to the order of
operators used in Eq. (\ref{operators0}).}
\label{fg:cormat}
\end{figure}
Let us consider a comparative study between the lattice correlators and the Fierz expansion of $O^{4q}$ operators.
From Eq. (\ref{opfq1}), we see that the first and second terms in the Fierz expansion represent $D\bar D^*$, while the
seventh term is similar to the $O^{MM}_{17}=\chi_{c1} ~ \sigma$. Hence we expect significant correlations between
these operators. This is indeed verified in \fgn{cormat}, showing the time averaged normalized ensemble averaged
correlation matrix
\beq
\tilde{\mathcal{C}}_{ij} = \frac19\sum_{t=2}^{10} \frac{\bar{\mathcal C}_{ij}(t)}{\sqrt{\bar{\mathcal C}_{ii}(t)\bar{\mathcal C}_{jj}(t)}}.
\eeq{TANCM}
With this normalization all the diagonal entries are forced to unity and all the off-diagonal entries to be less than unity.
The $\dad{c}{q}{c}{q}=O^{4q}_{19-22}$ have large correlations onto the $D\bar D^*=O^{MM}_{9,11,12}$ and
$\chi_{c1} ~ \sigma=O^{MM}_{17}$. The strong correlations between $O^{4q}$ and $O^{cc}$ operators can also be
explained by the $\chi_{c1} \sigma$ component in $O^{4q}$, where $\sigma$ couples to the vacuum.
\begin{figure*}[tbh]
\hspace{-1.0cm}
\parbox{.45\linewidth}{
\centering
\includegraphics[scale=0.72]{T1pp_I0.eps}
\hspace{0.9cm}
\parbox{.45\linewidth}{
\centering
\includegraphics[scale=0.72]{T1pp_I1.eps}
\caption{The spectra of states with $J^{PC}=1^{++}$ for the cases with $u/d$ valence quarks. The energies
$E_n=E^{lat}_n-m_{s.a.}^{lat}+m_{s.a.}^{exp}$ [\eqn{en}] are shown. The horizontal lines show energies
of noninteracting two-particle states (\ref{eni}) and experimental thresholds, indicating uncertainty
related to $\sigma$ width. In each subplot, the middle block shows the discrete spectrum
determined from our lattice simulation from the optimized basis [\eqn{optmbasis}]. The right-hand block shows the
spectrum we obtained from the optimized basis of operators with the $\dad{c}{q}{c}{q}$ operators excluded.
The gray marks, on the right-hand side of each pane, indicate the lowest three-meson threshold $m_{\eta_c}+2m_\pi $,
while the actual lowest $\eta_c\pi\pi$ level on the lattice appears higher due to $l=1$, which requires relative momenta.
The left-hand block shows the physical thresholds and possible experimental candidates (a) $\chi_{c1}$,
$X(3872)$ and $X(3940)$, (b) $Z_c^+$(4050) and $Z_c^+$(4250). The violet error bars for experimental candidates
show the uncertainties in the energy and the black error bars show its width. }
\label{fg:primresl}
\end{figure*}
\begin{figure}[tbp]
\centering
\includegraphics[scale=0.72]{T1pp_ss.eps}
\caption{The spectrum of states with $J^{PC}=1^{++}$ and hidden strange quarks. The possible
experimental candidates shown are $\chi_{c1}$, $X(3872)$, $Y(4140)$ and $Y(4274)$.
The gray marks, on the right-hand side of each pane, indicate the lowest three-meson threshold $m_{\eta_c}+2m_K $.
However, the actual lowest $\eta_cKK$ level on the lattice appears higher due to $l=1$, which requires relative momenta.
For further details see \fgn{primresl}.}
\label{fg:primress}
\end{figure}
\begin{figure}[tbp]
\centering
\includegraphics[scale=0.72]{T1pp_I0_SP.eps}\\
\caption{The spectrum of states (\eqn{en}) with $J^{PC}=1^{++}$ and quark content $\bar cc(\bar uu + \bar dd)$ \& $\bar cc$.
(i) Optimized basis (without $O^{MM}_{17}$), (ii) optimized basis without $\bar cc$ operators (and without $O^{MM}_{17}$)
and (iii) basis with only $\bar cc$ operators. Note that candidate for $X(3872)$ disappears when removing $\bar cc$ operators
although diquark-antidiquark operators are present in the basis, while it is not clear to infer on the dominant nature of this
state just from the third panel. The $O_{17}^{MM}=\chi_{c1}(0)\sigma(0)$ is excluded from the basis to achieve better signals and
clear comparison. }
\label{fg:i0sasa}
\end{figure}
\begin{figure*}[tbhp]
\begin{center}
\includegraphics[scale=0.6,clip]{I0_w4q.eps}
\end{center}
\begin{center}
\includegraphics[scale=0.6,clip]{I1_w4q.eps
\end{center}
\begin{center}
\includegraphics[scale=0.6,clip]{ss_w4q.eps
\end{center}
\caption{The overlap factors $Z_j^{(n)}=\langle \Omega|O_j|n\rangle$ [\eqn{gevp2}] shown in units of the maximal $|Z_j^m|$ for a given operator $j$ across all the eigenstates $m$.
These ratios are independent of the normalization of the interpolators $O_j$.
The horizontal axis corresponds to the complete basis of interpolators [Eq. (\ref{operators0})], where the optimized subsets [\eqn{optmbasis}] were employed. The levels are ordered from lowest to highest $E_n$ as in the middle pane of the spectrum in Figs. \ref{fg:primresl} and \ref{fg:primress}. The values are averages of the ratios over $4 \leq t \leq 13$ with
error bars due to jackknife sampling.}
\label{fg:zratios}
\end{figure*}
\section{Results\label{sec:Results}}
The discrete spectra in Figs. \ref{fg:primresl} and \ref{fg:primress} are the main results from our lattice calculation. They show the energies
\beq
E_n=E^{lat}_n-m_{s.a.}^{lat}+m_{s.a.}^{exp},\quad m_{s.a.} = \frac{1}{4}(m_{\eta_c}+3m_{J/\psi})~
\eeq{en}
of the states with $J^{PC}=1^{++}$ and three quark contents. The horizontal lines represent various two-meson noninteracting energies.
The states that have dominant overlap with two-meson scattering operators are represented by circles and the color coding
identifies the respective scattering channels based on the following criteria:
\begin{itemize}
\item The levels appear close to the expected two-meson noninteracting energies.
\item They have dominant overlaps $\langle\Omega|O_j^{M_1M_2}|n\rangle$ with corresponding $O_j^{M_1M_2}$. This is also verified based on the ratios $Z_j^n/max_m(Z_j^m)$,which are independent of normalization of operators and are shown in \fgn{zratios}.
\item If the corresponding two-meson interpolators are excluded from the basis, this eigenstate disappears or becomes too
noisy to be identified. This is determined by comparing the pattern of the effective masses and overlaps between the original basis and the basis after operator exclusion.
\end{itemize}
The remaining states, that are not attributed to the two-meson scattering channels, are represented by red squares.
Figures \ref{fg:primresl} and \ref{fg:primress} also compare the spectra between the two bases of operators, one with optimized operator set and another with the optimized set excluding $\dad{c}{q}{c}{q}$. In all three cases we see an almost negligible effect on the low lying states, while we do observe an improvement in the signals for higher lying states in the basis without $\dad{c}{q}{c}{q}$. The same conclusion applies for overlaps.
The employed irreducible representation $T_1^{++}$ contains the states $J^{PC}=1^{++}$ of interest,
as well as $J^{PC}=3^{++}$ states due to the broken rotational symmetry. Upon inclusion of the
interpolator $O_{8}^{\bar cc}$ to the basis [\eqn{optmbasis}] the spectra for both $I=0$ channels
remain essentially unchanged except for an additional level at $E\simeq 4.1-4.2~$GeV [\eqn{en}]. This is
where the earlier simulation on the same ensemble \cite{Mohler:2012na} and the simulation
\cite{Liu:2012ze} have identified the only $3^{++}$ state in the energy region of our interest.
In the following subsections, we present the spectra of $J^{PC}=1^{++}$ states in three flavor
channels for the basis (\eqn{optmbasis}), where $O_{8}^{\bar cc}$ is excluded.
\subsection{$I=0$ channel with flavor $\bar cc (\bar uu+\bar dd)$ and $\bar cc$ \label{ss:I0l}}
This is the channel where the experimental $X(3872)$ resides. We will argue that the energy levels affected by
this state are $n=2$ (red squares) and $n=6$ (blue circle) from \fgn{primresl}(a). The lowest state is the
conventional $\chi_{c1}(1P)$. The overlaps of the three low-lying levels represented by circles show dominant
$J/\psi(0)\omega(0)$, $\eta_c(1)\sigma(-1)$ and $\chi_{c1}(0)\sigma(0)$ Fock components. The
highest two states in \fgn{primresl}(a) have significant overlap with the $J/\psi(1)\omega(-1)$ and
$D_0(1) \bar D_0^*(-1)$ operators.
Now we focus on the eigenstates that are related to $X(3872)$. The $\bar cc$ interpolators alone give
an eigenstate close to $D\bar D^*$ threshold (right pane of \fgn{i0sasa}), but one cannot establish
whether this eigenstate is related to $X(3872)$ or to nearby two-meson states in this case. Therefore
we turn to the spectrum of the full optimized basis [midpane in \fgn{primresl}(a)], where levels
$n\!=\!2$ (red squares) and $n\!=\!6$ (blue circles) are found to have dominant overlap with the
$\bar{c}c$ and $D\bar D^*$ operators. Excluding either of these operators results in disappearance of
one level and a shift in the other level towards the $D\bar D^*$ threshold. We emphasize that one of the two
levels remains absent when $D\bar D^*$ and $O^{4q}$ are used and $O^{\bar cc}$ is not, as is
evident from the first and second panel from the left of \fgn{i0sasa}. This indicates that the $\bar cc$
Fock component is crucial for $X(3872)$, while the $\dad{c}{q}{c}{q}$ structure alone does not render
it. This also implies a combined dominance of $\bar{c}c$ and $D\bar D^*$ operators in determining
the position of these two levels, while their resulting energies are not significantly affected
whether $O^{4q}$ is used in addition or not.
\begin{table}[htb]
\betb{ c c | c c c | c c }
\hline
$X(3872)$ &&& $m_{X} - m_{s.a.}$ &&& $m_{X} - m_{D_0} - m_{D^*_0}$ \\ \hline
Lat. &&& 816(15) &&& -8(15) \\
Lat. - $O^{4q}$ &&& 815(8) &&& -9(8) \\
LQCD \cite{Prelovsek:2013cra} &&& 815(7) &&& -11(7) \\
LQCD \cite{Lee:2014uta} &&& - &&& -13(6) \\ \hline
Exp. &&& 803(1) &&& -0.11(21) \\
\hline
\eetb
\caption{Mass of $X(3872)$ with respect to $m_{s.a.}$ and the $D_0\bar D_0^*$ threshold. Our estimates are from the
correlated fits to the corresponding eigenvalues using single exponential fit form with and without diquark-antidiquark
operators. Results from previous lattice QCD simulations \cite{Prelovsek:2013cra,Lee:2014uta} and experiment are also presented.}
\label{tb:X3872psa}
\end{table}
We determine the $D\bar D^*$ scattering phase shift from levels $n=2,6$ via L\"uscher's relation
\cite{Luscher:1990ux} assuming elastic scattering. The phase shift is interpolated near threshold
using the effective-range approximation. The eigenstate $n\!=\!6$ (blue circle) is interpreted as the
$D(0)\bar D^*(0)$ scattering state, which is significantly shifted up due to a large negative
scattering length \cite{Sasaki:2006jn}. The resulting scattering matrix $T\propto 1/(\cot\delta(p) -i)$ has a pole just
below the threshold where $\cot \delta(p_B)=i$ is satisfied. We neglect possible effects of the
left-hand cut in the partial wave amplitude. The results confirm a shallow bound
state just below the $D\bar D^*$ threshold and the binding momentum $p_B$ renders the mass of the bound
state, interpreted as experimentally observed $X(3872)$. The resulting mass of $X(3872)$ and its
binding energy are provided in \tbn{X3872psa} and in \fgn{X3872psa}, which indicate that it is
insensitive to inclusion of diquark-antidiquark interpolators within errors. The mass of $X(3872)$
was extracted along these lines for the first time in Ref. \cite{Prelovsek:2013cra}, where this
channel was studied in a smaller energy range on the same ensemble without
diquark-antidiquark
interpolators. The error on the binding energy in the present paper is larger due to the larger
interpolator basis. These results are in agreement with a possible interpretation of X(3872), where its properties are due to the accidental alignment of a $\bar cc$ state with the $D^0\bar D^{*0}$ threshold \cite{Danilkin:2010cc,Takizawa:2012hy}, but we cannot rule out other options.
\begin{figure}[htb]
\includegraphics[scale=0.8]{X3872.eps}
\caption{Mass of $X(3872)$ with respect to $m_{s.a.}$ from the present simulation, previous lattice studies
\cite{Prelovsek:2013cra,Lee:2014uta} and experiment \cite{Agashe:2014kda}. }
\label{fg:X3872psa}
\end{figure}
With regard to the other experimentally observed charmonia-like states [e.g. $X(3940)$], which could appear in
this channel, we do not find any candidate in addition to the expected two-meson scattering levels. We also do not
find candidates for other $\bar cc$ states with $J^{PC}=1^{++}$ [e.g. $\chi_{1c}(nP)$] in the region between
the $D\bar D^*$ threshold and $4.2~$GeV.
\subsection{$I=1$ channel with flavor $\bar cc \bar du$ \label{i1}}
A careful analysis of this isospin channel is crucial due to the large branching ratio for the decay
$X(3872)\rightarrow J/\psi \rho$ and current experimental interests in search of a charged $X(3872)$.
With no disconnected diagrams allowed in the light quark propagation, the correlation matrix is constructed
purely of four-quark interpolators and connected Wick contractions in \fgn{Wickcont}(a).
The spectrum of eigenstates is shown in \fgn{primresl}(b), where all energies are close to noninteracting energy
levels. All the eigenstates have a dominant overlap with the two-meson interpolators. The spectrum shows very
little influence on the inclusion of $\dad{c}{q}{c}{q}$, which is evident from \fgn{primresl}(b). Given that
all the levels below 4.2 GeV can be attributed to the expected two-meson scattering states, we conclude that
our lattice simulation gives no evidence for $Z_c(4050)^+$ and $Z_c(4250)^+$.
Our results also do not support charged or neutral $X(3872)$ with $I=1$. There is no experimental indication
for charged $X$, while the neutral $X$ does have a large decay rate to $I=1$ final state $J/\psi \rho^0$. One popular
phenomenological explanation for this decay is that $X(3872)$ has $I=0$ and the isospin is broken in the decay
mechanism (due to the $D^+\bar D^{*-}$ vs $D^0\bar D^{*0}$ mass difference) \cite{Gamermann:2009fv,Takizawa:2012hy}. According to another explanation,
$X$ is a linear combination of $I=0$ and $I=1$ components, where the $I=1$ component vanishes in the isospin limit
\cite{Tornqvist:2004qy}. Our simulation is performed in the isospin limit $m_u=m_d$, so it
is perhaps not surprising that $X$ with $I\!=\!1$ is not observed. Future simulations with nondegenerate $u/d$ quarks
would be very welcome for this channel.
As pointed out in \scn{Interpolators}, $\rho$ in $J/\psi \rho$ is treated as stable, although $\rho(1)$ is kinematically
close to the decay channel $\pi(1)\pi(0)$. In the absence of a simulation of a three-meson system, it is disputable what `noninteracting' energy should be taken for the $\rho(1)$. An estimate
from the diagonal correlator $\rho(1)$ leads to `noninteracting' energy roughly 65 MeV below the eigenstate energy,
which is identified to have a dominant overlap with the $J/\psi(1) \rho(-1)$ interpolator. However, taking the resonance
position \cite{Lang:2011mn} brings the `noninteracting' level in agreement with the measured eigenenergy.
\subsection{$I=0$ channel with flavor $\bar cc \bar ss$ and $\bar cc$\label{i0s}}
Our goal in simulating this channel is to search for a possible presence of the $Y(4140)$ resonance, which was found in
$J/\psi \phi$ scattering in several experiments \cite{Aaltonen:2009tz}.
Our lattice simulation of $J/\psi\phi$ scattering takes into account the annihilation of the valence
strange quarks and thereby the mixing with $\bar cc$ flavor content.
With no strange quark effects in the sea, the study of this channel is based on the following
assumptions. We construct a basis with only $\bar{c}c$ and four-quark operators ($O^{MM}$, $O^{4q}$)
with valence hidden strange content for this analysis. We assume that these interpolators have negligible coupling to two-meson states with flavor content
$\bar cc(\bar uu+\bar dd)$. In other words, we assume that two-meson states like $D\bar D^*$ and
$J/\psi \omega$ will not appear in the spectrum based on the chosen interpolators.
The resulting spectrum in this channel confirms this assumption. We point out that $Y(4140)$ has been experimentally observed only in the $J/\psi \phi$ final state with valence strange content, but it has not been observed in $D\bar D^*$ and $J/\psi \omega$ final states. Although this ensemble does
not have strange quarks in the sea, we assume that the valence strange content could uncover hints
on the existence of the charm-strange exotics, if they exists.
Spectra in this channel are shown in \fgn{primress}. We identify the lowest two states, represented
by squares, to be $\chi_{c1}(1P)$ and the level related to $X(3872)$. The remaining four states
are identified with the expected $D_s\bar D_s^*$ and $J/\psi \phi$ scattering levels. Thus in the
energy region below 4.2 GeV, we find no levels that could be related to $Y(4140)$ or any other exotic
structure. Note that the existence of $Y(4140)$ is not yet finally settled from experiment, and
its quantum numbers, except for $C=+1$, are unknown. Therefore it is possible that its absence in
our simulation is related to the fact that we explored the channel $J^P=1^+$ only.
\subsection{Discussion}
The only exotic charmonium-like state found in our simulation is a $X(3872)$ candidate with $J^{PC}=1^{++}$ and $I\!=\!0$.
It is found as a bound state slightly below $D\bar D^*$ threshold and has a mass close to the experimental mass of $X(3872)$.
We point out that this mass corresponds to our $m_\pi\simeq 266~$MeV and was obtained from a rather small lattice volume,
while chiral and continuum extrapolations have not been performed. Precision determination of its mass with respect to
$D\bar D^*$ threshold will be a challenging task for future lattice simulation on larger volumes, which also should account
for its coupling with multiple open scattering channels involving two or more hadrons. Recent analytic studies
consider the quark mass dependence, the volume dependence and the effect from the isospin breaking relevant for future
lattice studies of $X(3872)$ \cite{Jansen:2013cba}
Candidates for no other ``exotic" charmonium-like states [except for $X(3872)$] are found in our exploration of the three $J^{PC}=1^{++}$ channels.
We list several possible reasons for the absence of the energy levels related to other possible exotic states in our simulation:
\begin{itemize}
\item The existence of $Y(4140)$, $Z_c^+$(4050), $Z_c^+$(4250) or any other exotic state in these
channels, is not yet settled experimentally. Even if they exist, only $C\!=\!+1$ is established
experimentally, while their $J^P$ is unknown. This could explain their absence in our simulation,
which probes only $J^P=1^+$.
\item Based on the experience, discussed in \scn{methodology}, we expect an additional energy level
if the exotic state is a resonance associated to a pole near the real axis in the unphysical Riemann
sheet. The absence of an additional energy level could also indicate a different origin of the
experimental peak, e.g., a coupled-channel threshold effect. Further analytical work and
lattice simulations are needed to settle the question whether an additional energy level is
expected in this case.
\item Finally, we cannot exclude the possibility that some exotic candidates could be absent due
to the relatively heavy pion mass $m_\pi\simeq 266~$MeV, isospin limit $m_u=m_d$, neglect of the
charm annihilation contributions, or the absence of the strange dynamical quarks in our simulation.
\end{itemize}
\section{Conclusions\label{sec:Conc}}
We present the spectra from a lattice QCD simulation of $J^{PC}=1^{++}$ channels with three different
quark contents: $\bar cc \bar du$, $\bar cc(\bar uu+\bar dd)$ and $\bar cc \bar ss$, where the later
two can mix with $\bar cc$. The pion mass in this study with $u/d$ dynamical quarks is $m_\pi\!\simeq\! 266~$MeV.
Using a large number of interpolating fields $\dadt{c}{q}{c}{q}$, $\dads{c}{q}{c}{q}$,
$(\bar cq)_{1_c}(\bar qc)_{1_c}$, $(\bar cc)_{1_c}(\bar qq)_{1_c}$ and $(\bar cc)_{1_c}$, we extract the spectra
up to $4.2~$GeV. We find evidence for $\chi_{c1}$ and $X(3872)$, while all the remaining eigenstates
are related to the expected two-meson scattering channels, which inevitably appear in the dynamical
QCD. The $\bar cc$ Fock component in $X(3872)$ appears to be more important than the $\dad{c}{q}{c}{q}$,
since we find a candidate for $X(3872)$ only when $\bar cc$ interpolating fields are used. The $D\bar D^*$
interpolators show a more prominent effect on the position of $X(3872)$ than the $\dad{c}{q}{c}{q}$.
Candidates for charged or neutral $X(3872)$ with $I=1$ are not found in our simulation with
$m_u\!=\!m_d$, and future simulations with broken isospin would be welcome for this channel.
We also do not find a candidate for $Y(4140)$ or any other exotic charmonium-like
structure. Our search for the exotic states assumes an appearance of an additional energy eigenstate on the lattice,
which is a typical manifestation for conventional hadrons. Further analytic work is needed to establish whether this working
assumption applies also for several coupled channels and all exotic structures of interest.
\acknowledgments\label{sec:Ackn}
We thank Anna Hasenfratz and the PACS-CS for providing the gauge configurations. We acknowledge the discussions
with R. Briceno, L. Leskovec, D. Mohler, S. Ozaki, S. Sasaki and C. DeTar. The calculations were performed on computing
clusters at the University of Graz (NAWI Graz), at the Vienna Scientific Cluster (VSC) and at Jozef Stefan
Institute. This work is supported in part by the Austrian Science Fund FWF:I1313-N27 and by the Slovenian
Research Agency ARRS Project No. N1-0020. S.P. acknowledges support from U.S. Department of Energy Contract No.
DE-AC05-06OR23177, under which Jefferson Science Associates, LLC, manages and operates Jefferson Laboratory.
|
\section{Introduction}\label{intro}
Pulsar Timing Arrays \cite[PTAs; e.g.,][]{manchester13,kramer13,mclaughlin13} seek to detect nanohertz gravitational waves from cosmological and extragalactic sources by looking for correlations between contemporaneously measured pulse arrival times from multiple radio pulsars \citep{hellings83}. The sensitivity of a PTA is limited by pulsar timing noise, i.e., stochastic wandering of pulse arrival times.
External noise sources include interstellar plasma turbulence, jitter noise and errors in terrestrial time standards; see \citet{cordes13} for a description of all dominant noise sources and an estimate of their magnitudes. Intrinsic noise sources have been attributed to microglitches \citep{cordes85,dalessandro95,melatos08}, post-glitch recovery \citep{johnston99}, magnetospheric state switching \citep[e.g.,][]{kramer06,lyne10}, fluctuations in the spin-down torque \citep{cheng87a,cheng87b,urama06}, variable coupling between the crust and core or pinned and corotating regions \citep{alpar86,jones90}, asteroid belts \citep{shannon13b} and superfluid turbulence \citep{greenstein70,link12a,melatos14}. Analyses of long-term millisecond pulsar timing data indicate that timing noise power spectra are typically white above some frequency and red below it \citep{kaspi94, shannon10, vanhaasteren11, shannon13a}.
Red timing noise power spectra cannot extend to arbitragerily low frequencies, as the infinite integrated noise-power implies divergent phase residuals and hence (if phase residuals arise from torque fluctuations) unphysical pulsar angular velocities. One therefore expects the spectrum to plateau, or even become blue, below some turn-over frequency $f_c$. A number of physical models naturally predict low-frequency plateaus, including superfluid turbulence \citep{melatos14} and asteroid belts \citep{shannon13b}. We discuss the former in detail below. A low-frequency plateau enhances prospects for the detection of a stochastic gravitational wave background. As the gravitational wave spectrum is a steep power law for most cosmological sources \citep[e.g.,][]{maggiore00,phinney01,grishchuk05}, it rises above the plateau below some frequency as long as it too does not have a low-frequency cut-off \citep[e.g.,][and discussion below]{sesana13b,ravi14a}.
In this article, we quantify how a low-frequency timing noise plateau affects the direct detection of gravitational waves with PTAs. Specifically, we calculate the minimum observation time for any individual pulsar to become sensitive to gravitational wave stochastic backgrounds from binary supermassive black holes (SMBHs) and cosmic strings.
We note this minimum observation time is only an indicative quantity for determining when a gravitational wave signal will dominate the timing residuals for an individual pulsar; it does not account for algorithms that correlate noise properties between pulsars, a point we discuss in more detail throughout.
We do this in two ways, firstly by parametrising the timing noise in a model independent way, and secondly by applying the superfluid turbulence model of \citet{melatos14}. In the first approach, we express this minimum time in terms of three pulsar observables; the amplitude and spectral index of the timing noise power spectral density and the turn-over frequency. The second approach is included as an example of how to relate PTA observables to neutron star internal properties in the context of one particular physical model with only two free parameters. It does not imply any theoretical preference for the superfluid turbulence model, and will be extended to other physical models in the future.
The paper is set out as follows. In section \ref{TN} we define a phenomenological model for timing noise, and review predictions for the power spectral density of the phase residuals induced by gravitational waves from SMBHs and cosmic strings. In section \ref{mintime} we calculate minimum observation times for hypothetical pulsars as a function of their timing noise spectral index, normalisation, and turn-over frequency. In section \ref{superfluidmodel} we apply the superfluid turbulence model to data and extract `by-eye' parameter estimates for various pulsars in the Parkes Pulsar Timing Array (PPTA). We then determine criteria for selecting `optimal' pulsars in section \ref{optimal} and conclude in section \ref{conclusion}.
\section{Power spectrum of the Phase residuals}\label{TN}
\subsection{Timing noise}
Let $\Phi_{\rm TN}(f)$ denote the Fourier transform of the autocorrelation function of the phase residuals, $\delta\phi(t)$, viz.
\begin{equation}
\Phi_{\rm TN}(f)=\int_{-\infty}^{\infty}d\tau\,{\rm e}^{2\pi i f\tau}\left<\delta\phi(t)\delta\phi(t+\tau)\right>.
\end{equation}
If the timing noise is stationary, $\left<\delta\phi(t)\delta\phi(t+\tau)\right>$ is independent of $t$, as is the mean-square phase residual
\begin{equation}
\left<\delta\phi(t)^{2}\right>=\frac{1}{\pi}\int^{\infty}_{0}df\,\Phi_{\rm TN}(f)\label{phaseresidual}.
\end{equation}
In practice, the time spent observing the neutron star, $T_{\rm obs}$, is finite. Hence, one must replace the lower terminal of the integral in the right-hand side of (\ref{phaseresidual}) by $f_{\rm obs}\equiv1/T_{\rm obs}$. In reality, fitting models to timing data implies PTAs are sensitive to $f<f_{\rm obs}$ [see \citet{coles11} and \citet{vanhaasteren13} for details of timing-model fits in the presence of red noise] implying the lower terminal in (\ref{phaseresidual}) depends on the PTA data analysis algorithm, with $f\lesssim f_{\rm obs}$.
Millisecond pulsar radio timing experiments measure $\Phi_{\rm TN}(f)\propto f^{-q}$ at low frequencies, $f\lesssim1\,{\rm yr}^{-1}$, with $q\geq0$ \citep[e.g.,][]{kaspi94, shannon10, vanhaasteren11, shannon13a}. However, the observed power law must roll over below some frequency, $f_{c}$, otherwise equation (\ref{phaseresidual}) implies divergent phase residuals. To capture this phenomenologically, we model the spectrum in its entirety by
\begin{equation}
\Phi_{\rm TN}(f)=\frac{A_{\rm TN}}{\left(1+f^{2}/f_{c}^{2}\right)^{q/2}} + A_{\rm W},\label{phi}
\end{equation}
which has the observed large-$f$ behaviour and is even in $f$. In equation (\ref{phi}), $A_{\rm TN}$ (with units of time) is the dc power spectral density, i.e. $\Phi_{\rm TN}\left(f\ll f_{c}\right)= A_{\rm TN}$, which cannot be measured directly in existing data sets
\citep{shannon13a}. In the regime where $\Phi_{\rm TN}(f)\propto f^{-q}$, we can express the more commonly used root-mean-square-induced pulsar timing residuals, $\sigma_R$, in terms of $A_{\rm TN}$ and $q$ as
\begin{equation}
\sigma_R
\frac{5.64}{\sqrt{q-1}}\left(\frac{A_{\rm TN}}{10^{-10}\,{\rm yr}}\right)^{1/2}\left(\frac{f_{\rm obs}}{1\,{\rm yr^{-1}}}\right)^{1/2}\left(\frac{P}{1\,{\rm ms}}\right)\,{\rm ns}.
\label{convert}
\end{equation}
where $P$ is the pulsar spin period. For completeness, we include a white noise component, $A_{\rm W}$, in equation (\ref{phi}), which is observed in all pulsars, dominates for $f\gtrsim1\,{\rm yr}^{-1}$, and is the only observed noise component in some objects. The white component contributes weakly to setting the minimum observation time for gravitational wave detection by PTAs, the key concern of this article.
Equation (\ref{phi}) can be compared against predictions of phase residuals from the cosmological gravitational wave background, $\Phi_{\rm GW}(f)$. The reciprocal of the frequency where the two curves intersect gives the minimum observation time, $T_{\rm min}$, required
before an individual pulsar becomes sensitive to a gravitational wave background,
\begin{equation}
\Phi_{\rm TN}\left(T_{\rm min}^{-1}\right)=\Phi_{\rm GW}\left(T_{\rm min}^{-1}\right).\label{Tmin}
\end{equation}
Equation (\ref{Tmin}) provides a quantitative method for determining when the gravitational wave signal will dominate the timing residual power spectrum. We emphasise that this is only an indicative threshold for detection; it is not a substitute for a careful signal-to-noise estimate given desired false alarm and false dismissal rates. Cross-correlation search algorithms look simultaneously at a range in $f$ \citep[e.g.,][]{hellings83,jenet05,anholm09,vanhaasteren09}. For example, our definition (\ref{Tmin}) is equivalent to the boundary between the `weak signal limit' and the `intermediate regime' as defined in \citet{siemens13}.
While \citet{siemens13} calculate a scaling of gravitational wave detection significance with time assuming only white timing noise, they also perform simulations with red noise assuming $q=-3$.
A future research project is therefore to introduce red noise with and without a low-frequency turn-over into the analytic calculations of \citet{siemens13}.
It is likely that the near future will see an increasing number of PTA pulsars satisfy the condition $\Phi_{\rm GW}(f)>\Phi_{\rm TN}(f)$, and that this will occur {\it before} a statistically significant detection is announced. Equation (\ref{Tmin}) and the analysis presented in this article therefore provide an important input into the time-scale on which this condition will be met by individual pulsars, as a prelude to a cross-correlation detection strategy.
\subsection{Cosmological gravitational wave background}\label{SGWB}
Pulsar timing arrays are sensitive to gravitational wave backgrounds generated by two cosmological sources\footnote{Relic gravitational waves from inflation, such as those purportedly seen by the BICEP2 experiment \citep{ade14}, are expected to be undetectably weak in the pulsar timing band, but may be relevant for Advanced LIGO; see \citet{aasi14} and references therein.}: binary supermassive black holes and vibrations from cosmic strings.
\subsubsection{Supermassive binary black holes}
At binary separations where gravitational radiation dominates the orbital dynamics, the SMBH background is parametrised as a power law
\begin{equation}
h_{c}(f)=A_{\rm GW}\left(\frac{f}{\mbox{yr}^{-1}}\right)^{\alpha},\label{hc}
\end{equation}
with $\alpha=-2/3$ \citep{phinney01}. The normalisation coefficient, $A_{\rm GW}$, is the subject of intense debate. We utilise the most recent predictions by \citet{sesana13} and \citet{ravi15}, quoted in table \ref{table}. These two predictions assume that gravitational wave emission has already circularized the binary orbits; at binary separations where energy loss to environments dominates instead, the SMBH wave-strain spectrum whitens \citep{sesana13b,ravi14a}. Whitening of $\Phi_{\rm GW}(f)$ at low frequencies increases $T_{\rm min}$.
The one-sided power spectral density of the pulsar phase residuals induced by $h_{c}(f)$ is given by
\begin{equation}
\Phi_{\rm GW}(f)=\frac{h_{c}(f)^{2}}{12\pi^{2}P^{2}f^{3}}. \label{PhiGW}
\end{equation}
$\Phi_{\rm GW}(f)$ has units of time and can be compared directly with $\Phi_{\rm TN}(f)$ as in equation (\ref{Tmin}).
\subsubsection{Cosmic strings}
Cosmic strings are topological defects that may form in phase transitions in the early Universe and produce strong bursts of gravitational radiation, which may be detectable in PTAs \citep{damour00,damour01,damour05}. A cosmic string-induced stochastic background of gravitational waves is characterised by three dimensionless parameters: the string tension, $G\mu$, the reconnection probability, $p$, and a parameter, $\epsilon$, related to the size of loops. The best quoted limit of $G\mu\lesssim1.2\times10^{-8}$ is derived from PTA limits of the stochastic gravitational wave background \citep{vanhaasteren11,vanhaasteren12}, although a more stringent constraint (still to be computed), is possible with existing data sets [see \citet{sanidas13} for projected constraints in the near future]. Combined observations using the ground-based Laser Interferometer Gravitational Wave Observatory (LIGO) and Virgo constrain the $\epsilon$ -- $G\mu$ plane to be $7\times10^{-9}<G\mu<1.5\times10^{-7}$ and $\epsilon<8\times10^{-11}$ \citep{abbott09,aasi14}. Limits on $G\mu$ are model dependent; the reconnection probability is inversely proportional to $\Phi_{\rm GW}(f)$, and smaller values of $\epsilon$ increase the minimum gravitational wave frequency emitted. This can take the maximum of the stochastic background out of the sensitivity band for PTAs \citep[e.g.,][]{siemens07,olmez10}.
Despite the above caveats, a power-law model for the characteristic strain spectrum from cosmic strings given by equation (\ref{hc}) with $-1\lesssim\alpha\lesssim0.8$ is a good approximation for the PTA frequency band \citep[][]{maggiore00}. The predicted range for $A_{\rm GW}$ is quoted in Table \ref{table}.
\begin{table*}
\centering
\begin{tabular}{|cccc|}
\hline\hline
Source & $A_{\rm GW}^{\rm min}$ & $A_{\rm GW}^{\rm max}$ & $\alpha$\\
\hline
SMBHs \citep[][68\%]{sesana13} & $3.5\times10^{-16}$ & $1.5\times10^{-15}$ & $-2/3$\\
SMBHs \citep[][95\%]{ravi15} & $5.1\times10^{-16}$ & $2.4\times10^{-15}$ & $-2/3$\\
Cosmic Strings & $10^{-16}$ & $10^{-15}$ & $-1\lesssim\alpha\lesssim-0.8$\\
\hline
\end{tabular}
\caption{\label{table}
Theoretical spectral parameters for the gravitational wave background from supermassive black hole binaries (SMBHs) and cosmic strings. Three predictions for the SMBH population are presented: the 68\% confidence interval from \citet{sesana13}, and the 95\% confidence interval
from \citet{ravi15}. The cosmic string models are from \citet{maggiore00}.
}
\end{table*}
\section{Model Independent Minimum Observation Time}\label{mintime}
To attain adequate sensitivity to gravitational waves at a frequency, $f$, in the phase residuals of an individual pulsar, we must have $\Phi_{\rm GW}(f)>\Phi_{\rm TN}(f)$ for that pulsar, subject to the caveats regarding specific data analysis algorithms expressed in the text following equation (\ref{Tmin}). If equation (\ref{hc}) applies across all relevant frequencies, and $\Phi_{\rm TN}(f)$ turns over below $f_{c}$, then $\Phi_{\rm GW}(f)>\Phi_{\rm TN}(f)$ is always satisfied for some $f=T_{\rm min}^{-1}$, as in equation (\ref{Tmin}).
In figure \ref{phifig} we plot $\Phi_{\rm TN}$ and $\Phi_{\rm GW}$ as functions of $f$. The coloured shaded regions and the region enclosed by the black dotted curves in the left-hand plot contain all the $\Phi_{\rm GW}$ curves in the parameter range in Table \ref{table} for SMBHs. The blue shaded region
is the 95\% confidence interval
from \citet{ravi15} as described in table \ref{table}. The black dotted curves enclose the 68\% confidence interval from \citet{sesana13}. The shaded grey region is the predicted range from \citet{ravi14a} that includes low-frequency-whitening of $\Phi_{\rm GW}(f)$ due to non-circular binaries. In the right-hand plot, the green shaded region is the parameter space enclosed by the cosmic string predictions from table \ref{table} with $\alpha=-1$. The dotted black curves are specific, representative calculations of the cosmic string background with $p=\epsilon=1$ and $G\mu=1.2\times10^{-8}$ (top curve) and $1.0\times10^{-10}$ (bottom curve)\footnote{Calculations used the GWPlotter website: \hyperref[http://homepages.spa.umn.edu]{http://homepages.spa.umn.edu/$\sim$gwplotter} }. The black, red and purple curves in each panel are indicative examples of pulsar timing noise as described by equation (\ref{phi}), with values of $f_{c}$, $A_{\rm TN}$, $q$ and $A_{\rm W}$ given in the caption to figure \ref{phifig}. The correspondingly coloured dashed curves extrapolate backwards the power-law scaling (equivalently assuming $f_{c}\rightarrow0$). Finally, the dot labelled `PPTA' in both panels marks the lowest limit on the stochastic gravitational wave background from \citet{shannon13a}.
\begin{figure*}
\includegraphics[angle=0,width=0.95\linewidth]{Phi.pdf}
\caption{\label{phifig} Power spectral density of timing noise phase residuals, $\Phi_{\rm TN}$, and gravitational wave phase residuals, $\Phi_{\rm GW}$ (both in s) as functions of frequency, $f$ (in Hz). In both plots, the black, red and purple curves are indicative examples of pulsars described by equation (\ref{phi}) with $(f_{c},\,A_{\rm TN},\,q)=(0.1 {\rm yr}^{-1},\,10^{-10}\,{\rm yr},\,4)$, $(0.2 {\rm yr}^{-1},\,10^{-10}\,{\rm yr},\,2)$ and $(0.2 {\rm yr}^{-1},\,10^{-8}\,{\rm yr},\,2)$ respectively and with a nominal white noise component of $A_{\rm W}=10^{-5}$\,s. The solid curves include the turnover at $f_c$ in equation (\ref{phi}), and the dashed curves extrapolate backwards the high-$f$ scaling $\propto f^{-q}$ (i.e., $f_{c}\rightarrow0$). The shaded regions and the dotted black curves encompass the regions covered by $\Phi_{\rm GW}$ for the parameter ranges quoted in Table \ref{table}, where the left-hand plot is for SMBHs and the right plot is for cosmic strings. In the left plot, the shaded blue region
is the 95\% confidence interval from \citet{ravi15}, the region enclosed by the black dotted curves is the 68\% confidence interval from \citet{sesana13} and the shaded grey region is the predicted range from \citet{ravi14a}. In the right plot, the shaded green region represents cosmic string models with $10^{-16}<A_{\rm GW}<10^{-15}$ and $\alpha=-1$. The dotted black curves have $p=\epsilon=1$ and $G\mu=1.2\times10^{-8}$ and $1.0\times10^{-10}$ for the top and bottom curves respectively. The black dot labelled `PPTA' is the lowest published limit on the stochastic background \citep{shannon13a}.
}
\end{figure*}
Figure \ref{phifig} illustrates the principal idea of this paper. If $\Phi_{\rm TN}$ is a simple power-law without a low-frequency turn-over, and for moderate values of $q$, timing noise masks the gravitational wave background down to low frequencies, and $T_{\rm min}$ is correspondingly long (we quantify this below). A turn-over in $\Phi_{\rm GW}(f)$ at some $f_{c}$ is therefore critical for practical PTA experiments with any millisecond pulsar that exhibits a steep timing noise spectrum with $q\gtrsim2$. The low-frequency plateau in $\Phi_{\rm GW}$ from elliptical binary SMBHs (the grey shaded region in the left-hand panel of figure \ref{phifig}) makes the need for a turn-over in $\Phi_{\rm TN}(f)$ even more acute.
In figure \ref{ObsTime} we plot the minimum observation time, $T_{\rm min}$, defined by equation (\ref{Tmin}), as a function of the asymptotic (high-$f$) timing noise spectral index, $q$, for a hypothetical pulsar with $P=10\,{\rm ms}$ and various values of $f_{c}$ and $A_{\rm TN}$ in each panel. The shaded regions
and dotted black curves delineate the ranges of $T_{\rm min}(q)$ for binary SMBHs and cosmic strings, following the same colour scheme as in figure \ref{phifig} and as detailed in the caption of figure \ref{ObsTime}. The coloured dashed curves give the limits on $T_{\rm min}(q)$ if $\Phi_{\rm TN}$ does not turn-over (i.e., $f_{c}\rightarrow0$). The horizontal dashed black line marks the PPTA observing time of 11.3\,yr used for the lowest limit on the stochastic background published to date \citep{shannon13a}.
To help interpret figures \ref{phifig} and \ref{ObsTime}, consider a hypothetical pulsar with $A_{\rm TN}=10^{-8}$ yr (i.e., the two left hand panels) and $q=2$. If the timing noise spectral density turns over at $f_{c}=1/5$ yr$^{-1}$ or $1/20$ yr$^{-1}$, the minimum observation time given the most optimistic scenario from \citet{ravi15} is $T_{\rm min}=20.5$ yr or $37.2$ yr, respectively. On the other hand, if $\Phi_{\rm TN}(f)$ does {\it not} turn over, then the dashed blue curves show that the pulsar
is insensitive to a gravitational wave signal
until $T_{\rm min}=70.7$ yr. The effect of a plateau in $\Phi_{\rm TN}(f)$ is therefore quite striking. Pulsars without a plateau and $q\gtrsim3$ (depending less sensitively on $A_{\rm TN}$) are relatively inferior as a tool for detecting gravitational waves.
\begin{figure*}
\includegraphics[angle=0,width=0.95\linewidth]{ObsTime.pdf}
\caption{\label{ObsTime} Minimum observation time, $T_{\rm min}$, defined by equation (\ref{Tmin}), as a function of the high-frequency timing noise spectral index, $q$, with different values of $f_{c}$ and $A_{\rm TN}$ in each panel. The shaded blue and grey regions represent the ranges of solution space for $T_{\rm min}(q)$ for binary SMBHs from \citet{ravi15} and \citet{ravi14a} respectively, while the dotted black curves encompass the solution space predicted by \citet{sesana13}.
The shaded green regions are cosmic string predictions. The coloured dashed curves are the corresponding limits where the timing noise spectrum does not plateau at $f\lesssim f_c$. The horizontal dashed black line is longest the PPTA observing time in published data, viz. $11.3$ yr \citep{shannon13a}.}
\end{figure*}
\section{Timing Noise from Superfluid Turbulence: A Worked Example}\label{superfluidmodel}
In section \ref{mintime}, the description of timing noise is model independent, in the sense that $\Phi_{\rm TN}$ is parametrised phenomenologically by equation (\ref{phi}), without reference to a specific underlying, physical model. In this section, we repeat the analysis in section \ref{mintime} for the timing noise model of \citet{melatos14} and \citet{melatos14b}, which attributes the fluctuating phase residuals to shear-driven turbulence in the interior of the neutron star. We emphasise that we do not express any theoretical preference for this model ahead of other models in the literature (see section \ref{intro}). We focus on it here only because (i) it is predictive, (ii) its results can be expressed in compact, analytic form and, (iii) the theoretical formula for $\Phi_{\rm TN}(f)$ depends on just three internal neutron star parameters, so it is easy to infer constraints on these parameters by combining the model with data.
Consider an idealised neutron star model in which the rigid crust is coupled to the charged electron-proton fluid which, in turn, couples through mutual friction to the inviscid neutron condensate. The electromagnetic braking torque creates a crust-core shear layer that excites turbulence in the high-Reynolds number superfluid \citep{peralta05,peralta06a,peralta06b,melatos07,peralta08}. The turbulent condensate reacts back to produce angular momentum fluctuations in the crust, which are observed as timing noise \citep{greenstein70,melatos10}. In particular, \citet{melatos14b} showed that the timing noise spectral density can be expressed as
\begin{eqnarray}
\Phi_{\rm TN}\left(f\right)\!\!\!\!\!\!&=\!\!\!\!\!\!&\frac{15\Gamma(q/2)}{8\pi^{1/2}\Gamma\left[(q-1)/2\right]\lambda^{2}\eta\left(R^{-1}\right)}\nonumber\\
\!\!\!\!\!\!&\times\!\!\!\!\!\!&\int^{\infty}_{2\pi}dx\, x^{-35/3}\left(x^{4}+3x^{2}+9\right)\left[1+\frac{4\pi^{2}f^{2}}{\eta\left(R^{-1}\right)^{2}x^{4/3}}\right]^{-q/2},\label{superfluidTN}
\end{eqnarray}
where $\Gamma(x)$ is the Gamma function. Equation (\ref{superfluidTN}) contains three free parameters: the non-condensate fraction of the moment of inertia, $\lambda=I_{c}/I_{0}$, the decorrelation time-scale, $\eta(R^{-1})^{-1}$, and $q$. Here, $I_{c}$ is the moment of inertia of the crust plus the rigidly rotating charged fluid plus entrained neutrons, $I_{0}$ is the total moment of inertia, and we define $\eta(R^{-1})=(2\pi)^{-1/2}\epsilon^{1/3}R^{-2/3}\gamma$, where $\epsilon$ is the energy dissipation rate per unit enthalpy (which, in general, is a function of the spin-down rate), $\gamma=\tau_{\rm eddy}/\tau_{\rm turb}\le1$ is the ratio of the eddy turnover time-scale to the characteristic time-scale over which turbulent structures change (which is longer in general due to pinning), and $R$ is the stellar radius.
The value of the exponent, $q$, in equation (\ref{superfluidTN}) depends on the form of the superfluid velocity two-point decorrelation function. \citet{melatos10} executed a first attempt to calculate the velocity correlation function numerically on the basis of Hall-Vinen-Bekarevich-Khalatnikov superfluid simulations \citep{peralta08}, but it is not well understood for terrestrial turbulence experiments, let alone for a neutron star interior, especially when stratification plays a role \citep[e.g.,][and references therein]{lasky13a}. An empirical choice is therefore made that reproduces the asymptotic power-law dependence from timing noise data, i.e., $\Phi_{\rm TN}\propto f^{-q}$ as $f\rightarrow\infty$ [for details see \citet{melatos14,melatos14b}]. We emphasise equation (\ref{superfluidTN}) is not a unique choice, nor can it be inverted uniquely to infer the underlying velocity correlation function \citep{melatos14b}.
In addition to the power-law scaling at high-frequencies, the superfluid turbulence model predicts a plateau at $f\lesssim f_{c}\approx\eta(R^{-1})$. For time intervals greater than $\sim1/\eta(R^{-1})$, turbulent motions throughout the star decohere, implying torque fluctuations exerted on the crust become statistically independent. By expanding equation (\ref{superfluidTN}) for $f\ll f_{c}$ and $f\gg f_{c}$, and evaluating the resultant expression in terms of equation (\ref{phi}), we find
\begin{eqnarray}
A_{\rm TN}\!\!\!\!\!\!&=&\!\!\!\!\!\!\frac{9\Gamma(q/2)}{16(2\pi)^{67/6}\sqrt{2}\Gamma[(q-1)/2]\eta(R^{-1})\lambda^{2}}\nonumber\\
&&\times\left(16\pi^{4}+\frac{120\pi^{2}}{13}+\frac{45}{8}\right),\label{superfluidTNapproxATN}\\
f_{c}\!\!\!\!\!\!&=&\!\!\!\!\!\!\frac{\eta(R^{-1})}{\left(2\pi\right)^{1/3}}\Big[-10\left(\frac{16\pi^{4}}{q-10}+\frac{12\pi^{2}}{q-13}+\frac{9}{q-16}\right)\nonumber\\
&&\,\,\,\,\,\,\,\,\,\,\,\,\,\,\,\,\,\,\,\,\times\left(16\pi^{4}+\frac{120\pi^{2}}{13}+\frac{45}{8}\right)^{-1}\Big]^{1/q}.\label{superfluidTNapproxfc}
\end{eqnarray}
Equations (\ref{superfluidTNapproxATN}) and (\ref{superfluidTNapproxfc}) relate the phenomenological model in section \ref{mintime} to the specific physical model in this section. A similar approach applies equally to other models.
In figure \ref{superfluidPSD1} we show four examples of millisecond pulsar phase residual power spectra measured by the PPTA \citep{manchester13}.
Overplotted on the data are reasonable `by-eye' fits generated by the superfluid turbulence model for $q=2$, 4 and $6$. The fits are neither unique nor optimal (e.g., in a least-squares sense), but they are representative. It is outside the scope of this paper to extract detailed fits and values for $\lambda$, $\eta(R^{-1})$, and $q$ for each pulsar\footnote{The amplitude and spectral index of red-noise in pulsar timing residuals are highly covariant, especially when only the lowest few frequency bins show evidence for red noise \cite[e.g.,][]{vanhaasteren09, vanhaasteren13}. Finding best-fit parameters for the superfluid turbulence model is therefore a non-trivial task that will be the subject of future work.}. We simply note that a broad range of parameters fit the phase residuals for any given pulsar.
The pulsars shown in figure \ref{superfluidPSD1} have been chosen as they appear to have moderate to high levels of timing noise, cf. other PPTA pulsars. All exhibit a relatively red spectrum. In the context of superfluid turbulence, they imply $f_{c}\gtrsim10^{-2}$ yr$^{-1}$, so that the plateau is potentially observable in the not-too-distant future\footnote{We note that PSR J1824$-$2452A resides in a globular cluster \citep{lyne87}, implying most of the timing noise is likely a result of motions within that cluster rather than superfluid turbulence. The curves shown in figure \ref{superfluidPSD1} therefore represent an upper limit on the contribution from superfluid turbulence.}.
In figure \ref{superfluidPSD2} we plot two further examples of millisecond pulsar phase residuals. These objects exhibit the lowest level of timing noise in the PPTA sample. For the superfluid turbulence model to remain consistent with these data, the objects must have long decorrelation time-scales, i.e., $f_{c}\lesssim10^{2}$ yr$^{-1}$. The data show the white noise component, $A_{\rm W}$, and the turbulence-driven red-component sits below $A_{\rm W}$. Under these circumstances, the turnover in $\Phi_{\rm TN}(f)$ occurs too low in frequency to be observed, and the main factor limiting PTA detection is $A_{\rm W}$.
\begin{figure}
\includegraphics[angle=0,width=0.95\columnwidth]{SuperfluidPSD.pdf}
\caption{\label{superfluidPSD1}
Phase residual power spectra, $\Phi(f)$ for four millisecond pulsars from the Parkes Pulsar Timing Array (thick black curves). Overplotted are theoretical curves generated by the superfluid turbulence model for the physical parameters $\left(\lambda,\,\eta,\,q\right)$ specified in the legend. The corresponding values of $A_{\rm TN}$ and $f_{c}$ for each curve are shown in figure \ref{fcAtn}.
}
\end{figure}
\begin{figure}
\includegraphics[angle=0,width=0.95\columnwidth]{SuperfluidPSD_0437.pdf}
\caption{\label{superfluidPSD2} Same as for figure \ref{superfluidPSD1}, but for PSRs J0437-4715 and J1909-3744, two pulsars from the PPTA catalogue with the lowest level of timing noise. Theoretical curves for the superfluid turbulence model require smaller values of $f_{c}\approx\eta(R^{-1})$ than in figure \ref{superfluidPSD1} to remain consistent with the data.
}
\end{figure}
\section{Optimal Pulsars}\label{optimal}
What pulsars are best placed to detect a gravitational wave background, given the longest time one is prepared to wait? In figure \ref{fcAtn} we plot $1/f_{c}$ against $A_{\rm TN}$, for different values of $q$ and $T_{\rm min}$ in each panel. The left-hand vertical axis displays the results for the model-independent form of $\Phi_{\rm TN}$ in equation (\ref{phi}). The right-hand vertical axis registers the decorrelation time $1/\eta(R^{-1})$, in the superfluid turbulence model in section \ref{superfluidmodel}. The dashed grey curves are curves of constant $\lambda$. Overplotted are the superfluid turbulence model `fits' to the PPTA pulsar data in figure \ref{superfluidPSD1}, where the open circles, closed circles, open squares and closed squares are PSRs J1024$-$0719, J1643$-$1224, J1824$-$2452A and J1939$-$2134 respectively.
\begin{figure*}
\includegraphics[angle=0,width=0.95\linewidth]{fcAtn.pdf}
\caption{\label{fcAtn} Reciprocal of the timing noise turnover frequency, $1/f_{c}$, (left-hand vertical axis), as a function of the normalisation, $A_{\rm TN}$, (horizontal axis) for equation (\ref{phi}), for different values $T_{\rm min}$ as defined by equation (\ref{Tmin}) and $q$. The right-hand vertical axis shows the decorrelation time-scale in the context of the superfluid turbulence model in section \ref{superfluidmodel}. The dashed grey curves are $\lambda$ contours with $\lambda=1$, 0.5, 0.2, 0.1, 0.05, 0.02, 0.01, 0.005 and 0.002 going from left to right in each panel. The shaded blue, green and grey regions, the thick blue curve and the dotted black curves represent the same theoretical limits for SMBHs and cosmic strings as in figure \ref{ObsTime}. Overplotted are the superfluid turbulence model fits to the PPTA observational data presented in figure \ref{superfluidPSD1}, where the open circles, closed circles, open squares and closed squares are PSRs J1024$-$0719, J1643$-$1224, J1824$-$2452A and J1939$-$2134 respectively.
}
\end{figure*}
Figure \ref{fcAtn} allows us to ask whether, for example, $20$ yr of timing a specific pulsar will allow for sensitivity to the most optimistic SMBH gravitational wave strain of $A_{\rm GW}=2.4\times10^{-15}$. In the middle set of panels, the latter strain limit appears as the
right-most boundary of the blue shaded region. A pulsar with timing noise below this curve is sensitive to a gravitational wave signal in $T_{\rm obs}\le20\,{\rm yr}$. Sensitivity depends on $q$ as illustrated in the three different panels running vertically. It also depends on $f_{c}$. For example, a hypothetical pulsar with $q=4$ and $A_{\rm TN}\approx10^{-11}$ yr is only sensitive to a gravitational wave background for $1/f_{c}\lesssim12$ yr. This is an interesting constraint: a pulsar
in a PTA that tolerates $T_{\rm min}\le20$ yr is sensitive to a gravitational wave background if $\Phi_{\rm TN}(f)$ exhibits a plateau after $\lesssim12$ yr of timing.
The superfluid turbulence model fits from figure \ref{superfluidPSD1} give an indication as to the usefulness of individual pulsars from the PPTA dataset. For example, consider PSR J1939$-$2134 (closed squares). If one again tolerates $T_{\rm min}\le20$ yr, the fits imply
a pulsar is sensitive to a conservative prediction for the gravitational wave background
for $q\lesssim4$, although for $q\approx4$ this requires the timing noise spectrum to plateau after approximately 15 yr of timing. We emphasise again that the model fits should only be taken as indicative; careful and detailed analysis is required to extract the true timing noise signal parameters from the data.
\section{Conclusion}\label{conclusion}
Pulsar Timing Array limits on the cosmological gravitational wave background are continually dropping to the point where they usefully constrain galaxy formation models \citep{shannon13a}.
Positive detections, on the other hand, require a cross-correlation algorithm to simultaneously analyse timing residuals from multiple pulsars.
Such a detection will likely occur when the gravitational wave background is the largest component in the unmodelled portion of many individual pulsar's timing residuals \citep{siemens13}.
If the timing noise spectrum is steeper asymptotically (at high $f$) than the gravitational wave spectrum, this is only possible if the timing noise spectrum flattens below some frequency, $f_c$.
In this paper, we calculate the minimum observation time required, given $f_c$, before the gravitational wave background rises above the timing noise plateau in any specific pulsar. We calculate this minimum observation time both in a model-independent way, and for timing noise arising from superfluid turbulence. The latter model is selected not because it is necessarily preferred physically, but because it is simple, predictive and analytically tractable and therefore provides a test-bed for repeating the calculation with other physical models in the future.
Our results rely on the timing noise spectrum whitening below some threshold frequency, $f_c$. This provides an observational diagnostic that can be used to infer
the effectiveness of an individual pulsar in a PTA.
If, upon observing a pulsar for some $T>1/f_c$, one finds that $\Phi_{\rm TN}(f)$ has not whitened below $f_c$, that
pulsar's capacity for assisting usefully in the detection of a gravitational wave background is severely diminished.
The $f_c$ for a given pulsar is a function of the rotational parameters of the pulsar, and the gravitational wave amplitude and spectral index. Therefore, using the prescription outlined in this paper, one can predict $f_c$ for a given pulsar and a given gravitational wave background.
In reality, measuring $f_{c}$ in a single pulsar is difficult. Firstly, the noise in a given pulsar timing power spectrum is large, and secondly, the power in the lowest-frequency bin is generally dominated by the fact that a quadratic polynomial is fit to the timing residuals [see \citet{vanhaasteren13}]. These two effects potentially mimic a low-frequency turn-over, implying multiple low-frequency bins are required to confirm the existence of a low-frequency cut-off.
Many data analysis algorithms simultaneously fit the timing model and the unknown noise contributions for any individual pulsar. In this sense, one can include a low-frequency plateau into gravitational-wave detection algorithms, e.g., by way of a Bayesian prior on the form of the power spectral density. Physically motivated models for timing noise, such as the superfluid turbulence model discussed herein, could be used to guide such priors.
\section*{acknowledgments}
We are grateful to the anonymous reviewer for the thoughtful and thorough review of the manuscript.
PDL and AM are supported by Australian Research Council (ARC) Discovery Project DP110103347. PDL is also supported by ARC DP140102578. VR is a recipient of a John Stocker Postgraduate Scholarship from the Science and Industry Endowment Fund. We thank Yuri Levin for comments on the manuscript and Ryan Shannon for comments on an earlier version. Calculations of the cosmic string stochastic background used the GWPlotter website: \hyperref[http://homepages.spa.umn.edu]{http://homepages.spa.umn.edu/$\sim$gwplotter}.
\bibliographystyle{mn2e}
\bsp
\label{lastpage}
\pagerange{\pageref{firstpage}--\pageref{lastpage}}
\maketitle
|
\section{The bulk solution}
We start our analysis with a brief introduction to the bulk construction, namely the Lifshitz solution in the context of New Massive Gravity (NMG)\cite{Bergshoeff}. The corresponding action in the NMG reads as \cite{AB,Bergshoeff},
\begin{equation}
S_{NMG}=\frac{1}{2\kappa}\int d^{3} x \sqrt{g}\Big[R-2\Lambda-\lambda\Big(R_{\mu\nu}R^{\mu\nu}-\frac{3}{8} R^2\Big)\Big].\label{E1}
\end{equation}
Here $\Lambda$ is the cosmological constant and $\lambda$ is the NMG coupling constant. The above action (\ref{E1}) admits a Lifshitz solution of the type,
\begin{equation}
ds^2=- \frac{r^{2z}}{l^{2z}} f(r) dt^2+\frac{dr^2}{g(r)}+\frac{r^2}{l^2} dx^2\label{E2}
\end{equation}
where \cite{AB}, $$f(r)=\frac{r^2}{l^2} g(r)=\Big(1-\frac{ M^2 l^2}{r^2}\Big)$$
together with, $$z=3,\quad \Lambda=-\frac{13}{2 l^2},\quad \lambda=-2 l^2\,.$$ The horizon is located at $r=r_{h}= M l\,.$ For convenience, we set $l=1$ in the subsequent calculations.
Interestingly in our analysis we observe that unlike the case for the usual higher derivative theories in $ AdS $ \cite{Buchel 3}, the equations of motion corresponding to the hydrodynamic (spin $ 0 $) fluctuations turn out to be fourth order linear differential equations which thereby invite a completely different as well as challenging platform to compute the hydrodynamic transports for the non relativistic $ z=3 $ plasma at strong coupling.
\section{Thermodynamics}
In this section we provide a brief discussion to the basic thermodynamic principles and/or identities which any fluid dynamical system must obey. These are the so called first law of fluid mechanics and the Euler relation \cite{LL}. Before we actually come to these issues, let us first note that the Wald entropy density ($ s $) for the above black hole solution (\ref{E2}) turns out to be \cite{AsS},
\begin{equation}
s=\frac{8 \pi r_{h}}{\kappa}\label{E3}
\end{equation}
which also matches with \cite{AB1} in the appropriate limit.
The energy density ($ \epsilon $) could be formally obtained by integrating over the Noether charge associated with the time like Killing symmetry\footnote{One can derive this result by using several other techniques for example, using the notion of renormalized holographic stress tensor \cite{stress, stress1} or dimensionally reducing to dilaton gravity \cite{dilaton}. Interestingly, all these methods yield the same result and therefore are claimed to be consistent with each other \cite{consistent}.} \cite{AB1},
\begin{equation}
\epsilon = \frac{r_{h}^4}{\kappa}.\label{E4}
\end{equation}
The Hawking temperature ($ T_h $) associated with the black hole solution (\ref{E2}) turns out to be \cite{AB}, \cite{AB1},
\begin{equation}
T_{h}=\frac{r_{h}^{z}}{2\pi}.\label{E5}
\end{equation}
It is now in fact a quite trivial task to check that the above entities (\ref{E3}), (\ref{E4}) and (\ref{E5}) trivially satisfy the first law \cite{LL},
\begin{equation}
d\epsilon= T_{h} d s.
\end{equation}
Next, we compute the pressure density ($ p $) using the Euler relation \cite{LL},
\begin{equation}
p=T_{h} s-\epsilon =\frac{3 r_{h}^4}{\kappa}.\label{E7}
\end{equation}
Finally, using (\ref{E4}) and (\ref{E7}) the velocity of sound ($ v_{s} $) turns out to be,
\begin{equation}
v_{s}=\sqrt{\frac{\partial p}{\partial \epsilon}}=\sqrt{3}.\label{E8}
\end{equation}
Eq.(\ref{E8}) is precisely the speed of sound propagation from the perspective of the boundary hydrodynamics. The fact that $ v_s >1 $ should not come as a surprise since we are dealing with a non relativistic system corresponding to $ z=3 $ fixed point where in principle there should not be any upper bound on the speed of light. Incidentally because of this reason, the (conjectured) universality bound on the charge diffusion does not in general hold for Lifshitz like fixed points \cite{Rtiz}.
Finally, using (\ref{E4}) and (\ref{E5}) the heat capacity turns out to be,
\begin{eqnarray}
C= \frac{4}{3 \kappa }\left( 2 \pi T_{h} \right)^{1/3} \label{E9}
\end{eqnarray}
which turns out to be positive definite. Eq.(\ref{E9}) is an important ingredient in order to estimate the speed of sound ($ v_s $) inside a fluid medium \cite{kittel}. We will make more specific comments about it in the concluding section.
\section{Hydrodynamics}
In the standard framework of fluid dynamics, the velocity for sound modes are obtained by knowing the pole(s) in the expression for the retarded correlator(s) of the type $<T_{tx}T_{tx}>$ \cite{Kovtun, Policastro}. As far as the current analysis is concerned, the stress tensor corresponding to the boundary non conformal fluid could be formally expressed as,
\begin{equation}
T_{\mu\nu}=(\epsilon+p) u_{\mu}u_{\nu}+p\eta_{\mu\nu}-\zeta(u_{\mu}u_{\nu}+\eta_{\mu\nu})\partial_{\lambda}u^{\lambda}
\end{equation}
where $u_{\mu}$ is the 2-velocity of the fluid satisfying $u_{\mu}u^{\mu}=-1$ and $\zeta$ is the coefficient of bulk viscosity which appears as a consequence of the broken conformal invariance in the boundary theory. At this stage it is noteworthy to mention that for fluids in $1+1$ dimensions, it is the coefficient of bulk viscosity ($ \zeta $) that turns out to be the only existing nontrivial transport coefficient in the context of first order viscous hydrodynamics.
In order to find pole(s) in the expression for the retarded correlator(s), we first perturb the fluid in its rest frame as \cite{David},
\begin{eqnarray}
T_{tt}&=&\epsilon+\delta T_{tt}\nonumber\\
T_{tx}&=&\delta T_{tx}\nonumber\\
T_{xx}&=&p+ \delta T_{xx}\nonumber\\
u_{t}&=&1,~~ u_{x}=\delta u_{x}.\label{E10}
\end{eqnarray}
Next, substituting the above perturbations (\ref{E10}) first into the conservation equations $ \partial_{\mu}T^{\mu\nu} =0$ and then performing the Fourier transform for each of the thermodynamic entities it is quite straightforward to obtain the following algebraic relation \cite{David},
\begin{equation}
\Big(-i\omega^2+i q^2 \frac{\partial p}{\partial \epsilon}+\frac{\zeta}{\epsilon+p}\omega q^2\Big)\delta T_{tx}=0.\label{E11}
\end{equation}
The dispersion relation that follows trivially from (\ref{E11}) could be formally expressed as,
\begin{equation}
\omega=\pm v_{s} q-i \frac{\zeta}{2 (\epsilon+p)}q^2 \label{E12}
\end{equation}
where, $v_{s}^2=\frac{\partial p}{\partial \epsilon}$ is the square of the sound velocity in the non conformal fluid. Using the Euler relation,
\begin{equation}
\epsilon + p=T_{h} s
\end{equation}
one can finally express (\ref{E12}) in a more sophisticated form as,
\begin{equation} \label{dis}
\omega=\pm v_{s} q-i \frac{\zeta}{2 T_{h} s}q^2.
\end{equation}
With the above set up in hand, our next task is twofold, first, to check the validity of the conjectured upper bound on the velocity of sound \cite{Cherman, Hohler} for non conformal fluids corresponding to $ z=3 $ Lifshitz fixed point and second, to explore the effect of \textit{non perturbative} higher derivative (NMG) corrections on the bulk viscosity to entropy ($ \zeta/s $) ratio in $ 1+1 $ dimensions and comment on its conjectured lower bound \cite{Buchel,David}.
\section{Sound Modes and bulk viscosity}
The first step towards computing the holographic sound modes is to turn on the scalar (with respect to boundary rotational symmetry) perturbations in the bulk namely \cite{Policastro},
\begin{eqnarray}
h_{\mu\nu}= \Big\{h_{tt},h_{tx},h_{xx}\Big\}
\end{eqnarray}
where we have fixed the gauge $ h_{\mu r} =0$. This gauge fixing does not fully exhaust all the gauge freedoms of the full theory. The equations of motion corresponding to the above scalar fluctuations exhibit residual gauge and/or diffeomorphism invariance \cite{Policastro}. As a result the most general gauge invariant fluctuation turns out to be,
\begin{eqnarray}
\mathcal{Z}&=-c_{1}(r) H_{tt}(r)+c_{2}(r) H_{tx}(r)+c_{3}(r) H_{xx}(r)\nonumber\\&-q^2 \Big(r^{2z}f(r) H_{tt}(r)\Big)^{'}+2 q \omega \Big(r^2 H_{tx}(r)\Big)^{'}\nonumber \\&+ \omega ^2 \Big(r^2 H_{xx}(r)\Big)^{'}\label{E16}
\end{eqnarray}
where the individual coefficients ($ c_i $) turn out to be,
\begin{widetext}
\begin{equation} \nonumber
\resizebox{.6\textwidth}{!}{$\displaystyle c_{1}(r)= \frac{\left(q^2 r^3 ( M-r) ( M+r) \left(q^2 r^2 \left(6 M^2-23 M r^2+18 r^4\right)+ \omega ^2 \left( M-2 r^2\right)\right)\right)}{\left( M-r^2\right) \left( \omega ^2+q^2 r^2 \left(2 M-3 r^2\right)\right)}$}\end{equation}
\begin{equation} \nonumber
\resizebox{.4\textwidth}{!}{$\displaystyle c_{2}(r)=2 q r^2 \omega \left(\frac{r}{ M-r^2}-\frac{4 q^2 r \left( M-3 r^2\right)}{ \omega ^2+q^2 r^2 \left(2 M-3 r^2\right)}-\frac{1}{r}\right)$}\end{equation}
\begin{equation}
\resizebox{.45\textwidth}{!}{$\displaystyle c_{3}(r)=\frac{ \left( r \omega ^2 \left(q^2 r^2 \left(6 M^2-23 M r^2+18 r^4\right)+ \omega ^2 \left( M-2 r^2\right)\right)\right)}{\left(r^2- M\right) \left( \omega ^2+q^2 r^2 \left(2 M-3 r^2\right)\right)}$}.
\end{equation}
\end{widetext}
In order to arrive at (\ref{E16}) we have used the following metric redefinitions namely,
\begin{eqnarray}
h_{xx}&=& e^{-i \omega t+ i q x} r^2 H_{xx}(r)\nonumber\\
h_{tx}&= &e^{-i \omega t+ i q x} r^2 H_{tx}(r)\nonumber\\
h_{tt}&=& -e^{-i \omega t+ i q x}r^{2z} f(r) H_{tt}(r).
\end{eqnarray}
Note that here prime denotes derivative with respect to the radial coordinate ($ r $).
Using the linearized equations of motion of individual scalar perturbations, it is indeed quite interesting to note that unlike the previous examples in the literature \cite{Policastro}, the gauge invariant combination (\ref{E16}) satisfies a sixth order linear differential equation of the type,
\begin{eqnarray}
&A_{0}(r,\omega,q) \mathcal{Z}^{''''''}+A_{1}(r,\omega,q) \mathcal{Z}^{'''''}+A_{2}(r,\omega,q) \mathcal{Z}^{''''}+\nonumber\\&A_{3}(r,\omega,q)\mathcal{ Z}^{'''}+A_{4}(r,\omega,q)\mathcal{ Z}^{''}+A_{5}(r,\omega,q) \mathcal{Z}^{'}+\nonumber\\&A_{6}(r,\omega,q) \mathcal{Z}=0\label{E19}
\end{eqnarray}
where details of the coefficients ($ A_i $) are too cumbersome and not quite illuminating.
Next, we explore (\ref{E19}) in two asymptotic regions namely, near the boundary ($ r\rightarrow \infty $) of the spacetime and at the horizon ($ r=r_h=1 $).
We first consider the following ansatz for $\mathcal{Z}$,
\begin{equation} \label{z}
\mathcal{Z}=\left(1-\frac{1}{r}\right)^{\alpha}F(r).
\end{equation}
Considering the near horizon expansion and considering the ingoing wave boundary condition \cite{Policastro}, the parameter $\alpha$ could be read off as,
\begin{eqnarray}
\alpha=-\frac{i \omega}{4\pi T_{h}}=-\frac{i \omega}{2}.
\end{eqnarray}
Next, considering the large $ r $ limit we expand $F(r)$ in the frequency ($ \omega $) as well as in the momentum ($q$),
\begin{equation}
F(r)=F_{0}(r)+ \omega F_{1}(r)+ q\, F_{2}(r)+\omega^2 F_{3} (r)+ \omega\, q F_{4}(r)+ q^2 F_{5}(r)
\end{equation}
and extract the finite piece in the limit $r\rightarrow \infty$. Keeping those finite pieces intact and considering the asymptotic normalization condition \cite{Policastro},
\begin{equation}
F(\infty)=0\label{E24}
\end{equation}
we finally arrive at the cherished dispersion relation of the following form \footnote{As the gauge invariant combination satisfies a six order differential equation (\ref{E19}) and there are in fact two boundary conditions namely, the in going wave boundary condition and the asymptotic boundary condition (\ref{E24}), therefore in principle we can uniquely fix two of the six unknown coefficients. Surprisingly the special combination of the remaining four coefficients get uniquely fixed once we identify the leading coefficient (associated with $ q $) on the r.h.s of (\ref{E25}) as the speed of sound propagation (\ref{E8}) for the boundary hydrodynamics. This automatically fixes the $ \zeta/s $ ratio too.},
\begin{equation}
\omega= \sqrt{3}q-\frac{i}{2} \Gamma q^2 \label{E25}
\end{equation}
where,
\begin{equation}
\Gamma T_{h}=\frac{\zeta}{s}=\frac{1}{4\pi}\left(1-\frac{3\lambda}{4}\right)\approx 0.1989 \label{E26}
\end{equation}
which clearly seems to preserve the conjectured lower bound \cite{Buchel, David} as $ \lambda$ is negative. Eq.(\ref{E26}) represents the full \textit{non perturbative} NMG correction to $ \zeta/s $ ratio in $ 1+1 $ dimensions. In \cite{Brustein:2009tk}, it has been argued for the first time that the presence of effective coupling for gravitons could modify the sound damping coefficient. However, the arguments of \cite{Brustein:2009tk} are mostly valid for conformal case, whereas on the other hand, our analysis might be regarded as the first non trivial example in the context of non conformal fluids.
\section{Summary and final remarks}
Let us now summarize the key findings of our analysis. The present analysis has two major outcomes. The first observation tells us that the speed of sound propagation ($ v_s $) inside a $z=3$ strongly coupled plasma in $ 1+1 $ dimensions clearly exceeds the conjectured upper bound. This observation is not surprising in the context of non relativistic hydrodynamics where the Lorentz boost symmetry is explicitly broken near the UV fixed point of the theory. Finally and most importantly, we note that the nonperturbative NMG correction to $\zeta/s$ ratio does not violate the conjectured lower bound in $1+1$ dimensions. This is the first nonperturbative higher derivative correction to $\zeta/s$ ratio to the best of our knowledge.
There are some real life condensed matter systems where $ z=3 $ fixed point is realized exactly \cite{conmat,conmat1}. Information regarding the speed of sound could in principle be inferred by utilizing certain specific data in a systematic manner. Heuristically, thermal energy is related to the density of states in the following manner \cite{kittel},
\begin{center}
\begin{equation} \label{eqn}
E_{TH}=\int d\omega \mathfrak{D}(\omega) <n(\omega)> \hbar \omega
\end{equation}
\end{center}
where the density of the states ($\mathfrak{D}(\omega) $) is related to the group velocity of the phonon excitations of the medium,
\begin{equation} \label{eqn1}
\mathfrak{D}(\omega) d\omega=\int \frac{dl_{\omega}}{v_{s}} d\omega
\end{equation}
where $l_{\omega}$ is the line element over the two dimensional surface in the momentum space . Combining (\ref{eqn}) and (\ref{eqn1}) one can in fact relate thermal energy with the speed of sound propagation ($ v_s $) as,
\begin{equation}
E_{TH}=\int d\omega \frac{dl_{\omega}}{v_{s}} <n(\omega)> \hbar \omega\,.
\end{equation}
Finally, the specific heat could be found as,
\begin{equation}
C=\frac{\partial E_{TH}}{\partial T}\label{E30}
\end{equation}
which therefore suggests that one should in principle be able to estimate the speed of sound ($ v_s $) from the knowledge of heat capacity ($ C $) (\ref{E30}). In other words, experimentally one can measure the velocity of sound by knowing the corresponding data on specific heat for the underlying system. There are in fact some systems where the dynamical exponent ($ z $) is very close to 3 \cite{conmat}, and the specific heat measurements are also performed for them. So we hope that the information regarding the speed of sound ($ v_s $) could in principle be estimated for fluids in $ 1+1 $ dimensions and this opens up the possibility to test various predictions of Gauge/gravity duality in near future.
\\
\textbf{Acknowledgements:} Authors acknowledge Aninda Sinha and Justin David for useful discussions. Authors also would like convey their special thanks to Alex Buchel for going through the draft and giving valuable comments. The authors also acknowledge the financial support from CHEP, IISc, Bangalore.
|
\section{Introduction}
Consider two probability distributions $P$ and $Q$ defined on a common
measurable space $(\ensuremath{\mathcal}{A}, \mathscr{F})$.
The Csisz\'{a}r-Kemperman-Kullback-Pinsker inequality (a.k.a. Pinsker's inequality)
states that
\begin{align}
\label{eq: Pinsker}
\tfrac12 \, |P-Q|^2 \, \log e \leq D(P \| Q)
\end{align}
where
\begin{align} \label{eq: relative entropy}
D(P \| Q) = \ensuremath{\mathbb{E}}_P\left[ \log \frac{\text{d}P}{\text{d}Q} \right]
= \int_{\ensuremath{\mathcal}{A}} \text{d}P(a) \, \log \frac{\text{d}P}{\text{d}Q} \, (a)
\end{align}
designates the relative entropy (a.k.a. the Kullback-Leibler divergence)
from $P$ to $Q$, and
\begin{align} \label{eq: TV distance}
|P-Q| = 2 \, \sup_{\mathcal{F} \in \mathscr{F}} |P(\mathcal{F}) - Q(\mathcal{F})|
\end{align}
is the total variation distance between $P$ and $Q$.
A ``reverse Pinsker inequality" providing an upper bound on the relative
entropy in terms of the total variation distance does not exist in general
since we can find distributions that are arbitrarily close in total variation
but with arbitrarily high relative entropy.
Nevertheless, it is possible to introduce constraints under which such
reverse Pinsker inequalities can be obtained.
In the case where the probability measures $P$ and $Q$ are defined
on a common discrete (i.e., finite or countable) set~$\ensuremath{\mathcal}{A}$,
\begin{align}
\label{eq: relative entropy - discrete}
& D(P \| Q) = \sum_{a \in \ensuremath{\mathcal}{A}} P(a) \, \log \frac{P(a)}{Q(a)}, \\
\label{eq: total variation - discrete}
& |P-Q| = \sum_{a \in \ensuremath{\mathcal}{A}} \bigl|P(a) - Q(a)\bigr|.
\end{align}
One of the implications of \eqref{eq: Pinsker} is that convergence
in relative entropy implies convergence in total variation distance. The total
variation distance is bounded $|P-Q| \leq 2$, whereas the relative
entropy is an unbounded information measure.
Improved versions of Pinsker's inequality were studied, e.g., in \cite{FedotovHT_IT03},
\cite{Gilardoni10}, \cite{OrdentlichW_IT2005}, \cite{ReidW11}, \cite{Vajda_IT1970}.
A ``reverse Pinsker inequality" providing an upper bound on the relative
entropy in terms of the total variation distance does not exist in general
since we can find distributions that are arbitrarily close in total variation
but with arbitrarily high relative entropy.
Nevertheless, it is possible to introduce constraints under which such
reverse Pinsker inequalities can be obtained.
In the case of a finite alphabet $\ensuremath{\mathcal}{A}$, Csisz\'ar and Talata
\cite[p.~1012]{CsiszarT_IT06} show that
\begin{align} \label{eq: CsTa}
D(P \| Q) \leq \left(\frac{\log e}{Q_{\min}} \right)
\cdot |P-Q|^2,
\end{align}
where
\begin{align} \label{eq: Q_min}
Q_{\min} \triangleq \min_{a \in \ensuremath{\mathcal}{A}} Q(a).
\end{align}
Recent applications of \eqref{eq: CsTa} can be found in \cite[Appendix~D]{KostinaV15} and
\cite[Lemma~7]{TomamichelT_IT13} for the analysis of the third-order asymptotics
of the discrete memoryless channel with or without cost constraints.
In addition to $Q_{\min}$ in \eqref{eq: Q_min}, the bounds in this paper involve
\begin{align}
\label{eq: beta1}
& \beta_1 = \min_{a\in \ensuremath{\mathcal}{A}} \frac{Q(a)}{P(a)}, \\
\label{eq: beta2}
& \beta_2 = \min_{a\in \ensuremath{\mathcal}{A}} \frac{P(a)}{Q(a)}
\end{align}
so, $\beta_1, \beta_2 \in [0,1]$.
In this paper, Section~\ref{section: Reksnip} derives a reverse Pinsker inequality
for probability measures defined on a common finite set, improving the bound in \eqref{eq: CsTa}.
The utility of this inequality is studied in Section~\ref{sec: applications}, and it
is extended in Section~\ref{section: RD} to R\'{e}nyi divergences
of an arbitrary non-negative order.
\section{A New Reverse Pinsker Inequality for Distributions on a Finite Set}
\label{section: Reksnip}
The present section introduces a strengthened version of \eqref{eq: CsTa},
followed by some remarks and an example.
\subsection{Main Result and Proof}
\begin{theorem} \label{thm: reksnip - improved}
{\em Let $P$ and $Q$ be probability measures defined on a common finite set $\ensuremath{\mathcal}{A}$, and assume
that $Q$ is strictly positive on $\ensuremath{\mathcal}{A}$. Then, the following inequality holds:
\begin{align}
\label{eq: UB-RE-FS1}
D(P \| Q) & \leq \log \left(1 + \frac{|P-Q|^2}{2 Q_{\min}} \right) - \frac{\beta_2 \log e}{2} \cdot |P-Q|^2 \\
\label{eq: UB-RE-FS2}
& \leq \log \left(1 + \frac{|P-Q|^2}{2 Q_{\min}} \right)
\end{align}
where $Q_{\min}$ and $\beta_2$ are given in \eqref{eq: Q_min} and
\eqref{eq: beta2}, respectively.}
\end{theorem}
\begin{proof}
Theorem~\ref{thm: reksnip - improved} is proved by obtaining upper and lower bounds on
the $\chi^2$-divergence from $P$ to~$Q$
\begin{align} \label{eq: chi-square}
\chi^2(P\|Q) \triangleq \sum_{a \in \ensuremath{\mathcal}{A}} \frac{(P(a)-Q(a))^2}{Q(a)}.
\end{align}
A lower bound follows by invoking Jensen's inequality
\begin{align}
\label{eq1: lb chi-square divergence}
\chi^2(P\|Q) & = \sum_{a \in \ensuremath{\mathcal}{A}} \frac{P(a)^2}{Q(a)} - 1 \\
\label{eq2: lb chi-square divergence}
& = \sum_{a \in \ensuremath{\mathcal}{A}} P(a) \, \exp\left( \log \frac{P(a)}{Q(a)} \right) - 1 \\
\label{eq3: lb chi-square divergence}
& \geq \exp\left(\sum_{a \in \ensuremath{\mathcal}{A}} P(a) \, \log \frac{P(a)}{Q(a)} \right) - 1 \\[0.1cm]
\label{eq4: lb chi-square divergence}
& = \exp\bigl( D(P \| Q) \bigr) - 1.
\end{align}
Alternatively, \eqref{eq4: lb chi-square divergence} can be obtained by combining the equality
\begin{align}
\chi^2(P\|Q) = \exp\bigl(D_2(P\|Q)\bigr) - 1
\end{align}
with the monotonicity of the R\'{e}nyi divergence $D_{\alpha}(P\|Q)$ in $\alpha$, which implies
that $D_2(P\|Q) \geq D(P\|Q)$.
A refined version of \eqref{eq4: lb chi-square divergence} is derived in the following. The starting
point is a refined version of Jensen's inequality in \cite[Lemma~1]{ISSV15}, generalizing a result from
\cite[Theorem~1]{Dragomir06}), which leads to (see \cite[Theorem~7]{ISSV15})
\begin{align}
& \min_{a \in \ensuremath{\mathcal}{A}} \frac{P(a)}{Q(a)} \cdot D(Q\|P) \nonumber \\[0.1cm]
\label{eq0: re and chi^2}
& \leq \log \bigl( 1 + \chi^2(P\|Q) \bigr) - D(P\|Q) \\[0.1cm]
\label{eq: re and chi^2}
& \leq \max_{a \in \ensuremath{\mathcal}{A}} \frac{P(a)}{Q(a)} \cdot D(Q\|P).
\end{align}
From \eqref{eq: re and chi^2}
and the definition of $\beta_2$ in \eqref{eq: beta2}, we have
\begin{align}
& \chi^2(P\|Q) \nonumber \\[0.1cm]
\label{eq1: refined lb chi^2 divergence}
& \geq \exp \Bigl( D(P\|Q) + \beta_2 \, D(Q \|P) \Bigr) - 1 \\[0.1cm]
\label{eq2: refined lb chi^2 divergence}
& \geq \exp \left( D(P \|Q) + \frac{\beta_2 \, \log e}{2} \cdot |P-Q|^2 \right) - 1
\end{align}
where \eqref{eq1: refined lb chi^2 divergence} follows from
\eqref{eq0: re and chi^2} and the definition of $\beta_2$ in \eqref{eq: beta2},
and \eqref{eq2: refined lb chi^2 divergence} follows from Pinsker's inequality
\eqref{eq: Pinsker}. Note that the lower bound in
\eqref{eq2: refined lb chi^2 divergence} refines the lower bound in
\eqref{eq4: lb chi-square divergence} since $\beta_2 \in [0,1]$.
An upper bound on $\chi^2(P\|Q)$ is derived as follows:
\begin{align}
\chi^2(P\|Q) & = \sum_{a \in \ensuremath{\mathcal}{A}} \frac{(P(a)-Q(a))^2}{Q(a)} \nonumber \\[0.1cm]
\label{eq1: ub1}
& \leq \frac{\sum_{a \in \ensuremath{\mathcal}{A}} \bigl(P(a)-Q(a)\bigr)^2}{Q_{\min}} \\[0.1cm]
\label{eq2: ub1}
& = \frac{|P-Q|}{Q_{\min}} \cdot \max_{a \in \ensuremath{\mathcal}{A}} |P(a)-Q(a)|
\end{align}
and, from \eqref{eq: TV distance},
\begin{align}
|P-Q| \geq 2 \max_{a \in \ensuremath{\mathcal}{A}} |P(a)-Q(a)|.
\label{eq: ub2}
\end{align}
Combining \eqref{eq2: ub1} and \eqref{eq: ub2} yields
\begin{align}
\chi^2(P\|Q) \leq \frac{|P-Q|^2}{2Q_{\min}}.
\label{eq: ub chi-square divergence}
\end{align}
Finally, \eqref{eq: UB-RE-FS1} follows by combining the upper and lower bounds
on the $\chi^2$-divergence in \eqref{eq2: refined lb chi^2 divergence} and
\eqref{eq: ub chi-square divergence}.
\end{proof}
\begin{remark}
{\em It is easy to check that Theorem~\ref{thm: reksnip - improved} strengthens the
bound by Csisz\'ar and Talata in \eqref{eq: CsTa} by at least a factor of~2
since upper bounding the logarithm in \eqref{eq: UB-RE-FS1} gives
\begin{align} \label{betterthanCT}
D(P \| Q) \leq \frac{(1-\beta_2 \, Q_{\min}) \log e}{2 Q_{\min}} \cdot {|P-Q|^2}.
\end{align}}
\end{remark}
\vspace*{0.2cm}
In the finite-alphabet case, we can obtain another upper bound on $D(P\|Q)$ as a function of the $\ell_2$ norm
$\| P - Q \|_2$:
\begin{align} \label{eq: UB-RE-FS}
D(P \| Q) \leq \log \left(1 + \frac{\|P-Q\|_2^2}{Q_{\min}} \right) - \frac{\beta_2 \log e}{2} \cdot \|P-Q\|_2^2
\end{align}
which follows by combining \eqref{eq2: refined lb chi^2 divergence}, \eqref{eq1: ub1}, and
$\|P-Q\|_2 \leq |P-Q|$. Using the inequality $\log (1 + x) \leq x \log e$ for $x \geq 0$
in the right side of \eqref{eq: UB-RE-FS}, and also loosening this bound by ignoring the
term $\frac{\beta_2 \log e}{2} \cdot \|P-Q\|_2^2$, we recover the bound
\begin{align} \label{eq: UB-RE-FS-looser}
D(P \| Q) \leq \frac{\|P-Q\|_2^2 \, \log e}{Q_{\min}}
\end{align}
which appears in the proof of Property~4 of \cite[Lemma~7]{TomamichelT_IT13},
and also used in \cite[(174)]{KostinaV15}.
\vspace*{0.1cm}
\begin{remark}
{\em The lower bounds on the $\chi^2$-divergence in \eqref{eq4: lb chi-square divergence}
and \eqref{eq2: refined lb chi^2 divergence} improve the one in \cite[Lemma~6.3]{CsiszarT_IT06}
which states that $D(P \| Q) \leq \chi^2(P\|Q) \log e$.}
\end{remark}
\vspace*{0.1cm}
\begin{remark}
{\em Reverse Pinsker inequalities have been also derived in quantum information theory
(\cite{AE1, AE2}), providing upper bounds on the relative entropy of two quantum
states as a function of the trace norm distance when the minimal eigenvalues of the
states are positive (c.f. \cite[Theorem~6]{AE1} and \cite[Theorem~1]{AE2}).
These type of bounds are akin to the weakend form in \eqref{eq: UB-RE-FS2}. When the variational
distance is much smaller than the minimal eigenvalue (see \cite[Eq.~(57)]{AE1}),
the latter bounds have a quadratic scaling in this distance, similarly to
\eqref{eq: UB-RE-FS2}; they are also inversely proportional to the minimal eigenvalue,
similarly to the dependence of \eqref{eq: UB-RE-FS2} in $Q_{\min}$.}
\end{remark}
\section{Applications of Theorem~\ref{thm: reksnip - improved}}
\label{sec: applications}
\subsection{The Exponential Decay of the Probability for a Non-Typical Sequence}
\label{subsection: typicality}
To exemplify the utility of Theorem~\ref{thm: reksnip - improved}, we bound the function
\begin{align}\label{def:ldeltaq}
L_{\delta} ( Q ) = \min_{P \not \in \mathcal{T}_\delta (Q )} D(P\|Q)
\end{align}
where we have denoted the subset of probability measures on $ (\ensuremath{\mathcal}{A}, \mathscr{F})$
which are $\delta$-close to $Q$ as
\begin{align}
\mathcal{T}_\delta (Q ) = \Bigl\{P
\colon \forall \, a \in \ensuremath{\mathcal}{A}, \; \; |P(a)-Q(a)| \leq \delta \, Q(a) \Bigr\}
\end{align}
Note that $(a_1, \ldots , a_n )$ is strongly $\delta$-typical according to $Q$ if its
empirical distribution belongs to $\mathcal{T}_\delta (Q )$. According to Sanov's theorem
(e.g. \cite[Theorem~11.4.1]{Cover_Thomas}), if the random variables are independent
distributed according to $Q$, then the probability that
$(Y_1, \ldots , Y_n)$, is not $\delta$-typical
vanishes exponentially with exponent $L_{\delta} ( Q )$.
\par
To state the next result, we invoke the following notions from \cite{OrdentlichW_IT2005}.
Given a probability measure $Q$, its \textit{balance coefficient} is given by
\begin{align} \label{balanceOW}
\beta_Q = \inf_{A\in \mathscr{F}\colon Q(A) \geq \frac12} Q(A).
\end{align}
The function $\phi\colon(0, \tfrac12] \to [\tfrac12 \log e, \infty)$ is given by
\begin{align}
\label{eq: phi refined pinsker}
\phi(p) = \left\{
\begin{array}{ll}
\frac1{4(1-2p)} \, \log \left( \frac{1-p}{p} \right), & p \in
\bigl(0, \tfrac12 \bigr), \\[0.2cm]
\tfrac12 \log e, & p=\tfrac12 .
\end{array}
\right.
\end{align}
\begin{theorem}
{\em If $Q_{\min}>0$, then
\begin{align}
\label{potalo}
\phi(1 - \beta_Q) \, Q_{\min}^2 \, \delta^2
&\leq L_{\delta} ( Q ) \\
&\leq \log \left(1 + 2 Q_{\min} \, \delta^2 \right)
\label{potaup}
\end{align}
where \eqref{potaup} holds if $\delta \leq Q_{\min}^{-1} -1$.}
\end{theorem}
\vspace*{0.2cm}
\begin{proof}
Ordentlich and Weinberger \cite[Section~4]{OrdentlichW_IT2005}
show the refinement of Pinsker's inequality:
\begin{align} \label{eq: OrdentlichW}
\phi(1-\beta_Q) \; |P-Q|^2 \leq D(P \| Q).
\end{align}
Note that if $Q_{\min} > 0$ then $\beta_Q \leq 1 - Q_{\min} < 1$, and
therefore $\phi(1-\beta_Q)$ is well defined and finite.
If $P \not\in \mathcal{T}_\delta (Q )$ the simple bound
\begin{align}
|P - Q| > \delta Q_{\min}
\end{align}
together with \eqref{eq: OrdentlichW} yields \eqref{potalo}.
\par
The upper bound \eqref{potaup} follows from \eqref{eq: UB-RE-FS2}
and the fact that if $\delta \leq Q_{\min}^{-1}-1$, then
\begin{align}
\min_{P \not\in \mathcal{T}_\delta (Q )} | P - Q | = 2 \delta Q_{\min}.
\end{align}
\end{proof}
If $\delta \leq Q_{\min}^{-1} - 1$, the ratio between the upper and lower bounds
in \eqref{potaup}, satisfies
\begin{align} \label{eq: exponents' ratio}
\frac1{Q_{\min}} \cdot \frac{\log e}{2 \, \phi(1-\beta_Q)} \cdot
\frac{\log \left(1 + 2 Q_{\min} \, \delta^2 \right)}{\tfrac12 \log e
\; Q_{\min} \, \delta^2} \leq \frac{4}{Q_{\min}}
\end{align}
where \eqref{eq: exponents' ratio} follows from the fact that its second
and third factors are less than or equal to~1 and~4, respectively. Note
that the bounds in \eqref{potalo} and \eqref{potaup} scale like $\delta^2$
for $\delta \approx 0$.
\subsection{Distance from Equiprobable}
\label{subs: equiprobable}
If $P$ is a distribution on a finite set $\ensuremath{\mathcal}{A}$, $H ( P )$ gauges the ``distance" from
$U$, the equiprobable distribution, since
\begin{align} \label{eq: RE-equiprobable}
H ( P ) = \log | \ensuremath{\mathcal}{A} | - D(P \| U).
\end{align}
Thus, it is of interest to explore the relationship between $H ( P ) $ and $|P-U|$.
Particularizing \eqref{eq: Pinsker}, \cite[(2.2)]{BretagnolleH79} (see also \cite[pp.~30--31]{Vapnik98}),
and \eqref{eq: UB-RE-FS2} we obtain
\begin{align}
\label{eq: 1st ubtv-uniform}
|P-U| &\leq \sqrt{ \frac{2}{\log e} \cdot \bigl( \log | \ensuremath{\mathcal}{A} | - H(P) \bigr) }, \\
\label{eq: 2nd ubtv-uniform}
|P-U| &\leq 2 \sqrt{1 - \frac1{|\ensuremath{\mathcal}{A}|} \cdot \exp\bigl( H(P) \bigr)}, \\
\label{eq: lbtv-uniform}
|P-U| &\geq \sqrt{ 2 \left( \exp\bigl( -H(P) \bigr) - \frac1{|\ensuremath{\mathcal}{A}|} \right) },
\end{align}
respectively.
\begin{figure}[here!]
\includegraphics[width=8.2cm]{equiprobable_alphabet_size_4-SV.pdf}
\includegraphics[width=8.2cm]{equiprobable_alphabet_size_16-SV.pdf}
\caption{\label{figure:equiprobable}
Bounds on $|P-U|$ as a function of $H(P)$ for $|\ensuremath{\mathcal}{A}| = 4$, and
$|\ensuremath{\mathcal}{A}| = 16$. The point $(H(P), |P-U|)=(0, 2(1-|\ensuremath{\mathcal}{A}|^{-1}) )$
is depicted on the $y$-axis. In the curves of the two plots, the bounds
(a), (b) and (c) refer, respectively, to \eqref{eq: 1st ubtv-uniform},
\eqref{eq: 2nd ubtv-uniform} and \eqref{eq: lbtv-uniform}.}
\end{figure}
The bounds in \eqref{eq: 1st ubtv-uniform}--\eqref{eq: lbtv-uniform} are illustrated
for $|\ensuremath{\mathcal}{A}| = 4, 16$ in Figure~\ref{figure:equiprobable}.
For $H(P)=0$, $|P - U| = 2(1-|\ensuremath{\mathcal}{A}|^{-1} )$ is shown for reference in Figure~\ref{figure:equiprobable};
as the cardinality of the alphabet increases,
the gap between $|P-U|$ and its upper bound is reduced (and this gap decays
asymptotically to zero).
\par
Results on the more general problem of finding bounds on $|H( P ) - H(Q)|$ based on $|P-Q|$
can be found in \cite[Theorem~17.3.3]{Cover_Thomas}, \cite{HoY_IT2010},
\cite{Prelov_PPI2008}, \cite{Sason_IT2013}, \cite[Section~1.7]{Verdu_book} and \cite{Zhang_IT2007}.
\section{Extension of Theorem~\ref{thm: reksnip - improved} to R\'{e}nyi Divergences}
\label{section: RD}
\begin{definition}
{\em The R\'{e}nyi divergence of order $\alpha \in [0, \infty]$ from $P$ to $Q$ is defined
for $\alpha \in (0,1) \cup (1, \infty)$ as
\begin{align}
& D_{\alpha}(P || Q) \triangleq \frac{1}{\alpha-1} \; \log \left( \sum_{a \in \mathcal{A}}
P^{\alpha}(a) \, Q^{1-\alpha}(a) \right).
\label{eq: Renyi divergence}
\end{align}
Recall that $D_{1}(P\|Q) \triangleq D(P\|Q)$ is defined to be the analytic extension
of $D_{\alpha}(P || Q)$ at $\alpha=1$ (if $D(P||Q) < \infty$, L'H\^{o}pital's rule
gives that $D(P||Q) = \lim_{\alpha \uparrow 1} D_{\alpha}(P || Q)$).
The extreme cases of $\alpha = 0, \infty$ are defined as follows:
\begin{itemize}
\item If $\alpha = 0$ then $D_0(P||Q) = -\log Q(\text{Support}(P))$,
\item If $\alpha = +\infty$ then
$$D_{\infty}(P||Q) = \log \left(\sup_{a \in \ensuremath{\mathcal}{A}} \frac{P(a)}{Q(a)}\right).$$
\end{itemize}}
\end{definition}
Pinsker's inequality was extended by Gilardoni \cite{Gilardoni10}
for a R\'{e}nyi divergence of order $\alpha \in (0,1]$
(see also \cite[Theorem~30]{ErvenH14}), and it gets the form
\begin{equation*}
\tfrac{\alpha}{2} \, |P-Q|^2 \, \log e \leq D_{\alpha}(P \| Q).
\end{equation*}
A tight lower bound on the R\'{e}nyi divergence of order
$\alpha > 0$ as a function of the total variation distance
is given in \cite{Sason_ISIT15}, which is consistent with
Vajda's tight lower bound for $f$-divergences in \cite[Theorem~3]{Vajda_1972}.
Motivated by these findings, we extend the upper bound on the relative entropy
in Theorem~\ref{thm: reksnip - improved} to R\'{e}nyi divergences of an arbitrary order.
\vspace*{0.2cm}
\begin{theorem}\label{thm: ub-RD-TV-FS}
{\em
Assume that $P, Q$ are strictly positive with minimum masses
denoted by $P_{\min}$ and $Q_{\min}$, respectively.
Let $\beta_1$ and $\beta_2$ be given in \eqref{eq: beta1} and \eqref{eq: beta2},
respectively, and abbreviate $\delta \triangleq \tfrac12 |P-Q| \in [0,1]$.
Then, the R\'{e}nyi divergence of order $\alpha \in [0,\infty]$ satisfies
\begin{align} \label{eq: ub-RD-TV-FS}
& D_{\alpha}(P \| Q) \nonumber \\
& \leq \left\{
\begin{array}{ll}
f_1, & \mbox{$\alpha \in (2, \infty]$} \\[0.3cm]
f_2, & \mbox{$\alpha \in [1,2]$} \\[0.3cm]
\min \left\{ f_2, f_3, f_4 \right\},
& \mbox{$\alpha \in \bigl(\tfrac12, 1 \bigr)$} \\[0.3cm]
\min \left\{ 2 \log\left(\frac1{1-\delta} \right), f_2, f_3, f_4 \right\},
& \mbox{$\alpha \in \bigl[0, \tfrac12\bigr]$}
\end{array}
\right.
\end{align}
where, for $\alpha \in [0, \infty]$,
\begin{align} \label{eq: f1}
& f_1(\alpha, \beta_1, \delta) \nonumber \\
& \triangleq \left\{
\begin{array}{ll}
\frac{1}{\alpha-1} \; \log\left(1 + \frac{\delta (\beta_1^{1-\alpha}-1)}{1-\beta_1}\right)
& \alpha \in [0,1) \cup (1, \infty) \\[0.3cm]
\frac{\delta}{1-\beta_1} \; \log \frac1{\beta_1}, & \alpha = 1, \\[0.3cm]
\log \frac1{\beta_1}, & \alpha = \infty \\[0.3cm]
\end{array}
\right.
\end{align}
for $\alpha \in [0,2]$
\begin{align}
\label{eq: f2}
& f_2(\alpha, \beta_1, Q_{\min}, \delta) \nonumber \\[0.2cm]
& \triangleq \min \left\{f_1(\alpha, \beta_1, \delta), \; \log\left(1+\frac{2\delta^2}{Q_{\min}}\right) \right\}
\end{align}
and, for $\alpha \in [0,1)$, $f_3$ and $f_4$ are given by
\begin{align}
\label{eq: f3}
& f_3(\alpha, P_{\min}, \beta_1, \delta) \nonumber \\[0.2cm]
& \triangleq
\left(\frac{\alpha}{1-\alpha}\right)
\left[ \log \left(1 + \frac{2 \delta^2}{P_{\min}} \right)
- 2 \beta_1 \delta^2 \, \log e \right], \\[0.3cm]
\label{eq: f4}
& f_4(\beta_2, Q_{\min}, \delta) \nonumber \\
& \triangleq \min\left\{ \log \left(1 + \frac{2 \delta^2}{Q_{\min}} \right)
- 2 \beta_2 \delta^2 \, \log e , \right.
\nonumber \\[0.1cm]
& \hspace*{1.2cm} \left. \log \left(1 + \frac{\min\{\delta, 2\delta^2\}}{Q_{\min}} \right) \right\}.
\end{align}}
\end{theorem}
\vspace*{0.3cm}
\begin{proof}
See \cite[Section~7.C]{ISSV15}.
\end{proof}
\begin{remark}
{\em A simple bound, albeit looser than the one in Theorem~\ref{thm: ub-RD-TV-FS} is
\begin{align} \label{eq: simple ub}
D_{\alpha}(P\|Q) \leq \log \left(1 + \frac{|P-Q|}{2 Q_{\min}} \right)
\end{align}
which is asymptotically tight as $\alpha \to \infty$ in the case of a binary alphabet
with equiprobable $Q$.}
\end{remark}
\begin{example}
{\em Figure~\ref{figure:compare_ub_RD_finite_graph1} illustrates the bound in
\eqref{eq: f1}, which is valid for all $\alpha \in [0, \infty]$ (see
\cite[Theorem~23]{ISSV15}), and the upper bounds of Theorem~\ref{thm: ub-RD-TV-FS}
in the case of binary alphabets.
\begin{figure}[here!]
\centerline{\includegraphics[width=9.8cm]{jagged.pdf}}
\caption{\label{figure:compare_ub_RD_finite_graph1}
The R\'{e}nyi divergence $D_{\alpha}(P\|Q)$ for $P$ and $Q$
which are defined on a binary alphabet with $P(0)=Q(1)=0.65$,
compared to (a) its upper bound in \eqref{eq: ub-RD-TV-FS},
and (b) its upper bound in \eqref{eq: f1} (see \cite[Theorem~23]{ISSV15}).
The two bounds coincide here when $\alpha \in (1, 1.291) \cup (2, \infty)$.}
\end{figure}}
\end{example}
\section{Summary}
We derive in this paper some ``reverse Pinsker inequalities" for probability measures
$P \ll Q$ defined on a common finite set, which provide lower bounds on the total
variation distance $P-Q$ as a function of the relative entropy $D(P\|Q)$ under the
assumption of a bounded relative information or $Q_{\min} > 0$. More general results
for an arbitrary alphabet are available in \cite[Section~5]{ISSV15}.
In \cite{ISSV15}, we study bounds among various $f$-divergences, dealing with arbitrary
alphabets and deriving bounds on the ratios of various distance measures.
New expressions of the R\'{e}nyi divergence in terms of the relative information
spectrum are derived, leading to upper and lower bounds on the R\'{e}nyi divergence
in terms of the variational distance.
\section*{Acknowledgment}
The work of I. Sason has been supported by the Israeli Science Foundation (ISF) under
Grant 12/12, and the work of S. Verd\'{u} has been supported by the US National Science
Foundation under Grant CCF-1016625, and in part by the Center for Science of Information,
an NSF Science and Technology Center under Grant CCF-0939370.
|
\section*{Acknowledgement}
The author S.M. is thankful to UGC for NET-JRF.
The author S.S. is thankful to UGC-BSR Programme of Jadavpur University for awarding Research Fellowship.
S.C. is thankful to UGC-DRS programme, Department of Mathematics, J.U.
|
\section{Motivations and main results}
The classical Euler gamma function may be defined for $\Re(z)>0$ by
\begin{equation*}
\Gamma(z)= \int_{0}^{\infty}t^{z-1}e^{-t} \td t.
\end{equation*}
The logarithmic derivative of the gamma function $\Gamma(z)$ is denoted by
\begin{equation*}
\psi(z)=\frac{\td}{\td z}[ \ln \Gamma(z)]=\frac{\Gamma'(z)}{\Gamma(z)}
\end{equation*}
and called the digamma function.
The derivatives $\psi'(z)$ and $\psi''(z)$ are called the trigamma and tetragamma functions respectively. As a whole, the functions $\psi^{(k)}(z)$ for $k\in\{0\}\cup\mathbb{N}$ are called the polygamma functions.
These functions are widely used in theoretical and practical problems in all branches of mathematical science. Consequently, many mathematicians were preoccupied to establish new results about the gamma function, polygamma functions, and other related functions.
\subsection{The first main result}
In 2007, Alzer and Batir~\cite[Corollary]{alz} discovered that the double inequality
\begin{equation}\label{alzer-batir-ineq-aml}
\sqrt{2\pi} \,x^x \exp \biggl[-x-\frac12\psi(x+\alpha)\biggr]< \Gamma(x) < \sqrt{2\pi} \,x^x \exp \biggl[-x-\frac12\psi(x+ \beta) \biggr]
\end{equation}
holds for $x>0$ if and only if $\alpha \ge\frac13$ and $\beta \le0$. For information on generalizations of results in the paper~\cite{alz}, please refer to~\cite{Convexity2CM.tex, CM-Tri-Gamma.tex} and references cited therein.
Motivated by the double inequality~\eqref{alzer-batir-ineq-aml}, Mortici~\cite{m1} proposed the asymptotic formula
\begin{equation*}
\Gamma(x) \sim \sqrt{2\pi}e^{-b}(x+b) ^{x} \exp \biggl[-x-\frac12\psi(x+c) \biggr], \quad x \to\infty
\end{equation*}
and then determined the optimal values of parameters $b$, $c$ in such a way that this convergence is the fastest possible.
\par
In 2004, Batir~\cite[Theorem 2.1]{batir0} presented the double inequality
\begin{multline} \label{bb}
\Gamma(c) \exp[\psi(x)e^{\psi(x)}-e^{\psi(x)}+1] \le \Gamma(x) \\*
\le \Gamma(c) \exp \biggl\{\frac{6e^{\gamma}}{\pi^2} \bigl[\psi(x) e^{\psi(x)}-e^{\psi(x)}+1 \bigr] \biggr\},
\end{multline}
for every $x \ge c$, where $c=1.461 \dotsc$ is the unique positive zero of the digamma function $\psi$ and $\gamma=0.577 \dotsc$ is the Euler-Mascheroni constant. In 2010, Mortici~\cite{m2} discovered the asymptotic formula
\begin{equation*}
\Gamma(x) \sim\frac{\sqrt{2\pi}}{e} \exp \bigl[\psi(x)e^{\psi(x)}-e^{\psi(x)}+1 \bigr], \quad x \to\infty.
\end{equation*}
In 2011, the double inequality~\eqref{bb} was generalized by~\cite[Theorem~2]{notes-best-simple-open-jkms.tex} to a monotonicity property which reads that the function
\begin{equation*}
f_{s,t}(x)=\begin{cases}\displaystyle
\frac{g_{s,t}(x)}{[g'_{s,t}(x)-1]\exp[g'_{s,t}(x)]+1},&x\ne c\\
\dfrac1{g''_{s,t}(c)},&x=c
\end{cases}
\end{equation*}
on $x\in(-\alpha,\infty)$ is decreasing for $\lvert t-s\rvert<1$ and increasing for $\lvert t-s\rvert>1$, where $s$ and $t$ are real numbers, $\alpha=\min\{s,t\}$, $c\in(-\alpha,\infty)$, and
\begin{equation*}
g_{s,t}(x)=\begin{cases}\displaystyle
\frac1{t-s} \int_c^x\ln\biggl[\frac{\Gamma(u+t)}{\Gamma(u+s)} \frac{\Gamma(c+s)}{\Gamma(c+t)}\biggr]\td u,&s\ne t\\
\displaystyle
\int_c^x[\psi(u+s)-\psi(c+s)]\td u,&s=t
\end{cases}
\end{equation*}
on $x\in(-\alpha,\infty)$.
\par
In 2000, Elezovi\'c, Giordano, and Pe\v{c}ari\'c~\cite{ele} found the single-sided inequality
\begin{equation} \label{ele}
\psi'(x)<e^{-\psi(x)}, \quad x>0.
\end{equation}
This inequality is closely related to the monotonicity and convexity of the function
\begin{equation*}
\mathcal{Q}(x)=e^{\psi(x+1)}-x
\end{equation*}
on $(-1,\infty)$. See also~\cite{MIA-1729.tex} and plenty of references therein.
By the way, as a conjecture posed in~\cite[Remark~3.6]{Infinite-family-Digamma.tex}, the complete monotonicity of the function $\mathcal{Q}(x)$ on $(0,\infty)$ still keeps open. An infinitely differentiable function $f$ is said to be completely monotonic on an interval $I$ if it satisfies $(-1)^kf^{(k)}(x)\ge0$ on $I$ for all $k\ge0$. This class of functions has applications in the approximation theory, asymptotic analysis, probability, integral transforms, and the like. See~\cite[Chapter~14]{Cheney-Light-AMS-09}, \cite[Chapter~XIII]{mpf-1993}, \cite[Chapter~1]{Schilling-Song-Vondracek-2nd}, and~\cite[Chapter~IV]{widder}.
For more information on the functions
\begin{equation*}
Q'(x-1)=\psi'(x)e^{\psi(x)}-1 \quad \text{and}\quad Q''(x-1)=\big\{\psi''(x)+[\psi'(x)]^2\bigr\}e^{\psi(x)},
\end{equation*}
please see the papers~\cite{Yang-Fan-2008-Dec-simp.tex, Infinite-family-Digamma.tex, Guo-Qi-Srivasta-Unique.tex, UPB-1635.tex, x-4-q-di-tri-gamma-rcsm.tex, deg4tri-tetra-short.tex, Mortici-monoburn.tex, BAustMS-5984-RV.tex, notes-best-simple-cpaa.tex, AAM-Qi-09-PolyGamma.tex, SCM-2012-0142.tex, notes-best.tex, Bukac-Sevli-Gamma.tex, x-4-di-tri-gamma-p(x)-Slovaca.tex, x-4-di-tri-gamma-upper-lower-combined.tex}, the expository and survey articles~\cite{bounds-two-gammas.tex, Gautschi-Kershaw-TJANT.tex, Wendel-Gautschi-type-ineq-Banach.tex, Wendel2Elezovic.tex-JIA}, and a number of references cited therein.
In 2011, Batir~\cite[Theorem~2.7]{batir} obtained the double inequality
\begin{equation} \label{ab}
(x+a^{\ast})e^{-2\psi(x+1)}<\psi'(x+1) \le(x+b^{\ast})e^{-2\psi(x+1)}, \quad x>0,
\end{equation}
where the constants $a^{\ast}=\frac12$ and $b^{\ast}=\frac{\pi^2}{6e^{2\gamma}}=0.518 \dotsc$ are the best possible.
In other words, Batir~\cite{batir} proposed the approximation formula
\begin{equation} \label{a1}
\psi'(x+1) \sim(x+a)e^{-2\psi(x+1)},
\end{equation}
where $a$ is a constant. A numerical computation shows that the approximation~\eqref{a1} gives a better result when choosing $a=a^{\ast}$, rather than the value $a=b^{\ast}$. This fact is somewhat expected as Batir obtained~\eqref{ab} as a result of the decreasing monotonicity of the function
\begin{equation*}
\theta(x)=\psi'(x+1)e^{2\psi(x+1)}-x
\end{equation*}
on $(0,\infty)$, with $\theta(0)=\frac{\pi^2}{6e^{2\gamma}}$ and $\theta(\infty)=\frac12$.
Hence, the function $\theta(x)$ becomes gradually closer to
$\theta(\infty)$, as $x$ approaches infinity.
The decreasing monotonicity of the function $\theta(x)$ may be rewritten and extended as that the function
\begin{equation*
\Theta(x)=\psi'(x)e^{2\psi(x)}-x+\frac12
\end{equation*}
is decreasing and positive, but not convex or concave, on $(0,\infty)$. In 2010, among other things, Qi and Guo~\cite[Corollary~2]{AAM-Qi-09-PolyGamma.tex} proved that the function
\begin{equation*
f_{p,q}(x)=e^{p\psi(x+1)}-q x
\end{equation*}
for $p \ne0$ and $q \in \mathbb{R}$ is strictly convex with respect to $x \in(-1,\infty)$ if and only if $p \ge1$ or $p<0$.
Since
\begin{equation}\label{mortici-20-eq}
\frac{\td e^{\alpha\psi(x)}}{\td x}= \alpha\psi'(x)e^{\alpha\psi(x)} \quad \text{and} \quad
\frac{\td^2e^{\alpha\psi(x)}}{\td x^2}
= \alpha \bigl\{\psi''(x)+ \alpha[\psi'(x)]^2 \bigr\}e^{\alpha\psi(x)},
\end{equation}
the above functions between~\eqref{ele} and~\eqref{mortici-20-eq} have something to do with the more general function
\begin{equation*}
f_{a,b,\alpha,\beta,\lambda}(x)=e^{a\psi(x+b)}+\alpha x^2+\beta x+\lambda
\end{equation*}
and its monotonicity and convexity, even its complete monotonicity.
\par
For obtaining accurate approximations of the type~\eqref{a1}, the value $a=\frac12$ should be used. However, approximations of the form
\begin{equation*}
\psi'(x+1) \sim[x+a(x)]e^{-2\psi(x+1)},
\end{equation*}
with $a(x) \to\frac12$ as $x$ tends to $\infty$, are much better. Furthermore, several experiments using some asymptotic expansions lead us to the claim that approximations of the type
\begin{equation*}
\psi'(x+1) \sim[x+a(x)] \exp \biggl[-2\psi(x+1)-\frac1{120x^4} \biggr]
\end{equation*}
are more accurate. Consequently, we obtain the following theorem.
\begin{theorem}\label{000-mortici-psy-thm1}
For $x \ge3$, the double inequality
\begin{multline} \label{aba}
[x+ \alpha(x)] \exp \biggl[-2\psi(x+1)-\frac1{120x^4} \biggr] \le\psi'(x+1) \\
\le[x+ \beta(x)] \exp \biggl[-2\psi(x+1)-\frac1{120x^4} \biggr]
\end{multline}
is valid, where
\begin{equation*}
\alpha(x)=\frac12+\frac1{90x^3}-\frac1{60x^4} \quad \text{and} \quad
\beta(x)=\frac12+\frac1{90x^3}.
\end{equation*}
\end{theorem}
\subsection{The second main result}
In 2013, Guo and Qi~\cite[Lemma 2]{Yang-Fan-2008-Dec-simp.tex} proved that
\begin{equation}\label{e-1-t-1}
\psi'(x)<e^{1/x}-1, \quad x>0.
\end{equation}
This inequality has been generalized to the complete monotonicity of the function
\begin{equation}\label{exp-trigamma-dif}
e^{1/x}-\psi'(x)-1, \quad x>0
\end{equation}
and others. For detailed information, please refer to~\cite{property-psi.tex, property-psi-ii-Munich.tex, Bessel-ineq-Dgree-CM-Simp.tex}, \cite[Theorem~3.1]{QiBerg.tex}, \cite[Theorem~1.1]{simp-exp-degree-revised.tex}, \cite[Theorem~1.1]{simp-exp-degree-new.tex}, and closely related references therein.
In some of these references, it was obtained that
\begin{equation*}
\lim_{x \to\infty} \bigl[e^{1/x}-\psi'(x) \bigr]=1.
\end{equation*}
The approximation $\psi'(x) \sim e^{1/x}-1$ indeed gives a good result for the large value of $x$.
Now we propose the improvement
\begin{equation*}
\psi'(x) \sim e^{1/x+ \mu(x)}-1,
\end{equation*}
where $\mu(x) \to0$ and $\mu(x)=o(x)$ as $x \to\infty$. Our main result may be stated in details as the following theorem.
\begin{theorem}\label{000-mortici-psy-thm2}
For $x \ge3$, we have
\begin{equation}\label{000-mortici-psy-ineq}
e^{m(x)}-1<\psi'(x)<e^{M(x)}-1,
\end{equation}
where
\begin{equation*}
m(x)=\frac1x-\frac1{24x^4}+\frac{7}{360x^6} \quad \text{and} \quad M(x)=m(x) +\frac1{90x^7}.
\end{equation*}
\end{theorem}
\subsection{The third main result}
In 2014, Yang, Chu, and Tao found~\cite[Theorem~1]{Yang-Chu-Tao-AAA-14} that the constants $p=1$ and $q=2$ are the best possible real parameters such that the double inequality
\begin{equation} \label{teta}
\theta(x,p)<\psi'(x+1)< \theta(x,q)
\end{equation}
holds for $x>0$, where
\begin{equation*}
\theta(x,m)=\frac{e^{m/(x+1)}-e^{-m/x}}{2m}.
\end{equation*}
Because
\begin{equation}\label{psi'(x+1)recur}
\psi'(x+1)=\psi'(x)-\frac1{x^2},
\end{equation}
the double inequality~\eqref{teta} may be reformulated as
\begin{equation}\label{theta-x-sq-eq}
\frac1{x^2}+\theta(x,1)<\psi'(x)<\frac1{x^2}+\theta(x,2), \quad x>0.
\end{equation}
The complete monotonicity of the function~\eqref{exp-trigamma-dif} implies that
\begin{equation*
e-\psi'(1)> e^{1/(x+1)}-\psi'(x+1)>1, \quad x>0
\end{equation*}
which may be rearranged as
\begin{equation}\label{x+1-eq}
e^{1/(x+1)}-e+\psi'(1)<\psi'(x+1)<e^{1/(x+1)}-1<\frac12\sinh\frac2{x}
\end{equation}
on $(0,\infty)$. When $t<1.6\dotsc$, the lower bound $e^{1/(x+1)}-e+\psi'(1)$ is better than the corresponding one in~\cite[Corollary~2]{Yang-Chu-Tao-AAA-14}. The upper bound $e^{1/(x+1)}-1$ in~\eqref{x+1-eq} and the upper bound $\theta(x,2)$ in~\eqref{teta} can not be compared with each other. However, the upper bound $\frac1{x^2}+\theta(x,2)$ in~\eqref{theta-x-sq-eq} is better than the upper bound $e^{1/x}-1$ in~\eqref{e-1-t-1}.
\par
In this paper, we will improve the double inequality~\eqref{teta} as follows.
\begin{theorem}\label{000-mortici-psy-thm3}
For $x\ge1$, we have
\begin{equation}\label{z1}
\theta(x,1) +\frac1{24x^5}-\frac{5}{48x^6}<\psi'(x+1)< \theta(x,1) +\frac1{24x^5}
\end{equation}
and
\begin{equation}\label{z2}
\theta(x,2) -\frac1{45x^7}<\psi'(x+1)
< \theta(x,2) -\frac1{45x^7}+\frac{7}{90x^8}.
\end{equation}
\end{theorem}
\section{Proofs of Theorems~\ref{000-mortici-psy-thm1} to~\ref{000-mortici-psy-thm3}}
Now we start out to prove our three main results.
\par
In 1997, Alzer~\cite[Theorem 8]{zzz0} proved that for $m,n \ge1$ the functions
\begin{equation*
F_{m}(x)= \ln \Gamma(x+1) -\biggl(x+\frac12\biggr) \ln x+x-\frac12 \ln2\pi- \sum_{i=1}^{2m}\frac{B_{2i}}{2i(2i-1) x^{2i-1}}
\end{equation*}
and
\begin{equation*
G_{n}(x)=- \ln \Gamma(x+1) +\biggl(x+\frac12\biggr) \ln x-x+\frac12 \ln2\pi+ \sum_{i=1}^{2n-1}\frac{B_{2i}}{2i(2i-1) x^{2i-1}}
\end{equation*}
are completely monotonic on $(0,\infty)$, where $B_j$ are the Bernoulli numbers given by the generating function
\begin{equation*}
\frac{t}{e^{t}-1}= \sum_{j=0}^{\infty}B_j\frac{t^{j}}{j!}
=1-\frac{t}2+\sum_{j=0}^{\infty}B_{2j}\frac{t^{2j}}{(2j)!}.
\end{equation*}
See also~\cite[Theorem~2]{Koumandos-jmaa-06}, \cite[Theorem~2.1]{Koumandos-Pedersen-09-JMAA}, and~\cite[Theorem~3.1]{Mortici-AA-10-134}. From $F_1'(x)<0$, $G_2'(x)<0$, $F_1''(x)>0$, $F_2''(x)>0$, $G_2''(x)>0$, and $G_3''(x)>0$, it follows that
\begin{equation}\label{22}
\ln x+\frac1{2x}-\frac1{12x^2}+\frac1{240x^4}-\frac1{252x^6}<\psi(x+1)
<\ln x+\frac1{2x}-\frac1{12x^2}+\frac{1}{240x^4},
\end{equation}
\begin{equation}\label{psi'(x+1)-l-power5}
\psi'(x+1) >\frac1x-\frac1{2x^2}+\frac1{6x^3}-\frac1{30x^5},
\end{equation}
\begin{multline}\label{33}
\frac1x-\frac1{2x^2}+\frac1{6x^3}-\frac1{30x^5}+\frac1{42x^7}-\frac1{30x^9}<\psi'(x+1)\\
<\frac1x-\frac1{2x^2}+\frac1{6x^3}-\frac1{30x^5}+\frac1{42x^7},
\end{multline}
and, also by~\eqref{psi'(x+1)recur},
\begin{multline}\label{power11-9-ineq}
\frac1x+\frac1{2x^2}+\frac1{6x^3}-\frac1{30x^5}+\frac1{42x^7}-\frac1{30x^9}<\psi'(x) \\
<\frac1x+\frac1{2x^2}+\frac1{6x^3}-\frac1{30x^5} +\frac1{42x^7}-\frac1{30x^9}+\frac{5}{66x^{11}}.
\end{multline}
\begin{proof}[Proof of Theorem~\ref{000-mortici-psy-thm1}]
By taking the logarithm, the double inequality~\eqref{aba} may be rearranged as
\begin{equation*}
F(x)= \ln[x+ \alpha(x)]-2\psi(x+1) - \ln\psi'(x+1) -\frac1{120x^4}<0
\end{equation*}
and
\begin{equation*}
G(x)=- \ln[x+ \beta(x)] + \ln\psi'(x+1) +2\psi(x+1) +\frac1{120x^4}<0
\end{equation*}
for $x \ge3$. By virtue of inequalities~\eqref{22} and~\eqref{33}, one may deduce that $F(x)<F_1(x)$ and $G(x)<G_1(x)$, where
\begin{align*}
F_1(x) &= \ln[x+ \alpha(x)] -2\biggl(\ln x+\frac1{2x}-\frac1{12x^2}+\frac1{240x^4} -\frac1{252x^6}\biggr) \\
&\quad - \ln\biggl(\frac1x-\frac1{2x^2}+\frac1{6x^3}-\frac1{30x^5} +\frac1{42x^7}-\frac1{30x^9}\biggr) -\frac1{120x^4}
\end{align*}
and
\begin{align*}
G_1(x) &=- \ln[x+ \beta(x)] + \ln\biggl(\frac1x-\frac1{2x^2} +\frac1{6x^3}-\frac1{30x^5}+\frac1{42x^7}\biggr) \\
&\quad +2\biggl( \ln x+\frac1{2x}-\frac1{12x^2}+\frac1{240x^4}\biggr)
+\frac1{120x^4}.
\end{align*}
Since
\begin{equation*}
F_1'(x)=\frac{A(x-3)}{105x^7(180x^5+90x^4+2x-3) B(x-3)}
\end{equation*}
and
\begin{equation*}
G_1'(x)=\frac{C(x-3)}{15x^5(90x^4+45x^3+1) D(x-3)},
\end{equation*}
where
\begin{align*}
A(x)&=84000 x^{13}+3753750 x^{12}+75109720 x^{11}+896904120 x^{10}\\
&\quad+7160223140 x^9+40457327085 x^8+166700796732 x^7+507517074474 x^6\\
&\quad+1141703970759 x^5+1873185114060 x^4+2175691772642 x^3\\
&\quad+1690072075536 x^2+783944661553 x+162974708124,\\
B(x)&=210 x^8+4935 x^7+50750 x^6+298305 x^5+1096193 x^4\\
&\quad+2578821 x^3+3792857 x^2+3188649 x+1173161,\\
C(x)&=23625 x^9+604200 x^8+6818475 x^7+44510545 x^6\\
&\quad+184933335 x^5+506070905 x^4+909421185 x^3\\
&\quad+1030441127 x^2+663679092 x+183168418,\\
D(x)&=210 x^6+3675 x^5+26810 x^4+104370 x^3+228683 x^2+267393 x+130352
\end{align*}
are polynomials with positive coefficients, the functions $F_1(x)$ and $G_1(x)$ are strictly increasing on $[3,\infty)$. Furthermore, since
\begin{equation*}
\lim_{x \to\infty}F_1(x)= \lim_{x \to\infty}G_1(x)=0,
\end{equation*}
it follows that $F_1(x)<0$ and $G_1(x)<0$ for $x\ge3$. As a result, we have $F(x)<F_1(x)<0$ and $G(x)<G_1(x)<0$. The proof of Theorem~\ref{000-mortici-psy-thm2} is complete.
\end{proof}
\begin{proof}[Proof of Theorem~\ref{000-mortici-psy-thm2}]
The inequality~\eqref{000-mortici-psy-ineq} can be written as
\begin{equation*}
m(x)< \ln[1+\psi'(x)]<M(x).
\end{equation*}
Considering the double inequality~\eqref{power11-9-ineq}, since $m(x) - \ln[1+\psi'(x)]<m_1(x)$ and $M(x) - \ln[1+\psi'(x)] >M_1(x)$, one may see that it suffices to show that
\begin{equation*}
m_1(x)\triangleq m(x) - \ln\biggl(1+\frac1x+\frac1{2x^2}+\frac1{6x^3} -\frac1{30x^5}+\frac1{42x^7}-\frac1{30x^9}\biggr)<0
\end{equation*}
and
\begin{equation*}
M_1(x)\triangleq M(x) - \ln\biggl(1+\frac1x+\frac1{2x^2}+\frac1{6x^3}-\frac1{30x^5} +\frac1{42x^7}-\frac1{30x^9}+\frac{5}{66x^{11}}\biggr)>0.
\end{equation*}
A straightforward computation gives
\begin{equation*}
m_1'(x)=\frac{E(x-3)}{60x^7G(x-3)}\quad \text{and}\quad M_1'(x)=-\frac{F(x-3)}{180x^8H(x-3)},
\end{equation*}
where
\begin{align*}
E(x)&=980 x^8+22485 x^7+221130 x^6+1212855 x^5+4032099 x^4\\
&\quad+8229303 x^3+9865371 x^2+6074127 x+1290163,\\
F(x)&=66495 x^{10}+2146155 x^9+31007240 x^8+263913573 x^7\\
&\quad+1464790565 x^6+5537745108 x^5+14437981040 x^4\\
&\quad+25626153678 x^3+29624987873 x^2+20135221233 x+6106987838,\\
G(x)&=210 x^9+5880 x^8+73185 x^7+531440 x^6+2481255 x^5+7724423 x^4\\
&\quad+16033731 x^3+21398207 x^2+16660569 x+5765861,\\
H(x)&=2310 x^{11}+78540 x^{10}+1213905 x^9+11258170 x^8+69614160 x^7\\
&\quad+301344043 x^6+931827204 x^5+2058324400 x^4+3182887290 x^3\\
&\quad+3281444518 x^2+2029943157 x+570820414
\end{align*}
are polynomials with positive coefficients. This means that $m_1'(x)>0$ and $M_1'(x)<0$ for $x\ge3$, that is, the function $m_1(x)$ is strictly increasing and the function $M_1(x)$ is strictly decreasing on $[3,\infty)$, with the limits
\begin{equation*}
\lim_{x \to\infty}m_1(x)= \lim_{x \to\infty}M_1(x)=0.
\end{equation*}
Accordingly, one obtain that $m_1(x)<0$ and $M_1(x)>0$ on $[3,\infty)$. The proof of Theorem~\ref{000-mortici-psy-thm2} is complete.
\end{proof}
\begin{proof}[Proof of Theorem~\ref{000-mortici-psy-thm3}]
The left hand side inequality in~\eqref{z1} can be written as
\begin{equation*}
c(x)\triangleq\frac1{2}e^{1/(x+1)}-\frac1{2}e^{-1/x}+\frac1{24x^5}-\frac{5}{48x^6}-\psi'(x+1)<0.
\end{equation*}
Since
\begin{equation*}
e^{-1/x}>\sum_{k=0}^7\frac{1}{k!}\biggl(-\frac1{x}\biggr)^k
\end{equation*}
and the inequality~\eqref{psi'(x+1)-l-power5} is valid, we have
\begin{equation*}
2c(x)<e^{1/(x+1)}-\frac{P(x)}{5040x^7},
\end{equation*}
where
\begin{equation*}
P(x)=5040 x^7+5040 x^6-2520 x^5+840 x^4+210 x^3-798 x^2+1057 x-1.
\end{equation*}
Hence, it is sufficient to prove that
\begin{equation*}
c_1(x)=\frac1{x+1}- \ln\frac{P(x)}{5040x^7}<0.
\end{equation*}
A straightforward differentiation yields
\begin{equation*}
c_1'(x)=\frac{7630 x^2+6329 x-7}{x(x+1)^2P(x)}=\frac{7630(x-1)^2+21589(x-1)+13952}{x(x+1)^2Q(x-1)},
\end{equation*}
where
\begin{align*}
Q(x)&=5040 x^7+40320 x^6+133560 x^5+240240 x^4\\
&\quad+255570 x^3+161112 x^2+56371 x+8868.
\end{align*}
This implies that the function $c_1(x)$ is strictly increasing on $[1,\infty)$. Furthermore, because of $\lim_{x \to\infty}c_1(x)=0$, it follows that $c_1(x)<0$ on $[1,\infty)$. The proof of the left hand side inequality in~\eqref{z1} is complete.
\par
Similarly, we may verify other inequalities in~\eqref{z1} and~\eqref{z2}. For the sake of saving the space and shortening the length of this paper, we do not repeat the processes. The proof of Theorem~\ref{000-mortici-psy-thm3} is complete.
\end{proof}
\section{Remarks}
Finally we give several remarks on our three main results.
\begin{remark}
The right hand side in~\eqref{aba} is better than the right hand side in~\eqref{ab}, but the left hand side in~\eqref{aba} is not better than the left hand side in~\eqref{ab}. This is because the inequalities
\begin{equation*}
(x+a^{\ast})e^{-2\psi(x+1)}>[x+ \alpha(x)] \exp \biggl[-2\psi(x+1)-\frac1{120x^4} \biggr]
\end{equation*}
and
\begin{equation*}
[x+ \beta(x)] \exp \biggl[-2\psi(x+1)-\frac1{120x^4} \biggr]<(x+b^{\ast})e^{-2\psi(x+1)},
\end{equation*}
which are equivalent to
\begin{equation}\label{01}
u(x)= \ln \biggl(x+\frac12+\frac1{90x^3}-\frac1{60x^4} \biggr) - \ln \biggl(x+\frac12 \biggr)-\frac1{120x^4}<0
\end{equation}
and
\begin{equation}\label{02}
v(x)= \ln \biggl(x+\frac12+\frac1{90x^3} \biggr)
-\frac1{120x^4}- \ln \biggl(x+\frac{\pi^2}{6e^{2\gamma}} \biggr)<0,
\end{equation}
are valid for $x \ge1$. The inequalities~\eqref{01} and~\eqref{02} may be verified as follows. A direct differentiation gives
\begin{equation*}
u'(x)=\frac{540 x^6+60 x^5+180 x^4+90x \bigl(x-\frac1{30} \bigr)^2+\frac{9}{10}x+2} {30x^5(2x+1)S\bigl(x-\frac1{10}\bigr)}
>0
\end{equation*}
for $x \ge\frac1{10}$ and
\begin{equation*}
v'(x)=\frac{Q(x-1)}{30x^5(90x^4+45x^3+1) (6e^{2\gamma}x+\pi^2)}>0
\end{equation*}
for $x \ge1$, where
\begin{equation*}
S(x)=180x^4+162x^3 +\frac{189}{5}x^2 +\frac{21}{50}x+\frac{226}{125}
\end{equation*}
and
\begin{align*}
Q(x)&=2700 \bigl(\pi^2-3e^{2\gamma} \bigr)x^8+21600 \bigl(\pi^2-3e^{2\gamma} \bigr)x^7+75600 \bigl(\pi^2-3e^{2\gamma} \bigr)x^6 \\
& \quad+180 \bigl(840\pi^2-2521e^{2\gamma} \bigr)x^5+630 \bigl(300\pi^2-901e^{2\gamma} \bigr)x^4 \\
& \quad+45 \bigl(3361\pi^2-10096e^{2\gamma} \bigr)x^3+45 \bigl(1683\pi^2-5044e^{2\gamma} \bigr)x^2 \\
& \quad+3 \bigl(7245\pi^2-21538e^{2\gamma} \bigr)x+2 \bigl(1373\pi^2-4002e^{2\gamma} \bigr)
\end{align*}
are polynomials with positive coefficients. This means that the functions $u(x)$ and $v(x)$ are strictly increasing on $[1,\infty)$. Further, from the limits
\begin{equation*}
\lim_{x \to\infty}u(x)= \lim_{x \to\infty}v(x)=0,
\end{equation*}
it follows that $u(x)<0$ on $(0,\infty)$ and $v(x)<0$ on $[1,\infty)$.
\end{remark}
\begin{remark}[The asymptotic series of $\psi'(x+1)e^{2\psi(x+1)}$]
Whenever an approximation formula $f(x) \sim g(x)$ is considered in the sense that the ratio $\lim_{x \to\infty}\frac{f(x)}{g(x)}=1$ (sometimes $\lim_{x\to\infty}[f(x) -g(x)]=0$), there is a tendency to improve it by adding new terms, or an entire series, of the form
\begin{equation*}
f(x) \sim g(x) + \sum_{k=1}^{\infty}\frac{a_k
}{x^k}.
\end{equation*}
Such a series is called an asymptotic series and it plays a central role in the theory of approximation. Although such a series is often divergent, by truncating at the $m$-th term, it provides approximations of any desired accuracy $\frac1{x^{m+1}}$. In other words, the formula
\begin{equation*}
f(x)\sim g(x) + \sum_{k=1}^{m}\frac{a_k}{x^k
}+O\biggl(\frac1{x^{m+1}}\biggr), \quad x\to\infty
\end{equation*}
is valid for every integer $m \ge1$.
\par
It is well known~\cite{ab} that the digamma and trigamma functions admit respectively the following asymptotic expansions:
\begin{equation}\label{0}
\psi(x+1)\sim \ln x+\frac1{2x}- \sum_{k=2}^{\infty}\frac{B_k}{kx^k}\quad \text{and}\quad
\psi'(x+1)\sim -\frac1{x^2}+ \sum_{k=1}^{\infty}\frac{B_{k-1}}{x^k}.
\end{equation}
The asymptotic expansions in~\eqref{0} can be written explicitly as
\begin{equation*}
\psi(x+1)\sim\ln x+\frac1{2x}-\frac1{12x^2}+\frac{1}{240x^4}-\frac1{252x^6} +\frac1{240x^8}-\frac1{132x^{10}}+ \dotsm
\end{equation*}
and
\begin{equation*}
\psi'(x+1)\sim\frac1x-\frac1{2x^2}+\frac{1}{6x^3}-\frac1{30x^5}+\frac1{42x^7} -\frac1{30x^9}+\frac{5}{66x^{11}}-\dotsm.
\end{equation*}
\par
In order to construct the asymptotic expansion of the function $\psi'(x+1)e^{2\psi(x+1)}$, we recall from the papers~\cite{Chen-Elezovic-Vuksic-JCA, Gould-AMS-1974} and the monographs~\cite[pp.~20--21]{Erdelyi-1956} and~\cite[pp.~539\nobreakdash--541]{Knopp} the following classical results from the theory of asymptotic series:
\begin{enumerate}
\item
if
\begin{equation*}
u(x)\sim \sum_{k=0}^{\infty}\frac{p_k}{x^k} \quad \text{and}\quad
v(x)\sim \sum_{k=0}^{\infty}\frac{q_k}{x^k}
\end{equation*}
as $x\to\infty$, then
\begin{equation}\label{mortici-prop-2}
u(x)v(x)\sim \sum_{k=0}^{\infty}\frac{r_k}{x^k},
\end{equation}
where
\begin{equation*}
r_k= \sum_{i+j=k}p_{i}q_j;
\end{equation*}
\item
if
\begin{equation*}
f(x)\sim\sum_{k=0}^{\infty}\frac{a_k}{x^k}, \quad x\to\infty,
\end{equation*}
then
\begin{equation}\label{mortici-prop-1}
\exp f(x)\sim\sum_{k=0}^{\infty}\frac{\alpha_k}{x^k}, \quad x \to\infty,
\end{equation}
where $\alpha_{0}= \exp a_{0}$ and
\begin{equation*}
\alpha_k= \sum_{j=1}^k\frac1{j!} \sum_{i_1+ \dotsm+i_j=k}\prod_{\ell=1}^ja_{i_\ell}
\end{equation*}
for $k \ge1$.
\end{enumerate}
\par
Now the asymptotic series of the function $\psi'(x+1)e^{2\psi(x+1)}$ can be computed in two steps. Firstly, by virtue of the formula~\eqref{mortici-prop-1}, we may transform the series~\eqref{0} to obtain the series of $e^{2\psi(x+1)}$. Secondly, with the help of the formula~\eqref{mortici-prop-2}, we multiply the series of $e^{2\psi(x+1)}$ and the second series in~\eqref{0}.
\par
Because the general term (in terms of the Bernoulli numbers) of the series of the function $\psi'(x+1)e^{2\psi(x+1)}$ has an unattractive form, we just write down the first few terms as follows:
\begin{equation}\label{sss}
\psi'(x+1)e^{2\psi(x+1)} =x+\frac{1}{2}+\frac1{90x^3}-\frac1{60x^4}+\frac{2}{567x^5} +\frac{43}{2268x^6}-O\biggl(\frac1{x^7}\biggr).
\end{equation}
Till now we can see immediately that the functions $\alpha(x)$ and $\beta(x)$ in Theorem~\ref{000-mortici-psy-thm1} are truncations of the series~\eqref{sss} at the first three or four terms.
\par
When more terms of~\eqref{sss} are considered, the conclusion of Theorem~\ref{000-mortici-psy-thm1} remains true and the estimates become more accurate.
\end{remark}
\begin{remark}
We conjecture that the function $e^{M(x)}-\psi'(x)-1$ is completely monotonic on $(0,\infty)$, but the function $\psi'(x)-e^{m(x)}+1$ is not.
\end{remark}
\begin{remark}
The reason why the double inequality~\eqref{teta} is of interest is because it provides much accurate estimates for the trigamma function $\psi'(x)$ for large values of $x$, that is,
\begin{equation}\label{mm}
\psi'(x+1) \sim \theta(x,m), \quad x \to\infty.
\end{equation}
From~\eqref{z1} and~\eqref{z2}, it is easily deduced that
\begin{equation*}
\psi'(x+1) - \theta(x,1)=O\biggl(\frac1{x^5}\biggr)\quad \text{and}\quad \psi'(x+1) - \theta(x,2)=O\biggl(\frac1{x^7}\biggr),
\end{equation*}
then it is easy to see that the best approximation of the form~\eqref{mm} may be obtained by taking $m=2$.
\end{remark}
\begin{remark}
Evidently, our inequalities~\eqref{z1} and~\eqref{z2} are much stronger than those in~\cite{Yang-Chu-Tao-AAA-14}. Our proofs for~\eqref{z1} and~\eqref{z2} are also much simpler than the original proof of~\cite[Theorem~1]{Yang-Chu-Tao-AAA-14}. This shows that the natural approach to solve problems of approximating some quantities for large values of the variable is the theory of asymptotic series. This method or approach has been utilized and applied in the papers~\cite{m8, m6, m7, m3, m4, m5}, for examples.
\end{remark}
\begin{remark}
In this paper, we essentially talk about the relations between inequalities, asymptotic approximations, and complete monotonicity.
\end{remark}
\subsection*{Acknowledgements}
The first author was in part supported by the NNSF of China under Grant No.~11361038.
The second author was partially supported by a Grant of the Romanian National Authority for Scientific Research, CNCS-UEFISCDI, with the Project Number PN-II-ID-PCE-2011-3-0087.
|
\section{Introduction}
In a server system, sharing resources among virtual machines (VMs) implies that the resources of the server are accessed simultaneously among different users. Users' concerns about physical device sharing, stem from the fact that their data resides in these shared devices, and virtualization technology has to provide the necessary mechanisms to ensure the security of user data. Protection of user data at different levels of architecture like CPU, memory and Input/Output (I/O) devices has to be provided, proved and assured to convince the users of the credibility of the system. The survey reports \cite{jithin2014virtual} \cite{pearce2013virtualization} \cite{rehman2013virtual} on the security of VMs, show that
\begin{itemize}
\item VMs should be as isolated as physical machines, i.e., the only means of communication between them should be exactly as between two different physical machines (Isolation Property of VMs)
\item Hypervisors cannot be trusted, due to the possibility of hypervisor based malware, and other infections possible on hypervisors
\end{itemize}
An isolated VM environment can be provided by the hypervisor, either through hardware modifications or through additional software modules that restrain a VM from accessing other VMs' data at various architectural levels \cite{jithin2014virtual}. Software modules to impose these restrictions would take additional CPU clock cycles and thereby reduce VM performance, implying that software solutions could provide security to VMs only at the cost of compromising VM performance. A hardware modification could impose these restraints with a lesser performance degradation, as it would use either a marginal number of additional clock cycles, or simply a negligible increase in the pipeline cycle time. Hence, we aim at designing enhancements to hardware that can achieve security for VMs without compromising on their performance, and also provide security in the presence of compromised hypervisors.
This paper reviews literature in the area of memory virtualization to identify open challenges that prevent hypervisors from providing a true isolated environment for VMs without losing out on performance. Some open challenges have been identified and a memory architecture model aimed at solving those challenges is proposed, which unlike existing memory models like nested paging and IOMMU, does not degrade performance. ASMI (Architectural Support for Memory Isolation), our proposed model, requires modification to hardware and is illustrated with the help of hardware currently in popular use for supporting VMs.
Memory virtualization techniques used in present day systems have been reviewed in the Section \ref{memory}, along with the open challenges in memory virtualization. Necessary features in any technology designed for improving VMs in terms of security and performance are identified in Section \ref{solution}, along with a description of the design of ASMI. Section \ref{conclusion} concludes the article with future directions of research.
\section{Memory Virtualization} \label{memory}
The normal paging mechanism cannot be used \cite{Smith2006} in a virtualized environment. In a normal computing environment, the paging hardware is accessed by the single operating system to get the physical address corresponding to the virtual address. In a virtual environment, a single physical memory is shared among different VMs or guest operating systems (hereafter referred as guest OS, in this paper).
If a guest OS accesses the paging hardware to translate its virtual address, there are possibilities that a guest OS may access a physical address which is previously accessed by another guest OS. As this is a serious security threat to the isolation property of VMs, the security of VMs at memory level, has to be assured by isolating the physical pages of each VM from others. A mechanism to differentiate the physical address of each guest OS is a mandatory requirement.
A memory virtualization mechanism that is popularly used in VM technology nowadays to achieve differentiation of logical addresses is called \textit{Nested Paging \cite{Smith2006}}. The following section describes nested paging.
\subsection{Nested Paging}
In nested paging \cite{Smith2006}, the virtual address of the process running in each VM is converted to a \textit{pseudo physical address} by the corresponding guest OS. This translation is achieved by looking up \emph{page tables}, which are maintained by the guest OS \cite{Smith2006}. In the next level of translation, the pseudo physical address stored in page tables of each guest OS is translated to the actual physical address by the hypervisor. This translation is stored in the \textit{real map table} maintained by the hypervisor \cite{Smith2006}. The hypervisor maintains a separate real map table for each guest OS. These two page tables, in combination, are known as a \textit{nested page table}. Thus, in a virtual environment, there are three layers of memory \cite{Smith2006}. They are \emph{Physical Memory} (Original device memory, visible to hypervisor), \emph{Pseudo Physical Memory} (Virtual view of physical memory to VMs, visible to guest OS) and \emph{Virtual Memory} (Logical memory for each
program, visible to process)
On contemporary platforms, page translation is supported by a combination of page table and a translation lookaside buffer (TLB). Currently there are two different types of hardware configurations available for address translation. One is an architectured page table. The second is an architectured TLB \cite{Smith2006}. Virtualizing these hardware configurations is a primary requirement for memory virtualization.
Each guest OS maintains its own page tables. These page tables represent the virtual to pseudo physical mapping that guest OS manages. To virtualize the architectured page table, a \textit{shadow page table} \cite{Smith2006} is used by the hypervisor, which contains the virtual address and the corresponding physical address. Each VM should have a shadow page table. When a context switch (among VMs) occurs, the shadow page table in use by the hypervisor changes according to the currently active VM \cite{Smith2006}. Shadow page tables are updated along with the nested page tables.
In an architectured TLB, TLB stores the recently occurred virtual to physical address translations. But in a virtual environment these entries will be different for each guest OS. So, when the context switch between VMs happens, the TLB has to be flushed to avoid the guest OS accessing the translations of another guest OS. TLB flush on each context switch is computationally expensive \cite{Smith2006} \cite{tai2013comparisons}. Moreover, failure to hit the TLB causes many extra pipeline cycles.
To overcome the problem, designers added an extra field named Address Space Identifier (ASID) to the TLB. The ASID field is used to distinguish the address of currently running process in a normal environment \cite{Smith2006}. The ASID field entry shows the owner (process) of the address translation entry in the TLB. In a virtualized system, a \textit{virtual TLB} is maintained by the hypervisor, which contains the virtual ASID field, virtual page field and the real page field. An \textit{ASID map table} is maintained by the hypervisor to map the \textit{virtual ASID}, of each process in each VM, to a unique \textit{real ASID} value. Only the address translations of the currently running process in the active guest OS, are allowed to access from the TLB (distinguished by the real ASID field) \cite{Smith2006}.
Nested page tables along with either virtualized versions of architectured page tables or with virtualized versions of architectured TLBs, create a virtual address space in VMs. However, there are issues with this addressing mechanism, from the perspective of performance and security. The nested paging technique compromises security while enabling Direct Memory Access (DMA) mechanism \cite{Smith2006} \cite{DarrenAbramson2006} .
The DMA mechanism is designed to work with the actual physical address space. If the DMA mechanism is enabled, the guest OS can reconfigure the device to access the memory of another VM through DMA mechanism \cite{szefer2012architectural}. A solution to this security threat, is to disable the DMA mechanism. Disabling DMA reduces performance because more CPU clock cycles would be required for data transfer between memory and I/O devices. Protecting memory by access from other VMs while enabling DMA mechanism in I/O devices, is called \textit{DMA isolation}. DMA isolation has to be achieved to enable the DMA mechanism and thereby improve the performance of VMs.
The requirements for better performance without compromising security led to the development of another address translation mechanism named I/O Memory Management Unit (IOMMU) \cite{ben2006utilizing} which is explained next.
\subsection{IOMMU}
The IOMMU architecture is illustrated here with the help of an Intel based technology named Intel VT-d \cite{vt-directed-io-spec}. Intel introduced a DMA remapping hardware in their chip set. A generalized IOMMU architecture is implemented inside this DMA remapping hardware. The process of converting the DMA address from one form to another (virtual address to physical address) is known as DMA remapping. The hardware for DMA remapping is called DMA remapping hardware.
Intel VT-d architecture divides the physical address space into different partitions. Each partition is a subset of the entire physical memory. Each partition is considered as a protection domain \cite{vt-directed-io-spec} \cite{DarrenAbramson2006}. I/O devices are allocated to a single protection domain by the hypervisor. I/O devices are not allowed to access any domain other than the one it is allocated. VMs are also allocated to a protection domain by the hypervisor. A VM is allowed to access only the I/O devices that are allocated to its own domain. VT-d enables hypervisor to allocate one or more I/O devices to a protection domain.
DMA isolation is achieved by restricting the access to a protection domain by the I/O devices not assigned to that domain. DMA isolation is implemented through two address translation tables used in this architecture \cite{vt-directed-io-spec} \cite{DarrenAbramson2006}. They are
\begin{itemize}
\item Root Entry Table (RET)
\item Context Entry Table (CET)
\end{itemize}
VT-d hardware treats the address in a DMA request as a \emph{DMA Virtual Address(DVA)} \cite{vt-directed-io-spec} \cite{DarrenAbramson2006}. DVA can be a guest physical address (pseudo physical address) of the VM to which the I/O device is assigned. Thus, I/O devices that are assigned to a protection domain can be provided a view of memory that is different from the host physical memory. VT-d transforms the DVA address to the host physical address with the help of RET and CET.
Each DMA address has three fields \cite{vt-directed-io-spec} \cite{DarrenAbramson2006}. They are \emph{Bus number}, \emph{Device number} and \emph{Function number}. The bus number field content is used to index into the RET. Each entry in the RET point to a CET. The device number and function number field contents are collectively used to index into the CET. Each entry in CET points to an address translation structure which is a multilevel page table similar to a IA-32 processor page table named \emph{hierarchical translation structure} \cite{vt-directed-io-spec}.
Since VMs are in different protection domains, I/O devices alloted to a VMs protection domain would not be able to access the memory of another VM. Thus, in VT-d architecture, the DMA mechanism can be enabled without compromising the security, thereby improving the performance of I/O devices, and consequently the performance of VMs.
It is observed from the two memory management techniques, nested paging and IOMMU, that nested paging would have the worse performance of the two, due to the inability to safely use DMA. IOMMU architecture allows the use of DMA and could achieve DMA isolation, as the I/O devices and VMs are permitted to access only one protection domain. Memory translation is a two level lookup and is hence slower than a normal environment.
Though the IOMMU architecture provides better DMA isolation than nested paging, it still stays vulnerable to many security threats. In order to illustrate the problem, we discuss the security threats and proposed solution in literature, in the succeeding paragraphs.
\subsection{Security Threats}
A survey on the security of VMs shows that there exist vulnerabilities at different architecture levels like CPU, memory and network, which would help a malicious VM to easily gain the control of the hypervisor \cite{jithin2014virtual}. In the IOMMU architecture, the hypervisor has access to all the memory locations including the entire allocated memory space for machines. This would result in the situation that a malicious VM infects a hypervisor or an infected hypervisor could attack or access the memory of another VM. Providing security in the presence of infected hypervisor is considered an open research problem in the area of the security of VMs \cite{szefer2012architectural}. For improved security in the context of infected hypervisors, solutions that move the protection of memory from hypervisor level to the hardware level are desirable \cite{szefer2012architectural}. The proposed work in \cite{szefer2012architectural}, called HyperWall, is
briefly discussed.
\subsubsection{HyperWall}
\textit{HyperWall} \cite{szefer2012architectural} is an architectural solution aimed at providing hardware for protecting guest VMs from a malicious hypervisors. The HyperWall hardware is proposed as an extension to the IOMMU hardware unit. The authors claim that the key feature of HyperWall is \emph{Confidentiality} and \emph{Integrity } protection to VMs. HyperWall architecture provides additional protection bits to each memory page without modifying the paging structure in the hypervisor or guest OS. Four protection bits are used by HyperWall associated with each physical page. These protection bits can represent four new protection modes to the physical page. They are \emph{Not assigned to any VM} (Hypervisor access only), \emph{Assigned to a VM with hypervisor and DMA allowed}, \emph{Assigned to a VM with hypervisor access denied} and \emph{Assigned to a VM with neither hypervisor nor DMA allowed.} The pages assigned to a VM are protected from other VMs and the hypervisor by applying these user
specification protection modes to each page.
According to our analysis, there exist security issues with this architectural solution. They are
\begin{itemize}
\item HyperWall architecture aims to protect only the confidentiality and integrity of a VMs data and not \textit{availability}. However, availability is also an important security property. The architecture does not provide the assurance of a minimum and fair amount of memory to all VMs at all point of their active lifetime. It allows the VM user to set the memory page with a protection level which denies access to DMA and hypervisor. Hence, any malicious VM can use this feature to create memory starvation for other VMs by locking several pages at a time.
\item HyperWall does not protect against covert channels and side channels attacks \cite{szefer2012architectural},which are the main threat towards the isolation property of VMs \cite{jithin2014virtual}. Most of the covert channel attacks are done through memory \cite{Xiao:2012:CCC:2382196.2382318}, \cite{barham2003xen} \cite{wu2011identification} \cite{Xu:2011:ELC:2046660.2046670}. Without preventing those attacks at the memory level, true isolation in the memory level cannot be achieved.
\end{itemize}
Apart from security threats, there are performance challenges to be met by virtualization mechanisms. In the next two paragraphs, we state the major reasons for memory slowdown that result in under-performance, as identified in our survey.
\subsection{Underutilization of Memory}
Our survey on the challenges in memory architecture shows that a huge volume of unutilized memory is allocated to VMs \cite{agmon2014ginseng} \cite{hwang2014mortar}. It is because, in the current VM environment, the requested physical memory is divided and allocated to VMs when they are created. That memory would not be used by the assigned VMs if they do not need that memory. But, it cannot be assigned to other VMs that require memory. A fairer allocation of physical memory is required.
\textit{Ginseng} \cite{agmon2014ginseng} is a solution to the problem of underutilization of VM memory. Ginseng runs in the hypervisor, and allocates physical memory to a guest OS via a balloon driver. The balloon driver is installed in each guest OS. A balloon controller is installed in hypervisor. The communication between balloon driver in guest OS and balloon controller in hypervisor is done through \textit{libvirt} \cite{agmon2014ginseng}, an application programming interface (API) for managing the hypervisor.
There are two types of communication modes between the balloon driver and balloon controller, \textit{Inflation} and \textit{Deflation}. In inflation, the balloon driver transfers memory from guest OS to the hypervisor. The hypervisor keeps such memory from each guest together as a memory pool, available to guest OS when its assigned memory is over-utilized, through deflation. By inflation and deflation, Ginseng solves the problem of memory starvation effectively.
\textit{Mortar} \cite{hwang2014mortar} is another technique which helps to utilize this underutilized memory as a cache. There are two uses for this cache. The first one is to use it as a cache for pre-fetching disk blocks. The next is to use it as a an application level distributed cache that follows the \textit{memcached} \cite{wang2012streaming} protocol. Rather than allocating the underutilized memory to guest OS, this architecture pools together the spare memory on each guest OS and exposes it as a volatile cache. Whenever the guest OS require its memory, the memory cache objects are freed and given back to the guest OS. This architecture also efficiently uses the underutilized memory.
The main difference between Ginseng and Mortar is that Ginseng pools the unallocated memory and gives it to other VMs, while Mortar pools the unallocated memory and uses it as an application level cache.
Our analysis is that both Ginseng and Mortar are susceptible to covert channel based attacks by colluding VMs on the same server, thereby risking user program and data on the VMs.
\subsection{Memory Access Times}
Another major challenge in VM technology is to improve the access time of memory for processes. Both nested paging technique and IOMMU architecture use two level memory address translation. In nested paging architecture as well as in IOMMU architecture, address translation hardware is virtualized in such a way that the address translation hardware is directly accessible to hypervisor and not to guest OS \cite{Smith2006} \cite{DarrenAbramson2006}. HyperWall architecture does not offer any improvement in VM performance over IOMMU or nested paging either.
Existing literature also includes illustrations of challenges and solutions in deduplication \cite{chiang2013introspection} \cite{chen2014cmd} and double-paging \cite{arya2014tesseract} features of VMs. Deduplication and double-paging are implemented in hypervisors, by sharing the memory pages among VMs and hypervisor. Those features are considered as a threat to memory isolation \cite{jithin2014virtual}, and are hence not advisable for use in clouds.
Enabling the access of address translation hardware directly by the guest OS removes one level of indirection in address translation process. This would improve the performance of VMs by improving the average memory access time. Clearly the need of the day is to have a memory virtualization infrastructure that supports the following properties even in the presence of a malicious hypervisors.
\begin{itemize}
\item Isolated physical memory region to each VM
\item Guest OS should be able to use the address translation hardware directly
\item Fair allocation of physical memory among VMs
\end{itemize}
\section{Architectural Support for Memory Isolation} \label{solution}
The analysis presented in the previous section clearly establishes that a novel mechanism has to be introduced, such that it would be able to protect the memory of a VM from a malicious hypervisor, address the problem of underutilization of memory in VMs, and allow the user programs to use the address translation hardware directly, for improved VM performance. We propose \textit{ASMI, Architectural Support for Memory Isolation}, that can satisfy the above requirements. Figure \ref{proposal} illustrates the proposed architecture.
\begin{figure*}[ht!]
\centering
\includegraphics[width=100mm]{Proposal}
\caption{ASMI: Architectural Support for Memory Isolation}
\label{proposal}
\end{figure*}
We propose ASMI as a generic solution, aimed at providing hardware for enabling direct access to physical memory by the VMs. We illustrate and expand the concept on an Intel platform with Intel-VT capabilities \cite{vt-directed-io-spec}, as details about the existing technology are available in literature. Hence, hereafter, in this paper, ASMI refers to the design enhancement we describe in the succeeding paragraphs.
\paragraph{Description}In an Intel 64 bit architecture, the address translation mechanism disables segmentation and only paging mechanism exists \cite{64-ia-32-arch-manual-325462}. In ASMI, segmentation is enabled and utilized to improve isolation. The physical memory is partitioned into physical segments. Each segment should be of a fixed length. Each segment contain fixed number of pages. A new hardware unit named \textit{Pro-mem} controls the entire physical memory.
A new register named VMIDR is introduced and its contents managed by each processor to store the identity of the currently running VM on that processor. A unique ID is assigned and stored in VMIDR register by the Pro-mem unit when a new VM is created. Switch between VMs on a processor causes a changes in the VMIDR value, done by the Pro-mem. Moving the control of execution from VM to hypervisor and vice versa would be informed to the Pro-mem by the \emph{VM Exit} and \emph{VM Entry} instructions.
\emph{VM Entry} and \emph{VM Exit} are the instructions in Intel VT architecture \cite{vt-directed-io-spec} that move the execution to and fro between VMs and hypervisor. In the proposed architecture, these instructions are modified to inform the Pro-mem unit about the control transfer among VMs and hypervisor in an atomic manner.
\emph{SegMax} is introduced, which is maintained by Pro-mem and contains the maximum number of segments(MSEG) that can be assigned to a VM when the entire physical memory is full. The value stored in MSEG is calculated by dividing the total number of physical segments (TSEG) in primary memory with TOT, where \emph{TOT} is the sum of total number of VMs and hypervisor running above the hardware.
SegMax, MSEG, and TOT values are not fixed. They are changed when a new VM is created or when an old one is destroyed. TSEG is fixed at boot time, by the hypervisor, depending on the page size that the hypervisor is designed to work with. TSEG values cannot be changed until next reboot. Initially at boot time, MSEG and TOT values are zero. When a hypervisor is loaded or a VM is created, TOT value is incremented by one and MSEG is recalculated according to the new TOT value.
A new data structure named \textit{Memory Protection Table (MPT)} is also proposed, and it is managed by the Pro-mem. MPT contains the segment ID (SegID) and its corresponding virtual Machine ID (VMID). Each segment can be assigned to a single VM. MPT would be stored in the primary memory as shown in Figure \ref{proposal}. This primary memory portion of MPT will be inaccessible to any software module. This table is accessible only to the Pro-mem unit. A VMID value zero in MPT is used to indicate the segments that belongs to the hypervisor.
Initially, when the physical machine boots, the MPT will be empty. When the hypervisor is started by the BIOS, Pro-mem stores a unique ID to the VMIDR register. Similarly when a VM is created, a unique ID is stored in VMIDR register by Pro-mem.
When the \textit{VM exit} instruction is executed, VMIDR contents would get stored in the initial address of the first memory segment of the corresponding VM and the VMIDR is loaded with the initial address of the first memory segment of the hypervisor. Similarly when the \textit{VM entry} instruction executes, the VMIDR value is stored in the initial address of the first segment of hypervisor and the VMIDR value is loaded from the initial address of the first segment of the running VM. These load and store operations are executed by Pro-mem as atomic operations to \textit{VM exit} and \textit{VM entry} instructions.
\paragraph{Allocation}
When a VM requests for a memory page, Pro-mem would check for available free pages in the allotted segments. If a free page is available in the allotted segment, it would be assigned to the VM. If no free pages are available, Pro-mem would check for a free segment. If a free segment available, it is allotted to the current VM, the MPT updated and the page assigned to the VM.
When no free segments are available, Pro-mem would check whether any of the VMs has been allotted more than MSEG number of segments. The number of segments by which it exceeds MSEG would be informed to the corresponding VM by the Pro-mem, to make it free through swapping. Then those pages would free be used by the requesting VM. If no VM is using more than MSEG segments, Pro-mem will send a memory full exception to the requesting VM. This exception can be used by the guest OS to swap its allotted memory pages and load the new one, and thereby continue its execution.
The above techniques ensure a minimum number of physical segments or a minimum amount of physical memory to each VM when the physical memory need is the maximum. It simultaneously ensures that physical memory is not be left free or unutilized when a VM requires it. Hence, an optimum utilization of physical memory can be achieved by this architecture.
\paragraph{Access}
Access to segments would be validated by Pro-mem with the help of VMIDR value and MPT table. Pro-mem would raise an exception if a VM or hypervisor tries to access the segment which is not allotted to it. The main advantage of this architecture is that it can provide memory isolation among VMs and hypervisor irrespective of the hypervisor security.
In this version, we propose Pro-mem to be installed in between hardware paging architecture and primary memory. The paging unit would be modified to include Pro-mem functionality. Each guest OS maintains a page table in memory, which contains the virtual address and its corresponding physical address. The guest OS gives the virtual/linear address to the paging unit. The paging unit would give this linear address to the Pro-mem unit. Pro-mem would return the physical address, within the guest OS alloted segment, to the paging unit. The paging unit would return this physical address to the guest OS.
\paragraph{Performance and Security}
The physical address obtained in guest OS, by the address translation described above, would be the actual physical address within the allotted physical segment. Only a single level address translation is required in this architecture. It would improve the performance of VMs, as it improves over two level translation. DMA can be enabled because the memory allotted to a VM can not be accessed by the other VMs as they are separated by different segments. Separating the physical memory of each VMs at the hardware level could provide hypervisor-independent security to VMs.
Operating systems in a normal environment use the paging hardware to translate linear addresses to physical addresses. In a virtual environment, the guest operating system uses the paging hardware with Pro-mem to translate its linear address to the physical address, and the physical address range is controlled by the Pro-mem. Thus, the guest operating system directly uses the address translation hardware (Paging unit with Pro-mem) similar to the normal (not virtualized) environment. Hence, no modification to guest OS is required. This enables the use of native operating systems, and a performance much better than existing virtualized systems.
\section{Conclusions and Future Work} \label{conclusion}
A memory architecture model named ASMI, has been proposed and described in the paper. ASMI provides an isolated memory region to each VM and the hypervisor. ASMI has been illustrated in this paper on an Intel Platform with hardware enhancements to implement the design. ASMI is designed to provide memory isolation to VMs, irrespective of the integrity of hypervisor.
Our architecture includes an address translation unit named Pro-mem. Pro-mem is shared among VMs in time division multiplexing without compromising security, to improve address translation performance. The design of Pro-mem solves the problem of underutilization of memory, as Pro-mem allocates memory to VMs only on individual requests and does not pre-allocate. ASMI can thus provide confidentiality, integrity and availability to VMs at memory level irrespective of hypervisor security without compromising performance.
Implementing ASMI through kernel level simulations and comparison of the security and performance of VMs with the currently available memory architectures like IOMMU and nested paging are planned as future tasks in our research. ASMI is a generic solution. Studying the suitability of ASMI on other architectures like MIPS, and other RISC variants, is also included in our research agenda.
\bibliographystyle{IEEEtran}
|
\section{Introduction}
At asymptotically high temperature $T$, QCD-like theories are in a weakly
coupled state known as the Quark-Gluon Plasma (QGP). In it semi-classical solitons -- instantons and their
constituents, monopoles etc -- have large action $S=O( 1/\alpha_s)\gg 1$. Their semi-classical
treatment is parametrically reliable, but their density is exponentially suppressed by $e^{-S}$. As a
result their effects are small.
However, as the temperature decreases the semi-classical action $S$ decreases. Since the soliton
density grows as a power of $1/T$ their contribution to the QCD partition increases.
At a critical density fixed by $T_c$, confinement sets in, and the near-zero expectation value
of the Polyakov line $\left<L\right>\approx 0$ switches off the quark component of the QGP, as well as the
(non-diagonal) gluons. Below the critical temperature $T_c$, the solitons dominate the field ensemble.
The major questions at the transition point are: (i) Are these objects still made of strong enough fields,
allowing for a semi-classical analysis; (ii) Are their interactions weak enough to preserve their
individual identity; (iii) Are the semi-classical interactions in the thermal ensemble amenable
to known methods of many-body theory. As we will argue below, two first questions
will be answered in the affirmative, and the third also, provided the ensemble is dense enough.
The instanton liquid model developed in the 1980$^\prime$s is an example of such a semi-classical treatment.
In vacuum at $T=0$, the action per typical SU(3) instanton was found to be large with $S\sim 12$,
and the inter-instanton and anti-instanton interactions tractable. The non-perturbative
vacuum topological fluctuations are related to the explicit violation of the axial U(1), and
the formation of fermionic zero modes. The collectivization of the fermionic zero modes leads to
the spontaneous breaking of flavor chiral symmetry~\cite{ALL} (and references therein). More recently,
instanton-induced effects were found to be important
for hadronic spin physics~\cite{SPIN}.
However, around the critical temperature $T\sim T_c$,
instantons should know about the non-vanishing of the Polyakov
line expectation value, also referred to as a non-trivial holonomy.
Instantons with non-trivial holonomies were found in~\cite{KVLL}.
The key discovery was that large holonomies split
instantons into $N_c$ constituents, the selfdual instanton-dyons.
Since these objects have nonzero Euclidean electric and magnetic charges and source
Abelian (diagonal) massless gluons, the corresponding ensemble is
an ``instanton dyon plasma"
We remark that the electric dyon field is real in Euclidean space-time but
imaginary in Minkowski space. The instanton-dyons are also referred to as
instanton-monopoles or instanton-quarks. However, the notion of a non-zero holonomy and all the instanton-related
contructions do not exist outside of the Euclidean finite-$T$ formulation. On the lattice, both the electric and magnetic charges
of the instanton-dyons are observable by standard Gaussian surface integrals.
Diakonov and Petrov~\cite{DP}
emphasized that, unlike the (topologically protected) instantons, the dyons interact directly with
the holonomy field. They
further suggested that since such dyon (anti-dyon) fields become significant
at low temperature, they may be at the origin of a vanishing of the mean Polyakov line, or confinement.
This mechanism is similar to the Berezinsky-Kosterlitz-Thouless-like transition
of instantons into fractional ``instanton quarks" suggested earlier by Zhitnitsky and others~\cite{ZHITNITSKY}, inspired by
the fractionalization of the topological charge in 2-dimensional CPN models~\cite{FATEEV}, although it
is substantially different in details.
It is also different from the random dyon-anti-dyon ensemble suggested earlier by Simonov
and others~\cite{SIMONOV}. It is not yet clear how this Euclidean mechanism relates to the
the quantum condensation of magnetic monopoles suggested initially by t$^\prime$ Hooft~\cite{THOOFT}
and Mandelstam~\cite{MAND}, and subsequently supported in the supersymmetric model discussed by
Seiberg and Witten~\cite{SEIBERG}. In many ways, it is similar to the 3-dimensional monopole plasma
discussed by Polyakov~\cite{POLYAKOV}.
Unsal and Yaffe~\cite{UNSAL1} , using a double-trace
deformation of Yang-Mills at large $N$ on $S^1\times R^3$, argued that it prevents the spontaneous
breaking of center symmetry. A similar trace deformation was used in the context of two-dimensional (confining) QED with unequal
charges on $S^1\times R$~\cite{HOLGER} to analyze the nature of center symmetry and its spontaneous breaking.
This construction was extended to QCD with adjoint fermions by Unsal~\cite{UNSALALL},
and by Unsal and others~\cite{UNSAL} to a class of deformed supersymmetric theories with soft supersymmetry breaking.
While the setting includes a compactification on a small circle, with weak coupling and
an exponentially $small$ density of dyons, the minimum at the confining holonomy
value is induced by the repulsive interaction in the dyon-anti-dyon pairs (called
$bions$ by the authors). The supersymmetry is needed to calculate the contribution of the dyon-anti-dyon pairs ,
and, even more importantly, for the
cancellation of the perturbative Gross-Pisarski-Yaffe-Weiss (GPYW) holonomy potential \cite{WEISS}.
Shuryak and Sulejmanpasic \cite{TIN}
have argued that induced by the ``repulsive cores" in dyon-antidyon channel also generates confinement,
explaining it in a simple model. The first numerical study of the
classical interaction of the dyons with anti-dyons has been recently carried in~\cite{LARSEN-SHURYAK}.
The streamline configurations were found by a gradient flow method, and their action assessed.
This classical interaction will be included -- for the first time -- in our paper.
Another major non-pertubarive phenomenon in QCD-like theories is spontaneous chiral symmetry breaking.
Shuryak and Sulejmanpasic~\cite{SHURYAK} have analyzed a number of phenomena induced by
the fermionic zero modes of the instanton-dyons such as the formation of clusters
(molecules or bions) at high temperature and their collectivization, generating spontaneous breaking of chiral symmetry
at low temperature. Faccioli and Shuryak~\cite{FACCIOLI} have started numerical simulations of the
dyon ensemble with light fermions to understand the nature of the fermionic collectivization. We will
provide an analytical analysis of these effects in the second paper of this series.
Before we get into the details of the various approximations to our analysis,
let us try to provide some qualitative answers
to the three generic questions formulated above: (i) At $T\sim T_c$, we will consider the action per dyon
(anti-dyon) to be still large or $S\sim 4$ whatever $N_c$;
(ii) The dyon interactions will be of the order of
$\Delta S_{\rm int}\sim 1 \ll S$. The quantum (one-loop) interactions are several times smaller and naively can be considered small.
However they are quite non-trivial and the repulsion they provide would be our key finding.
(iii) In general, the dyon plasma is strongly coupled and it is hard to treat it analytically.
However we will argue below that in some window of temperatures (below $T_c$)
one can still use the Debye-Huckel plasma theory.
A major contribution to the understanding of the one-loop dyon interaction has been made by Diakonov and others~\cite{DP,DPX}.
They have found that at $T>T_c$ their interaction with the surrounding QGP leads to a linear (confining) potential
between the dyons, proportional to the perturbative Debye mass. Since in this work we will only
consider the opposite case $T<T_c$, this will not be included in what follows. Key to the one-loop effect is
the explicit quantum weight of the KvBLL instantons in terms of the collective coordinates of the
constitutive dyons at all separations. The self-dual sector is characterized by a moduli space with a hyper-Kahler metric.
Its volume element is given by the determinant of Coulomb-like matrix. We will refer to it as Diakonov determinant.
In his first attempts to treat the dyonic plasma, Diakonov kept only the one loop determinant,
the volume of the moduli space, ignoring the
QGP screening effects and -- as we will discuss in detail -- the even larger classical
dyon-anti-dyon interaction. Furthermore,
he assumed that the attractive and repulsive terms
induced by the determinant cancel out on average. We disagree on this conclusion as we detail below.
Indeed, Bruckmann and others in~\cite{LATTICE}
tried to generate configurations of randomly placed dyons
using the determinantal measure, and observed that for the physically
relevant dyonic densities, the determinantal measure develops negative eigenvalues. This
makes no sense if the measure is to account for the volume of the dyonic moduli space.
We will show that this issue may become resolved in a strongly correlated ensemble.
It is well known that the separate treatment of self dual and anti-selfdual sectors is only justified
in the context of supersymmetry where self-duality is dual to holomorphy.
In QCD-like theories, the interaction between self dual and anti-selfdual sectors is strong
and not factorizable. It is described semi-classically by a ``streamline" with a
classical inter-particle potential of order $1/ \alpha_s$, which is larger than the
1-loop quantum induced potential of order $\alpha_s^0$. Furthermore, configurations with too strongly
overlapping objects with small action, are not subject to the semiclassical treatment. To account for that
one usually relies on the use of a ``repulsive core" as in the instanton liquid model for instance.
As we will discuss in detail, classical dyon-antidyon interaction
\cite{LARSEN-SHURYAK}
is about an order of magnitude stronger than the one-loop Coulomb effects. It
generically leads to the dyon plasma in the strongly coupled regime, with $e^{-V_{D\bar D}}\gg 1$.
We will however focus on the very dense regime of such plasma, in which
screening is strong enough that statistical mechanics of the ensemble can be treated by a variant of the
Debye-Huckel mean field plasma theory.
In such case the screening length is short enough
to fence the system from strong coupling correlations and
molecular-type instabilities induced by the streamline.
The more dilute systems such as those appearing at $T>T_c$,
will not be discussed in this work, as they need more powerful many-body methods, such as e.g.
strongly coupled
Coulomb plasmas many-body physics re-summations~\cite{DH,CHO} (and references therein).
As we will show, in this case the free energy has a minimum at the ``confining" holonomy value
$v=\pi T$.
In this paper we will detail the strongly coupled nature of the dyonic plasma. Our original results
consist of (i) introducing the strong correlations between dyons and anti-dyons as described
by the streamline~\cite{LARSEN-SHURYAK}; (ii) showing that the determinantal interactions induced
by the moduli space for dyons or anti-dyons are mostly repulsive causing the moduli volume to vanish
for randomly distributed dyons; (iii) showing that suitably organized dyons to account for screening
correlations yield finite moduli volumes; (iv) deriving an explicit 3 dimensional effective action that
account exactly for the screening of dyons and anti-dyons on the moduli space with strong
inter-dyon-anti-dyon streamline interactions; (v) showing explicitly that the strongly coupled dyonic
plasma is center symmetric and thus confining; (vi) deriving the Debye-Huckel corrections induced
by the dyons and anti-dyons to the leading Pressure for the dyonic plasma and using it
to asses the critical temperature for the SU(2) plasma; (vi) providing the explicit
results for the gluon topological susceptibility and compressibility near the critical temperature in the
center symmetric phase; (vii) deriving the scalar and charged structure factors of the dyonic plasma
showing explicit screening of both electric and magnetic charges at large distances
with explicit predictions for the electric and magnetic masses; (viii) showing that
the strongly coupled dyonic plasma supports both electric and magnetic confinement.
This paper is organized as follows:
In section 2 we review the key elements of the dyon and anti-dyon measure derived
in~\cite{DP,DPX} using the KvBLL instanton. The dyon-anti-dyon measure is then composed of the
product of two measures with streamline interactions between the dyons and anti-dyons.
We briefly detail the exact re-writing of the 3-dimensional grand-partition function in terms of
a 3-dimensional effective theory in the SU(2) case. We also show that the ground state of this
effective theory is center symmetric. In sections 3-6 we show that in the linearized screening
approximation the dyon-anti-dyon liquid still screens both electric and magnetic charges, generates a
linearly rising potential between heavy charges and confines the large spatial Wilson loops.
The t$^\prime$ Hooft loop in the dyon-anti-dyon ensemble is shown to be 1 modulo
${\cal O}(\alpha_s)$ self-energy corrections which are perimeter-like in section 7.
Our conclusions are in section 8.
\section{Interacting Dyon-Anti-Dyon Ensemble}
\subsection{The setting}
The first step is the introduction of the nonzero expectation value of the 4-th component
of the gauge field, which is gauge invariant since at finite temperature it enters the holonomy
integral over the time period, known also as the Polyakov line. Working in a gauge in which
$\left<A_4\right>$ belongs to the diagonal and traceless sub-algebra of $N_c-1$ elements, one
observes the standard Higgsing via the adjoint field. All gluons except the diagonal ones
become massive.
We will work with the simplest case of two color gauge theory $N_c=2$, in which
there is only one diagonal matrix and the
VEV of the gauge field (holonomy) is normalized as follows
\begin{eqnarray}
\left<A^3_4\right> =v{\tau^3\over 2}= 2\pi T \nu {\tau^3\over 2}
\end{eqnarray}
where $ {\tau^3/2}$ is the only diagonal color generator of SU(2).
At high $T$ it is trivial with $\nu\rightarrow 0$, and at
low $T<T_c$ it takes the confining value $\nu=1/2$.
With this definition, the only dimensional quantity in the classical approximation is the temperature $T$,
while the quantum effects add to the running coupling and its $\Lambda$ parameter. Since we are
working near and below $T_c$, we will follow the lattice practice and we use the latter as our main unit.
In the semi-classical approximation, the Yang-Mills partition function is assumed to be dominated by an interacting ensemble of
dyons (anti-dyons)~\cite{DP,DPX}. For large separations or a very dilute ensemble, the semi-classical interactions are mostly Coulombic,
and are encoded in the collective or moduli space of the ensemble. For multi-dyons a plausible moduli space was argued starting
from the KvBLL caloron~\cite{KVLL} that has a number of pertinent symmetries, among which permutation symmetry, overall charge neutrality,
and clustering to KvBLL at high temperature. Since the underlying calorons are self-dual, the induced metric on the moduli space was shown to be
hyper-Kahler.
The
SU(2) KvBLL instanton (anti-instanton) is composed of a pair of dyons labeled by $L,M$ (anti-dyons by $\overline L,\overline M$)
in the notations of~\cite{DP}. Specifically $M$ carries $(+,+)$ and $L$ carries $(-,-)$ for (electric-magnetic) charges, with
fractional topological charges $v_m=\nu$ and $v_l=1-\nu$ respectively. Their corresponding actions are $S_L=2\pi v_m/\alpha_s$ and $S_M=2\pi v_l/\alpha_s$.
The statistical measure for a correlated ensemble of dyons and anti-dyons is
\be
d\mu_{D\overline D}[K]\equiv &&\,e^{-V_{D\overline D}(x-y)}\\
&&\times\prod_{m=1}^N\prod_{i=1}^{K_m}\,\frac {f\,d^3{x_{mi}}}{K_m !}\,{\rm det}(G_{mi}[x])\nonumber
\\&&\times \prod_{n=1}^{ N}\prod_{j=1}^{{\overline K}_n}\,\frac {f\,d^3{y_{nj}}}{{\overline K}_n !}\,{\rm det}(G_{nj}[y])\,
\nonumber
\label{MEASUREDD}
\ee
The streamline interactions induced by the potential $V_{D\bar D}$ correlate
the two otherwise statistically independent dyon and anti-dyon sectors.
(Note that by the potential we mean the extra action and not the energy, thus no extra $1/T$).
Asymptotically,
\begin{eqnarray}
V_{D\overline D}(x-y)\rightarrow \sum_{mn,ij}\,\frac {C_D/2}{\alpha_s\,T} \,\frac{Q_{mi}{\overline Q}_{nj}}{|x_{mi}-y_{nj}|}
\label{VDD}
\end{eqnarray}
is a Coulomb-like classical interaction between dyons and anti-dyons.
Here $x_{mi}$ and $y_{nj}$ are the 3-dimensional coordinate of the i-dyon of m-kind
and j-anti-dyon of n-kind.
At shorter separations the streamline stops at certain distance $a_{D\bar D}$,
we will refer to it as a ``core size". While the interaction is more complex than just electric Coulomb, it is proportional to
the electric charges $Q,\overline{Q}$. In general those are the (Cartan) roots of $SU(N_c)$ supplemented by
the affine root. They satisfy
\begin{eqnarray}
Q_{mi}{\overline Q}_{nj}\equiv -\left(2\delta_{mn}-\delta_{m,n+1}-\delta_{m,n-1}\right)
\label{QQ}
\end{eqnarray}
The dimensionality of $G[x]$ is
$(K_1+...+K_N)^2$ and similarly for $G[y]$. Their explicit form can be found in~\cite{DP,DPX}.
In the SU(2) case there is only one electric charge.
The semiclassical 3-density of all dyon species $n_D\equiv n_L+n_M+n_{\bar L}+n_{\bar M}$ is
\begin{eqnarray}
n_D ={dN \over d^3x} = \frac {C T^3 \,e^{-\frac{\pi}{\alpha_s}}}{\alpha_s^2}
\label{NDDD}
\end{eqnarray}
where $C$ is a constant to be determined below (see (\ref{NDX})). (\ref{NDDD})
can be re-written using the asymptotic freedom formula for SU(2) pure gauge theory
with $2\pi/\alpha_s(T)=(22/3)\, {\rm ln}\,(T/\Lambda)$
in terms of the
scale parameter $\Lambda$. The
dimensionless density
\begin{eqnarray} {n_D \over T^3} \sim \left({\Lambda \over T}\right)^{11/3}
\end{eqnarray}
is small at high $T$ but increases as $T$ decreases.
With the exception of section \ref{sec_deconf}, where we will estimate the critical deconfinement temperature
by including perturbative ${\cal O}(\alpha_s^0))$ effects in the dimensionless pressure,
we will always assume the temperature to be small enough, so that the dyons effect are the dominant ones.
The dyon fugacity $f$ is
\begin{eqnarray}
f\approx \frac{n_D}{8\pi}
\label{FUGA}
\end{eqnarray}
to order ${\cal O}(n_D^{3/2})$ in the dyon density (see below).
The absolute value of the parameter $\Lambda$ appearing in the semiclassical formulae
can be related to standard parameters like $\Lambda_{\overline{MS}}$, but this has
no practical value since the accuracy with which they are known is too low to give an accurate
value of the dyonic density. In practice it is obtained from
the fit to the lattice instanton
data performed in~\cite{SHURYAK} in the range $0.5<T/T_c<3$. The caloron action -- the sum of
$S_L$ and $S_M$ -- is then writen as
\begin{eqnarray}
S_{L+M} (T)\equiv \frac{2\pi }{\alpha_s(T)}\equiv \frac{22}3{\rm ln}\left(\frac {T}{0.36\,T_c}\right)
\label{RUN}
\end{eqnarray}
We will use this fit as a basis for our running coupling. In particular,
the action of the SU(2) caloron at $T_c$ $S_{\rm L+M} (T_c)\approx 7.47$ translates to the value of the coupling
$\alpha_s(T_c)=0.84$. Since in this paper we only work in the confining regime of the
holonomy with all dyon actions identical, the action per dyon is about 3.75.
The repulsive linear interaction between unlike dyons (anti-dyons) found in \cite{DPX}
acts as a linearly confining force in the center asymmetric
phase, favoring the molecular or KvBLL configuration at $T>T_c$.
This interaction stems from QGP thermal quanta scattering on the dyons.
However,
we will be interested in this paper in the center symmetric phase at $T>T_c$,
in which there is no QGP,
we do not include this interaction.
Since the classical $V_{D\bar D}\sim 1/\alpha_s$ it dominates the quantum determinants, which include Coulomb
interaction of order
$\alpha_s^0$. On this point we differ from the argument presented in~\cite{DP} regarding the re-organization of (\ref{VDD})
in an extended quantum determinant. At large relative separations
between all particles the measure (\ref{MEASUREDD}) is exact. It is also exact when each bunch of dyons or anti-dyons coalesce into a KvBLL instanton or anti-instanton at all separations.
The above notwithstanding, the grand-partition function associated with the measure (\ref{MEASUREDD})
\begin{eqnarray}
{\cal Z}_{D\overline D}[T]\equiv \sum_{[K]} \,\int d\mu_{D\overline D}[K]
\label{ZDD}
\end{eqnarray}
describes a highly correlated ensemble of dyon-anti-dyons which is no longer integrable in the presence of the streamline. The case $V_{D\overline D}=0$ amounts
to ${\cal Z}_{D\overline D}\rightarrow {\cal Z}_{D}{\cal Z}_{\overline D}$ where each factor can be exactly re-written in terms of a 3-dimensional
effective theory. We now analyze (\ref{ZDD}) for the SU(2) case following and correcting the arguments in~\cite{DP}.
\be
{\cal Z}_{D\overline D}[T]&&\equiv \sum_{[K]}\prod_{i_L=1}^{K_L} \prod_{i_M=1}^{K_M} \prod_{i_{\bar L}=1}^{K_{\bar L}} \prod_{i_{\bar M}=1}^{K_{\bar M}}\nonumber\\
&&\times \int\,\frac{fd^3x_{Li_L}}{K_L!}\frac{fd^3x_{Mi_M}}{K_M!}
\frac{fd^3y_{{\bar L}i_{\bar L}}}{K_{\bar L}!}\frac{fd^3y_{{\bar M}i_{\bar M}}}{K_{\bar M}!}\nonumber\\
&&\times {\rm det}(G[x])\,{\rm det}(G[y])\,\,e^{-V_{D\overline D}(x-y)}
\label{SU2}
\ee
with $G[x]$ a $(K_L+K_M)^2$ matrix and $G[y]$ a $(K_{\bar L}+K_{\bar M})^2$ matrix whose explicit form are given in~\cite{DP,DPX}.
\subsection{Classical dyon-antidyon interactions}
The explicit form of the Coulomb asymptotic in (\ref{SU2}) for the SU(2) case is
\be
&&V_{D\overline{D}}(x-y)\rightarrow-\frac {C_D}{\alpha_s\, T}\nonumber\\
&&\times\left(\frac 1{|x_M-y_{\overline{M}}|}+\frac 1{|x_L-y_{\overline{L}}|}-\frac 1{|x_M-y_{\overline{L}}|}-\frac 1{|x_L-y_{\overline{M}}|}
\right)\nonumber\\
\label{DDXX}
\ee
The strength of the Coulomb interaction in (\ref{DDXX}) is $C_D/\alpha_s$ and is of order $1/\alpha_s$. It follows from the asymptotics
of the streamline configuration. In Fig.~\ref{fig_pot} we show the attractive potential for the SU(2)
streamline configuration in the $M\bar M$ channel~\cite{LARSEN-SHURYAK}. The solid curve is a numerical fit to the
data given by
\begin{eqnarray}
V_{D\bar{D}}(r) \equiv s_{D\bar D}\,V(r)=s_{D\bar D}\,{A v\over g^2}{(r \cdot v-B)^2 \over (r \cdot v)^3+C}
\label{VMM}
\end{eqnarray}
with $s_{M\bar M}=-1$ in units of the critical temperature
$T_c$ and $g$ set to 1 and $A=30.9, B=0.9072, C=15.795$. The dashed line corresponds to the Coulomb
asymptotics
\begin{eqnarray}
V_{M\bar{M}}(r)\approx -\frac {C_D}{\alpha_s r}
\label{VMMC}
\end{eqnarray}
with $C_D=A/4\pi=2.46$. We recall that in the uncombed $D\bar D$ potential, the
asymptotic Coulomb interaction corresponds to
$C_D=2$. The attraction in the streamline is stronger asymptotically
owing to the relative combing between the dyons. Fig.~\ref{fig_pot} shows that the $D\bar D$ core is
about $a_{D\bar D}\approx 1/T$. The second observation is that one should not use the Coulomb asymptotic (the lower dashed curve)
but the actual potential which correctly takes care of the dyons, as extended charged objects rather than
point charges.
Below the core value of $a_{D\bar D}$, the streamline configuration annihilates
into perturbative gluons making the parametrization (\ref{VMM}) arbitrary. Throughout, we will
parametrize the core by a constant, replacing (\ref{VMM}) by
\begin{eqnarray}
V_{D\bar D}(r)\equiv s_{D\bar D}\left(V(r)\theta(r-a_{D\bar D})+V(a_{D\bar D})\theta(a_{D\bar D}-r)\right)\nonumber\\
\label{STEP}
\end{eqnarray}
with $s_{M\bar M}=s_{L\bar L}=-1$ (attractive) and $s_{L\bar M}=s_{L\bar M}=+1$ (repulsive).
The ensemble (\ref{SU2}) can be viewed as a 4-component dense and strongly coupled liquid. The quantity in the
exponent, known as the classical plasma parameter
\begin{eqnarray}
\Gamma_{D\bar D}=V(a_{D\bar D})\approx \frac{C_D/\alpha_s a_{D\bar D}}{3T_c}\approx 1
\label{GDD}
\end{eqnarray}
is not small. Its exponent $e^{\Gamma_{D\bar D}}$ can be even large.
This implies
that the ``dyonic plasma" we want to study belongs to a class of {\it strongly coupled} plasmas,
with non-negligible correlations between the particles. So a priori,
this problem should be studied by methods more powerful than the
usual mean field approximations, such as the Debye-Huckel theory.
However, we will show below that when the dyonic densities are sufficiently large (and that
implies the overall $T$ of the ensemble to be sufficiently low), the screening mass gets large enough
to put the effective -- screened -- interaction inside the domain in which the analytic Debye-Huckel theory
becomes justified.
Furthermore, as we will detail below, the treatment of the repulsive core
is in fact a rather sensitive issue.
We chose the ``most smooth" version of the potential, shown by the solid curve in Fig.~\ref{fig_pot}.
Its Fourier transform provides a smooth form factor in momentum space. We note that the actual streamline
was only found for distances $r> a_{D\bar D}\approx 1.2$ (about $4/v$ in the dyon units).
The upper (blue) dashed curve is an example of an arbitrary parameterization discussed in~\cite{LARSEN-SHURYAK}, extending it to smaller
values of $r$. If one uses it, or even cut off the small $r< a_{D\bar D}$ region
completely -- the approach known as hard core or excluded volume -- the Fourier transform of the potential develops large oscillations.
In this case the instability of the Debye-Huckel theory becomes stronger and its applicability domain shrinks.
The use of (\ref{VMM}) in the repulsive channels $M\bar L$ and $L\bar M$ approximates a smaller repulsion than Coulomb
at shorter distances. A numerical investigation of these channels would be welcome. Note that both the measure in
(\ref{SU2}) and the asymptotic (\ref{DDXX}) do not include the quantum corrections around the streamline configuration. Both of
which should add more repulsion to the interaction between $D$ and $\bar D$. A leading quantum correction to the
asymptotic (\ref{DDXX}) follows by analogy from the Coulomb corrections emerging from the $DD$ and $\bar D\bar D$ determinantal
interactions. In our case they are repulsive and amount to the shift
\begin{eqnarray}
C_D\rightarrow C_D -\frac{2\alpha_s}\pi +{\cal O}(\alpha_s^2)
\label{SHIFT}
\end{eqnarray}
in the Coulomb constant. The relevance of this correction will be briefly discussed below.
\begin{figure}
\begin{center}
\includegraphics[width=8cm]{potential.jpg}
\caption{ (Color online)
Black solid line is the SU(2) $D\bar D$ (dimensionless) potential versus the distance $r$ (in units of $1/T$).
Upper (blue) dashed line is the parameterization proposed in Ref.\cite{LARSEN-SHURYAK}, the lower (red)
(dashed) line is the Coulomb asymptotics. }
\label{fig_pot}
\end{center}
\end{figure}
\subsection{Qualitative effect of the one-loop moduli space}
The volume element of the moduli space of
self-dual SU(2) dyons is given by $\sqrt{g_{HK}}\equiv {\rm det}\,G$ with $g_{HK}$
its associated Hyper-Kahler metric~\cite{DP}.
As we already mentioned in the introduction, the one-loop determinant in the measure
(\ref{MEASUREDD}) must be positive definite for all configurations.
Furthermore, the positivity of all eigenvalues is required,
since they have the meaning of the volume element in the corresponding subspace.
As noted in \cite{LATTICE}, this is not the case
for ensembles with randomly placed dyons. These ensembles get denser
and the positivity condition is only fulfilled for a very small fraction of the configurations.
In fact one of the main issues of the dyonic ensembles is the non-trivial character of the
one-loop interaction induced by the Diakonov determinant.
Before we show how this carries to our case through various
fermionization and bosonization and diagrammatic re-summations, it is instructive to provide a qualitative
understanding of the issues using simple explicit examples.
Although it is well known, for completeness let us start with
the simplest case of two dyons in the SU(2) theory with symmetric
holonomy $\nu=\bar\nu=1/2$. Omitting the overall factors, Diakonov $2\times 2$ matrix $G$ reads
\be
G_{2\times 2}[x]\sim
\begin{pmatrix}
\,&1\pm\frac 1{v x_{12}} & \mp \frac 1 {v x_{12}} \\
\,&&\,&\,\\
\,& \mp \frac 1{v x_{12}}& 1 \pm \frac 1{ v x_{12}}
\end{pmatrix}
\label{GMATRIX}
\ee
with $x_{12}\equiv |{\vec x}_{(1)}-{\vec x}_{(2)}|$ the distance between the dyons in units of $1/v=1/\pi T$. The
upper signs are for different (ML) dyons, and the lower for similar (MM, LL) pairs.
The metric-induced potential is thus
$V(x_{12})\equiv -{\rm ln det}\,G= -{\rm ln}\,(1\pm 2/(v x_{12}))\approx \mp 2/(v x_{12})$
is Coulomb-like at large distances.
(At short distances the induced
potential is proportional to ${\rm ln}(1/r)$ and not $1/r$.
There is no divergence in the partition function.)
Let us now consider an ensemble of several ($N=8$) dyons
with $N_M=N_L=4$ and set them randomly in a cube of size $a$. We then evaluate all inter-dyon
distances and calculate ${\rm det}\, G[x]$ (which is now an $8\times 8$ matrix)
as a function of the Coulomb parameter $\epsilon=1/(\pi aT)$. For each sampling,
the determinant is a polynomial of $\epsilon$ of degree $N$. The results of 10 random samplings
are displayed in Fig.\ref{fig_8dyons} by the dashed lines. For small $\epsilon$ the determinant deviates
from 1 in a non-uniform way. Some configurations are Coulomb attractive with ${\rm det}\,G>1$, while
some others are repulsive with ${\rm det}\,G<1$ for small $\epsilon$. To first order, they
average to zero for a large number of charges as there are equal number of positive and negative ones.
At next order, the attraction is to win thanks to the general theorem of second order perturbation theory.
However, we observe that already for $\epsilon=1/(\pi aT)\sim 0.2$ the repulsive trend is dominant
and ${\rm det}\,G< 0$ for some samplings. This means that the moduli space of these configurations
vanishes at the corresponding density. This sets an upper limit on the density of random ensembles
of dyons
\begin{eqnarray}
n<n_{\rm max}=8\,(0.2\,\pi T)^3\sim 1.98\,T^3
\end{eqnarray}
The lesson: Diakonov determinantal interaction for randomly placed dyons is strongly repulsive, reducing dramatically
the moduli space all the way to zero size for small $\epsilon$. It amounts to a strong effective core
of order $\alpha_s^0$.
However this is not the end of the story. Let us look at
the opposite case of a well ordered arrangement of dyons in the unit box.
For that we pre-arrange
the 8 dyons of the previous ensemble in a salt-like or fcc configuration on the unit cube, and assess
the corresponding ${\rm det}\,G$. The result is shown in Fig.~\ref{fig_8dyons} by the solid line.
While the qualitative trend is the same -- attraction at some interval of densities, changing to repulsion
and then reaching zero at some density -- the value of the maximal density to be reached is
changed by a large factor of about $5^3=125$. Here is lesson number 2: the moduli space can be made
much larger for the same inter-particle Coulomb strength $\epsilon$, if the correlations between charges are
correctly taken into account.
The overall lesson we get from those examples is the following: Diakonov's original suggestion that
attraction and repulsion would always cancel out is incorrect. Our analysis shows that
ultimately the repulsion
always wins and at some density the volume of the moduli space always goes to zero. However,
correctly implemented correlations between charges to maximize screening locally, can increase this critical density
by about two orders of magnitude.
\begin{figure}
\begin{center}
\includegraphics[width=7cm]{8dyons.jpg}
\caption{ (Color online) ${\rm det}\,{G}$ as a function of $\epsilon=1/(\pi aT)$. The dashed lines are for 8 dyons
randomly placed in a cube of size $a\equiv 1$. The solid line is for correlated dyons in a salt-like or fcc configuration
also in a unit cube. }
\label{fig_8dyons}
\end{center}
\end{figure}
\subsection{Fermionization and Bosonization}
Following~\cite{DP} each determinant in (\ref{SU2}) can be fermionized using 4 pairs of ghost fields $\chi^\dagger_{L,M},\chi_{L,M}$ for the dyons
and 4 pairs of ghost fields $\chi^\dagger_{{\bar L},{\bar M}},\chi_{{\bar L},{\bar M}}$ for the anti-dyons. The ensuing Coulomb factors from the determinants are then bosonized using 4 boson fields $v_{L,M},w_{L,M}$ for the dyons and similarly for
the anti-dyons. The result is a doubling of the 3-dimensional free actions obtained in~\cite{DP}
\be
&&S_{1F}[\chi,v,w]=-\frac {T}{4\pi}\int d^3x\nonumber\\
&&\left(|\nabla\chi_L|^2+|\nabla\chi_M|^2+\nabla v_L\cdot \nabla w_L+\nabla v_M\cdot \nabla w_M\right)+\nonumber\\
&&\left(|\nabla\chi_{\bar L}|^2+|\nabla\chi_{\bar M}|^2+\nabla v_{\bar L}\cdot \nabla w_{\bar L}+\nabla v_{\bar M}\cdot \nabla w_{\bar M}\right)
\label{FREE1}
\ee
For the interaction part $V_{D\bar D}$, we note that
the pair Coulomb interaction in (\ref{DDXX}) between the dyons and anti-dyons can also be bosonized using
the standard trick~\cite{POLYAKOV} in terms of $\sigma$ and $b$ fields. Here $\sigma$ and $b$
are the un-Higgsed long range U(1) parts of the original magnetic field $F_{ij}$ and electric potential
$A_4$ (modulo the holonomy) respectively. As a result each dyon species acquire additional
fugacity factors such that
\begin{eqnarray}
M:e^{-b-i\sigma}\,\,\,\,\, L:e^{b+i\sigma}\,\,\,\,\, \bar M: e^{-b+i\sigma}\,\,\,\,\, \bar L :e^{b-i\sigma}
\end{eqnarray}
These assignments are consistent with those suggested in~\cite{UNSAL,TIN} using different arguments.
As a result there is an additional contribution to the free part (\ref{FREE1})
\be
&&S_{2F}[\sigma, b]=T\int d^3x\, d^3y\\
&&\times\left(b(x)V^{-1}(x-y) b(y)+ \sigma(x)V^{-1}(x-y)\sigma(y)\right)\nonumber
\label{FREE2}
\ee
with $V(r)$ defined in~(\ref{VMM}). The interaction part is now
\be
&&S_I[v,w,b,\sigma,\chi]=-\int d^3x \nonumber\\
&&e^{-b+i\sigma}f\left(4\pi v_m+|\chi_M -\chi_L|^2+v_M-v_L\right)e^{w_M-w_L}+\nonumber\\
&&e^{+b-i\sigma}f\left(4\pi v_l+|\chi_L -\chi_M|^2+v_L-v_M\right)e^{w_L-w_M}+\nonumber\\
&&e^{-b-i\sigma}f\left(4\pi v_{\bar m}+|\chi_{\bar M} -\chi_{\bar L}|^2+v_{\bar M}-v_{\bar L}\right)e^{w_{\bar M}-w_{\bar L}}+\nonumber\\
&&e^{+b+i\sigma}f\left(4\pi v_{\bar l}+|\chi_{\bar L} -\chi_{\bar M}|^2+v_{\bar L}-v_{\bar M}\right)e^{w_{\bar L}-w_{\bar M}}
\label{FREE3}
\ee
In terms of (\ref{FREE1}-\ref{FREE3}) the dyon-anti-dyon partition function (\ref{ZDD}) can be exactly re-written as an interacting
effective field theory in 3-dimensions,
\begin{eqnarray}
{\cal Z}_{D\overline D}[T]\equiv \int D[\chi]\,D[v]\,D[w]\,D[\sigma]\,D[b]\,e^{-S_{1F}-S_{2F}-S_{I}}
\label{ZDDEFF}
\end{eqnarray}
In the absence of the screening fields $\sigma, b$ (\ref{ZDDEFF}) reduces to the 3-dimensional effective field theory
discussed in~\cite{DP} which was found to be integrable. In the presence of $\sigma, b$ the integrability is lost as the
dyon-anti-dyon screening upsets the hyper-Kahler nature of the moduli space. We will investigate them by linearizing
the screening effects in the symmetric state.
Since the effective action in (\ref{ZDDEFF}) is linear in the $v_{M,L,\bar M,\bar L}$, the latters are auxiliary fields that
integrate into delta-function constraints. However and for convenience, it is best to shift away
the $b,\sigma$ fields from (\ref{FREE3}) through
\begin{eqnarray}
&&w_M-b+i\sigma\rightarrow w_M\nonumber\\
&&w_{\bar M}-b-i\sigma\rightarrow w_{\bar M}
\label{SHIFT}
\end{eqnarray}
which carries unit Jacobian and no anomalies, and recover them in the pertinent arguments of the delta function constraints as
\be
&&-\frac{T}{4\pi}\nabla^2w_M+f e^{w_M-w_L}-f e^{w_L-w_M}=\frac {T}{4\pi}\nabla^2(b-i\sigma)\nonumber\\
&&-\frac{T}{4\pi}\nabla^2w_L+f e^{w_L-w_M}-f e^{w_M-w_L}=0
\label{DELTA}
\end{eqnarray}
and similarly for the anti-dyons.
In~\cite{DP} it was observed that the classical solutions
to (\ref{DELTA}) can be used to integrate the $w^\prime$s in (\ref{ZDDEFF}) to one loop. The resulting bosonic determinant
was shown to cancel against the fermionic determinant after also integrating over the $\chi^\prime$s in (\ref{ZDDEFF}). This
holds for our case as well. However, the presence of $\sigma, b$ makes the additional parts of (\ref{ZDDEFF}) still very
involved in 3 dimensions.
After inserting the constraints in the 3-dimensional effective action in (\ref{ZDDEFF}), the ground state corresponds
to constant fields because of translational invariance. Thus, the potential per unit 3-volume $V_3$ following from(\ref{FREE3})
after the shifts (\ref{SHIFT}) is
\be
-{\cal V}/V_3&&=4\pi f\, \left(v_me^{w_M-w_L}+v_le^{w_L-w_M}\right)\nonumber\\
&&+4\pi f\, \left(v_{\bar m}e^{w_{\bar M}-w_{\bar L}}+v_{\bar l}e^{w_{\bar L}-w_{\bar M}}\right)\nonumber\\
\label{POT}
\ee
Note that if we did not perform the shift (\ref{SHIFT}) then both the potential
(\ref{POT}) and the constraints (\ref{DELTA}) depend on $b$ and $\sigma$ making
the extrema search for ${\cal V}$ more involved. Of course the results should be the same.
For fixed holonomies $v_{m,l}$, the constant $w^\prime$s are real by (\ref{DELTA})
as all right hand sides vanish, and the extrema of (\ref{POT}) occur for
\be
e^{w_M-w_L}=\pm \sqrt{v_l/v_m}\nonumber\\
e^{w_{\bar M}-w_{\bar L}}=\pm \sqrt{v_{\bar l}/v_{\bar m}}
\label{EXT}
\ee
(\ref{EXT}) is only consistent with (\ref{DELTA}) if and only if
$v_l=v_m=1/2$ and $v_{\bar l}=v_{\bar m}=-1/2$. That is for confining holonomies or a center
symmetric ground state. However and because of the constraint (\ref{DELTA}) there is no effective potential
for the holonomies in the interacting dyon-anti-dyon liquid. Indeed, by enforcing (\ref{DELTA}) before variation
we have ${\cal V}/V_3=-n_D$,
whatever the $v^\prime$�s. On this point we differ from the arguments and corresponding results
made in~\cite{DP} where the constraints (\ref{DELTA}) were not enforced before the variational derivation
of the holonomy potential. Note that the alternative argument in~\cite{DP} in favor of the holonomy potential
fixes the number of dyon species $K_i$ to be equal a priori, while (\ref{SU2}) fixes it only on the average.
\section{Linearized Screening Approximation in Center Symmetric State}
For the center symmetric ground state of the 3-dimensional effective theory, we may assess the correction to
the potential ${\cal V}$ to one-loop
in the $b, \sigma$ fields. This is achieved by linearizing the constraints (\ref{DELTA}) around the ground state solutions.
Specifically
\be
&&\left(-\frac{T}{4\pi}\nabla^2+2 f\right)\,w_M-2fw_L\approx \frac {T}{4\pi}\nabla^2(b-i\sigma)\nonumber\\
&&\left(-\frac{T}{4\pi}\nabla^2+2f \right)\,w_L-2fw_M\approx 0
\label{DELTAL}
\end{eqnarray}
and similarly for the anti-dyons.
The one-loop correction to ${\cal V}$ follows by inserting (\ref{DELTAL})
in (\ref{ZDDEFF}). The ensuing quadratic contributions before integrations are
\begin{eqnarray}
S_{1L}=
{\cal V}-4\pi f\int \frac{d^3p}{(2\pi)^3}\frac{(\frac{T}{4\pi} p^2)^2}{(\frac{T}{4\pi}p^2+4f)^2}\left(b(p)^2-\sigma(p)^2\right)\nonumber\\
\label{quad}
\end{eqnarray}
The coefficient of the $b$ field appears tachyonic but is momentum dependent and vanishes at zero momentum.
\subsection{Pressure}
Carrying the Gaussian integration in $b,\sigma$ in (\ref{quad}) yields to one-loop
\be
{\rm ln} Z_{\rm 1L}/V_3=-{\cal V}-\frac{1}{2}\int \frac{d^3p}{(2\pi)^3}
{\rm ln}\left|1-\frac {V^2(p)}{16}\frac{p^8 M^4}{(p^2+M^2)^4}\right|\nonumber\\
\label{1loop}
\ee
with $V(p)$ the Fourier transform of (\ref{VMM})
\begin{eqnarray}
V(p)=\frac{4\pi}{p^2}\int_0^\infty dr\,{\rm sin}\,r\,V_{D\bar D}(r/p)
\label{VP}
\end{eqnarray}
and the screening mass $M=\sqrt{2n_D/T}$ with $|Q^2|=2$ for SU(2).
In Fig.~\ref{fig_FF} we show the form factor (\ref{VP}) in dots line
in units of $T_c$. A simple parametrization is shown in solid line of the form
\begin{eqnarray}
V(p)\approx {4\alpha}\,\frac{e^{-p\,a_{D\bar D}}}{p^2}\,\,{\rm cos}(p\,a_{D\bar D})
\label{FFP}
\end{eqnarray}
with $\alpha=\pi C_D/\alpha_s$ and a core $a_{D\bar D}\approx 1/T_c$.
Inserting (\ref{FFP}) into (\ref{1loop}) and setting $\tilde p=p/M$ yield
\be
{\rm ln} Z_{\rm 1L}/V_3=-{\cal V}-\frac{M^3}{2}\int \frac{d^3{\tilde p}}{(2\pi)^3}
{\rm ln}\left|1-{\tilde\alpha}^2(\tilde p)\frac{{\tilde p}^4}{({\tilde p}^2+1)^4}\right|\nonumber\\
\label{1loopXX}
\ee
with
\begin{eqnarray}
\tilde\alpha(\tilde p)\equiv \alpha\,e^{-{Ma_{D\bar D}\tilde p}}\,{\rm cos}\,({Ma_{D\bar D}\tilde p})
\label{ALPHAP}
\end{eqnarray}
The dominant contribution to the integral in (\ref{1loop}) comes from the region $\tilde p\approx 1$
for which (\ref{ALPHAP}) can be approximated by $\tilde\alpha(1)\equiv\tilde\alpha$. As a result (\ref{1loop})
can be done approximately by fixing $\tilde\alpha$ and we have the classical contribution to the pressure
\begin{eqnarray}
\frac{{\cal P}_{\rm cl}}T\equiv {\rm ln} Z_{\rm 1L}/V_3\approx n_D+\kappa(\tilde\alpha)\frac{M^3}{12\pi}
\label{PRESSURE}
\end{eqnarray}
with
\begin{eqnarray}
\kappa(\tilde\alpha)=\frac{2+\frac 52 \tilde\alpha+\frac 12 \tilde\alpha^2}{\sqrt{1+\frac{\tilde\alpha}4}}+
\frac{2-\frac 52 \tilde\alpha+\frac 12 \tilde\alpha^2}{\sqrt{1-\frac{\tilde\alpha}4}}-4
\label{KAPPA}
\end{eqnarray}
(\ref{KAPPA}) is seen to vanish for $\tilde\alpha=0$ or in the absence of $D\bar D$
interactions. Near $T_c$ the screening mass is $M\approx \sigma_E/T_c$ (see below), thus
\begin{eqnarray}
\tilde\alpha\equiv (\pi C_D/\alpha_s)\,e^{-{Ma_{D\bar D}}}\,{\rm cos}\,({Ma_{D\bar D}})\approx -0.52
\label{ALPHAXX}
\end{eqnarray}
For $|\tilde\alpha|<4$ the 1-loop contribution to the pressure from the charged $D\bar D$ dyons
is real with no dimer or molecular instability. The large core produced by the form factor (\ref{ALPHAP})
is considerably screened by the large dyon density as captured by the large dielectric constant
$1/\kappa(-0.52)\approx 5.26$ in (\ref{PRESSURE}).
\begin{figure}
\begin{center}
\includegraphics[width=8cm]{PastedGraphic-1.jpg}
\caption{ (Color online)
The dots show the form factor, the ratio $V(p)\cdot(p^2/4\pi)$ of the Fourier transform of (\ref{VMM}) to that of a pure Coulomb law
versus $p/T$. The thin line is its parameterization. See text.}
\label{fig_FF}
\end{center}
\end{figure}
The correction in (\ref{PRESSURE}) to the free contribution is a Debye-Huckel
correction~\cite{DH} (and references therein). A simple but physical way to understand it is to note that a
screened Coulomb charge carries a lower constant energy
\begin{eqnarray}
\frac {e^{-M|x|}}{4\pi |x|}\approx \frac 1{4\pi |x|}-\frac{M}{4\pi}+...
\label{DHXX}
\end{eqnarray}
The Debye-Huckel as a mean-field estimate for the pressure follows
\begin{eqnarray}
\frac{{\cal P}_{DH}}{T}\approx \frac{n_DM}{4\pi T} =\frac{M^3}{8\pi} \rightarrow \frac{M^3}{12\pi}
\label{STANDARD}
\end{eqnarray}
where $n_D=M^2T/2$ is the density of charged particles (see below). The standard Debye-Huckel limiting result
for a multi-component ionic plasma in 3 spatial dimensions
is shown on the right-most side of (\ref{STANDARD}).
The correction in (\ref{PRESSURE}) is considerably reduced by the large screening
through the effective dielectric constant played by $1/\kappa(\tilde \alpha)\approx 32/(15{\tilde\alpha}^2)$
for $\tilde\alpha\ll 1$. In particular $1/\kappa (-0.52)\approx 5.26\gg 1$ as noted earlier. It can be recast in the form
\be
\frac{{\cal P}_{\rm cl}}{T^4}={\tilde n}_D+\frac{\kappa (\tilde\alpha)}{3\pi\sqrt{2}}\,{\tilde n}_D^{\frac 32}
\label{EOS1}
\ee
with ${\tilde n}_D=n_D/T^3$.
Using
$Ma_{D\bar D}\approx \sigma_E/T_c^2\approx 1/(0.71)^2$ for SU(2) we have
${\tilde n}_D\approx 1$, so that
${\cal P}_{\rm cl}/T^4\approx (1+0.01))$. The screening corrections are small of the order of 1\%
thanks to the large dyonic densities.
The limitations of the Debye-Huckel approximation are readily seen from (\ref{1loop}).
In Fig.~\ref{fig_log}a we plot the argument of the logarithm in the last term of (\ref{1loop}).
The different curves from top to bottom follow from $Ma_{D\bar D}=1.5, 1, 0.7, 0.56$ respectively.
The smaller the Debye mass $M$ the stronger the dip. For $Ma_{D\bar D}<0.56$, the argument
of the logarithm becomes negative resulting into an $i\pi$ contribution to the pressure and thus an instability.
This is a clear indication of a well known phenomenon:
the Debye-Huckel approximation is in general inapplicable for strongly coupled plasmas,
and the interaction mediated by the streamline is strong. Only a large enough density
of dyons, producing sufficiently strong screening, allows for the use of the Debye-Huckel theory.
In Fig.~\ref{fig_log}b we show how the total integrated contribution to the free energy
changes as a function of the dimensionless Debye mass $Ma_{D\bar D}$
\begin{eqnarray}
(M a_{D\bar D})^3\int_0^\infty dp \,p^2 {\rm ln}\left|1-\frac {V^2(p M)}{16}\frac{p^8 }{(p^2+1)^4}\right|
\label{INTEGRAND}
\end{eqnarray}
The main lesson is that beyond the critical
value of the screening, this contribution becomes rapidly very small. This is
consistent with the analytical estimate above. This justifies the use of the
Debye-Huckel mean-field analysis in general, and the use of the semi-classical expansion
in particular.
\begin{figure}
\begin{center}
\includegraphics[width=7cm]{inside_log.jpg}
\includegraphics[width=8cm]{integral.jpg}
\caption{ (Color online)
(a) The argument of the logarithm in the last term of (\ref{1loop}) versus the dimensionless momentum $p$, for
different values of the dimensionless Debye mass $M a_{D\bar D}=1.5,1,0.7,0.56$, top to bottom. As the screening mass decreases
to its critical value,
the lower (green) curve touches zero. The smaller values of $M$ leads to a negative argument of the logarithm, thus
an instability.
(b) A semi-logarithmic plot of the integral entering in (\ref{1loop}) as defined in (\ref{INTEGRAND})
as a function of $M a_{D\bar D}$. The decrease is steady from its maximum at the critical value of the screening mass
or $Ma_{D\bar D}=0.56$.
}
\label{fig_log}
\end{center}
\end{figure}
\subsection{Beyond the Debye-Huckel theory}
The unravelling of the Debye-Huckel approximation may be due
to corrections to an interacting Coulomb system, such as
1/ core corrections; 2/ dimer, tetramer and so on many-body interactions. The large core corrections were already
identified and discussed above and yield a substantial reduction in the Debye-Huckel contribution
near the critical value of $Ma_{D\bar D}\sim 0.56$.
Bound state corrections in the form of electrically charged $L\bar L$ or $M\bar M$ dimers, or electrically neutral
$L\bar L \,M\bar M$ tetramers commonly referred to as instanton-anti-instanton molecules, can bind through the
streamline interaction
(\ref{DDXX}-\ref{VMM}). The combinations $L\bar M$ and $M\bar L$ are repulsive. The binding energy in a dimer
is $\epsilon_{D\bar D}\approx (C_{D}/3)/(\alpha_s a_{D\bar D})=T$. The dimer enhancement is expected to be of order
$e^{\epsilon_{D\bar D}/T}\approx e^{(C_D/3)/\alpha_s}\approx 2.72$ for $T\approx T_c$
using the reduced effective Coulomb coupling. As we noted earlier, this enhancement becomes
substantially larger at high temperature as $\alpha_s$ decreases with the onset of dimerization set at about
$\alpha_{s,{\rm crit}}=\pi (C_D/3)/4\approx 0.67$. At this coupling which occurs above $T_c$, the Coulomb dimer enhancement factor
is $e^{C_D/3\alpha_{s,{\rm crit}}}\approx 3.57$.
In sum, the dyons and anti-dyons form a
Coulomb liquid with strong short range correlations induced by both the finite cores and bindings.
The liquid supports center symmetry and confines. The deconfinement transition is characterized by
clustering into charged dimers and possibly uncharged and topologically neutral tetramers, forming
mixtures with the restoration of center symmetry. Coulomb mixtures present rich phase diagrams~\cite{FISHER}.
\subsection{Dyonic densities}
(\ref{1loop}) can be readily used to assess the dyon densities $K_M$ and $K_L$ (and similarly for $K_{\bar M}$ and $K_{\bar L}$)
in the center symmetric vacuum with screening
dyons-anti-dyons. For that we need to change $f\rightarrow \sqrt{f_Mf_L}$ and take derivatives
of (\ref{1loop}) with respect
to ${\rm ln}f_{M,L}$ separately and then setting them equal by bulk charge neutrality. The result per species is
\begin{eqnarray}
K= \frac 14 n_D+\kappa(\tilde\alpha)\frac {M^3}{32\pi}
\label{DENSITY}
\end{eqnarray}
for all dyon and anti-dyon species.
Each dyon (anti-dyon) is characterized by an SU(2) core of size $\rho\approx 1/(2\pi T\,\nu)
\approx 0.33$ fm in the center symmetric phase with $\nu=1/2$ at $T=1/{\rm fm}$. The Debye length
$\lambda_D=1/M\approx \sqrt{T/2n_D}\approx 0.70\,
{\rm fm}$ is about twice the core size. The classical Coulomb ratio for the ${DD,\bar D\bar D}$ pairs with a core of $2\rho$ is about
\begin{eqnarray}
\Gamma_{DD,\bar D\bar D}\equiv \frac {1}{2\pi (2\rho)T}\approx \frac {\nu}2=\frac 14
\end{eqnarray}
is small. Recall from (\ref{GDD}) that $\Gamma_{D\bar D}\approx 1$.
The Coulomb ${DD,\bar D\bar D}$ interactions are quantum and of order $\alpha_s^0$ with strength $1/\pi$
as can be seen by expanding the exponential form of the determinantal interaction in (\ref{GMATRIX}).
The dyon-anti-dyon ensemble is close to a strongly coupled 4-component Coulomb liquid. Since the measure for the unlike dyons is exact, it is valid even in the dense configuration. It is only asymptotically exact for like dyons. For the dyons and anti-dyons the streamline is numerically exact
at all separations outside its core. However its corresponding quantum determinant was not calculated. Only a qualitative correction was argued
in (\ref{SHIFT}).
\subsection{Gluon condensates and susceptibilities}
The topological charge fluctuates locally in this
dyon-anti-dyon model. The topological susceptibility at 1-loop follows from (\ref{1loop}) through the substitution
$f\rightarrow f{\rm cos}(\theta/2)$ both in ${\cal V}$ and also $M\rightarrow M\sqrt{\rm cos(\theta/2)}$.
At finite vacuum angle $\theta$ and in leading order we have
\be
&&\left<F\tilde F\right>_\theta\equiv - \frac{T}{V_3}\frac{\partial {\rm ln}Z_{1L}}{\partial\theta}=\nonumber\\
&&{\rm sin}(\theta/2)\left(\frac 12 n_DT+\kappa(\tilde\alpha)\frac {M^3T}{16\pi}\sqrt{{\rm cos}(\theta/2)}\right)
\nonumber\\
\label{TOP0}
\ee
Thus the topologicl susceptibility
\be
\chi_T\equiv \frac {V_3}{T}\left<(F\tilde F)^2\right>_0\approx \left(\frac 12\right)^2 (n_DT)
\label{TOP1}
\end{eqnarray}
in leading order.
Since the dyons carry half the topological charge (\ref{TOP1}) shows
that the topological fluctuations are Poissonian to order ${\cal O}(n_D^{3/2})$.
The behavior of $\chi_T/T^4$ versus $T/T_c$ is shown in Fig.~\ref{fig_xit} with
$n_D$ defined in (\ref{NDX}) below.
\begin{figure}
\begin{center}
\includegraphics[width=7cm]{xi_top.jpg}
\caption{Topological susceptibility in units of $T$ versus $T/T_c$}
\label{fig_xit}
\end{center}
\end{figure}
The gluon condensate to 1-loop in the screening approximation follows from
\be
&&\frac 1{16\pi^2}\left<F^2\right>_0\equiv -\frac {T}{4\pi V_3}\frac{\partial\, {\rm ln} Z_{1L}}{\partial \,{1/\alpha_s}}\nonumber\\
&&\approx -\frac{T}{4\pi}\left(\frac 2{\alpha_s}-\pi\right)\left(n_D+\frac{\kappa(\tilde\alpha)M^3}{8\pi}\right)
\label{TOP2}
\ee
which is non-Poissonian because of the scale anomaly.
The compressibility of the ground state is
\be
&&\sigma_\chi\equiv \frac{V_3}{T}\left<\left(\frac{F^2}{16\pi^2}\right)^2\right>_c\\
&&\approx \frac{T}{16\pi^2}\left(2\left(n_D+\frac{\kappa M^3}{8\pi}\right)+\left(\frac 2{\alpha_s}-\pi\right)^2
\left(n_D+\frac{3\kappa M^3}{16\pi}\right)\right)\nonumber
\ee
for the connected correlator.
We can use (\ref{PRESSURE}) and (\ref{TOP2}) to extract the electric $\left<E^2\right>_0$ and magnetic $\left<B^2\right>_0$
condensates in the dyonic ensemble. For that we note that the energy per volume in Euclidean space follows from (\ref{PRESSURE})
through
\begin{eqnarray}
\frac 1{8\pi}{\left<{B^2-E^2}\right>_0}=T^2\frac{\partial}{\partial T}\frac{P_{\rm cl}}{T}
\label{EEV3}
\end{eqnarray}
The results are
\be
&&\frac{\left<B^2\right>_0}{4\pi T}=\,\,
\left(+3-(1+\frac {2\alpha_s^\prime}{\alpha_s^2})\left(\frac 1{\alpha_s}-\frac \pi2\right)\right)
\left(n_D+\frac{\kappa M^3}{8\pi}\right)\nonumber\\
&&\frac{\left<E^2\right>_0}{4\pi T}=\left(-3
-(1-\frac{2\alpha_s^\prime}{\alpha_s^2})\left(\frac 1{\alpha_s}-\frac \pi2\right)\right)
\left(n_D+\frac{\kappa M^3}{8\pi}\right)\nonumber\\
\ee
with $\alpha_s^\prime=\partial\alpha_s/\partial\, {\rm ln} T$. In Fig.~\ref{fig_EEBB} we show the behavior of
the chromo-electric condensate $\left<E^2\right>$ (solid-black), the chromo-magnetic condensate
$\left<B^2\right>$ (dashed-blue) and the (Euclidean) energy density $\left<B^2-E^2\right>$ (dot-dashed-brown)
in units of $T$ versus $T/T_c$ in the center symmetric phase. We used the dyon density fixed in
(\ref{NDX}) below. The chromo-magnetic condensate is about constant in the range of $0.6<T/T_c<1$
while the chromo-electric condensate decreases monotoneously. The condensates are about equal and opposite
near $T_c$ a point supported by the lattice extracted condensates in~\cite{ADAMI}.
\begin{figure}
\begin{center}
\includegraphics[width=7cm]{EEBB.jpg}
\caption{The electric $\left<E^2\right>$ (solid-black), magnetic $\left<B^2\right>$ (dashed-blue) and
(Euclidean) energy density $\left<B^2-E^2\right>$ (dot-dashed-brown)
in units of $T$ versus $T/T_c$. See text.}
\label{fig_EEBB}
\end{center}
\end{figure}
\subsection{Electric and magnetic screening masses}
The center symmetric phase of the dyon-anti-dyon liquid screens the long-range U(1) gauge fields left un-Higgsed
by the holonomy $A_4(\infty)/2\pi T= \nu T^3/2$. The electric and magnetic correlations in these Abelian U(1) charges can be
obtained by introducing the U(1) Abelian sources $\eta_{m,e}$ for the magnetic and electric charge densities
\begin{eqnarray}
\rho_{m,e}(x)=\sum_iQ_{m,e,i}\,\delta^3( x-{x}_i)
\end{eqnarray}
with $|Q_{m,e}|=1$, and shifting the U(1) fields $\sigma\rightarrow \sigma+\eta_m$ and $b\rightarrow b+\eta_e$
in the 3-dimensionsl effective action. To 1-loop the generating functional for the charge density correlators is
\begin{eqnarray}
Z_{1L}[\eta_m,\eta_e]=\int D[\sigma] D[b]\,e^{-S_{2F}[\sigma,b]-S_{1L}[\sigma+\eta_m,b+\eta_e]}
\label{ZEM}
\end{eqnarray}
which is Gaussian in the sources and therefore readily integrated out. Thus
\begin{eqnarray}
{\rm ln}{Z_{1L}[\eta_m,\eta_e]}=
{-\int \frac{d^3p}{(2\pi)^3}\sum_{i=e,m}\eta_i(p){\bf G}_i(p)\eta_i(-p)}
\end{eqnarray}
with the electric and magnetic density correlators following by variation,
\be
{\bf G}_{m,e}(p)\equiv \frac 1{V_3}\left<\left|\rho_{m,e}(p)\right|^2 \right>\approx \frac 14
\frac {TM^2p^4}{(p^2+M^2)^2\pm \tilde\alpha \,M^2p^2}\nonumber\\
\label{PAIR}
\ee
The upper sign is for magnetic and the lower sign for electric.
In x-space, (\ref{PAIR}) can be inverted by Fourier transforms.
The result for the electric correlator in spatial coordinates is
\be
\label{GEX}
&&-\frac{TM^4}{16\pi |x|}\,e^{-\sqrt{1-\frac {\tilde\alpha}4}\,M|x|}\nonumber\\
&&[
{\rm cos}\left(\frac{\sqrt{\tilde\alpha}}2\,M|x|\right)\,(\tilde\alpha -2) \nonumber\\&&+
{\rm sin}\left(\frac{\sqrt{\tilde\alpha}}2\,M|x|\right)\,\frac{1-2\tilde\alpha+\frac{{\tilde\alpha}^2}2}{\sqrt{\tilde\alpha\,(1-\frac{\tilde\alpha}4)}}]
\ee
The magnetic correlator follows by analytical continuation through the substitution $\tilde\alpha\rightarrow -\tilde\alpha$
in (\ref{GEX}). The electric screening masses $M_{M,E}$ follow from the large distance asymptotics. Using
our estimate of $\tilde\alpha\approx -0.52<0$ from the Debye-Huckel analysis above, we have
\be
&&\frac{{M_E}}{M}\approx \left(\sqrt{1+\frac{|\tilde\alpha|}4}-\frac{\sqrt{|\tilde\alpha|}}2\right)\approx 0.70\nonumber\\
&&\frac{M_M}{M}\approx \left(\sqrt{1-\frac{|\tilde\alpha|}4} \right)\approx 0.93
\label{SCREEME}
\ee
with $M^2=2n_D/T$. Also the arguments below show that
$M=\sigma_E/T$. Combining these two results allow us to fix $C$ in (\ref{NDDD}) above.
Indeed, at $T_c$ the SU(2) lattice results give $T_c/\sqrt{\sigma_E}\approx 0.71$. So (\ref{NDDD}) now reads
\begin{eqnarray}
\frac{n_D}{T^3}\approx 2\,\frac{\alpha^2_s(T_c)}{\alpha^2_s(T)}\,e^{\frac \pi{\alpha_s(T_c)} -\frac \pi{\alpha_s(T)}}
\label{NDX}
\end{eqnarray}
which gives $M_{E}\approx 1.4\,T_c$
and $M_M\approx 1.8\,T_c$, both of which are remarkably close to the reported SU(2) lattice results
in the vicinity of the critical temperature~\cite{BORN}. In Fig.~\ref{FIGEM}
we display the results (\ref{SCREEME}) for $M_{E,M}/T$ in the range $(0.5-1)\,T_c$
versus $T/T_c$. The points at $T>T_c$ are shown for comparison. We
note that the electric mass drops down at $T_c$. In the region we study $M_M>M_E$, while above $T_c$, in a more familiar
QGP region, $M_M<M_E$. This switching of the magnitude of the two screening masses
is better documented in lattice works with the $SU(3)$ gauge group. It has a simple explanation in our case.
Since at $T>T_c$ the dyon density drops it follows that $M$ decreases as well. As a result, the form factor in Fig.~\ref{fig_FF}
is probed at smaller momentum $p\approx M$ (larger distances) making $\tilde\alpha(p\approx M)$ in (\ref{ALPHAP})
switch from negative ($T<T_C$) to positive ($T>T_C$).
From (\ref{PAIR}) it follows that the expressions for $M_{E,M}$ in (\ref{SCREEME}) are now switched with
$M_M$ lighter than $M_E$. A simple estimate of the critical temperature at the crossing follows from the vanishing
of (\ref{ALPHAXX}) or $Ma_{D\bar D}=\pi/2$. This translates to a critical dyon density
$n_D^C\approx \pi^2T_c^3/8$ which is consistent
with our estimate of $T_c$ below (see (\ref{TCC})).
Finally, we note that
the value of $\alpha_s(T_c)\approx 0.84$ extracted from the cooled caloron data in~\cite{SHURYAK}
is also consistent with the reported value from bulk thermodynamics in~\cite{ALPHA}.
In the dyon-anti-dyon Coulomb liquid the correlators are modified
at intermediate distances as we now detail in terms of the static structure factors.
\begin{figure}
\centerline{
\includegraphics[width=6cm]{EE_EM}.jpg}
\caption{ (Color online) The electric $M_E/T$ (dashed line) and magnetic $M_M/T$ (solid line) screening masses in (\ref{SCREEME}) versus $T/T_c$.
The points are SU(2) lattice data from ~\cite{BORN} shown for comparison, (blue) circles are $M_E/T$, (red) squares are $M_M/T$.}
\label{FIGEM}
\end{figure}
\subsection{Static structure factors}
The charged structure factor between pair of magnetic or electric charges is
(\ref{PAIR}) which can be re-written as
\begin{eqnarray}
{\bf G}_{M,E}(p)\equiv \left<\frac 1{N_{m,e}}\left| \sum_{j=1}^{N_{m,e}}Q_{m,e,j}e^{ik\cdot x_j}\right|^2\right>
\end{eqnarray}
Thus
\begin{eqnarray}
{\bf G}_{M,E}(p)=\frac {{\bf G}_{m,e}(p)}{n_D/2}
\equiv \frac{\tilde p^4}{(\tilde p^2+1)^2\pm \tilde{\alpha}{\tilde p}^2}
\label{PAIRME}
\end{eqnarray}
with $\tilde p=p/M$.
We note that the pre-factor in (\ref{PAIRME}) involves two static electric or
magnetic exchanges with an identical screening mass $M$.
The charged structure factors vanish as ${\bf G}_{M,E}(p)\approx {\tilde p}^4$.
For large momenta or $\tilde p \gg 1$ both structure factors asymptote one from below
as shown in Fig.~\ref{FIGME}. The magnetic hole is slightly smaller than the electric one
around the same pairs. The absence of oscillations in the structure factor,
is a consequence of our linearized approximation.
\begin{figure}
\centerline{
\includegraphics[width=6cm]{GEM.jpg}}
\caption{ (Color online) The electric and magnetic structure factors (\ref{PAIR}) as a function of $p/M$.}
\label{FIGME}
\end{figure}
To characterize further the 4-component plasma of dyons and anti-dyons we define the scalar
static pair correlation function
\begin{eqnarray}
{\bf G}_S(x)=\left<\frac 1N\sum^N_{i\neq j}\delta^3(x+x_i-x_j)\right>
\label{PAIRS}
\end{eqnarray}
normalized to the total number of particles $N$. (\ref{PAIRS}) defines
the probability to find two particles a distance $|x|$ apart. Its Fourier transform
\begin{eqnarray}
{\bf G}_S(p)=\left<\frac 1N\left| \sum_{j=1}^{N} e^{ip\cdot x_j}\right|^2\right>
\label{PAIRSK}
\end{eqnarray}
is the scalar structure factor.
(\ref{PAIRSK}) can be evaluated by switching $f\rightarrow f+\delta f (x_i)$ in (\ref{SU2}) and then
linearizing the resulting effective action around the mean-density. Specifically, we can re-write the
linearized constraint (\ref{DELTAL}) formally as
\begin{eqnarray}
(w_M-w_L)=\frac{1}{\Delta_0+4\delta f}\left(\frac{T\nabla^2}{4\pi}\right)(b-i\sigma)
\label{PERT}
\end{eqnarray}
with $\Delta_0=-T\nabla^2/4\pi+4f$ and use perturbation theory to expand the denominator in
(\ref{PERT}) to order ${\cal O}(\delta f^3)$.
The result can be formally written as
\begin{eqnarray}
(w_M-w_L)(p)=\int\frac{d^3k}{(2\pi)^3}G(p,k)(b-i\sigma)(k)
\label{PERT1}
\end{eqnarray}
Inserting (\ref{PERT1}) into the potential (\ref{POT})
yields a quadratic action in $b$ and $\sigma$. Integrating over the latters yields the 1-loop determinant or
the effective action for $\delta\equiv \delta f/f$. Specifically
\begin{eqnarray}
{\rm det}(1+{\bf S}[\delta ])=e^{{\rm Tr ln}(1+{\bf S}[\delta ])}\approx e^{{\rm Tr}\, {\bf S}[\delta ]}
\end{eqnarray}
with the quadratic effective action for the scalar fluctuations as
\begin{eqnarray}
{\rm Tr}\, {\bf S}[\delta ]={-\int \frac{d^3p}{(2\pi)^3}\,\delta (p){\bf G}^{-1} (p)\delta (-p)}
\label{TSCA}
\end{eqnarray}
with to order ${\cal O}({\tilde\alpha}^3)$
\be
{\bf G}^{-1}(p)&&\approx n_D +
4\,{\tilde\alpha}^2M^8\int \frac{d^3k}{(2\pi)^3}\nonumber\\
&&[
\,\,\frac{k^4}{(k^2+M^2)^4}\frac{1}{((k+p)^2+M^2)^2}\nonumber\\
&&+\frac{2k^2}{(k^2+M^2)^3}\frac{1}{((k+p)^2+M^2)^2}\,\,]
\label{GSCA}
\ee
The $n_D$ contribution in (\ref{GSCA}) follows from the expansion of the leading contribution ${\cal V}$
using arguments similar to those used for the derivation of the dyonic densities above.
The scalar structure factor follows from (\ref{GSCA}) through the normalization
\begin{eqnarray}
{\bf G}_S(p)\equiv \frac {n_D}{V_3}\left<\left|\delta (p)\right|^2\right>=n_D\,{\bf G}(p)
\label{NSCA}
\end{eqnarray}
We note that the small momentum fluctuations in $\delta f$ couple
to the sound-like modes. Specifically,
\begin{eqnarray}
{\bf G}_S(p) \approx \frac{p^2}{c_s^2p^2}
\end{eqnarray}
is dominated by a massless pole at zero momentum
with
\begin{eqnarray}
c_s^2\approx 1 +8\,{\tilde\alpha}^2\left(\frac MT\right)
\int\frac{d^3k}{(2\pi)^3}\left[ \frac {k^4}{(k^2+1)^6}+\frac{2k^2}{(k^2+1)^5}\right]\nonumber\\
\label{CS2}
\end{eqnarray}
Alternatively from the pressure (\ref{PRESSURE}) we expect
\begin{eqnarray}
c_s^2\equiv \frac{\partial {\cal P}_{\rm cl}}{T\partial n_D}\approx
1 +\frac{\kappa(\tilde\alpha)}{4\pi}\left(\frac MT\right)
\end{eqnarray}
with $\kappa(\tilde\alpha)\approx 15\,{\tilde\alpha}^2/32$ in leading order and in total
agreement with (\ref{CS2}). Also, at large momentum (\ref{NSCA}) asymptotes
${\bf G}_S(\infty)=1$. The slight super-luminal character of (\ref{CS2}) reflects
on the fact that dyons are in essence Euclidean configurations with no physical
particle realization.
\subsection{Estimate of the critical $T_c$}
\label{sec_deconf}
The total thermodynamical pressure of the dyon-anti-dyon liquid consists of the classical and non-perturbative
contribution (\ref{EOS1}) plus the perturbative holonomy potential known as
Gross-Pisarski-Yaffe-Weiss (GPYW) potential
\cite{WEISS}, plus the purely perturbative
black-body contribution (ignoring the higher order ${\cal O}(\alpha_s)$ quantum corrections). Specifically $(N_c=2$)
\be
\frac{{\cal P}_{\rm tot}}{T^4}&&\approx {\tilde n}_D+\frac{\kappa (\tilde\alpha)}{{3\pi\sqrt{2}}}\,{\tilde n}_D^{\frac 32}
\nonumber\\
&&-\frac{\pi^2}{45}\left(N_c^2-\frac 1{N_c^2}\right)+\frac{\pi^2}{45}\left(N_c^2-1\right)
\label{EOS2}
\ee
The Debye-Huckel contribution is of order $N_c^{3}$, while the leading
classical contribution is of order $N_c^2$. So screening and large $N_c$ are not commutative.
For the SU(2) case of interest, the transition temperature $T_c$
from the disordered phase ($\nu=1/2$) to the ordered phase ($\nu=0$)
occurs when the first three contributions in (\ref{EOS2}) cancel out. Thus
\begin{eqnarray}
{\tilde n}_D+\frac{\kappa (\tilde\alpha)}{{3\pi\sqrt{2}}}\,{\tilde n}_D^{\frac 32}
\approx \frac{\pi^2}{12}
\label{TCC}
\end{eqnarray}
For $\kappa(-0.52)\approx 0.19$, the critical density is
$n^C_{D}\approx {\pi^2 T_c^3}/{12}\approx 0.88$.
Since the SU(2) electric string tension
is $\sigma_E=TM=\sqrt{2n_DT}$ (see below),
it follows that $T_c/\sqrt{\sigma}=6^{\frac 14}/\sqrt{\pi}\approx 0.88$ which is somehow
larger than the SU(2) lattice result $T_c/\sqrt{\sigma_E}=0.71$~\cite{TEPER}.
\section{Polyakov lines}
To probe the confining nature of the dyon-anti-dyon liquid in the 3-dimensional effective theory we will compute explicitly
the expectation of a heavy quark through the traced Polyakov line and the corelator of a heavy quark-anti-quark pair through
the correlator of the traced Polyakov line and its conjugate at fixed spatial separation. The insertion of these charges in the
dyon-anti-dyon liquid modifies the ground state through solitonic solutions around these sources.
In this section we present
a new derivation of the pertinent solitonic equations for the SU(2) case
that makes explicit use of the presence of the long range U(1) $b$ and $\sigma$
fields. In the linearized screening approximation, we show that the solitonic equations for the heavy source probes are in
agreement with those established in~\cite{DP} using different arguments.
\subsection{$\left<L\right>$}
\label{LSECTION}
In the SU(2) case the Polyakov line consists of inserting a heavy quark whose free energy consists of its Coulomb interactions
with all the Coulomb charged dyons and anti-dyons. Specifically, the traced Polyakov line before averaging is
\be
\label{L1X}
&&L(x_1)=\\
&&e^{2\pi i\mu_M+\frac{i}{2T}\sum_{i}\left(\frac{1}{|x_1-x_{Mi}|}+ \frac{1}{|x_1-x_{\bar Mi}|}-\frac{1}{|x_1-x_{Li}|}-\frac{1}{|x_1-x_{\bar Li}|}\right)}+\nonumber\\
&&e^{{2\pi i\mu_L}+\frac i{2T}\sum_{i}\left(\frac{1}{|x_1-x_{Li}|}+ \frac{1}{|x_1-x_{\bar Li}|}-\frac{1}{|x_1-x_{Mi}|}-\frac{1}{|x_1-x_{\bar Mi}|}\right)}
\nonumber
\ee
with $\mu_L-\mu_M=v_m$.
When averaging using the ensemble (\ref{MEASUREDD}) it is clear that each of the contributors to the string of factors in (\ref{L1X})
will match its analogue from the measure and re-exponentiate.
For instance the first contribution in (\ref{L1X}) re-exponentiates through the substitution
\be
e^{-b\pm i\sigma} \rightarrow &&e^{-b\pm i\sigma}e^{\frac{i}{2T|x_1-x|}}
\ee
The extra Coulomb factors can be re-defined away by shifting
\begin{eqnarray}
b\rightarrow b+\frac{i}{2T|x_1-x|}
\end{eqnarray}
thereby changing the constraint equation (\ref{DELTA}) to
\be
&&-\frac{T}{4\pi}\nabla^2w_M+f(e^{w_M-w_L}-e^{w_L-w_M})\nonumber\\
&&=\frac{T}{4\pi}\nabla^2(b-i\sigma)+\frac{i}{2}\delta^3(x-x_1)\nonumber\\
&&-\frac{T}{4\pi}\nabla^2 w_L+f(e^{w_L-w_M}-e^{w_M-w_L})=0
\label{SOLM}
\ee
and similarly for the second contribution in (\ref{L1X}) with $L\leftrightarrow M$.
The effect of the first contribution in the Polyakov line (\ref{L1X}) is to add a source
term to the constraint equation for $w_{M}$. It is in agreement with~\cite{DP}
after setting $b=\sigma=0$. (\ref{SOLM}) is a Poisson-Boltzmann type equation.
It is also referred to as an elliptic and periodic Toda lattice~\cite{DP,TODA}.
The solution is a local Debye-like cloud around the inserted heavy quark
\begin{eqnarray}
(w_M-w_L)(x)\approx \frac{2\pi i}{T}\int \frac{d^3p}{(2\pi)^3}\frac{e^{ip\cdot (x-x_1)}}{p^2+M^2}
\end{eqnarray}
This causes almost no change in the vacuum holonomies $v_{m,l}$. Thus, after the shift
\begin{eqnarray}
\left<L(x_1)\right>\approx e^{i2\pi\mu_M}+e^{i2\pi\mu_L}=0
\label{L00}
\end{eqnarray}
\subsection{$\left<LL^\dagger\right>$}
The preceding analysis can also be applied to the correlator of two heavy quarks through $LL^\dagger$
which consists now of 4 contributions before averaging
\begin{eqnarray}
L (x_1)L^\dagger(x_2)=\sum_{m,n=M,L}e^{2\pi i (\mu_{m}-\mu_{n})}e^{\frac{i}{2T}(F_{m}(x_1)-F_{n} (x_2))}\nonumber\\
\label{LL1}
\end{eqnarray}
with the pertinent Coulomb free energies $F_{m}(x_{1,2})$ following from (\ref{L1X}).
When averaged over the measure (\ref{MEASUREDD}), each of the factors in (\ref{LL1})
can be matched with its analogue in the measure. The preceding observations
show that the Coulomb factors in the probing correlator can be paired with
\be
e^{-b\pm i\sigma}\rightarrow e^{-b\pm i\sigma}e^{\frac{i}{2T|x-x_1|}-\frac{i}{2T|x-x_2|}}
\label{LL1X}
\ee
A rerun of the preceding arguments shows that the constraint equations acquire now two source contributions, one for each of the
heavy quark inserted
\be
&&-\frac{T}{4\pi}\nabla ^2w_M+f(e^{w_M-w_L}-e^{w_L-w_M})\nonumber \\
&&=\frac{T}{4\pi}\nabla^2(b-i\sigma)+\frac{i}{2} [\delta^3(x-x_1)-\delta^3 (x-x_2)] \nonumber\\
&&-\frac{T}{4\pi}\nabla^2w_L+f(e^{w_L-w_M}-e^{w_M-w_L})=0
\label{LL2}
\ee
Since $\nabla^21/|x-x_2|=-4\pi\delta^3(x-x_2)$,
we can symmetrize (\ref{LL2}) by shifting $\delta^3(x-x_2)$ from the first to the second equation
through
\begin{eqnarray}
w_{M,L}\rightarrow w_{M,L}+\frac{i/T}{2|x-x_2|}
\label{LL2X}
\end{eqnarray}
with unit Jacobian. The symmetrized (\ref{LL2}-\ref{LL2X}) equations are in agreement with those established in
\cite{DP} for the SU(2) case after setting $b=\sigma=0$. In this case the solution is peaked around the heavy quark sources
\begin{eqnarray}
(w_M-w_L)(x)\approx \frac{2\pi i}{T}\int \frac{d^3p}{(2\pi)^3}\frac{e^{ip\cdot (x-x_1)}-e^{ip\cdot (x-x_2)}}{p^2+M^2}
\end{eqnarray}
Inserting this back in the expectation value of the
correlator (\ref{LL1}) yields asymptotically
\begin{eqnarray}
\left<L (x_1)L^\dagger(x_2)\right>\approx e^{-M|x_1-x_2|}
\label{LLCONF}
\end{eqnarray}
in the 3-dimensional effective theory in agreement with the result in~\cite{DP}. In 4-dimensions
(\ref{LLCONF}) translates to confinement of the
electric charges with the electric string tension $\sigma_E=MT$.
The additional Coulomb screening in (\ref{MEASUREDD}) does not affect the asymptotics
of the linearly rising heavy quark potential to leading order.
The dyon-anti-dyon Coulomb liquid still electrically confines in the center symmetric phase.
\section{Single-Winding Wilson loop}
\label{WILSONSECTION}
To study the large spatial Wilson loops we use the same observations made above in the
presence of the U(1) fields $\sigma$ and $b$. As an observable the traced spatial Wilson loop
of area $S$ supported by the spatial contour $\partial S=C$ reads
\begin{eqnarray}
{\rm Tr}\,W(C)= e^{i\int_{S} B_{+}\cdot dS} +e^{i\int_{S} B_-\cdot dS}
\label{WC1}
\end{eqnarray}
and sources the static magnetic field
\begin{eqnarray}
B_{\pm\,\mu}= \pm\sum_i Q_{i} \frac{(x-x_i)_{\mu}}{|x-x_i|^3}
\label{WC2}
\end{eqnarray}
When averaged using (\ref{MEASUREDD}), the spatial Wilson loop (\ref{WC1}) modifies the
additional U(1) fugacity factors in the dyon sector. Their contribution follows again by shifting
$b\mp i\sigma\rightarrow b\mp i(\sigma-\eta_{\pm})$ in the constraint equations with
\begin{eqnarray}
\eta_{\pm}(x)=\pm \int_{S} dS_{y}\cdot\frac{x-y}{2|x-y|}
\end{eqnarray}
As a result, (\ref{DELTA}) in the presence of (\ref{WC1}) are now modified to read
\be
&&-\frac{T}{4\pi}\nabla^2w_M+f(e^{w_M-w_L}-e^{w_L-w_M})=\frac{iT}{4\pi} \nabla^2\eta_+(x) \nonumber
\\&&-\frac{T}{4\pi}\nabla^2w_L+f(e^{w_L-w_M}-e^{w_M-w_M})=0
\label{WC3}
\end{eqnarray}
for the first contribution and similarly
for the second contribution in (\ref{WC1}) with $\eta_+\rightarrow\eta_- $. After choosing the spatial Wilson loop to lay in the x-y plane
through $\nabla^2\eta_{\pm}=\pm 4\pi\delta^\prime(z)$,
the results (\ref{WC3}) are in agreement with those derived in \cite{DP} for the SU(2) case
but without the long range U(1) $\sigma$ and $b$ fields in the leading order approximation. Thus
$\left<{\rm Tr}\,W(C)\right>\approx e^{-\sigma_M \,S}$ is saturated by the pinned soliton,
with the magnetic string tension $\sigma_M=\sigma_E=MT$.
This result is expected from the equality of the
electric and magnetic masses in (\ref{SCREEME}).
A simple understanding of this result is as follows: while a heavy quark sources an electric field, a
large spatial Wilson loop sources a magnetic field by Ampere$^\prime$s law
which is classically composed of all the magnetic poles fluxing $S$
as is explicit in (\ref{WC2}). The typical contribution to (\ref{WC1})
for a planar surface in the xy-plane is then
\begin{eqnarray}
\left<e^{i\int_{S} B\cdot dS}\right>\approx e^{-\frac S2\int_S \left<B_z(x,y)B_z(0,0)\right>dS}
\end{eqnarray}
by keeping only the first cumulant in the average and using translational invariance for
large $S$. In this limit, $S$ acts as a uniformly charged magnetic sheet made of magnetic
dyons classically, so that
\begin{eqnarray}
\left<B_z(x,y)B_z(0,0)\right>\approx \left<\left(\frac{Q_M}{S}\right)^2\right>\approx \frac{\left<Q_M\right>}{S^2}
\end{eqnarray}
where the variance in the magnetic charge is assumed Poissonian. The magnetic charge density per unit
4-volume is $(TM)^2/2$. The typical magnetic charge per unit area is then about its square root or
$\left<Q_M\right>/S\approx TM$. Thus
\begin{eqnarray}
\left<e^{i\int_{S} B\cdot dS}\right>\approx e^{-\frac 12 MT\,S}
\end{eqnarray}
which is the expected behavior up to a factor of order one in the string tension.
\section{Double-Winding Wilson Loop}
Recently it was pointed out in~\cite{GREEN} that a co-planar and double winding Wilson loop in the SU(2)
pure gauge theory version of the model discussed by Diakonov and Petrov~\cite{DP} shows an exponentl
fall-off with the sum of the areas. In contrast lattice SU(2) simulations appear to show an exponential fall off
with the difference of the areas. The main observation in~\cite{GREEN} is that the solitonic configuration
contributing to the single-winding spatial Wilson loop as for instance from our linearized version with
$b=\sigma=0$ in (\ref{WC3}), factors out in the the double-winding and co-planar Wilson loop.
For two
identical loops with $C_1=C_2=C$, we have
the formal SU(2) identities~\cite{GREEN1} (and references therein)
\be
\left({\rm Tr}\, W(C)\right)^2 =&&{\rm Tr}_S\left(W(C)\right)+{\rm Tr}_A\left(W(C)\right)
\nonumber\\
{\rm Tr}\left(W(C)^2\right) =&&{\rm Tr}_S\left(W(C)\right)-{\rm Tr}_A\left(W(C)\right)
\label{DW1}
\end{eqnarray}
The simple trace ${\rm Tr}$ is carried over the fundamental representation of N-ality
$k=1$ as in (\ref{WC1}), and ${\rm Tr}_{S,A}$ are carried over the symmetric and anti-symmetric of N-ality $k=2$ (modulo 2)
representations of SU(2) respectively. The identities (\ref{DW1}) are commensurate with the Young-Tableau
decomposition. In the dyonic plasma considered here, the k-string tensions $\sigma_k$
in the linearized plasma approximation are
identical to those derived in~\cite{DP} with $\sigma_k/\sigma_1={\rm sin}\,{k\pi/2}$ for SU(2)
with $\sigma_1=\sigma_E$.
For $k=2$ we have $\sigma_2=0$ and the second identity in (\ref{DW1}) implies for large loops
\be
\left<{\rm Tr}\left(W(C)^2\right)\right> =\left<{\rm Tr}_S\left(W(C)\right)\right> -1
\label{DW1}
\ee
We have set all self-energies to zero for simplicity as they depend on the subtraction procedure.
(\ref{DW1}) is consistent with the doubly traced Wilson loop as dominated by the $k=2$ modulo
2 colorless di-quark-like $(qq)$ or baryon-like configuration in SU(2). In the dyonic plasma, the double
Wilson loop with $C_1=C_2$ is a bound colorless state with zero-size that is strongly
correlated within the dyons cores and therefore is consistent with the arguments presented
in~\cite{GREEN}.
For largely separated loops $C_{1,2}$ of arbitrary sizes but still lying in the spatial directions,
clearly
\begin{eqnarray}
\left<{\rm Tr}(W(C_1)W(C_2))\right>\approx e^{-\sigma_E(A_1+A_2)}
\end{eqnarray}
for $(A_1+A_2)<A_{12}$ where $A_{1,2}$ are the planar areas supported by $C_{1,2}$ separately,
and $A_{12}$ is the minimal area with boundaries $C_1$ and $C_2$.
The main issue is what happens for the same doubly wound SU(2)
spatial Wilson loops when $A_{12}<(A_1+A_2)\,? $ Here we note that $L\bar L$ and $M\bar M$
dimers carrying $(-2,0)$ and $(+2,0)$ ${\rm (electric, magnetic)}$ charge assignments could
cluster around the probe $qq$ (baryon) and $\bar q \bar q$ (anti-baryon) configurations respectively,
to form neutral molecular bound states with masses that scale with $A_{12}$ instead of $(A_1+A_2)$.
They are commensurate with the
massive off-diagonal and charged gluons Higgsed by the holonomy and dropped in the dyon liquid
analysis. These configurations were not retained in~\cite{DP}.
\section{t$^\prime$ Hooft Loops}
In an important study of the nature of confinement in gauge theories, t$^\prime$ Hooft~\cite{HLOOP} has introduced
the concept of a disorder operator or t$^\prime$ Hooft loop to quantify confinement in the Hilbert space
of gauge configurations. The t$^\prime$ Hooft loop is a canonical operator much like the Wilson loop. In
a Lorentz invariant confining vacuum, t$^\prime$ Hooft has argued that the temporal
Wilson loop and the t$^\prime$
Hooft loop cannot exhibit an area law simultaneously. The temporal Wilson loop obeys an area law, while the
t$^\prime$ Hooft loop obeys a perimeter law.
Physically, the Wilson loop corresponds to a color
charge in the fundamental representation running around a closed loop and measuring the
the chromo-magnetic flux across the loop. The t$^\prime$ Hooft loop corresponds to a dual charge
in the center of the gauge group running around a closed loop and
measuring the chromo-electric flux across the loop. The t$^\prime$ Hooft loop is the dual of the
Wilson loop.
In the temperature range $0.5\,T_c<T<T_c$ confinement is still at work and we expect the
temporal Wilson and t$^\prime$ Hooft loops to exhibit behaviors similar to those in the vacuum state.
In section~\ref{WILSONSECTION} we have explicitly checked that the closed spatial Wilson loop
obeys an area law. The temporal Wilson loop is not amenable to our dimensionally reduced
and Euclideanized effective field theory.
The t$^\prime$ Hooft loop $V(C)$ enforces a gauge transformation $\Omega_C$ which is singular
on a closed curve $C$. If a curve $C^\prime$ winds $n_{CC^\prime}$ times around $C$ then
\be
V^\dagger (C)W(C^\prime)V(C)=e^{i\frac {2\pi}{N_c}n_{CC^\prime}}W(C^\prime)
\label{TH1}
\ee
$V(C)$ amounts to a multi-valued gauge transformation on the loop $C$,
\begin{eqnarray}
\Omega_C (\theta = 2\pi )=e^{i\frac {2\pi}{N_c} n_{CC^\prime}}\Omega_C (\theta =0)
\label{TH0}
\end{eqnarray}
with $\theta$ an affine parameter along $C$. A simple choice is
\begin{eqnarray}
\Omega_C(x)=e^{i\frac {2\pi}{N_c}Q\varphi_C(x)}
\label{TH2}
\end{eqnarray}
where $\varphi_C(x)$ is a multi-valued scalar potential for the magnetic field $\vec B_C$ generated by a loop
of current $\vec j_C$ running along $C$, and $Q=(1, 1, ..., -N_c+1)$ a Cartan generator of SU(N$_c$). An alternative
construction using a discontinuous solid angle
was discussed in~\cite{KOSTAS,HUGO}. The effects of (\ref{TH2}) on an Abelianized Wilson loop is
\begin{eqnarray}
\Omega_C^\dagger \left(e^{i\int_{C^\prime}\,ds\cdot A}\right) \Omega_C
=e^{i\int_{C^\prime} ds\cdot (A-\frac {2\pi}{N_c}QB_C)}
\label{TH3}
\end{eqnarray}
with $\vec B_C=-\vec\nabla\varphi_C$. Note that since $\varphi_C$ is {multivalued} we have
$\vec\nabla\times \vec B_C=4\pi \vec j_C$. If we normalize the loop current $\vec j_C$ such that
\begin{eqnarray}
\int_{C^\prime}ds\cdot B_C=4\pi\int_{A(C^\prime)}dS\cdot j_C=-n_{CC^\prime}
\label{TH4}
\end{eqnarray}
then (\ref{TH3}) reduces to
\begin{eqnarray}
\Omega_C^\dagger \left(e^{i\int_{C^\prime}\,ds\cdot A}\right) \Omega_C
=e^{i\frac {2\pi}{N_c}n_{CC^\prime}}e^{i\int_{C^\prime}\,ds\cdot A}
\label{TH5}
\end{eqnarray}
In the space of gauge configurations, the gauge transformation $\Omega_C$ is inforced through
\begin{eqnarray}
V(C)=e^{i\frac {2\pi}{gN_c} \int d^3x {\rm Tr}\left( E_i D_i (Q\varphi_C)\right)}
\label{TH6}
\end{eqnarray}
For SU(2) we have
\begin{eqnarray}
V(C)=e^{i\frac{2\pi}{g} \int d^3x \vec{E^3}\cdot \vec B_C}\rightarrow e^{-\frac{2\pi}{g} \int d^3x \vec{E^3}\cdot \vec B_C}
\label{TH7}
\end{eqnarray}
where the latter substitution
$E\rightarrow iE$ follows in Euclidean space. With this in mind, the expectation value
of the t$^\prime$ Hooft loop in the dyonic ensemble involves a string of sources to be inserted in
(\ref{SU2}). In leading order
\begin{eqnarray}
V(C)\rightarrow &&\prod_{i=1}^{N+\bar N}
e^{\frac{2\pi}{g}\int d^3x\,B_C\cdot\nabla\frac {Q_{Ei}}{|x-x_i|}}\nonumber\\
=&&\prod_{i=1}^{N+\bar N}
e^{-\frac{2\pi}{g}\int d^3x\,\nabla\cdot B_C\frac {Q_{Ei}}{|x-x_i|}}=1
\end{eqnarray}
Thus $\left<V(C)\right>=1$ modulo ${\cal O}(\alpha_s)$ Coulomb-like self-energy corrections which are
perimeter-like in general.
Finally, the Polyakov line as a Wilson loop around the periodic temporal direction
has a dual Polyakov loop with a dual magnetic charge in the center. In the confined phase,
the temporal component of the gauge field $A_4$ asymptotes fixed electric-type holonomies, while its
dual $\tilde{A}_4$ asymptotes zero dual magnetic-type holomies thanks to parity. A rerun of the arguments
in~\ref{LSECTION} shows that while $\left<L({\bf x})\right>=0$ in (\ref{L00})
as expected in an Euclidean and confining thermal state,
its dual does not vanish, i.e.
\begin{eqnarray}
\left<{\tilde L}({\bf x})\equiv {\rm Tr}\left(e^{i\frac{4\pi}{gN_c}\int_0^\beta Q\tilde{A}^Q_4(x) d\tau}\right)\right>=1
\label{HH1}
\end{eqnarray}
again modulo ${\cal O}(\alpha_s)$ Coulomb corrections. This behavior is consistent with the one reported on the lattice
for $N_c=2,3$~\cite{DIGIA}.
\section{Conclusions}
The central theme in this paper is non-perturbative gauge theory for temperatures in the range $(0.5-1)\,T_c$
modeled by a dense plasma of instanton-dyons. The new element in our discussion
is the introduction of the leading classical ${\cal O}(1/\alpha_s)$
interactions between the dyons and anti-dyons as recently obtained in~\cite{LARSEN-SHURYAK}
using the classical ``streamline" set of configurations
for $M\bar{M},L\bar{ L}$ pairs. We have assumed that
the $M\bar{ L}, L\bar{ M}$ channels are repulsive and opposite in sign to
the streamline interaction. While carrying this work, this assumption has now been confirmed
numerically~\cite{PRIVATE}. Another important element of our analysis is the one-loop measure of the
dyon and anti-dyon moduli space, in the form proposed by Diakonov and Petrov \cite{DP}.
It leads to a small moduli space volume and thus repulsive interaction at higher density,
which however can be made much less repulsive by introducing correlaltions between the charges.
On general grounds, an ensemble of instanton-dyons is a strongly coupled
plasma, with significant correlations between the particles. Therefore, the statistical mechanics
of a generic instanton-dyon ensemble is very nontrivial and remains unsolved.
However -- and this is the main argument of the paper -- when the plasma is dense enough
for temperatures below $T_c$, it generates a
large screening mass $M$ which screens the interaction. A standard weak coupling
plasma theory, in a form similar to the Debye-Huckel theory is then applicable . The dimensionless
3-density of each dyon species $n_D/4$
in the regime considered is in the range of $n_D/4\approx T^3/4$, in agreement with the
qualitative arguments in~\cite{SHURYAK}.
Using it, we get a number of results concerning the details of the non-perturbative gauge fields,
in the temperature range $(0.5-1)\,T_c$. First , in the presence of strong screening the minimum of the
free energy is still at the confining (center symmetric) value of $\nu=1/2$, with
a vanishing Polyakov line $\left<L\right>\approx {\rm cos}(2\pi\nu)=0$ . Second, a re-summation of the the
linearized screening effects yields Debye-Huckel type corrections to
the pressure and dyonic densities. We have also analyzed the topological susceptibility,
the gluonic compressibility, and the electric and magnetic gluonic condensates in this linearized approximation.
We have calculated also the electric and magnetic screening masses, generated by the dyon ensemble.
We have found that the latter are larger than the former in the confined phase. This is qualitatively consistent with the existing lattice data,
which however are much better measured for the SU(3) gauge theory rather than the SU(2) one we have studied here.
Finally, we have calculated the structure factors
in the electric and magnetic sector in the linearized screening approximation as well.
For an estimate of the transition temperature from $\nu =1/2$ (confinement) to $\nu =0$ (deconfinement)
we have switched the perturbative (GPYW) holonomy potential~\cite{WEISS}
in section~\ref{sec_deconf}. For SU(2) the transition is observed to take place at $T_c/\sqrt{\sigma_E}\approx 0.88$.
In the dyonic plasma the large spatial Wilson loops exhibit area law, while the spatial t$^\prime$ Hooft loops
are found to be 1 modulo ${\cal O}(\alpha_s)$ Coulomb-like self-energy corrections. These dual behaviors
were argued in~\cite{HLOOP} for confining gauge theories at zero temperature. We found them to hold
in the confining dyon-ensemble in the regime $0.5<T<T_c$.
Needless to say, that all these predictions can and should be confronted with the lattice data
in the corresponding temperature range.
Finally, let us speculate about the dyon ensemble beyond the validity domain of the Debye-Huckel
approximation. First of all, strongly coupled Coulomb plasmas are tractable by certain analytic and/or
numerical (molecular dynamics) methods, see Refs~\cite{DH,CHO} for similar development.
Another option is to use brut force numerical simulations of the dyon ensemble \cite{LS_stat}.
Qualitatively, sufficiently strongly coupled plasmas develop either (i) correlations between particles,
resembling either a
liquid with crystal-like correlations (``molten salt"), or (ii) particular neutral clusters, the simplest of which
can be the $LM$ instantons themselves or $LM\bar{L}\bar{M}$ ``instanton molecules". Recent (unquenched) lattice simulations indicate that the instantons and
anti-instantons recombine into topologically neutral molecules across the transition temperature
~\cite{SHARMA,MOLECULE}. At much higher temperature,
the perturbative gluons dwarf all classical gauge configurations forcing the holonomy to zero.
One obvious extension of this work should be into the large number of colors
$N_c$. Strong correlations can appear, since $\Gamma_{D\bar D}\approx 1/\alpha_s\approx N_c\gg 1$.
Similar mechanism, leading to crystallization
appears to take place in dense holographic matter where the baryons as instantons
in the holographic direction split into a pair of dyons and re-arrange in salt crystals~\cite{SIN}.
Another obvious extension of this work is to include fermions, which we turn to in the second paper of the series~\cite{SECOND}.
\section{Acknowledgements}
We would like to thank Jeff Greensite, Tin Suleijmanpasic, Mithat Unsal and Ariel Zhitnitsky
for their comments on the manuscript after it was posted.
This work was supported by the U.S. Department of Energy under Contracts No.
DE-FG-88ER40388.
|
\section{Introduction}
\label{sec:intro}
Our systems today are vulnerable to Sybil attacks, in which an attacker injects multiple fake accounts into the system to compromise security and privacy~\cite{sybil}.
Recently, the increasing popularity of online social networks have made them attractive targets for Sybil attacks. It is estimated that tens of millions of Sybil accounts exist in popular social networks such as Twitter and Facebook~\cite{Facebooksybil}~\cite{benevenuto2010detecting}.
Attackers can leverage Sybil accounts to compromise system security via propagating social malware, as well as system privacy via learning users' private information~\cite{Facebooksybil}-\cite{Thomas11-imc}
An important thread of research proposes to mitigate Sybil attacks using social network-based trust relationships. The key insight of this line of defense is that it is hard for attackers to establish trust relationships with benign users. That is, the number of edges between benign users and Sybil identities (called attack edges) is limited. Systems such as SybilGuard~\cite{Yu06}, SybilLimit~\cite{Yu08}, SybilInfer~\cite{Danezis09}, SybilRank~\cite{sybilrank}, and SybilBelief~\cite{Gong13} exploit the limited number of attack edges to detect Sybil identities using graph-theoretic techniques.
While these systems as well as related works have pioneered the use of social network structure for Sybil defense, the actual deployment of these ideas in real world networks remains controversial. Yang et al.~\cite{Yang11-sybil} showed that network structure-based Sybil defenses failed in identifying Sybil
accounts in RenRen, a popular social network in China. This is because structure-based defense mechanisms make assumptions of \emph{strong trust} relationships between users, such that the number of attack edges is limited~\cite{Yu06}-\cite{Gong13}. These assumptions do not hold in networks with weak trust
relationships, which enables an adversary to create a large number of attack edges. Ghosh et al.~\cite{Ghosh12} showed that on Twitter, a link farming phenomenon is wide spread and poisonous, in which certain benign accounts blindly accept follow requests. Thus, in such weak-trust social networks,
previous structure-based Sybil defenses have limited applicability and performance.
In this paper, we focus on the problem of mitigating Sybil attacks in social networks with weak trust, i.e., when the number of attack edges is large. We propose SybilFrame, an approach that provides defense-in-depth against Sybil attacks. SybilFrame uses a multi-stage classification mechanism that is able to incorporate heterogeneous sources and types of information about the social network. In the first stage, SybilFrame leverages fine grained local information about
users and edges in the social network to design classifiers for predicting whether users or edges are benign or malicious. In the second stage, SybilFrame combines information from local classifiers with global structural properties of social networks (even ones with weak trust properties).
Our approach leverages the results of local classification about users and edges as prior probabilities in a pairwise Markov Random Field model~\cite{Cross83},
and uses Loopy Belief Propagation~\cite{Murphy99} to make probabilistic inferences.
We experimentally evaluate the performance of SybilFrame using both synthetic and Facebook
network topologies.
We show that local node classifiers that are better than random (e.g., false positive/negative rates as high as $40\%$), can significantly improve the Sybil detection accuracy when combined with global structural information. Similarly, local edge classifiers with even a small predictive capability, provide synergistic
information to global structural inference, and improve detection accuracy. Our approach is resilient to seed targeting attacks and a high number of attack edges which are common in social networks with weak trust.
We test SybilFrame on a large scale Twitter dataset with over $20M$ nodes and $265M$ edges. We obtain information about which accounts in this dataset were suspended by Twitter, and use this as ground truth for Sybil attacks. This dataset is typical of social networks with weak social trust, as the attacker has more than $18M$ attack edges for about $145,000$ Sybil identities. Even in this challenging setting with very large number of attack edges, SybilFrame is able to detect 51\% Sybil identities with $4.2\%$ false positives, with an overall accuracy of $95.4\%$.
In contrast, state-of-the-art approaches such as SybilBelief predict all nodes to be Sybil and thus completely fail on this dataset.
SybilFrame can also be used as a mechanism to rank user accounts. In the top 1K accounts ranked by SybilFrame (in increasing order of being benign), SybilFrame identifies $55\%$ Sybil accounts, which is 1-2 orders of magnitude better than state-of-the-art approaches. Furthermore, we manually examine the profile of the top 100 ranked users, of which 71 are suspended and 29 are active, and find that 24 active accounts are highly likely to be malicious. Thus, SybilFrame is able to uncover a large fraction (24/29) of suspicious accounts that Twitter fails to detect.
\section{Background}
\label{sec:background}
First, we give a formal definition of the Sybil defense problem in online social systems, and discuss state-of-the-art approaches.
Then, we introduce our design goals.
\subsection{Sybil Defense in Online Social Systems}
Consider a network topology $G = (V, E)$, comprising a set $V$ of nodes with a set $E$ of edges. In social network topologies, a node $v \in V$ denotes a user on the network, and an edge $(u, v) \in E$ denotes a friendship relationship between two users $u$ and $v$. Here we only consider mutual relationships, hence $(u, v) \in E$ is equivalent to $(v, u) \in E$ and $G$ is an undirected graph. Every node $v \in V$ in the network is either a benign node, or a Sybil identity. %
Figure~\ref{fig:sybil_attack} depicts the \emph{Sybil attack} problem. We denote the subnetwork containing all benign nodes to be the \emph{benign} region, and denote the subnetwork containing all Sybil nodes to be the \emph{Sybil} region. The edges that connect the benign region and the Sybil region are called \emph{attack edges}. Following the established convention in the literature, we do not impose any constraints on the size or the shape of the Sybil region. Attackers can create an unlimited number of Sybil nodes and set up edges between them arbitrarily. The main goal of Sybil defense is to design a mechanism to detect as many Sybil nodes as possible, while minimizing the number of benign nodes that are misdetected, i.e., a low false positive rate.
\begin{figure}[!htp]
\vspace{-0.4cm}
\centering
\includegraphics[width=0.3\textwidth]{sybil_attack.pdf}
\caption{Sybil attack problem.}
\label{fig:sybil_attack}
\vspace{-0.5cm}
\end{figure}
\subsection{State-of-the-art Approaches}
\myparatight{Content-based approaches} Content-based approaches seek to filter Sybil accounts by
analyzing the associated content information, such as news feeds and wall posts on Facebook and tweets and hashtags on Twitter~\cite{Thomas11-imc}. These approaches span a large category of mechanisms, including blacklisting, whitelisting, URL filtering, as well as various machine learning methods, such as Bayesian Reasoning, Support Vector Machines and Clustering~\cite{Wang10}-\cite{spam:acsac10}.
A major problem of these approaches is that attackers can mimic the behaviors of benign users and produce similar content
information, thus making content-based approaches less effective.
\myparatight{Structure-based approaches} Structure-based approaches,
seek to exploit graph-theoretic differences between benign and Sybil identities. %
The key insight is that in a social graph where edges represent strong trust relationships between users, it is hard for attackers to set up links to benign users. %
As a result, the number of attack edges is relatively small. Such networks preserve a strong level of homophily, i.e., two linked nodes are
likely to have similar attributes.
SybilGuard~\cite{Yu06} and SybilLimit~\cite{Yu08}, rely on the insight that it is easy for short random walks starting from a benign user to quickly reach other benign users, while hard for random walks starting from Sybil identities to enter into the benign region.
SybilInfer~\cite{Danezis09}, relies on random walks and a combination of Bayesian inference and Monte-Carlo sampling and aims to directly detect the bottleneck cut between benign and Sybil identities. SybilRank~\cite{sybilrank},
uses short random walks to distribute initial scores from a set of trusted benign seeds, and rely on the insight that benign users tend to have larger degree-normalized scores than Sybil identities. Criminal account Inference Algorithm (CIA)~\cite{Yang12-spam}, similar to SybilRank, starts random walks and distributes scores from Sybil seeded users and allows the restart from initial probability distribution with certain probability. Researchers have shown that despite considerable differences, the above schemes rely on identifying local communities around a trust node~\cite{Viswanath10}. SybilBelief~\cite{Gong13}, on the other hand, models the distribution of labels of the nodes as a pairwise Markov Random Field. Similar to SybilFrame, it adopts Loopy Belief Propagation to estimate probabilities of users being benign.
\'{I}ntegro~\cite{integro}, is an extension to SybilRank by incorporating victim predictions using content features, thus not purely structure-based.
We note that all of the above-mentioned structure-based methods are based on two key assumptions. First, the benign region is fast mixing~\cite{mohaisen:imc10}, which presumes the existence of a well-connected, giant community structure of benign users. Second, the social network is a strong trust network, where the number of attack edges is relatively small~\cite{Viswanath10}. Given the two assumptions, these structure-based approaches have been shown to provide reliable performance.
\subsection{Assumptions vs. Reality}
We claim that the above mentioned assumptions oversimplify social network structure, and do not hold well on all real-world social graphs.
First, benign users tend to form multiple small communities~\cite{Viswanath10} driven by different purposes (e.g., geographical location, education and career). This multi-community structure prohibits the existence of a giant community component and hence results in a longer mixing time. Mohaisen et al. \cite{mohaisen:imc10} measured the mixing time of real-world social graphs and found that the actual mixing time is longer than the theoretical anticipated value.
Second, real-world social networks may not necessarily represent strong trust networks. Yang et al. showed that RenRen, the largest social networking platform in China, does not follow this assumption~\cite{Yang11-sybil}. Another typical example is the Twitter network. The Twitter network is a directed network, on which links are established by the action of ``follow". Unlike Facebook, users in Twitter often use a pseudonym, which makes them less serious about whom they choose to follow.
Ghosh et al.~\cite{Ghosh12} showed that on Twitter, the notable phenomenon of link farming is wide spread, and that a majority of attack edges are farmed from a small fraction of Twitter users. Those users, the social capitalists, are benign users who are seeking to increase their social power and links by following back anyone who follows them. Even normal users, who are not as athirst for social power as social capitalists, are also likely to follow back strangers because they want to read their tweets or just by courtesy. On such weak trust social networks like Twitter, a large number of attack edges exist and the benign region may not be easily separable from the Sybil region. As a result, all of these structure-based Sybil defense mechanisms are limited in their
performance.
\subsection{Design Goals}
We aim to design a scheme that works even when the fast-mixing and strong trust assumptions are relaxed.
Our overall design goals are as follows:
\textbf{1) Defense-in-depth:} The scheme should provide multi-layered protection, and be robust to different attack strategies.
\textbf{2) Accuracy:} The scheme should have reliable detection accuracy when applied to a wide range of social network topologies, including both strong trust and weak trust social networks.
\textbf{3) Scalability:} The scheme should be scalable to large social networks, and be amenable to parallel deployment.
We propose SybilFrame, a defense-in-depth framework that adopts a multi-stage classification mechanism for incorporating heterogeneous sources and types of information about the social network.
\section{The SybilFrame Framework}
\label{sec:framework}
In this section, we give a detailed description of SybilFrame framework.
\subsection{Framework Overview}
\begin{figure}[!htp]
\centering
\includegraphics[width=0.49\textwidth]{framework.pdf}
\caption{SybilFrame framework.}
\label{fig:sybilframe}
\vspace{-0.3cm}
\end{figure}
Figure~\ref{fig:sybilframe} shows the general framework of SybilFrame. SybilFrame is a multi-stage classification approach that leverages the attributes of an individual node and correlation between connected nodes to make a combined classification of networked data. SybilFrame has two stages of inference. Once the raw data has been fed into the framework, \emph{Stage 1} will explore the dataset and extract useful information, to compute node prior information and edge prior information (Section~\ref{subsec:prior}). This prior information, together with a small set of nodes whose labels are known, i.e., trust seeds, will be fed into \emph{Stage 2}. \emph{Stage 2} is the posterior inference layer. To represent the correlation between nodes, we model the problem as a pairwise Markov Random Field (Section~\ref{subsec:mrf}). We adopt Loopy Belief Propagation (Section~\ref{subsec:lbp}) to make inferences about the posterior information. This posterior information will then be used to classify and rank Sybil identities (Section~\ref{subsec:classify}).
\subsection{Prior Information}
\label{subsec:prior}
\emph{Stage 1} in Figure~\ref{fig:sybilframe} takes the raw dataset as input, and outputs the prior information of all nodes and edges. We now formalize our notion of priors.
For a node $v \in V$, we denote $Prior_v$ as the node prior of $v$. $Prior_v$ is a real number in the range $[0, 1]$, that quantifies the probability that node $v$ takes a benign label. The larger $Prior_v$ is, the more likely that $v$ is a benign node. Specifically, $Prior_v > 0.5$ means that $v$ is more likely to take a benign label rather than a Sybil label. Similarly, $Prior_v < 0.5$ means that $v$ is more likely to take a Sybil label, and $Prior_v = 0.5$ means that $v$ takes a benign or Sybil label with equal probability. If $v$'s label is known, then $Prior_v = 1$ for a benign trust seed, and $Prior_v = 0$ for a Sybil trust seed.
For two nodes $u$ and $v$ that are connected by an edge, we denote $Prior_{u, v}$ as the edge prior of $(u, v) \in E$. $Prior_{u, v}$ is a real number in the range $[0, 1]$, that quantifies the likelihood that node $u$ and node $v$ take the same label. The larger $Prior_{u, v}$ is, the more likely that $u$ and $v$ take the same label. Specifically, $Prior_{u, v} > 0.5$ means that $u$ and $v$ are more likely to take the same label than different labels. Similarly, $Prior_{u, v} < 0.5$ means that $u$ and $v$ are more likely to take different labels, and $Prior_{u, v} = 0.5$ means that $u$'s label has no influence on $v$'s label, and vice versa. Generally, $Prior_{u, v}$ models the level of coupling strength between $u$ and $v$. $Prior_{u, v} > 0.5$ refers to a positive coupling relationship, and $Prior_{u, v} < 0.5$ refers to a negative coupling relationship, and $Prior_{u, v} = 0.5$ means that there is no coupling between $u$ and $v$.
The derivation of node priors and edge priors is based on the dataset we are given. We can leverage heterogeneous information sources to make inferences. To compute node priors, we can leverage the structural information and explore differences of local structure between benign and Sybil nodes. We can extract useful features and build a machine learning classifier that supports probability estimates, and use these probability outputs as node priors.
To infer edge priors, we want to assign lower scores to attack edges, and assign higher scores to edges between benign accounts. We do not care about edges between Sybil accounts, since attacker has a complete control over the Sybil region and can change it arbitrarily. This makes our approach robust to high number of attack edges and distinguishes SybilFrame from previous approaches. Since benign nodes tend to behave similarly and Sybil nodes tend to behave differently from benign nodes, a straightforward way is to explore similarities of two connected nodes under different metrics and obtain a scaled overall similarity score. This overall score can then be used as an edge prior.
We note that although we propose a structure-based scheme, as will be demonstrated and evaluated later, our framework can definitely incorporate content information. For example, we can analyze news feeds of each Facebook account and tweets of each Twitter account, and identify spam keywords and abnormal actions. We can then build a content-based classifier and compute node priors. The philosophy also works for content-based edge priors. The reason why we tend to use structural information is that it is harder for an attacker to alter the overall graph structure than mimic the content behaviors of benign users.
In Section~\ref{sec:evaluation_facebook} and Section~\ref{sec:twitter}, we will explore ways to compute node priors and edge priors on real-world, large-scale social graphs.
\subsection{Markov Random Field}
\label{subsec:mrf}
A Markov Random Field (MRF)~\cite{Cross83}, is a probabilistic graphical model over an undirected graph.
Nodes in MRF are random variables, and edges are used to model correlation between those random variables. For each node $v \in V$ on graph $G = (V, E)$, we associate it with a binary random variable $X_v$, that indicates the label of $v$. $X_v = 1$ refers to a benign label, and $X_v = -1$ refers to a Sybil label. To quantify the correlation, we use a set of functions called \emph{clique potentials}. A \emph{clique potential} is a function defined over a set of random variables, which maps any joint assignment of these random variables to a real number, which indicates how favorable this joint assignment is. Let $\Psi$ denote the set of potential functions. Specifically, if we only consider cliques comprising at most two connected nodes, $\Psi$ can be divided into the following two types of functions.
\begin{equation}
\setlength{\abovedisplayskip}{2pt}
\setlength{\belowdisplayskip}{2pt}
\psi_v(X_v)=\begin{cases}
Prior_v, & \text{if }X_v = 1 \\
1 - Prior_v, & \text{if }X_v = -1
\end{cases}
\end{equation}
\begin{equation}
\setlength{\abovedisplayskip}{2pt}
\setlength{\belowdisplayskip}{2pt}
\psi_{u, v}(X_u, X_v)=\begin{cases}
Prior_{u,v}, & \text{if }X_uX_v = 1 \\
1 - Prior_{u, v}, & \text{if }X_uX_v = -1
\end{cases}
\end{equation}
As defined in Section~\ref{subsec:prior}, $Prior_v$ is the prior information of node $v$, and $Prior_{u, v}$ is the prior information of edge $(u, v)$. We denote function $\psi_v$ as the \emph{node potential}, and function $\psi_{u, v}$ as the \emph{edge potential}. $(G, \Psi)$ then defines a \emph{pairwise Markov Random Field}.
Given a pairwise MRF $(G, \Psi)$, where $G = (V, E)$ and $\Psi = (\psi_v, \psi_{u, v})$, the full joint probability distribution is specified as
\begin{equation}
\setlength{\abovedisplayskip}{2pt}
\setlength{\belowdisplayskip}{2pt}
P(X_V) = \frac{1}{Z}\prod_{v \in V}\psi_v(X_v) \prod_{(u, v) \in E} \psi_{u, v}(X_u, X_v)
\end{equation}
Here, $X_V$ denotes a particular joint assignment of all random variables in set $V$, and $Z$ is the partition function given by
\begin{equation}
\setlength{\abovedisplayskip}{2pt}
\setlength{\belowdisplayskip}{2pt}
Z = \sum_{X_V}\prod_{v \in V}\psi_v(X_v) \prod_{(u, v) \in E} \psi_{u, v}(X_u, X_v)
\end{equation}
\subsection{Infer Posteriors}
\label{subsec:lbp}
Given the pairwise MRF $(G, \Psi)$, which contains prior information of trust seeds and other nodes and edges, for each node $v\in V$, we want to infer the posterior probability of random variable $X_v$.
\begin{equation}
\setlength{\abovedisplayskip}{2pt}
\setlength{\belowdisplayskip}{2pt}
P(X_v) = \frac{1}{Z}\sum_{X_{V\backslash v}}\prod_{s \in V}\psi_s(X_s) \prod_{(u, s) \in E} \psi_{u, s}(X_u, X_s)
\end{equation}
Exact inference is computationally difficult, and not scalable on large dataset. Therefore, we adopt \emph{Loopy Belief Propagation} to make approximate inferences. Loopy Belief Propagation~\cite{Murphy99} is an iterative process in which neighboring variables pass messages or beliefs to each other. Algorithm~\ref{alg:lbp} gives the Loopy Belief Propagation algorithm for the pairwise MRF $(G, \Psi)$.
\begin{algorithm}
\DontPrintSemicolon
\KwData{node potentials $\psi_v(X_v)$, edge potentials $\psi_{u, v}(X_u, X_v)$}
\KwResult{marginal beliefs $bel_v(X_v)$}
Initialize beliefs $bel_v(X_v) = 1$ for all nodes $v$\;
Initialize message $m_{u\rightarrow v} (X_v)= 1$ for all edges $u\rightarrow v$\;
\Repeat{
number of iterations $>$ threshold $d$}{
Messages update $m_{u\rightarrow v} (X_v) = \sum_{X_u}\left( \psi_u(X_u)\psi_{u, v}(X_u, X_v)\prod_{s\in Nbd(u)\backslash v}m_{s \rightarrow u}(X_s) \right)$\;
Beliefs update $bel_v(X_v) \propto \psi_v(X_v)\prod_{u\in Nbd(v)}m_{u\rightarrow v}(X_v)$\;
}
\caption{Loopy Belief Propagation Algorithm}\label{alg:lbp}
\end{algorithm}
We note that for social networks with loops, LBP approximates the posterior probability distribution without theoretical convergence guarantees. However, LBP has been widely used and demonstrated good results in practical applications~\cite{Murphy99}. Through our experiments, we find that setting $d$ to be 5$\sim$6 achieves good results.
\myparatight{Scalability} The complexity of LBP is $O(md)$ where $m$ is the number of edges and $d$ is the number of iterations. For sparse social networks, $O(md) = O(nd)$, where $n$ is the number of nodes. LBP is essentially parallelizable, and we will discuss related implementation issues in Section~\ref{subsec:bp_parallel}.
\subsection{Sybil Accounts Prediction and Ranking}
\label{subsec:classify}
We use posteriors obtained in Section~\ref{subsec:lbp} to predict the label of each node. For a node $v$ whose label is unknown, we predict the label $L_v$ using the following rule.
\begin{equation}
\setlength{\abovedisplayskip}{2pt}
\setlength{\belowdisplayskip}{2pt}
L_v = \sign(bel_v - 0.5)
\end{equation}
\noindent where $L_v = 1$ means that $v$ is predicted as a benign node, and $L_v = -1$ means that $v$ is predicted as a Sybil node.
We can also rank nodes in ascending order of its posterior, and produce a ranking list. Sybil nodes are likely to have lower posteriors, thus occur more in the front part. OSN operators can then go through the list from the beginning, and check a fixed number of nodes. More effective posteriors will let OSN operators detect more Sybil accounts within a certain amount of time.
\section{Security Evaluation on Synthetic Networks}
\label{sec:evaluation_synthetic}
In this section, we evaluate SybilFrame on different network structures. For comparison, we use SybilBelief~\cite{Gong13}, which
takes a similar probabilistic inference approach as SybilFrame. Since Gong et al.~\cite{Gong13} have demonstrated that SybilBelief outperforms other structure-based methods on trust networks, we limit our space here to only compare with SybilBelief. Later in Section~\ref{sec:evaluation_facebook} and Section~\ref{sec:twitter}, we will compare with other methods such as SybilLimit~\cite{Yu08}, SybilInfer~\cite{Danezis09}, and SybilRank~\cite{sybilrank}. We do not compare with \'{I}ntegro~\cite{integro} since it leverages network-specific content information for victim predictions.
\myparatight{Basic experimental setup} We adopt the Preferential Attachment (PA)~\cite{Barabasi99} model to generate both benign region and Sybil region. The size of benign region is 1000, and the size of Sybil region is 400. The average degree of both benign region and Sybil region is 10. We randomly add 1000 attack edges between the two regions. We only use 1 benign trust seed and 1 Sybil trust seed. For default node priors, we set 0.9 for benign trust seeds, and 0.1 for Sybil trust seeds, and 0.5 for others if we do not have any external priors fed in. For default edge priors, we set it to 0.9 in order to model homophily.
We will study the impact of different factors. When we study one factor, we fix the other factors to be the same as in the basic setup, and only vary the studied one. Under each setting, we run 100 experiments. In each experiment, we randomly generate 1 benign trust seed and 1 Sybil trust seed, configure prior information, and run SybilFrame and SybilBelief. We store results of SybilFrame and SybilBelief correspondingly, and take the average over 100 experiments in the end to be our final results.
\myparatight{Evaluation metrics} Following the convention, we denote Sybil nodes as positive examples and benign nodes as negative examples. Thus, we have TP (Sybil $\rightarrow$ Sybil), TN (benign $\rightarrow$ benign), FP (benign $\rightarrow$ Sybil) and FN (Sybil $\rightarrow$ benign). We use the following four evaluation metrics:
\textbf{1) Accuracy:} $(TP+TN)/(TP+TN+FP+FN)$
\textbf{2) Number of rejected benign nodes:} $FP$
\textbf{3) Number of accepted Sybil nodes:} $FN$
\textbf{4) Area Under the Receiver Operating Characteristic Curve (AUC)~\cite{auc}:} The probability that a randomly selected benign nodes ranks higher than a randomly selected Sybil node, given the ranking of posteriors of all nodes from the smallest to the largest.
\subsection{Influence of Node Priors}
\label{subsec:node_prior}
We want to explore SybilFrame when only incorporating external node priors. Since we are experimenting with synthetic networks, we need to figure out a way to obtain node priors that are able to model the real case. A straightforward solution is to use false positive rate (FPR) and false negative rate (FNR) to model the performance of an external node classifier. By setting up different FPR and FNR, we can generate prior scores that are able to model the level of noise, and use them to evaluate SybilFrame. Due to limited space, we list our \emph{Node Prior Generator} algorithm (Algorithm~\ref{alg:node_prior}) in Appendix~\ref{subsec:prior_generator}.
In Algorithm~\ref{alg:node_prior}, we set prior for benign/Sybil trust seeds to be 0.9/0.1, not 1/0 as discussed in Section~\ref{subsec:prior} in order to run LBP successfully.
\myparatight{Varying FPR and FNR}
\label{subsubsec:fpr_fnr_f}
First, we evaluate SybilFrame given node priors with different levels of noise. We tune $FPR=FNR$ from 0 to 0.5, i.e., from perfect classification to random guess, and fix everything else as in the basic setup. In addition to comparison with SybilBelief, we also compare with the performance of external node classifier, i.e., compare with priors, and explore whether there are improvements.
Figure~\ref{fig:node_prior_1} shows the results. As $FPR=FNR$ increases, the performance of external node classifier degrades linearly.
Besides, SybilFrame performs better than SybilBelief when $FPR=FNR\leq 0.4$, in terms of all four metrics. When $FPR=FNR\leq 0.3$, SybilFrame can achieve near optimal performance. This means that SybilFrame is resilient to prior noise with FPR and FNR as high as 40\%.
\begin{figure}
\begin{subfigure}[H]{0.24\textwidth}
\includegraphics[width=\linewidth]{node_prior_Accuracy_1.pdf}
\caption{Accuracy}
\label{fig:node_prior_accuracy_1}
\end{subfigure}%
\begin{subfigure}[H]{0.24\textwidth}
\includegraphics[width=\linewidth]{node_prior_AUC_1.pdf}
\caption{AUC}
\label{fig:node_prior_auc_1}
\end{subfigure}%
\begin{subfigure}[H]{0.24\textwidth}
\includegraphics[width=\linewidth]{node_prior_Benign_rej_1.pdf}
\caption{Rejected benign nodes}
\label{fig:node_prior_benign_rej_1}
\end{subfigure}%
\begin{subfigure}[H]{0.24\textwidth}
\includegraphics[width=\linewidth]{node_prior_Sybil_acc_1.pdf}
\caption{Accepted Sybil nodes}
\label{fig:node_prior_sybil_acc_1}
\end{subfigure}%
\caption{
Vary FPR=FNR (node prior)}
\label{fig:node_prior_1}
\vspace{
-0.7cm}
\end{figure}
\myparatight{Varying the number of attack edges}
\label{subsubsec:n_att}
Second, we evaluate SybilFrame when the number of attack edges changes. We set FPR and FNR to be 0.3, and vary the number of attack edges from 0 to 1000. Figure~\ref{fig:node_prior_2} shows the results. We find that both SybilFrame and SybilBelief have good performance with less than 200 attack edges. However, when the number of attack edges increases, SybilBelief degrades its performance while SybilFrame still has stable and near optimal detection accuracy.
\begin{figure}
\begin{subfigure}[H]{0.24\textwidth}
\includegraphics[width=\linewidth]{node_prior_Accuracy_2.pdf}
\caption{Accuracy}
\label{fig:node_prior_accuracy_2}
\end{subfigure}%
\begin{subfigure}[H]{0.24\textwidth}
\includegraphics[width=\linewidth]{node_prior_AUC_2.pdf}
\caption{AUC}
\label{fig:node_prior_auc_2}
\end{subfigure}%
\begin{subfigure}[H]{0.24\textwidth}
\includegraphics[width=\linewidth]{node_prior_Benign_rej_2.pdf}
\caption{Rejected benign nodes}
\label{fig:node_prior_benign_rej_2}
\end{subfigure}%
\begin{subfigure}[H]{0.24\textwidth}
\includegraphics[width=\linewidth]{node_prior_Sybil_acc_2.pdf}
\caption{Accepted Sybil nodes}
\label{fig:node_prior_sybil_acc_2}
\end{subfigure}%
\caption{
Vary the number of attack edges (node prior)}
\label{fig:node_prior_2}
\vspace{-0.7cm}
\end{figure}
\myparatight{Varying the size of the Sybil region}
\label{subsubsec:size_of_sybil}
Furthermore, we evaluate SybilFrame when attacker changes the size of Sybil region. We will not consider the case when the Sybil region is too small, since it has limited utility to perform large-scale attacks. We set FPR and FNR to be 0.3, and vary the size of Sybil region from 400 to 1000.
Figure~\ref{fig:node_prior_3} shows the results. When there are more Sybil nodes, both SybilFrame and SybilBelief improve performance. This is because when both the benign and Sybil region are large, the internal homophily is strong enough to overcome the influence of attack edges. However, SybilFrame still performs better than SybilBelief.
\begin{figure}[!htb]
\vspace{-0.5cm}
\begin{subfigure}[H]{0.24\textwidth}
\includegraphics[width=\linewidth]{node_prior_Accuracy_3.pdf}
\caption{Accuracy}
\label{fig:node_prior_accuracy_3}
\end{subfigure}%
\begin{subfigure}[H]{0.24\textwidth}
\includegraphics[width=\linewidth]{node_prior_AUC_3.pdf}
\caption{AUC}
\label{fig:node_prior_auc_3}
\end{subfigure}%
\begin{subfigure}[H]{0.24\textwidth}
\includegraphics[width=\linewidth]{node_prior_Benign_rej_3.pdf}
\caption{Rejected benign nodes}
\label{fig:node_prior_benign_rej_3}
\end{subfigure}%
\begin{subfigure}[H]{0.24\textwidth}
\includegraphics[width=\linewidth]{node_prior_Sybil_acc_3.pdf}
\caption{Accepted Sybil nodes}
\label{fig:node_prior_sybil_acc_3}
\end{subfigure}%
\caption{
Vary the size of Sybil region (node prior)}
\label{fig:node_prior_3}
\vspace{-0.5cm}
\end{figure}
\subsection{Influence of Edge Priors}
\label{subsec:edge_prior}
We want to explore SybilFrame when only incorporating external edge priors. Similarly, we use FPR and FNR to model the performance of an external edge classifier, which makes predictions of attack edges and other edges. We list our \emph{Edge Prior Generator} algorithm (Algorithm~\ref{alg:edge_prior}) in Appendix~\ref{subsec:prior_generator}.
\myparatight{Varying FPR and FNR}
\label{subsubsec:edge_fpr_fnr_f}
First, we \emph{tune $FPR=FNR$ for edge priors from 0 to 0.5}. We run SybilFrame with default node priors and compare with SybilBelief. From Figure~\ref{fig:edge_prior_1}, with $FPR=FNR\leq 0.3$, SybilFrame performs better than SybilBelief.
Figure~\ref{fig:edge_prior_1_2} in Appendix~\ref{syn_edge_fpr0.1fpr} shows the results when we \emph{set $FPR=0.1$ and tune $FNR$ from 0 to 0.5}. As we can see, SybilFrame has good performance and outperforms SybilBelief even when $FNR$ is 0.5. This means that as long as the external edge classifier has some power to detect attack edges, incorporating edge priors into SybilFrame gives better performance.
\begin{figure}
\begin{subfigure}[H]{0.24\textwidth}
\includegraphics[width=\linewidth]{edge_prior_Accuracy_1.pdf}
\caption{Accuracy}
\label{fig:edge_prior_accuracy_1}
\end{subfigure}%
\begin{subfigure}[H]{0.24\textwidth}
\includegraphics[width=\linewidth]{edge_prior_AUC_1.pdf}
\caption{AUC}
\label{fig:edge_prior_auc_1}
\end{subfigure}%
\begin{subfigure}[H]{0.24\textwidth}
\includegraphics[width=\linewidth]{edge_prior_Benign_rej_1.pdf}
\caption{Rejected benign nodes}
\label{fig:edge_prior_benign_rej_1}
\end{subfigure}%
\begin{subfigure}[H]{0.24\textwidth}
\includegraphics[width=\linewidth]{edge_prior_Sybil_acc_1.pdf}
\caption{Accepted Sybil nodes}
\label{fig:edge_prior_sybil_acc_1}
\end{subfigure}%
\caption{
Vary FPR=FNR (edge prior)}
\label{fig:edge_prior_1}
\vspace{-0.5cm}
\end{figure}
\myparatight{Varying the number of attack edges}
\label{subsubsec:edge_n_att}
Second, we set FPR to be 0.1 and FNR to be 0.5, and vary the number of attack edges from 0 to 1000.
Figure~\ref{fig:edge_prior_2} shows the results. As the number of attack edges increases, both SybilFrame and SybilBelief degrade performance. However, SybilFrame still outperforms SybilBelief. Notice that the performance of SybilFrame depends the detection accuracy of external classifier. If we have a classifier with 0.1 FPR and 0.1 FNR, SybilFrame will have near optimal performance.
\begin{figure}
\begin{subfigure}[H]{0.24\textwidth}
\includegraphics[width=\linewidth]{edge_prior_Accuracy_2.pdf}
\caption{Accuracy}
\label{fig:edge_prior_accuracy_2}
\end{subfigure}%
\begin{subfigure}[H]{0.24\textwidth}
\includegraphics[width=\linewidth]{edge_prior_AUC_2.pdf}
\caption{AUC}
\label{fig:edge_prior_auc_2}
\end{subfigure}%
\begin{subfigure}[H]{0.24\textwidth}
\includegraphics[width=\linewidth]{edge_prior_Benign_rej_2.pdf}
\caption{Rejected benign nodes}
\label{fig:edge_prior_benign_rej_2}
\end{subfigure}%
\begin{subfigure}[H]{0.24\textwidth}
\includegraphics[width=\linewidth]{edge_prior_Sybil_acc_2.pdf}
\caption{Accepted Sybil nodes}
\label{fig:edge_prior_sybil_acc_2}
\end{subfigure}%
\caption{
Vary the number of attack edges (edge prior)}
\label{fig:edge_prior_2}
\vspace{-0.1cm}
\end{figure}
\myparatight{Varying the size of the Sybil region}
\label{subsubsec:edge_size_of_sybil}
Furthermore, we evaluate SybilFrame with edge priors when attacker changes the size of the Sybil region. We set FPR to be 0.1 and FNR to be 0.5, and vary the size of Sybil region from 400 to 1000.
From Figure~\ref{fig:edge_prior_3}, SybilFrame improves its performance when there are more Sybil nodes, and still outperforms SybilBelief.
\begin{figure}
\begin{subfigure}[H]{0.24\textwidth}
\includegraphics[width=\linewidth]{edge_prior_Accuracy_3.pdf}
\caption{Accuracy}
\label{fig:edge_prior_accuracy_3}
\end{subfigure}%
\begin{subfigure}[H]{0.24\textwidth}
\includegraphics[width=\linewidth]{edge_prior_AUC_3.pdf}
\caption{AUC}
\label{fig:edge_prior_auc_3}
\end{subfigure}%
\begin{subfigure}[H]{0.24\textwidth}
\includegraphics[width=\linewidth]{edge_prior_Benign_rej_3.pdf}
\caption{Rejected benign nodes}
\label{fig:edge_prior_benign_rej_3}
\end{subfigure}%
\begin{subfigure}[H]{0.24\textwidth}
\includegraphics[width=\linewidth]{edge_prior_Sybil_acc_3.pdf}
\caption{Accepted Sybil nodes}
\label{fig:edge_prior_sybil_acc_3}
\end{subfigure}%
\caption{
Vary the size of Sybil region (edge prior)}
\label{fig:edge_prior_3}
\vspace{-0.6cm}
\end{figure}
\subsection{Resilient Against Seed Targeting Attacks}
\label{subsec:targeted_attack}
We are interested in the impact of seed targeting attacks, i.e., when the known labeled nodes are end points of attack edges. We consider the following cases:
\textbf{1) SI:} Benign (Sybil) trust seeds are not end points of attack edges.
\textbf{2) SII:} Benign (Sybil) trust seeds are end points of attack edges.
\begin{figure}[!htb]
\begin{subfigure}[H]{0.24\textwidth}
\includegraphics[width=\linewidth]{targeted_node_Accuracy_sen.pdf}
\caption{Accuracy (node prior)}
\label{fig:targeted_node_accuracy}
\end{subfigure}%
\begin{subfigure}[H]{0.24\textwidth}
\includegraphics[width=\linewidth]{targeted_edge_Accuracy_sen.pdf}
\caption{Accuracy (edge prior)}
\label{fig:targeted_edge_accuracy}
\end{subfigure}%
\caption{
Accuracy of SybilFrame under seed targeting attacks. (a) Given node priors. (b) Given edge priors.
}
\label{fig:targeted_accuracy}
\vspace{-0.3cm}
\end{figure}
Figure~\ref{fig:targeted_accuracy} shows the accuracy as a function of the number of attack edges for four scenario combinations of trust seeds, in the node prior experiment (FPR=FNR=0.3) and edge prior experiment (FPR=0.1, FNR=0.5). We find that the location of trust seeds have no influence on the detection accuracy. Due to limited space, we list the results of AUC, FP and FN in Appendix~\ref{subsec:seed_targeted_appendix}. We find that SybilFrame is
resilient against seed targeting attacks, and we can simply select trust seeds uniformly at random.
\subsection{Summary}
\label{subsec:summary}
In summary, we have the following observations:
\textbf{1)} SybilFrame outputs near optimal results when incorporating node priors with FPR and FNR less than 0.3.
\textbf{2)} SybilFrame outperforms SybilBelief when incorporating node priors with FPR and FNR less than 0.4.
\textbf{3)} When incorporating edge priors, as long as the edge priors has a low FPR (0.1) and some level of FNR (less than 0.5), SybilFrame outperforms SybilBelief.
\textbf{4)} SybilFrame is robust to different attack strategies and resilient against seed targeting attacks.
\section{Evaluation on Facebook Network}
\label{sec:evaluation_facebook}
We evaluate SybilFrame on semi-real Facebook network, and compare with state-of-the-art Sybil defense mechanisms: SybilLimit, SybilInfer, SybilRank and SybilBelief. We find that SybilFrame performs orders of magnitudes better than other methods, especially when the number of attack edges is large. Furthermore, the performance of SybilFrame is stable and near optimal.
\subsection{Dataset Description}
\label{subsec:facebook_dataset_description}
The dataset we use is the ego-Facebook dataset obtained from Stanford Network Analysis Project (SNAP)~\cite{facebookdata}. The Facebook graph contains 4,039 nodes and 88,234 edges. In this graph, nodes are Facebook accounts and edges are friendship relationships. The graph is connected and undirected, with a diameter 8 and average clustering coefficient 0.6055.
\subsection{Experimental Setup}
\label{subsec:facebook_exp_setup}
We construct the network topology as follows. We use this Facebook dataset as both the benign region and Sybil region, and randomly add attack edges between the two regions.
We vary the number of attack edges from 1000 to 20000, and evaluate the performance of SybilFrame, as well as SybilLimit, SybilInfer, SybilRank and SybilBelief. We randomly select 1 benign trust seed and 1 Sybil trust seed, and perform the experiments 100 times and then take the average.
\subsection{Compute Prior Information}
\label{subsec:facebook_prior}
To run SybilFrame, we need prior information. Since the benign region is identical to the Sybil region, we are not able to collect distinguishable node priors. Thus, we only explore ways to compute edge priors.
As discussed in Section~\ref{subsec:prior}, we can leverage
similarity between two connected nodes, and use it as a prior for the edge between them. Intuitively, connected benign nodes are similar and connected benign and Sybil nodes are not similar. Therefore, attack edges should have a lower score than non-attack edges. We adopt the Jaccard index here as a measure of similarity. For an edge $(u, v)\in E$, the Jaccard index~\cite{Liben-Nowell2007} of it is defined as $\frac{\left|\Gamma(u)\cap \Gamma(v)\right|}{\left|\Gamma(u)\cup \Gamma(v)\right|}$, where $\Gamma(u)$ denotes the set of one-hop neighbors of node $u$, and $\left|\Gamma(u) \cap \Gamma(v)\right|$ denotes the number of common neighbors of $u$ and $v$. For edges that connect two trust seeds, we set the prior of it to 0.1 if the edge is an attack edge, and set to 0.9 if the edge is a non-attack edge. For other edges, we compute the corresponding Jaccard index. We scale these indices into the range $[0.1, 0.9]$. These scaled Jaccard scores will then be used as priors.
We can also use other similarity metrics, such as Cosine index~\cite{Lu2011} or Adamic-Adar index~\cite{Liben-Nowell2007}
, and combine them to obtain an overall similarity score. A possible approach is to use these raw similarity scores as features for an edge, and obtain a feature matrix for all edges on the graph. We can then adopt a supervised learning approach by leveraging existing tools, such as Logistic Regression~\cite{hastie01statisticallearning} and Support Vector Machine~\cite{hastie01statisticallearning}, and make probabilistic predictions of each edge being a non-attack/attack edge. These probabilistic outputs can then be used as overall prior scores.
\subsection{Results}
\label{subsec:facebook_result}
Figure~\ref{fig:facebook} shows the performance of SybilFrame and other Sybil defense mechanisms as we vary the number of attack edges from 1000 to 20000. Since SybilRank is a ranking scheme and it is very hard to directly use the degree normalized scores to make predictions, we only compare with SybilRank in terms of AUC. Also, since SybilLimit and SybilInfer output binary predictions rather than belief scores, we do not include them into AUC comparison. We find that: 1) As the number of attack edges increases, the performance of precious methods degrades, with a lower accuracy (SybilLimit, SybilInfer, SybilBelief), and lower AUC (SybilRank). 2) The speed of performance degradation is fast. With more than 3000 attack edges, the detection accuracy of SybilBelief is less than 0.5, worse than a random guess, and SybilLimit and SybilInfer predict all Sybil nodes to be benign, thus losing the detection capability. Thus, SybilBelief, SybilLimit and SybilInfer do not work on weak trust networks with a large number of attack edges. 3) The performance of SybilFrame is stable and near optimal in all cases. By incorporating edge prior information, SybilFrame is able to restrict the amount of message passing across the attack edges. Thus, SybilFrame is able to successfully handle the situation when the number of attack edges is large, and performs orders of magnitudes better than other approaches.
\begin{figure}[!t]
\begin{subfigure}[H]{0.24\textwidth}
\includegraphics[width=\linewidth]{facebook_Accuracy.pdf}
\caption{Accuracy}
\label{fig:facebook_accuracy}
\end{subfigure}%
\begin{subfigure}[H]{0.24\textwidth}
\includegraphics[width=\linewidth]{facebook_AUC.pdf}
\caption{AUC}
\label{fig:facebook_auc}
\end{subfigure}%
\begin{subfigure}[H]{0.24\textwidth}
\includegraphics[width=\linewidth]{facebook_Benign_rej.pdf}
\caption{Rejected benign nodes}
\label{fig:facebook_benign_rej}
\end{subfigure}%
\begin{subfigure}[H]{0.24\textwidth}
\includegraphics[width=\linewidth]{facebook_Sybil_acc.pdf}
\caption{Accepted Sybil nodes}
\label{fig:facebook_sybil_acc}
\end{subfigure}%
\caption{
Performance comparison on Facebook network}
\label{fig:facebook}
\vspace{-0.6cm}
\end{figure}
\section{Evaluation on Real-World Large-Scale Twitter Network}
\label{sec:twitter}
In this section, we will evaluate SybilFrame on real-world large-scale Twitter network comprising over $20M$ nodes and $256M$ edges. We will explore ways to compute prior information, and incorporate it to SybilFrame.
\subsection{Collecting Twitter Dataset}
\label{subsec:twitter_analysis}
We obtained a snapshot of the Twitter follower network which was crawled by Kwak et al.~\cite{Kwak10}.
\myparatight{Pre-processing} Originally, the Twitter network is \textbf{directed}. Since it is easy for attackers to manipulate one-way directed edges,
we transform this directed network to an \textbf{undirected} one by retaining an undirected edge between $u$ and $v$ if both directed edges $(u,v)$ and $(v,u)$ exist. Furthermore, we select the largest connected component of the transformed network since all investigated algorithms require the networks to be connected. The largest connected component contains 21,297,772 nodes, and 265,025,545 edges, with average degree 24.9.
We note that some previous works remove nodes with degrees smaller than a threshold from the social networks. For instance, SybilLimit~\cite{Yu08} removes nodes with degree smaller than 5 and SybilInfer~\cite{Danezis09} removes nodes with degree smaller than 3. Mohaisen et al.~\cite{mohaisen:imc10} found that such pre-processing will prune a large portion of nodes. Indeed, social networks often have a long-tail degree distribution (e.g., power-law degree distribution~\cite{Clauset09} and lognormal degree distribution~\cite{Gong12-imc}), in which most nodes have very small degrees. Thus, a large portion of nodes are pruned by such pre-processing. Such pre-processing could result in high FPR or high FNR depending on how the OSN operator treats the pruned nodes.
If the OSN operator treats all the pruned nodes whose degrees are smaller than a threshold as benign nodes, then an attacker can create many malicious nodes with degree smaller than the threshold, resulting in high FNR, otherwise a large fraction of benign nodes will be treated as malicious nodes, resulting in high FPR. Therefore, we do not perform such pre-processing to the Twitter network.
\myparatight{Collecting ground truth} To evaluate the approaches, we need ground truth for the nodes in the Twitter network. The collected Twitter network includes users' Twitter IDs. Therefore, we re-crawled every account using Twitter's API, which tells us the status (i.e., active, suspended or deleted) of each account.
In summary, we found that 145,156 nodes (i.e., 0.7\%) are suspended, 1,911,482 nodes (i.e., 9.0\%) are deleted, and the rest of the nodes are still active. We take the suspended accounts as Sybil nodes and the active ones as benign nodes.
\subsection{Measuring Twitter Structure}
We find that: 1) Many Sybil nodes are isolated from other Sybils. 2) The number of attack edges is very large. This means that using existing structure-based Sybil detection approaches will achieve limited performance.
\myparatight{No community structure} We adopt \emph{modularity}~\cite{Newman04}, ranging from -0.5 to 1, to quantify if a partition of a network (i.e., the partition in our case consists of the benign and Sybil regions in the Twitter network) can be viewed as two communities. Clauset et al.~\cite{Clauset04} concluded, via a large amount of empirical experiments on real networks, modularity $>$ 0.3 indicates significant community structure. However, we find that the partition consisting of the benign and malicious regions only has modularity 0.0042.
Thus, the benign and Sybil regions cannot be viewed as two separate communities. Next, we show two reasons: half of the Sybil nodes are isolated and the number of attack edges per Sybil node is high.
\begin{figure}[!htb]
\centering
{\includegraphics[width=0.4\textwidth]{twitter_component-size-distribution.pdf}}
\caption{Distribution of connected Sybil component sizes
}
\label{component-size}
\vspace{-0.5cm}
\end{figure}
\begin{figure}
\begin{subfigure}[H]{0.24\textwidth}
\includegraphics[width=\linewidth]{twitter_deg-honest.pdf}
\caption{Benign nodes}
\label{benign-deg}
\end{subfigure}%
\begin{subfigure}[H]{0.24\textwidth}
\includegraphics[width=\linewidth]{twitter_deg-sybil.pdf}
\caption{Sybil nodes}
\label{sybil-deg}
\end{subfigure}%
\caption{CDFs for a) benign and b) Sybil nodes
}
\label{deg-dis}
\vspace{-0.5cm}
\end{figure}
\textbf{1) Half of the Sybil accounts are isolated:} In total, we find 77,917 connected components in the Sybil region (i.e., the subgraph including all malicious nodes and edges between them). Figure~\ref{component-size} shows the distribution of sizes of these components. First, around 50\% of Sybil nodes are \emph{isolated}, i.e., they only link to benign nodes. Second, we find that there exists a large connected component including 45\% of all malicious nodes. Specifically, this component consists of 65,579 nodes and 931,287 edges, resulting in an average degree of 28.40. Thus, the large component is even denser than the benign region whose average degree is 21.62.
We might wonder if this large connected component can be viewed as a community. However, we find that the modularity of the partition consisting of the benign region and the largest connected component is still only 0.0046, which means that even this large connected component cannot be viewed as a community. %
Third, the rest of nodes are in connected components whose sizes are less than 20.
\textbf{2) Large number of attack edges:} We observe that there are 18,414,469 attack edges, which means each Sybil node successfully attacks around 127 benign nodes on average. Figure~\ref{deg-dis} further characterizes how attack edges are distributed among the benign and malicious regions. We can draw several conclusions from this figure.
First, the benign and Sybil regions are structurally similar. Specifically, the number of all neighbors of both benign and Sybil nodes follow long-tail distributions. In fact, such long-tail distributions are also widely observed in other OSNs such as LiveJournal~\cite{Mislove07} and Google+~\cite{Gong12-imc}. We speculate that Sybil nodes are imitating the benign region %
to evade automatic detection.
Second, from Figure~\ref{benign-deg}, we find that around 90\% of benign nodes are not connected to any Sybil node. Moreover, the number of Sybil neighbors of benign nodes also follows a long-tail distribution. This implies that, although around 10\% of benign nodes link to malicious nodes, most attack edges concentrate on a smaller number of benign nodes. For instance, we find that 90\% of attack edges concentrate on only 3\% of benign nodes. We speculate that such nodes are celebrities that tend to follow back to any user who follows them.
Third, from Figure~\ref{sybil-deg}, we observe that around 2\% of Sybil nodes do not link to any benign node. Again, the number of benign neighbors of Sybil nodes follows a long-tail distribution, which implies that most attack edges are produced by a small portion of Sybil nodes. For instance, we find that 90\% of attack edges are produced by only 16\% of Sybil nodes.
Note that the structural properties (i.e., many Sybil nodes are isolated and there are a large number of attack edges per Sybil node) of the Sybil nodes in our Twitter dataset match those in another large-scale Twitter network~\cite{Ghosh12} and those in the RenRen social network~\cite{Yang11-sybil}, which indicates the representativeness of our observations. Figure~\ref{twitter_structure} gives a snapshot of the structure.
\myparatight{Summary} We observe that the reason why structure-based Sybil detection approaches fail is that the assumptions they require are not satisfied. Specifically, the benign and Sybil regions cannot be viewed as two separate communities. One reason is that a significant portion of the Sybil nodes are isolated, and the other reason is that the number of attack edges per Sybil node is high.
\begin{figure}[!t]
\centering
{\includegraphics[width=0.35\textwidth]{twitter_structure.pdf}}
\caption{
Structure of Twitter network
}
\label{twitter_structure}
\vspace{-0.6cm}
\end{figure}
\subsection{Computing Node Priors}
\label{subsec:twitter_nodeprior}
We now discuss ways to compute node priors. The idea is to collect features and train a classifier that outputs probabilistic scores. Since we do not know whether deleted accounts are benign or Sybil, we will not include them in the training, prediction and evaluation process. We just set the priors for them to be 0.5.
\subsubsection{Collecting node features}
\label{subsubsec:node_feature}
We compute the following three features. We compute Feature 1) and 2) for all nodes on the original directed network, and map to the corresponding nodes on the undirected largest connected component. We directly compute Feature 3) on the undirected topology.
\textbf{1) Incoming requests accepted ratio:} The insight is that a Sybil identity is more likely to accept incoming requests than benign users, in order to quickly propagate spam. Hence on average, Sybil identities shall have a higher incoming requests accepted ratio. Since we only have \textbf{structural} information, we decide to use the incoming and outgoing edges associated with a node to \textbf{model} the ratio. For a node $v$ on the original directed Twitter graph, we denote $\Gamma_{in}(v)$ as the set of all incoming edges of $v$, and denote $\Gamma_{out}(v)$ as the set of all outgoing edges of $v$. Hence $\Gamma_{in}(v) \cap \Gamma_{out}(v)$ is the set of edges that are both incoming and outgoing edges of $v$. The \emph{incoming requests accepted ratio} is modeled as
\begin{equation}
Req_{in} = \frac{\left| \Gamma_{in}(v) \cap \Gamma_{out}(v) \right|}{\left| \Gamma_{in}(v)\right|}
\end{equation}
\noindent where $\left| \Gamma(v)\right|$ denotes the cardinality of the set $\Gamma(v)$.
\textbf{2) Outgoing requests accepted ratio:} The insight is that a benign user is more reliable and hence the outgoing friend requests send from him/her are more likely to be accepted. Hence on average, benign users have a higher outgoing requests accepted ratio than Sybil identities. Similarly, we \textbf{model} the \emph{outgoing requests accepted ratio} for a node $v$ as
\begin{equation}
Req_{out} = \frac{\left| \Gamma_{in}(v) \cap \Gamma_{out}(v) \right|}{\left| \Gamma_{out}(v)\right|}
\end{equation}
\noindent where $\left| \Gamma(v)\right|$ denotes the cardinality of the set $\Gamma(v)$.
\textbf{3) Clustering coefficient:} The clustering coefficient for a vertex is a graph metric that measures how close its neighbors are to being a complete graph. For a node $v$ on the undirected graph $G = (V, E)$, its local \emph{clustering coefficient}~\cite{Gong12-imc} is given as.
\begin{equation}
C = \frac{2\left| \{ (i, j): i, j\in V, (i, j)\in E\} \right|}{k_v(k_v - 1)}
\end{equation}
\noindent where $k_v$ is the degree of $v$, $i$ and $j$ are both friends of $v$. The insight is that benign users tend to have well-connected social cliques, and users in such cliques share some attributes in common and are likely to be friends themselves. Therefore, benign users are likely to have a higher clustering coefficient than Sybil identities.
\begin{figure}
\begin{subfigure}[H]{0.24\textwidth}
\includegraphics[width=\linewidth]{twitter_node_feature_scatter_req_ratio.jpg}
%
\caption{Scatter plot}
\label{fig:twitter_node_feature_scatter}
\end{subfigure}%
\begin{subfigure}[H]{0.24\textwidth}
\includegraphics[width=\linewidth]{twitter_node_feature_cluster_coeff.pdf}
\caption{CDF}
\label{fig:twitter_node_feature_cc}
\end{subfigure}%
\caption{Analysis of the node features.}
\label{fig:twitter_node_feature}
\vspace{-0.5cm}
\end{figure}
Figure~\ref{fig:twitter_node_feature} shows the scatter plot of the outgoing requests accepted ratio versus incoming requests accepted ratio, as well as the CDF plot for the clustering coefficient for benign nodes and Sybil nodes. As expected, benign users tend to have a higher outgoing requests accepted ratio, a lower incoming requests accepted ratio, and a higher coefficient. Besides, only using any one of the three features is not able to obtain a clear decision boundary for the classification. Therefore, we need to leverage these three features together by building a machine learning classifier.
\subsubsection{Training a SVM classifier}
\label{subsubsec:svm}
We adopt the \emph{LIBSVM}~\cite{Chang2011} tool to build a Support Vector Machine (SVM)~\cite{hastie01statisticallearning} classifier. We sample a training set comprising 10,000 benign nodes and 10,000 Sybil nodes, and use the remaining nodes as testing. We train a SVM classifier with RBF kernel, whose parameters $c$ and $\gamma$ are obtained from \emph{Grid Search}. The overall prediction accuracy is 90.5\%, with 9.4\% FPR and 31.8\% FNR. Considering the fact that half of Sybil nodes are isolated, it is essentially hard for previous approaches
to detect more than half of total Sybil nodes. Thus, the 68.2\% Sybil nodes detection capability of SybilFrame is impressive.
Some applications may require a lower FPR. A natural way is to assign a higher penalty term to the benign class and a lower penalty term to Sybil class. In this way, we can reduce the FPR of our node classifier to be 8.5\% with 41.6\% FNR.
\subsubsection{Output node priors}
\label{subsubsec:node_outputs}
To output priors, LIBSVM has an internal scheme to allow for probability estimates by fitting a logistic curve and conducting a cross validation procedure~\cite{Chang2011}. We can use the same parameters obtained from grid search, and enable the probability outputs. These output scores are then used as node priors for SybilFrame.
\subsection{Computing Edge Priors}
\label{subsec:twitter_edgeprior}
We now explore ways to compute edge priors. As discussed in Section~\ref{subsec:facebook_prior}, we can leverage well-known similarity metrics. For each edge $(u, v)\in E$ on graph $G = (V, E)$, we compute the following similarity metrics:
\textbf{Number of Common Neighbors}~\cite{Liben-Nowell2007} $S_{uv} = \left| \Gamma(u) \cap \Gamma(v) \right|$
\textbf{Cosine Similarity Index}~\cite{Lu2011} $S_{uv} = \frac{\left| \Gamma(u)\cap \Gamma(v)\right|}{\sqrt{k_uk_v}}$
\textbf{Jaccard Similarity Index}~\cite{Liben-Nowell2007} $S_{uv} = \frac{\left|\Gamma(u)\cap\Gamma(v)\right|}{\left|\Gamma(u)\cup\Gamma(v)\right|}$
\textbf{Adamic-Adar Similarity Index}~\cite{Liben-Nowell2007} $S_{uv} = \sum_{s\in\Gamma(u)\cap\Gamma(v)}\frac{1}{k_s}$.
Following a similar procedure, we scale the features and train a RBF-SVM classifier. As a result, we can successfully detect 18\% attack edges, with FPR 10\%. To improve the performance, we may include more complex similarity metrics, i.e. Katz Index~\cite{Lu2011} and Leicht-Holme-Newman Index~\cite{Lu2011}, which may cost a longer time to compute. Another way is to use node priors to infer edge priors. Generally, for an edge whose end nodes have different predicted labels, we can assign a lower score to indicate a higher possibility to be an attack edge; otherwise, we set the score to be default 0.9 to model homophily. Since our node priors work much better than edge priors, we adopt this inference procedure.
\subsection{Scalable Implementation}
\label{subsec:bp_parallel}
We adopt the \emph{GraphLab} parallel framework~\cite{graphlab} to implement Loopy Belief Propagation in parallel. The parallel framework distributes nodes to multiple processors, and each processor passes and updates messages for the nodes that are assigned to. Essentially, computing node priors and edge priors is also parallelizable.
\subsection{Results}
\label{subsec:results}
We now present our experimental results. We compare SybilFrame with SybilBelief in terms of detection accuracy, FPR and FNR. We compare with SybilBelief and SybilRank in terms of relative ranking of Sybil nodes. If Sybil nodes tend to rank before benign nodes, OSN operators can leverage crowdsourcing (i.e. Amazon Mechanical Turks~\cite{Wang13}) to manually screen and label suspicious accounts. Since SybilLimit and SybilInfer do not scale to large datasets, we do not compare with them.
\myparatight{Detection results}
\label{subsubsec:induce_results}
We randomly select 1000 benign and Sybil seeds, and run SybilBelief and SybilFrame. Table~\ref{tab:detection_results} shows the results. We draw several conclusions: 1) Due to large number of attack edges, SybilBelief predicts all nodes to be Sybil thus completely losing detection capabilities. (We validated the implementation and results with the authors of SybilBelief.) 2) Node prior classifier of SybilFrame has certain detection power, which is able to detect 68.2\% Sybil nodes at maximum. By assigning different penalty terms, the FPR can be reduced to 8.5\% (Node classifier - II). 3) Incorporating node priors into SybilFrame can reduce FPR to 4.2\%. (SybilFrame - II), and achieve a better accuracy 95.4\%. 4) Since we label our ground truth based on whether the accounts are suspended by Twitter, it is possible that Twitter fails to detect some Sybil accounts, which are labeled by SybilFrame as positive examples. Thus, the true FPR should be \emph{lower} than our estimates. 5) Considering that half of Sybil nodes are isolated, the detection capability of 68.2\% Sybil accounts is impressive.
\begin{table}[!h]
\caption{Detection Performance on Twitter}\label{tab:detection_results}
\centering
\begin{tabular}{|l|c|c|c|}
\hline
~ &Accuracy &FPR &FNR\\\hline
SybilBelief &0.7\% &99.3\% &0.00\\\hline
Node classifier &90.5\% &9.4\% &31.8\%\\\hline
SybilFrame &91.8\% &8.0\% &33.5\%\\\hline
Node classifier - II &91.2\% &8.5\% &41.6\%\\\hline
SybilFrame - II & 95.4\% &4.2\% &48.9\%\\\hline
\end{tabular}
\vspace{-0.2cm}
\end{table}
\myparatight{Ranking results}
\label{subsubsec:rank_results}
We rank the posteriors score generated from SybilFrame, as well as scores of SybilBelief and SybilRank, in ascending order. We then compute the portion of Sybil identities in the 1K, 10K, 50K, 100K, 1M and 10M lowest-ranked users.
\begin{figure}[!htp]
\centering
\includegraphics[width=0.4\textwidth]{twitter_sybil_portion.pdf}
\caption{Portion of Sybil identities}
\label{fig:twitter_sybil_portion}
\vspace{-0.4cm}
\end{figure}
Figure~\ref{fig:twitter_sybil_portion} shows the results of four schemes: random guess, SybilRank, SybilBelief and SybilFrame. We draw several conclusions: 1) In the first 1K users, SybilFrame is able to rank over 500 Sybil accounts, 12 times better than SybilBelief, 35 times better than SybilRank and 72 times better than random guess. 2) There exists a significant descending trend of portions in SybilFrame, while SybilRank and SybilBelief do not have such trend. This means that SybilFrame has much more power to rank Sybil nodes in the top part of the ranking list, while SybilRank and SybilBelief roughly distribute the Sybil accounts evenly. 3) Given the same amount of time and human resource, OSN operators can use SybilFrame to detect more Sybil nodes than using SybilRank or SybilBelief.
\myparatight{Problem with Twitter's detection policy}
\label{subsubsec:top_100}
Recall that we obtained our ground truth based on whether the account was active or suspended by Twitter. Thus, it is possible that some accounts are actually Sybil but evade Twitter's detection policy. To test this, first we re-crawl the top 1K accounts, and find that 7 additional accounts have been suspended by Twitter since our first crawl. Next, we manually examine the top 100 accounts, of which 71 are suspended and 29 are active. We examine the profile of each of the 29 active accounts, and find that only 3 accounts are likely to be real, with a long timeline and diverse tweets. Besides, 24 accounts are highly likely to be fake, with common characteristics such as same account images and few tweets. Furthermore, most of their tweets are published around 7/5/2009, and they all contain URLs and are about making money.
Thus, we suspect that these 24 accounts were created by attackers and belong to the Sybil category.
The remaining 2 accounts have less than two tweets and a protected profile, which are marked as suspicious. We give a complete list of these 29 active accounts in Appendix~\ref{subsec:29_active}.
From the above analysis, we conclude that: 1) Twitter's Sybil detection policy is not optimal. 2) SybilFrame is able to uncover a large fraction (24/29) of suspicious accounts that Twitter fails to detect. Hence, the true FPR of SybilFrame should be lower than our estimates.
\subsection{Summary}
\label{subsec:twitter_summary}
In this section, we discussed ways to compute priors and implemented SybilFrame in parallel. We evaluated SybilFrame on real-world, large-scale Twitter network, and have following observations:
\textbf{1)} In terms of detection performance, SybilFrame performs orders of magnitudes better than SybilBelief. Even when the dataset is noisy and the number of attack edges is large, SybilFrame can detect 68.2\% Sybil nodes at maximum. By tuning parameters, SybilFrame can achieve 4.2\% FPR with 51\% detection rate.
\textbf{2)} In terms of ranking performance, SybilFrame performs orders of magnitudes better than SybilBelief and SybilRank. Among the first 1K users, SybilFrame is able to successfully rank 552 Sybil accounts, 12 times better than SybilBelief and 35 times better than SybilRank.
\textbf{3)} SybilFrame is able to uncover large fraction of suspicious accounts that Twitter fails to detect.
\section{Discussion}
\label{sec:discussion}
\myparatight{Defense-in-depth} We discuss the resilience of our approach to attackers that aim to mimic the features we use in \emph{Stage 1}.
Specifically for Twitter experiment, if the attacker wants to mimic the features and let more Sybils bypass the node classifier, he/she needs to control Sybil identities to establish more connections between themselves and form Sybil clusters, in order to have a lower $Req_{in}$, a higher $Req_{out}$ and a higher clustering coefficient. As a result, Sybils will be much more densely connected, and the edge classifier in \emph{Stage 1} together with LBP in \emph{Stage 2} will be more effective to detect them. This is the basic idea of SybilFrame's multi-layered protection and defense-in-depth.
Also, it remains to be discussed whether the attacker wants to spend time in performing such a complex strategy, which consumes both a lot of time and resource. In our Twitter experiment, we found that Sybil identities were often less intelligent (e.g. half of them are isolated and they share common characteristics as discussed in Section~\ref{subsubsec:top_100}), which makes it very easy for a human expert to identify them. However, Twitters fails to detect a significant fraction of Sybils but SybilFrame is able to uncover them.
Furthermore, we recall that although we collected structural features only to evaluate SybilFrame, SybilFrame is an open framework that is able to incorporate content information. Similarly, we can extract content features of each node and edge, and build a content-based classifier, or even combine structural and content features together to build a more powerful general classifier.
\myparatight{Lower FPR} Our experiment considers suspended accounts in Twitter as a ground truth for Sybil identities. Correspondingly, accounts that were not suspended were labeled as benign accounts. We note that this evaluation is conservative: our analysis considers accounts that are labeled as malicious by SybilFrame, but not suspended by Twitter as false positives. It is very well possible that these labeled false positives are actually malicious, however not detected by Twitter. As experimented in Section~\ref{subsubsec:top_100}, Twitter's current detection policy is far from optimal, and there is a large fraction of malicious accounts that Twitter fails to suspend. Therefore, our $4.2\%$ FPR should be essentially lowered.
\myparatight{Sybil detection capability in Twitter} Recall that SybilFrame was able to detect 66.5\% Sybil identities (Table~\ref{tab:detection_results}, SybilFrame). By tuning parameters, SybilFrame was able to reduce FPR to 4.2\% while still detecting 51\% Sybil identities (Table~\ref{tab:detection_results}, SybilFrame - II). We believe that this result is close to the optimal that any structure-based approach could achieve. On the Twitter graph, half of the Sybil identities form a connected component, and the remaining half are isolated and only connect to benign nodes. Since previous structure-based approaches are mostly based on detecting local communities, they are limited in their ability to detect those isolated Sybil nodes.
\myparatight{Broader applicability} Our approach of defense-in-depth, and using a multi-stage classification framework that is able to incorporate prior information about nodes and edges has broad applicability for network security. For example, the area of botnet detection can benefit from similar techniques that combine host-level information with network structure-based information.
\section{Conclusion}
\label{sec:conclusion}
In this paper, we proposed SybilFrame, a defense-in-depth framework, for structure-based Sybil detection in online social systems. SybilFrame uses a multi-stage classification mechanism, which is able to incorporate heterogeneous sources and types of information about the social network. By leveraging the fine grained local information about users and edges, SybilFrame transforms local information into beliefs of labels, and then propagate those beliefs to make collective inferences.
We experimentally evaluated the accuracy of our approach using both synthetic and real-world social network topologies. We evaluated SybilFrame on a large-scale Twitter dataset. Our results demonstrate that SybilFrame is resilient to high number of attack edges, and performs an order of magnitude better than previous structure-based approaches.
Future work includes collecting and incorporating local content information, and enforcing more fine grained control on the belief propagation rules.
\bibliographystyle{IEEEtran}
|
\section{Introduction}
The goal of this paper is to develop a KAM theory for
the functional equation:
\begin{equation}\label{eq:main}
v(\psi + \Omega) +
v(\psi - \Omega)
- 2 v(\psi) + W((\psi, \eta) + \beta v(\psi)) +\lambda = 0
\end{equation}
where $W:\torus^d \rightarrow \real$, $\Omega \in \real^{d-1}$, $\beta \in \real^{d}$
are given, $\psi$ is a variable in $\torus^{d-1}$ and we can
think of $\eta \in \torus$ as a parameter. We are to find $v:\torus^{d-1}\times\torus^{1}\rightarrow \real$ as a function of $(\psi,\eta)$
and $\lambda \in \real$ as a function of the parameter $\eta$. We will refer to \eqref{eq:main} as the ``equilibrium equation''.
The equation \eqref{eq:main} was derived in
the paper \cite{SuZL15} as
the equation is satisfied by hull functions of quasi-periodic
equilibria in Frenkel-Kontorova models of deposition on
quasi-crystals when the frequency of the equilibrium solution is
resonant with the frequencies of the substratum. The variable $\eta$ is an angle which has the meaning of a transversal phase.
Roughly, the model describes particles interacting with their
neighbors and with a substratum which is quasi-periodic. The configuration describing the state of the system is parameterized by the hull function $v$. We try to place
the particles with a frequency (inverse of the density of particles) which resonates with the frequencies of the medium.
$\Omega$ represents the \emph{intrinsic frequencies}. Since the medium is
resonating with the frequency of the configuration, the positions of the particles are parameterized by $d-1$ angles, i.e. they cover densely a $d-1$ dimensional torus which is indexed by $\eta$.
The $W$ represents the forces of the particles with the substratum and the
$\lambda$ is an external force.
We refer to the paper \cite{SuZL15} for the discussion of the physical
motivations (there are several physical motivations for the Frenkel-Kontorova model)
and for a formal analysis. From the strictly logical point of
view, this paper and \cite{SuZL15} are completely independent even if they are motivated by the same physical problem. They also rely on very different techniques.
To avoid repetition, we refer to \cite{SuZL15} for
references to the previous literature on the problem as well as for physical motivations.
The main goal of this paper is to develop a KAM theory for the equation
\eqref{eq:main}, but we will have to add
a one dimensional extra parameter to it.
The main source of the difficulty to implement a
Newton method -- as needed in KAM theory -- is that the equation \eqref{eq:inv_lin} (the linearization of
the equilibrium equation \eqref{eq:main}) is
not easily analyzable in a way that leads to tame estimates. We will deal with this problem by adding an extra auxiliary equation which implies that \eqref{eq:inv_lin} can be solved with tame estimates. The addition of an extra equation that allows to solve the linearization is similar in spirit to the introduction of the reducibility in KAM theory \cite{Moser'67}.
Nevertheless, our auxiliary equation is very different from the one in reducibility.
\subsection{The method of adding extra parameters to equations}
The main observation that allows us
to develop a KAM theory is that if we are allowed to adjust a
one dimensional parameter in the potential, then the linearized equilibrium equation
admits a very nice structure (it can be factorized into two
first order equations and the factorization allows to develop an iterative procedure which is quadratic convergent).
Hence, in Section \ref{sec:modified equilibrium equation}, we will add an extra parameter to the left hand side of \eqref{eq:main} to obtain the modified equilibrium equation \eqref{external force} such that its linearization can be factorized. Then, we supplement \eqref{external force} with another equation \eqref{modified factorization} in Section \ref{sec:factorization equation} (we call it \emph{the factorization equation})
which encodes that the linearization of the equilibrium equation can be solved.
\subsubsection{The modified equilibrium equation}\label{sec:modified equilibrium equation}
For each fixed $\eta \in \mathbb{T}^1$
we will look for a function
$v:\mathbb{T}^{d-1}\rightarrow \mathbb{R}$ and for numbers
$\lambda, \sigma$ in such a way that we have
\begin{equation}\label{external force}
\begin{split}
&\mathscr{E}[v, \sigma,\lambda](\psi, \eta)\\
\equiv &
v(\psi +\Omega) + v(\psi -\Omega) -
2v(\psi) + W((\psi, \eta) + \beta v(\psi)) + v(\psi) \sigma+
\lambda\\
= & 0.
\end{split}
\end{equation}
In the rest of the paper, we will call this modified equation \eqref{external force} the equilibrium equation. The equation \eqref{external force} has a symmetry that makes the solutions not unique (this corresponds to a gauge symmetry related to the choice of origin of the phase in the original problem). Hence, to obtain local uniqueness, we supplement \eqref{external force} with the following normalization:
\begin{equation}\label{normalization}
\int_{\mathbb{T}^{d-1}} v(\psi) \, d \psi = 0.
\end{equation}
We consider $\mathscr{E}$ as a functional that given a function $v$ and two numbers $\sigma, \lambda$ produces another function given by the second line in \eqref{external force}. And we treat the equation \eqref{external force} as searching for zeros of the functional $\mathscr{E}$.
Since $v,\sigma,\lambda$ all depend on the parameter $\eta$, for convenience, we will write $v_{\eta}$ when we need to emphasize the fact that $v$ depends on $\eta$, and similarly for all the other functions. We will also obtain $\sigma$, $\lambda$ as a function of $\eta$ (and maybe of frequency $\Omega$, but we will not discuss dependence on $\Omega$ in this paper).
Depending on the physical solutions, we may impose the value of one of these variables and determine the others. For example, if we are imposing an external force we may want to fix $\lambda$ or if the material is constrained to have a certain density, we may fix $\Omega$.
It is important to notice that, once we have established the KAM theorem, eliminating some variables in terms of the others is just an application of the finite dimensional implicit function theorem.
\begin{Remark}
Variants of the idea of adding external parameters and then setting them
to zero, has appeared in many guises. In perturbative expansions in
Physics, it is called \emph{the method of counterterms}
\cite{BogoliubovS80, Gallavotti85}. In differential equations, it is
called \emph{Cesari's alternative method} (Chapter IX of \cite{Hale80},\cite{ChowH82}).
Closer to us, in KAM theory, it was introduced in \cite{Moser'67}.
It was realized in \cite{Llave86,Broer96,Yoccoz92, Sevryuk99} that it provided a
good way to deal with degenerate problems. A very systematic
treatment of dependence on parameters (including
parameters taking values in nowhere dense sets) appears in \cite{Vano02}.
\end{Remark}
\begin{Remark}
The exact form of the counterterm added is not that important.
We could have put other counterterms
$\sigma F(v)$ for almost any function $F\nequiv0$.
The way of thinking geometrically of the counterterms is
that there is codimension $1$ set of potentials for
which the solutions move differentiably with respect to
parameters. The counterterm is a projection that moves to keep the problem in this manifold. We could have
taken any other family of corrections to the codimension one
manifold where the solutions are found. We refer to Section \ref{sec:extensions} for a precise formulation and more details.
\end{Remark}
\begin{Remark}
Even if we could consider \eqref{external force} as a functional equation for
each value of $\eta$, we will show in the next section that the
symmetries of the equation involve mixing the $\psi$ and $\eta$
dependence. Relatedly, we note that the equation
\eqref{external force} for a fixed value of $\eta$ does not have a variational
principle.
It will be important to mention that, because $\beta$ has components
both in the $\psi$ and the $\eta$ directions, the equation
\eqref{external force} cannot be considered just as a parameterized version
of the equations considered in \cite{SuL1}.
\end{Remark}
\subsubsection{The factorization equation}\label{sec:factorization equation}
For the KAM treatment of the equilibrium equation \eqref{external force},
we will find it useful to supplement \eqref{external force} with another equation which we call the factorization equation
\begin{equation}\label{modified factorization}
\mathscr{F}[v,\sigma,c](\psi,\eta) \equiv [-c(\psi)+2-\partial_{\beta}W((\psi,\eta)+\beta v_\eta(\psi))-\sigma]\ c(\psi+\Omega)-1=0.
\end{equation}
This should be considered as a functional equation for functions $v,c:\mathbb{T}^{d-1}\rightarrow \mathbb{R}$ and number $\sigma$ when $\eta$ is fixed.
The equation \eqref{modified factorization} gives a condition which ensures that the
linearization of the equilibrium equation \eqref{external force} has a nice structure which allows to implement a Newton method with tame estimates (namely, that it factorizes into two first order difference equations).
See Section \ref{sec:motivation} for a discussion of \eqref{modified factorization} as a tool for solving \eqref{external force}.
The equation~\eqref{modified factorization} is hard to solve exactly,
but we will be able to develop a quasi-Newton method for \eqref{modified factorization}.
The main idea of the paper is that, even if we do not know how to
carry out a KAM theory for the equilibrium equation
\eqref{external force} alone, we can carry out a KAM procedure for
the pair of equations \eqref{external force} and \eqref{modified factorization}.
As discussed in Section~\ref{sec:reducibility}, similar things (a functional equation supplemented by another auxiliary one that makes the associated linearized equation solvable) have happened in
classical problems in KAM theory \cite{Moser'67,Eliasson88}.
\begin{Remark}
Notice that both the equilibrium equation \eqref{external force} and the factorization equations \eqref{modified factorization}
are coupled because $v$ is an unknown in both equations.
Nevertheless, the effect of the factorization equation on the
equilibrium equation appears only through the counterterm $\sigma$ and is a
mild effect (linear in $\sigma$). (See Section \ref{sec:solve the linearization} for a detailed treatment.) On the other hand, the effect of the variables
of the equilibrium equation is very strong. Hence, one can
think of the pair \eqref{external force} and \eqref{modified factorization} almost as an upper triangular system of equations. Indeed, the perturbative treatment of the pair \eqref{external force} and \eqref{modified factorization} has a skew product structure. See Section~\ref{sec:perturbative}.
\end{Remark}
\subsubsection{Comparison of the method in this paper with the application of reducibility}
\label{sec:reducibility}
The method of adding an extra equation so that the linearization
of the equilibrium equation is solvable, has already appeared
in KAM theory.
In the theory of perturbation of lower dimensional elliptic
tori, the classical treatment is to try to
reduce the linearized equation to an equation with constant coefficients
\cite{Moser'67, Melnikov}. This requires extra non-resonance conditions
and, in principle extra parameters. See, in particular \cite{ChengW99,HanY06,Treschev91},
which study the problem of breakdown of resonant tori in Hamiltonian systems.
In the present case, the situation is completely different
in the details
(since we do not seek reducibility but rather factorization
into two second order equations) as well as in the concepts (in \cite{Moser'67}, the parameters are related to initial conditions or the characteristic numbers of the linearized equation). Hence the parameter count of the present method is
very different from what one could expect from reducibility. Also the reducibility equations have a very different geometric meaning from the factorization equations.
The factorization method has analogies in higher dimensional
systems and in elliptic PDE \cite{Kozlov'83,Moser'88}.
One can think of factorization as an analogue of putting the PDE in divergence form. The transformation to divergence form is achieved in \cite{Kozlov'83,Moser'88} for elliptic operators taking advantage of an identity (which is analogous to the Ward identities in gauge theory). Here, on the other
hand, we have to do a KAM theorem to obtain an auxiliary function that gives the factorization.
The present method gives an a-posteriori theorem and smooth dependence on parameters, justifies the perturbation theory and leads to efficient numerical implementations.
\subsection{Properties of
the equilibrium equations \eqref{external force} and its associated factorization equation \eqref{modified factorization}}
Before embarking on the hard analysis, in this section, we derive some
identities and symmetries of the equations which are only soft
analysis. This section can be skipped by readers interested only in the KAM methods.
Many of the symmetries and elementary properties
derived for the equilibrium equation in \cite{SuZL15} lift
straightforwardly to the factorization equation. Surprisingly, the formal perturbation theory developed here
for the pair \eqref{external force} and \eqref{modified factorization} is more efficient than the perturbation theory for \eqref{eq:main} alone developed in \cite{SuZL15}. The perturbation theory developed in \cite{SuZL15} was only for perturbation around integrable solutions, but the perturbation theory for the pair \eqref{external force} and \eqref{modified factorization} is developed around any solution of both equations \eqref{external force} and \eqref{modified factorization}. See Section \ref{sec:extensions}.
The
expansions around zero found in \cite{SuZL15} are a
particular case of the expansions found here since
the linearization of the equilibrium equation around zero admits a trivial factorization.
\subsubsection{The symmetries of
the equilibrium equations \eqref{external force}. }
We note that the symmetries for the equilibrium equation found in
\cite{SuZL15} extend to
the factorization equation.
We have that if $v_\eta, \sigma(\eta), \lambda(\eta)$
is a solution of \eqref{external force}, for any function $\iota(\eta)$
so is:
\begin{equation}\label{symmetry}
\begin{split}
\tilde v_\eta(\psi) & =
v_{\eta + \iota(\eta)\beta_\eta}\big(\psi + \iota(\eta) \beta_\psi\big) + \iota(\eta),\\
\tilde \sigma(\eta) &= \sigma\big(\eta + \iota(\eta) \beta_\eta \big),\\
\tilde \lambda(\eta) &= \lambda \big(\eta + \iota(\eta)\beta_\eta\big)
- \iota(\eta) \ \sigma\big( \eta + \iota(\eta) \beta_\eta\big) .
\end{split}
\end{equation}
Here we use $\beta_{\psi}$ to denote the first $d-1$ components of $\beta$ and $\beta_{\eta}$ to denote the last component of $\beta$.
Since the symmetry \eqref{symmetry} involves changes of
arguments, giving a $v_\eta$, finding the $\iota(\eta)$ that accomplishes the normalization involves solving the implicit
equation
\begin{equation}\label{implicit}
I(\eta + \beta_\eta \iota(\eta) ) + \iota(\eta) = 0
\end{equation}
where $I(\eta) \equiv \int_{\mathbb{T}^{d-1} } v_\eta(\psi)\, d \psi$.
Applying the finite dimensional implicit function theorem, we can
solve \eqref{implicit}
if $I$ and its derivative with respect to $\iota(\eta)$ are both small.
In contrast, in the non-resonant case treated in \cite{SuL1}, the normalization of
the function $v$ could always be solved explicitly.
As we will prove in Section~\ref{sec:proof}, the solutions
of \eqref{external force} that satisfy the normalization \eqref{normalization}
will be locally unique.
\section{Preliminaries}\label{spaces and linear estimates}
To formulate the KAM results (as well as to make quantitative the
Lindstedt series) we need to define precisely the norms of
analytic functions. In this section, we collect the
definitions and some standard properties of spaces of analytic functions.
In Section \ref{spaces}, we collect several standard definitions of
spaces and present some preliminary results on these spaces. In
Section \ref{Diophantine properties} we present definitions of the
Diophantine properties we will use in this paper. In Section
\ref{cohomology} we present well known estimates for cohomology
equations, which are the basis of the KAM procedure. Besides the customary constant coefficient equations, we study first order cohomology equations with non-constant coefficients in section \ref{sec:twisted}, which were also studied in \cite{Herman83}.
\subsection{Spaces of functions we will use}\label{spaces}
We will use a variation on the same spaces
of analytic functions which have been used very often in
KAM theory since \cite{Moser'67}. We will use the same notations as in
\cite{Rafael'08,CallejaL'10, SuL1, SuL2}.
We denote by
\begin{equation*}
D_\rho\equiv \{~ \psi \in \mathbb{C}^{d-1}/\mathbb{Z}^{d-1} ~| \quad
|\mathrm{Im}(\psi_j)|<\rho ~\}.
\end{equation*}
We denote the Fourier expansion of a periodic function $v(\psi)$ on
$D_\rho$ by
\begin{equation*}
v(\psi)=\sum_{k\in \mathbb{Z}^{d-1}} \hat{v}_k e^{2\pi i k\cdot \psi},
\end{equation*} where $\cdot$ is the Euclidean scalar product in
$\mathbb{C}^{d-1}$ and $\hat{v}_k$ are the Fourier coefficients of $v$.
We denote by $\mathscr{A}_\rho$ the Banach space of analytic
functions on $D_\rho$ which are real for real argument and extend
continuously to $\overline{D_\rho}$. We make $\mathscr{A}_\rho$ a
Banach space by endowing it with the supremum norm:
\begin{equation*}
\|v\|_\rho =\sup_{\psi\in \overline{D_\rho}}|v(\psi)|.
\end{equation*}
The spaces of analytic functions $\mathscr{A}_\rho$ are the same
spaces as in \cite{Moser'67} and that some of their elementary
properties used in the argument were discussed in
\cite{CallejaL10,SuL1}. Notably:
\begin{itemize}
\item
Interpolation inequalities (Hadamard three circle theorem):
\begin{equation}\label{interpolation inequalities}
\|v\|_{\theta\rho +(1-\theta) \rho'} \leq
\|v\|_{\rho}^{\theta} \
\|v\|_{\rho'}^{1-\theta}.
\end{equation}
\item
Cauchy inequalities:
\begin{align}
&\|D^l v\|_{\rho-\delta} \leq C(l,d)\
\delta^{-l}\ \|v\|_\rho, \nonumber\\
& |\hat{v}_k| \leq e^{-2\pi
\ |k|\ \rho}\ \|v\|_\rho.\nonumber
\end{align}
\item
The regularity of the composition:
\begin{prop}\label{composition}
Let $f$ be an analytic function in a domain $\mathscr{D} \subseteq \mathbb{C}, v\in \mathscr{A}_\rho$.
Assume $v(\mathbb{T}^d_\rho)\subseteq \mathscr{D}, \text{dist} (v(\mathbb{T}^d_\rho), \mathbb{C}-\mathscr{D}) \geq \xi>0$.
Then,
\begin{itemize}
\item [(1)] $f\circ v \in \mathscr{A}_\rho$;
\item [(2)] If $\|\tilde{v}\|_\rho < \frac{\xi}{2}$, we have
\[
\| f(v+ \tilde{v}) - f(v) - f'(v) \tilde{v}\|_\rho \leq C \| \tilde{v}\|_\rho^2.
\]
\end{itemize}
\end{prop}
\end{itemize}
The cohomology equations we will have to consider are
different from those studied before and we will present the results in
Section~\ref{sec:twisted}.
\begin{Remark}
The method of proof works also for spaces of functions with finite differentiability.
Indeed, KAM theory is often formulated as an abstract implicit function
theorem for the functional $\mathscr{E}$ acting on spaces of functions
that satisfy some mild properties \cite{Schwartz60, Zehnder75}.
(The paper \cite{CallejaL10} presents an implicit function theorem
well suited for the method in this paper.)
In particular, the method in this paper works as well when
$v$ is considered in Sobolev spaces of high enough
regularity. For simplicity, we will not formulate the finite differentiable version of the results.
\end{Remark}
\subsection{Diophantine condition}\label{Diophantine properties}
We will require that $\Omega$ satisfies the Diophantine condition in $\mathbb{R}^{d-1}$:
\begin{equation}\label{Diophantine}
|k\cdot \Omega-m| \geq \kappa |\tilde{k}|^{-\tau} \qquad \forall~
k\in \mathbb{Z}^{d-1}-\{0\},~m\in \mathbb{Z}.
\end{equation}
Here $\kappa,\tau$ are positive numbers.
We denote by
$\mathscr{D}(\kappa,\tau)$ the
set of $\Omega$ which satisfy \eqref{Diophantine}.
We also denote $\mathscr{D}(\tau) = \cup_{\kappa > 0} \mathscr{D}(\kappa,\tau)$.
It is well known that the set
$\mathscr{D}(\tau)$ with $\tau > d-1$ is of
full $d-1$ dimensional
Lebesgue measure in $\mathbb{R}^{d-1}$.
\subsection{Cohomology equations}\label{cohomology}
It is standard in KAM theory to solve for $v$ given
$\phi$ with zero average in such a way that:
\begin{equation}\label{Cohomology equation}
v(\psi+\Omega)-v(\psi)=\phi(\psi),
\end{equation}
where $\Omega\in\mathscr{D}(\kappa,\tau)$.
We will use equations similar to \eqref{Cohomology equation} frequently with some fixed frequency
$\Omega$. To simplify our notations, we will denote $v(\psi+\Omega)$
and $v(\psi-\Omega)$ as $v_+(\psi)$ and $v_-(\psi)$, respectively. Similar
notations will be used for other functions. We also use $T$ to
represent the translation operators,
i.e. $[T_{\Omega}v](\psi)=v(\psi+\Omega)$.
Estimates for \eqref{Cohomology equation} for Diophantine frequencies
were proved in \cite{Russmann'75, Russmann'76}.
The crucial point of these estimates is that the solution is bounded
in smaller domains and that there are bounds on the solution in the smaller
domains. These estimates have large constants if the loss of domain is small, but the constants can be chosen to be a power of the domain loss (tame estimates). See \eqref{cohomology bounds for analytic fixed}.
Tame estimates (and hence Diophantine conditions) are used in the convergence proofs and
in the KAM theory developed here.
\begin{lemma}\label{estimate lemma for cohomology equation}
Let $\phi \in\mathscr{A}_\rho(\mathbb{T}^{d-1})$ be such that
\begin{equation}\label{normalization phi}
\int_{\mathbb{T}^{d-1}}\phi (\psi)d\psi=0.
\end{equation}
Assume that $\Omega \in \mathscr{D}(\kappa,\tau)$.
Then, there
exists a unique solution $v$ of
\eqref{Cohomology equation} which satisfies
\begin{equation}\label{normalization equation}
\int_{\mathbb{T}^{d-1}} v (\psi)d\psi = 0.
\end{equation}
The solution $v$ is in $\mathscr{A}_{\rho'} $ for
any $0<\rho' < \rho$, and
\begin{equation}\label{cohomology bounds for analytic fixed}
\|v\|_{\rho^\prime}\leq C(d,\tau)\ \kappa^{-1}\
(\rho-\rho^\prime)^{-\tau}\|\phi \|_\rho.
\end{equation}
Furthermore, any distribution solution of \eqref{Cohomology
equation} differs from the solution claimed before by a constant.
If $\phi$ is such that it takes real values for real arguments,
so does $v$.
\medskip
Similarly if we consider analytic functions $\phi\in \mathscr{A}_\rho(\mathbb{T}^d)$ satisfying $\int_{\mathbb{T}^{d-1}} \phi (\psi,\eta)d\psi=0$, then for each $\eta$, we can solve
\begin{equation}\label{Cohomology equation2}
v(\psi+\Omega,\eta)-v(\psi,\eta)=\phi(\psi,\eta).
\end{equation}
The solution $v\in\mathscr{A}_{\rho}(\mathbb{T}^d)$ is analytic in $(\psi,\eta)$ and we have
\begin{equation}\label{cohomology bounds for analytic2}
\|v\|_{\rho^\prime}\leq C(d,\tau)\ \kappa^{-1}\
(\rho-\rho^\prime)^{-\tau}\|\phi\|_\rho.
\end{equation}
\end{lemma}
We note that, as it is well known that obtaining $v$
solving \eqref{Cohomology equation} for
given $\phi$ is very explicit in
terms of Fourier coefficients.
If
\[
\phi(\psi) = \sum_{k \ne 0} \hat \phi_k e^{2 \pi i (k \cdot \psi)}
\]
then, $v$ is given by
\[
v(\psi) = \sum_{k\ne 0} \hat \phi_k (e^{2 \pi i k \cdot \Omega} -1)^{-1} e^{2 \pi i (k \cdot \psi)}.
\]
The above formula for
the solution makes it clear that if $\phi$ is real for real values of its
arguments, so is $v$. Note also that if the function $\phi$ is
discretized in terms of $N$ Fourier coefficients, the computation of
the Fourier coefficients of $v$ takes only $N$ operations.
It also makes it clear how to obtain the tame estimates (with a worse exponent)
for Diophantine frequencies or the results for \eqref{subexponential}.
\subsection{Cohomology equations with non-constant coefficients}
\label{sec:twisted}
In this section we consider a generalization of the above theory for
equations with non-constant coefficients. These equations have also been called
{\sl twisted cohomology equations} in \cite{Herman83}.
Our strategy is similar to \cite{Herman83}, but we will need more details about the dependence of the solutions on the coefficient.
We will be considering equations of the form
\begin{equation}
\label{twisted}
a(\psi) v(\psi + \Omega) - b(\psi) v(\psi) = \lambda + \phi(\psi)
\end{equation}
where $\psi\in\mathbb{T}^{d-1}$ and $a,b$ are fixed functions in $\mathscr{A}_\rho$.
We consider that $\phi\in\mathscr{A}_{\rho}$ is the known data of the equation and
$\lambda\in\mathbb{R}$, $v\in\mathscr{A}_{\rho-\delta}$ are the unknowns we seek for. We will refer to \eqref{twisted} as ``twisted cohomology equations".
We will show that, given some non-degeneracy conditions in
$a$ and $b$, we can obtain estimates for $\lambda$ and $v$.
The idea, already present in \cite{Herman83}, is that we
can rewrite the equation \eqref{twisted} into a constant coefficient equation. We will go through the procedure in details because we will need rather detailed estimates on the effect of $a$ and $b$.
\begin{lemma}\label{lm:twisted}
Suppose
$||a-1||_{\rho}<r<1,~ ||b-1||_{\rho}<r<1$ and that $\Omega$ satisfies the Diophantine condition \eqref{Diophantine}. Then there exist positive real valued functions $\gamma_a, \gamma_b \in \mathscr{A}_{\rho-\delta/2}$ and
real numbers $\bar a$, $\bar b$ such that
\begin{equation} \label{auxiliary1}
\begin{split}
a(\psi) &= {\bar a} \frac{\gamma_a(\psi + \Omega)}{\gamma_a(\psi)}, \\
b(\psi) &= {\bar b} \frac{\gamma_b(\psi)}{\gamma_b(\psi + \Omega)}. \\
\end{split}
\end{equation}
In addition, for the twisted cohomology equation \eqref{twisted}, there exist a unique solution $\lambda\in\mathbb{R}$ and $v\in\mathscr{A}_{\rho-\delta}$ for \eqref{twisted} satisfying the normalization
\begin{equation}\label{normalization for twist cohomology}
\int_{\mathbb{T}^{d-1}}v d\psi=0
\end{equation}
such that
\begin{equation}\label{twist estimate}
\begin{split}
|\lambda|&\leq \frac{||\phi||_{\rho}||\gamma_a(\gamma_b)_+||_{\rho-\delta/2}}{\int_{\mathbb{T}^{d-1}} \gamma_a(\gamma_b)_+d\psi}\\
||\gamma_a\gamma_bv||_{\rho-\delta}&\leq C\delta^{-\tau}
||(\lambda+\phi)\gamma_a(\gamma_b)_+||_{\rho}\qquad \forall~ \delta>0.
\end{split}
\end{equation}
Therefore,
\begin{equation}\label{twist final estimate}
||v||_{\rho-\delta}\leq C||\gamma_a||_{\rho}\left\|1/\gamma_a \right\|_{\rho-\delta}
||\gamma_b||_{\rho}\left\|1/\gamma_b \right\|_{\rho-\delta }\delta^{-\tau}|| ( \lambda+\phi ) ||_{\rho}\qquad \forall~ \delta>0.
\end{equation}
\end{lemma}
\begin{proof}
To show \eqref{auxiliary1}, it suffices to observe that, taking logarithms, \eqref{auxiliary1} is
equivalent to:
\begin{equation} \label{auxiliarylog}
\begin{split}
(\log a)(\psi) &= \log( {\bar a}) + \log(\gamma_a)(\psi + \Omega)
- \log( \gamma_a) (\psi) \\
(\log b)(\psi) &= \log({\bar b})
+ \log( \gamma_b) (\psi) - \log(\gamma_b)(\psi + \Omega) \\
\end{split}
\end{equation}
which are cohomology equations with constant coefficients. Applying Lemma~\ref{estimate lemma for cohomology equation}, we get solutions $\gamma_a,\gamma_b\in\mathscr{A}_{\rho-\delta/2}$, $\bar{a},\bar{b}\in\mathbb{R}$ and estimates
\begin{equation} \label{estimates on log}
\begin{split}
||\log \gamma_a||_{\rho-\delta/2} &\leq C(d,\tau,\kappa)\delta^{-\tau}||\log a-\log\bar{a}||_{\rho}, \\
||\log \gamma_b||_{\rho-\delta/2} &\leq C(d,\tau,\kappa)\delta^{-\tau}||\log b-\log\bar{b}||_{\rho}.\\
\end{split}
\end{equation}
Once we solve the constant coefficient cohomology equation \eqref{auxiliarylog} for $\log{a}, \log{b}, \log{\gamma_a}, \log{\gamma_b}$ we obtain $a, b, \gamma_a, \gamma_b$ by taking exponentials.
This ensures that they are positive for real values of the argument.
We refer to $\bar{a},\bar{b}$ as the average coefficients of the cohomology equation.
Once we have the solution of
\eqref{auxiliary1}, we realize that the
equation \eqref{twisted} is equivalent to
\begin{equation} \label{twisted modified}
\bar{a}(\gamma_a\gamma_b v)_+-\bar{b}(\gamma_a\gamma_b v)=(\lambda+\phi)\gamma_a(\gamma_b)_+
\end{equation}
which is a cohomology equation with constant coefficients. Let us denote $m=\gamma_a\gamma_b v$ for simplicity.
When $\bar{a}=\bar{b}$, we can solve \eqref{twisted modified} using Lemma \ref{estimate lemma for cohomology equation} and get estimates \eqref{twist estimate}.
Note if $\bar{a}\neq \bar{b}$, the equation \eqref{twisted modified} is easier to solve since no small divisors appear.
\begin{itemize}
\item If $\bar{a}> \bar{b}$, then \eqref{twisted modified} is equivalent to
\[
m_+ - \frac{\bar{b}}{\bar{a}} m = \frac{\lambda+\phi}{\bar{a}}\gamma_a(\gamma_b)_+,
\]
namely, $m_+= \frac{\lambda+\phi}{\bar{a}}\gamma_a(\gamma_b)_+ + \frac{\bar{b}}{\bar{a}} m$.
Therefore,
\[
m(\psi)
=\sum_{n=0}^\infty \left(\frac{\bar{b}}{\bar{a}}\right)^{n} \frac{\lambda+\phi\big(\psi - (n+1)\Omega\big)}{\bar{a}}\gamma_a\big(\psi-(n+1)\Omega\big) \gamma_b(\psi-n\Omega)
\]
is a solution of \eqref{twisted modified}. Hence, $\|m\|_\rho \leq C(\bar{a}, \bar{b}) \
||(\lambda+\phi)\gamma_a(\gamma_b)_+||_{\rho}$.
\item If $\bar{a} < \bar{b}$, then \eqref{twisted modified} is equivalent to
\[
\frac{\bar{a}}{\bar{b}}m_+ - m = \frac{\lambda+\phi}{\bar{b}}\gamma_a(\gamma_b)_+,
\]
namely, $m= \frac{\bar{a}}{\bar{b}} m_+ - \frac{\lambda+\phi}{\bar{b}}\gamma_a(\gamma_b)_+ $.
Therefore,
\[
m(\psi)
=- \sum_{n=0}^\infty \left(\frac{\bar{a}}{\bar{b}}\right)^{n} \frac{\lambda+\phi (\psi + n\Omega)}{\bar{b}}\gamma_a(\psi +n\Omega) \gamma_b\big(\psi+(n+1)\Omega\big)
\]
is a solution of \eqref{twisted modified}. Hence, $\|m\|_\rho \leq C(\bar{a}, \bar{b}) \
||(\lambda+\phi)\gamma_a(\gamma_b)_+||_{\rho}$.
\end{itemize}
Note that when $\bar{a}\neq \bar{b}$ we do not have any loss of domain, but the estimates depend on $\bar{a},\bar{b}$. To get estimates uniformly in $a,b$, we need to use the Fourier method.
\end{proof}
The choice of the parameter $\lambda$ so as to achieve the normalization deserves some discussion. In the case $\bar{a}=\bar{b}$ we see that we have to choose $\lambda$ in such a way that the right hand side of \eqref{twisted modified} has zero average. In this case, however, we can choose the average of the solution of \eqref{twisted modified} arbitrarily.
In the case $\bar{a}\neq \bar{b}$, given any $\lambda$, the solution of \eqref{twisted modified} will be unique. Furthermore, the solution will be an affine function of $\lambda$ and, hence, so will be its average.
We see that the derivative of the average of the solution with respect to $\lambda$ is
\[
\frac{d}{d\lambda} \langle m \rangle = \frac{\langle \gamma_a (\gamma_b)_+\rangle}{\bar{a} - \bar{b}}.
\]
In summary, in the equal average coefficient case, the equation requires adjusting one parameter to be solvable, but it gives back one free parameter of solutions.
In the different average coefficient case, we do not require any parameter to ensure the solutions, but the solution is unique.
In both cases, the solutions of twisted cohomology equations require as many parameters as they give back. This allows us to discuss the solutions of the equations which factorize into twisted cohomology equations and they require as many parameters as they give back.
\begin{Remark}
It will be important for subsequent applications that the estimates \eqref{twist estimate} are formulated in terms of $\gamma$. We could have attempted to formulate them in terms of $a$, but this is not practical for subsequent applications. On the other hand, we have formulated the normalization condition \eqref{normalization for twist cohomology} in a way that it is independent of $\gamma$. This will be important in our iterative scheme because $\gamma$ will change from step to step.
In subsequent applications, we will be using Lemma \ref{lm:twisted} when the $a,b$ are changing. It will be important to track the changes of $\gamma_a, \gamma_b$ when we change $a, b$. See estimates \eqref{estimate for log_sigma_c}.
\end{Remark}
\section{The KAM theorem}
\label{KAM}
In this section, we will state precisely the main result of this paper.
\subsection{Statement of the main result}
For the following theorem, we will fix the parameter $\eta$ and omit the subscript $\eta$.
We denote $\langle g\rangle$ the average of some function $g$ and $\tilde{g}=g-\langle g\rangle$.
We also use
the notation:
\[
\mathscr{L} = T_{\Omega}+T_{-\Omega}-2+\partial_{\beta}W((\psi,\eta)+\beta v)+\sigma.
\]
\begin{theorem}\label{analytic KAM theorem}
Let $\alpha\in\mathbb{R}^d$ such that
$\alpha\cdot j\neq0,~j\in\mathbb{Z}^d \setminus \{0\}$ and $\omega\in \mathbb{R}$ be such that $\omega \alpha$ is resonant. Let $W$ be an analytic function defined in a domain $\mathscr{D}\subset\mathbb{C}^d /\mathbb{Z}^d$. Take $\rho>0$ and $0<s<\rho/2$.
Denote:
\begin{eqnarray}
\label{translation}
\mathscr{E}(v,\sigma,\lambda)&=&v_++v_--2v+W((\psi,\eta)+\beta v)+\sigma v + \lambda,\\
\mathscr{F}(v,\sigma,c)&=&(-c+2-\partial_{\beta}W((\psi,\eta)+\beta v)-\sigma)c_+-1.
\end{eqnarray}
We assume:
\begin{itemize}
\item[(H1)] Initial guesses: Let $(v^0, c^0, \sigma^0, \lambda^0)$ be an approximate solution such that $\|\mathscr{E}[v^0,\sigma^0,\lambda^0]\|_\rho \le \epsilon$ and $\|\mathscr{F}[v^0,\sigma^0,c^0]\|_\rho \le \epsilon$.
\item [(H2)] Diophantine properties: There exists a matrix $B\in SL(d, \mathbb{Z})$,
$\Omega \in \mathbb{R}^{d-1}$, $L \in \mathbb{Z}^d$ such that
$B\omega\alpha = (\Omega, 0) + L $ with
\begin{equation*}
|l\cdot \Omega-n| \geq \kappa |l|^{-\tau} \qquad \forall~
l\in \mathbb{Z}^{d-1}-\{0\},~n\in \mathbb{Z}.
\end{equation*}
\item [(H3)] Non-degeneracy conditions:
\begin{equation*}
\begin{split}
&||c^0-1||_{\rho}<M_1<1,\\
&|\sigma^0|<M_2<1,\\
&||W_{v^0}||_{C^2(\mathscr{D})}< M_3<1\\
&||v^0||_{\rho}<m<1,
\end{split}
\end{equation*}
where $W_{v^0}$ is short for $W((\psi,\eta) + \beta v^0)$, $M_1+M_2+M_3<1$ and $m$ depends on $d,\tau,\kappa,M_i,s$.
\item[(H4)] Composition condition: Denote
\begin{equation*}
\mathscr{R}_{v^0}=\{(\psi,\eta)+\beta v^0(\psi),(\psi,\eta)\in\mathbb{T}^d_{\rho}\}.
\end{equation*}
We assume $\mathscr{R}_{v^0}\subset\mathscr{D}$ and $\dist(\mathscr{R}_{v^0},\partial\mathscr{D})\geq 2\xi>0$, where $\xi$ depends on $d,\tau,\kappa,M_i,s,|\beta|_1$.
\end{itemize}
Let $\rho'=\rho-s-\delta<\rho-s$. Let $\epsilon$ be such that
\[
\epsilon\leq\epsilon^*\delta^{4\tau},
\]
where $\epsilon^*$ depends on $d,\tau,\kappa,M,s,|\beta|_1$ and will be specified in the proof.
Then,
there exist functions $v^*,c^*\in\mathscr{A}_{\rho'}$ and numbers
$\sigma^*,\lambda^*\in\mathbb{R}$ such that
\begin{equation}\label{distance}
\begin{split}
& \mathscr{E}[v^*,\sigma^*,\lambda^*]=0,\\
& \mathscr{F}[v^*,\sigma^*,c^*]=0.
\end{split}
\end{equation}
In addition,
\begin{equation}\label{distance2}
\begin{split}
& || v^0 - v^* ||_{\rho'} \le C\delta^{-4\tau}\epsilon,\\
& | \lambda^0 - \lambda^* |
\le C \delta^{-2\tau}\epsilon,\\
& | \sigma^0 - \sigma^* |
\le C \delta^{-2\tau}\epsilon,\\
& || c^0 - c^* ||_{\rho'} \le C \delta^{-4\tau} \epsilon,
\end{split}
\end{equation}
where $C$ depends on $d,\tau,\kappa,M,s,|\beta|_1$. In the rest of the paper, we will use $C$ to denote any constant depending on these parameters.
We also have local uniqueness: Suppose $(v_1,\sigma_1,\lambda_1,c_1)$ and $(v_2,\sigma_2,\lambda_2,c_2)$ satisfying
\begin{align}
&\mathscr{E}[v_1,\sigma_1,\lambda_1]=\mathscr{E}[v_2,\sigma_2,\lambda_2]=0,\\
&\mathscr{F}[v_1,\sigma_1,c_1]=\mathscr{F}[v_2,\sigma_2,c_2]=0,
\end{align}
and the normalization condition
\begin{equation}
\int_{\mathbb{T}^{d-1}}v_1d\psi=\int_{\mathbb{T}^{d-1}}v_2d\psi=0.
\end{equation}
If
\begin{equation}
\max\{||v_2-v_1||_{3\rho/2},|\sigma_2-\sigma_2|,|\lambda_2-\lambda_1|\}<\epsilon^*{\rho},
\end{equation}
then, we have
\begin{equation}
(v_1,c_1,\sigma_1,\lambda_1)=(v_2,c_2,\sigma_2,\lambda_2).
\end{equation}
Finally, we have that the solution depends on the parameter $\eta$ analytically.
\end{theorem}
The Theorem \ref{analytic KAM theorem} is in an ``a-posteriori" format. Given an approximate solution which satisfies some non-degeneracy condition, we establish that there is a true solution nearby.
For the experts in KAM theory, we note that we have two parameters $s,\delta$ measuring the domain loss. The parameter $s$ measures the domain loss in the first step and the parameter $\delta$ measures the domain loss in the subsequent steps. The first step sets up several quantities which affect subsequent steps and are estimated in a different way from the other steps. The steps after the first are all very similar, and the estimates are very similar to the standard Nash-Moser scheme, even if the iterative step is very different.
To produce the approximate solution, we can use a variety of methods.
In the case that $W$ is small, we can take as approximate solution $(v^0, c^0, \sigma^0, \lambda^0) = (0,1,0,0)$ (the solution corresponding to $W=0$) and then (H1) just becomes smallness condition on $W$.
We can use the result of a Lindstedt series as approximate solutions. Then, we obtain a validation of the Lindstedt procedure, and a consequence for the (complex) differentiability of the solutions and the convergence of Lindstedt series (see Section \ref{sec:extensions}). See also \cite{SuZL15}.
Even if we will not discuss it in this paper, one could take as an approximate solution the outcome of a numerical computation and Theorem~\ref{analytic KAM theorem} gives a validation of the numerical result (provided that we check a few ``condition numbers"). Indeed, the proof leads to an efficient algorithm.
\section{Proof of Theorem \ref{analytic KAM theorem}}\label{sec:proof}
\subsection{Outline of the proof}\label{sec:outline of the proof}
The proof of Theorem \ref{analytic KAM theorem} is based on a quadratically convergent iterative method. The convergence of the method will be established by a Nash-Moser argument. Here we give an outline of the proof.
The linearization of the equilibrium equation is a linear second order difference equation. We will solve this linearized equilibrium equation using that it factorizes into two first order difference equations. In our case, we cannot write down the factorization in closed form (as in \cite{SuL1,Rafael'08}). So we will impose an auxiliary equation for the coefficients of the factorized equation. We will call this equation the factorization equation. We will not need to solve exactly the factorization equation at each step, but we will require that there is an approximate solution of the factorization equation with an accuracy comparable to the accuracy of the approximate solution of the equilibrium equation. Then, we will derive an iterative method that improves both of the equilibrium equation and the factorization equation. This is reminiscent of the procedure to study an elliptic torus and its reducibility at the same time. See \cite{Moser'67, Poschel89}. Note without the parameter $\sigma$, we can not find coefficients satisfying the factorization equation.
Even with this novelty, we still have difficulty that the equilibrium equation and the factorization equation are coupled. We do not know how to establish convergence if we solve one equation after another. Therefore we will derive a way to solve the two coupled linearized equations simultaneously. After we factorize the linearized equilibrium equation, we will face three twisted cohomology equations (difference equations with non-constant coefficients), two for the linearized equilibrium equation and one for the linearized factorization equation. We will solve them using the technique in Section \ref{sec:motivation}. This completes one step of iteration.
One small technical complication is that the first step is estimated in a different way from the subsequent steps.
In the first step of the iteration, we compute some auxiliary function related to factorization (the auxiliary functions $\gamma$ related to the twisted cohomology equation) and estimate them from the original data. In subsequent steps, even if we compute the same auxiliary functions, we do not estimate them from the data but we estimate the change on the auxiliary functions induced by the (rather small) changes in the approximate solution.
Under some minor non-degeneracy conditions, we can repeat the process indefinitely. Finally, we will prove that the iterative procedure converges with suitably chosen domains under the analytic norm. This is very standard in KAM theory.
The local uniqueness will be obtained in Section \ref{sec:uniqueness} by showing that the linearized equation admits unique solutions using a standard technique. See \cite{CallejaL'10}.
We also note that the main part of Theorem \ref{analytic KAM theorem} has what is called ``a-posteriori" format in numerical analysis. We show that if there is an approximate enough solution that satisfies some non-degeneracy conditions, then, there is a true solution nearby. It is well known that this a-posteriori format leads to smooth dependence on parameters, bootstrap of regularity and several other consequences.
Finally, the analyticity with respect to parameters(we will use either $\epsilon$ or $\eta$ as parameters) will be a corollary of the existence of the perturbative expansions (see Section \ref{sec:perturbative}). Using the a-posteriori format of Theorem \ref{analytic KAM theorem} and the local uniqueness, it follows that the solution is complex differentiable in the parameters. Hence, it is analytic on the parameters. Furthermore, since the perturbative expansions are the Taylor expansions of an analytic function they converge.
We note that the iterative process described above leads to an efficient algorithm, which we have formulated in Section \ref{sec:algorithm}.
\subsection{Motivation for the iterative step}\label{sec:motivation}
Our goal in this section is to devise a procedure that given an approximate solution produces another approximate
solution with much smaller error and not much worse non-degeneracy conditions. This procedure is done for the equilibrium and factorization equations simultaneously.
\subsubsection{The equilibrium equation}
We consider the initial guess $(v^0,\sigma^0, \lambda^0)$ which solves equation \eqref{external force} with a small error $e$, where $||e||_{\rho}<\epsilon$, i.e.
\begin{equation}
\label{eq:inv}
v^0_++v^0_--2v^0+W((\psi,\eta)+\beta v^0)+\sigma^0v^0 + \lambda^0=e.
\end{equation}
The Newton procedure for the approximate solutions of the equilibrium equation \eqref{eq:inv} requires to find an update $(\hat{v},\hat{\sigma},\hat{\lambda})$ satisfying
\begin{equation}
\label{eq:inv_lin}
\hat{v}_++\hat{v}_--2\hat{v}+\partial_{\beta}W_{v^0}\hat{v}+\hat{\sigma}v^0+\sigma^0\hat{v}+\hat{\lambda}=-e,
\end{equation}
where we denote $\partial_{\beta}W=\beta\cdot \nabla W$ for simplicity and we use $W_{v}$ to indicate $W((\psi,\eta)+\beta v)$.
\subsubsection{The factorization equation}
The equation \eqref{eq:inv_lin} is not easy to solve directly. Nevertheless, we impose that it can be factorized into two first order difference equations with non-constant coefficients. If we accomplish this, we can solve \eqref{eq:inv_lin} using the theory of twisted cohomology equations developed in Section \ref{sec:twisted}.
Therefore we want that the operator $\mathscr{L}$ (recall that we denote $\mathscr{L}=T_{\Omega}+T_{-\Omega}-2+\partial_{\beta}W_{v^0}+\sigma$) is factorized into two first order operators:
\begin{equation} \label{original equilibrium equation}
\mathscr{L}=\mathscr{A}_+\mathscr{A}_-,
\end{equation}
where
\begin{align}
\label{eq:firstordereq}
\mathscr{A}_+&=a(\psi)T_{\Omega}-b(\psi)\\
\mathscr{A}_-&=c(\psi)-d(\psi)T_{-\Omega}
\end{align}
and $a,b,c,d$ are the new unknowns.
A direct calculation shows
\begin{equation}
\mathscr{A}_+\mathscr{A}_-\hat{v}=a(\psi)c_+(\psi)\hat{v}_+-[a(\psi)d_+(\psi)+b(\psi)c(\psi)]\hat{v}+b(\psi)d(\psi)\hat{v}_-.
\end{equation}
Hence we want to choose $a,b,c,d$ satisfying the following equations
\begin{align}\label{eq:coeff}
a(\psi)c_+(\psi)&=1\nonumber,\\
-[a(\psi)d_+(\psi)+b(\psi)c(\psi)]&=-2+\partial_{\beta}W_{v^0}+\sigma^0,\\
b(\psi)d(\psi)&=1.\nonumber
\end{align}
Note that the problem of factorization has always many solutions. If $\mathscr{A}_+\mathscr{A}_-$ is a solution and $g$ is any invertible function, then $\tilde{\mathscr{A}}_+\tilde{\mathscr{A}}_-$ is also a solution, where
\begin{equation}
\label{eq:change}
\begin{split}
\tilde{\mathscr{A}}_+&=\mathscr{A}_+g,\\
\tilde{\mathscr{A}}_-&=g^{-1}\mathscr{A}_-.
\end{split}
\end{equation}
We can use this non-uniqueness to impose some extra normalization condition. In this paper, we will take the normalization
\begin{equation}
\label{eq:normalized coeff}
b(x)=1.
\end{equation}
With condition \eqref{eq:normalized coeff} we can simplify the system (\ref{eq:coeff}) to (after eliminating $a$):
\begin{equation}
(-c+2-\partial_{\beta}W_{v^0}-\sigma^0)c_+=1.
\end{equation}
This is how the operator $\mathscr{F}$ comes into play. Note it is a non-linear, non-local equation which we will have to solve iteratively.
By assumption (H1), we can solve this equation with an error $f$ using some initial guess $c^0$, i.e.
\begin{equation}\label{eq:factorization}
(-c^0+2-\partial_{\beta}W_{v^0}-\sigma^0)c^0_+-1=f
\end{equation}
and we have $||f||_{\rho}<\epsilon$.
Then the Newton procedure for \eqref{eq:factorization} requires solving the following equation for $\hat{c}$:
\begin{equation}
\label{eq:F}
-c^0_+\hat{c}+(-c^0+2-\partial_{\beta}W_{v^0}-\sigma^0)\hat{c}_+-c^0_+\hat{\sigma}-\partial_{\beta}\partial_{\beta}W_{v^0} c^0_+\hat{v}=-f.
\end{equation}
Then $c+\hat{c}$ would be a more accurate solution.
We replace $\mathscr{L}$ with $\mathscr{A}_+\mathscr{A}_-$ in the original equilibrium equation \eqref{eq:inv_lin} to get
\begin{equation}
\label{eq:I}
\mathscr{A}_+\mathscr{A}_-\hat{v}+\hat{\sigma}v^0+\hat{\lambda}=-e.
\end{equation}
\begin{Remark}
Note that the equation \eqref{eq:I} is slightly different from \eqref{eq:inv_lin}, so the $(\hat{v},\hat{\sigma},\hat{\lambda},\hat{c})$ we find is not exactly the solution of the linearized equation of \eqref{external force}. But the only difference between \eqref{eq:I} and \eqref{eq:inv_lin} is a term $f\hat{v}(c_+^0)^{-1}$, which we will show is quadratic in $e,f$ (since $\hat{v}$ will be of the order of $e$).
\end{Remark}
\begin{Remark}
To obtain quadratic convergence, the full Newton step will have to improve both the equilibrium and the factorization equation at the same time. (A moment's reflection shows that improving one equation and then the other does not lead to quadratic convergence.)
We will prove our Newton procedure using \eqref{eq:I} instead of \eqref{eq:inv_lin} and improving the factorization still converges quadratically to a solution of \eqref{external force}.
\end{Remark}
In the following section, we will develop a method to solve \eqref{eq:I} and \eqref{eq:F} simultaneously. In Section \ref{sec:estimation of errors} we will obtain estimates for the size of the error. Then, we will show in Section \ref{sec:convergence} that they lead to an improved solution and that the process can be iterated and converges to a solution.
\subsection{Solving the linearized equations}\label{sec:solve the linearization}
Our goal in this section is to solve \eqref{eq:I} and \eqref{eq:F} simultaneously.
The difficulty arises because the equations \eqref{eq:I} and \eqref{eq:F} are coupled (the unknown $\hat{\sigma}$ appears in both of them). Observe that since $\hat{\sigma}$ appears in an affine way, we can guess that the solutions will be affine in $\hat{\sigma}$. This allows to uncouple the equations.
When we find the corrections $(\hat{v},\hat{\sigma},\hat{\lambda},\hat{c})$, we can write down the updated solution of the equilibrium equation \eqref{external force} and the factorization equation \eqref{eq:factorization}:
\begin{equation}\label{eq:improved solution}
(v^1,\sigma^1,\lambda^1,c^1)=(v^0+\hat{v},\sigma^0+\hat{\sigma},\lambda^0+\hat{\lambda},c^0+\hat{c}),
\end{equation}
which can be used in the next iterative step. We will show that the improved solution \eqref{eq:improved solution} is indeed more approximate (in a smaller domain). Then the procedure can be repeated indefinitely and converges to a solution if the errors are small enough.
Now we write
\begin{align}
\label{eq:v_affine}
\hat{v}&=A+\hat{\sigma}B,\\
\hat{\lambda}&=G+\hat{\sigma}D,\label{eq:lambda_affine}
\end{align}
where $A$ and $B$ are functions, $G$ and $D$ are numbers. All these quantities will be determined shortly.
Then the equation (\ref{eq:I}) becomes
\begin{align}
\label{eq:cohom_inv1}
\mathscr{A}_+\mathscr{A}_-A+G&=-e,\\
\label{eq:cohom_inv2}
\mathscr{A}_+\mathscr{A}_-B+D&=-v^0.
\end{align}
The Newton equation for factorization \eqref{eq:F} after substitution \eqref{eq:cohom_inv1} and \eqref{eq:cohom_inv2} becomes
\begin{equation}
\label{eq:cohom_fac}
-c^0_+\hat{c}+(-c^0+2-\partial_{\beta}W_{v^0}-\sigma^0)\hat{c}_++(-c^0_+-\partial_{\beta}\partial_{\beta} W_{v^0} c_+^0B)\hat{\sigma}=\partial_{\beta}\partial_{\beta}W_{v^0} c^0_+A-f.
\end{equation}
Hence the Newton step improving simultaneously the invariance and factorization equations consists in solving \eqref{eq:cohom_inv1}, \eqref{eq:cohom_inv2}, \eqref{eq:cohom_fac}.
Notice that this system has an upper triangular structure, we can solve \eqref{eq:cohom_inv1}, \eqref{eq:cohom_inv2} for $A,B, G, D$ and then solve \eqref{eq:cohom_fac} for $\hat{c}$ and $\hat{\sigma}$. Once we have $\hat{\sigma}$, we can determine $\hat{v}$ and $\hat{\lambda}$ using \eqref{eq:v_affine} and \eqref{eq:lambda_affine}.
To solve \eqref{eq:cohom_inv1} and \eqref{eq:cohom_inv2}, we observe that each of them can be obtained by solving two cohomology equations with non-constant coefficients.
We can find the zero average solution of \eqref{eq:cohom_inv1} and \eqref{eq:cohom_inv2}. In fact, $G$ is determined so that $A$ has zero average and $D$ is determined so that $B$ has zero average. This ensures that
$\langle \hat{v} \rangle =0$.
We note that the choice of the constants in the solution of two consecutive twisted cohomology equations is a simple extension of the arguments developed at the end of Section \ref{sec:twisted}.
In the case that $\mathscr{A}_+,\mathscr{A}_-$ both have different average coefficients, we see that the average of the solution as a function of $G$ is the composition of two affine functions, hence affine and if the linear part of each of them is not zero, we obtain that the composition of the two affine functions has non-zero derivative.
In the case that $\mathscr{A}_+$ has equal average coefficient, but $\mathscr{A}_-$ does not, we choose $G$ so that $\mathscr{A}_+$ is solvable and choose the average of its solution so that the solution of $\mathscr{A}_-$ satisfies the normalization.
In the case that $\mathscr{A}_+$ has different average coefficient but $\mathscr{A}_-$ has the same average coefficient, we choose $G$ so that the solution produced by $\mathscr{A}_+$ satisfies the compatibility conditions for $\mathscr{A}_-$. Then, the solution of $\mathscr{A}_-$ is defined up to a constant which can be chosen in a unique way to satisfy the normalization condition.
In the case that both $\mathscr{A}_+,\mathscr{A}_-$ have the same average coefficients--which is the case that appears in standard KAM theory--we choose $G$ so that $\mathscr{A}_+$ is solvable. Then, we use the free parameter of the solution of $\mathscr{A}_-$ to ensure the solvability of $\mathscr{A}_-$ and use the free constant in the solution of $\mathscr{A}_-$ to adjust the normalization.
Therefore, we obtain the solution of \eqref{eq:cohom_inv1}. Similar procedure can be done for \eqref{eq:cohom_inv2}.
Now that we have solved \eqref{eq:cohom_inv1} and \eqref{eq:cohom_inv2} we turn to solving \eqref{eq:cohom_fac}.
As long as
\begin{equation}\label{transversal to factorization manifold}
\int_{\mathbb{T}^{d-1} } \left(-c_+^0-\partial_{\beta}\partial_{\beta} W_{v^0} c_+^0B\right)\ d\psi \neq 0,
\end{equation}
which can be proved under some non-degeneracy conditions on the initial guesses, equation (\ref{eq:cohom_fac}) is a twisted cohomology equation. Provided that $-c^0_+$ and $-c^0+2-\partial_{\beta}W_{v^0}-\sigma^0$ are away from zero, we can apply Lemma \ref{lm:twisted} to solve (\ref{eq:cohom_fac}) after $A$ and $B$ are found.
To make the procedure clear, we will write the full algorithm in Section~\ref{sec:algorithm}. In Section \ref{sec:estimation}, we will carefully estimate the errors after one iterative step to show that the errors of the equilibrium and the factorization equation are both reduced quadratically after performing the corrections in the iterative step. Finally, we will prove the convergence rigorously in Section~\ref{sec:convergence}.
\subsection{Formulation of the iterative step}
\label{sec:algorithm}
\begin{alg}
\begin{enumerate}
\item [(1)] Given initial guesses $v^0,\sigma^0, \lambda^0, c^0$, set $a^0=(c^0_+)^{-1}$.
\item [(2)] Calculate the errors $e=\mathscr{E}(v^0,\sigma^0,\lambda^0)$ and $f=\mathscr{F}(v^0,\sigma^0,c^0)$.
\item [(3)] Find $\gamma_{a^0}$ and $\bar{a}^0$ satisfying
\begin{equation*}
\log a^0=\log \bar{a}^0+\log(\gamma_{a^0})_+-\log(\gamma_{a^0}).
\end{equation*}
Also find $\gamma_{c^0}$ and $\bar{c}^0$ satisfying
\begin{equation*}
\log c^0=\log\bar{c}^0+\log(\gamma_{c^0})-\log(\gamma_{c^0})_-
\end{equation*}
\item [(4)] Compute $\bar{a}_+, \bar{b}_+, \gamma_{a_+}, \gamma_{b_+}$ (the average coefficients and the auxiliary functions) for $\mathscr{A}_+$ and $\bar{a}_-, \bar{b}_-, \gamma_{a_-}, \gamma_{b_-}$ for $\mathscr{A}_-$.
\item [(5a)] If $\bar{a}_+\neq\bar{b}_+, \bar{a}_-\neq \bar{b}_-$, compute
\[
\alpha= -\mathscr{A}_+^{-1} \mathscr{A}_-^{-1} e,\
\beta= -\mathscr{A}_+^{-1} \mathscr{A}_-^{-1} 1.
\]
Set $G=-\frac{\langle \alpha \rangle}{\langle \beta \rangle}$. Set $A= \alpha + G\beta$.
\item [(5b)] If $\bar{a}_+=\bar{b}_+, \bar{a}_-\neq \bar{b}_-$, choose $G= -\langle e\rangle$.
Set $\tilde{\alpha} = - \mathscr{A}_+^{-1}(e+ G)$ and $\langle \tilde{\alpha} \rangle =0$.
Set $\beta_1 = \mathscr{A}_-^{-1}\tilde{\alpha}, \beta_2= -\mathscr{A}_-^{-1} 1$.
Set $A= \beta_1 - \beta_2 \frac{\langle \beta_1 \rangle}{\langle \beta_2 \rangle}$.
\item [(5c)] If $\bar{a}_+\neq\bar{b}_+, \bar{a}_-= \bar{b}_-$, choose $\alpha_1 = - \mathscr{A}_+^{-1} e, \alpha_2 = - \mathscr{A}_+^{-1} 1$.
Set $G= \frac{\langle \alpha_1 \rangle}{\langle \alpha_2 \rangle}$.
Set $A=\mathscr{A}_-^{-1}(\alpha_1 - G\alpha_2)$ and $\langle A\rangle =0$.
\item [(5d)] If $\bar{a}_+ = \bar{b}_+, \bar{a}_-= \bar{b}_-$, choose $G= - \langle e\rangle$.
Set $\alpha=\mathscr{A}_+^{-1} (e- \langle e\rangle)$ and $\langle \alpha\rangle=0$.
Set $A= \mathscr{A}_-^{-1} \alpha$ and $\langle A\rangle =0$.
\begin{comment}
\item Choose $G=-\langle e\gamma_{a^0}\rangle/\langle\gamma_{a^0}\rangle$, find $Y$ in
\begin{equation*}
\bar{a}^0(\gamma_{a^0}Y)_+-\gamma_{a^0}Y=(-e-G)\gamma_{a^0}
\end{equation*}
\item Set $\tilde{Y}=Y-\langle Y\gamma_{c^0}\rangle/\langle\gamma_{c^0}\rangle$. Find $A$ in
\begin{equation*}
\bar{c}\gamma_{c^0}A-(\gamma_{c^0}A)_-=\tilde{Y}(\gamma_{c^0})_-
\end{equation*}
where we set $\langle A\rangle=0.$
\end{comment}
\item [(6)] Find $B$ and $D$ in a similar way as we found $A$ and $G$. Also set $\langle B\rangle=0.$
\item [(7)] Find $\hat{\sigma}$ and $\hat{c}$ by solving (\ref{eq:cohom_fac}).
\item [(8)] Set $\hat{v}=A+\hat{\sigma}B$ and $\hat{\lambda}=G+\hat{\sigma}D$.
\item [(9)] Update $u,\lambda, \sigma$ and $c$ and repeat the steps.
\end{enumerate}
\end{alg}
\begin{Remark}
When we apply repeatedly the iterative steps to obtain estimates, it would be advantageous in Step 3 to use the information we have on $\gamma_a$ computed in the previous steps, because it leads to better estimates.
In the first step, we can only use Cauchy estimates on the initial data. See Section \ref{sec:outline of the proof}.
\end{Remark}
\begin{Remark}
The iterative method described above achieves
quadratic convergence (see Section \ref{sec:convergence}). It only entails performing algebraic operations among functions, composing them,
taking derivatives and solving the cohomology equations.
If we discretize a function by the values of the point in a grid of $N$ points and
(redundantly) $N$ Fourier coefficients,
we note that each of the operations above requires $N$ operations either in the
grid representation or in the Fourier space representation. If we obtain either the Fourier or
the real space representation for a function, we can obtain the other representation using the
FFT algorithm that requires $N\log(N)$ operations.
Hence the method in Section~\ref{sec:proof} achieves quadratic convergence but no matrix inversion (or storage) is
required. It only requires $O(N)$ storage and $O(N \log(N))$ operations per step.
This has already been observed in \cite{CallejaL09} in the periodic case
(both for short range and for long range interactions) and \cite{SuL1}
for the quasi-periodic non-resonant case. Numerical implementations in the
non-resonant periodic case were carried out in \cite{BlassL}.
We have not implemented the above algorithm, but we think it would be interesting to do so.
\end{Remark}
\subsection{Estimates on the corrections}
\label{sec:estimation}
Denote $V^0=-c^0+2-\partial_{\beta}W_{v^0}-\sigma^0$. In the first step, we argue as follows. From \eqref{estimates on log}, it is clear that all the quantities $||\gamma_{c^0}||_{\rho-\delta/4},||\gamma^{-1}_{c^0}||_{\rho-\delta/4},||\gamma_{(c^0)^{-1}}||_{\rho-\delta/4},$ $||\gamma^{-1}_{(c^0)^{-1}}||_{\rho-\delta/4},||\gamma_{V^0}||_{\rho-\delta/4}$ and $||\gamma_{V^0}^{-1}||_{\rho-\delta/4}$ are bounded by some constant which depends only on $d,\tau,\kappa,M,s$. We denote the bound by $E$.
In subsequent steps, we will just assume that we have the bounds $E$ for the above auxiliary quantities. These bounds will be derived from the bounds in the first step.
Apply the estimate \eqref{twist final estimate} of twisted cohomology equations for equations (\ref{eq:cohom_inv1}) and (\ref{eq:cohom_inv2}), also recall the averages of $A$ and $B$ both vanish we have
\begin{align}
||A||_{\rho-\delta/2}&\le CE^4\delta^{-2\tau}||\tilde{e}||_{\rho}=C\delta^{-2\tau}||\tilde{e}||_{\rho},\\
||B||_{\rho-\delta/2}&\le CE^4\delta^{-2\tau}||\tilde{v}^0||_{\rho}=C\delta^{-2\tau}||\tilde{v}^0||_{\rho},\\
|G| & \le CE^4||e||_{\rho}\le C\epsilon,\\
|D| & \le CE^4||v^0||_{\rho}\le C.
\end{align}
Choose $m$ small, we can make $||B||_{\rho-\delta/2}<1$. Then
\begin{equation}
||(-c^0_+-\partial_{\beta}\partial_{\beta}W_{v^0}c^0_+B)^{-1}||_{\rho-\delta/4}\le (1-M_1)^{-1}(1-M_2)^{-1}.
\end{equation}
This guarantees $\int_{\mathbb{T}^{d-1}} \left(\right.-c^0_+-\partial_{\beta} \partial_{\beta} W_{v^0}c^0_+B \left.\right) d\psi\neq 0$. Therefore the cohomology equation (\ref{eq:cohom_fac}) is solvable.
The estimate for (\ref{eq:cohom_fac}) gives
\begin{equation}\label{eq:estimate for sigma_hat}
\begin{split}
|\hat{\sigma}|&\le C(||A||_{\rho-\delta/2}+||f||_{\rho})\\
&\le C\delta^{-2\tau}\epsilon+C\epsilon\\
&\le C\delta^{-2\tau}\epsilon,
\end{split}
\end{equation}
and
\begin{equation}\label{estimate for c_hat}
\begin{split}
||\hat{c}||_{\rho-\delta}&\le CE^4\delta^{-\tau}(||A||_{\rho-\delta/2}+||f||_{\rho})\\
&\le C\delta^{-3\tau}\epsilon.
\end{split}
\end{equation}
From (\ref{eq:v_affine}), we have the estimate
\begin{equation}\label{eq:estimate for v_hat}
\begin{split}
||\hat{v}||_{\rho-\delta/2}&\le ||A||_{\rho-\delta/2}+|\hat{\sigma}|||B||_{\rho-\delta/2}\\
&\le C\delta^{-4\tau}\epsilon.
\end{split}
\end{equation}
Also, from the estimates for $G,D$ and $\sigma$, we have
\begin{equation}\label{estimate for v_hat}
|\lambda|\le C\epsilon+C\delta^{-2\tau}\epsilon\le C\delta^{-2\tau}\epsilon .
\end{equation}
\subsection{Estimates on the improved error}\label{sec:estimation of errors}
Next we estimate the new error for both equilibrium and factorization equation after one iteration step.
We need the Assumption (H4) to make sure the compositions can be done in appropriate domains. Recall that $\|W_{v^0}\|_{\mathscr{D}}<M$. Also note $\hat{v}$ is relatively smaller than $v^0$. From \eqref{estimate for v_hat}, we see that as long as $\epsilon$ is small enough the range of $v+\hat{v}$ is inside the domain of $W$. So we can choose proper $\xi$ such that $\dist(\mathscr{R}_{v^0},\partial\mathscr{D})\geq \xi>0$.
For the subsequent iterative step, we note that if $\delta^{-4\tau}\epsilon$ is small enough, $||\hat{v}||_{\rho-\delta}$ will also be small. So that under smallness condition on $\delta^{-4\tau}\epsilon$, we can ensure that $v+\hat{v}$ is well inside the domain of definition of $W$. The linear estimates are valid for all $\delta$'s, but in order to ensure that we can apply the non-linear estimates, we have to choose $\delta$'s (the domain loss) in such a way that $\delta^{-4\tau}\epsilon$ is small.
It is standard in KAM theory (and we will do it later in section \ref{sec:convergence}) that one can choose domain losses $\delta_n$ in such a way that the composition condition is met, such that $\delta_n$'s go to zero fast enough, so there is still a domain left. Therefore, the composition will remain in the proper domain for all iterative steps and the procedure will converge in a non-trivial domain.
Using the Taylor expansions (see Proposition \ref{composition}) and the equations \eqref{eq:inv} and \eqref{eq:factorization} for the initial guesses, we have:
\begin{equation}
\begin{split}
\mathscr{E}[v^0+&\hat{v},\sigma^0+\hat{\sigma},\lambda^0+\hat{\lambda}] \\
=& \mathscr{E}[v^0,\sigma^0,\lambda^0]+\hat{v}_++\hat{v}_--2\hat{v}+\hat{\sigma}v^0+\sigma^0\hat{v}+\hat{\sigma}\hat{v}+\hat{\lambda}-W_{v^0}+W_{v^0+\hat{v}}\\
=& e+\mathscr{A}_+\mathscr{A}_-\hat{v}-f\hat{v}-(\partial_{\beta}W_{v^0}+\sigma^0)\hat{v}+\hat{\sigma}v^0+\sigma^0\hat{v}+\hat{\sigma}\hat{v}+\hat{\lambda}\\
&-W_{v^0}+W_{v^0+\hat{v}}\\
=& -f\hat{v}+\hat{\sigma}\hat{v}+(W_{v^0+\hat{v}}-W_{v^0}-\partial_{\beta}W_{v^0}\hat{v}),
\end{split}
\end{equation}
and
\begin{equation}
\begin{split}
\mathscr{F}[v^0+&\hat{v},\sigma^0+\hat{\sigma},c^0+\hat{c}] \\
= & \mathscr{F}[v^0,\sigma^0,c^0]+[-(c^0+\hat{c})+2-\partial_{\beta}W_{v^0}-(\sigma^0+\hat{\sigma})]\hat{c}_+\\
&+[-\hat{c}-\partial_{\beta}W_{v^0+\hat{v}}+\partial_{\beta}W_{v^0}-\hat{\sigma}]c^0_+\\
= & [-\hat{c}-(\partial_{\beta}W_{v^0+\hat{v}}-\partial_{\beta}W_{v^0})-\hat{\sigma}]\hat{c}_+\\
&-[(\partial_{\beta}W_{v^0+\hat{v}}-\partial_{\beta}W_{v^0})c^0_+-\partial_{\beta}\partial_{\beta}W_{v^0}c^0_+\hat{v}].
\end{split}
\end{equation}
Take $0<\delta<\rho$, we have
\begin{align}
||\mathscr{E}[v^0+\hat{v},&\sigma^0+\hat{\sigma},\lambda^0+\hat{\lambda}]||_{\rho-\delta}&\\\nonumber
\le & ||f\hat{v}||_{\rho-\delta}+||\hat{\sigma}\hat{v}||_{\rho-\delta}+||W_{v^0+\hat{v}}-W_{v^0}-\partial_{\beta}W_{v^0}\hat{v}||_{\rho-\delta}\\\nonumber
\le & ||f||_{\rho-\delta}||\hat{v}||_{\rho-\delta}+||\hat{\sigma}||_{\rho-\delta}||\hat{v}||_{\rho-\delta}+C||\hat{v}||^2_{\rho-\delta}.
\end{align}
Also,
\begin{align}
||\mathscr{F}[v^0+\hat{v},&\sigma^0+\hat{\sigma},c^0+\hat{c}]||_{\rho-\delta}&\\\nonumber
\le & (||\hat{c}||_{\rho-\delta}+C||\hat{v}||_{\rho-\delta}+||\hat{\sigma}||_{\rho-\delta})||\hat{c}||_{\rho-\delta}+C||\hat{v}||^2_{\rho-\delta}||c^0||_{\rho-\delta}.
\end{align}
Now it is clear the new errors are quadratic in $f,\hat{v},\hat{\sigma},\hat{c}$, multiplied by the domain loss to a negative power. This is what are called ``tame estimates'' in KAM theory.
The final estimates for both equilibrium and factorization equations on the updated solutions are
\begin{equation}
||\mathscr{E}[v^0+\hat{v},\sigma^0+\hat{\sigma},\lambda^0+\hat{\lambda}]||_{\rho-\delta}\le C\delta^{-8\tau}\epsilon^2
\end{equation}
and
\begin{equation}
||\mathscr{F}[v^0+\hat{v},\sigma^0+\hat{\sigma},c^0+\hat{c}]||_{\rho-\delta}\le C\delta^{-8\tau}\epsilon^2.
\end{equation}
Therefore we get updates for $c,v,\sigma$ and $\lambda$ which reduce the error of both the equilibrium equation and the factorization equation quadratically under the condition that all the bounds we derived before still hold for all iterative steps. We can prove the convergence following standard procedure, which we give in the following.
\subsection{Proof of convergence}
\label{sec:convergence}
The main effect of the iterative step is to reduce the error in \eqref{eq:I} and \eqref{eq:F}. But it could also deteriorate the constants in the non-degeneracy assumptions when it improves the solution.
Therefore the first goal of this section is to show that the constants in non-degeneracy assumptions of the solution do not deteriorate much and that the deterioration can be estimated by the error. In this case, as we will see, the convergence can be proved by choosing suitable domains (this is the choice we will do first) for each iterative step as is standard in KAM theory.
We will use
subscript $n$ to denote the quantities $\rho, \delta$ and $\epsilon$ after application of the iterative step $n$ times, while we use superscript $n$ for $v,c,\sigma,\lambda$ and $B$. We take
\begin{equation}
\rho_0=\rho,\rho_1=\rho-s-\delta_0,\delta_n=\delta_0\cdot 2^{-n} \quad \text{and}\quad \rho_{n+1}=\rho_{n}-\delta_n.
\end{equation}
Denote $\epsilon_n=||\mathscr{E}(v^n,\delta^n,\lambda^n)||_{\rho_n}$ and $\tilde{\epsilon}_n=||\mathscr{F}(v^n,\sigma^n,c^n)||_{\rho_n}$. We will prove the following holds for all iterative steps by induction.
\begin{description}
\item[(B1)] All the quantities $||v^n||_{\rho_n},||c^n-1||_{\rho_n},|\sigma^n|,||W_{v^0}||_{\rho_n},||B^n||_{\rho_n},||\gamma_{c^n}||_{\rho_n},$\\$||\gamma_{c^n}^{-1}||_{\rho_n},||\gamma_{V^n}||_{\rho_n},||\gamma_{V^n}^{-1}||_{\rho_n}$, and $||(-c^0_+-\partial_{\beta}\partial_{\beta}W_{v^0}c^0_+B)^{-1}||_{\rho_n}$ are still bounded uniformly in $n$. To be precise, for any quantity $A^0<E$, we will prove $|A^n-A^{n-1}|<E\cdot2^{-n} ) $, therefore, $A^n<2E$ for any $n$.
\item[(B2)] Denote $\mathscr{R}_{v^n}=\{(\psi,\eta)+\beta v^n,(\psi,\eta)\in \mathbb{T}^d_{\rho_n}\}$. Then $\dist(\mathscr{R}_{v^n},\partial\mathscr{D})$ is also bounded by $\xi^*$.
\item[(B3)] $\epsilon_{n+1}\le (C\epsilon_0)^{2^{n+1}},\quad \tilde{\epsilon}_{n+1}\le (C\tilde{\epsilon}_0)^{2^{n+1}}$.
\end{description}
Note that as a consequence of (B3), and the choices of $\delta_n$, we obtain that the composition assumption holds
if $\epsilon_0$ is small enough.
The first step is already shown in Section \ref{sec:estimation}. Now we assume the first $n$ steps are proved.
We first prove (B1). We only prove the bound
\begin{equation}
\label{eq:bounds}
||\gamma_{c^n}-\gamma_{c^{n-1}}||_{\rho_n-\delta/4}<E\cdot2^{-n}.
\end{equation}
The proof of the other bounds is similar (up to changing the symbols).
From \eqref{estimates on log}, we have
\begin{equation}\label{estimate for log_sigma_c}
||\log \gamma_{c^{n}}-\log \gamma_{c^{n-1}}||_{\rho_{n}-\delta_n/4}\leq C\delta_n^{-\tau}||\log c^{n}-\log c^{n-1}||_{\rho_n}.
\end{equation}
From estimate \eqref{estimate for c_hat} for $\hat{c}$, we have
\begin{equation}
||\hat{c}^{n}||_{\rho_{n}}\le C\delta_n^{-3\tau}\epsilon_n.
\end{equation}
As a result, we have
\begin{equation}\label{estimate for log_c}
||\log c^{n}-\log c^{n-1}||_{\rho_{n}}\le C\delta_n^{-3\tau}\epsilon_n.
\end{equation}
Combining \eqref{estimate for log_sigma_c} and \eqref{estimate for log_c}, we have
\begin{equation}
||\log \gamma_{c^{n}}-\log \gamma_{c^{n-1}}||_{\rho_{n}-\delta_n/4}\le C\delta_n^{-4\tau}\epsilon_n.
\end{equation}
Therefore we have (using that $\gamma_{c^{n}}$ are uniformly bounded)
\begin{equation}
||\gamma_{c^{n}}-\gamma_{c^{n-1}}||_{\rho_{n}-\delta_n/4}\le C\delta_n^{-4\tau}\epsilon_n.
\end{equation}
Note for (B2), we only need bounds for $v^n$, which is easy to prove.
Now recall $\epsilon_n<(C\epsilon_0)^{2^n}$. Therefore if we choose $\epsilon_0$ such that $\delta^{-4\tau}\epsilon_0$ is small enough, we can guarantee \eqref{eq:bounds} (i.e. (B2)). (B3) can be shown as follows
\begin{equation}
\begin{split}
\epsilon_{n}&\le C\delta_{n-1}^{-8\tau}\epsilon^2_{n-1}\\
&=C\delta_0^{-8\tau}(2^{8\tau})^{n-1}\epsilon_{n-1}^2\\
&\le (C\delta_0^{-8\tau})^{1+2+\cdots+2^{n-1}}(2^{8\tau})^{(n-1)+(n-2)\cdot 2+\cdots+1\cdot (n-1)}\epsilon_0^{2^{n}}\\
&\le (C\delta_0^{-8\tau}2^{8\tau}\epsilon_0)^{2^{n}}\\
&= (C2^{8\tau}\epsilon_0)^{2^{n}}.
\end{split}
\end{equation}
Similar estimates also hold for $\tilde{\epsilon}_n$.
\subsection{Proof of local uniqueness}\label{sec:uniqueness}
Suppose we have two solutions $(v_1,\sigma_1,\lambda_1,c_1)$ and $(v_2,\sigma_2,\lambda_2,c_2)$ of both the equilibrium and the factorization equations, which also satisfy the non-degeneracy conditions. We have
\begin{equation}
\mathscr{E}[v_1,\sigma_1,\lambda_1]=\mathscr{E}[v_2,\sigma_2,\lambda_2]=0.
\end{equation}
Then we can write
\begin{equation}\label{identity to prove uniqueness}
\begin{split}
\mathscr{E}[v_2,\sigma_2,\lambda_2]&=\mathscr{E}[v_1,\sigma_1,\lambda_1]+D\mathscr{E}[v_1,\sigma_1,\lambda_1]\cdot (v_2-v_1,\sigma_2-\sigma_1,\lambda_2-\lambda_1)+R^2\\
&=D\mathscr{E}[v_1,\sigma_1,\lambda_1]\cdot (v_2-v_1,\sigma_2-\sigma_1,\lambda_2-\lambda_1)+R^2=0,
\end{split}
\end{equation}
where $R^2$ is the Taylor remainder for $\mathscr{E}$.
Another way to read this identity \eqref{identity to prove uniqueness} is to say that $v_1-v_1,\lambda_2-\lambda_1,\sigma_2-\sigma_1$ is a solution of the equation
\begin{equation}
\label{eq:rq:I_2}
\mathscr{A}_+\mathscr{A}_-(v_2-v_1)+(\sigma_2-\sigma_1)v^1+(\lambda_2-\lambda_1)=-R^2.
\end{equation}
Similarly, we have
\begin{equation}
\label{eq:rq:F_2}
-c^1_+(c_2-c_1)+(-c^1+2-\partial_{\beta}W_{v^1}-\sigma^1)(c_2-c_1)_+-c^1_+(\sigma_2-\sigma_1)-\partial_{\beta}\partial_{\beta}W_{v^1} c^1_+(v_2-v_1)=-\tilde{R}^2.
\end{equation}
We have shown in Section \ref{sec:estimation} that the solutions of \eqref{eq:rq:I_2} and \eqref{eq:rq:F_2} satisfying the normalization condition are unique and satisfy the bounds \eqref{eq:estimate for sigma_hat} and \eqref{eq:estimate for v_hat}. So we have
\begin{equation}
||v_2-v_1||_{\frac{\rho}{2}}+||c_2-c_1||_{\frac{\rho}{2}}+|\sigma_2-\sigma_1|\leq C\rho^{-4\tau}||R^2||_{\rho}\leq C\rho^{-4\tau}(||v_2-v_1||_{\rho}+||c_2-c_1||_{\rho}+|\sigma_2-\sigma_1|)^2 .
\end{equation}
Since $||v_2-v_1||_{\rho}$, $||c_2-c_1||_{\rho}$ and $|\sigma_2-\sigma_1|$ are all small, we have
\begin{equation}
\begin{split}
||v_2-v_1||_{\frac{\rho}{2}}\leq &C \rho^{-4\tau} ||v_2-v_1||^2_{\rho}\\
\leq & C \rho^{-4\tau} ||v_2-v_1||_{\frac{\rho}{2}}||v_2-v_1||_{\frac{3\rho}{2}}.
\end{split}
\end{equation}
The last inequality uses the interpolation inequality \eqref{interpolation inequalities}.
Therefore, as long as $\rho^{-4\tau}||v_2-v_1||_{\frac{3\rho}{2}}$ is sufficiently small (depending on the properties of the auxiliary function $c$ in the factorization equation), we have $||v_1-v_2||_{\frac{\rho}{2}}=0$, which also implies $\sigma_2=\sigma_1$ and $\lambda_2=\lambda_1$.
\section{Consequence of Theorem~\ref{analytic KAM theorem} and its
proof}\label{sec:extensions}
\subsection{Perturbative series around any solution}
\label{sec:perturbative}
In this section we present methods to compute formal power
series expansions. In Section~\ref{sec:convergenceseries} we will show that, under Diophantine
conditions on the frequency, they converge on a sufficiently small domain.
Assume:
\begin{description}
\item [{\bf A0}] We are given a family of interaction potentials $W^\mu$
indexed
by an external parameter $\mu$ which can be complex.
We assume that $W^\mu$ is analytic in both its arguments $\psi,\eta$ and the parameter $\mu$.
\item
[{\bf A1}] For some parameter $\mu_0$ we have a solution
$v^0, \sigma^0, \lambda^0 $ of $\eqref{external force}$.
We assume that $v^0 \in \mathscr{A}_\rho$ for some $\rho > 0$.
\item
[{\bf A2}] The operator $\mathscr{L}_{v^0}$, the linearization
of the equilibrium equation factorizes at $\mu_0$ into two first order operators.
\end{description}
Our goal is to find a perturbative expansion of the solutions of \eqref{external force} in formal
power series of $\mu - \mu_0$. Later, we will show
that these perturbative expansions are convergent following
an argument of \cite{Moser'67} which is made much easier by the
a-posteriori format of Theorem~\ref{analytic KAM theorem}. See Section~\ref{sec:convergenceseries}.
We seek
\begin{equation}\label{generalexpansion}
\begin{split}
& v^\mu = v^0 + \sum_{n > 0} (\mu - \mu_0)^n v^n \\
& \sigma^{\mu} = \sigma^0+\sum_{n > 0}(\mu-\mu_0)^n\sigma^n\\
& \lambda^\mu = \lambda^0 + \sum_{n > 0} (\mu - \mu_0)^n \lambda^n
\end{split}
\end{equation}
in such a way that, when we substitute it in \eqref{external force}
and expand (formally) in powers of $(\mu - \mu_0)^n$ we obtain that the coefficients of same powers match.
Note that this generalizes the standard Lindstedt series, which is a particular case
of the expansion in the case that the family is just $W^\mu = \mu W$ and that we
expand near $\mu_0 = 0$.
If we substitute \eqref{generalexpansion}
in \eqref{external force}, expand in powers of
$\mu - \mu_0$, the coefficient of order $n$ has the form:
\begin{equation}\label{generalordern}
\begin{split}
&v^n_\eta(\psi + \Omega) +
v^n_\eta(\psi - \Omega)
+ v^n_\eta(\psi) (-2 + \partial_\beta W^{\mu_0}((\psi, \eta) ) +\sigma^0v^n+\sigma^nv^0+\lambda^n = R_n
\end{split}
\end{equation}
where $R_n$ is a polynomial expression in $v^0,\ldots, v^{n-1}$.
The main observation is that the equation \eqref{generalordern} is
precisely the Quasi-Newton equations \eqref{eq:I} which can be solved
by factorization. Note that we get the perturbative series to all
orders only assume that $W^{\mu}$ factorizes at $\mu=\mu_0$ and the
expansion series we get has $\sigma=0$ to all orders.
We note that to solve the cohomology equations without any quantitative
estimates, as shown in \cite{SuZL15}, it suffices to assume
\begin{equation} \label{subexponential}
\lim_{N\to\infty} \frac{1}{N} \sup_{|\tilde{k}| \le N, m \in \mathbb{Z}} \bigg| \ln |\tilde{k}\cdot \Omega-m| \bigg| = 0,
\end{equation}
which is weaker than Diophantine and, indeed weaker than Brjuno conditions.
We have, therefore established the following:
\begin{lemma} \label{generalexpansionexists}
Under the assumptions {\bf A0}, {\bf A1}
and that $\Omega$ satisfies the quantitative conditions.
Then, we can find formal power series as in \eqref{generalexpansion}
such that for any
$N \in \mathbb{N}$ and any $\rho'$, $0 < \rho' < \rho$, we have
\[
\left\| \mathscr{E}_\mu \left[ \sum_{n \le N} v^n (\mu -\mu_0)^n , \sum_{n \le N} \lambda^n (\mu -\mu_0)^n \right]\right\|_{\rho'}
\le C_{N,\rho'} |\mu - \mu_0|^{N+1}.
\]
Furthermore, it is possible to find
a formal power series
that satisfies the normalization
\[
\langle v^n_\eta \rangle = 0 \quad \quad n =1,\ldots N.
\]
We refer to it as the \emph{``normalized perturbative expansion''}.
This normalized perturbative expansion is unique.
\end{lemma}
Similarly, we can obtain existence of perturbation expansions
for both the equilibrium and the factorization equations.
In analogy with \eqref{generalexpansion}, we seek expansions
\begin{equation}\label{generalexpansion2}
\begin{split}
& v^\mu = v^0 + \sum_{n > 0} (\mu - \mu_0)^n v^n \quad \quad
\lambda^\mu = \lambda^0 + \sum_{n > 0} (\mu - \mu_0)^n \lambda^n \\
& c^\mu = c^0 + \sum_{n > 0} (\mu - \mu_0)^n c^n \quad \quad
\sigma^\mu = \sigma^0 + \sum_{n > 0} (\mu - \mu_0)^n \sigma^n \\
\end{split}
\end{equation}
in such a way that the equilibrium and factorization equations are solved.
The equations for order $n$ are
\begin{equation}\label{ordernboth}
\begin{split}
&v^n_++v^n_-+(-2+\partial_{\beta}W^{\mu_0}_{v^0}+\sigma^0)v^n+v^0\sigma^n+\lambda^n=R_n \\
&-c_+^0c^n+(-c^0+2-\partial_{\beta}W_{v^0}^{\mu^0}-\sigma^0)c^n_+-c_+^0\sigma^n-\partial_{\beta}\partial_{\beta}W_{v^0}^{\mu_0}c_+^0v^n=\tilde{R_n}.
\end{split}
\end{equation}
From the proof of the KAM theorem, it's clear these can be solved to all orders assuming that $W^{\mu}$ factorizes at $\mu=\mu_0$. The series we get do not have $\sigma=0$, which is different from the series when we do not require the factorization for all orders.
We have, therefore established the following:
\begin{lemma} \label{generalexpansionexists2}
Under the assumptions {\bf A0}, {\bf A1}, {\bf A2}
and that $\Omega$ satisfies the quantitative conditions.
Then, we can find formal power series as in \eqref{generalexpansion2}
such that for any
$N \in \mathbb{N}$ and any $\rho'$, $0 < \rho' < \rho$, we have
\[
\left\| \mathscr{E}_\mu \left[ \sum_{n \le N} v^n (\mu -\mu_0)^n , \sum_{n \le N} \sigma^n (\mu -\mu_0)^n, \sum_{n \le N} \lambda^n (\mu -\mu_0)^n \right]\right\|_{\rho'}
\le C_{N,\rho'} |\mu - \mu_0|^{N+1}
\]
and
\[
\left\| \mathscr{F}_\mu \left[ \sum_{n \le N} v^n (\mu -\mu_0)^n , \sum_{n \le N} \sigma^n (\mu -\mu_0)^n, \sum_{n \le N} c^n (\mu -\mu_0)^n \right]\right\|_{\rho'}
\le C_{N,\rho'} |\mu - \mu_0|^{N+1}.
\]
Furthermore, it is possible to find
a formal power series
that satisfies the normalization
\[
\langle v^n_\eta \rangle = 0 \quad \quad n =1,\ldots N.
\]
We refer to it as the \emph{``normalized perturbative expansion''}.
This normalized perturbative expansion is unique.
\end{lemma}
The above results could be interpreted in a more geometric way. We have proved convergence for the perturbative expansion involving also the counterterms that ensure the factorization. Given a potential $W$ small, the main result shows that we can find a locally unique counterterm $\sigma$, which is a functional on $W$ so that the equation factorizes.
The set $\{W(v) + \sigma(W) v\}$ can be interpreted as a codimension 1 manifold (the factorization manifold) in the space of potentials.
The perturbative expansions in Lemma \ref{convergence of the formal expansion} can be interpreted as the perturbative expansions for a path of potentials indexed by $\epsilon$ but requiring that $W$ stays on the factorization manifold.
In this geometric interpretation, the condition \eqref{transversal to factorization manifold} that among $\sigma$ we can achieve the factorization can be interpreted geometrically as saying that the direction (in the space of potentials) given by $\sigma v$ are transversal to the factorization manifold $F$. Note that this factorization manifold may depend on $\omega$.
\subsection{Convergence of Lindstedt series for the equilibrium and factorization equations}
\label{sec:convergenceseries}
\begin{theorem}\label{convergence of the formal expansion}
Assume that the conditions of Lemma~\ref{generalexpansionexists2} hold
and that $\Omega$ is Diophantine. Then, the normalized formal series
obtained in Lemma~\ref{generalexpansionexists2}
is convergent.
\end{theorem}
Note that a particular case of the above result is the convergence of the Lindstedt series
when the frequency is Diophantine.
We need the frequency to be Diophantine because, as we will see, the proof
uses repeatedly the KAM theorem. It seems quite possible that for the frequencies
that satisfy \eqref{subexponential} but not \eqref{Diophantine}, for many perturbations, it is
possible to obtain perturbative expansions to all orders, which nevertheless do not converge.
\begin{proof}
Without loss of generality, we assume that $v^0$ satisfies \eqref{normalization}.
Using the KAM theorem (Theorem \ref{analytic KAM theorem}),
we obtain that for all $\mu$ in a small ball centered on $\mu_0$ there exists a (unique) normalized
solution. (It suffices to note that $v^0$ is a sufficiently approximate solution for
all $\mu$ close to $\mu_0$.)
Now, given any $\tilde \mu$ in this small ball, we can obtain a perturbative expansion
in powers of $\mu - \tilde \mu$.
We consider the first term of the expansion $v^1_{\tilde \mu}$.
We remark that it will be uniformly bounded in $|| \cdot||_{\rho'}$.
We note that because
\[
||\mathscr{E}_\mu[ v^0_{\tilde \mu} + (\mu - \tilde \mu) v^1_{\tilde \mu}, \lambda^0_{\tilde \mu} + (\mu - \tilde \mu) \lambda^1_{\tilde \mu} ]||_{\rho'} \le
C |\mu - \tilde \mu |^2
\]
we can apply the KAM theorem (note that the non-degeneracy conditions of
the KAM theorem are satisfied with uniform bounds when the ball is
considered small enough) and obtain that there is a normalized solution $v^*_\mu, \lambda^*_\mu$
of \eqref{external force} for any value of $\mu$ in a ball around $\tilde \mu$ and that it
satisfies
\begin{equation}\label{newsolution}
\begin{split}
&|| v^*_\mu - v^0_{\tilde \mu} + (\mu - \tilde \mu) v^1_{\tilde \mu} ||_{\rho''} \le C|\mu - \tilde\mu|^2\\
&|| \lambda^*_\mu - \lambda^0_{\tilde \mu} + (\mu - \tilde \mu) \lambda^1_{\tilde \mu} ||_{\rho''} \le C|\mu - \tilde\mu|^2.
\end{split}
\end{equation}
Using the uniqueness obtained in the KAM we obtain that $v^*_\mu = v_\mu^0$, $\lambda^*_\mu = \lambda^0_\mu$.
Hence, \eqref{newsolution} means that $v^1_{\tilde \mu} $ is the derivative at $\mu = \tilde \mu$
of the mapping that
to $\mu$ associates $v^0_\mu, \lambda^0_\mu$ if we give $v$ the topology in $\mathscr{A}_{\rho''}$.
We recall that the Cauchy-Goursat theorem shows that any complex function which is differentiable
at every point, is analytic \cite{Ahlfors}. This argument also works for functions taking values in Banach
spaces.
Alternatively, it is not difficult to show that the mapping $\mu \mapsto v^1_\mu$ is continuous
(a quick way is to show that the graph of the map is closed because of the uniqueness
and that, since it uniformly bounded, it is compact by Montel's theorem \cite{Ahlfors}).
Once we have that the function is analytic, we know its Taylor series converges, but the Taylor series
has to be the one given by the formal series expansion.
\end{proof}
Notice as a corollary of the dependence on parameters we can obtain that the solution is analytic in the parameter $\eta$.
Since in the physical applications $\eta\in\mathbb{T}^1$ is important to discuss the periodicity in $\eta$ of the solutions thus obtained, we remark that , if we start with an approximate solution which is periodic in $\eta$, we will obtain a solution which is also periodic in $\eta$.
This can be seen in two ways. One can observe that , applying Theorem \ref{analytic KAM theorem} we can obtain solutions in small enough intervals of $\eta$. They will be analytic in these small intervals and therefore they give a global analytic solution. Furthermore, we observe that if the approximate solution corresponding to $\eta=1$ is close to the solution corresponding to $\eta=0$, they have to agree because of uniqueness. Hence, we obtain that the solution is periodic.
We could also argue that, the proof is based on an iterative step and that, by examining the proof, all the steps preserve the periodicity in $\eta$ of the approximate solutions. Hence, the KAM procedure that we describe for a fixed $\eta$ lifts to a procedure for periodic functions of $\eta$.
The lifting of the problem to a space of functions of $\eta$ also gives a direct proof of smooth dependence on parameters.
\section*{Acknowledgements}
We thank Professor T. Blass for discussions.
R. L. and L. Z. have been supported by DMS-1500943.
The hospitality of
JLU-GT Joint institute for Theoretical Sciences for the three authors
was instrumental in finishing the work.
R.L also acknowledges the hospitality of
the Chinese Acad. of Sciences.
X. Su is supported by both National
Natural Science Foundation of China (Grant No. 11301513) and ``the Fundamental Research Funds for the Central Universities".
\bibliographystyle{alpha}
|
\section{Introduction}
If $(X,0)$ is a germ of an analytic space of pure dimension $ \dim (X,0)$, we denote by $m(X,0) $ its multiplicity and by $edim(X,0)$ its embedding dimension.
Minimal singularities were introduced by J. Koll\'ar in \cite{K} as the germs of analytic spaces $(X,0)$ of pure dimension which are reduced, Cohen-Macaulay, whose tangent cone is reduced and whose multiplicity is minimal in the sense that Abhyankar's inequality
$$m(X,0) \geq edim(X,0) - \dim (X,0) + 1$$
is an equality (see \cite[Section 3.4]{K} or \cite[Section 5] {BL}).
In this paper, we only deal with normal surfaces. In this case, minimality can be defined as follows (\cite[Remark 3.4.10]{K}): a normal surface singularity $(X,0)$ is {\it minimal} if it is rational with a reduced minimal (also called fundamental) cycle.
Minimal surface singularities play a key role in resolution theory of normal complex surfaces since they appear as central objects in the two main resolution algorithms: the resolution obtained as a finite sequence of normalized Nash transformations (\cite{S}), and the one obtained by a sequence of normalized blow-up of points (\cite{Z1939}). The question of the existence of a duality between these two algorithms, asserted by D. T. L\^e in \cite[Section 4.3]{L0} (see also \cite[Section 8]{BL}) remains open, and the fact that minimal singularities seem to be the common denominator between them suggests the need of a better understanding of this class of surface germs.
In this paper, we study minimal surface singularities from the point of view of their Lipschitz geometries, and we show that they are characterized by a remarkable metric property: they are Lipschitz normally embedded. Let us explain what this means.
If $(X,0)$ is a germ of a
complex variety, then any embedding $\phi\colon(X,0)\hookrightarrow
(\C^n,0)$ determines two metrics on $(X,0)$: the outer metric
$$d_{out}(x_1,x_2):=||\phi(x_1)-\phi(x_2)||\quad\text{
(i.e., distance in $\C^n$)}$$ and the inner metric
$$d_{inn}(x_1,x_2):=\inf\{\operatorname{length}(\phi\circ\gamma):
\gamma\text{ is a rectifyable path in }X\text{ from }x_1\text{ to
}x_2\}\, ,$$
using the riemannian metric on $X \setminus \{0\}$ induced by the hermitian metric on $\C^n$. For all $x,y \in X, \ d_{inn}(x,y) \geq d_{out}(x,y)$, and the outer metric determines the inner metric. Up to
bilipschitz equivalence both these metrics are independent of the
choice of complex embedding. We speak of the (inner or outer)
\emph{Lipschitz geometry} of $(X,0)$ when considering these metrics up
to bilipschitz equivalence.
\begin{definition} A germ of a complex normal variety $(X,0)$ is {\it Lipschitz normally embedded} if inner and outer metrics coincide up to bilipschitz equivalence, i.e., there exists a neighbourhood $U$ of $0$ in $X$ and a constant $K\geq 1$ such that for all $x,y \in U$
$$\frac{1}{K} d_{inn}(x,y) \leq d_{out}(x,y).$$
\end{definition}
It is a classical fact that the topology of a germ of a complex variety $(X,0)\subset(\C^n,0)$ is locally homeomorphic to the cone over its link $X^{(\epsilon)}=\Bbb S^{2n-1}_{\epsilon} \cap X$, where $\Bbb S^{2n-1}_{\epsilon}$ denotes the sphere with small radius $\epsilon$ centered at the origin in $\C^n$.
If $(X,0)$ is a curve germ then it is in fact bilipschitz equivalent to the metric cone over its link with respect to the inner metric, while the data of its Lipschitz outer geometry is equivalent to that of the embedded topology of a generic plane projection (see \cite{PT, F, NP1}). Therefore, an irreducible complex curve is Lipschitz normally embedded if and only if it is smooth. Our main result shows that this is not true in higher dimension:
any minimal surface singularity is Lipschitz normally embedded.
In
section \ref{sec:rational + NE implies minimal} we also prove a converse to this among rational singularities, so:
\begin{theorem} \label{theorem:main}
A rational surface singularity is Lipschitz normally embedded if and
only if is minimal.
\end{theorem}
The paper is organized as follows. In Section
\ref{sec:genericity}, we give basic definitions about generic
projections of a normal surface germ and their polar curves and
discriminants. In Section \ref{outergeometry}, we recall the
geometric decomposition with rates of a normal surface germ given in
\cite{NP}, which completely describes the inner Lipschitz geometry
and an important part of the outer geometry. The proof of theorem
\ref{theorem:main} is based on two
preliminary results. The first one is a characterization of
Lipschitz normally embedded surface singularities (Theorem \ref{prop:characterization of normal embedding}). The second one is a
complete description of the geometric decomposition of a minimal
singularity given in Section \ref{sec:inner} by using results of
\cite{BNP} and the explicit description of the polar and
discriminant curves of minimal surface singularities given in
\cite{S} and \cite{B1, B2}. Finally, one direction of Theorem
\ref{theorem:main} (minimal singularities are normally embedded) is proved in Section \ref{sec: main proof} and illustrated through an
example in Section \ref{sec: example liftings}, and the other
direction is proved in Section \ref{sec:rational + NE implies minimal}.
\smallskip\noindent{\bf Acknowledgments.}
Neumann was supported by NSF grant
DMS-1206760. Pichon was supported by ANR-12-JS01-0002-01 SUSI. We are also grateful for the
hospitality and support of the following institutions: Columbia University, Institut de Math\'ematiques de Marseille and Aix Marseille
Universit\'e, and IAS Princeton.
\section{Generic projections} \label{sec:genericity}
We denote by $\mathbf G(k,n)$ the grassmannian of $k$-dimensional
subspaces of $\C^n$. For $\cal D\in \mathbf G(n-2,n)$ let $\ell_{\cal D} \colon \C^n \to
\C^2$ be the linear projection with kernel $\cal D$.
\begin{definition}
Let $(\gamma,0) \subset (\C^n,0)$ be a complex curve germ. Let $V \subset
\mathbf G(n-2,n)$ be the open dense subset such that for each $\cal D \in V$, $\cal D$ does
not contain any limit of bisecant lines to $\gamma$. The projection
$\ell_{\cal D}$ is \emph{generic for $(\gamma,0)$} if $\cal D \in V$. (See \cite{BGG} for the definition of the cone of limits of bisecants.)
\end{definition}
\begin{definition} \label{def:generic} (\cite[(2.2.2)]{LT0} and \cite[V. (1.2.2)]{T3}.) \label{def:generic linear projection}
Let $(X,0)\subset (\C^n,0)$ be a normal surface germ. We assume that the restriction $\ell_{\cal D} |_{X}$ is a finite morphism (this is true for $\cal D$ in an open dense set of $\mathbf G(n-2,n)$).
Let $\Pi_{\cal D}\subset X$ be the polar curve of this projection,
i.e., the closure in $(X,0)$ of the singular locus of the
restriction of $\ell_{\cal D}$
to $X \setminus \{0\}$,
and let
$\Delta_{\cal D}=\ell_{\cal D}(\Pi_{\cal D})$ be the discriminant
curve.
There exists an open dense subset $\Omega \subset
\mathbf G(n-2,n)$ such that
\begin{enumerate}
\item \label{generic1} {for each $\cal D$ in $\Omega$, the projection $\ell_{\cal D}$ is generic for its polar curve $\Pi_{\cal D}$;}
\item \label{generic2} $\{(\Pi_{\cal D},0): \cal D \in
\Omega\}$ forms an equisingular family of curve germs in terms of
strong simultaneous resolution.
\end{enumerate}
We say the projection $\ell_{\cal D} \colon \C^n
\to \C^2$ is \emph{generic} for $(X,0)$ if $\cal D \in \Omega$.
\end{definition}
\begin{remark} For each $\cal D\in \Omega$ the restriction $\ell_{\cal D} |_{X} \colon (X,0) \to (\C^2,0)$ is a finite cover whose degree equals the multiplicity of $(X,0)$ and $\cal D \cap C_{X,0} = \{0\}$ where $C_{X,0} $ denotes the tangent cone to $(X,0)$. In fact, for any $\cal D$, these two properties are equivalent (\cite[Remarque 2.2]{BL}).
\end{remark}
\section{Lipschitz geometry and geometric decomposition of a normal surface singularity}\label{outergeometry}
In this section, we describe the geometric decomposition of $(X,0)$
introduced in \cite{NP}, which completely determines the inner
geometry and is an invariant of the outer geometry. We give first
the description through resolution as presented in \cite[Section
9]{NP} and we give an alternative but equivalent description through
carrousel decomposition at the end of the section. See also \cite{NP}
for more details.
We need to define the contact exponent between two germs of curves.
Let $(C,0)$ and $(D,0)$ be two irreducible plane curve
germs intersecting only at $0$. Let us choose coordinates $x$ and
$y$ in $\C^2$ so that $C$ and $D$ admit Puiseux parametrizations respectively
$$y=\alpha(x^{1/m}) = \sum_{i\geq m} a_i x^{i/m}
\hskip0,3cm \hbox{ and } \hskip0,3cm y=\beta(x^{1/n}) = \sum_{j\geq n} b_j x^{j/n},$$
where $m$ and $n$ are the multiplicities of $C$ and $D$.
Replacing $x$ by $\omega x$ in $\alpha$ (resp.~$\beta$) where $\omega^m=1$ (resp.~$\omega^{n}=1$), we get all the Puiseux parametrizations of $C$ (resp.~$D$).
Denote by $\big\{ \alpha_i(x^{1/m}) \big\}_{i=1}^m$ the set of Puiseux parametrizations of $C$ and $\big\{ \beta_j(x^{1/n}) \big\}_{j=1}^n$ that of $D$.
\begin{definition} The {\it contact exponent} between $C$ and $D$ is the rational number defined by:
$$q_{C,D} = \max_{1\leq i \leq m, 1\leq j \leq n} \big\{ ord_x(\alpha_i(x^{1/m}) - \beta_j(x^{1/n}) ) \big\}$$
More generally, if $C,D$ are two germs of
curves in $(\C^n,0)$ intersecting only at $0$, we define the {\it
contact exponent} $q_{C,D}$ between $C$ and $D$ as the contact exponent between $\ell(C)$ and $\ell (D)$ where $\ell \colon \C^n \to \C^2$ is a generic projection for $C \cup D$.
\end{definition}
Notice that $q_{C,D}$ depends neither on the choice of $\ell $ nor on that of the coordinates $x$ and $y$. Notice also that the contact exponent between two smooth curves is an integer. It is in fact the minimal number of blow-ups of points necessary to separate their strict transforms.
In order to define the geometric decomposition of $(X,0)$, we consider the minimal good
resolution $\pi_0\colon (\widetilde X_0,E) \to (X,0)$ with the following two properties:
\begin{enumerate}
\item it resolves the basepoints of a general linear system of
hyperplane sections of $(X,0)$ (i.e., it factors through the
normalized blow-up of the maximal ideal of $X$);
\item it resolves the basepoints of the family of polar curves of
generic plane projections (i.e., it factors through the
Nash modification of $X$).
\end{enumerate}
This resolution is obtained from the minimal good resolution of $(X,0)$ by
blowing up further until the basepoints of the two kinds are
resolved.
We denote by $\Gamma_0$ the dual resolution
graph of $\pi_0$.
\begin{definition}\label{def:LP-curve} An \emph{\L-curve} is an
exceptional curve in $\pi_0^{-1}(0)$ which intersects the strict
transform of a generic hyperplane section. The vertex of $\Gamma_0$
representing an \L-curve is an \emph{\L-node}.
A \emph{\P-curve} (\P \ for ``polar") will be an exceptional curve
in $\pi_0^{-1}(0)$ which intersects the strict transform of the polar
curve of any generic linear projection. The vertex of $\Gamma_0$
representing this curve is a \emph{\P-node}.
A vertex of $\Gamma_0$ is called a \emph{node} if it is
an \L- or \P-node or has valency $\ge 3$ or represents an exceptional curve of
genus $>0$.
A \emph{string} of a resolution graph is a connected
subgraph whose vertices have valency $2$ and are not nodes, and a
\emph{bamboo} is a non-node vertex of valency
$1$ union a string attached to it.
\end{definition}
Now, consider the resolution $\pi \colon \widetilde X \to X$
obtained from $\widetilde X_0$ by blowing up each intersection point
of pairs of curves of $\pi^{-1}(0)$ which correspond to nodes of $\Gamma_0$. We then obtain a resolution satisfying (1) and (2) and such that there are no adjacent nodes in its resolution graph. Let $\Gamma$ be the resolution graph of $\pi$. Denote by $E_1,\ldots,E_r$ the exceptional curves in $E=\pi^{-1}(0)$ and
by $v_k$ the vertex of $\Gamma$ corresponding to $E_k$.
For each $k=1,\ldots,r$, let $N(E_k)$ be a
small closed tubular neighbourhood of $E_k$ and let $${\cal N}(E_{k})
= \overline{N(E_k) \setminus \bigcup_{k' \neq k} N(E_{k'})}.$$ For any
subgraph $\Gamma'$ of $\Gamma$ define:
$$N(\Gamma'):= \bigcup_{v_k \in\Gamma'}N(E_k)\quad\text{and}\quad
\Nn(\Gamma'):=
\overline{N(\Gamma)\setminus \bigcup_{v_k \notin \Gamma'}N(E_k)}\,.$$
We now describe the geometric decomposition of $(X,0)$. It is a
decomposition of $(X,0)$ as a union of semi-algebraic pieces of three
types: $B(1)$, $B(q)$ with $q>1$ and $A(q,q')$ where the $q<q'$ are
rational numbers $\geq 1$. The pieces $B(1)$ are metrically conical,
i.e., bilipschitz equivalent to a strict metric cone in the
inner metric. For each
piece $B(q)$ with $q>1$, $B(q)\setminus\{0\}$ fibers over a punctured
disk $D^2\setminus\{0\}$ with $2$-manifold fibers having
diameter of order $O(t^q)$ at distance $t$ from the origin. We call $q$ the \emph{rate} of $B(q)$. Each $A(q,q')$ is an intermediate piece between a $B(q)$ and a
$B(q')$ piece and is topologically the cone on a toral annulus $T^2\times I$. For a more precise definition of pieces see \cite[Section 2]{NP}.
If $v_j$ is a vertex of $\Gamma$, we denote by $\Gamma_j$ the subgraph of $\Gamma$ consisting of $v_j$ union any attached bamboos.
\begin{proposition}\cite[Proposition 9.3]{NP} \label{prop:decomposition} The pieces of the geometric decomposition of $(X,0)$ are as follows:
\begin{enumerate}
\item the $B(1)$-pieces are the sets $\pi_1({\cal N}(\Gamma_{j}))$ where
$v_j$ is an \L-node;
\item each $B(q)$-piece for $q>1$ is a set $\pi_1({\cal N}(\Gamma_{j}))$ where
$v_j$ is a node which is not an \L-node;
\item the $A(q,q')$-pieces (which have $1\le q < q'$) are the $\pi_1({
N}(\sigma))$ where $\sigma$ is a maximal string between two
nodes.
\end{enumerate}
In both cases (1) and (2), the rate $q$ is the contact exponent between the $\pi_1$-images of two curvettes of $E_j$ meeting $E_j$ at distinct points.
\end{proposition}
If $E'$ is an irreducible component of a normal crossing divisor $E$ in a complex smooth surface $S$, we call {\it curvette} of $E'$ a small smooth curve on $S$ transversal to $E'$ at a smooth point of $E$.
\begin{remark} The geometric decomposition of $(X,0)$ is a refinement of the thick-thin decomposition of $(X,0)$ introduced in \cite{BNP}. Namely the thick part of $(X,0)$ is the union of $B(1)$ pieces plus adjacent $A(1,q)$-pieces.
\end{remark}
The geometric decomposition can be encoded in the dual resolution graph $\Gamma$ of $\pi$ decorated with a rate $q$ at each node. See the left graph in Example \ref{example}.
In \cite[Definition 8.6]{NP}, we define an equivalence relation
between pieces by saying that two pieces with same rate $q$ are
\emph{equivalent}, if they can be made equal by attaching a ``$q$-collar'' (a
$B(q)$ piece which is topologically the cone on $T^2\times I$) at each
boundary component.
Similarly, two $A(q,q')$-pieces are equivalent if they can
be made equal by
removing a $q$-collars at the outer boundaries and removing $q'$-collars at the inner boundaries.
\begin{proposition} \cite[Proposition 8.7]{NP} \label{prop:unicity}
The geometric decomposition is unique up to equivalence of the
pieces and it is an invariant of the outer Lipschitz geometry.
\end{proposition}
Let us now explain how the geometric decomposition is related (and even built) from the geometry of the discriminant of a generic plane projection.
We first construct a decomposition of the germ $(\C^2,0)$ into $B(q)$- and $A(q,q')$-pieces based on a resolution of $\Delta$.
Let $\rho \colon Y \to \C^2$ be the minimal sequence of blow-ups
starting with the blow-up of $0 \in \C^2$ which resolves the
basepoints of the family of images $\ell(\Pi_{\cal D})$ by $\ell$ of
the polar curves of generic plane projections and let $\Delta$ be
some $\ell(\Pi_{\cal D})$. We set $\rho^{-1}(0) = \bigcup_{k=1}^m
C_k$, where $C_1$ is the exceptional curve of the first blow-up.
Denote by $R$ the dual graph of $\rho$, so $v_{1}$ is its root
vertex. We call a \emph{$\Delta$-curve} an exceptional curve in
$\rho^{-1}(0)$ intersecting the strict transform of $\Delta$, and a
\emph{$\Delta$-node} a vertex of $R$ which represents a
$\Delta$-curve. We call any vertex of $R$ which is either $v_{1}$
or a $\Delta$-node or a vertex with valency $\geq 3$ a \emph{node of
$R$}. A {\it string} is, as in Definition \ref{def:LP-curve}, a connected
subgraph whose vertices have valency $2$ and are not nodes, and a \emph{bamboo} is again a non-node vertex of valency
$1$ union a string attached to it.
If two nodes are adjacent, we blow up the intersection points of the
two corresponding curves in order to create a string between them. Denote $\rho'\colon Y'\to \C^2$ the obtained resolution.
The decomposition of $(\C^2,0)$ is as follows:
\begin{enumerate}
\item \label{B(1)} the single $B(1)$-piece is the set $\rho'({\cal N}(C_{1}))$;
\item \label{Delta} the $B(q)$-pieces for $q>1$ are the sets $\rho'({\cal N}(R_k))$ where $R_k$ is a subgraph of $R$ consisting of a
node $v_k$ which is not the root $v_{1}$ plus any attached bamboo. The rate
$q$ is the contact exponent between the $\rho'$-images of two curvettes of the exceptional curve corresponding to $v_k$.
\item \label{A} the $A(q,q')$-pieces are the sets
$\rho'(N(\sigma))$ where $\sigma$ is a maximal string between two
nodes.
\end{enumerate}
\begin{definition} We call this decomposition of $(\C^2,0)$ into $B$- and $A$-pieces {\it the carrousel decomposition of $(\C^2,0)$ with respect to $\Delta$.}
\end{definition}
The pieces of the carrousel decomposition can also described in terms of truncated Puiseux parametrizations of the components of $\Delta$. See \cite[Section 12]{BNP} or \cite[Sections 3, 7]{NP} for details.
Let us now describe how the carrousel decomposition of $(\C^2,0)$ with respect to $\Delta$ and the geometric decomposition of $(X,0)$ are related.
\begin{definition}
Let $\ell \colon (X,0)\to (\C^2,0)$ be a generic plane projection of
$(X,0)$ as defined in Section \ref{sec:genericity}. Let $\Pi$ be its
polar curve and $\Pi^*$ its strict transform in the resolution
$\pi\colon\widetilde X\to X$. A {\it polar wedge} about $\Pi$ is a neighborhood of $\Pi$
saturated by the $\pi$-images of neighbouring curvettes of $\Pi^*$. A {\it $\Delta$-wedge} is the $\ell$-image of a polar wedge.
(For details see \cite{BNP}).
\end{definition}
According to \cite[Proposition 3.4]{BNP}, $\ell$ is a Lipschitz map for the
inner metric outside a polar wedge $A$ about $\Pi$. Moreover, $A$ and the $\Delta$-wedge $\ell(A)$ are union of $B$-pieces with trivial topology, i.e., the fibers are $2$-disks (we call $D$-pieces such trivial $B$-pieces). More precisely, if $A_0 \subset A$ is a $D(s)$-piece, then $\ell(A_0)$ is also a $D(s)$-piece (with same rate $s$).
As a consequence of this, each piece of the carrousel decomposition
of $(\C^2,0)$ just constructed lifts to a union of $A$ and $B$-pieces of the same type in $(X,0)$. After absorption of the polar-wedges and of the $D(q)$-pieces which do not contain components of the polar curve (see \cite[section 13]{BNP}), one obtains the geometric decomposition
of $(X,0)$ previously described. In the case of minimal singularities, no absorption will be needed so we omit the details of absorption here.
This correspondence between the geometric decomposition of $(X,0)$ and the carrousel decomposition of $(\C^2,0)$ can be read through the correspondence between the resolutions $\rho'$ and $\pi$ given by the Hirzebruch-Jung resolution process. This will be a key argument in the proof of Theorem \ref{theorem:main} in Section \ref{sec: main proof}. A full example is given in Example \ref{example}.
\section{Characterization of Lipschitz normally embedded surface singularities}\label{sec:normal embedding}
\begin{definition} Let $\delta$ be a singular plane curve. We say {\it rate} of $\delta$ for any rational number which is either a characteristic Puiseux exponent of a branch of $\delta$ or the contact exponent between two branches of $\delta$. If $\delta$ is smooth its set of rates is empty.
\end{definition}
The following easy remark will be useful in the induction of Section \ref{sec: main proof}.
\begin{remark}\label{lem:outer rates bl-up} Let $e$ be the blow-up of
the origin of $\C^2$ and let $(\delta,0) \subset (\C^2,0)$ be a
singular plane curve germ. If $\delta$ is irreducible and its set of rates $A$ consists of numbers $\geq 2$, then the set of rates of $(\delta^*,p)$ is $A-1$, where $^*$ denotes strict transform by $e$. If $\delta_1$ and $\delta_2$ are two components of $\Delta$ whose strict transforms by $e$ meet
$e^{-1}(0)$ at a single point $p$, and if $q$ is the contact exponent between $\delta_1 $ and $ \delta_2$, then the contact exponent between $\delta_1^*$ and $\delta_2^*$ equals $q-1$.
\end{remark}
Let $(X,0)$ be the germ of a normal complex surface and let $\ell \colon X \to \C^2$ be a generic plane projection. Let $\Pi$ be the polar curve of $\ell$ and $\Delta=\ell(\Pi)$ be its discriminant curve. Let $\rho \colon Y \to \C^2$ be the resolution of $\Delta$ introduced in Section \ref{outergeometry} and let $R$ be the dual graph of $\rho$.
\begin{definition} \label{def:test}
A {\it test curve} is an irreducible curve $(\gamma,0) \subset (\C^2,0)$ which is the projection by $\rho$ of a curvette $\gamma^*$ of an exceptional curve in $\rho^{-1}(0)$ represented by a node of $R$ such that $\gamma^*$ and the strict transforms of $(\Delta,0)$ by $\rho$ do not intersect.
\end{definition}
\begin{definition} Let $(\gamma,0) \subset (\C^2,0)$ be a test curve. Let $C $ be the component of $\rho^{-1}(0)$ which intersects the strict transform $\gamma^*$. We define the {\it inner rate} $q_{\gamma}$ of $\gamma$ as the contact exponent between the $\rho$-images of two generic curvettes of $C$.
Let $\widetilde{\ell} \colon (\C^n,0) \to (\C^2,0)$ be a plane projection which is generic for the curve $\ell^{-1}(\gamma)$. We define the \emph{outer rates} of $\gamma$ as the rates of the plane curve $\widetilde{\ell}(\ell^{-1}(\gamma))$. Notice that the outer rates of $\gamma$ do not depend on the choice of $\widetilde{\ell}$.
\end{definition}
\begin{theorem}\label{prop:characterization of normal embedding}
$(X,0)$ is Lipschitz normally embedded if and only if for any test curve $(\gamma,0)$, the outer rates of $(\gamma,0)$ satisfy the following conditions:
\begin{itemize}
\item[$(\ast 1)$] Any outer rate of $\gamma$ which is a characteristic exponent equals $q_{\gamma}$.
\item[$(\ast 2)$] For any pair $\delta_1, \delta_2$ of components of $\ell^{-1}(\gamma)$, let $\pi'\colon X' \to X$ be $\pi$ composed with a sequence of successive blow-ups of points which resolves the curve $\delta_1 \cup \delta_2$. Let $E_1$ and $E_2$ be the components of $\pi'^{-1}(0)$ which intersect the strict transforms of $\delta_1$ and $\delta_2$, and let $q_0$ be the maximum among minimum of inner rates of vertices along paths joining the vertices $v_1$ and $v_2$ in the resolution graph of $\pi'$. Then the contact exponent between the generic projections $\widetilde{\ell}(\delta_1)$ and $\widetilde{\ell}(\delta_2)$ equals $q_0$.
\end{itemize}
\end{theorem}
\begin{remark}\label{rk:outer} In fact, an outer rate of $\gamma$ which is a
characteristic exponent is always $\geq q_{\gamma}$. Indeed, let
$C_{\nu}$ be the exceptional curve of $\rho^{-1}(0)$ such that $\gamma^*$ is a curvette of $C_{\nu}$. Then $B=\rho(\cal N(C_{\nu}))$ is a
$B(q_{\gamma})$-piece in the sense of \cite[Section 2]{NP} (see
also \cite[Section 11]{BNP}). Let $\tilde{\ell} \colon (X,0) \to
(\C^2,0)$ be another generic projection. Note that $\tilde{\ell} (\ell^{-1}(B))$ is also a
$B(q_{\gamma})$-piece, so any characteristic exponent of a
complex curve inside it is $\geq q_{\gamma}$.
Similar arguments show that the contact exponent between $\widetilde{\ell}(\delta_1)$ and $\widetilde{\ell}(\delta_2)$, as in Condition ($\ast2$), is always $\geq q_0$.
\end{remark}
\begin{remark}
Notice that Condition $(\ast 2)$ for $\gamma$ does not depend on the
choice of the resolution $\pi'$. In fact, one could take for $\pi'$
the resolution $\pi_{\rho} \colon \tilde{X} \to X$ of $(X,0)$
obtained by taking the pull-back of $\rho$ by $\ell$, then normalizing
and then resolving the remaining quasi-ordinary singularities. It is
easy to prove that $\pi_{\rho}$ is a good resolution for the lifting
by $\ell$ of any test curve of $\rho$.
\end{remark}
\begin{proof}[Proof of Theorem \ref{prop:characterization of normal embedding}] The proof will use some arguments already presented in \cite{NP}, in particular in the proof of \cite[20.1]{NP}.
Assume that there exists a test curve $(\gamma,0) \subset (\C^2,0)$ which does not satisfy Condition $(\ast 1)$. Let $\delta$ be a component of the lifting $\ell^{-1}(\gamma)$ such that $\widetilde{\ell}(\delta)$ admits a characteristic exponent $q > q_{\gamma}$.
Let $(x(w),y(w))=(w^n,\sum_{i\ge n}b_iw^i)$ be a parametrization of $\gamma$.
Fix $w_0 \in \C^*$ and consider the algebraic arc $p\colon [0,1) \to \gamma$ defined by $p(t) = (x(tw_0),y(tw_0))$. Then $p$ lifts to two semi-algebraic arcs $\gamma_1, \gamma_2 \colon [0,1)\to \delta$ such that for all $t$, $\ell(\gamma_1(t))=\ell(\gamma_2(t))=p(t)$ and $d_{out}(\gamma_1(t),\gamma_2(t)) = O(t^{nq})$.
Since $\gamma_1(t)$ and $\gamma_2(t)$ belong to different sheets of
the cover $\ell$, then for any path $\sigma$ between them, the loop
$\ell(\sigma)$ will have to travel through the $B(q_{\gamma})$-piece
containing the point $p(t)$, so
$\operatorname{length}(\ell(\sigma))\geq O(t^{nq_{\gamma}})$. Since
$\ell$ is Lipschitz for the inner metric (outside polar-wedges which
can be avoided by multiplying the length of $\ell(\sigma)$ by a factor
of $\pi$), we also have that the length of $\sigma$ is $\geq O(t^{nq_{\gamma}})$. Therefore, we get $d_{inn}(\gamma_1(t),\gamma_2(t)) \geq O(t^{nq_{\gamma}})$. So
$$\lim_{t \to 0} \frac{ d_{out}(\gamma_1(t),\gamma_2(t))}{d_{inn}(\gamma_1(t),\gamma_2(t))} = 0,$$
which implies that $(X,0)$ is not Lipschitz normally embedded.
Assume now that $(\gamma,0) \subset (\C^2,0)$ does not satisfy Condition $(\ast 2)$ of Theorem \ref{prop:characterization of normal embedding}, i.e., there exists two components $\delta_1$ and $\delta_2$ of the lifting $\ell^{-1}(\gamma)$ such that $q>q_0$, where $q_0$ is defined as in Condition ($\ast2$) and $q$ equals the contact exponent $q_{\widetilde{\ell}(\delta_1), \widetilde{\ell}(\delta_2)}$. Now lift the arc $p$ defined before to two semi-algebraic arcs $\gamma_1 \colon [0,1)\to \delta_1$ and $\gamma_2 \colon [0,1)\to \delta_2$. We then have
$d_{out}(\gamma_1(t),\gamma_2(t)) = O(t^{nq})$. On the other hand, any path $\sigma$ in $X$ from $\gamma_1(t)$ to $\gamma_2(t)$ corresponds to a path $\hat{\sigma}$ in the resolution graph $\Gamma$ joining the vertices $v_1$ and $v_2$. By the same argument as before, the length of a minimal $\sigma$ will be $ O(t^{nq_0})$ where $q_0$ is the minimal rate of the vertices on $\hat{\sigma}$. Therefore $d_{inn}(\gamma_1(t),\gamma_2(t)) = O(t^{nq_0})$ and we conclude as before that $(X,0)$ is not Lipschitz normally embedded.
Therefore, if $(X,0)$ is Lipschitz normally embedded, then any test curve $(\gamma,0) \subset (\C^2,0)$ satisfies conditions $(\ast 1)$ and $(\ast 2)$.
We now want to prove that, conversely, if any test curve satisfies
conditions $(\ast 1)$ and $(\ast 2)$, then $(X,0)$ is Lipschitz normally embedded.
Let $\ell_1, \ell_2 $ and $\ell_3 \colon (X,0) \to (\C^2,0)$ be three distinct generic projections for $(X,0)$ and for $i=1,2,3$, let $A_i$ be a polar wedge for $\ell_i$ such that $A_1, A_2$ and $A_3$ are pairwise disjoint outside the origin. Then for any pair of points $\underbar{p},\underbar{q} \in X \setminus \{0\}$ sufficiently close to $0$, $\underbar{p}$ and $\underbar{q}$ are both outside $A_i$ for at least one of $i \in \{1,2,3\}$. Choose such an $i$ and set $\ell_{(\underbar{p},\underbar{q})} = \ell_i$.
By a straightforward adaptation of the argument in the last page of \cite{NP}, one shows that there exist a neighborhood $V$ of $0$ in $X$ and a constant $L \geq 1$ such that for any pair $\underbar{p}, \underbar{q} \in V$,
$$\frac{1}{L} \bigg( d_{inn}(\underbar{p},\underbar{q}') + d_{out} (\underbar{q}', \underbar{q}) \bigg) \leq d_{out}(\underbar{p}, \underbar{q}) $$
where $\underbar{q}'$ is the extremity of the lifting by $\ell= \ell_{(\underbar{p}, \underbar{q} )}$ of the segment $[\ell(\underbar{p}),\ell(\underbar{q})]$ with origin $\underbar{p}$. In particular, $\ell(\underbar q' )= \ell(\underbar q)$.
Thus the result follows from the following Lemma \ref{lem:aligned}, which is proved later.\end{proof}
\begin{lemma} \label{lem:aligned} Let $\ell \colon (X,0) \to (\C^2,0)$ be a generic projection for $(X,0)$ and let $(A,0)$ be a polar wedge for $\ell$. Assume that any test curve for $\ell$ satisfies conditions $(\ast 1)$ and $(\ast 2)$. Then there exists $\epsilon>0$ and a constant $K \geq 1$ such that for any $\underbar p\in B^4_\epsilon\setminus\{0\}$ and any pair of distinct points $\underbar p_1, \underbar p_2 \in X \setminus A$ such that $\ell(\underbar p_1)=\ell ( \underbar p_2)$, we have $d_{inn}(\underbar p_1,\underbar p_2)\leq K d_{out}(\underbar p_1,\underbar p_2)$.
\end{lemma}
In order to prove Lemma \ref{lem:aligned}, we need two preliminary lemmas \ref{lem:iso} and \ref{lem:sigma}.
\begin{lemma}\label{lem:iso}
Let $\gamma$ be a test curve. Then Condition $(\ast 1)$ implies that the restriction of $\ell$ to any component $\delta$ of $\ell^{-1}(\gamma)$ is an isomorphism.
\end{lemma}
\begin{proof}
Let $C_{\nu}$ be the irreducible component of $\rho^{-1}(0)$ such that $\gamma^*$ is a generic curvette of $C_{\nu}$. Then $B=\rho(\cal N(C_{\nu}))$ is a $B(q_{\gamma})$-piece. Let $\widetilde{B}$ be the component of $\ell^{-1}(B)$ which contains $\delta$ (by a component of a semi-algebraic germ $(Z,0)$, we mean the closure of a connected component of $Z \setminus \{0\}$).
Let $\ell' \colon (X,0) \to (\Bbb C^2,0)$ be another generic projection and let $\Delta'$ be its discriminant curve. Then $\ell'(\widetilde{B})$ is up to higher order a $B(q_{\gamma})$-piece of a carrousel decomposition of $\Delta'$. Choose a generic test curve $\gamma'$ in it and let $\delta'$ be the component of the lifting $\ell'^{-1}(\gamma')$ inside $\widetilde{B}$.
Let $y=\sum_{q_i\geq 1} a_{i} x^{q_i}$ be a Puiseux expansion of $\gamma$. Since the curve $\ell(\delta')$ is inside $B$ and since $B$ is a $q_{\gamma}$-neighbourhood of $\gamma$, then the Puiseux expansion of $\ell(\delta')$ coincides with that of $\gamma$ up to exponent $q_i=q_{\gamma}$, with a distinct non zero coefficient for $x^{q_{\gamma}}$.
Assume $\ell \mid_{\delta} \colon \delta \to \gamma$ is not an isomorphism, then $mult(\delta) = k\ mult(\gamma)$ where $k$ is an integer $\geq 2$. Since the curves $\delta$ and $\delta'$ are isomorphic, then $mult(\delta) = mult(\delta')$. Since $\ell$ is generic for $\delta'$, then $mult(\ell(\delta'))= mult(\delta')$. So we get $mult(\ell(\delta')) = k \ mult(\gamma)$. Since $q_{\gamma}$ is the greatest characteristic Puiseux exponent of $\gamma$ and since the Puiseux expansions of $\gamma$ and $\ell(\delta')$ coincide up to exponent $q_{\gamma}$, we then obtain that $\ell(\delta')$ admits a characteristic exponent $q > q_{\gamma}$. Since $\delta$ and $\delta'$ are isomorphic, $q$ is an outer rate of $\gamma$ and Condition $(\ast 1)$ is not satisfied.
\end{proof}
Let $\pi \colon \widetilde{X} \to X$ be a resolution of $(X,0)$
described after Definition \ref{def:LP-curve}. It factors through
the Nash modification $\nu \colon \widecheck{X} \to X$ and through
the blow-up of the origin and no two nodes of its resolution
graph $\Gamma$ are adjacent. Let $\sigma \colon \widetilde{X} \to
\mathbf G(2,n)$ be the map induced by the projection $p_2 \colon
\widecheck X \subset X \times \mathbf G(2,n) \to \mathbf G(2,n)$.
The map $\sigma$ is well defined on $E=\pi^{-1}(0)$ and its
restriction to $E$ is constant on any connected component of the
complement of \P-curves in $E$ (\cite[Section 2]{GS}, \cite[Part III,
Theorem 1.2]{S}). The connected subgraphs of $\Gamma$ obtained
by removing all \P-nodes and adjacent edges are called
\emph{\P-Tjurina components}.
\begin{lemma} \label{lem:sigma}
Let $\gamma$ be a test curve and $\delta_1$ and $\delta_2$ two components of $\ell^{-1}(\gamma)$ whose strict transforms meet $E_1$ and $E_2$ at smooth points $\underbar p_1$ and $\underbar p_2$ of $E$.
Assume $E_1$ and $E_2$ are not $\cal P$-curves and that there exists a path in the graph of $\pi$ joining $v_1$ to $v_2$ through vertices with rates $\geq q_{\gamma}$. Then Condition $(\ast 2)$ implies $\sigma(\underbar p_1 )\neq \sigma(\underbar p_2)$.
\end{lemma}
\begin{proof} Consider a semi-algebraic arc $p \colon [0,1) \to \gamma$ such that $||p(t)||= O(t)$ and two distinct liftings $\gamma_1 \colon [0,1) \to \delta_1$ and $\gamma_2 \colon [0,1) \to \delta_2$ of $p$ by $\ell$. Then $d_{out}(\gamma_1(t),\gamma_2(t)) = O(t^q)$ where $q$ is an outer rate of $\gamma$.
For $q' > q_{\gamma}$ sufficiently close to $q_{\gamma}$, there exists an $A(q_{\gamma}, q')$-piece $A$ such that $\gamma$ is in the outer boundary of $A$ and $A \cap \Delta = \emptyset$. Let $A_1$ and $A_2$ be the liftings of $A$ containing respectively $\delta_1$ and $\delta_2$. $A_1$ and $A_2$ are inside $N(\Gamma_1)$ and $ N(\Gamma_2)$ where $\Gamma_1$ and $\Gamma_2$ are the $\cal P$-Tjurina components containing $v_{1}$ and $v_{2}$. Assume $\sigma(\underbar p_1) = \sigma(\underbar p_2 )$. Adapting the arguments of the proof of Lemma 11.7 in \cite{NP} inside the pieces $A_1$ and $A_2$, one obtains
$d_{inn} (\gamma_1(t),\gamma_2(t)) = O(t^{q_{\gamma}})$ and $$\lim_{t \to 0} \frac{d_{inn}(\gamma_1(t),\gamma_2(t))}{d_{out}(\gamma_1(t),\gamma_2(t))} = \infty.$$
This contradict Condition $(\ast 2)$ (here $q_0 = q_{\gamma}$).
\end{proof}
\begin{remark}\label{rk:tjurina}
A consequence of Lemma \ref{lem:sigma} is that for any $\cal P$-Tjurina component $\Gamma'$, $\pi(N (\Gamma'))$ contains at most one component of $\ell^{-1}(\gamma)$. \end{remark}
\begin{proof}[Proof of Lemma \ref{lem:aligned}]
Consider a $B(q)$-piece $B$ of the carrousel decomposition of $\C^2$ with respect to the
discriminant curve $\Delta$ of $\ell$. Let $N_1$ and $N_2$ be two
components of $\ell^{-1}(B)$, so $N_1$ and $N_2$ are $B(q)$-pieces,
possibly with $N_1 = N_2$. Let $q_0$ be the maximum among minimum of
inner rates of pieces along paths joining $N_1 \setminus \{0\}$ to
$N_2 \setminus \{0\}$ in $X \setminus \{0\}$, so $q_0 \leq q$ and
$q_0 = q$ if and only if $N_1=N_2$ or $N_1$ or $N_2$ can be joined
by a path through pieces with higher rates.
\vskip0,2cm \noindent
{\bf Claim 1.} There exists $K_1, K_2 >0$ and $\epsilon >0$ such that for all $\underbar p \in (B_{\epsilon}^4 \setminus \{0\}) \cap B$ and for all any pair of distinct points ${\underbar p_1}, \underbar p_2 $ such that $\underbar p_1 \in (N_1 \setminus A) \cap \ell^{-1}(B_{\epsilon}^4)$, $\underbar p_2 \in (N_2 \setminus A) \cap \ell^{-1}(B_{\epsilon}^4)$ and $\ell(\underbar p_1)=\ell(\underbar p_2)=\underbar p$, we have
$$d_{inn}(\underbar p_1,\underbar p_2) \leq K_1 ||\underbar p||^{q_0},$$
and
$$ K_2 ||\underbar p||^{q_0} \leq d_{out}(\underbar p_1,\underbar p_2).$$
\vskip0,2cm \noindent
A straightforward consequence of Claim 1 is that for all $\underbar p \in (B_{\epsilon}^4 \setminus \{0\}) \cap B$ and $\underbar p_1,\underbar p_2$ as before,
$$\frac{d_{inn}(\underbar p_1,\underbar p_2)} {d_{out}(\underbar p_1,\underbar p_2)} \leq \frac{K_1}{K_2}.$$
\begin{proof}[Proof of Claim 1] Choose coordinates $(x,y,z,\ldots)$ in $\C^n$ such that $\ell=(x,y)$ and the piece $B$ is foliated by test-curves $\gamma_{\alpha}$ with Puiseux expansions
$$y = \sum_{i =1}^k a_i x^{p_k} + \alpha x^q$$
where $p_1<p_2<\ldots <p_k<q$ and where $\alpha$ is in a compact set $W \subset \Bbb C$.
For any $q' \geq 1$, if a projection $\ell'$ is generic for a curve $\delta$ then it is generic for any curve in a $q'$-neighbourhood of $\delta$. Since $W$ is compact, one can choose a finite number of projections $\ell_1,\ldots,\ell_s$ and a decomposition $W = W_1 \cup \ldots \cup W_s$ into compact sets such that for any $\alpha \in W_j$, $\ell_j$ is generic for $\ell^{-1}(\gamma_{\alpha})$. We will assume $s=1$ since the proof is similar for $s \geq 2$ taking minimums of bounds. So we choose $\ell_1$ such that for each $\alpha \in W$, the projection $\ell_1$ is generic for $\ell^{-1}(\gamma_{\alpha})$ and we choose the coordinates $(x,y,z,\ldots )$ in $\Bbb C^n$ so that $\ell_1=(x,z)$.
Let $\alpha \in W$ and let $\underbar{p} \in \gamma_{\alpha} \cap (B^4_{\epsilon}\setminus\{0\})$. By Lemma \ref{lem:iso}, two distinct points $\underbar p_1$ and $\underbar p_2$ such that $\ell(\underbar p_1) = \ell(\underbar p_2) = \underbar p$ belong to distinct connected components of the lifting $\ell^{-1}(\gamma_{\alpha})$.
Let $\delta_{\alpha}^{(1)} \subset N_1$ and $\delta_{\alpha}^{(2)} \in N_2$ be two distinct irreducible components of the lifting $\ell^{-1}(\gamma_{\alpha})$ such that $\delta_{\alpha}^{(1)} \cap A = \delta_{\alpha}^{(2)} \cap A = \{0\}$ and $\delta_{\alpha}^{(1)} \neq \delta_{\alpha}^{(2)}$ if $N_1=N_2$.
Since $\ell_1$ is generic for $\delta_{\alpha}^{(1)} \cup \delta_{\alpha}^{(2)}$, Condition $(\ast 2)$ implies that the curves $\ell_1(\delta_{\alpha}^{(1)})$ and $\ell_1(\delta_{\alpha}^{(2)})$ have Puiseux expansions respectively:
$$ z = \sum_{j=1}^m b_j(\alpha) x^{r_j} + b^{(1)}_{q_0}(\alpha) x^{q_0} + h.o, $$
and
$$ z = \sum_{j=1}^m b_j(\alpha) x^{r_j} + b^{(2)}_{q_0}(\alpha) x^{q_0} + h.o, $$
where
\begin{itemize}
\item $r_1<r_2<\ldots<r_m<q_0$,
\item $b_j(\alpha), j=1\ldots,m, \ b^{(1)}_{q_0}(\alpha)$ and $b^{(2)}_{q_0}(\alpha)$ depend continuously on $\alpha$,
\item for all $\alpha \in W, b^{(1)}_{q_0}(\alpha) \neq b^{(2)}_{q_0}(\alpha)$,
\end{itemize}
and where ``$+h.o.$" means plus higher order terms.
Let $\underbar p =(x,y) \in \gamma_{\alpha}$ and $\underbar p_1 \in \delta_{\alpha}^{(1)}$ and $\underbar p_2 \in \delta_{\alpha}^{(2)}$ such that $\ell(\underbar p_1) = \ell(\underbar p_2) = \underbar p$. Since $|| \underbar p || = O(|x|)$, there exists $L \geq 1$ such that:
$$\frac{1}{L} |b^{(1)}_{q_0}(\alpha) - b^{(2)}_{q_0}(\alpha)| \ ||\underbar p||^{q_0} \leq || \ell_1(\underbar p_1) -\ell_1(\underbar p_2)|| \leq d_{out}(\underbar p_1,\underbar p_2).$$
Since $W$ is compact, there exists $K'_2>0$ such that for all $\alpha \in W$, $K'_2 \leq \frac{1}{L} |b^{(1)}_{q_0}(\alpha) - b^{(2)}_{q_0}(\alpha)|$. Taking the minimum $K_2$ of $K'_2$ among all pairs $\delta_{\alpha}^{(1)}, \delta_{\alpha}^{(2)}$ as before, there exists $\epsilon >0$ such that for all $\underbar p \in (B_{\epsilon}^4 \setminus \{0\}) \cap B$ and for $\underbar p_1 $ and $ \underbar p_2$ as before,
$$K_2 ||\underbar p||^{q_0} \leq d_{out}(\underbar p_1,\underbar p_2)$$
Let us now bound $ d_{inn}(\underbar p_1,\underbar p_2)$
We consider the carrousel decomposition of $\C^2$ with respect
the the discriminant curve $\Delta$ of $\ell$ and we decompose $X$
into pieces consisting of components of inverse images by $\ell$ of
pieces of the decomposition of $\C^2$ (see also end of Section \ref{outergeometry}).
Since $\underbar p_1$ and $\underbar p_2$ are on different sheets of
the cover $\ell$, a path from $\underbar p_1$ to $\underbar p_2$
with minimal length will have to travel through a $B(q_0)$-piece $N$
of $(X,0)$ and it only travels through pieces with rates $\geq
q_0$. Now $\ell(N)$ is a $B(q_0)$-piece in $(\Bbb
C^2,0)$ and there exists $a>0$ such for $\epsilon >0$ sufficiently
small, if $(x,y_1)$ and $(x,y_2)$ are in $\ell(N) \cap
B^4_{\epsilon}$ then $|y_1-y_2|\leq a |x|^{q_0}$.
Let $m=m(X,0)$ (so $m$ is at most the order of the cover
$\ell|_N\colon N\to \ell(N)$), and let $\kappa$ be the local
bilipschitz constant of $\ell$ outside the polar wedge $A$. Then
$d_{inn}(\underbar p_1, \underbar p_2)$ is less than or equal to $m
\kappa$ times the diameter of $\ell(N)$
providing one avoids the $\Delta$-wedge $\ell(A)$. So we obtain
$$d_{inn}(\underbar p_1, \underbar p_2) \leq m(\pi a) \kappa ||p||^{q_0},$$
where we use the factor $ \pi a$ instead of $a$ to allow replacing a segment by a path avoiding $\ell(A)$. Setting $K_1 = m(\pi a) \kappa$ completes the proof of Claim 1.
\end{proof}
Let us now consider an $A(q,q')$-piece $A_0$ of the carrousel decomposition of $\C^2$ with respect to $\Delta$ with $q<q'$, so $A_0 \cap \Delta = \emptyset$. According to Lemma \ref{lem:iso} and Remark \ref{rk:tjurina}, if $A'$ is a component of $\ell^{-1}(A_0)$, the restriction $\ell|_{A'} \colon (A',0) \to (A_0,0)$ is a homeomorphism and two distinct components of $\ell^{-1}(A_0)$ have distinct corresponding $\cal P$-Tjurina components.
Let $A_1$ and $A_2$ be two distinct components of $\ell^{-1}(A_0)$,
so $A_1$ and $A_2$ are $A(q,q')$-pieces. Let $q_0$ be the maximum
among minimum of inner rates of pieces along paths joining $A_1
\setminus \{0\}$ to $A_2 \setminus \{0\}$ in $X \setminus \{0\}$. (So
$q_0 \leq q$ and $q_0 = q$ if and only if $A_1$ or $A_2$ can be
joined by a path through pieces with rates $\geq q_0$.)
\medskip \noindent
{\bf Claim 2.} There exists $K >0$ and $\epsilon >0$ such that for all $\underbar p \in (B_{\epsilon}^4 \setminus \{0\}) \cap A_0$ and for all $\underbar p_1 \in A_1 \cap \ell^{-1}(B_{\epsilon}^4)$ and $\underbar p_2 \in A_2 \cap \ell^{-1}(B_{\epsilon}^4)$ such that $\ell(\underbar p_1)=\ell(\underbar p_2)=\underbar p$, we have
$$\frac{d_{inn}(\underbar p_1,\underbar p_2)} {d_{out}(\underbar p_1,\underbar p_2)} \leq K.$$
\begin{proof}[Proof of Claim 2] Let $\Gamma_1$ and $\Gamma_2$ be the
$\cal P$-Tjurina components such that $A_i \subset
\pi(N(\Gamma_i))$, $i=1,2$. Let $P_1$ (resp.\ $P_2$) be the values
of $\sigma$ on $\bigcup_{v_j \in \Gamma_1} E_j $ (resp.\
$\bigcup_{v_j \in \Gamma_2} E_j $).
In suitable coordinates the semi-algebraic set $A_0$ is defined by inequalities
$$ \alpha |x|^{q'} \leq |y-\sum_{i=1}^k a_ix^{p_i}| \leq \alpha |x|^{q},$$
where $\alpha >0$ and $1 \leq p_1<p_2<\ldots<p_k=q$.
Let $n= \operatorname{lcm} (\denom(q'), \denom(p_i)$, $i=1,\ldots,k)$. Then $A_0$ is the union of the images of the maps $\phi_{\xi_1,\xi_2} \colon [0,1] \times [0,1) \to (\C^2,0)$ parametrized by $(\xi_1,\xi_2) \in \Bbb S^1 \times \Bbb S^1$ and defined by
$$ \forall (s,t) \in [0,1] \times [0,1), \ \phi_{\xi_1,\xi_2}(s,t)= (x(t),y(s,t))$$
with
$$ x(t) = \xi_1 t^n \ \ \hbox{ and } \ \ y(s,t) = \sum_{i=1}^k a_i x(t)^{p_i} + \alpha \xi_2 \bigg( s x(t)^{q'} +(1-s)x(t)^{q}\bigg) $$
Notice that $|| \phi_{\xi_1,\xi_2}(s,t) || = O(t^n)$.
\noindent{\bf Case 1.} Assume first that $A_1$ or $A_2$ can be joined by a path through pieces with rates $\geq q'$.
Fix $\xi_1,\xi_2 \in \Bbb S^1 \times \Bbb S^1$ and consider two liftings $\phi_1 \colon [0,1] \times [0,1) \to A_1$ and $\phi_2 \colon [0,1] \times [0,1) \to A_2$ of $\phi$ by $\ell$, i.e., $\phi_1 = (\ell |_{A_1})^{-1} \circ \phi_{\xi_1,\xi_2} $ and $\phi_2 = (\ell |_{A_2})^{-1} \circ \phi_{\xi_1,\xi_2} $. Let $\big(c_{s,t}\big) $ be a continuous family of paths in $X \setminus \{0\} $ parametrized by $(s,t) \in [0,1]\times(0,1)$ such that,
\begin{itemize}
\item $c_{s,t}$ has origin $\phi_1(s,t)$ and extremity $\phi_2(s,t)$
\item $c_{s,t}$ consists of the path $\phi_1([s,1]\times\{t\})$ followed by a path $c'_t$ from $\phi_1(1,t)$ to $\phi_2(1,t)$ (independent of $s$) through pieces with rates $>q'$ followed by the reversed $\phi_2([s,1]\times\{t\})$.
\end{itemize}
Fix $s \in [0,1)$. As $t$ tends to $0$, the ratio $\frac{\operatorname{length}(c'_t)}{\operatorname{length}(\phi_i([s,1]\times\{t\}))}$ tends to zero for $i=1,2$ and the path $c_{s,t}$ tends to the union of two segments whose angles with the kernel of $\ell$ depends only on $P_1$ and $P_2$ (see Figure \ref{fig1:claim2}). Since $P_1 \neq P_2$ (Lemma \ref{lem:sigma}) and since the projection $\ell$ is generic, we obtain that for all $s \in [0,1)$
$$\lim_{t \to 0} \frac{\operatorname{length}(c_{s,t})} {d_{out}(\phi_1(s,t), \phi_2(s,t))} =a, $$
where $a>0$ just depends on $P_1$ and $P_2$ (so it is independent of $\xi_1,\xi_2$ and $s$). Since $d_{inn}( \phi_1(s,t), \phi_2(s,t)) \leq \operatorname{length}(c_{s,t})$, this proves Claim 2 in that case.
\begin{figure}[ht]
\centering
\begin{tikzpicture}
\draw[line width=0.3pt] (0,2)--(4,0.2);
\draw[line width=0.3pt] (0,-1.5)--(4,-0.2);
\draw[line width=1pt] (1.5,1.33)--(4,0.2);
\draw[line width=1pt] (1.5,-1)--(4,-0.2);
\draw[line width=1pt] (4,0.2).. controls (4.4,0.1) and (4.4,-0.1)..(4,-0.2);
\draw[fill] (0,2)circle(2pt);
\draw[fill] (0,-1.5)circle(2pt);
\node(a)at(0,2.3){$\phi_2(0,t)$};
\node(a)at(0,-1.8){$\phi_1(0,t)$};
\draw[line width=0.6pt,dotted] (0,2)--(0,-1.5);
\draw[fill] (1.5,1.33)circle(2pt);
\draw[fill] (1.5,-1)circle(2pt);
\node(a)at(1.7,1.7){$\phi_2(s,t)$};
\node(a)at(1.7,-1.3){$\phi_1(s,t)$};
\draw[line width=0.6pt,dotted] (1.5,1.33)--(1.5,-1);
\draw[fill] (4,0.2)circle(2pt);
\draw[fill] (4,-0.2)circle(2pt);
\node(a)at(4.2,0.5){$\phi_2(1,t)$};
\node(a)at(4.2,-0.5){$\phi_1(1,t)$};
\node(a)at(4.7,0){$c'_{t}$};
\node(a)at(3,1){$c_{s,t}$};
\end{tikzpicture}
\caption{}
\label{fig1:claim2}
\end{figure}
\noindent{\bf Case 2.} Assume now that any path from $A_1$ or $A_2$ goes through pieces with rates $\leq q$. Let $B$ (resp.\ $B'$) be the $B(q)$-piece (resp.\ $B(q')$-piece) attached to the outer (resp.\ inner) boundary of $A_0$ and let $\epsilon$, $K_1$ and $K_2$ (resp.\ $K'_1$ and $K'_2$) be constants associated to $B$ (resp.\ $B'$) as in Claim 1.
Fix $\xi_1,\xi_2 \in \Bbb S^1 \times \Bbb S^1 $ and consider again $\phi_1$ and $\phi_2$ as defined in Case 1.
We have for all $(s,t) \in [0,1]\times(0,1)$,
\begin{align*}
d_{inn}(\phi_1(s,t),\phi_2(s,t)) \leq d_{inn}(&\phi_1(0,t),\phi_2(0,t)) +\\
&\operatorname{length}(\phi_1([0,s]\times\{t\})) + \operatorname{length}(\phi_2([0,s]\times\{t\}))
\end{align*}
Let $\epsilon>0$ be sufficiently small and let $\kappa >0$ be a bound for the local bilipschitz constant of the restriction $\ell \colon \overline{X \cap \ell^{-1}(B_{\epsilon}) \setminus A} \to B_{\epsilon} $, where $A$ is a polar wedge around the polar curve of $\ell$. We then have for $t$ sufficiently small, i.e., such that $\phi_{\xi_1,\xi_2}(s,t) \in B_{\epsilon}$,
$$d_{inn}(\phi_1(s,t),\phi_2(s,t)) \leq d_{inn}(\phi_1(0,t),\phi_2(0,t)) + 2 \kappa \alpha s (t^{nq} - t^{n q'}) \leq K''_1 t^{nq_0}, $$
where $K''_1 = K_1 + 2 \kappa \alpha$. Notice that $K''_1$ is independent of $\xi_1, \xi_2$ and $s$. Since $|| \phi_{\xi_1,\xi_2}(s,t) || = O(t^n)$, we then have proved that there exist a constant $C>0$ and $\epsilon>0$ such that for any $\underbar{p} \in B_{\epsilon} \cap A_0$,
$$d_{inn}(\underbar{p}_1,\underbar{p}_2) \leq C ||\underbar{p}||^{q_0}$$
where $\underbar{p}_1$ and $\underbar{p}_2$ are the liftings of $\underbar{p}$ to $A_1$ and $A_2$ respectively.
Let us now deal with the outer distance. As $t$ tends to $0$, the two arcs $\phi_1([0,1]\times\{t\})$ and $\phi_2([0,1]\times\{t\})$ tend to the union of two coplanar segments which are opposite sides of a trapezoid (Figure \ref{fig2:claim2}). Then for $t>0$ sufficiently small and for any $s \in [0,1]$,
$$d_{out} (\phi_1(s,t),\phi_2(s,t)) \geq (1-\eta) \min \bigg(d_{out}
(\phi_1(0,t),\phi_2(0,t)),d_{out} (\phi_1(1,t),\phi_2(1,t))\bigg)$$
with $\eta$ small (the constant $1-\eta$ is for the case when $P_1=P_2$ in the previous notation, i.e., the trapezoid is a rectangle).
Applying Claim 1 to the pieces $B$ and $B'$ adjacent to $A_0$, there exist $K_2>0$ and $K'_2>0$ such that for all $t>0$ sufficiently small,
$$K_2 ||\phi_{\xi_1,\xi_2}(0,t)||^{q_0} \leq d_{out} (\phi_1(0,t),\phi_2(0,t)),$$
and
$$K'_2 ||\phi_{\xi_1,\xi_2}(1,t)||^{q_0} \leq d_{out} (\phi_1(1,t),\phi_2(1,t)).$$
Since for all $s$, $|| \phi_{\xi_1,\xi_2}(s,t) || = O(t^n)$, we then obtain that there exists $C'>0$ and $\epsilon >0$ such that for any $\underbar{p} \in B_{\epsilon} \cap A_0$,
$$ C' ||\underbar{p}||^{q_0} \leq d_{out}(\underbar{p}_1,\underbar{p}_2). $$
where $\underbar{p}_1$ and $\underbar{p}_2$ are the liftings of $\underbar{p}$ to $A_1$ and $A_2$ respectively. This proves Claim 2 in Case 2.
\begin{figure}[ht]
\centering
\begin{tikzpicture}
\draw[line width=0.6pt] (0,2)--(4,1);
\draw[line width=0.6pt] (0,-1.5)--(4,-0.2);
\draw[fill] (0,2)circle(2pt);
\draw[fill] (0,-1.5)circle(2pt);
\node(a)at(0,2.3){$\phi_2(0,t)$};
\node(a)at(0,-1.8){$\phi_1(0,t)$};
\draw[line width=0.6pt,dotted] (0,2)--(0,-1.5);
\draw[fill] (1.5,1.62)circle(2pt);
\draw[fill] (1.5,-1)circle(2pt);
\node(a)at(1.7,2){$\phi_2(s,t)$};
\node(a)at(1.7,-1.3){$\phi_1(s,t)$};
\draw[line width=0.6pt,dotted] (1.5,1.61)--(1.5,-1);
\draw[fill] (4,1)circle(2pt);
\draw[fill] (4,-0.2)circle(2pt);
\draw[line width=0.6pt,dotted] (4,1)--(4,-0.2);
\node(a)at(4.2,1.4){$\phi_2(1,t)$};
\node(a)at(4.2,-0.5){$\phi_1(1,t)$};
\end{tikzpicture}
\caption{}
\label{fig2:claim2}
\end{figure}
\end{proof}
Claim 1 and Claim 2 imply Lemma \ref{lem:aligned}.
\end{proof}
\section{ The geometric decomposition of a minimal singularity} \label{sec:inner}
The aim of this section is to describe the geometric decomposition with rates of a minimal surface singularity germ $(X,0)$ and its correspondence with the carrousel decomposition of $(\C^2,0)$ with respect to the discriminant curve $\Delta$ of a generic projection of $(X,0)$ and the resolution $\rho$ of $\Delta$ (see Section \ref{outergeometry}).
Let us first recall the definition of the minimal (also called fundamental) cycle $Z_{min}$ of a normal surface singularity $(X,0)$. We refer to \cite{N} for details. Let ${\pi}\colon{(\widetilde X,E)}\to{(X,0)}$ be a resolution and let $E_1,\ldots,E_r$ be the irreducible components of the exceptional divisor $E=\pi^{-1}(0)$. The {\it minimal cycle} $Z_{min}$ is the minimal element of the set of divisors $Z=\sum_{i=1}^r m_i E_i$ whose coefficients $m_i$ are strictly positive integers and such that $\forall j=1,\ldots,r,$ $Z\cdot E_j \leq 0$. A {\it reduced} minimal cycle means that $Z_{min}=\sum_i^r E_i$, i.e., $m_i = 1$ for all $i=1,\ldots,r$.
If ${f}\colon{(X,0)}\to{(\C,0)}$ is an analytic function, then its total transform $(f)=(f \circ \pi)^{-1}(0)$ decomposes into $(f) = Z(f) + f^*$ where $f^*$ is the strict transform and $Z(f)$ a positive divisor with support on $E$. For each $j=1,\ldots,r$, one has $(f)\cdot E_j = 0$. Hence $Z(f)\cdot E_j \leq 0$ for all $j=1,\ldots,r$. If $h \colon (X,0) \to (\C,0)$ is a generic linear form, then $Z(h)$ is the minimal element among divisors $Z(f)$, and $Z_{min}\leq Z(h)$. For any rational singularity, a fortiori for minimal, the minimal resolution resolves the basepoints of the family of generic linear forms and $Z(h)=Z_{min}$ (see \cite{A,N}). So, for a rational singularity, the $\cal L$-nodes in a resolution graph, and then, the thick-thin decomposition, are topologically determined.
We now restrict to minimal singularities. In order to describe the geometric decomposition of $(X,0)$, we will use the description by Spivakovsky (\cite{S}) of the minimal resolution $\pi$ of the pencil of polar curves of generic plane projections $ (X,0)\to (\C^2,0)$ and the description by Bondil (\cite{B1,B2}) of the resolution $\rho$ of the family of projected polars $\ell(\Pi_{\cal D})$, where $\ell \colon (X,0)\to (\C^2,0)$ is a generic plane projection.
In \cite{S}, Spivakovsky gives the following combinatorial characterization of the dual resolution graphs of minimal singularities which immediately furnishes the $\cal L$-nodes. Let $ \pi' \colon X' \to X$ be the the minimal good resolution of $(X,0)$ and let $\Gamma'$ be its dual graph. Denote by $V(\Gamma')$ the set of vertices of $\Gamma'$. If $v \in V(\Gamma')$, we denote by $E_v$ the corresponding irreducible component of the exceptional divisor $(\pi')^{-1}(0)$, we set $w(v)=E_v^2$ and we denote by $\nu(v)$ the valence of $v$, i.e., the number of edges adjacent to $v$.
\begin{proposition} \cite{S} A surface singularity is minimal if and only if $\Gamma'$ is a tree of rational curves and for all vertices $v\in V(\Gamma')$,
$-w(v)\geq \nu(v)$.
\end{proposition}
\begin{remark} \label{rk:Lnodes} Since the minimal cycle is reduced, a vertex of $\Gamma'$ is an $\cal L$-node if and only if
$-w(v)> \nu(v)$.
\end{remark}
Spivakovsky introduced the function $s \colon{V(\Gamma')}\to{\N}$ defined as follows: $s(v)$ is
the number of vertices on the shortest path from $v$ to an $\cal L$-node. So $s(v)=1$ if and only if $v$ is an $\cal L$-node.
Since minimal singularities
are rational they can be resolved by only blowing up points, as
Tjurina showed in \cite{Tyu}, and $s(v)$ is the number of blow-ups it
takes before $E_v$ appears in the successive exceptional
divisors.
We now state Theorem 5.4 in Chapter III of \cite{S} in a formulation inspired by Bondil in \cite{B1}.
\begin{theorem}\cite[Chap.~III, Theorem 5.4]{S} \label{thm:Pi} Let $(X,0)$ be a minimal surface singularity. Let $\ell \colon (X,0) \to (\C^2,0)$ be a generic linear projection and let $\Pi$ be its polar curve. Let ${\pi}' \colon (X',0)\to (X,0) $ be the minimal resolution of $(X,0)$. Consider the cycle $S:=\sum s(v)E_v$, where the $E_v$ are the irreducible components of $(\pi')^{-1}(0)$.
Then the strict transform $\Pi^*$ of $\Pi$ by ${\pi}'$ is smooth. It consists of exactly $-(S+E_v)\cdot E_v-2$ curvettes of each $E_v$ and one component through each intersection point $E_v \cap E_w$ for which $s(v)=s(w)$. Moreover, the latter intersection points are the only basepoints of the family of generic polars $\Pi_{\cal D}$, and they are simple, i.e., they are resolved by one blow-up.
\end{theorem}
Following the terminology of \cite{S}, an edge of $\Gamma'$ between two vertices $v$ and $w$ is \emph{central} if $s(v)=s(w)$, and a vertex $v$ is \emph{central} if there are at least two neighboring vertices $w,w'$ such that $s(v)-1=s(w)=s(w')$. Using this, the above theorem says that for each central edge there is one component of $\Pi^*$ through the intersection point of the corresponding curves and that for each central vertex $v$, there is at least one component of $\Pi^*$ which is a curvette of $E_v$. Any other components of $\Pi^*$ go through $\cal L$-nodes.
In \cite{B1}, Bondil shows that the minimal resolution of $(X,0)$ obtained by only blowing up points is also the minimal resolution of $\Pi$ just described. Then, blowing up the points corresponding to central edges, we get the resolution $\pi_0\colon (\widetilde X_0,E) \to (X,0)$ introduced in Section \ref{outergeometry}, i.e., the minimal resolution which factors through the blow-up of the origin and through Nash blow-up.
We then know the resolution graph $\Gamma_0$ of $\pi_0$ together with its nodes. So we topologically know the geometric decomposition of $(X,0)$ from the resolution graph. We now need to determine the rate $q$ of each node. In order to do this, we will use a more precise description of the polar curve and of the discriminant curve presented by Bondil in \cite{B1}.
An $A_n$-\emph{curve} is a germ of an analytic curve isomorphic to the plane curve $y^2+x^{n+1}=0$. If $n$ is odd, then $A_n$ consists of a pair of smooth curves with contact exponent $\frac{n+1}{2}$ while if $n$ is even, $A_n$ is an irreducible curve
\begin{theorem}[\cite{B1,B2}] \label{genericpolar}
Let $(X,0)$ be a minimal singularity, and let $\Pi$ be the polar of a generic linear projection. Then
\begin{enumerate}
\item $\Pi$ decomposes as a union of $A_{n_i}$-curves $\Pi=\bigcup_i C_i$ and each $C_i$ meet a single irreducible component $E_{v_i}$ of the exceptional divisor of $\pi_0$
\item If $E_{v_i}^2=-1$ (i.e., $E_{v_i}$ comes from blowing up a central edge), then $C_i$ is an (irreducible) $A_{ 2s(v_i)-2}$-curve. Otherwise $C_i$ consists of two smooth curves forming an $A_{2s(v_i)-1}$-curve
\item The contact exponent between $C_i$ and $C_j$ equals the minimal value of $s(v)$ on the shortest path in $\Gamma_0$ between the vertices ${v_i}$ and ${v_j}$.
\end{enumerate}
\end{theorem}
Using the fact that each branch of $\Pi$ is isomorphic to a plane curve and that the restriction $\ell |_{\Pi} \colon \Pi \to \Delta$ is generic, Bondil deduces from Theorem \ref{genericpolar} the following description of the discriminant curve:
\begin{proposition}[\cite{B1,B2}]\label{prop:discriminant}
The discriminant curve $\Delta$ of a generic projection $\ell$ of $(X,0)$ is a union of $A_{n}$-curves in one-to-one correspondence with the curves $C_i$ of Proposition \ref{genericpolar}, and their pairwise contact exponents equal that of the corresponding $C_i$'s. Moreover, the minimal resolution of $\Delta$ is the resolution $\rho \colon Y \to \C^2$ which resolves the base points of the family of generic polar curves $(\ell(\Pi_{\cal D}))_{\cal D \in \Omega}$.
\end{proposition}
We deduce from this the rates of the pieces $B(q)$ of the geometric decomposition of $(X,0)$:
\begin{corollary}\label{innergeomertythm}
Let $(X,0)$ be a minimal surface singularity and let $\Gamma_0$ be the dual resolution graph of the resolution $\pi_0$ described above. The rate $q_v$ of a node $v$ of $\Gamma$ is given by:
\begin{align*}
q_v=\begin{cases}
s(v) &\text{ if }E_v^2<-1\\
s(v)-1/2 &\text{ if }E_v^2=-1,
\end{cases}
\end{align*}
\end{corollary}
\begin{proof}
The rate $q_v$ is the contact exponent between the $\pi$-images of two generic curvettes of $E_v$.
When $v$ is a node such that $E_v^2<-1$, the images of two generic curvettes of $E_v$ form a $A_{2s(v)-1}$-curve so their contact exponent equals $s(v)$. When $E_v^2=-1$ the images of two generic curvettes of $E_v$ are $A_{2s(v)-2}$-curves whose contact exponent equals $s(v)-\frac{1}{2}$.
\end{proof}
\begin{example}\label{example}
Let $(X,0)$ be a minimal singularity with the following resolution graph:
\begin{center}
\begin{tikzpicture}
\draw[thin ](-1,0)--(4,0);
\draw[thin ](1,0)--(1,-3);
\draw[fill=black ] (-1,0)circle(2pt);
\draw[fill=white ] (0,0)circle(2pt);
\draw[fill=white ] (1,0)circle(2pt);
\draw[fill=black] (2,0)circle(2pt);
\draw[fill=white ] (3,0)circle(2pt);
\draw[fill =black ] (4,0)circle(2pt);
\draw[fill=white ] (1,-1)circle(2pt);
\draw[fill=white ] (1,-2)circle(2pt);
\draw[fill =black ] (1,-3)circle(2pt);
\node(a)at(-1,0.3){ $-4$};
\node(a)at(0,0.3){ $-2$};
\node(a)at(1,0.3){ $-3$};
\node(a)at(2,0.3){ $-3$};
\node(a)at(3,0.3){ $-2$};
\node(a)at(4,0.3){ $-2$};
\node(a)at(0.6,-1){ $-2$};
\node(a)at(0.6,-2){ $-2$};
\node(a)at(0.6,-3){ $-2$};
\end{tikzpicture}
\end{center}
The negative weights are the self-intersections of the exceptional curves. The $\cal L$-nodes are the vertices $v$ such that $-w(v)>\nu(v)$ (Remark \ref{rk:Lnodes}). They are in black in the graph.
The following graph on the left shows two different things. First, the arrows represent the components of the polar curve of a generic plane projection. The gray node represents a curve obtained by blowing up the minimal resolution at the intersection point of two exceptional curves corresponding to a central edge. There are four pairs $C_1,\ldots,C_4$ of smooth components, and one component $C_5$ with multiplicity $2$. Secondly, it shows the geometric decomposition of $(X,0)$. The rational numbers in italics are the rates of the nodes.
The graph on the right is the resolution graph of the discriminant curve $\Delta=\ell(\Pi)$. The arrows represent the components of $\Delta = \bigcup_{i=1}^5$ where $\Delta_i=\ell(\Pi_i)$, $i=1,\ldots,5$. The root-vertex is the black vertex and each vertex is weighted by the corresponding rate.
\begin{center}
\begin{tikzpicture}
\draw[thin ](-2,0)--(4,0);
\draw[thin ](1,0)--(1,-3);
\draw[thin,>-stealth,->](-2,0)--+(-0.8,0.8);
\draw[thin,>-stealth,->](-2,0)--+(-0.8,-0.8);
\draw[thin,>-stealth,->](-2,0)--+(-1,-0.5);
\draw[thin,>-stealth,->](-2,0)--+(-1,0.5);
\node(a)at(-3.2,0.8){ $C_1$};
\node(a)at(-3.2,-0.8){ $C_2$};
\draw[thin,>-stealth,->](0,0)--+(-0.2,-1.1);
\node(a)at(-0.5,-1.1){ $C_5$};
\draw[thin,>-stealth,->](1,-1)--+(1.1,-0.2);
\draw[thin,>-stealth,->](1,-1)--+(1.1,0.2);
\node(a)at(2.1,-1.5){ $C_4$};
\draw[thin,>-stealth,->](3,0)--+(-0.2,-1.1);
\draw[thin,>-stealth,->](3,0)--+(0.2,-1.1);
\node(a)at(3.6,-1.1){ $C_3$};
\draw[fill=black ] (-2,0)circle(2pt);
\draw[fill=white ] (-1,0)circle(2pt);
\draw[fill=gray ] (0,0)circle(2pt);
\draw[fill=white ] (1,0)circle(2pt);
\draw[fill=black] (2,0)circle(2pt);
\draw[fill=white ] (3,0)circle(2pt);
\draw[fill =black ] (4,0)circle(2pt);
\draw[fill=white ] (1,-1)circle(2pt);
\draw[fill=white ] (1,-2)circle(2pt);
\draw[fill =black ] (1,-3)circle(2pt);
\node(a)at(-2,0.3){ $\it 1$};
\node(a)at(0,0.3){ $\it 5/2$};
\node(a)at(1,0.3){ $\it 2$};
\node(a)at(2,0.3){ $\it 1$};
\node(a)at(3,0.3){ $\it 2$};
\node(a)at(4,0.3){ $\it 1$};
\node(a)at(0.7,-3){ $\it 1$};
\node(a)at(0.7,-1){ $\it 3$};
\draw[thin ](7,0)--(7,-3);
\draw[thin ](7,0)--(8.5,0);
\draw[thin ](7,-1.5)--(6,-0.5);
\draw[thin ](7,-3)--(8,-2);
\draw[thin,>-stealth,->](7,-3)--+(-0.8,0.8);
\draw[thin,>-stealth,->](7,-3)--+(-0.8,-0.8);
\draw[thin,>-stealth,->](7,-3)--+(-1,-0.5);
\draw[thin,>-stealth,->](7,-3)--+(-1,0.5);
\node(a)at(5.8,-2.2){ $\Delta_1$};
\node(a)at(5.8,-3.6){ $\Delta_2$};
\draw[thin,>-stealth,->](8,-2)--+(-0.2,1.1);
\draw[thin,>-stealth,->](8,-2)--+(0.2,1.1);
\node(a)at(8.6,-1){ $\Delta_3$};
\draw[thin,>-stealth,->](6,-0.5)--+(-0.2,1.1);
\draw[thin,>-stealth,->](6,-0.5)--+(0.2,1.1);
\node(a)at(6,1){ $\Delta_4$};
\draw[thin,>-stealth,->](7,0)--+(-0.2,1.1);
\node(a)at(7.2,1){ $\Delta_5$};
\draw[fill=black ] (7,-3)circle(2pt);
\draw[fill=gray ] (7,0)circle(2pt);
\draw[fill=white ] (7,-1.5)circle(2pt);
\draw[fill=white ] (6,-0.5)circle(2pt);
\draw[fill=white ] (8,-2)circle(2pt);
\draw[fill=gray ] (8.5,0)circle(2pt);
\node(a)at(5.7,-0.5){ $\it 3$};
\node(a)at(6.6,0){ $\it 5/2$};
\node(a)at(8.8,0){ $\it 3$};
\node(a)at(7.3,-1.5){ $\it 2$};
\node(a)at(7.3,-3){ $\it 1$};
\node(a)at(8.3,-2){ $\it 2$};
\end{tikzpicture}
\end{center}
\end{example}
\section{Minimal implies Lipschitz normally embedded} \label{sec: main proof}
The aim of this section is to prove one direction of Theorem \ref{theorem:main}: any minimal surface singularity is Lipschitz normally embedded. We first state and prove the key Proposition \ref{lem:generic2}.
Let $(X,0)$ be a normal surface germ and let $\ell \colon (X,0) \to
(\C^2,0)$ be a generic projection. Let $U$ be an open neighborhood of
$0$ in $\C^2$ and let $e\colon U' \to U $ be the blow-up of the
origin.
Let $\widehat X$ be the pull-back of $\ell$ and $e$ and let
$\widehat\ell \colon \widehat X \to U'$ and $\widehat e \colon \widehat X \to
X$ be the two projections. Let $n\colon X'\to \widehat X$ be the
normalization of $\widehat X$. By \cite[Prop.~2.15]{BL}, $e'=\widehat e\circ n$ is the normalized blowup of the maximal ideal of $(X,0)$.
We then have a commutative diagram:
$$\xymatrix{X'\ar@{->}[drr]^{e'}\ar@{->}[dr]^(.7)n\ar@{->}[ddr]_{\ell'}\\
&\widehat X\ar@{->}[r]_{\widehat e}\ar@{->}^{\widehat \ell}[d]&X\ar@{->}^{\ell}[d]\\
&U'\ar@{->}^e[r]&U
}
$$
When $(X,0)$ is rational, a fortiori when minimal, $e'$ is the blowup of the maximal ideal; no normalization is needed (\cite{Tyu}).
\begin{proposition} \label{lem:generic2} Let $(X,0)$ be a minimal surface singularity. Let $\Pi$ be the polar curve of a generic projection $\ell\colon(X,0)\to (\C^2,0)$ and let
$E'=e'^{-1}(0)$. Choose $p \in E'$ such that if it is a smooth point of $X'$, then it is not on the strict transform $ \Pi^*$ of $\Pi$.
Set $q := \ell'(p)$. Then
\begin{itemize}
\item[(i)] \label{(i)} the germ $(X',p)$ is a minimal singularity with multiplicity the number of irreducible components of $E'$ through $p$;
\item[(ii)] \label{(ii)} the map germ $\ell' \colon (X',p) \to (U',q)$ is a generic projection for $(X',p)$.
\end{itemize}
\end{proposition}
\begin{remark} If $(X',p)$ is smooth and $p \in \Pi^*$, then, according to \cite{B1} either $(\Pi^*,p)$ is the strict transform of a moving polar (i.e., $p$ is not a basepoint of the family of generic polars) or $(X,0)$ is the singularity $A_2: x^2 + y^2 + z^3=0$. In both cases, it is easy to see that the degree of $\ell'$ at $(X',p)$ equals $2$ while the multiplicity of $(X,p')$ is $1$. So $\ell' \colon (X',p) \to (U',q)$ is not generic.
\end{remark}
\begin{remark} The fact the $(X',p)$ is minimal is \cite[Th\'eor\`eme 5.9]{BL}. The authors prove it there without using the existence of a resolution of $(X,0)$. We give here a short proof using this fact.
\end{remark}
\begin{proof}[Proof of Proposition \ref{lem:generic2}] Let $\pi \colon Y \to X$ be the minimal resolution of $(X,0)$ and let
$\Gamma$ be its resolution graph. Since $(X,0)$ is rational, then
$\pi$ factors through the blow-up of the maximal ideal
(\cite{A}). Assume $(X',p)$ is not smooth. Then $(X',p)$ has minimal resolution graph one of the connected components
$\Gamma'$ of $\Gamma$ minus the $\cal L$-nodes. So $\Gamma'$ is a rational graph and $(X',p)$ is rational. Moreover, since the minimal cycle of $(X,p)$ is reduced, the minimal cycle of $(X',p)$ is also reduced and the multiplicity $m'$ of $(X',p)$ equals the
number of $\cal L$-nodes adjacent to $\Gamma'$, i.e., the number of
irreducible components of $E'$ containing $p$.
Assume now that $(X',p)$ is smooth. Then by hypothesis, $p \not\in \Pi^*$. Since there is a branch of $\Pi^*$
through any singular point of $E'$, it implies that $p$ is a smooth point of $E'$. So the number of branches of $E'$
through $p$ equals $1$. This proves (i).
In order to prove (ii), we have to check that $\ell' \colon (X',p) \to (U',q)$ satisfies Conditions \eqref{generic1} and \eqref{generic2} of Definition \ref{def:generic}.
The map $\ell' \colon X' \to U'$ is a branched cover with degree the
multiplicity $m=m(X,0)$ of $(X,0)$. Its discriminant locus is included in
the strict transform $\Delta^*$ of $\Delta$ by $e$ union the
exceptional curve $C=e^{-1}(0)$. This divisorial discriminant is
computed in \cite[Proposition 6.1]{BL} for any normal surface germ
$(X,0)$: it equals $\Delta^* + (m -r) C$ where $r$ the number of branches of the generic
hyperplane section of $(X,0)$. In our case, $r=m$ since $(X,0)$ is
minimal (\cite[Lemma 5.4 and Theorem 5.8 ]{BL}). So, the discriminant
of $\ell'$ is just the reduced curve $\Delta^*$ and the branching locus of $\ell'$ is $\Pi^*$. In particular the polar curve of $\ell' \colon (X',p) \to (U',q)$ is the germ $(\Pi^*,p)$.
Assume $(X',p)$ is smooth and $p\notin \Pi^*$. Then $\ell' \colon (X',p) \to (U',q)$ has empty polar curve, so it is an
isomorphism and $\ell' \colon (X',p) \to (U',q)$ is a generic projection. This proves (ii) in that case.
Assume now $(X',p)$ is not smooth. Then $p \in \Pi^*$ since $(X,0)$ is resolved by a sequence of blowing-ups
of points on the successive strict transforms of $\Pi$ (\cite[Lemma 3.1]{B1}. Since the projection $\ell$ is generic for its polar curve (Condition \eqref{generic1} of Definition \ref{def:generic}), then the following Lemma \ref{lem: curve generic} (which will be used again later) implies that $\ell' \colon (X',p) \to (U',q)$ is a generic projection for its polar curve $(\Pi^*,p)$, i.e., it satisfies Condition \eqref{generic1} of Definition \ref{def:generic}.
\begin{lemma} \label{lem: curve generic} Denote by $e_0\colon U_0 \to
\C^N$ the blow-up of the origin of $\C^N$ and let $(\gamma,0)\subset
(\C^N,0)$ be a curve germ whose strict transform $\gamma^*$ by $e_0$
intersects $e_0^{-1}(0)$ at a unique point $p$. Let $\ell \colon
\C^N \to \C^2$ be a linear projection which is generic for the curve
germ $(\gamma,0)$.
We may choose coordinates so that $\ell$ is the
map $(z_1,\dots, z_N)\mapsto (z_1,z_2)$. Denote by $\mathbb P^{n-3}$ the
subset of $e_0^{-1}(0)$ given by $z_1=z_2=0$.
Let $e \colon U'\to U$ be the blow-up of the origin of $\C^2$, where
$U$ is a neighborhood of $0$. Then there is a map $\ell_0\colon
(U_0\setminus \mathbb P^{n-3}) \to U'$ with $e \circ \ell_0=\ell \circ
e_0 $. By genericity of $\ell$, the point $p$ is in $U_0\setminus \mathbb P^{n-3}$.
Set $q=\ell_0 (p)$. Then the map germ
$\ell_0 \colon (U_0\setminus\mathbb P^{n-3},p)\to (U',q)$ is generic for the curve germ
$(\gamma^*,p)$.
\end{lemma}
\begin{proof}
We will use the criterion of genericity introduced in the proof of
Theorem 5.1 in \cite{NP1}.
Let us first assume that $(\gamma,0)$ is irreducible and that the coordinates are chosen so that $(\gamma,0)$ admits a Puiseux parametrization of the form
$$\omega \mapsto (z_1(\omega),\ldots,z_N(\omega)) = (\omega^n,
\sum_{j\geq n}
a_{2j}\omega^j, \dots, \sum_{j\geq n} a_{Nj}\omega^j)$$
Set $A:=\{j: \exists i, a_{ij}\ne 0\}$ and call an exponent $j\in A\setminus\{n\}$ an
{\it essential integer exponent} if
$$ \gcd\{i\in \{n\}\cup A: i\le j\}<\gcd\{i\in \{n\}\cup
A:i<j\}.$$
Denote by $B$ the set of essential integer exponents of $(\gamma,0)$.
\medskip\noindent
{\bf Genericity criterion (\cite[Section 5]{NP1})} The projection $\ell$ is generic for the curve germ $(\gamma,0)$ if and only if $B \subset \{ j, \ a_{2j}\neq 0\}$.
\medskip We can assume our coordinates are chosen so that $(\gamma,0)$ is
tangent to the $z_1$-axis and then $A \subset \{j : j > n\}$ We
consider for $e_0$ and $e$ the chart over $z_1 \neq 0$ so that
writing $(z_1,v_2,v_3,\ldots,v_N)$ the corresponding local
coordinates of $U_0$ and $(z_1,v_2)$ that of $U'$, we have:
$e_{0}(z_1,v_2,v_3,\ldots,v_N)=(z_1,z_1v_2,z_1v_3,\ldots,z_1v_N)$ and
$e(z_1,v_2)=(z_1,z_1v_2)$. Then $q$ is the origin of the local
coordinates of $U_0$ and the strict transform $\gamma^*$ of $\gamma$
by $e_0$ has the following Puiseux parametrization in the coordinates
$(z_1,v_2,v_3,\ldots,v_N)$:
$$\omega \mapsto (\omega^n, \sum_{j\geq n}
a_{2j}\omega^{j-n},\dots , \sum_{j\geq n} a_{Nj}\omega^{j-n} )$$
Since $B \subset \{a_{2,j}\neq 0\}$, the set of essential integer exponents of $\gamma^*$ is $\{ j-n; j \in B\}$. Since $\ell_0$ is given by $\ell_0(z_1,v_2,v_3,\ldots,v_N)=(z_1,v_2)$, then, according to the above genericity criterion, $\ell_0$ is generic for $(\gamma^*,q)$.
The proof when $(\gamma,0)$ is reducible is essentially the same using
the extension of the genericity criterion in \cite[Section 5]{NP1} taking account of the contact exponents between branches.
\end{proof}
Let us now prove that Condition \eqref{generic2} of Definition \ref{def:generic} is satisfied. Let $\pi \colon \widetilde{X} \to X$ be the resolution introduced in Section \ref{outergeometry}. By definition, it factors through the blow-up $e'$. Consider the map ${\pi}' \colon \widetilde X \to X'$ defined by ${\pi}=e' \circ {\pi}' $. According to Theorem \ref{thm:Pi}, its restriction over $(X',p)$ is a resolution of $(X',p)$ which factors through the normalized Nash transform of $(X',p)$ and the $\cal P$-curves of $(X,0)$ and $(X',p)$ over $p$ coincide.
Now, take any $\cal D \subset \Omega$, where $(\ell_{\cal D} \colon (X,0) \to (\C^2,0))_{\cal D \in \Omega}$ denotes the family of generic projections of $(X,0)$. Let $\ell'_{\cal D}\colon (X',p) \to (\C^2,0)$ be the projection defined by $e \circ \ell'_{\cal D} = \ell_{\cal D} \circ e'$. We know that the polar curve $\Pi'_{\cal D}$ of $\ell'_{\cal D}$ equals the germ $(\Pi_{\cal D}^*,p')$ where $^*$ means strict transform by $e'$. Therefore the family of polars $(\Pi'_{\cal D})_{\cal D \in \Omega}$ of projections $\ell'_{\cal D}$ coincide with the family of germs $(\Pi_{\cal D}^*,p')_{\cal D \in \Omega}$, which is equisingular in terms of strong simultaneous resolution. This shows that Condition \eqref{generic2} of Definition \ref{def:generic} is satisfied for the family $(\ell'_{\cal D})_{\cal D \in \Omega} $.
\end{proof}
We now prove the ``if'' direction of Theorem \ref{theorem:main}.
Let $(X,0)$ be a minimal surface singularity with generic projection
$\ell \colon (X,0)\to (\C^2,0)$. Let $\rho \colon Y \to \C^2$ be the
sequence of blow-ups which resolves the base points of the family of
curves $(\ell(\Pi_{\cal D}))_{\cal D \in \Omega}$ and let $R$ be its
resolution graph. We have to check Conditions $(\ast1)$ and $(\ast2)$
of Theorem \ref{prop:characterization of normal embedding} for any
test curve $(\gamma,0)$ associated to a node of $R$. In fact, we will
check these conditions for the $\rho$-image of a curvette of any
irreducible curve in $\rho^{-1}(0)$ (so any vertex of $R$, not only
nodes). In the proof we say test curve for such a curve even if it correspond to a non node vertex. By Proposition \ref{prop:discriminant}, the discriminant curve $\Delta$ of $\ell$ is a union of $A_{n}$-curves, and $\rho \colon Y \to \C^2$ is the minimal resolution of $\Delta$. We consider the following two cases:
\begin{itemize}
\item[Case 1.] $\gamma$ is the $\rho$-image of a curvette of $\rho^{-1}(0)$ whose inner rate $q_\gamma$ is an integer $n$ (in particular, $\gamma$ is smooth);
\item[Case 2.] $\gamma$ is the $\rho$-image of a curvette of $\rho^{-1}(0)$ such that $q_\gamma=n+1/2$ with $n \geq 1$ an integer;
\end{itemize}
\smallskip\noindent
{\bf Case 1.} We will proceed by induction on $q_\gamma$, so
assume first $q_{\gamma}=1$, i.e., $\gamma$ is a generic line through the origin of $\C^2$, so $(\ell^{-1}(\gamma),0)$ is a generic hyperplane section of $(X,0)$. Since $(X,0)$ is minimal, the generic hyperplane section $(\ell^{-1}(\gamma),0)$ also has a minimal singularity (\cite[Lemma 3.4.3]{K}) so it is a union of $m(X,0)$ smooth transversal curves, where $m(X,0)$ denotes the multiplicity of $(X,0)$. Therefore $\gamma$ has a single outer rate which equals $1$ and Conditions $(\ast1)$ and $(\ast2)$ are satisfied.
Let $n$ be an integer $\geq 2$. Assume that for any minimal singularity, any test curve with inner rate $n-1$ satisfies Conditions $(\ast1)$ and $(\ast2)$. Let $\gamma$ be the $\rho$-image of a curvette with inner rate $q_{\gamma}=n$. We use again the notations $\ell$, $e$, $\ell'$ and $e'$ introduced for Proposition \ref{lem:generic2} and we set $C=e^{-1}(0)$.
Consider the point $q = \gamma^* \cap C $, where $ ^*$ means strict transform by $e$. Since $n \geq 2$, the strict transform $\Delta^*$ contains $q$. Since $\ell$ is generic for its polar curve and since $e'$ is the blow-up of the origin, then the fiber $\ell'^{-1}(q)$ contains a unique point $p$ which belongs to the strict transform $\Pi^*$ of $\Pi$ by $e'$.
\noindent
{\bf Claim 1. } $\gamma$ satisfies condition $(\ast 1)$.
{\it Proof.} Let $\sigma$ be a component of $\ell^{-1}(\gamma)$. We have to prove that $(\sigma,0)$ is smooth.
Assume first that the strict transform $\sigma^*$ of $\sigma$ by $e'$ meets $E' = e'^{-1}(0)$ at a point $p' \in \ell'^{-1}(q)$ distinct from $p$. Then $p' $ does not belong to the strict transform $\Pi^*$ of $\Pi$ by $e'$. Therefore $(X',p')$ is smooth, $p'$ is a smooth point of $E' = e'^{-1}(0)$ by (i) of Proposition \ref{lem:generic2} and the map germ $\ell'\colon (X',p') \to (U',q)$ is an isomorphism by (ii) of Proposition \ref{lem:generic2}. Since $\gamma^*$ is a smooth curve transverse to $C$ at $q$, then $(\ell'^{-1}(\gamma^*),p') = (\sigma^*,p')$ is a curvette of $E'$. Since $(X,0)$ is minimal, the multiplicity of a generic linear form on $(X,0)$ has multiplicity $1$ along $E'$. By \cite[1.1]{GSL}, this implies that $\sigma$ is a smooth curve of $(X,0)$.
Assume now that $\sigma^* \cap E =p$. According to Proposition
\ref{lem:generic2}, the map germ $\ell' \colon (X',p)\to (U',q)$ is a
generic projection of $(X',p)$. Moreover, its discriminant and polar
curves are respectively the strict transform $(\Delta^*,q)$ of
$(\Delta,0)$ by $e$ and the strict transform $(\Pi^*,p)$ of $(\Pi,0)$
by $e'$. Since $\gamma$ has rate $n\geq 2$, then its strict transform
$(\gamma^*,q)$ by the blow-up of $0$ is a test curve with inner rate
$n-1$ (Remark \ref{lem:outer rates bl-up}). Taking $(\gamma^*,q)$ as test curve for $(X',p)$, we apply the induction assumption: $(\gamma^*,q)$ satisfies Condition $(\ast1)$, i.e., $(\gamma^*,q)$ is a smooth curve on $(X',p')$. Let $\pi_0 \colon (X_0,E) \to (X,0)$ be the minimal resolution of $(X,0)$. It factors through $e'$. Let $\pi'\colon X_0 \to X'$ be the resolution of $X'$ such that $\pi_0 = e' \circ \pi'$. Since $\sigma^*$ is a smooth curve on $(X',p')$, then by \cite[1.1]{GSL}, its strict transform $\sigma''$ by $\pi'$ is a curvette at a smooth point of $\pi'^{-1}(p')$ (along which the multiplicity of the maximal cycle is $1$). The curve $\sigma''$ is also the strict transform of $\sigma$ by $\pi_0$, and since the maximal cycle of $(X,0)$ is reduced, then again by \cite[1.1]{GSL} (we use here the converse statement) then $(\sigma,0)$ is a smooth curve on $(X,0)$.
\noindent
{\bf Claim 2. } $\gamma$ satisfies Condition $(\ast 2)$.
{\it Proof.} Let us write $\ell^{-1}(\gamma)$ as the union $\ell^{-1}(\gamma) = \eta_1 \cup \eta_2$ where $\eta_2$ is the union of components of $\ell^{-1}(\gamma)$ whose strict transforms by $e'$ contain $p$.
Let $\sigma$ be a component of $\eta_1$ and let $p' = \sigma^* \cap E'$. The strict transform by $e'$ of another component $\sigma_1$ of $\ell^{-1}(\gamma)$ meets $E'$ at a point in $\ell'^{-1}(q)$ different from $p'$. Therefore $\sigma$ and $\sigma_1$ have distinct tangent lines so their contact exponent equals $1$. This proves that any component of $\eta_1$ has contact exponent $1$ with any other component of $\ell^{-1}(\gamma)$.
It remains to prove Condition $(\ast 2)$ for two components $\delta_1$ and $\delta_2$ of $\eta_2$. Let $q_0$ the rate associated to $\delta_1$ and $\delta_2$ as in Condition $(\ast 2)$. The strict transforms $\delta_1^*$ and $\delta_2^*$ of $\delta_1$ and $\delta_2$ by $e'$ are two components of the liftings $\ell'^{-1}(\gamma^*)$ of the test curve $(\gamma^*,q)$ for the surface germ $(X',p')$, and the rate associated to $\delta_1^*$ and $\delta_2^*$ as in Condition $(\ast 2)$ for $(X',p')$ equals $q_0-1$. We now use the induction assumption: since $\gamma^*$ satisfies Condition $(\ast 2)$ as a test curve of $(X',p')$ with rate $n-1$, then $q_0-1=n-1$, so $q_0=n$. This proves Claim 2.
\noindent
{\bf Case 2.} We now assume $\gamma$ is a curvette of
$\rho^{-1}(0)$ with inner rate $q_{\gamma} = n + 1/2$ where $n$ is
an integer $\geq 1$. Then, in suitable coordinates $x$ and $y$,
$\gamma$ is a curve with a Puiseux expansion of the form $y=a
x^{n+\frac{1}{2}}$ and there is a unique component $\Delta'$ of
$\Delta$ with same type $y=a' x^{n+\frac{1}{2}}+$ higher order. Let
$q = \gamma^* \cap C = \Delta'^* \cap C$ as in Case 1, and let $\Pi' \subset \Pi$ such that $\ell (\Pi') = \Delta'$ and $p = \ell'^{-1}(q) \cap \Pi'^*$.
We will proceed again by induction on $n$, using similar arguments as in Case 1.
Assume first $n=1$, i.e., $q_{\gamma}=3/2$. Then $\Delta'$ and $\gamma$ are $3/2$-cusps, i.e., equisingular to $u^2-v^3=0$, with contact exponent $3/2$. The strict transforms $\Delta'^*$ and $\gamma^*$ by $e$ are smooth curves meeting $e^{-1}(0)$ at the same point $q$.
Let $p' \in \ell'^{-1}(q)$ be distinct from $p$. Then the map germ $\ell' \colon (X',p') \to (U',q)$ is an isomorphism, so the germ $(\ell'^{-1}(\gamma^*),p')$ is a smooth curve tangent to $E'$. Therefore, there is a unique component $\sigma$ of $\ell^{-1}(\gamma)$ whose strict transform by $e'$ contains $p'$, and it has multiplicity $2$. So $\sigma$ is a plane curve, and since it is smooth after one blow-up, it is a $3/2$-cusp. Moreover, a similar argument as in the proof of Claim 1 shows that $\sigma$ has contact exponent $1$ with any other component of $\ell^{-1}(\gamma)$.
Let us now deal with $(\ell'^{-1}(\gamma^*),p)$. According to Theorem \ref{genericpolar}, $p$ is a smooth point of an exceptional curve obtained by blowing-up the intersection point between two exceptional curves of $\ell'^{-1}(C)$ corresponding to a central edge in the resolution graph, and $(\ell'^{-1}(\gamma^*),p)$ consists of the strict transform of a component of $\ell^{-1}(\gamma)$ which is equisingular to $\Pi'$. So this component is a cusp, i.e., its unique rate is $3/2$. This implies that $\gamma$ satisfies Conditions $(\ast 1)$ and $(\ast 2)$.
The rest of the induction uses the same arguments as in Case 1.
This completes the proof of the ``if'' direction of Theorem \ref{theorem:main}.
\section{Explicit example of lifting of test curves}\label{sec: example liftings}
The aim of this section is to give in an explicit example a flavor of Conditions $(\ast 1)$ and $(\ast 2)$ of Theorem \ref{prop:characterization of normal embedding} in the case of a minimal singularity. We return back to Example \ref{example}, and we give for some examples of test curves $(\gamma,0)$, the resolution graph of the lifting $\ell^{-1}(\gamma)$ on the left and the resolution of its generic projection $\ell'(\ell^{-1}(\gamma))$ on the right. Figure \ref{fig:example1} is for the test curve given by a generic line. Figures \ref{fig:example2} and \ref{fig:example3} are for two test curves which are $\rho$-images of curvettes of the exceptional curves corresponding respectively to the vertices $v_1$ and $v_2$.
\begin{figure}[h]
\begin{tikzpicture}
\draw[thin ](-2,0)--(4,0);
\draw[thin ](1,0)--(1,-3);
\draw[thin,>-stealth,->](-2,0)--+(-1,-0.5);
\draw[thin,>-stealth,->](-2,0)--+(-1,0.5);
\draw[thin,>-stealth,->](-2,0)--+(-1.2,0);
\draw[thin,>-stealth,->](4,0)--+(1,0.3);
\draw[thin,>-stealth,->](2,0)--+(0.8,-0.5);
\draw[thin,>-stealth,->](1,-3)--+(1,0.3);
\draw[fill=black ] (-2,0)circle(2pt);
\draw[fill=white ] (-1,0)circle(2pt);
\draw[fill=gray ] (0,0)circle(2pt);
\draw[fill=white ] (1,0)circle(2pt);
\draw[fill=black] (2,0)circle(2pt);
\draw[fill=white ] (3,0)circle(2pt);
\draw[fill =black ] (4,0)circle(2pt);
\draw[fill=white ] (1,-1)circle(2pt);
\draw[fill=white ] (1,-2)circle(2pt);
\draw[fill =black ] (1,-3)circle(2pt);
\node(a)at(-2,0.3){ $\it 1$};
\node(a)at(0,0.3){ $\it 5/2$};
\node(a)at(1,0.3){ $\it 2$};
\node(a)at(2,0.3){ $\it 1$};
\node(a)at(3,0.3){ $\it 2$};
\node(a)at(4,0.3){ $\it 1$};
\node(a)at(0.6,-1){ $\it 3$};
\node(a)at(0.6,-3){ $\it 1$};
\draw[thin ](7,0)--(7,-3);
\draw[thin ](7,0)--(8.5,0);
\draw[thin ](7,-1.5)--(6,-0.5);
\draw[thin ](7,-3)--(8,-2);
\draw[thin,>-stealth,->](7,-3)--+(-0.8,0.8);
\draw[thin,>-stealth,->](7,-3)--+(-0.8,-0.8);
\draw[thin,>-stealth,->](7,-3)--+(-1,-0.5);
\draw[thin,>-stealth,->](7,-3)--+(-1,0.5);
\draw[thin,>-stealth,->](7,-3)--+(-1.1,-0.2);
\draw[thin,>-stealth,->](7,-3)--+(-1.1,0.2);
\draw[fill=black ] (7,-3)circle(2pt);
\draw[fill=white ] (7,0)circle(2pt);
\draw[fill=white ] (7,-1.5)circle(2pt);
\draw[fill=white ] (6,-0.5)circle(2pt);
\draw[fill=white ] (8,-2)circle(2pt);
\draw[fill=white ] (8.5,0)circle(2pt);
\node(a)at(5.7,-0.5){ $\it 3$};
\node(a)at(6.6,0){ $\it 5/2$};
\node(a)at(8.8,0){ $\it 3$};
\node(a)at(7.3,-1.5){ $\it 2$};
\node(a)at(7.3,-3){ $\it 1$};
\node(a)at(8.3,-2){ $\it 2$};
\end{tikzpicture}
\caption{}
\label{fig:example1}
\end{figure}
\begin{figure}[h]
\centering
\begin{tikzpicture}
\draw[thin ](-2,0)--(4,0);
\draw[thin ](1,0)--(1,-3);
\draw[thin,>-stealth,->](-2,0)--+(-1,-0.5);
\draw[thin,>-stealth,->](-2,0)--+(-1,0.3);
\draw[thin,>-stealth,->](-1,0)--+(0.8,0.8);
\draw[thin,>-stealth,->](4,0)--+(1,0.3);
\draw[thin,>-stealth,->](1,0)--+(0.8,-0.5);
\draw[thin,>-stealth,->](1,-2)--+(1,0.3);
\draw[fill=black ] (-2,0)circle(2pt);
\draw[fill=white ] (-1,0)circle(2pt);
\draw[fill=gray ] (0,0)circle(2pt);
\draw[fill=white ] (1,0)circle(2pt);
\draw[fill=black] (2,0)circle(2pt);
\draw[fill=white ] (3,0)circle(2pt);
\draw[fill =black ] (4,0)circle(2pt);
\draw[fill=white ] (1,-1)circle(2pt);
\draw[fill=white ] (1,-2)circle(2pt);
\draw[fill =black ] (1,-3)circle(2pt);
\draw[thin ](7,0)--(7,-3);
\draw[thin ](7,0)--(8.5,0);
\draw[thin ](7,-1.5)--(6,-0.5);
\draw[thin ](7,-3)--(8,-2);
\draw[thin,>-stealth,->](7,-3)--+(-1,-0.5);
\draw[ thin,>-stealth,->](7,-3)--+(-1.1,0);
\draw[thin,>-stealth,->](7,-3)--+(-1,0.5);
\draw[thin,>-stealth,->](7,-1.5)--+(1,0.5);
\draw[ thin,>-stealth,->](7,-1.5)--+(1.1,0);
\draw[thin,>-stealth,->](7,-1.5)--+(1.1,0.25);
\draw[fill=black ] (7,-3)circle(2pt);
\draw[fill=white ] (7,0)circle(2pt);
\draw[fill=white ] (7,-1.5)circle(2pt);
\draw[fill=white ] (6,-0.5)circle(2pt);
\draw[fill=white ] (8,-2)circle(2pt);
\draw[fill=white ] (8.5,0)circle(2pt);
\node(a)at(6.7,-1.7){ $v_1$};
\end{tikzpicture}
\caption{}
\label{fig:example2}
\end{figure}
\begin{figure}[h]
\centering
\begin{tikzpicture}
\begin{scope}[yshift=-4.3cm]
\draw[thin ](-2,0)--(4,0);
\draw[thin ](1,0)--(1,-3);
\draw[thin,>-stealth,->](-2,0)--+(-1,-0.3);
\draw[thin,>-stealth,->](-2,0)--+(-1,0.3);
\draw[thin,>-stealth,->](-2,0)--+(-1.1,0);
\draw[thin,>-stealth,->](3,0)--+(0.3,1);
\draw[thin,>-stealth,->](3,0)--+(-0.3,1);
\draw[thin,>-stealth,->](1,-3)--+(0.8,0.8);
\draw[fill=black ] (-2,0)circle(2pt);
\draw[fill=white ] (-1,0)circle(2pt);
\draw[fill=gray ] (0,0)circle(2pt);
\draw[fill=white ] (1,0)circle(2pt);
\draw[fill=black] (2,0)circle(2pt);
\draw[fill=white ] (3,0)circle(2pt);
\draw[fill =black ] (4,0)circle(2pt);
\draw[fill=white ] (1,-1)circle(2pt);
\draw[fill=white ] (1,-2)circle(2pt);
\draw[fill =black ] (1,-3)circle(2pt);
\draw[thin ](7,0)--(7,-3);
\draw[thin ](7,0)--(8.5,0);
\draw[thin ](7,-1.5)--(6,-0.5);
\draw[thin ](7,-3)--(8,-2);
\draw[thin,>-stealth,->](7,-3)--+(-1,-0.5);
\draw[ thin,>-stealth,->](7,-3)--+(-1.1,0.2);
\draw[ thin,>-stealth,->](7,-3)--+(-1.1,-0.2);
\draw[thin,>-stealth,->](7,-3)--+(-1,0.5);
\draw[thin,>-stealth,->](8,-2)--+(1,0.5);
\draw[thin,>-stealth,->](8,-2)--+(1.1,0.25);
\draw[fill=black ] (7,-3)circle(2pt);
\draw[fill=white ] (7,0)circle(2pt);
\draw[fill=white ] (7,-1.5)circle(2pt);
\draw[fill=white ] (6,-0.5)circle(2pt);
\draw[fill=white ] (8,-2)circle(2pt);
\draw[fill=white ] (8.5,0)circle(2pt);
\node(a)at(8,-1.7){ $v_2$};
\end{scope}
\end{tikzpicture}
\caption{}
\label{fig:example3}
\end{figure}
\section{Rational and Lipschitz normally embedded implies minimal}\label{sec:rational + NE implies minimal}
In this section, we prove the other direction of Theorem \ref{theorem:main}: any rational surface singularity which is Lipschitz normally embedded is minimal.
\begin{remark} A Lipschitz normally embedded surface singularity is
not necessarily minimal. A counter-example is given by the (non
rational) hypersurface in $\Bbb C^3$ with equation $xy(x+y) +
z^4=0$. It is a superisolated singularity. The graph of its minimal
resolution factorizing through Nash has four vertices. It consists
of a central vertex and three bamboos of length one, these three
leaves being the $\cal L$-nodes, and the central vertex the single
$\cal P$-node.
\end{remark}
\begin{proof}
Let $(\widetilde{X},E)$ be the minimal resolution of $(X,0)$, $Z$ the minimal
cycle and $E=\bigcup E_i$. The multiplicity of $Z$ at any $\cal L$-curve is $1$,
since $(X,0)$ is Lipschitz normally embedded. Consider Laufer's algorithm for
finding $Z$ (\cite[Proposition 4.1]{La}), and let $E_j\subset E$ be the last curve one adds
in the algorithm before one obtains $Z$. Assume that $E_j$ is not an
$\cal L$-curve, so $Z\cdot E_j=0$. Let $Z'$ be the penultimate cycle obtained by Laufer's algorithm. Then $Z'=Z-E_j$ and
$Z'\cdot E_j= -E_j^2>1$ which contradicts $(X,0)$ being rational by
Laufer's criterion \cite[Theorem 4.2]{La}. So the last curve added by Laufer's algorithm is
always an $\cal L$-curve.
One can always run Laufer's algorithm such that each curve is
added once, before any curve is added a second time. So unless $Z=\sum
E_i$ there
would be an $\cal L$-curve with multiplicity $>1$, which is a contradiction. Thus $(X,0)$ is minimal.
\end{proof}
|
\section{Introduction}
For two natural numbers $n,m \in \bbn,$ let
$$
U_{m,n}(\bbc):=
\{
X \in M_{m,n}(\bbc)
\,\, | \,\,
X^{*} X = \lambda I_n \text{ for some } \lambda \in \bbc - \{0\}
\},
$$
which may be regarded as a generalization of a unitary group.
It plays an important role in studying the princiapl $U(n)$ bundles $U(n) \ra U(n+m)/U(m) \ra G_{n,m}$ over Grassmannian manifolds,
where, for $k \in \bbn,$ $U(k)$ has a metric, related to the Killing-Cartan form, given by
$$
\langle A,B \rangle
= \tfrac{1}{k} \text{Re}\big(\text{Tr}(A^{*}B)\big),
\qquad A,B \in \mathfrak{u}(k),
$$
and each quotient space has the induced metric which makes the projection a riemannian submersion.
Consider the Hopf fibration $S^1\ra S^3\ra S^2$. Let $\gamma$ be a
simple closed curve on $S^2$. Pick a point in $S^3$ over
$\gamma(0)$, and take the unique horizontal lift $\wt\gamma$ of
$\gamma$. Since $\gamma(1)=\gamma(0)$, $\wt\gamma(1)$ lies in the
same fiber as $\wt\gamma(0)$ does. We are interested in
understanding the difference between $\wt\gamma(0)$ and
$\wt\gamma(1)$. The following equality was already known
\cite{Pin}:
$$
V(\gamma)=e^{\frac{1}{2} A(\gamma) i},
$$
where $V(\gamma)$ is the holonomy
displacement along $\gamma$, and
$A(\gamma)$ is the area of the region surrounded by $\gamma$.
In this paper, we generalize this fact to the following higher dimensional Stiefel bundle over the Grassmannian manifold through $U_{m,n}(\bbc)$
$$
U(n) \ra U(n+m)/U(m) \stackrel{\pi}\ra G_{n,m},
$$
where $G_{n,m}=U(n+m)/\big(U(n) \times U(m)\big).$
\noindent
The main results are stated as follows:
For $\hat{X} \in {\mathfrak u}(n+m)$ which is induced by
$X \in U_{m,n}(\bbc)$,
consider a 2-dimensional subspace
$\frakm' \subset \frakm $ $ \subset \mathfrak{u}(m+n) $
with $\hat{X} \in \frakm'.$
Assume either
$$\frakm' = \text{Span}_{\bbr}\{\hat{X},\hat{Y}\}$$
for some $Y \in U_{m,n}$ with $X^{*}Y = \mu I_n$ for some $\mu \in \bbr$
or
$$\frakm' = \text{Span}_{\bbr} \{\hat{X},\widehat{iX}\}.$$
Then $\frakm'$ gives rise to a complete totally geodesic surface $S$ in the base
space. Furthermore,
let $\gamma$
be a piecewise smooth, simple closed curve on $S$
parametrized by $0\leq t\leq 1$, and $\wt\gamma$ its horizontal lift on
the bundle
$U(n) \ra \pi^{-1}(S) \stackrel{\pi}{\rightarrow} S,$
which is immersed in
$U(n) \ra U(n+m)/U(m) \stackrel{\pi}\ra G_{n,m} $.
Then
$$
\wt\gamma(1)= \wt\gamma(0) \cdot ( e^{i \theta} I_n)
\text{\hskip24pt or\hskip12pt } \wt\gamma(1)= \wt\gamma(0),
$$
depending on whether the immersed bundle is flat or not,
where $A(\gamma)$ is the area of the region on the surface $S$
surrounded by $\gamma$ and
$\theta= 2 \cdot \tfrac{n+m}{2n} A(\gamma).$
See {\rm Theorem \ref{thm-sphere}}.
\section{
The bundle
$
S^1 \ra SU(2) \ra \bbc P^1
$
} \label{hopf}
It will be studied not only the holonomy displacement of the bundle
$S^1 \ra SU(2) \ra \bbc P^1$
but also its isomorphic equivalence to the one
$$
S\big( U(1) \times U(1) \big)
\ra
SU(1+1)
\ra
SU(1+1) / S\big(U(1) \times U(1)),
$$
not the isometric equivalence.
In fact, a conformal map $h: SU(1+1) / S\big(U(1) \times U(1)) \ra \bbc P^1$ will be constructed such that the identity map on $SU(2)$ is the bundle map covering it.
The latter bundle will play an important role for the case
$\frakm' = \text{Span}_{\bbr} \{\hat{X},\widehat{iX}\}.$
Of course,
$$
S^3\cong SU(2)=\{A\in\GL(2,\bbc)\ :\ A^* A=I\text{ and } \det(A)=1\}
$$
for $S^3 = \{(z_1, z_2) \big| |z_1|^2 + |z_2|^2 =1\}$
under the map
$$
(z_1, z_2)
\mapsto
\left[\begin{array}{cccc}
\bar{z}_1 & - \bar{z}_2\\
z_2 & z_1\\
\end{array}
\right]
: S^3 \ra SU(2).
$$
From now on, we use the convention of
$\frak{gl}(k,\bbc)\subset\frak{gl}(2k,\bbr)$ by
$$
\left[\begin{array}{cccc}
z_{11} & z_{12}\\
z_{21} & z_{22}\\
\end{array}\right]
\lra
\left[\begin{array}{cccc}
x_{11} +i y_{11} & x_{12} +i y_{12} \\
x_{21} +i y_{21} & x_{22} +i y_{22} \\
\end{array}\right]
\lra
\left[\begin{array}{rrrrrrrr}
x_{11} &-y_{11} &x_{12} &-y_{12}\\
y_{11} & x_{11} &y_{12} & x_{12}\\
x_{21} &-y_{21} &x_{22} &-y_{22}\\
y_{21} & x_{21} &y_{22} & x_{22}\\
\end{array}\right],
$$
which is an isometric monomorphism with respect to the metric
on $GL(k, \bbc)$ and on $GL(2k, \bbr),$ given by
$$
\langle A,B \rangle
= \tfrac{1}{k}\text{Re}\big(\text{Tr}(A^{*}B) \big),
\qquad A,B \in \frak{gl}(k,\bbc)
$$
and
$$
\langle C,D \rangle
= \tfrac{1}{2k}\text{Tr}(C^{t}D), \qquad C,D \in \frak{gl}(2k,\bbr),
$$
respectively.
\bigskip
The group $SU(2)$ has the following natural representation into
$\GL(4,\bbr)$:
$$
w=
\left[
\begin{array}{rrrr}
\fm w_1 &w_2 &-w_3 &-w_4\\
-w_2 &w_1 &w_4 &-w_3\\
w_3 &-w_4 &w_1 &-w_2\\
w_4 &w_3 &w_2 &w_1
\end{array}
\right]
$$
with the condition $w_1^2 + w_2^2 + w_3^2 + w_4^2 =1$.
In fact, the map
$$
w_1 + w_2 i + w_3 j + w_4 k
\longmapsto
w
$$
is an isometric monomorphism from the unit quaternions into $\gl(4,\bbr)$.
The circle group
$$
S\big( U(1) \times U(1) \big) =
\left\{
\left[
\begin{array}{ll}
e^{-iz} &0 \\
0& e^{iz} \\
\end{array}\right]\ :\ 0\leq z\leq 2\pi
\right\}
$$
\noindent
is a subgroup of $SU(2)$, and acts on $SU(2)$ as right translations,
freely with quotient
$SU(1+1) / S\big(U(1) \times U(1)),$
which is an affine symmetric space and produces a principal circle bundle
$$
S\big( U(1) \times U(1) \big)
\ra
SU(1+1)
\ra
G_{1,1}= SU(1+1) / S\big(U(1) \times U(1)).
$$
\bigskip
Let $\wt w$ be the ``$i$-conjugate'' of $w$ (replace $w_2$ by $-w_2$). That is,
$$
\wt w=
\left[
\begin{array}{rrrr}
\fm w_1 &-w_2 &-w_3 &-w_4\\
w_2 &w_1 &w_4 &-w_3\\
w_3 &-w_4 &w_1 &w_2\\
w_4 &w_3 &-w_2 &w_1
\end{array}
\right].
$$
Then,
$$
w \wt w=
{\tiny
\left[
\begin{array}{rrrr}
w_1^2+w_2^2-w_3^2-w_4^2 &0 &-2 (w_1 w_3+w_2 w_4) &2 w_2 w_3-2 w_1 w_4\\
0 &w_1^2+w_2^2-w_3^2-w_4^2 &-2 w_2 w_3+2 w_1 w_4 & -2 (w_1 w_3+w_2 w_4)\\
2 (w_1 w_3+w_2 w_4) &2 w_2 w_3-2 w_1 w_4 &w_1^2+w_2^2-w_3^2-w_4^2 &0\\
-2 w_2 w_3+2 w_1 w_4 &2 (w_1 w_3+w_2 w_4) &0 &w_1^2+w_2^2-w_3^2-w_4^2
\end{array}
\right]
}
$$
and
$$
(w_1^2+w_2^2-w_3^2-w_4^2)^2+(2 w_1 w_3+2 w_2 w_4)^2 +(-2 w_2 w_3+2 w_1
w_4)^2=1.
$$
Clearly, $\bbc P^1$ can be identified with the following
$$
\bbc P^1 =
\left\{
\left[
\begin{array}{rrrr}
x & 0 &-y &-z \\
0 & x & z &-y\\
y &-z & x & 0\\
z & y & 0 & x
\end{array}\right]
:\ x^2+y^2+z^2=1
\right\},
$$
which is a subset of $SU(2)$ such that $i$-conjugate on $\bbc P^1$ is the identity map of $\bbc P^1$.
And the map
$$
p: SU(2)\lra \bbc P^1
$$
defined by
$$
p(w)=w \wt w
$$
has the following properties:
\begin{align*}
p(wv)&=w p(v) \wt w\quad\text{for all } w,v\in SU(2)\\
p(wv)&=p(w) \quad\text{if and only if}\quad
v\in S\big(U(1) \times U(1)) \cong S^1
\end{align*}
under the convention of
$S\big(U(1) \times U(1)) \hookrightarrow GL(4,\bbr).$
This shows that the map $p$ is, indeed, the orbit map of the
principal bundle
$$S^1\lra SU(2) \stackrel{p}\lra \bbc P^1.$$
But we have to be careful that the inclusion map $\bbc P^1 \hookrightarrow SU(2)$ is not a cross-section in this bundle. In fact,
$p(v) = v^2 \in \bbc P^1$ for any $ v \in \bbc P^1.$
\bigskip
Define a map $h: SU(2)/S\big(U(1) \times U(1)\big) \ra \bbc P^1$ by
$$h(v H) = v^2 = p(v) \qquad v \in \bbc P^1,$$
where $H = S\big(U(1) \times U(1)\big).$
Then, the identity map of $SU(2)$ is a trivially isomorphic bundle map which covers the map $h.$
Under the identification
$
(x,y,z) =
\left[
\begin{array}{rrrr}
x & 0 &-y &-z \\
0 & x & z &-y\\
y &-z & x & 0\\
z & y & 0 & x
\end{array}\right]
: S^2 \cong \bbc P^1,
$
give the metric $\langle \cdot , \cdot \rangle _{S^2}$ of $S^2$ to $\bbc P^1$ and consider a metric space
$\big( \bbc P^1, \langle \cdot , \cdot \rangle _{S^2} \big) .$ Will $h$ be an isometry?
\bigskip
The Lie group $SU(2)$ will have a left-invariant Riemannian metric
given by the
following orthonormal basis on the Lie algebra $\mathfrak{su}(2) $
$$
E_1 =
\left[
\begin{array}{cc}
0 & -1 \\ 1 & 0
\end{array}
\right]\ ,
\quad
E_2 =
\left[
\begin{array}{cc}
0 & i \\ i & 0
\end{array}
\right]\ ,
\quad
E_3 =
\left[
\begin{array}{cc}
-i & 0 \\ 0 & i
\end{array}
\right]\ ,
$$
which correspond to
\[
e_{1}= \left( \begin{array}{rrrr}
0 & 0 & -1& 0 \\
0 & 0 & 0 & -1\\
1 & 0 & 0 &0\\
0 & 1 & 0 &0
\end{array}\right),\;\;
e_{2}= \left( \begin{array}{rrrr}
0 & 0 & 0& -1 \\
0 & 0 & 1 & 0\\
0 & -1 & 0 &0\\
1 & 0 & 0 &0
\end{array}\right) ,\;\;
e_{3}=
\left( \begin{array}{rrrr}
0 & 1 & 0& 0 \\
-1 & 0 & 0 & 0\\
0 & 0 & 0 &-1\\
0 & 0 & 1 &0\\
\end{array}\right)
\]
in $\frak{gl}(2k,\bbr),$ respectively.
Notice that
$[e_1,e_2]=2 e_3$.
\bigskip
In order to understand
the map $h$ between base spaces and the projection map $p$ better, consider the subset of $SU(2)$:
\begin{align*}
T&=
\left\{
\left[\begin{array}{ll}
\cos x &-(\sin x)e^{-i y}\\
(\sin x)e^{i y} &\cos x\\
\end{array}\right]\ :\ 0\leq x\leq \pi,\ 0\leq y\leq 2\pi
\right\}\\
&=
\left\{
\left[\begin{array}{rrrrrrrr}
\cos x &0 &-(\sin x)(\cos y) &-(\sin x)(\sin y)\\
0 &\cos x & (\sin x)(\sin y) &-(\sin x)(\cos y)\\
(\sin x)(\cos y) &-(\sin x)(\sin y) &\cos x &0 \\
(\sin x)(\sin y) & (\sin x)(\cos y) &0 &\cos x\\
\end{array}\right]
\right\} \\
\end{align*}
which is the exponential image of
$$
\frakm=
\left\{\left[\begin{array}{cc}
0 & -\bar{\xi}^{t} \\ \xi &0
\end{array} \right]\ :\ \xi \in {\bbc} \right\}.
$$
Furthermore, it is exactly same as $\bbc P^1,$ so the map $p$ restricted to $T$ is just the squaring map; that is,
$$
p(w)=w^2,\quad w\in T.
$$
\medskip
To check $h$ is a conformal map:
given
$$
w =\big(\cos{x}, (\sin{x})(\cos{y}), (\sin{x})(\sin{y}) \big) \in T = \bbc P^1,$$
\begin{align*}
\big| D_1(wH) \big|
&= \big|(D_1 w)^{\text{h}} \big| \\
&= \big|
\big(
(\cos{y}) {L_w}_{*} e_1 + (\sin{y}) {L_w}_{*} e_2
\big)^{\text{h}}
\big| \\
&= 1
\end{align*}
and
\begin{align*}
&\big| D_2(wH) \big| = \big|(D_2 w)^{\text{h}} \big| \\
&= \big|
\big( -\tfrac{1}{2} (\sin{2x})(\sin{y}) {L_w}_{*} e_1
+ \tfrac{1}{2} (\sin{2x})(\cos{y}) {L_w}_{*} e_2
- (\sin^2 {x}) {L_w}_{*} e_3
\big)^{\text{h}}
\big| \\
&= \tfrac{1}{2} \big| \sin{2x} \big|,
\end{align*}
while, under the expression $\langle a,b,c \rangle$ of vectors in $\bbr ^3,$
\begin{align*}
\big| D_1 \, h(wH) \big|
&= \big| D_1 \, w^2 \big| \\
&= \big|
\langle
-2\sin{2x}, \, 2 (\cos{2x}) (\cos{y}), \, 2 (\cos{2x}) (\sin{y})
\rangle
\big| \\
&= 2
\end{align*}
and
\begin{align*}
\big| D_2 \, h(wH) \big|
&= \big| D_2 \, w^2 \big| \\
&= \big|
\langle
0, \, - (\sin{2x}) (\sin{y}), \, (\sin{2x}) (\cos{y})
\rangle
\big| \\
&= \big| \sin{2x} \big|.
\end{align*}
Thus $h$ is a conformal map.
\bigskip
\begin{theorem}[\cite{Pin}]
\label{area-u2}
Let $S^1\ra SU(2)\ra \Big( \bbc P^1, \langle \cdot , \cdot \rangle _{S^2} \Big)$ be the natural fibration.
Let $\gamma$ be a piecewise smooth, simple closed curve on $\bbc P^1$.
Then the holonomy displacement along $\gamma$ is given by
$$
V(\gamma)= e^{\tfrac{1}{2} A(\gamma) i}
= e^{2 \cdot A( h^{-1} \circ \gamma) \Phi}
\in S^1 \cong S\big( U(1) \times U(1) \big)
$$
where $A(\gamma)$ is the area of the region on $\bbc P^1$
enclosed by $\gamma$ and
$$
\Phi =
\left[
\begin{array}{cc}
i & 0 \\ 0 & -i
\end{array}
\right].
$$
\end{theorem}
\begin{proof}
Let $\gamma(t)$ be a closed loop on $\bbc P^1$ with $\gamma(0)=p(I_4)$.
Therefore,
{\tiny
\begin{align*}
\gamma(t)=
\left[\begin{array}{rrrrrrrr}
\cos 2x(t) &0 &-\sin 2x(t)\cos y(t) &-\sin 2x(t)\sin y(t)\\
0 &\cos 2x(t) & \sin 2x(t)\sin y(t) &-\sin 2x(t)\cos y(t)\\
\sin 2x(t)\cos y(t) &-\sin 2x(t)\sin y(t) &\cos 2x(t) &0 \\
\sin 2x(t)\sin y(t) & \sin 2x(t)\cos y(t) &0 &\cos 2x(t)\\
\end{array}\right]
\end{align*}}
Let
\begin{align*}
{\tiny
\wt\gamma(t)=
\left[\begin{array}{rrrrrrrr}
\cos x(t) &0 &-\sin x(t)\cos y(t) &-\sin x(t)\sin y(t)\\
0 &\cos x(t) & \sin x(t)\sin y(t) &-\sin x(t)\cos y(t)\\
\sin x(t)\cos y(t) &-\sin x(t)\sin y(t) &\cos x(t) &0 \\
\sin x(t)\sin y(t) & \sin x(t)\cos y(t) &0 &\cos x(t)\\
\end{array}\right]
}
\end{align*}
with $0\leq x(t)\leq\pi/2$
so that $p(\wt\gamma(t))=\gamma(t)$ ($\wt\gamma$ is a lift of $\gamma$),
and let
\[
\omega(t)=
\left[
\begin{array}{rrrr}
\cos z(t) & -\sin z(t) & 0& 0 \\
\sin z(t) & \cos z(t) & 0 & 0\\
0& 0 & \cos z(t) & \sin z(t) \\
0& 0 & -\sin z(t) & \cos z(t) \\
\end{array}\right].
\]
Put
$$
\eta(t)=\wt\gamma(t)\cdot\omega(t).
$$
Then still $p(\eta(t))=\gamma(t)$, and $\eta$ is another lift of
$\gamma$. We wish $\eta$ to be the horizontal lift of $\gamma$.
That is, we want $\eta'(t)$ to be orthogonal to the fiber at $\eta(t)$.
The condition is that
$\angles{\eta'(t),(\ell_{\eta(t)})_{*}(e_{3})}=0$, or equivalently,
$\angles{(\ell_{\eta(t)\inv})_*\eta'(t),e_3}=0$.
That is,
$$
\eta(t)^{-1} \cdot \eta'(t)= \alpha_{1}e_1 + \alpha_{2}e_{2}
$$
for some $\alpha_1,\alpha_2 \in\bbr$.
From this, we get the following equation:
\begin{equation}
\label{shpere-z}
z'(t)=\sin^2 x(t) y'(t).
\end{equation}
Since any piecewise smooth curve can be approximated by a sequence of
piecewise linear curves which are sums of boundaries of rectangular
regions,
it will be enough to
prove the statement for a particular type of curves as follows ~\cite{CL}:
Suppose we are given a rectangular region in the $xy$-plane
\begin{align*}
p \leq x \leq p+a,\qquad
q \leq y \leq q+b.
\end{align*}
Consider the image $R$ of this rectangle in $\bbc P^1$ by the map
$$
(x,y)\mapsto \bfr(x,y)=(\cos 2x, (\sin 2x)(\cos y), (\sin 2x)(\sin y)).
$$
\noindent Then $||\bfr_x\x\bfr_y||=
2 \sin 2x$, (because $0\leq x\leq \pi/2$). Thus, the area of $R$ is
\begin{align*}
\int_q^{q+b}\int_p^{p+a} 2\sin 2x\ dx dy=2 b(\sin^2(p+a)-\sin^2(p)).
\end{align*}
On the other hand, the change of $z(t)$ along the boundary $\gamma(t)$ of this region
can be calculated using condition (\ref{shpere-z}). Let $\gamma(t)$ be represented by
$(p+4at,q)$ for $t\in [0, \frac{1}{4}]$,
$(p+a,q+b(4t-1)) $ for $t\in [\frac{1}{4}, \frac{1}{2}]$, $(p+a(3-4t),q+b)$ for $ t\in [\frac{1}{2}, \frac{3}{4}]$,
$(p,q+b(4-4t)) $ for $t\in [\frac{3}{4}, 1]$. Then
$$
z(1)-z(0)
=\int_0^1 z'(t) dt
=b\cdot\sin^2 (p+a)- b\cdot\sin^2 (p).
$$
Thus the total vertical change of $z$-values, $z(1)-z(0)$, along the perimeter of this rectangle
is
$$
b\cdot(\sin^2 (p+a)-\sin^2 (p))
$$
which is $\frac{1}{2}$ times the area. Hence we get the conclusion.
\end{proof}
\bigskip
\section{
The bundle
$U(n)\lra U(n+m)/U(m)\lra G_{n,m}$
}
To deal with the bundle
\[ U(n) \rightarrow U(n+m)/U(m) {\rightarrow} G_{n,m}\, , \]
we investigate the bundle
\[ U(n) \times U(m) \rightarrow U(n+m) {\rightarrow} G_{n,m}\, . \]
The Lie algebra of $U(n+m)$ is $\fraku(n+m)$, the skew-Hermitian
matrices, and has the following canonical decomposition:
$$
\frakg=\frakh +\frakm,
$$
where
$$
\frakh=\fraku(n)+\fraku(m)=\left\{
\left[\begin{array}{cc}
A & 0\\ 0 & B \end{array}\right]\ :\
A\in \fraku(n),\ B\in \fraku(m)\right\}
$$
and
$$
\frakm=
\left\{
\hat{X} \!:=
\left[
\begin{array}{cc}
0 & - X^* \\ X &0
\end{array}
\right]\
:\ X \in M_{m,n}(\bbc)
\right\}.
$$
\vspace{0.3cm}
Define an Hermitian inner product $h : \bbc ^m \rightarrow \bbc$ by
$$h(v,w) = v^*\, w,$$
where $v$ and $w$ are regarded as column vectors.
\begin{lemma}
\label{lambda}
If a matrix $X \in M_{m,n}$ satisfies $X^{*}X = \lambda I_n$ for some $\lambda \in \bbc,$ then $\lambda$ will be a nonnegtive real number and $\lambda=0$ only if $X$ is trivial.
\end{lemma}
\begin{proof}
Given any column vector $v$ of $X$, $\lambda = v^{*}v = h(v,v) \ge 0$ and the equality holds only if $v=0,$ which shows the claim.
\end{proof}
\bigskip
From the lemma \ref{lambda}, we obtain that
\begin{align*}
U_{m,n}(\bbc)
&=
\{
X \in M_{m,n}(\bbc)
\,\, | \,\,
X^{*} X = \lambda I_n \text{ for some } \lambda \in \bbc - \{0\}
\}
\\
&=
\{
X \in M_{m,n}(\bbc)
\,\, | \,\,
X^{*} X = \lambda I_n \text{ for some } \lambda > 0
\}.
\end{align*}
\bigskip
\begin{lemma}
\label{calculation}
Let
$$
X = \Big(a^r_k + i b^r_k \Big) \, ,
Y = \Big(c^r_k + i d^r_k \Big)
\in M_{m,n}(\bbc)$$
for $r=1, \cdots, m$, and $ k=1, \cdots, n$. Suppose that for their induced
$\hat{X}, \, \hat{Y} \in \frakm$,
$$
[[\hat{X}, \, \hat{Y}], \hat{X}] = \hat{Z} \in \frakm
$$
for some
$
Z =
\Big(\alpha ^r_k \Big) \in M_{m,n}(\bbc)
$ for $r=1, \cdots, m$, and $k=1, \cdots, n$.
Then we have
$$
\alpha ^r_k =
\sum^{n}_{j=1} (a^r_j + i b^r_j) \big(\!-2 h(Y_j, X_k) + h (X_j, Y_k)\big)
+ \sum^{n}_{j=1} (c^r_j + i d^r_j) \, h(X_j, X_k),
$$
where $X_k$ and $Y_k$ are $k$-column vectors of $X$ and $Y$ for $k= 1, \cdots, n.$
\end{lemma}
\begin{proof}
It is easily obtained from
$$
[[\hat{X},\hat{Y}], \hat{X}] =
\hat{X}(2 \hat{Y}\hat{X}-\hat{X}\hat{Y})
- \hat{Y}\hat{X}\hat{X}.$$
\end{proof}
Recall the following proposition, which gives the clue for the holonomy displacement in the principal $U(n)$ bundles over Grassmaniann manifolds
$U(n) \ra U(n+m)/U(m) \stackrel{\pi} \ra G_{n.m}$.
\begin{proposition} \cite{KN} \label{affine}
Let $(G,H,\sigma)$ be a symmetric space and
$\mathfrak{g} = \mathfrak{h} + \mathfrak{m}$
the canonical decomposition. Then there is a natural one-to-one correspondence between the set of linear subspaces $\mathfrak{m}'$ of $\mathfrak{m}$ such that
$[[\mathfrak{m}', \mathfrak{m}'], \mathfrak{m}'] \subset \mathfrak{m}'$
and the set of complete totally geodesic submanifolds $M'$ through the origin $0$ of the affine symmetric space $M=G/H,$ the correspondence being given by
$\mathfrak{m}' = T_0 (M').$
\end{proposition}
Note that $\mathfrak{m}'$ in the Proposition \ref{affine} will make a bunch of complete totally geodesic submanifolds, each of which is obtained from another one by a translation, in the affine symmetric space $G/H.$
\bigskip
The role of $U_{m,n}(C)$ in this paper will be seen from now on.
\bigskip
\begin{theorem} \label{easy}
Given $X \in U_{m,n}(\bbc)$ and
the natural fibration
$
U(n) \times U(m) \ra U(n+m) \ra G_{n,m}(\bbc),
$
assume a 2-dimensional subspace $\frakm' = \text{Span}_{\bbr} \{ \hat{X}, \hat{Y}\} $ of
$\frakm\subset {\mathfrak u}(n+m)$
satisfies
\begin{align}
X^* \, X = \lambda I_n, \quad X^* \, Y = \mu I_n, \qquad
\mu \in \bbc \label{STAR}
\end{align}
for $ Y \in M_{m,n}(\bbc).$
Then $\frakm'$ gives rise to a complete totally geodesic surface $S$ in
$G_{n,m}(\bbc)$
if and only if
$
\text{Im}\,{\mu} = 0
$
and $Y \in U_{m,n}(\bbc)$
\end{theorem}
\begin{proof}
To begin with, note that $\lambda >0.$
Assume that $\frakm'$ gives rise to a complete totally geodesic surface $S$ in $G_{n,m}(\bbc)$.
By a translation, without loss of generality, we can assume that $S$ passes through the origin of the affine symmetric space
$G_{n,m}(\bbc) = U(n+m)/\left(U(n) \times U(m)\right)$
To show $\text{Im}\mu = 0$ by contradiction, suppose that
$\text{Im}\mu \not= 0$. Let $e_k \in \bbc^m, \: k=1, \cdots, m,$ be an elementary vector which has all components 0 except for the $k$-component with 1. Then
$$h(X_k, Y_j) = h(X e_k, Y e_j) = e_k ^* (X^* Y) e_j ,$$
so the condition (\ref{STAR}) is equivalent to
$$
h(X_k, Y_k) = \mu, \quad
h(X_k, X_k) = \lambda, \quad
h(X_k, X_j) =0 , \quad
h(X_k, Y_j) =0
$$
\noindent
for $k \not= j $ in $\{1, \cdots ,n \}$. From $ h(X_k, Y_k) = \mu $,
we obtain
$$
-2h(Y_k, X_k) + h(X_k, Y_k) =
- \text{Re} \mu + 3i \text{Im} \mu.
$$
\noindent
Thus Lemma ~\ref{calculation}, Proposition \ref{affine} and the hypothesis of totally geodesic say that \\
\begin{align*}
a \hat{X} + b \hat{Y}
=
[[\hat{X}, \hat{Y}], \hat{X}]
&= (-\text{Re}\mu + 3i \text{Im}\mu) \hat{X} + \lambda \hat{Y}
\\
&= 3 \text{Im}\mu (i\hat{X})
+ (-\text{Re}\mu \hat{X} + \lambda \hat{Y}) .
\end{align*}
for some $a, b \in \bbr$.
Since $\text{Im}\mu \not= 0$, $i \hat{X}$ will lie in $\text{Span}_{\bbr} \{\hat{X}, \hat{Y} \} = \frakm' \subset \mathfrak{u}(n+m)$, and then
$$
-i\hat{X} = - (i \hat{X}) = (i\hat{X})^{*} = -i \hat{X}^{*} = i \hat{X},
$$
\noindent
which implies $\hat{X}=O_{n+m}$, a contradiction.
From $\text{Im}\mu = 0$,
$$
-X^* Y + Y^* X = - X^* Y + (X^* Y)^*
= - 2 i \text{Im}\mu \, I_n = O_n,
$$
\noindent
so
$$
[\hat{X}, \hat{Y}] =
\left[
\begin{array}{cc}
O_n & 0 \\ 0 & - X Y^* + Y X^*
\end{array}
\right]\
\in \mathfrak{u} (m) \subset \mathfrak{u} (n+m).
$$
Let $M = - X Y^* + Y X^*$. Then
$$
[\hat{X}, \hat{Y}] =
\left[
\begin{array}{cc}
O_n & 0 \\ 0 & M
\end{array}
\right]\
$$
and
$ [[\hat{Y}, \hat{X}], \hat{Y}] = - \widehat{MY} \in \frakm'$
from the hypothesis of the condition of totally geodesic and from
Proposition \ref{affine}.
Note that
$$
-MY = XY^*Y - Y X^*Y = XY^*Y - Y \mu I_n
= XY^*Y- (\text{Re}u) Y.
$$
Thus $XY^*Y = aX + bY$ for some $a, b \in \bbr $. Then
$
\lambda Y^*Y = X^*(XY^*Y)
= X^*( aX + bY) = (a \lambda + b \text{Re}\mu) I_n
$
and so
$$
Y^*Y= \tfrac{ a \lambda + b \text{Re}\mu}{\lambda}I_n, \quad
\tfrac{ a \lambda + b \text{Re}\mu}{\lambda} \in \bbr.
$$
Since $\mathfrak{m}'= \text{Span}_{\bbr} \{ \hat{X}, \hat{Y}\}$ is 2-dimensional, $Y$ is not a zero matrix and so from Lemma \ref{lambda},
$Y \in U_{m,n}(\bbc)$.
Conversely, assume the necessary part holds and let $Y^{*}Y = \eta I_n$, where $\eta > 0$.
Then, the condition $\text{Im}\mu = 0$ says that
$$
[\hat{X}, \hat{Y}] =
\left[
\begin{array}{cc}
O_n & 0 \\ 0 & M
\end{array}
\right]\ , \,\,
[[\hat{X}, \hat{Y}], \hat{X}] = \widehat{MX} \,\,
\text{ and } \,\,
[[\hat{Y}, \hat{X}], \hat{Y}] = -\widehat{MY},
$$
where $M = - X Y^* + Y X^*$. It suffices to show that $[[\hat{X}, \hat{Y}], \hat{X}] \in \frakm'$ and
$[[\hat{Y}, \hat{X}], \hat{Y}] \in \frakm'$.
Since
$$
MX = - XY^*X + Y X^*X = - X \bar{\mu} I_n + Y \lambda I_n
= - \text{Re}\mu X + \lambda Y,
$$ \noindent we get $[[\hat{X}, \hat{Y}], \hat{X}] \in \frakm'$.
We also get $[[\hat{Y}, \hat{X}], \hat{Y}] \in \frakm'$ since
$$
-MY = XY^*Y - Y X^*Y
= X \eta I_n - Y \mu I_n
= \eta X - \text{Re}\mu Y.
$$
Hence we get the conclusion.
\end{proof}
\begin{corollary}
Given $X,Y \in U_{m,n}(\bbc)$ and given the natural fibration
$
U(n) \times U(m) \ra U(n+m) \ra G_{n,m}(\bbc),
$
assume $\frakm' = \text{Span}_{\bbr} \{ \hat{X}, \hat{Y}\} $ produce a 2-dimensional subspace of $\frakm\subset {\mathfrak u}(n+m).$
If $X^* \, Y = \mu I_n \text{ for some } \mu \in \bbr,$
then
$\frakm'$ will give rise to a complete totally geodesic surface $S$ in
$G_{n,m}(\bbc)$
\end{corollary}
\begin{remark}
Given $X \in U_{m,n}(\bbc), $ if $n \le m, $ then
$X: \bbc^n \ra \bbc^m$ is a conformal one-one linear map. In view of
$\hat{X} \in {\mathfrak u}(n+m) \subset \text{End}(\bbc^{n+m})$,
$\hat{X}$ sends the subspace $\bbc^n$ to its orthogonal subspace $\bbc^m$ conformally. And the condition of the relation between $X$ and $Y$ in Theorem ~\ref{easy} says that
$$
h_{\bbc^m}(Xv, Yw) = \mu \: h_{\bbc^n}(v,w) \qquad
\text{ for } v, w \in \bbc^n,
$$
where $h_{\bbc^k}$ is an Hermitian on
$\bbc^k, \: k= 1,2, \cdots$, given by
$$
h_{\bbc^k}(u_1,u_2) = u_1^* u_2 \quad \text{ for } u_1, u_2 \in \bbc^k.$$
\end{remark}
\bigskip
When $n=1$, the condition (\ref{STAR}) is satisfied automatically for any two vectors in $\bbc^m$ by identifying
$M_{m,1}(\bbc)$ with $\bbc^m$.
So we get
\begin{corollary}
\label{geod-cond-cpn}
A 2-dimensional subspace $\frakm'$ of
$\frakm\subset {\mathfrak u}(m+1)$
gives rise to a complete totally geodesic submanifold in the affine symmetric space
${\bbc}P^m = U(1+m)/ \left(U(1) \times U(m) \right)$
if
$\frakm'$ has two linearly independent tangent vectors $\hat{v}$ and $\hat{w}$ such that
$\text{Im}h_{\bbc^m}(v,w) =0$.
\end{corollary}
\bigskip
We return to the bundle
$U(n) \ra U(n+m)/U(m) \stackrel{\pi}\lra G_{n,m}$.
Any submanifold $A \subset G_{n,m}$ induces a bundle
$U(n) \ra \pi^{-1}(A) \ra A$,
which is immersed in the original bundle and diffeomorphic to the pullback bundle with respect the inclusion of $A$ into $G_{n,m}$. In fact, in the bundle
$U(n)\times U(m) \ra U(n+m) \stackrel{\tilde{\pi}}\lra G_{n,m}$, the induced distribution in $\tilde{\pi}^{-1}(A)$
from ${\mathfrak u}$(m) in $U(n+m)$ is integrable and preserved by the right multiplication of $U(n),$ so this induces the bundle $U(n) \ra \pi^{-1}(A) \ra A$.
\begin{theorem}
\label{geod-cond-sphere1}
Given a complete totally geodesic surface $S$ in $G_{m,n}$ which is induced by a 2-dimensional subspace $\mathfrak{m}' \subset \mathfrak{m}$ with the necessary condition in Theorem ~\ref{easy} satisfied,
the bundle
$U(n) \ra \pi^{-1}(S) \ra S$,
which is immersed in the original bundle
$U(n) \ra U(n+m)/U(m) \stackrel{\pi}\lra G_{n,m}$,
is flat.
\end{theorem}
\begin{proof}
By a left translation, without loss of generality, assume that $S$ passes through the origin of the affine symmetric space $G_{n,m}.$
Consider the bundle
$U(n)\times U(m) \ra U(n+m) \stackrel{\tilde{\pi}}\lra G_{n,m}.$
Then $S$ induces a bundle
$U(n)\times U(m) \ra \tilde{\pi}^{-1}(S) \ra S$.
Totally geodesic condition says that the distribution induced from
$\text{Span}_{\bbr} \{ \hat{X}, \hat{Y}, [\hat{X},\hat{Y}] \}$ is integrable.
Since $[\hat{X},\hat{Y}]$ is contained in the Lie algebra $ {\mathfrak u}(m) $ of $U(m)$ from the proof of Theorem \ref{easy}, the conclusion is obtained.
\end{proof}
\bigskip
\begin{theorem}
\label{easy_generalizion}
\label{geod-cond-sphere}
Given $X \in U_{m,n}(\bbc)$ and the natural fibration
$
U(n) \times U(m) \ra U(n+m) \stackrel{\tilde{\pi}}\lra G_{n,m}(\bbc),
$
consider the 2-dimensional subspace $\frakm' = \text{Span}_{\bbr} \{ \hat{X}, \widehat{iX}\} $. Then,
\begin{enumerate}
\item
$\frakm'$ gives rise to a complete totally geodesic surface $S$ in
$G_{n,m}(\bbc),$
\item
$\frakm'$ induces a $U(1)$-subbundle of a bundle
$$U(n) \times U(m) \ra \tilde{\pi}^{-1}(S) \ra S,$$
which is an immersion of the bundle
$$
S\big( U(1) \times U(1) \big)
\ra
SU(1+1)
\ra
SU(1+1) / S\big(U(1) \times U(1))
$$
into
$$U(n) \times U(m) \ra U(n+m) \stackrel{\tilde{\pi}}\lra G_{n,m},$$
such that it is isomorphic to
the Hopf bundle $S^1 \ra S^3 \ra S^2,$
\item
the immersion is conformal, and isometric in case of $n=m.$ In fact,
$$ \big| \tilde{f}_{*}v \big| = \sqrt{\tfrac{2n}{n+m}} \, |v|$$
under the expression $\tilde{f}: SU(2) \ra U(n+m)$ for the immersion.
\end{enumerate}
\end{theorem}
\begin{proof}
From Lemma \ref{lambda}, let $X^* \, X = \lambda I_n$ for some $\lambda>0$.
By a left translation, without loss of generality, assume that $S$ passes through the origin of the affine symmetric space $G_{n,m}.$
Note that,
for
$
K =
\left[
\begin{array}{cc}
- i \lambda I_n & 0 \\ 0 & iXX^{*}
\end{array}
\right]\
\!\in {\mathfrak {u}}(n) \times {\mathfrak {u}}(m),
$
$$
[\hat{X},\widehat{iX}]=2 K, \quad
[K,\hat{X}]=2 \lambda \widehat{iX}, \quad
[K,\widehat{iX}]=-2 \lambda \hat{X},
$$
which implies $[[\frakm',\frakm'],\frakm'] \subset \frakm'$ and the conclusion (1).
Consider an orthonormal basis of
${\mathfrak {su}}(1+1)$:
$$
E_1 =
\left[
\begin{array}{cc}
0 & -1 \\ 1 & 0
\end{array}
\right]\ ,
\quad
E_2 =
\left[
\begin{array}{cc}
0 & i \\ i & 0
\end{array}
\right]\ ,
\quad
E_3 =
\left[
\begin{array}{cc}
-i & 0 \\ 0 & i
\end{array}
\right]\ ,
$$
and a Lie algbra monomorphism
$f : {\mathfrak su}(1+1) \ra {\mathfrak u}(n+m)$, given by
$$
f(aE_1+bE_2+cE_3)
= \frac{a}{\sqrt{\lambda}} \hat{X}
+ \frac{b}{\sqrt{\lambda}} \widehat{iX}
+ \frac{c}{\lambda} K
$$
\noindent
for $a,b,c \in \bbr,$ from
$$[E_1,E_2]=2 E_3, \quad [E_3,E_1]=2E_2, \quad [E_3,E_2]=-2E_1.$$
For any $\theta \in \bbr,$
$$
e^{\theta E_3} =
\left[
\begin{array}{cc}
e^{-i \theta} & 0 \\ 0 & e^{i \theta}
\end{array}
\right]\
\in S \big( U(1) \times U(1) \big).
$$
Thus $f$ will induce a Lie group monomorphism
$\tilde{f} : SU(1+1) \ra U(n+m)$ with
$
\tilde{f}\Big({S \big( U(1) \times U(1) \big)}\Big)
\subset
U(n) \times U(m)
$
since $SU(2)$ is simply connected and
$S\big(U(1) \times U(1)\big)$ is connected.
Furthermore, it is the bundle map from
$$
S\big( U(1) \times U(1) \big)
\ra
SU(1+1)
\ra
G_{1,1}= SU(1+1) / S\big(U(1) \times U(1))
$$
to
$$U(n) \times U(m) \ra U(n+m) \stackrel{\tilde{\pi}}\ra G_{n,m},$$
so the connected component of the integral manifold of the distribution induced by
$\text{Span}_{\bbr}\{K,\hat{X}, \widehat{iX}\},$
which is the image of $\tilde{f},$
shows (2).
Note that
$
\{
\frac{1}{\sqrt{\lambda}} \hat{X} ,
\frac{1}{\sqrt{\lambda}} \widehat{iX} ,
\frac{1}{\lambda} K
\}
$
is an orthogonal basis of the image of $\tilde{f}$ such that
$$
\sqrt{\tfrac{2n}{n+m}}
= \Big| \tfrac{1}{\sqrt{\lambda}} \hat{X} \Big|
= \Big| \tfrac{1}{\sqrt{\lambda}} \widehat{iX} \Big|
= \Big| \tfrac{1}{\lambda} K \Big|,
$$
which shows (3).
\end{proof}
\bigskip
\begin{remark} \label{fiber}
Let $\hat{\theta} = \tfrac{\theta}{\lambda}.$
Then, for $\Phi = -E_3,$
$$
\tilde{f}(e^{\theta \Phi})
= \tilde{f}(e^{-\theta E_3})
= e^{-\hat{\theta} K}
=
\left[
\begin{array}{cc}
e^{i \theta} I_n & 0 \\
0 & I_m + \tfrac{e^{-i \theta} -1}{\lambda} XX^{*}
\end{array}
\right]\
$$
from
$$
(-i \hat{\theta} XX^{*})^{j}
= \Big(\tfrac{-i\theta}{\lambda}\Big)^{j} X(X^{*}X)^{j-1}X^{*}
= \tfrac{(-i \theta)^{j}}{\lambda} XX^{*}
$$
for $j=1,2, \cdots .$
Furthermore,
\begin{align*}
&\Big(I_m + \tfrac{e^{-i \theta} -1}{\lambda} XX^{*}\Big)
\Big(I_m + \tfrac{e^{-i \phi} -1}{\lambda} XX^{*}\Big) \\
&=
I_m + \tfrac{e^{-i \theta} + e^{-i \phi} -2}{\lambda} XX^{*}
+ \tfrac{e^{-i(\theta +\phi)}-e^{-i \theta}-e^{-i \phi} +1}{\lambda ^2}
X(X^{*}X)X^{*} \\
&=
I_m + \tfrac{e^{-i(\theta +\phi)} -1}{\lambda} XX^{*},
\end{align*}
from which it is also obtained that
$$
I_m =
\Big(I_m + \tfrac{e^{-i \theta} -1}{\lambda} XX^{*}\Big)
\Big(I_m + \tfrac{e^{-i \theta} -1}{\lambda} XX^{*}\Big)^{*} .
$$
\end{remark}
\bigskip
We return to the bundle
$U(n) \ra U(n+m)/U(m) \stackrel{\pi}\lra G_{n,m}$.
In fact, Remark \ref{fiber} implies that the immersed $U(1)$-subbundle, which is the image of $\tilde{f},$ gives two $U(1)$-bundles, one of which is an immersed $U(1)$-subbundle in the bundle
$U(n) \ra U(n+m)/U(m) \stackrel{\pi}\lra G_{n,m}$
and the other one is an immersed $U(1)$-subbundle in the bundle
$U(m) \ra U(n+m)/U(n) \stackrel{\hat{\pi}}\lra G_{n,m}.$
\bigskip
\begin{theorem}
\label{thm-sphere}
Assume the same condition for a complete totally geodesic surface
$S$ of either Theorem \ref{easy} or Theorem ~\ref{easy_generalizion},
and consider the immersed bundle
$U(n) \ra \pi^{-1}(S) \stackrel{\pi}{\rightarrow} S$
in the bundle
$U(n) \ra U(n+m)/U(m) \stackrel{\pi} \ra G_{n,m}.$
Let $\gamma$ be a piecewise smooth, simple closed curve on $S$.
Then the holonomy displacement along $\gamma$,
$$\wt\gamma(1)= \wt\gamma(0) \cdot V(\gamma),$$
is given by the right action of
$$
V(\gamma)=e^{i \theta} I_n \quad \text{or} \quad e^{0i} I_n \in U(n),
$$
depending on whether the immersed bundle is flat or not,
where $A(\gamma)$ is the area of the region on the surface $S$ surrounded
by $\gamma$ and $\theta= 2 \cdot \tfrac{n+m}{2n} A(\gamma) .$
Especially, $\theta = 2\cdot A(\gamma)$ in case of $n=m.$
\end{theorem}
\begin{proof}
If the immersed bundle is flat, then it is obvious that the holonomy displacement is trivial. \\
Assume the condition of Theorem ~\ref{easy_generalizion} for the immersed $U(1)$-subbundle, which is the image of $\tilde{f}.$
Consider the induced map $\hat{f}: B \ra S \subset G_{m,n}$ between base spaces from the bundle map $\tilde{f}: SU(2) \ra \text{Im}(\tilde{f}) \subset U(n+m),$ which is a monomorphism, where $B = SU(2)/S(U(1) \times U(1)).$
Let
$\alpha = \sqrt{\tfrac{2n}{n+m}},$
$
\theta
= 2\cdot \tfrac{n+m}{2n} A(\gamma)
= \tfrac{\alpha ^{-2}}{8} A(\gamma)
$
and
$\hat{\theta} = \tfrac{\theta}{\lambda}.$
The Theorem ~\ref{easy_generalizion}, Theorem \ref{area-u2} and
Remark \ref{fiber} say that
the holonomy displacement of $\gamma$ in the bundle
$U(n) \times U(m) \ra \pi^{-1}(S) \stackrel{\pi}{\rightarrow} S,$
which is immersed in the bundle
$U(n) \times U(m) \ra U(n+m) \stackrel{\pi} \ra G_{n,m},$
is given by the right action of
\begin{align*}
V(\gamma)
&= \tilde{f} \big( V(\hat{f}^{-1} \circ \gamma) \big) \\
&= \tilde{f}
\big( e^{2 \cdot A(\hat{f}^{-1} \circ \gamma) \Phi} \big) \\
&= \tilde{f} \big( e^{\theta \Phi} \big) \\
&= \left[
\begin{array}{cc}
e^{i \theta} I_n & 0 \\
0 & I_m + \tfrac{e^{-i \theta} -1}{\lambda} XX^{*}
\end{array}
\right]\ .
\end{align*}
Thus in the bundle $U(n) \ra \pi^{-1}(S) \stackrel{\pi}{\rightarrow} S,$ which is immersed in the bundle
$U(n) \ra U(n+m)/U(m) \stackrel{\pi} \ra G_{n,m},$
the holonomy displacement is given by the right action of
$$
V(\gamma) = e^{i \theta} I_n .
$$
\end{proof}
\bigskip
\begin{remark}
For $n=1$, we have the following Hopf bundle
$S^1\ra S^{2m+1} \ra \bbc P^m, $ where $\bbc P^m$ is given by the quotient metric, so the projection is a Riemannian submersion.
Let $S$ be a complete totally geodesic
surface in $\bbc P^m$
and $\gamma$ be a piecewise smooth, simple closed curve on $S$.
Identify $\bbc ^m \cong M_{m,1}(\bbc).$
If $S$ is induced by $\text{Span} \{ v, w \} \subset \bbc ^m $ with
$\text{Im}h_{\bbc ^m}(v,w)=0,$ then the holonomy displacement along $\gamma$ is trivial.
See Corollary \ref{geod-cond-cpn} and Theorem \ref{geod-cond-sphere1}.
If $S$ is induced by a two dimensional subspace with complex structure in
$\bbc ^m,$
then the holonomy displacement depends not only on the area of the region surrounded by $\gamma$ but also on $m$ unless $m=1.$ In case of $m=1,$ here, $\bbc P^m$ is isometric to $S^2 \Big( \tfrac{1}{2} \Big).$ Refer to the map $h$ defined in Section \ref{hopf}.
\end{remark}
\medskip
\begin{remark}
Let $U(m) \ra U(n+m)/U(n) \stackrel{\hat{\pi}} \ra G_{n,m}$ be the natural
fibration.
Assume the same condition for a complete totally geodesic surface
$S$ of Theorem ~\ref{easy_generalizion},
and consider the bundle
$U(m) \ra \hat{\pi}^{-1}(S) \stackrel{\hat{\pi}}{\rightarrow} S$. Let $\gamma$ be a piecewise smooth, simple closed curve on $S$.
Then the holonomy displacement along $\gamma$ is given by the right action of
$$
V(\gamma)=I_m + \tfrac{e^{-i \theta} -1}{\lambda} XX^{*} \in U(m),
$$
which depends on $X,$ not only on $n$ and $m,$
where $\theta= 2 \cdot \tfrac{n+m}{2n} A(\gamma).$
\end{remark}
\bibliographystyle{amsplain}
|
\section{Introduction}
\subsection{The main Theorem}
Let $M$ be a connected $d_{M}$-dimensional, compact, complex projective manifold and $(A,h)$ be an ample positive Hermitian line bundle on $M$.
We may assume that the curvature form of the unique compatible connection $\nabla_{A}$ is $\Theta=-2i\omega$, where $\omega$ is a
K\"{a}hler form. Let $dV_{M}$ be the volume form $\frac{\omega^{\wedge d_{M}}}{d_{M}!}$ associated with $(M,\omega)$.
We put $h=g-i\omega$ where $g$ is the induced Riemannian structure. Suppose given two connected compact Lie groups $G$ and $T$, with $T$ a torus, of dimension $d_{G}$ and $d_{T}$, respectively, and commuting holomorphic and Hamiltonian actions $\mu^{G}:G\times M\rightarrow M$ and $\mu^{T}:T\times M \rightarrow M$.
Thus $\mu_{g}^{G}\circ\mu_{t}^{T}=\mu_{t}^{T}\circ\mu_{g}^{G}$, $\forall (g,t)\in G\times T$ with moment maps $\Phi_{G},\Phi_{T}$.
Assume that both actions unitarily linearize to $A$, that is, that they admit metric preserving lifting $\tilde{\mu}^{G},\tilde{\mu}^{T}$.
Let $\widehat{G}$ be the collection of irreducible characters of $G$ and for any
$\nu_{G}\in\widehat{G}$ let $V_{\nu_{G}}$ be the corresponding irreducible unitary representation.
The action of $G$ on $A$ dualizes to an action on the dual line bundle $A^{\vee}$ and the $G$-invariant Hermitian metric $h$ on $A$ naturally induces an Hermitian metric on $A^{\vee}$ also denoted by $h$.
Let $X\subseteq A^{\vee}$ be the unit circle bundle, with projection $\pi:X\rightarrow M$. Then $X$ is a contact manifold,
with contact form given by the connection 1-form $\alpha$.
Since $G$ and $T$ preserve the Hermitian metric $h$ on $A^{\vee}$, they act on $X$. Furthermore,
as both linearized actions preserve the unique compatible connection
$\nabla_{A}$, both actions leave $\alpha$ invariant.
The actions of $G$ and $T$ on $X$ preserve the volume form $dV_{X}=\alpha\wedge \pi^{*}\big(\frac{dV_{M}}{2\pi}\big)$ on $X$, whence they induce commuting unitary representations of $G$ and $T$ on $L^{2}(X)$, which preserve the Hardy space $H(X)=L^2(X)\cap \mathop{}\!\mathrm{Ker}{(\overline{\partial}_{b})}$.
By virtue of the Peter-Weyl Theorem, we may then unitarily and equivariantly decompose $H(X)$ over the irreducible representations of $G$ and $T$, respectively.
For every $\nu_{G}\in \widehat{G}$ we define $H(X)_{\nu_{G}}^{G}\subseteq H(X)$ be the maximal subspace equivariantly isomorphic to a direct sum of copies of $V_{\nu_{G}}$. In the same way we define $H(X)_{\nu_{T}}^{T}$. So decomposing the Hardy space of $X$ unitarily and equivariantly over the irreducible representations of $T$ and $G$, we have:
\begin{equation}
\label{pacbell}
H(X)=\bigoplus_{\nu_{T}\in\mathbb{Z}^{d_{T}}}H(X)_{\nu_{T}}^{T}=\bigoplus_{\nu_{G}\in\widehat{G}}H(X)_{\nu_{G}}^{G}.
\end{equation}
Similarly, under the previous assumptions there is an holomorphic Hamiltonian action of the product $P=G\times T$, and a corresponding unitary representation, so that we also have:
\begin{equation}
\label{kayak3}
H(X)=\bigoplus_{\nu_{G}\in\widehat{G},\nu_{T}\in\mathbb{Z}^{d_{T}}}H(X)_{\nu_{G},\nu_{T}}^{G\times T},
\end{equation}
\noindent
where $H(X)_{\nu_{G},\nu_{T}}^{G\times T}=H(X)_{\nu_{G}}^{G}\cap H(X)_{\nu_{T}}^{T}$.
Under the assumption that $\mathbf{0}\not\in\Phi_{T}$, we have that $\mathop{}\!\mathrm{dim}(H(X)_{\nu_{T}}^{T})<+\infty$ for each $\nu_{T}\in\mathbb{Z}^{d_{T}}$.
\begin{definition}
Given a pair of irreducible weights $\nu_{G}$ and $\nu_{T}$ for $G$ and $T$, respectively, we shall denote by $\widetilde{\Pi}_{\nu_{G},\nu_{T}}:L^{2}(X)\rightarrow H(X)_{\nu_{G},\nu_{T}}$ the orthogonal projector, and refer to its Schwartz kernel as the level $(\nu_{G},\nu_{T})$-Szeg\"{o} projector of $X$ (with the two actions understood).
In terms of an orthonormal basis $\left\{s_{j}^{(\nu_{G},\nu_{T})}\right\}_{j=1}^{N_{\nu_{G},\nu_{T}}}$ of $H(X)_{\nu_{G},\nu_{T}}$, it is given by:
\begin{equation}
\label{schwartz}
\widetilde{\Pi}_{\nu_{G},\nu_{T}}(x,y)=\sum_{j}\widehat{s}_{j}^{(\nu_{G},\nu_{T})}(x)\overline{\widehat{s}_{j}^{(\nu_{G},\nu_{T})}(y)}.
\end{equation}
\end{definition}
In this paper we shall consider the local asymptotics of the equivariant Szeg\"{o} kernels $\widetilde{\Pi}_{\nu_{G},k\nu_{T}}$, where the irreducible representation of $T$
tends to infinity along a ray, and the irreducible representation of $G$ is held fixed. To this end, we shall use a combination of the techniques in
$\cite{primo}$ and $\cite{quarto}$.\\
\begin{observation}
The smooth function $x\mapsto \widetilde{\Pi}_{\nu_{G},k\nu_{T}}(x,x)$ descends to a smooth function on $M$.
\end{observation}
\begin{remark}
We shall use the following notation for the various equivariant Szeg\"{o} kernels coming into play:
\begin{itemize}
\item[1] $\widetilde{\Pi}_{\nu_{G},k\nu_{T}}$ in the general case, under the action of $G\times T$;
\item[2] $\widetilde{\Pi}_{k\nu_{T}}$ in the case of $G$ trivial;
\item[3] $\Pi_{k}$ in the case of $G$ trivial and $T=S^{1}$ with $\Phi_{T}=1$;
\item[4] $\Pi_{\nu_{G},k}$ in the case of $T=S^{1}$ and $\Phi_{T}=1$;
\item[5] $\widetilde{\Pi}_{\nu_{G},k}$ and $\widetilde{\Pi}_{k}$ also for the case $T=S^{1}$ and not necessarily $\Phi_{T}=1$.
\end{itemize}
\end{remark}
\noindent
A key tool used in the proofs are the Heisenberg local coordinates centered at $x\in X$ (see $\cite{terzo}$).
We denote this system of coordinates by:
$$ \gamma_{x}:(\theta,\mathop{}\!\mathrm{v})\in (-\pi,+\pi)\times B_{2d_{M}}(\mathbf{0},\delta)\mapsto x+(\theta,\mathop{}\!\mathrm{v})\in X, $$
\noindent
here $B_{2d_{M}}(\mathbf{0},\delta)$ is the open ball of $\mathbb{R}^{2d_{M}}$ centered at the origin with radius $\delta>0$.
We have that $\theta$ is an angular coordinate along the circle fiber and $\mathop{}\!\mathrm{v}\in T_{m}M$ a local coordinate on $M$. We shall also set $x+\mathop{}\!\mathrm{v}=x+(0,\mathop{}\!\mathrm{v})$.
Given the choice of HLC centered at $x\in X$, there are induced unitary isomorphisms $T_{x}X\cong \mathbb{R}\oplus \mathbb{R}^{2d_{M}}$ and $T_{m}M \cong \mathbb{R}^{2d_{M}} \cong \mathbb{C}^{d_{M}}$.
Therefore, each equivariant Szeg\"{o} kernel $\widetilde{\Pi}_{\nu_{G},\nu_{T}}$ is a smoothing operator, with $\mathcal{C}^{\infty}$ Schwartz kernel given by $(\ref{schwartz})$.
We shall make the following three transversality assumptions on the moment maps:
\begin{itemize}
\item[1] $\mathbf{0}\not\in \Phi_{T}(M)$ and $\Phi_{T}$ is transversal to the ray $\mathbb{R}_{+}\cdot\nu_{T}$, so that $M_{\nu_{T}}^{T}=\Phi_{T}^{-1}(\mathbb{R}_{+}\cdot\nu_{T})\subseteq M$ is a compact, $G\times T$-invariant and connected submanifold of dimension $2d_{M}+1-d_{T}$. This is equivalent to requiring that the action of $T$ on $X$ be locally free on the inverse of $M^{T}_{\nu_{T}}$ (see $\cite{quarto}$);
\item[2] $\mathbf{0}\in\mathfrak{g}^{\vee}$ is a regular value of $\Phi_{G}$, so that $M_{0}^{G}=\Phi_{G}^{-1}(\mathbf{0})\subseteq M$ is a compact, $G\times T$-invariant and connected submanifold of dimension $2d_{M}-d_{G}$;
\item[3] the two submanifolds $M_{\nu_{T}}^{T}$ and $M_{0}^{G}$ are mutually transversal.
\end{itemize}
These conditions imply the following (which is what we shall really be using). Since the two actions commute, they give rise to an action of
the product group $P=G\times T$, which is also holomorphic and Hamiltonian, with moment map:
$$ \Phi_{P}=(\Phi_{G},\Phi_{T}):M\rightarrow \mathfrak{g}^{\vee}\oplus\mathfrak{t}^{\vee}\cong \mathfrak{p}; $$
\noindent
then $\mathbf{0}\not\in \Phi_{P}(M)$, and $\Phi_{P}$ is transversal to the ray $\mathbb{R}_{+}\cdot \big(\mathbf{0},\nu_{T}\big)$. In particular,
$$ M_{0,\nu_{T}}=\Phi^{-1}_{P}\left(\mathbb{R}_{+}\cdot(\mathbf{0},\nu_{T})\right) $$
is a smooth connected submanifold of $M$ with codimension $d_{G}+d_{T}-1$.
Leaving aside that here $G$ is not required to be a torus, these hypothesis are similar in nature to the hypothesis in $\cite{quarto}$, applied however to $P$ rather than $T$. Unlike $\cite{quarto}$, where the local scaling asymptotics for representations along a ray $k\nu_{T}$ are considered, we shall study the local scaling asymptotics of doubly equivariant pieces of $\Pi$ associated to pair of representations $(\nu_{G},k\nu_{T})$, where only one of the representations drifts to infinity, while the other is held fixed. Let us also remark that when $T=S^{1}$ and $\Phi_{T}=1$, we are reduced to considering the isotypical components of the spaces of holomorphic global sections $H^{0}(M,A^{\otimes k})$ under the action of $G$, as in $\cite{primo}$ and $\cite{quindicesimo}$.
We will find asymptotic expansions that generalize and combine the previous cases.
Let $N_{m}$ we denote the normal bundle to $\Phi_{P}^{-1}\big(\mathbb{R}_{+}(\mathbf{0},\nu_{T})\big)$ then $N_{m}$ is naturally isomorphic to $J_{m}\big(\mathop{}\!\mathrm{Ker}{(\Phi_{P}(m))}\big)$ (see $\cite{quarto}$ section 2.2).
The transversality condition is equivalent to require the injectivity of the evaluation map (see as before $\cite{quarto}$). Now for every $m\in M_{0,\nu_{T}}$ we have two Euclidean structures on
$$\mathop{}\!\mathrm{Ker}{(\Phi_{P}(m))}=\mathfrak{g}\times\mathop{}\!\mathrm{Ker}{(\Phi_{T}(m))}\subseteq\mathfrak{g}\oplus\mathfrak{t},$$
\noindent
one induced from $\mathfrak{g}\oplus\mathfrak{t}$ and the second from $T_{m}M$.
Let $D(m)$ the matrix representing the latter Euclidean product on $N_{m}$, with respect to an orthonormal basis.
Then $D(m)$ is indipendent of the choice of an orthonormal basis for $\mathfrak{g}\times\mathop{}\!\mathrm{Ker}{(\Phi_{T}(m))}$,
and it determines a positive smooth function on $M_{0,\nu_{T}}$. As in $\cite{quarto}$ we have the following definition.
\begin{definition}
Define $\mathcal{D}\in\mathcal{C}^{\infty}(M_{0,\nu_{T}})$, with $m\in M_{0,\nu_{T}}$, by setting
$$ \mathcal{D}(m)=\sqrt{\det{D(m)}}.$$
\end{definition}
We consider $M_{0,\nu_{T}}$ and for the decomposition of the tangent space $T_{m}M$ we have that:
\begin{equation}
\label{labb}
T_{m}M=H_{m}\oplus V_{m}\oplus N_{m}
\end{equation}
\noindent
where, given $J_{m}:T_{m}M\rightarrow T_{m}M$ the complex structure, we have that:
\begin{equation}
\label{antoniosalierii}
V_{m}=\mathfrak{g}_{M}(m)\oplus\mathop{}\!\mathrm{val}{(\mathop{}\!\mathrm{Ker}{(\Phi_{T}(m))})}, \ \ N_{m}=J_{m}(V_{m}), \ \ H_{m}=[V_{m}\oplus N_{m}]^{\perp},
\end{equation}
\noindent
are respectively the vertical, the transversal and the horizontal part. Given $m\in M_{0,\nu_{T}}$ and $\mathop{}\!\mathrm{v}\in T_{m}M$ we can decompose $\mathop{}\!\mathrm{v}$ uniquely as $\mathop{}\!\mathrm{v}=\mathop{}\!\mathrm{v}_{h}+\mathop{}\!\mathrm{v}_{v}+\mathop{}\!\mathrm{v}_{t}$ with $\mathop{}\!\mathrm{v}_{v}\in V_{m}$, $\mathop{}\!\mathrm{v}_{t}\in N_{m}$ and $\mathop{}\!\mathrm{v}_{h}\in H_{m}$.
The scaling asymptotics of the equivariant Szeg\"{o} kernels, that we will see later, are controlled by a quadratic exponent in the components $\mathop{}\!\mathrm{v}_h,\mathop{}\!\mathrm{v}_v,\mathop{}\!\mathrm{v}_t$ of a tangent vector at a given $m\in M_{0,\nu_T}$ (viewed as a small displacement from $m$).
\begin{definition}
Let $x\in X$ and $v_{l}=(\theta_{l},\mathop{}\!\mathrm{v}_{l})\in T_{x}X$ with $l=1,2$. We define $H: TX\oplus TX\rightarrow \mathbb{C}$ as
\begin{equation*}
\begin{multlined}[t][12.5cm]
H(v_{1},v_{2})=\\
\lambda_{\nu_{T}}\left(i[\omega_{m}\left(\mathop{}\!\mathrm{v}_{1v},\mathop{}\!\mathrm{v}_{1t}\right)-\omega_{m}\left(\mathop{}\!\mathrm{v}_{2v},\mathop{}\!\mathrm{v}_{2t}\right)]+i\omega_{m}\left(\frac{(\theta_{2}-\theta_{1})}{\|\Phi_{T}(m)\|}\eta_{Mh}(m),\mathop{}\!\mathrm{v}_{1h}+\mathop{}\!\mathrm{v}_{2h}\right)-\right.\\
\left.-i\omega_{m}\left(\mathop{}\!\mathrm{v}_{1h},\mathop{}\!\mathrm{v}_{2h}\right)-\|\mathop{}\!\mathrm{v}_{1t}\|^2-\|\mathop{}\!\mathrm{v}_{2t}\|^2-\frac{1}{2}\left\|\mathop{}\!\mathrm{v}_{1h}-\frac{(\theta_{2}-\theta_{1})}{\|\Phi_{T}(m)\|}\eta_{Mh}(m)-\mathop{}\!\mathrm{v}_{2h}\right\|^2\right)
\end{multlined}
\end{equation*}
\noindent
with $\eta_{Mh}(m)$ the unitary generator of $\mathop{}\!\mathrm{Ker}{(\Phi_{P}(m))}^{\perp}$ such that $\langle\eta,\Phi_{P}(m)\rangle=\|\Phi_{T}(m)\|$ and $\lambda_{\nu_{T}}=\frac{\|\nu_{T}\|}{\|\Phi_{T}(m)\|}$.
\end{definition}
\begin{theorem}[main Theorem]
\label{teo:secondo}
Under the previous assumptions fix $\nu_{G}\in\widehat{G}$ and consider
$\nu_{T}\in\mathbb{Z}^{d_{T}}$, assume that $\Phi_{P}$ is transversal to the ray $\mathbb{R}_{+}\cdot(\mathbf{0},\nu_{T})$. We have:
\begin{itemize}
\item[$1)$] If $C,\delta>0$, and
$$\max{\{\mathop{}\!\mathrm{dist}_{M}(\pi(x),M_{0,\nu_{T}}),\mathop{}\!\mathrm{dist}_{M}(\pi(y),M_{0,\nu_{T}})\}}\geq Ck^{\delta-\frac{1}{2}},$$
\noindent
then $\widetilde{\Pi}_{\nu_{G},k\nu_{T}}=O(k^{-\infty})$.
\item[$2)$] Uniformly in $x\in X_{0,\nu_{T}}$ and $v_{l}\in T_{x}X$ with $\|\mathop{}\!\mathrm{v}_{l}\|\leq Ck^{\frac{1}{9}}$, as $k\rightarrow +\infty$ we have:
\begin{equation*}
\begin{multlined}[t][12.5cm]
\widetilde{\Pi}_{\nu_{G},k\nu_{T}}\left(x+\frac{\mathop{}\!\mathrm{v}_{1}}{\sqrt{k}},x+\frac{\mathop{}\!\mathrm{v}_{2}}{\sqrt{k}}\right)\sim \frac{1}{(\sqrt{2}\pi)^{d_{T}-1}}d_{\nu_{G}}2^{\frac{d_{G}}{2}}\cdot \\
\cdot\left(\frac{k}{\pi}\|\nu_{T}\|\right)^{d_{M}-\frac{d_{P}}{2}+\frac{1}{2}}\left(\sum_{j=1}^{N_{x}}\chi_{\nu_{G}}(g_{j}^{-1})e^{-ik\vartheta_{j}\nu_{T}}e^{H(v_{1}^{j},v_{2})}\right)\cdot \frac{e^{-i\sqrt{k}(\theta_{2}-\theta_{1})\lambda_{\nu_{T}}}}{\mathcal{D}(m)}\cdot \\
\cdot\frac{1}{\|\Phi_{T}\|^{d_{M}+1-\frac{d_{P}}{2}+\frac{1}{2}}}\left(1+\sum_{l\geq 1}R_{\nu_{G},l}(m,v_{1}^{j},v_{2})k^{-\frac{l}{2}}\right)
\end{multlined}
\end{equation*}
where $d_{P}=d_{G}+d_{T}$, $v_{1}^{j}$, $\mathop{}\!\mathrm{v}^{j}$ denote the monodromy representation
$F_{x} \rightarrow GL(T_{m}M)$, such that for every $j=1,\cdots, N_{x}$, $\mathop{}\!\mathrm{v}\in T_{m}M$ we have $p_{j} \mapsto d_{m}\mu^{P}_{p_{j}}(\mathop{}\!\mathrm{v})=\mathop{}\!\mathrm{v}^{(j)}\in T_{m}M$. Where $F_{x}$ is the stabilizator of $P$ in $x$ and $R_{\nu_{G},l}$ are polynomials in $v_{1}^{j},v_{2}$ with coefficients depending on $x$,$\nu_{G}$ and $\nu_{T}$.
\item[$3)$]More in general, for every $p_{0}\in P$, denoting $P\cdot x$ the orbit of $x\in X_{0,\nu_{T}}$, then the following expansion holds for $k\rightarrow +\infty$:
\begin{equation}
\label{generalizationcase2}
\begin{multlined}[t][12.5cm]
\widetilde{\Pi}_{\nu_{G},k\nu_{T}}\left(x+\frac{u_{1}}{\sqrt{k}},p_{0}\cdot\left(x+\frac{u_{2}}{\sqrt{k}}\right)\right)\sim \frac{1}{(\sqrt{2}\pi)^{d_{T}-1}}d_{\nu_{G}}2^{\frac{d_{G}}{2}}\cdot\\
\cdot\left(\frac{k}{\pi}\|\nu_{T}\|\right)^{d_{M}-\frac{d_{P}}{2}+\frac{1}{2}}\sum_{j=1}^{N_{x}}\overline{\chi_{\nu_{P}}(p_{j}p^{-1}_{0})}e^{H(\mathop{}\!\mathrm{v}^{j}_{1},\mathop{}\!\mathrm{v}_{2})}\cdot \frac{e^{-i\sqrt{k}(\theta_{2}-\theta_{1})\lambda_{\nu_{T}}}}{\mathcal{D}(m)}\cdot \\
\cdot\frac{1}{\|\Phi_{T}\|^{d_{M}+1-\frac{d_{P}}{2}+\frac{1}{2}}}\left(1+\sum_{l\geq 1}R_{\nu_{G},l}(m,\mathop{}\!\mathrm{v}^{j}_{1},\mathop{}\!\mathrm{v}_{2})k^{-\frac{l}{2}}\right),
\end{multlined}
\end{equation}
\noindent
where $p_{j}\in F_{x}$ and $u_{j}=(\theta_{j},\mathop{}\!\mathrm{v}_{j})$ for $j=1,2$.
\end{itemize}
\end{theorem}
The previous result describes the asymptotics of $\widetilde{\Pi}_{\nu_{G},k\nu_{T}}$ in a shrinking neighborhood of the orbit $P\cdot x$, where $x\in X_{0,\nu_{T}}$. It is complemented by the following:
\begin{proposition}
\label{prop:complement}
Suppose $x\in X_{0,\nu_{T}}$ and $\varepsilon,D>0$. Then uniformly for $\mathop{}\!\mathrm{dist}_{X}(y,P\cdot x)\geq Dk^{\varepsilon-1/2}$ we have
$$\widetilde{\Pi}_{\nu_{G},k\nu_{T}}(x,y)=O(k^{-\infty}).$$
\end{proposition}
\subsection{Special cases and relation to prior work}
Before continuing our exposition, it is in order to digress on the relation of our results to prior work in this area. Let us focus on the following two special cases:
\begin{itemize}
\item[a)] $T=S^{1}$ acts in the standard manner (with $\Phi_{T}=1$);
\item[b)] $G$ is trivial.
\end{itemize}
Let us first consider the case a), and to fix ideas let us start with the case where $G$ is trivial.
Let $\rho(\cdot,\cdot)$ be a system of Heisenberg local coordinates for $X$ centered at $x$.
We have for $X$ centered at $x$, inducing a unitary isomorphism $(T_{m}M,h_{m})$ with $\mathbb{C}^{d_{M}}$ the standard Hermitian structure.
In Theorem 3.1 of $\cite{terzo}$ and in $\cite{dodicesimo}$, for $\mathop{}\!\mathrm{v}_{1},\mathop{}\!\mathrm{v}_{2}\in B(\mathbf{0},1)\subseteq \mathbb{C}^{d_{M}}$, $\theta\in (-\pi,\pi)$ and $k\rightarrow +\infty$ the following expansion has been determined for the level $k$ of Szeg\"{o} kernel $\Pi_{k}$ (see also $\cite{tredicesimo}$):
\begin{equation}
\label{standardcaseTYZ55}
\begin{multlined}[t][12.5cm]
\Pi_{k}\left(\rho\left(\theta,\frac{\mathop{}\!\mathrm{v}_{1}}{\sqrt{k}}\right),\rho\left(\theta',\frac{\mathop{}\!\mathrm{v}_{2}}{\sqrt{k}}\right)\right)\sim\left(\frac{k}{\pi}\right)^{d_{M}}\cdot e^{ik(\theta-\theta')+\psi_{2}(\mathop{}\!\mathrm{v}_{1},\mathop{}\!\mathrm{v}_{2})}\cdot\\ \cdot\left(1+\sum_{j\geq 1}a_{j}(x,\mathop{}\!\mathrm{v}_{1},\mathop{}\!\mathrm{v}_{2})k^{-\frac{j}{2}}\right)
\end{multlined}
\end{equation}
\noindent
where
$$\psi_{2}(\mathop{}\!\mathrm{v}_{1},\mathop{}\!\mathrm{v}_{2})=\mathop{}\!\mathrm{v}_{1}\cdot\overline{\mathop{}\!\mathrm{v}_{2}}-\frac{1}{2}(\|\mathop{}\!\mathrm{v}_{1}\|^2+\|\mathop{}\!\mathrm{v}_{2}\|^2)$$
\noindent
and $a_{j}$ are polynomials in $\mathop{}\!\mathrm{v}_{1}$ and $\mathop{}\!\mathrm{v}_{2}$.
\begin{observation}
Another way to write $\psi_{2}$ is:
$$ \psi_{2}(\mathop{}\!\mathrm{v}_{1},\mathop{}\!\mathrm{v}_{2})=-i\omega_{m}\left(\mathop{}\!\mathrm{v}_{1},\mathop{}\!\mathrm{v}_{2}\right)-\frac{1}{2}\|\mathop{}\!\mathrm{v}_{1}-\mathop{}\!\mathrm{v}_{2}\|^{2},$$
\noindent
in this form we can see directly the real and the imaginary part of $\psi_{2}$ observing that it is responsible to the exponential decay near the diagonal.
\end{observation}
Now let us consider the Hamiltonian action of a compact Lie group $G$ on $M$ and suppose that $\mathbf{0}\in\mathfrak{g}^{\vee}$ is a regular value of the moment map. Then $\Pi_{\nu_{G},k}(x,x)$ is rapidly decreasing away from $\Phi_{G}^{-1}(\mathbf{0})$, and assuming $\Phi_{G}(\pi(x))=\mathbf{0}$, under the standard action of $S^{1}$ the following asymptotic expansion holds with $m=\pi(x)$:
\begin{equation}
\label{groooup}
\begin{multlined}[t][12.5cm]
\Pi_{\nu_{G},k}\left(x+\frac{v_{1}}{\sqrt{k}},x+\frac{v_{2}}{\sqrt{k}}\right) \\
\sim \left(\frac{k}{\pi}\right)^{d_{M}-\frac{d_{G}}{2}}e^{\left[Q\left(v_{1v}+v_{1t},v_{2v}+v_{2t}\right)\right]}\sum_{g\in G_{m}}e^{\psi_{2}\left(v_{1gh},v_{2h}\right)}\cdot A_{\nu_{G},k}(g,x)\cdot\\
\cdot\left(1+\sum_{j\geq 1}R_{\nu_{G},j}(m,v_{1g},v_{2})k^{-\frac{j}{2}}\right)
\end{multlined}
\end{equation}
\noindent
where $Q\left(v_{1v}+v_{1t},v_{2v}+v_{2t}\right)=-\|v_{2t}\|^2-\|v_{1t}\|^2+i[\omega_{m}(v_{1v},v_{1t})-\omega_{m}(v_{2v},v_{2t})]$,
$G_{m}=\{g\in G:\mu_{g}(m)=m\}$, $R_{\nu_{G},j}$ are polynomials in $v_{1},v_{2}$ and
$$A_{\nu_{G},k}(g,x)=2^{\frac{d_{G}}{2}}\frac{\mathop{}\!\mathrm{dim}(V_{\nu_{G}})}{V_{\mathop{}\!\mathrm{eff}}{(\pi(x))}}\frac{1}{|G_{\pi(x)}|}\chi_{\nu_{G}}(g)h_{g}^{k},$$
\noindent
where $V_{\mathop{}\!\mathrm{eff}}{(\pi(x))}$ is the volume of the fiber above $m$ in $\Phi^{-1}_{G}(\mathbf{0})$ (for more details on the effective potentials see $\cite{venticinquesimo}$) and here we have set
$$ v_{1g}=d_{m}\mu^{G}_{g_{j}^{-1}}(v_{1}) $$
\noindent
with $g_{j}$ in the stabilizator of $G$. Obviously $(\ref{groooup})$ reduces to $(\ref{standardcaseTYZ55})$ for trivial $G$.
Let us consider case b). Thus assume that there is a holomorphic Hamiltonian action of a compact torus $T$, and that the moment map determining the linearization is nowhere zero. To fix ideas, let us first consider the case where $T$ is one-dimensional. If $\xi_{M}$ and $\xi_{X}$ are vector fields on $M$ and on $X$ induced by $\mu^{T}$ and $\widetilde{\mu}^{T}$, we have that in the Heisenberg local coordinates $\xi_{X}(x)=(-\Phi_{T}(m),\xi_{M}(m))$ with $m=\pi(x)$. Let $\xi_{X}(x)^{\perp}\subseteq T_{x}X$ be the orthocomplement of $\xi_{X}(x)$. In view of Theorem 1 of $\cite{quarto}$, again working in a system of HLC centered at $x$ and that $v_{l}=(\theta_{l},\mathop{}\!\mathrm{v}_{l})\in T_{x}X\cong \mathbb{R}\times T_{m}M$ satisfying $\mathop{}\!\mathrm{v}_{l}\in\xi_{X}(x)^{\perp}$, $\|\mathop{}\!\mathrm{v}_{l}\|\leq Ck^{1/9}$, as $k\rightarrow +\infty$ we have:
\begin{equation}
\label{tooooruscase}
\begin{multlined}[t][12.5cm]
\widetilde{\Pi}_{k}\left(x+\frac{\mathop{}\!\mathrm{v}_{1}}{\sqrt{k}},x+\frac{\mathop{}\!\mathrm{v}_{2}}{\sqrt{k}}\right)\\
\sim \left(\frac{k}{\pi}\right)^{d_{M}}\Phi_{T}(m)^{-(d_{M}+1)}e^{i\sqrt{k}\frac{(\theta_{1}-\theta_{2})}{\Phi_{T}(m)}}\cdot\left(\sum_{t\in T_{m}}t^{k}e^{E(d_{x}\tilde{\mu}^{T}_{t^{-1}}(\mathop{}\!\mathrm{v}_{1}),\mathop{}\!\mathrm{v}_{2})}\right)\cdot\\
\cdot\left(1+\sum_{j\geq 1}R_{j}(m,v_{1},v_{2})k^{-\frac{j}{2}}\right)
\end{multlined}
\end{equation}
\noindent
for certain smooth functions $R_{j}$, polynomial in the $v_{l}$'s, with
\begin{equation*}
\begin{multlined}[t][12.5cm]
E(v_{1},v_{2})=\frac{1}{\Phi_{T}(m)}\left\{i\left[ \frac{(\theta_{2}-\theta_{1})}{\Phi_{T}(m)}\omega_{m}(\xi_{M}(m),\mathop{}\!\mathrm{v}_{1}+\mathop{}\!\mathrm{v}_{2})-\omega_{m}(\mathop{}\!\mathrm{v}_{1},\mathop{}\!\mathrm{v}_{2})\right]-\right.\\
\left.-\frac{1}{2}\left\|\mathop{}\!\mathrm{v}_{1}-\mathop{}\!\mathrm{v}_{2}-\frac{(\theta_{2}-\theta_{1})}{\Phi_{T}(m)}\xi_{M}(m)\right\|^2\right\}.
\end{multlined}
\end{equation*}
This last result can be generalize to a $d_{T}$-dimensional torus as in Theorem 2 of $\cite{quarto}$. In this Theorem
we have a result similar to the previous but with the appearance of an additional important invariant, which plays a role
analogous to the effective potential in $(\ref{standardcaseTYZ55})$.
Suppose that $\Phi_{T}$ is transversal to the ray $\mathbb{R}_{+}\cdot \nu_{T}$. Then the normal space to the inverse image $N_{m}$ at any $m\in M_{\nu_{T}}=\Phi_{T}^{-1}(\mathbb{R}_{+}\cdot \nu_{T})$ is $N_{m}\cong J_{m}\left(\mathop{}\!\mathrm{Ker}{(\Phi_{T}(m))}\right)$ and the evaluation map $\mathop{}\!\mathrm{val}:\mathop{}\!\mathrm{Ker}{(\Phi_{T}(m))}\rightarrow T_{m}M$ is injective. Therefore, we have on $\mathop{}\!\mathrm{Ker}{(\Phi_{T}(m))}$ two Euclidean products, and given two orthonormal basis $\mathcal{B}_{1},\mathcal{B}_{2}$ we can consider the matrix $D(m)$ whose determinant is independent of the choice of the basis. Thus we can let $\mathcal{D}(m)=\sqrt{\det{D(m)}}$.
Considering $\nu_{T}\in \mathbb{Z}^{d_{T}}$, as $k\rightarrow +\infty$ we have:
\begin{equation}
\label{tooooruscase2}
\begin{multlined}[t][12.5cm]
\widetilde{\Pi}_{k\nu_{T}}\left(x+\frac{\mathop{}\!\mathrm{v}_{1}}{\sqrt{k}},x+\frac{\mathop{}\!\mathrm{v}_{2}}{\sqrt{k}}\right)\\
\sim \left(\frac{1}{(\sqrt{2}\pi)^{d_{T}-1}}\right)\left(\|\nu_{T}\|\frac{k}{\pi}\right)^{d_{M}+\frac{1-d_{T}}{2}}\frac{1}{(\|\Phi_{T}\|)^{d_{M}+1+\frac{1-d_{T}}{2}}\mathcal{D}(m)}\\
e^{i\sqrt{k}\frac{(\theta_{1}-\theta_{2})}{\Phi_{T}(m)}}\cdot\left(\sum_{t\in T_{m}}t^{k}e^{H_{m}(d_{x}\tilde{\mu}^{T}_{t^{-1}}(v_{1}),v_{2})}\right)\cdot\\
\cdot\left(1+\sum_{j\geq 1}R_{j}(m,v_{1},v_{2})k^{-\frac{j}{2}}\right)
\end{multlined}
\end{equation}
\noindent
with
\begin{equation*}
H_{m}(v_{1},v_{2})=\frac{\|\nu_{T}\|}{\|\Phi_{T}\|}\left[-i\omega_{m}(\mathop{}\!\mathrm{v}_{1},\mathop{}\!\mathrm{v}_{2})-\|\mathop{}\!\mathrm{v}_{1}\|^2-\|\mathop{}\!\mathrm{v}_{2}\|^2\right].
\end{equation*}
In this paper, we shall pair these situations. More precisely, we shall assume given actions of $G$ and $T$ as above, compatible in the sense that they commute, and consider the resulting asymptotics relative to a pair $(\nu_{G}, k\nu_{T})$ of irreducible characters, where $\nu_{G}$ is held fixed, and $k\nu_{T}\rightarrow +\infty$ along an integral ray.
We consider the case of a $d_{T}$-dimensional torus. We have shown at the beginning the fundamental result of this work. Now we present some observations.
\begin{observation}
If $G$ is trivial, $d_{G}=0$ the leading term is:
\begin{equation*}
\begin{split}
\frac{e^{-ik\vartheta_{j}\nu_{T}}}{(\sqrt{2}\pi)^{d_{T}-1}}\left(\frac{\|\nu_{T}\|k}{\pi}\right)^{d_{M}-\frac{d_{T}-1}{2}}\cdot\frac{1}{\|\Phi_{T}(m)\|^{d_{M}+1-\frac{d_{T}}{2}+\frac{1}{2}}\mathcal{D}(m)}e^{\lambda_{\nu_{T}}\left[-i\omega_{m}(\mathop{}\!\mathrm{v}_{1}^{j},\mathop{}\!\mathrm{v}_{2})-\frac{1}{2}\|\mathop{}\!\mathrm{v}_{1}^{j}-\mathop{}\!\mathrm{v}_{2}\|^2\right]}
\end{split}
\end{equation*}
\noindent
and we're back to the equation $(\ref{tooooruscase2})$ when $\theta_{1}=\theta_{2}=0$.
\end{observation}
\begin{observation}
In the case $T=S^{1}$ with the standard action we have $\lambda_{\nu_{T}}=1$,
and when $\theta_{1}=\theta_{2}=0$ the result is the formula $(\ref{groooup})$.
\end{observation}
Now we present a Theorem that is the diagonal version, without scaling of the point $2)$ of the main Theorem.
\begin{theorem}
\label{teo:diagonal}
Under the hypothesis of the main Theorem, for $m=\pi(x)\in M_{0,\nu_{T}}$, as $k\rightarrow +\infty$ we have:
\begin{equation}
\label{principal44}
\begin{multlined}[t][12.5cm]
\widetilde{\Pi}_{\nu_{G},k\nu_{T}}(x,x)\\
\sim\frac{d_{\nu_{G}}2^{d_{G}/2}}{\left(\sqrt{2}\pi\right)^{d_{T}-1}}\cdot\left(\frac{\|\nu_{T}\|k}{\pi}\right)^{d_{M}+\frac{1-d_{P}}{2}}\cdot\sum_{j=1}^{N_{x}}\chi_{\nu_{G}}(g_{j}^{-1})e^{-ik\vartheta_{j}\nu_{T}}\\
\cdot\frac{1}{\mathcal{D}(m)\|\Phi_{T}(m)\|^{d_{M}+1+\frac{1-d_{P}}{2}}}\cdot\left(1+\sum_{l\geq 1}B_{l}(m)k^{-l}\right)
\end{multlined}
\end{equation}
\noindent
with $B_{l}$ that are smooth functions on $M_{0,\nu_{T}}$.
\end{theorem}
\begin{corollary}
\label{cor:quarto}
Under the assumptions of Theorem $~\ref{teo:secondo}$,
\begin{equation*}
\begin{split}
\lim_{k\rightarrow +\infty}\left(\frac{\pi}{\|\nu_{T}\|k}\right)^{d_{M}-d_{P}+1}&\mathop{}\!\mathrm{dim}{\left(H(X)_{\nu_{G},k\nu_{T}}\right)}=\\
=& \frac{d_{\nu_{G}}^{2}}{(2\pi)^{d_{T}-1}}\cdot\int_{M_{0,\nu_{T}}}\frac{\|\Phi_{T}(m)\|^{-(d_{M}+1)+d_{P}-1}}{\mathcal{D}(m)}dV_{M}(m).
\end{split}
\end{equation*}
\end{corollary}
As a very special example, we observe that when $d_{T}=1$ and $T^{1}=S^1$ acts trivially on $M$ with moment map $\Phi_{T}=1$, we have $H(X)_{k}$ the $k$-th isotypical component for the standard $S^1$-action on $X$, which is naturally and unitarily isomorphic to $H^{0}(M,A^{\otimes k})$.
In this case we have the celebrated Tian-Yau-Zelditch expansion. For this result we refer to the work of Zelditch in $\cite{diciannovesimo}$ (see also $\cite{ventitreesimo}$, $\cite{ventiquattresimo}$, $\cite{trentunesimo}$, $\cite{trentaseiesimo}$ and, for its near diagonal rescaled generalizations see $\cite{dodicesimo}$, $\cite{
terzo}$, $\cite{tredicesimo}$ and $\cite{quattordicesimo}$).
\subsection{Applications to Toeplitz operator kernels}
By way of application, motivated by the standard Berezin-Toeplitz quantization of a classical observable (see $\cite{trentesimo}$, $\cite{diciottesimo}$, $\cite{trentaquattresimo}$, $\cite{ventottesimo}$, $\cite{ventinovesimo}$, $\cite{trentatreesimo}$, $\cite{trentottesimo}$, $\cite{trentacinquesimo}$, $\cite{trentaduesimo}$ and $\cite{trentasettesimo}$), let us consider the scaling asymptotics of the equivariant components of certain Toeplitz operators (we will consider Toeplitz operators in the sense of $\cite{undicesimo}$). Given $f\in\mathcal{C}^{\infty}(M)$ and assuming for simplicity that $f$ is invariant under the action of the product group $P=G\times T$, we can consider the Toeplitz operators $T_{\nu_{G},k\nu_{T}}[f]=\widetilde{\Pi}_{\nu_{G},k\nu_{T}}\circ M_{f}\circ \widetilde{\Pi}_{\nu_{G},k\nu_{T}}$, where $M_{f}$ denotes multiplication by $f\circ \pi$. Then $T_{\nu_{G},k\nu_{T}}[f]$ is a self-adjoint endomorphisms of $H(X)_{\nu_{G},k\nu_{T}}$.
Given that $\mathbf{0}\not\in\Phi_{P}$, the equivariant Toeplitz operator $T_{\nu_{G},k\nu_{T}}[f]$ is smoothing, and its distributional kernel is given by the following two alternative expressions:
\begin{equation}
\label{toeplitz}
\begin{multlined}[t][12.5cm]
T_{\nu_{G},k\nu_{T}}[f](x,x')=\int_{X}\widetilde{\Pi}_{\nu_{G},k\nu_{T}}(x,y)f(y)\widetilde{\Pi}_{\nu_{G},k\nu_{T}}(y,x')dV_{X}(y)\\
=\sum_{j}T_{\nu_{G},k\nu_{T}}[f](s_{j}^{k}(x))\overline{(s_{j}^{k}(x'))}
\end{multlined}
\end{equation}
\noindent
with $x,x'\in X_{0,\nu_{T}}$ and $s_{j}^{k}$ an orthonormal basis of $H(X)_{\nu_{G},k\nu_{T}}$.
We will see that $T_{\nu_{G},k\nu_{T}}[f](x,x')$ has asymptotic expansions near the diagonal similar to the one for $\widetilde{\Pi}_{\nu_{G},k\nu_{T}}$.
Note that with $f(y)$ we denote $f(\pi(y))$ and that every $f\in \mathcal{C}^{\infty}(M)$ lifts to an invariant function $f(x)$ on $X$.
For the sake of simplicity, we shall focus on points of the form $(x+n,x+n)$ (with rescaling), as usual, in a system of Heisenberg local coordinates centered at $x$, where $n$ is a normal vector to the $P$-orbit of $x$ and we shall make the extra assumption that the stabilizer of $x$ in $P$ is trivial. Notice that any point sufficiently close to $P\cdot x$ may be written in this manner, possibly replacing $x$ with $p\cdot x$ for some $p\in P$.
\begin{theorem}
\label{teo:app2}
Assume that $\mathbf{0}\not\in \Phi_{P}$, $f\in\mathcal{C}^{\infty}(M_{0,\nu_{T}})$ is $\mu^{P}$-invariant and that the stabilizer of $P$ in $x$ is trivial. Suppose $x\in X_{0,\nu_{T}}$ and fix a system of HLC centered at $x$. Let $m=\pi(x)$. Then we have:
\begin{itemize}
\item[$1)$]If $C,\delta>0$ and
$$\max{\{\mathop{}\!\mathrm{dist}_{M}(\pi(x),M_{0,\nu_{T}}),\mathop{}\!\mathrm{dist}_{M}(\pi(y),M_{0,\nu_{T}})\}}\geq Ck^{\delta-\frac{1}{2}},$$
\noindent
then $T_{\nu_{G},k\nu_{T}}[f](x,x')=O(k^{-\infty})$.
\item[$2)$]Uniformly in $n_{1}\in N_{x}^{P}=T_{x}(P\cdot x)^{\perp}$ as $k\rightarrow +\infty$:
\begin{equation}
\label{Toeplitz7}
\begin{multlined}[t][12.5cm]
T_{\nu_{G},k\nu_{T}}[f]\left(x+\frac{n_{1}}{\sqrt{k}},x+\frac{n_{1}}{\sqrt{k}}\right)\\
\sim \frac{1}{(\sqrt{2}\pi)^{d_{T}-1}}d_{\nu_{G}}2^{\frac{d_{G}}{2}}\left(\frac{k}{\pi}\|\nu_{T}\|\right)^{d_{M}-\frac{d_{P}}{2}+\frac{1}{2}}f(m)e^{-2\lambda_{\nu_{T}}\|\mathop{}\!\mathrm{t}_{1}\|^2}\cdot\\
\cdot \frac{1}{\mathcal{D}(m)}\cdot\frac{1}{\|\Phi_{T}\|^{d_{M}+1-\frac{d_{P}}{2}+\frac{1}{2}}}\left(\sum_{l\geq 0}k^{-\frac{l}{2}}R_{l}(n_{1},m)\right)
\end{multlined}
\end{equation}
\noindent
with $R_{l}(n_{1},m)$ a polynomial in $n_{1}$ and $\mathop{}\!\mathrm{t}_{1}\in N_{m}=J_{m}(\mathop{}\!\mathrm{val}_{m}(\mathop{}\!\mathrm{Ker}{(\Phi_{P}(m))}))$.
\end{itemize}
\end{theorem}
\begin{corollary}
\label{cor:quinto}
Under the assumptions of Theorem $~\ref{teo:app2}$,
\begin{equation*}
\begin{split}
\lim_{k\rightarrow +\infty}\left(\frac{\pi}{\|\nu_{T}\|k}\right)^{d_{M}-d_{P}+1}&\mathfrak{T}\left(T_{\nu_{G},k\nu_{T}}[f]\right)=\\
=& \frac{d_{\nu_{G}}^{2}}{(2\pi)^{d_{T}-1}}\cdot\int_{X_{0,\nu_{T}}}\frac{f(\pi(x))\|\Phi_{T}(\pi(x))\|^{-(d_{M}+2-d_{P})}}{\mathcal{D}(\pi(x))}dV_{X}(x),
\end{split}
\end{equation*}
\noindent
where $\mathfrak{T}\left(T_{\nu_{G},k\nu_{T}}[f]\right)$ is the trace of the Toeplitz operator.
\end{corollary}
\section{Examples}
The main Theorem predicts that the diagonal restriction $\widetilde{\Pi}_{\nu_{G},k\nu_{T}}(x,x)$ of the equivariant Szeg\"{o} kernel (which descends to a function on $M$) is rapidly decreasing away from the locus $M_{0,\nu_{T}}$, and grows like $k^{d_{M}+\frac{1-d_{P}}{2}}$ there. Let us illustrate this explicitly by two examples (cfr the computations in $\cite{primo}$). Recall from $\cite{dodicesimo}$ that for $k=1,2,\cdots$ an orthonormal basis of
$H^{0}(\mathbb{P}^{n},\mathcal{O}_{\mathbb{P}^{n}}(k))$ is $\{s_{J}^{k}\}_{|J|=k}$, where:
\begin{equation}
s_{J}^{k}=\sqrt{\frac{(k+n)!}{\pi^{n}J!}}z^{J}
\end{equation}
\noindent
and where $J!=\prod_{l=0}^{n}j_{l}!$, $z^{J}=\prod_{l=0}^{n}z_{l}^{j_{l}}$.
In the next example we consider a particular product action and we show that outside of $M_{0,\nu_{T}}$ we have the exponential decay of the Szeg\"{o} kernel.
\begin{example}
Let us make $M=\mathbb{P}^{1}$. Let us consider the action of $G=T^{1}$ on $M$ induced by the representation on $\mathbb{C}^{2}$ given by $\mu^{G}(z_{0},z_{1})=w\cdot(z_{0},z_{1})=(w^{-1}z_{0},wz_{1})$, and the action of $T=T^{1}$ induced by the representation given by $\mu^{T}(z_{0},z_{1})=(s^{-1}z_{0},s^{-2}z_{1})$. These actions are holomorphic and Hamiltonian, with moment maps:
$$\Phi_{G}(z_{0},z_{1})=\frac{|z_{0}|^2-|z_{1}|^2}{|z_{0}|^2+|z_{1}|^2}$$
\noindent
and
$$\Phi_{T}(z_{0},z_{1})=\frac{|z_{0}|^2+2|z_{1}|^2}{|z_{0}|^2+|z_{1}|^2}.$$
Then we have:
$$\Phi_{G}^{-1}(0)=\{[z_{0}:z_{1}]:|z_{0}|=|z_{1}|\}$$
\noindent
and placing $X=S^{3}\subseteq \mathbb{C}^{2}$ we have
$$X_{0}=\pi^{-1}\left(\Phi_{G}^{-1}(0)\right)=\left\{(z_{0},z_{1}):|z_{0}|=|z_{1}|=\frac{1}{\sqrt{2}}\right\}\cong S^{1}\times S^{1}$$
\noindent
with a free action of $S^{1}$ on $X_{0}$.
We have $\nu_{T}=1\in\mathbb{Z}$, $\Phi_{P}^{-1}(\mathbb{R}_{+}\cdot(0,1))=\Phi_{G}^{-1}(0)=\{(z_{0},z_{1}):|z_{0}|=|z_{1}|\}$ and the action of $P$ is given by:
$$\mu^{P}([z_{0}:z_{1}])=(w,s)\cdot (z_{0},z_{1})=\left((ws)^{-1}z_{0},ws^{-2}z_{1}\right).$$
If $|z_{0}|=|z_{1}|$ ($\not=0$) and $(w,s)\cdot (z_{0},z_{1})=(z_{0},z_{1})\Rightarrow ws=1, ws^{-2}=1 \Rightarrow s=s^{-2}$ so $s=e^{\frac{2}{3}\pi ji}$ with $j=0,1,2$ and $w=\frac{1}{s}$ then the action is locally free. We are in the hypothesis of the main Theorem.
We have $s\cdot(z_{0}^{a}z_{1}^{b})=(sz_{0})^{a}(s^{2}z_{1})^{b}=s^{a+2b}z_{0}^{a}z_{1}^{b}$ and then
$$\widetilde{H}^{T}(X)_{k}=\mathop{}\!\mathrm{span}{\{z_{0}^{a}z_{1}^{b}: a+2b=k\}}.$$
In the other side we have $w\cdot(z_{0}^{a}z_{1}^{b})=(wz_{0})^{a}(w^{-1}z_{1})^{b}=w^{a-b}z_{0}^{a}z_{1}^{b}$ and then
$$\widetilde{H}^{G}(X)_{\nu_{G}}=\mathop{}\!\mathrm{span}{\{z_{0}^{a}z_{1}^{b}: a=b+\nu_{G}\}}.$$
Thus
$$ \widetilde{H}^{P}(X)_{\nu_{G},k}=\mathop{}\!\mathrm{span}{\left\{z_{0}^{a}z_{1}^{b}: a=b+\nu_{G},a+2b=k\right\}}$$
\noindent
then $a+2b=k \Rightarrow b+\nu_{G}+2b=k \Rightarrow 3b=k-\nu_{G}$ and
$$ \mathop{}\!\mathrm{dim}{\left(\widetilde{H}^{P}(X)_{\nu_{G},k}\right)}=\begin{cases} 0 &\mbox{if} \ \ k\equiv\nu_{G} \mod{3}\\
1 & \mbox{if} \ \ k\not\equiv \nu_{G}\mod{3} \end{cases}. $$
If $k=\nu_{G}+3b$ we have:
$$ \widetilde{H}^{P}_{\nu_{G},\nu_{G}+3b}(X)=\mathop{}\!\mathrm{span}{\left\{z_{0}^{b+\nu_{G}}z_{1}^{b}\right\}}, $$
\noindent
the corresponding Szeg\"{o} projector is:
\begin{equation}
\label{loopspaces}
\begin{multlined}[t][12.5cm]
\widetilde{\Pi}^{P}_{\nu_{G},\nu_{G}+3b}\left((z_{0},z_{1}),(u_{0},u_{1})\right)=\frac{(2b+\nu_{G}+1)!}{\pi(b+\nu_{G})!b!}(z_{0}\overline{u}_{0})^{b+\nu_{G}}(z_{1}\overline{u}_{1})^{b}.
\end{multlined}
\end{equation}
Now consider $z_{j}=u_{j}$ with $|z_{0}|^2+|z_{1}|^2=1$ and we set $x=|z_{0}|^{2}$, $y=y_{b}=\frac{b+\nu_{G}}{2b+\nu_{G}}\rightarrow \frac{1}{2}$ as $b\rightarrow +\infty$.
Using Stirling approximation:
$$ n!\sim \sqrt{2\pi n}\frac{n^{n}}{e^{n}}$$
\noindent
and the projector:
$$
\widetilde{\Pi}^{P}_{\nu_{G},\nu_{G}+3b}\left((z_{0},z_{1}),(z_{0},z_{1})\right)=\frac{(2b+\nu_{G}+1)!}{\pi(b+\nu_{G})!b!}|z_{0}|^{2(b+\nu_{G})}|z_{1}|^{2b}
$$
\noindent
we can find the following asymptotic for the coefficient:
\begin{equation}
\label{loopspaces2}
\begin{multlined}[t][12.5cm]
\frac{(2b+\nu_{G}+1)!}{\pi(b+\nu_{G})!b!}\sim \frac{2}{\sqrt{\pi}}\sqrt{b}\left(\frac{1}{y_{b}}\right)^{b+\nu_{G}}\left(\frac{1}{1-y_{b}}\right)^{b}
\end{multlined}
\end{equation}
\noindent
and for the projector:
\begin{equation}
\label{loopspaces3}
\begin{multlined}[t][12.5cm]
\widetilde{\Pi}^{P}_{\nu_{G},\nu_{G}+3b}\left((z_{0},z_{1}),(z_{0},z_{1})\right)\sim 2\sqrt{\frac{b}{\pi}}\left(\frac{x}{y}\right)^{\nu_{G}}e^{bF(x,y)},
\end{multlined}
\end{equation}
\noindent
where we set $F(x,y)=\log{x}+\log{(1-x)}-\log{y}-\log{(1-y)}=f(x)-f(y)$ with $f(t)=\log{t}+\log{(1-t)}$ and $0<t<1$.
We observe that for $t\rightarrow 0^{+},1^{-}$ we obtain $f(t)\rightarrow -\infty$ and that the derivative:
$$ f'(t)=\frac{1}{t}-\frac{1}{1-t}=0 \Leftrightarrow 1-t= t \Leftrightarrow t=\frac{1}{2} $$
\noindent
with $f(1/2)=-\log{4}$. Thus for $b\gg 0$ and $y=y_{b}$ we have $f(y)=-\log{4}-\delta(b)$ with $\delta(b)>0$ and $\delta(b)\rightarrow 0$ as $b\rightarrow +\infty$. If $x \not=\frac{1}{2}$ we have $f(x)=-\log{4}-\delta(x)$ ( with $\delta(x)>0$ fixed).
Then $F(x,y_{b})=-\delta(x)+\delta(b)\leq -\frac{\delta(x)}{2}$.
Now
\begin{equation}
\label{loopspaces44}
\begin{multlined}[t][12.5cm]
\left|\widetilde{\Pi}^{P}_{\nu_{G},\nu_{G}+3b}\left((z_{0},z_{1}),(z_{0},z_{1})\right)\right|\sim 2\sqrt{\frac{b}{\pi}}\left(\frac{x}{y}\right)^{\nu_{G}}e^{bF(x,y)}\\
\leq 2\sqrt{\frac{b}{\pi}}\left(\frac{x}{y}\right)^{\nu_{G}}e^{-b\frac{\delta(x)}{2}}=O(b^{-\infty})
\end{multlined}
\end{equation}
\noindent
but $x=\frac{1}{2}$ that is $|z_{0}|=|z_{1}|=\frac{1}{\sqrt{2}}$ and we have:
\begin{equation}
\label{loopspaces55}
\begin{multlined}[t][12.5cm]
\left|\widetilde{\Pi}^{P}_{\nu_{G},\nu_{G}+3b}\left((z_{0},z_{1}),(z_{0},z_{1})\right)\right|\sim 2\sqrt{\frac{b}{\pi}}\left(\frac{x}{y}\right)^{\nu_{G}}e^{bF(1/2,y_{b})}\\
\leq 2\sqrt{\frac{b}{\pi}}\left(\frac{1}{2y_{b}}\right)^{\nu_{G}}e^{b\delta(b)}
\end{multlined}
\end{equation}
\noindent
and considering that we have for $y_{b}$:
\begin{equation}
\label{loopspaces56}
\begin{multlined}[t][12.5cm]
y_{b}=\frac{b+\nu_{G}}{2b+\nu_{G}}=\frac{1+\frac{\nu_{G}}{b}}{2\left(1+\frac{\nu_{G}}{2b}\right)}=\frac{1}{2}+\frac{\nu_{G}}{4b}+O\left(\frac{1}{b^{2}}\right),
\end{multlined}
\end{equation}
\noindent
then $f(y_{b})=-\log{4}+O\left(\frac{1}{b^{2}}\right)$ (because $f'(1/2)=0$) and so follows that $b\delta(b)=O\left(\frac{1}{b}\right)\rightarrow 0$ as $b\rightarrow +\infty$. Thus
\begin{equation}
\label{loopspaces5566}
\begin{multlined}[t][12.5cm]
\left|\widetilde{\Pi}^{P}_{\nu_{G},\nu_{G}+3b}\left((z_{0},z_{1}),(z_{0},z_{1})\right)\right|\sim 2\sqrt{\frac{b}{\pi}}.
\end{multlined}
\end{equation}
\end{example}
Another possible variation similar to the previous is the following.
\begin{example}
Let us make $M=\mathbb{P}^{2}$. Let us consider the actions of $G=T^{2}$ on $M$ induced by the representation on $\mathbb{C}^{3}$ given by $\mu^{G}(z_{0},z_{1},z_{2})=(w_{1}^{-1}z_{0},w_{1}w_{2}^{-1}z_{1},w_{2}z_{2})$, and the action of $T=T^{1}$ induced by the representation given by $\mu^{T}(z_{0},z_{1},z_{2})=(s^{-1}z_{0},s^{-2}z_{1},s^{-3}z_{2})$. These actions are holomorphic and Hamiltonian, with moment maps:
$$\Phi_{G}(z_{0},z_{1},z_{2})=\left(\frac{|z_{0}|^2-|z_{1}|^2}{|z_{0}|^2+|z_{1}|^2},\frac{|z_{1}|^2-|z_{2}|^2}{|z_{1}|^2+|z_{2}|^2}\right)$$
\noindent
and
$$\Phi_{T}(z_{0},z_{1},z_{2})=\frac{|z_{0}|^2+2|z_{1}|^2+3|z_{2}|^{2}}{|z_{0}|^2+|z_{1}|^2+|z_{2}|^{2}}.$$
Then
$$\Phi_{G}^{-1}(\mathbf{0})=\{[z_{0}:z_{1}:z_{2}]:|z_{0}|=|z_{1}|=|z_{2}|\}$$
\noindent
and placing $X=S^{5}\subseteq \mathbb{C}^{3}$ we have:
$$X_{0}=\pi^{-1}\left(\Phi_{G}^{-1}(\mathbf{0})\right)=\left\{(z_{0},z_{1},z_{2}):|z_{0}|=|z_{1}|=|z_{2}|=\frac{1}{\sqrt{3}}\right\}\cong S^{1}\times S^{1}\times S^{1}$$
\noindent
with a free action of $G$ on $X_{0}$.
We have $\nu_{T}=1\in\mathbb{Z}$ and
$\Phi_{P}^{-1}(\mathbb{R}_{+}\cdot(\mathbf{0},1))=\Phi_{G}^{-1}(\mathbf{0})=\{(z_{0},z_{1},z_{2}):|z_{0}|=|z_{1}|=|z_{2}|\}$ and the action of $P$ is given by:
$$\mu^{P}([z_{0}:z_{1}:z_{2}])=(w,s)\cdot (z_{0},z_{1},z_{2})=\left((w_{1}s)^{-1}z_{0},w_{1}w_{2}^{-1}s^{-2}z_{1},w_{2}s^{-3}z_{2}\right).$$
If $|z_{0}|=|z_{1}|=|z_{2}|$ ($\not=0$) and $(w,s)\cdot (z_{0},z_{1},z_{2})=(z_{0},z_{1},z_{2})\Rightarrow w_{1}s=1, w_{1}w_{2}^{-1}s^{-2}=1, w_{2}s^{-3}=1 \Rightarrow s^{6}=1$ so $s=e^{\frac{2}{6}\pi ji}$ with $j=0,1,2,4,5$ and $w_{1}=\frac{1}{s},w_{2}=\frac{1}{s^{3}}$ then the action is locally free. The hypothesis of the main Theorem are satisfied.
We have $s\cdot(z_{0}^{a}z_{1}^{b}z_{2}^{c})=(sz_{0})^{a}(s^{2}z_{1})^{b}(s^{3}z_{2})^{c}=s^{a+2b+3c}z_{0}^{a}z_{1}^{b}z_{2}^{c}$ and then
$$\widetilde{H}^{T}(X)_{k}=\mathop{}\!\mathrm{span}{\{z_{0}^{a}z_{1}^{b}z_{2}^{c}: a+2b+3c=k\}}.$$
In analogue way we obtain that
$$\widetilde{H}^{G}(X)_{\nu_{G}}=\mathop{}\!\mathrm{span}{\{z_{0}^{a}z_{1}^{b}z_{2}^{c}: (a-b,b-c)=(\nu_{1},\nu_{2})=\nu_{G}\}}.$$
Thus
$$\widetilde{H}^{P}(X)_{\nu_{G},k}=\mathop{}\!\mathrm{span}{\left\{z_{0}^{a}z_{1}^{b}z_{2}^{c}: (a-b,b-c)=(\nu_{1},\nu_{2}),a+2b+3c=k\right\}}$$
\noindent
and
$$\mathop{}\!\mathrm{dim}{\left(\widetilde{H}^{P}(X)_{\nu_{G},k}\right)}=\begin{cases} 0 &\mbox{if } k\equiv\nu_{1} \mod{6}\\
1 & \mbox{if } k\not\equiv \nu_{1}\mod{6} \end{cases}.$$
If $k=6c+\nu_{1}+3\nu_{2}$ we have:
$$\widetilde{H}^{P}(X)_{(\nu_{1},\nu_{2}),\nu_{1}+3\nu_{2}+6c}=\mathop{}\!\mathrm{span}{\left\{z_{0}^{c+\nu_{1}+\nu_{2}}z_{1}^{c+\nu_{2}}z_{2}^{c}\right\}}.$$
Now consider $z_{j}=u_{j}$ with $|z_{0}|^2+|z_{1}|^2+|z_{2}|^2=1$ and we set $x=|z_{0}|^{2}$, $y=|z_{1}|^2, \ \ z=|z_{2}|^2=1-x-y$.
As before, using Stirling approximation for the projector:
\begin{equation}
\label{loopspaces88}
\begin{multlined}[t][12.5cm]
\widetilde{\Pi}^{P}_{\nu_{G},6c+\nu_{1}+3\nu_{2}}\left((z_{0},z_{1},z_{2}),(z_{0},z_{1},z_{2})\right)\\
=\frac{(3c+\nu_{1}+2\nu_{2}+2)!}{\pi^2(c+\nu_{2}+\nu_{1})!(c+\nu_{2})!c!}(|z_{0}|)^{2(c+\nu_{2}+\nu_{1})}(|z_{1}|)^{2(c+\nu_{2})}(|z_{2}|)^{2c},
\end{multlined}
\end{equation}
\noindent
the following asymptotic holds:
\begin{equation}
\label{loopspaces2acca}
\begin{multlined}[t][12.5cm]
\frac{(3c+\nu_{1}+2\nu_{2}+2)!}{\pi^2(c+\nu_{2}+\nu_{1})!(c+\nu_{2})!c!}\\
\sim \frac{9\sqrt{3}c}{2\pi^3}\left[\frac{3c+\nu_{1}+2\nu_{2}}{c+\nu_{2}+\nu_{1}}\right]^{c+\nu_{2}+\nu_{1}}\left[\frac{3c+\nu_{1}+2\nu_{2}}{c+\nu_{2}}\right]^{c+\nu_{2}}\left[\frac{3c+\nu_{1}+2\nu_{2}}{c}\right]^c
\end{multlined}
\end{equation}
\noindent
and, for the projector:
\begin{equation}
\label{loopspaces3acca}
\begin{multlined}[t][12.5cm]
\widetilde{\Pi}^{P}_{\nu_{G},6c+\nu_{1}+3\nu_{2}}\left((z_{0},z_{1},z_{2}),(z_{0},z_{1},z_{2})\right)\sim \frac{9\sqrt{3}c}{2\pi^3}x^{\nu_{1}+\nu_{2}}y^{\nu_{2}}e^{cF(x,y)},
\end{multlined}
\end{equation}
\noindent
with $F(x,y)= \log{x}+\log{y}+\log{(1-x-y)}$, $0<x,y$ and $x+y<1$.
Studing the partial derivatives we find that we have a critical point $x=y=\frac{1}{3}$ with $F\left(\frac{1}{3},\frac{1}{3}\right)=-\log{27}$.
So we have $F\left(x,y\right)=-\log{27}-\delta(z)$ with $\delta(z)>0$. If $x=y\not=\frac{1}{3}$ we have:
\begin{equation}
\label{loopspaces3141}
\begin{multlined}[t][12.5cm]
\left|\widetilde{\Pi}^{P}_{\nu_{G},6c+\nu_{1}+3\nu_{2}}\left((z_{0},z_{1},z_{2}),(z_{0},z_{1},z_{2})\right)\right|\\
\leq \frac{9\sqrt{3}c}{2\pi^3}x^{\nu_{1}+\nu_{2}}y^{\nu_{2}}e^{-c\delta(z)}=O(c^{-\infty})
\end{multlined}
\end{equation}
\noindent
but $x=y=\frac{1}{3}$, $\delta(z)=0$ and
\begin{equation}
\label{loopspaces3141}
\begin{multlined}[t][12.5cm]
\left|\widetilde{\Pi}^{P}_{\nu_{G},6c+\nu_{1}+3\nu_{2}}\left((z_{0},z_{1},z_{2}),(z_{0},z_{1},z_{2})\right)\right|
\sim\frac{9\sqrt{3}c}{2\pi^3}\left(\frac{1}{3}\right)^{\nu_{1}+\nu_{2}}\left(\frac{1}{3}\right)^{\nu_{2}}.
\end{multlined}
\end{equation}
\end{example}
\section{Preliminaries}
It is know that if $G$ and $T$ both act on a symplectic manifold $M$ in an Hamiltonian fashion with moment maps $\Phi_{G}$ and $\Phi_{T}$ and these actions commute,
then $P=G\times T$ act on $M$ and the moment map is $\Phi_{P}=\Phi_{G}\oplus\Phi_{T}:M\rightarrow \mathfrak{g}^{\vee}\oplus\mathfrak{t}^{\vee}$.
We give an explicit expression for $H(X)_{\nu_{G},\nu_{T}}$ as
$$H(X)_{\nu_{G},\nu_{T}}=\left\{s\in H(X)_{\nu_{G}}: s(\tilde{\mu}^{T}_{t^{-1}}(x))=t^{\nu_{T}}s(x), \forall x\in X,\forall t\in T^{d_{T}},\forall \nu_{G}\in\widehat{G}\right\}.$$
\subsection{The geometric setting}
We remember that the matrix $D(m)$ represents the Euclidean product on $N_{m}$ with an orthonormal basis.
It determines a positive smooth function $\mathcal{D}$ on $M_{0,\nu_{T}}$ defined above.
\begin{remark}
Note that there is a relation between the $D(m)$ matrix and the $C(m)$ matrix used in Theorem $\ref{teo:secondo}$. We have in fact that $D(m)=C(m)^{t}\cdot C(m)$ and so
$\mathcal{D}(m)=\sqrt{\det{D(m)}}=|\det{C(m)}|$.
\end{remark}
Let us consider the symplectic cone $\Sigma \subseteq TX^{\vee}\setminus\{0\}$ sprayed by the connection form $\alpha$:
$$ \Sigma=\{ (x,r\alpha_{x}): x\in X, r>0\}.$$
This cone is important for the microlocal description of Szeg\"{o} kernel (as in $\cite{secondo}$) and in the theory of Toeplitz operators (see $\cite{undicesimo}$).
We have that the wave front set of
$\Pi$ is the anti-diagonal:
$$ \Sigma^{\#}=\{(x,r\alpha_{x},x,-r\alpha_{x}): x\in X , r>0\}. $$
Notice that $\Sigma \cong X\times \mathbb{R}_{+}$ in a natural manner.
Let $\omega_{\Sigma}$ the restriction to $\Sigma$ of the symplectic structure on $TX^{\vee}$. Let $r$ be the cone coordinate on $\Sigma$ and
$\theta$ be the circle coordinate on $X$, locally defined, and pulled-back to $\Sigma$. Then $\omega_{\Sigma}=d\lambda=dr\wedge \alpha +2r\omega$,
with $\lambda=r\alpha$. Let $\widetilde{\xi}_{f}$ be the contact lift to $X$ of the Hamiltonian vector field $\xi_{f}$ on $(M,2\omega)$. Then the cotangent flow restricted to $\Sigma$ is generated by $(\widetilde{\xi}_{f},0)$. Thus the cotangent flow on $\Sigma\cong X\times \mathbb{R}$ is $\phi^{\Sigma}_\tau=\phi^{X}_\tau\times id_\mathbb{R}$. It is follows that if $f$ and $g$ Poisson commute on $M$, then their flows on $M$, $X$ and $\Sigma$ also commute, and conversely.
About the product action (referring to $\cite{quarto}$), since by assumption $\Phi_{P}$ is transverse to $\mathbb{R}_{+}\cdot (\mathbf{0},\nu_{T})$, the action of $P$ on $X_{0,\nu_{T}}$ is locally free. We have also that $X_{0,\nu_{T}}$ is invariant for $G\times T$. In fact we have that $\mathbb{R}_{+}\cdot(\mathbf{0},\nu_{T})$ is invariant for the coadjoint action and, with the fact that $\Phi_{P}$ is equivariant, we conclude.
\subsection{Adapted Heisenberg local coordinates}
A key tool used in the proofs are the Heisenberg local coordinates centered at $x\in X$ defined in $\cite{terzo}$.
We choose an adapted holomorphic coordinate system $(z_{1},\cdots , z_{d_{M}})$ for $M$ centered at $\pi(x)$ so $\omega$
espressed in $z_{i}$'s at $\pi(x)$ is the standard symplectic structure on $\mathbb{C}^{d_{M}}$ that is
$\omega\left(\pi(x)\right)=\frac{i}{2}\sum_{j=1}^{d_{M}}dz_{j}\wedge d\overline{z}_{j}$. By the choice of the adapted holomorphic coordinate
system we have the unitary isomorphism $T_{\pi(x)}M\cong\mathbb{C}^{d_{M}}$. Now we choose a preferred local frame $\sigma_{A}$ for $A$ at $\pi(x)$, in the sense of
$\cite{terzo}$.
The problem now is to find the espression of Heisenberg coordinates under the action of $P=G\times T$.
The expression of $\tilde{\mu}_{g^{-1}}^{G}\circ\tilde{\mu}_{-\frac{\vartheta}{\sqrt{k}}}^{T}\left(x_{1k}\right)$ in Heisenberg coordinates, where
$x_{1k}=x+\frac{\mathop{}\!\mathrm{v}_{1}}{\sqrt{k}}$, $g=\exp_{G}\frac{\varsigma}{\sqrt{k}}$
is an element of the group $G$, $e^{i\vartheta}$ an element of the torus $T$ and $\exp_{G}$ is the exponential map, are given by the following lemma:
\begin{lemma}
\label{lemma:hei}
Suppose $x\in X$, $\Phi_{G}\circ\pi(x)=\mathbf{0}$, and fix a system of Heisenberg local coordinates centered at $x$. Suppose in addition that
$F_{x}=\{(g_{j},t_{j}):j=1,\cdots, N_{x}\}$ is not trivial.
Then there exist $\mathcal{C}^{\infty}$ functions $\widetilde{B}_{3},\widetilde{B}_{2}:\mathbb{R}^{d_{T}}\times\mathbb{C}^{d_{M}}\times \mathbb{R}^{d_{G}} \rightarrow \mathbb{C}^{d_{M}}$,
vanishing at the origin to third and second order, respectively, such that the following holds.
For $\mathop{}\!\mathrm{v}_{1}-\vartheta\xi_{M}(m)\in T_{\pi(x)}M$, where $\vartheta=(\vartheta_{1},\cdots,\vartheta_{d_{T}})$ and $\xi_{M}(m)=\mathop{}\!\mathrm{val}_{m}{\xi}$
with $\xi\in \mathfrak{t}$ and $\varsigma_{M}(m)$ the valutation of $\varsigma\in\mathfrak{g}$. As $k \rightarrow +\infty$ the Heisenberg local coordinates of:
$$ \tilde{\mu}^{G}_{e^{-\frac{\varsigma}{\sqrt{k}}}}\circ\tilde{\mu}^{T}_{-\frac{\vartheta}{\sqrt{k}}}\left(\tilde{\mu}^{G}_{g_{j}^{-1}}\circ \tilde{\mu}^{T}_{t_{j}^{-1}}(x_{1k})\right) $$
\noindent
are given by
{ \footnotesize
\begin{equation}
\label{autoscuolavisconti}
\begin{multlined}[t][12.5cm]
\left(\frac{1}{\sqrt{k}}\left(\vartheta\Phi_{T}(m)+\theta_{1}\right)+\frac{1}{k}\omega_{m}\left(\vartheta\xi_{M}(m),\mathop{}\!\mathrm{v}_{1}^{j}\right)+\frac{1}{k}\omega_{m}\left(\varsigma_{M}(m),\mathop{}\!\mathrm{v}_{1}^{j}\right)+\widetilde{B}_{3}\left(\frac{\vartheta}{\sqrt{k}},\frac{\mathop{}\!\mathrm{v}}{\sqrt{k}},\frac{\varsigma}{\sqrt{k}}\right),\right.\\
,\left.\frac{1}{\sqrt{k}}\left(\mathop{}\!\mathrm{v}_{1}^{j}-\vartheta\xi_{M}(m)-\varsigma_{M}(m)\right)+\widetilde{B}_{2}\left(\frac{\vartheta}{\sqrt{k}},\frac{\mathop{}\!\mathrm{v}}{\sqrt{k}},\frac{\varsigma}{\sqrt{k}}\right)\right).
\end{multlined}
\end{equation}
}\end{lemma}
\emph{Proof.} We apply corollary 2.2 of $\cite{quarto}$ and we have $(\ref{autoscuolavisconti})$.
\hfill $\Box$\\
\begin{remark}
Note that here $\vartheta=(\vartheta_{1},\cdots,\vartheta_{d_{T}})$ with $-\pi<\vartheta_{i}<\pi$ and $\xi=\left.\frac{\partial}{\partial \vartheta}\right|_{0}$. We have:
$$ \Phi_{l}=\langle\Phi,\xi_{l}\rangle, \ \ \vartheta\cdot\Phi=\sum_{l=1}^{d_{T}}\vartheta_{l}\Phi_{l},
\ \ \vartheta\cdot\xi=\sum_{l=1}^{d_{T}}\vartheta_{l}\xi_{l}.$$
\end{remark}
\section{Proof of the main Theorem $\bf{\ref{teo:secondo}}$}
$\mathop{}\!\mathit{Proof}.$
Proof of $1)$. We consider $(\rho_{\nu_{G}},V_{\nu_{G}})$ an unitary irreducible representation of $G$ and
we define $\rho_{\nu_{G},k\nu_{T}}:G\times T\rightarrow GL(V_{\nu_{G}})$ as $\rho_{\nu_{G},k\nu_{T}}(g,t)=t^{k\nu_{T}}\rho_{\nu_{G}}(g)$.
We have that $(\rho_{\nu_{G},k\nu_{T}},V_{\nu_{G}})$ is an unitary irreducible representation of $G\times T$ with character
$\chi_{\nu_{G},k\nu_{T}}(g,t)=\mathop{}\!\mathrm{Tr}{(\rho_{\nu_{G},k\nu_{T}}(g,t))}=t^{k\nu_{T}}\mathop{}\!\mathrm{Tr}{(\rho_{\nu_{G}}(g))}=t^{k\nu_{T}}\chi_{\nu_{G}}(g)$.
Assuming that $\mathbf{0}\not\in \Phi_{T}(M)$ we have that $H(X)_{k\nu_{T}}$ is finite dimensional, then $H(X)_{\nu_{G},k\nu_{T}}\subseteq H(X)_{k\nu_{T}}$
and $\widetilde{\Pi}_{\nu_{G},k\nu_{T}}\in\mathcal{C}^{\infty}(X\times X)$. We want study the asymptotic behavior of $\widetilde{\Pi}_{\nu_{G},k\nu_{T}}$ with $k\rightarrow +\infty$.
Since $\widetilde{\Pi}_{\nu_{G},k\nu_{T}}$ is the the composition of $\Pi:L^2(X)\rightarrow H(X)$ and the orthogonal projector of $H(X)$ onto:
\begin{equation}
\label{riemann}
\widetilde{\Pi}_{\nu_{G},k\nu_{T}}\left(x,y\right)=\frac{d_{\nu_{G}}}{(2\pi)^{d_{T}}}\int_{G}\int_{T}\chi_{\nu_{G}}(g^{-1})t^{-k\nu_{T}}\Pi\left(\tilde{\mu}^{G}_{g^{-1}}\circ\tilde{\mu}^{T}_{t^{-1}}\left(x\right),y\right)dtdg,
\end{equation}
\noindent
where $d_{\nu_{G}}=\mathop{}\!\mathrm{dim}{(V_{\nu_{G}})}$ and $dg$, $dt$ are the associated measure for $G$ and
$T$ such that $\int_{G}dg=1$ and $\int_{T}dt=1$. We start considering the diagonal case, so we have:
\begin{equation}
\label{scented}
\begin{multlined}[t][12.5cm]
\widetilde{\Pi}_{\nu_{G},k\nu_{T}}\left(x,x\right)=d_{\nu_{G}}\int_{G}\int_{T}\chi_{\nu_{G}}(g^{-1})t^{-k}\Pi\left(\tilde{\mu}^{G}_{g^{-1}}\circ\tilde{\mu}^{T}_{t^{-1}}\left(x\right),x\right)dt dg\\
=\frac{d_{\nu_{G}}}{2\pi}\int_{G}\int_{(-\pi,+\pi)^{d_{T}}}\chi_{\nu_{G}}(g^{-1})e^{-ik\nu_{T}\cdot\vartheta}\Pi\left(\tilde{\mu}^{G}_{g^{-1}}\circ\tilde{\mu}^{T}_{-\vartheta}\left(x\right),x\right)d\vartheta dg,
\end{multlined}
\end{equation}
\noindent
where $\vartheta\in(-\pi,\pi)^{d_{T}}$. For the moment suppose $x\in X$ generic and fixed, and denote $F_{x}\subseteq G\times T$ the stabilizer of $x$.
For $\varepsilon>0$ we set
$$ A=\{(g,t)\in G\times T: \mathop{}\!\mathrm{dist}_{G\times T}{((g,t),F_{x})}<2\varepsilon\} $$
\noindent
and
$$ B=\{(g,t)\in G\times T: \mathop{}\!\mathrm{dist}_{G\times T}{((g,t),F_{x})}>\varepsilon\} $$
\noindent
so we have $G\times T=A\cup B$ and we can consider a partition of the unity $\gamma_{1}+\gamma_{2}=1$ associated to the covering $\{A,B\}$.
We observe that the function:
\begin{equation}
\label{tea}
(g,t)\mapsto \gamma_{2}(g,t)\chi_{\nu_{G}}(g^{-1})\Pi\left(\tilde{\mu}^{G}_{g^{-1}}\circ\tilde{\mu}^{T}_{-\vartheta}\left(x\right),x\right)
\end{equation}
\noindent
is $\mathcal{C}^{\infty}$ because the singular support of $\Pi$ is included in the diagonal of $X\times X$.
Then
\begin{equation}
\label{tea}
t\mapsto \int_{G}\gamma_{2}(g,t)\chi_{\nu_{G}}(g^{-1})\Pi\left(\tilde{\mu}^{G}_{g^{-1}}\circ\tilde{\mu}^{T}_{-\vartheta}\left(x\right),x\right)dg
\end{equation}
\noindent
is infinitely smooth and the Fourier transform is rapidly decreasing. Thus the contribution coming from $B$ is rapidly decreasing and we can multiply the integrand by $\gamma_{1}$. So we can only consider:
\begin{equation}
\label{store}
\begin{multlined}[t][12.5cm]
\widetilde{\Pi}_{\nu_{G},k\nu_{T}}\left(x,x\right)\\\sim \frac{d_{\nu_{G}}}{2\pi}\int_{G}\int_{(-\pi,\pi)^{d_{T}}}\gamma_{1}(g,\vartheta)\chi_{\nu_{G}}(g^{-1})e^{-ik\nu_{T}\cdot\vartheta}\Pi\left(\tilde{\mu}^{G}_{g^{-1}}\circ\tilde{\mu}^{T}_{-\vartheta}\left(x\right),x\right)d\vartheta dg.
\end{multlined}
\end{equation}
Now if $\gamma_{2}(g,\vartheta)\not=0$ then $\tilde{\mu}^{G}_{g^{-1}}\circ\tilde{\mu}^{T}_{-\vartheta}(x),x$ are near and we can represent $\Pi$ as
Fourier integral operator as in $\cite{secondo}$:
\begin{equation}
\label{scotch}
\Pi(y,y')=\int_{0}^{+\infty}e^{it\psi(y,y')}s(y,y',t)dt,
\end{equation}
\noindent
where $\Im{(\psi)}\geq 0$ and $s$ is a semiclassical symbol admitting an asymptotic expansion $s(y,y',t)\sim\sum_{j=0}^{+\infty}t^{n-j}s_{j}(y,y')$.
Inserting $(\ref{scotch})$ in $(\ref{store})$ we obtain:
\begin{equation}
\label{sharktale}
\begin{multlined}[t][12.5cm]
\widetilde{\Pi}_{\nu_{G},k\nu_{T}}\left(x,x\right)\\\sim \frac{d_{\nu_{G}}}{2\pi}\int_{G}\int_{(-\pi,+\pi)^{d_{T}}}\int_{0}^{+\infty}\gamma_{1}(g,\vartheta)\chi_{\nu_{G}}(g^{-1})e^{i[t\psi(\tilde{\mu}^{G}_{g^{-1}}\circ\tilde{\mu}^{T}_{-\vartheta}\left(x\right),x)-k\nu_{T}\cdot\vartheta]}\cdot\\
\cdot s\left(\tilde{\mu}^{G}_{g^{-1}}\circ\tilde{\mu}^{T}_{-\vartheta}\left(x\right),x,t\right) dt d\vartheta dg,
\end{multlined}
\end{equation}
\noindent
and performing the change of variables $t\rightarrow kt$, we get:
\begin{equation}
\label{sharktale2}
\begin{multlined}[t][12.5cm]
\widetilde{\Pi}_{\nu_{G},k\nu_{T}}\left(x,x\right)\sim \frac{d_{\nu_{G}}}{2\pi}k\cdot\int_{G}\int_{(-\pi,+\pi)^{d_{T}}}\int_{0}^{+\infty}\gamma_{1}(g,\vartheta)\chi_{\nu_{G}}(g^{-1})e^{ik\Psi(t,g,\vartheta,x)}\cdot\\
\cdot s\left(\tilde{\mu}^{G}_{g^{-1}}\circ\tilde{\mu}^{T}_{-\vartheta}\left(x\right),x,kt\right) dt d\vartheta dg,
\end{multlined}
\end{equation}
\noindent
where we have set $\Psi(t,g,\vartheta,x)=t\psi(\tilde{\mu}^{G}_{g^{-1}}\circ\tilde{\mu}^{T}_{-\vartheta}\left(x\right),x)-\nu_{T}\cdot\vartheta$. We shall now use integration by parts in $\vartheta$ to prove that only a rapidly decreasing contribution to the asymptotic is lost, if the integrand in $(\ref{sharktale2})$
is multiplied by a suitable cut-off function. In local coordinates we have $\tilde{\mu}^{G}_{g^{-1}}\circ\tilde{\mu}^{T}_{-\vartheta}\left(x\right)= x+O(\varepsilon)$
with $\varepsilon>0$ very small, because $\left(g,e^{i\vartheta}\right)\in U$ with $U$ a small neighborhood of $F_{x}$. Thus we have that:
$$d_{(\tilde{\mu}^{G}_{g^{-1}}\circ\tilde{\mu}^{T}_{-\vartheta}\left(x\right),x)}\psi=d_{(x,x)}\psi+O(\varepsilon)=\left(\alpha_{x},-\alpha_{x}\right)+O(\varepsilon),$$
\noindent
with $\partial_{\vartheta}\Psi= t\Phi_{T}(m)-\nu_{T}+O(\varepsilon)$. Therefore, since $\Phi_{T}(m)\not=\mathbf{0}$ and $\nu_{T}\not=\mathbf{0}$ we have for $t\gg 0$ that
$$\|\partial_{\vartheta}\Psi\| \geq Ct,$$
\noindent
for some $C>0$. In a similar way for $0<t\ll 1$ we have
$$\|\partial_{\vartheta}\Psi\|\geq C_{1}>0,$$
\noindent
for some $C_{1}>0$. Therefore by integration by parts in $d\vartheta$, we have that the asymptotics for $k\rightarrow +\infty$ is unchanged. We multiply the integrand by $\rho(t)$, where $\rho\in\mathcal{C}^{\infty}_{0}\left(\frac{1}{2D},2D\right)$ and $\rho \equiv 1$ on $\left(\frac{1}{D},D\right)$, so that the integral in $dt$ is now compactly supported.
We shall now use integration by parts in $dt$ to show that only a rapidly decreasing contribution is lost, if the integration in $(g,\vartheta)$ is restricted to a
tubular neighborhood of $F_{x}$ of radius $O(k^{\delta-\frac{1}{2}})$. We have that $\partial_{t}\Psi=\psi(\tilde{\mu}^{G}_{g^{-1}}\circ\tilde{\mu}^{T}_{-\vartheta}\left(x\right),x)$. If $\mathop{}\!\mathrm{dist}{\left(\left(g,e^{i\vartheta}\right),F_{x}\right)}\geq Ck^{\delta-\frac{1}{2}}$, then
$$\mathop{}\!\mathrm{dist}{\left(\tilde{\mu}^{G}_{g^{-1}}\circ\tilde{\mu}^{T}_{-\vartheta}\left(x\right),x\right)}\geq Ck^{\delta-\frac{1}{2}}$$
\noindent
and so:
\begin{equation}
\label{intttt}
\left|\psi(\tilde{\mu}^{G}_{g^{-1}}\circ\tilde{\mu}^{T}_{-\vartheta}\left(x\right),x)\right|\geq \Im{\psi(\tilde{\mu}^{G}_{g^{-1}}\circ\tilde{\mu}^{T}_{-\vartheta}\left(x\right),x)}\geq C_{2}k^{2\delta-1}
\end{equation}
\noindent
(see Corollary 2.3 of $\cite{secondo}$). Introducing the operator
$$L_{t}=\left[\psi(\tilde{\mu}^{G}_{g^{-1}}\circ\tilde{\mu}^{T}_{-\vartheta}\left(x\right),x)\right]^{-1}\partial_{t},$$
\noindent
we have that:
$$e^{ik\Psi}=-\frac{i}{k}L_{t}\left(e^{ik\Psi}\right).$$
We can now mimick the standard proof of the Stationary Phase Lemma: iteratively integrating by parts, we obtain at each step in view of $(\ref{intttt})$ a factor of order $O\left(k^{-2\delta}\right)$, and then after $N$ steps a factor of order $O\left(k^{-2N\delta}\right)$.
This proves that the contribution to the asymptotics coming from the locus where $\mathop{}\!\mathrm{dist}{\left(\left(g,e^{i\vartheta}\right),F_{x}\right)}\geq Ck^{\delta-\frac{1}{2}}$ is rapidly decreasing. We can now prove that $(\ref{sharktale2})$ is rapidly decreasing in $k$ for $\mathop{}\!\mathrm{dist}_{X}{(x,X_{0,\nu_{T}})}\geq Ck^{\delta-\frac{1}{2}}$.
Now we consider a bump function $\rho_{1}:P\rightarrow \mathbb{R}$ supported in a small neighborhood of $F_{x}$ and $\equiv 1$ near to $F_{x}$. The function $\rho_{1}$ is defined as $\rho_{1}=\rho_{1}(f,\xi)$ with $f\in F_{x}$ and $\xi$ the normal coordinate to $F_{x}$. We can multiply the integrand of $(\ref{sharktale2})$ by $\rho_{1}\left(f,k^{\frac{1}{2}-\delta}\xi\right)$ losing only an $O(k^{-\infty})$. Then if $\rho_{1}(g,t)\not=0$ we have $\tilde{\mu}^{G}_{g^{-1}}\circ\tilde{\mu}^{T}_{-\vartheta}\left(x\right)=x + O\left(k^{\delta-\frac{1}{2}}\right)$. Therefore:
$$d_{\left(\tilde{\mu}^{G}_{g^{-1}}\circ\tilde{\mu}^{T}_{-\vartheta}(x),x\right)}\psi
=d_{(x,x)}\psi+O\left(k^{\delta-\frac{1}{2}}\right)=\left(\alpha_{x},-\alpha_{x}\right)+O(k^{\delta-\frac{1}{2}}),$$
\noindent
and $\partial_{(\varsigma,\vartheta)}\psi=t\Phi_{P}(m)-\nu_{T}+ O\left(k^{\delta-\frac{1}{2}}\right)$. Here $(\varsigma,\vartheta)$ are local coordinates on $P$ induced by the exponential map $\exp_{P}$. Then if $\mathop{}\!\mathrm{dist}_{X}{\left(x,X_{0,\nu_{T}}\right)}\geq C_{3}k^{\delta-\frac{1}{2}}$ we have that:
$$\|\partial_{(\varsigma,\vartheta)}\psi\|\geq C_{4}k^{\delta-\frac{1}{2}}.$$
Thus we find a differential operator $L_{\varsigma,\vartheta}$ with $\left|L_{\varsigma,\vartheta}\right|\geq C_{5}k^{\delta-\frac{1}{2}}$ where $\mathop{}\!\mathrm{dist}_{X}{\left(x,X_{0,\nu_{T}}\right)}\geq O\left(k^{\delta-\frac{1}{2}}\right)$ such that $L_{\varsigma,\vartheta}\left(e^{ik\Psi}\right)=ike^{ik\Psi}$. Iterating the integration by parts, in view of the scaling factor we have at each step a factor $O\left(k^{-2\delta}\right)$. This proves that $\widetilde{\Pi}_{\nu_{G},k\nu_{T}}(x,x)=O(k^{-\infty})$ for $\mathop{}\!\mathrm{dist}_{X}{\left(x,X_{0,\nu_{T}}\right)}\geq Ck^{\delta-\frac{1}{2}}$. Let us consider $(x,y)\in X\times X$ with
$$\max{\{\mathop{}\!\mathrm{dist}_{X}(x,X_{0,\nu_{T}}),\mathop{}\!\mathrm{dist}_{X}(y,X_{0,\nu_{T}})\}}\geq Ck^{\delta-\frac{1}{2}}$$
\noindent
for every $\delta$ fixed and using the Cauchy-Schwarz inequality we have:
\begin{equation}
\label{antonio}
\left|\widetilde{\Pi}_{\nu_{G},k\nu_{T}}\left(x,y\right)\right|\leq \sqrt{\widetilde{\Pi}_{\nu_{G},k\nu_{T}}\left(x,x\right)}\cdot\sqrt{\widetilde{\Pi}_{\nu_{G},k\nu_{T}}\left(y,y\right)},
\end{equation}
\noindent
so $\widetilde{\Pi}_{\nu_{G},k\nu_{T}}\left(x,y\right)=O(k^{-\infty})$. This complete the proof of $1)$.
Let us now consider the proof of $2)$. Now setting $x_{jk}=x+\frac{(\theta_{j},\mathop{}\!\mathrm{v}_{j})}{\sqrt{k}}$ for $j=1,2$, using FIO representation as before and changing variables $t\rightarrow kt$, we get:
\begin{equation}
\label{newton}
\begin{multlined}[t][12.5cm]
\widetilde{\Pi}_{\nu_{G},k\nu_{T}}\left(x_{1k},x_{2k}\right)=\frac{kd_{\nu_{G}}}{(2\pi)^{d_{T}}}\int_{W}\chi_{\nu_{G}}(g^{-1})e^{ik\Psi^{(1)}(t,\vartheta,x)}\cdot\\
\cdot s\left(\tilde{\mu}^{G}_{g^{-1}}\circ\tilde{\mu}^{T}_{-\vartheta}\left(x_{1k}\right),x_{2k},kt\right)dV_{W}(w),
\end{multlined}
\end{equation}
\noindent
where
\begin{equation}
W=G\times (-\pi,\pi)^{d_{T}}\times (0,+\infty) \ \ \ \ dV_{W}(w)=dg d\vartheta dt,
\end{equation}
\noindent
and
\begin{equation}
\label{jacobi}
\Psi^{(1)}(t,\vartheta,x)=t\psi\left(\tilde{\mu}^{G}_{g^{-1}}\circ\tilde{\mu}^{T}_{-\vartheta}\left(x_{1k}\right),x_{2k}\right)-\nu_{T}\cdot \vartheta.
\end{equation}
Here $t=(t_{1}, \cdots, t_{d_{T}})=\left(e^{i\vartheta_{1}},\cdots , e^{i\vartheta_{d_{T}}}\right)=e^{i\vartheta}$. Let $F_{m}\subseteq P$, $F_{m}=\{p_{j}\}=\{(g_{j},t_{j})\}$ the finite stabilizer of $x\in X_{0,\nu_{T}}$. We introduce a bump function $\rho=\sum_{j=1}^{N_{x}}\rho_{j}$ with support of $\rho_{j}$ in a neighborhood of $p_{j}=(g_{j},t_{j})$. As consequence we have:
\begin{equation}
\label{eulero}
\widetilde{\Pi}_{\nu_{G},k\nu_{T}}\left(x_{1k},x_{2k}\right)\sim \sum_{j}\Pi_{\nu_{G},k\nu_{T}}\left(x_{1k},x_{2k}\right)^{(j)},
\end{equation}
\noindent
where each addend of $(\ref{eulero})$ is given by $(\ref{newton})$ multiplied by $\rho_{j}$.
In the support of each $p_{j}$ we write $g=g_{j}\exp_{G}{\left(\frac{\gamma}{\sqrt{k}}\right)}$ and $t=t_{j}e^{\frac{i\vartheta}{\sqrt{k}}}$, where with $\exp_{G}$ we denote the exponential map from $\mathfrak{g}\rightarrow G$ and $\gamma,\vartheta$ are coordinates respectively on $\mathfrak{g}\cong \mathbb{R}^{d_{G}}$,$\mathfrak{t}\cong \mathbb{R}^{d_{T}}$ associated with the respective orthonormal basis. Omitting $\rho_{j}$ in the integrand we have:
\begin{equation}
\label{leibnitz}
\begin{multlined}[t][12.5cm]
\widetilde{\Pi}_{\nu_{G},k\nu_{T}}\left(x_{1k},x_{2k}\right)\sim\frac{k^{1-\frac{d_{P}}{2}}d_{\nu_{G}}}{(2\pi)^{d_{T}}}\int_{W'}\chi_{\nu_{G}}\left(g^{-1}_{j}\exp_{G}{\left(-\frac{\gamma}{\sqrt{k}}\right)}\right)t_{j}^{-1}e^{ik\Psi^{(2)}(t,\vartheta,x)}\cdot\\ \\
\cdot s\left(\tilde{\mu}^{G}_{-\frac{\gamma}{\sqrt{k}}}\circ\tilde{\mu}^{T}_{-\frac{\vartheta}{\sqrt{k}}}\left(x_{1k}\right),x_{2k},kt\right)dV_{W'}(w),
\end{multlined}
\end{equation}
\noindent
where
\begin{equation}
W'=\mathbb{R}^{d_{G}}\times \mathbb{R}^{d_{T}}\times (0,+\infty) \ \ \ \ dV_{W'}(w)=d\gamma d\vartheta dt
\end{equation}
\noindent
and
\begin{equation}
\label{cantor}
\Psi^{(2)}(t,\vartheta,x)=t\psi\left(\tilde{\mu}^{G}_{-\frac{\gamma}{\sqrt{k}}}\circ\tilde{\mu}^{T}_{-\frac{\vartheta}{\sqrt{k}}}\circ\tilde{\mu}^{G}_{g^{-1}_{j}}\circ\tilde{\mu}^{T}_{t^{-1}_{j}}\left(x_{1k}\right),x_{2k}\right)-\nu_{T}\cdot \frac{\vartheta}{\sqrt{k}}.
\end{equation}
We write $x_{1k}^{j}=\tilde{\mu}^{G}_{g^{-1}_{j}}\circ\tilde{\mu}^{T}_{t^{-1}_{j}}\left(x_{1k}\right)=x+\frac{1}{\sqrt{k}}(\theta_{1},\mathop{}\!\mathrm{v}_{1}^{j})$, for a particular choice of $\sigma$ adapted section in the definition of HLC. We assume that the orthonormal basis of $\mathfrak{t}$ is taken as $(w_{1},\cdots,w_{d_{T}})$ with $(w_{1},\cdots,w_{d_{T}-1})$ an orthonormal basis for $\mathop{}\!\mathrm{Ker}{(\Phi_{T}(m))}$ and $\langle \Phi_{T}(m),w_{d_{T}}\rangle=\|\Phi_{T}(m)\|$.
So we have that, if $(v_{1},\cdots ,v_{d_{G}})$ is the orthonormal basis for $\mathfrak{g}$, an orthonormal basis for $\mathfrak{p}=\mathop{}\!\mathrm{Lie}{(P)}$ is of the form:
$$ (v_{1},\cdots,v_{d_{G}},w_{1},\cdots,w_{d_{T}}=\eta). $$
We call $(a_{1},\cdots, a_{d_{P}}=b)$ the corresponding linear coordinates on $\mathfrak{p}$ such that $a=(a_{1},\cdots, a_{d_{P}}=b)\in \mathbb{R}^{d_{P}-1}\cong \mathop{}\!\mathrm{Ker}{(\Phi_{P}(m))}$ and $a_{M}(m)\in \mathbb{R}^{2d_{M}}\cong T_{m}M$ is his injective valutation. Considering $(\ref{cantor})$, we write $(\gamma,\vartheta)\in \mathfrak{p}\cong \mathbb{R}^{d_{P}}\cong \mathbb{R}^{d_{G}}\times \mathbb{R}^{d_{T}}$ as
\begin{equation}
\label{zermelo}
(\gamma,\vartheta)=(\gamma,\vartheta')+\vartheta\eta=a+b\eta,
\end{equation}
\noindent
remember that $\nu_{T}=\lambda\cdot \Phi_{T}(m) \Rightarrow (\mathbf{0},\nu_{T})=\lambda\Phi_{P}(m)$, then we have:
$$\nu_{T}\cdot \vartheta=(\mathbf{0},\nu_{T})\cdot (a+b\eta)=b\lambda\|\Phi_{T}(m)\|.$$
Here $\lambda=\lambda_{\nu_{T}}$ is such that $\nu_{T}=\lambda_{\nu_{T}}\cdot\Phi_{T}(m)$. Thus we have:
\begin{equation}
\label{hilbert}
\Psi^{(2)}(t,b,x)=t\psi\left(\tilde{\mu}^{P}_{-\frac{a+b\eta}{\sqrt{k}}}\left(x_{1k}^{j}\right),x_{2k}\right)-\lambda_{\nu_{T}}\frac{\|\Phi_{T}(m)\|b}{\sqrt{k}}.
\end{equation}
Now let $p=a+b\eta$ we have that $a\in\mathop{}\!\mathrm{Ker}{(\Phi_{P}(m))}$ and $\eta\in\mathop{}\!\mathrm{Ker}{(\Phi_{P}(m))}^{\perp}$, with $\|\eta\|=1$, $\langle\Phi_{P}(m),\eta\rangle=\langle\Phi_{T}(m),\eta\rangle=\|\Phi_{T}(m)\|$.
We get:
\begin{equation}
\label{godel}
\begin{multlined}[t][12.5cm]
\tilde{\mu}^{P}_{-\frac{p}{\sqrt{k}}}\left(x_{1k}^{j}\right)=\tilde{\mu}^{P}_{-\frac{p}{\sqrt{k}}}\left(x+\frac{1}{\sqrt{k}}(\theta_{1},\mathop{}\!\mathrm{v}_{1}^{j})\right)=\\ \\
=x+\left(\frac{\theta_{1}+\langle\Phi_{T}(m),p\rangle}{\sqrt{k}}+\frac{1}{k}\omega_{m}(a_{M}(m)+b\eta_{M}(m),\mathop{}\!\mathrm{v}_{1}^{j})+ \widetilde{B}_{3}\left(\frac{a_{M}(m)}{\sqrt{k}},\frac{\mathop{}\!\mathrm{v}}{\sqrt{k}},\frac{b}{\sqrt{k}}\right),\right.\\
\left.,\frac{\theta_{1}}{\sqrt{k}}(\mathop{}\!\mathrm{v}_{1}^{j}-a_{M}(m)-b\eta_{M})+\widetilde{B}_{2}\left(\frac{a_{M}(m)}{\sqrt{k}},\frac{\mathop{}\!\mathrm{v}}{\sqrt{k}},\frac{b}{\sqrt{k}}\right)\right)
\\
\\=x+(\mathcal{A}_{j,k}(\vartheta,v_{1}^{j},v_{2}),\mathcal{B}_{j,k}(\vartheta,v_{1}^{j},v_{2})),
\end{multlined}
\end{equation}
\noindent
with $\widetilde{B}_{2},\widetilde{B}_{3}$ that vanish at the origin to third and second order.
We obtain that:\\
\\
\noindent
$t\psi\left(\tilde{\mu}^{P}_{-\frac{p}{\sqrt{k}}}\left(x_{1k}^{j}\right),x_{2k}\right)-\frac{\vartheta}{\sqrt{k}}$
\begin{equation}
\label{einstein}
\begin{multlined}[t][12.5cm]
=it\left[1-e^{i\left(\mathcal{A}_{j,k}(\vartheta,v_{1}^{j},v_{2})-\frac{\theta_{2}}{\sqrt{k}}\right)}\right]\\ -\frac{it}{k}\psi_{2}\left(\mathop{}\!\mathrm{v}_{1}^{j}-a_{M}(m)-b\eta_{M}(m),\mathop{}\!\mathrm{v}_{2}\right)e^{i\left(\mathcal{A}_{j,k}(\vartheta,v_{1}^{j},v_{2})-\frac{\theta_{2}}{\sqrt{k}}\right)}-\frac{\vartheta}{\sqrt{k}}+\\
+it R_{3}^{\psi}\left(\frac{1}{\sqrt{k}}\left(\mathop{}\!\mathrm{v}_{1}^{j}-\vartheta\xi_{M}(m)-\varsigma_{M}(m)\right),\frac{\mathop{}\!\mathrm{v}_{2}}{\sqrt{k}}\right)e^{i\left(\mathcal{A}_{j,k}(\vartheta,v_{1}^{j},v_{2})-\frac{\theta_{2}}{\sqrt{k}}\right)},
\end{multlined}
\end{equation}
\noindent
where $R_{3}^{\psi}$ vanishes to third order at the origin and
$$ \psi_{2}(r,s)=-i\omega_{m}(r,s)-\frac{1}{2}\|r-s\|^2 \ \ \ \ (r,s\in\mathbb{C}^{n}).$$
Now
\begin{equation*}
\begin{multlined}[t][12.5cm]
i\left(\mathcal{A}_{j,k}(\vartheta,v_{1}^{j},v_{2})-\frac{\theta_{2}}{\sqrt{k}}\right)=\\
=\frac{i}{\sqrt{k}}(\theta_{1}-\theta_{2}+b\|\Phi_{T}\|)+\frac{i}{k}\omega_{m}(a_{M}+b\eta_{M},\mathop{}\!\mathrm{v}_{1}^{j})+ \widetilde{B}_{3}'\left(\frac{a_{M}(m)}{\sqrt{k}},\frac{\mathop{}\!\mathrm{v}}{\sqrt{k}},\frac{b}{\sqrt{k}}\right),
\end{multlined}
\end{equation*}
\noindent
then
\begin{equation*}
\begin{multlined}[t][12.5cm]
1-e^{i\left(\mathcal{A}_{j,k}(\vartheta,v_{1}^{j},v_{2})-\frac{\theta_{2}}{\sqrt{k}}\right)}=\\
=1-\left\{ 1+ \frac{i}{\sqrt{k}}(\theta_{1}-\theta_{2}+b\|\Phi_{T}\|)+ \frac{i}{k}\omega_{m}\left(a_{M}(m)+b\eta_{M}(m),\mathop{}\!\mathrm{v}_{1}^{j}\right)-\right.\\
\left.-\frac{1}{2k}(\theta_{1}-\theta_{2}+b\|\Phi_{T}\|)^2+\widetilde{B}_{3}''\left(\frac{a_{M}(m)}{\sqrt{k}},\frac{\mathop{}\!\mathrm{v}}{\sqrt{k}},\frac{b}{\sqrt{k}}\right)\right\}\\
=- \frac{i}{\sqrt{k}}(\theta_{1}-\theta_{2}+b\|\Phi_{T}\|)-\frac{i}{k}\omega_{m}\left(a_{M}(m)+b\eta_{M}(m),\mathop{}\!\mathrm{v}_{1}^{j}\right)+\\
+\frac{1}{2k}(\theta_{1}-\theta_{2}+b\|\Phi_{T}\|)^2+\widetilde{B}_{3}'''\left(\frac{a_{M}(m)}{\sqrt{k}},\frac{\mathop{}\!\mathrm{v}}{\sqrt{k}},\frac{b}{\sqrt{k}}\right)
\end{multlined}
\end{equation*}
\noindent
and
\begin{equation*}
\begin{multlined}[t][12.5cm]
it\left[1-e^{i\left(\mathcal{A}_{j,k}(\vartheta,v_{1}^{j},v_{2})-\frac{\theta_{2}}{\sqrt{k}}\right)}\right]=\\
=\frac{t}{\sqrt{k}}(\theta_{1}-\theta_{2}+b\|\Phi_{T}\|)+\frac{1}{k}t\omega_{m}\left(a_{M}(m)+b\eta_{M}(m),\mathop{}\!\mathrm{v}_{1}^{j}\right)+\\
+\frac{i}{2k}t(\theta_{1}-\theta_{2}+b\|\Phi_{T}\|)^2+it\widetilde{B}_{3}'''\left(\frac{a_{M}(m)}{\sqrt{k}},\frac{\mathop{}\!\mathrm{v}}{\sqrt{k}},\frac{b}{\sqrt{k}}\right).
\end{multlined}
\end{equation*}
Thus as conseguence:
\begin{equation*}
\label{weyl}
\begin{multlined}[t][12.5cm]
\Psi^{(2)}(t,b,x)=\frac{t}{\sqrt{k}}(\theta_{1}-\theta_{2}+b\|\Phi_{T}\|)+\\
+\frac{1}{k}t\omega_{m}\left(a_{M}(m)+b\eta_{M}(m),\mathop{}\!\mathrm{v}_{1}^{(j)}\right)+\frac{i}{2k}t(\theta_{1}-\theta_{2}+b\|\Phi_{T}\|)^2-\lambda_{\nu_{T}}\frac{\|\Phi_{T}(m)\|b}{\sqrt{k}}-\\
-\frac{it}{k}\psi_{2}\left(\mathop{}\!\mathrm{v}_{1}^{j}-a_{M}(m)-b\eta_{M}(m),\mathop{}\!\mathrm{v}_{2}\right)+it\widetilde{B}_{3}'''\left(\frac{a_{M}(m)}{\sqrt{k}},\frac{\mathop{}\!\mathrm{v}}{\sqrt{k}},\frac{b}{\sqrt{k}}\right)\\
\end{multlined}
\end{equation*}
\begin{equation*}
\begin{multlined}[t][12.5cm]
=\frac{1}{\sqrt{k}}\left(t(\theta_{1}-\theta_{2}+b\|\Phi_{T}\|)-\lambda_{\nu_{T}}\|\Phi_{T}(m)\|b\right)+\\
+\frac{1}{k}\left[t\omega_{m}\left(a_{M}(m)+b\eta_{M}(m),\mathop{}\!\mathrm{v}_{1}^{(j)}\right)+\frac{i}{2}t(\theta_{1}-\theta_{2}+b\|\Phi_{T}\|)^2-\right.\\
\left.-it\psi_{2}\left(\mathop{}\!\mathrm{v}_{1}^{j}-a_{M}(m)-b\eta_{M}(m),\mathop{}\!\mathrm{v}_{2}\right) \right]+it\widetilde{B}_{3}'''\left(\frac{a_{M}(m)}{\sqrt{k}},\frac{\mathop{}\!\mathrm{v}}{\sqrt{k}},\frac{b}{\sqrt{k}}\right)
\end{multlined}
\end{equation*}
\noindent
and \\
\\
\noindent
$ik\Psi^{(2)}(t,b,x)$
\begin{equation*}
\label{weyl}
\begin{multlined}[t][12.5cm]
=i\sqrt{k}\left(t(\theta_{1}-\theta_{2}+b\|\Phi_{T}\|)-\lambda_{\nu_{T}}\|\Phi_{T}(m)\|b\right)+\\
+\left[it\omega_{m}\left(a_{M}(m)+b\eta_{M}(m),\mathop{}\!\mathrm{v}_{1}^{(j)}\right)-\frac{1}{2}t(\theta_{1}-\theta_{2}+b\|\Phi_{T}\|)^2+\right.\\
\left.+t\psi_{2}\left(\mathop{}\!\mathrm{v}_{1}^{j}-a_{M}(m)-b\eta_{M}(m),\mathop{}\!\mathrm{v}_{2}\right) \right]-kt\widetilde{B}_{3}'''\left(\frac{a_{M}(m)}{\sqrt{k}},\frac{\mathop{}\!\mathrm{v}}{\sqrt{k}},\frac{b}{\sqrt{k}}\right).
\end{multlined}
\end{equation*}
Continuing calculations we can rewrite the $j$-term as:
\begin{equation}
\label{weil}
\begin{multlined}[t][12.5cm]
\widetilde{\Pi}_{\nu_{G},k\nu_{T}}(x_{1k},x_{2k})^{(j)} \\
\sim k^{1-\frac{d_{P}}{2}}\frac{d_{\nu_{G}}}{(2\pi)^{d_{T}}}\cdot\int_{\mathbb{R}^{d_{P}-1}}da\cdot\left[\int_{0}^{+\infty}dt\int_{-\infty}^{+\infty}db e^{i\sqrt{k}\Upsilon(t,b)}e^{A(m,\theta,\mathop{}\!\mathrm{v},v,a)}B(j)\right],
\end{multlined}
\end{equation}
\noindent
where
\begin{equation}
\label{dedekind}
\Upsilon(t,b)=t\left(b\|\Phi_{T}(m)\|+\theta_{1}-\theta_{2}\right)-\lambda\|\Phi_{T}\|b,
\end{equation}
\begin{equation}
\begin{multlined}[t][12.5cm]
A(m,\theta,\mathop{}\!\mathrm{v},v,a)\\
=-\frac{t}{2}(b\|\Phi_{T}(m)\|+\theta_{1}-\theta_{2})^2+it\omega_{m}(a_{M}(m)+b\eta_{M}(m),\mathop{}\!\mathrm{v}_{1}^{j})+\\
+t\psi_{2}(\mathop{}\!\mathrm{v}_{1}^{j}-a_{M}(m)-b\eta_{M}(m),\mathop{}\!\mathrm{v}_{2})e^{i\mathcal{A}_{j,k}(\vartheta,v_{1}^{j},v_{2})}
\end{multlined}
\end{equation}
\noindent
and
\begin{equation}
\label{degiorgi}
B(j)=\chi_{\nu_{G}}(g_{j}^{-1}) e^{-ik\vartheta_{j}\nu_{T}}.
\end{equation}
The internal integral in $(\ref{weil})$ is oscillatory in $\sqrt{k}$ with phase $\Upsilon$. The phase has critical points $(t_{0},b_{0})=\left(\lambda_{\nu_{T}},\frac{\theta_{2}-\theta_{1}}{\|\Phi_{T}(m)\|}\right)$. The Hessian is:
$$H(\Upsilon)(P_{0})=\left(\begin{array}{cc} 0 & \|\Phi_{T}(m)\| \\ \|\Phi_{T}(m)\| & 0 \end{array}\right)$$
\noindent
and
$$\sqrt{\left|\det{\frac{\sqrt{k}H}{2\pi}}\right|}=\frac{\sqrt{k}}{2\pi}\|\Phi_{T}(m)\|.$$
Using the Stationary Phase Lemma $\cite{settimo}$ we have:
\begin{equation}
\label{pitagora}
\begin{multlined}[t][12.5cm]
\widetilde{\Pi}_{\nu_{G},k\nu_{T}}(x_{1k},x_{2k})^{(j)}\sim \frac{d_{\nu_{G}}}{(2\pi)^{d_{T}-1}}B(j)k^{\frac{1}{2}-\frac{d_{P}}{2}}\cdot\int_{\mathbb{R}^{d_{P}-1}}e^{-i\sqrt{k}(\theta_{2}-\theta_{1})\lambda_{\nu}}\cdot\\
\cdot e^{A(a,t_{0},b_{0})}\cdot\frac{1}{\|\Phi_{T}(m)\|}\left(\frac{k\lambda_{\nu_{T}}}{\pi}\right)^{d_{M}}da,
\end{multlined}
\end{equation}
\noindent
where
\begin{equation*}
\begin{multlined}[t][12.5cm]
A(a,t_{0},b_{0})=\lambda_{\nu_{T}}\left[i\omega_{m}\left(a_{M}(m)+b_{0}\eta_{M},\mathop{}\!\mathrm{v}_{1}^{j}\right)+\right.\\
\left.+\psi_{2}\left(\mathop{}\!\mathrm{v}_{1}^{j}-a_{M}(m)-b_{0}\eta_{M}(m),\mathop{}\!\mathrm{v}_{2}\right)\right],
\end{multlined}
\end{equation*}
\noindent
and we have to evaluate the integral $\int_{\mathbb{R}^{d_{P}-1}}da e^{A(a,t_{0},b_{0})}$. In order to do this we define the following spaces:
$$V_{m}=\mathop{}\!\mathrm{val}_{m}{\left(\mathfrak{g}\oplus\mathop{}\!\mathrm{Ker}{(\Phi_{T}(m))}\right)}=\mathop{}\!\mathrm{val}_{m}{(\mathop{}\!\mathrm{Ker}{(\Phi_{P}(m))})} \subseteq T_{m}M,$$
$$N_{m}= J_{m}(V_{m}),$$
$$H_{m}=[V_{m}\oplus N_{m}]^{\perp},$$
\noindent
where $V_{m},H_{m}$ are complex subspaces of $T_{m}M$ and $N_{m}$ is the normal space to $M_{0,\nu_{T}}$. Decomposing $\mathop{}\!\mathrm{v}\in T_{m}M$ as $\mathop{}\!\mathrm{v}=\mathop{}\!\mathrm{v}_{h}+\mathop{}\!\mathrm{v}_{v}+\mathop{}\!\mathrm{v}_{t}$, with respectively $\mathop{}\!\mathrm{v}_{h}\in H_{m}$, $\mathop{}\!\mathrm{v}_{v}\in V_{m}$ and $\mathop{}\!\mathrm{v}_{t}\in N_{m}$, we have that $a_{M}=a_{Mv}$, $\eta_{M}=\eta_{Mh}+\eta_{Mv}$ and\\
\\
\noindent
$i\omega_{m}\left(a_{M}(m)+b_{0}\eta_{M}(m),\mathop{}\!\mathrm{v}_{1}^{j}\right)+i\psi_{2}\left(\mathop{}\!\mathrm{v}_{1}^{j}-a_{M}(m)-b_{0}\eta_{M}(m),\mathop{}\!\mathrm{v}_{2}\right)$
\begin{equation}
\label{fermat2}
\begin{multlined}[t][12.5cm]
=i\omega_{m}(a_{M}(m),\mathop{}\!\mathrm{v}_{1t}^{j}+\mathop{}\!\mathrm{v}_{2t})-\frac{1}{2}\|\mathop{}\!\mathrm{v}_{1v}-a_{M}(m)-b_{0}\eta_{Mv}(m)-\mathop{}\!\mathrm{v}_{2v}\|^{2}\\
+ib_{0}\omega_{m}(\eta_{Mh}(m),\mathop{}\!\mathrm{v}_{1h}^{j}+\mathop{}\!\mathrm{v}_{2h})+ib_{0}\omega_{m}(\eta_{Mv}(m),\mathop{}\!\mathrm{v}_{1t}^{j}+\mathop{}\!\mathrm{v}_{2t})-i\omega_{m}(\mathop{}\!\mathrm{v}_{1h}^{j},\mathop{}\!\mathrm{v}_{2h})+\\
-\frac{1}{2}\|\mathop{}\!\mathrm{v}_{1h}^{j}-b_{0}\eta_{Mh}(m)-\mathop{}\!\mathrm{v}_{2h}\|^{2}-i\omega_{m}(\mathop{}\!\mathrm{v}_{1v}^{j},\mathop{}\!\mathrm{v}_{2t})-\frac{1}{2}\|\mathop{}\!\mathrm{v}_{1t}^{j}-\mathop{}\!\mathrm{v}_{2t}\|^{2}-i\omega_{m}(\mathop{}\!\mathrm{v}_{1t}^{j},\mathop{}\!\mathrm{v}_{2v}).
\end{multlined}
\end{equation}
We define $r_{M}(m)\in \mathbb{R}^{d_{P}-1}$ translated by $a_{M}(m)$ such that:
$$ r_{M}(m)=\mathop{}\!\mathrm{v}_{1v}^{j}-a_{M}(m)-b_{0}\eta_{Mv}(m)-\mathop{}\!\mathrm{v}_{2v}, $$
\noindent
so $ a_{M}(m)=\mathop{}\!\mathrm{v}_{1v}^j-r_{M}(m)-b_{0}\eta_{Mv}(m)-\mathop{}\!\mathrm{v}_{2v}$ and we get:
\begin{equation}
\label{gaylussac2}
\begin{multlined}[t][12.5cm]
\omega_{m}(a_{M}(m),\mathop{}\!\mathrm{v}_{1t}^{j}+\mathop{}\!\mathrm{v}_{2t})=\omega_{m}(\mathop{}\!\mathrm{v}_{1v}^{j},\mathop{}\!\mathrm{v}_{1t}^{j})+\omega_{m}(\mathop{}\!\mathrm{v}_{1v}^{j},\mathop{}\!\mathrm{v}_{2t})-\omega_{m}(\mathop{}\!\mathrm{v}_{2v},\mathop{}\!\mathrm{v}_{1t}^{j})-\omega_{m}(\mathop{}\!\mathrm{v}_{2v},\mathop{}\!\mathrm{v}_{2t})-\\
-\omega_{m}(r_{M}(m),\mathop{}\!\mathrm{v}_{1t}^{j}+\mathop{}\!\mathrm{v}_{2t})-\omega_{m}(b_{0}\eta_{Mv}(m),\mathop{}\!\mathrm{v}_{1t}^{j}+\mathop{}\!\mathrm{v}_{2t}).
\end{multlined}
\end{equation}
Putting $(\ref{gaylussac2})$ in $(\ref{fermat2})$ and deleting the opposite terms, we obtain:\\
\\
\noindent
$i\omega_{m}\left(a_{M}(m)+b_{0}\eta_{M}(m),\mathop{}\!\mathrm{v}_{1}^{j}\right)+i\psi_{2}\left(\mathop{}\!\mathrm{v}_{1}^{j}-a_{M}(m)-b_{0}\eta_{M}(m),\mathop{}\!\mathrm{v}_{2}\right)$
\begin{equation*}
\begin{multlined}[t][12.5cm]
=i\left[\omega_{m}(\mathop{}\!\mathrm{v}_{1v}^{j},\mathop{}\!\mathrm{v}_{1t}^{j})-\omega_{m}(\mathop{}\!\mathrm{v}_{2v},\mathop{}\!\mathrm{v}_{2t})\right]-\frac{1}{2}\|\mathop{}\!\mathrm{v}_{1t}^j-\mathop{}\!\mathrm{v}_{2t}\|^{2} +\\
+ib_{0}\omega_{m}(\eta_{Mh}(m),\mathop{}\!\mathrm{v}_{1h}^{j}+\mathop{}\!\mathrm{v}_{2h})-i\omega_{m}(\mathop{}\!\mathrm{v}_{1h}^{j},\mathop{}\!\mathrm{v}_{2h})-\frac{1}{2}\left\|\mathop{}\!\mathrm{v}_{1h}^{j}-b_{0}\eta_{Mh}(m)-\mathop{}\!\mathrm{v}_{2h}\right\|^2-\\
-i\omega_{m}(r_{M}(m),\mathop{}\!\mathrm{v}_{1t}^{j}+\mathop{}\!\mathrm{v}_{2t})-\frac{1}{2}\|r_{M}(m)\|^2.
\end{multlined}
\end{equation*}
Let $C$ the matrix of $\mathop{}\!\mathrm{val}_{m}:\mathop{}\!\mathrm{Ker}{(\Phi_{T}(m))}\oplus\mathfrak{g}\rightarrow V_{m}$. Changing variable $r'=Cr$ we have that $dr=\det{C}^{-1}dr'$. We have to evaluate:
\begin{equation}
\label{nash}
\frac{1}{\det{C}}\int_{\mathbb{R}^{d_{P}-1}}e^{\lambda_{\nu_{T}}\left[-i\omega_{m}(r',\mathop{}\!\mathrm{v}_{1t}^{j}+\mathop{}\!\mathrm{v}_{2t})-\frac{1}{2}\|r'\|^2\right]}dr'.
\end{equation}
We make the substitution $s=\sqrt{\lambda_{\nu_{T}}}r'$, so we obtain:
\begin{equation}
\label{nash2}
=\frac{\lambda_{\nu_{T}}^{-\frac{1}{2}(d_{P}-1)}}{\det{C}}\int_{\mathbb{R}^{d_{P}-1}}e^{\left[-i\omega_{m}(s,\sqrt{\lambda_{\nu_{T}}}(\mathop{}\!\mathrm{v}_{1t}^{j}+\mathop{}\!\mathrm{v}_{2t}))-\frac{1}{2}\|s\|^2\right]}ds=\frac{(2\pi)^{\frac{d_{P}-1}{2}}}{\lambda^{\frac{1}{2}(d_{P}-1)}}e^{-\frac{1}{2}\lambda_{\nu_{T}}\|\mathop{}\!\mathrm{v}_{1t}^{j}+\mathop{}\!\mathrm{v}_{2t}\|^{2}}.
\end{equation}
Thus the exponential factor in the asymptotic expansion is $e^{H(v_{1}^{j},v_{2})}$ with
\begin{equation}
\begin{multlined}[t][12.5cm]
H(v_{1}^{j},v_{2})=\lambda_{\nu_{T}}\left(-i\omega_{m}\left(\mathop{}\!\mathrm{v}_{1h}^{j},\mathop{}\!\mathrm{v}_{2h}\right)-\|\mathop{}\!\mathrm{v}_{1t}\|^2-\|\mathop{}\!\mathrm{v}_{2t}\|^2+ib_{0}\omega_{m}\left(\eta_{Mh}(m),\mathop{}\!\mathrm{v}_{1h}^{j}+\mathop{}\!\mathrm{v}_{2h}\right)\right.\\
\left.+i[\omega_{m}\left(\mathop{}\!\mathrm{v}_{1v}^{j},\mathop{}\!\mathrm{v}_{1t}^j\right)-\omega_{m}\left(\mathop{}\!\mathrm{v}_{2v},\mathop{}\!\mathrm{v}_{2t}\right)]-\frac{1}{2}\left\|\mathop{}\!\mathrm{v}_{1h}^{j}-b_{0}\eta_{Mh}(m)-\mathop{}\!\mathrm{v}_{2h}\right\|^2\right),
\end{multlined}
\end{equation}
\noindent
where $b_{0}=\frac{\theta_{2}-\theta_{1}}{\|\Phi_{T}(m)\|}$ and the principal term is of the form:
\begin{equation}
\label{colzanileonardo}
\begin{multlined}[t][12.5cm]
\frac{d_{\nu_{G}}2^{\frac{d_{G}}{2}}}{(\sqrt{2}\pi)^{d_{T}-1}}\left(\frac{k\lambda_{\nu_{T}}}{\pi}\right)^{d_{M}-\frac{d_{G}}{2}-\frac{(d_{T}-1)}{2}}\cdot\chi_{\nu_{G}}(g_{j}^{-1}) e^{-ik\vartheta_{j}\nu_{T}}\cdot\\
\cdot\frac{e^{i\sqrt{k}(\theta_{2}-\theta_{1})\lambda_{\nu}}\cdot e^{H}}{\|\Phi_{T}(m)\|\mathcal{D}(m)}.
\end{multlined}
\end{equation}
This complete the proof of $2)$.
Let us now describe the necessary changes to the previous argument to prove $3)$.
Instead of considering a neighborhood of $(x,x)$ we consider the asymptotic in a neighborhood of $(x,p_{0}\cdot x)$.
We set $y=p\cdot x$ and we assume given the local system of Heisenberg coordinates in a neighborhood of $x$ and $y$. We may also assume without loss that the Heisenberg coordinate system centered at $y$ is obtained from the one centered at $x$ by a $p$-translation, that is,
$$y+(\theta,\mathop{}\!\mathrm{v})=p\cdot \big(x+(\theta,\mathop{}\!\mathrm{v})\big).$$
We must evaluate:
$$ \widetilde{\Pi}_{\nu_{G},k\nu_{T}}\left(x_{1k},y_{2k}\right),$$
\noindent
where $x_{1k}=x+\frac{u_{1}}{\sqrt{k}}$, $y_{2k}=y+\frac{u_{2}}{\sqrt{k}}$ and $u_{j}=\left(\theta_{j},\mathop{}\!\mathrm{v}_{j}\right)$. Now with the preceding interpretation we have that $y_{2k}=p_{0}\left(x+\frac{u_{2}}{\sqrt{k}}\right)=p_{0}x_{2k}$. Proceeding as in the previous case we have:
\begin{equation}
\label{riemannnnn1}
\begin{multlined}[t][12.5cm]
\widetilde{\Pi}_{\nu_{G},k\nu_{T}}\left(x_{1k},y_{2k}\right)\\=\frac{d_{\nu_{G}}}{(2\pi)^{d_{T}}}\int_{G}\int_{T}\chi_{\nu_{G}}(g^{-1})t^{-k\nu_{T}} \cdot\Pi\left(\tilde{\mu}^{G}_{g^{-1}}\circ\tilde{\mu}^{T}_{t^{-1}}\left(x\right),p_{0}\cdot x_{2k}\right)dtdg,
\end{multlined}
\end{equation}
\noindent
and due to the unitary action:
\begin{equation}
\label{riemannnnn2}
\begin{multlined}[t][12.5cm]
\widetilde{\Pi}_{\nu_{G},k\nu_{T}}\left(x_{1k},y_{2k}\right)\\=\frac{d_{\nu_{G}}}{(2\pi)^{d_{T}}}\int_{G}\int_{T}\chi_{\nu_{G}}(g^{-1})t^{-k\nu_{T}}
\cdot\Pi\left(\tilde{\mu}^{P}_{p_{0}^{-1}}\circ\tilde{\mu}^{G}_{g^{-1}}\circ\tilde{\mu}^{T}_{t^{-1}}\left(x\right), x_{2k}\right)dtdg.
\end{multlined}
\end{equation}
Let $p_{0}=\left(g_{0},t_{0}\right)=\left(g_{0},e^{i\vartheta_{0}}\right)\in P$, then
$$ \tilde{\mu}^{P}_{p_{0}^{-1}}\circ\tilde{\mu}^{P}_{(g,t)^{-1}}(x_{1k})=\tilde{\mu}_{g_{0}^{-1}g^{-1}}^{G}\circ\tilde{\mu}_{t_{0}^{-1}t^{-1}}^{T}(x_{1k}), $$
\noindent
and changing variables $g'=gg_{0}$, $t'=tt_{0}$, we have $g=g'g_{0}^{-1}$, $t=t't_{0}^{-1}$ and $\vartheta=\vartheta'-\vartheta_{0}$.
Then
\begin{equation}
\label{riemannnnnstacippa}
\begin{multlined}[t][12.5cm]
\widetilde{\Pi}_{\nu_{G},k\nu_{T}}\left(x_{1k},y_{2k}\right)\\
=\frac{d_{\nu_{G}}}{(2\pi)^{d_{T}}}\int_{G}\int_{T}\overline{\chi_{\nu_{G}}(g'g_{0}^{-1})}e^{-ik\nu_{T}(\vartheta'-\vartheta_{0})}\Pi\left(\tilde{\mu}^{P}_{p_{0}^{-1}}\circ\tilde{\mu}^{G}_{g'^{-1}}\circ\tilde{\mu}^{T}_{t'^{-1}}\left(x\right), x_{2k}\right)dt'dg',
\end{multlined}
\end{equation}
\noindent
that will be the same as before with the difference that in the $j$-addendum we will make the substitution: $$\overline{\chi_{\nu_{G}}\left(g_{j}\right)}t_{j}^{-k\nu_{T}}=\overline{\chi_{\nu_{g},k\nu_{T}}(p_{j})}\mapsto \overline{\chi_{\nu_{G}}\left(g_{j}g_{0}^{-1}\right)}(t_{j}t_{0}^{-1})^{-k\nu_{T}}=\overline{\chi_{\nu_{G},k\nu_{T}}(p_{j}p_{0}^{-1})}.$$
This complete the proof of $3)$ and complete the proof of the main Theorem.
\hfill $\Box$
\section{Proof of Theorem $\bf{~\ref{teo:diagonal}}$}
$\mathop{}\!\mathit{Proof}.$
On the diagonal we have:
\begin{equation}
\label{arnold}
\begin{multlined}[t][12.5cm]
\widetilde{\Pi}_{\nu_{G},k\nu_{T}}(x,x)\\
=d_{\nu_{G}}\int_{G}\int_{T}\chi_{\nu_{G}}(g^{-1})t^{-k\nu_{T}}\Pi\left(\tilde{\mu}_{g^{-1}}^{G}\circ\tilde{\mu}_{t^{-1}}^{T}(x),x\right)dt dg\\
=\frac{d_{\nu_{G}}}{(2\pi)^{d_{T}}}\int_{G}\int_{(-\pi,\pi)^{d_{T}}}\chi_{\nu_{G}}(g^{-1})e^{-ik\nu_{T}\cdot\vartheta}\Pi\left(\tilde{\mu}_{g^{-1}}^{G}\circ\tilde{\mu}_{-\vartheta}^{T}(x),x\right)d\vartheta dg,
\end{multlined}
\end{equation}
\noindent
where $\vartheta=(\vartheta_{1},\cdots\cdots, \vartheta_{d_{T}})$, $\nu_{T}=(\nu_{1},\cdots,\nu_{d_{T}})\in\mathbb{Z}^{d_{T}}$ and $\nu_{T}\cdot\vartheta=\sum_{j=1}^{d_{T}}{\nu_{T}}_{j}\vartheta_{j}$.
We consider $F_{x}=\left\{(g_{1},t_{1}),\cdots ,(g_{N_{x}},t_{N_{x}})\right\}$ the stabilizer, with $|F_{x}|=N_{x}$.
Let $\varepsilon >0$ and we consider the following open subsets of $P=G\times T$:
$$A=\{(g,t)\in G\times T : \mathop{}\!\mathrm{dist}_{G\times T}{((g,t),F_{x})}<2\varepsilon\}$$
$$B=\{(g,t)\in G\times T : \mathop{}\!\mathrm{dist}_{G\times T}{((g,t),F_{x})}>\varepsilon\}.$$
Then $P=A\cup B$, and we have choose a partition of unity $\gamma_1+\gamma_2=1$ subordinate to the open cover $\{A,B\}$.
Then for $(g,t)\in \mathop{}\!\mathrm{supp} (\gamma_2)$ we have
$$\mathop{}\!\mathrm{dist}_{X}\left(\tilde{\mu}_{g^{-1}}^{G}\circ\tilde{\mu}_{t^{-1}}^{T}(x),x\right)\geq C\varepsilon$$
\noindent
for some constant $C>0$. Therefore, the map
$$(g,t)\in P\mapsto \gamma_{2}(g,t)\chi_{\nu_{G}}(g^{-1})\Pi\left(\tilde{\mu}_{g^{-1}}^{G}\circ\tilde{\mu}_{t^{-1}}^{T}(x),x\right)$$
\noindent
is $\mathcal{C}^\infty$ because the singular support of $\Pi$ is included in the diagonal $X\times X$. As consequence the function:
$$ t\in T\mapsto \int_{G}\gamma_{2}(g,t)\chi_{\nu_{G}}(g^{-1})\Pi\left(\tilde{\mu}_{g^{-1}}^{G}\circ\tilde{\mu}_{t^{-1}}^{T}(x),x\right)dg $$
\noindent
is $\mathcal{C}^{\infty}$ and so its Fourier transform evaluated at $k\nu_{T}$ is rapidly decreasing for $k\rightarrow+\infty$, since by assumption $\nu_{T}\neq 0$.
We set $\gamma_{1}(g,\vartheta)=\sum_{j=1}^{N_{x}}\rho_{j}(g,\vartheta)$, with each $\rho_{j}$ supported in a neighborhood of $(g_{j},\vartheta_{j})$, and consider
$$ \widetilde{\Pi}_{\nu_{G},k\nu_{T}}(x,x)\sim \sum_{j=1}^{N_{x}}\widetilde{\Pi}_{\nu_{G},k\nu_{T}}(x,x)^{(j)} $$
\noindent
where
\begin{equation}
\label{arnold2}
\begin{multlined}[t][12.5cm]
\widetilde{\Pi}_{\nu_{G},k\nu_{T}}(x,x)^{(j)}\\
\sim \frac{d_{\nu_{G}}}{(2\pi)^{d_{T}}}\int_{G}\int_{(-\pi,\pi)^{d_{T}}}\rho_{j}(g,\vartheta)\chi_{\nu_{G}}(g^{-1})e^{-ik\nu_{T}\cdot\vartheta}\Pi\left(\tilde{\mu}_{g^{-1}}^{G}\circ\tilde{\mu}_{-\vartheta}^{T}(x),x\right)d\vartheta dg.
\end{multlined}
\end{equation}
Let us now examine the asymptotics of each integrand separately.
On the support of $\gamma_1$, $\tilde{\mu}_{g^{-1}}^{G}\circ\tilde{\mu}_{t^{-1}}^{T}(x)$ is close to $x$, and therefore we may replace $\Pi$ by its representation as a Fourier integral, perhaps after disregarding a smoothing term which contributes negligibly to the asymptotics. After rescaling in $t$ we have:
\begin{equation}
\label{arnold3}
\begin{multlined}[t][12.5cm]
\widetilde{\Pi}_{\nu_{G},k\nu_{T}}(x,x)^{(j)}\\
\sim \frac{d_{\nu_{G}}}{(2\pi)^{d_{T}}}k\int_{G}\int_{(-\pi,\pi)^{d_{T}}}\int_{0}^{+\infty}\chi_{\nu_{G}}(g^{-1})e^{ik\Psi(x,t,g,\vartheta)}\\
\cdot s\left(\tilde{\mu}_{g^{-1}}^{G}\circ\tilde{\mu}_{-\vartheta}^{T}(x),x,kt\right)\gamma_{j}(g,\vartheta)d t d\vartheta dg
\end{multlined}
\end{equation}
\noindent
with
\begin{equation}
\label{arnold4}
\begin{multlined}[t][12.5cm]
\Psi(x,t,g,\vartheta)=t\psi\left(\tilde{\mu}_{g^{-1}}^{G}\circ\tilde{\mu}_{-\vartheta}^{T}(x),x\right)-\nu_{T}\cdot\vartheta.
\end{multlined}
\end{equation}
Let us regard $\ref{arnold3}$ as an oscillatory integral with a complex phase $\Psi$ of positive type. Let us look for critical points of $\Psi$.
We have that $\partial_{t}\Psi(x,t,g,\vartheta)=\psi\left(\tilde{\mu}_{g^{-1}}^{G}\circ\tilde{\mu}_{-\vartheta}^{T}(x),x\right)=0$ if and only if $\vartheta=\vartheta_{j}$ and $g=g_{j}$. We set $ \vartheta=\eta+\vartheta_{j}$ with $\eta\sim \mathbf{0}$. In the neighborhood of $g_{j}$, we can write $g=g_{j}\exp_{G}{\xi}$, where $\xi\in \mathfrak{g}$ is close to the origin. With abuse of notation we shall write $\tilde{\mu}_{-\varsigma}^{G}\circ\tilde{\mu}_{-\vartheta}^{T}(x)=\tilde{\mu}_{-\xi}^{G}\circ\tilde{\mu}_{-\eta}^{T}(x)$. Upon choosing orthonormal basis for the Lie algebras, we shall identify them with $\mathbb{R}^{d_{T}}$ and $\mathbb{R}^{d_{P}}$, respectively. We consider now $\left.\partial_{(\varsigma,\vartheta)}\Psi(x,t,\varsigma,\vartheta)\right|_{\xi=0,\eta=0}=t\Phi_{T}-\nu_{T}$. Thus in the new coordinates at any critical point $\xi=0$, $\eta=0$. The critical points are of the form:
$$P_{0}=(t_{0},(\varsigma_{0},\vartheta_{0}))=\left(\frac{\|\nu_{T}\|}{\|\Phi_{T}(m)\|},(\varsigma_{j},\vartheta_{j})\right).$$
Considering the second derivatives, $\partial_{tt}^{2}\Psi=0$, and using Heisenberg coordinates we write:
\begin{equation}
\label{noscalll}
\begin{multlined}[t][12.5cm]
\Psi(x,t,\xi,\eta)=t\psi\left(x+\left(\xi\Phi_{G}(m)+\eta\Phi_{T}(m)+ O(\|(\xi,\eta)\|^3),\right.\right.\\
\left.\left.,-\xi_{M}(m)-\eta_{M}(m)+ O(\|(\xi,\eta)\|^2)\right),x\right)-\nu_{T}\cdot\eta-\nu_{T}\cdot\vartheta_{j}.
\end{multlined}
\end{equation}
Here with abuse of language we have identified $\eta_{M}(m)$ and $\xi_{M}(m)$ with their representation in local coordinates.
We note that $m\in M_{0,\nu_{T}}$ so
\begin{equation}
\label{noscalll2}
\begin{multlined}[t][12.5cm]
\Psi(x,t,\xi,\eta)=t\psi\left(x+\left(\eta\Phi_{T}(m)+ O(\|(\xi,\eta)\|^3),\right.\right.\\
\left.\left.,-\xi_{M}(m)-\eta_{M}(m)+ O(\|(\xi,\eta)\|^2)\right),x\right)-\nu_{T}\cdot\eta-\nu_{T}\cdot\vartheta_{j}\\
=-\nu_{T}\cdot\eta-\nu_{T}\cdot\vartheta_{j}+it\left\{\left[1-e^{i\eta\Phi_{T}(m)}\right]\right.\\ \left.+\frac{1}{2}\|\eta_{M}(m)+\xi_{M}(m)\|^{2}e^{i\eta\Phi_T(m)}
+ O(\|(\xi,\eta)\|^3)\right\}
\end{multlined}
\end{equation}
\noindent
and setting $\gamma=(\gamma_1,\cdots,\gamma_{d_P})=(\xi_{1},\cdots,\xi_{d_{G}},\eta_{1},\cdots,\eta_{d_{T}})$ and $\sigma=(\varsigma,\vartheta)$, we have $\sigma_{0}=(\varsigma_{0},\vartheta_{0})$ and
\begin{equation}
\label{noscalll2}
\begin{multlined}[t][12.5cm]
\left.\partial_{t\sigma}^{2}\Psi(x,t,\sigma)\right|_{(t_{0},\sigma_{0})}=\Phi_{T}(m).
\end{multlined}
\end{equation}
Thus, we have:
\begin{equation}
\label{noscalll3}
\begin{multlined}[t][12.5cm]
\left.\partial_{\sigma_{l}\sigma_{k}}^{2}\Psi(x,t,\sigma)\right|_{(t_{0},\sigma_{0})}=i\lambda_{\nu_{T}}\left[\Phi_{l}(m)\Phi_{k}(m)+\langle\gamma_{l},\gamma_{k}\rangle_{m}\right],
\end{multlined}
\end{equation}
\noindent
with $\lambda_{\nu_{T}}=t_{0}$, we are in the same case of $\cite{quarto}$ in the proof of Theorem 2, so we have that:
\begin{equation}
\label{superhessian}
\det{H(t_{0},\sigma_{0})}=i^{d_{P}+1}\lambda_{\nu_{T}}^{d_{P}-1}\|\Phi_{T}(m)\|^2\det{C(m)}.
\end{equation}
Where $C(m)$ is a scalar product on $\mathop{}\!\mathrm{Ker}{(\Phi_{P}(m))}$ and
\begin{equation}
\label{superhessian}
\det{\left[\frac{k}{2\pi i}H(t_{0},\sigma_{0})\right]}=\left(\frac{k}{2\pi}\right)^{d_{P}+1}\lambda_{\nu_{T}}^{d_{P}-1}\|\Phi_{T}(m)\|^2\mathcal{D}(m)^2.
\end{equation}
The principal term is:
\begin{equation}
\label{principal4}
\frac{d_{\nu_{G}}e^{-ik\vartheta_{j}\cdot\nu_{T}}2^{d_{G}/2}\chi_{\nu_{G}}(g_{j}^{-1})}{\left(\sqrt{2}\pi\right)^{d_{T}-1}}\cdot\left(\frac{\|\nu_{T}\|k}{\pi}\right)^{d_{M}+\frac{1-d_{P}}{2}}\cdot\frac{1}{\mathcal{D}(m)\|\Phi_{T}(m)\|^{d_{M}+1+\frac{1-d_{P}}{2}}}.
\end{equation}
This complete the proof of the Theorem.
\hfill $\Box$
\section{Proof of Corollary $\bf{~\ref{cor:quarto}}$}
$\mathop{}\!\mathit{Proof}.$
We start considering the dimension of $H(X)_{\nu_{G},k\nu_{T}}$:
$$\mathop{}\!\mathrm{dim}{\left(H(X)_{\nu_{G},k\nu_{T}}\right)}=\int_{X}\widetilde{\Pi}_{\nu_{G},k\nu_{T}}(x,x)dV_{X}(x).$$
Now let us observe that $\widetilde{\Pi}_{\nu_{G},k\nu_{T}}(x,x)$ is naturally $S^{1}$-invariant as a function of $x$, and therefore descends to a function
on $M$, that we shall denote by $\widetilde{\Pi}_{\nu_{G},k\nu_{T}}(m,m)$ with abuse of language. Thus by integrating first along the fibers the previous integral may be naturally interpreted as an integral over $M$, that we shall write in the form:
$$\mathop{}\!\mathrm{dim}{\left(H(X)_{\nu_{G},k\nu_{T}}\right)}=\int_{M}\widetilde{\Pi}_{\nu_{G},k\nu_{T}}(m,m)dV_{M}(m).$$
Now by the above $\widetilde{\Pi}_{\nu_{G},k\nu_{T}}(m,m)$ is rapidly decreasing away from a shrinking neighborhood of $M_{0,\nu_{T}}$. So, using a smoothly varying system of adapted coordinates centered at points $m\in M_{0,\nu_{T}}$, we can locally parametrize a neighborhood $U$ of $M_{0,\nu_{T}}$ in the form $m+\mathop{}\!\mathrm{v}$, where $m\in M_{0,\nu_{T}}$ and
$\mathop{}\!\mathrm{v}\in N_{m}$. This parametrization is only valid locally in $m$, since we may not expect to find a single $\mathcal{C}^\infty$ family of adapted coordinates $\gamma_{m}$ ($m\in M_{0,\nu_{T}}$). Hence to make this argument strictly rigorous we should introduce a partition of unity on $M_{0,\nu_{T}}$ subordinate to an appropriate open cover. However, we shall simplify notation and leave this point implicit.
\begin{equation*}
\mathop{}\!\mathrm{dim}{\left(H(X)_{\nu_{G},k\nu_{T}}\right)}=\int_{M_{0,\nu_{T}}}\int_{\mathbb{R}^{d_{P}-1}}\widetilde{\Pi}_{\nu_{G},k\nu_{T}}\left(m+\mathop{}\!\mathrm{v},m+\mathop{}\!\mathrm{v}\right)d\mathop{}\!\mathrm{v} dV_{M}(m).
\end{equation*}
In view of Theorem $~\ref{teo:secondo}$ the asymptotics of the previous integral are unchanged, if the integrand is multiplied by a cut-off of the form $\varrho \big(k^{\frac{7}{18}}\|\mathop{}\!\mathrm{v}\|\big)$, where $\varrho\in \mathcal{C}^\infty_{0}(\mathbb{R})$ is identically equal to $1$ in some neighborhood of $0$.
\begin{equation*}
\mathop{}\!\mathrm{dim}{\left(H(X)_{\nu_{G},k\nu_{T}}\right)}=\int_{M_{0,\nu_{T}}}\int_{\mathbb{R}^{d_{P}-1}}\widetilde{\Pi}_{\nu_{G},k\nu_{T}}\left(m+\mathop{}\!\mathrm{v},m+\mathop{}\!\mathrm{v}\right)\varrho \big(k^{\frac{7}{18}}\|\mathop{}\!\mathrm{v}\|\big)d\mathop{}\!\mathrm{v} dV_{M}(m).
\end{equation*}
Let us now operate the rescaling $\mathop{}\!\mathrm{v}=\frac{\mathop{}\!\mathrm{u}}{\sqrt{k}}$. We can now make use of the asymptotic expansion in Theorem $~\ref{teo:diagonal}$, with $\mathop{}\!\mathrm{u}=\mathop{}\!\mathrm{u}_{t}$ (that is, $\mathop{}\!\mathrm{u}_{v}=\mathop{}\!\mathrm{u}_{h}=0$). We obtain:
\begin{equation}
\label{telufficio5731}
\begin{multlined}[t][12.5cm]
\mathop{}\!\mathrm{dim}{\left(H(X)_{\nu_{G},k\nu_{T}}\right)}=k^{-\frac{d_{P}-1}{2}}\int_{M_{0,\nu_{T}}}\int_{\mathbb{R}^{d_{P}-1}}\widetilde{\Pi}_{\nu_{G},k\nu_{T}}\left(m+\frac{\mathop{}\!\mathrm{u}}{\sqrt{k}},m+\frac{\mathop{}\!\mathrm{u}}{\sqrt{k}}\right)\varrho \big(k^{-\frac{1}{9}}\|\mathop{}\!\mathrm{u}\|\big)d\mathop{}\!\mathrm{u} dV_{M}\\
=k^{-\frac{d_{P}-1}{2}}\cdot\int_{M_{0,\nu_{T}}}\frac{2^{\frac{d_{G}}{2}}d_{\nu_{G}}^{2}}{(\sqrt{2})^{d_{T}-1}\pi^{d_{T}-1}}\left(\frac{\|\nu_{T}\|k}{\pi}\right)^{d_{M}-\frac{d_{P}-1}{2}}\cdot\\
\cdot\int_{\mathbb{R}^{d_{P}-1}}\frac{1}{\|\Phi_{T}\|^{d_{M}-\frac{d_{P}-1}{2}+1}\det{C(m)}}e^{-\lambda_{\nu_{T}}2\|\mathop{}\!\mathrm{u}\|^2}\varrho \big(k^{-\frac{1}{9}}\|\mathop{}\!\mathrm{u}\|\big)d\mathop{}\!\mathrm{u} dV_{M}(m) + \cdots,
\end{multlined}
\end{equation}
\noindent
where $d_{P}=d_{G}+d_{T}$ and the dots denote lower order terms.
Now we evaluate the Gaussian integral, let us operate the change of variables $q=\sqrt{2\lambda_{\nu_{T}}}\mathop{}\!\mathrm{u}$, we have that:
\begin{equation}
\label{boiaboiagausss}
\begin{multlined}[t][12.5cm]
\int_{\mathbb{R}^{d_{P}-1}} e^{-2\lambda_{\nu_{T}}\|\mathop{}\!\mathrm{u}\|^{2}}d\mathop{}\!\mathrm{u}=\frac{1}{\lambda_{\nu_{T}}^{\frac{d_{P}-1}{2}}(\sqrt{2})^{d_{P}-1}}\int_{\mathbb{R}^{d_{P}-1}} e^{-\|q\|^{2}}dq\\
=\frac{\|\Phi_{T}(m)\|^{\frac{d_{P}-1}{2}}\pi^{\frac{d_{P}-1}{2}}}{(\sqrt{2})^{d_{P}-1}\|\nu_{T}\|^{\frac{d_{P}-1}{2}}}
\end{multlined}
\end{equation}
\noindent
and substituting in $(\ref{telufficio5731})$ we obtain the following expression:
\begin{equation*}
\begin{multlined}[t][12.5cm]
\mathop{}\!\mathrm{dim}{\left(H(X)_{\nu_{G},k\nu_{T}}\right)}=\frac{d_{\nu_{G}}^{2}}{2^{d_{T}-1}\pi^{d_{T}-1}}\left(\frac{\|\nu_{T}\|k}{\pi}\right)^{d_{M}-d_{P}+1}\cdot\\
\cdot\int_{M_{0,\nu_{T}}}\frac{\|\Phi_{T}(m)\|^{-d_{M}+d_{P}-2}}{\det{C(m)}}dV_{M}(m)+ \cdots.
\end{multlined}
\end{equation*}
The proof is complete.
\hfill $\Box$
\section{Proof of Proposition $\bf{~\ref{prop:complement}}$}
$\mathop{}\!\mathit{Proof}.$
We assume that $\mathop{}\!\mathrm{dist}_{X}{(y,p\cdot x)}\geq Dk^{\varepsilon-1/2}$, for every $p\in P$. The method consist to use iteratively integration by parts to deduce the rapidly decreasing behavior of the kernel. We start following the previous situations using the standard representation as Fourier integral operator. So, by performing the change of variables $t\mapsto kt$, we obtain the following expression for the Szeg\"{o} kernel:
\begin{equation}
\label{geometriainbicocca2015}
\begin{multlined}[t][12.5cm]
\widetilde{\Pi}_{\nu_{G},k\nu_{T}}\left(x,y\right)=\frac{kd_{\nu_{G}}}{(2\pi)^{d_{T}}}\int_{G}\int_{T}\int_{0}^{+\infty}\chi_{\nu_{G}}(g^{-1})e^{ik\Psi(t,\vartheta,x)}\cdot\\
\cdot s\left(\tilde{\mu}^{G}_{g^{-1}}\circ\tilde{\mu}^{T}_{-\vartheta}\left(x\right),y,kt\right)dtdgd\vartheta,
\end{multlined}
\end{equation}
\noindent
where
\begin{equation}
\label{geometriainbicocca20152}
\Psi(t,\vartheta,x)=t\psi\left(\tilde{\mu}^{G}_{g^{-1}}\circ\tilde{\mu}^{T}_{-\vartheta}\left(x\right),y\right)-\nu_{T}\cdot \vartheta
\end{equation}
\noindent
is the phase of the oscillatory integral. First we observe that:
$$ \|\partial_{\vartheta}\Psi\|=\|t\psi\left(\tilde{\mu}^{G}_{g^{-1}}\circ\tilde{\mu}^{T}_{-\vartheta}\left(x\right),y\right)-\nu_{T}\|\geq C$$
\noindent
for $0<t\ll 1$ and $C>0$. In a similar way
$$ \|\partial_{\vartheta}\Psi\|\geq C_{1}t $$
\noindent
for $t\gg 0$ and $C_{1}>0$. Using integration by parts in $d\vartheta$, the asymptotics for $k\rightarrow +\infty$ is unchanged. We multiply the integrand by $\gamma(t)$, where $\gamma\in \mathcal{C}^{\infty}_{0}\left(\frac{1}{2D},2D\right)$, $\gamma\equiv 1$ on $\left(\frac{1}{D},D\right)$ and $\gamma \equiv 0$ outside of $\left(\frac{1}{2D},2D\right)$. The integral now is compactly supported in $dt$. Taking the partial derivative respect $t$, we deduce that:
$$\partial_{t}\Psi(t,\vartheta,x)=\psi\left(\tilde{\mu}^{G}_{g^{-1}}\circ\tilde{\mu}^{T}_{-\vartheta}\left(x\right),y\right)$$
\noindent
and by the assumption we find that:
$$\left\|\partial_{t}\Psi(t,\vartheta,x)\right\|=\left|\psi\left(\tilde{\mu}^{G}_{g^{-1}}\circ\tilde{\mu}^{T}_{-\vartheta}\left(x\right),y\right)\right|\geq \Im{\psi\left(\tilde{\mu}^{G}_{g^{-1}}\circ\tilde{\mu}^{T}_{-\vartheta}\left(x\right),y\right)}\geq D'k^{2\varepsilon-1}.$$
Now we introduce the differential operator:
$$ L_{t}=\left[\psi\left(\tilde{\mu}^{G}_{g^{-1}}\circ\tilde{\mu}^{T}_{-\vartheta}\left(x\right),y\right)\right]^{-1}\partial_{t} $$
\noindent
and observing that $e^{ik\Psi}=-\frac{i}{k}L_{t}\left(e^{ik\Psi}\right)$ we can apply iteratively the integration by parts. So step by step we obtain a factor of order $O(k^{-2N\varepsilon})$.
The proof is complete.
\hfill $\Box$
\section{Proof of Theorem $\bf{~\ref{teo:app2}}$}
$\mathop{}\!\mathit{Proof}.$
In view of the equality on the first line of $(\ref{toeplitz})$ we observe that $1)$ follows immediately from the point $1)$ of the main Theorem.
Let us now consider the proof of $2)$.
Let $f\in C^\infty(M)$, we consider the associated Toeplitz operator:
\begin{equation}
\label{Toeplitz1}
\begin{multlined}[t][12.5cm]
T_{\nu_{G},k\nu_{T}}[f]\left(x+\frac{\mathop{}\!\mathrm{v}}{\sqrt{k}},x+\frac{\mathop{}\!\mathrm{v}}{\sqrt{k}}\right)\\=\int_{X}\widetilde{\Pi}_{\nu_{G},k\nu_{T}}\left(x+\frac{\mathop{}\!\mathrm{v}}{\sqrt{k}},y\right)f(y)\widetilde{\Pi}_{\nu_{G},k\nu_{T}}\left(y,x+\frac{\mathop{}\!\mathrm{v}}{\sqrt{k}}\right)dV_{X}(y)
\end{multlined}
\end{equation}
\noindent
with $x\in X_{0,\nu_{T}}$, where $f(y)=f(\pi(y))$. Now in view of Proposition $\ref{prop:complement}$ only a shrinking neighborhood of the orbit $P\cdot x$ contributes non-negligibly to the asymptotics. Therefore, the asymptotics are unchanged if the integrand in $(\ref{Toeplitz1})$ is multiplied by a cut-off function $\varrho_{k}(y)$, where $\varrho_{k}=1$ for $\mathop{}\!\mathrm{dist}_{X}(y,P\cdot x)\leq Dk^{\delta-1/2}$ (for example concretely $\delta$ equal $1/9$) and $\varrho_{k}=0$ for $\mathop{}\!\mathrm{dist}_{X}{(y,P\cdot x)}\geq 2Dk^{\delta-1/2}$. We shall make a more explicit choice of $\varrho_{k}$ below.
Let $x+(\theta,\mathop{}\!\mathrm{v})$ be a system of Heisenberg local coordinates on $X$ centered at $x$. This determines for every $p\in P$ a system of HLC centered at $p\cdot x$, by setting
$$p\cdot x+(\theta,\mathop{}\!\mathrm{v})= p\cdot \big(x+(\theta,\mathop{}\!\mathrm{v})\big).$$
In this manner we have a unique smoothly varying family of HLC systems centered at points of $P\cdot x$, and identifications $T_{p\cdot x}X\cong T_{x}X\cong
\mathbb{R}\times\mathbb{R}^{2d_{M}}$, $T_{p\cdot m}M\cong T_{m}M\cong \mathbb{R}\times\mathbb{R}^{2d_{M}}$. Furthermore, the action of $P$ preserves the contact and CR structures of $M$, and the decomposition of the tangent spaces in $h$-, $v$-, and $t$-components. This means that the corresponding decomposition is preserved under the identification $T_{p\cdot m}M\cong T_{m}M$. If $y\in P\cdot x$, let $N^{P}_{y}$ be the normal space to $P\cdot x$ in $X$ at $y$; then we have natural unitary isomorphisms $N^{P}_{y}\cong N^{P}_{x}$.
With this identification implicit and some abuse of language, we may then parametrize a suitably small open neighborhood of $P\cdot x$ by the map
$$(p,n)\in P\times N^{P}_{x}\mapsto p\cdot x+n.$$
We set $y=p\cdot \left(x+n\right)$, where $p=(g,t)\in P$ and $n$ is a tangent vector normal to the orbit.
For simplicity we suppose that the stabilizator of $x$ in $P$ is trivial. We have a diffeomorphism $P\times N(\varepsilon)\rightarrow X'$, with $X'$ a $\varepsilon$-tubular neighborhood of $P\cdot x$ in $X$ and $N(\varepsilon)$ is a ball of radious $\varepsilon$ in the normal space of the orbit in $x$ (a real vector space of dimension $2d_{M}+1-d_{P}$). This diffeomorphism doesn't preserve the volume form. Thus in coordinates $(p,n)\in P\times N_{x}^{P}$ (where $N_{x}^{P}$ denote the normal space to $P\cdot x$ in $x$):
\begin{equation}
\label{newcoordinates}
dV_{X}(y)=\mathfrak{D}(p,n)dV_{P}(p)dn,
\end{equation}
\noindent
where $dV_{P}(p)=\frac{d\vartheta}{(2\pi)^{d_{T}}}dV_{G}(g)$ is the Haar measure of $P$, $dn=d\mathcal{L}(n)$ the Lebesgue measure on $N_{x}^{P}$ (unitarily identified with $\mathbb{R}^{2d_{M}+1-d_{P}}$) and $\mathfrak{D}(p,0)=\mathfrak{R}_{x}(p)$ with $\mathfrak{R}_{x}:P\rightarrow \mathbb{R}_{>0}$ a distortive function defined as follow.
Let $\mathcal{B}_{0}$ an orthonormal basis of $\mathfrak{p}$ and let $\mathop{}\!\mathrm{val}_{x}:\mathfrak{p}\rightarrow T_{x}X$ the valutation map. Let $D_{x}(p)$ the matrix associated to $\mathop{}\!\mathrm{val}_{x}^{\vee}(g_{X}):\mathfrak{p}\times \mathfrak{p}\rightarrow \mathbb{R}$ respect to $\mathcal{B}_{0}$ and
\begin{equation}
\label{distortionf}
\mathfrak{R}_{x}(p)=\sqrt{\det{D_{x}(p)}},
\end{equation}
\noindent
here $g_{X}$ is the Riemannian metric on $X$ and $\mathop{}\!\mathrm{val}_{x}^{\vee}(g_{X})$ is the pull back of such metric to $\mathfrak{p}=\mathop{}\!\mathrm{Lie}{(P)}$ using the valutation
$\mathop{}\!\mathrm{val}_{x}:\gamma \mapsto \gamma_{X}(x)$.
We observe that $\mathfrak{R}_{x}(p)$ is constant along the orbit, because $P$ acts by isometries.
We put $\mathfrak{R}_{x}(p)=r_{x}$.
We obtain that:
\begin{equation}
\label{Toeplitz2}
\begin{multlined}[t][12.5cm]
T_{\nu_{G},k\nu_{T}}[f]\left(x+\frac{\mathop{}\!\mathrm{v}}{\sqrt{k}},x+\frac{\mathop{}\!\mathrm{v}}{\sqrt{k}}\right)\\
\\
=\int_{X}\widetilde{\Pi}_{\nu_{G},k\nu_{T}}\left(x+\frac{\mathop{}\!\mathrm{v}}{\sqrt{k}},y\right)f(y)\widetilde{\Pi}_{\nu_{G},k\nu_{T}}\left(y,x+\frac{\mathop{}\!\mathrm{v}}{\sqrt{k}}\right)\varrho_{k}(y)dV_{X}(y)\\
=\int_{P}\int_{N_{x}}\widetilde{\Pi}_{\nu_{G},k\nu_{T}}\left(x+\frac{\mathop{}\!\mathrm{v}}{\sqrt{k}},p\left(x+n\right)\right)f\left(p\cdot\left(x+n\right)\right)\cdot\\
\cdot\widetilde{\Pi}_{\nu_{G},k\nu_{T}}\left(p\left(x+n\right),x+\frac{\mathop{}\!\mathrm{v}}{\sqrt{k}}\right)\mathfrak{D}(p,n)\varrho_{k}\left(p\cdot\left(x+n\right)\right)dV_{P}(p)dn.
\end{multlined}
\end{equation}
Rescaling $n\mapsto \frac{n}{\sqrt{k}}$, we observe that $\mathop{}\!\mathrm{Rank}{(N)}=\mathop{}\!\mathrm{dim}{X}-\mathop{}\!\mathrm{dim}{P}=2d_{M}+1-d_{P}=2\left[d_{M}+\frac{1-d_{P}}{2}\right]$ from which we have that
$dn \rightarrow k^{-\left[d_{M}+\frac{1-d_{P}}{2}\right]}dn$ and we obtain that:
\begin{equation}
\label{Toeplitz3}
\begin{multlined}[t][12.5cm]
T_{\nu_{G},k\nu_{T}}[f]\left(x+\frac{\mathop{}\!\mathrm{v}}{\sqrt{k}},x+\frac{\mathop{}\!\mathrm{v}}{\sqrt{k}}\right)\\ \\=k^{-\left[d_{M}+\frac{1-d_{P}}{2}\right]}\int_{X}\widetilde{\Pi}_{\nu_{G},k\nu_{T}}\left(x+\frac{\mathop{}\!\mathrm{v}}{\sqrt{k}},y\right)f(y)\widetilde{\Pi}_{\nu_{G},k\nu_{T}}\left(y,x+\frac{\mathop{}\!\mathrm{v}}{\sqrt{k}}\right)\varrho_{k}(y)dV_{X}(y)\\
=k^{-\left[d_{M}+\frac{1-d_{P}}{2}\right]}\int_{P}\int_{N_{x}}\widetilde{\Pi}_{\nu_{G},k\nu_{T}}\left(x+\frac{\mathop{}\!\mathrm{v}_{1}}{\sqrt{k}},p\left(x+\frac{n}{\sqrt{k}}\right)\right)f\left(p\left(x+\frac{n}{\sqrt{k}}\right)\right)\cdot\\
\cdot\widetilde{\Pi}_{\nu_{G},k\nu_{T}}\left(p\left(x+\frac{n}{\sqrt{k}}\right),x+\frac{\mathop{}\!\mathrm{v}}{\sqrt{k}}\right)\mathfrak{D}\left(p,\frac{n}{\sqrt{k}}\right)\varrho_{k}\left(p\cdot \left(x+\frac{n}{\sqrt{k}}\right)\right)dV_{P}(p)dn.
\end{multlined}
\end{equation}
Here $\varrho_k(p\cdot x+n)=\varrho_k\left(p\cdot x+k^{1/2-\varepsilon}n\right)$ and $\varepsilon=1/9$. $n\in N_{x}^{P}$ by construction and $v\in N^{P}_{x}$ by assumption. We have that:
$$N_{x}^{P}=[\mathop{}\!\mathrm{val}_{x}(\mathop{}\!\mathrm{Ker}{(\Phi_{P}(m))})\oplus\mathop{}\!\mathrm{span}{(\eta_{X}(x))}]^{\perp}.$$
Let $\eta\in \mathfrak{t}$ be the unique element such that
$$
\eta\in \ker\Phi_{T}(m)^\perp,\,\,\,\,\langle \Phi_{T}(m),\eta\rangle=\|\Phi_{T}(m)\|.
$$
In particular we have $\eta\in \mathfrak{t}$, and $\eta$ has unit norm. Then
\begin{equation}
\label{HLCinx}
\eta_{X}(x)=(\|\Phi_{P}(m)\|,-\eta_{M}(m)).
\end{equation}
So in terms of the isomorphism $T_{x}X\cong \mathbb{R}\times T_{m}M$ we have:
\begin{equation*}
p_{X}(x)=(\{0\}\times\mathop{}\!\mathrm{val}_{m}(\mathop{}\!\mathrm{Ker}{(\Phi_{P}(m))}))\oplus\mathop{}\!\mathrm{span}_{\mathbb{R}}{((\|\Phi(m)\|,-\eta_{M}(m)))},
\end{equation*}
\noindent
and so in the notation of $(\ref{labb})$ and $(\ref{antoniosalierii})$
\begin{equation*}
\begin{multlined}[t][12.5cm]
p_{X}(x)^{\perp}=(\{0\}\times V_{m})^{\perp}\cap\mathop{}\!\mathrm{span}_{\mathbb{R}}{(\|\Phi(m)\|,-\eta_{M}(m))}^{\perp}\\
=\left[\mathbb{R}\times(H_{m}\oplus N_{m})\right]\cap[\mathop{}\!\mathrm{span}_{\mathbb{R}}{((\|\Phi(m)\|,-\eta_{M}(m)))}]^{\perp},
\end{multlined}
\end{equation*}
\noindent
with $V_{m}=\mathop{}\!\mathrm{val}_{m}(\mathop{}\!\mathrm{Ker}{(\Phi_{P}(m))})$, $N_{m}=J_{m}(V_{m})$ and $H_{m}=(V_{m}\oplus N_{m})^{\perp}$.
We have that $\eta_{Mt}(m)=0$, because $M_{0,\nu_{T}}$ is $P$-invariant. Thus if $(\lambda,\mathop{}\!\mathrm{h}+\mathop{}\!\mathrm{t})\in \mathbb{R}\times(H_{m}\oplus N_{m})$ we have:
\begin{equation*}
g_{X}((\lambda,\mathop{}\!\mathrm{h}+\mathop{}\!\mathrm{t}),(\|\Phi(m)\|,-\eta_{M}(m)))=\lambda\|\Phi_{P}(m)\|-g_{X}(\mathop{}\!\mathrm{h},\eta_{Mh}(m)).
\end{equation*}
Denoting $\Phi=\Phi_{P}(m)$ and $\eta_{M}=\eta_{M}(m)$ we have:
\begin{equation}
\label{normalspaaace}
\begin{multlined}[t][12.5cm]
N_{x}^{P}=\left\{(\lambda,\mathop{}\!\mathrm{h}+\mathop{}\!\mathrm{t})\in\mathbb{R}\times(H_{m}\oplus N_{m}): \lambda=\frac{g_{X}(\mathop{}\!\mathrm{h},\eta_{Mh})}{\|\Phi\|}\right\}\\
=[\{0\}\times N_{m}]\oplus\left\{\left(\frac{1}{\|\Phi\|}g_{X}(\mathop{}\!\mathrm{h},\eta_{Mh}),\mathop{}\!\mathrm{h}\right):\mathop{}\!\mathrm{h}\in H_{m}\right\}.
\end{multlined}
\end{equation}
We write $v=n_{1}=\left(\frac{1}{\|\Phi\|}g_{X}(\mathop{}\!\mathrm{h}_{1},\eta_{Mh}),\mathop{}\!\mathrm{h}_{1}+\mathop{}\!\mathrm{t}_{1}\right)$, $n=\left(\frac{1}{\|\Phi\|}g_{X}(\mathop{}\!\mathrm{h},\eta_{Mh}),\mathop{}\!\mathrm{h}+\mathop{}\!\mathrm{t}\right)$. Recalling that $\Pi(x_1,x_2)=\overline{\Pi(x_2,x_1)}$ we obtain:
\begin{equation}
\label{Toeplitz4}
\begin{multlined}[t][12.5cm]
T_{\nu_{G},k\nu_{T}}[f]\left(x+\frac{n_{1}}{\sqrt{k}},x+\frac{n_{1}}{\sqrt{k}}\right)=\\
k^{-\left[d_{M}+\frac{1-d_{P}}{2}\right]}\int_{P}\int_{N_{x}}\left|\widetilde{\Pi}_{\nu_{G},k\nu_{T}}\left(x+\frac{n_{1}}{\sqrt{k}},p\left(x+\frac{n}{\sqrt{k}}\right)\right)\right|^{2}f\left(p\left(x+\frac{n}{\sqrt{k}}\right)\right)\cdot\\
\cdot\mathfrak{D}\left(p,\frac{n}{\sqrt{k}}\right)\varrho_{k}\left(p\cdot \left(x+\frac{n}{\sqrt{k}}\right)\right)dV_{P}(p)dn.
\end{multlined}
\end{equation}
Let us now make use of the asymptotic expansion of $\widetilde{\Pi}_{\nu_{G},k\nu_{T}}$ from point 3) of the Theorem $\ref{teo:secondo}$ and, using the Taylor expansion for $f\left(p\cdot x+\frac{n}{\sqrt{k}}\right)$ and for $\mathfrak{D}$
\begin{equation}
\label{Toeplitz6}
\begin{multlined}[t][12.5cm]
T_{\nu_{G},k\nu_{T}}[f]\left(x+\frac{n_{1}}{\sqrt{k}},x+\frac{n_{1}}{\sqrt{k}}\right)\sim k^{\left[d_{M}+\frac{1-d_{P}}{2}\right]}r_{x}\cdot f\left(p\cdot x\right)C(m,\nu_{P})\\
\int_{P}\int_{N_{x}}\left|\chi_{\nu_{P}}\left(p\right)\right|^2\cdot e^{H(n_{1},n)+H(n,n_{1})}\varrho_{k}\left(p\cdot \left(x+\frac{n}{\sqrt{k}}\right)\right)dV_{P}(p)dn(1+\cdots)
\end{multlined}
\end{equation}
\noindent
where we have set:
$$ C(m,\nu_{P})=\left[\frac{d_{\nu_{G}}2^{\frac{d_{G}}{2}}}{(\sqrt{2}\pi)^{d_{T}-1}}
\cdot\frac{\left(\frac{\|\nu_{T}\|}{\pi}\right)^{d_{M}+\frac{1-d_{P}}{2}}}{\mathcal{D}(m)\|\Phi_{T}\|^{d_{M}+1+\frac{1-d_{P}}{2}}}\right]^2 $$
\noindent
and the dots stand for terms of less degree.
Here the exponent is as follows: let
$$n=\mathop{}\!\mathrm{v}+\mathop{}\!\mathrm{t}+\mathop{}\!\mathrm{h},\,\,\,\,\, n_1=\mathop{}\!\mathrm{v}_{1}+\mathop{}\!\mathrm{t}_{1}+\mathop{}\!\mathrm{h}_{1}$$
be the decomposition as in $(\ref{labb})$ and $(\ref{antoniosalierii})$. We call $R_{x}=N_{x}^{P}$ and $r,r_{1}$ in place of $n_{1},n$. Then
\begin{equation}
\label{exxponent1}
\begin{multlined}[t][12.5cm]
H(r_{1},r)+H(r,r_{1})=-2\lambda_{\nu_{T}}\left(\|\mathop{}\!\mathrm{t}_{1}\|^2+\|\mathop{}\!\mathrm{t}\|^2\right)-\lambda_{\nu_{T}}\left\|\mathop{}\!\mathrm{h}_{1}-\mathop{}\!\mathrm{h}-\frac{(\mathop{}\!\mathrm{h}-\mathop{}\!\mathrm{h}_{1},\eta_{Mh})}{\|\Phi_{T}(m)\|}\eta_{Mh}\right\|^2.
\end{multlined}
\end{equation}
We evaluate the Gaussian integral $\int_{R_{x}}e^{H(r_{1},r)+H(r,r_{1})}dr$. So we have that:
\begin{equation}
\label{Gaussian11}
\int_{R_{x}}e^{H(r_{1},r)+H(r,r_{1})}dr=e^{-2\lambda_{\nu_{T}}\|\mathop{}\!\mathrm{t}_{1}\|^2}\int_{N_{m}}e^{-2\lambda_{\nu_{T}}\|\mathop{}\!\mathrm{t}\|^{2}}d\mathop{}\!\mathrm{t}\cdot\int_{H_{m}}e^{-\lambda_{\nu_{T}}\left\|\mathop{}\!\mathrm{h}_{1}-\mathop{}\!\mathrm{h}-\frac{(\mathop{}\!\mathrm{h}-\mathop{}\!\mathrm{h}_{1},\eta_{Mh})}{\|\Phi_{T}(m)\|}\eta_{Mh}\right\|^2}d\mathop{}\!\mathrm{h}.
\end{equation}
Let $d_{N_{m}}=d_{P}-1$ the dimension of $N_{m}$. Let us first consider the first Gaussian integral in $(\ref{Gaussian11})$.
\begin{equation}
\label{Gaussian22}
\int_{N_{m}}e^{-2\lambda_{\nu_{T}}\|\mathop{}\!\mathrm{t}\|^{2}}d\mathop{}\!\mathrm{t}=\frac{1}{\left(2\sqrt{\lambda_{\nu_{T}}}\right)^{d_{N_{m}}}}\int_{N_{m}}e^{-\frac{1}{2}\|s\|^{2}}ds=\left(\frac{\pi}{2\lambda_{\nu_{T}}}\right)^{(d_P-1)/2}.
\end{equation}
To compute the second Gaussian integral in $(\ref{Gaussian11})$, let us operate the change of variable
\begin{equation}
\label{samsung44}
-\mathop{}\!\mathrm{h}_{1}+\mathop{}\!\mathrm{h}+\frac{(\mathop{}\!\mathrm{h}-\mathop{}\!\mathrm{h}_{1},\eta_{Mh})}{\|\Phi_{T}(m)\|}\eta_{Mh}=\mathop{}\!\mathrm{w}.
\end{equation}
We differentiate the previous expression $d\mathop{}\!\mathrm{h}=\det{\left(\frac{\partial \mathop{}\!\mathrm{w}}{\partial \mathop{}\!\mathrm{h}}\right)}^{-1}d\mathop{}\!\mathrm{w}$ and in order to determine $\det{\left(\frac{\partial \mathop{}\!\mathrm{w}}{\partial \mathop{}\!\mathrm{h}}\right)}$ we observe that:
$$\det{\left(\frac{\partial \mathop{}\!\mathrm{w}}{\partial \mathop{}\!\mathrm{h}}\right)}=1+\frac{\|\eta_{Mh}\|^2}{\|\Phi_{T}\|}.$$
So we have:
\begin{equation}
\label{Gaussian33}
\begin{multlined}[t][12.5cm]
\int_{H_{m}}e^{-\lambda_{\nu_{T}}\left\|\mathop{}\!\mathrm{h}_{1}-\mathop{}\!\mathrm{h}-\frac{(\mathop{}\!\mathrm{h}-\mathop{}\!\mathrm{h}_{1},\eta_{Mh})}{\|\Phi_{T}(m)\|}\eta_{Mh}\right\|^2}d\mathop{}\!\mathrm{h}\\
=\frac{\|\Phi_{T}\|}{\|\Phi_{T}\|+\|\eta_{Mh}\|^2}\int_{H_{m}}e^{-\frac{1}{2}\|s\|^2}\frac{ds}{\left(2\lambda_{\nu}\right)^{\frac{d_{H_{m}}}{2}}}=\frac{\|\Phi_{T}\|}{\|\Phi_{T}\|+\|\eta_{Mh}\|^2}\left(\frac{\pi}{\lambda}\right)^{d_{H_{m}}/2},
\end{multlined}
\end{equation}
\noindent
with $d_{H_{m}}=2(d_{M}+1-d_{P})$ the dimension of $H_{m}$.
Inserting $(\ref{Gaussian33})$ and $(\ref{Gaussian22})$ in $(\ref{Gaussian11})$,
we obtain
\begin{equation}
\label{Gaussianfinalll}
\begin{multlined}[t][12.5cm]
\int_{R_{x}}e^{H(r_{1},r)+H(r,r_{1})}dr\\
=e^{-2\lambda_{\nu_{T}}\|\mathop{}\!\mathrm{t}_{1}\|^2}\cdot\frac{\|\Phi_{T}\|}{\|\Phi_{T}\|+\|\eta_{Mh}\|^2}\left(\frac{\pi}{\lambda_{\nu_{T}}}\right)^{\frac{d_{H_{m}}+d_{N_{m}}}{2}}2^{-\frac{d_{N_{m}}}{2}}.
\end{multlined}
\end{equation}
Let us now insert this in $(\ref{Toeplitz6})$. We obtain a leading term depending on $r_x$ and $f(x)$. We can then determine $r_x$ using that for $f=1$ this must reduce to the leading term in 2) of Theorem $\ref{teo:secondo}$. Thus, recalling that $\int_{P}\left|\chi_{\nu_{P}}\left(p\right)\right|^2dV_{P}(p)=1$ (see $\cite{ventisettesimo}$ Theorem 4.11) and noting that $d_{H_{m}}+d_{N_{m}}=2d_{M}+1-d_P$ we obtain that the leading order term in $(\ref{Toeplitz6})$ is given by:
\begin{equation}
\label{costanteperduta}
\begin{multlined}[t][12.5cm]
k^{\left[d_{M}+\frac{1-d_{P}}{2}\right]}\cdot C(m,\nu_{P})\cdot\left(\frac{\pi}{\lambda_{\nu_{T}}}\right)^{\frac{d_{H_{m}}+d_{N_{m}}}{2}}\cdot 2^{-\frac{d_{N_{m}}}{2}}\\
\cdot e^{-2\lambda_{\nu_{T}}\|\mathop{}\!\mathrm{t}_{1}\|^2}\cdot\frac{\|\Phi_{T}\|}{\|\Phi_{T}\|+\|\eta_{Mh}\|^2} \cdot r_{x}=\frac{1}{(\sqrt{2}\pi)^{d_{T}-1}}d_{\nu_{G}}2^{\frac{d_{G}}{2}}\left(\frac{k}{\pi}\|\nu_{T}\|\right)^{d_{M}-\frac{d_{P}}{2}+\frac{1}{2}}e^{-2\lambda_{\nu_{T}}\|\mathop{}\!\mathrm{t}_{1}\|^2}\cdot\\
\cdot \frac{1}{\mathcal{D}(m)}\cdot\frac{1}{\|\Phi_{T}\|^{d_{M}+1-\frac{d_{P}}{2}+\frac{1}{2}}},
\end{multlined}
\end{equation}
\noindent
and we find that
\begin{equation}
\label{costanteperduta2}
\begin{multlined}[t][12.5cm]
r_{x}=\frac{\pi^{d_{T}-1}\mathcal{D}(m)(\|\Phi_{T}\|+\|\eta_{Mh}\|^2)}{d_{\nu_{G}}\left(\sqrt{2}\right)^{-d_{G}-d_{T}+1-d_{N_{m}}}}.
\end{multlined}
\end{equation}
The leading term become:
\begin{equation}
\label{Toeplitz7}
\begin{multlined}[t][12.5cm]
\frac{1}{(\sqrt{2}\pi)^{d_{T}-1}}d_{\nu_{G}}2^{\frac{d_{G}}{2}}\left(\frac{k}{\pi}\|\nu_{T}\|\right)^{d_{M}-\frac{d_{P}}{2}+\frac{1}{2}}f(m)e^{-2\lambda_{\nu_{T}}\|\mathop{}\!\mathrm{t}_{1}\|^2}\cdot\\
\cdot \frac{1}{\mathcal{D}(m)}\cdot\frac{1}{\|\Phi_{T}\|^{d_{M}+1-\frac{d_{P}}{2}+\frac{1}{2}}}.
\end{multlined}
\end{equation}
This complete the proof of $2)$ and of the Theorem.
\hfill $\Box$
\section{Proof of Corollary $\bf{~\ref{cor:quinto}}$}
$\mathop{}\!\mathit{Proof}.$
We start considering the trace of $T_{\nu_{G},k\nu_{T}}[f]$:
$$\mathfrak{T}\left(T_{\nu_{G},k\nu_{T}}[f]\right)=\int_{X}T_{\nu_{G},k\nu_{T}}[f](x,x)dV_{X}(x).$$
Now we observe that $T_{\nu_{G},k\nu_{T}}(x,x)$ is rapidly decreasing away from a shrinking neighborhood of $X_{0,\nu_{T}}$. So, using a smoothly varying system of adapted coordinates centered at points $x\in X_{0,\nu_{T}}$, we can locally parametrize a neighborhood $U$ of $X_{0,\nu_{T}}$ in the form $x+\mathop{}\!\mathrm{t}$, where $x\in X_{0,\nu_{T}}$ and
$\mathop{}\!\mathrm{t}\in N_{m}$. This parametrization is only valid locally in $x$. We introduce a partition of unity on $X_{0,\nu_{T}}$ subordinate to an appropriate open cover and we simplify the notation leaving this point implicit.
\begin{equation*}
\mathfrak{T}\left(T_{\nu_{G},k\nu_{T}}[f]\right)=\int_{X_{0,\nu_{T}}}\int_{\mathbb{R}^{d_{P}-1}}T_{\nu_{G},k\nu_{T}}[f]\left(x+\mathop{}\!\mathrm{t},x+\mathop{}\!\mathrm{t}\right)d\mathop{}\!\mathrm{t} dV_{X}(x).
\end{equation*}
In view of Theorem $~\ref{teo:app2}$ the asymptotics of the previous integral are unchanged, if the integrand is multiplied by a cut-off of the form $\varrho \big(k^{\frac{7}{18}}\|\mathop{}\!\mathrm{t}\|\big)$, where $\varrho\in \mathcal{C}^\infty_{0}(\mathbb{R})$ is identically equal to $1$ in some neighborhood of $0$.
\begin{equation*}
\mathfrak{T}\left(T_{\nu_{G},k\nu_{T}}[f]\right)=\int_{X_{0,\nu_{T}}}\int_{\mathbb{R}^{d_{P}-1}}T_{\nu_{G},k\nu_{T}}\left(m+\mathop{}\!\mathrm{v},m+\mathop{}\!\mathrm{v}\right)\varrho \big(k^{\frac{7}{18}}\|\mathop{}\!\mathrm{t}\|\big)d\mathop{}\!\mathrm{t} dV_{X}(x).
\end{equation*}
Let us now operate the rescaling $\mathop{}\!\mathrm{t}=\frac{\mathop{}\!\mathrm{u}}{\sqrt{k}}$. We can now make use of the asymptotic expansion in Theorem $~\ref{teo:app2}$, with $n_{1}=\mathop{}\!\mathrm{u}$. We obtain:
\begin{equation}
\label{telufficio57314}
\begin{multlined}[t][12.5cm]
\mathfrak{T}\left(T_{\nu_{G},k\nu_{T}}[f]\right)=k^{-\frac{d_{P}-1}{2}}\int_{X_{0,\nu_{T}}}\int_{\mathbb{R}^{d_{P}-1}}T_{\nu_{G},k\nu_{T}}\left(x+\frac{\mathop{}\!\mathrm{u}}{\sqrt{k}},x+\frac{\mathop{}\!\mathrm{u}}{\sqrt{k}}\right)\varrho \big(k^{-\frac{1}{9}}\|\mathop{}\!\mathrm{u}\|\big)d\mathop{}\!\mathrm{u} dV_{X}\\
=k^{-\frac{d_{P}-1}{2}}\cdot\frac{2^{\frac{d_{G}}{2}}d_{\nu_{G}}^{2}}{(\sqrt{2})^{d_{T}-1}\pi^{d_{T}-1}}\left(\frac{\|\nu_{T}\|k}{\pi}\right)^{d_{M}-\frac{d_{P}-1}{2}}\cdot\\
\int_{X_{0,\nu_{T}}}\int_{\mathbb{R}^{d_{P}-1}}\frac{f(\pi(x))}{\|\Phi_{T}\|^{d_{M}-\frac{d_{P}-1}{2}+1}\mathcal{D}(\pi(x))}e^{-\lambda_{\nu_{T}}2\|\mathop{}\!\mathrm{u}\|^2}\varrho \big(k^{-\frac{1}{9}}\|\mathop{}\!\mathrm{u}\|\big)d\mathop{}\!\mathrm{u} dV_{X}(x) + \cdots,
\end{multlined}
\end{equation}
\noindent
where $d_{P}=d_{G}+d_{T}$ and the dots denote lower order terms.
Now we evaluate the Gaussian integral, that is the same of $(\ref{boiaboiagausss})$ in the Corollary $\ref{cor:quarto}$.
Substituting the result in $(\ref{telufficio57314})$ we obtain the following expression:
\begin{equation*}
\begin{multlined}[t][12.5cm]
\mathfrak{T}\left(T_{\nu_{G},k\nu_{T}}[f]\right)=\frac{d_{\nu_{G}}^{2}}{2^{d_{T}-1}\pi^{d_{T}-1}}\left(\frac{\|\nu_{T}\|k}{\pi}\right)^{d_{M}-d_{P}+1}\cdot\\
\cdot\int_{X_{0,\nu_{T}}}\frac{f(\pi(x))\|\Phi_{T}(\pi(x))\|^{-d_{M}+d_{P}-2}}{\mathcal{D}(\pi(x))}dV_{X}(x)+ \cdots.
\end{multlined}
\end{equation*}
The proof is complete.
\hfill $\Box$
\begin{center}
\textbf{Acknowledgments}
\end{center}
This article is based on my doctoral dissertation at the University of Milano Bicocca under the supervision of Professor Roberto Paoletti. I am grateful to my advisor for introducing me to this area of mathematics and guiding me patiently. I am also endebted to Professor Andrea Loi for carefully reading this manuscript and suggesting many significants improvements in exposition and organization.
|
\section{Introduction}
The goal of this paper is to formulate the theory of resonant equilibria in quasi-periodic Frenkel-Kontorova models. We argue that these equilibria play an important role in the phenomenon of pinning (the equilibria that survive after an external force is applied). In periodic Frenkel-Kontorova models, the role of resonances in pinning is well established.
We recall that in Frenkel-Kontorova models\cite{BK04, Selke2}, one considers configurations given by a sequence of real numbers (think of the position of a sequence of particles deposited on a 1-D material). The (formal) energy of the system is the sum of a term of interaction between nearest neighbors of the deposited material and a term modeling interaction with the media. In the quasi-periodic Frenkel-Kontorova models studied here, the interacting potential will be a quasi-periodic function of the position reflecting that the medium is quasi-periodic. We will be interested in equilibria, i.e., configurations such that the derivatives of the (formal) energy with respect to the position of each of the particles vanish. We note that even if the energy is a formal sum, the equilibrium equations are well defined. More details of the models will be discussed in Section~\ref{models}.
In \cite{SuL1, SuL2}, one can find a rigorous mathematical theory of
quasi-periodic solutions whose frequency is not resonant (indeed
Diophantine) with the frequencies of the substratum. The rigorous
theory of \cite{SuL1, SuL2} also leads to efficient algorithms that
can compute these quasi-periodic solutions arbitrarily close to their
breakdown. Implementations of these algorithms and investigation of
the phenomena at breakdown appear in \cite{BlassL}. The paper \cite{SuL1},
studies models with nearest neighbor interaction while
\cite{SuL2} studies the case of long range interactions.
Our motivation is to study the phenomena of ``depinning". When we add an external force (no matter how small) to the model, many quasi-periodic solutions disappear. Nevertheless, there are still other quasi-periodic solutions survive. This physically corresponds to the deposited material rearranges itself to withstand the force. It is well known in the periodic Frenkel-Kontorova models that the solutions which persist under forcing are resonant with the media. The papers \cite{SuL1, SuL2} also show that, in the quasi-periodic case, the smooth non-resonant solutions do not exist when there is an external force. Hence we are interested in resonant solutions in the quasi-periodic models.
In this paper, we show that there exist at least two formal power series in the amplitude of the coupling describing resonant solutions even in the presence of external forces. The delicate analytic question of convergence of these series will be pursued in \cite{ZhangSL15}. We also develop a dynamical interpretation of the equilibria as orbits of a dynamical system. Using this dynamical interpretation, we show that the phonon gap of quasi-periodic equilibria is zero.
We hope that this paper can lay the ground work for future explorations. We will use rigorous mathematical tools to study the convergence of the formal series developed here in \cite{ZhangSL15}. We will also use numerical methods to explore in a non-rigorous but more quantative way some of the phenomena discussed here.
Equilibria in quasi-periodic media with a resonant frequency have been investigated numerically in
\cite{vanErp'99, vanErp'01, vanErp'02}. These papers also studied the
phonon gap and found it to vanish when there are smooth solutions
(in agreement with the results here).
The variational and topological methods that have been proved useful in the periodic case do not extend to the quasi-periodic case in its full strength. Several interesting counterexamples are in \cite{LS'03, Federer'74}. Partial results are in \cite{Gambaudo, Aliste10, GPT13, KunzeO13}.
\section{Models considered and formulation of the problem}
\label{models}
We consider models of deposition in a quasi-periodic one-dimensional
medium.
If $x_n\in\mathbb{R}$ denotes the position of the $n$-th particle of the deposited material,
the state of the system is specified by a configuration, i.e. a sequence $\{x_n\}_{n\in\mathbb{Z}}$.
We associate the following formal energy to a configuration of
the system
\begin{equation}\label{formal energy}
\mathscr{S}(x) = \frac{1}{2}\sum_{n\in\mathbb{Z}} (x_{n+1} - x_n - a )^2 -
V(x_n \alpha) - \lambda x_n
\end{equation} where $V:\mathbb{T}^d\rightarrow \mathbb{R}$ is an
analytic function, $\alpha\in \mathbb{R}^d$ is an irrational
vector and $a,\lambda$ are some real numbers.
A natural example often used is
\begin{equation}\label{formal energy:example}
\mathscr{S}(x) = \frac{1}{2}\sum_{n\in\mathbb{Z}} (x_{n+1} - x_n - a )^2 +A\cos x_n+B\cos(\sqrt{2}x_n+\phi)- \lambda x_n
\end{equation} which corresponds to $\alpha=(1,\sqrt{2})$ and $V(\theta_1,\theta_2)=A\cos\theta_1+B\cos(\theta_2+\phi)$.
The term $( x_{n+1} - x_n -a )^2$ represents the interaction among
neighboring deposited atoms. The term $V(x_n\alpha)$ represents the
interaction with the substratum. The interaction at position $x\in\mathbb{R}$ is the quasi-periodic function $V(x\alpha)$. This models that the substratum is quasi-periodic.
The term $\lambda x_n$ has the interpretation of a constant field
applied to the model. In the case of deposited materials, we
can imagine that the sample is tilted and $\lambda$ is
the component of the gravity. The physical meaning of the equilibria corresponding to $\lambda\neq 0$ is that, when we apply a small external field, the configurations rearrange themselves so that they can respond to the force and do not slide. Of course, when the force is strong enough, no rearrangement is possible and all the configurations slide. This is the microscopic origin of static friction.
In the case of periodic media, the pinned solutions are known to correspond to resonant frequencies. Since the resonant tori do not survive, it is reasonable to consider resonant frequencies.
The existence of external forces $\lambda$
is a very important novelty with respect to
the previous papers \cite{SuL1, SuL2}. It was shown in \cite{SuL1, SuL2} that if there is
non-resonant quasi-periodic solution, then
$\lambda = 0$. In our case, we will show how to construct quasi-periodic equilibria with nontrivial $\lambda$ and will show how to compute perturbatively the range of such $\lambda$ for which solutions with a prescribed resonant frequency exist.
Without any loss of generality, we can assume that
\begin{equation}\label{alpha nonresonant}
k\cdot\alpha\notin \mathbb{N}\quad \forall ~k\in \mathbb{Z}^d-\{0\}.
\end{equation}
If there existed a resonance $k\cdot \alpha=0$, we could just use
less frequencies to express the quasi-periodic function.
\subsection{Equilibrium equations}
A configuration is in equilibrium if the forces acting on all the
particles vanish. Equivalently, the derivatives of the energy with
respect to the position of the particles vanish. That is,
\[
\frac{\partial \mathscr{S}}{\partial x_n}(x) = 0 \qquad \forall ~n\in
\mathbb{Z}.
\]
In the model \eqref{formal energy}, the equilibrium equations are
\begin{equation}\label{equilibrium}
x_{n+1} + x_{n-1} - 2 x_n + \partial_\alpha V(x_n\alpha) + \lambda = 0 \quad
\forall~ n\in \mathbb{Z}
\end{equation}
where $\partial_\alpha = \alpha\cdot \nabla$ and $\nabla$ is the usual gradient.
Note that even if the energy \eqref{formal energy} is just a formal
sum, the equilibrium equations \eqref{equilibrium} are well defined
equations.
It is very tempting to consider \eqref{equilibrium} as a dynamical system, so that we obtain $x_{n+1}$ as a function of $x_n$ and $x_{n-1}$. This system has very unusual properties. This will be pursued in Section \ref{sec:dynamical}.
\subsection{Quasi-periodic configurations, hull functions}
\label{sec:hull}
In this paper, we will be interested in quasi-periodic solutions of
frequency $\omega\in\mathbb{R}$.
These are configurations of the form
\begin{equation}\label{hull}
x_n = n\omega + h(n\omega\alpha),
\end{equation}where $h:\mathbb{T}^d\rightarrow \mathbb{R}$.
A configuration given by a hull function \eqref{hull} satisfies the
equilibrium equation \eqref{equilibrium} if and only if the hull
function $h$ satisfies
\begin{equation}\label{hull equilibrium}
h(n\omega \alpha + \omega\alpha) + h(n\omega \alpha - \omega\alpha)
- 2 h(n\omega \alpha) + \partial_\alpha V(n\omega\alpha + \alpha
h(n\omega\alpha)) +\lambda = 0.
\end{equation}
The equation \eqref{hull equilibrium} was considered in
\cite{SuL1,SuL2} when $\omega\alpha$ is Diophantine (in particular,
$n\omega\alpha$ is dense in the torus $\mathbb{T}^d$).
In our case, $n\omega \alpha$ will not be dense on the $d$-dimensional
torus (see Section~\ref{sec:resonances})
and the equilibrium equations we will derive are different from those in \cite{SuL1,SuL2}.
\subsection{Resonances}
\label{sec:resonances}
The goal of this paper is to study situations when there are
$k\in\mathbb{Z}^d-\{0\}$ and $m\in\mathbb{Z}$ such that
\begin{equation}\label{resonance}
k\cdot\omega\alpha - m = 0.
\end{equation}
When \eqref{resonance} holds we say that $(k,m)$ is a discrete
resonance for $\omega\alpha$ and we refer to the pair $(k,m)$ as a
resonance.
\begin{Remark}
Note that these discrete resonances \eqref{resonance}
are different from the resonances of the media
we excluded before ($k\cdot \alpha\neq 0$, $\forall ~k\in\mathbb{Z}^d-\{0\}$).
\end{Remark}
\begin{Remark}
If \[k\cdot\alpha\neq0~\quad\forall
~k\in\mathbb{Z}^d\setminus \{0\},\]
given any $k_0 \in\mathbb{Z}^d \setminus \{0\}$, $m \in \mathbb{Z}$
we have that $\omega = -m/(k_0 \cdot \alpha)$ is a resonant frequency.
Since $k_0 \cdot \alpha$ can be arbitrarily large, we see that the
set of resonant frequencies is dense on the real line. Of course, once we fix $\alpha$, the set of resonant $\omega$ is a countable set.
\end{Remark}
\subsubsection{Multiplicity of a resonance}
Clearly, if $(k, m)$, $(\tilde{k},
\tilde{m})$ are discrete resonances so is $(k+\tilde{k},
m+\tilde{m})$.
In mathematical language,
\[
\mathscr{M}_{\omega\alpha} = \left\{(k,m)\in
\mathbb{Z}^d\times\mathbb{Z}~:~k\cdot \omega\alpha-m =0 \right\}
\]
is a $\mathbb{Z}$-module called the resonance module for $\omega$.
We denote by $l(\omega) =dim(\mathscr{M}_{\omega\alpha})$ the dimension of
the resonance module and we call it the multiplicity of
the resonance. The meaning of $l(\omega)$ is the number of independent
resonances. We can find $(k_1, m_1),\ldots, (k_l,m_l)$ in such a way
that all resonances can be expressed as combinations of the basic
resonances (and also no other set of basic resonances with smaller
number of elements will allow to express all the resonances).
\subsubsection{Only resonances of multiplicity 1 appear in the models \eqref{formal energy}}
\label{multiplicity 1}
In Hamiltonian mechanics for systems with $d$ degrees of
freedom, one can find resonances of all
multiplicities up to $d$. As we will see later, in Section \ref{sec:dynamical}, one can give a dynamical interpretation of the equilibrium equations as a dynamical system in $d+1$ dimensions. Nevertheless, in our models only $l=1$ appears
independently of the number of degrees of freedom.
This highlights that the problem here is different from the Hamiltonian problem.
\begin{prop}
If $\omega\alpha$ is resonant, i.e. $\mathscr{M}_{\omega\alpha}\neq \{0\}$, then $l(\omega)=1$.
\end{prop}
\begin{proof}
Note that
\[
k_1\cdot \omega\alpha - m_1 = k_2\cdot \omega\alpha - m_2 =0
\]
implies (because $m_1\neq 0, ~m_2\neq 0$ because of \eqref{alpha
nonresonant})
\[
\omega = \frac{m_1}{k_1\cdot \alpha} = \frac{m_2}{k_2\cdot \alpha}
\] and therefore
\[
\alpha\cdot (k_1 m_2 - k_2 m_1) = 0
\] and, because $\alpha$ is non-resonant \eqref{alpha nonresonant} we have
\[
k_1 m_2 = k_2 m_1.
\]
Therefore, the two resonant vectors are related.
\end{proof}
\subsubsection{The intrinsic frequencies}
When $\omega\alpha$ is resonant, we can find a matrix $B\in SL(d, \mathbb{Z})$,
$\Omega \in \mathbb{R}^{d-1}$, $L \in \mathbb{Z}^d$ in such a way that
\begin{equation}\label{Bdefined}
B\omega\alpha = (\Omega, 0) + L \quad \text{with}\quad \Omega\cdot
k\notin \mathbb{Z} \text{ for } k\in
\mathbb{Z}^{d-1}-\{0\}.
\end{equation}
We will refer to $\Omega$'s as the intrinsic frequencies.
They are essentially unique, i.e., unique up to changes of basis in $\mathbb{R}^{d-1}$ given by a matrix in $SL(d-1, \mathbb{Z})$.
In this case, the set $\{n\omega\alpha\}_{n\in\mathbb{Z}}$ has a
closure which is a $d-1$ dimensional torus. This torus is invariant
under the translation $T_{\omega\alpha}$. If we stay in this $d-1$ dimensional torus, $T_{\omega\alpha}$ can be described as $T_{\Omega}$. The torus $\mathbb{T}^d$ is
foliated by these $\mathbb{T}^{d-1}$ indexed by another parameter
$\eta\in\mathbb{T}^1$. We will write a point in $\mathbb{T}^d$ as
$(\psi,\eta)$ where $\psi$ is the coordinate corresponding to the
position in $\mathbb{T}^{d-1}$. The coordinate $\eta$ selects the
$d-1$ torus we are considering.
\subsection{Quasi-periodic equilibria
with resonant frequencies}
\label{resonanthull}
The natural notion of the hull functions in the resonant case would be
to assume that the equilibrium solutions have the form
\begin{equation} \label{hull resonant}
x_n = n \omega + v(n \Omega)
\end{equation}
with $v:\mathbb{T}^{d-1} \rightarrow \mathbb{R}$.
Note that the physical meaning of $\omega$ is still the mean spacing
of the solutions (i.e., an inverse density). The term $v(n\Omega)$
represents fluctuations that can be parameterized in terms of the
intrinsic frequency $\Omega$. Of course, we could represent them in
terms of the original frequencies, but it is more natural to change
variables so that they become a part of the equation.
We will refer to the $\eta$ variable as the \emph{transversal phase}.
The resonant solutions considered here, cover densely a torus
of codimension one.
The one-dimensional variable $\eta$ measures
the position of these codimension-one tori on the configuration space $\mathbb{T}^d$ corresponding to the internal phases of $V$.
If we substitute the parameterization \eqref{hull resonant}
into the equilibrium equation, we obtain that
the equilibrium equation \eqref{equilibrium} is equivalent to:
\begin{equation} \label{equilibrium resonant hull}
v(n \Omega + \Omega) +
v(n \Omega - \Omega)
- 2 v(n \Omega) + \partial_\alpha V(n \omega \alpha + \alpha v(n \Omega)) +\lambda = 0.
\end{equation}
If we furthermore introduce the notation
$\partial_\alpha V(\theta) =W(B \theta)$ and $B \alpha = \beta$,
and observe that the $n \Omega$
is dense on $\mathbb{T}^{d-1}$, we see that for continuous functions $v$, \eqref{equilibrium resonant hull}
is equivalent to:
\begin{equation}\label{equilibrium resonant hull3}
v(\psi + \Omega) +
v(\psi - \Omega)
- 2 v(\psi) + W((\psi, 0) + \beta v(\psi)) +\lambda = 0.
\end{equation}
Note that we can also consider solutions of the form
\begin{equation}\label{hull2}
x_n=n\omega+v(n\Omega+\xi_1)+\xi_2
\end{equation}
for any fixed $\xi_1$ and $\xi_2$. This will give us freedom to add the transversal phase $\eta$ as an additional parameter in \eqref{equilibrium resonant hull3}. So the equilibrium equation we will consider is
\begin{equation}\label{equilibrium resonant hull2}
v(\psi + \Omega) +
v(\psi - \Omega)
- 2 v(\psi) + W((\psi, \eta) + \beta v(\psi)) +\lambda = 0.
\end{equation}
By simple calculations, it can be shown that \eqref{equilibrium resonant hull2} is equivalent to \eqref{equilibrium} when the hull function is of the form \eqref{hull2} and $B\alpha\xi_2=(\xi_1,\eta)$.
\begin{Remark}
Because $\beta$ has components
both in the $\psi$ and the $\eta$ directions, the equation
\eqref{equilibrium resonant hull2} cannot be considered as a parameterized version
of the equations considered in \cite{SuL1}. As we will see,
the symmetries of the equation involve transformations that mix
the dependence in $\psi$ and in $\eta$.
\end{Remark}
\subsection{The symmetries of the invariance equation \eqref{equilibrium resonant hull2}}
The equation \eqref{equilibrium resonant hull2} possesses remarkable symmetries
that make the solutions not unique. These symmetries lead to Ward identities.
In contrast with the case of non-resonant solutions, the group of symmetries
is infinite dimensional. In \cite{Rafael'08, SuL1, SuL2} these symmetries
are used to develop a KAM method.
The main observation is that
if $(v, \lambda)$ is a solution of \eqref{equilibrium resonant hull2},
then, for every $\iota(\eta): \mathbb{T}^1 \rightarrow \mathbb{R}$,
the pair $(\tilde v, \tilde\lambda)$ is also a solution of \eqref{equilibrium resonant hull2}
where we denote $\beta = (\beta_\psi, \beta_\eta)$ and $\tilde v, \tilde \lambda$ are defined
by:
\begin{equation}\label{symmetry}
\begin{split}
\tilde v(\psi,\eta) & =
v\big((\psi,\eta) + \iota(\eta) \beta\big) + \iota(\eta),\\
\tilde \lambda(\eta) &= \lambda \big(\eta + \iota(\eta)\beta_\eta\big).
\end{split}
\end{equation}
Notice that the symmetry \eqref{symmetry} involves changing not only
the argument $\psi$ but also the argument $\eta$. Note the space of symmetries of the equation is not just a finite dimensional space but rather an infinite dimensional space of functions.
\subsection{A normalization of the solutions of the invariance equation \eqref{equilibrium resonant hull2}}
For later applications, it will be useful to have local
uniqueness of the solutions (e.g. to discuss smooth dependence on
parameters, perturbative expansions on parameters). Hence we impose the normalization
\begin{equation}\label{normalization}
\int_{\mathbb{T}^{d-1}} v(\psi,\eta) \, d \psi = 0.
\end{equation}
Since the symmetry \eqref{symmetry} involves changes of
arguments, giving a $v_\eta$, finding the $\iota(\eta)$ that accomplishes the normalization involves solving the implicit
equation
\begin{equation}\label{implicit}
I(\eta + \beta_\eta \iota(\eta) ) + \iota(\eta) = 0
\end{equation}
where $I(\eta) \equiv \int_{\mathbb{T}^{d-1} } v(\psi,\eta)\, d \psi$.
If $I$ and its derivative are small, one can solve eqref{implicit} using implicit function theorem.
\subsection{Diophantine condition}\label{Diophantine properties}
In contrast with KAM theory, we will not need very delicate estimates on the solutions
and hence, we can deal with very general Diophantine conditions.
We will assume that $\Omega$ satisfies
\begin{equation} \label{subexponential}
\lim_{N\to\infty} \frac{1}{N} \sup_{|k| \le N, m \in \mathbb{Z}} \bigg| \ln |k\cdot \Omega-m| \bigg| = 0.
\end{equation}
Note that the condition \eqref{subexponential} is much weaker than the usual
Diophantine conditions and even than the Bjruno-R\"ussmann conditions.
The condition\eqref{subexponential} is the natural condition in
the study of existence of series to all orders.
The following proposition shows that the sets of frequencies we are considering are abundant.
Fix a vector $k \in \integer^d \setminus \{0\}$, $m \in \integer\setminus\{0\}$ and assume without loss of generality that there is no common divisor in the components of $k$. For any $\alpha \in \real^d$ satisfying $\alpha\cdot k\neq 0$, we can find a unique $\omega$
such that $\alpha \cdot k \omega - m = 0$. Fix a $B_k\in SL(d,\mathbb{Z})$ such that the last row of $B$ equals $k$ (It's possible to find such $B$ since the components of $k$ has no common divisor). Then, let $B_k\alpha\omega=(\Omega,0)+m$. Hence, for any $k, m$ and $\alpha$ satisfying $\alpha\cdot k\neq 0$, we can
define $\Omega$ as a function of $\alpha$. Denote $\Omega = F_{k,m}(\alpha)$.
\begin{prop}
\label{prop:diophantine}
The set of $\alpha$ for which
$F_{k,m}(\alpha)$ satisfies \eqref{subexponential} for all $k,m$ is of full measure in $\mathbb{R}^d$. In other words, for a full measure set of medium frequencies, we can find a countable many resonant frequencies that lead to intrinsic frequencies satisfying \eqref{subexponential}.
\end{prop}
\begin{proof}
Since countable intersections of sets of full measure are of full measure, to prove Proposition~\ref{prop:diophantine} it suffices to show that for a fixed $k,m$ as above, the set $\{\alpha \in \real^d|\alpha\cdot k\neq 0\ \mathrm{and}\ F_{k,m}(\alpha)\ \mathrm{satisfies}\ \eqref{subexponential}\}$ is of full measure.
Because the set of $\Omega$'s which satisfy \eqref{subexponential} is of full measure on $\real^{d -1}$ and the linear map $B_k$ is differentiable and surjective, the preimage of the set of $\Omega$'s that satisfy \eqref{subexponential} under $B_k$ is also of full measure in the hyperplane $\Gamma=\{\gamma\mid \gamma\cdot k-m=0\}$. Denote the set of preimages as $\Gamma'$. Then any nonzero scaling of an element of $\Gamma'$ will give an $\alpha$ we want, which also form a full measure set in $\mathbb{R}^d$.
\end{proof}
\section{Function spaces and linear estimates}\label{cohomology}
The main tool that we will use to construct perturbation theories is
the solution of cohomology equations.
We denote
\[
D_\rho = \{ \theta \in \complex^d/\integer^d \mid | \text{Im}(\theta_i)| < \rho\}
\]
and denote the Fourier expansion of a periodic mapping $v(\psi,\eta)$ on
$D_\rho$ by
\begin{equation*}
v(\psi,\eta)=\sum_{k\in \mathbb{Z}^d} v_k e^{2\pi i k\cdot (\psi,\eta)},
\end{equation*} where $\cdot$ is the Euclidean scalar product in
$\mathbb{C}^d$ and $v_k$ are the Fourier coefficients.
We denote by $\mathscr{A}_\rho$ the Banach space of analytic
functions on $D_\rho$ which are real for real argument and extend
continuously to $\overline{D_\rho}$. We make $\mathscr{A}_\rho$ a
Banach space by endowing it with the supremum norm:
\begin{equation*}
\|v\|_\rho =\sup_{(\psi,\eta)\in \overline{D_\rho}}|v(\psi,\eta)|.
\end{equation*}
These Banach spaces of analytic functions
are the same spaces as in \cite{Moser'67}.
We will consider equations of the form
\begin{equation}\label{Cohomology equation}
v(\psi+\Omega,\eta)-v(\psi,\eta)=\phi(\psi,\eta),
\end{equation}
where $\psi \in \torus^{d-1}$.
To simplify our notations, we will denote $v(\psi+\Omega)$ and $v(\psi-\Omega)$ as $v_+$ and $v_-$, respectively. Similar notations will be used for other functions. We also use $T$ to represent the translation operators, i.e., $T_{\Omega}v(\psi)=v(\psi+\Omega)$.
\begin{lemma}\label{estimate lemma for cohomology equation}
Let $\phi \in\mathscr{A}_\rho(\mathbb{T}^d)$ be such that
\begin{equation}\label{normalization phi}
\int_{\mathbb{T}^{d-1}}\phi (\psi,\eta)d\psi=0,
\end{equation}
for all $\eta$.
Assume that $\Omega$ satisfies the assumption
\eqref{subexponential}.
Then, for a fixed $\eta$, there
exists a unique solution $v_\eta$ of
\eqref{Cohomology equation} which satisfies
\begin{equation}\label{normalization equation}
\int_{\mathbb{T}^{d-1}} v (\psi,\eta)d\psi = 0.
\end{equation}
The solution $v \in \mathscr{A}_{\rho'} $ for
any $\rho' < \rho$
and we have
\[
|| v ||_{\rho'} \le C(\rho,\rho') || \phi_\eta||_\rho.
\]
Furthermore, any distribution solution of \eqref{Cohomology
equation} differs from the solution claimed before by a constant.
If $\phi$ is such that it takes real values for real arguments,
so does $v$.
If we consider now the dependence in $\eta$, we
have that
$v \in \mathscr{A}_{\rho'}(\mathbb{T}^d)$
and
\[
||v||_{\rho'} \le C(\rho, \rho') ||\phi||_\rho.
\]
\end{lemma}
\begin{proof}
We note that, as it is well known that obtaining $v$
solving \eqref{Cohomology equation} for
given $\phi$ is very explicit in
terms of Fourier coefficients.
If
\[
\phi(\psi,\eta) = \sum_{k\ne 0}\hat \phi_k(\eta) e^{2 \pi i k \cdot \psi}
= \sum_{k \ne 0 , m} \hat \phi_{k,m} e^{2 \pi i (k \cdot \psi + m \eta)}
\]
then, $v$ is given by
\[
v(\psi,\eta) =
\sum_{k \ne 0} \hat \phi_k(\eta)(e^{2 \pi i k \cdot \Omega} -1)^{-1}
e^{2 \pi i k \cdot \psi}
= \sum_{k\ne 0, m} \hat \phi_{k,m} (e^{2 \pi i k \cdot \Omega} -1)^{-1} e^{2 \pi i (k \cdot \psi + m \eta)}.
\]
Using Cauchy estimates for the Fourier coefficients
$|\hat \phi_{k,m} | \le \exp( -2 \pi \rho ( |k| + |m|)) ||\phi||_\rho$
and that
$| e^{2 \pi k \cdot \Omega} -1|^{-1} \le C \dist(k \cdot \Omega, \integer)^{-1} $
and the assumption~\eqref{subexponential}, we obtain
that
\[
\begin{split}
||v||_{\rho'}
& \le C \sum_{k\ne 0, m} \exp( -2 \pi \rho ( |k| + |m|)) ||\phi||_\rho
\dist(k \cdot \Omega, \integer)^{-1}
||e^{2 \pi i (k\cdot \psi + m \eta)} ||_{\rho'} \\
& \le C ||\phi||_\rho\sum_{k\ne 0, m} \exp( -2 \pi \rho ( |k| + |m|))
{\rm dist}(k \cdot \Omega, \integer)^{-1}
\exp( 2 \pi \rho' ( |k| + |m|))\\
& \le C(\rho,\rho')||\phi||_{\rho}.
\end{split}
\]
\end{proof}
\section{Lindstedt series for quasi-periodic solutions with resonant frequencies}
\label{Lindstedt}
The goal of this section is to study \eqref{equilibrium resonant hull2}
perturbatively when the non-linear term is small. Hence, we will
write \eqref{equilibrium resonant hull2} with a small parameter $\eps$
\begin{equation}\label{equilibrium revised}
\begin{split}
v(\psi +\Omega,\eta) + v(\psi -\Omega,\eta) -
2 v(\psi,\eta) + \epsilon W((\psi, \eta) + \beta v(\psi,\eta)) + \lambda(\eta)&=0 .\\
\end{split}
\end{equation}
We will find $v(\psi,\eta), \lambda(\eta)$
solving \eqref{equilibrium revised} and \eqref{normalization} in the sense of
formal power
series in $\epsilon$.
In this paper, we will not consider the problem of whether these series converge or represent a function. This will be studied in more details in \cite{ZhangSL15}.
Since one may want to find solutions correspond to $\lambda=0$ (or $\lambda=\lambda^*$ with $|\lambda^*|$ small), it is important for us to keep track of
$\frac{\partial \lambda}{\partial \eta}(\eta,\epsilon)$ in order to solve $\lambda(\eta,\epsilon)=0$ by implicit function theorem.
Following the standard perturbative procedure we will write
\begin{equation}
\label{seriesexpansion}
\begin{split}
v & = \sum_{n= 0}^\infty \epsilon^n v^n,\\
\lambda &= \sum_{n= 0}^\infty \epsilon^n \lambda^n.
\end{split}
\end{equation}
Here $v^n$ and $\lambda^n$ are coefficients of $\epsilon^n$, not powers of $v$ or $\lambda$. Substitute \eqref{seriesexpansion} in \eqref{equilibrium revised}
and equate powers of $\epsilon$.
Of course, carrying out this procedure for $n\leq N$ will require
that $\Omega$ satisfies some Diophantine properties as well as some differentiability
assumptions.
Equating the coefficients of $\epsilon^0$ in \eqref{equilibrium revised}
we obtain
\begin{equation}
\begin{split}
v ^ 0(\psi+ \Omega,\eta) + v^0(\psi - \Omega,\eta) -
2v^0(\psi,\eta) + \lambda^0(\eta) &=0.\\
\end{split}
\end{equation}
Hence, if $\Omega$ satisfies the condition \eqref{subexponential} we see that $v^0$ is constant,
$\lambda^0 = 0$ and imposing the normalization \eqref{normalization} we obtain
$v^0 = 0$.
Matching coefficients of $\epsilon^1$ in both sides of \eqref{equilibrium revised}
we obtain
\begin{equation}\label{order1}
\begin{split}
v^1(\psi+\Omega,\eta) + v^1(\psi-\Omega,\eta) -
2v^1(\psi,\eta) + W (\psi, \eta) + \lambda^1(\eta) &= 0.
\end{split}
\end{equation}
We see that, using the theory in Section ~\ref{cohomology}, to have analytic
$v^1$ solving \eqref{order1}, it is necessary and sufficient to
have
\begin{equation}\label{lambda1}
\begin{split}
\lambda^1(\eta) &= -\int_{\mathbb{T}^{d-1}} W (\psi, \eta) d\psi.
\end{split}
\end{equation}
Then, $v^1$, $\lambda^1$ can be determined uniquely up to a constant from
\eqref{order1}. In fact, in Fourier series, the equation for $v^1, \lambda^1 $
is
\begin{equation}
\begin{split}
v_k ^1 2 (\cos (2\pi k \Omega) -1) & = - W_k - \delta_{0, k} \lambda^1,\\
\end{split}
\end{equation}
where $\delta_{0,k}$ is the Kronecker delta.
In particular, the constant in $v^1$ is determined by the normalization \eqref{normalization}.
Proceeding to higher order follows the same pattern. We see that
matching the terms of order $\epsilon^n$ in \eqref{equilibrium revised} we obtain
\begin{equation}\label{ordern}
\begin{split}
v^n (\psi +\Omega,\eta) + v^n (\psi -\Omega,\eta) -
2v^n (\psi,\eta) + R^n(\psi,\eta) + \lambda^n(\eta) &=0,\\
\end{split}
\end{equation}where $R^n$ is a polynomial expression in $v^1,\ldots,
v^{n-1}$ with coefficients which are derivatives with respect
to $\psi$ of $W((\psi, \eta) + \beta v(\psi,\eta))$.
This polynomial can be computed
explicitly because it is given by
\begin{equation}\label{R^n}
R^n = \frac{1}{(n-1)!}\frac{d^{n-1}}{d\epsilon^{n-1}} W
\bigg((\psi,\eta) + \beta \sum_{j=0}^{n-1} v^j
(\psi,\eta)\bigg)\bigg|_{\epsilon=0}
\end{equation} and these are well known formulae. We also note that,
from the algorithmic point of view there are efficient ways to
compute $R^N$ using methods of ``automatic
differentiation'' \cite{Harodifferentiation, BCHNN06}.
Since $R^n$ can be computed explicitly, \eqref{ordern} can be solved the same way as \eqref{order1}.
We have therefore established
\begin{theorem}\label{LindstedtN}
Assume that $\Omega$ satisfies \eqref{subexponential} and that $W:D_{\rho} \rightarrow \mathbb{C}$
is an analytic function. We can find formal power series solutions in $\epsilon$ of the form
\eqref{seriesexpansion} solving the equation \eqref{equilibrium revised}. For any $0<\delta<\rho$, each of the terms $v^n(\psi,\eta)$ is analytic in $D_{\rho-\delta}$. If $W$ takes real values for real values,
then so do $v^n$, and $\lambda$ is real.
\end{theorem}
\subsection{The auxiliary equation}
Now, we turn to the problem of studying the equation
\begin{equation}\label{auxiliary}
\lambda(\eta, \epsilon) = \lambda^*.
\end{equation}
We expect to obtain a solution $\eta^\ast(\epsilon)$ provided that \eqref{auxiliary} satisfies some non-degeneracy conditions.
Having solution of \eqref{auxiliary} to order 1 in $\epsilon$, amounts to
\[
\lambda^1(\eta) =0.
\]
That is, we need to find $\eta$ such that
\begin{equation}\label{auxiliaryfirst}
\int_{\mathbb{T}^{d-1}} W((\psi, \eta)) d\psi =0.
\end{equation}
\begin{theorem}
\label{prop:twosolutions}
The equation \eqref{auxiliaryfirst} has always two solutions.
\end{theorem}
\begin{proof}
Since
\[
\int_{\mathbb{T}^{d-1}} W((\psi, \eta)) d\psi =
\int_{\mathbb{T}^{d-1}} (\partial_\alpha V)\big(B^{-1}(\psi,\eta) \big)
d\psi,
\]if we integrate again with respect to $\eta$ we obtain
\begin{equation}
\int_{\mathbb{T}}\int_{\mathbb{T}^{d-1}} W((\psi, \eta) d\psi\ d\eta = \int_{\mathbb{T}^{d}} (\partial_\alpha V)\big(B^{-1}(\psi,\eta)\big)
d\psi\ d\eta = 0.
\end{equation}
Hence the function of $\eta$ given by $\int_{\mathbb{T}^{d-1}}
W((\psi, \eta) d\psi$ is a continuous periodic function of $\eta$
with zero average. Therefore, it has at least two zeros. We also note that there are open sets of perturbations where there are 4,6,$\cdots$ zeros.
\end{proof}
Denote one of these solutions of \eqref{auxiliaryfirst} as
$\eta^\ast$.
A sufficient condition that ensures that we can solve the equation
\eqref{auxiliary} to all orders is that
\begin{equation}\label{nondeg}
\frac{\partial}{\partial \eta} \lambda^1(\eta, \epsilon)
\left|_{\eta=\eta^\ast,\epsilon=0}\right. \neq 0.
\end{equation}
More explicitly,
\begin{equation}\label{nondegeneracy}
\int_{\mathbb{T}^d} \frac{\partial}{\partial \eta} (\partial_\alpha V)\big(B^{-1}(\psi,\eta)\big)
d\psi\ d\eta \neq 0.
\end{equation}
Then, the implicit function theorem for power series
\cite{Cartan95,Dieudonne71} gives us that we can indeed find $\eta^\ast(\epsilon)$.
Similarly, we can solve the equation $\lambda(\eta) = \lambda^*$ provided that $|\lambda^*|$ is
sufficiently small.
Therefore, we have established
\begin{theorem}\label{secondLindstedt}
Assume that $\Omega\in \mathscr{D}(\nu,\tau)$ as defined in
\eqref{subexponential} , that $W$ is an analytic function,
and that \eqref{nondegeneracy} holds,
we can find formal power series $\eta_\epsilon$ in $\epsilon$
so that $v_{\eta_\epsilon}$ is the solution of \eqref{equilibrium revised}.
\end{theorem}
Clearly, since the function $\lambda^n(\eta)$ are bounded, if
$\lambda^*$ -- the physical force -- is large enough, there is no solution.
This has a clear physical meaning. If we increase the external force but keep it small,
the system can react by changing the transversal phase. If the force increases beyond
a threshold, the system cannot react by adapting the phase. Hence, the equilibrium breaks
down. In this paper, we are not considering the dynamics of
the model, only the equilibria (our models
for the energy include only the potential energy of
the configuration and not any kinetic energy).
One can, however, expect that, if there was some dynamics,
the equilibria considered here could slide.
Of course, the sufficient condition \eqref{nondegeneracy} is
far from being necessary and there are many other conditions that
are enough.
\begin{prop}\label{secondLindstedtdegenerate}
Assume that $\Omega$ satisfies \eqref{subexponential} , that $W((\psi, \eta) + \beta v(\psi,\eta)) $ is an analytic function,
and that \eqref{nondegeneracy} holds.
Assume that $\eta^*$ is such that for some $m \in \mathbb{N}$
we have
\begin{equation}
\begin{split}
& \lambda^i(\eta^*) = 0, \quad i = 1,\ldots 2 m\\
& \lambda^{2 m +1} (\eta^*) \ne 0.
\end{split}
\end{equation}
Then, we can find formal power series $\eta_\epsilon$ in $\epsilon$
so that $v(\psi,\eta_\epsilon)$ is the solution of \eqref{equilibrium revised}.
\end{prop}
The proof is again an application of the implicit function theorem for power series.
\section{A dynamical interpretation of the equilibrium equations of Frenkel-Kontorova models}
\label{sec:dynamical}
In this section, we present a dynamical interpretation of the
equilibrium equations \eqref{equilibrium} in Frenkel-Kontorova models.
Even if the dynamical interpretation is possible for finite
range interactions, we see that adding another small interaction of
longer range is a singular perturbation (even the dimension of the phase space changes). Whereas, for the methods in
this paper, adding a small term in the longer range is a regular perturbation of the same order.
A straightforward way of transforming the equilibrium equation
\begin{equation}
x_{n+1} + x_{n-1} - 2 x_n + \epsilon\partial_\alpha V(x_n\alpha) + \lambda = 0 \quad
\forall~ n\in \mathbb{Z}
\end{equation}
into a
dynamical system is setting
\begin{equation}\label{evolution}
\begin{split}
y_n&=(x_n, x_{n-1})\\
y_{n+1}&= (2y_n^1- y_n^2 - \epsilon\partial_\alpha V(\alpha y_n^1)-\lambda, y_n^1).
\end{split}
\end{equation}
However, \eqref{evolution} is not very useful because we have to
consider it as a map of $\mathbb{R}^2$ and the term $\partial_\alpha V
(\alpha y_n ^1)$ does not make apparent that it is periodic in $\alpha
y_n^1$.
A more natural formulation is obtained by observing that the equation
\eqref{equilibrium} is equivalent to the system on $\mathbb{T}^d\times
\mathbb{R}$
\begin{equation}\label{standard}
\begin{split}
p_{n+1} &= p_n - \epsilon\partial_\alpha V(q_n) - \lambda\\
q_{n+1} &= q_n + \alpha p_{n+1},
\end{split}
\end{equation}where $q_n\in\mathbb{T}^d$, $p_n\in \mathbb{R}$.
(Just multiply \eqref{equilibrium} by $\alpha$ and use the substitution $p_n=x_n-x_{n-1}$, $q_n=\alpha x_n$. Note that \eqref{equilibrium} is equivalent to
\[
(x_{n+1} - x_n) - (x_n - x_{n-1}) + \epsilon\partial_\alpha V(\alpha x_n) + \lambda =0
\]
hence, we obtain the first equation.)
We will write the mapping~\eqref{standard} as
\begin{equation}\label{standardmap}
(p_{n+1}, q_{n+1}) = F_{\eps,\lambda}( p_n, q_n).
\end{equation}
Note that \eqref{standard} is
typographically very similar to the standard map \cite{Chirikov}
or to analogues introduced for volume preserving maps.
Nevertheless, there are significant differences
(besides the different dimensions).
A very crucial difference between \eqref{standardmap}
and the generic volume preserving maps is that
$q_{n+1} - q_n$ is always a multiple of $\alpha$ (see \eqref{standard}).
So that the two dimensional leaves
\begin{equation} \label{constraints}
\mathcal{M}_{q_0} = \{ (p, q_o + \alpha t) ~ |~ p, t \in \mathbb{R}\}
\end{equation}
are preserved.
Note that each of the leaves $\mathcal{M}_{q_0}$ is dense in the $d+1$
dimensional phase space.
The mapping \eqref{standard} clearly preserves the volume form
$dp\wedge dq_1 \wedge \ldots \wedge q_d$ since it is the composition of
\begin{equation}\label{map1}
\begin{split}
p_{n+1} &= p_n - \epsilon\partial_\alpha V(q_n) - \lambda\\
q_{n+1} &= q_n
\end{split}
\end{equation}
and
\begin{equation}\label{map2}
\begin{split}
p_{n+1} &= p_n \\
q_{n+1} &= q_n + \alpha p_{n+1}.
\end{split}
\end{equation}
We recall that, in our context,
a volume preserving map
is exact when
$F^* (p dq_1 \wedge d q_2 \wedge \ldots \wedge d q_d) =
p dq_1 \wedge d q_2 \wedge \ldots \wedge d q_d + d P
$ where $P$ is $d-1$ form.
Indeed, \eqref{standardmap} is an exact volume preserving map if and only if
$\lambda = 0$, since it is easy to observe that, when $\lambda = 0$, both \eqref{map1}
and \eqref{map2} are exact.
When $\eps = 0$, $\lambda = 0$, the map \eqref{standard}
is integrable. The codimension-one tori given by $p = \text{cte.}$
are invariant and the motion in them is a rotation.
The volume preserving KAM theory leads us to expect that for $\epsilon\ll 1,\lambda=0$, the tori in which the frequency of the motion is Diophantine survive. We also expect that the tori with resonance, breaks down into lower dimensional tori. The lower dimensional tori can survive for $|\lambda|\ll 1$ (depending on $\epsilon$).
In this paper we have quantitative (but formal and non-rigorous) prediction of these phenomenon based on perturbative expansions. We hope that some of them may be either verified by rigorous results or explored numerically.
\subsection{On the global geometry of the constraints given
by \eqref{constraints} }
Integrable systems with constraints have been studied extensively
in geometric mechanics. Nevertheless, the systems we consider here
have some unusual properties that we would like to highlight.
It is customary to classify the constraints in holonomic when the
distributions are integrable (in the sense that they foliate the phase
space with a smooth quotient) and non-holonomic when the distributions
are not integrable and they violate the
hypothesis of Frobenius Theorem \cite{Souriau,Audin,Holm1}.
The constraints \eqref{constraints} escape this dichotomy.
They are locally integrable (they do satisfy the hypothesis of
Frobenius Theorem and are locally given by invariant manifolds that give rise
to a foliation) but nevertheless, the manifolds are dense, so that
they do not give
a nice quotient manifold.
Hence, even if we have holonomic constraints locally (and the infinitesimal results
about holonomic systems are applicable), some global aspects such as symplectic reduction \cite{Meyer73,
MarsdenW74,MarsdenW01} cannot be applied to \eqref{standard}.
\subsection{Lyapunov exponents and phonon localization}\label{sec:phonon}
In this section we study the so called \emph{phonon gap} around the equilibria of \eqref{equilibrium resonant hull}
given by a hull function.
Let us start by recalling some standard definitions.
The main idea is that sound waves are defined by the propagation of
infinitesimal disturbances around an equilibrium equation.
If we linearize around an equilibrium solution $x=\{x_n\}_{n\in\mathbb{Z}}$, we
obtain the dynamics of the infinitesimal perturbations
$\xi_n$ is given by
\begin{equation} \label{evolutionlinear}
\ddot \xi_n = \xi_{n+1} + \xi_{n-1} - 2 \xi_n + (\partial_\alpha)^2 V(\alpha x_n)
\xi_n \equiv (\mathcal{L}_{x} \xi)_n.
\end{equation}
It is clear that the propagation properties of sound waves will
be affected by the spectral properties of the operator $\mathcal{L}_x$.
Note that the operator $\mathcal{L}$ is a one-dimensional
Schr\"odinger operator with a position dependent potential.
The dependence will be given by the dynamics of the $x_n$. In particular, for
the solutions given by a hull function, we will be considering quasi-periodic
potentials.
The mathematical theory of the spectrum of
quasi-periodic Schr\"odinger operators is well developed
\cite{PasturF92,HaroL}. In particular, it is known that the spectrum is
independent of the $\ell^p$ space in which it is considered,
and, more important for us, that the spectrum can be characterized
by the existence of approximate eigenfunctions. In the
dynamical interpretation in this section,
the spectrum corresponds to the Lyapunov exponents of the
solution \cite{AubryKB92}.
In the case of \eqref{standard}, we can study the Lyapunov spectra for any orbit using the
geometric constraints \eqref{constraints}.
\begin{prop}
Let $(p_n,q_n)$ be an orbit of the mapping given by \eqref{standard}. Assume that it is an orbit in the full measure set that Osledets
Theorem applies. Then, $d-1$ Lyapunov exponents are
zero. Also, the sum of all the Lyapunov exponents is zero.
\end{prop}
\begin{proof}
Consider $\tilde F$, the lift of the map $F$ in \eqref{standardmap}.
Let $s$ be a vector perpendicular to $\alpha$. It is a simple computation
to show that:
\[
\tilde F( \tilde M_{q_0 + s} ) = \tilde M_{q_0} + s.
\]
Then it is clear that the $d-1$ vectors in the directions perpendicular to $s$ do
not grow.
The fact that the sum of the Lyapunov exponents for orbits of
a volume preserving map is
zero is well known since the sums of the Lyapunov exponents is
the rate of growth of the determinant of iterates of the map.
\end{proof}
Of course, the dynamical system \eqref{standard} is straightforward to implement numerically and allows study of statistical properties of depinning.
\section*{Acknowledgements}
We thank Dr. T. Blass for discussions.
R. L. and L. Z. have been supported by DMS-1500943.
The hospitality of
JLU-GT Joint institute for Theoretical Sciences for the three authors
was instrumental in finishing the work.
R.L also acknowledges the hospitality of
the Chinese Acad. of Sciences.
X. Su is supported by both National
Natural Science Foundation of China (Grant No. 11301513) and ``the Fundamental Research Funds for the Central Universities".
\bibliographystyle{alpha}
|
\section{Introduction}
The on-demand single electron source (SES) based on a driven mesoscopic capacitor, \cite{Feve07} has allowed to achieve interferometric experiments with individual electron and hole wave-packets propagating ballistically along integer quantum Hall (IQH) channels and opened the way to electron quantum optics.\cite{Bocquillon13b}
It relies on a quantum dot both tunnel-coupled to a quantum Hall edge channel and capacitively coupled to a periodically modulated gate.
In its optimal regime of operation, this periodic source emits exactly one electron in the first half-period and one hole in the second half-period. \cite{Mahe08, Grenier11}
This is achieved for an intermediate transparency of the point contact connecting the dot and the edge, when the gate is biased with a square voltage whose amplitude is equal to the dot level spacing.
A complete characterization of this SES can be given in terms of a non-interacting picture in which the electron-electron Coulomb interaction on the dot can be effectively taken into account in terms of a renormalization of the level spacing. This allows to model the action of the mesoscopic capacitor through the Floquet scattering matrix theory. \cite{Moskalets02, Moskalets07, Moskalets08}
The SES allowed to perform Hanbury Brown and Twiss\cite{HBT} (HBT) experiments with single electrons, as well as Hong-Ou-Mandel\cite{HOM} (HOM) collisions between two electrons propagating on opposite edge channels of the quantum Hall effect.
\begin{figure}[h]
\centering
\includegraphics[scale=0.50]{Antidot_Fig1.pdf}
\caption{Artistic view of a (strongly asymmetric) antidot embedded into an Hall fluid and coupled with edge channels through tunneling amplitudes $t_{L}$ and $t_{R}$ respectively ($|t_{L}|\gg |t_{R}|$. Dark area represents the Hall fluid while the bright ones are in correspondence of the edges of the Hall bar (black arrows) and of the antidot (with black circles indicating the its energy levels). Blue dots indicated quasiparticles. The antidot is also pierced by a time dependent magnetic flux $\Phi$}
\label{fig1}
\end{figure}
While it is becoming clearer that electron-electron interactions in quantum Hall interferometric devices dramatically affect the nature of the electronic excitations leading to their fractionalization,\cite{Bocquillon13a, Wahl14, Ferraro14} it is natural to contemplate the fascinating possibility of operating such electronic interferometers in the fractional quantum Hall (FQH) regime,\cite{FQHE} where interactions are at their strongest level. There, one would not be dealing with electrons, but rather with emergent Laughlin quasiparticles\cite{Laughlin83} that carry fractional charge, and obey fractional statistics (anyons). \cite{Stern08}
The detection of fractionally charged quasiparticles in the FQH effect has been theoretically predicted in calculations of the DC shot noise characteristics in a tunneling geometry between two counterpropagating edge states.\cite{kane_fisher} Experiments gave a confirmation of these predictions.\cite{glattli_depicciotto} Alternatively, capacitive measurements using antidot geometries seem to also point toward the detection of fractional charges.\cite{goldman_science}
Unfortunately there is so far no proposal for emitting such single quasiparticles along fractional edge channels: the simple picture of the source as a dot coupled via tunneling to the edge cannot be easily generalized to the case of quasiparticle (QP) and quasihole (QH) emission in the FQH regime. Indeed, for dot transparencies in the optimal regime of emission, the Hall fluid is predicted to be so much depleted that only electrons can tunnel between the dot and the edge of the Hall bar. \cite{Leicht11} Conversely, at higher transparencies, the dot and the edge are strongly coupled so that the output signal reduces to the response of the edge to an applied voltage and is therefore not quantized in general. Granted, in this case, a properly designed Lorentzian voltage pulse could provide an efficient way to realize individual QP injection,\cite{Jonckheere05, Keeling06, Dubois13} at least for the simplest fractional states belonging to the Laughlin sequence. \cite{Laughlin83}
The main purpose of this work is to characterize the operation of a single Laughlin QP emitter which bears analogies with the driven mesoscopic capacitor source, functioning with a driven antidot which is embedded in the fractional quantum Hall fluid.
An antidot consists of a region of total depletion realized in the Hall fluid (see Fig. \ref{fig1}), and it can be driven either by modulating the Hall magnetic field or with a back gate. As mentioned above, this geometry has been considered as a tool to extract information about both the charge and the statistics of the QP excitations.\cite{goldman_science} In particular, earlier theoretical works investigated the adiabatic pumping of the tunneling amplitudes connecting the antidot with the edges of the Hall bar as a way to emit a perfectly quantized fractional charge per cycle. \cite{Simon00, Das08} In the stationary regime this peculiar geometry also offers the possibility to obtain a persistent current induced by the piercing magnetic field. \cite{Geller97, Geller00} Moreover, noise and higher order current cumulants present features able to disentangle the universal effects associated with the filling factor and the non-universal ones related to the action of the external environment. \cite{Braggio06, Merlo07, Braggio12} More recently, the same geometry has been discussed in the case of two dimensional topological insulators, \cite{Qi11} where the asymmetry of the antidot configuration turns out to be crucial in order to induce a spin polarized current across the sample. \cite{Dolcetto13} On the experimental side, the periodicity of the conductance peaks in this geometry has been measured as a function of magnetic field and back-gate voltage. This allows to extract charge and exchange statistics of the emergent excitations of the Hall fluid,\cite{Franklin96, Maasilta00, Goldman01, Goldman05} even if these results have been longly debated. \cite{Kane03}
In this paper, we show that a periodically driven antidot, can either behave as a single-QP source (SQS) or as a SES in the Laughlin sequence of the FQH regime. The setup is shown in Fig.~\ref{fig1} where, in order to ensure the injection of fractionally charged QPs into the left ($L$) edge channel of the Hall bar only, a strong asymmetry is assumed. Here, for working purposes, the periodic drive is obtained by means of a time-dependent modulation of the magnetic field piercing the sample (equivalent modulation of the antidot levels can be achieved with a periodically biased voltage gate). In the present context, the oscillations imposed by the drive constitute only small perturbations with respect to the Hall quantizing magnetic field in order not to deviate from the fractional plateaus. The analysis of this system is carried out using the master equation approach,\cite{Furusaki98} however, in a different manner from what is usually discussed in the literature. Indeed, we need to face two new relevant problems: on the one hand the non-adiabatic time dependence of the drive has to be taken into account, and on the other hand we need to properly characterize the transient regime which is the physically relevant one for our purpose.
By tuning the amplitude of the magnetic field oscillations it is possible to induce the emission into the edge channel of a periodic train either of QPs and QHs, or of electrons and holes. The first case only involves two charge states of the antidot and can be solved analytically. In particular, it is possible to compute the emitted charge per half-period as well as the associated current fluctuations, both these quantities showing remarkable resemblance with the integer quantum Hall (IQH) mesoscopic capacitor setup. \cite{Mahe10, Albert10, Parmentier12} The second case requires a full numerical treatment, due to the large number of charge states involved for the antidot. There, we observe a properly quantized electron charge per half-period, and a vanishing noise at zero frequency.
However, the noise calculated at the drive frequency exceeds what would be expected for a SES, a consequence of spurious charge emissions randomly occurring during the half-period. All these tunneling processes do not affect the average charge or its fluctuations at zero frequency, but are inherent to the emission process of several quasiparticles in this system. Despite the great number of processes involved, in the regime of electron emission it is also possible to define an effective escape time for the excitations which can be experimentally extracted from the measurement of the first harmonic of the current.
The paper is divided as follows. In Sec. \ref{Model} we discuss the model of a finite length chiral Luttinger liquid describing the antidot in the Laughlin regime of the FQH effect, coupled to the edges of the same fractional Hall bar. Sec. \ref{Transport} is devoted to the derivation of the tunneling rates (for the master equation approach) as well as the relevant physical quantities, such as the occupation number of the antidot, the current and noise along the edge, which are all essential in order to characterize the performance of the device in both the SQS and the SES regimes. A discussion of the parameters for the optimal emission regime in the two cases is then carried out. In Sec. \ref{SQS} we consider the optimal regime of emission of QPs, where a simple analytical treatment is possible, characterizing the device as a perfect SQS. We investigate the emitted charge, as well as the current fluctuations, showing that the considered setup behaves in exactly the same way as the SES realized in the integer regime, up to the renormalization of the charge of the carriers to account for their fractional nature, $e^{*}= \nu e$. The optimal regime of emission of electrons and holes (SES) is investigated in Sec. \ref{SES} using a full numerical approach. A perfect quantization of the average emitted charge
is obtained also in this case, even though additional tunneling processes are required in order to properly describe the system. These can be detected in the finite frequency noise and lead to additional complications in defining the escape time of the electron (hole) from the antidot. In Sec. \ref{sec:estimates}, we perform estimates of the various physical parameters appearing in our SQS calculations in order to compare them with their counterpart in actual experimental realizations in the IQH regime, and we justify the feasibility of our proposed setup. Sec. \ref{Conclusions} is devoted to conclusions, while an Appendix discusses the connection between zero noise contribution at zero frequency and the absence of charge fluctuations during a period.
\section{Model} \label{Model}
The starting point of our discussion is the antidot geometry proposed in Refs.~\onlinecite{Braggio06, Merlo07}, and schematically illustrated in Fig.~\ref{fig1}. The FQH fluid is depleted by means of an electrostatic gate creating a circular ``empty'' region into the Hall bar. According to Wen's hydrodynamical approach \cite{Wen95} for the description of FQH states belonging to the Laughlin sequence \cite{Laughlin83} with filling factor $\nu=1/(2n+1)$ ($n\in \mathbb{N}$), the Hamiltonian associated with the boundaries (right edge $R$, left edge $L$ and antidot $ad$) is quadratic in terms of the edge-magnetoplasmon creation ($a^{\dagger}_{l, s}$) and annihilation ($a_{l,s}$) operators and can be written in the form ($\hbar=1$)
\begin{equation}
H^{0}_{l}= \epsilon \sum^{+\infty}_{s=1} s a^{\dagger}_{l, s}a_{l,s}+E_{c}^{l} (N_{l}-N_{\Phi})^{2}
\label{H_zero}
\end{equation}
where $l=R, L, ad$ labels the various elements of the device, $\epsilon= 2 \pi v/L_{l}$ is the energy associated with the plasmonic modes and $E^{l}_{c}= \pi v \nu /L_{l}$ is the charging energy associated with the zero modes, $L_{l}$ being the length of the $l$-th edge. The propagation velocity $v$ along the edges is assumed constant throughout all the setup.
$N_{l}$ is the number of QPs enclosed by the edge $l$ (with respect to a fixed background $N^{0}_{l}$) and $N_{\Phi}=\Phi/\Phi_{0}$ is the number of elementary flux quanta $\Phi$ of magnetic field piercing the antidot section ($\Phi_{0}=2\pi/|e|$ the elementary flux quantum). Notice that the last term in Eq.~(\ref{H_zero}) is reminiscent of the minimal Aharonov-Bohm coupling $\textbf{j} \cdot \textbf{A}$, with $\textbf{j}$ the current density along the edge of the antidot and $|\textbf{A}|= \Phi/ L$ the vector potential felt by the antidot itself. \cite{Geller97} In the following, we consider a finite length $L_{ad}=L$ only for the antidot with a consequent non-zero energy $E^{ad}_{c}=E_{c}$, while we assume the thermodynamic limit $L_{R}, L_{L}\rightarrow+\infty$ for the $R$ and $L$ edges (thus yielding $E^{R}_{c}, E^{L}_{c}\rightarrow 0$). Under these conditions the zero mode contribution plays a relevant role only for the dynamics of the antidot, while the other edges are only described in terms of their plasmonic modes.
As long as the Hall fluid is present between the antidot and the boundaries of the Hall bar (see Fig.~\ref{fig1}), the dominant tunneling process involves single-QPs with charge $e^{*} = \nu e$ ($e$ the electron charge). \cite{Kane92} Therefore the local tunneling Hamiltonian connecting the antidot with the edges of the Hall bar is given by
\begin{equation}
H^{T}_{j}= v \left[t_{j}\Psi^{\dagger}_{ad}(x_{j}) \Psi_{j}(0)+H. c.\right] \quad \text{with } j= L, R.
\end{equation}
Here, the tunneling amplitudes $t_{j}$ are related to the overlap between the Laughlin wave-functions \cite{Levkivskyi10} on the different edges and decay exponentially with their distance. Tunneling processes occur at points $x_{j}$ (see Fig.~\ref{fig1}), whose precise location is not relevant as long as the tunneling is assumed to be local. \cite{Chevallier10, Dolcetto12} The vertex operator associated with the annihilation of one single-QP can be written, according to the standard bosonized description,\cite{Miranda03} as
\begin{equation}
\Psi_{l}(x)= \frac{1}{\sqrt{ 2\pi \alpha}} e^{i \varphi_{l}(x)} e^{i \pi \nu \frac{x}{L_{l}}}.
\label{Psi}
\end{equation}
The bosonic field $\varphi_{l}(x)$ appearing in the exponent can be naturally decomposed into the sum of a plasmonic ($\varphi^{p}_{l}$) and a zero mode ($\varphi^{0}_{l}$) contribution given respectively by
\begin{eqnarray}
\varphi^{p}_{l}(x)&=& \sqrt{\frac{2 \pi \nu }{L_{l}}}\sum_{k_{l}>0}\left(\frac{1}{\sqrt{k_{l}}} a_{l, k_{l}}e^{i k_{l} x} +H.c.\right)e^{-k_{l}\alpha/2} \nonumber\\
\varphi^{0}_{l}(x)&=&\frac{2 \pi }{L_{l}} \nu N_{l}x-\chi_{l}
\label{modes}
\end{eqnarray}
with $\alpha$ a finite length cut-off. The operator $\chi_{l}$ satisfies
\begin{equation}
[\chi_{l}, N_{l'}]= i \delta_{l, l'}
\end{equation}
and, once exponentiated, plays the role of a Klein factor, which is essential to provide the correct exchange statistical properties between excitations from different edges. \cite{Braggio06, Merlo07, Guyon02} Notice that the last phase factor in Eq.~(\ref{Psi}) has been introduced in order to satisfy the boundary conditions
\begin{equation}
\Psi_{l}(x+L_{l})= \Psi_{l}(x) e^{i 2\pi \nu N_{l}}
\end{equation}
counting the number of fractional excitations enclosed by the edge. \cite{Geller97, Geller00}
\section{Transport properties} \label{Transport}
\subsection{Tunneling rates}
To lowest order in the tunneling Hamiltonian the transport properties depend on the tunneling rates which, by exploiting the periodicity associated with the antidot and assuming the plasmon modes fully relaxed to thermal equilibrium, can be written as \cite{Braggio06, Merlo07, Braggio01, Braggio03, Cavaliere04}
\begin{equation}
\tilde{ \Gamma}_{j}(E)=\sum_{p=-\infty}^{+\infty} w_{p}\gamma_{j}(E-p\epsilon) ,
\end{equation}
with $\epsilon$ defined in Sec. \ref{Model}.
This combines the standard expression for the tunneling rate at finite temperature and infinite length of the edges \cite{Roddaro03, Martin05}
\begin{align}
\gamma_{j}(\xi) =& |t_{j}|^{2} \frac{\omega_{c}}{(2\pi)^{2}} \left(\frac{\beta \omega_{c}}{2 \pi} \right)^{1-\nu} \nonumber \\
& \qquad \times B\left[\frac{\nu}{2} +i\frac{\beta \xi}{2\pi}, \frac{\nu}{2} -i\frac{\beta \xi}{2\pi} \right]e^{\beta \xi/2},
\label{rate_gamma}
\end{align}
with the correction accounting for the finite length of the antidot\cite{Braggio00}
\begin{equation}
w_{p}= \left(\frac{\epsilon}{\omega_{c}} \right)^{\nu}e^{-p \epsilon/\omega_{c}} \frac{\Gamma(\nu +p)}{\Gamma(\nu)p!}\Theta(p+0^{+})
\label{wp}
\end{equation}
evaluated analytically in the very low temperature limit $\beta \epsilon \gg 1$ ($\beta$ being the inverse temperature) and for $\epsilon/\omega_{c} \ll 1$. Here, we introduced the energy cutoff $\omega_{c}=v/\alpha$, and used $B[a, b]$ and $\Gamma(a)$ as Euler's Beta and Gamma functions respectively.
These tunneling rates present a peaked structure near $E/E_{c}\approx 0$, with a maximum value decaying exponentially for negative energy (with a scale set by temperature), and as a power-law for positive one. Similar peaks also appear at $E/E_{c}\approx 2p/\nu$ ($p\in \mathbb{N}^{0}$) as a consequence of plasmonic excitations, only with a less pronounced amplitude due to the damping factor $w_p$.
When the tunneling rates are small compared to both the temperature and the charging energy ($\tilde{\Gamma}_{j}<\beta^{-1}<E_{c}$), it is possible to restrict the analysis to the sequential regime, \cite{Beenakker91, Furusaki98} where only single-QP tunneling processes involving one excitation (incoming or outgoing with respect to the antidot) contribute to the dynamics of the system. In this approximation the relevant rates are the transition probabilities between the initial antidot state with $N$ excitations at time $t=0$ and the final one with $N\pm 1$ excitations at time $t$.
\subsection{Master equation}
A simple and useful way to generalize this idea involves the master equation approach, which allows to characterize the time evolution of the probability of occupation of the antidot. \cite{Braggio06, Merlo07, Furusaki98, Geller00} In particular, the probability $\mathcal{P}_{N}(t)$ of having a fixed number $N$ of QPs in a strongly asymmetric antidot ($|t_{L}|\gg |t_{R}|$) at a given time $t$ satisfies the first order differential equation
\begin{align}
\frac{d {\cal P}_N}{d t} =& \sum_{N'} \left[ \tilde{\Gamma} \left( E^{N' \to N}\right) {\cal P}_{N'} - \tilde{\Gamma} \left( E^{N \to N'}\right) {\cal P}_{N} \right] ,
\label{master_general}
\end{align}
(dropping the subscript $j$ in $\tilde{\Gamma}_{j}$ because of the strong asymmetry), where the transition energies are given by
\begin{equation}
E^{N \to N'} = E_c \left[ \left( N - N_\phi \right)^2 - \left( N' - N_\phi \right)^2 \right].
\end{equation}
Since we want to tune the number of QPs on the antidot, we allow modifications of $N_\phi$, with respect to a reference value,\cite{Note2} chosen to be a half-integer, so that
\begin{equation}
N_\phi = N_0 + \frac{1}{2} + \delta(t) .
\label{eq:Nphi}
\end{equation}
Setting $n = N-N_0$, $p = N'-N_0$ and introducing $\hat{{\cal P}}_n = \left. {\cal P}_N \right|_{N = n+N_0}$, the master equation can be conveniently rewritten under a matrix form as
\begin{equation}
\frac{d \hat{{\cal P}} (t)}{d t} = \hat{\Gamma} \hat{{\cal P}} (t) ,
\label{eq:mastermatrix}
\end{equation}
where $\hat{{\cal P}} (t)$ is a column vector whose elements are the occupation probabilities $\hat{{\cal P}}_n (t)$, and $\hat{\Gamma}$ is a square matrix whose elements are given by
\begin{align}
\hat{\Gamma}_{n p} =& \tilde{\Gamma} \left( -(n-p) (n+p-1-2 \delta(t)) E_c \right) \nonumber\\
&- \left[ \sum_{q \in \mathbb{Z}} \tilde{\Gamma} \left( (n-q) (n+q-1-2 \delta(t)) E_c \right) \right] \delta_{n p}.
\label{eq:gammamatrix}
\end{align}
In practice, while various possibilities are foreseeable, we will mostly focus on a square drive, defined over one drive period $T$ as $\delta (t) = \delta\times \text{Sgn} \left( \frac{T}{2} - t\right)$.
\subsection{Occupation number, current and noise}
The occupation ${\cal N} (t)$ of the antidot is readily obtained upon summing up the occupation probabilities solution of the master equation weighted by their corresponding number of QPs
\begin{equation}
{\cal N} (t) = \sum_{n \in \mathbb{Z}} n \hat{{\cal P}}_n (t).
\label{eq:dot_occ}
\end{equation}
Of course, this occupation is defined with respect to the background reference set by $N_0$ (cf. Eq.~\eqref{eq:Nphi}).
The total charge on the antidot is readily obtained from the occupation as $Q (t) = e^* {\cal N}(t)$. It follows that the current flowing from the antidot to the edge is simply given by
\begin{equation}
I (t) = - \frac{d Q (t)}{dt} = - e^* \sum_{n \in \mathbb{Z}} \sum_{p \in \mathbb{Z}} n \hat{\Gamma}_{n p} \hat{{\cal P}}_p (t) .
\end{equation}
While the current provides crucial information on the operation of the antidot source, relevant information can also be obtained from its noise characteristics. Indeed, the finite frequency signal of the current current correlations allows a finer characterization of the operating conditions of the QP or electron emitter, especially with regard to its dependence on the escape time. \cite{Mahe10,Parmentier12}
When the escape time is much smaller than the period of the drive, the antidot emits
QP or QH (alternatively, electrons and holes) in essentially a periodic manner and the main contribution to the noise is due to the uncertainty of the emission time within each period. This constitutes the regime of so-called ``phase noise'', which is due to the random jitter of triggering of the drive and the actual emission time. On the opposite, when the escape time of QP (QH) (alternatively, electrons or holes) is much larger than the period of the drive, nothing guarantees that the escape from the antidot really occurs, and these rare events give rise to a shot noise like contribution.
The master equation which is employed here allows to describe both regimes as well as the crossover between the two. However, prior noise experiments performed with the mesoscopic capacitor as the emitter tend (for convenience) to measure the noise at a frequency close to that of the drive, and chose instead to modify the escape time by tuning the transmission of the capacitor coupled to the edge, in order to explore the range of parameters.
We define the current-current correlation as:
\begin{equation}
{\cal C}_I (t, t') = \langle \delta I (t) \delta I (t+t') \rangle ,
\end{equation}
where $\delta I (t) = I(t) - \langle I(t) \rangle$.
Because of the periodic drive, the current-current correlation ${\cal C}_I (t, t')$ depends on both $t$ and $t'$, and is $T-$periodic in time $t$. Since we are only interested in the behavior with respect to the time difference $t'$, we consider a time-averaged quantity defined as
\begin{equation}
{\cal S}_I (t') = 2 \int_0^T \frac{dt}{T} \langle \delta I (t) \delta I (t+t') \rangle = 2 \overline{\langle \delta I (t) \delta I (t+t') \rangle} ,
\end{equation}
and the corresponding quantity in frequency space
\begin{equation}
{\cal S}_I (\omega) = \int dt' {\cal S}_I (t') e^{i \omega t'} .
\end{equation}
As it turns out, the simplest way to derive the current noise is to first access the charge noise as the two are trivially related: ${\cal S}_I (\omega) = \omega^2 {\cal S}_Q (\omega)$.\cite{Albert10} Let us first focus on the charge correlation function $\langle Q(t) Q(t+t') \rangle$ and consider for simplicity that $t' > 0$.
The charge correlation is only finite when the dot is occupied both at time $t$ and at time $t+t'$. The actual value of the charge correlation is obtained by summing over all possible occupations $n_1$ and $n_2$, multiplied by the joint probability of having $n_1$ QPs at time $t$ and $n_2$ QPs at time $t+t'$. \cite{Korotkov94} The latter is further written, using conditional probabilities, as the product of the probability $\hat{{\cal P}} (n_1,t)$ of having $n_1$ QPs occupying the antidot at time $t$, and the conditional probability $\hat{{\cal P}}(n_2,t+t' | n_1, t)$ of having $n_2$ QPs at time $t+t'$ given that there were $n_1$ at time $t$. One is left with
\begin{equation}
\langle Q(t) Q(t+t') \rangle = {e^*}^2 \sum_{n_1, n_2} n_1 n_2 \hat{{\cal P}}(n_1,t) \hat{{\cal P}} (n_2,t+t' | n_1, t).
\end{equation}
Numerically, this conditional probability is obtained by propagating the condition $\hat{{\cal P}} (n_1,t) = 1$ through time, using the master equation in Eq.~(\ref{eq:mastermatrix}).
Performing the same calculation for negative values of $t'$, and accounting for the average charge, we have
\begin{align}
{\cal C}_Q (t, t') &= \langle \delta Q (t) \delta Q (t+t') \rangle \nonumber \\
&= {e^*}^2 \sum_{n_1, n_2} n_1 n_2
\left\{ \theta(t') \hat{{\cal P}}(n_1,t) \Delta \hat{{\cal P}} (n_2,t+t' | n_1,t) \right. \nonumber \\
& + \left. \theta(-t') \hat{{\cal P}}(n_2,t+t') \Delta \hat{{\cal P}} (n_1,t | n_2,t+t') \right\}
\label{eq:cqtt}
\end{align}
with $\Delta \hat{{\cal P}} (n_1,t_1 | n_2,t_2) = \hat{{\cal P}}(n_1,t_1 | n_2, t_2) - \hat{{\cal P}}(n_1,t_1)$.
Computing the average over $t$ and taking the Fourier transform, this ultimately leads to the following expression for the frequency-dependent current noise
\begin{align}
{\cal S}_I (\omega) &= 4 {e^*}^2 \omega^2 \int_0^{\infty} dt' \cos \left( \omega t' \right) \int_0^T \frac{dt}{T} \nonumber \\
& \times \sum_{n_1, n_2} n_1 n_2 \hat{{\cal P}}(n_1,t) \Delta \hat{{\cal P}} (n_2,t+t' | n_1,t) .
\end{align}
\subsection{Tuning the drive}
We now want to determine the optimal regime of operation of the source.
First, we need to ensure that the antidot has sufficient time to emit/absorb QPs every half-period of the drive. To meet that goal, let us consider the situation of a rather large tunneling rate, say $\gamma_0 T \simeq 100$ where
\begin{equation}
\gamma_0 =\left( \frac{ 2 E_c}{\nu \omega_c} \right)^\nu |t_L|^2 \frac{\omega_c}{(2 \pi)^2} \left( \frac{\beta \omega_c}{2 \pi} \right)^{1-\nu} \frac{\Gamma \left( \frac{\nu}{2}\right)^2}{\Gamma(\nu)},
\end{equation}
is the maximum tunneling rate between any two energy levels, and follow the evolution of the dot occupation over time for different values of the drive amplitude $\delta$.
The results for the occupation ${\cal N} (t)$ are presented in Fig.~\ref{fig2} at filling factor $\nu = 1/3$, for a given set of parameters satisfying the constraints of our model, namely $\beta E_c \gg 1$ and $E_c/\omega_c \ll 1$.
These two plots clearly represent the two regimes of interest for the source. The upper panel shows variations of the antidot occupation between 0 and 1, corresponding to the emission/absorption of a single QP. The lower panel shows variations between $-1$ and 2, corresponding to the emission/absorption of an electron charge (in the form of 3 QPs, consistently with the standard picture for the fractional states in the Laughlin sequence). Notice that the negative occupation is an artifact of the choice of the background reference, as the two extrema of the occupation are symmetric with respect to $N_\phi - N_0 = 1/2$.
The optimal regime of operation corresponds to a situation where the occupation is close to an integer value at the end of every half-period, while behaving monotonously over each half-period. From the results of Fig.~\ref{fig2}, it seems the most appropriate choice for the drive amplitude is $\delta = -0.2$ and $\delta = -1.2$ for the SQS and the SES regime respectively.
Strikingly, the evolution of the two types of occupation in the upper and lower panel (QP and electrons) bear strong similarities. When the drive amplitude is larger than the optimal drive in both cases, one sees that upon imposing the AC drive the occupation of the antidot overshoots the zero or single QP or electron occupation: for QP (electrons), the trigger of the drive brings the antidot in a configuration which is slightly less than zero, while at the half period, the occupation is on average larger than one. This overshoot is slightly higher in the electron case than in the QP case. In the optimal case ($\delta = -0.2$ and $\delta = -1.2$ respectively), the response to the drive bears strong similarities with that of the mesoscopic capacitor.
\begin{figure}[htb]
\centering
\includegraphics[scale=0.62]{Antidot_Fig2.pdf}
\caption{Occupation of the antidot over time as a function of the drive amplitude $\delta$ for the SQS (top) and the SES (bottom) regime. Other parameters are: $\nu = 1/3$, $E_c/\omega_c = 0.01$, $\beta E_c = 20$ and $\gamma_0 T = 100$. }
\label{fig2}
\end{figure}
The remainder of this paper is devoted to the detailed study of these two specific cases. There, we will focus on the "stationary" state, still submitted to a periodic drive but independent of the initial conditions, namely the state of the system when the drive was turned on.
\section{antidot as a SQS} \label{SQS}
We solve numerically the master equation in Eq.~(\ref{master_general}) over several drive periods, considering a vector column $\hat{{\cal P}}$ of 31 elements ($n=-15,...,15$) using a matrix version of the fourth-order Runge-Kutta scheme. As it turns out, most occupation probabilities are vanishingly small in this case, and one can focus on a very much reduced set of equations in order to properly describe the behavior of the system, allowing for an analytic treatment.
In practice, one thus only needs to keep track of the probabilities for having a singly occupied or an empty antidot, namely $\hat{{\cal P}}_1 (t)$ and $\hat{{\cal P}}_0 (t)$ respectively. Focusing on a single half-period, corresponding e.g. to the emission of a QP, their dynamics is captured by the following set of equations
\begin{align}
\frac{d \hat{{\cal P}}_1 (t)}{d t} &= -\gamma \hat{{\cal P}}_1 (t) , \\
\frac{d \hat{{\cal P}}_0 (t)}{d t} &= \gamma \hat{{\cal P}}_1 (t) ,
\end{align}
where $\gamma = -\hat{\Gamma}_{11} = \tilde{\Gamma} (-2 E_c \delta)$. The other half-period, corresponding to the absorption of a QP, is described by similar equations, only exchanging $\hat{{\cal P}}_0$ and $\hat{{\cal P}}_1$. Note that by definition, $\gamma (\delta =0) = \gamma_0$ and that $\gamma$ then rapidly decreases as one raises $|\delta|$.
This set of equations is simple enough to be propagated analytically. The occupation probability $\hat{{\cal P}}_1 (t)$ is thus given by
\begin{align}
\hat{{\cal P}}_1 (t) &= \hat{{\cal P}}_1 (0) e^{-\gamma t} & \text{for~} 0 \leq t \leq \frac{T}{2} \nonumber \\
\hat{{\cal P}}_1 (t) &= 1 + e^{-\gamma t} \left( \hat{{\cal P}}_1 (0) - e^{\gamma T/2} \right) & \text{for~} \frac{T}{2} \leq t \leq T.
\label{eq:solSQS}
\end{align}
while $\hat{{\cal P}}_0 (t) = 1 - \hat{{\cal P}}_1(t)$ at all times. The stationary state requires that $\hat{{\cal P}}_1(0) = \hat{{\cal P}}_1(T)$, so that $\hat{{\cal P}}_1(0) = \frac{\exp (\gamma T/4)}{2 \cosh (\gamma T/4)}$. Numerical results for the evolution of these two probabilities over time are presented in Fig.~\ref{fig3} for a given set of parameters at filling factor $\nu=1/3$.
\begin{figure}[t]
\centering
\includegraphics[scale=0.62]{Antidot_Fig3.pdf}
\caption{Occupation probability of the two relevant levels of the antidot ($\hat{\mathcal{P}}_{0}$ and $\hat{\mathcal{P}}_{1}$) for the optimal regime of the SQS ($\delta = -0.2$). Other parameters are: $\nu = 1/3$, $E_c/\omega_c = 0.01$, $\beta E_c = 20$ and $\gamma_0 T = 100$. The relevant scale entering the master equation is then $\gamma \simeq 17.4512 ~T^{-1}$.}
\label{fig3}
\end{figure}
\subsection{Dot occupation and current}
\begin{figure}[htb]
\centering
\includegraphics[scale=0.62]{Antidot_Fig4.pdf}
\caption{Occupation of the antidot (top) and current flowing between the source and the edge in units of $\frac{e^*}{T}$ (bottom), in the optimal regime of SQS ($\delta=-0.2$) for different values of $\gamma_0 T$. Other parameters are: $\nu = 1/3$, $E_c/\omega_c = 0.01$ and $\beta E_c = 20$.}
\label{fig4}
\end{figure}
The occupation of the dot is obtained directly from the occupation probabilities and Eq.~\eqref{eq:dot_occ}. Here this trivially reduces to the much simpler form
\begin{equation}
{\cal N} (t) = \hat{{\cal P}}_1 (t) ,
\end{equation}
and similarly for the current
\begin{equation}
I (t) = - e^* \frac{d \hat{{\cal P}}_1 (t)}{dt} .
\end{equation}
The evolution over time of these two quantities, evaluated numerically for different values of $\gamma_0 T$, is presented in Fig.~\ref{fig4}.
Note that, because of the simple correspondence between occupation of the antidot and probability of single occupation, the escape time $\tau$ defined as the typical time associated with the emission/absorption process also corresponds to the typical time-scale governing the evolution of the two relevant occupation probabilities. In other words, one trivially has $\tau = \gamma^{-1}$.
The solution, Eq.~\eqref{eq:solSQS}, also allows to estimate the average charge transferred to the edge during every half-period (in absolute value)
\begin{align}
\overline{Q} &= e^* \left| {\cal N} (T/2) - {\cal N} (0) \right| \nonumber \\
&= e^* \tanh \left( \frac{\gamma T}{4} \right) .
\end{align}
confirming that the ideal operating regime is the one where $\gamma T \gg 1$ ($\overline{Q}\rightarrow e^{*}$).
\subsection{Charge fluctuations}
The time-averaged charge fluctuations can readily be derived from the computation of the conditional occupation probabilities following Eq.~\eqref{eq:cqtt}.
The results are presented in Fig.~\ref{fig5}. They show that the time-averaged charge correlation $\overline{{\cal C}_Q (t,t')}$ vanishes exponentially with the same characteristic time scale as the dot occupation, i.e. $\gamma^{-1}$. The value taken at $t'=0$ can be readily estimated from the expression for $\hat{{\cal P}}_1 (t)$ and reads
\begin{equation}
\overline{{\cal C}_Q (t,0)} = {e^*}^2 \frac{\tanh (\gamma T/4)}{\gamma T} .
\end{equation}
One can verify that the charge fluctuations are thus given by the following exponentially decaying form
\begin{equation}
\overline{{\cal C}_Q (t,t')} = {e^*}^2 \frac{\tanh (\gamma T/4)}{\gamma T} e^{- |\gamma t'|} ,
\end{equation}
which is very reminiscent of what was obtained for the SES in the integer quantum Hall regime. \cite{Mahe10, Albert10, Parmentier12}
\begin{figure}[ht]
\centering
\includegraphics[scale=0.62]{Antidot_Fig5.pdf}
\caption{Time-averaged charge correlation $\overline{{\cal C}_Q (t,t')}$ in units of ${e^*}^2$, for the optimal regime of the SQS ($\delta= -0.2$) and different values of $\gamma_0 T$. Other parameters are: $\nu = 1/3$, $E_c/\omega_c = 0.01$ and $\beta E_c = 20$.}
\label{fig5}
\end{figure}
\subsection{Noise at the drive frequency}
Experimentally, the most accessible fluctuation-related quantity is the current noise probed at the frequency of the drive. \cite{Mahe10, Parmentier12} We computed this frequency-dependent current noise ${\cal S}_I (\omega)$ and evaluated it at the drive frequency $\Omega = \frac{2 \pi}{T}$ for different values of $\gamma_0 T$, or equivalently different values of the escape time $\tau = \gamma^{-1}$. The results are provided in Fig.~\ref{fig6}.
\begin{figure}[htb]
\centering
\includegraphics[scale=0.62]{Antidot_Fig6.pdf}
\caption{Frequency-dependent current noise ${\cal S}_I$ in units of $\frac{{e^*}^2}{T}$, computed at the frequency $\Omega$ of the external drive as a function of the escape time $\tau$ from the antidot, for the optimal regime of the SQS ($\delta = -0.2$). The result is compared to the analytical expectation transposed from the known results of the SES in the IQH case, assuming a fractional charge $e^{*}= \nu e$. Other parameters are: $\nu = 1/3$, $E_c/\omega_c = 0.01$ and $\beta E_c = 20$.}
\label{fig6}
\end{figure}
Again, it makes sense here to compare the obtained behavior to the one expected for the SES in the IQH regime, as the two are governed by a similar set of equations in the present regime of operation. \cite{Albert10, Parmentier12} Not surprisingly, the current noise obtained here reproduces exactly the result of the SES upon changing $e$ into $e^*$, namely
\begin{equation}
{\cal S}_I (\Omega) = \frac{4 {e^*}^2}{T} \tanh \left( \frac{T}{4 \tau} \right) \frac{\left( \Omega \tau \right)^2 }{1+ \left( \Omega \tau \right)^2} .
\label{eq:noiseSQS}
\end{equation}
This excellent agreement means that there are no spurious processes in the system, such as missed emissions compensated by double emissions, as well as QP/QH pairs. The source
of the observed noise for $\tau/T <1$ is the incertitude in the time of emission of the excitations during the half-period, what is usually referred to as \emph{phase} or \emph{jitter} noise. \cite{Mahe10}
The considered emission process is therefore noiseless at long time-scales (see App. \ref{AppA}).
\section{The antidot as a SES} \label{SES}
We now turn to the case of emission/absorption of a single electron, by increasing the amplitude of the drive compared to the previous case of the SQS. Our results still rely on the numerical solution of the master equation over several drive periods, considering a vector column $\hat{{\cal P}}$ of 31 elements ($n=-15...15$) and using a matrix version of the fourth-order Runge-Kutta scheme.
Interestingly, this regime is much more complicated than the SQS. While one could have hoped to deal with only two occupation probabilities just as before (say $\hat{{\cal P}}_2$ and $\hat{{\cal P}}_{-1}$), it turns out not to be sufficient to properly describe the SES, which here cannot be reduced to a simple analytic treatment.
In practice, even in the best case scenario (very large value of $\gamma_0 T$) one needs to keep track of 8 different occupation probabilities in order to account for the behavior of the source over one drive period, and 6 if one focuses on only a half-period (the other 2 being recovered by symmetry). Numerical results for the time evolution of these 8 probabilities are presented in Fig.~\ref{fig8}, for $\gamma_0 T = 100$.
Focusing on a single half-period, corresponding e.g. to the emission of an electron, their dynamics is captured by the following set of equations
\begin{align}
\label{eq:pm3el}
\frac{d \hat{{\cal P}}}{dt} &= \mathcal M \hat{{\cal P}} \quad \quad \quad \quad \mbox{ with }\\
\mathcal M & = \left ( \begin{array}{cccccc}
\hat{\Gamma}_{-3,-3} & 0 & 0 & 0 & 0 & \hat{\Gamma}_{-3,2} \\
\hat{\Gamma}_{-2,-3} & \hat{\Gamma}_{-2,-2} & 0 & 0 & \hat{\Gamma}_{-2,1} & \hat{\Gamma}_{-2,2}\\
\hat{\Gamma}_{-1,-3} & \hat{\Gamma}_{-1,-2} & 0 & \hat{\Gamma}_{-1,0} & \hat{\Gamma}_{-1,1} & \hat{\Gamma}_{-1,2}\\
\hat{\Gamma}_{0,-3} & \hat{\Gamma}_{0,-2} & 0 & \hat{\Gamma}_{0,0} & \hat{\Gamma}_{0,1} & \hat{\Gamma}_{0,2}\\
\hat{\Gamma}_{1,-3} & 0 & 0 & 0 & \hat{\Gamma}_{1,1} & \hat{\Gamma}_{1,2}\\
0 & 0 & 0 & 0 & 0 & \hat{\Gamma}_{2,2}
\end{array}
\right)
\label{eq:p2el}
\end{align}
where we omitted the time dependence for notational convenience. Here, the matrix elements $\hat{\Gamma}_{n p}$ were defined in Eq.~\eqref{eq:gammamatrix}, and we kept only the non-vanishing contributions. The other half-period, corresponding to the emission of an electron, is readily obtained from the same set of equations upon exchanging $\hat{{\cal P}}_{-3} \leftrightarrow \hat{{\cal P}}_{3}$, $\hat{{\cal P}}_{-2} \leftrightarrow \hat{{\cal P}}_{4}$, $\hat{{\cal P}}_{-1} \leftrightarrow \hat{{\cal P}}_{2}$ and $\hat{{\cal P}}_{0} \leftrightarrow \hat{{\cal P}}_{1}$ (corresponding to the exchange between similar line styles in the top and bottom panels of Fig.~\ref{fig7}).
\begin{figure}[htb]
\centering
\includegraphics[scale=0.62]{Antidot_Fig7.pdf}
\caption{Occupation probability of the relevant levels for the optimal regime of the SES ($\delta = -1.2$), with $\gamma_0 T = 100$. Other parameters are: $\nu = 1/3$, $E_c/\omega_c = 0.01$ and $\beta E_c = 20$.}
\label{fig7}
\end{figure}
A few comments are in order from the observation of this reduced set of equations:
\begin{itemize}
\item the apparent scattering that results in having the occupation of the antidot oscillating between 2 and $-1$ does not involve the direct exchange of 3 quasiparticles, but rather processes involving between 1 and 5 QPs,
\item during the emission process, $\hat{{\cal P}}_2$(t) follows an exponential decay, which introduces the characteristic time-scale $\tau_2 = \hat{\Gamma}_{2,2}^{-1}$,
\item the above set of equations can be solved by successively substituting the solution of the known probabilities into the next, according to the following order: $\hat{{\cal P}}_{2} \to \hat{{\cal P}}_{-3} \to \hat{{\cal P}}_{1} \to \hat{{\cal P}}_{-2} \to \hat{{\cal P}}_{0} \to \hat{{\cal P}}_{-1}$. This means in particular that all relevant occupation probabilities can be viewed as a sum of exponential terms of the form $\exp \left( \hat{\Gamma}_{n,n} t \right)$ [note that all $\hat{\Gamma}_{n,n}$ are negative according to Eq.~(\ref{eq:gammamatrix})].
\end{itemize}
This picture gets even more involved as one reduces the value of $\gamma_0 T$. For $\gamma_0 T = {\cal O} (1)$, about 10 occupation probabilities are required in order to properly describe the electron source, signaling the importance of processes which involve transferring up to 9 QPs.
\subsection{Dot occupation and current}
The occupation of the dot is obtained directly from the occupation probabilities and Eq.~\eqref{eq:dot_occ}, and similarly for the current. The evolution over time of these two quantities, evaluated for different values of $\gamma_0 T$, is presented in Fig.~\ref{fig8}. Like the individual occupation probabilities, the antidot occupation can be written as a weighted sum of exponential terms of the form $\exp (\hat{\Gamma}_{n,n} t)$, as a direct consequence of Eq.~\eqref{eq:dot_occ}.
\begin{figure}[htb]
\centering
\includegraphics[scale=0.62]{Antidot_Fig8.pdf}
\caption{Occupation of the antidot (top) and current flowing between the source and the edge in units of $\frac{e^*}{T}$ (bottom), in the optimal regime of SES ($\delta=-1.2$) for different values of $\gamma_0 T$. Other parameters are: $\nu = 1/3$, $E_c/\omega_c = 0.01$ and $\beta E_c = 20$.}
\label{fig8}
\end{figure}
Note that, in contrast with the SQS, there is no proper way of unambiguously defining the escape time from the antidot. Following Eqs.~\eqref{eq:pm3el}-\eqref{eq:p2el}, one might be tempted to introduce the characteristic time-scale $\tau_{2}$ associated with the exponential decay of $\hat{{\cal P}}_2$, the probability for single electron occupation of the antidot. However, these equations are not sufficient in the regime of low $\gamma_0 T$, and $\hat{{\cal P}}_2$ no longer follows quite the same simple exponential decay. Similarly, one might want to define the escape time from the antidot $\tau_\text{ad}$ directly from the antidot occupation, e.g. from its initial rapid decay. However, since the occupation is a weighted sum of different exponential terms, one can only extract estimates from the early rapid decay.
A more convenient way is to follow the experimental procedure and extract the escape time from the first harmonics of the current. Indeed, one can write the first harmonics $I_\Omega = \frac{1}{T} \int_0^T dt I(t) e^{i\Omega t}$ as $I_\Omega = |I_\Omega| e^{i \phi}$ and an exponential decay of the current leads back to $\tan \phi = \Omega \tau$. Transposing this to the present quantities, one can define the escape time
\begin{align}
\tau_\Omega = - \frac{T}{2 \pi} \frac{\int_0^T dt ~{\cal N}(t) \cos \left( \Omega t \right)}{\int_0^T dt~ {\cal N} (t) \sin \left( \Omega t \right)}.
\end{align}
The average charge $\overline{Q}$ transfered to the edge during every half-period was computed numerically from the obtained solution to the master equation and is presented in Fig.~\ref{fig9}. It qualitatively behaves as in the case of the SQS, i.e. it shows a plateau at an integer value (here 3 corresponding to the fact that one electron is equal to a bunch of $3$ QPs in the Laughlin regime at $\nu=1/3$) for the lowest escape times then decreases rapidly, confirming that the ideal operating regime is the one where $\tau \ll T$. Note however that it cannot be written as simply as before, and the form $e \tanh(T/(4 \tau))$ (with any of the three definitions of $\tau$) though providing the good qualitative behavior, does not fit exactly the numerical data.
\begin{figure}[tb]
\centering
\includegraphics[scale=0.62]{Antidot_Fig9.pdf}
\caption{Average charge $\overline{Q}$ transfered during one half-period, in units of $e^*$, as a function of the escape time (corresponding to different values of $\tau_\text{dot}$, $\tau_2$ and $\tau_\Omega$) and for the optimal regime of the SES ($\delta = -1.2$). Other parameters are: $\nu = 1/3$, $E_c/\omega_c = 0.01$ and $\beta E_c = 20$.}
\label{fig9}
\end{figure}
\subsection{Charge fluctuations}
The time-averaged charge fluctuations are derived from the computation of the conditional occupation probabilities. The results are presented in Fig.~\ref{fig10}.
\begin{figure}[b]
\centering
\includegraphics[scale=0.62]{Antidot_Fig10.pdf}
\caption{Time-averaged charge correlation $\overline{{\cal C}_Q (t,t')}$ in units of ${e^*}^2$, for the optimal regime of the SES ($\delta= -1.2$) and different values of $\gamma_0 T$. Other parameters are: $\nu = 1/3$, $E_c/\omega_c = 0.01$ and $\beta E_c = 20$.}
\label{fig10}
\end{figure}
As already observed in the SQS case, the time-averaged charge correlation $\overline{{\cal C}_Q (t,t')}$ vanishes rapidly with a characteristic time-scale which seems to be similar in value to $\tau_\text{dot}$, $\tau_2$ and $\tau_{\Omega}$, without being quite exactly equal to any of those. This can be understood from Eq.~\eqref{eq:cqtt}, as $\overline{{\cal C}_Q (t,t')}$ appears as a sum of various exponential contributions, themselves obtained by combining two different terms of the form $\exp(\hat{\Gamma}_{n,n} t)$.
\subsection{Noise at the drive frequency}
We consider now the frequency-dependent current noise ${\cal S}_I (\omega)$ associated with the SES and evaluate it at the drive frequency $\Omega = \frac{2 \pi}{T}$ for different values of $\gamma_0 T$, or equivalently different values of the escape time $\tau$ (evaluated in three separate ways: $\tau_\text{dot}$, $\tau_2$ and $\tau_\Omega$). The results are provided in Fig.~\ref{fig11}.
\begin{figure}[t]
\centering
\includegraphics[scale=0.62]{Antidot_Fig11.pdf}
\caption{Frequency-dependent current noise ${\cal S}_I$ in units of $\frac{{e^*}^2}{T}$ computed at the frequency $\Omega$ of the external drive as a function of the escape time $\tau$ from the antidot, for the optimal regime of the SES ($\delta = -1.2$). Other parameters are: $\nu = 1/3$, $E_c/\omega_c = 0.01$ and $\beta E_c = 20$. The results are compared to the analytical expectation transposed from the SQS regime, Eq.~\eqref{eq:noiseSQS}, multiplied by 9 to account for the fact that the emitted charge now corresponds to one electron (3 QPs).}
\label{fig11}
\end{figure}
It makes sense here to compare the obtained behavior to the one observed in the case of the SQS, correcting for the increased charge of the carriers. Since the noise is sensible to the charge squared, we plot for comparison the result obtained in Eq.~\eqref{eq:noiseSQS} for QP injection in the previous section, multiplied by 9 (to account for the fact that the source now emits/absorbs 3 QPs per half-period). While the curves have qualitatively the same overall shape, the current noise of the SES remains much larger than the one for the SQS source, even when accounting for the increased transfered charge. This constitutes a clear indication of the importance of tunneling processes involving multiple QPs in this regime.
Indeed, as argued when deriving the set of equations for the occupation probabilities, Eqs~\eqref{eq:pm3el}-\eqref{eq:p2el}, a proper account of the behavior of the SES requires to consider tunneling processes involving up to 9 QPs whereas we only emit an average of 3 QPs or less per half-period. Although these processes are responsible for the significant increase of the current noise at finite frequency, thus contributing to the \emph{phase} (finite frequency) noise, they do not affect the quantization of the emitted charge per half-period (see Appendix \ref{AppA}) so that the numerically evaluated frequency-dependent noise still vanishes at zero frequency.
\section{Estimating the physical parameters of the SQS} \label{sec:estimates}
For our simulations, we have chosen specific values for the various parameters which satisfy the constraints of our model, but may otherwise look random. Here we show through a simple estimate of the scales at play in an actual realization of our setup that these quantities are actually compatible with what is typically observed in electron quantum optics experiments carried out in the integer Hall case.\cite{Feve07, Mahe10, Bocquillon13b} In what follows, we reintroduce the proper factors of $\hbar$ and $k_B$ when needed.
\subsection{Energy scales}
\label{scales}
Focusing on the state at filling factor $\nu=1/3$, assuming a small antidot of circumference $L\approx 1$ $\mu$m \cite{Franklin96} and a propagation velocity along the edge $v\approx 10^{5}$ m/s, one has the charging energy
\begin{equation}
E_{c}/k_{B}\approx 0.8 \text{K}
\end{equation}
(expressed in units of temperature) which is smaller, but of the same order of magnitude as the energy gap observed in the SES in the IQH regime ($\Delta\approx 1.4$ to $4.2$ K). \cite{Feve07, Bocquillon13a, Bocquillon13b}
The plasmonic energy is larger
\begin{equation}
\epsilon/k_{B}\approx 4.8 \text{K},
\end{equation}
but plays only a marginal role in the dynamics of the device.
The temperature of the system, as assumed in the simulations, is
\begin{equation}
\Theta= 0.05 E_{c}/k_{B}\approx 40 \text{mK}.
\end{equation}
It is comparable to the one obtained in experiments involving the antidot geometry \cite{Franklin96, Goldman01} and not out of reach from nowadays experimental techniques which allow to carry out electron quantum optics measurements ($\Theta\approx 60$ to $100$ mK). \cite{Bocquillon13a, Bocquillon13b}
The optimal operation regime of the SQS discussed in the text is reached for a drive amplitude $\delta = -0.2$, which corresponds to a magnetic field fluctuation of
\begin{equation}
\Delta B= 0.2 ~\frac{4 \pi \Phi_{0}}{L^{2}}\approx 10^{-2} \text{T}.
\end{equation}
This represents a small and experimentally achievable fluctuation with respect to the magnetic field needed to realize the FQH states \cite{Goldman01, Goldman05} (typically of the order of $10\text{T}$, depending on the two dimensional electron density of the considered sample).
The high-frequency cutoff $\omega_{c}= v/\alpha$ represents the highest energy scale in the model. We assume that the short length cutoff $\alpha$ is typically set by the magnetic length of the system, so that $\alpha\approx 10$ nm. This in turn leads to $\omega_{c} \approx 10$ THz, or equivalently
\begin{equation}
\hbar \omega_{c}/k_B \approx 76 \text{K}
\end{equation}
from which one readily sees that $E_c/\omega_c \simeq 0.01$ as used in our simulations.
\subsection{Time scales}
In order to operate the antidot in the optimal regime of emission, the typical time-scales of the device need to satisfy the relation
\begin{equation}
\tau_{0} < \gamma^{-1} < \frac{T}{2}
\label{time_constraints}
\end{equation}
where $\tau_{0}=L/v$ is the time required to make a loop around the antidot.
The first inequality ensures the validity of the continuous limit assumed in the master equation approach, as $\tau_{0}$ corresponds to the discretization time associated with the semiclassical processes. \cite{Albert10} The second one guarantees that the emission of a QP (QH) is achieved during one half-period. According to the previous estimation one has
\begin{equation}
\tau_{0}=\frac{L}{v}\approx 10 \text{ps}.
\end{equation}
Assuming a driving frequency $f=500$ MHz, close to the GHz regime investigated in Refs. \onlinecite{Feve07, Mahe10}, the drive period is given by
\begin{equation}
T= \frac{1}{f}\approx 2 \text{ns}
\end{equation}
so that the half-period is of the order of the nanosecond.
This enforces the emission time from the SQS to vary in the range $10$ to $1000\text{ps}$, compatible with what is currently observed in similar experiments ($\gamma^{-1}\approx 60$ to $900\text{ps}$ in the optimal regime for the SES in the IQH case).\cite{Feve07, Bocquillon13a, Bocquillon13b} Note that any value of the emission time $\gamma^{-1}$ in this range also satisfies the condition
\begin{equation}
\hbar \gamma < k_{B} \theta < E_{c}
\end{equation}
necessary to fulfill the sequential tunneling approximation.\cite{Beenakker91, Furusaki98}
The maximum value of the emitted current is directly related to the emission time, as one has
\begin{equation}
I_\text{max} = e^* \gamma \frac{\exp(\gamma T/4)}{2 \cosh (\gamma T/4)}
\end{equation}
which corresponds to a value of the current between $10\text{pA}$ and $1\text{nA}$, a reasonable range in comparison with what is experimentally measured for the SES in the integer regime\cite{Feve07} and for the continuous current measurement in the weak backscattering regime at fractional filling factor.\cite{Chung03}
Finally, using our energy-scale estimates, we can propagate these bounds on the emission time and obtain
\begin{equation}
10^{-4} < |t_{L}|^{2} < 10^{-2},
\end{equation}
for the range of transmission amplitude between the antidot and the nearby edge channel.
\section{Conclusions} \label{Conclusions}
In this paper we have analyzed a strongly asymmetric antidot geometry realized through depletion of the Hall fluid, and periodically driven in time by a modulated magnetic flux. Through a master equation approach, we discussed the possibility to use this kind of setup both as a SQS and a SES. In the first case, only two charge states of the antidot are involved and the dynamics of the system allows for a tractable analytic treatment. It shows the quantization of a fractional charge $e^{*}=\nu e$ emitted during each half-period and fluctuations analogous to the ones observed for the electron emission in the integer quantum Hall case. The electron emission regime proves more complicated as various charge states of the antidot are involved, requiring a full numerical treatment.
Here, we observe that we can reach a regime where the emitted charge over a half-period is precisely $e$ (with the precise opposite charge emitted over the second half of the period).
However, the noise measured at the drive frequency is strongly enhanced with respect to what is observed in the integer regime due to the random emissions of additional excitations having zero mean charge (quasiparticle-quasihole pairs, etc.) which provide an essential contribution to the dynamics of the system at finite frequency.
In spite of the presence of these additional tunneling events, it is possible to extract information about the escape time of the electron by looking at the first current harmonic as is usually carried out in experiments.
It is worth underlying that in previous works, discussing similar geometries realized with Luttinger liquids \cite{Braggio03} (typically in CNT \cite{Cavaliere04} or two dimensional topological insulators \cite{Dolcetto13}) the role of the magnetic flux is played by an external gate voltage. The possibility to exploit a fluctuation of the same gate voltage used to realize the antidot also to induce the QP injection could be fruitful from the experimental point of view, an external electrostatic gate being easier to tune than a magnetic field. However, in this case, modifications of the antidot geometry are expected and their effects on the functionality of the presented device have to be carefully taken into account.
The present study could prove quite relevant for the implementation of both theoretical and experimental investigations of interferometric phenomena in HBT\cite{HBT} and HOM\cite{HOM} setups in the context of electron quantum optics. Indeed, the present SQS or SES source could be embedded in a quantum Hall bar which is divided in two by a central quantum point contact, where either the HBT partitioning of the source could be analyzed in order to quantify the production of spurious excitations, or alternatively with two antidot sources where two QP collisions could be achieved in order to probe the overlap of QP wave packets. In the near future, we are determined to model such experiments assuming that either ideal QP wavepackets or ideal electron wavepackets have been deposited on the fractional Hall edge,\cite{rech_fqhe} but a proper description of the source, such as presented in the context of this work, will ultimately be necessary to bring the description sufficiently close to experimental reality.
Furthermore, the present study was achieved assuming the weak coupling hypothesis where tunneling rates are smaller than the electronic temperature. The numbers we provided in Sec. \ref{scales} seem to point out that this hypothesis is justified in the context of present experimental working conditions ($k_B\Theta\simeq 50mK$). However, upon either increasing the tunneling rates (by bringing the antidot closer to the edge, which would increase the QP tunneling amplitude) or by working at much lower temperatures, a coherent description of tunneling will be eventually required. This constitutes a truly challenging task, which may have to rely on Keldysh non equilibrium Green's function formalism in order to describe the time dependence of the current, while taking into account the (finite) QP occupation of the antidot in its evolution.
\acknowledgements
Early discussions with D. C. Glattli are gratefully acknowledged.
The authors also acknowledge the support of Grants No. ANR-2010-BLANC-0412 (``1 shot'') and No. ANR-2014-BLANC (``one shot reloaded'').
This work was granted access to the HPC resources of Aix-Marseille Universit\'e financed by the project Equip@Meso (Grant No. ANR-10-EQPX-29-01) and has been carried out in the framework of the Labex ARCHIMEDE (Grant No. ANR-11-LABX-0033) and of the AMIDEX project (Grant No. ANR-11-IDEX-0001-02), all funded by the ``investissements d'avenir'' French Government program managed by the French National Research Agency (ANR).
|
\section{Introduction}
The light curves of Type Ia supernovae (SNe~Ia) are of general interest, in
particular for applications in cosmology. The optical
light curves of SNe~Ia admit a lightcurve shape relation
\citep{philm15,rpk96,philetal99,goldhetal01} which allows them to be
used as ``correctable candles'' and hence, their distances can be
determined from the observed photometry.
The underlying physical explanation for the light curve shape
relation is that the total \ensuremath{\nuc{56}{Ni}}\xspace mass produced in the explosion
determines to first
order the temperature
and thus, more \ensuremath{\nuc{56}{Ni}}\xspace leads to higher temperatures, which leads to
higher opacities, increasing the diffusion time
\citep{nugseq95,kmh93,hkw95,PE00a,PE00b,PE01,mazzetal01}. It has been
questioned whether the primary effect of the \ensuremath{\nuc{56}{Ni}}\xspace mass is on the diffusion
time or rather on the color evolution of the spectra \citep{KW07}, but
the ultimate source for the observed luminosity of a SN~Ia is the
decay of \ensuremath{\nuc{56}{Ni}}\xspace, and thus to zero order the luminosity should be
correlated with the mass of \ensuremath{\nuc{56}{Ni}}\xspace.
Using the light curves of
SNe~Ia for distance measurements, led to the discovery of the phenomenon
that the Universe's expansion is accelerated, i.e, the discovery of
the dark energy
\citep{riess_scoop98,perletal99}. The nature of the dark energy is of
fundamental importance for cosmology.
Many physical details of Type Ia supernovae events are still not fully
understood \citep[see][for reviews]{hnar00,parrent14}. One of the most
important open questions is
the nature of the progenitor(s) of SN Ia explosions.
The progenitor is believed to be a CO white dwarf in a binary system. Its companion
star could be a main sequence star, an evolved star, or another white
dwarf. The primary white dwarf grows in mass via accretion, which can
be slow in the Chandrasekhar mass scenario, or rapid in scenarios that
involve the merging of two white dwarfs.
\citep[see][for reviews of the suggested possibilities]{levanon15,maoz14}.
In fact, SNe Ia may come from a variety of progenitor systems, and
detailed observations of nearby SNe Ia may lead to the understanding
of the nature of these progenitor systems \citep{nug_11fe_11,dilday11kx12}.
There is further
ongoing discussion
about the explosion mechanism of Type Ia supernovae.
An instantaneous detonation in Chandrasekhar-mass WDs has
been ruled out, because it fails to produce intermediate mass elements (IME),
which are observed in the
spectra of SNe Ia \citep{arnett69}. In sub-Chandrasekhar-mass WDs
an instantaneous detonation can produce IME due to the lower
central densities
\citep{shigeyama92,sim10}. However, the mass of the WD is rather
tightly constrained to produce both iron group elements and IME and the low densities
will not produce observed stable iron group elements.
Several different explosion models such as a pure deflagration,
delayed detonations,
gravitationally confined detonations, violent mergers, and collisions have been suggested
\citep{nomoto84,khok91a,PCL04,jordangcd08,jordan12,pakmor12,pakmor13,rosswog09,kushnir13,long14},
which could
explain the appearance of the intermediate mass elements observed in
SNe Ia spectra.
To understand the physical processes in a Type Ia
supernova further, different numerical codes have been developed and
used to model the spectra and light curves of SNe Ia
\citep{branch85,mazzali93,maeda14,wollaeger14,h03a,kasen06a,sim13,dessart10,jack11}.
We focus here on the light curves in the near-infrared wavelength
region. The light curves of the near-infrared bands $I$, $J$, $H$, and $K$
very often show a secondary maximum \citep{elias81,HGFS99by02}. In his
detailed study of near-infrared light curves of SNe Ia,
\citet{kasen06b} showed that the secondary maxima are caused by the
recombination of iron peak elements, which comes from a drastic
opacity change between ionization stages III to II of the iron
elements. This observation has been confirmed with other theoretical
radiative transfer calculations \citep{jack12,dessart14b,gall12}.
\citet{jack12} were able to assign specific features of Fe II and Co
II that cause the individual secondary maxima in the $I$, $J$, $H$ and $K$
bands. However, their investigation was based only on theoretical light
curve calculations. \citet{dessart14b} found that [Co~II] lines play a
strong role, but this result has not been confirmed in other
calculations and may be due to the way that levels and superlevels are
coupled in their approximations.
In this work, we will present observational evidence that an Fe II
feature, produced by a blend of Fe~II lines,
causes the secondary maximum in the $I$-band and additionally, we
present an investigation of
the plateau phase of the $R$-band light curve.
In \S~\ref{sec:methods}, we present the details of the observations
of SN~2014J and describe
the methods used to calculate the theoretical light curves with the \texttt{PHOENIX}\xspace\ code. In \S~\ref{sec:secmax}, we present the results
obtained for the secondary maximum in the $I$-band and the plateau phase
of the $R$-band.
\section[]{Methods}\label{sec:methods}
We obtained a time series of observations of high resolution spectra
of the bright SN~Ia, 2014J in M82 with the TIGRE
telescope. We fitted the model light curves calculated with the \texttt{PHOENIX}\xspace\
code to the observed light curves in the near-infrared $I$ and $R$-bands.
The observed and theoretical spectra are then
compared to determine the features that cause the secondary maximum in
the $I$-band and the plateau phase in the $R$-band.
SN~2014J was discovered by \citet{fossey14} in the nearby galaxy M82
approximately 1~week after explosion. We use for the explosion time
January 14.75 UT as found by \citet{zheng14}.
\subsection{Observations}
We observed SN 2014J in M82 with the TIGRE telescope situated in
Central Mexico close to the city of Guanajuato \citep{schmitt14}. The
telescope TIGRE is a 1.2 m mirror telescope that operates completely
robotically. Its original design was to monitor the stellar activity
of solar-like stars, but other observational campaigns are also
feasible. The TIGRE telescope is equipped with the HEROS spectrograph,
which has a high resolution $R \sim 20,000$ and covers the
wavelength range from about 3800 \AA\ to about 8800 \AA. The
spectrum is divided into two channels (red and blue). The data
reduction pipeline is also fully automatic.
Beginning with the first observation during the night of 23rd of January 2014,
we were able to observe a spectrum of SN 2014J in M82 almost every
night following the discovery. See \citet{jack15} for a
complete presentation of the time series of SN~2014J spectra
observed with the TIGRE telescope.
We used an exposure time of 3 hours for every observed spectrum.
The spectra were obtained with the full spectral resolution $R \sim
20,000$. Since a supernova spectrum
shows only very broad features and such a high resolution is not
necessarily required, we binned the spectrum down to a resolution of
10 \AA.
The binning significantly improves the signal to noise ratio.
For this work, we are interested in the near-infrared wavelength range,
so that we will only use the spectra of the red channel (5800 \AA\ to
8800 \AA). SN 2014J showed
significant reddening so that the obtained spectra in the blue channel
(3800 \AA\ to 5800 \AA) observed at such high resolution have a much
lower signal to noise.
\subsection{Theoretical light curves}
In order to model the theoretical light curves in different bands of
SN 2014J, we used the time dependent extension of the \texttt{PHOENIX}\xspace\ code as
described in detail in \citet{jack09,jack11}. The code follows the
evolution of the supernovae envelope after the explosion enforcing energy
conservation. This includes energy deposition due to
$\gamma$-rays from the radioactive decay of $^{56}$Ni and $^{56}$Co.
Cooling due to the adiabatic expansion of the envelope, which is
assumed to be homologous, is also included. For the transport of
energy by radiation throughout the envelope, we solve the spherically
symmetric special relativistic radiative transfer equation
\citep{hauschildt92,hbjcam99}.
The envelope was divided into 128 layers. About 1,000 time
steps are performed for each point in the light curve.
We usually calculate a point in the light curve every 1 or 2 days.
All light curves have been calculated
assuming local thermodynamic equilibrium (LTE) and using about 2000
wavelength points.
The final spectra presented in this work have been calculated
using a higher resolution of about 20,000 wavelength points.
\texttt{PHOENIX}\xspace uses the update Kurucz database \footnote{http://kurucz.harvard.edu/linelists/gfall/}
which contains about 80 million
atomic line transition.
For the initial input structure, we used the
results of the W7 deflagration model \citep{nomoto84}, which also
includes the specific non-homogeneous abundances.
In this work, we focus on the secondary maxima in the
near-infrared wavelength region. We seek to obtain the best fit
to the individual observed light curves of SN 2014J in the respective
bands. As described in \citet{jack12}, with the above assumptions, it
is necessary to
vary the equivalent two-level atom thermalization parameter when solving the
radiative transfer problem. The source function of the radiative
transfer equation including scattering for an equivalent two level
atom can be written as
\begin{equation}
S_{\lambda}=(1-\epsilon_{\lambda})J_{\lambda}+\epsilon_{\lambda}B_{\lambda}.
\end{equation}
$S_{\lambda}$ is the source function, $B_{\lambda}$ is the Planck
function, and $J_{\lambda}$ is the mean intensity. All these
quantities are wavelength dependent. For the LTE \texttt{PHOENIX}\xspace\ calculations,
it is possible to set a wavelength independent factor
$\epsilon=\epsilon_{\lambda}=$ constant to approximate LTE line
scattering over the whole wavelength range. We use this method here
for computational expediency. Since the SN Ia envelope becomes
thinner during its free expansion phase, scattering becomes more and
more important, and the thermalization parameter $\epsilon$
decreases. Applying this method of a decreasing thermalization
parameter, we obtained fits to the observed light curves of SN 2014J.
\section{Secondary Maximum}\label{sec:secmax}
We compare the spectra of SN~2014J observed with the TIGRE
telescope with the theoretical spectra calculated with \texttt{PHOENIX}\xspace\ focusing on
the secondary maximum of the $I$-band light curve and the plateau phase
observed in the $R$-band.
All observed spectra and light curves have been dereddened with the values
$E(B-V)=1.33$ and $R_{\rm V}=1.3$ found by
\citet{amanullah14} in their study of the extinction law of SN 2014J.
\subsection{$I$-band}
We obtained a time series of spectra of SN 2014J during the
rise of the light curve towards the secondary maximum in the $I$-band.
Unfortunately, the HEROS spectrograph does not cover the whole range of
the $I$-band filter, meaning that we could not reconstruct a light curve
from our observed spectra. Therefore, we used the observed light curve
of the AAVSO database\footnote{www.aavso.org} to find the best fit of the model
light curve calculated with \texttt{PHOENIX}\xspace. As described above, we varied the
thermalization parameter $\epsilon$ to obtain the best fit to the
observed light curve in the $I$-band. Although this is not a physically
correct method to calculate a light curve, the overall trend of a
decreasing $\epsilon$ is physically motivated. However, the goal here
is to use this
fit to study the spectral evolution of a SN Ia towards the secondary
maximum.
\begin{figure}
\begin{center}
\resizebox{\hsize}{!}{\includegraphics{I_band.eps}}
\caption{The circles represent the dereddened AAVSO observed light curve in absolute magnitude of SN 2014J in the $I$-band.
The solid line shows the best fit obtained from \texttt{PHOENIX}\xspace\ models.
The
values for $\epsilon$ are indicated along the
bottom of the plot and the vertical dotted lines demarcate the range.}
\label{fig:lc_I}
\end{center}
\end{figure}
In Figure \ref{fig:lc_I}, the filled circles represent the light curve
of SN 2014J in the $I$-band obtained from AAVSO observation data.
The observed light curve has been dereddened.
We also plotted the AAVSO light curve in absolute magnitude using the distance
of 3.5 Mpc to M82, which results in a distance modulus of 27.7~mag,
which is comparable with the value of 27.6~mag found by \citet{foley14}.
The
solid line represents the best fit of the theoretical light curve
calculated with \texttt{PHOENIX}\xspace\ models.
As stated above, the models were calculated with different values
of $\epsilon$ for different epochs. This causes the kinks
in the model light curve.
The rise towards the secondary maximum
is clearly visible and starts around 32 days after the explosion. The
secondary maximum is reached at around 45 days. These are the epochs
that are interesting for studying the spectral evolution and
identifying the feature that causes this secondary maximum.
\begin{figure}
\begin{center}
\resizebox{\hsize}{!}{\includegraphics{phx_spec_I.eps}}
\caption{Theoretical spectra at different epochs after the explosion calculated with \texttt{PHOENIX}\xspace.
At $\sim 7500$\;\AA, a new feature
of blended Fe II lines appears and causes the secondary maximum.
The dotted line indicates the $I$-band filter response.}
\label{fig:phx_spec_I}
\end{center}
\end{figure}
In Figure~\ref{fig:phx_spec_I}, we show a plot of the theoretical
spectra calculated with \texttt{PHOENIX}\xspace\ at different epochs obtained from the best
fit to the observed light curve.
The spectra are shown in flux received at Earth if the models would be at
the distance of SN 2011fe (6.6~Mpc) to be able to compare the fluxes directly.
The dotted line shows the $I$-band
filter response.
At that epoch the spectrum does not show many clear
features. The brightness decreases after the $I$ maximum. The
Ca II IR triplet feature at $8500 - 9000$~\AA\ becomes
prominent. However, this feature,
does not show a change in brightness during the rise to the
secondary maximum. From the beginning of the rise (32 days) up to the
secondary maximum (45 days) there appears a new feature of Fe II at a
wavelength of $\sim 7500$ \AA.
In Fe II there are numerous transitions from relatively high excited
states in this wavelength region, for example transitions from
$3d^6(^3D)4s-3d^6(^5D)4p$, $3d^5(^4P)4s4p(^3P^\circ)-3d^6(^5D)6s$, and
similar transitions, which have Einstein coefficient $A_{21}$ values $> 1\times 10^7$~s$^{-1}$.
Thus, this feature causes the observed
secondary maximum in the $I$-band of SNe Ia.
The observed spectra also clearly show the appearance of the Fe II
feature mentioned above.
\begin{figure}
\begin{center}
\resizebox{\hsize}{!}{\includegraphics{tigre_spec_I.eps}}
\caption{Spectra of SN 2014J observed on two different days before and during the secondary maximum. The appearance of the feature
of blended Fe~II lines can be seen in the observations at $\sim 7500$\;\AA\ as well.
The dotted line represents the $I$-band filter function.}
\label{fig:tigre_spec_I}
\end{center}
\end{figure}
Figure \ref{fig:tigre_spec_I} shows two observed spectra of SN 2014J
in the same wavelength range as was used for the theoretical spectra in Figure~\ref{fig:phx_spec_I}.
The dotted line represents the $I$-band filter response.
The spectrum from the 16th of February was taken directly before the
beginning of the rise towards the secondary maximum.
The spectrum from the 1st of March was observed during the secondary maximum. These observed spectra correspond to the theoretical
spectra of day 32 and day 45 in Fig. 2.
The feature of Fe II at around 7500 \AA\ also appears in the observed spectra. Note that there is a strong telluric absorption feature
around 7600~\AA, which contaminates the spectrum in this region.
However, we can clearly confirm that this Fe II feature actually
causes the secondary maximum in the $I$-band of SN~2014J, and thus, it
is likely the cause in most SNe~Ia.
Since our observed spectra of SN~2014J do not cover the whole wavelength
range of the $I$-band, we used archival data to obtain additional evidence
that the Fe~II feature causes the secondary maximum.
\begin{figure}
\begin{center}
\resizebox{\hsize}{!}{\includegraphics{sn2011fe_I.eps}}
\caption{Spectra of SN 2011fe observed. The appearance of the feature
of blended Fe~II lines can be seen in the observations at $\sim 7500$\;\AA\ as well.
The dotted line represents the $I$-band filter function.}
\label{fig:sn2011fe_I}
\end{center}
\end{figure}
In Figure \ref{fig:sn2011fe_I} we show two spectra of SN 2011fe. The spectra were
observed by the Nearby Supernova Factory \citep{pereira13} in a time
series of spectra. The spectrum
from the 24th of September 2011 coincides
with the minimum before the rise towards the secondary maximum in the $I$-band.
The secondary maximum was observed around the 8th of October
2011. Unfortunately,
there does not exist a spectrum of that day. Therefore, we used the spectrum
of the 4th of October. In the plot one can clearly see that a feature
around 7500 \AA\ arises. This gives more observational
evidence that this Fe~II feature causes the secondary maximum
in the $I$-band of SNe~Ia.
\subsection{$R$-band}
The light curve of a typical type Ia supernovae in the $R$-band does not
show a secondary maximum, but usually a short plateau phase. Since all
other light curves in the optical wavelength range show a steady
decline after maximum light, there should also be a feature that
causes the effective $R$-band photosphere to ``hang up''.
We used the same
method to identify this feature as in the study of the $I$-band
secondary maximum. Using a varying line thermalization parameter
$\epsilon$, we obtained a best fit to the observed (dereddened)
$R$-band light curve of SN 2014J.
\begin{figure}
\centering
\resizebox{\hsize}{!}{\includegraphics{R_band.eps}}
\caption{
The circles represent the dereddened AAVSO observed light curve in absolute magnitude of SN 2014J in the $R$-band.
The solid line shows the best fit obtained from \texttt{PHOENIX}\xspace\ models. The
values for $\epsilon$ are indicated along the
bottom of the plot and the vertical dotted lines demarcate the range.}
\label{fig:lc_R}
\end{figure}
\begin{figure}
\begin{center}
\resizebox{\hsize}{!}{\includegraphics{lc_compare_Rband.eps}}
\caption{Here we show the \texttt{PHOENIX}\xspace\ $R$-band light curve calculated with
the $\epsilon$ values for the $I$-band and $R$-band.}
\label{fig:compare}
\end{center}
\end{figure}
In Figure \ref{fig:lc_R}, the filled circles show the observed AAVSO
light curve in the $R$-band.
Between about 30 and 40 days after explosion the $R$-band
light curve of SN 2014J shows a short plateau phase, which we
also modeled with the theoretical light curve of \texttt{PHOENIX}\xspace.
The solid line represents the best fit
obtained from \texttt{PHOENIX}\xspace\ models by varying the thermalization
parameter. Note that the variations of $\epsilon$ are
slightly different from those used to compute the $I$-band light
curve.
The differences, while small, represent a limitation in our
modeling approach.
However, the overall trend of
a decreasing $\epsilon$ is the same for all near-infrared bands.
Scattering becomes more and more important as the SN Ia envelope
expands.
In Figure \ref{fig:compare} we compare the $R$-band
light curves that
have been calculated using the values for $\epsilon$
used for the $R$-band and for the $I$-band.
One can see some clear differences during the phase of the minumum before
the plateau phase.
However, since $\epsilon$ is wavelength dependent it is not possible
to distinguish between a possible inaccuracy of the models or
a real difference in the value of $\epsilon$ for the different
wavelength regions.
\begin{figure}
\begin{center}
\resizebox{\hsize}{!}{\includegraphics{phx_spec_R.eps}}
\caption{Spectra of the best fit to the $R$-band light curve.
Two features of blended Co~II lines and blended Fe~II lines contribute
to the flux that causes the $R$-band plateau phase.
The dotted line represents the $R$-band filter function.
The model luminosity has been scaled by the distance to SN 2011fe.}
\label{fig:phx_spec_R}
\end{center}
\end{figure}
We then studied the spectral evolution during this observed plateau
phase. In Figure \ref{fig:phx_spec_R}, we show the theoretical
spectra calculated with \texttt{PHOENIX}\xspace\ at the beginning of the plateau phase
(30 days) and at the end of the $R$-band plateau (40 days).
The dotted line represents the $R$-band filter function. After the $R$ maximum
(16 days), the brightness decreases in the whole wavelength range. At
30 days, which marks the beginning of the plateau phase, the
characteristic P-Cygni Si~II feature around 6000 \AA\ to 6500 \AA\
strengthens.
This feature changes between 30 and 40 days during the
plateau phase. However, the contribution of this feature to the brightness
in the $R$-band decreases and can, therefore, not cause the plateau phase.
In the models, blends of high excitation Co~II
lines appear in the wavelength region $6500-7000$~\AA, however the
total flux in the models is not large enough to produce the observed plateau.
The feature of Fe II at 7500~\AA\ that
causes the secondary maximum in the $I$-band is also present in the
$R$-band filter, but again the filter sensitivity is falling off here so
the total flux is not large enough to produce the observed plateau.
While nickel holes are seen in SNe~Ia \citep{fesen07}, the nickel hole
in W7 is too large to be compatible with galactic nucleosynthesis and
thus, the center of W7 has too much stable Ni and Fe, which will
reduce the total amount of Co. However, the Co~II lines seen in the
models are not easily identified in the observations (see below).
\begin{figure}
\begin{center}
\resizebox{\hsize}{!}{\includegraphics{tigre_spec_R.eps}}
\caption{Spectra of SN 2014J observed on two different days at the beginning
and the end of the $R$-band plateau phase. The Fe~II features is clearly visible
while the Co~II feature of the models has not been identified.
The dotted line represents the $R$-band filter function.}
\label{fig:tigre_spec_R}
\end{center}
\end{figure}
Figure~\ref{fig:tigre_spec_R} shows the observed spectra of SN~2014J
in the $R$-band using the same wavelength range as for the theoretical
spectra of Fig.~\ref{fig:phx_spec_R}. The dotted line once again
represents the $R$-band
filter function. We plotted the spectrum from 14th of February, which
corresponds to the beginning of the plateau phase (30 days). The end
of the plateau phase at 40 days is represented by the observed spectrum
from the 24th of February. One can clearly see the change of the expansion
velocity of the characteristic Si~II feature as the photosphere moves
inwards. There is no clear feature
arising in the $R$-band that obviously causes the short plateau phase.
Of course, the Fe~II feature at around 7500
\AA\ also contributes somewhat to the light curve in the $R$-band. As
discussed above Co~II lines in the $6500-7000$~\AA\ range are seen in
the models, but their signatures are not strong in the observed spectra.
\section{Conclusions}
We observed a time-series of spectra of the nearby Type Ia supernova
2014J in M82 with the TIGRE telescope. This time-series also
covered the secondary maximum in the $I$-band and the plateau phase in
the $R$-band. We calculated the respective theoretical light curves
with the time dependent extension of the \texttt{PHOENIX}\xspace\ code by varying the
line thermalization parameter to obtain a best fit to the observed light
curves. For the $I$-band light curve of SN~2014J, we find that the
secondary maximum is caused by a blend of Fe~II high excitation lines because of
the recombination of iron peak elements from ionization stage III to
stage II. This confirms previous work studying near-infrared
light curves of
SNe Ia \citep{kasen06b,jack12,
dessart14b}. Studying the spectral evolution of SN~2014J, we find
observational evidence that the Fe~II feature at around
7,500 \AA\ causes the secondary maximum in the $I$-band. While W7 is
only a parameterized model of a real SN~Ia, the interior composition
is fairly generic due to the nuclear physics and thus, our conclusion
should be fairly general. However, the details are a bit more complicated,
and recent explosion calculations show
quite some diversity in the ejected material \citep{fink14,woosley11,moll14}.
As far as we know, there exists no detailed study of the $R$-band light
curves of Type Ia supernovae. We observed SN~2014J, and it ---
like most other SNe Ia --- shows a short plateau phase in the $R$-band between
30 and 40 days after the explosion. We reproduced the observed
light curve with our \texttt{PHOENIX}\xspace\ models to be able to identify the feature
that causes this short plateau phase.
Co~II lines in the range $6500-7000$~\AA\ do increase in flux in the
models, and the Fe~II lines to the red may also contribute somewhat,
but the total flux from these features does not appear to be large
enough in the models and the Co~II signatures are not clearly seen in
the observations.
For future work, it will be very interesting to study the spectral
evolution of a Type Ia supernova in the other near-infrared bands $J$, $H$,
$Y$, and $K$. With this one might be able to identify the features which
cause the secondary maxima in these bands as proposed
by \citet{jack12}. A further, more detailed study of the $R$-band is
important, since understanding the detailed composition and ionization
structure of the ejecta with time will help to constrain explosion models.
\section*{Acknowledgments}
We would like to thank the TIGRE team for having made possible the
observations of SN 2014J on short notice. We acknowledge with thanks
the variable star observations from the AAVSO International Database
contributed by observers worldwide and used in this research. We thank
Kevin Krisciunas for helpful discussion on the $R$-band light curve.
The work has been supported in part by support for
programs HST-GO-122948.04-A provided by NASA
through a grant from the Space Telescope Science Institute, which is
operated by the Association of Universities for Research in Astronomy,
Incorporated, under NASA contract NAS5-26555. This work was also
supported in part by the NSF grant AST-0707704. This research
used resources of the National Energy Research Scientific Computing
Center (NERSC), which is supported by the Office of Science of the
U.S. Department of Energy under Contract No. DE-AC02-05CH11231; and
the H\"ochstleistungs Rechenzentrum Nord (HLRN). We thank both these
institutions for a generous allocation of computer time.
\bibliographystyle{mn2e}
|
\section{Introduction}
\label{introduction}
The Generalized Gram-Charlier (GGC) series expands an unknown \textit{pdf}
as a linear combination of the increasing order differentiations of a reference \textit{pdf},
where the coefficients of expansion involve cumulant differences between those of an unknown \textit{pdf} and a reference \textit{pdf}.
The GGC expansions are used to approximate \textit{pdf} and functions of \textit{pdf} in Statistics, Machine Learning,
Economics, Chemistry, Astronomy and other application areas.
There have been used Poisson's distribution \citep{aroian1937}, log-normal distribution \citep{GGC2pdf47}, binomial distribution \citep{bookReitz27}, gamma distribution \citep{bowers1966expansion} and others as reference \textit{pdf}s.
But, the Gram-Charlier (GC) expansion with Gaussian density as a reference \textit{pdf}
is the most popular and identified as the Gram-Charlier A (GCA) series.
Specifically, the GCA series is used for test of Gaussianity or near Gaussian \textit{pdf} approximations \citep{GGC2pdf47,GGCchem94,girolaminegH96,GGCpdf01,GCDnd09}, for entropy measure and independence measure approximations \citep{GGCHAmari96,ICAbook01}, for optima analysis through derivatives of \textit{pdf} \citep{ICA03Boscolo}, for time-frequency analysis \citep{genGramC98} and others. The rearrangement of the terms in GCA series results into the Edgeworth series with better convergence property.
\hspace{0.2 in} There exists multiple ways to derive univariate GCA series, as reported by \citet{HistoryGGS00,hald2002history}. It's generalization to univariate GGC series is derived by \citet{Schleher77,genGramC98,genGramC07,genGramC11}.
\hspace{0.2 in} The multivariate GGC or GCA series derivation requires multivariate representations of the Taylor series, the increasing order differentiations of a reference \textit{pdf} and the cumulants.
To signify the representation issue in required multivariate extensions; it is worth quoting
\citet{ndHermite02} that says, ` $\ldots$ though the generalizations to higher dimensions may be considered straightforward, the methodology and the mathematical notations get quite cumbersome when dealing with the higher order derivatives of the characteristic function or other functions of a random vector $\ldots$ '.
\hspace{0.2 in} Conventionally, the higher order differentiations of a multivariate \textit{pdf}; and therefore, the multivariate cumulants and multivariate Hermite polynomials; require multilinear representations.
It is known and acknowledged historically that going from matrix like notations to tensor notations for multivariate cumulants and multivariate Hermite polynomials have made the representation more transparent and the calculations more simpler \citep{Tensorbook87}.
Though the tensor notations have advantages over the matrix notations, they require componentwise separate representations and are more tedious compare to the vector notations.
As a unified and comprehensive solution to this, there has been used an approach based only on elementary calculus of several variables by \citet{ndHermite02}. The approach uses a specific Kronecker product based differential operator, identified as the `K-derivative operator', to achieve vectorization of the Jacobian matrix of a multivariate vector function. The successive applications of this K-derivative operator achieves vectorization of the higher order derivatives also. Using this approach,
there have been derived the multivariate Taylor series and the higher order cumulants \citep{ndHermite02,SankhyaCumVec06}; as well, the multivariate Hermite polynomials \citep{ndHermite02} in vector notations resulting into more transparent representations. In fact, it could be noticed that the same approach has been first used to derive vector Hermite polynomials by \citet{dnHermite96}; and later\footnote{Somehow, the citation for \citep{dnHermite96} is not found in article \citep{ndHermite02}.} it has been formalized and generalized by \citet{ndHermite02}.
\hspace{0.2 in} There exists various approaches deriving multivariate GCA series with various representations.
They include GCA series representation using multi-element matrix notations for moments and cumulants by \citet{sauer1979convenient}; using
recursive formula for Hermite polynomials by \citet{ndHPGCberkowitz70}; using tensor representation for cumulants and Hermite polynomials by \citet[Chapter 5]{Tensorbook87}; using tensor representation for cumulants and involving multivariate Bell polynomials by \citet{withers2014dual}; using vector moments and vector Hermite polynomials by \citet{dnHermite96} and others.
There exists various
derivations for multivariate Edgeworth series
\citep{ndEdge76,ndEdgeworth86,amari1983differential,Tensorbook87,unifieddensities98,withers2014dual}.
There also exists multivariate GGC series, derived by \citet[Chapter 5]{Tensorbook87}, in tensor notations.
But, as per the author's knowledge, there exists neither the multivariate GGC series nor the multivariate Edgeworth series in vector notations. For the ease of the readers in following and comparing the various representations;
the existing representations of multivariate GGC series and multivariate GCA series are shown and compared in \ref{historyGGCreps}.
\hspace{0.2 in} Overall, to take advantages due to the recent advancement in representation, this article extends a specific derivation for univariate GGC series by \citet{genGramC07} to multivariate;
using only elementary calculus of several variables instead of Tensor calculus.
As a by product, it also derives mutual relations between vector cumulants and vector moments of a random vector; integral form of the multivariate \textit{pdf}; integral form of the multivariate vector Hermite polynomials
and the multivariate GCA series. All the derived multivariate expressions are more elementary as using vector notations and more comprehensive as apparently more nearer to their counterparts for univariate; compare to their coordinatewise tensor notations. The intermediate theoretical results, in the article, are verified using suitable known examples.
\hspace{0.2 in}
Towards the aim of the article, the next Section \ref{Kronprod} briefs some necessary background on the Kronecker product and a way to obtain vectorization of the higher order differentiations of a multivariate \textit{pdf}. It also obtains the required multivariate Taylor series expansion using the derived notations.
After the preliminary background,
this article follows almost the same sequence for multivariate as that in \citep{genGramC07} for univariate.
The Section \ref{chfn} uses the characteristic function and the generating functions to derive
cumulants and moments of a random vector in vector notations with their mutual relationships.
The Section \ref{pdfcum} obtains multivariate \textit{pdf} in terms of its vector cumulants. The expressions for derivatives of multivariate Gaussian density and vector Hermite polynomials are derived in Section \ref{DerGauss}. The Section \ref{ndGCAseries} derives multivariate GCA series by representing an unknown \textit{pdf} in terms of the Gaussian \textit{pdf} as a reference. The Section \ref{ndGGC} derives GGC Expansion, representing an unknown \textit{pdf} in terms of the
a known reference \textit{pdf}. The Section \ref{chGGC} derives an unknown characteristic function of a random vector in terms of a reference characteristic function. The Section \ref{compactGGC} derives the same GGC expansion in a more compact way that summarizes the approach of the whole derivation. Finally, Section \ref{conclusion} concludes the article. For the sake of clarity; the calculation details, the proofs and the expressions for existing multivariate expansions are kept in appendix at the end of the article.
\section{Vectorization of the higher order differentiations}
\label{Kronprod}
The section briefs the Kronecker product and the way it can be applied to achieve vectorization of the higher order differentiations of a multivariate \textit{pdf}. Based on it, the multivariate Taylor series is obtained in vector notations.
More details can be found
on the Kronecker Product in \citep[Chapter 2]{MagnusBook99},
on achieving vectorization of the higher order differentiations in \citep{ndHermite02,SankhyaCumVec06} and on the commutation matrices in \citep[Chapter 3, Section 7]{MagnusBook99}.
\begin{definition}[Kronecker Product Operator $(\otimes)$]
The Kronecker Product Operator $(\otimes)$ between matrices $\mathbf{A}$ with size $p \times q$ and $\mathbf{B}$ with size $m \times n$ is defined as:\\
\begin{align}
\mathbf{A}\otimes\mathbf{B} = \left[
\begin{array}[]{c c c c}
a_{11}\mathbf{B} & a_{12}\mathbf{B} & \cdots & a_{1q}\mathbf{B} \\
a_{21}\mathbf{B} &a_{22}\mathbf{B} &\cdots &a_{2q}\mathbf{B} \\
\vdots & \vdots & \ddots&\vdots \\
a_{p1}\mathbf{B} &a_{p2}\mathbf{B} &\cdots &a_{pq}\mathbf{B}
\end{array}
\right]
\end{align}
\end{definition}
The resultant matrix is of dimension $pm \times qn$.
As a further example; let $\mathbf{A}$ is with size $p \times 1$ and $\mathbf{B}$ is with size $m \times 1$, then $\mathbf{A} \otimes \mathbf{B}'$ is\footnote{The symbol ' stands for Transpose of a matrix} a matrix with size $p \times m $. $\mathbf{A} \otimes \mathbf{A}$ is symbolically represented as $\mathbf{A}^{\otimes 2}$ and has size $p^2 \times 1$. In general, $\mathbf{A} \otimes \mathbf{A} \otimes \ldots \otimes \mathbf{A} \mbox{ (n times)}$ is symbolically represented as $\mathbf{A}^{\otimes n}$ and has size $p^n \times 1$.
\begin{definition}[Jacobian Matrix]
Let $\bs{\lambda} = (\lambda_1, \lambda_2, \ldots, \lambda_d)'$, $\bs{\lambda} \in \mathbb{R}^d$ and $\mathbf{f}(\bs{\lambda}) = $ \\
$(f_1(\lambda), f_2(\lambda), \ldots, f_m(\lambda))' \in \mathbb{R}^{m}$ be a differentiable m-component vector function. Then, Jacobian matrix of $\mathbf{f}(\bs{\lambda})$ $(\mathbf{J}(\mathbf{f}))$ is an $m \times d$ matrix defined as under:
\begin{align}
\mathbf{J}(\mathbf{f}(\bs{\lambda})) = \frac{d \mathbf{f}}{d {\bs{\lambda}}} = \left[ \frac{\partial \mathbf{f}}{\partial \lambda_1}, \frac{\partial \mathbf{f}}{\partial \lambda_2},\ldots, \frac{\partial \mathbf{f}}{\partial \lambda_d} \right] = \left[
\begin{array}[]{c c c c}
\frac{\partial f_1}{\partial \lambda_1} & \frac{\partial f_1}{\partial \lambda_2} & \cdots & \frac{\partial f_1}{\partial \lambda_d} \\
\frac{\partial f_2}{\partial \lambda_1} & \ddots & &\vdots \\
\vdots & & \ddots &\vdots \\
\frac{\partial f_m}{\partial \lambda_1} & \frac{\partial f_m}{\partial \lambda_2} & \cdots & \frac{\partial f_m}{\partial \lambda_d}
\end{array}
\right]
\end{align}
\end{definition}
Let the vector differential operator be defined as a column vector
$\mathbf{D}_{\bs{\lambda}} = \left( \frac{\partial}{\partial \lambda_1}, \frac{\partial}{\partial \lambda_2},\ldots, \frac{\partial}{\partial \lambda_d} \right)'$, then
the Jacobian matrix, in terms of the $\mathbf{D}_{\bs{\lambda}}$, can be re-written as:
\begin{align}
\mathbf{J}(\mathbf{f}(\bs{\lambda})) = \mathbf{D}_{\bs{\lambda}}(\mathbf{f}) = \mathbf{f}(\bs{\lambda})\mathbf{D}'_{\bs{\lambda}} = \left( f_1(\bs{\lambda}), f_2(\bs{\lambda}), \ldots, f_m(\bs{\lambda}) \right)' \left( \frac{\partial}{\partial \lambda_1}, \frac{\partial}{\partial \lambda_2},\ldots, \frac{\partial}{\partial \lambda_d} \right)
\end{align}
This implies that to match the definition of differentiation from matrix calculus, the vector differential operator should be applied from the right to the left. This is same as the requirement to be satisfied on generalization of vector derivative to matrix derivative as discussed by \citet{Magnus10}.
So, applying vector derivative operator from right to the left, has been kept as a rule throughout the article.
\begin{definition}[The K-derivative Operator]
Let $\bs{\lambda} = (\lambda_1, \lambda_2, \ldots, \lambda_d)'$, $\bs{\lambda} \in \mathbb{R}^d$; the vector differential operator $\mathbf{D}_{\bs{\lambda}} = \left( \frac{\partial}{\partial \lambda_1}, \frac{\partial}{\partial \lambda_2},\ldots, \frac{\partial}{\partial \lambda_d} \right)'$ and a differentiable m-component vector function
$\mathbf{f}(\bs{\lambda}) = (f_1(\lambda), f_2(\lambda), \ldots, f_m(\lambda))' \in \mathbb{R}^{m}$. Then, the Kronecker product between $\mathbf{D}_{\bs{\lambda}}$ and $\mathbf{f}(\bs{\lambda})$ is given as under:
\begin{align}
\mathbf{D}_{\bs{\lambda}}^\otimes \mathbf{f}(\bs{\lambda}) &=
\left[
\begin{array}[]{c }
f_1(\bs{\lambda}) \\ f_2(\bs{\lambda}) \\ \vdots \\ f_m(\bs{\lambda})
\end{array}
\right]
\otimes
\left[
\begin{array}[]{c }
\frac{\partial }{\partial \lambda_1} \\ \frac{\partial }{\partial \lambda_2} \\ \vdots \\ \frac{\partial }{\partial \lambda_d}
\end{array}
\right] =
Vec\left[
\begin{array}[]{c c c c}
\frac{\partial f_1}{\partial \lambda_1} & \frac{\partial f_1}{\partial \lambda_2} & \cdots & \frac{\partial f_1}{\partial \lambda_d} \\
\frac{\partial f_2}{\partial \lambda_1} & \ddots & &\vdots \\
\vdots & & \ddots &\vdots \\
\frac{\partial f_m}{\partial \lambda_1} & \frac{\partial f_m}{\partial \lambda_2} & \cdots & \frac{\partial f_m}{\partial \lambda_d}
\end{array}
\right]' \\
\Rightarrow \mathbf{D}_{\bs{\lambda}}^\otimes \mathbf{f}(\bs{\lambda}) &= Vec\left(\frac{\partial \mathbf{f}}{\partial \bs{\lambda}'}\right)' = Vec \left( \frac{\partial}{\partial \bs{\lambda}}\mathbf{f}' \right)
\end{align}
where, the $Vec$ operator converts $m \times d$ matrix into an $ md \times 1$ column vector by stacking the columns one after an other.
The operator $\mathbf{D}_{\bs{\lambda}}^{\otimes } $ is called Kronecker derivative operator or simply, K-derivative operator.
\end{definition}
Thus, the Kronecker product with the vector differential operator, obtains vectorization of the transposed Jacobian of a vector function. Corresponding to the definition, the $k^{th}$ order differentiation is given by:
\begin{align}
\mathbf{D}_{\bs{\lambda}}^{\otimes k} \mathbf{f} = \mathbf{D}_{\bs{\lambda}}^{\otimes}\left( \mathbf{D}_{\bs{\lambda}}^{\otimes k-1}\mathbf{f} \right) = [f_1(\lambda), f_2(\lambda), \ldots, f_m(\lambda)]' \otimes \left[ \frac{\partial }{\partial \lambda_1}, \frac{\partial }{\partial \lambda_2}, \cdots, \frac{\partial }{\partial \lambda_d} \right]^{'\otimes k}
\end{align}
The $\mathbf{D}_{\bs{\lambda}}^{\otimes k}\mathbf{f}$ is a column vector of dimension $md^k \times 1$.
Some important properties of the K-derivative operator, those are useful in the further derivations, are listed in Appendix \ref{KderProp}.
\subsection{Application of the K-derivative operator to the multivariate Taylor series}
Let $\mathbf{x} = \left(X_1, X_2,..., X_d\right)'$ be a d-dimensional column vector and $f(\mathbf{x})$ be the function of several variables differentiable in each variable. Using the defined K-derivative operator, the Taylor series for $f(\mathbf{x})$, expanding it at origin, is given as:
\begin{equation}
\label{ndTaylor}
f(\mathbf{x}) = \sum_{m=0}^{m=\infty}\frac{1}{m!}\mathbf{c}(m,d)'\mathbf{x}^{\otimes m}
\end{equation}
where, $\mathbf{c}(m,d)$ is the vector of dimension $d^m \times 1$ and given in terms of the derivative vector $\mathbf{D}_{\mathbf{x}} = \left(\frac{\partial}{\partial x_1}, \frac{\partial}{\partial x_2}, \cdots, \frac{\partial}{\partial x_d}\right)'$ as
\begin{equation}
\mathbf{c}(m,d) = \left( \mathbf{D}_{\mathbf{x}}^{\otimes m} f(\mathbf{x})\right)\vline_{\mathbf{x}=\mathbf{0}}
\end{equation}
The Taylor series expansion near $\bs{\lambda} = \mathbf{0}$, called the Maclaurian series, of some required functions based on the Equation \eqref{ndTaylor} are derived in appendix \ref{Tayexpand}.
\section{Moments, cumulants and characteristic function of a random vector}
\label{chfn}
Let $\mathbf{x} = \left(X_1, X_2,..., X_d\right)'$ be a d-dimensional random vector and $f(\mathbf{x})$ be its joint \textit{pdf} differentiable in each variable.
\hspace{0.2 in} The Characteristic function ($\mathcal{F}$) of $\mathbf{x}$ is defined as the expected value of $\mbox{e}^{i\mathbf{x}'\bs{\lambda}}$, where $\bs{\lambda} = (\lambda_1, \lambda_2, \ldots, \lambda_d)'$, $\bs{\lambda} \in \mathbb{R}^d$.
Also, both the characteristic function and the \textit{pdf} are the Fourier Transform ($ \mathsf{F}$) of each other, in the sense they are dual.
\begin{align}
\mathcal{F}_{\mathbf{x}}(\bs{\lambda}) = E\left\{\mbox{e}^{ i (\mathbf{x}'\bs{\lambda}) } \right\} = \mathsf{F}(f(\mathbf{x})) = \int_{-\infty}^{\infty}\cdots \int_{-\infty}^{\infty} f(\mathbf{x})\mbox{e}^{i(\mathbf{x}'\bs{\lambda})}d\mathbf{x}
\end{align}
Expanding $\mbox{e}^{i\mathbf{x}'\bs{\lambda}}$ using its Maclaurian series in Equation \eqref{Tayeax} in appendix \ref{Tayexpand},
we get:
\begin{align}
\mathcal{F}_{\mathbf{x}}(\bs{\lambda}) = \sum_{k=0}^{\infty} \mathbf{m}(k,d)'\frac{(i\bs{\lambda})^{\otimes k}}{k!}
\end{align}
where, $\mathbf{m}(k,d)$ is the $k^{th}$ order moment vector of dimension $d^k \times 1$ and given by
\begin{align}
\mathbf{m}(k,d) &= \int_{\mathbb{R}^{d}}\mathbf{x}^{\otimes k}f(\mathbf{x})d\mathbf{x} \nonumber \\
\label{deltafx}
\mbox{Also, } f(\mathbf{x}) = \mathsf{F}^{-1}(\mathcal{F}(\lambda)) &= \mathsf{F}^{-1}\left(\sum_{k=0}^{\infty} \mathbf{m}(k,d)'\frac{(i\bs{\lambda})^{\otimes k}}{k!} \right) \\
&= \sum_{k=0}^{\infty} \frac{\mathbf{m}(k,d)'}{k!} \left( \frac{1}{(2\pi)^d} \int_{\mathbb{R}^{d}}(i\bs{\lambda)}^{\otimes k} \mbox{e}^{-i\mathbf{x}'\bs{\lambda}}d\bs{\lambda} \right) \nonumber \\
\label{fxdelta}
& = \sum_{k=0}^{\infty} (-1)^k \frac{\mathbf{m}(k,d)'}{k!} \mathbf{D}^{(k)}\delta (\mathbf{x}) \\
& \quad ( \because \mbox{Proof in Appendix \ref{prfderdel}}) \nonumber
\end{align}
The Moment Generating Function (MGF) of $f(\mathbf{X})$ is given as
\begin{align}
\mathbf{M}(\bs{\lambda}) & = E\left\{\mbox{e}^{ \mathbf{x}'\bs{\lambda} } \right\} = \int_{\mathbb{R}^{d}}f(\mathbf{X})\mbox{e}^{\mathbf{x}'\bs{\lambda}}d\mathbf{X} \\
\label{ndMGF}
& = \sum_{k=0}^{\infty} \mathbf{m}(k,d)'\frac{\bs{\lambda}^{\otimes k}}{k!} \quad \mbox{ ( $\because$ Expanding $\mbox{e}^{\mathbf{x}'\bs{\lambda}}$)}
\end{align}
Assuming $\mathbf{M}(\bs{\lambda})$ and $\mathcal{F}(\bs{\lambda})$ are expanded using Taylor series,
\begin{equation}
\label{momkd}
\mathbf{m}(k,d) = \mathbf{D}_{\bs{\lambda}}^{\otimes k} \mathbf{M}(\bs{\lambda})\vline_{\bs{\lambda}=\mathbf{0}} = (-i)^k \mathbf{D}_{\bs{\lambda}}^{\otimes k} \mathcal{F}_{\mathbf{x}}(\bs{\lambda})\vline_{\bs{\lambda}=\mathbf{0}}
\end{equation}
The Cumulant Generating Function (CGF) of $f(\mathbf{X})$ is given by,
\begin{equation}
\label{ndCGF}
\mathbf{C}(\bs{\lambda}) = \ln \mathbf{M}(\bs{\lambda}) = \sum_{k=1}^{\infty}\mathbf{c}(k,d)'\frac{\bs{\lambda}^{\otimes k}}{k!}
\end{equation}
where, $\mathbf{c}(k,d)$ is the $k^{th}$ order cumulant vector of dimension $d^k \times 1$. \\
The Cumulant Generating Function (CGF) of $f(\mathbf{X})$ can also be defined using the characteristic function, as under:
\begin{equation}
\label{ndCGF2}
\mathcal{C}(\bs{\lambda}) = \ln \mathcal{F}(\bs{\lambda}) = \sum_{k=1}^{\infty}\mathbf{c}(k,d)'\frac{(i\bs{\lambda})^{\otimes k}}{k!}
\end{equation}
Assuming $\mathbf{C}(\bs{\lambda})$ and $\mathcal{C}(\bs{\lambda})$ have been expanded using Taylor series,
\begin{equation}
\label{cumkd}
\mathbf{c}(k,d) = \mathbf{D}_{\bs{\lambda}}^{\otimes k} \mathbf{C}(\bs{\lambda})\vline_{\bs{\lambda}=\mathbf{0}} = (-i)^k \mathbf{D}_{\bs{\lambda}}^{\otimes k} \mathcal{C}_{\mathbf{x}}(\bs{\lambda})\vline_{\bs{\lambda}=\mathbf{0}}
\end{equation}
Significantly, this section has derived the moments and the cumulants of a random vector in vector notations.
\subsection{Relation between the cumulant vectors and the moment vectors}
\label{RelCumMom}
The relation between the moments and the cumulants is given by combining Equation \eqref{ndMGF} and Equation \eqref{ndCGF} as below:
\begin{equation}
\label{momcumeq}
\mathbf{M}(\bs{\lambda}) = \sum_{k=0}^{\infty} \mathbf{m}(k,d)'\frac{\bs{\lambda}^{\otimes k}}{k!} = \exp \left( \sum_{k=1}^{\infty}\mathbf{c}(k,d)'\frac{\bs{\lambda}^{\otimes k}}{k!}\right)
\end{equation}
For $k=1$, using Equation \eqref{momkd}, we get:
\begin{align*}
\mathbf{m}(1,d) = \mathbf{D}_{\bs{\lambda}}^{\otimes 1} \mathbf{M}(\bs{\lambda})\vline_{\bs{\lambda}=\mathbf{0}}
\end{align*}
Applying K-derivative $(\mathbf{D}_{\bs{\lambda}}^{\otimes })$ to Equation \eqref{momcumeq},
\begin{align*}
\sum_{p=1}^{\infty} \mathbf{m}(p,d)'\frac{ \left( \bs{\lambda}^{\otimes (p-1)} \otimes \mathbf{I}_d \right) }{(p-1)!}\vline_{\bs{\lambda} = \mathbf{0}} &= \sum_{p=1}^{\infty}\mathbf{c}(p,d)'\frac{ \left( \bs{\lambda}^{\otimes (p-1)} \otimes \mathbf{I}_d \right) }{(p-1)!} \nonumber \\
& \quad \otimes \exp \left( \sum_{q=1}^{\infty}\mathbf{c}(q,d)'\frac{\bs{\lambda}^{\otimes q}}{q!}\right)\vline_{\bs{\lambda}= \mathbf{0}} \\
\Rightarrow \mathbf{m}(1,d) &= \mathbf{c}(1,d)
\end{align*}
Similarly, based on Equation \eqref{momkd}, for any k taking $k^{th}$ order K-derivative of above Equation \eqref{momcumeq} on both sides relates $\mathbf{m}(k,d)$ with $\mathbf{c}(k,d)$. For example, the cases for $k=2$ and $k=3$ are shown in Appendix \ref{Apmomcum}. Overall, the vector moments in terms of the vector cumulants can be summarized as under:
\begin{equation}
\label{eqndmom2cum}
\begin{aligned}
\mathbf{m}(1,d) &= \mathbf{c}(1,d) \\
\mathbf{m}(2,d) &= \mathbf{c}(2,d) + \mathbf{c}(1,d)^{\otimes 2} \\
\mathbf{m}(3,d) &= \mathbf{c}(3,d) + 3\mathbf{c}(2,d)\otimes \mathbf{c}(1,d) + \mathbf{c}(1,d)^{\otimes 3} \\
\mathbf{m}(4,d) &= \mathbf{c}(4,d) + 4\mathbf{c}(3,d)\otimes \mathbf{c}(1,d) + 3\mathbf{c}(2,d)^{\otimes 2} + 6 \mathbf{c}(2,d) \\
& \quad \otimes \mathbf{c}(1,d)^{\otimes 2}+ \mathbf{c}(1,d)^{\otimes 4} \\
\mathbf{m}(5,d) &= \mathbf{c}(5,d) + 5\mathbf{c}(4,d)\otimes \mathbf{c}(1,d)+ 10\mathbf{c}(3,d)\otimes \mathbf{c}(2,d)
\\
& \quad + 10\mathbf{c}(3,d) \otimes \mathbf{c}(1,d)^{\otimes 2} + 15\mathbf{c}(2,d)^{\otimes 2}\otimes \mathbf{c}(1,d) \\
& + 10\mathbf{c}(2,d)\otimes \mathbf{c}(1,d)^{\otimes 3} + \mathbf{c}(1,d)^{\otimes 5} \\
\mathbf{m}(6,d) &= \mathbf{c}(6,d) + 6\mathbf{c}(5,d)\otimes \mathbf{c}(1,d) + 15 \mathbf{c}(4,d) \otimes \mathbf{c}(2,d) \\
& \quad + 15 \mathbf{c}(4,d) \otimes \mathbf{c}(1,d)^{\otimes 2} + 10 \mathbf{c}(3,d)^{\otimes 2} + 60\mathbf{c}(3,d)\otimes \mathbf{c}(2,d) \\
& \quad \otimes \mathbf{c}(1,d) + 20\mathbf{c}(3,d)\otimes \mathbf{c}(1,d)^{\otimes 2}
+ 15\mathbf{c}(2,d)^{\otimes 3} \\
& + 45\mathbf{c}(2,d)^{\otimes 2}\otimes \mathbf{c}(1,d)^{\otimes 2} + 15\mathbf{c}(2,d)\otimes \mathbf{c}(1,d)^{\otimes 4} + \mathbf{c}(1,d)^{\otimes 6}
\end{aligned}
\end{equation}
The above set of equations can be represented through more compact formulas as under. The Equation \eqref{ndmomcum} gives generalized $k^{th}$ order d-variate cumulant vector in terms of the moment vectors and the
Equation \eqref{ndcummom} gives the vice-a-versa.
\begin{align}
\label{ndmomcum}
\mathbf{m}(k,d) = \sum_{p=0}^{k-1} \binom{k-1}{p} \mathbf{K}^{-1}_{\mathfrak{p} h \leftrightarrow l}\mathbf{c}(k-p,d) \otimes \mathbf{m}(p,d) \\
\label{ndcummom}
\mathbf{c}(k,d) = \mathbf{m}(k,d) - \sum_{p=1}^{k-1} \binom{k-1}{p} \mathbf{K}^{-1}_{\mathfrak{p} h \leftrightarrow l}\mathbf{c}(k-p,d) \otimes \mathbf{m}(p,d)
\end{align}
where, $\mathbf{K}^{-1}_{\mathfrak{p} h \leftrightarrow l}$ is a specific commutation matrix with corresponding dimensions that changes the place of the cumulants for Kronecker product such that the expression has decreasing order cumulants from left to the right, i.e. the higher order cumulant vector on the left and the lower order cumulant vector on the right. As the Kronecker products are non-commutative, without using the commutation matrices it would have been impossible to derive the compact formula.
The derived multivariate expressions reduce to the following expressions for dimension $d=1$ and are exactly same as those derived in \citep{genGramC07}.
\begin{equation}
\begin{aligned}
c(k,1) &= m(k,1) - \sum_{p=1}^{k-1} \binom{k-1}{p} c(k-p,1) m(p,1) \\
\mbox{or more simply, } c_k & = m_k - \sum_{p=1}^{k-1} \binom{k-1}{p} c_{k-p} m_p
\end{aligned}
\end{equation}
Thus, the derived multivariate expressions in Equation \eqref{eqndmom2cum} are elementary vector extensions to those for univariate.
\section{Multivariate \textit{pdf} representation in terms of the cumulants}
\label{pdfcum}
From Equation \eqref{deltafx} and Equation \eqref{ndCGF2}, the multivariate \textit{pdf} $f(\mathbf{x})$ can be written as:
\begin{align}
\label{ndpdf}
f(\mathbf{x}) = \mathsf{F}^{-1}(\mbox{e}^{\mathcal{C}(\bs{\lambda})}) = \left(\frac{1}{2\pi}\right)^d\int_{\mathbb{R}^{d}}\exp \left( \sum_{k=1}^{\infty}\mathbf{c}(k,d)'\frac{(i\bs{\lambda})^{\otimes k}}{k!} \right) \exp{(-i \mathbf{x}'\bs{\lambda})}d \bs{\lambda}
\end{align}
As \textit{pdf} is a real function and $\mbox{Re}(e^{A+iB}e^{-iC}) = e^A\cos(B-C)= e^A\cos(C-B)$, the Equation \eqref{ndpdf} can be re-written as:
\begin{align}
\label{ndpdfreal}
f(\mathbf{x}) &= \frac{1}{(2\pi)^d}\int_{\mathbb{R}^{d}} \exp{\left( \sum_{k=1}^{\infty}\frac{\mathbf{c}(2k,d)'}{2k!} {(i\bs{\lambda})^{\otimes 2k}} \right) } \nonumber \\
& \quad \cos\left( \mathbf{x}'\bs{\lambda} + \sum_{k=1}^{\infty}\frac{\mathbf{c}(2k-1,d)'}{(2k-1)!} (i)^{2k} (\bs{\lambda})^{\otimes (2k-1)} \right) d \bs{\lambda}
\end{align}
The integrand in this equation is an even function. So,
\begin{align}
\label{ndpdfeven}
f(\mathbf{x}) &= \frac{1}{(\pi)^d}\int_{(\mathbb{R}^{+})^d} \exp \left( \sum_{k=1}^{\infty}\frac{\mathbf{c}(2k,d)'}{2k!} {(i\bs{\lambda})^{\otimes 2k}} \right) \nonumber \\
& \quad \cos\left( \mathbf{x}'\bs{\lambda} + \sum_{k=1}^{\infty}\frac{\mathbf{c}(2k-1,d)'}{(2k-1)!} (i)^{2k}(\bs{\lambda})^{\otimes (2k-1)}\right) d \bs{\lambda}
\end{align}
where, $\mathbb{R}^{+} = \{ x \in \mathbb{R} : x \geq 0 \}$.
The Equation \eqref{ndpdfreal} and the Equation \eqref{ndpdfeven} give a multivariate \textit{pdf} in terms of the cumulants. As they are derived using Taylor series expansion, the infinite order differentiability is an implicit assumption.
\hspace{0.2 in} The equations can be verified using known \textit{pdf} examples with finite number of moments and cumulants.
Let say, the
impulse delta density function has only the first order cumulant being non-zero and all other higher order cumulants are zero. Using this knowledge in Equation \eqref{ndpdfreal},
\begin{align}
f(\mathbf{x}) &= \frac{1}{(2\pi)^d}\int_{\mathbb{R}^{d}} \cos\left( \left( \mathbf{x} - \mathbf{c}(1,d)\right)'\bs{\lambda} \right) d \bs{\lambda} \\
&= \delta (\mathbf{x} - \mathbf{c}(1,d)) \quad \mbox{(shifted impulse delta function)} \nonumber
\end{align}
Let's take another example, the Gaussian density function has first two order cumulants nonzero and all other order cumulants are zero. Using this knowledge in Equation \eqref{ndpdfreal},
\begin{equation}
\label{Gpdf}
\begin{aligned}
f(\mathbf{x}) &= \frac{1}{(2\pi)^d}\int_{\mathbb{R}^{d}} \exp \left( - \frac{\mathbf{c}(2,d)'}{2}\bs{\lambda}^{\otimes 2} \right) \cos\left( \left( \mathbf{x} - \mathbf{c}(1,d)\right)'\bs{\lambda} \right) d \bs{\lambda} \\
&= G(\mathbf{x}) \quad \mbox{ ( See Appendix \ref{Gaussint} for proof)}
\end{aligned}
\end{equation}
\section{The multivariate Hermite polynomials in integral form}
\label{DerGauss}
An interesting application of the integral form of multivariate \textit{pdf} representation is achieved in this section.
The multivariate Gaussian expressed as in Equation \eqref{Gpdf} is used to derive it's differentials and Hermite polynomials in a simple way. Taking $k^{th}$ K-derivative of $G(\mathbf{x})$,
\begin{align}
\label{derkGx}
G^{(k)}(\mathbf{x})
& := \mathbf{D}_{\mathbf{x}}^{\otimes k}G(\mathbf{x})
= \frac{1}{(2\pi)^d}\int_{\mathbb{R}^{d}} \bs{\lambda}^{\otimes k} \exp \left( -\frac{\mathbf{c}(2,d)'}{2}\bs{\lambda}^{\otimes 2} \right) \nonumber \\
& \quad \cos\left( \left( \mathbf{x} - \mathbf{c}(1,d) \right)'\bs{\lambda} + \frac{k\pi}{2} \right) d \bs{\lambda}
\end{align}
The multivariate Hermite polynomials defined by \citet{dnHermite96} are defined as under:
\begin{align}
\label{hermitend}
\mathbf{H}_k(\mathbf{x}; \mathbf{0},\mathbf{C}_{\mathbf{x}}) &= [G(\mathbf{x};\mathbf{0},\mathbf{C}_{\mathbf{x}})]^{-1} (-1)^{k} \left( \mathbf{C}_{\mathbf{x}} \mathbf{D}_x \right)^{\otimes k} G(\mathbf{x};\mathbf{0},\mathbf{C}_{\mathbf{x}}) \\
\mbox{where, } G(\mathbf{x};\mathbf{0},\mathbf{C}_{\mathbf{x}}) &= \left|\mathbf{C}_{\mathbf{x}} \right|^{-1/2}(2\pi)^{-d/2}\exp\left(- \frac{1}{2}\mathbf{x}' \mathbf{C}_{\mathbf{x}}^{-1} \mathbf{x} \right) \nonumber \\ & = \left|\mathbf{C}_{\mathbf{x}} \right|^{-1/2}(2\pi)^{-d/2}\exp\left( - \frac{1}{2} \left( Vec \mbox{ } \mathbf{C}_{\mathbf{x}}^{-1}\right)' \mathbf{x}^{\otimes 2} \right)
\end{align}
This is equivalent to the 1-dimensional definition of Hermite polynomials\footnote{This is the 'probabilists' Hermite polynomials and not the 'physicists' Hermite polynomials used by \citet{genGramC07}.} by Rodrigues's formula in Equation \eqref{hermite1d}, except the introduction of matrix $\mathbf{C}_{\mathbf{x}}$.
\begin{align}
\label{hermite1d}
H_k(x) = [G(x)]^{-1}(-1)^k \frac{d^k}{dx^k}G(x) \mbox{ where, } G(x) = \frac{1}{2\pi}\mbox{e}^{-\frac{1}{2}x^2}
\end{align}
Using Equation \eqref{derkGx}, the Equation \eqref{hermitend} for multivariate Hermite polynomials is rewritten as:
\begin{align}
\label{hermitenew}
\mathbf{H}_k & (\mathbf{x}; \mathbf{0},\mathbf{C}_{\mathbf{x}} )
= (2\pi)^{d/2} \left|\mathbf{C}_{\mathbf{x}}\right|^{1/2} (-1)^k \left(\mathbf{C}_{\mathbf{x}} \right)^{\otimes k} \exp \left( \frac{1}{2} \left(Vec \mbox{ } \mathbf{C}_{\mathbf{x}}^{-1}\right)' \mathbf{x}^{\otimes 2} \right) \frac{1}{(2\pi)^d} \nonumber \\
&\quad \int_{\mathbb{R}^{d}} \bs{\lambda}^{\otimes k} \exp \left( -\frac{\mathbf{c}(2,d)'}{2}\bs{\lambda}^{\otimes 2} \right) \cos\left( (\mathbf{x} - \mathbf{c}(1,d))'\bs{\lambda} + \frac{k\pi}{2} \right) d \bs{\lambda}
\end{align}
Taking $\mathbf{C}_{\mathbf{x}} = \mathbf{I}_d$, where $\mathbf{I}_d$ is $d \times d$ identity matrix; using the property $(Vec \mbox{ } \mathbf{I}_d )'\mathbf{x}^{\otimes 2} = \mathbf{x}'\mathbf{x}$ and using change of variable as $\bs{\lambda}/\sqrt{2} = \mathbf{u}$ - the integral form of multivariate Hermite polynomials is obtained as under:
\begin{align}
\label{hermitend01}
\mathbf{H}_k(\mathbf{x}; \mathbf{0},\mathbf{I}_d) & = (2)^{ \frac{k+d+1}{2} } (\pi)^{-d/2}
\exp \left( \frac{1}{2}\mathbf{x}' \mathbf{x}\right) \nonumber\\
& \quad \int_{\mathbb{R}^{d}} \mathbf{u}^{\otimes k}
\exp \left( -\mathbf{u}'\mathbf{u} \right) \cos\left( \sqrt{2}\mathbf{x}'\mathbf{u} - \frac{k\pi}{2} \right) d \mathbf{u}
\end{align}
\hspace{0.2 in} The result in Equation \eqref{derkGx} can also be obtained using Equation \eqref{Gpdf} and applying the derivative property of Fourier transform ($\mathsf{F}$). The $k^{th}$ derivative of $G(x)$ is given by,
\begin{align}
\label{eqFgk}
G^{(k)}(\mathbf{x}) &= \mathsf{F}^{-1}\left( \mathsf{F}(G^{(k)}(\mathbf{x}))\right) \\
&= \frac{1}{(2\pi)^d} \int_{\mathbb{R}^{d}} \left( i\bs{\lambda} \right)^{\otimes k}\mathsf{F}(G(\mathbf{x})) \exp^{-i\mathbf{x}'\bs{\lambda}} d \bs{\lambda} \\
&= \frac{1}{(2\pi)^d}\int_{\mathbb{R}^{d}} \bs{\lambda}^{\otimes k} \exp \left( -\frac{\mathbf{c}(2,d)'}{2}\bs{\lambda}^{\otimes 2} \right) \cos\left( \left( \mathbf{x} - \mathbf{c}(1,d) \right)'\bs{\lambda} + \frac{k\pi}{2} \right) d \bs{\lambda}
\end{align}
\section[Multivariate Gram-Charlier A series ]{Multivariate Gram-Charlier A series}
\label{ndGCAseries}
Till now, the article has derived - an unknown \textit{pdf} expressed in terms of its cumulants in Equation \eqref{ndpdfreal}; the Gaussian density function expressed in terms of its cumulants in Equation \eqref{Gpdf} and the Hermite polynomials in Equation \eqref{hermitend01}. Based on them, the multivariate Gram Charlier A series that expresses an unknown \textit{pdf} using Gaussian density as a reference can be obtained. The expansion assumes first and second order cumulants being same for both the unknown \textit{pdf} and the reference \textit{pdf}. Using the expansion $\exp(A+B)\cos(C+D) = \exp(A)\exp(B)(\cos C \cos D -\sin C \sin D)$, the Equation \eqref{ndpdfreal} can be re-written as:
\begin{align}
\label{ndpdfGx}
f(\mathbf{x}) &= \frac{1}{(2\pi)^d}\int_{\mathbb{R}^{d}}
\exp \left(- \frac{\mathbf{c}(2,d)'}{2}\bs{\lambda}^{\otimes 2} \right)
\exp{ \left( \sum_{k=2}^{\infty}\frac{\mathbf{c}(2k,d)'}{2k!} {(i\bs{\lambda})^{\otimes 2k}} \right) } \nonumber \\
& \quad \left\{
\cos( ( \mathbf{x} - \mathbf{c}(1,d) )'\bs{\lambda} ) \cos \left( \sum_{k=2}^{\infty} \frac{ \mathbf{c}(2k-1,d)' }{(2k-1)!} (i)^{2k} (\bs{\lambda})^{\otimes (2k-1)} \right) \right. \nonumber \\
& \quad \left. - \sin( (\mathbf{x} - \mathbf{c}(1,d))'\bs{\lambda}) \sin \left( \sum_{k=2}^{\infty}\frac{\mathbf{c}(2k-1,d)'}{(2k-1)!} (i)^{2k} (\bs{\lambda})^{\otimes (2k-1)} \right) \right\} d \bs{\lambda}
\end{align}
Using the expansions in Appendix \ref{Tayexpand}, parts of the Equation \eqref{ndpdfGx} can be approximated upto maximum $6^{th}$-order statistics
as under:
\begin{align}
\label{excoscum}
\exp & \left( \sum_{k=2}^{\infty}\frac{\mathbf{c}(2k,d)'}{2k!} (i\bs{\lambda})^{\otimes 2k} \right) \cos \left( \sum_{k=2}^{\infty} \frac{ \mathbf{c}(2k-1,d)' }{(2k-1)!} (i)^{2k} (\bs{\lambda})^{\otimes (2k-1)} \right) \nonumber \\
& \quad = 1 + \left( \frac{\mathbf{c}(4,d)'\bs{\lambda}^{\otimes 4}}{4!} - \frac{\mathbf{c}(6,d)'\bs{\lambda}^{\otimes 6}}{6!}\right) - \frac{1}{2}\left( \frac{\mathbf{c}(3,d)'\bs{\lambda}^{\otimes 3}}{3!} \right)^{\otimes 2} + \ldots \\
\label{exsincum}
\exp & \left( \sum_{k=2}^{\infty} \frac{\mathbf{c}(2k,d)'}{2k!} (i\bs{\lambda})^{\otimes 2k} \right) \sin \left( \sum_{k=2}^{\infty}\frac{\mathbf{c}(2k-1,d)'}{(2k-1)!} (-i)^{2k} (\bs{\lambda})^{\otimes (2k-1)} \right) \nonumber \\
&\quad = \frac{\mathbf{c}(3,d)'\bs{\lambda}^{\otimes 3}}{3!} - \frac{\mathbf{c}(5,d)'\bs{\lambda}^{\otimes 5}}{5!} + \ldots
\end{align}
Using above Equation \eqref{excoscum} and Equation \eqref{exsincum}, the Equation \eqref{ndpdfGx} can be re-written as:
\begin{align}
f(\mathbf{x}) &= \frac{1}{(2\pi)^d} \int_{\mathbb{R}^{d}} \left\{
\exp \left( - \frac{\mathbf{c}(2,d)'}{2}\bs{\lambda}^{\otimes 2} \right) \cos( ( \mathbf{x} - \mathbf{c}(1,d) )'\bs{\lambda} ) \right. \nonumber \\
& \quad \left. \left( 1 + \frac{\mathbf{c}(4,d)'\bs{\lambda}^{\otimes 4}}{4!} - \frac{1}{6!} \left( \mathbf{c}(6,d) - 10 \mathbf{c}(3,d)^{\otimes 2} \right)'{\bs{\lambda}^{\otimes 6} } + \ldots \right) \right\} d \bs{\lambda} \nonumber \\
& \quad - \frac{1}{(2\pi)^d}\int_{\mathbb{R}^{d}} \left\{
\exp \left(- \frac{\mathbf{c}(2,d)'}{2}\bs{\lambda}^{\otimes 2} \right) \cos \left( \left(\mathbf{x} - \mathbf{c}(1,d)\right)'\bs{\lambda} - \frac{\pi}{2}\right) \right. \nonumber \\
& \quad \left. \left( \frac{\mathbf{c}(3,d)'\bs{\lambda}^{\otimes 3}}{3!} - \frac{\mathbf{c}(5,d)'\bs{\lambda}^{\otimes 5}}{5!} + \ldots
\right) \right\} d \bs{\lambda}
\end{align}
Using the Equation \eqref{derkGx} for derivatives of Gaussian defined, the above Equation can be simplified as:
\begin{align}
\label{GCA}
f(\mathbf{x}) & = G(\mathbf{x}) - \frac{\mathbf{c}(3,d)'}{3!} G^{(3)}(\mathbf{x}) + \frac{\mathbf{c}(4,d)'}{4!}G^{(4)}(\mathbf{x}) - \frac{\mathbf{c}(5,d)'}{5!} G^{(5)}(\mathbf{x}) \nonumber \\
& \quad + \frac{\mathbf{c}(6,d)' + 10 \mathbf{c}(3,d)^{\otimes 2'}}{6!} G^{(6)}(\mathbf{x}) + \ldots
\end{align}
The Equation \eqref{GCA} is the Gram-Charlier A series expressed directly in terms of the cumulants and the derivatives of the Gaussian \textit{pdf}. Usually, the GCA is represented in terms of the Hermite polynomials. So, the GCA expansion (Equation \eqref{GCA}) in terms of the Hermite polynomials; either using definition in Equation \eqref{hermitend} or using $\mathbf{H}_k( \mathbf{x}; \mathbf{0}, \mathbf{C}_{\mathbf{x}}^{-1})$ derived in Equation \eqref{hermitenew}; can be re-written as:
\begin{align}
\label{GCAhermitecx}
f(\mathbf{x}) & = G(\mathbf{x}) \left[ 1 + \frac{\mathbf{c}(3,d)'}{3!} \left( \mathbf{C}_{\mathbf{x}}^{-1} \right)^{\otimes 3} \mathbf{H}_3(\mathbf{x}; \mathbf{0},\mathbf{C}_{\mathbf{x}} ) + \frac{\mathbf{c}(4,d)'}{4!}\left( \mathbf{C}_{\mathbf{x}}^{-1} \right)^{\otimes 4}\mathbf{H}_4(\mathbf{x}; \mathbf{0},\mathbf{C}_{\mathbf{x}} ) \right. \nonumber \\
& \quad \left. + \frac{\mathbf{c}(5,d)'}{5!} \left( \mathbf{C}_{\mathbf{x}}^{-1} \right)^{\otimes 5}\mathbf{H}_5(\mathbf{x}; \mathbf{0},\mathbf{C}_{\mathbf{x}} )
+ \frac{\mathbf{c}(6,d)' + 10 \mathbf{c}(3,d)^{\otimes 2'}}{6!} \left( \mathbf{C}_{\mathbf{x}}^{-1} \right)^{\otimes 6} \right. \nonumber \\
& \quad \left. \mathbf{H}_6(\mathbf{x}; \mathbf{0},\mathbf{C}_{\mathbf{x}} ) + \ldots \right] \\
\label{GCAhermite}
f(\mathbf{x}) & = G(\mathbf{x}) \left[ 1 + \frac{\mathbf{c}(3,d)'}{3!} \mathbf{H}_3(\mathbf{x}; \mathbf{0},\mathbf{I}_d ) + \frac{\mathbf{c}(4,d)'}{4!}\mathbf{H}_4(\mathbf{x}; \mathbf{0},\mathbf{I}_d ) + \frac{\mathbf{c}(5,d)'}{5!} \mathbf{H}_5(\mathbf{x}; \mathbf{0},\mathbf{I}_d ) \right. \nonumber \\
& \quad \left. + \frac{\mathbf{c}(6,d)' + 10 \mathbf{c}(3,d)^{\otimes 2'}}{6!} \mathbf{H}_6(\mathbf{x}; \mathbf{0},\mathbf{I}_d ) + \ldots \right]
\end{align}
Finally, the GCA series, in vector notations, can be expressed either using $k^{th}$ order derivative of Gaussian $(G_k(\mathbf{x}))$ or using $k^{th}$ order vector Hermite polynomials $(\mathbf{H}_k(\mathbf{x}))$ as under:
\begin{align}
\label{ndGCAshort}
f(\mathbf{x}) &= \sum_{k=0}^{\infty} (-1)^k\frac{\mathbf{c}(k,d)'}{k!}G^{(k)}(\mathbf{x}) \\
\label{ndGCAhershort}
& = \sum_{k=0}^{\infty} \frac{\mathbf{c}(k,d)'}{k!}\mathbf{H}_k(\mathbf{x}; \mathbf{0},\mathbf{I}_d)
\end{align}
\section{Multivariate Generalized Gram-Charlier series}
\label{ndGGC}
To derive the generalized Gram-Charlier series, an unknown \textit{pdf} $f(\mathbf{x})$ need be represented in terms of any known reference \textit{pdf} $\psi(\mathbf{x})$, where both the \textit{pdf}s are represented in terms of their cumulants.
Let the $k^{th}$ order cumulant vector of the reference \textit{pdf} $\psi(\mathbf{x})$ be $\mathbf{c}_r(k,d)$. Then, the $k^{th}$ order cumulant difference vector $\bs{\delta}(k,d)$ is: $\bs{\delta}(k,d) = \mathbf{c}(k,d) - \mathbf{c}_r(k,d), \forall k$. Using $\bs{\delta}(k,d)$, the Equation \eqref{ndpdfreal} can be re-written
as under:
\begin{align}
f(\mathbf{x}) &= \frac{1}{(2\pi)^d}\int_{\mathbb{R}^{d}} \exp{\left( \sum_{k=1}^{\infty}\frac{\mathbf{c}_r(2k,d)'}{2k!} {(i\bs{\lambda})^{\otimes 2k}} \right) }\exp{\left( \sum_{k=1}^{\infty}\frac{\bs{\delta}(2k,d)'}{2k!} {(i\bs{\lambda})^{\otimes 2k}} \right) } \nonumber \\
& \quad \cos\left( \mathbf{x}'\bs{\lambda} + \left( \sum_{k=1}^{\infty}\frac{\mathbf{c}_r(2k-1,d)}{(2k-1)!} + \sum_{k=1}^{\infty}\frac{\bs{\delta}(2k-1,d)}{(2k-1)!} \right)' (i)^{2k} (\bs{\lambda})^{\otimes (2k-1)} \right) d \bs{\lambda} \\
\label{ndpdfcrdelta}
& =\frac{1}{(2\pi)^d}\int_{\mathbb{R}^{d}} \exp{\left( \sum_{k=1}^{\infty}\frac{\mathbf{c}_r(2k,d)'}{2k!} {(i\bs{\lambda})^{\otimes 2k}} \right) } \exp{\left( \sum_{k=1}^{\infty}\frac{\bs{\delta}(2k,d)'}{2k!} {(i\bs{\lambda})^{\otimes 2k}} \right) } \nonumber \\ & \quad
\left\{ \cos\left( \mathbf{x}'\bs{\lambda} +
\sum_{k=1}^{\infty}\frac{\mathbf{c}_r(2k-1,d)'}{(2k-1)!} (i)^{2k} (\bs{\lambda})^{\otimes (2k-1)} \right)
\cos\left( \sum_{k=1}^{\infty}\frac{\bs{\delta}(2k-1,d)'}{(2k-1)!} \right. \right. \nonumber \\
& \left. \left. (i)^{2k} (\bs{\lambda})^{\otimes (2k-1)} \right)
- \sin\left( \mathbf{x}'\bs{\lambda} +
\sum_{k=1}^{\infty}\frac{\mathbf{c}_r(2k-1,d)'}{(2k-1)!} (i)^{2k} (\bs{\lambda})^{\otimes (2k-1)} \right)
\right. \nonumber \\
& \quad \left. \sin\left( \sum_{k=1}^{\infty}\frac{\bs{\delta}(2k-1,d)'}{(2k-1)!}
(i)^{2k} (\bs{\lambda})^{\otimes (2k-1)} \right) \right\}d\bs{\lambda}
\end{align}
Using the expansions in Appendix \ref{Tayexpand}, parts of the Equation \eqref{ndpdfcrdelta} can be approximated upto maximum $6^{th}$-order statistics
as under:
\begin{align}
\label{expcos}
\exp & {\left( \sum_{k=1}^{\infty}\frac{\bs{\delta}(2k,d)'}{2k!} {(i\bs{\lambda})^{\otimes 2k}} \right) }\cos\left( \sum_{k=1}^{\infty}\frac{\bs{\delta}(2k-1,d)'}{(2k-1)!} (i)^{2k} (\bs{\lambda})^{\otimes (2k-1)} \right) \nonumber \\
&= 1 + \left( \sum_{k=1}^{\infty}\frac{\bs{\delta}(2k,d)'}{2k!} {(i\bs{\lambda})^{\otimes 2k}} \right)+ \frac{1}{2}\left( \sum_{k=1}^{\infty}\frac{\bs{\delta}(2k,d)'}{2k!} {(i\bs{\lambda})^{\otimes 2k}} \right)^{\otimes 2} \nonumber \\
& \quad - \frac{1}{2}
\left( \sum_{k=1}^{\infty}\frac{\bs{\delta}(2k-1,d)'}{(2k-1)!} (i)^{2k} (\bs{\lambda})^{\otimes (2k-1)} \right)^{\otimes 2} - \ldots \nonumber \\
& = 1 + \left( - \frac{\bs{\delta}(2,d)'}{2!}\bs{\lambda}^{\otimes 2} + \frac{\bs{\delta}(4,d)'}{4!}\bs{\lambda}^{\otimes 4} - \frac{\bs{\delta}(6,d)'}{6!}\bs{\lambda}^{\otimes 6} \right) + \frac{1}{2} \left( - \frac{\bs{\delta}(2,d)'}{2!}\bs{\lambda}^{\otimes 2} \right)^{\otimes 2}
\nonumber \\
&\quad - \frac{1}{2}\left( - \bs{\delta}(1,d)'\bs{\lambda}
+ \frac{\bs{\delta}(3,d)'}{3!}\bs{\lambda}^{\otimes 3} \right)^{\otimes 2} + \frac{1}{6}\left( - \frac{\bs{\delta}(2,d)'}{2!}\bs{\lambda}^{\otimes 2} \right)^{\otimes 3}
- \frac{1}{2}\left( \left( - \frac{\bs{\delta}(2,d)'}{2!}\bs{\lambda}^{\otimes 2} \right. \right.
\nonumber \\
& \quad \left. \left. + \frac{\bs{\delta}(4,d)'}{4!}\bs{\lambda}^{\otimes 4} \right)\otimes \left( \bs{\delta}(1,d)'\bs{\lambda}\right)^{\otimes 2} \right)
- \frac{1}{4}\left( \left(- \frac{\bs{\delta}(2,d)'}{2!}\bs{\lambda}^{\otimes 2}\right)^{\otimes 2} \otimes \left( \bs{\delta}(1,d)'\bs{\lambda}\right)^{\otimes 2}\right) \nonumber \\
& \quad + \frac{1}{4!} \left( - \frac{\bs{\delta}(2,d)'}{2!}\bs{\lambda}^{\otimes 2} \right)^{\otimes 4}
- \frac{1}{4!}\left( - \bs{\delta}(1,d)'\bs{\lambda} \right)^{\otimes 4} - \left( \frac{\bs{\delta}(2,d)'}{2!}\bs{\lambda}^{\otimes 2} \otimes \bs{\delta}(1,d)'\bs{\lambda}^{\otimes 4} \right) +
\ldots \nonumber \\
& = 1 - \frac{1}{2}\left( \bs{\delta}(1,d)^{\otimes 2} + \bs{\delta}(2,d)\right)'\bs{\lambda}^{\otimes 2}
+ \frac{1}{4!} \left( \bs{\delta}(1,d)^{\otimes 4}+ 6 \bs{\delta}(2,d)\otimes \bs{\delta}(1,d)^{\otimes 2} \right. \nonumber \\ & \quad \left. + 3\bs{\delta}(2,d)^{\otimes 2}
+ 4\bs{\delta}(3,d)\otimes \bs{\delta}(1,d) + \bs{\delta}(4,d)\right)'\bs{\lambda}^{\otimes 4}
- \frac{1}{6!}\left( \bs{\delta}(1,d)^{\otimes 6} + 15\bs{\delta}(2,d)^{\otimes 3} \right. \nonumber \\ & \quad \left.
+ 10 \bs{\delta}(3,d)^{\otimes 2} +
15 \bs{\delta}(4,d) \otimes \bs{\delta}(2,d)
+ 15 \bs{\delta}(4,d)
\otimes \bs{\delta}(1,d)^{\otimes 2}
+ 20 \bs{\delta}(3,d)\otimes \bs{\delta}(1,d)^{\otimes 2} \right. \nonumber \\ &\quad \left.
+ 15\bs{\delta}(2,d)\otimes \bs{\delta}(1,d)^{\otimes 4} + 45\bs{\delta}(2,d)^{\otimes 2}\otimes \bs{\delta}(1,d)^{\otimes 2}
+ 6 \bs{\delta}(5,d)\otimes \bs{\delta}(1,d) \right. \nonumber \\ &\quad \left.
+ 60\bs{\delta}(3,d)\otimes \bs{\delta}(2,d)\otimes \bs{\delta}(1,d) + \bs{\delta}(6,d)\right)'\bs{\lambda}^{\otimes 6} + \ldots
\end{align}
Similarly,
\begin{align}
\label{expsin}
\exp & {\left( \sum_{k=1}^{\infty}\frac{\bs{\delta}(2k,d)'}{2k!} {(i\bs{\lambda})^{\otimes 2k}} \right) }\sin\left( \sum_{k=1}^{\infty}\frac{\bs{\delta}(2k-1,d)'}{(2k-1)!} (-i)^{2k} (\bs{\lambda})^{\otimes (2k-1)} \right) \nonumber \\
&= \left( - \bs{\delta}(1,d)'\bs{\lambda} + \frac{ \bs{\delta}(3,d)'}{3!}\bs{\lambda}^{\otimes 3} - \frac{ \bs{\delta}(5,d)'}{5!}\bs{\lambda}^{\otimes 5} \right)
+ \left( \frac{ \bs{\delta}(2,d)'\bs{\lambda}^{\otimes 2}}{2!} \otimes \frac{ \bs{\delta}(1,d)'\bs{\lambda}}{1} \right. \nonumber \\
& \quad \left. + \frac{ \bs{\delta}(3,d)'\bs{\lambda}^{\otimes 3}}{3!}\otimes \frac{ \bs{\delta}(2,d)'\bs{\lambda}^{\otimes 2}}{2!} + \frac{ \bs{\delta}(4,d)'\bs{\lambda}^{\otimes 4}}{4!}\otimes \frac{ \bs{\delta}(1,d)'\bs{\lambda}^{\otimes 1}}{1!} \right) \nonumber \\
& \quad - \frac{1}{2}\left( \frac{ \bs{\delta}(2,d)'\bs{\lambda}^{\otimes 2} }{2!} \right)^{\otimes 2}\otimes \frac{ \bs{\delta}(1,d)'\bs{\lambda}^{\otimes 1}}{1!} - \left( - \bs{\delta}(1,d)'\bs{\lambda}\right)^{\otimes 3} + \ldots\nonumber \\
& = \bs{\delta}(1,d)'\bs{\lambda} + \frac{1 }{3!}\left( \bs{\delta}(3,d) + 3\bs{\delta}(2,d)\otimes \bs{\delta}(1,d) + \bs{\delta}(1,d)^{\otimes 3} \right)' \bs{\lambda}^{\otimes 3}
- \frac{ 1}{5!}\left( \bs{\delta}(5,d) \right. \nonumber \\ &\quad \left.
+ 5\bs{\delta}(4,d)\otimes \bs{\delta}(1,d)+ 15\bs{\delta}(2,d)^{\otimes 2}\otimes \bs{\delta}(1,d) + 10\bs{\delta}(3,d)\otimes \bs{\delta}(2,d) \right. \nonumber \\
& \quad \left. + 10 \bs{\delta}(2,d)\otimes \bs{\delta}(1,d)^{\otimes 3} + 10 \bs{\delta}(3,d)\otimes \bs{\delta}(1,d)^{\otimes 2} + \bs{\delta}(1,d)^{\otimes 5} \right)'\bs{\lambda}^{\otimes 5} \ldots
\end{align}
Now, $\mathbf{D}_{\mathbf{x}}^{\otimes k} \psi(\mathbf{x})$ can be obtained by taking $k^{th}$-order K-derivative of Equation \eqref{ndpdfreal} as under:
\begin{align}
\label{derkref}
\psi^{(k)}(\mathbf{x}) & = \mathbf{D}_{\mathbf{x}}^{\otimes k} \psi(\mathbf{x}) = \frac{1}{(2\pi)^d}\int_{\mathbb{R}^{d}} (\bs{\lambda})^{\otimes k} \exp{\left( \sum_{k=1}^{\infty}\frac{\mathbf{c}_r(2k,d)'}{2k!} {(i\bs{\lambda})^{\otimes 2k}} \right) } \nonumber \\
& \quad \quad \cos\left( \mathbf{x}'\bs{\lambda} + \sum_{k=1}^{\infty}\frac{\mathbf{c}_r(2k-1,d)'}{(2k-1)!} (i)^{2k} (\bs{\lambda})^{\otimes (2k-1)} + \frac{k\pi}{2} \right) d \bs{\lambda}
\end{align}
The above Equation \eqref{derkref} with the previous results on expansions in Equation \eqref{expcos} and Equation \eqref{expsin} can be used to simplify the Equation \eqref{ndpdfcrdelta}.
This derives the Generalized Gram-Charlier (GGC) series expressing an unknown \textit{pdf} $f(\mathbf{x})$ in terms of the cumulant difference vectors $(\bs{\delta}(k,d))$ and derivatives of a reference \textit{pdf} $\psi^{(k)}(\mathbf{x})$ as under:
\begin{align}
\label{GGC}
f(\mathbf{x}) = \sum_{k=0}^{\infty} (-1)^k\frac{\bs{\alpha}(k,d)'}{k!}\psi^{(k)}(\mathbf{x})
\end{align}
where,
\begin{equation}
\label{delta2alpha}
\begin{aligned}
\bs{\alpha}(0,d) &= 1 \\
\bs{\alpha}(1,d) &= \bs{\delta}(1,d) \\
\bs{\alpha}(2,d) &= \bs{\delta}(2,d) + \bs{\delta}(1,d)^{\otimes 2} \\
\bs{\alpha}(3,d) &= \bs{\delta}(3,d) + 3\bs{\delta}(2,d)\otimes \bs{\delta}(1,d) + \bs{\delta}(1,d)^{\otimes 3} \\
\bs{\alpha}(4,d) &= \bs{\delta}(4,d) + 4\bs{\delta}(3,d)\otimes \bs{\delta}(1,d) + 3\bs{\delta}(2,d)^{\otimes 2} + 6 \bs{\delta}(2,d) \\
& \otimes \bs{\delta}(1,d)^{\otimes 2} + \bs{\delta}(1,d)^{\otimes 4} \\
\bs{\alpha}(5,d) &= \bs{\delta}(5,d) + 5\bs{\delta}(4,d)\otimes \bs{\delta}(1,d)+ 10\bs{\delta}(3,d)\otimes \bs{\delta}(2,d) \\
& \quad + 10\bs{\delta}(3,d) \otimes \bs{\delta}(1,d)^{\otimes 2}
+ 15\bs{\delta}(2,d)^{\otimes 2}\otimes \bs{\delta}(1,d) \\
& + 10\bs{\delta}(2,d)\otimes \bs{\delta}(1,d)^{\otimes 3}+ \bs{\delta}(1,d)^{\otimes 5} \\
\bs{\alpha}(6,d) &= \bs{\delta}(6,d) + 6\bs{\delta}(5,d)\otimes \bs{\delta}(1,d) + 15 \bs{\delta}(4,d) \otimes \bs{\delta}(2,d) \\
& \quad + 15 \bs{\delta}(4,d) \otimes \bs{\delta}(1,d)^{\otimes 2} + 10 \bs{\delta}(3,d)^{\otimes 2} + 60\bs{\delta}(3,d)\otimes \bs{\delta}(2,d) \\
& \quad \otimes \bs{\delta}(1,d) + 20\bs{\delta}(3,d) \otimes \bs{\delta}(1,d)^{\otimes 2} + 15\bs{\delta}(2,d)^{\otimes 3} + 45\bs{\delta}(2,d)^{\otimes 2} \\
& \quad \otimes \bs{\delta}(1,d)^{\otimes 2} + 15\bs{\delta}(2,d)\otimes \bs{\delta}(1,d)^{\otimes 4} + \bs{\delta}(1,d)^{\otimes 6}
\end{aligned}
\end{equation}
The above set of equations \eqref{delta2alpha} has exact resemblance with that expressing moments in terms of the cumulants in Section \ref{RelCumMom}.
This must happen, as Equation \eqref{GGC} for GGC expansion with $\delta(\mathbf{x})$ as a reference \textit{pdf}
is matching Equation \eqref{fxdelta}.
This matching proves that $\bs{\alpha}(k,d)$ is related in same way to $\bs{\delta}(k,d)$, as $\mathbf{m}(k,d)$ to $\mathbf{c}(k,d)$. That is,:
\begin{align}
\label{alphadeltaeq}
\sum_{k=0}^{\infty} \bs{\alpha}(k,d)'\frac{\bs{\lambda}^{\otimes k}}{k!} = \exp \left( \sum_{k=1}^{\infty}\bs{\delta}(k,d)'\frac{\bs{\lambda}^{\otimes k}}{k!}\right)
\end{align}
Further, the $\bs{\alpha}(k,d)$ in Equation \eqref{GGC} recursively can be obtained in terms of the cumulant difference vector $\bs{\delta}(k,d)$ as under:
\begin{align}
\label{ndalphadelta}
\bs{\alpha}(k,d) = \sum_{p=0}^{k-1} \binom{k-1}{p} \mathbf{K}^{-1}_{\mathfrak{p} h \leftrightarrow l}\bs{\delta}(k-p,d) \otimes \bs{\alpha}(p,d)
\end{align}
where, $\mathbf{K}^{-1}_{\mathfrak{p} h \leftrightarrow l}$ is a specific commutation matrix; as described previously; to
change the order of the cumulants for Kronecker product such that the expression has decreasing order cumulants from left to the right.
\hspace{0.2 in} The verification of the derived GGC can be obtained by taking Gaussian density as a reference \textit{pdf}. With Gaussian density as a reference, $\bs{\delta}(1,d)= \mathbf{0}$ and $\bs{\delta}(2,d) = \mathbf{0}$. So, the coefficients $\bs{\alpha}(k,d)$ in Equation \eqref{GGC} can be derived as under:
\begin{equation}
\begin{aligned}
\bs{\alpha}(0,d) &= 1 \\
\bs{\alpha}(1,d) &= \mathbf{0} \\
\bs{\alpha}(2,d) &= \mathbf{0} \\
\bs{\alpha}(3,d) &= \bs{\delta}(3,d) = \mathbf{c}(3,d) \\
\bs{\alpha}(4,d) &= \bs{\delta}(4,d) = \mathbf{c}(4,d) \\
\bs{\alpha}(5,d) &= \bs{\delta}(5,d) = \mathbf{c}(5,d) \\
\bs{\alpha}(6,d) &= \bs{\delta}(6,d) + 10 \bs{\delta}(3,d)^{\otimes 2} = \mathbf{c}(6,d) + 10 \mathbf{c}(3,d)^{\otimes 2}
\end{aligned}
\end{equation}
Thus, the GGC series is derived and verified using known examples.
\section{Characteristic function of an unknown random vector in terms of a reference characteristic function}
\label{chGGC}
The GGC derived as in Equation \eqref{GGC} can be used to give the characteristic function of an unknown \textit{pdf}, in terms of the characteristic function of a reference \textit{pdf}. For that taking Fourier transform $(\mathsf{F})$ of Equation \eqref{GGC}, we get:
\begin{align}
\mathcal{F}_{\mathbf{x}}(\bs{\lambda}) &= \sum_{k=0}^{\infty} (-1)^k\frac{\bs{\alpha}(k,d)'}{k!}\mathsf{F}( \psi^{(k)}(\mathbf{x})) \\
\mathcal{F}_{\mathbf{x}}(\bs{\lambda}) & = \left[ \sum_{k=0}^{\infty}\bs{\alpha}(k,d)' \frac{(i\bs{\lambda})^{\otimes k}}{k!}\right] \mathsf{F}(\psi(\mathbf{x})) \quad \mbox{($\because$ differentiation property of $\mathsf{F}$ )}\\
\label{GGCch}
& = \exp\left[ \sum_{k=1}^{\infty}\bs{\delta}(k,d)' \frac{(i\bs{\lambda})^{\otimes k}}{k!}\right]\mathcal{F}_r(\mathbf{x}) \quad (\because \mbox{ Equation }\eqref{alphadeltaeq})
\end{align}
where, $\mathcal{F}_r$ is the characteristic function of the reference \textit{pdf}.
\section[Compact derivation for the GGC expansion]{Compact derivation for the Generalized Gram-Charlier expansion}
\label{compactGGC}
The compact derivation of Equation \eqref{GGC} follows as under:
\begin{align}
\mathcal{F}_{\mathbf{x}}(\bs{\lambda}) & = \exp\left[ \sum_{k=1}^{\infty}\mathbf{c}(k,d)' \frac{(i\bs{\lambda})^{\otimes k}}{k!}\right] \quad \mbox{( $\because$ definition in Equation \eqref{ndCGF2} )}\\
& = \exp\left[ \sum_{k=1}^{\infty}\bs{\delta}(k,d)' \frac{(i\bs{\lambda})^{\otimes k}}{k!}\right]\exp\left[ \sum_{k=1}^{\infty}\mathbf{c}_r(k,d) \frac{(i\bs{\lambda})^{\otimes k}}{k!}\right] \\
& \quad (\because \bs{\delta}(k,d) = \mathbf{c}(k,d) - \mathbf{c}_r(k,d) ) \nonumber \\
& = \left[ \sum_{k=0}^{\infty}\bs{\alpha}(k,d)' \frac{(i\bs{\lambda})^{\otimes k}}{k!}\right]\mathsf{F}(\psi(\mathbf{x}))&
\end{align}
Taking inverse Fourier transform of the above equation brings
\begin{align}
f(\mathbf{x}) &= \sum_{k=0}^{\infty}\bs{\alpha}(k,d)' \frac{(-1)^k}{k!}\bs{\delta}^{(k)}(\mathbf{x}) * \psi(\mathbf{x}) \\
\mbox{ or }f(\mathbf{x}) &= \sum_{k=0}^{\infty}\bs{\alpha}(k,d)' \frac{(-1)^k}{k!}\psi^{(k)}(\mathbf{x})
\end{align}
where, $*$ indicates convolution. Thus, the Equation \eqref{GGC} is obtained in a more compact way.
\section{Conclusion}
\label{conclusion}
The article has derived multivariate Generalized Gram-Charlier (GGC) expansion in Equation \eqref{GGC}; combined with Equation \eqref{ndalphadelta}; that expresses an unknown multivariate \textit{pdf} in terms of vector cumulants and vector derivatives of a reference \textit{pdf}.
The multivariate Gram-Charlier A series is derived in Equation \eqref{ndGCAshort} and Equation \eqref{ndGCAhershort} representing an unknown multivariate \textit{pdf} in terms of its vector cumulants and vector Hermite polynomials. There has been also derived compact formulas for obtaining multivariate vector moments from vector cumulants in Equation \eqref{ndmomcum} and vise-a-verse in Equation \eqref{ndcummom}; the integral form of multivariate \textit{pdf} representation in Equation \eqref{ndpdfreal} and the integral form of multivariate vector Hermite polynomials in Equation \eqref{hermitenew}, as well, in Equation \eqref{hermitend01}. The expressions are derived using only
elementary calculus of several variables in vector notations through Kronecker product based derivative operator. Thus, they are more transparent and more comprehensive compare to their corresponding multi-linear matrix representations or tensor representations.
|
\section{Introduction}
The study of twists of curves can be a very useful tool for understanding some arithmetic problems. For example, it has been proved to be really helpful for exploring the Sato-Tate conjecture \cite{FKRS}, \cite{FLS}, \cite{FS2}, \cite{FS3},\cite{thesis}. As well as for solving some Diophantine equations \cite{PSS} or computing $\mathbb{Q}-$curves realizing certain Galois representations \cite{BFGL}, \cite{diff}.
The twists of curves of genus $\leq 2$ are well-known.
While the genus $0$ and $1$ cases date back to a long time ago \cite{Sil}, the genus $2$
case is due to the work of Cardona and Quer over number fields \cite{Cart}, \cite{Carp}, and to Cardona over finite fields \cite{CarFF}. All the
genus $0$, $1$ or $2$ curves are hyperelliptic (at least in the sense that they are not non-hyperelliptic, since genus $0$ and $1$ curves are not usually called hyperelliptic). However, for genus greater than $2$
almost all the curves are non-hyperelliptic. Only few twists of genus $3$ curves over number fields have been previously computed \cite{diff}, \cite{PSS}. Over finite fields, more twists of genus $3$ has been computed \cite{MT10}, but, in this case, equations are not given.
We devote the present paper to show a method for computing twists of smooth curves of genus greater than $0$, and in
the particular case of non-hyperelliptic curves we show how to compute equations for the twists. The method is not completely original since it is based on well-known results, but as far as we know this is the first time that all the strategies used for computing twists are joined together and all the gaps are filled in order to produce a systematic method. In particular, when the field of definition of the curve has characteristic different from zero, the method gives rise to an algorithm.
In a forthcoming paper \cite{Lor14}, this method will be useful for computing the twists of all non-hyperelliptic genus $3$ curves defined over any number field.
\subsection{Outline} The structure of this paper is as follows. Section \ref{11} establishes a correspondence between the set of twists of any smooth and irreducible genus $g>0$ curve $C$ defined over a perfect field $k$ and the set of solutions to a Galois embedding problem, see Theorem \ref{main}. In Section \ref{12}, we show how to compute equations of the twists in the particular case in which the curve $C$ is non-hyperelliptic. We do this by studying the action of the Galois group of a certain extension of the field of definition of the curve $C$, in the vector space of regular differentials $\Omega^1(C)$. Section \ref{13} describes in detail the method obtained for computing the twists of non-hyperelliptic curves. First step is computing a canonical model of the curve. The second one is posing the corresponding Galois embedding problem, whose solutions are in bijection with the set of twists, and solving it. In general, if $k$ is a number field, there is not known method for solving a Galois embedding problem over $k$, and this step should be treated case-by-case. We compute the solutions to an infinity family of such problems in Proposition \ref{q}. Nevertheless, if $k$ is a finite field, any Galois embedding problem over $k$ is known how to be solved (e.g. \cite[Chapter 1]{Ser}). The third and last step is computing equations for the twists. Finally, in Section \ref{14} we illustrate the method by computing all the twists of the non-hyperelliptic genus $6$ curve $x^7-y^3z^4-z^7=0$ when it is considered to be defined over a number field such that $[k(\zeta_{21}):k]=12$.
\subsection{Notation}
We now fix some notation and conventions that will be valid through the paper. For any field $F$, we denote by $\bar{F}$ an algebraic closure of $F$, and by $G_F$ the absolute Galois group $\operatorname{Gal}(\bar{F}/F)$. We recurrently consider the action of $G_F$ on several sets, and this action is in general denoted by left exponentiation. For a field $F$, let $\operatorname{GL}_n(F)$ (resp. $\operatorname{PGL}_n(F)$) be the ring of $n$ by $n$ invertible matrices with coefficients in F (resp. that are projective).
By $k$ we always mean a perfect field. All field extensions of $k$ that we consider are contained in a fixed algebraic closure $\bar{k}$. We write $\zeta_n$ to refer to a primitive $n-$th root of unity in $\bar{k}$. When $k$ is a number field, we denote by $\mathcal{O}_k$ the ring of integers of $k$.
Given a projective, smooth and geometrically irreducible curve $C/k$ we denote by $\operatorname{Aut}(C)$ the grup of automorphisms of $C$ defined over $\bar{k}$. By $K$ we denote the minimal extension $K/k$ where all the automorphism of $C$ can be defined. The $k-$vector space of regular differentials of $C$ is denoted by $\Omega^1(C)$.
When we work with groups, we usually use the SmallGroup Library-GAP \cite{GAP}. Where the group $<N,r>$ denotes the group of order $N$ that appears in the $r-$th position in such library. By $\text{ID}(G)$, we mean the corresponding GAP notation for the group $G$.
\subsection{Aknowledgments} The author would like to thank to Joan-Carales Lario for bringing this problem to her attention and to Francesc Fit\'e for a careful reading of the manuscript and useful comments and suggestions.
\section{Galois embedding problems}
\label{11}
Let $k$ be a perfect field and $C/k$ be a projective curve of genus $g>0$. Let us denote by $K$ the minimal field where all the automorphisms of $C$ can be defined. Let us define the twisting group $\Gamma := \operatorname{Aut}(C)\rtimes \operatorname{Gal}(K/k)$, where $\operatorname{Gal}(K/k)$ acts naturally on $\operatorname{Aut}(C)$, and the multiplication rule is $(\alpha,\sigma)(\beta,\tau)=(\alpha\,^{\sigma}\beta,\sigma\tau)$ \cite{FL}.
Let us define the following sets:
\begin{equation}\label{e3}
\operatorname{Twist}_k(C):=\left\{ C'/k\operatorname{curve}\mid\exists\:\overline{k}\operatorname{-isomorphism}\:\phi\colon C'\to C\right\} /k\operatorname{-isomorphism},
\end{equation}
\begin{equation}\label{e1}
\operatorname{H}^{1}(G_k,\operatorname{Aut}(C)):=\left\{ \xi\colon G_k \to \operatorname{Aut}(C)\,\,\text{continuous} \mid \xi_{\sigma\tau}=\xi_{\sigma}{}^{\sigma}\xi_{\tau} \right\}/\sim ,
\end{equation}
where the topology in $G_k$ is the profinite one, and we consider the discrete topology in $\operatorname{Aut}(C)$. Two cocycles are cohomologous $\xi\sim\xi '$, if and only if, there is $\varphi\in\operatorname{Aut}(C)$ such that $\xi_{\sigma}'=\varphi\cdot\xi_{\sigma}\cdot{}^{\sigma}\varphi^{-1}$. We also define
\begin{equation}\label{e2}
\widetilde{\operatorname{Hom}}(G_k,\Gamma):=\left\{ \Psi\colon G_k\to\Gamma\mid\Psi\:\operatorname{epi}_{2}-\operatorname{morphism} \right\}/\sim ,
\end{equation}
the meaning of $\operatorname{epi}_{2}-\operatorname{morphism}$ is that $\Psi$ is a continuous group homomorphism (again with the profinite and discrete topologies respectively) such that the composition $\pi_2 \cdot \Psi :\, G_k\to\Gamma\to\operatorname{Gal}(K/k)$ is surjective where $\pi_2:\,\Gamma\to\operatorname{Gal}(K/k)$ (resp. $\pi_1$) is the natural projection on the second (resp. first) component of the elements of $\Gamma$. We say that $\Psi\sim\Psi'$ are equivalent if there is $(\varphi,1)\in\Gamma$ such that $\Psi_{\sigma}'=(\varphi,1)\Psi_{\sigma}(\varphi,1)^{-1}$.
\begin{definition} With notation above, we say that $L$ is the splitting field of the twist $\phi:\,C'\rightarrow C$, if $L$ is the minimal field where, for all $\alpha\in\text{Aut}(C)$, the isomorphisms $\alpha\circ \phi$ are defined. Similarly, we define the splitting field of a cocycle $\xi$ as the field $L$ that satisfies the condition $$\text{Gal}(\bar{k}/L)=\bigcup_{\xi\sim\xi'}\text{Ker}(\xi').$$ Since $\xi$ is continuous, $L$ is well-defined.
For an element $\Psi\in\widetilde{\operatorname{Hom}}(G_k,\Gamma)$, we define its splitting field as the field $L$ such that $\text{Gal}(\bar{k}/L)=\text{Ker}(\Psi)$.
\end{definition}
Notice that the previous splitting fields are all of them finite extensions of $k$ since we are considering curves of genus greater than $0$, and then the group $\text{Aut}(C)$ is finite.
\begin{theorem}\label{main} There are natural one-to-one correspondences between the following three sets:
$$
\operatorname{Twist}_k(C)\longrightarrow
\operatorname{H}^{1}(G_k,\operatorname{Aut}(C))\longrightarrow
\widetilde{\operatorname{Hom}}(G_k,\Gamma)$$
These correspondences send $\phi$ to $\xi_{\sigma}=\phi\cdot{}^{\sigma}\phi^{-1}$, and $\xi$ to $\Psi_{\sigma}=(\xi_{\sigma},\overline{\sigma})$, where $\overline{\sigma}$ denotes the projection of $ \sigma\in G_k$ onto $\operatorname{Gal}(K/k)$. Moreover, the splitting fields of elements in these three sets are preserved by these correspondences.
\end{theorem}
\begin{proof} The correspondence between the set (\ref{e3}) of twists $\operatorname{Twist}_k(C)$ and the first Galois cohomology set (\ref{e1}) is well-known and can be found in \cite[$X.2$ Theorem $2.2$]{Sil}. The statement about the splitting fields follows by definition. So, it only remains to prove that the map between the sets (\ref{e1}) and (\ref{e2}) is a correspondence. Let us first prove that they are well-defined. Clearly, given $\xi\in\operatorname{Twist}_k(C)$, we have that $\Psi$ defined by $\Psi_{\sigma}:=(\xi_{\sigma},\overline{\sigma})$ defines an element in $\widetilde{\operatorname{Hom}}(G_k,\Gamma)$. Conversely, given an element $\Psi\in\widetilde{\operatorname{Hom}}(G_k,\Gamma)$, we have that $\xi:=\pi_1(\Psi)$ defines an element in $\operatorname{Twist}_k(C)$. Finally, it is a straightforward computation to check that this two maps are one inverse to the other and that they preserve the equivalence relations defined in both sets.
\end{proof}
\begin{remark} Notice that any element $\Psi\in\widetilde{\operatorname{Hom}}(G_k,\Gamma)$ can be reinterpreted as a solution to the following Galois embedding problem:
\begin{align}
\xymatrix{
& & & G_k \ar@{-->>}_-{\Psi}[dl] \ar@{->>}[d] & \\
1 \ar[r] & \operatorname{Aut}(C)\,\ar@{^{(}->}[r] & \Gamma \ar[r]_-{\pi} & \operatorname{Gal}(K/k) \ar[r] & 1
} \label{gep}
\end{align}
Reciprocally, every solution $\Psi$ of the above embedding problem is an element in $\widetilde{\operatorname{Hom}}(G_k,\Gamma)$ and gives rise to a twist of $C$.
In order to keep track of the equivalence classes of twists we must here consider two solutions $\Psi$ and $\Psi'$ equivalent only under the restricted conjugations allowed in the definition of the set $\widetilde{\operatorname{Hom}}(G_k,\Gamma)$, that is slightly different from the standard one \cite[Section $9.4$]{probe}.
\end{remark}
\section{Equations of the twists}
\label{12}
First of all, remark that a twist is not a curve, it is an equivalence class of curves, so when we say that we compute equation for a twist, what we mean is that we compute equations for some particular curve in the equivalence class. Secondly, notice that a curve can have different models, and a particular model for a non-hyperelliptic curve is its canonical model. The method that we present in this section, it is a method for computing the canonical model of a curve in the equivalence class of a twist defined by a cocycle.
This method is a generalization of the one used by Fern\'andez, Gonz\'alez and Lario \cite{diff}. They used it for computing equations of twists of some particular non-hyperelliptic genus $3$ curves, special case for which the canonical model coincides with the plane model.
Notice that, in our context, finding equations for a twist that is given by a cocycle $\xi\in\operatorname{H}^{1}(G_k,\operatorname{Aut}(C))$, is actually equivalent to computing an inverse map for the correspondence in Theorem \ref{main}
$$
\operatorname{Twist}_k(C)\longrightarrow
\operatorname{H}^{1}(G_k,\operatorname{Aut}(C)).
$$
Let $\Omega^1(C)$ be the $k-$vector space of regular differentials of $C$. Let $\omega_1,...,\omega_g$ be a basis of $\Omega^1(C)$, where $g$ is the genus of $C$ (the existence of such a basis can be deduced from the fact that there always exists a canonical divisor defined over the definition field $k$ of the curves, which is a consequence of \cite[II, Lemma 5.8.1]{Sil}). Given a cocycle $\xi:\, \text{G}_k\to \text{Aut}(C)$ and its splitting field $L$, we consider the extension of scalars $\Omega^1_L(C)=\Omega^1(C)\otimes_{k}L$ which is a $k-$vector space of dimension $g[L:k]$. We can then see the elements of $\Omega_{L}^{1}(C)$ as sums $\sum \lambda_{i}\omega_{i}$ where $\lambda_{i}\in L$. For every $\sigma\in\operatorname{Gal}(L/k)$, we consider the twisted action on $\Omega_L^1(C)$ defined as follows:
$$
(\sum\lambda_{i}\omega_{i})^{\sigma}_{\xi}:=\sum{}^{\sigma}\lambda_{i}\xi_{\sigma}^{*-1}(\omega_{i}).
$$
Here, $\xi_{\sigma}^{*}\in\operatorname{End}_K(\Omega^1(C))$ denotes the pull-back of $\xi_{\sigma}=\phi\cdot^{\sigma}\phi^{-1}\in\operatorname{Aut}_K(C)$.
One readily checks that
$$
\rho_{\xi}\colon\operatorname{Gal}(L/k)\to\operatorname{GL}(\Omega^1_L(C)),\:\:\:\rho_{\xi}(\sigma)(\omega):=\omega_{\xi}^{\sigma}
$$
is a $k-$linear representation. Indeed, since $\xi_{\sigma\tau}^{*}=\,^{\sigma}\xi_{\tau}^{*}\cdot\xi^{*}_{\sigma}$, we have
\begin{align}
\rho_{\xi}(\sigma\tau)(\sum\lambda_{i}\omega_{i})&=\sum{}^{\sigma\tau}\lambda_{i}\xi_{\sigma\tau}^{*-1}(\omega_{i}) \notag\\
&=\sum{}^{\sigma\tau}\lambda_{i}\xi_{\sigma}^{*-1}\cdot^{\sigma}\xi_{\tau}^{*-1}(\omega_{i})\notag \\
&= \rho_{\xi}(\sigma)(\sum{}^{\tau}\lambda_{i}\xi_{\tau}^{*-1}(\omega_{i}))\notag \\
&= \rho_{\xi}(\sigma)\rho_{\xi}(\tau)(\sum\lambda_{i}\omega_{i}).\notag
\end{align}
\begin{lemma}\label{iso} Let $\phi:\,C\to C'$ be a twist such that $\phi\circ^{\sigma}\phi^{-1}=\xi_{\sigma}$. Then, the following $k-$vector spaces are isomorphic as $G_k-$modules:
$$
\Omega_{L}^{1}(C)_{\xi}^{\operatorname{Gal}(L/k)}\simeq \Omega^{1}(C').
$$
\end{lemma}
\begin{proof} We claim that the map $\Omega_{L}^{1}(C)_{\xi}^{\operatorname{Gal}(L/k)}\rightarrow \Omega^{1}(C'):\,\omega\rightarrow \phi^{*}(\omega)$ is an isomorphism of $G_k-$modules. The only non-trivial fact is the surjectivity. But this is a consequence of the equivalent result for function fields.
Recall that the function field $k(C')$ may be reinterpreted as the fixed field $\overline{k}(C)_{\xi}^{G_{k}}$ where the action of the Galois group $G_{k}$ on $\overline{k}(C)$ is twisted by $\xi$ according to $f_{\xi}^{\sigma}:=f\cdot\xi_{\sigma}$ \cite[X.$2$]{Sil}.
\end{proof}
We identify the previous vector spaces via an isomorphism as in Lemma \ref{iso}, so, for explicit computations, we can use
\begin{align}\label{CalcBase}
\Omega^{1}(C')=\bigcap_{\sigma\in\operatorname{Gal}(L/k)}\operatorname{Ker}(\rho_{\xi}(\sigma)-\operatorname{Id}).\end{align}
Consider the canonical morphism and the canonical model $\phi_K:\,C\rightarrow \mathcal{C}\subset \mathbb{P}^{g-1}$ given by the basis $\{\omega_1,...,\omega_g\}$ of $\Omega^1(C)$. Let
$$
\mathcal{C}:\:\left\{F_{h}(\omega_1,...,\omega_g)=0\right\}_h
$$
be a set of equations defining the canonical model in $\mathbb{P}^{g-1}$.
Let $\{\sum_{i=1}^{g}\mu_{j}^{i}\omega_{i}\}_j$ be a basis of $\Omega_{L}^{1}(C)_{\xi}^{\operatorname{Gal}(L/k)}$. We can then take a basis $\omega_{j}' =\sum_{i=1}^{g}\mu_{j}^{i}\omega_{i}$ of $\Omega^{1}(C')$ via an isomorphism as in Lemma \ref{iso}. Thus, we can write
$$
\omega_{i}=\sum_{j=1}^{g}\eta_{j}^{i}\omega_{j}'
$$
for some $\eta_{j}^{i}\in L$. We then obtain equations for the canonical model $\mathcal{C}'$, given by the basis $\{\omega_{j}' \}$, of the twist $C'$ via the substitution
$$
\mathcal{C}':\:\left\{F_{h}(\sum_{j=1}^{g}\eta_{j}^{1}\omega_{j}',...,\sum_{j=1}^{g}\eta_{j}^{g}\omega_{j}')=F'_{h}(\omega'_1,...\omega'_g)=0\right\}_h.
$$
Notice that the projective matrix $\eta=(\eta^{i}_{j})_{ij}$ defines an isomorphism of canonical models $\eta:\,\mathcal{C}'\rightarrow \mathcal{C}$, and that $\eta\cdot^{\sigma}\eta^{-1}=(\xi_{\sigma}^{*})^{-1}$. In general, on a canonical models level, any morphism of curves is given by a matrix, since a morphism of curves induces a linear morphism on the regular differential vector spaces.
\begin{remark} Notice that, for non-hyperelliptic curves, Lemma \ref{iso} is equivalent to prove that the dimension of $\Omega_{L}^{1}(C)_{\xi}^{\operatorname{Gal}(L/k)}$ is equal to $n$, that is, to find a matrix in $\eta\in\text{GL}_{g}(L)$ such that $\eta\cdot^{\sigma}\eta^{-1}=(\xi_{\sigma}^{*})^{-1}$. But this is a consequence of Hilbert $90$th Problem since $(\xi^{*})^{-1}\in\text{H}^1(\text{Gal}(L/k),\text{GL}(\Omega^{1}_{L}(C)))$.
\end{remark}
\section{Description of the method}
\label{13}
Let $C$ be a smooth non-hyperelliptic genus $g$ curve defined
over a perfect field $k$. Assume that its automorphism group $\operatorname{Aut}(C)$ is known and let us denote by $K$ the minimal field where $\operatorname{Aut}(C)$ is defined. We now proceed to describe a method for computing the set of twists of the curve $C$. In each step, we will compute different things:
\subsection*{Step 1: a canonical model} Firstly, we take a basis of $\Omega^{1}(C)$, and via this basis we
obtain a canonical model $\mathcal{C}/k$ as the image of the canonical morphism $C\hookrightarrow\mathbb{P}^{g-1}$. Again, the existence of a canonical divisor defined over $k$ implies that we can take the canonical model $\mathcal{C}$ also defined over $k$. Hence, $\mathcal{C}$ and
$C$ belong to the same class in $\mathrm{Twist}_k\left(\mathcal{C}\right)$
and $\mathrm{Twist}_k\left(\mathcal{C}\right)=\mathrm{Twist}_k\left(C\right)$.
In addition, the automorphisms group $\mathrm{Aut}\left(\mathcal{C}\right)$
can be viewed in a natural way as a subgroup of $\operatorname{PGL}_{g}\left(K\right)$ (via the induced automorphism in $\mathbb{P}^{g-1}$ by the canonical morphism).
Indeed, we can see it as a subgroup of $\operatorname{GL}_{g}\left(K\right)$ if we look at its action on $\Omega^1(C)\otimes_{k}K$ as a $K-$vector space. Furthermore, any isomorphism $\phi:\,\mathcal{C}'\rightarrow \mathcal{C}$ can be also viewed as a matrix in $\operatorname{PGL}_{g}\left(\bar{k}\right)$.
\subsection*{Step 2: the set $\operatorname{Twist}_k(C)$} We will first compute the set $\widetilde{\mathrm{Hom}}\left(G_{k},\Gamma\right)$.
From this set, we will compute $\mathrm{H}^{1}\left(G_{k},\mathrm{Aut}\left(\mathcal{C}\right)\right)$ via the correspondence in Theorem \ref{main}.
Given an element $\Psi\in\widetilde{\mathrm{Hom}}\left(G_{k},\Gamma\right)$,
let $L$ be its splitting field. We have the following isomorphisms: $\Psi(G_K)\simeq\operatorname{Gal}(L/K)$ and $\Psi(G_k)\simeq\operatorname{Gal}(L/k)$. Hence, we can see $\Psi$
as a proper solution to the Galois embedding problem
\begin{align*}
\xymatrix{
& & & G_k \ar@{-->>}_-{\Psi}[dl] \ar@{->>}[d] & \\
1 \ar[r] &\Psi(G_K) \ar[r] & \Psi(G_k) \ar[r] & \operatorname{Gal}(K/k) \ar[r] & 1
}
\end{align*}
As it was noticed in Section \ref{11}, we have $\mathrm{Gal}\left(L/k\right)\simeq\mathrm{Image}\left(\Psi\right)\subseteq\Gamma$
and $\mathrm{Gal}\left(L/K\right)\simeq\Psi\left(G_{K}\right)\subseteq\mathrm{Aut}\left(\mathcal{C}\right)\rtimes\left\{ 1\right\} $.
Hence, we can break the computation of $\widetilde{\mathrm{Hom}}\left(G_{k},\Gamma\right)$, i.e., the solutions (proper or not) to the Galois embedding problem (\ref{gep}), into the computation of the proper solutions to some Galois embedding problems attached to a pair $(G,H)$ as follows
\begin{align}
\xymatrix{
& & & G_k \ar@{-->>}_-{\Psi}[dl] \ar@{->>}[d] & \\
1 \ar[r] &H \ar[r] & G \ar[r] & \operatorname{Gal}(K/k) \ar[r] & 1
} \label{pGH}
\end{align}
where we consider all the pairs $(G,H)$ such that $G\subseteq\Gamma$,
$H=G\cap\mathrm{Aut}\left(C\right)\rtimes\left\{ 1\right\} $ and
$\left[G:H\right]=\left|\mathrm{Gal}\left(K/k\right)\right|$ (up to
conjugacy by elements $\left(\varphi,1\right)\in\Gamma$).
Every proper solution to a Galois embedding problem (\ref{pGH}) can be lifted to a solution to
the Galois embedding problem (\ref{gep}).
Notice that the same field
$L$ can appear as the splitting field of more than one solution $\Psi$ corresponding to a pair $(G,H)$. This is because given an automorphism $\alpha$ of $\operatorname{Gal}(L/k)$ that leaves $\operatorname{Gal}(K/k)$ fixed, $\alpha\Psi$ is other solution that has $L$ as splitting field. Two such solutions are equivalent if and only if there exists $\beta\in \operatorname{Aut}(\mathcal{C})$ such that $\alpha\Psi=\beta\Psi\beta^{-1}$. So, the number of non-equivalent solutions with splitting field $L$ and $\Psi(\operatorname{G}_k)=G$ is the cardinality $n_{(G,H)}$ of the group \cite{Carp}:
\begin{align}
\mathrm{Aut}_{2}\left(G\right)/\operatorname{Inn}_{G}\left(\mathrm{Aut}\left(\mathcal{C}\right)\rtimes\left\{ 1\right\} \right),\label{numero}
\end{align}
where $\mathrm{Aut}_{2}\left(G\right)$ is the group of automorphisms
of $G$ such that leave the second coordinate invariant and $\operatorname{Inn}\left(\mathrm{Aut}\left(\mathcal{C}\right)\rtimes\left\{ 1\right\} \right)$
is the group of inner automorphisms of $\mathrm{Aut}\left(\mathcal{C}\right)\rtimes\left\{ 1\right\} $
lifted in the natural way to $\mathrm{Aut}\left(G\right)$.
We can then divide this step in two:
\subsubsection*{Step 2a: computing the pairs $(G,H)$} The pairs $(G,H)$ and the number $n_{(G,H)}$ defined above, can be, for example, computed with magma \cite{magma} (c.f. \cite[Appendix]{thesis} for an implemented code).
\subsubsection*{Step 2b: computing the proper solutions to the Galois embedding problems (\ref{pGH})} The solutions should be computed case-by-case for each pair $(G,H)$. If $k$ is a finite field this is know how to be done (e.g. \cite[Theorem $1.1$]{Ser}), and the method described in this paper becomes then an algorithm. Unfortunately, if $k$ is a number field there is not known systematical method for solving these problems
Next proposition, that is a generalization of \cite[Lemma $9.6$]{cox} for $q=3$, will be useful for solving some of these Galois embedding problems.
\begin{proposition}\label{q} Let be $q=p^r$, where $p$ is a prime number, let $k$ be a number field, and let $\zeta$ be a fixed $q$-th primitive root of the unity in $\bar{k}$. We denote $K=k(\zeta)$ and we assume $\left[ k(\zeta ):k\right] =p^{r-1}(p-1)$. Let us define $G_q:=\mathbb{Z}/q\mathbb{Z}\rtimes (\mathbb{Z}/q\mathbb{Z})^*$ where the action of $(\mathbb{Z}/q\mathbb{Z})^*$ on $\mathbb{Z}/q\mathbb{Z}$ is given by the multiplication rule $(a,b)(a',b')=(a+ba',bb')$. Let us consider the Galois embedding problem:
\[
\xymatrix{
& & & G_k \ar@{->>}[d]_-{\pi} & \\
1 \ar[r] &\mathbb{Z}/q\mathbb{Z} \ar[r] & G_q \ar[r] & (\mathbb{Z}/q\mathbb{Z})^* \ar[r] & 1
},
\]
where the horizontal morphisms are the natural ones, and the projection $\pi$ is given by $\pi(\sigma)=(0,b)$ if $\sigma(\zeta)=\zeta^{b}$. Then, the splitting fields of the proper solutions to this Galois embedding problem are of the form $L=K(\sqrt[q]{m})$ where $m\in \mathcal{O}_k$ is an integer that is not a $p$-power. Moreover, every such field is the splitting field for a solution $\Psi$ to the above Galois embedding problem.
\end{proposition}
\begin{proof} Notice first that there exist proper solutions $\Psi$ to the Galois embedding problem. Given a field $L=K(\sqrt[q]{m})$ with $m\in k$ and not a $p$-power, there is a natural isomorphism $\operatorname{Gal}(L/k)\simeq G_q$ compatible with the projection $G_q \rightarrow (\mathbb{Z}/q\mathbb{Z})^*$. The natural projection $G_k\twoheadrightarrow \operatorname{Gal}(L/k)$ then provides a solution to the Galois embedding problem above.
Now, let $\Psi$ be any proper solution to the problem, and let us denote by $L$ its splitting field. Let $G$ be the subgroup of $G_q$ that contains all the elements of the form $(0,b)$, and let $\sigma\in G_k$ be such that $\Psi(\sigma)=(1,1)$.
Let $\alpha$ be a primitive element of the extension $L^G/k$ that moreover is an algebraic integer. We then have that $L=K(\alpha)$. This is because $\left[K:k\right]=p^{r-1}(p-1)$, $\left[L^G:k\right]=q$ and $L^G\cap K=k$. Let us now define for $i=0,1,...,q-1$ the numbers:
$$
u_i=\alpha+\zeta^i\sigma^{-1}(\alpha)+\zeta^{2i}\sigma^{-2}(\alpha)+...+\zeta^{(q-1)i}\sigma^{-(q-1)}(\alpha).
$$
Then $\sigma(u_i)=\zeta^{i}u_i$ and for any $\tau\in G_k$ such that $\Psi(\tau)=(0,b)$ we have $\Psi(\tau\sigma^{j})=(0,b)(j,1)=(bj,1)(0,b)=\Psi(\sigma^{bj}\tau)$, so $\tau(u_i)=u_i$ . In particular, we have that $u_0,\,u_1^q,...,\,u_{q-1}^q\in\mathcal{O}_k$. Hence, if $u_j\neq 0$ for some $j>0$, we have that $L=K(u_j)$, since $L^G=k(u_j)$. So, if we put $m=u_j^q\in \mathcal{O}_k$, we get $L=K(\sqrt[q]{m})$. Otherwise, that is, if $u_1=u_2=...=u_{q-1}=0$, then $u_0=u_0+u_1+...+u_{q-1}=q\alpha\in \mathcal{O}_k$, what is a contradiction with $\alpha$ being a primitive element of the extension $L^G/k$.
\end{proof}
For each proper solution $\Psi$ to a Galois embbeding problem (\ref{pGH}) attached to a pair $(G,H)$, we trivially compute the corresponding cocycle $\xi$ via the correspondence between the sets (\ref{e1}) and (\ref{e2}) in Theorem \ref{main}.
\subsection*{Step 3: Equations} We want to compute equations for a twist corresponding to a given cocycle $\xi$.
For this purpose we use the method explained in Section \ref{12}. Computing equations for a twist turns out to be equivalent to computing an isomorphism $\phi:\,C'\rightarrow C$, that is, to explicitly computing the inverse map to the correspondence between sets (\ref{e3}) and (\ref{e1}) in Theorem \ref{main}.
\section{An example}
\label{14}
In order to illustrate the method, we will apply it to the smooth non-hyperelliptic genus $6$ curve which admits the affine plane model
$$
C:\,x^7-y^3-1=0.
$$
As the point at infinity is singular, the projectivization of this plane model is not smooth. However, there is a unique curve, up to $\mathbb{Q}$-isomorphism, which is smooth and birationally equivalent to $C$. So, they have the same function field. We will apply the method for this smooth curve, which is non-hyperelliptic and has genus equal to $6$.
\subsection*{Step 1} First, we must find a canonical model by the usual procedure: finding a basis of holomorphic differentials. Let us call $X=x/z$ and $Y=y/z$. One has
$$
\operatorname{div}(X)=(0:-1:1)+(0:-\zeta_3:1)+(0:-\zeta_3^2:1)-3(0:1:0)=P_1+P_2+P_3-3\infty,
$$
$$
\operatorname{div}(Y)=Q_1+Q_2+Q_3+Q_4+Q_5+Q_6+Q_7-7\infty,
$$
where $Q_i=(\zeta_7^i:0:1)$. Then, $dX$ is an uniformizer for all points except for the $Q_i$'s, because the tangent space to the curve at these points have equation $X-\alpha=0$ for some $\alpha\in\bar{k}$. Then, for the points $Q_i$'s we have to use the expression
$$
dX=-\frac{3y^2}{7x^6}dY
$$
Thus, by \cite[Proposition $4.3$]{Sil}, we finally get
$$
\operatorname{div}(dX)=2(Q_1+Q_2+Q_3+Q_4+Q_5+Q_6+Q_7)-4\infty.
$$
We obtain the following basis of holomorphic differentials:
$$
\omega_1=\frac{dX}{Y^2},\,\,\omega_2=\frac{XdX}{Y^2},\,\,\omega_3=\frac{X^2dX}{Y^2},\,\,\omega_4=\frac{dX}{Y},\,\,\omega_5=\frac{X^3dX}{Y^2},\,\,\omega_6=\frac{XdX}{Y}.
$$
We consider the rational map
$$C\rightarrow\mathbb{P}^5:\,(x:y:z)\rightarrow (z^3:xz^2:x^2z:yz^2:x^3:xyz)$$
The ideal of the image of this map clearly contains the homogeneous polynomials:
$$
\begin{aligned}
f_1&=\omega_1\omega_6-\omega_2\omega_4,&f_2&=\omega_{2}^{2}-\omega_1\omega_3,&
f_3&=\omega_2\omega_3-\omega_1\omega_5,&f_4&=\omega_2\omega_5-\omega_{3}^{2}\\
f_5&=\omega_2\omega_6-\omega_3\omega_4,&f_6&=\omega_3\omega_6-\omega_4\omega_5,&
f_7&=\omega_{4}^{3}-\omega_{3}^{2}\omega_5+\omega_{1}^{3},&
f_8&=\omega_{5}^{3}-\omega_{4}\omega_{6}^{2}-\omega_{1}\omega_{2}^{2}.
\end{aligned}
$$
Now, we claim that the ideal generated by these polynomials gives a smooth curve. To see this, note that, if $\omega_1\neq 0$, the deshomogenization of this ideal with respect to $\omega_1$ gives the affine curve $C$. Now, we isolate from $f_2$ and $f_3$ the variables $\omega_3$ and $\omega_5$ and we plug them into $f_7$. Therefore, $C$ is birationally equivalent to $\mathcal{C}\cap\{ \omega_1\neq 0\}$. Next, if $\omega_1=0$, then the vanishing locus of $f_2,f_4,f_7,f_8$ is the point $(0:0:0:0:0:1)$. To check that $\mathcal{C}$ is non-singular at this point we consider the partial derivatives of the polynomials: $f_1,f_5,f_6,f_8$. Thus, $\mathcal{C}$ is a canonical model of the initial smooth non-hyperelliptic genus $6$ curve.
The automorphism group $\operatorname{Aut}(C)$ is generated by the automorphisms \cite{Swi}:
$$
(x:y:z)\rightarrow (x:\zeta_3y:z)\,\operatorname{and}\, (x:y:z)\rightarrow (\zeta_7x:y:z).
$$
Then, the automorphism group of the canonical model $\mathcal{C}$ is generated by the matrices in $\operatorname{PGL}_{6}\left(\bar{\mathbb{Q}}\right)$:
\small
$$
r=\left(\begin{array}{cccccc} \zeta_{3}&0&0&0&0&0\\ 0&\zeta_3&0&0&0&0\\0&0&\zeta_{3}&0&0&0\\0&0&0&\zeta_{3}^{2}&0&0\\0&0&0&0&\zeta_3&0\\0&0&0&0&0&\zeta_{3}^{2}\end{array}\right),\, s=\left( \begin{array}{cccccc} \zeta_{7}&0&0&0&0&0\\ 0&\zeta_{7}^{2}&0&0&0&0\\0&0&\zeta_{7}^{3}&0&0&0\\0&0&0&\zeta_7&0&0\\0&0&0&0&\zeta_{7}^{4}&0\\0&0&0&0&0&\zeta_{7}^{2}\end{array}\right).
$$
\normalsize
\subsection*{Step 2a} Let $k$ be a number field and consider the curve $\mathcal{C}/k$. We want to compute its twists over $k$. Let $K=k(\zeta_7,\,\zeta_3)$ and assume that $\left[K:k\right]=12$. Then, we compute using MAGMA the following possibilities for the pairs $(G,H)$:
\vskip 0.5truecm
\begin{center}
\begin{tabular}{|c|c|c|c|c|}
\hline
& $ \operatorname{ID}(G) $ \rule[0.4cm]{0cm}{0cm} & $ \operatorname{ID}(H)$ & $\operatorname{gen}(H)$ &$n_{(G,H)}$\tabularnewline
\hline
$1$ & $<12,5>$ & $<1,1>$ & $1$ & $1$\tabularnewline
\hline
$2$ & $<36,12>$ & $<3,1>$ & $r$ & $2$\tabularnewline
\hline
$3$ & $<84,7>$ & $<7,1>$ & $s$& $6$ \tabularnewline
\hline
$4$ & $<252,26>$ & $<21,2>$ & $r,\,s$ & $12$\tabularnewline
\hline
\end{tabular}
\end{center}
\vspace{3mm}
The fourth column in this table exhibits generators of the group $H$. In all the cases $G$ is the group generated by the elements $(g,1)$, for $g$ in $H$, together with the elements $(1,\tau_1)$ and $(1,\tau_2) $, where $\tau_1$ is the element in $\operatorname{Gal}(K/k)$ which sends $\zeta_3$ into $\zeta_3^2$ and $\zeta_7$ into $\zeta_7$, and $\tau_2$ is the element which sends $\zeta_3$ into $\zeta_3$ and $\zeta_7$ into $\zeta_7^3$. The fifth column exhibits the cardinality of the set in Formula (\ref{numero}) for each pair $(G,H)$.
\subsection*{Step 2b} Now, we have to find the proper solutions to the Galois embedding problems associated to each of the pairs $(G,H)$.
\begin{enumerate}
\item[1.] The first case is clear: $L=K$.
\item[2.] For the second one, note that $L=k(\zeta_7)M$, where $M/k$ is a solution to the Galois embedding problem in Proposition \ref{q} with $q=3$. Hence, $L=k(\zeta_3,\,\zeta_7,\,\sqrt[3]{m})$, for some $m\in\mathcal{O}_k$ that is not a $3$-power.
\item[3.] In this case, we can write $L=k(\zeta_3)M$, where $M/k$ is a solution to the Galois embedding problem in Proposition \ref{q} with $q=7$. Hence, $L=k(\zeta_3,\,\zeta_7,\,\sqrt[7]{n})$, for some $n\in\mathcal{O}_k$ that is not a $7$-power.
\item[4.] In the last case, $L=M_1M_2$, where $M_i/k$ is a solution to the Galois embedding problem in Proposition \ref{q} with $q=3,7$. Hence, $L=k(\zeta_3,\,\zeta_7,\,\sqrt[3]{m},\,\sqrt[7]{n})$, for some $m,n\in\mathcal{O}_k$, where $m$ is not a $3$-power and $m$ is not a $7$-power.
\end{enumerate}
\subsection*{Step 3} Fir each previous field $L$, we will compute equations of a twist that has $L$ as splitting field. The other twists, with splitting field $L$, will be then easily computed by considering symmetries. Let us consider a solution $\Psi$ (that is, a particular twist) to the Galois embedding problem with pair $(G,H)$ and splitting field $L$ by fixing an isomorphism between the group $H$ and the group $\text{Gal}(L/K)$:
$$
(r,1):\,\sqrt[3]{m},\,\sqrt[7]{n}\rightarrow \zeta_3\sqrt[3]{m},\,\sqrt[7]{n},
$$
$$
(s,1):\,\sqrt[3]{m},\,\sqrt[7]{n}\rightarrow \sqrt[3]{m},\,\zeta_7\sqrt[7]{n}.
$$
Now, we compute equations for a twist in each case:
\begin{enumerate}
\item[1.] Clearly, this solution gives us the trivial twist, so we have the curve $\mathcal{C}/k$.
\item[2.] The correspondence between the sets (\ref{e1}) and (\ref{e2}) gives us the cocycle given by $\xi_{\tau_1}=1$, $\xi_{\tau_2}=1$ and $\xi_{(r,1)}=r$. If we take the basis of $\Omega_L^1(\mathcal{C})$ given by $\left\{(a,b,c,i)\right\}:=\left\{ \sqrt[3]{m^a}\zeta_3^b\zeta_7^c\omega_i\right\}$ where $a,b\in\{0,1,2\}$, $c\in\{0,1,...,6\}$ and $i\in\{1,...,6\}$, we obtain the twisted action of $\operatorname{Gal}(L/k)$ on $\Omega_L^1(\mathcal{C})$ given in Section \ref{12}:
$$
\tau_1(a,b,c,i)=(a,2b,c,i),\,\tau_2(a,b,c,i)=(a,b,3c,i)
$$
$$
(r,1)(a,b,c,i)=\begin{cases}(a,a+b+2,c,i)&\text{if}\,\,\,i=4,6\\ (a,a+b+1,c,i)&\text{otherwise}\end{cases}
$$
Now, we use formula (\ref{CalcBase}) and get a basis of $\Omega^{1}(\mathcal{C}')\simeq\Omega_{L}^{1}(\mathcal{C})_{\xi}^{\operatorname{Gal}(L/k)}$ given by:
$$
\left\{\sqrt[3]{m^2}\omega_1,\,\sqrt[3]{m^2}\omega_2,\,\sqrt[3]{m^2}\omega_3,\,\sqrt[3]{m}\omega_4,\,\sqrt[3]{m^2}\omega_5,\,\sqrt[3]{m}\omega_6 \right\}.
$$
So we get the generators of the ideal defining the twist:
$$
\begin{aligned}
\omega_1\omega_6&-\omega_2\omega_4,&\omega_{2}^{2}&-\omega_1\omega_3,&
\omega_2\omega_3&-\omega_1\omega_5,&\omega_2\omega_5&-\omega_{3}^{2},\\
\omega_2\omega_6&-\omega_3\omega_4,&\omega_3\omega_6&-\omega_4\omega_5,&
m\omega_{4}^{3}-\omega_{3}^{2}&\omega_5+\omega_{1}^{3},&
\omega_{5}^{3}-m\omega_{4}\omega_{6}^{2}-&\omega_{1}\omega_{2}^{2}
\end{aligned}
$$
We obtain generators for the other solution $\Psi$ that has $L$ as splitting field by exchanging $m$ by $m^2$.
\item[3.] In this case, the correspondence between the sets (\ref{e1}) and (\ref{e2}) gives us the cocycle given by $\xi_{\tau_1}=1$, $\xi_{\tau_2}=1$ and $\xi_{(s,1)}=s$. If we take the basis of $\Omega_L^1(\mathcal{C})$ given by $\left\{(a,b,c,i)\right\}:=\left\{ \sqrt[7]{n^a}\zeta_3^b\zeta_7^c\omega_i\right\}$, where $a,c\in\{0,1,...,6\}$, $b\in\{0,1,2\}$ and $i\in\{1,...,6\}$, we obtain the twisted action of $\operatorname{Gal}(L/k)$ on it given in Section \ref{12}:
$$
\tau_1(a,b,c,i)=(a,2b,c,i),\,\tau_2(a,b,c,i)=(a,b,3c,i)
$$
$$
(r,1)(a,b,c,i)=\begin{cases}(a,a+b+1,c,i)&\text{if}\,\,\,i=1,4\\ (a,a+b+2,c,i)&\text{if}\,\,\,i=2,6\\ (a,a+b+3,c,i)&\text{if}\,\,\,i=3\\ (a,a+b+4,c,i)&\text{if}\,\,\,i=5\end{cases}
$$
Now, we use formula (\ref{CalcBase}) again and get a basis of $\Omega^{1}(\mathcal{C}')\simeq\Omega_{L}^{1}(\mathcal{C})_{\xi}^{\operatorname{Gal}(L/k)}$ given by
$$
\left\{\sqrt[7]{n^6}\omega_1,\,\sqrt[7]{n^5}\omega_2,\,\sqrt[7]{n^4}\omega_3,\,\sqrt[7]{n^6}\omega_4,\,\sqrt[7]{n^3}\omega_5,\,\sqrt[7]{n^5}\omega_6 \right\}.
$$
Then, we get the set of generators of the ideal defining the twist:
$$
\begin{aligned}
\omega_1\omega_6&-\omega_2\omega_4,&\omega_{2}^{2}&-\omega_1\omega_3,&
\omega_2\omega_3&-\omega_1\omega_5,&\omega_2\omega_5&-\omega_{3}^{2},\\
\omega_2\omega_6&-\omega_3\omega_4,&\omega_3\omega_6&-\omega_4\omega_5,&
\omega_{4}^{3}-n\omega_{3}^{2}&\omega_5+\omega_{1}^{3},&
n\omega_{5}^{3}-\omega_{4}\omega_{6}^{2}-&\omega_{1}\omega_{2}^{2}
\end{aligned}
$$
We compute generators for the other solutions $\Psi$ that have splitting field equal to $L$ by exchanging $n$ by $n^2, \,n^3,\,n^4,\,n^5,\,n^6$.
\item[4.] In the last case, we have the cocycle given by $\xi_\tau=1$, $\xi_{(r,1)}=r$ and $\xi_{(s,1)}=s$. We take the basis of $\Omega_L^1(\mathcal{C})$ given by $\left\{(a,b,c,d,i)\right\}:=\left\{ \sqrt[3]{m^a}\sqrt[7]{n^b}\zeta_3^c\zeta_7^d\omega_i\right\}$ where $a,c\in\{0,1,2\}$, $b,d\in\{0,...,6\}$ and $i\in\{1,...,6\}$, and we consider on $\Omega_L^1(\mathcal{C})$ the twisted action of $\operatorname{Gal}(L/k)$ given in Section \ref{12}.
Thus, formula (\ref{CalcBase}) provides a basis of $\Omega^{1}(\mathcal{C}')\simeq\Omega_{L}^{1}(\mathcal{C})_{\xi}^{\operatorname{Gal}(L/k)}$ given by:
$$
\left\{\sqrt[3]{m^2}\sqrt[7]{n^6}\omega_1,\,\sqrt[3]{m^2}\sqrt[7]{n^5}\omega_2,\,\sqrt[3]{m^2}\sqrt[7]{n^4}\omega_3,\,\sqrt[3]{m}\sqrt[7]{n^6}\omega_4,\,\sqrt[3]{m^2}\sqrt[7]{n^3}\omega_5,\,\sqrt[3]{m}\sqrt[7]{n^5}\omega_6 \right\}.
$$
Then, we get the set of generators of the ideal defining the twist
$$
\begin{aligned}
\omega_1\omega_6&-\omega_2\omega_4,&\omega_{2}^{2}&-\omega_1\omega_3,&
\omega_2\omega_3&-\omega_1\omega_5,&\omega_2\omega_5&-\omega_{3}^{2},\\
\omega_2\omega_6&-\omega_3\omega_4,&\omega_3\omega_6&-\omega_4\omega_5,&
m\omega_{4}^{3}-n\omega_{3}^{2}&\omega_5+\omega_{1}^{3},&
n\omega_{5}^{3}-m\omega_{4}\omega_{6}^{2}-&\omega_{1}\omega_{2}^{2}
\end{aligned}
$$
We compute generators for the other solutions $\Psi$ that have $L$ as splitting field by exchanging $m$ and $n$ by $m,\,m^2$ and $n,\,n^2, \,n^3,\,n^4,\,n^5,\,n^6$.
\end{enumerate}
We can summarize these results as follows:
\begin{proposition} The twists of the curve $\mathcal{C}/k$ defined above where $k$ is a number field such that $\left[k(\zeta_{21}):k\right]=12$, are in one-to-one correspondence with the curves given by the ideals generated by the following homogeneous polynomials:
$$
\begin{aligned}
\omega_1\omega_6&-\omega_2\omega_4,&\omega_{2}^{2}&-\omega_1\omega_3,&
\omega_2\omega_3&-\omega_1\omega_5,&\omega_2\omega_5&-\omega_{3}^{2},\\
\omega_2\omega_6&-\omega_3\omega_4,&\omega_3\omega_6&-\omega_4\omega_5,&
m\omega_{4}^{3}-n\omega_{3}^{2}&\omega_5+\omega_{1}^{3},&
n\omega_{5}^{3}-m\omega_{4}\omega_{6}^{2}-&\omega_{1}\omega_{2}^{2}
\end{aligned}
$$
where $m\in\mathcal{O}_{k}^{*}/(\mathcal{O}_{k}^{*})^3$ and $n\in\mathcal{O}_{k}^{*}/(\mathcal{O}_{k}^{*})^7$.
Equivalently, we can consider the (singular) plane models
$$
nx^7-my^3z^4-z^7=0.
$$
\end{proposition}
|
\section{Introduction}\label{sec:intro}
Polycyclic aromatic hydrocarbons (PAHs) are believed to be widely distributed throughout the interstellar medium (ISM) in both the gas and solid phases \citep{Tielens08,Salama08}. The observation of two sets of unidentified spectral features, the diffuse interstellar bands (DIBs) and the unidentified infrared bands (UIRs), is routinely attributed to the presence of PAHs (in absorption and emission, respectively).
The DIBs -- weak unidentified absorption features in the near ultraviolet (UV), visible, and near infrared (IR) region of the electromagnetic spectrum -- were first observed in the 1920s \citep{Heger22,Merrill34}. These bands are observed on lines of sight containing sufficiently high column densities, such as those traversing diffuse interstellar clouds.
Additionally, the UIR emission features at 3.3, 6.2, 7.7, 8.6, and 11.3~$\mu$m, first observed in the 1970s \citep[e.g.][]{Gillett73}, are observed in the IR spectra of almost all objects. Although no individual carrier molecule has been identified with certainty, the group of polyaromatic hydrocarbons
(PAHs) is believed to contribute to these bands \citep[e.g.][and references therein]{Allamandola89,Salama96}. The development of the so-called ``PAH hypothesis'', which holds PAHs responsible for the UIRs \citep{Leger84} due to UV-pumped IR fluorescence, was very quickly followed by an hypothesis suggesting that ionised PAHs were likely carriers of the DIBs \citep{Leger85,Allamandola85}.
Both experimental and theoretical studies support the identification of PAHs, and more specifically, PAH cations, as the carriers of the DIBs and UIRs \citep[e.g.][]{Pino99,Salama99,Mattioda05,Pathak08,Garkusha11,Gredel11}. Models have shown that UV pumping \citep[e.g.][]{Schutte93}, or even NIR absorption (in areas of low UV flux \citet{Mattioda08}), can drive MIR emission from PAHs, effectively linking DIBs with UIRs. One of the remaining points of debate is the nature of the PAH cations: whether they are radical cations (PAH$^+$) or protonated species (H$^+$PAH). Radical cations have been shown experimentally to be highly reactive and thus will react readily with hydrogen to form protonated species \citep{Snow98}; PAHs are, however, stripped of H atoms by interaction with photons \citep{Tielens08}.
More recent IR studies have suggested that the sub-group of nitrogen-substituted PAHs, or PANHs, exhibit spectral features similar to those of PAHs, and may also contribute to unidentified spectral bands \citep{Hudgins05, Bernstein05, Mattioda08, Alvaro10}. One study in particular illustrates the importance of PANHs in photon-dominated regions (PDRs), as the spectra of PANH cations were not merely useful, but required, to fit the 6.2 and 11~$\mu$m emission features observed towards NGC~7023 \citep{Boersma13}. This is not an unexpected result, as the abundance of N relative to C in the ISM is $\sim$~0.25 \citep{Spitzer78} and thus PANHs are likely relatively abundant interstellar species.
Additionally, \citet{Hudgins05} conclude that the substitution of N in PAHs could account for the variations observed in the peak position of the 6.2~$\mu$m UIR.
These spectroscopic studies in the IR have strengthened the case for the presence of charged PANHs in the ISM and their contribution to the UIRs, but very few studies have been made of the electronic spectra of gas phase PANH cations for comparison with the DIBs \citep{Dryza12}.
PANHs have a high proton affinity, so it is highly likely that protonated PANHs exist in the ISM, particularly in ionised environments. It has been illustrated that, for a small sample of PANH molecules, the IR spectra of protonated PANH cations (H$^+$PANH) better reproduce the 6.2 and 8.6~$\mu$m UIRs than the spectra of ionised PANH radical cations (PANH$^+$) \citep{Alvaro10}. However, to date no studies have determined the electronic spectra of H$^+$PANH for comparison with the DIBs.
In this study, we provide the electronic spectra of gas phase protonated cations acridine ([C$_{13}$H$_{9}$N]H$^+$, hereafter AcH$^+$) and phenanthridine ([C$_{13}$H$_{9}$N]H$^+$, hereafter PhH$^+$), nitrogen-substituted versions of the PAHs anthracene and phenanthrene (see Appendix~\ref{numbering}). We show that these protonated molecules present vibrationally resolved electronic states in the visible or near UV spectral region where DIBs are observed.
\section{Methods}
\subsection{Experimental methods}
The electronic spectra of the protonated aromatic PANH were obtained via parent ion photo-fragment spectroscopy in a cryogenically-cooled quadrupole ion trap (Paul Trap from Jordan TOF Products, Inc.) \citep{Alata13}.
The setup is similar to the one developed in several groups based on the original design by \citet{Wang08}.
The protonated ions are produced in an electrospray ionisation source built at Aarhus University \citep{Anderson04}. At the exit of the capillary, ions are trapped in an octopole trap for 90~ms. They are extracted by applying a negative pulse of \textit{ca.} 50~V and are further accelerated to 190~V by a second pulsed voltage just after the exit electrode. This timed sequence of pulsed voltages produces ion packets with durations of between 500~ns and 1~$\mu$s. The ions are driven by electrostatic lenses towards the Paul trap biased at 190~V so that the ions enter the trap gently, avoiding fragmentation induced by collisions. A mass gate placed at the entrance of the trap allows selection of the parent ion. The Paul trap is mounted on the cold head of a cryostat (Coolpak Oerlikon) connected to a water-cooled He compressor. Helium gas is injected in the trap using a pulsed valve (General Valve) triggered 1~ms before the ions enter the trap, as previously reported by \citet{Kamrath10}. The ions are trapped and thermalised at a temperature between 20 and 50~K through collisions with the cold He buffer gas. The ions are held in the trap for several tens of ms before the photodissociation laser is triggered. This delay is necessary both to ensure thermalisation of ions and the efficient pumping of the buffer gas from the trap in order to avoid collision induced dissociation of the ions during subsequent extraction.
The photo-dissociation laser is an optical parametric oscillator (OPO) laser, which has a 10~Hz repetition rate, 10~ns pulse width, a resolution of $\sim$~8~cm$^{-1}$ and a scanning step of 0.02~nm (laser from EKSPLA uab). The laser is shaped to a 1~mm$^2$ spot to fit the entrance hole of the trap and the laser power is around 20~mW in the UV spectral region.
After laser excitation, the ions are stored in the trap for a delay that can be varied between 20 and 90~ms before extraction into the 1.5~m time-of flight mass spectrometer. The full mass spectrum is recorded on a micro channel plates (MCP) detector with a digitising storage oscilloscope interfaced to a PC. The photofragmentation yield detected on each fragment is normalised to the parent ion signal and the laser power.
\subsection{Calculations}
\textit{Ab initio} calculations were performed with the TURBOMOLE programme package \citep{Ahlrichs89}, making use of the resolution-of-the-identity (RI) approximation for the evaluation of the electron-repulsion integrals \citep{Hattig03}. The equilibrium geometries of the protonated species in their electronic ground states (S$_0$) were determined at the MP2 (M{\o}ller-Plesset second order perturbation theory) level. Adiabatic excitation energies of the lowest electronic excited singlet states (S$_1$) were determined at the RI-ADC(2)(second order Algebraic Diagrammatic Construction level \citep{Schirmer82}). Calculations were performed with the correlation-consistent polarised valence double-zeta (cc-pVDZ) basis set \citep{Woon93}. The vibrations in the ground and excited states were calculated for protonated acridine and phenanthridine and the electronic spectra simulated using the PGOPHER software \citep{Western} for Franck Condon analysis.
\section{Results and discussion}
\subsection{Protonated acridine}
\begin{figure*}[!tb]
\includegraphics[width=\textwidth]{figure1new.eps}
\caption{Photofragmentation spectrum of protonated acridine compared to the simulated spectrum: a) the experimental spectrum plotted as the fragment ion signal (arbitrary units) as a function of the excitation energy; b) the simulated spectrum plotted as the Franck-Condon factors as a function of the excitation energy, calculated using the calculated ground and excited state frequencies. The main vibrational progressions (indicated in blue in panel a) involve, essentially, the two symmetric modes $\nu_{10}$ at 383~cm$^{-1}$ (ring breathing) and $\nu_{18}$ at 587~cm$^{-1}$. These vibrational modes are illustrated schematically above the experimental spectrum, with arrows representing the direction and relative magnitude of the displacement of each atom. The intensities of the out of plane modes $\nu_{1}$ and $\nu_{3}$ are overestimated in the simulation.}\label{fig:acridine}
\end{figure*}
The protonated acridine photofragmentation spectrum is presented in Fig.~\ref{fig:acridine}a. The first electronic transition shows clear vibrational progressions in the visible starting at 442.24~nm (22612~cm$^{-1}$/ 2.80~eV) and extending to $\sim$~400~nm, this absorption region being in agreement with the absorption recorded in water solution at low pH \citep{Ryan97}. These progressions involve the 383~cm$^{-1}$ vibrational mode starting from the 0-0 transition ($\nu_0$), and a symmetric mode at 587~cm$^{-1}$ ($\nu_{18}$). The vibronic bands are narrow, with a FWHM of 10~$\pm$~1 cm$^{-1}$ that represents the convolution of the laser band width ($\sim$~8~cm$^{-1}$) with a low temperature rotational contour \textit{i.e.} the apparent line broadening due to the fact that the spacing between rotational lines is narrow compared to the laser width.
The main fragmentation channels correspond to loss of H, 2H/H$_2$ and m/z~28 (probably corresponding to loss of H$^+$HCN/H$^+$HNC, \citet{Johansson11}) complemented by weaker fragmentation channels: loss of m/z~26, 27, 29 and loss of m/z~51, 52 and 53 (see mass spectra in Appendix~\ref{mass_spectra}).
Ground and first excited state calculations were performed, both with DFT/TD-DFT (Density Functional Theory/ Time Dependent-DFT) and MP2/ADC(2) methods, for the isomer protonated on the nitrogen atom, which is by far the most stable. The ground state energies of the isomers protonated on carbon atoms were also calculated using DFT with the B3LYP functional and the cc-pVDZ basis set, and they all lie more than 1.5~eV higher than the isomer protonated on N. The molecule stays planar in both states. The first excited state is calculated vertically at 3.13~eV, in agreement with previous CASPT2 (complete active space perturbation theory) calculations \citep{Rubio01}, and corresponds to a $\pi\pi$* HOMO-LUMO (highest occupied molecular orbital-lowest unoccupied molecular orbital) transition ($^1$L$_a$ in Platt's notation). Optimisation leads to an adiabatic transition at 2.75~eV (2.60~eV when the difference in zero point energy is included), in agreement with the experiment (2.80~eV), which corroborates the location of the proton on the nitrogen atom.
The changes in geometry between the ground and excited states are minor: for the outer rings, there is an alternation of long/short C-C bonds in S$_0$, which becomes short/long in S$_1$, with a maximum difference in bond length of 0.053~{\AA} for the C1C2 and C7C8 bonds (see Appendix~\ref{numbering} for the S$_0$ and S$_1$ optimised geometries and the numbering of atoms); for the center ring bearing the N atom, the NH bond is unchanged, N10C11 \& N10C14 increase by 0.031~{\AA}, C9C12 \& C9C13 increase by 0.035~{\AA}, while C11C12 \& C13C14 decrease by 0.008~{\AA}. The angles differ by at most 2$^{\circ}$.
The spectrum was simulated (see Fig.~\ref{fig:acridine}b) using the ground and excited state frequencies calculated and the PGOPHER software for Franck Condon analysis \citep{Western}. The agreement between the experimental and simulated spectra is quite good for the band positions, and allows assignment of the main vibrational progressions to the symmetric breathing mode $\nu_{10}$ (contraction/elongation of the rings along the long axis) alone or in combination with another symmetric mode $\nu_{18}$ (elongation of the center ring along the short axis with contraction of the outer rings). The scheme of the vibrations is shown at the top of Fig.~\ref{fig:acridine}. Two out of plane modes are weakly active in the experimental spectrum and overestimated in the simulation: $\nu_{1}$ is the butterfly mode and $\nu_{3}$ mostly an out of plane motion of the N10-H and C9-H atoms (seesawing motion).
\subsection{Protonated phenanthridine}
\begin{figure*}[!tb]
\includegraphics[width=\textwidth]{figure2new.eps}
\caption{Photofragmentation spectrum of protonated phenanthridine compared to the simulated spectrum: a) the experimental spectrum plotted as the fragment ion signal (arbitrary units) as a function of the excitation energy; b) the simulated spectrum plotted as the Franck-Condon factors as a function of the excitation energy, calculated using the calculated ground and excited state frequencies. The main vibrational progressions (indicated in red in panel a) $\nu_{5}$ and $\nu_{17}$, are illustrated schematically below the simulated spectrum.}\label{fig:phenanthridine}
\end{figure*}
The protonated phenanthridine photofragmentation spectrum (presented in the upper panel of Fig.~\ref{fig:phenanthridine}) shows clear vibrational progressions in the near UV spectral region, with a first transition at 397.28~nm (25171 cm$^{-1}$/ 3.12~eV). This transition is blue shifted by 2500~cm$^{-1}$ compared to the AcH$^+$ transition. Vibrational progressions on a mode of 240~cm$^{-1}$ starting on the 0-0 transition ($\nu_0$), and on a symmetric mode at 533~cm$^{-1}$ ($\nu_{17}$), are observed for several quanta, and the bands are narrow, with a FWHM of 12~$\pm$~1~cm$^{-1}$ (see Appendix~\ref{PhH+vibronic}). As for AcH$^+$, the main fragmentation channels correspond to loss of H, 2H/H$_2$ and H$^+$HCN/H$^+$HNC (see mass spectrum in Appendix~\ref{mass_spectra}).
Ground and first excited state calculations performed for the isomer protonated on the nitrogen atom show that PhH$^+$ stays planar in both states (see Appendix~\ref{numbering}). As for AcH$^+$, this isomer is by far the most stable, the calculated ground state energies of the isomers protonated on carbon atoms (DFT/ B3LYP/ cc-pVDZ) all lie more than 1.6~eV higher than the isomer protonated on N. The first excited state is a $\pi\pi$* state calculated vertically at 3.45~eV and optimisation leads to an adiabatic transition at 3.22~eV, in agreement with the experiment (3.12~ev).
\subsection{Three-ringed protonated PANHs}
For both protonated acridine and phenanthridine, the proton is located on the nitrogen atom, as shown by the excellent agreement between experiments and calculations: the calculated excited state transitions are within 0.2~eV from the experimental band origins and the spectra simulated with the calculated frequencies reproduce well the experimental progressions. Protonated acridine and phenanthridine are planar in the ground and excited states, which was not the case for protonated anthracene and phenanthrene \citep{Alata10b}. The line widths are narrow (taking into account the laser bandwidth), indicating long lived excited states.
These two protonated PANHs absorb in the visible or near UV spectral region, this absorption region being in agreement with the absorption recorded in water solution at low pH \citep{Ryan97}. This is the same absorption region as previously recorded for protonated PAHs with two to six aromatic rings, protonated naphthalene \citep{Alata10a}, anthracene \citep{Alata10b,Garkusha11}, phenanthrene \citep{AlataThesis,Garkusha11}, fluorine \citep{Alata12b}, tetracene \citep{Alata10b}, pyrene \citep{AlataThesis,Garkusha11,Hardy13} and coronene \citep{Garkusha11,Rice14}.
\begin{table*}
\caption{Transition origin (eV) for selected neutral and protonated PAHs and PANHs.}\label{table:transitions}
\begin{tabular}{cccc}
\hline\hline
PAH & Transition & PANH & Transition \\
& origin (eV) & & origin (eV) \\
\hline
Anthracene & 3.43$^{1}$ & Acridine & 3.22$^{2,3}$\\
AnthraceneH$^+$ & 2.52$^{4}$ & AcridineH$^+$ & 2.80$^{5}$\\
\hline
Phenanthrene & 3.63$^{6}$ & Phenanthridine & 3.64$^{7}$\\
PhenanthreneH$^+$ & 2.08$^{8,9}$ & PhenanthridineH$^+$ & 3.12$^{5}$\\
\hline
\end{tabular}
\tablebib{
(1) \citet{Lambert84}; (2) \citet{Prochorow98}; (3) \citet{Rubio01}; (4) \citet{Alata10b}; (5) this work; (6) \citet{Hager88}; (7) \citep{Prochorow04}; (8) \citet{AlataThesis}; (9) \citet{Garkusha11}.
}
\end{table*}
Protonation of aromatic heterocycles shifts the absorption spectra to lower energy, but to a smaller extent than for the fully carbonated analogues: for example acridineH$^+$ is red shifted by 0.4~eV from acridine, while anthraceneH$^+$ is red shifted by 0.9~eV from neutral anthracene (see Table~\ref{table:transitions}). This results in a blue shift of the transition of protonated acridine and phenanthridine compared to protonated anthracene and phenanthrene. The reverse was observed in the case of radical cations: ionised quinoline and isoquinoline have their transitions red shifted compared to ionised naphthalene \citep{Dryza12}.
\section{Astrophysical Implications}
\begin{table*}
\caption{Comparison of observed DIB wavelengths and equivalent widths with experimental AcridineH$^+$ wavelengths (in nm).}\label{table:wavelengths}
\begin{tabular}{ccccc}
\hline\hline
\multicolumn{2}{c}{DIB} & \multicolumn{3}{c}{Experimental AcH$^+$} \\
wavelength & FWHM/{\AA} & wavelength & intensity\tablefootmark{a} & Assignment\tablefootmark{b} \\
\hline
442.883$^{1}$/442.819$^{2}$/442.888$^{3-5}$ & 22.5$^{1,2}$/12.3$^3$ & 442.24 & 100 & 0$_0^0$ \\
& & 439.54 & 15 & 1$_0^2$ \\
437.173$^{1}$/436.386$^{2}$ & 1.03/0.46$^{1,2}$ & & & \\
& & 434.86 & 170 & 10$_0^1$ \\
& & 433.38 & 30 & \\
& & 432.74 & 85 & 3$_0^2$ \\
& & 432.18 & 15 & \\
& & 431.42 & 30 & \\
& & 431.04 & 130 & 18$_0^1$ \\
& & 430.30 & 55 & \\
& & 429.52 & 95 & \\
& & 428.9 & 20 & \\
& & 428.3 & 65 &\\
& & 427.62 & 150 & 10$_0^2$ \\
& & 427.14 & 25 & \\
& & 426.22 & 20 & \\
425.901$^{2}$ & 1.05$^{1,2}$& 425.62 & 55 & \\
& & 424.32 & 20 & \\
& & 423.98 & 110 & 10$_0^1$18$_0^1$ \\
& & 423.28 & 25 & \\
& & 423.02 & 50 & \\
& & 422.46 & 55 & \\
& & 420.56 & 60 & 10$_0^3$ \\
& & 420.24 & 55 & \\
& & 418.94 & 25 & \\
417.65$^{3-5}$ & 23.3$^{3}$& 417.62 & 50 & 10$_0^2$18$_0^1$ \\
& & 417.16 & 35 & \\
& & 416.24 & 35 & \\
& & 415.46 & 60 & \\
& & 414.84 & 65 & \\
& & 413.83 & 85 & 10$_0^1$18$_0^2$ \\
& & 413.5 & 45 & \\
& & 412.72 & 35 & \\
& & 411.08 & 25 & \\
\hline
\end{tabular}
\tablebib{
(1) \citet{Hobbs09}; (2) \citet{Hobbs08}; (3) \citet{Jenniskens94}; (4) \citet{Krelowski95}; (5) \citet{Tuairisg00}.
}
\tablefoot{
\tablefoottext{a}{The intensity of the vibronic bands relative to the origin band are given (as percentages), but the absolute values should be used for guidance only, as there are possible saturation effects in the experimental method which are not well controlled.}
\tablefoottext{b}{The notation $v_0^n$ corresponds to a transition from the
vibrationless level in the ground electronic state, S$_0$, to a level
with $n$ quanta of the vibrational mode $v$ in the S$_1$ electronic excited
state. At the low temperature of the experiment, only the
vibrationless level is populated in the S$_0$ ground state.}}
\end{table*}
In this study we have reported the electronic spectra for two protonated three-ringed PANH cations, derived both experimentally and theoretically. These are the first such experimental spectra, and allow us to draw several astrophysically-relevant conclusions for H$^+$PANH ions. Protonated PAHs have been suggested to be among the potential candidates for diffuse interstellar band (DIB) carriers \citep{Salama96,LePage97,Snow98,Pathak08,Hammonds09} and this study was designed to determine if protonated PANHs could be DIBs carriers by comparing the laboratory gas phase absorptions at low temperature to astronomical data \citep{Hobbs08,Hobbs09}.
The first important conclusion is that H$^+$PANH ions represent good candidates for the source of the DIBs. The substitution of a N atom into the H$^+$PAH skeleton introduces a blue shift in the UV-visible electronic spectrum, resulting in bands at higher energies than those observed for H$^+$PAHs. It has previously been shown that both N-substitution \citep{Hudgins05} and protonation of PAHs \citep{Alvaro10} introduce blue shifts in the IR emission features, giving a better fit to the UIR features attributed to PAHs. A direct comparison of our H$^+$PANH spectra with H$^+$PAH spectra of anthracene-9H and phenanthene-9H \citep{Garkusha11} reveals blueshifts of $\sim$~11 and 133~nm, respectively, due to the presence of N in the ion. The destabilising effect of N on the ion in its excited state is particularly marked for phenanthridine. For the small PANH investigated here, this blue shift results in the majority of their absorption features occurring at higher energies than those observed for DIBs, but this does not preclude the bands of larger, more astrophysically-relevant PANHs falling within the DIBs frequency range.
Importantly, the absorption lines in our spectra have FWHM of approximately 10~cm$^{-1}$ ($\sim$~15 -- 20~{\AA}), which (upon correction for the $\sim$~8~cm$^{-1}$ bandwidth of the laser) are comparable to observed DIB line widths of 2 -- 3~cm$^{-1}$. Our data are directly comparable to all of the DIBs in the 445 -- 410 nm range, two of which have FWHM of $\sim$~ 23~{\AA} ($>$ 100~cm$^{-1}$) and three of which are on the order of 1~{\AA} (2 -- 5 cm$^{-1}$, see Table~\ref{table:wavelengths}). The narrow bands of these H$^+$PANH species are in direct contrast with those of PAH$^+$ radical cations, which are much wider ($\sim$~20 -- 30~cm$^{-1}$, \citet{Hudgins05}) due to their intrinsically shorter excited state lifetimes. Due to the closed-shell nature of the H$^+$PANH, they are also more photochemically stable than open-shell radical cation species, and therefore would be expected to be more stable in, for example, PDRs. It should be noted here that neutral PAHs (which are themselves closed-shell species) also exhibit narrow absorption lines of $\sim$~3~cm$^{-1}$ \citep{Tan05} and are likely to be major contributors to the weak DIBs.
The first electronic transition that we observe in the spectrum of protonated acridine is centred at 4422.4~{\AA} and has a bandwidth-corrected FWHM of $<$~15~{\AA}. The strongest broad DIB is centred at 4428.2~{\AA} and has a FWHM of 22.50~{\AA} \citep{Hobbs08,Jenniskens94}. The AcH$^+$ band is shifted by $\sim$ 6~{\AA} compared to this unusually broad DIB, and thus we conclude that it is not responsible for the interstellar feature. The other strong bands in the AcH$^+$ spectrum do not match known DIBs either (see Table~\ref{table:wavelengths}).
Moreover, the poor correlation between DIBs precludes the contribution of molecular systems with high vibrational activity to these bands. Although the protonated PANHs studied here present fewer active modes than their fully carbonated analogues, they do show vibrational progressions. The same behaviour seems to exist in small PANH radical cations, which present long vibrational progressions in the visible \citep{Dryza12}. Nonetheless, the progressions that we observe in our small H$^+$PANHs would likely become less pronounced in larger species as they derive from vibrational modes which would be damped by the larger aromatic structure. Additionally, PANHs bend less than the corresponding PAHs, therefore there is more chance of observing isolated bands. Protonation of the studied PANHs occurs preferentially on the N atom, and both species contain N substituted in the PAH exoskeleton. The most favourable protonation orientation was thus H$^+$ in the plane of the molecule, maintaining the planar nature of the species upon ionisation. This would not necessarily be the case for endoskeletal N as \textit{(i)} the presence of N already breaks the planar nature of the molecule and \textit{(ii)} if protonation occurs on the N, it would occur above or below the planar ring structure. The resulting non-planar structure would most likely have additional vibration in electronically excited states, resulting in vibrational progressions in the electronic spectra. Thus, exoskeletal H$^+$PANH probably represent better candidates for the DIBs.
However, as discussed above, it is important to remember that small H$^+$PAH and H$^+$PANH molecules will not be stable in interstellar environments such as PDRs or diffuse HI clouds, and these experiments must be extended to species with larger skeletons. Previous studies suggest that the 6.2~$\mu$m band is best reproduced by PAHs with 60 -- 90 C atoms \citep{Schutte93} and with 2 -- 3 N atoms \citep{Hudgins05}. Such species may be photostable and their electronic transitions should be more to the red, where there is a larger number of DIBs. As mentioned in \S~\ref{sec:intro}, the interstellar abundance of N is $\sim$~0.25 that of C \citep{Spitzer78}. Thus, from a purely statistical point of view, all PAHs large enough to be stable in the ISM (approximately $>$ 50 C) should contain more than $\sim$~ 10 N atoms. Any deviation from these statistical abundances therefore reveals evidence of selectivity in the formation routes to such species. \citet{Hudgins05} conclude that ``most'' PAH species contributing to the UIR features contain nitrogen atoms, with an estimated $>$ 1.2~\% of interstellar N contained in PANHs. When compared to the fraction of elemental C locked up in PAHs (approximately 3.5~\% \citep{Tielens08}), the relative abundance of N appears slightly more than that expected from the statistical N:C ratio, but this could be due to the difference in derivation methods for PAH and PANH abundances.
On the basis of the results of IR \citep{Hudgins05} and UV-visible (this work) studies into PANHs, it is clear that such species are good candidates for inclusion in the ``PAH hypothesis''. Although the spectra of the small H$^+$PANHs studied in this work do not, themselves, correspond to observed DIBs, this preliminary study points to the need for further examination of this class of molecular ion. In particular, the effects of larger PANHs, substitution of larger numbers of N (or other heteroatoms), and endoskeletal heteroatom substitution on the UV-visible and IR spectra of protonated PANH molecules should be investigated.
\begin{acknowledgements}
J.~A. Noble is a Royal Commission for the Exhibition of 1851 Research Fellow. We acknowledge the use of the computing facility cluster GMPCS of the LUMAT federation (FR LUMAT 2764).
\end{acknowledgements}
|
\section{Introduction}
Adiabatic process is aimed at stabilizing a parameter-varying quantum
system at its eigenstate. This process has many applications in the
engineering of quantum systems
\cite{Bergmann98,Wu13,Ribeiro13,Zhang13,Wang12}, and in particular plays
the fundamental role in adiabatic quantum computation (AQC)
\cite{Farhi00,Farhi01,Sarandy05}. The adiabatic theorem
\cite{Born28,Kato50} states that a system will undergo adiabatic evolution
given that the system parameter varies slowly.
Quantifying the applicability of adiabatic approximations is an interesting topic of current research efforts. On the one hand, this kind of research has been spurred by so-called shortcuts to adiabaticity \cite{Torrontegui13}, and on the other hand recent insights from thermodynamics haven put adiabatic processes back into focus \cite{Acconcia15,Sivak12}. In particular, the validity of the adiabatic theorem has been under intensive studies both theoretically and experimentally since it was proposed, and much of these efforts were devoted to the rigorous
description of the sufficient quantitative conditions of adiabatic theorem, and the estimation of the error accumulated over a
long time \cite{Kato50,Nenciu93,Avron12,Cao12}. Once the exact knowledge on the
adiabatic process is available, it is straightforward to apply the results to the optimal
design of adiabatic control on specific systems \cite{Rezakhani09,Wilson12}. The most interesting
progress is that the validity of the adiabatic theorem itself has been challenged in
the recent decade \cite{Marzlin04,Tong05,Tong07,Du08,Zhao08,Comparat09,Ortigoso12,Rigolin12,Zhang14}, both by strict
analysis and counter-examples. According to these findings, the errors induced by the adiabatic
approximation could accumulate over time despite certain quantitative condition is satisfied \cite{Marzlin04,Tong05,Tong07,Comparat09,Ortigoso12}, e.g., when
there exists an additional perturbation or driving that is resonant with the system. Particularly as indicated in \cite{Comparat09}, it is not new that resonant driving can cause population transfer between eigenstates. Also, a proof can be found in \cite{Ortigoso12} stating that only a resonant perturbation whose amplitude gradually decays to
zero can result in a violation of a well-known sufficient condition.
In this paper we consider the following process: the process starts at $t=0$. The system Hamiltonian at $t=0$
is $H_1$, and the system Hamiltonian at $t=T$ is $H_2=H_1+\lambda\Delta H,\lambda>0$. $\lambda$ is a dimensionless quantity. $\Delta H$ is a fixed operator and so the direction of the variation is fixed. We assume $H_1$, $H_2$, and $\Delta H$ are bounded operators throughout this
paper. $T$ is the evolution time. The transition of the system from $H_1$ to $H_2$ can be described using an interpolating function $f(t)$ so that
\begin{eqnarray}\label{ad1}
H(t)=H_1+f(t)(H_2-H_1)=H_1+\lambda f(t)\Delta H,
\end{eqnarray}
with $f(0)=0$ and $f(T)=1$. We work under the condition that a valid
perturbative analysis of the system evolution is available. This often
means $\lambda$ should be smaller than a threshold value. It is worth mentioning that the classical adiabatic theorem was proved also using a perturbative analysis, which cannot be applied directly to a large variation of Hamiltonian. Therefore, our
analysis in this paper is not concerned with the adiabatic evolution for a
large variation of Hamiltonian. However, our analysis provides a rigorous
estimation of the error accumulated during this
small-variation evolution for an
arbitrarily given interpolation.
Our work is different from the previous works in two ways. First, instead
of studying the evolution of the eigenstates and their corresponding
probability amplitudes, the mean value of a Hermitian operator is defined
as a measure of the error. For example, in the context of adiabatic quantum
computation where one wants to prepare the ground state of a target
Hamiltonian $\hat{H}_2\geq0$ whose ground-state energy is $0$, $\epsilon=\langle \hat{H}_2\rangle_{\rho_t}$ serves
as a good measure of the distance between the real-time state $\rho_t$ and the
ground state. This measure resembles the excitation energy or excess work performed during the process, as studied in thermodynamics \cite{Acconcia15}. In this paper we only consider the error accumulated over the
entire process, which means we are only interested in $\epsilon=\langle
\hat{H}_2\rangle_{\rho_T}$. The second difference is that the error, or the excitation energy or excess work performed during the process, can be estimated
with a sufficient precision for arbitrarily given interpolating
functions. As a result, the parameters which are related to the suppression
of the error can be easily identified. For example, we have
$\epsilon=O(\frac{\lambda^2}{T^2\lambda_2^3})$ as $\lambda\rightarrow0$ in
the case of linear
interpolation. Here $\lambda_2$ is the energy gap between the ground and first-excited states of the initial Hamiltonian. However for the interpolation in the counterexample
\cite{Marzlin04,Ortigoso12}, the scaling of $\epsilon$ is not so simple.
This paper is organized as follows. In Section \ref{secdef}, we introduce the model of this paper. In Section \ref{seclinear}, we give the estimation of the error for linear interpolation. In section \ref{secnonlinear}, we present the general algorithm to estimate the error for an arbitrarily given interpolating function. We discuss three examples in Section \ref{secexam}. Conclusion is given in section \ref{seccon}.
\section{Definitions and Preliminaries}\label{secdef}
The system is defined on an $N$-dimensional Hilbert space. We set Dirac constant $\hbar=1$. $||\cdot||$ denotes the matrix norm. Two real functions $f_1(x)$ and $f_2(x)$ can be denoted as $
f_1(x)=O(f_2(x)),\ x\rightarrow\infty,$ if and only if there exists a positive real number $M$ and a real number $x_0$ such that $|f_1(x)|\leq M|f_2(x)|,\ x\geq x_0$, where $|\cdot|$ denotes the absolute value.
Let $\{\omega_i:i=1,2,...N\}$ be the monotonically increasing sequence of
eigenvalues of $H_1$, so that $\omega_i\geq\omega_j$ when $i>j$, and
$\{|i\rangle\}$ be the corresponding eigenstates. We denote the energy gap
between the $i$th eigenstate and the ground state
as $\lambda_i=\omega_i-\omega_1$. Similarly, we define
the increasing sequence of eigenvalues of $H_2$, $\{\omega_i^{'}:i=1,2,...N\}$ and $\{\lambda_i^{'}\},\{|i^{'}\rangle\}$ correspondingly.
For convenience, we also introduce two offset Hamiltonians, $\hat{H}_1$ and
$\hat{H}_2$. The Hamiltonian $\hat{H}_1$ is defined as
$\hat{H}_1=H_1-\omega_1$, i.e., by offsetting the Hamiltonian of the system at
$t=0$ by a constant operator $\omega_1$ so that
$\hat{H}_1\geq0$.
By $\hat{H}_1\geq0$ we mean $\hat{H}_1$ is positive semidefinite and its
the smallest eigenvalue of $\hat{H}_1$ is zero. Similarly, we define
$\hat{H}_2=H_2-\omega^{'}_1\geq0$ by offsetting
the system Hamiltonian by a constant operator $\omega^{'}_1$. Let $\rho_t$ denote the system state at time $t$ and let $\rho_g$
be the initial state of the system at $t=0$. We
always assume that $\rho_g$ is the ground state of $\hat{H}_1$, and so we
have $\langle \hat{H}_1\rangle_{\rho_g}=0$.
The measure of adiabaticity is proposed as follows
\begin{Def}
The distance between the final state and the ground state of $H_2$ is measured by
\begin{equation}\label{nde}
\epsilon=\langle \hat{H}_2\rangle_{\rho_T}.
\end{equation}
\end{Def}
Obviously, if the evolution is adiabatic, i.e., $\rho_T$ is the ground
state of $H_2$, then we have $\epsilon=0$. In particular, $\epsilon$ is closely related to the fidelity of the final state and ground state in the Schr{\"o}dinger picture (See Appendix \ref{appendix3}). A small error $\epsilon$ implies a large fidelity.
In this paper we also call $\epsilon$ the adiabatic approximation error, as $\epsilon$ reflects how well we can approximate the evolution as a perfect adiabatic process.
In this paper we only consider $\lambda$ such that $\rho_t,\ t\in[0,T]$ can be expanded using Magnus series in the interaction picture.
For more details about the expansion in the interaction picture, please
refer to Appendix \ref{appendix1}. If the series expansion is valid in the
interaction picture, we can transform back to the Schr\"{o}dinger picture
and write the evolution of the state as (see Appendix \ref{appendix1})
\begin{eqnarray}\label{new1}
\rho_t=e^{-\mbox iH_1t}(\rho_g+R(t)+\mbox{i}[\rho_g,\lambda\int_0^tdt^{'}e^{\mbox iH_1t^{'}}f(t^{'})\Delta He^{-iH_1t^{'}}])e^{\mbox{i}H_1t},
\end{eqnarray}
where we have $||R(t)||=O(\lambda^2)$. A sufficient condition for the Magnus series to converge is given by (see Appendix \ref{appendix1})
\begin{equation}
\lambda<\frac{\pi}{\|\Delta H\|\int_0^Tf(t)dt}.
\end{equation}
Our aim is to estimate an asymptotic behaviour of $\epsilon$ provided
$\lambda\to 0$. Furthermore, we will use the obtained estimate
to analyze several cases of the adiabatic theorem including those where
some difficulties with adiabatic approximation have been encountered.
\section{Adiabatic approximation under linear interpolation of the Hamiltonian}\label{seclinear}
The Heisenberg evolution of the expectation of an observable is written as
\begin{equation}\label{heidy}
\frac{d}{dt}\langle X(t)\rangle_{\rho_g}=\langle-\mbox{i}[X(t),H]\rangle_{\rho_g},
\end{equation}
where $H$ is the system Hamiltonian. Recall that $\rho_g=|1\rangle\langle 1|$. Since $H_1|1\rangle=\omega_1|1\rangle$, $\langle X(t)\rangle_{\rho_g}$ is a constant of motion under the action of $H_1$:
\begin{equation}\label{ini}
\frac{d}{dt}\langle X(t)\rangle_{\rho_g}=\langle-\mbox{i}[X(t),H_1]\rangle_{\rho_g}=0=\langle-\mbox{i}[X(t),\hat{H}_1]\rangle_{\rho_g}
\end{equation}
for any Hermitian operator $X(t)$.
We will need to study the dynamics of $\langle \hat{H}_2\rangle_{\rho_t}=\langle \hat{H}_2(t)\rangle_{\rho_g}$ in order to solve for $\epsilon$. The time evolution of $\langle \hat{H}_2\rangle_{\rho_t}$ is determined by its generator $\frac{d}{dt}\langle \hat{H}_2\rangle_{\rho_t}=\langle-\mbox{i}[\hat{H}_2,H(t)]\rangle_{\rho_{t}}$. For linear interpolating function $f(t)=\frac{t}{T}$, integration of $\frac{d}{dt}\langle \hat{H}_2\rangle_{\rho_t}$ over $[0,T]$ results in the following expression (See details in Appendix \ref{appendix2}):
\begin{eqnarray}
&&\langle \hat{H}_2\rangle_{\rho_T}-\langle \hat{H}_2\rangle_{\rho_g}=\int_0^T(\langle-\mbox{i}[\hat{H}_2,H(t)]\rangle_{\rho_{t}})dt\nonumber\\
&=&\int_0^Tdt[-2(1-f(t))\sum_{i\neq 1}(\omega_i-\omega_1)\langle 1|\hat{H}_2|i\rangle\langle i|\hat{H}_2|1\rangle\int_0^tdt^{'}f(t^{'})\cos((\omega_i-\omega_1)(t^{'}-t))]\nonumber\\
&+&\int_0^Tdt\tr\{-\mbox{i}e^{\mbox{i}H_1t}[\hat{H}_2,(1-f(t))\hat{H}_1]e^{-\mbox{i}H_1t}R(t)\}.\label{ft}\\
&=&\sum_{i\neq 1}(-\frac{1}{\lambda_i}+\frac{4\sin^2(\lambda_iT/2)}{T^2\lambda_i^3})\langle 1|\hat{H}_2|i\rangle\langle i|\hat{H}_2|1\rangle\nonumber\\
&+&\int_0^Tdt\tr\{-\mbox{i}e^{\mbox{i}H_1t}[\hat{H}_2,(1-\frac{t}{T})\hat{H}_1]e^{-\mbox{i}H_1t}R(t)\}\label{ad9}
\end{eqnarray}
As we noted before, $\langle \hat{H}_2\rangle_{\rho_T}$ is exactly zero
if $\rho_T$ is the ground state of $\hat{H}_2$. If $\rho_T$ is not the
ground state of $\hat{H}_2$, we can determine the bound on $\epsilon=\langle
\hat{H}_2\rangle_{\rho_T}$ from the following equality
\begin{eqnarray}\label{int1}
\langle \hat{H}_2\rangle_{\rho_T}-\langle \hat{H}_2\rangle_{\rho_g}&=&\int_0^{T}\langle-\mbox{i}[\hat{H}_2,H(t)]\rangle_{\rho_{t}}dt\nonumber\\
&=&\int_0^{T}\langle-\mbox{i}[H_2,H(t)]\rangle_{\rho_{t}}dt=\langle H_2\rangle_{\rho_T}-\langle H_2\rangle_{\rho_g}.
\end{eqnarray}
Since
\begin{equation}
\hat{H}_2=H_2-\omega_1^{'},
\end{equation}
The error $\epsilon$ can be expressed as
\begin{equation}\label{int3}
\epsilon=\langle H_2\rangle_{\rho_T}-\langle H_2\rangle_{\rho_g}-[\omega_1^{'}-\langle H_2\rangle_{\rho_g}].
\end{equation}
With the aid of (\ref{ad9}), we can investigate the rate of convergence of $\epsilon$ to zero as $\lambda$ tends to zero in the
case where $f(t)$ defines a linear interpolation, as summarized in the
following proposition:
\begin{prop}\label{theorem1}
Assume $\lambda_2>0$ (the ground state of $H_1$ is non-degenerate) and
suppose $f(t)=t/T$, which corresponds to the linear interpolation of the
Hamiltonian. The estimation of $\epsilon$ is given by $\sum_{i\neq 1}\frac{4\lambda^2\sin^2(\lambda_iT/2)|\langle 1|\Delta H|i\rangle|^2}{T^2\lambda_i^3}+O(\lambda^3)$, which is of the order
$O(\frac{\lambda^2}{T^2\lambda_2^3})$ as $\lambda\rightarrow0$.
\end{prop}
\begin{proof}
Referring to (\ref{int3}) and (\ref{int1}), we need to compute the difference between (\ref{ad9}) and $\omega_1^{'}-\langle 1|H_2|1\rangle$. First we write (\ref{ad9}) as
\begin{eqnarray}
\langle H_2\rangle_{\rho_T}-\langle H_2\rangle_{\rho_g}&=&\sum_{i\neq 1}(-\frac{1}{\lambda_i}+\frac{4\sin^2(\lambda_iT/2)}{T^2\lambda_i^3})\langle 1|\hat{H}_2|i\rangle\langle i|\hat{H}_2| 1\rangle\label{aad1}\\
&+&\int_0^Tdt\tr\{-\mbox{i}e^{\mbox{i}H_1t}[\hat{H}_2,(1-f(t))\hat{H}_1]e^{-\mbox{i}H_1t}R(t)\}\nonumber\\
&=&-\sum_{i\neq 1}\frac{\lambda^2|\langle 1|\Delta H|i\rangle|^2}{\lambda_i}+\sum_{i\neq 1}\frac{4\lambda^2\sin^2(\lambda_iT/2)|\langle 1|\Delta H|i\rangle|^2}{T^2\lambda_i^3}\nonumber\\
&+&\int_0^Tdt\tr\{-\mbox{i}e^{\mbox{i}H_1t}[\hat{H}_2,(1-f(t))\hat{H}_1]e^{-\mbox{i}H_1t}R(t)\},\label{ad10}
\end{eqnarray}
by noting that
\begin{eqnarray}\label{ad11}
\sum_{i\neq 1}\frac{|\langle 1|\hat{H}_2|i\rangle|^2}{\lambda_i}=\sum_{i\neq 1}\frac{|\langle 1|H_1+\lambda\Delta H-\omega^{'}_1|i\rangle|^2}{\lambda_i}=\sum_{i\neq 1}\frac{\lambda^2|\langle 1|\Delta H|i\rangle|^2}{\lambda_i}.
\end{eqnarray}
Moreover, by the definition of the notation $O(\cdot)$ in Section~\ref{secdef} we can write $\sum_{i\neq 1}\frac{4\lambda^2\sin^2(\lambda_iT/2)|\langle 1|\Delta H|i\rangle|^2}{T^2\lambda_i^3}=O(\frac{\lambda^2}{T^2\lambda_2^3})$.
Denote $\bar{H}=\max_{f(t)\in(0,1)}||H(t)||$. Since
\begin{eqnarray}\label{ad12}
||\int_0^Tdt\tr\{-\mbox{i}e^{\mbox{i}H_1t}[\hat{H}_2,(1-f(t))\hat{H}_1]e^{-\mbox{i}H_1t}R(t)\}||\leq\frac{T\lambda}{2}\bar{H}^2||R(t)||
\end{eqnarray}
is $O(\lambda^3)$, we can further write (\ref{ad10}) as
\begin{equation}\label{twoerror}
\langle H_2\rangle_{\rho_T}-\langle H_2\rangle_{\rho_g}=-\sum_{i\neq 1}\frac{\lambda^2|\langle 1|\Delta H|i\rangle|^2}{\lambda_i}+\sum_{i\neq 1}\frac{4\lambda^2\sin^2(\lambda_iT/2)|\langle 1|\Delta H|i\rangle|^2}{T^2\lambda_i^3}+O(\lambda^3).
\end{equation}
Next we will calculate $\omega_1^{'}-\langle 1|H_2|1\rangle$. We have
\begin{equation}\label{ad13}
\langle 1|H_2|1\rangle=\langle 1|H_1+\lambda\Delta H|1\rangle=\omega_1+\lambda\langle 1|\Delta H|1\rangle.
\end{equation}
The smallest eigenvalue $\omega_1^{'}$ of $H_2$ can be calculated using the first-order time-independent perturbation theory for non-degenerate system. Assume $H_1$ is the unperturbed Hamiltonian and the perturbation is $\lambda\Delta H$, then the lowest eigenvalue of the perturbed Hamiltonian $H_1+\lambda\Delta H$ can be written as series in terms of $\lambda$ and $\omega_1$ \cite{Griffiths95}:
\begin{equation}\label{ad14}
\omega_1^{'}=\omega_1+\lambda\langle 1|\Delta H|1\rangle-\lambda^2\sum_{i\neq 1}\frac{|\langle 1|\Delta H|i\rangle|^2}{\lambda_i}+O(\lambda^3).
\end{equation}
Thus we conclude
\begin{equation}\label{ad15}
\omega_1^{'}-\langle H_2\rangle_{\rho_g}=\omega_1^{'}-\langle 1|H_2|1\rangle=-\lambda^2\sum_{i\neq1}\frac{|\langle1|\Delta H|i\rangle|^2}{\lambda_i}+O(\lambda^3).
\end{equation}
Comparing (\ref{twoerror}) and (\ref{ad15}), the terms $-\lambda^2\sum_{i\neq1}\frac{|\langle1|\Delta H|i\rangle|^2}{\lambda_i}$ cancel and so the error $\epsilon$ is estimated by
\begin{eqnarray}\label{prop1error}
\epsilon&=&\sum_{i\neq 1}\frac{4\lambda^2\sin^2(\lambda_iT/2)|\langle 1|\Delta H|i\rangle|^2}{T^2\lambda_i^3}+O(\lambda^3)\nonumber\\
&=&O(\frac{\lambda^2}{T^2\lambda_2^3}),\quad \lambda\rightarrow0.
\end{eqnarray}
\end{proof}
\section{Error Estimation for Arbitrary Interpolations}\label{secnonlinear}
The approach derived in the previous section can be easily generalized for
arbitrary given continuous interpolating functions. The generalization can
simply be done by replacing the linear interpolation function with the
given continuous function $f(t)$ and then recalculating the double
integration
\begin{equation}
A_i(T)=-2\int_0^Tdt\int_0^tdt^{'}(1-\frac{t}{T})\lambda_if(t^{'})\cos(\lambda_i(t^{'}-t))
\end{equation}
in (\ref{ft}). The error estimation can easily be obtained from the proof of Proposition \ref{theorem1}:
\begin{prop}\label{theorem2}
For an arbitrarily given $f(t)$, the error estimation is given by
\begin{equation}\label{thm2e}
\epsilon=\lambda^2\sum_{i\neq1}A_i(T)|\langle1|\Delta H|i\rangle|^2+\lambda^2\sum_{i\neq1}\frac{|\langle1|\Delta H|i\rangle|^2}{\lambda_i}+O(\lambda^3)
\end{equation}
as $\lambda\rightarrow0$.
\end{prop}
\begin{proof}
$\epsilon$ is still calculated by (\ref{int3}), using $\langle H_2\rangle_{\rho_T}-\langle H_2\rangle_{\rho_g}$ and $\omega_1^{'}-\langle H_2\rangle_{\rho_g}$. We have
\begin{equation}
\langle H_2\rangle_{\rho_T}-\langle H_2\rangle_{\rho_g}=\sum_{i\neq1}A_i(T)\lambda^2|\langle1|\Delta H|i\rangle|^2+O(\lambda^3)
\end{equation}
and
\begin{equation}
\omega_1^{'}-\langle H_2\rangle_{\rho_g}=-\lambda^2\sum_{i\neq1}\frac{|\langle1|\Delta H|i\rangle|^2}{\lambda_i}+O(\lambda^3).
\end{equation}
\end{proof}
It must be pointed out that $A(T)$ is very easy to calculate with the aid of any softwares that can perform symbolic integration, and therefore it is straightforward to apply Proposition \ref{theorem2} to find the error estimation for a given interpolating function, as we are going to do in the next section.
\section{Examples}\label{secexam}
\subsection{Linear Interpolation:$\ f(t)=t/T$}
By Proposition \ref{theorem1}, the error estimation is $\epsilon=\sum_{i\neq 1}\frac{4\sin^2(\lambda_iT/2)}{T^2\lambda_i^3}|\langle 1|\Delta H|i\rangle|^2\lambda^2+O(\lambda^3)$ as $\lambda\rightarrow0$. Since $\sin^2(\lambda_iT/2)$ and $\Delta H$ are bounded, this error term is primarily determined by $\frac{\lambda}{T}$ which is the average speed of the variation of the system Hamiltonian, and $\frac{1}{\lambda_i}$ which is the inverse of the energy gap between the ground and $i$-th eigenstates of $H_1$, as $\lambda\rightarrow0$. In particular, we have
\begin{equation}
\lim_{{\lambda}\rightarrow0}\frac{\epsilon}{(\frac{\lambda}{T})^2}=\sum_{i\neq 1}\frac{4\sin^2(\lambda_iT/2)}{\lambda_i^3}|\langle 1|\Delta H|i\rangle|^2.
\label{VU.lininterp}
\end{equation}
Therefore, when the inverse of the energy gaps $\frac{1}{\lambda_i}$ are fixed values, the approximation error $\epsilon$ is estimated to be proportional to the square of the average speed of the variation of the Hamiltonian, which is $(\frac{\lambda}{T})^2$, as $\lambda\rightarrow0$.
\subsection{Quadratic Interpolation:$\ f(t)=t^2/T^2$}
Replace $f(t)$ with a nonlinear function $f(t)=\frac{t^2}{T^2}$ in
(\ref{ft}) and we recalculate the integral to be
\begin{eqnarray}\label{nonlinear1}
\sum_{i\neq1}A_i(T)=\sum_{i\neq 1}(-\frac{1}{\lambda_i}+\frac{16\sin^2(\frac{T\lambda_i}{2})+4T^2\lambda_i^2-8T\lambda_i\sin(T\lambda_i)}{T^4\lambda_i^5})|\langle1|\Delta H|i\rangle|^2.\nonumber\\
\end{eqnarray}
By Proposition \ref{theorem2}, for sufficiently small $\lambda$, the error is estimated to be of order of $\lambda^2$:
\begin{eqnarray}
\epsilon_{quad}&=&\lambda^2\sum_{i\neq 1}\frac{16\sin^2(\frac{T\lambda_i}{2})+4T^2\lambda_i^2-8T\lambda_i\sin(T\lambda_i)}{T^4\lambda_i^5}|\langle1|\Delta H|i\rangle|^2+O(\lambda^3)\nonumber\\
&=&(\frac{\lambda}{T})^2\sum_{i\neq
1}[\frac{16\sin^2(\frac{T\lambda_i}{2})}{T^2\lambda_i^5}+\frac{4}{\lambda_i^3}-\frac{8\sin(T\lambda_i)}{T\lambda_i^4}]|\langle1|\Delta H|i\rangle|^2+O(\lambda^3).\nonumber\\
\label{VO.quadinterp}
\end{eqnarray}
That is, in contrast to the linear interpolation case, we have
\begin{equation}
\lim_{\lambda\rightarrow0}\frac{\epsilon_{quad}}{(\frac{\lambda}{T})^2}=\sum_{i\neq
1}[\frac{16\sin^2(\frac{T\lambda_i}{2})}{T^2\lambda_i^5}+\frac{4}{\lambda_i^3}-\frac{8\sin(T\lambda_i)}{T\lambda_i^4}]|\langle1|\Delta H|i\rangle|^2.
\end{equation}
This calculation shows that if the evolution speed is infinitely slow, then
the system dynamics is adiabatic during $t\in[0,T]$. However, the scaling
of $\epsilon_{quad}$ with respect of the square of the average evolution speed $\frac{\lambda}{T}$ is not as simple
as in the linear case, where the scaling of $\epsilon$ with respect of $(\frac{\lambda}{T})^2$ is primarily determined by the inverse of the energy gaps as $\lambda\rightarrow0$.
In the quadratic case, this scaling is primarily determined by a complex factor
$[\frac{16\sin^2(\frac{T\lambda_i}{2})}{T^2\lambda_i^5}+\frac{4}{\lambda_i^3}-\frac{8\sin(T\lambda_i)}{T\lambda_i^4}]$
which depends mainly on the inverse of the energy gaps $\{\lambda_i\}$ and the inverse of the evolution time $T$.
\subsection{Interpolation with Decaying Resonant Terms}
Here we assume a linear interpolating function with an additional
oscillating term that gradually decays to zero. That is,
\[
f(t)=\frac{t}{T}+g(1-\frac{t}{T})\sin(\lambda_ct),
\]
where $\lambda_c$ is the oscillating frequency of the
perturbation. Ortigoso observed in \cite{Ortigoso12} the inconsistency
in the applicability of the adiabatic theorem when the Hamiltonian contains
resonant terms whose amplitudes go asymptotically to zero.
Replace $f(t)$ with $f(t)=\frac{t}{T}+g(1-\frac{t}{T})\sin(\lambda_ct)$ in
(\ref{ft}) and we recalculate the integral to be
\begin{equation}\label{nonlinear4}
\sum_{i\neq1}A_i(T)=\sum_{i\neq 1}\frac{Q_1(g,T,\lambda_i,\lambda_c)}{T^2(2\lambda_i^{11}-8\lambda_i^9\lambda_c^2+12\lambda_i^7\lambda_c^4-8\lambda_i^5\lambda_c^6+2\lambda_i^3\lambda_c^8)}.
\end{equation}
$Q_1$ is a function of four parameters. In particular, we note that each term in (\ref{nonlinear4}) is well defined for all $\lambda_c$, including $\lambda_c=\lambda_i$, since as $\lambda_c\to\lambda_i$, the $i$-th term in (\ref{nonlinear4}) approaches
\begin{eqnarray}\label{nonlinear5}
&&[-128\sin^2(\frac{T\lambda_i}{2})-16g\sin(T\lambda_i)+8g\sin(2T\lambda_i)+g^2\nonumber\\
&+&g^2(2\sin^2(T\lambda_i)-1)-32T^2\lambda^2-16gT\lambda_i+2g^2T^4\lambda_i^4\nonumber\\
&+&16gT^2\lambda_i^2\sin(T\lambda_i)-16gT\lambda_i(2\sin^2(\frac{T\lambda_i}{2})-1)\nonumber\\
&-&2g^2T^2\lambda_i^2(2\sin^2(T\lambda_i)-1)-2g^2T\lambda_i\sin(2T\lambda_i)]/32T^2\lambda_i^3\nonumber\\
&=&-\frac{1}{\lambda_i}+\frac{g^2}{16}T^2\lambda_i+\frac{g}{2\lambda_i}\sin(T\lambda_i)-\frac{g^2}{16\lambda_i}(2\sin^2(T\lambda_i)-1)+Q(T),
\end{eqnarray}
where $Q(T)$ is a complicated fraction with $T$ being in its denominator. The error resulting
from the $i$-th term is given by
\begin{eqnarray}\label{counte}
&&\epsilon_i\nonumber\\
&=&|\langle1|\Delta H|i\rangle|^2[\frac{g^2T^4\lambda_i}{16}+\frac{gT^2\sin(T\lambda_i)}{2\lambda_i}-\frac{g^2T^2(2\sin^2(T\lambda_i)-1)}{16\lambda_i}+T^2Q(T)](\frac{\lambda}{T})^2\nonumber\\
&+&O(\lambda^3)
\end{eqnarray}
as $\lambda\rightarrow0$. We have
\begin{equation}
\lim_{\lambda\rightarrow0}\frac{\epsilon_i}{(\frac{\lambda}{T})^2}=|\langle1|\Delta H|i\rangle|^2[\frac{g^2T^4\lambda_i}{16}+\frac{gT^2\sin(T\lambda_i)}{2\lambda_i}-\frac{g^2T^2(2\sin^2(T\lambda_i)-1)}{16\lambda_i}+T^2Q(T)].
\label{VO.oscinterp}
\end{equation}
The scaling of $\epsilon_i$ with respect of $(\frac{\lambda}{T})^2$ is additionally determined by $T^2$ and $T^4$, as compared to the quadratic case. This is
where adiabatic approximation error may not be small if the average evolution speed is slow. In particular by (\ref{VO.oscinterp}), if one chooses a comparably large value for $T$ in an adiabatic evolution experiment,
the adiabatic approximation error may not decrease as expected when one applies a slow evolution speed $\frac{\lambda}{T}$.
In order to further illustrate this point, we can heuristically compare the speed of convergence of $\epsilon$ to zero observed in this case and the quadratic case, as the speed of the adiabatic process ($\lambda/T$) reduces and the evolution horizon $T$ increases. The difference in the speed of convergence can be clearly seen using the ratio
\begin{equation}\label{ratio}
\lim_{T\to\infty}\left(\lim_{(\lambda/T)\to 0}\frac{\epsilon_i}{\epsilon_{quad}}\right)=\infty.
\end{equation}
Therefore, the rate of convergence considered in this subsection is slower than that in the quadratic
or linear case. i.e., $\epsilon$ goes to zero as $\lambda\to 0$ at a much slower
rate than in the linear interpolation case or the quadratic interpolation case
if $T$ is large. Furthermore, the larger $T$ is, the slower
the convergence.
\section{Conclusion}\label{seccon}
In this paper we provide a rigorous analysis of the time-dependent
evolution of Hamiltonian-varying quantum systems. As we calculated, the adiabatic approximation error
is not proportional to the average speed of the variation of the system Hamiltonian and the inverse
of the energy gaps in many cases. The results in this paper may provide guidelines when applying complicated interpolation for adiabatic evolution.
|
\section{Introduction}
\label{sec:intro}
Materials such as turbid water, white paint, and egg shells are considered opaque because multiple scattering by the randomly placed constituent scatterers in the medium frustrates the passage of light \cite{ishimaru1999wave}. The seminal papers by Dorokhov \cite{dorokhov1982transmission}, Barnes and Pendry \textit{et al. } \cite{pendry1990maximal,barnes1991multiple}, and others \cite{mello1988macroscopic,beenakker2009applications} postulate that even if a normally incident wavefront barely propagates through a thick slab of such media, there will generically exist a few highly-transmitting wavefronts that will propagate through the slab with a transmission coefficient close to $1$, \textit{i.e}, they will be nearly perfectly-transmitting. These perfectly transmitting eigen-wavefronts are the right singular vectors of the modal transmission matrix and are optimized to the specific random medium.
These seminal papers inspired the breakthrough experiments by Vellekoop and Mosk \cite{vellekoop2008phase,vellekoop2008universal}, and others \cite{popoff2010measuring,kohlgraf2010transmission,shi2010measuring,kim2012maximal,van2011optimal,aulbach2011control,cui2011high,cui2011parallel,stockbridge2012focusing} provide credence to the hypothesis that there generally exist (nearly) perfectly transmitting eigen-wavefronts in highly scattering random media composed of a larger number of non-absorbing scatterers. Recently, we verified this hypothesis \cite{jin2012iterative,cjin2013} for 2-D systems with periodic boundary conditions composed of hundreds of thousands of non-absorbing scattering using numerically rigorous simulations.
The perfect-transmission supporting universal transmission coefficient distribution postulated by Dorokhov, Mello, Pereyra, Kumar, Pendry, and Barnes \cite{dorokhov1982transmission,pendry1990maximal,barnes1991multiple,mello1988macroscopic}, was derived assuming that the medium was deep enough so that the scattering matrix obeyed a physically consistent (\textit{i.e.} obeying reciprocity and time-reversal conditions) maximum-entropy law. Their analysis does not provide a principled and mathematically grounded framework for reasoning about whether, when or the sense in which a deep medium composed of a large number of randomly placed point-like scatterers with an arbitrary distribution of refractive indices can be expected to have a perfect transmission-supporting transmission coefficient distribution.
In this paper, we use modern random matrix theory to revisit the problem of predicting the transmission coefficient distribution of 2-D random media with periodic boundary conditions composed of a larger number of randomly placed point-like scatterers with an arbitrary refractive index distribution. We provide a characterization of the transmission coefficient distribution that explicitly depends on the refractive index distribution, the number of propagating modes and the depth of the medium for layered random media (in a sense we will make precise) composed of a large number of point-like scatterers.
The critical part of our derivation relies on the development of an isotropic random matrix model for the modal transfer matrix of a single randomly placed point-like scatterer. The random transfer matrix has singular value distribution that matches the singular values of the physical transfer matrix of a randomly-placed point-like scatterer. However, the left and right singular vectors of our random transfer matrix construction are modeled as independent and isotropically random. This allows us to use tools from free probability theory to approximate the transmission coefficient distribution of a layered random media composed of layers containing point-like scatterers.
We show that the derived distribution agrees remarkably well with results obtained using a numerically rigorous spectrally convergent simulation that utilizes spectrally accurate methodologies. This justifies the use of our isotropic model for reasoning about the properties of the derived distribution. Analysis of the resulting distribution brings into sharp focus the universal, \textit{i.e.}, scatterer-property independent, aspects of the distribution and provides the strongest principled justification yet of why we should expect perfect transmission in such deep random media regardless of the refractive index distribution of the constituent scatterers. The analysis brings into focus a sparsity condition under which random media can be expected to exhibit a perfect transmission-supporting universal transmission coefficient distribution in the deep medium limit.
We describe the setup and define the transmission coefficient distribution in Section \ref{sec:Setup} . We highlight some pertinent properties of the system modal transfer matrix in Section \ref{sec:transfer matrix}, and employ them in Section \ref{sec:transfer matrix model} to formulate a isotropically random model for the transfer matrix of a single point-like scatterer. In Section \ref{sec:tcoeff distribution}, we describe the pertinent free probabilistic tools from random matrix theory that allow us to analytically characterize the limiting transmission coefficient distribution of a medium composed of many scatterers from the eigen-distribution of the isotropic transfer matrix of a single point-like scatterer. We analyze the properties of the limiting transmission coefficient distribution thus obtained in Section \ref{sec:tcoeff properties}, and bring into sharp focus its universal, \textit{i.e.}, scatterer property independent, aspects. We validate our theoretical predictions using numerically rigorous simulations in Section \ref{sec:NumSim}. Details of some computations have been relegated to the Appendix.
\section{Setup}
\label{sec:Setup}
\begin{figure}[h]
\centering
\scalebox{0.5}{\input{ScatSysTCD5.pspdftex}}
\caption{Setup.}
\label{fig:setup}
\end{figure}
We study scattering from a two-dimensional (2D) random slab of thickness $L$ and periodicity $D$; the slab's unit cell occupies the space $0 \leq x < D$ and $0 \leq y < L$ (Fig. \ref{fig:setup}). The slab contains $\Nlay$ infinite and $z$-invariant circular cylinders of radius $r$ that are placed randomly within the cell, as described shortly. The cylinders are assumed to be dielectric with refractive index $n_{d}$; care is taken to ensure the cylinders do not overlap. The radius of the cylinders is chosen to be much smaller than the wavelength $\lambda$ so they, in effect, act like point scatterers.
For $i_c = 1, 2, \ldots, \Nlay$, the $x$ and $y$ position of the center of the $i_c$-th cylinder are $u_{x,i_c}, u_{y,i_c} + (i_c -1) \ell)$, respectively where $u_{x,i_c}$ and $u_{y,i_c}$ s are i.i.d, random variables with uniform distribution on $[r,D-r]$ and $[r,\ell -r]$, respectively. Here $\ell= L/\Nlay$ is depth of each ``layer''; $\ell$ is chosen to be larger than $\sqrt{D \lambda}$. Each cylinder's refractive index $n_{i_c}$ is drawn independently from the same distribution of refractive indices $\eta(n)$.
Fields are $\sf TM_{\mbox{$z$}}$ polarized: electric fields in the $y<0$ $(i=1)$ and $y>L=\Nlay\,\ell$ $(i=2)$ halfspaces are denoted $\underline{e}_{i}(\underline{\rho})=e_{i}(\underline{\rho})\hat{z}$. The field amplitude $e_{i}(\underline{\rho})$ can be decomposed in terms of $+y$ and $-y$ propagating waves as $e_{i}(\underline{\rho}) = e_{i}^{+}(\underline{\rho}) + e_{i}^{-}(\underline{\rho})$, where
\begin{equation}
\label{eq:IncidentWave}
e^{\pm}_{i}(\underline{\rho}) = \displaystyle \sum_{n=-N}^{N} h_{n} a^{\pm}_{i,n} e^{-j\underline{k}^{\pm}_{n} \cdot \underline{\rho}}\,.
\end{equation}
In the above expression, $\underline{\rho}=x\hat{x}+y\hat{y}\equiv(x,y)$, $\underline{k}^{\pm}_{n} = k_{n,x}\hat{x} \pm k_{n,y}\hat{y} \equiv (k_{n,x},\pm k_{n,y})$, $k_{n,x}=2\pi n/D$, $k_{n,y} = 2\pi \sqrt{(1/\lambda)^{2} - (n/D)^{2}}$, $\lambda$ is the wavelength, and $h_{n}=\sqrt{\| \underline{k}^{\pm}_{n} \|_{2} / k_{n,y}}$ is a power-normalizing coefficient; a time dependence $e^{j\omega t}$ is assumed and suppressed. We assume $N=\lfloor D/\lambda \rfloor$, \textit{i.e. } we only model propagating waves and denote $M=2N+1$. The modal coefficients $a^{\pm}_{i,n}$, $i=1,2$; $n=-N,\ldots,N$ are related by the scattering matrix
\begin{equation}\label{eq:scat matrix}
\left[\begin{array}{c}\aleftm\\\arightp\\\end{array}\right] = \underbrace{\left[ \begin{array}{cc} S_{11} & S_{12} \\ S_{21} & S_{22} \end{array} \right]}_{=:S} \left[\begin{array}{c}\aleftp\\\arightm\\\end{array}\right],
\end{equation}
where $\underline{a}^{\pm}_{i} = \begin{bmatrix} a^{\pm}_{i,-N} & \ldots a^{\pm}_{i,0} & \ldots a^{\pm}_{i,N}\end{bmatrix}^{T}$. In what follows, we assume that the slab is only excited from the $y<0$ halfspace; hence, $\arightm=0$. For a given incident field amplitude $e^{+}_{1}(\underline{\rho})$, we define the transmission coefficient as \begin{align}
\label{eq:tcoeff vec}
\tau(\aleftp) := \dfrac{\| S_{21}\cdot \aleftp \|_{2}^{2}}{\| \aleftp \|_{2}^{2}}.
\end{align}
We denote the transmission coefficient of a normally incident wavefront by $\tnorm = \tau( \begin{bmatrix} 0 & \cdots & 1 & \cdots &0 \end{bmatrix}^{T} )$; here $^{T}$ denotes transposition.
\subsection{The transmission coefficient distribution}\label{sec:tcoeff}
The problem of designing an incident wavefront $\aopt$ that maximizes the transmitted power can be stated as
\begin{equation}\label{eq:optimization problem}
\aopt= \argmax_{\aleftp} \tau(\aleftp) = \argmax_{\aleftp} \dfrac{\| S_{21}\cdot \aleftp \|_{2}^{2}}{\| \aleftp \|_{2}^{2}} = \argmax_{\parallel \aleftp\parallel_{2} = 1} \| S_{21} \cdot \aleftp \|_{2}^{2}
\end{equation}
where $\parallel \aleftp \parallel_{2} = 1$ represents an incident power constraint.
Let $S_{21}= \sum_{i=1}^{M} \sigma_{i}\, \underline{u}_i \cdot \underline{v}_{i}^H$ denote the singular value decomposition (SVD) of $S_{21}$; $\sigma_{i}$ is the singular value associated with the left and right singular vectors $\underline{u}_i$ and $\underline{v}_i$, respectively. By convention, the singular values are arranged so that ${\sigma}_{1} \geq \ldots \geq {\sigma}_{M}$ and $^H$ denotes complex conjugate transpose. Then via a well-known result for the variational characterization of the largest right singular vector \cite[Theorem 7.3.10]{horn1990matrix} we have that
\begin{equation}\label{eq:aopt s21}
\aopt = {\underline{v}}_{1}.
\end{equation}
When the optimal wavefront $\aopt$ is excited, the transmitted power is $ \topt :=\tau(\aopt) = \sigma_{1}^{2}$. When the wavefront associated with the $i$-th right singular vector $\underline{v}_{i}$ is transmitted, the transmitted power is $\tau_{i} := \tau(\underline{v}_{i}) = \sigma_{i}^{2}$, which we refer to as the transmission coefficient of the $i$-th eigen-wavefront of $S_{21}$. We are interested in the limiting transmission coefficient distribution whose p.d.f. is defined as
\begin{equation}\label{eq:pdf tau}
f(\tau) = \lim_{M, \Nlay \to \infty} \mathbb{E} \left[\dfrac{1}{M} \sum_{i=1}^{M} \delta\left(\tau-\tau(\underline{v}_{i})\right) \right] = \lim_{M, \Nlay\to \infty} \mathbb{E}\left[\dfrac{1}{M} \sum_{i=1}^{M} \delta\left(\tau-\sigma_{i}^2)\right)\right],
\end{equation}
where we assume that $\Nlay/M \to c \in (0, \infty)$ as $M, \Nlay \to \infty$. The Dorokhov-Mello-Pereyra-Kumar (henceforth, DMPK) distribution \cite{dorokhov1982transmission,mello1988macroscopic} has density given by
\begin{equation}\label{eq:dmpk}
f_{\sf DMPK} (\tau) = \dfrac{l_{\sf free}}{2L} \dfrac{1}{\tau \sqrt{1-\tau}}, \qquad \textrm{ for } 4 \exp(-L/2l_{\sf free}) \lessapprox \tau \leq 1,
\end{equation}
where $l_{\sf free}$ is the mean-free path in the medium. The DMPK distribution is posited \cite{dorokhov1982transmission,pendry1990maximal,barnes1991multiple,mello1988macroscopic,beenakker2009applications}
to be the \textit{universal limiting distribution} for systems comprised of many scatterers in the limit where $L \gg M$.
Assuming a scattering regime where the DMPK distribution holds, Eq. (\ref{eq:dmpk}) predicts the existence of highly-transmitting eigen-wavefronts that achieve (nearly) perfect transmission. Since the DMPK distribution was derived under a maximum-entropy type assumption (which we shall revisit shortly), the material properties of the scatterers, such as the distribution of refractive indices, do not explicitly appear in the expression in Eq. (\ref{eq:dmpk}) for its p.d.f. but instead are encoded implicitly via the $l_{\sf free}$ parameter. Our objective is to theoretically predict $f(\tau)$ in Eq. (\ref{eq:pdf tau}) and explicitly characterize its dependence on the refractive index distribution $\eta(n)$, $\Nlay$, and $M$, assuming we are in a regime where each scatterer is small enough so that it effectively acts as an isotropic point scatterer. Our mathematically-derived framework permits reasoning about the conditions under which we might expect a universal limiting distribution and the existence of the (nearly) perfectly transmitting eigen-wavefronts.
\section{Background: the transfer matrix and its pertinent properties}\label{sec:transfer matrix}
The scattering matrix $S$ in Eq. (\ref{eq:scat matrix}) describes the relationship between the modal coefficients of incoming and outgoing waves. Rearranging the terms in Eq. (\ref{eq:scat matrix}) relates the modal coefficients in $i=1$ and $i=2$ halfspace via the transfer matrix $T$
\begin{equation}
{ \begin{bmatrix} \underline{a}_{2}^{+} \\ \\ \underline{a}_{2}^{-} \end{bmatrix} = \underbrace{\begin{bmatrix} S_{21} - S_{22} \cdot S_{12}^{-1}\cdot S_{11} & S_{22} \cdot S_{12}^{-1} \\ & \\ - {{S_{12}}}^{-1} \cdot S_{11} & S_{12}^{-1} \end{bmatrix}}_{=: \, T} \cdot \begin{bmatrix} \underline{a}_{1}^{+} \\ \\ \underline{a}_{1}^{-} \end{bmatrix}},
\end{equation}
where we have assumed that the $S_{12}$ matrix is invertible. Rewriting the transfer matrix as
\begin{equation}\label{eq:Tmatrix block}
T = \begin{bmatrix} S_{22} & 0 \\ 0 & S_{12}^{-1} \end{bmatrix} \cdot \begin{bmatrix} S_{22}^{-1} \cdot S_{21} \cdot S_{11}^{-1} - S_{12}^{-1} & I \\
- I & S_{12} \end{bmatrix} \cdot \begin{bmatrix} S_{11} & 0 \\ 0 & S_{12}^{-1} \end{bmatrix},
\end{equation}
allows us to easily verify that $\det(T) = \det(T^{H}\cdot T) = 1$. In the lossless setting when $S^{H} \cdot S = I$, and $S_{12}$ is invertible, it is shown in Appendix \ref{sec:appendix transfer} that the $2M$ eigenvalues of $T^{H}\cdot T$ denoted by $\lambda_{1} \geq \ldots \geq \lambda_{2M}$ are
\begin{equation} \label{eq:tth eigs} \lambda_{i} = \dfrac{2 - \tau_i + 2 \sqrt{1 - \tau_i}}{\tau_i}\,\,\, \textrm{and}\,\, \lambda_{2M - i + 1} = \dfrac{2 - \tau_i - 2 \sqrt{1 - \tau_i}}{\tau_i} \qquad \textrm{ for } i = 1, \ldots, M.
\end{equation}
Note that $\lambda_{i} \cdot \lambda_{2M - i + 1} = 1$ so that the $2M$ eigenvalues of $T^{H}\cdot T$ come in reciprocal pairs. From Eq. (\ref{eq:tth eigs}), we have that
$$ \lambda_{i} + \lambda_{2M - i + 1} = \dfrac{4}{\tau_i} - 2,$$
so that
\begin{equation}\label{eq:taui lambdai}
\tau_{i} = \dfrac{4}{\lambda_{i} + 1/\lambda_{i} + 2}.
\end{equation}
Substituting $\lambda_{i} = \exp(2 \,x_i)$ in Eq. (\ref{eq:taui lambdai}) yields
$$ \tau_i = \dfrac{4}{\exp(2 x_i) + \exp(-2 x_i) + 2} = \dfrac{1}{ (\exp(x_i) + \exp(-x_i)/2)^2} = \dfrac{1}{\cosh^2(x_i)}.
$$
Equivalently, since $x_{i} = 0.5 \, \ln \lambda_{i}$, we have
\begin{equation}\label{eq:conversion}
\tau_i = \dfrac{1}{\cosh^2(0.5 \ln \lambda_i)} \leftrightarrow \lambda_i = \exp( 2 \cosh^{-1}(1/\sqrt{\tau_i})),
\end{equation}
and we have obtained a direct relationship between the eigenvalues of $T^H\cdot T$ and the transmission coefficients. Let $h(\lambda)$ denote the limiting eigenvalue distribution of the transfer matrix defined as
\begin{equation}\label{eq:pdf lambda}
h(\lambda) = \lim_{M, \Nlay \to \infty} \mathbb{E}\left[ \dfrac{1}{2M} \sum_{i=1}^{2M} \delta\left(\lambda-\lambda_i\right)\right].
\end{equation}
Then, a direct consequence of Eq. (\ref{eq:conversion}) is that once we know $h(\lambda)$, a simple change of variables yields the transmission coefficient distribution $f(\tau)$ as
\begin{align}\label{eq:ftau1}
f(\tau) & = h(\lambda) \dfrac{1}{{| \partial \tau}/{\partial \lambda |}} \bigg|_{\lambda \textrm{ in Eq. } (\ref{eq:conversion}) } = \,\, h(\lambda) \dfrac{(\lambda+1)^{3}}{4 |\lambda-1|} \bigg|_{\lambda = \exp( 2 \cosh^{-1}(1/\sqrt{\tau}))}.
\end{align}
Since the eigenvalues of $T^{H}\cdot T$ come in reciprocal pairs, $h(\lambda)$ for $\lambda \leq 1$ uniquely determines $h(\lambda)$ for $\lambda >1$. Thus we can rewrite Eq. (\ref{eq:ftau1}) as
\begin{equation}\label{eq:ftau2}
f(\tau) = 2 h(\lambda) \, \mathbb{I}_{\lambda \leq 1} \dfrac{(\lambda+1)^{3}}{4 | \lambda-1 |} \bigg|_{\lambda = \exp( 2 \cosh^{-1}(1/\sqrt{\tau}))},
\end{equation}
where $\mathbb{I}_{\lambda \leq 1}$ denotes the indicator function on the set $\lambda \leq 1$. From Eq. (\ref{eq:taui lambdai}), we have that
$$ \dfrac{1}{\tau} = \dfrac{(\lambda+1)^{2}}{4\,\lambda} \quad \textrm{ and } \qquad \dfrac{1}{{1-\tau}} = \dfrac{(\lambda+1)^{2}}{(\lambda-1)^{2}},
$$
so that rearranging terms on the right hand side of Eq. (\ref{eq:ftau2}), yields
\begin{align}\label{eq:ftau conversion}
f(\tau) & =\dfrac{1}{\tau \sqrt{1- \tau}} \cdot \{ 2 \, h(\lambda) \lambda \, \mathbb{I}_{\lambda \leq 1}\}\bigg|_{\lambda = \exp( 2 \cosh^{-1}(1/\sqrt{\tau}))}.
\end{align}
We note that Eq. (\ref{eq:ftau conversion}) is an exact relationship between the eigenvalue distribution of the transfer matrix and the transmission coefficient distribution. Comparing Eqs. (\ref{eq:ftau conversion}) and (\ref{eq:dmpk}) reveals the important insight that the DMPK distribution arises under the assumption that in the limit of deep random media, $h(\lambda) = l_{\sf free}/(4 L \lambda) $, or equivalently, that $h(\lambda)$ is a \textit{log-uniform distribution}. This is the \textit{maximum-entropy assumption} that yields the DMPK distribution for deep random media. Our goal is to analytically characterize $h(\lambda)$ and hence $f(\tau)$, via Eq. (\ref{eq:ftau conversion}) as a function of $\Nlay$, $M$ and the refractive index distribution of the scatterers for the setup in Fig. \ref{fig:setup}.
\section{An isotropically random model for the transfer matrix of a single point-like scatterer}\label{sec:transfer matrix model}
Let $T_{i}$ denote the transfer matrix of a layer containing a single scatterer (Fig. \ref{fig:setup}) and let $S^{(i)}$, $S_{11}^{(i)}$, $S_{22}^{(i)}$, and $S_{21}^{(i)}$ denote its scattering matrix and subblocks thereof, respectively. When the scatterers are point-like and $D$ is large, then $S_{11}$ and $S_{22}$ are well approximated by a rank one matrix whose largest singular value $\alpha\in[0,1)$ we will refer to as the scattering strength. This is obviously true for $D \to \infty$, and remains remarkably accurate for smaller $D$ as well. Since $S$ is unitary for lossless media, we have that $S_{11}^{H}\cdot S_{11} + S_{21}^{H} \cdot S_{21} = I$. Hence, the $S_{21}$ matrix must have an SVD of the form
\begin{align*}
S_{21}^{(i)} = U_i \cdot \Diag (1,\ldots, 1, \sqrt{1-\alpha^{2}}) \cdot V_i^{H},
\end{align*}
where $U_{i}$ and $V_{i}$ are the left and right singular vectors of $S_{21}^{(i)}$, which encode the physics of the scattering system. Consequently, by Eq. (\ref{eq:tcoeff vec}), the transmission coefficients of $S_{21}^{(i)}$ are approximately
$$\tau_{1} \approxeq \ldots \approxeq \tau_{M-1} \approxeq 1, \textrm{ and } \tau_{M} \approxeq 1-\alpha^{2}.$$
From Eq. (\ref{eq:tth eigs}), we can conclude that the $2M-2$ eigenvalues of $T^{H}\cdot T$ will equal one. The remaining two eigenvalues will $\lambda_{M}$ and $\lambda_{M+1} = 1/\lambda_{M}$ where
\begin{equation}\label{eq:theta alpha}
\lambda_{M} = \dfrac{2- (1-\alpha^{2}) + 2 \sqrt{1-(1-\alpha^2)}}{1-\alpha^{2}} = \dfrac{1+ \alpha^{2} + 2\,\alpha}{(1-\alpha)(1+\alpha)} = \dfrac{1+\alpha}{1- \alpha}=: \theta.
\end{equation}
This implies that the transfer matrix will have an SVD of the form
\begin{equation}\label{eq:transfer matrix svd}
T_{i} = \widetilde{U}_{i} \, {\rm diag}(\underbrace{1, 1, \cdots, 1, 1}_{2M - 2 \, \textrm{entries}}, \sqrt{\theta}, 1/\sqrt{\theta}) \, \widetilde{V}_{i}^{H},
\end{equation}
where $\widetilde{U}_{i}$ and $\widetilde{V}_{i}$ are the left and right singular vectors of $T_{i}$, which again encode the physics of the scattering systems. The refractive index distribution $\eta(n)$ induces a distribution $f_\theta(t)$ on $\theta$ which we assume to known and obtained either a via a change of variables as
\begin{equation}\label{eq:point change of variables}
\alpha \approx 9 \sqrt{\dfrac{ r^{4} (2\pi/\lambda)^{3} \pi^2 (n_d^2 - 1)^2}{16 D}},
\end{equation}
under a point scatterer assumption for the large $D$, $r \ll \lambda$, and $n_d \approx 1 $ regime, or using computational electromagnetic techniques.
The transfer matrix of the entire system in Fig. \ref{fig:setup} is obtained from those of the layers as
\begin{equation}\label{eq:transfer product}
T = \prod_{i=1}^{\Nlay} T_{i}.
\end{equation}
Each of the transfer matrices are independent and identically distributed. Fig. \ref{fig:incoherence} plots the expected values of the squared magnitude of the (bistochastic) correlation matrix formed by the inner product of the right singular vectors of a transfer matrix associated with a single randomly placed scatterer and the left singular vectors of an independent transfer matrix associated with another randomly placed scatter, averaged over $100,000$ independent realizations. If the singular vectors were independent and isotropically random (or Haar distributed) then we would get an empirically averaged matrix with all of its entries close to $1/M$. From Fig. \ref{fig:incoherence}, we can conclude that the singular vectors of two independent transfer matrices are not isotropically random with respect to each other. However, most of the entries of the correlation matrix have entries `close' to $1/M$.
Very recently, Anderson and Farrell \cite{anderson2014asymptotically} rigorously showed that the product of independent (Hermitian) random matrices with independent eigenvectors having a correlation matrix whose entries have squared magnitude entries exactly equal to $1/M$ will have the same limiting distribution as the product of independent random matrices with the same eigenvalue distribution but isotropically random eigenvectors. Here too, we have a situation where we are interested in analyzing the singular value distribution of products of random matrices with independent singular vectors. However, the correlation matrix of the singular vectors has entries whose squared magnitude is not exactly $1/M$, as would be the case if the left and right singular vectors were isotropically random, but instead \textit{close to} $1/M$. This leads to our conjecture that the singular value distribution of the matrix in Eq. (\ref{eq:transfer product}) can be `well approximated by' the singular value distribution of independent random matrices with the same per-matrix singular value distribution but isotropically random left and right singular vectors.
Motivated by this conjecture, we now consider an isotropic random matrix model for the transfer matrices whose singular values are specified by Eq. (\ref{eq:transfer matrix svd}) but whose left and right singular vectors are independent and isotropically random. We then use tools from free probability theory to analytically characterize the transmission coefficient distribution that arises due to this isotropic model for the transfer matrix of a point-like scatterer. Numerically rigorous physical simulations in Section \ref{sec:NumSim} will validate our conjecture. A mathematically rigorous treatment of this conjecture, including a quantification of the approximation error, remains an open problem.
\begin{comment}
\cite[Section 15,4, pp. 316]{van2012light} for very thin cylinders. Case I has E field parallel to cylinder. Case II has H field parallel to cylinder.
$$ Q_{1} = \dfrac{\pi^{2}}{8} \cdot \left(\dfrac{2\,\pi\,r}{\lambda}\right)^3 \cdot (n^2-1)$$
$$ Q_{2} = \dfrac{\pi^{2}}{4} \cdot \left(\dfrac{2\,\pi\,r}{\lambda}\right)^3 \cdot \dfrac{(n^2-1)^{2}}{(n^2+1)^{2}}$$
\textcolor{red}{Justify the random matrix assumption here.}
\end{comment}
\begin{figure}[H]
\centering
\label{fig:incoherence}
\subfloat[3D plot.]{
\includegraphics[width=0.5\textwidth]{TSVC.png}
}
\subfloat[Top view.]{
\includegraphics[width=0.5\textwidth]{TSVCpcolor.png}
}
\caption{Relationship between singular vectors plotted in $10\log_{10}$ scale. The absolute value squared of the correlation matrix between the singular vectors averaged over 100000 trials. The settings were $n=3.1, \alpha=0.9, \theta = 0.053, r = 0.05\lambda, \ell = 11.66\lambda, D = 50.43\lambda, M=101$.}
\end{figure}
\section{Analytically characterizing the transmission coefficient distribution}\label{sec:tcoeff distribution}
We now discuss some preliminaries required to compute the transmission coefficient distribution under the isotropic transfer matrix assumption. Let $X_M$ be an $M \times M$ symmetric (or Hermitian) random matrix whose ordered eigenvalues are denoted by $\lambda_{1}(X_M) \geq \cdots \geq \lambda_{n}(X_M)$. Let $h_{X_{M}}$ be the empirical eigenvalue distribution, \textit{i.e. }, the probability distribution with density
\[
h_{X_{M}}(z) = \frac{1}{M} \sum_{i=1}^{M} \delta\left(z-\lambda_{i}(X_M)\right).
\]
Now suppose that $A_M$ and $B_M$ are two independent $M \times M$ matrices whose empirical eigenvalue distributions converge as $M \longrightarrow \infty$ to non-random distributions having densities $h_{A}$ and $h_{B}$, respectively. A natural question then is: how is the limiting eigenvalue distribution of the matrix $B_{M}^{H} \cdot A_{M}^{H} \cdot A_M \cdot B_M $ related to the limiting eigenvalue distributions of $A_{M}$ and $B_{M}$?
Free probability theory \cite{voiculescu1991limit,dv1992free} states that if we know $h_A$ and $h_B$ and the matrices $A_M$ and $B_M$ are \textit{asymptotically free}, we can compute the limiting eigenvalue distribution of $A_M \cdot B_M$ from the limiting eigenvalue distributions of $A$ and $B$. Specifically, in this setting, $h_{A \cdot B}$ is given by the free multiplicative convolution of $h_A$ and $h_B$, denoted by $h_{A} \boxtimes h_{B}$ which is computed as described next. We first define the {\it $S$-transform}\footnote{Denoted here by $\psi(\cdot)$ to avoid any confusion with the $S$ (or scattering) matrix.}, which is given by
\begin{equation}\label{eq:S transform}
\psi_X(z):= \dfrac{1+z}{z } \cdot \dfrac{1}{\xi_X^{-1}(z)},
\end{equation}
where
\begin{equation}\label{eq:xiz defn}
\xi_X(z) =\int\dfrac{t}{z-t} h_X(t) dt =-1 + z \,g_X(z),
\end{equation}
and
\begin{equation}
g_{X} (z) = \int\dfrac{1}{z-t} h_X(t) dt,
\end{equation}
is the Cauchy transform of $h_X$. Then $S$-transform of $h_{A\cdot B}$ is
\begin{equation}
\psi_{A\cdot B}(z)= \psi_{A}(z)\psi_{B}(z).
\end{equation}
Note, that given the Cauchy transform $g_X(z)$, we can recover the density via the inversion formula
\begin{equation}\label{eq:cauchy inversion}
h_X(z) = - \dfrac{1}{\pi} \lim_{\epsilon \to 0} \Imag \, g_{X}(z+j \,\epsilon).
\end{equation}
A sufficient condition for the asymptotic freeness of two random matrices is that their singular vectors are independent and isotropically random \cite{hiai2000semicircle}. Consequently, under the isotropic transfer matrix assumption, the transfer matrices of successive layers are asymptotically free, by construction. Hence, we can use free multiplicative convolution machinery to characterize the limiting eigenvalue distribution of the transfer matrix of a multi-layered scattering system as depicted in Fig. \ref{fig:setup}, since, by Eq. (\ref{eq:transfer product}), the composite transfer matrix is the product of $\Nlay$ independent (and asymptotically free) random transfer matrices each having independent, isotropically random left and right singular vectors and singular values given by Eq. (\ref{eq:transfer matrix svd}).
To that end, we first compute the empirical eigenvalue distribution of $T_{i}^{H} \cdot T_{i}$ which is
\begin{equation}\label{eq:h kequals1}
h_{i}(\lambda) = \left(1-\dfrac{2}{2M}\right) \delta(\lambda-1) + \dfrac{1}{2M} \delta(\lambda- \theta) + \dfrac{1}{2M} \delta(\lambda- 1/\theta).\end{equation}
Its Cauchy transform is given by
\begin{equation}
g_{i}(z) = \left(1-\dfrac{1}{M}\right) \dfrac{1}{z-1} + \dfrac{1}{2M} \left( \dfrac{1}{z-\theta}+ \dfrac{1}{z-1/\theta}\right),
\end{equation}
and
\begin{align}\label{eq:xiz}
\begin{split}
\xi_{i}(z) &= -1 + \left(1-\dfrac{1}{M}\right) \dfrac{z}{z-1} + \dfrac{1}{2M} \left( \dfrac{z}{z-\theta} + \dfrac{z}{z-1/\theta}\right)\\
& = -1 + \dfrac{z}{z-1} - \dfrac{1}{M} \left(\dfrac{z}{z-1} - \dfrac{0.5 \, z}{z-\theta} - \dfrac{0.5\,z}{z-1/\theta} \right)\\
& = \underbrace{\dfrac{1}{z-1}}_{=: \,\xi_{0}(z)} + \dfrac{1}{M} \underbrace{\left( \dfrac{0.5 \, z}{z-\theta} + \dfrac{0.5\,z}{z-1/\theta} -\dfrac{z}{z-1} \right)}_{=: \,\widetilde{\xi}(z)}.
\end{split}
\end{align}
Repeating the computation for the setting where the $\theta_i$'s are random with pdf $f_{\theta}(\cdot)$ yields
\begin{equation}\label{eq:xi general}
\widetilde{\xi}(z)= \displaystyle z \int \left[ \dfrac{0.5}{z-t} + \dfrac{0.5}{z-1/t} -\dfrac{1}{z-1} \right] f_{\theta}(t) dt .
\end{equation}
To compute $\psi_{i}(z)$ using Eq. (\ref{eq:S transform}) we need to compute $\xi_{i}^{-1}(z)$. A standard application of perturbation theory (see, e.g., \cite{schwartz1977classical}) yields
\begin{equation}\label{eq:finversion}
\xi_i^{-1}(z) = \xi_{0}^{-1}(z) - \dfrac{1}{M} \, \dfrac{\widetilde{\xi}(x)}{\partial_{x} \xi_{0}(x)} \bigg|_{x = \xi_{0}^{-1}(z)} + O\left(\dfrac{1}{M^{2}}\right).
\end{equation}
Substituting $\xi_{0}^{-1}(z) = (z+1)/z$ gives
\begin{equation}\label{eq:finversion}
\xi_i^{-1}(z) = \dfrac{z+1}{z} - \dfrac{1}{M} \, \dfrac{\widetilde{\xi}(x)}{\partial_{x} \xi_{0}(x)} \bigg|_{x = 1+1/z} + O\left(\dfrac{1}{M^{2}}\right),
\end{equation}
or equivalently
$$ \dfrac{z}{z+1} \, \xi_i^{-1}(z) = 1 +\dfrac{1}{M} \dfrac{z}{z+1} \cdot \dfrac{\widetilde{\xi}(1+\tfrac{1}{z})}{z^2}
+ O\left(\dfrac{1}{M^{2}}\right),$$
so that by Eq. (\ref{eq:S transform}),
$$ \psi_{i}(z) = 1 - \dfrac{1}{M} \dfrac{\widetilde{\xi}(1+\tfrac{1}{z}) }{z(z+1)}
+ O\left(\dfrac{1}{M^{2}}\right).
$$
Then,
\begin{align}
\psi_{h}(z) & = \prod_{i=1}^{\Nlay} \psi_{i}(z) =\left[ 1 - \dfrac{1}{M} \dfrac{\widetilde{\xi}(1+\tfrac{1}{z}) }{z(z+1)}
+ O\left(\dfrac{1}{M^{2}}\right) \right]^{\Nlay}
\end{align}
In the regime where $M, \Nlay \to \infty$ with $\Nlay/M \to c$ we obtain
\begin{align}\label{eq:psi final}
\begin{split}
\psi_{h}(z; c) & = \lim_{M \to \infty} \left[ 1 - \dfrac{1}{M} \dfrac{\widetilde{\xi}(1+\tfrac{1}{z}) }{z(z+1)}
+ O\left(\dfrac{1}{M^{2}}\right) \right]^{M \cdot \frac{\Nlay}{M}}\\
& = \left[\exp\left( - \dfrac{\widetilde{\xi}(1+\tfrac{1}{z}) }{z(z+1)}\right)\right]^{c} = \exp\left( -c \cdot \dfrac{\widetilde{\xi}(1+\tfrac{1}{z}) }{z(z+1)}\right).
\end{split}
\end{align}
We next discuss how to obtain the distribution from $\psi_{h}(z;c)$. Inserting $z=\xi_h(y)$ in Eq. (\ref{eq:S transform}), we get
\begin{equation}
\dfrac{1+\xi_h(y)}{\xi_h(y) } \dfrac{1}{y} = \psi_h(\xi_{h}(y)).
\end{equation}
Substituting in the expression for $\xi_{h}(y)$ from Eq. (\ref{eq:xiz defn}), we get
\begin{equation}
\dfrac{1-1 + y \,g_h(y)}{-1 + y \,g_h(y) } \dfrac{1}{y} = \psi_h(-1 + y \,g_h(y)).
\end{equation}
Therefore, we get the fixed-point equation
$$\dfrac{g_{h}(z)}{z g_{h}(z) - 1} = \psi_h(zg_{h}(z)-1).$$
Substituting Eq. (\ref{eq:psi final})
\begin{equation}
\label{eq:fixedptEq}
\dfrac{g_{h}(z)}{z g_{h}(z) - 1} = \exp\left[ -c \cdot \dfrac{\widetilde{\xi}(\frac{z g_h(z)}{z\,g_h(z) -1} ) }{z g_h(z)\,(z \,g_h(z)+1)}\right]
\end{equation}
The density $h(\lambda)$ can be recovered from the Cauchy transform $g_h(\lambda)$ using Eq. (\ref{eq:cauchy inversion}) after solving the fixed-point equation. The transmission coefficient distribution is then obtained by Eq. (\ref{eq:ftau conversion}). Note that $\widetilde{\xi}(z)$ in Eq. (\ref{eq:xi general}) explicitly encodes the portion of the limiting distribution that depends on the scatterer-dependent properties via $f_\theta(t)$, where $\theta$ is related to the scattering strength $\alpha$ of a single scatterer via Eq. (\ref{eq:theta alpha}) and $\alpha$ is related to the scatterer-dependent properties via Eq. (\ref{eq:point change of variables}).
\section{Properties of the limiting transmission coefficient distribution}\label{sec:tcoeff properties}
We now analyze the properties of the distributions characterized by Eq. (\ref{eq:fixedptEq}). The mean of $f(\tau)$ is
\begin{equation}\label{eq:mean}
\mathbb{E}[\tau] = \int \tau f(\tau) d\tau = \displaystyle \int \dfrac{4 }{\lambda + \lambda^{-1} + 2} h(\lambda) d\lambda = 4 \displaystyle\int \dfrac{\lambda}{(\lambda+1)^{2}} h(\lambda) d\lambda
\end{equation}
where we have used Eq. (\ref{eq:taui lambdai}) to express $\mathbb{E}[\tau]$ with respect to $h(\lambda)$. From Eq. (\ref{eq:xiz defn}), we note that
\begin{equation}\label{eq:dxiz}
\xi_{h}'(z) := \partial_{z} \xi_{h}(z) = - \displaystyle\int \dfrac{\lambda}{(z-\lambda)^{2}} h(\lambda) d\lambda.
\end{equation}
Thus by comparing the righthand sides of Eqs. (\ref{eq:mean}) and (\ref{eq:dxiz}), we have that
\begin{equation}\label{eq:mean2}
\mathbb{E}[\tau] = -4\, \xi_{h}'(-1).
\end{equation}
From the computation in Appendix \ref{sec:first moment}, we obtain the closed-form expression
\begin{align}
\begin{split}\label{eq:first moment closed form}
\mathbb{E}[\tau] = \dfrac{1}{1+c\, \underbrace{\displaystyle\int \left( \dfrac{1- t}{1 + t} \right)^2 f_{\theta}(t) dt}}_{=:B_{2}} = \dfrac{1}{1+cB_{2}}.
\end{split}
\end{align}
Here, we call $B_{2}$ the normalizing factor, and it represents the average scattering strength of a single layer. The normalizing factor can be used to homogenize two different materials by giving measures to calculate the effective lengths, and its specific usage will be discussed in Section \ref{sec:NumSim}. We now compute the second moment of $f(\tau)$, which is given by
\begin{equation}\label{eq:second moment}
\mathbb{E}[\tau^{2}] = \int \tau^{2} f(\tau) d\tau = \displaystyle \int \dfrac{16 }{(\lambda + \lambda^{-1} + 2)^{2}} h(\lambda) d\lambda = 16 \displaystyle\int \dfrac{\lambda^{2}}{(\lambda+1)^{4}} h(\lambda) d\lambda .
\end{equation}
Note that
$$ \xi_{h}''(z) := \partial_{z} \xi_{h}'(z) = -2 \displaystyle\int \dfrac{\lambda}{(\lambda-z)^{3}} h(\lambda) d\lambda,$$
and
$$\xi_{h}'''(z) := \partial_{z} \xi_{h}''(z) = -6 \displaystyle \int \dfrac{\lambda}{(\lambda-z)^{4}} h(\lambda) d\lambda,$$
so that
\begin{equation}\label{eq:xi second moment}
\dfrac{1}{6} \xi_{h}'''(z) - \dfrac{1}{2} \xi_{h}''(z) = \displaystyle\int \left[- \dfrac{\lambda}{(\lambda-z)^{4}}+ \dfrac{\lambda}{(\lambda-z)^{3}} \right] h(\lambda) d\lambda = \displaystyle\int \dfrac{\lambda^{2} - \lambda\,z- \lambda}{(\lambda-z)^{4}} h(\lambda) d\lambda.
\end{equation}
Comparing Eqs. (\ref{eq:second moment}) and (\ref{eq:xi second moment}) gives us the relationship
\begin{equation}\label{eq:second moment}
\mathbb{E}[\tau^{2}] = 16 \left[ \dfrac{1}{6} \xi_{h}'''(-1) - \dfrac{1}{2} \xi_{h}''(-1) \right].
\end{equation}
The closed-from expression for the second moment is lengthy and derived in Appendix \ref{sec:second moment}.
From Eq. (\ref{eq:Etau sq calc}) and Eq. (\ref{eq:mean3}) we obtain
\begin{subequations}
\begin{align}
\dfrac{\mathbb{E}[\tau^2]}{\mathbb{E}[\tau]} &= \dfrac{16}{-4} \dfrac{\dfrac{1}{6} \xi_{h}'''(-1) - \dfrac{1}{2} \xi_{h}''(-1) }{\xi_{h}'(-1)}\\
&= -4 \dfrac{\dfrac{1}{6} \dfrac{3\left( \xi_{h}^{-1''} \right)^2 - \xi_{h}^{-1'} \xi_{h}^{-1'''}}{\left(\xi_{h}^{-1'}\right)^5} + \dfrac{1}{2} \dfrac{\xi_{h}^{-1''}}{\left( \xi_{h}^{-1'} \right)^3} }{ \dfrac{1}{\xi_{h}^{-1'} }}\\
\label{eq:Ratio}&= \dfrac{-2}{3} \dfrac{ 3\left( \xi_{h}^{-1''} \right)^2 - \xi_{h}^{-1'} \xi_{h}^{-1'''} + 3(\xi_{h}^{-1'})^2 \xi_{h}^{-1''}}{ (\xi_{h}^{-1'})^{4} }.
\end{align}
\end{subequations}
The exact (cumbersome) expression for the ratio can be obtained by plugging in Eqs. (\ref{eq:xiInv1new}), (\ref{eq:xiInv2}) and (\ref{eq:xiInv3}) into Eq. (\ref{eq:Ratio}).
\subsection{Universal aspects of the limiting distribution}
We now consider the $c \rightarrow \infty$ properties of the limiting distribution. Consider the ratio $\mathbb{E}[\tau^2]/\mathbb{E}[\tau]$.
To that end, we isolate the highest order term of $c$ in the denominator and numerator and obtain
\begin{align}
\dfrac{\mathbb{E}[\tau^2]}{\mathbb{E}[\tau]} = \dfrac{2}{3} \dfrac{2^{12} c^4 B_8 + O(c^3)}{2^{12} c^4 B_8 +O(c^3)}.
\end{align}
We arrived at this expression by manipulating the expressions for $\mathbb{E}[\tau]$ and $\mathbb{E}[\tau^2]$ given by Eq. (\ref{eq:mean3}) and Eq. (\ref{eq:second moment2}) respectively, that involved the terms in Eqs. (\ref{eq:xiInv1new})- (\ref{eq:xiInv3}). Therefore, \begin{equation}\label{eq:limiting ratio}
\lim_{c \rightarrow \infty} \dfrac{\mathbb{E}[\tau^2]}{\mathbb{E}[\tau]} = \dfrac{2}{3}.
\end{equation}
This limiting ratio is universal in the sense that it does not depend on $f_{\theta}$ and coincides with the answer obtained by integrating the DMPK distribution \cite{mello1988macroscopic,beenakker1997random,vellekoop2008universal}
We will now compute the first two moments of the DMPK distribution in Eq. (\ref{eq:dmpk}). Let us suppose that the eigenvalues of the transfer matrix are log-uniformly distributed so that $h(\lambda) \lambda = \kappa \mathcal{I}_{[\epsilon,1/\epsilon]}$ for some small positive $\epsilon$ such that $\epsilon \ll 1$. Then Eq. (\ref{eq:mean}) gives us
$$ \mathbb{E}[\tau] = 4 \,\kappa \,\displaystyle\int_{0}^{\infty} \dfrac{1}{(\lambda+1)^{2}} d\lambda + O(\epsilon)= 4 \kappa + O(\epsilon),$$
whereas Eq. (\ref{eq:second moment}) gives us
$$ \mathbb{E}[\tau^{2}] = 16 \,\kappa \,\displaystyle\int_{0}^{\infty} \dfrac{\lambda}{(\lambda+1)^{4}} d\lambda + O(\epsilon) = \dfrac{16}{6} \kappa + O(\epsilon),$$
so that
$$
\dfrac{\mathbb{E}[\tau^2]}{\mathbb{E}[\tau]} = \dfrac{16/6}{4} = \dfrac{2}{3} + O(\epsilon),$$
and for $\epsilon \ll 1$, we get the universal limiting ratio predicted Eq. (\ref{eq:limiting ratio}). Thus in the $c \to \infty$ limit the DMPK distribution exhibits the same universal ratio of the first and second moments as the limiting distribution we have derived using random matrix theoretic arguments.
The discussion in Section \ref{sec:k points} suggests that whenever the medium is `sparse' in the sense that $k_M \, \Nlay/M \to c$, we can expect to get a distribution of the form posited by the DMPK theory irrespective of the material properties of the individual scatterers. The natural next step in this line of inquiry is to analyze the large $c$ asymptotics of the transmission coefficient distribution via its implicit characterization in Eq. (\ref{eq:fixedptEq}) to answer finer questions about the existence of a density at $\lambda = 1$ (equivalently $\tau = 1$) for all $ c \in (0, \infty)$. We leave these for future work.
\subsection{Multiple-point-scatterer-per-layer scenarios that lead to same limiting distribution}\label{sec:k points}
We now consider multiple-point-scatterer-per-layer scenarios that lead to the same limiting distribution - this will suggest a sparsity condition for the existence of the perfect transmission-supporting universal limiting transmission coefficient distribution. Consider the setting similar to that in Fig. \ref{fig:setup} except with $k$ randomly placed point scatterers per layer. Then if $D \gg r$ is large, we expect the $S_{11}$ and $S_{22}$ matrix to be approximately rank $k$, by neglecting the scatterer-scatterer interaction related terms. Consequently, we can model the empirical eigenvalue distribution of $T_i^{H}\cdot T_{i}$ as
\begin{equation}\label{eq:h general k}
h_{i}(\lambda) = \left(1-\dfrac{2k}{2M}\right) \delta(\lambda-1) + \dfrac{k}{2M} \sum_{j=1}^{k}\left[\delta(\lambda- \theta_j) + \delta(\lambda- 1/\theta_j) \right].\end{equation}
Retracing the steps after Eq. (\ref{eq:h kequals1}), we observe that we arrive at the same limiting distribution encoded in Eq. (\ref{eq:fixedptEq}) except now with $k \, \Nlay/M \to c$ and
$$f_{\theta}(t) := \dfrac{1}{k} \sum_{j=1}^{k} f(\theta_j).$$
Now, suppose that we are in the setting where the rank of the $S_{11}$ and $S_{22}$ matrices depends on $M$. Let us make this dependence explicit by denoting it as $k_M$. Suppose that $k_M/M \to 0$. Then, following the argument following Eq. (\ref{eq:h general k}), we will arrive at the same limiting distribution encoded in Eq. (\ref{eq:fixedptEq}) whenever $k_M \, \Nlay/M \to c$ and
$$f_{\theta}(t) := \dfrac{1}{k_M} \sum_{j=1}^{k_M} f(\theta_j).$$
Our analysis thus suggests the sparsity condition $k_M/M \to 0$ and $k_M \, \Nlay/M \to c$ for the emergence of the perfect transmission-supporting universal transmission coefficient distribution.
\section{Numerical simulations}
\label{sec:NumSim}
To validate the predicted transmission coefficient distribution, we adopt the numerical simulation protocol described in \cite{cjin2013}. Specifically, we compute the scattering matrices in Eq. (\ref{eq:scat matrix}) via a spectrally accurate, T-matrix inspired integral equation solver that characterizes fields scattered from each cylinder in terms of their traces expanded in series of azimuthal harmonics. As in \cite{cjin2013}, interactions between cylinders are modeled using 2D periodic Green's functions. The method constitutes a generalization of that in \cite{mcphedran1999calculation}, in that it does not force cylinders in a unit cell to reside on a line but allows them to be freely distributed throughout the cell. As in \cite{cjin2013}, all periodic Green's functions/lattice sums are rapidly evaluated using a recursive Shank's transform using the methods described in \cite{singh1991use,sidi2003practical}. Our method exhibits exponential convergence in the number of azimuthal harmonics used in the description of the field scattered by each cylinder. As in \cite{cjin2013}, in the numerical experiments below, care was taken to ensure 11-digit accuracy in the entries of the computed scattering matrices.
We now describe how the simulations were performed. We generated a random scattering system with $r = 0.05\lambda, \ell = 12.63\lambda, D = 25.75\lambda$, and $M=51$. The locations of the scatterers were selected randomly as described in Section \ref{sec:Setup}. For a given $\Nlay$, the number of layers in the scattering system, we numerically compute the scattering matrices. We then compute the empirical transmission coefficient distribution over $200$ Monte-Carlo trials and compare it to the analytically predicted transmission coefficient distribution obtained as a fixed point of Eq. (\ref{eq:fixedptEq}) for $c = \Nlay/M$ and an appropriate choice of $f_{\theta}$.
We first consider the setting where all the randomly placed cylinders have the same refractive index. Plugging in $f_{\theta}(t) = \delta(t - \theta)$ into Eq. (\ref{eq:xi general}) yields the expression
\begin{equation}\label{eq:xi nonrandom}
\widetilde{\xi}(z) = z\left( \dfrac{1}{z-\theta} + \dfrac{1}{z-\theta^{-1}} - \dfrac{2}{z-1} \right).
\end{equation}
For $n = 2.08$, we get $\alpha = 0.33$ and $\theta = 0.5$. Plugging in $\theta = 0.5$ into Eq. (\ref{eq:xi nonrandom}) and solving Eq. (\ref{eq:fixedptEq}) yields the transmission coefficient as a function of $c$. Fig. \ref{fig:nonrand} shows the agreement between the physically rigorous empirical distribution and the analytically predicted distribution. Note in particular, the agreement for $c=2$ where the distribution is far from the characteristically bimodal DMPK distribution.
We now consider the setting where with probability $p_1$ a cylinder has refractive index $n_1$ and with probability $p_2$ it has a refractive index $n_2$. Plugging in $f_{\theta}(t) = p_1\delta(t-\theta_1) + p_2\delta(t-\theta_2)$ into Eq. (\ref{eq:xi general}) yields the expression
\begin{equation}\label{eq:xi two atom}
\widetilde{\xi}(z) = z\left( \dfrac{p_{1}}{z-\theta_{1}} + \dfrac{p_{1}}{z-\theta_{1}^{-1}} + \dfrac{p_{2}}{z-\theta_{2}} + \dfrac{p_{2}}{z-\theta_{2}^{-1}} - \dfrac{2}{z-1} \right).
\end{equation}
For $n_1 = 1.28$ and $n_2 = 2.89$ we get $\alpha_1 = 0.05$, $\theta_1 = 0.9$, and $\alpha_2 = 0.82$, $\theta_2 = 0.1$. Plugging in these values into Eq. (\ref{eq:xi two atom}) with $p_1 = 0.8$ and $p_2=0.2$ and solving Eq. (\ref{eq:fixedptEq}) yields the transmission coefficient as a function of $c$. Fig. \ref{fig:atomic} shows the agreement between the numerically obtained empirical distribution and the analytically predicted distribution. Note in particular, the agreement for $c=2$ where the predicted distribution is supported on two intervals and agrees with the empirical results.
Finally, we consider the setting corresponding to $f_{\theta}(t) = \dfrac{1}{\theta_2-\theta_1} \mathbb{I}_{\theta_1 \leq t \leq \theta_2}$. We generated the scattering system by mapping each random realization of $\theta$ to a random realization of the refractive index. Plugging this choice into Eq. (\ref{eq:xi general}) yields the expression
\begin{equation}\label{eq:xi uniform}
\widetilde{\xi}(z) = \dfrac{z}{\theta_{2}-\theta_{1}} \left( \log\left(\dfrac{z-\theta_{1}}{z-\theta_{2}}\right) + \dfrac{\theta_{2}-\theta_{1}}{z} + \dfrac{1}{z^2}\log\left( \dfrac{\theta_{2} z - 1}{\theta_{1} z -1} \right) - \dfrac{2}{z-1} \right).
\end{equation}
The choice of $\theta_1 = 0.1$ and $\theta_2 = 0.9$ corresponds to a refractive index of $n_1 = 2.89$ (with $\alpha_1 = 0.82$) and a refractive index $n_2 = 1.28$ (with $\alpha_2 = 0.9$). Plugging these values of $\theta_1$ and $\theta_2$ into Eq. (\ref{eq:xi uniform}) and solving Eq. (\ref{eq:fixedptEq}) yields the transmission coefficient as a function of $c$. Fig. \ref{fig:uniform} shows the agreement between the physically rigorous empirical distribution and the analytically predicted distribution. Note in particular, the agreement for $c=2$ where the distribution is far from the characteristically bimodal DMPK distribution.
Appendix \ref{movies} contains some movies that shows the evolution of the transmission coefficient distribution with $c$ for each of the three scenarios discussed. As expected, for large enough $c$ the distribution eventually becomes characteristically bimodal as predicted by the DMPK theory. The behavior for small values of $c$ is accurately predicted by our theory.
For the three settings described above, we analytically compute $\mathbb{E}[\tau]$ from the associated $f_\theta$ via Eq. (\ref{eq:first moment closed form}). The computation involves the normalizing factor $B_2$, which for the three settings is given by
\begin{subequations}
\begin{align}\label{eq:B2 analytical}
B_{2}^{\rm nonrandom} &= \left( \dfrac{1-\theta}{1+\theta} \right)^2,\\
B_{2}^{\rm atomic} &= p_1 \left( \dfrac{1-\theta_1}{1+\theta_1} \right)^2 + p_2 \left( \dfrac{1-\theta_2}{1+\theta_2} \right)^2,\\
B_{2}^{\rm uniform} &= 1 - \dfrac{4}{\theta_2-\theta1}\log\left( \dfrac{\theta_2+1}{\theta_1+1} \right) + \dfrac{4}{(\theta_{1}+1)(\theta_{2}+1)}.
\end{align}
\end{subequations}
The closed-from expression for $\mathbb{E}[\tau^2]$ is lengthy and therefore omitted here. It can be obtained using the calculations in Appendix \ref{sec:second moment}.
Fig. \ref{fig:moment sims} compares the empirical moments with the predicted moments and shows the good agreement for a range of values of $c$.
Finally, we numerically validate the analytical prediction in Eq. (\ref{eq:limiting ratio}). To that end, we generated a random scattering system with $D=197\lambda, r=0.11\lambda, \widetilde{L} = 3.4 \times 10^5 \lambda, N_{c} = 430,000, n_{d} = 1.3$, and $M = 395$. The locations of the scatterers were selected randomly and produced a system with $\overline{l}=6.69\lambda$, where $\overline{l}$ is the average distance to the nearest scatterer. Let $L$ denote the thickness of the scattering system we are interested in analyzing. We vary $L$ from $\lambda$ to $\widetilde{L}$ and for each value of $L$ we compute the scattering matrices associated with only the scatterers contained in the $(0,L)$ portion of the $(0,\widetilde{L})$ system we have generated. This construction ensures that the average density per ``layer'' of the medium is about the same. We computed the first and second moment of the empirical transmission coefficient distribution by averaging over $1700$ random realizations of the scattering system and computed ratio as a function of $c = M/\widetilde{L}$. Fig. \ref{fig:ratio} shows that the empirical result validate our theoretical prediction.
\section*{Acknowledgements}
This work supported by a DARPA Young Faculty Award D14AP00086, AFOSR Young Investigator Award FA9550-12-1-0266, ONR Young Investigator Award N00014-11-1-0660, US Army Research Office (ARO) under grant W911NF-11-1-0391 and NSF grant CCF-1116115.
\begin{figure}
\centering
\subfloat[C=2.]{
\includegraphics[trim = 22 0 30 20, clip=true,width=0.45\textwidth]{PhysicsTheoryNonRandK1C2.pdf}
}
\subfloat[C=41.]{
\includegraphics[trim = 22 0 30 20, clip=true,width=0.45\textwidth]{PhysicsTheoryNonRandK1C41.pdf}
}\\
\subfloat[C=2: Zoom-in.]{
\includegraphics[trim = 130 0 20 150, clip=true,width=0.70\textwidth]{PhysicsTheoryNonRandK1C2.pdf}
}\\
\subfloat[C=41: Zoom-in.]{
\includegraphics[trim = 40 0 20 200, clip=true,width=0.70\textwidth]{PhysicsTheoryNonRandK1C41.pdf}
}
\caption{The transmission coefficient distribution for the setting where $f_{\theta}(t) = \delta(t - \theta)$. The red line is the theoretical prediction - the histograms are from the physically rigorous simulation averaged over $100$ trials. Note the agreement with theory in the $C=2$ where the distribution is far from the DMPK distribution.}
\label{fig:nonrand}
\end{figure}
\begin{figure}
\centering
\subfloat[C=2.]{
\includegraphics[trim = 22 0 30 20, clip=true,width=0.45\textwidth]{PhysicsTheoryRandK1C2.pdf}
}
\subfloat[C=33.]{
\includegraphics[trim = 22 0 30 20, clip=true,width=0.45\textwidth]{PhysicsTheoryRandK1C33.pdf}
}\\
\subfloat[C=2: Zoom-in. The predicted distribution is supported on two intervals.]{
\includegraphics[trim = 40 0 40 200, clip=true,width=0.90\textwidth]{PhysicsTheoryRandK1C2.pdf}
}\\
\subfloat[C=33: Zoom-in.]{
\includegraphics[trim = 40 0 20 200, clip=true,width=0.90\textwidth]{PhysicsTheoryRandK1C33.pdf}
}
\caption{The transmission coefficient distribution for the setting where $f_{\theta}(t) = 0.8 \delta(t-0.9) + 0.2 \delta(t-0.1)$. The red line is the theoretical prediction - the histograms are from the physically rigorous simulation averaged over $100$ trials. Note the agreement with theory in the $C=2$ where the distribution is far from the DMPK distribution.}
\label{fig:atomic}
\end{figure}
\begin{figure}
\centering
\subfloat[C=2.]{
\includegraphics[trim = 22 0 30 20, clip=true,width=0.45\textwidth]{PhysicsTheoryUniformK1C2.pdf}
}
\subfloat[C=25.]{
\includegraphics[trim = 22 0 30 20, clip=true,width=0.45\textwidth]{PhysicsTheoryUniformK1C25.pdf}
}\\
\subfloat[C=2: Zoom-in.]{
\includegraphics[trim = 40 0 30 200, clip=true,width=0.90\textwidth]{PhysicsTheoryUniformK1C2.pdf}
}\\
\subfloat[C=25: Zoom-in.]{
\includegraphics[trim = 40 0 30 200, clip=true,width=0.90\textwidth]{PhysicsTheoryUniformK1C25.pdf}
}
\caption{The transmission coefficient distribution for the setting where $f_{\theta}(t) = \dfrac{1}{\theta_2-\theta_1} \mathbb{I}_{\theta_1 \leq t \leq \theta_2}$. The red line is the theoretical prediction - the histograms are from the physically rigorous simulation averaged over $100$ trials. Note the agreement with theory in the $C=2$ where the distribution is far from the DMPK distribution.}
\label{fig:uniform}
\end{figure}
\begin{figure}
\subfloat[The first moment.]{
\includegraphics[trim = 0 0 0 0, clip=true,width=0.45\textwidth]{Mom1LogA.pdf}
}
\subfloat[The second moment.]{
\includegraphics[trim = 0 0 0 0, clip=true,width=0.45\textwidth]{Mom2LogA.pdf}
}
\caption{The first moment versus $c$ for the settings corresponding to Fig. \ref{fig:nonrand}, Fig. \ref{fig:atomic} and Fig. \ref{fig:uniform} respectively. The results of the physical simulations were averaged over $100$ trials.
}\label{fig:moment sims}
\end{figure}
\begin{figure}
\centering
\includegraphics[trim = 0 0 0 0, clip=true,width=0.95\textwidth]{RatioBig.pdf}
\caption{The ratio of $\mathbb{E}[\tau^2]/\mathbb{E}[\tau]$ as a function of $c$. The $2/3$ line corresponds to the prediction in Eq. (\ref{eq:limiting ratio}) for the large $c$ limit of this ratio.}
\label{fig:ratio}
\end{figure}
|
\section{Introduction}
{ We have been monitoring the central region of M31 with {\em Chandra}, averaging $\sim$1 observation per month over 1999--2012,} in order to discover X-ray transients. Since then, we have reduced our observation rate to 5 per year. { Particularly bright or otherwise interesting} transients are followed up with two {\em HST} ACS observations, the first is taken a few weeks after outburst, and the second observation is normally taken $\sim$6 months later; this allows us to identify the counterpart via difference imaging \citep[see e.g.][ and references within]{barnard2012b}. We summarized the results of the first 12 transients (labeled T1--T12) found via this effort in \citet{barnard2012b}, and performed this analysis for two further transients: XMM J004243.6+412519 \citep[known as M31 ULX2, ][]{barnard13c}, and CXOM31 004252.457+411631.17 \citep[referred to as T13, ][]{barnard14b}.
In this work we report our findings for CXOM31 004205.77+411330.43, hereafter referred to as T14. We note that T14 is source number 71 in our variability survey of 528 X-ray sources within $\sim$20$'$ of { the galaxy nucleus} where we made $\sim$200 new X-ray binary { (XB)} identifications \citep{barnard14a}; { the 0.3--10 keV luminosities for these sources ranged over $\sim$1--6000$\times 10^{35}$ erg s$^{-1}$}. The initial outburst of T14 was in 2011 July; however, the second observations was made almost 3 years later, in 2014 June.
X-ray transients within the central region of M31 are most likely to contain a low mass secondary, as the majority of stars there are old { \citep[see e.g.][]{williams03}}. Low mass X-ray binaries { (LMXBs)} may be transient X-ray sources due to instabilities in their accretion disks; the disk has two stable phases (hot and cold), and an unstable intermediate phase--- matter accumulates in the disk in the cold phase, and is rapidly dumped onto the compact object in the hot phase \citep[see e.g.][]{lasota2001}. However, the X-rays produced by accretion from the hot disk prevent the disk from cooling; the X-ray luminosity decays exponentially if the whole disk is ionised, and linearly if only part of the disk is ionised\citep {king98}.
\citet{vp94} found an empirical relation between the ratio of X-ray and optical luminosities of Galactic X-ray binaries and their orbital periods, suggestive that the optical emission is dominated by reprocessed X-rays in the disk; this relation holds over a 10 magnitude range in optical luminosity, and appears to be independent of inclination. Their chosen X-ray band was 2--10 keV. For an irradiated accretion disk with radius $a$, X-ray luminosity $L_{\rm X}$, optical luminosity $L_{\rm opt}$, and temperature $T$, $T^4$ $\propto$ $L_{\rm X}$/$a^2$, while the surface brightness of the disk, $S$, $\propto$ $T^2$ for typical XBs \citep{vp94}. Since $L_{\rm opt}$ $\propto$ $Sa^2$, $L_{\rm opt}$ $\propto$ $L_{\rm X}^{1/2} a$; also $a$ $\propto$ $P^{2/3}_{\rm orb}$, where $P_{\rm orb}$ is the orbital period.
\citet{vp94} defined $\Sigma$ = $\left(L_{\rm X}/L_{\rm EDD}\right)^{1/2}\left(P_{\rm orb}/1 {\rm hr}\right)^{2/3}$, choosing $L_{\rm EDD}$ = 2.5$\times 10^{38}$ erg s$^{-1}$ as a normalizing constant, and found
\begin{equation}
M_{\rm V} = 1.57(\pm0.24) - 2.27(\pm 0.32) \log \Sigma.
\end{equation}
However, \citet{vp94} sampled a mixture of neutron star and black hole binaries, in various spectral states. A cleaner sample was obtained by A. Moss et al. (2015, in prep), who used only black hole transients at the peaks of their outburst, and found
\begin{equation}
M_{\rm V} = 0.84(\pm0.30) - 2.36(\pm0.30) \log \Sigma.
\end{equation}
We note that these two relations only differ significantly in normalization, caused by black hole X-ray binaries having larger disks than neutron star binaries with the same period. We have period estimates for 14 M31 transients (T1--T12, ULX2, and T13) observed by {\em {\em Chandra}} and {\em HST} \citep{barnard2012b, barnard13b, barnard14b}. { These period estimates appear to be slightly shorter over all than for Galactic transients, ranging from $<$1 hr to 140$^{+50}_{-40}$ hr, and mostly $<$30 hr \citep{barnard2012b, barnard13b, barnard14b}.}
\section{Observations and data reduction}
We observed the region containing T14 in 101 {\em Chandra} observations (60 ACIS, 41 HRC).
We observed T14 in outburst for 5.2 ks with the HST/ACS on 2011 July 21 using the F435W filter and the WFC mode (or ``aperture'') (H1, jbm105010, PI M. Garcia); the 4.8 ks observation in quiescence occurred on 2014 June 26 (H2, jc6b02010, PI R. Barnard), using the WFC1-CTE aperture. The WFC1-CTE aperture improves the charge transfer efficiency by moving the target to the corner of a detector, allowing point sources to be more circular, but does not effect sensitivity; hence the level of background observed in H2 is unaffected by this change. We also analyzed { a 2001 June 29 $\sim$60 ks {\em XMM-Newton} observation containing a previous outburst} that had sufficient counts for spectral fitting (0109270101, PI K. Mason).
All optical analysis was performed with {\sc pc-iraf} Revision 2.14.1, except where noted. The {\em Chandra} and {\em XMM-Newton} observations were analyzed with {\sc ciao} version 4.6.3 and SAS version 13.0 respectively; X-ray spectra from both telescopes were treated with {\sc xspec} version 12.8.2b.
\subsection{{\em HST} analysis}
Each {\em HST} observation included four flat-fielded (FLT) images, and one drizzled (DRZ) image. The flat-fielded images are corrected for instrumental effects, but not background subtracted; the total number of counts in each pixel is given. The native ACS resolution is comparable to the FWHM of the PSF \citep{fruchter09}. The drizzled image combines the flat-fielded images, removes any cosmic rays, and subtracts the sky background; it is normalized to give the number of counts per second per pixel. We used the DRZ images from H1 and H2 to create a difference image; however, we used the H2 FLT images for our aperture photometry because the DRZ images sometimes contain pixels with slightly negative values, and this can cause problems when estimating the number of photons in a region.
\subsubsection{Creating a difference image}
We reprojected the H2 DRZ image into the coordinates of the H1 DRZ image, to produce an accurate difference image. To do this, we first registered the H1 and H2 images to the LGS Field 5 image with {\sc ccmap}, using unsaturated stars that were close to the target. Then, we used the {\sc iraf} task {\sc wregister} to make the pixel orientation of the H2 image match that of the H1 image. We registered the H2 image to the H1 image before mapping to reduce the noise during image subtraction. The difference image was produced by subtracting H2 from H1 using the {\sc ftools} task {\sc farith}.
\subsubsection{Measuring the optical counterpart}
For H1 and H2 we found the number of photons for each FLT image within various extraction regions, yielding $C_{\rm tot}$ photons in total over $T$ seconds. We subtracted the H2 total from the H1 total to get $C_{\rm net}$ source photons.
We converted this to Vega $B$ magnitude via
\begin{equation}\label{4s}
B \simeq -2.5 \log\left[ C_{\rm net}/T \right] + ZP,
\end{equation}
having obtained the zero point ($ZP$ = 25.77) from the ACS Zero Point calculator\footnote{http://www.stsci.edu/hst/acs/analysis/zeropoints/zpt.py}; we see from \citet[][Equation 12 and Table 18]{sirianni05} that the conversion from F435W counts to B magnitude is within 3$\sigma$ of our ZP for B-V = $-$0.09 \citep[assuming a typical disk spectrum][]{liu01}. We also estimated the 4$\sigma$ upper threshold for source detection from the total photon count from H2 ($C_{\rm tot, H2}$):
\begin{equation}
B \simeq -2.5 \log\left[ 4\times C_{\rm tot,H2}^{0.5}/T \right] + ZP.
\end{equation}
We can convert from $B$ magnitude to $M_{\rm V}$ via
\begin{equation}\label{conv}
M_{\rm V} = B + 0.09 - N_{\rm H}\times\left(1+1/3\right)/1.8\times10^{21} - 24.47,
\end{equation}
where $N_{\rm H}$ is the line of sight absorption; this accounts for the difference in B and V magnitudes of a typical accretion disc, a relationship between B band extinction and measured line-of-sight absorption towards the object, and the distance to M31 \citep[see][and references within]{barnard2012b}.
\subsection{{\em XMM-Newton} analysis}
The {\em XMM-Newton} observation suffered substantial { flares in the particle background due to soft Solar protons}. We selected good time intervals by making a 10--12 keV pn lightcurve from the whole image with 100 s binning, { using standard filters}. { We extracted source source and background spectra from the pn, MOS1 and MOS2 data: source spectra from a circular region that was optimized by the software { (17$''$ radius)}, and background spectra from a nearby circular region on the same CCD with no point source { (35$''$ radius)}. We grouped the source spectra to give a minimum of 20 counts per bin, created the corresponding response matrixes and ancillary response files, then fitted the spectra with {\sc xspec}. The pn, MOS1, and MOS2 spectra were fitted simultaneously with constants of normalization to account for differences in detector responses.}
{ We obtained uncertainties in the best fit parameters by simulating 1000 spectra with the {\sc multifake} command, with deviations in simulated spectra being drawn from the properties of the observed spectrum. The best fit parameters for these spectra were each ranked in ascending order, with 1$\sigma$ uncertainties derived from the 160$^{th}$ and 840$^{th}$ data points.}
\begin{figure*}
\includegraphics[scale=0.43]{fig1.eps}
\caption{{\em Left:} Detail of the 2011 July 21 HST/ACS WFC (H1) image of T14 during outburst. {\em Middle:} 2014 June (H2) image of T14 in quiescence. {\em Right: } Difference image, where white points were brighter in H1, and dark points were brighter in H2. North is up, East is left. The 3$\sigma$ uncertainty in X-ray source position (0.9$"$ RA, 1.2$"$ Dec) is indicated by an ellipse. We highlight the counterpart with two lines. }\label{him}
\end{figure*}
\subsection{ {\em Chandra} Analysis}
\subsubsection{Locating the X-ray source}
We used 27 X-ray bright globular clusters (GCs) to register a combined $\sim$350 ks ACIS image (supplied by Z. Li) to the B band Field 5 image of M31 provided by the Local Galaxy group Survey (LGS) \citep{massey06}. We used {\sc pc-iraf} v2.14.1 to perform the registration, following the same procedure as described in \citet{barnard2012b}. The X-ray and LGS positions of the 27 X-ray bright GCs were determined using {\sc imcentroid}; the equivalent FK5 coordinates were calculated for the X-ray and optical position of each GC using {\sc xy2sky} v2.0, distributed with {\sc ftools}. The X-ray positions of each GC were altered to match the LGS positions, allowing the registration of the merged {\em Chandra} image to Field 5 using the {\sc iraf} task {\sc ccmap}. This registration yielded 1$\sigma$ position uncertainties of 0.11$''$ in R.A., and 0.09$''$ in Dec \citep{barnard2012b}.
The final uncertainties in the X-ray position of T14 combine the position uncertainties in the X-ray image, and the uncertainties in registering the merged {\em Chandra} image to the M31 Field 5 LGS image.
{ \subsubsection{Spectral fitting}
Even the best {\em Chandra} ACIS spectrum for T14 yielded few counts. We modeled the spectrum in XSPEC in two ways: i) using a small number of bins with $\ge$20 counts per bin and $\chi^2$ statistics; and ii) using a larger number of bins with $\ge$5 counts per bin and Cash statistics. We estimated the uncertainties for each spectral fit using the same method as for our analysis of the {\em XMM-Newton} spectrum.}
\subsubsection{Estimating luminosities}
For each of our {\em Chandra} ACIS observations we extracted source and background 0.3--7.0 keV spectra from circular regions with 10$''$ radii. We then created a response matrix using {\sc mkacisrmf}, and obtained an ancillary response file from {\sc mkarf}. Spectral analysis was performed using {\sc xspec}, giving unabsorbed 0.3--10 keV luminosities for each observation, allowing us to create a long-term luminosity lightcurve.
{ Most of the {\em Chandra} observations of T14 yielded too few photons for spectral fitting. }However, we are able to estimate the 0.3--10 keV flux (and therefore luminosity) by assuming a particular emission model. If this model is a good approximation, then the luminosities of closely-spaced ACIS and HRC observations should agree; however, if the model is inappropriate, then the ACIS and HRC luminosities should be systematically offset, due to differences in instrumental response.
For each faint ACIS spectrum we used XSPEC to estimate the unabsorbed 0.3--10 keV flux equivalent to 1 count s$^{-1}$, assuming the best fit emission model for the {\em XMM-Newton} pn spectrum of T14. Multiplying this number by the background-subtracted intensity yields a flux that is corrected for instrumental effects, and we obtain the luminosity by assuming a distance of 780 kpc \citep{stanek98}. It was important to do this for every observation because the roll angle was not constrained, meaning that the source appeared in different parts of the detectors, at various off-axis angles.
For HRC observations, we included only PI channels 48--293, thereby reducing the instrumental background. We used the WebPIMMS tool to find the unabsorbed luminosity equivalent to 1 count s$^{-1}$, assuming the same emission model as for the ACIS observations. We created a 1 keV exposure map for each observation, and compared the exposure within the source region with that of an identical on-axis region, in order to estimate the exposure correction. We multiplied the background subtracted, corrected source intensity by the correction factor to get the 0.3--10 keV luminosity.
\section{Results}
\subsection{Locating and measuring the optical counterpart}
The uncertainty in X-ray position for T14 is 0.3$"$ in R.A. and 0.4$"$ in Dec \citep{barnard14a}. Registering H1 to the LGS M31 Field 5 image provided by \citet{massey06} yielded r.m.s. offsets of 0.04$"$ in R.A. and 0.02$"$ in Dec. The uncertainties arising from HST are negligible in comparison with the X-ray uncertainties. In Figure~\ref{him} we present details of the HST/ACS WFC F435W images of T14 from H1 (left), and H2 (middle), along with the difference image (right). For each panel, an ellipse represents the 3$\sigma$ uncertainty in X-ray position with respect to the LGS image (0.9$"$ in R.A., and 1.2$"$ in Dec.). For the difference image, white stars were brighter in H1, while dark stars were brighter in H2. We find evidence for a faint optical counterpart at 00:42:05.818 +41:13:30.00, with respect to the LGS Field 5 B band image. { This is $\sim$8$'$ offset from the centres of the {\em Chandra} images.}
In \citet{barnard14b} we used a circle with 3 pixel radius (3$\times$FWHM) to estimate the flux for T13; however, this was impossible for T14 because this region was severely contaminated by cosmic rays in the H2 image, and the H2 intensity (7.6$\pm$0.08 count s$^{-1}$) was substantially higher than the H1 intensity (5.5 count$\pm$0.04 s$^{-1}$). Extracting counts from a 3$\times 3$ pixel region yielded very similar result to extraction from a 2$\times$2 pixel region: background-subtracted intensities of 0.11$\pm$0.02 and 0.10$\pm$0.02 count s$^{-1}$ respectively. All uncertainties in this work are quoted at the 1$\sigma$ level.
For H1, the 3$\times$3 pixel region yielded 6688 photons over 5240 seconds, while the 2$\times$2 region yielded 3406 photons; for H2 the 3$\times$3 region produced 4263 photons over 4790 seconds, while the 2$\times2$ region gave 2658 photons. These results allow us to estimate the B magnitude of T14 to be 28.21$\pm$0.16 in H1 (using Equation 3) , with a 4$\sigma$ detection limit of 28.6 magnitudes measured for H2 from the 3$\times3$ pixel region (Equation 4).
The faintest counterpart that we detected previously had an apparent B magnitude of 24.87$\pm$0.09, but we measured 4$\sigma$ upper limits of B $>$28.7 for two other transients; the detection limit is strongly location-dependent, ranging from B $\sim$26 in the bulge to B $\sim$29 further out \citep{barnard2012b, barnard13b}.
We note that the H1 DRZ file, which is background-subtracted, yields an intensity of 0.16 count s$^{-1}$ for the 3$\times$3 region, and 0.10 count s$^{-1}$ for the 2$\times$2 region, which are consistent with our intensities obtained by subtracting H2 from H1
\subsection{Characterising the X-ray spectrum}
\begin{figure}
\includegraphics[scale=0.4]{fig2.eps}
\caption{Unfolded {\em XMM-Newton} { pn, MOS1 and MOS2 spectra} for T14 from the 2001 June 29 observation of T14, fitted with the best fit absorbed power law model. The y axis shows the model multiplied by channel energy. { For this model, $N_{\rm H}$ = 2.6$\pm$0.3$\times 10^{21}$ atom cm$^{-2}$ and $\Gamma$ = 1.84$\pm$0.12, $\chi^2$/dof = 42/54; the MOS1 and MOS2 spectra were normalized by factors of 1.02$\pm$0.03 and 1.15$\pm$0.03 respectively in order to account for differences in detector response. }}\label{spec}
\end{figure}
\subsubsection{XMM-Newton}
The {\em XMM-Newton} pn spectrum of T14 contained $\sim$550 net source counts, { with an intensity of 0.0209$\pm$0.0009 count s$^{-1}$}, { while the MOS1 and MOS2 spectra contained 242 and 323 counts respectively}; this is sufficient for simple emission models. { For the simultaneous spectral fits, $k_1$ and $k_2$ represent the normalizations for MOS1 and MOS2 respectively. } X-ray binaries exhibit characteristic emission spectra that depend on the accretor and spectral state. The hard state is common to all X-ray binaries, whether the accretor is a neutron star or black hole \citep{vdk94}; its emission is dominated by inverse Comptonization of cool photons on hot electrons, and may be represented by a power law with photon index ($\Gamma$) = 1.4--2.1 \citep[see e.g.][]{vdk94,remillard06}. This hard state is observed at luminosities below $\sim$10\% of the Eddington limit \citep{gladstone07,tang11}. Black hole transients often exhibit a thermally dominated state during outburst, characterized by a multi-temperature disk blackbody with inner disk temperature ($kT_{\rm in}$) $\sim$1 keV \citep{remillard06}. The other states of NS and BH XBs are more complex, with multiple emission components, so we do not attempt to fit them here.
{ Fitting an absorbed power law model yielded a best fit line of sight absorption ($N_{\rm H}$) of { 2.6$\pm$0.3$\times 10^{21}$ atom cm$^{-2}$} and $\Gamma$ = 1.84$\pm$0.12, $k_1$ = 1.03$\pm$0.03, $k_2$ = 1.15$\pm$0.03; the 0.3--10 keV unabsorbed luminosity was 1.89$\pm$0.12$\times 10^{37}$ erg s$^{-1}$, and $\chi^2$/dof = 42/54. }
{ The best fit disk blackbody model yielded $N_{\rm H}$ = 0.09$\pm$0.03$\times 10^{21}$ atom cm$^{-2}$ and $kT_{\rm in}$ = 1.460$\pm$0.13 keV, $k_1$ = 1.00$\pm$0.03, $k_2$ = 1.15$\pm$0.04; the 0.3--10 keV unabsorbed luminosity was 1.25$\pm$0.08 $\times 10^{37}$ erg s$^{-1}$ and $\chi^2$/dof = 56/54.}
The unabsorbed luminosities obtained from these emission models are inconsistent at the 3$\sigma$ level; this is due to the lower column density from the disk blackbody model, caused by the natural downward curvature of the model at low energies. The fit may be improved by using a power law model rather than a disk blackbody model: $\Delta\chi^2$ = 7 for 3 free parameters, and the probability that this improvement is genuine is 93\%, i.e. not 3$\sigma$. When we consider that the luminosity is consistent with the hard state of any XB (NS or BH accretor), then we find the power law emission model to be most likely. We present the unfolded pn spectrum for T14 assuming the best fit power law emission model in Fig.~\ref{spec}; the y-axis shows the spectrum multiplied by channel energy in order to show the distribution of flux.
\subsubsection{ Chandra}
{
The best {\em Chandra} ACIS observation of T14 yielded $\sim$130 counts in the 0.3--7.0 keV range, with no photons detected in the background region over the $\sim$4 ks exposure time. This produced a 6 bin spectrum for $\chi^2$ fitting, and a 25 bin spectrum for fitting with Cash statistics.
The best $\chi^2$ fit absorbed power law model yielded $N_{\rm H}$ = 8$\pm$6$\times 10^{21}$ atom cm$^{-2}$, $\Gamma$ = 1.8$\pm$0.5, and a 0.3--10 keV luminosity of 9$\pm$3$\times 10^{37}$ erg s$^{-1}$. The best Cash statistic fit yielded $N_{\rm H}$ = 4$\pm$3$\times 10^{21}$ atom cm$^{-2}$, $\Gamma$ = 1.5$\pm$0.3, and 0.3--10 keV luminosity 7.1$\pm$1.1$\times 10^{37}$ erg s$^{-1}$. These fits are consistent with those obtained from the XMM-Newton spectrum.
Fitting an absorbed disk blackbody model with $\chi^2$ statistics resulted in $N_{\rm H}$ = 3$\pm$3$\times 10^{21}$ atom cm$^{-2}$, k$T$ = 2.1$\pm$0.7, and $L$ = 5.2$\pm$0.8$\times 10^{37}$ erg s$^{-1}$. Cash statistics yielded $N_{\rm H}$ = 2$\pm$2$\times 10^{21}$ atom cm$^{-2}$, and kT = 2.0$\pm$0.7, with a 0.3--10 keV luminosity of 3.1$\pm$0.6$\times 10^{37}$ erg s$^{-1}$.
}
\subsection{The X-ray lightcurve}
\begin{figure}
\includegraphics[scale=0.4]{fig3.eps}
\caption{Unabsorbed 0.3--10 keV lightcurve for T14, from ACIS (filled circles), HRC (hollow circles) and {\em XMM-Newton} (triangle) observations, assuming an absorbed power law with { $N_{\rm H}$ = 2.6$\times 10^{21}$ atom cm$^{-2}$ and $\Gamma$ = 1.84}. We also show indicate times of the HST observations with downward arrows. }\label{lc}
\end{figure}
We obtained a long-term 0.3--10 keV unabsorbed luminosity lightcurve for T14 from the ACIS and HRC observations by assuming an absorbed power law emission model with { $N_{\rm H}$ = 2.6$\times 10^{21}$ atom cm$^{-2}$ and $\Gamma$ = 1.84}. In Fig.~\ref{lc} we present luminosities from the ACIS (filled circles) and HRC (hollow circles) observations; we indicate the times of the HST observations with downward arrows, and also include the {\em XMM-Newton} observation (triangle).
We see that T14 exhibited several outbursts; they do not appear to follow the traditional exponential or linear decay seen in Galactic transients \citep{king98}; we discuss a possible reason for this below. The luminosities obtained from {\em XMM-Newton}, ACIS and HRC are all consistent for the first outburst despite the very different instrumental responses, and the ACIS and HRC luminosities appear to be in good agreement for the whole lightcurve, showing that the emission model is probably a good approximation to the true spectrum.
However, we caution that we did not observe any period where the luminosity exceeded 2$\times$10$^{37}$ erg s$^{-1}$ ($\sim$10\% Eddington for a 1.4 $M_\odot$ NS) with more than one detector. Hence, we cannot determine whether the spectrum evolved as the luminosity increased.
\subsection{Period estimation}
\begin{figure}
\includegraphics[scale=0.4]{fig4.eps}
\caption{Close-up of the outburst examined by {\em Chandra} and HST. The axes and symbols are the same as for Fig.~\ref{lc}.}\label{closeup}
\end{figure}
Assuming an apparent B magnitude of 28.21 for T14 during H1 and line-of-sight absorption equivalent to 2.6$\times 10^{21}$ H atom cm$^{-2}$, Equation 5 yields an absolute V magnitude of { +1.9} for T14 in outburst. The highest 2--10 keV luminosity observed from T14 during the outburst was 2.7$\pm$0.2$\times 10^{37}$ erg s$^{-1}$ for our assumed emission model. We estimated the orbital period using Equations 1 and 2.
We included the uncertainties in $L_{\rm X}$, $M_{\rm V}$ as well as the uncertainties in the relations themselves by simulating 1000 instances using a Park-Miller random number generator to obtain numbers from 0 to 1, and the Box-Muller technique to convert these numbers into a normal distribution: if we obtain two random numbers $a$ and $b$ between 0 and 1, then $\sqrt{-2\ln a} \cos\left(2\pi b\right)$ and $\sqrt{-2\ln a} \sin\left(2\pi b\right)$ are two independent random numbers with a normal distribution. For each parameter, we added a normally distributed random number multiplied by the 1$\sigma$ uncertainty to the best fit value in order to obtain the value for that instance.
For each empirical relation between X-ray to optical flux ratio and period (Equation 1 and Equation 2), we obtained 1000 estimates of the period and ranked these estimates in ascending order; we estimated the mean for each period as ($P_{840}$ + $P_{160}$)/2 and the 1$\sigma$ uncertainty as ($P_{840}$ $-$ $P_{160}$)/2, where $P_{160}$ and $P_{840}$ are the 160$^{th}$ and 840$^{th}$ ranked periods. { The estimated period from Equation 1 is { 3.0$^{+1.8}_{-1.2}$ hr}, while Equation 2 yields a period of 1.1$^{+0.7}_{-0.4}$ hr; these uncertainties include the uncertainties in the relations described in Equations 1 and 2 as well as uncertainties in X-ray and optical fluxes}.
\section{Summary and conclusions}
We conducted a $\sim$13 year programme of {\em Chandra} observations of the central region of M31, monitoring for outbursts of X-ray transients. { Transients that were particularly bright or otherwise interesting} were followed up with HST/ACS, using the F435W filter that approximates the B band
to search for optical counterparts. CXOM31 004205.77+411330.43 (T14) is one such transient; the outburst was observed by HST in 2011 July, but the follow-up HST observation occurred only recently, in 2014 June. Difference imaging yielded a counterpart with B magnitude 28.21$\pm$0.16.
Modeling an {\em XMM-Newton} spectrum from a previous outburst { provided a good fit} from an absorbed power law model with { $N_{\rm H}$ = 2.6$\pm$0.3$\times 10^{21}$ atom cm$^{-2}$ and $\Gamma$ = 1.84$\pm$0.12, with $\chi^2$/dof = 42/54.} { Fitting the best {\em Chandra} ACIS spectrum yielded results consistent with the {\em XMM-Newton} fits. Furthermore, using the {\em XMM-Newton} mode} for ACIS and HRC observations yielded consistent results, suggesting that this is a good description of the spectrum, as instrumental differences would result in substantial systematic offsets in luminosity if the spectrum was very different. Using the best fit column density, and assuming a typically flat accretion disk spectrum, we estimated the absolute V band magnitude to be { $\sim$1.9}.
The highest observed 2--10 keV luminosity during outburst was { 2.92$\pm$0.06$\times 10^{37}$} erg s$^{-1}$ with our assumed emission model. Using two empirical relations between the ratio of X-ray to optical flux and the orbital period (Equations 1 and 2), we estimate the orbital period of T14 to be { 3.0$^{+1.7}_{-1.2}$ hr and 1.1$^{+0.7}_{-0.4}$ hr} respectively.
We note that Equation 1 was derived from a mixture of NS and BH XBs in several different spectral states, while Equation 2 was obtained from a more selective sample, including only BH XBs in the thermal dominant state. For several of our transients, the resulting periods varied by a factor $>$3; however, for others, including T14, the periods are consistent within uncertainties \citep{barnard2012b}.
We also note that relations between optical to X-ray flux ratio and orbital period best apply to the peak of the outburst, because the optical decay rate is systematically a factor $\sim$2 longer than the X-ray decay \citep{chen97}. If the peak was unobserved, then the X-ray to optical ratio would be higher, resulting in a smaller orbital period.
It is therefore likely that the orbital period for T14 is { $<$4 hr for a BH accretor}. This short period may explain the rather erratic behaviour during the decay, as shown in Fig.~\ref{closeup}, because accretion may be modulated by the precession of an asymmetric disk. { Such behaviour was first observed in cataclysmic variables (CVs) with low mass ratios (and short orbital periods).}
\citet{osaki89} proposed that the super-outbursts observed in CVs with low mass ratios (i.e. short periods) are enhanced by a tidal instability that occurs when the outer disk crosses the 3:1 resonance with the secondary; the additional tidal torque causes the disk to elongate and precess, and also greatly enhances the loss of angular momentum (and therefore the accretion rate). The disk precession is prograde in the rest frame, and the secondary repeats its motion with respect to the disk on the beat period between the orbital period and the precession period, a few percent longer than the orbital period. The secondary modulates the disk's viscous dissipation on this period, giving rise to the maxima in the optical lightcurve known as superhumps. Some short period, persistently bright CVs exhibit permanent superhumps \citep{patterson99,retter00}.
We note that \citet{haswell01} found that X-ray binaries with periods $\la$4 hr are likely to have asymmetric, precessing disks.
The X-ray binary associated with the M31 globular cluster B158 is observed at a high inclination, and exhibits periodic 10017$\pm$50 s ($\sim$2.8 hr) variation in the X-ray lightcurve in some observations \citep{trud02} but not others \citep{barnard06}. The long-term unabsorbed 0.3--10 keV luminosity lightcurve of XB158 varied between $\sim$4 and $\sim$20$\times$10$^{37}$ erg s$^{-1}$; we suggested that this modulation was due to changes in mass transfer during the disk precession cycle \citep{barnard13}.
In \citet{barnard06} we conducted three dimensional smooth particle hydrodynamical modeling of XB158, and predicted a disk precession period of 29$\pm$1 times the orbital period, or $\sim$81$\pm$3 hr. This motivated us to observe XB158 30 times over $\sim$30 days with the Swift XRT in order to search for modulations in the intensity over timescales of a few days. We observed approximately sinusoidal variation of the 0.3--10 keV unabsorbed luminosity over $\sim$4--20$\times 10^{37}$ erg s$^{-1}$ on a 5.65$\pm$0.05 day period \citep{barnard15}.
The outburst of T14 studied with {\em Chandra} and HST has a decay profile (Fig.~\ref{closeup}) that is unlike those of Galactic transient XBs, which exhibit exponential decay when the whole disk is irradiated, and linear decay in case of partial irradiation; the decay can change from exponential to linear when the X-ray luminosity is no longer sufficient to irradiate the whole disk \citep{king98}. We propose that the X-ray luminosity may be modulated by variations in mass transfer from an asymmetrical, precessing disk.
\section*{Acknowledgments}
This research
has made use of data obtained from the {\em {\em Chandra}} satellite,
and software provided by the {\em {\em Chandra}} X-Ray Center (CXC). We thank Z. Li for creating the merged {\em Chandra} image used to register the X-ray image.
We also include analysis of public archival data from {\em {\em XMM-Newton}}, an ESA
science mission with instruments and contributions directly
funded by ESA member states and the US (NASA) R.B. was
funded by the {\em {\em Chandra}} grant GO3-14096X,
along with the {\em HST} grant GO-13111.05-A. Support for program \#13111 was provided by NASA through a grant from the Space Telescope Science Institute, which is operated by the Association of Universities for Research in Astronomy, Inc., under NASA contract NAS 5-26555. M.R.G. and S.S.M are partially supported by NASA contract NAS8-03060.
\label{lastpage}
|
\section{Introduction}
The recent discovery of the $\sim$125 GeV scalar particle at the Large Hadron Collider (LHC)
\cite{:2012gk,:2012gu} perfectly fills the vacancy of the Higgs boson necessary for
the completion of the Standard Model (SM) at the Fermi scale. Surprisingly, the SM with the Higgs boson in this mass range becomes formally self-consistent up to the Planck scale. In the absence of any signal of physics beyond the SM, this fact drastically strengthens the position of this model as the theoretical basis of particle physics.
\quad Although the new observed scalar state has so far all the properties expected of the SM
Higgs boson, it is still possible that it could be a light scalar in a multi-Higgs extension of the SM, or a light supersymmetric
Higgs boson, or a Higgs boson coming from a strongly interacting dynamics,
where the theory becomes nonperturbative above the Fermi scale and the
breaking is achieved through some condensate. Now the priority of the LHC
experiments will be to measure precisely the couplings of the observed scalar
to fermions and gauge bosons, and to establish its quantum numbers in order
to identify it with one of these or some other options. On the other hand, searches for new particles beyond the SM are an essential task of the LHC experiments \cite{PDG,Brooijmans:2014eja,Carena:2013qia,Bechtle:2013wla,Heinemeyer:2014uoa,Dev:2014yca,Carena:2014nza,Bhattacharyya:2014oka}.
\quad In this paper, we consider a multi-Higgs extension of the SM, with an arbitrary number $N$ of the Higgs electroweak doublets. Our goal is to study possible bounds on the number of Higgs doublets from the precision measurements of the oblique $T$ and $S$ parameters.
\quad We assume that the $N$ Higgs $SU(2)$
doublets are identical, with hypercharge equal to $1$. Some features such as
the relation between the mass and gauge eigenstates in the scalar sector and
the relation of the Higgs vacuum expectation values with the symmetry breaking scale
$v\approx 246 $ GeV presented in the two Higgs doublet model are still fulfilled when the number of Higgs doublets is increased \cite{2N-model}.
\quad The paper is organized as follows. In Sec. \ref{model}, we
briefly describe the theoretical structure of the $N$ Higgs doublet model (NHDM). In Sec. \ref{TandSparameters}, we compute the one-loop contribution to the $T$ and $S$ parameters in the NHDM. The bounds on the number of Higgs doublets coming from $T$ and $S$ parameter
constraints at $95\%$C.L. are computed in Sec. \ref{T-S-exp bounds}. In
Sec. \ref{Conclusions}, we summarize our results.
\section{The Model}
\label{model} We consider an extension of the SM with $N$ copies of the complex $%
SU(2)_{L}$ weak doublet scalar Higgs fields with hypercharge $Y=1$ (NHDM). The model scalar potential, invariant with respect to the SM gauge group, is
\begin{equation} \label{V-scalar}
V=\frac{1}{2}\sum_{i,j=1}^{N}\mu _{ij}^{2}\Phi _{i}^{\dagger }\Phi _{j}+
\frac{1}{4}\sum_{i,j,k,l=1}^{N}\lambda _{ij,kl}\left( \Phi _{i}^{\dagger
}\Phi _{j}\right) \left( \Phi _{k}^{\dagger }\Phi _{l}\right)
+\sum_{i,j,k,l=1}^{N}\sigma _{ij,kl}\left( \Phi _{i}\tau^{2} \Phi
_{j}\right) \left( \Phi _{k} \tau^{2} \Phi _{l}\right)^{\dagger} .
\end{equation}
where $\tau ^{2}$ is a Pauli matrix in the $SU(2)_{L}$ space and
\begin{equation}
\sigma _{ij,kl}=-\sigma _{ji,kl}=-\sigma _{ij,lk},\hspace{2cm}
\end{equation}
For simplicity, we assume all the parameters in the scalar potential to be real. Then
the Hermiticity of the scalar potential (\ref{V-scalar}) implies
\begin{eqnarray} \label{Prop-2}
\lambda _{ij,kl}=\lambda _{ji,lk},\hspace{5mm} \sigma_{ij,kl}=\sigma
_{ji,lk}, \hspace{5mm} \mu _{ij}=\mu _{ji},\hspace{2cm}\hspace{2cm}
\end{eqnarray}
The minimum of the scalar potential is parametrized by N vacuum expectation
values
\begin{equation}
\left\langle \Phi _{l}\right\rangle =\left(
\begin{array}{c}
0 \\
\frac{v_{l}}{\sqrt{2}}%
\end{array}%
\right) ,\hspace{2cm}\hspace{2cm}l=1,2,\cdots ,N.
\end{equation}%
We decompose the Higgs fields around this minimum as
\begin{equation}
\Phi _{l}=\left(
\begin{array}{c}
\phi _{l}^{+} \\
\frac{1}{\sqrt{2}}\left( v_{l}+\rho _{l}+i\eta _{l}\right)%
\end{array}%
\right) =\left(
\begin{array}{c}
\frac{1}{\sqrt{2}}\left( \omega _{l}+i\xi _{l}\right) \\
\frac{1}{\sqrt{2}}\left( v_{l}+\rho _{l}+i\eta _{l}\right)%
\end{array}%
\right) \label{doublets}
\end{equation}%
where
\begin{equation}
\left\langle \rho _{l}\right\rangle =\left\langle \eta _{l}\right\rangle
=\left\langle \omega _{l}\right\rangle =\left\langle \xi _{l}\right\rangle
=0,\hspace{2cm}\hspace{2cm}l=1,2,\cdots ,N.
\end{equation}%
Then the covariant derivative acting on the Higgs doublets takes the form
\begin{eqnarray}
D_{\mu }\Phi _{l} &=&\partial _{\mu }\Phi _{l}-\frac{i}{2}gW_{\mu }^{a}\tau
^{a}\Phi _{l}-\frac{i}{2}g^{\prime }Y_{l}B_{\mu }\Phi _{l} \notag \\
&=&\left(
\begin{array}{c}
\frac{1}{\sqrt{2}}\partial _{\mu }\omega _{l}+\frac{1}{2\sqrt{2}}\left[
gW_{\mu }^{1}\eta _{l}-gW_{\mu }^{2}\left( v_{l}+\rho _{l}\right) +\left(
gW_{\mu }^{3}+g^{\prime }Y_{l}B_{\mu }\right) \xi _{l}\right] \\
\frac{1}{\sqrt{2}}\partial _{\mu }\rho _{l}+\frac{1}{2\sqrt{2}}\left(
gW_{\mu }^{1}\xi _{l}+gW_{\mu }^{2}\omega _{l}-gW_{\mu }^{3}\eta
_{l}+g^{\prime }Y_{l}B_{\mu }\eta _{l}\right)%
\end{array}%
\right) \notag \\
&&+\frac{i}{2\sqrt{2}}\left(
\begin{array}{c}
2\partial _{\mu }\xi -\left( gW_{\mu }^{3}+g^{\prime }Y_{l}B_{\mu }\right)
\omega _{l}-\left[ gW_{\mu }^{1}\left( v_{l}+\rho _{l}\right) +gW_{\mu
}^{2}\eta _{l}\right] \\
2\partial _{\mu }\eta _{l}-gW_{\mu }^{1}\omega _{l}+gW_{\mu }^{2}\xi
_{l}-\left( -gW_{\mu }^{3}+g^{\prime }Y_{l}B_{\mu }\right) \left( v_{l}+\rho
_{l}\right)%
\end{array}%
\right) ,
\end{eqnarray}%
where the $\tau ^{a}$ are the ordinary $SU(2)_{L}$ Pauli matrices and $%
Y_{l}=1$.\newline
The NHDM scalar-gauge boson interactions are given by
\begin{eqnarray}
\dsum\limits_{l=1}^{N}\left( D_{\mu }\Phi _{l}\right) \left( D^{\mu }\Phi
_{l}\right) ^{\dag } &=&\frac{1}{8}\dsum\limits_{l=1}^{N}\left\{ 2\partial
_{\mu }\omega _{l}+\left[ gW_{\mu }^{1}\eta _{l}-gW_{\mu }^{2}\left(
v_{l}+\rho _{l}\right) +\left( gW_{\mu }^{3}+g^{\prime }Y_{l}B_{\mu }\right)
\xi _{l}\right] \right\} ^{2} \notag \\
&&+\frac{1}{8}\dsum\limits_{l=1}^{N}\left\{ 2\partial _{\mu }\xi _{l}-\left(
gW_{\mu }^{3}+g^{\prime }Y_{l}B_{\mu }\right) \omega _{l}-\left[ gW_{\mu
}^{1}\left( v_{l}+\rho _{l}\right) +gW_{\mu }^{2}\eta _{l}\right] \right\}
^{2} \notag \\
&&+\frac{1}{8}\dsum\limits_{l=1}^{N}\left\{ 2\partial _{\mu }\rho _{l}+\left[
gW_{\mu }^{1}\xi _{l}+gW_{\mu }^{2}\omega _{l}-\left( gW_{\mu
}^{3}-g^{\prime }Y_{l}B_{\mu }\right) \right] \eta _{l}\right\} ^{2} \notag
\\
&&+\frac{1}{8}\dsum\limits_{l=1}^{N}\left\{ 2\partial _{\mu }\eta
_{l}-gW_{\mu }^{1}\omega _{l}+gW_{\mu }^{2}\xi _{l}-\left( -gW_{\mu
}^{3}+g^{\prime }Y_{l}B_{\mu }\right) \left( v_{l}+\rho _{l}\right) \right\}
^{2}. \label{GBSInt}
\end{eqnarray}%
The connection between the interaction and mass scalar eigenstates is
explained in what follows. The charged scalar fields of Eq. (\ref{doublets})
are linear combinations of the charged Goldstone bosons and the charged physical
scalars. The imaginary parts of the neutral component of the scalar doublets
of Eq. (\ref{doublets}) are linear combinations of the neutral Goldstone
bosons and of the CP-odd neutral scalar fields. The real parts of the neutral
component of the scalar doublets of Eq. (\ref{doublets}) are linear
combinations of the CP-odd neutral scalar fields. Within this framework we
consider a scenario where the interaction and mass eigenstates are related
in the way analogous to the two Higgs doublet model (2HDM) \cite{2N-model}
\begin{equation} \label{Neutral}
\rho _{l}=\sum_{j=1}^{N}R_{lj}H_{j}^{0},\hspace{1.5cm}\hspace{1.5cm}\eta
_{l}=Q_{l1}\pi ^{0}+\sum_{j=2}^{N}Q_{lj}A_{j-1}^{0},\hspace{2cm}l=1,2,\cdots
,N.
\end{equation}%
\begin{equation} \label{Comp-Charged}
\omega _{l}=Q_{l1}\pi ^{1}+\sum_{j=2}^{N}Q_{lj}H_{j-1}^{1},\hspace{1.5cm}%
\hspace{1.5cm}\xi _{l}=Q_{l1}\pi ^{2}+\sum_{j=2}^{N}Q_{lj}H_{j-1}^{2},%
\hspace{2cm}l=1,2,\cdots ,N.
\end{equation}%
where:%
\begin{equation} \label{vev}
v_{l}=vQ_{l1},\hspace{1.2cm}l=1,2,\cdots ,N,\hspace{1.2cm}%
v^{2}=\sum_{i=1}^{N}v_{l}^{2},\hspace{1.2cm}\sum_{l=1}^{N}R_{li}R_{lj}=%
\delta _{ij},\hspace{1.2cm}\sum_{l=1}^{N}Q_{li}Q_{lj}=\delta _{ij}.
\end{equation}
Here $v\approx 246$ GeV is the conventional electroweak symmetry breaking
scale. The fields $H_{i}^{0}$ ($i=1,2,\cdots ,N$) and $A_{j}^{0}$ ($%
j=1,2,\cdots ,N-1$) are the CP-even and CP-odd neutral Higgs bosons,
respectively. Similarly to the $W^{\pm}$ gauge bosons which are defined in terms of $W^1$ and $W^2$, the charged Higgs and Goldstone bosons are related to the component fields in (\ref{Comp-Charged}) as
\begin{eqnarray} \label{Charged}
H^{\pm}_{j} = \frac{H^{1}_{j} \mp i H^{2}_{j}}{\sqrt{2}}, \ \ \ \ \pi^{\pm}
= \frac{\pi^{1} \mp i \pi^{2}}{\sqrt{2}}
\end{eqnarray}
Thus we assumed the following:
\begin{enumerate}
\item The rotation matrix $Q$, which relates the neutral Goldstone boson $%
\pi^{0}$ and the CP odd neutral Higgses $A^{0}_{j}$ with the interaction
eigenstate scalars $\eta_l$ ($l=1,2,\cdots ,N$) in Eq. (\ref{Neutral}), is
the same as the one that relates the components of the charged Goldstone bosons $%
\pi^{1,2}$ and Higgses $H^{1,2}_{j}$ with the corresponding interaction
eigenstates $\omega_{l}, \xi_{l}$, ($l=1,2,\cdots ,N$) in Eqs. (\ref{Comp-Charged}), (\ref{Charged}).
\item The vacuum expectation values of $N$ Higgs fields $v_{l}$ ($%
l=1,2,\cdots ,N$) are related to the common symmetry breaking scale $%
v\approx 246$ GeV through the first relation in Eq. (\ref{vev}).
\end{enumerate}
Both assumptions are generalizations of the corresponding relations of the
2HDM \cite{2N-model}. In the case of NHDM, these relations are not true
everywhere in the parametric space but only in a certain part of it. Adopting the above assumptions, we limit
ourselves to a region in the parametric space of the NHDM, which is motivated
(hinted) by the 2HDM.
\section{One-loop contribution to the $T$ and $S$ parameters.}
\label{TandSparameters} In this section we calculate one-loop contributions to
the oblique parameters $T$ and $S$ defined as \cite{Peskin:1991sw,Peskin:1991sw2,epsilon-approach,epsilon-approach2,Barbieri:2004,Barbieri-book}%
:
\begin{equation}
T=\frac{\Pi _{33}\left( q^{2}\right) -\Pi _{11}\left( q^{2}\right) }{\alpha
_{EM}(M_{Z})M_{W}^{2}}\biggl|_{q^{2}=0},\ \ \ \ \ \ \ \ \ \ \ S=\frac{2\sin 2%
{\theta }_{W}}{\alpha _{EM}(M_{Z})}\frac{d\Pi _{30}\left( q^{2}\right) }{%
dq^{2}}\biggl|_{q^{2}=0}. \label{T-S-definition}
\end{equation}%
Here $\Pi _{11}\left( 0\right) $, $\Pi _{33}\left( 0\right) $, and $\Pi
_{30}\left( q^{2}\right) $ are the vacuum polarization amplitudes with $%
\{W_{\mu }^{1},W_{\mu }^{1}\}$, $\{W_{\mu }^{3},W_{\mu }^{3}\}$ and $\{W_{\mu }^{3},B_{\mu }\}$ external gauge bosons, respectively, where $q$ is
their momentum. Let us note that, in the aforementioned definitions of the oblique $T$ and $S$ parameters, it is assumed that the new physics is not light compared to $M_W$ and $M_Z$.
\subsection{$T$ parameter}
\label{subsection T-parameter}
The interaction Lagrangian, relevant for the computation of one-loop
contributions to the $T$ parameter in Eq. (\ref{T-S-definition}), is
\begin{eqnarray} \label{T-Lagrangian}
\tciLaplace _{int}^{\left( T\right) } &=&\frac{gg^{\prime }v}{2}%
\pi^{1}W^{1\mu }B_{\mu }+\frac{gg^{\prime }v}{2}
\sum_{i=1}^{N}P_{i1}H_{i}^{0}W^{3\mu }B_{\mu }+\frac{g}{2}\left(
\pi^{0}\partial _{\mu }\pi ^{1}-\pi ^{1}\partial _{\mu }\pi ^{0}\right)
W^{1\mu}+ \frac{g}{2}\sum_{i=1}^{N-1}\left( A_{i}^{0}\partial
_{\mu}H_{i}^{1}-H_{i}^{1}\partial _{\mu }A_{i}^{0}\right) W^{1\mu } \notag
\\
&&+\frac{g}{2}\sum_{i=1}^{N}P_{i1}\left( \pi ^{2}\partial
_{\mu}H_{i}^{0}-H_{i}^{0}\partial _{\mu }\pi ^{2}\right) W^{1\mu }+\frac{g}{2%
} \sum_{i=1}^{N}\sum_{j=1}^{N-1}P_{i,j+1}\left( H_{j}^{2}\partial
_{\mu}H_{i}^{0}-H_{i}^{0}\partial _{\mu }H_{j}^{2}\right) W^{1\mu } \notag
\\
&&+\frac{g}{2}\left( \pi ^{2}\partial _{\mu }\pi ^{1}-\pi ^{1}\partial
_{\mu}\pi ^{2}\right) W^{3\mu }+ \frac{g}{2}\sum_{i=1}^{N-1}\left(H_{i}^{2}%
\partial _{\mu }H_{i}^{1}-H_{i}^{1}\partial _{\mu }H_{i}^{2}\right) W^{3\mu
}+\frac{g}{2}\sum_{i=1}^{N}P_{i1}\left( H_{i}^{0}\partial _{\mu }\pi
^{0}-\pi ^{0}\partial _{\mu }H_{i}^{0}\right) W^{3\mu } \notag \\
&&+\frac{g}{2}\sum_{i=1}^{N}\sum_{j=1}^{N-1}P_{i,j+1}\left(H_{i}^{0}\partial
_{\mu }A_{j}^{0}-A_{j}^{0}\partial _{\mu }H_{i}^{0}\right) W^{3\mu }.
\end{eqnarray}
where $P_{ij}$ is given by
\begin{equation} \label{P-par}
P_{ij}=\sum_{l=1}^{N}R_{li}Q_{lj}.
\end{equation}
By definition it satisfies the inequality
\begin{eqnarray} \label{P-lim}
0 \leq P_{ij}\leq 1
\end{eqnarray}
As seen from Eq. (\ref{T-Lagrangian}), the $T$ parameter
(\ref{T-S-definition}) at one-loop level receives contributions from the diagrams shown in
Fig.~\ref{diag1}.
\begin{figure}[tbh]
\includegraphics[width=15cm,height=20cm,angle=0]{DiagT.pdf}
\vspace{-35mm} \vspace{5mm}
\caption{One-loop Feynman diagrams contributing to the $T$ parameter.}
\label{diag1}
\end{figure}
Their partial contributions, assuming the cutoff $\Lambda $ to be
much larger than the masses of the scalar particles, are
\begin{equation}
T_{\left( \pi ^{1}B\right) }\simeq -\frac{3}{16\pi \cos ^{2}\theta _{W}}\ln
\left( \frac{\Lambda ^{2}}{m_{W}^{2}}\right) ,
\end{equation}%
\begin{equation}
\sum_{i=1}^{N}T_{\left( H_{i}^{0}B\right) }\simeq \frac{3}{16\pi \cos
^{2}\theta _{W}}\sum_{i=1}^{N}P_{i1}^{2}\ln \left( \frac{\Lambda ^{2}}{%
m_{H_{i}^{0}}^{2}}\right) ,
\end{equation}%
\begin{equation}
\sum_{i=1}^{N}\sum_{j=1}^{N-1}T_{\left( H_{i}^{0}A_{j}^{0}\right) }\simeq\frac{1}{%
16\alpha _{EM}(M_{Z})\pi ^{2}v^{2}}\sum_{i=1}^{N}%
\sum_{j=1}^{N-1}P_{i,j+1}^{2}F\left( \Lambda
^{2},m_{H_{i}^{0}}^{2},m_{A_{j}^{0}}^{2}\right) ,
\end{equation}%
\begin{equation}
\sum_{i=1}^{N-1}T_{\left( H_{i}^{1}H_{i}^{2}\right) }\simeq\frac{1}{16\alpha
_{EM}(M_{Z})\pi ^{2}v^{2}}\sum_{i=1}^{N-1}G\left( \Lambda ^{2},m_{H_{i}^{\pm
}}^{2}\right) ,
\end{equation}%
\begin{equation}
\sum_{i=1}^{N}\sum_{j=1}^{N-1}T_{\left( H_{i}^{0}H_{j}^{2}\right) }\simeq-\frac{1%
}{16\alpha _{EM}(M_{Z})\pi ^{2}v^{2}}\sum_{i=1}^{N}%
\sum_{j=1}^{N-1}P_{i,j+1}^{2}F\left( \Lambda
^{2},m_{H_{i}^{0}}^{2},m_{H_{j}^{\pm }}^{2}\right) ,
\end{equation}%
\begin{equation}
\sum_{i=1}^{N-1}T_{\left( H_{i}^{1}A_{i}^{0}\right) }\simeq-\frac{1}{16\alpha
_{EM}(M_{Z})\pi ^{2}v^{2}}\sum_{i=1}^{N-1}F\left( \Lambda ^{2},m_{H_{i}^{\pm
}}^{2},m_{A_{i}^{0}}^{2}\right).
\end{equation}
The subscripts in $T_{ab}$ denote the internal lines of the diagrams in Fig.~\ref{diag1}.
The functions $F\left( \Lambda ^{2},m_{1}^{2},m_{2}^{2}\right) $ and
$G\left( \Lambda ^{2},m^{2}\right) $ are defined as
\begin{equation}
F\left( \Lambda ^{2},m_{1}^{2},m_{2}^{2}\right)=\Lambda ^{2}-\frac{m_{1}^{4}}{m_{1}^{2}-m_{2}^{2}}\ln \left( \frac{\Lambda ^{2}}{m_{1}^{2}}\right) -\frac{m_{2}^{4}}{m_{2}^{2}-m_{1}^{2}}\ln \left( \frac{\Lambda^{2}}{m_{2}^{2}}\right) ,
\end{equation}%
\begin{equation}
G\left( \Lambda ^{2},m^{2}\right) =\lim_{m_{1},m_{2}\rightarrow m}F\left(
\Lambda ^{2},m_{1},m_{2}\right) =\allowbreak \Lambda ^{2}-2m^{2}\ln \left( \frac{\Lambda ^{2}}{m^{2}}\right) +m^2
\end{equation}%
Collecting all the contributions together, we find the one-loop contribution to
the $T$ parameter coming from the scalar sector of the NHDM:
\begin{eqnarray}
T = \sum_{ab}T_{ab} &\simeq &-\frac{3}{16\pi \cos ^{2}\theta _{W}}\sum_{i=1}^{N}P_{i1}^{2}\ln
\left( \frac{m_{H_{i}^{0}}^{2}}{m_{W}^{2}}\right) +\frac{1}{16\alpha
_{EM}(M_{Z})\pi ^{2}v^{2}}\sum_{i=1}^{N-1}\left[ m_{H_{i}^{\pm
}}^{2}-h\left( m_{A_{i}^{0}}^{2},m_{H_{i}^{\pm }}^{2}\right) \right] \notag
\\
&&+\frac{1}{16\alpha _{EM}(M_{Z})\pi ^{2}v^{2}}\sum_{i=1}^{N}%
\sum_{j=1}^{N-1}P_{i,j+1}^{2}\left[ h\left(
m_{H_{i}^{0}}^{2},m_{A_{j}^{0}}^{2}\right) -h\left(
m_{H_{i}^{0}}^{2},m_{H_{j}^{\pm }}^{2}\right) \right] \notag \\
&=&-\frac{3}{16\pi \cos ^{2}\theta _{W}}\ln \left( \frac{m_{h}^{2}}{m_{W}^{2}%
}\right) +\frac{3\left( 1-P_{11}^{2}\right) }{16\pi \cos ^{2}\theta _{W}}\ln \left( \frac{m_{H^{0}}^{2}}{%
m_{h}^{2}}\right) \notag \\
&&+\frac{N-1}{16\alpha _{EM}(M_{Z})\pi ^{2}v^{2}}\left[ m_{H_{i}^{\pm
}}^{2}-h\left( m_{A_{i}^{0}}^{2},m_{H_{i}^{\pm }}^{2}\right) \right] \notag
\\
&&+\frac{1}{16\alpha _{EM}(M_{Z})\pi ^{2}v^{2}}\sum_{i=2}^{N}%
\sum_{j=1}^{N-1}P_{i,j+1}^{2}\left[ h\left(
m_{H_{i}^{0}}^{2},m_{A_{j}^{0}}^{2}\right) -h\left(
m_{H_{i}^{0}}^{2},m_{H_{j}^{\pm }}^{2}\right) \right] , \label{Th}
\end{eqnarray}%
where we identified the lightest CP-even Higgs $H_{1}^{0}=h$ with the LHC
Higgs-like particle with the mass $m_{h}=125$ GeV.
The function $h\left( m_{1}^{2},m_{2}^{2}\right) $ is given by:
\begin{equation}
h\left( m_{1}^{2},m_{2}^{2}\right) =\frac{m_{1}^{2}m_{2}^{2}}{%
m_{1}^{2}-m_{2}^{2}}\ln \left( \frac{m_{1}^{2}}{m_{2}^{2}}\right) ,\hspace{%
1.5cm}\hspace{1.5cm}\lim_{m_{2}\rightarrow m_{1}}h\left(
m_{1}^{2},m_{2}^{2}\right) =m_{1}^{2}.
\end{equation}%
We can split the $T$ parameter as $T=T_{SM}+\Delta T$, where $T_{SM} $ is
the contribution from the SM, while $\Delta T$ contain all the
contributions involving the heavy scalars:
\begin{equation}
T_{SM}=-\frac{3}{16\pi \cos ^{2}\theta _{W}}\ln \left( \frac{m_{h}^{2}}{%
m_{W}^{2}}\right) , \label{Tt}
\end{equation}%
\begin{eqnarray}
\label{Delta-T-1}
\Delta T &\simeq &-\frac{3}{16\pi \cos ^{2}\theta _{W}}%
\sum_{i=2}^{N}P_{i1}^{2}\ln \left( \frac{m_{H_{i}^{0}}^{2}}{m_{h}^{2}}%
\right) +\frac{1}{16\pi ^{2}v^{2}\alpha _{EM}(M_{Z})}\sum_{i=1}^{N-1}\left[
m_{H_{i}^{\pm }}^{2}-h\left( m_{A_{i}^{0}}^{2},m_{H_{i}^{\pm }}^{2}\right) %
\right] \\
\nonumber
&&+\frac{1}{16\pi ^{2}v^{2}\alpha _{EM}(M_{Z})}\sum_{i=1}^{N}%
\sum_{j=1}^{N-1}P_{i,j+1}^{2}\left[ h\left(
m_{H_{i}^{0}}^{2},m_{A_{j}^{0}}^{2}\right) -h\left(
m_{H_{i}^{0}}^{2},m_{H_{j}^{\pm }}^{2}\right) \right] .
\end{eqnarray}
\subsection{$S$ parameter}
\label{subsection S-parameter}
The interaction Lagrangian relevant for the computation of the one-loop
contribution to the $S$ parameter in Eq. (\ref{T-S-definition}) is
\begin{eqnarray}
\tciLaplace _{int}^{\left( S\right) } &=&\frac{g}{2}\left( \pi ^{2}\partial
_{\mu }\pi ^{1}-\pi ^{1}\partial _{\mu }\pi ^{2}\right) W^{3\mu }+\frac{g}{2}%
\sum_{i=1}^{N-1}\left( H_{i}^{2}\partial _{\mu }H_{i}^{1}-H_{i}^{1}\partial
_{\mu }H_{i}^{2}\right) W^{3\mu } \notag \\
&&+\frac{g^{\prime }}{2}\left( \pi ^{2}\partial _{\mu }\pi ^{1}-\pi
^{1}\partial _{\mu }\pi ^{2}\right) B^{\mu }+\frac{g^{\prime }}{2}%
\sum_{i=1}^{N-1}\left( H_{i}^{2}\partial _{\mu }H_{i}^{1}-H_{i}^{1}\partial
_{\mu }H_{i}^{2}\right) B^{\mu } \notag \\
&&+\frac{g}{2}\sum_{i=1}^{N}P_{i1}\left( H_{i}^{0}\partial _{\mu }\pi
^{0}-\pi ^{0}\partial _{\mu }H_{i}^{0}\right) W^{3\mu }+\frac{g}{2}%
\sum_{i=1}^{N}\sum_{j=1}^{N-1}P_{i,j+1}\left( H_{i}^{0}\partial _{\mu
}A_{j}^{0}-A_{j}^{0}\partial _{\mu }H_{i}^{0}\right) W^{3\mu } \notag \\
&&-\frac{g^{\prime }}{2}\sum_{i=1}^{N}P_{i1}\left( H_{i}^{0}\partial _{\mu
}\pi ^{0}-\pi ^{0}\partial _{\mu }H_{i}^{0}\right) B^{\mu }-\frac{g^{\prime }%
}{2}\sum_{i=1}^{N}\sum_{j=1}^{N-1}P_{i,j+1}\left( H_{i}^{0}\partial _{\mu
}A_{j}^{0}-A_{j}^{0}\partial _{\mu }H_{i}^{0}\right) B^{\mu }.
\end{eqnarray}
As follows from this Lagrangian and the definition (\ref{T-S-definition}),
the $S$ parameter at one-loop level receives contributions from the diagrams shown in
Fig.~\ref{diag-2}.
\begin{figure}[tbh]
\centering
\vspace{-30mm} \vspace{0mm}%
\includegraphics[width=15cm,height=20cm,angle=0]{DiagS.pdf}
\vspace{-60mm} \vspace{10mm}
\caption{One-loop Feynman diagrams contributing to the $S$ parameter.}
\label{diag-2}
\end{figure}
Their partial contributions, assuming the cutoff $\Lambda $ to be
much larger than the masses of the scalar particles, are
\begin{equation}
S_{\left( \pi ^{1}\pi ^{2}\right) }\simeq \frac{1}{12\pi }\ln \left( \frac{%
\Lambda ^{2}}{m_{W}^{2}}\right) ,
\end{equation}%
\begin{equation}
\sum_{i=1}^{N-1}S_{\left( H_{i}^{1}H_{i}^{2}\right) }\simeq \frac{1}{12\pi }%
\sum_{i=1}^{N-1}\ln \left( \frac{\Lambda ^{2}}{m_{H_{i}^{\pm }}^{2}}\right) ,
\end{equation}%
\begin{equation}
\sum_{i=1}^{N}S_{\left( H_{i}^{0}\pi ^{0}\right) }\simeq -\frac{1}{12\pi }%
\sum_{i=1}^{N}P_{i1}^{2}\ln \left( \frac{\Lambda ^{2}}{m_{H_{i}^{0}}^{2}}%
\right) ,
\end{equation}%
\begin{eqnarray}
\sum_{i=1}^{N}\sum_{j=1}^{N-1}S_{\left( H_{i}^{0}A_{j}^{0}\right) } &\simeq
&-\frac{1}{12\pi }\sum_{i=1}^{N}\sum_{j=1}^{N-1}\frac{P_{i,j+1}^{2}}{\left(
m_{A_{j}^{0}}^{2}-m_{H_{i}^{0}}^{2}\right) {}^{3}}\left\{ m_{A_{j}^{0}}^{6}%
\left[ \ln \left( \frac{\Lambda ^{2}}{m_{A_{j}^{0}}^{2}}\right) +\frac{5}{6}%
\right] -m_{H_{i}^{0}}^{6}\left[ \ln \left( \frac{\Lambda ^{2}}{%
m_{H_{i}^{0}}^{2}}\right) +\frac{5}{6}\right] \right. \notag \\
&&+\left. 3m_{H_{i}^{0}}^{2}m_{A_{j}^{0}}^{2}\left[ m_{H_{i}^{0}}^{2}\left[
\ln \left( \frac{\Lambda ^{2}}{m_{H_{i}^{0}}^{2}}\right) +\frac{3}{2}\right]
-m_{A_{j}^{0}}^{2}\left[ \ln \left( \frac{\Lambda ^{2}}{m_{A_{j}^{0}}^{2}}%
\right) +\frac{3}{2}\right] \right] \right\} .
\end{eqnarray}
As before, the subscripts in $S_{ab}$ denote the internal lines of the diagrams in
Fig.~\ref{diag-2}.
Then, the 1-loop Higgs contribution to the $S$ parameter in the NHDM is
\begin{eqnarray}
S = \sum_{ab}S_{ab}&\simeq &\frac{1}{12\pi }\left[ \sum_{i=1}^{N}P_{i1}^{2}\ln \left( \frac{
m_{H_{i}^{0}}^{2}}{m_{W}^{2}}\right)
+\sum_{i=1}^{N}\sum_{j=1}^{N-1}P_{i,j+1}^{2}K\left(
m_{H_{i}^{0}}^{2},m_{A_{j}^{0}}^{2},m_{H_{j}^{\pm }}^{2}\right) \right]
\notag \\
&=&\frac{1}{12\pi }\ln \left( \frac{m_{h}^{2}}{m_{W}^{2}}\right) +\frac{1}{%
12\pi }\left[ \sum_{i=2}^{N}P_{i1}^{2}\ln \left( \frac{m_{H_{i}^{0}}^{2}}{%
m_{h}^{2}}\right) +\sum_{i=1}^{N}\sum_{j=1}^{N-1}P_{i,j+1}^{2}K\left(
m_{H_{i}^{0}}^{2},m_{A_{j}^{0}}^{2},m_{H_{j}^{\pm }}^{2}\right) \right],
\label{Sh}
\end{eqnarray}
where we identified the lightest CP-even Higgs $H_{1}^{0}=h$ with the LHC
Higgs-like particle with the mass $m_{h}=125$ GeV. We defined a function
\begin{eqnarray}
K\left( m_{1}^{2},m_{2}^{2},m_{3}^{2}\right) &=&\frac{1}{\left(
m_{2}^{2}-m_{1}^{2}\right) {}^{3}}\left\{ m_{1}^{4}\left(
3m_{2}^{2}-m_{1}^{2}\right) \ln \left( \frac{m_{1}^{2}}{m_{3}^{2}}\right)
-m_{2}^{4}\left( 3m_{1}^{2}-m_{2}^{2}\right) \ln \left( \frac{m_{2}^{2}}{%
m_{3}^{2}}\right) \right. \notag \\
&&-\left. \frac{1}{6}\left[ 27m_{1}^{2}m_{2}^{2}\left(
m_{1}^{2}-m_{2}^{2}\right) +5\left( m_{2}^{6}-m_{1}^{6}\right) \right]
\right\} ,
\end{eqnarray}%
with the properties
\begin{eqnarray}
\lim_{m_{1}\rightarrow m_{2}}K(m_{1}^{2},m_{2}^{2},m_{3}^{2})
&=&K_{1}(m_{2}^{2},m_{3}^{2})=\ln \left( \frac{m_{2}^{2}}{m_{3}^{2}}\right) ,
\notag \\
\lim_{m_{2}\rightarrow m_{3}}K(m_{1}^{2},m_{2}^{2},m_{3}^{2})
&=&K_{2}(m_{1}^{2},m_{3}^{2})=\frac{%
-5m_{1}^{6}+27m_{1}^{4}m_{3}^{2}-27m_{1}^{2}m_{3}^{4}+6\left(
m_{1}^{6}-3m_{1}^{4}m_{3}^{2}\right) \ln \left( \frac{m_{1}^{2}}{m_{3}^{2}}%
\right) +5m_{3}^{6}}{6\left( m_{1}^{2}-m_{3}^{2}\right) ^{3}}, \notag \\
\lim_{m_{1}\rightarrow m_{3}}K(m_{1}^{2},m_{2}^{2},m_{3}^{2})
&=&K_{2}(m_{2}^{2},m_{3}^{2}).
\end{eqnarray}
We can split the $S$ parameter as $S=S_{SM}+\Delta S$, where $S_{SM}$\ is
the contribution from the SM, while $\Delta S$ contain all the
contributions involving the heavy scalars:
\begin{equation}
S_{SM}=\frac{1}{12\pi }\ln \left( \frac{m_{h}^{2}}{m_{W}^{2}}\right) ,
\end{equation}
\begin{equation}
\label{Delta-S-1}
\Delta S\simeq \frac{1}{12\pi }\left[ \sum_{i=2}^{N}P_{i1}^{2}\ln \left(
\frac{m_{H_{i}^{0}}^{2}}{m_{h}^{2}}\right)
+\sum_{i=1}^{N}\sum_{j=1}^{N-1}P_{i,j+1}^{2}K\left(
m_{H_{i}^{0}}^{2},m_{A_{j}^{0}}^{2},m_{H_{j}^{\pm }}^{2}\right) \right] .
\end{equation}
\section{T and S bounds on NHDM}
\label{T-S-exp bounds}
\begin{figure}[tbh]
\centering
\includegraphics[width=10cm,height=10cm,angle=0]{Ellipse.pdf}
\caption{The interior of the ellipse in $\Delta S-\Delta T$ plane is the experimentally allowed region at 95\%C.L. from Ref. \protect\cite{Baak:2012kk}. The
reference point $\Delta S=\Delta T=0$ is conventionally taken to the Standard Model value of $\Delta S$ and $\Delta T$, at $m_h=125.7$GeV and $m_t=173.18GeV$}
\label{diag2}
\end{figure}
The inclusion of the extra scalar particles modifies the SM predictions for the oblique parameters $T$ and $S$, and therefore
their values extracted from high
precision measurements can be used to constrain
the $N$ Higgs extension of the SM. Our goal is to examine if these measurements are able to restrict the number $N$ of Higgs electroweak doublets.
The experimental results on $T$ and $S$ restrict the deviations
$\Delta T$ and $\Delta S$ from the SM predictions to lie inside a region in
the $\Delta S-\Delta T$
plane. At the $95\%$C.L.,
this region is the elliptic contour shown in
Fig.~\ref{diag2}, taken from Ref.~\cite{Baak:2012kk}. The reference point $\Delta S=\Delta T=0$ is conventionally taken to be the SM value of $\Delta S$ and $\Delta T$ at $m_h=125.7$GeV and $m_t=173.18GeV$.
In view of the complexity of the general case of the $N$ Higgs doublet model, we consider several benchmark scenarios described below.
\subsection{All the heavy Higgses are degenerate.}
\label{subsect: Senario-1} This is the most simple case of the Higgs
spectrum with the lightest CP-even Higgs $H_{1}^{0}=h$ identified with the LHC
Higgs-like particle, with a mass $m_{h}=125.7$ GeV and all the other heavier
Higgses degenerate having a common mass $m_{H}$. Thus,
\begin{eqnarray}
&&m_{H_{j}^{\pm }}=m_{A_{j}^{0}}=m_{H},\hspace{4.4cm}j=1,2,\cdots ,N-1,
\label{Scenario-1} \\
\nonumber
&&m_{H_{1}^{0}}\ =m_{h}=125.7GeV,\hspace{0.5cm}m_{H_{i}^{0}}=m_{H},\hspace{%
1.5cm}i=2,\cdots ,N, \\
\nonumber
&&m_{H}> m_{h}.
\end{eqnarray}%
For this spectrum, Eqs. (\ref{Delta-T-1}) and (\ref{Delta-S-1}) for the $\Delta T$ and $\Delta S$ parameters
are drastically simplified and take the form
\begin{eqnarray}
\Delta T &=&-\frac{3\left( 1-P_{11}^{2}\right) }{16\pi \cos ^{2}\theta _{W}}
\ln \left( \frac{m_{H}^{2}}{m_{h}^{2}}\right) , \label{Delta S-S1} \\
\Delta S &=&\frac{1-P_{11}^{2}}{12\pi }\left[ \ln \left( \frac{m_{H}^{2}}
{m_{h}^{2}}\right) +K_{2}\left( m_{h}^{2},m_{H}^{2}\right) \right].
\end{eqnarray}
Thus, in this scenario neither of the two parameters $S$ and $T$ depends on $N$. Therefore, the spectrum in Eq. (\ref{Scenario-1}) does not constrain the
number of Higgs doublets.
\subsection{Degeneracy inside the groups of the heavy CP-even, CP-odd and charged Higgses.}
\label{subsect: Scenario-2}
{{\bf Subscenario B1:} \it The CP-even and CP-odd neutral Higgses are degenerate.}\\
The next-to-simplest scenario that we consider has the following Higgs spectrum:
\begin{eqnarray}
\label{spectrum-B1}
&&m_{A_{j}^{0}}=m_{H},\hspace{9mm}\hspace{1.5cm}m_{H_{j}^{\pm
}}=m_{H}+\Delta ,\hspace{1.5cm}\hspace{1.5cm}j=1,2,\cdots ,N-1, \label{s2a}
\\
&&m_{H_{1}^{0}}\ =m_{h}=125.7GeV,\hspace{1.4cm}m_{H_{i}^{0}}=m_{H},\hspace{%
3.7cm}i=2,\cdots ,N, \notag \\
&&m_{H}>m_{h}. \notag
\end{eqnarray}
This spectrum, using Eqs. (\ref{Delta-T-1}) and (\ref{Delta-S-1}), leads to the expressions
\begin{figure}[tbh]
\centering
\includegraphics[width=15cm,height=10cm,angle=0]{NmaxvsDeltacaso1.pdf}
\caption{Upper bound $N\leq N_{max}$ on the number $N$ of Higgs doublets, obtained from $T$ and
$S $ at $95\%$C.L., using the experimental constraints indicated in Fig.~\ref{diag2} for the Higgs spectrum in Eq.~(\ref{spectrum-B1}). }
\label{diagcaso1}
\end{figure}
\begin{eqnarray}
\label{Delta-spectrum-B1}
\Delta T &=&-\frac{3\left( 1-P_{11}^{2}\right) }{16\pi \cos ^{2}\theta _{W}}%
\ln \left( \frac{m_{H}^{2}}{m_{h}^{2}}\right) +\frac{N-1}{16\pi
^{2}v^{2}\alpha _{EM}(M_{Z})}\left[ m_{H^{\pm }}^{2}+m_{H}^{2}-2h\left(
m_{H}^{2},m_{H^{\pm }}^{2}\right) \right] \\
&&+\frac{1-P_{11}^{2}}{16\pi ^{2}v^{2}\alpha _{EM}(M_{Z})}\left[ h\left(
m_{h}^{2},m_{H}^{2}\right) -h\left( m_{h}^{2},m_{H^{\pm }}^{2}\right)
-m_{H}^{2}+h\left( m_{H}^{2},m_{H^{\pm }}^{2}\right) \right] , \notag \\
\Delta S &=&\frac{1}{12\pi }\left\{ \left( 1-P_{11}^{2}\right) \left[ \ln
\left( \frac{m_{H}^{2}}{m_{h}^{2}}\right) +K\left(
m_{h}^{2},m_{A}^{2},m_{H^{\pm }}^{2}\right) -K_{1}\left( m_{H}^{2},m_{H^{\pm
}}^{2}\right) \right] +\left( N-1\right) K_{1}\left( m_{H}^{2},m_{H^{\pm
}}^{2}\right) \right\} , \notag
\end{eqnarray}%
with
\begin{equation}
P_{11}=\sum_{l=1}^{N}R_{l1}Q_{l1}=\frac{1}{v}\sum_{l=1}^{N}R_{l1}v_{l},
\end{equation}
where we used the first relation in Eq. (\ref{vev}).
Now, using Eqs.~(\ref{Delta-spectrum-B1}), we find the maximal values $N_{max}$ of the Higgs doublets $N$ compatible with the precision data in Fig~\ref{diag2}. We scan the parameter space within
\begin{eqnarray}\label{scan-sp-2-1}
&& 0 \leq P_{11} \leq 1, \ \ \ 600 \mbox{GeV} \leq m_{H}\leq 1 \mbox{TeV}.
\end{eqnarray}
In Fig.~\ref{diagcaso1}, we show the resulting values $N_{max}$ in a function of the splitting parameter $\Delta$.
It is noteworthy that the maximum number of Higgs doublets
decreases when the mass splitting $\Delta$ between the heavy physical scalars is increased.
This behavior follows from the fact that increasing the number of Higgs doublets yields an increase of the $T$ and $S$ oblique parameters.
\quad In the limit $\Delta \rightarrow 0$, we find $N_{max} \rightarrow \infty$, corresponding to
no limits on $N$, which is consistence with the scenario (\ref{Scenario-1}). \\[3mm]
{{\bf Subscenario B2:} \it The CP-even neutral and charged Higgses are
degenerate:}
\begin{eqnarray}
\label{spectrum-B2}
&&m_{H_{j}^{\pm }}=m_{H},\hspace{9mm}\hspace{1.5cm}m_{A_{j}^{0}}=m_{H}+%
\Delta ,\hspace{1.5cm}\hspace{1.5cm}j=1,2,\cdots ,N-1, \label{s2b} \\
&&m_{H_{1}^{0}}\ =m_{h}=125.7GeV,\hspace{1.4cm}m_{H_{i}^{0}}=m_{H},\hspace{%
3.7cm}i=2,\cdots ,N, \notag \\
&&m_{H} > m_{h}. \notag
\end{eqnarray}
\begin{figure}[tbh]
\centering
\includegraphics[width=15cm,height=10cm,angle=0]{NmaxvsDeltacaso2.pdf}
\caption{The same as in Fig.~\ref{diagcaso1}, but for the spectrum in
Eq.~(\ref{spectrum-B2}). }
\label{diagcaso2}
\end{figure}
From Eqs. (\ref{Delta-T-1}) and (\ref{Delta-S-1}), we find for this spectrum
\begin{eqnarray}
\label{Delta-spectrum-B2}
\Delta T &=&-\frac{3\left( 1-P_{11}^{2}\right) }{16\pi \cos ^{2}\theta _{W}}%
\ln \left( \frac{m_{H}^{2}}{m_{h}^{2}}\right) +\frac{1-P_{11}^{2}}{16\pi
^{2}v^{2}\alpha _{EM}(M_{Z})}\left[ h\left( m_{h}^{2},m_{A}^{2}\right)
-h\left( m_{h}^{2},m_{H}^{2}\right) +m_{H}^{2}-h\left(
m_{H}^{2},m_{A}^{2}\right) \right] , \notag \\
\Delta S &=&\frac{1}{12\pi }\left\{ \left( 1-P_{11}^{2}\right) \left[ \ln
\left( \frac{m_{H}^{2}}{m_{h}^{2}}\right) +K\left(
m_{h}^{2},m_{A}^{2},m_{H}^{2}\right) \right] +\left( N-2+P_{11}^{2}\right)
K_2\left( m_{H}^{2},m_{A}^{2}\right) \right\}\label{DeltaTS}.
\end{eqnarray}
Scanning the parameter space in the region (\ref{scan-sp-2-1}), we find the maximal values on the number $N$ of Higgs doublets compatible with the data in Fig.~\ref{diag2}. The results are shown in Fig.~\ref{diagcaso2}. As seen, this spectrum is significantly less restrictive for
$N$ than that in Eq. (\ref{spectrum-B1}). This is mainly because of the fact that only the $S$ parameter depends on $N$ in the present case, while in the case of the spectrum
(\ref{spectrum-B1}) both $T$ and $S$ are $N$-dependent.\\[3mm]
{{\bf Subscenario B3:} \it The CP-odd neutral and charged Higgses are
degenerate:}
\begin{eqnarray}
\label{spectrum-B3}
&&m_{H_{j}^{\pm }}=m_{H},\hspace{9mm}\hspace{1.5cm}m_{A_{j}^{0}}=m_{H},%
\hspace{1.5cm}\hspace{1.5cm}j=1,2,\cdots ,N-1, \label{s2c} \\
&&m_{H_{1}^{0}}\ =m_{h}=125.7GeV,\hspace{1.4cm}m_{H_{i}^{0}}=m_{H}+\Delta ,%
\hspace{3.7cm}i=2,\cdots ,N, \notag \\
&&m_{H} > m_{h}.\notag
\end{eqnarray}
In this case we find from Eqs. (\ref{Delta-T-1}) and (\ref{Delta-S-1})
\begin{eqnarray}
\label{Delta-spectrum-B3}
\Delta T &\simeq &-\frac{3\left( 1-P_{11}^{2}\right) }{16\pi \cos ^{2}\theta
_{W}}\ln \left( \frac{m_{H^{0}}^{2}}{m_{h}^{2}}\right) \notag \\
\Delta S &\simeq &\frac{1}{12\pi }\left\{ \left( 1-P_{11}^{2}\right) \left[
\ln \left( \frac{m_{H^{0}}^{2}}{m_{h}^{2}}\right) +K_2\left(
m_{h}^{2},m_{H}^{2}\right) \right] +\left( N-2+P_{11}^{2}\right)
K_2\left( m_{H^{0}}^{2},m_{H}^{2}\right) \right\} .
\end{eqnarray}
\begin{figure}[tbh]
\centering
\includegraphics[width=15cm,height=10cm,angle=0]{NmaxvsDeltacaso3.pdf}
\caption{The same as in Fig.~\ref{diagcaso1}, but for the spectrum in
Eq.~(\ref{spectrum-B3}). }
\label{diagcaso3}
\end{figure}
Applying the same procedure as previously, we scan the parameter space in the region (\ref{scan-sp-2-1}) and find the maximal values on the number $N$ of the Higgs doublets compatible with the data of Fig.~\ref{diag2}. The results are shown in Fig.~\ref{diagcaso3}.
Again, as in the case (\ref{spectrum-B2}), the parameter $T$ is independent of $N$.
As a consequence, the limits on $N$ for the spectrum (\ref{spectrum-B3}) are significantly weaker than for (\ref{spectrum-B1}).\\[3mm]
{{\bf Subscenario B4:} \it Split groups with the degenerate interior:}
\begin{eqnarray}
&\mbox{Case 1}: & m_{A_{j}^{0}}=m_{H}+\Delta ,\hspace{9mm}\hspace{1.5cm}
m_{H_{j}^{\pm}}=m_{H}+a\Delta ,\hspace{1.5cm} j=1,2,\cdots ,N-1, \notag \\
&&m_{H_{1}^{0}}\ =m_{h}=125.7GeV,\hspace{1.4cm}m_{H_{i}^{0}}=m_{H},
\hspace{3.7cm}i=2,\cdots ,N, \notag \\
&&m_{H} > m_{h}. \label{spectrum-B4}\\[3mm]
&\mbox{Case 2}: & m_{H_{i}^{0}} =m_{H}+\Delta ,\hspace{9mm}\hspace{1.5cm}m_{H_{j}^{\pm}}=m_{H}+a\Delta ,\hspace{1.5cm} j=1,2,\cdots ,N-1, \notag \\
&&m_{H_{1}^{0}}\ =m_{h}=125.7GeV,\hspace{1.4cm}m_{A_{j}^{0}}=m_{H},
\hspace{3.7cm}i=2,\cdots ,N, \notag \\
&&m_{H} > m_{h}. \label{spectrum-B5}\\[3mm]
&\mbox{Case 3}: & m_{H_{i}^{0}} =m_{H}+\Delta ,\hspace{9mm}\hspace{1.5cm}
m_{A_{j}^{0}}=m_{H}+a\Delta ,\hspace{1.5cm}
j=1,2,\cdots ,N-1, \notag \\
&&m_{H_{1}^{0}}\ =m_{h}=125.7GeV,\hspace{1.4cm}m_{H_{j}^{\pm }}=m_{H},
\hspace{3.7cm}i=2,\cdots ,N, \notag \\
&&m_{H} > m_{h}. \label{spectrum-B6}
\end{eqnarray}
For all these cases we find from Eqs. (\ref{Delta-T-1}) and (\ref{Delta-S-1})
\begin{eqnarray}
\label{Delta-spectrum-B4}
\Delta T &=&-\frac{3\left( 1-P_{11}^{2}\right) }{16\pi \cos ^{2}\theta _{W}}%
\ln \left( \frac{m_{H^{0}}^{2}}{m_{h}^{2}}\right) + \notag \\
&&+\frac{1-P_{11}^{2}}{16\pi ^{2}v^{2}\alpha _{EM}(M_{Z})}\left[ h\left(
m_{H^{0}}^{2},m_{H^{\pm }}^{2}\right) -h\left(
m_{H^{0}}^{2},m_{A}^{2}\right) +h\left( m_{h}^{2},m_{A}^{2}\right) -h\left(
m_{h}^{2},m_{H^{\pm }}^{2}\right) \right] + \notag \\
&&+\frac{N-1}{16\pi ^{2}v^{2}\alpha _{EM}(M_{Z})}\left[ m_{H^{\pm
}}^{2}-h\left( m_{A}^{2},m_{H^{\pm }}^{2}\right) +h\left(
m_{H^{0}}^{2},m_{A}^{2}\right) -h\left( m_{H^{0}}^{2},m_{H^{\pm
}}^{2}\right) \right] , \notag \\[0.12in]
\Delta S &=&\frac{1}{12\pi }\left\{ \left( 1-P_{11}^{2}\right) \left[ \ln
\left( \frac{m_{H^{0}}^{2}}{m_{h}^{2}}\right) +K\left(
m_{h}^{2},m_{A}^{2},m_{H^{\pm }}^{2}\right) \right] +\left(
N-2+P_{11}^{2}\right) K\left( m_{H^{0}}^{2},m_{A}^{2},m_{H^{\pm
}}^{2}\right) \right\}.
\end{eqnarray}
As seen from Eq. (\ref{Delta-spectrum-B4}), now both parameters $T$ and $S$ depend
linearly on $N$, the number of Higgs doublets.
Scanning the parameter space in the region (\ref{scan-sp-2-1}), we find the results for several sample values of the parameter $a = 0.5, 1, 2$ shown in Figs.~\ref{diagpdg1}-\ref{diagpdg3}.
Note the general tendency: the large splitting, corresponding to the larger values of $\Delta$ and
$a$, leads to more stringent constraints on $N$.
\begin{figure}[tbh]
\centering
\includegraphics[width=9cm,height=9cm,angle=0]{NmaxvsDeltapdgcase1ab.pdf}%
\includegraphics[width=9cm,height=9cm,angle=0]{NmaxvsDeltapdgcase1bc.pdf}
\hspace{2cm}{\footnotesize {$a=1$}}\hspace{9cm}{\footnotesize {$a=0.5,2$}}%
\hspace{0cm}\newline
\caption{The same as in Fig.~\ref{diagcaso1}, but for the spectrum in
Eq.~(\ref{spectrum-B4}). The left panel is for $a = 1$; the right panel is for $a = 0.5$ (dashed line), $a = 2$ (solid line). }
\label{diagpdg1}
\end{figure}
\begin{figure}[tbh]
\centering
\includegraphics[width=9cm,height=9cm,angle=0]{NmaxvsDeltapdgcase2ab.pdf}%
\includegraphics[width=9cm,height=9cm,angle=0]{NmaxvsDeltapdgcase2bc.pdf}
{\footnotesize {$a=1$}}\hspace{9cm}{\footnotesize {$a=0.5,2$}}\hspace{0cm}%
\caption{The same as in Fig.~\ref{diagpdg1}, but for the spectrum in
Eq.~(\ref{spectrum-B5}).
}
\label{diagpdg2}
\end{figure}
\begin{figure}[tbh]
\centering
\includegraphics[width=15cm,height=10cm,angle=0]{NmaxvsDeltapdgcase3.pdf}
\caption{The same as in Fig.~\ref{diagcaso1}, but for the spectrum in
Eq.~(\ref{spectrum-B6}). The curves from the bottom to the top
correspond to $a =0.5,1,2$, respectively. }
\label{diagpdg3}
\end{figure}
\label{Scenario-3}
\begin{figure}[tbh]
\centering
\hspace{-20mm}%
\includegraphics[width=15cm,height=10cm,angle=0]{NmaxvsDeltagen.pdf}
\caption{The same as in Fig.~\ref{diagcaso1}, but for the spectrum Eq.~(\ref{No-Deg}).
The curves from the bottom to the top correspond to $n= 20,10,2$,
respectively.}
\label{diagnodeg}
\end{figure}
\subsection{No degeneracy with a particular structure of the spectrum}
Finally, let us consider a benchmark scenario in which all the Higgses are
nondegenerate. Since the general case can hardly be analyzed, we consider a particular structure of the Higgs spectrum:
\begin{eqnarray}
\label{spectrum-C}
m_{H_{1}^{0}} &=&m_{h}=125.7GeV,\hspace{1.5cm}\hspace{3mm}m_{H_{i}^{0}}\
=m_{H}+2\left( i-2\right) \Delta ,\hspace{1.5cm}i=2,\cdots ,N,
\label{No-Deg} \\
\nonumber
m_{A_{j}^{0}} &=&m_{H}+\left( 2j-1\right) \Delta ,\hspace{1.5cm}%
m_{H_{j}^{\pm }}=m_{H}+2\left( j-1\right) \Delta +\delta ,\hspace{8mm}%
j=1,2,\cdots ,N-1,\\
\nonumber
m_{H} &>& m_{h}.
\end{eqnarray}
This is a spectrum equidistant in each of the three groups of Higgses
$\{A^{0}_{j}\}, \{H^{0}_{i}\}, \{H^{\pm}_{j}\}$ with a step $2\Delta$. The bands of these groups overlap with each other. The spectrum is characterized by two parameters $\Delta$
and $\delta$.
For simplicity, we assume
\begin{equation}
P_{i,j}=\delta _{ij},\hspace{1.5cm}\hspace{1.5cm}i,j=i=1,2,\cdots ,N
\end{equation}
Then from Eqs. (\ref{Delta-T-1}) and (\ref{Delta-S-1}) we get
\begin{eqnarray}
\Delta T &\simeq &\frac{1}{16\pi ^{2}v^{2}\alpha _{EM}(M_{Z})}%
\sum_{i=1}^{N-1}\left[ m_{H_{i}^{\pm }}^{2}-h\left(
m_{A_{i}^{0}}^{2},m_{H_{i}^{\pm }}^{2}\right) +h\left(
m_{H_{i+1}^{0}}^{2},m_{A_{i}^{0}}^{2}\right) -h\left(
m_{H_{i+1}^{0}}^{2},m_{H_{i}^{\pm }}^{2}\right) \right] , \label{Delta S-S3}
\\
\Delta S &\simeq &\frac{1}{12\pi }\sum_{i=1}^{N-1}K\left(
m_{H_{i+1}^{0}}^{2},m_{A_{i}^{0}}^{2},m_{H_{i}^{\pm }}^{2}\right) .
\end{eqnarray}
Scanning the parameter space in the region (\ref{scan-sp-2-1}), we find the results for several sample values of the parameter $\delta = \Delta/n$
shown in Fig.~\ref{diagnodeg}.
The curves from the bottom to the top correspond to
$n=20,10,2$, respectively.
With the larger value $n=50$, we find
$N_{max} \sim 570$ for $\Delta =20$ GeV.
\section{Summary and Conclusions}
\label{Conclusions}
We have considered an $N$ Higgs $SU(2)$ doublet model (NHDM) with arbitrary number
$N$. In this model, we calculated the one-loop contributions $\Delta S$ and $\Delta T$ of the Higgs doublets to the electroweak oblique parameters $S$ and $T$. The calculated contribution depends on the number $N$ of Higgs doublets, and, therefore, our results can be used to constrain $N$ from data on the precision measurements of the parameters $T$ and $S$.
Within the generic case of the NHDM, due to the large number of free parameters, this program can hardly be realized. For this reason, we have analyzed several benchmark scenarios with particular mass spectra [Eqs.~(\ref{Scenario-1}), (\ref{spectrum-B1}), (\ref{spectrum-B2}), (\ref{spectrum-B3}), (\ref{spectrum-B4})-(\ref{spectrum-B6}), and (\ref{No-Deg})] of the physical scalars of the NHDM, including some other simplifying assumptions, inspired by the well-known case of the 2HDM, about
the physical Higgs mixing and the vacuum structure of the model. These scenarios correspond to certain domains of the NHDM parameter space. We have shown that, except for a very particular ``fine-tuned'' case with all the physical heavy Higgses degenerate [Eq.~(\ref{Scenario-1})], these scenarios imply constraints on the number of Higgs doublets $N$, in order to be compatible with the existing data on the precision measurements of $T$ and $S$.
We presented our results on $N\leq N_{max}$
in Figs.~\ref{diagcaso1}-\ref{diagnodeg} as functions of the mass splitting parameter $\Delta$.
The general feature of our results is that the maximal number $N_{max}$ of Higgs doublets is a monotonically increasing function for small values of the splitting
\mbox{$\Delta \leq 20$ GeV} and a monotonically decreasing one for larger values. Thus, the data on $T$ and $S$ are able to accommodate an arbitrary large number $N$ with decreasing splitting between the masses of the physical scalars, and, vice versa, $N$ becomes stringently constrained in those parts of the NHDM parameter space with large mass splitting in the scalar sector. The same tendency is demonstrated by the plots with respect to their dependence on the two other additional parameters $a$ and $\delta$, characterizing the Higgs mass spectrum: a smaller mass splitting corresponds to a larger number $N$ compatible with the analyzed data and visa versa. This is the main message of the present study. Also worth mentioning is that the maximal number $N_{max}$ of Higgs doublets is exactly the same when the charged Higgses are degenerate between either the CP-even or the CP-odd neutral Higgses, as shown in Figs.~\ref{diagcaso2},\ref{diagcaso3}, and larger than the obtained in the scenario of CP-odd - CP-even neutral Higgses degeneracy. Consequently, the tightest $T$ and $S$ oblique parameter constraints arise in the scenario where the charged Higgses are split in mass between either of the neutral CP-even or CP-odd Higgses, as shown in Fig.~\ref{diagcaso1}. This indicates that making the charged Higgses degenerate between either the neutral CP-even or the neutral CP odd Higgses helps avoid constraints from electroweak precision observables, a feature already present in the 2HDM.
\quad Our analysis cannot exclude a deviation from this tendency in certain parts of the NHDM parameter space; however, in our opinion, this should be related with certain ``fine-tuning'' of the parameters, as in the case of the spectrum (\ref{Scenario-1}). Naturally, the concrete limit on $N$ depends on a particular scenario within the generic NHDM framework. We hope our results will help examine such scenarios regarding their consistence with the present and future data on the precision measurements of the electroweak oblique parameters $T$ and $S$.
\quad As a final remark, let us note that the maximum number of Higgs multiplets is constrained from the requirement of perturbative unitarity in the scattering amplitudes of two transversely polarized $W$ bosons into a scalar pair, as shown in detail in Ref. \cite{Hally:2012pu}. From the results given in Ref. \cite{Hally:2012pu}, it follows that the maximum number of Higgs doublets consistent with the aforementioned perturbative unitarity requirement is 2307, which is larger than our obtained upper bounds on the number of Higgs doublets, from oblique $T$ and $S$ parameter constraints, for the several benchmark scenarios considered in this paper. Consequently, our
upper bounds on the number of Higgs doublets
are consistent with the
perturbative unitarity in the scattering amplitudes of two transversely polarized $W$ bosons into a scalar pair.
\begin{acknowledgments}
This work has been partially supported by \mbox{FONDECYT} Projects No. 11130115, No. 1150792 and No. 110287, Centro-Cient\'{\i}fico-Tecnol\'{o}gico de Valpara\'{\i}so and DGIP internal Grant No. 111458. We thank an anonymous referee for very valuable comments.
\end{acknowledgments}
\section*{Abstract (Not appropriate in this style!)}%
\else \small
\begin{center}{\bf Abstract\vspace{-.5em}\vspace{\z@}}\end{center}%
\quotation
\fi
}%
}{%
}%
\@ifundefined{endabstract}{\def\endabstract
{\if@twocolumn\else\endquotation\fi}}{}%
\@ifundefined{maketitle}{\def\maketitle#1{}}{}%
\@ifundefined{affiliation}{\def\affiliation#1{}}{}%
\@ifundefined{proof}{\def\proof{\noindent{\bfseries Proof. }}}{}%
\@ifundefined{endproof}{\def\endproof{\mbox{\ \rule{.1in}{.1in}}}}{}%
\@ifundefined{newfield}{\def\newfield#1#2{}}{}%
\@ifundefined{chapter}{\def\chapter#1{\par(Chapter head:)#1\par }%
\newcount\c@chapter}{}%
\@ifundefined{part}{\def\part#1{\par(Part head:)#1\par }}{}%
\@ifundefined{section}{\def\section#1{\par(Section head:)#1\par }}{}%
\@ifundefined{subsection}{\def\subsection#1%
{\par(Subsection head:)#1\par }}{}%
\@ifundefined{subsubsection}{\def\subsubsection#1%
{\par(Subsubsection head:)#1\par }}{}%
\@ifundefined{paragraph}{\def\paragraph#1%
{\par(Subsubsubsection head:)#1\par }}{}%
\@ifundefined{subparagraph}{\def\subparagraph#1%
{\par(Subsubsubsubsection head:)#1\par }}{}%
\@ifundefined{therefore}{\def\therefore{}}{}%
\@ifundefined{backepsilon}{\def\backepsilon{}}{}%
\@ifundefined{yen}{\def\yen{\hbox{\rm\rlap=Y}}}{}%
\@ifundefined{registered}{%
\def\registered{\relax\ifmmode{}\r@gistered
\else$\m@th\r@gistered$\fi}%
\def\r@gistered{^{\ooalign
{\hfil\raise.07ex\hbox{$\scriptstyle\rm\RIfM@\expandafter\text@\else\expandafter\mbox\fi{R}$}\hfil\crcr
\mathhexbox20D}}}}{}%
\@ifundefined{Eth}{\def\Eth{}}{}%
\@ifundefined{eth}{\def\eth{}}{}%
\@ifundefined{Thorn}{\def\Thorn{}}{}%
\@ifundefined{thorn}{\def\thorn{}}{}%
\def\TEXTsymbol#1{\mbox{$#1$}}%
\@ifundefined{degree}{\def\degree{{}^{\circ}}}{}%
\newdimen\theight
\@ifundefined{Column}{\def\Column{%
\vadjust{\setbox\z@=\hbox{\scriptsize\quad\quad tcol}%
\theight=\ht\z@\advance\theight by \dp\z@\advance\theight by \lineskip
\kern -\theight \vbox to \theight{%
\rightline{\rlap{\box\z@}}%
\vss
}%
}%
}}{}%
\@ifundefined{qed}{\def\qed{%
\ifhmode\unskip\nobreak\fi\ifmmode\ifinner\else\hskip5\p@\fi\fi
\hbox{\hskip5\p@\vrule width4\p@ height6\p@ depth1.5\p@\hskip\p@}%
}}{}%
\@ifundefined{cents}{\def\cents{\hbox{\rm\rlap c/}}}{}%
\@ifundefined{tciLaplace}{\def\tciLaplace{\ensuremath{\mathcal{L}}}}{}%
\@ifundefined{tciFourier}{\def\tciFourier{\ensuremath{\mathcal{F}}}}{}%
\@ifundefined{textcurrency}{\def\textcurrency{\hbox{\rm\rlap xo}}}{}%
\@ifundefined{texteuro}{\def\texteuro{\hbox{\rm\rlap C=}}}{}%
\@ifundefined{euro}{\def\euro{\hbox{\rm\rlap C=}}}{}%
\@ifundefined{textfranc}{\def\textfranc{\hbox{\rm\rlap-F}}}{}%
\@ifundefined{textlira}{\def\textlira{\hbox{\rm\rlap L=}}}{}%
\@ifundefined{textpeseta}{\def\textpeseta{\hbox{\rm P\negthinspace s}}}{}%
\@ifundefined{miss}{\def\miss{\hbox{\vrule height2\p@ width 2\p@ depth\z@}}}{}%
\@ifundefined{vvert}{\def\vvert{\Vert}}{
\@ifundefined{tcol}{\def\tcol#1{{\baselineskip=6\p@ \vcenter{#1}} \Column}}{}%
\@ifundefined{dB}{\def\dB{\hbox{{}}}}{
\@ifundefined{mB}{\def\mB#1{\hbox{$#1$}}}{
\@ifundefined{nB}{\def\nB#1{\hbox{#1}}}{
\@ifundefined{note}{\def\note{$^{\dag}}}{}%
\defLaTeX2e{LaTeX2e}
\ifx\fmtnameLaTeX2e
\DeclareOldFontCommand{\rm}{\normalfont\rmfamily}{\mathrm}
\DeclareOldFontCommand{\sf}{\normalfont\sffamily}{\mathsf}
\DeclareOldFontCommand{\tt}{\normalfont\ttfamily}{\mathtt}
\DeclareOldFontCommand{\bf}{\normalfont\bfseries}{\mathbf}
\DeclareOldFontCommand{\it}{\normalfont\itshape}{\mathit}
\DeclareOldFontCommand{\sl}{\normalfont\slshape}{\@nomath\sl}
\DeclareOldFontCommand{\sc}{\normalfont\scshape}{\@nomath\sc}
\fi
\def\alpha{{\Greekmath 010B}}%
\def\beta{{\Greekmath 010C}}%
\def\gamma{{\Greekmath 010D}}%
\def\delta{{\Greekmath 010E}}%
\def\epsilon{{\Greekmath 010F}}%
\def\zeta{{\Greekmath 0110}}%
\def\eta{{\Greekmath 0111}}%
\def\theta{{\Greekmath 0112}}%
\def\iota{{\Greekmath 0113}}%
\def\kappa{{\Greekmath 0114}}%
\def\lambda{{\Greekmath 0115}}%
\def\mu{{\Greekmath 0116}}%
\def\nu{{\Greekmath 0117}}%
\def\xi{{\Greekmath 0118}}%
\def\pi{{\Greekmath 0119}}%
\def\rho{{\Greekmath 011A}}%
\def\sigma{{\Greekmath 011B}}%
\def\tau{{\Greekmath 011C}}%
\def\upsilon{{\Greekmath 011D}}%
\def\phi{{\Greekmath 011E}}%
\def\chi{{\Greekmath 011F}}%
\def\psi{{\Greekmath 0120}}%
\def\omega{{\Greekmath 0121}}%
\def\varepsilon{{\Greekmath 0122}}%
\def\vartheta{{\Greekmath 0123}}%
\def\varpi{{\Greekmath 0124}}%
\def\varrho{{\Greekmath 0125}}%
\def\varsigma{{\Greekmath 0126}}%
\def\varphi{{\Greekmath 0127}}%
\def{\Greekmath 0272}{{\Greekmath 0272}}
\def\FindBoldGroup{%
{\setbox0=\hbox{$\mathbf{x\global\edef\theboldgroup{\the\mathgroup}}$}}%
}
\def\Greekmath#1#2#3#4{%
\if@compatibility
\ifnum\mathgroup=\symbold
\mathchoice{\mbox{\boldmath$\displaystyle\mathchar"#1#2#3#4$}}%
{\mbox{\boldmath$\textstyle\mathchar"#1#2#3#4$}}%
{\mbox{\boldmath$\scriptstyle\mathchar"#1#2#3#4$}}%
{\mbox{\boldmath$\scriptscriptstyle\mathchar"#1#2#3#4$}}%
\else
\mathchar"#1#2#3#
\fi
\else
\FindBoldGroup
\ifnum\mathgroup=\theboldgroup
\mathchoice{\mbox{\boldmath$\displaystyle\mathchar"#1#2#3#4$}}%
{\mbox{\boldmath$\textstyle\mathchar"#1#2#3#4$}}%
{\mbox{\boldmath$\scriptstyle\mathchar"#1#2#3#4$}}%
{\mbox{\boldmath$\scriptscriptstyle\mathchar"#1#2#3#4$}}%
\else
\mathchar"#1#2#3#
\fi
\fi}
\newif\ifGreekBold \GreekBoldfalse
\let\SAVEPBF=\pbf
\def\pbf{\GreekBoldtrue\SAVEPBF}%
\@ifundefined{theorem}{\newtheorem{theorem}{Theorem}}{}
\@ifundefined{lemma}{\newtheorem{lemma}[theorem]{Lemma}}{}
\@ifundefined{corollary}{\newtheorem{corollary}[theorem]{Corollary}}{}
\@ifundefined{conjecture}{\newtheorem{conjecture}[theorem]{Conjecture}}{}
\@ifundefined{proposition}{\newtheorem{proposition}[theorem]{Proposition}}{}
\@ifundefined{axiom}{\newtheorem{axiom}{Axiom}}{}
\@ifundefined{remark}{\newtheorem{remark}{Remark}}{}
\@ifundefined{example}{\newtheorem{example}{Example}}{}
\@ifundefined{exercise}{\newtheorem{exercise}{Exercise}}{}
\@ifundefined{definition}{\newtheorem{definition}{Definition}}{}
\@ifundefined{mathletters}{%
\newcounter{equationnumber}
\def\mathletters{%
\addtocounter{equation}{1}
\edef\@currentlabel{\arabic{equation}}%
\setcounter{equationnumber}{\c@equation}
\setcounter{equation}{0}%
\edef\arabic{equation}{\@currentlabel\noexpand\alph{equation}}%
}
\def\endmathletters{%
\setcounter{equation}{\value{equationnumber}}%
}
}{}
\@ifundefined{BibTeX}{%
\def\BibTeX{{\rm B\kern-.05em{\sc i\kern-.025em b}\kern-.08em
T\kern-.1667em\lower.7ex\hbox{E}\kern-.125emX}}}{}%
\@ifundefined{AmS}%
{\def\AmS{{\protect\usefont{OMS}{cmsy}{m}{n}%
A\kern-.1667em\lower.5ex\hbox{M}\kern-.125emS}}}{}%
\@ifundefined{AmSTeX}{\def\AmSTeX{\protect\AmS-\protect\TeX\@}}{}%
\def\@@eqncr{\let\@tempa\relax
\ifcase\@eqcnt \def\@tempa{& & &}\or \def\@tempa{& &}%
\else \def\@tempa{&}\fi
\@tempa
\if@eqnsw
\iftag@
\@taggnum
\else
\@eqnnum\stepcounter{equation}%
\fi
\fi
\global\@ifnextchar*{\@tagstar}{\@tag}@false
\global\@eqnswtrue
\global\@eqcnt\z@\cr}
\def\@ifnextchar*{\@TCItagstar}{\@TCItag}{\@ifnextchar*{\@TCItagstar}{\@TCItag}}
\def\@TCItag#1{%
\global\@ifnextchar*{\@tagstar}{\@tag}@true
\global\def\@taggnum{(#1)}%
\global\def\@currentlabel{#1}}
\def\@TCItagstar*#1{%
\global\@ifnextchar*{\@tagstar}{\@tag}@true
\global\def\@taggnum{#1}%
\global\def\@currentlabel{#1}}
\def\QATOP#1#2{{#1 \atop #2}}%
\def\QTATOP#1#2{{\textstyle {#1 \atop #2}}}%
\def\QDATOP#1#2{{\displaystyle {#1 \atop #2}}}%
\def\QABOVE#1#2#3{{#2 \above#1 #3}}%
\def\QTABOVE#1#2#3{{\textstyle {#2 \above#1 #3}}}%
\def\QDABOVE#1#2#3{{\displaystyle {#2 \above#1 #3}}}%
\def\QOVERD#1#2#3#4{{#3 \overwithdelims#1#2 #4}}%
\def\QTOVERD#1#2#3#4{{\textstyle {#3 \overwithdelims#1#2 #4}}}%
\def\QDOVERD#1#2#3#4{{\displaystyle {#3 \overwithdelims#1#2 #4}}}%
\def\QATOPD#1#2#3#4{{#3 \atopwithdelims#1#2 #4}}%
\def\QTATOPD#1#2#3#4{{\textstyle {#3 \atopwithdelims#1#2 #4}}}%
\def\QDATOPD#1#2#3#4{{\displaystyle {#3 \atopwithdelims#1#2 #4}}}%
\def\QABOVED#1#2#3#4#5{{#4 \abovewithdelims#1#2#3 #5}}%
\def\QTABOVED#1#2#3#4#5{{\textstyle
{#4 \abovewithdelims#1#2#3 #5}}}%
\def\QDABOVED#1#2#3#4#5{{\displaystyle
{#4 \abovewithdelims#1#2#3 #5}}}%
\def\tint{\mathop{\textstyle \int}}%
\def\tiint{\mathop{\textstyle \iint }}%
\def\tiiint{\mathop{\textstyle \iiint }}%
\def\tiiiint{\mathop{\textstyle \iiiint }}%
\def\tidotsint{\mathop{\textstyle \idotsint }}%
\def\toint{\mathop{\textstyle \oint}}%
\def\tsum{\mathop{\textstyle \sum }}%
\def\tprod{\mathop{\textstyle \prod }}%
\def\tbigcap{\mathop{\textstyle \bigcap }}%
\def\tbigwedge{\mathop{\textstyle \bigwedge }}%
\def\tbigoplus{\mathop{\textstyle \bigoplus }}%
\def\tbigodot{\mathop{\textstyle \bigodot }}%
\def\tbigsqcup{\mathop{\textstyle \bigsqcup }}%
\def\tcoprod{\mathop{\textstyle \coprod }}%
\def\tbigcup{\mathop{\textstyle \bigcup }}%
\def\tbigvee{\mathop{\textstyle \bigvee }}%
\def\tbigotimes{\mathop{\textstyle \bigotimes }}%
\def\tbiguplus{\mathop{\textstyle \biguplus }}%
\def\dint{\mathop{\displaystyle \int}}%
\def\diint{\mathop{\displaystyle \iint}}%
\def\diiint{\mathop{\displaystyle \iiint}}%
\def\diiiint{\mathop{\displaystyle \iiiint }}%
\def\didotsint{\mathop{\displaystyle \idotsint }}%
\def\doint{\mathop{\displaystyle \oint}}%
\def\dsum{\mathop{\displaystyle \sum }}%
\def\dprod{\mathop{\displaystyle \prod }}%
\def\dbigcap{\mathop{\displaystyle \bigcap }}%
\def\dbigwedge{\mathop{\displaystyle \bigwedge }}%
\def\dbigoplus{\mathop{\displaystyle \bigoplus }}%
\def\dbigodot{\mathop{\displaystyle \bigodot }}%
\def\dbigsqcup{\mathop{\displaystyle \bigsqcup }}%
\def\dcoprod{\mathop{\displaystyle \coprod }}%
\def\dbigcup{\mathop{\displaystyle \bigcup }}%
\def\dbigvee{\mathop{\displaystyle \bigvee }}%
\def\dbigotimes{\mathop{\displaystyle \bigotimes }}%
\def\dbiguplus{\mathop{\displaystyle \biguplus }}%
\if@compatibility\else
\RequirePackage{amsmath}
\fi
\def\makeatother\endinput{\makeatother\endinput}
\bgroup
\ifx\ds@amstex\relax
\message{amstex already loaded}\aftergroup\makeatother\endinput
\else
\@ifpackageloaded{amsmath}%
{\if@compatibility\message{amsmath already loaded}\fi\aftergroup\makeatother\endinput}
{}
\@ifpackageloaded{amstex}%
{\if@compatibility\message{amstex already loaded}\fi\aftergroup\makeatother\endinput}
{}
\@ifpackageloaded{amsgen}%
{\if@compatibility\message{amsgen already loaded}\fi\aftergroup\makeatother\endinput}
{}
\fi
\egroup
\typeout{TCILATEX defining AMS-like constructs in LaTeX 2.09 COMPATIBILITY MODE}
\let\DOTSI\relax
\def\RIfM@{\relax\ifmmode}%
\def\FN@{\futurelet\next}%
\newcount\intno@
\def\iint{\DOTSI\intno@\tw@\FN@\ints@}%
\def\iiint{\DOTSI\intno@\thr@@\FN@\ints@}%
\def\iiiint{\DOTSI\intno@4 \FN@\ints@}%
\def\idotsint{\DOTSI\intno@\z@\FN@\ints@}%
\def\ints@{\findlimits@\ints@@}%
\newif\iflimtoken@
\newif\iflimits@
\def\findlimits@{\limtoken@true\ifx\next\limits\limits@true
\else\ifx\next\nolimits\limits@false\else
\limtoken@false\ifx\ilimits@\nolimits\limits@false\else
\ifinner\limits@false\else\limits@true\fi\fi\fi\fi}%
\def\multint@{\int\ifnum\intno@=\z@\intdots@
\else\intkern@\fi
\ifnum\intno@>\tw@\int\intkern@\fi
\ifnum\intno@>\thr@@\int\intkern@\fi
\int
\def\multintlimits@{\intop\ifnum\intno@=\z@\intdots@\else\intkern@\fi
\ifnum\intno@>\tw@\intop\intkern@\fi
\ifnum\intno@>\thr@@\intop\intkern@\fi\intop}%
\def\intic@{%
\mathchoice{\hskip.5em}{\hskip.4em}{\hskip.4em}{\hskip.4em}}%
\def\negintic@{\mathchoice
{\hskip-.5em}{\hskip-.4em}{\hskip-.4em}{\hskip-.4em}}%
\def\ints@@{\iflimtoken@
\def\ints@@@{\iflimits@\negintic@
\mathop{\intic@\multintlimits@}\limits
\else\multint@\nolimits\fi
\eat@
\else
\def\ints@@@{\iflimits@\negintic@
\mathop{\intic@\multintlimits@}\limits\else
\multint@\nolimits\fi}\fi\ints@@@}%
\def\intkern@{\mathchoice{\!\!\!}{\!\!}{\!\!}{\!\!}}%
\def\plaincdots@{\mathinner{\cdotp\cdotp\cdotp}}%
\def\intdots@{\mathchoice{\plaincdots@}%
{{\cdotp}\mkern1.5mu{\cdotp}\mkern1.5mu{\cdotp}}%
{{\cdotp}\mkern1mu{\cdotp}\mkern1mu{\cdotp}}%
{{\cdotp}\mkern1mu{\cdotp}\mkern1mu{\cdotp}}}%
\def\RIfM@{\relax\protect\ifmmode}
\def\RIfM@\expandafter\text@\else\expandafter\mbox\fi{\RIfM@\expandafter\RIfM@\expandafter\text@\else\expandafter\mbox\fi@\else\expandafter\mbox\fi}
\let\nfss@text\RIfM@\expandafter\text@\else\expandafter\mbox\fi
\def\RIfM@\expandafter\text@\else\expandafter\mbox\fi@#1{\mathchoice
{\textdef@\displaystyle\f@size{#1}}%
{\textdef@\textstyle\tf@size{\firstchoice@false #1}}%
{\textdef@\textstyle\sf@size{\firstchoice@false #1}}%
{\textdef@\textstyle \ssf@size{\firstchoice@false #1}}%
\glb@settings}
\def\textdef@#1#2#3{\hbox{{%
\everymath{#1}%
\let\f@size#2\selectfont
#3}}}
\newif\iffirstchoice@
\firstchoice@true
\def\Let@{\relax\iffalse{\fi\let\\=\cr\iffalse}\fi}%
\def\vspace@{\def\vspace##1{\crcr\noalign{\vskip##1\relax}}}%
\def\multilimits@{\bgroup\vspace@\Let@
\baselineskip\fontdimen10 \scriptfont\tw@
\advance\baselineskip\fontdimen12 \scriptfont\tw@
\lineskip\thr@@\fontdimen8 \scriptfont\thr@@
\lineskiplimit\lineskip
\vbox\bgroup\ialign\bgroup\hfil$\m@th\scriptstyle{##}$\hfil\crcr}%
\def\Sb{_\multilimits@}%
\def\endSb{\crcr\egroup\egroup\egroup}%
\def\Sp{^\multilimits@}%
\let\endSp\endSb
\newdimen\ex@
\[email protected]
\def\rightarrowfill@#1{$#1\m@th\mathord-\mkern-6mu\cleaders
\hbox{$#1\mkern-2mu\mathord-\mkern-2mu$}\hfill
\mkern-6mu\mathord\rightarrow$}%
\def\leftarrowfill@#1{$#1\m@th\mathord\leftarrow\mkern-6mu\cleaders
\hbox{$#1\mkern-2mu\mathord-\mkern-2mu$}\hfill\mkern-6mu\mathord-$}%
\def\leftrightarrowfill@#1{$#1\m@th\mathord\leftarrow
\mkern-6mu\cleaders
\hbox{$#1\mkern-2mu\mathord-\mkern-2mu$}\hfill
\mkern-6mu\mathord\rightarrow$}%
\def\overrightarrow{\mathpalette\overrightarrow@}%
\def\overrightarrow@#1#2{\vbox{\ialign{##\crcr\rightarrowfill@#1\crcr
\noalign{\kern-\ex@\nointerlineskip}$\m@th\hfil#1#2\hfil$\crcr}}}%
\let\overarrow\overrightarrow
\def\overleftarrow{\mathpalette\overleftarrow@}%
\def\overleftarrow@#1#2{\vbox{\ialign{##\crcr\leftarrowfill@#1\crcr
\noalign{\kern-\ex@\nointerlineskip}$\m@th\hfil#1#2\hfil$\crcr}}}%
\def\overleftrightarrow{\mathpalette\overleftrightarrow@}%
\def\overleftrightarrow@#1#2{\vbox{\ialign{##\crcr
\leftrightarrowfill@#1\crcr
\noalign{\kern-\ex@\nointerlineskip}$\m@th\hfil#1#2\hfil$\crcr}}}%
\def\underrightarrow{\mathpalette\underrightarrow@}%
\def\underrightarrow@#1#2{\vtop{\ialign{##\crcr$\m@th\hfil#1#2\hfil
$\crcr\noalign{\nointerlineskip}\rightarrowfill@#1\crcr}}}%
\let\underarrow\underrightarrow
\def\underleftarrow{\mathpalette\underleftarrow@}%
\def\underleftarrow@#1#2{\vtop{\ialign{##\crcr$\m@th\hfil#1#2\hfil
$\crcr\noalign{\nointerlineskip}\leftarrowfill@#1\crcr}}}%
\def\underleftrightarrow{\mathpalette\underleftrightarrow@}%
\def\underleftrightarrow@#1#2{\vtop{\ialign{##\crcr$\m@th
\hfil#1#2\hfil$\crcr
\noalign{\nointerlineskip}\leftrightarrowfill@#1\crcr}}}%
\def\qopnamewl@#1{\mathop{\operator@font#1}\nlimits@}
\let\nlimits@\displaylimits
\def\setboxz@h{\setbox\z@\hbox}
\def\varlim@#1#2{\mathop{\vtop{\ialign{##\crcr
\hfil$#1\m@th\operator@font lim$\hfil\crcr
\noalign{\nointerlineskip}#2#1\crcr
\noalign{\nointerlineskip\kern-\ex@}\crcr}}}}
\def\rightarrowfill@#1{\m@th\setboxz@h{$#1-$}\ht\z@\z@
$#1\copy\z@\mkern-6mu\cleaders
\hbox{$#1\mkern-2mu\box\z@\mkern-2mu$}\hfill
\mkern-6mu\mathord\rightarrow$}
\def\leftarrowfill@#1{\m@th\setboxz@h{$#1-$}\ht\z@\z@
$#1\mathord\leftarrow\mkern-6mu\cleaders
\hbox{$#1\mkern-2mu\copy\z@\mkern-2mu$}\hfill
\mkern-6mu\box\z@$}
\def\qopnamewl@{proj\,lim}{\qopnamewl@{proj\,lim}}
\def\qopnamewl@{inj\,lim}{\qopnamewl@{inj\,lim}}
\def\mathpalette\varlim@\rightarrowfill@{\mathpalette\varlim@\rightarrowfill@}
\def\mathpalette\varlim@\leftarrowfill@{\mathpalette\varlim@\leftarrowfill@}
\def\mathpalette\varliminf@{}{\mathpalette\mathpalette\varliminf@{}@{}}
\def\mathpalette\varliminf@{}@#1{\mathop{\underline{\vrule\@depth.2\ex@\@width\z@
\hbox{$#1\m@th\operator@font lim$}}}}
\def\mathpalette\varlimsup@{}{\mathpalette\mathpalette\varlimsup@{}@{}}
\def\mathpalette\varlimsup@{}@#1{\mathop{\overline
{\hbox{$#1\m@th\operator@font lim$}}}}
\def\stackunder#1#2{\mathrel{\mathop{#2}\limits_{#1}}}%
\begingroup \catcode `|=0 \catcode `[= 1
\catcode`]=2 \catcode `\{=12 \catcode `\}=12
\catcode`\\=12
|gdef|@alignverbatim#1\end{align}[#1|end[align]]
|gdef|@salignverbatim#1\end{align*}[#1|end[align*]]
|gdef|@alignatverbatim#1\end{alignat}[#1|end[alignat]]
|gdef|@salignatverbatim#1\end{alignat*}[#1|end[alignat*]]
|gdef|@xalignatverbatim#1\end{xalignat}[#1|end[xalignat]]
|gdef|@sxalignatverbatim#1\end{xalignat*}[#1|end[xalignat*]]
|gdef|@gatherverbatim#1\end{gather}[#1|end[gather]]
|gdef|@sgatherverbatim#1\end{gather*}[#1|end[gather*]]
|gdef|@gatherverbatim#1\end{gather}[#1|end[gather]]
|gdef|@sgatherverbatim#1\end{gather*}[#1|end[gather*]]
|gdef|@multilineverbatim#1\end{multiline}[#1|end[multiline]]
|gdef|@smultilineverbatim#1\end{multiline*}[#1|end[multiline*]]
|gdef|@arraxverbatim#1\end{arrax}[#1|end[arrax]]
|gdef|@sarraxverbatim#1\end{arrax*}[#1|end[arrax*]]
|gdef|@tabulaxverbatim#1\end{tabulax}[#1|end[tabulax]]
|gdef|@stabulaxverbatim#1\end{tabulax*}[#1|end[tabulax*]]
|endgroup
\def\align{\@verbatim \frenchspacing\@vobeyspaces \@alignverbatim
You are using the "align" environment in a style in which it is not defined.}
\let\endalign=\endtrivlist
\@namedef{align*}{\@verbatim\@salignverbatim
You are using the "align*" environment in a style in which it is not defined.}
\expandafter\let\csname endalign*\endcsname =\endtrivlist
\def\alignat{\@verbatim \frenchspacing\@vobeyspaces \@alignatverbatim
You are using the "alignat" environment in a style in which it is not defined.}
\let\endalignat=\endtrivlist
\@namedef{alignat*}{\@verbatim\@salignatverbatim
You are using the "alignat*" environment in a style in which it is not defined.}
\expandafter\let\csname endalignat*\endcsname =\endtrivlist
\def\xalignat{\@verbatim \frenchspacing\@vobeyspaces \@xalignatverbatim
You are using the "xalignat" environment in a style in which it is not defined.}
\let\endxalignat=\endtrivlist
\@namedef{xalignat*}{\@verbatim\@sxalignatverbatim
You are using the "xalignat*" environment in a style in which it is not defined.}
\expandafter\let\csname endxalignat*\endcsname =\endtrivlist
\def\gather{\@verbatim \frenchspacing\@vobeyspaces \@gatherverbatim
You are using the "gather" environment in a style in which it is not defined.}
\let\endgather=\endtrivlist
\@namedef{gather*}{\@verbatim\@sgatherverbatim
You are using the "gather*" environment in a style in which it is not defined.}
\expandafter\let\csname endgather*\endcsname =\endtrivlist
\def\multiline{\@verbatim \frenchspacing\@vobeyspaces \@multilineverbatim
You are using the "multiline" environment in a style in which it is not defined.}
\let\endmultiline=\endtrivlist
\@namedef{multiline*}{\@verbatim\@smultilineverbatim
You are using the "multiline*" environment in a style in which it is not defined.}
\expandafter\let\csname endmultiline*\endcsname =\endtrivlist
\def\arrax{\@verbatim \frenchspacing\@vobeyspaces \@arraxverbatim
You are using a type of "array" construct that is only allowed in AmS-LaTeX.}
\let\endarrax=\endtrivlist
\def\tabulax{\@verbatim \frenchspacing\@vobeyspaces \@tabulaxverbatim
You are using a type of "tabular" construct that is only allowed in AmS-LaTeX.}
\let\endtabulax=\endtrivlist
\@namedef{arrax*}{\@verbatim\@sarraxverbatim
You are using a type of "array*" construct that is only allowed in AmS-LaTeX.}
\expandafter\let\csname endarrax*\endcsname =\endtrivlist
\@namedef{tabulax*}{\@verbatim\@stabulaxverbatim
You are using a type of "tabular*" construct that is only allowed in AmS-LaTeX.}
\expandafter\let\csname endtabulax*\endcsname =\endtrivlist
\def\endequation{%
\ifmmode\ifinner
\iftag@
\addtocounter{equation}{-1}
$\hfil
\displaywidth\linewidth\@taggnum\egroup \endtrivlist
\global\@ifnextchar*{\@tagstar}{\@tag}@false
\global\@ignoretrue
\else
$\hfil
\displaywidth\linewidth\@eqnnum\egroup \endtrivlist
\global\@ifnextchar*{\@tagstar}{\@tag}@false
\global\@ignoretrue
\fi
\else
\iftag@
\addtocounter{equation}{-1}
\eqno \hbox{\@taggnum}
\global\@ifnextchar*{\@tagstar}{\@tag}@false%
$$\global\@ignoretrue
\else
\eqno \hbox{\@eqnnum
$$\global\@ignoretrue
\fi
\fi\fi
}
\newif\iftag@ \@ifnextchar*{\@tagstar}{\@tag}@false
\def\@ifnextchar*{\@TCItagstar}{\@TCItag}{\@ifnextchar*{\@TCItagstar}{\@TCItag}}
\def\@TCItag#1{%
\global\@ifnextchar*{\@tagstar}{\@tag}@true
\global\def\@taggnum{(#1)}%
\global\def\@currentlabel{#1}}
\def\@TCItagstar*#1{%
\global\@ifnextchar*{\@tagstar}{\@tag}@true
\global\def\@taggnum{#1}%
\global\def\@currentlabel{#1}}
\@ifundefined{tag}{
\def\@ifnextchar*{\@tagstar}{\@tag}{\@ifnextchar*{\@tagstar}{\@tag}}
\def\@tag#1{%
\global\@ifnextchar*{\@tagstar}{\@tag}@true
\global\def\@taggnum{(#1)}}
\def\@tagstar*#1{%
\global\@ifnextchar*{\@tagstar}{\@tag}@true
\global\def\@taggnum{#1}}
}{}
\def\tfrac#1#2{{\textstyle {#1 \over #2}}}%
\def\dfrac#1#2{{\displaystyle {#1 \over #2}}}%
\def\binom#1#2{{#1 \choose #2}}%
\def\tbinom#1#2{{\textstyle {#1 \choose #2}}}%
\def\dbinom#1#2{{\displaystyle {#1 \choose #2}}}%
\makeatother
\endinput
|
\section{Introduction}
\label{sec:intro}
Clusters comprising $N$ individual particles occur in widely different areas of physics ranging from the atomic world
~\cite{Wales_review} to nanoparticles~\cite{review_nanoclusters}, colloids~\cite{review_colloids,Pine_Science_2003,Glotzer,Kraft_Wittkowski_PRE_2013,Nemeth_Lowen_JPCM_1998,RoyallJPCM2009,Wales_Nano12,BrennerPNAS2011,Grzybowski}
and to macroscopic granulates~\cite{granular_clusters}. In the simplest case, the particles interact via a
pairwise potential, such as a Lennard-Jones potential~\cite{References_pure_LJ} or a hard-sphere-dipole
interaction~\cite{DietrichPRE94,Miller,Farrell,Patey,Holm,Kantorovic,Klapp,Messina_DipoleCluster,RehbergArxiv}, and the equilibrium
groundstate structure of the cluster is obtained by minimization of the total potential energy.
Even for small $N$ the structure can be nontrivial and differs substantially from a cutout of a simple crystal.
Here, we consider clusters where the constituents are active or self-propelled particles
which are swimming in a viscous solvent at low Reynolds number.
Such active particles or microswimmers can be artificially realized as colloidal Janus particles
exposed to a thermal gradient~\cite{Sano_PRL2010} close to a solvent phase transition~\cite{Bechinger} or a
chemical reactant catalyzed at one part of the Janus
particle~\cite{Paxton,Baraban_ACSnano,Kaiser_PRE2013Janus,SenPNAS13,BocquetPRL2010,Takagi,Poon_SM14,Ebbens_EPL14,DietrichSwimmer}.
This brings the particle into motion along its symmetry axis, thereby creating a nonequilibrium situation
where the minimization principle of the potential energy breaks down.
An ``active cluster'' composed of aggregated self-propelled
particles will therefore exhibit a characteristic structure, and a characteristic motion.
In particular, as the self-propulsion force grows stronger, the cluster can break revealing an
activity-induced fission.
The occurrence of clusters within active suspensions has been studied frequently in the last years through experiments~\cite{Poon,Bialke_PRL2013,Palacci_science,BocquetPRL12},
theory~\cite{CatesClusterTheory,BialkeReview}, and simulations~\cite{MarchettiPRL12,ZoettelStarkPRL14,CR_SM14,PohlStarkPRL14}
considering purely repulsive as well as slightly attractive particle interactions~\cite{BaskaranPRE13}.
However, controlled fission of such clusters has not been a focus of research, with first forrays in this field made
only recently by Soto and Golestanian~\cite{Soto_Golestanian_PRL_2014,Soto_Golestanian_2015}.
We describe these phenomena by a simple model combining
Weeks-Chandler-Andersen and dipole pair potentials with an intrinsic effective self-propulsion
force~\cite{BtHNatComm,BtHJPCM2014} along the dipole moment~\cite{KlappSwimmer}.
Our model differs from the one employed by Soto and Golestanian~\cite{Soto_Golestanian_PRL_2014,Soto_Golestanian_2015},
which focuses on clusters of self-propelled particles induced by chemotaxis and characterized by their violation of the actio-reactio principle.
In order to keep the model as simple as possible, we consider in this paper only a small number $N$ of dipoles
for the trivial cases $N=1,2$ up to $N=5$. The particles are soft spheres with an embedded dipole
moment that is oriented along the direction of self-propulsion. The situation with $N=4$ and $N=5$ dipoles already reveals
a quite complex dynamical behaviour.
Our initial configuration is either a linear or ring-like structure which represents the groundstate of the
system~\cite{Miller,Messina_DipoleCluster,Kantorovic} or a metastable structure such as various compact and branched
clusters, in particular a Y-junction of $N=4$ dipoles~\cite{SafranNatMat,Rovigatti}.
We then introduce the self-propulsion and follow the dynamics of the particles. When doing so,
we distinguish between two different protocols: an instantaneous switching to a given velocity and an ``adiabatic''
slow increase of the self-propulsion.
As a result, we find that the cluster's center-of-mass generally moves along a helical trajectory.
The details of the helix are governed by the magnetization of the initial cluster.
Moreover, we find a plethora of final cluster states, which depend on the model parameters and the enacted protocol.
For strong self-propulsion and a nonlinear initial state
with nonvanishing dipole moment, fission of the cluster occurs for instantaneous switching. However, interestingly,
for the same parameters, there is no fission in the adiabatic switching case.
Our results can be verified for colloidal Janus particles which a strong dipole moment
along their symmetry axis. At vanishing self-propulsion
these particles have been considered quite a lot in the context of ferrogels and
ferrofluids~\cite{Patey,Holm,Holm2,Kantorovic,Klapp,Messina_DipoleCluster}. Dipolar self-propelled particles have been
prepared recently by Baraban and coworkers~\cite{Baraban_Nano2013,Baraban_ACSNano2013} and aggregation into clusters has indeed
been observed~\cite{LB_pc}.
Our paper is organized as follows: In Sec.~\ref{Sec:model} we specify our model for active dipolar particles.
The different types of motion of clusters with and without vanishing total dipole moment are presented in Sec.~\ref{Sec:Trajectories}. In Sec.~\ref{Sec:Clusters} we study the fission induced by activity in detail and we conclude in Sec.~\ref{Sec:conc}.
\section{Model}
\label{Sec:model}
We consider $N$ spherical dipoles in three spatial dimensions. The position of the $i$th dipole will be denoted
as $ {\bf r} _i=[x_i,y_i,z_i]$ and
the dipole moment ${\bf m}_i=m \hat{\bf u} _i$ is directed along the unit vector
$ \hat{\bf u} _i = (\sin \theta_i \cos \phi_i, \sin\theta_i \sin \phi_i, \cos \theta_i)$.
The total pairwise interaction potential $U_{ij}$ is the sum of a
Weeks-Chandler-Andersen $U^{\text{WCA}}$, which
describes a repulsive soft core~\cite{WCA} and a point dipole potential $U^{\text{D}}$
\begin{eqnarray}
U^{\text{WCA}}_{ij} &=& \begin{cases} 4\epsilon \left[ \left( \frac{\sigma}{r_{ij}} \right)^{12} - \left( \frac{\sigma}{r_{ij}} \right)^6 \right] + \epsilon, &r \leq r_c,\\0, &r > r_c, \end{cases}\\
U^{D}_{ij} &=& \frac{m^2}{r_{ij}^3} \left[ \hat{\bf u} _i \cdot \hat{\bf u} _j - \frac{3 ( \hat{\bf u} _i \cdot {\bf r} _{ij}) ( \hat{\bf u} _j \cdot {\bf r} _{ij})}{r_{ij}^2} \right],
\label{eq:Potential}
\end{eqnarray}
where $ {\bf r} _{ij}= {\bf r} _i- {\bf r} _j$ is the position of particle $j$ relative to particle $i$ and $r_{ij}$ the respective
distance, see Fig.~\ref{f1}.
Here, the cut-off length is $r_{c} = 2 ^{1/6} \sigma$.
We introduce the self-propulsion by means of an effective internal force ${\bf F}_i=F \hat{\bf u} _i$~\cite{BtHComment,BtHJPCM2014}
which is directed along all dipole moments ${\bf m}_i$, leading to a constant self-propulsion velocity
$ {\bf v} _i = v \hat{\bf u} _i$ for an individual single particle. The velocity is given by $v=F/f_t$, and $f_t$ denotes the
translational Stokes friction coefficient.
As units of energy and length we choose the parameters $\epsilon$ and $\sigma$ from the WCA-potential.
Time is measured in units of $\tau = f_t \sigma^2 / \epsilon$.
\begin{figure}[tbhp]
\begin{center}
\includegraphics[clip=,width= \columnwidth]{SketchDipolePair}
\caption{ \label{f1} Sketch of a pair of dipoles in three spatial dimensions with center-of-mass distance $ {\bf r} _{ij}$.
The self-propulsion and the dipole moment are directed along the unit vector
$ \hat{\bf u} _i= (\sin \theta_i \cos \phi_i, \sin\theta_i \sin \phi_i, \cos \theta_i)$.}
\end{center}
\end{figure}
The motion of microswimmers is restricted to the low Reynolds number regime and the corresponding overdamped
equations of motion for the positions $ {\bf r} _i$ and orientations $ \hat{\bf u} _i$ are
\begin{eqnarray}
f_t \cdot \partial_t {\bf r} _i (t) &=& -\nabla_{{ {\bf r} }_i} U+ f_t v \hat{\bf u} _i (t),\\
f_r \cdot \partial_t \hat{\bf u} _i (t) &=& - {\bf T}_i \times \hat{\bf u} _i (t),
\label{eq:EOMs}
\end{eqnarray}
where ${\bf T}_i = \hat{\bf u} _i (t) \times \nabla_{ \hat{\bf u} _i} U$ is the torque on particle $i$
and $U = 1/2 \sum_{i,j\neq i} {U_{ij}}$ the total interaction potential. The
rotational friction coefficient $f_r$ of the spherical particles
is assumed to be linked to $f_t$ via the equilibrium relation $f_r = f_t \sigma^2 / 3$.
We neglect any thermal fluctuations and any solvent-induced hydrodynamic interactions between the particles in our model.
Simulations are performed in three spatial dimensions using a simple Euler integration
algorithm with time steps $\Delta t = 10^{-4} \tau$ for simulation times $t=400\tau$,
which are sufficiently large enough to allow the active dipole clusters to achieve steady state structures.
The dipole strength is varied in the range $0 \leq m^2 / (\epsilon \sigma^3) < 6$.
Starting configurations are gathered by an energy minimization for the respective parameter sets for passive dipoles, $v=0$.
We consider two protocols of how the activity is applied to the dipoles.
The self-propulsion is either instantaneously increased
for the starting configurations, or we slowly increase the activity
stepwise by a velocity increase of $\Delta v =0.1 \sigma / \tau$
subsequent after a long waiting time of $t=400\tau$.
This corresponds to slow or 'adiabatic' switching.
\section{Active dipole clusters: Simple cases and center-of-mass trajectories}
\label{Sec:Trajectories}
\subsection{The special cases $N=1,2,3$}
Let us now discuss the simple cases of very few particles $N=1,2,3$. First of all, a single particle $N=1$ will trivially
move with velocity $v$ on a straight line along its orientation $\hat {\bf u}$.
Next, the ground state of two dipoles ($N=2$) is a head-tail-configuration where the two magnetic moments possess the same
orientational direction and are aligned along the separation vector of the two spheres. This is shown in
Fig.~\ref{Trajectories}(c). Putting a drive along the magnetic moments will lead to joint motion of this linear cluster
with the same velocity $v$ as that for the individual spheres. This behaviour does not depend on the protocol of the drive.
Third, a magnetic triplet ($N=3$) will form a metastable ring-like structure where the sphere centers are on an equilateral triangle
and the orientations have relative angle differences of $2\pi/3$, see Fig.~\ref{Trajectories}(b). However, the
groundstate is a linear chain~\cite{Messina_DipoleCluster}. This ring-like cluster will rotate
around its non-moving center-of-mass when a self-propulsion is turned on. The distance between the spheres and the angular
velocity increase with increasing drive.
\subsection{Center-of-mass trajectories}
Next, we consider the center-of-mass coordinate of dipole clusters
as defined by
\begin{eqnarray}
{\bf R}_c (t)= \frac{1}{N}\sum_{i=1}^N {\bf r} _i (t)\,.
\label{eq:Rcom}
\end{eqnarray}
Due to the reciprocal interactions between the individual dipoles, the motion of the center-of-mass ${\bf R}_c$
can be determined by the orientations $ \hat{\bf u} _i$ of the active particles,
\begin{eqnarray}
\partial_t{\bf R}_c (t)= \frac{1}{N}\sum_{i=1}^N \partial_t {\bf r} _i (t) =\frac{f_tv}{N}\sum_{i=1}^N \hat{\bf u} _i(t) \,,
\label{eq:eomRc}
\end{eqnarray}
such that the center-of-mass velocity $v_c$ is proportional to the total
magnetic moment
\begin{eqnarray}
M= \left| \sum_{i=1}^{N} {\bf m}_i \right| = m \left| \sum_{i=1}^{N} \hat{\bf u} _i \right| \,.
\label{eq:dipmom}
\end{eqnarray}
The motion of the center-of-mass can be easily classified as summarized in Fig.~\ref{Trajectories}.
The simplest situation is an initial {\it ring cluster\/} as shown for $N=3,4,5$ in Fig.~\ref{Trajectories}(b)
whose center-of-mass is in the ring center. This ring has a vanishing total dipole moment, $M=0$.
Based on a simple analysis of the equation of motions and on symmetry arguments,
an initial ring cluster will just rotate around its center-of-mass such that the center-of-mass is non-moving.
This is valid for any $N$, for any self-propulsion strength $v$ and for any protocol (similar to the
special case $N=3$ discussed above)~\cite{SuppMat}. Note, however,
that the radius of the rotating ring and the corresponding angular velocity do increase with increasing drive $v$.
The result of a non-moving center-of-mass can also be obtained for other initial clusters with vanishing initial $M$.
An example for a compact three-dimensional cluster different from a ring is shown in Fig.~\ref{Trajectories}(b) for $N=4$.
This tetrahedral cluster is mechanically stable (but energetically metastable)
in equilibrium $(v=0)$ and has no spontaneous magnetization $(M=0)$.
\begin{figure}[thbp]
\begin{center}
\includegraphics[clip=,width=1\columnwidth]{trajMay
\caption{ \label{Trajectories} (a) Trajectories of the center-of-mass for different initial
active dipole clusters composed of $N=1,\ldots,5$ dipoles.
(b) Ring-like clusters perform a rotation around the center-of-mass such that the center-of-mass is non-moving.
(c) Chains, as well as a Y-junction, move along a straight line and (d) a compact cluster of $N=5$ dipoles propagates
on a helix. Please note that the relative positions of the cluster particles change during the cluster propagation
for an initial Y-junction and the compact cluster shown in~(d). For all cluster conformations the respective total potential energy $U/\epsilon$ is given for $m^2/(\epsilon\sigma^3) = 1$.}
\end{center}
\end{figure}
Obviously, linear chains of $N \geq 1$ dipoles move on a straight line and do not change their relative shape,
see Fig.~\ref{Trajectories}(c),
since all dipole moments and consequently the self-propulsion
velocities are directed along the main symmetry axis.
There are more complex structures
like a Y-junction with $N=4$ dipoles, see again Fig.~\ref{Trajectories}(c), whose center-of-mass moves on a straight line.
The relative positions of these cluster particles, however, can change. The details of the relative motion
will be considered in the next section.
Finally, a moving cluster of constant shape generally performs a helical motion~\cite{WittkowskiPRE12}.
The helical motion is generated by a non-vanishing magnetization $M\not= 0$ which provides
the translational force and a non-vanishing torque on the cluster. A typical example
is an initially compact cluster with non-vanishing magnetization as shown
for $N=5$ particles in Fig.~\ref{Trajectories}(d). This compact cluster has the dipoles arranged at the corners of a pyramid.
\section{Reorganization and fission of active dipole clusters}
\label{Sec:Clusters}
We now turn to clusters composed of four ($N=4$) and five ($N=5$) particles. These cases reveal nontrivial and interesting
dynamical behaviour caused by self-propulsion. In particular, some structures will reorganize, reassemble and split due to the
self-propulsion and this depends explicitly on the protocol applied to turn on the self-propulsion. To be specific we first consider
an initial metastable Y-junction for $N=4$ particles and then study a compact metastable initial cluster for $N=5$ particles.
\subsection{Y-junction $(N=4)$}
\begin{figure*}[tbh]
\begin{center}
\includegraphics[clip=,width=2\columnwidth]{FinalPhaseMay}
\caption{ \label{PhaseDiag} (a) Emerging state diagram for an initial Y-junction for varied reduced dipole
strength $m^2/(\epsilon \sigma^3)$ and self-propulsion velocity $v$. Each symbol represents
a different type of the final cluster configuration, as shown by snapshots (middle).
The black lines indicate an ideal hard sphere scaling. (b) Rescaled state diagram, showing that state transitions are
proportional to $m^2/(F\sigma^4)$ for small self-propulsion and dipole strengths. (c) Trajectories of the individual dipoles
observed in the moving center-of-mass frame~\cite{SuppMat}.
(d) Comparison of the emerging clusters for fixed reduced dipole strength $m^2/(\epsilon \sigma^3)$ and varied reduced
self-propulsion $v\tau/\sigma$ for an instantaneous and an adiabatic switching protocol.}
\end{center}
\end{figure*}
\begin{figure}[thbp]
\begin{center}
\includegraphics[clip=,width=\columnwidth]{AnglesYMay}
\caption{ \label{angles} (a) Schematic representation of the final angles $\alpha$ and $\phi$
and (b),(c) their dependence of the self-propulsion velocity $v$, here rescaled with the fixed dipole strength.
Respective conformations characterized by the order parameters are indicated by sketches.}
\end{center}
\end{figure}
An initial metastable Y-junction built up by $N=4$ dipoles is shown in Fig.~\ref{Trajectories}(c).
We study its dynamical behaviour for varying dipole strength, self-propulsion velocity $v$ and switching protocol. In this case
the motion is two-dimensional in the plane set by the initial configuration.
Figure~\ref{PhaseDiag}(a) shows the cluster state diagram after a long simulation time, $t=400\tau$, in the two-dimensional
parameter space of dipole strength $m^2$ and an instantaneously applied self-propulsion velocity $v$.
As a result, seven different final configurations can be discriminated by suitable order parameters, as explained in detail below.
Clearly, for small self-propulsion the cluster retains its shape as it is metastable but exhibits a drift along its
total magnetization {\bf $M$}. This is indicated by the (red) filled circle. Obviously one finds this situation
when the self-propulsion velocity is small. For very large drives $v$, on the other hand, {\it fission\/} of the cluster shows up
(black plus symbols) where the separating dipoles fly
away pointing forward. In between these two extremes, there are five other states:
i) a Y-junction where the orientations of the 'head dipoles' now point backwards (brown filled square),
ii) a kite cluster where these dipoles shift back towards the center of the whole cluster (yellow filled diamond),
iii) a configuration which we denote as anti-Y-junction since
the former 'head particles' are located in the back (green filled pentagon),
iv) a state where two clusters occur, the backbone and two parallel
dipoles moving in the opposite direction (blue non-filled squares) and finally another
fission state where the separating dipoles burst away pointing backwards (black cross).
The corresponding trajectories of the dipoles during rearrangement of the clusters can be seen in Fig.~\ref{PhaseDiag}(c),
in the center-of-mass frame of reference~\cite{SuppMat}.
It is instructive to compare the scaling behaviour of our findings with that of hard sphere dipoles.
Since hard spheres only bear a length scale (such as the diameter $\sigma$)
and no energy scale, the only ratio which matters,
is that of the dipole-dipole interaction force, proportional to $m^2/\sigma^4$, relative to the internal driving force $F$.
Therefore, in order to test this scaling, we have plotted in Fig.~\ref{PhaseDiag}(b)
the same diagram with the scaled dipolar strength $m^2/(F\sigma^4)$. If hard sphere scaling would hold,
the transition lines should be horizontal in this plot. Deviations from this scaling then need to be attributed
to the softness of the WCA-potential. Indeed, as is revealed in Fig.~\ref{PhaseDiag}(b), the transition
lines between different states are almost horizontal. The largest deviation occurs for the transitions where the
kite-like structure (yellow filled diamond) and the backward Y-junctions (brown filled square) are involved.
Next, we change the protocol to adiabatic switching, see Fig.~\ref{PhaseDiag}(d).
By slowly increasing the self-propulsion velocity, the initial Y-junction deforms into a kite cluster. Further increase of $v$
leads to temporal detaching from the backbone and formation of an anti-Y-junction, but fission does not occur.
This can intuitively be attributed to the fact that the attraction of the dipole cluster has a larger impact on the cluster connectivity
if the particles are perturbed smoothly over time.
We now discuss the details of the different structures with the help of suitable order parameters. Let us first introduce the
positional angle $\alpha$ of one of the 'head dipoles' position relative to the backbone of the Y-junction, see Fig.~\ref{angles}(a).
We further define an angle $\phi$ of the orientation of the 'head dipole' relative to the backbone, see again Fig.~\ref{angles}(a).
In general, the angles $\alpha$ and $\phi$ are different.
A third order parameter is provided by the distance $d$ of the head-dipole to the next backbone particle.
We define connectivity of the cluster by the simple distance criterion: if $d<r_c$ the cluster is connected, else it is split.
Figure~\ref{angles}(b) shows the final angle $\alpha$ for fixed dipole strength
and varied self-propulsion velocity $v$ in the case of the instantaneous switching protocol.
In line with the state diagram discussed previously, for increasing $v$,
the angle $\alpha$ broadens and the initial Y-junction $(\alpha < 90^{\circ})$ becomes a kite cluster
($90^{\circ} \leq \alpha < 120^{\circ})$, which reorganizes into an anti-Y-junction
($120^{\circ} \leq \alpha < 180^{\circ})$. Here, the nearest neighbour is now the dipole in the rear of the backbone,
before finally fission occurs. In the limit of small dipolar interactions, an arbitrary self-propulsion leads to fission,
see again Fig.~\ref{PhaseDiag}(a).
The test of hard-sphere scaling, i.e.\ a scaled plot versus $m^2/(F\sigma^4)$ [see Fig.~\ref{PhaseDiag}(b)], reveals a
Master curve up to $m^2 / (\epsilon\sigma^3) < 3$.
This shows that the hard sphere limit $(\epsilon \rightarrow \infty)$ is reached quickly and that the
state diagram essentially depends on the parameter $F\sigma^4 / m^2 = v\tau\epsilon\sigma^2 / m^2$ for
$m^2 / (\epsilon\sigma^3) < 3$, which is consistent with Fig.~\ref{PhaseDiag}.
Likewise, in Fig.~\ref{angles}(c) the angle $\phi$ is shown, which provides a better resolution of structural details.
The initial passive Y-junction has an angle $\phi_0 = 53.4^{\circ}$. Self-propulsion first induces a broadening of $\phi$
in the Y-junction in the range $(\phi_0 \leq \phi < 90^{\circ})$ (red filled circles). Then the orientation of the
outermost particles flips to the opposite direction $(90^{\circ} \leq \phi < 180^{\circ})$ (brown filled squares)
until the kite cluster occurs, where the head dipoles turn until they are
reversed relative to the backbone magnetization,
$\phi \approx 180^{\circ}$. This conformation only shows up for small dipolar strengths $m^2$.
For even larger $v$, the Y-junction temporarily splits into a remaining backbone and two
single dipoles. These dipoles will be attracted again to the rear of the backbone,
and reattach, being orientated again in the direction of motion $(270^{\circ} < \phi < 360^{\circ})$. This reassembled structure
is denoted as an anti-Y-junction. If $v$ is increased further, these detached dipoles may still be attracted by
the backbone. However they may just collide in its wake and align with each other $(\phi=180^{\circ})$, leading to a
configuration of two clusters propagating in opposing directions.
If $v$ exceeds a critical threshold,
a fission into several units will ultimately occur.
Again, we can discriminate between a state where the detached dipoles move in the opposite
$(90^{\circ} \leq \phi < 180^{\circ})$ or in the same direction $(\phi_0 \leq \phi < 90^{\circ})$ as the backbone of the
initial cluster.
As a final remark, the hard sphere scaling has a similar performance in the angle $\phi$ as shown in Fig.~\ref{angles}(c).
\begin{figure}[tbhp]
\begin{center}
\includegraphics[clip=,width= \columnwidth]{vcm}
\caption{ \label{Vcom} Reduced center-of-mass velocity $v_{c}/v$ of an initial Y-junction as a function of time
for fixed instantaneously applied self-propulsion velocity $v$ for each emerging cluster conformation using the color coding of
Fig.~\ref{PhaseDiag} and fixed dipole strength $m^2/(\epsilon\sigma^3)=1$.}
\end{center}
\end{figure}
Finally, the center-of-mass velocity $v_c = |$d${\bf R}_c(t) / $d$t|$ (see Eq.~(\ref{eq:eomRc}))
is shown in Fig.~\ref{Vcom} for all emerging cluster conformations.
While the kite cluster (yellow filled diamond) and the split-into-two-units state
(blue open squares) have a vanishing center-of-mass
velocity $v_c$, it only vanishes temporarily during rearrangement into the anti-Y-junction (green filled pentagon).
\subsection{Compact cluster $(N=5)$}
\begin{figure*}[thbp]
\begin{center}
\includegraphics[clip=,width=2\columnwidth]{5phase}
\caption{ \label{PhaseDiag5er} (a) Emerging state diagram for an initially compact cluster of $N=5$ dipoles for
varied reduced dipole strength $m^2/(\epsilon\sigma^3)$ and reduced self-propulsion velocity $v\tau/\sigma$.
As in Fig.~\ref{PhaseDiag}, each symbol represents a type of final cluster configuration,
shown through snapshots. (b) Comparison between instantaneous switching and adiabatic switching of the self-propulsion velocity
$v$ for fixed dipole strength $m^2/(\epsilon\sigma^3)=1$.}
\end{center}
\end{figure*}
\begin{figure}[thbp]
\begin{center}
\includegraphics[clip=,width=1\columnwidth]{GyrDipMom}
\caption{ \label{DipMom} (a) Inverse squared radius of gyration $1/R^2$ and (b) reduced total dipole moment $M/m$
of an active cluster of $N=5$ dipoles for fixed dipole strength $m^2$ and varied self-propulsion velocity $v$. The dashed line
in (a) shows the inverse squared cut-off radius of the WCA potential $1/r_c^2$.}
\end{center}
\end{figure}
We finally study an initial three-dimensional cluster with a nonvanishing dipole moment
for $N=5$ dipoles. Such a compact cluster of passive dipoles is metastable only
for $m^2/(\epsilon\sigma^3) < 4.5$, see Fig.~\ref{PhaseDiag5er}(a). For $m^2/(\epsilon\sigma^3) > 4.5$ this cluster will spontaneously
transform into a ring-like structure which is the corresponding ground state~\cite{Miller}.
If self-propulsion is turned on instantaneously, this cluster will either stay stable and move
on a helical trajectory as discussed in Sec.~\ref{Sec:Trajectories} (for $m^2/(\epsilon\sigma^3) < 4.5$)
or spontaneously transform into a ring (for $m^2/(\epsilon\sigma^3) > 4.5$).
The emerging state diagram for the final cluster is presented in Fig.~\ref{PhaseDiag5er}(a).
On top of the helical cluster motion and the ring-like structure, fission occurs at high self-propulsion $v$.
As a function of increasing $v$, the compact cluster first transforms into a ring before it splits apart.
There is a critical drive $v=v_c \approx 6 \sigma/\tau $ beyond which the compact initial cluster looses its stability.
For fixed drive $v<v_c$, an interesting reentrant scenario
for increasing dipolar strength $m$ occurs.
There is fission at small $m$, then the ring-structure is emerging as a final state, subsequently the compact cluster
is the final state and then the ring structure is getting stable again.
The comparison of different protocols is shown in
Fig.~\ref{PhaseDiag5er}(b). In qualitative accordance with the case of $N=4$ particles
discussed above there are fewer (in this case only two) final states.
The fission is missing completely for adiabatic switching and the final cluster conformation for large
self-propulsion velocities $v$ is a two-dimensional ring.
We have documented the fission of the cluster by monitoring the cluster radius $R$ of gyration defined via
\begin{eqnarray}
R^2= \frac{1}{N} \sum_{i=1}^{N} \left( {\bf r} _i - {\bf R}_c\right)^2\,.
\end{eqnarray}
Results for $R$ for an instantaneous switching protocol are shown in Fig.~\ref{DipMom}(a).
Here the inverse of $R^2$ is plotted such that a cluster explosion is indicated by a vanishing $1/R^2$.
The dashed line indicates a size comparable to the cut-off distance $r_c$.
A comparison of $R$ relative to this line clearly reveals the
transition from a compact cluster towards fission.
For completeness we have also shown the cluster magnetization $M$ in Fig.~\ref{DipMom}(b)
which is a good order parameter for a ring-like structure as $M$ vanishes there. These results
confirm the full phase diagram of Fig.~\ref{PhaseDiag5er}(a).
\section{Conclusion}
\label{Sec:conc}
We have studied the dynamical response of initially passive and metastable soft-sphere dipole clusters to the
onset of an internal self-propulsion.
The latter is modelled by an internal effective force along the dipole moment
and the dynamics is completely overdamped in a solvent at low Reynolds number. Even though the cluster is small,
a wealth of different types of motion is obtained, which depend on the interaction parameters, the strength of the
self-propulsion, and the self-propulsion switch-on protocol.
Moving cluster structures emerge which are not stable
in equilibrium. Interestingly, if the self-propulsion is very large and applied quickly (``instantaneous switching''), the cluster
shows a permanent fission but there is no fission if the drive is applied slowly (``adiabatic switching'').
The center-of-mass of the cluster moves on linear, circular, or
helical paths, depending on the initial magnetization of the cluster.
Self-propelled dipolar particles can be realized as Janus particles~\cite{Kretzschmar},
with an embedded dipole moment~\cite{SaguesPRL08,JanusDielectric,MagneticJanusColloids,2012GranickNature} where the
self-propulsion \cite{Baraban_Nano2013,Baraban_ACSNano2013} is generated by self-diffusiophoresis.
Cluster aggregation has indeed been observed~\cite{LB_pc}.
Furthermore, dipolar particles driven by external magnetic fields are conceivable as experimental
realizations at microscopic~\cite{snezhko_prl} as well as macroscopic length scales~\cite{Grzybowski2000}. These particles
can readily be observed in real-space such that our predictions are, in principle, verifiable.
Another realization of self-propelled dipoles are active droplets which are filled with a liquid crystal.
These possess an inner topological defect, shifted from their center
of mass, which induces an electric dipole moment. Here, the strength of the dipole moment is proportional to the self-propulsion, which may
lead to new interesting effects~\cite{1988Lavrentovich,2014Herminghaus}.
Our work can and should be extended towards several directions for future research. First of all,
soft spheres with a non-central dipole moment have been considered recently in the passive case and
the stable cluster structure was found to be different to that of central dipoles~\cite{Kantorovic,Holm2}.
Moreover, one can imagine that the two directions of dipole moment and self-propulsion are not collinear
which is expected to lead to even more complex dynamics. In this case, fission is expected also for slow switching.
For microswimmers, Brownian fluctuations induced by the solvent needs to be considered and incorporated into the
dynamics~\cite{SvT2008,WittkowskiPRE12}.
A more detailed modelling which we shall pursue in the future is an inclusion of hydrodynamic
interactions by using more complicated friction tensors. The friction tensor
corresponding to the direct forces (or body forces) is derived from the pairwise interparticle potential. It
can be treated by the Oseen or Rotne-Prager tensor \cite{PuseyLH,RexLowenPRL08,Dhontbook}, as would be done
for passive particles. On this level, strictly speaking a dumbbell has another
friction tensor than of two spheres.
It is important to note that these friction tensors do not affect
the initial equilibrium structure of the cluster. The swimming process, however, needs to be modelled
with a tensor that decays faster with interparticle distance than the Oseen tensor
~\cite{Ramaswamy02,Ramaswamy04}. We expect that solving the coupled equation of motion
could give rise to new unexpected cluster dynamics as both interactions
(the hydrodynamic and the dipolar one) are long-ranged and therefore compete. Moreover, the behaviour
will depend on the hydrodynamic boundary conditions. An unbounded solvent around the cluster will be described
by a different friction tensor than the motion of dipolar particles on a substrate~\cite{LaugaPRL13}
or on a pending air-liquid interface which will make
the full problem even more complicated.
Finally, particle shapes different from that of a sphere can be studied
like C- or L-shaped particles~\cite{BtHNatComm,Kaiser2carrier,WensinkPRE14},
which tend to perform a circular motion if they are self-propelled even in absence of any dipolar interaction.
\vspace{0.1cm}
\acknowledgments
We thank Eugene Terentjev and Peet Cremer for helpful discussions and Sandra Held for proofreading the manuscript.
This work was supported by the science priority program
SPP 1681 of the German Science Foundation (DFG) and by the ERC
Advanced Grant INTERCOCOS (Grant No. 267499).
|
\section{Introduction}
Large and balanced datasets are normally crucial for learning classifiers.
In real-world scenarios, however, one always struggles to find adequate
amounts of labeled data. Even with the help of crowdsourcing, e.g.,
Amazon Mechanical Turk (AMT), it is often difficult to collect a large
quantity of labeled instances with high quality that is necessary for
training a classifier for a real-world problem.
In terms of quantity, it has been shown that the amount of
available training data, per object class, roughly follows a Zipf distribution \cite{torralba2011app_share}. That means a small number of
object classes account for most of the available training instances.
In terms of quality, some domains, such as the analysis of satellite images
(e.g. the comet images from Rosetta), require extensive and detailed
expert user annotation \cite{remotesensing2013,ZX:14b}. Large volume
of LiDAR point cloud data have to be labeled before they can be used
to train some classifiers \cite{ZX:14}. Such labeling process usually is very time consuming and requires expert-level labeling efforts or expensive equipments. Practically only a very limited portion of the data points
can be obtained.
To solve the problem of lacking enough training samples,
attributes \cite{lampert2009zeroshot_dat,palatucci2009zero_shot,farhadi2009attrib_describe}
have been introduced to transfer the knowledge held by majority classes to instances in minority classes.
Nevertheless, for certain tasks, such shared
attributes~\cite{fu2012attribsocial,Dhar2011cvpr,liu2011action_attrib,datta2011face_attrib,Yucatergorylevel}
may simply be unavailable or nontrivial to define.
In contrast, rather than using such a `learning to learn' \cite{Thrun96learningto} framework, humans can
generalize and associate the similar patterns from images. This
ability inspires us to circumvent the problem of lacking enough training data and solve it from a different
angle: utilizing the synthetic data (e.g. the synthetic roof edges in Fig.
\ref{fig: roof real and synI}) associated with real data (e.g.
real roof edges in Fig. \ref{fig: roof real and synI}) in order to learn a better classifier.
\begin{figure}
\centerline{ %
\begin{tabular}{cc}
\resizebox{0.23\textwidth}{!}{\rotatebox{0}{ \includegraphics{RealSamples}}}
&
\resizebox{0.23\textwidth}{!}{\rotatebox{0}{ \includegraphics{TmpltSamples}}}
\tabularnewline
(a) Real roof edges & (b) Synthetic roof edges
\tabularnewline
\end{tabular}} \protect\caption{Examples of (a) real roof edge vs. corresponding (b) synthetic roof
edge images. The synthetic data is generated by the algorithms in Sec.
4. The examples are randomly drawn from the SRC dataset.}
\label{fig: roof real and synI}
\end{figure}
The idea of associating synthetic data with real data has a long
history and is associated with the development of cognitive psychology,
artificial intelligence, and computer vision. For example, cognitive psychology
studied a case that an infant learns to understand and imitate a facial
expression from parents' examples. In the computing domain, exemplar SVM \cite{examplarSVM}
tries to associate images with training exemplars. Different
from these previous works, we create synthetic images to associate them with
the real images whilst previous works associate `new' real
data with `old' real data. By contrast, our approach is a 'free lunch'
in the sense that the proposed approach does not need any human annotation of real data,
thus we could easily amplify the dataset used in training.
Learning a classifier from synthetic data is unfortunately extremely
challenging due to the following reasons. Firstly, the feature distribution
of synthetic data generated will shift away from that of real data.
Such distribution shift is termed \emph{synthetic gap} and illustrated
in Fig \ref{fig: tsne_vis}. The synthetic gap is
a major obstacle in using synthetic data to help learning classifiers,
since synthetic data may fail to simulate the potential useful patterns
of real data for training classifiers. To our knowledge, this synthetic
gap problem has never been formally identified nor addressed in the literature.
Secondly, since practically a small amount of labeled images may be
available, it is necessary to jointly learn from synthetic and real
data. The learning process must be automatically leveraged between
synthetic data and real data.
\textcolor{red}{}
\begin{figure}
\centering
%
\begin{tabular}{cc}
\includegraphics[scale=0.29]{Real2-crop}
&
\includegraphics[scale=0.29]{Reconst2-crop}\tabularnewline
(a) & (b)\tabularnewline
\end{tabular}\protect\caption{t-SNE visulization of synthetic gap using the data from SRC dataset. (a) synthetic gap of real and
synthetic data; (b) MCAE bridges the synthetic gap.}
\label{fig: tsne_vis}
\end{figure}
To better learn a classifier from synthetic data, we propose a novel
framework --Multichannel Autoencoder (MCAE) which is an extension of sparse autoencoder. The training step of
MCAE is a process of bridging the synthetic gap between the real and
the synthetic data by learning the mapping from (1) synthetic
to real data and (2) real to real data. Critically, such mapping try to keep the real data
while enforce MCAE to learn a transfer from the synthetic data to the real data.
We thus can generate more synthetic data which will simulate the real data when the learned mapping is applied to them.
To facilitate the study on satellite image analysis,
we introduce a new benchmark satellite roof classification (SRC) image dataset.
The SRC dataset needs expert-level
labeling and has unique challenges, such as satellite image blurring,
building shadows, and extremely imbalanced roof class
instances. To demonstrate the generality of the proposed approach, we use an additional
handwritten digit dataset from the UCI machine learning repository \cite{Bache+Lichman:2013}.
In both datasets, synthetic data is generated using a parametric model of derived from real data that roughly mimics real data in terms of appearances and basic structure. Experimental results using these datasets demonstrate that better classification results can
be obtained by training a classifier using the synthetic data when used
by the proposed approach.
We thus highlight three contributions in this paper: (1) To the best
of our knowledge, this is the first attempt to address the problem of synthetic gap, by solving which we demonstrate that the synthetic data could be used to improve the performance of classifiers. (2) We propose a Multichannel Autoencoder (MCAE) model to bridge
the synthetic gap and jointly learn from both real and synthetic data.
(3) Also, a novel benchmark dataset -- Satellite Roof Classification (SRC) is
introduced to the vision community. Such dataset is of expert-level
label annotations as well as great challenges for satellite image
analysis.
\section{Related Work}
\textbf{3D image analysis. }Synthetic data has been used for several
3D image analysis applications, but not for helping learn classifiers. A large number of synthetic 3D meshes
in \cite{ZX:12} were created by a series of mesh editing steps including
subdivision, simplification, smooth, adding noise and Poisson reconstruction,
in order to automatically evaluate the subjective visual quality of
a 3D object. Recently, to circumvent the point labeling difficulty
in a building roof classification problem using LiDAR point cloud,
\cite{ZX:14} explicitly indicated semantic roof points on synthetically
created roof point clouds and compute point features from the synthetic
point clouds. Techniques such as point cloud resampling, size normalization
and mesh erosion are employed to reduce the differences between real
roof and synthetic ones in data space.
\textbf{Generating synthetic data. } Previous method generate synthetic
data in data space using tools including geometrical transformation
and degradation models: In \cite{VT:03}\cite{VT:04}, to help off-line
recognition of handwritten text, a perturbation model combined with
morphological operation is applied to real data. They showed that
when a moderate transformation is added to the real data, the resulting
synthetic training set boost the performance. To enhance the quality
of degraded document, in \cite{BG:08} degradation models such as
brightness degradation, blurring degradation, noise degradation and
texture-blending degradation were used to create a training dataset
for a handwritten text recognition problem. The synthetic minority
oversampling technique (SMOTE) \cite{CNV:02} and its variants \cite{HH:05}\cite{HH:08}
are also powerful methods that have shown many success in various applications.
However, these previous methods are relatively limited to one particular type of dataset, whilst
we propose a more general methodology of generating synthetic data in this paper.
We show that our methodology can be used both for SRC and handwritten digits dataset.
\textbf{Transfer Learning} aims to extract the knowledge from one
or more source tasks and applies the knowledge to a target task. Transfer
learning has been found helpful in many real world problems, such
as in sentiment classification \cite{Blitzer07Biographies}, web page
classification\textcolor{red}{{} }\cite{Sarinnapakorn:2007:CST:1313047.1313197}
and zero-shot classification of image and video data \cite{lampert13AwAPAMI,lampert2009zeroshot_dat,yanweiPAMIlatentattrib,yao2011action_part,yanweiembedding,yanweiBMVC,RohrbachCVPR12,rohrbach2010semantic_transfer,RichardNIPS13}.
Transfer learning is categorized to three classes \cite{pan2009transfer_survey}:
inductive transfer learning, transductive transfer learning and unsupervised
transfer learning. The work in this paper falls into a framework of
domain adaptation \cite{Ben-David:2010:TLD:1745449.1745461,Weinberger:2009:FHL:1553374.1553516}
in the transductive transfer learning. Nonetheless, different from
previous domain adaptation tasks of different source and target domains,
the synthetic gap is caused by the shifted feature distribution of
synthetic data from real data. To solve this problem, our MCAE is
developed from the idea of autoencoder.
\textbf{Autoencoder} is one type of neural network and its output
vectors have the same dimensionality as the input vectors \cite{vincent2008ICML}.
The hidden representation obtained by training a sparse autoencoder
followed by a parameters fine tuning is useful in pre-training a deeper
neural network. Recently autoencoder with its different variants \cite{MarginalizedDenoisingAutoencoders2012ICML,Glorot11domainadaptation}
also exhibit the success in learning and transferring sharing knowledge
among data source from different domains \cite{BP:12,BY:12,DJ:13},
thus benefit other machine learning tasks.
\section{Multichannel Autoencoder (MCAE) }
In this section, we introduce the MCAE model as illustrated in Fig.
\ref{fig: YAE-structure}. It can (1) bridge synthetic gap by minimizing
the discrepancy between real and synthetic data; and (2) preserve
and emphasize the potential useful patterns existed in both real and
synthetic data in order to generate the better feature representations
used for learning classifiers.
Essentially, synthetic and real data should have similar patterns, a natural
idea of bridging synthetic gap is to learning a mapping from the synthetic data to the real data using an autoencoder, and vice versa. MCAE, hence, provides a more flexible way to learn this mapping due to the specific structure of the MCAE. There are two channels in MCAE, left one and right one. Each channel basically is an SAE, however, two channels share the same hidden layer. With this structure, MCAE basically learns two tasks in the same time. By setting different types of input and out data such as the one in denoising autoencoder \cite{VP:10}, MCAE is capable for many applications. In our work, to bridge the gap between synthetic data and real data, we set the task in left channel as one that takes synthetic data as input and real data as \textit{reconstruction target}, while the task in right channel use real data in both input and \textit{reconstrution target}. This configuration actually is essentially meaningful that by keeping the \textit{reconstruction target} identical in two channels, MCAE attempts to transform inputs in two channels towards the same target, thus minimize the discrepancy between two input dataset which are synthetic data and real data in our work.
\begin{figure}[ht]
\centerline{ %
\begin{tabular}{cc}
\resizebox{0.21\textwidth}{!}{\rotatebox{0}{ \includegraphics{StackedMAE}}} & \resizebox{0.24\textwidth}{!}{\rotatebox{0}{ \includegraphics{Multi-tasking_AE}}} \tabularnewline
(a) & (b) \tabularnewline
\end{tabular}} \protect\protect\caption{(a) Illustration of the proposed MCAE model in a stacked autoencoder
structure, where black edge between two layers are linked to and shared
by two tasks, red and blue links are separately connected to left
and right task respectively. (b) A zoom in structure of MCAE. }
\label{fig: YAE-structure}
\end{figure}
\subsection{Problem setup }
Our MCAE is built on the sparse autoencoder (SAE). A basic autoencoder
is a fully connected neural network with one hidden layer and can
be decomposed into two parts: an encoding and a decoding process.
Assume an input dataset with $n$ instances $X=\{x_{i}\}_{i=1}^{n}$
where $x_{i}\in{\mathbb{R}^{m}}$ and $m$ is the dimension of each
instance. Encoding typically transforms input data to hidden layer
representation using an affine mapping squashed by a sigmoid function:
\begin{equation}
h_{e}(x_{i})=f(W_{e}x_{i}+b_{e})\label{eq:encoder}
\end{equation}
where $f(\cdot)$ is a sigmoid function and $\theta_{e}=\{W_{e},b_{e}\}$,
$W_{e}\in\mathbb{R}^{k\times m},b_{e}\in\mathbb{R}^{k}$ is a set
of unknown parameters in encoding with $k$ nodes in hidden layer.
While in decoding, with parameters $\theta_{d}=\{W_{d},b_{d}\}$,
$W_{d}\in\mathbb{R}^{m\times k},b_{d}\in\mathbb{R}^{m}$, autoencoder
attempts to reconstruct the input data at the output layer by imposing
another affine mapping followed by nonlinearity to hidden representation
$h_{e}(x_{i})$:
\begin{equation}
h_{d}(x_{i})=f(W_{d}h_{e}(x_{i})+b_{d})\label{eq:decoding}
\end{equation}
In above equation $h_{d}(x_{i})$ is viewed as a reconstruction of
input $x_{i}$. Normally, we impose $h_{d}(x_{i})\approx x_{i}$.
Here $x_{i}$ play a role of \textit{reconstruction target} in this
expression and we use notation $\langle\mathfrak{i}\text{:}X_{i},\:\mathfrak{t}\text{:}X_{i}\rangle$
to denote the configuration of input data short for $\mathfrak{i}$
and \textit{reconstruction target} short for $\mathfrak{t}$ in an
autoencoder. $X_{s}$ and $X_{r}$ indicate synthetic and real data
respectively. By minimizing the reconstruction errors of all data
instances, we have following objective function:
\begin{equation}
J(\theta_{e},\theta_{d})=\frac{1}{n}\sum_{i=1}^{n}(h_{d}(x_{i})-x_{i})^{2}+\lambda W\label{eq:reconstruction_error_basic}
\end{equation}
where $W=({\sum W_{e}^{2}+\sum W_{d}^{2}})/2$ is a weight decay term
added to improve generalization of the autoencoder and $\lambda$
leverages the importance of this term.
To avoid learning identity mapping in autoencoder, a regularization
term $\Theta=\sum_{i=1}^{k}{\delta\text{log}\frac{\delta}{\hat{\delta}_{i}}+(1-\delta)\text{log}\frac{1-\delta}{1-\hat{\delta}_{i}}}$
that penalizes over-activation of the nodes in the hidden layer is
added. $\delta$ is a sparsity parameter and is set by users and $\hat{\delta}_{i}=\frac{1}{k}\sum_{i=1}^{k}h_{e}(x_{i})$.
\begin{equation}
J(\theta_{e},\theta_{d})=\frac{1}{n}\sum_{i=1}^{n}(\hat{x}_{i}-x_{i})^{2}+\lambda W+\rho\Theta\label{Equ: SparsityObjective}
\end{equation}
$\rho$ controls sparsity of representation in hidden layer.
Note that directly applying sparse autoencoder to our problem does
not work well. For example, we can train an autoencoder purely
by placing synthetic data in input layer and real data in output layer
denoted as $\langle\mathfrak{i}\text{:}X_{s},\:\mathfrak{t}\text{:}X_{r}\rangle$
which however can not bridge the synthetic gap in our problem. Such
way of reconstruction is only to complement the missing information
in synthetic data from real data but not vice versa%
\footnote{Please refer to supplemenatry maerial for the validation%
}.
A better representation should be reconstructed by using the information
from both real and synthetic data simultaneously. Specifically, we
aim at two tasks: one is $\langle\mathfrak{i}\text{:}X_{s},\:\mathfrak{t}\text{:}X_{r}\rangle^L$
which reconstructs synthetic data towards real data, and the other
one is $\langle\mathfrak{i}\text{:}X_{r},\:\mathfrak{t}\text{:}X_{r}\rangle^R$
which uses identical real data for input and \textit{reconstruction
target}, where $\langle\cdot\rangle^{L}$ and $\langle\cdot\rangle^{R}$
indicate the left and right channel of MCAE.
\subsection{MCAE model}
We propose a multichannel autoencoder that uses a balance regularization
to leverage the learning between two tasks, i.e. $\langle\mathfrak{i}\text{:}X_{s},\:\mathfrak{t}\text{:}X_{r}\rangle^L$
and $\langle\mathfrak{i}\text{:}X_{r},\:\mathfrak{t}\text{:}X_{r}\rangle^R$.
The structure of this new autoencoder is shown in Fig. \ref{fig: YAE-structure}.
In this new structure, tasks of two channels will share the same parameters
$\theta_{e}$ in encoding process which will enforce autoencoder to
reconstruct common structure in both tasks. However, in decoding process,
we divide autoencoder to two separate channels that two tasks will
have their own parameters $\theta_{d}^{L}$ and $\theta_{d}^{R}$.
Dividing autoencoder to two channels at decoding layer enable a more
flexible control between the two tasks. Thus autoencoder better leverage
the common knowledge from the two tasks.
With two channels in the MCAE, we target to minimize the reconstruction
error of two tasks together while taking into account the balance
between two channels. The new objective function of the MCAE is given
in the following:
\begin{equation}
E=J^{L}(\theta_{e},\theta_{d}^{L})+J^{R}(\theta_{e},\theta_{d}^{R})+\gamma\Psi\label{Equ: YAE-objective}
\end{equation}
where
\begin{equation}
\Psi=\frac{1}{2}(J^{L}(\theta_{e},\theta_{d}^{L})-J^{R}(\theta_{e},\theta_{d}^{R}))^{2}
\end{equation}
is a regularization added to balance the learning rate between two
channels. This regularization will have two effects on the MCAE. First,
$\Psi$ accelerates the speed of optimizing Eq. \ref{Equ: YAE-objective},
since minimizing $\Psi$ requires both $J^{L}(\theta_{e},\theta_{d}^{L})$
and $J^{R}(\theta_{e},\theta_{d}^{R})$ are small which in turn cause
$E$ decreases faster. Second, $\Psi$ penalizes a situation more
when difference of learning error between two channels are large,
so as to avoid imbalanced learning between two channels.
The minimization of Eq. \ref{Equ: YAE-objective} is achieved by back propagation
and stochastic gradient descent using Quasi-Newton method. Since the
regularization term is added to leverage the balance of different
tasks, we have to compute the gradient of parameters $\theta_{e}$
and $\theta_{d}^{L},\theta_{d}^{R}$ in MCAE. Please refer to the
supplementary material for the detailed computation of gradients.
\subsection{The advantages of MCAE over the alternative Configurations}
Our MCAE enforces autoencoder to learn useful class patterns from
the two tasks simultaneously. Thus it helps with capturing a common
structure of synthetic and real images. Another alternative way is
to concatenate the input and target of the two tasks $\langle\mathfrak{i}\text{:}X_{s}X_{r},\:\mathfrak{t}\text{:}X_{r}X_{r}\rangle$
for autoencoder. We annotate the usage of this autoencoder as Concatenate-Input
Autoencoder (CIAE), since this autoencoder learns concatenated tasks at
the same time. Such configurations however may
result in an unbalanced optimization for these two tasks: the optimization
process of one task will take over the process of the other one. It
results in a biased reconstructed hidden layer of the autoencoder
and thus a limited classification performance. Our experiments also
validate this point in Sec. \ref{Sec: Results}.
\section{Generating Synthetic Data}
It is an important and yet less studied topic of how to generate synthetic
data. This section discusses the methodology of generating synthetic data
used in our experiments. Such synthetic data have some similarities
and differences with the augmented data used in deep learning e.g.
\cite{KrizhevskySH12}. Both of synthetic data and augmented data
aim at improving the generalisation capacity of classifiers. Nevertheless,
the methodology of generating synthetic data brings more deformed
patterns than the simply label-preserving transformations used in
data augmentation.
Synthetic data are created to highlight the potential useful pattern
existed in real images. We have two stages of generating synthetic
data. In the first stage, for each real data used to train MCAE, a
synthetic version that best matching appearance of the real data is
generated; thus pairs of corresponding real and synthetic data can
be used to train the MCAE. In the second stage, more synthetic data
could be derived using synthetic data generated in the first stage
by both interpolation and extrapolation. To distinguish the set of
synthetic data used in these two stages, we use abbreviation \textit{Syn
I} and \textit{Syn II} to represent them respectively.
In the proposed approach, the synthetic data are represented as a
parametric model of a set of control points and edges associated to
these points in the images. From the control points, the synthetic
images could be generated to simulate the real images in terms of
having the same structure or a similar appearance. Initially, the control
points are selected from a synthetic prototype that generalize all images in
the same class. Then the locations of the control points are iteratively
optimized until convergence in order to minimize the distance between
synthetic images generated by control points and the real image. We
annotate the control points and edges associated to them as $\textbf{S}=\{\textbf{P},\textbf{E}\}$,
where $\textbf{P}=\{p_{i}\}_{i=1}^{n}$ is the set of the control points, and
$\textbf{E}=\{(p_{i},p_{j})\},1\leq i,j\leq n$ is the set of edges
connecting control points. A generalized algorithm of getting the
best matching synthetic image is provided in Algorithm \ref{Alg: MatchingSynthetic}.
\begin{algorithm}[htb]
\small
\caption{ Get Matching Synthetic Image.}
\label{Alg: MatchingSynthetic}
\begin{algorithmic}[1] \REQUIRE ~~\\
$\bullet$ A real image $U$. \\
$\bullet$ A set of control points
$\textbf{S}=\{\textbf{P}, \textbf{E}\}$
with all control points $p_i\in \textbf{P}$ set to their initial positions.\\
$\bullet$ A prototype image $V$ generated using the initial $\textbf{S}$. \\
\WHILE{\textbf{S} is not converged}
\STATE \textbf{S} = OptimizeControlPoints($U, V, \textbf{S}$).
\STATE Generate $V$ using $\textbf{S}$.
\ENDWHILE
\STATE Generate synthetic image $I$ using \textbf{S}.
\RETURN $I$.
\end{algorithmic}
\end{algorithm}
The synthetic prototype could be manually design or learning from given data in our work given different tasks. We will show how these two methods produce synthetic data in following two sections respectively.
\subsection{Explicitly Design of the Synthetic Prototype}
The generation of the synthetic prototype and control points in this scenario is inspired by the approach proposed by
Zhang \textit{et al.} \cite{ZX:14}. In their work, given enough pre-knowledge about the 3D objects,
a synthetic prototype of 3D objects is explicitly designed and built. By adjusting the control points of the prototype, various kinds of 3D objects are generated. In this work, essentially, our data is very similar to theirs in a sense that roof images share a lot of characteristics such as ridge lines, valley lines and intersections between these lines, which make it possible to manually design the synthetic prototypes that characterize these patterns. Based on this observation, a synthetic roof prototype could be generated
by setting the control points at the intersections of the ridge or
valley lines and drawing segments connecting these control points.
\footnote{In our experiments, classification of the roof images is essentially
similar to that of \cite{ZX:14}. Our approach recognizes the style
of the roofs based on edges extracted from the roof images. For more
visualisation results, please refer to our supplementary material.%
}.
In this scenario, the OptimizeControlPoints($U,V,\textbf{S}$) function of Alg. \ref{Alg: MatchingSynthetic}
turns out to be a process that searches for optimal control point locations which results in a synthetic image minimizing the discrepancy between the real image and the synthetic image. A coordinate descent framework is employed to accelerate the search process. We summarize this method in Alg. \ref{Alg: OCP1}.
\begin{algorithm}[htb]
\small
\caption{ OptimizeControlPoints($U, V, \textbf{S}$) Case 1}
\label{Alg: OCP1}
\begin{algorithmic}[1] \REQUIRE ~~\\
$\bullet$ A real image $U$. \\
$\bullet$ A prototype of the synthetic image $\textbf{S}=\{\textbf{P}, \textbf{E}\}$.\\
$\bullet$ A synthetic image $V$ generated using $\textbf{S}$. \\
\FOR{$p_i\in \textbf{P}, 1\leq i \leq n$} \STATE Update $\textbf{S}$ by moving $p_i$ by one unit.
\STATE Generate $V$ using \textbf{S}.
\IF {$\textbf{S}$ does not reduce Dist($U, V$)}
\STATE Cancel the last move of $p_i$.
\STATE Generate $V$ using \textbf{S}.
\ENDIF
\ENDFOR
\RETURN \textbf{S}.
\end{algorithmic}
\end{algorithm}
\subsection{Learning Synthetic Prototype from Data }
In hand written digit dataset used in this work, we learn a synthetic prototype from given data.
A digit prototype is generated for all images with the same digit. Congealing
algorithm proposed in \cite{ME:00} is employed in this step to produce the synthetic prototypes for digits. In congealing,
the project transformations are applied to images to minimize a joint
entropy. Thus the prototype is considered to be an average image of all
images after congealing.
Then control points are evenly sampled from the boundary detected
from the prototype image. The control points needs to be mapped
to each digit image in order to generate a synthetic image. To find this mapping we implement an
approach that migrates the control points from the prototype images to destination image.
This point migration algorithm is based on a series of intermediate
images generated in between synthetic prototype and destination image. To
generate the intermediate images, we binarize all the images and the
distance transformed images\cite{GB:86} of the synthetic prototype
and the real image are generated. Given the number of steps, an intermediate
image then is generated as a binarized image of linear interpolation
between two distance transformed images. In each
step, the control points are snapped to the closest boundary pixels
of the intermediate image. The algorithm of OptimizeControlPoints($U,V,\textbf{S}$)
in this situation is given in Algorithm \ref{Alg: OCP2}, we fix the
number of steps to $10$ in this algorithm.
\begin{figure}[ht]
\centerline{ %
\begin{tabular}{ccccccc}
\resizebox{0.045\textwidth}{!}{\rotatebox{0}{ \includegraphics{prototypeDT}}}
&
\resizebox{0.045\textwidth}{!}{\rotatebox{0}{ \includegraphics{step1}}}
&
\resizebox{0.045\textwidth}{!}{\rotatebox{0}{ \includegraphics{step2}}}
&
\resizebox{0.045\textwidth}{!}{\rotatebox{0}{ \includegraphics{step3}}}
&
\resizebox{0.045\textwidth}{!}{\rotatebox{0}{ \includegraphics{step4}}}
&
\resizebox{0.045\textwidth}{!}{\rotatebox{0}{ \includegraphics{step5}}}
&
\resizebox{0.045\textwidth}{!}{\rotatebox{0}{ \includegraphics{realDT}}} \tabularnewline
\end{tabular}} \caption{Illustrations of the migration of control points and intermediate
synthetic images generated using control points in each step. The
distance transform images of the synthetic prototype and real images
are shown as the left most and right most images respectively.}
\label{fig: roofexamples}
\end{figure}
\begin{algorithm}[htb]
\small
\caption{ OptimizeControlPoints($U, V, \textbf{S}$) Case 2}
\label{Alg: OCP2}
\begin{algorithmic}[1] \REQUIRE ~~\\
$\bullet$ A real image $U$. \\
$\bullet$ A prototype of the synthetic image $\textbf{S}=\{\textbf{P}, \textbf{E}\}$.\\
$\bullet$ A synthetic image $V$.
\\
\STATE $steps=10$.
\STATE Compute distance transform image of $U, V$ as $U', V'$.
\FOR{$i=1$ to $steps$}
\STATE $I=(1-\frac{i}{steps})U'+\frac{i}{steps}V'$.\\
\STATE $I$=Binarize($I$).\\
\STATE Update $\textbf{S}$ by snapping to the closest boundary pixel on $I$.\\
\ENDFOR \STATE Set the status of $\textbf{S}$ to be converged.
\RETURN \textbf{S}.
\end{algorithmic}
\end{algorithm}
To generate the $SynII$ dataset, we either interpolate or extrapolate
between control points of randomly choose two synthetic images from
$SynI$ dataset. The weights used in interpolation and extrapolation
is uniformly drawn from $0$ to $1$.
\section{Experiments and Results}
\label{Sec: Results} We validate the proposed MCAE dataset on several
applications in this section. This section is organised as follows.
First, in Sec 5.1, we introduce a new benchmark dataset -- Satellite
Roof Classification (SRC) dataset to vision community. This dataset
is of high quality satellite roof class labels for the satellite images. We
also briefly summarizes the handwritten digits dataset used in our
paper. We explain the experimental settings in Sec. 5.2 and discuss
the experimental results in Sec. 5.3.
\subsection{Experiment Datasets}
\subsubsection{Satellite Roof Classification (SRC) Dataset}
One particular interesting problem of learning classifiers from synthetic
data is to analyze satellite images of the Earth. Such problems generally
need very high quality (expert-level) labeled data. However, there
is no previous dataset for such research purposes. To facilitate the
study, a new benchmark Satellite Roof Classification (SRC) Dataset
is created and used in our experiments. Given a satellite image, we employ
a method described in \cite{ZX:14b} to crop roof images by
registering artificial building footprints with the satellite image.
Later, all roof images are aligned using their footprint principal
directions using a method proposed in \cite{ZQY:08} and then are
scaled to images with resolution of $128\times256$. Two experts are
invited to contribute the labels of 6 different roof styles: flat,
gable, gambrel, halfhip, hip and pyramid. Example instances of SRC
dataset are shown in Fig~\ref{fig: roofexamples}.
This dataset is of great challenges for the task of visual analysis.
First, qualities of the some satellite images are degraded because
of significant image blurring occurred when capturing the satellite
images. Second, roofs in these images are covered by various kinds
of equipments such as air conditioners chimneys and water tanks, and
most of roofs in our dataset are partially occluded by shadows cast
by trees and some other stuffs. Such covering and shadows are great
obstacles to robust visual analysis algorithms. Furthermore, the class
instances of SRC dataset are naturally extreme imbalance, since some
particular types of roofs (such as gambrel and pyramid) are far less
than the other types in the real world. Such unbalanced distributions
of data are compared in in Table \ref{Tab: datadistribution}.
\begin{figure*}[t]
\centering
\includegraphics[scale=0.32]{AllRoofsExamples}\caption{The illustration of the SRC dataset.}
\label{fig: roofexamples}
\end{figure*}
\begin{table}[h]
\centering %
\begin{tabular}{@{}lccc@{}}
\toprule
\textbf{Styles} & \textbf{Training \#} & \textbf{Testing \#} & \textbf{Total \#} \tabularnewline
\midrule
\textbf{Flat} & 1232 & 1748 & 3080 \tabularnewline
\textbf{Gable} & 1111 & 1665 & 2776 \tabularnewline
\textbf{Gambrel} & 156 & 232 & 388 \tabularnewline
\textbf{Halfhip} & 268 & 400 & 668 \tabularnewline
\textbf{Hip} & 960 & 1440 & 2400 \tabularnewline
\textbf{Pyramid} & 133 & 199 & 332 \tabularnewline
\bottomrule
\end{tabular}\protect\protect\caption{The distribution of the roof styles used in the experiments.}
\label{Tab: datadistribution}
\end{table}
Classification of the roof styles in the experiments are based on
recognizing edges detected from the roof images. We employed the adaptive
Otsu edge detection method \cite{otsu:79} to extract edges from the
roof images. We create synthetic prototype to characterize primary
ridge lines or valley lines in a certain type of roof style. Examples
of the synthetic prototypes are shown in Fig. \ref{fig: roofprototype}.
Real roof edge images and matching \textit{Syn I} images are shown
in Fig. \ref{fig: roof real and synI}. To create \textit{Syn II}
images of this dataset, $2000$ synthetic images are produced by interpolation
and extrapolation between images in \textit{Syn I} for each roof style.
\begin{figure}[ht]
\centerline{ %
\begin{tabular}{ccc}
\resizebox{0.14\textwidth}{!}{\rotatebox{0}{ \includegraphics{Gable-crop}}} & \resizebox{0.14\textwidth}{!}{\rotatebox{0}{ \includegraphics{Hip-crop}}} & \resizebox{0.14\textwidth}{!}{\rotatebox{0}{ \includegraphics{Pyramid-crop}}} \tabularnewline
\end{tabular}} \protect\caption{The illustration of synthetic roof prototype we used in the process
of finding matching synthetic data for each real data. There are two
types of control points: red ones and blue ones. Blue control points
are constraint to move along the boundary only. The area of point's
movements masked using light blue. }
\label{fig: roofprototype}
\end{figure}
\subsubsection{Handwritten Digits Dataset}
We also validate our framework on handwritten digits dataset from
UCI machine learning repository \cite{Bache+Lichman:2013} which totally
has $5620$ instances. The handwritten digits from $0$ to $9$ in
this dataset are collected from $43$ people: $30$ contributed to
the training set and the other $13$ to the test set. In the experiments,
the \textit{Syn I} data are generated using Algorithm \ref{Alg: OCP2}.
The \textit{Syn II} data of this dataset is generated using interpolation
and extrapolation as described in Sec 4.
\subsection{Experimental Settings}
We fix the configuration of MCAE as $\langle\mathfrak{i}\text{:}X_{s},\:\mathfrak{t}\text{:}X_{r}\rangle^{L}$
(left channel) and $\langle\mathfrak{i}\text{:}X_{r},\:\mathfrak{t}\text{:}X_{r}\rangle^{R}$
(right channel). Specifically, the left
channel is the reconstruction process from synthetic data to real data,
while the right channel works in the same way as a standard SAE. Our
experimental results will show that the representations learned in
such way greatly benefit the performance of classifiers we compared.
In the experiments%
\footnote{All codes (including our MCAE and creating synthetic data) will be
released once accepted.%
}, two different classifiers of utilizing learned representations from
MCAE (from Sec. 3) are compared. In the first scenario, MCAE encodes
input data to a representation (feature) in the hidden layer and a
SVM using RBF kernel is employed in this case to show the performance
of the classification. In the second scenario, MCAE takes the input
images and produces the reconstructed images at the output layer.
Features, in this case, are images, therefore can be fed to Convolutional
Neural Network (CNN) for classification. In our experiments we build
a LeNet-5 \cite{LY:98} which is originally created for digit recognition.
We show that using the same number of input data, the performance
of the CNN prefers to the data produced by the MCAE.
We summarize all evaluations and comparisons using F-1 score, which
is defined as:
\begin{equation}
F_{1}=2\cdot\frac{\text{Precision}\cdot\text{Recall}}{\text{Precision}+\text{Recall}}
\end{equation}
\subsection{Evaluation%
\footnote{Due to the page limit, please refer to our supplementary material
for the additional experimental results.%
}}
\begin{figure*}
\centering
\begin{tabular}{cc}
\includegraphics[scale=0.45]{original}
&
\includegraphics[scale=0.45]{reconst}\tabularnewline
(a) & (b) \tabularnewline
\end{tabular}
\caption{\label{fig:t-SNE-visualization-of}t-SNE \cite{tsne}visualization
of synthetic gap bridged by MCAE. (a) Data distributions of each class
of SRC dataset. For many data instances, the (circle) real and (dot
points) synthetic data are not overlapping. This is synthetic gap.
(b) Data distributions of the reconstructed images by MCAE for each
class of SRC dataset. The reconstructed images of all the real (circle)
and synthetic (dot points) are almost overlapped. It means that our
MCAE can bridge the synthetic gap.}
\end{figure*}
\textbf{\textcolor{black}{MCAE is better than CIAE and SAE. }} \quad{}To
better evaluate the performance of the proposed MCAE, we compare MCAE
with\textbf{\textcolor{red}{{} }}Concatenate-Input Autoencoder (CIAE)
\cite{multitask_autoencoder2011ICML} and Sparse Autoencoder (SAE)
\cite{VP:10}. In these experiments, we evaluate the performance on
two classifiers: a CNN using reconstructed images and SVM using encoded
hidden layer representation. We present the results of these comparisons
in Table \ref{Tab: RoofBetterAE} and Table \ref{Tab: DigitBetterAE}
for SRC and handwritten digit datasets respectively. It could be observed
from these two tables that although the performance of the CIAE is
close to MCAE, the proposed MCAE gets a better performance almost
in all the comparisons.
\begin{table}[h]
\centering %
\begin{tabular}{@{}lccc@{}}
\toprule
& %
\begin{tabular}{@{}c@{}}
\textbf{Data to train} \tabularnewline
\textbf{autoencoder} \tabularnewline
\end{tabular} & %
\begin{tabular}{@{}c@{}}
\textbf{CNN} \tabularnewline
\textbf{Reconstructed} \tabularnewline
\end{tabular} & %
\begin{tabular}{@{}c@{}}
\textbf{SVM} \tabularnewline
\textbf{Encoded} \tabularnewline
\end{tabular}\tabularnewline
\midrule
\textbf{MCAE} & %
\begin{tabular}{@{}c@{}}
$\langle\mathfrak{i}\text{:}\textit{Syn I},\:\mathfrak{t}\text{:Real}\rangle^{L}$ \tabularnewline
$\langle\mathfrak{i}\text{:}\text{Real},\:\mathfrak{t}\text{:Real}\rangle^{R}$ \tabularnewline
\end{tabular} & \textbf{0.68 } & \textbf{0.80 }\tabularnewline
\midrule
\textbf{CIAE} & %
\begin{tabular}{@{}c@{}}
$\langle\mathfrak{i}\text{:}\textit{Syn I}+\text{Real}$,\tabularnewline
$\;\;\mathfrak{t}\text{:}\textit{Syn I}+\text{Real}\rangle$ \tabularnewline
\end{tabular} & 0.68 & 0.78 \tabularnewline
\midrule
\textbf{SAE} & $\langle\mathfrak{i}\text{:}\textit{Syn I},\:\mathfrak{t}\text{:}\textit{Syn I}\rangle$ & 0.63 & 0.59 \tabularnewline
\midrule
\textbf{SAE} & $\langle\mathfrak{i}\text{:Real},\:\mathfrak{t}\text{:Real}\rangle$ & 0.62 & 0.62 \tabularnewline
\bottomrule
\end{tabular}\protect\caption{F1-score of roof style classification using reconstructed images (in
CNN) and encoded image features (in SVM). Second column shows the
data used to train the autoencoder in the first column. In classification,
Real+\textit{Syn II} are used in the training of CNN and SVM. ; $\textit{Syn I}+\text{Real}$
means that we use concatenation of the Syn I and real images as the input for
the corresponding autoencoders.}
\label{Tab: RoofBetterAE}
\end{table}
\begin{table}[h]
\centering %
\begin{tabular}{@{}lccc@{}}
\toprule
& %
\begin{tabular}{@{}c@{}}
\textbf{Data to train} \tabularnewline
\textbf{autoencoder} \tabularnewline
\end{tabular} & %
\begin{tabular}{@{}c@{}}
\textbf{CNN} \tabularnewline
\textbf{Reconstructed} \tabularnewline
\end{tabular} & %
\begin{tabular}{@{}c@{}}
\textbf{SVM} \tabularnewline
\textbf{Encoded} \tabularnewline
\end{tabular}\tabularnewline
\midrule
\textbf{MCAE} & %
\begin{tabular}{@{}c@{}}
$\langle\mathfrak{i}\text{:}\textit{Syn I},\:\mathfrak{t}\text{:Real}\rangle^{L}$ \tabularnewline
$\langle\mathfrak{i}\text{:}\text{Real},\:\mathfrak{t}\text{:Real}\rangle^{R}$ \tabularnewline
\end{tabular} & \textbf{0.98 } & \textbf{0.96 }\tabularnewline
\midrule
\textbf{CIAE} & %
\begin{tabular}{@{}c@{}}
$\langle\mathfrak{i}\text{:}\textit{Syn I}+\text{Real}$,\tabularnewline
$\;\;\mathfrak{t}\text{:}\textit{Syn I}+\text{Real}\rangle$ \tabularnewline
\end{tabular} & 0.97 & 0.96 \tabularnewline
\midrule
\textbf{SAE} & $\langle\mathfrak{i}\text{:}\textit{Syn I},\:\mathfrak{t}\text{:}\textit{Syn I}\rangle$ & 0.94 & 0.91 \tabularnewline
\midrule
\textbf{SAE} & $\langle\mathfrak{i}\text{:Real},\:\mathfrak{t}\text{:Real}\rangle$ & 0.95 & 0.65 \tabularnewline
\bottomrule
\end{tabular}\protect\caption{F1-score of handwritten digit recognition.}
\label{Tab: DigitBetterAE}
\end{table}
\textbf{Synthetic data help learning a better classifier}. \quad{}We
designed another group of experiments. In these experiments three
different configurations of data are either reconstructed and encoded
using the proposed MCAE, then used to train a CNN or a SVM in the
experiments. All results from these experiments are compared in Table
\ref{Tab: RoofBetterData} and Table \ref{Tab: DigitBetterData} respectively.
An interesting thing to notice is that in experiments, using synthetic
data can only achieve the same result as using a combination of real
and synthetic data. This result proves that the distribution of the
real data in this case is almost overlapping with the distribution
of the synthetic data.
\begin{table}[h]
\centering %
\begin{tabular}{@{}lcccc@{}}
\toprule
& \textbf{Feature type} & \textbf{Real} & \textbf{\textit{Syn II}} & \textbf{Real+}\textbf{\textit{Syn II}} \tabularnewline
\midrule
\textbf{CNN} & Reconstructed & 0.65 & 0.68 & 0.68 \tabularnewline
\textbf{SVM} & Encoded & 0.77 & 0.78 & 0.80 \tabularnewline
\bottomrule
\end{tabular}\protect\caption{F1-score of roof style classification by classifier (CNN and SVM)
using different set of data reconstructed of encoded using the proposed
MCAE.}
\label{Tab: RoofBetterData}
\end{table}
\begin{table}[h]
\centering %
\begin{tabular}{@{}lcccc@{}}
\toprule
& \textbf{Feature type} & \textbf{Real} & \textbf{\textit{Syn II}} & \textbf{Real+}\textbf{\textit{Syn II}} \tabularnewline
\midrule
\textbf{CNN} & Reconstructed & 0.94 & 0.96 & 0.96 \tabularnewline
\textbf{SVM} & Encoded & 0.96 & 0.96 & 0.98 \tabularnewline
\bottomrule
\end{tabular}\protect\caption{F1-score of handwritten digit recognition.}
\label{Tab: DigitBetterData}
\end{table}
\textbf{MCAE bridges the synthetic gap. } \quad{}We compare the correlation
defined as:
\begin{equation}
\text{Corr}=\frac{\text{Cov}(X,Y)}{\text{Var}(X)\text{Var}(Y)}
\end{equation}
between real and \textit{Syn I} data before and after being reconstructed
by the MCAE. The intention of these comparisons is to show that real
synthetic images become much more alike each other in terms of the
appearance after being reconstructed by the MCAE. The results are
shown in Fig. \ref{fig: correlation}. It is shown that our method
almost achieves $100\%$ correlation between real and \textit{Syn
I} when both data are reconstructed by the proposed MCAE. That means
the proposed MCAE bridges the synthetic gap between the real data
and the synthetic data. The results are shown in Fig.~\ref{fig:t-SNE-visualization-of}.
It intuitively shows that our MCAE can help bridge the synthetic gap
between real and synthetic data.
\begin{figure}[ht]
\centerline{ %
\begin{tabular}{c}
\resizebox{0.45\textwidth}{!}{\rotatebox{0}{ \includegraphics{One2OneCorrelation}}} \tabularnewline
\tabularnewline
\resizebox{0.45\textwidth}{!}{\rotatebox{0}{ \includegraphics{One2OneCorrelationRoof}}} \tabularnewline
\end{tabular}} \protect\caption{Correlation between real and corresponding best matching \textit{Syn
I} data.}
\label{fig: correlation}
\end{figure}
\section{Conclusion}
In this paper we identify the problem of synthetic gap. By solving this problem, in our experiments, we demonstrate that the synthetic data could be used to improve the performance of classifiers. To better learn classifiers from synthetic data, we have proposed a novel Multichannel autoencoder (MCAE) model. MCAE has multiple channels in its structure and is an extension from standard autoencoder. We show that MCAE not only bridges the synthetic gap between real data and synthetic data, it also jointly learns from both real and synthetic data, thus can provide more robust representation for both data. To facilitate the study on satellite image analysis, we introduce a novel benchmark dataset -- SRC as one dataset used in our experiments. The proposed method has been validated on SRC and handwritten digits datasets.
{\small{}\bibliographystyle{abbrv}
|
\section{Introduction}
One of the most influent result in class field theory is Chebotarev's density theorem. As it is well known, this result is a deep generalization of the Theorem of Dirichlet about equidistribution of rational primes in arithmetic progression and gives a complete understanding of the distribution of primes in a fixed Galois number field extension with respect to their splitting behavior (for an interesting discussion of the theorem and its original proof see \cite{LS96}).
In the function field case, the parallel statement is carried over by the Sato-Tate conjecture for curves, which studies the distribution of the Frobenius endomorphism of the reduction modulo $p$ of a fixed curve, when the prime $p$ varies.
In order to complement this research line in other directions, several mathematicians were led to consider the following new general problem:
given a family of curves, of genus $g$ over $\mathbb{F}_{q}$, satisfying certain properties, understand the distribution of the Frobenius endomorphism of the curves of the family. This is sometimes called the \emph{vertical Sato-Tate conjecture}, since the prime $p$ is fixed and the curve varies in the family. We can study the limiting distribution in two different ways, depending on whether we let the genus $g$ or the cardinality $q$ of the field tend to infinity. It is then interesting to compare both limit results.
When $g$ is fixed and $q$ goes to infinity the problem can be solved thanks to Deligne's equidistribution theorem (cf. \cite{KS99}) while for the complementary case different techniques are applied depending on the particular family considered. The fluctuation in the number of points at the $g$-limit has been studied for different families of curves, such as:
\begin{itemize}
\item
Hyperelliptic curves , cf. \cite{KR09}, \cite{Betal09},
\item
Cyclic trigonal curves (i.e. cyclic 3-covers of the projective line), cf. \cite{Betal09}, \cite{Xiong2010},
\item
General trigonal curves, cf. \cite{Wood2012},
\item
$p$-fold cover of the projective line, \cite{Betal2011},
\item
$\ell$-covers of the projective line, cf. \cite{Betal09}, \cite{Betal2015}.
\end{itemize}
In the present paper, we study the distribution of the number of points over $\mathbb{F}_{q}$ for a genus $g$ curve $C$ defined over $\mathbb{F}_{q}$ which is a quartic non-cyclic cover of the projective line $\mathbb{P}^{1}_{\mathbb{F}_{q}}$, at the $q$-limit (for a genus $g$ fixed) and at the $g$-limit (with $q$ fixed). This is the first time that a family of non-cyclic abelian covers is studied. The distribution obtained is different to the product of probabilities for the family of hyperelliptic curves, what at first sight could be guessed. Therefore, the study of this family seems to be the first natural step is order to understand the general abelian case.
Let $\mathcal{B}_{g}(\mathbb{F}_q)$ be the family of genus $g$ quartic non-cyclic cover of the projective line $\mathbb{P}^{1}_{\mathbb{F}_{q}}$, and consider the following decomposition
$$
\mathcal{B}_{g}(\mathbb{F}_q)=\bigcup_{g_{1}+g_{2}+g_{3}=g}\mathcal{B}_{(g_{1},g_{2},g_{3})}(\mathbb{F}_q)
$$
where $\mathcal{B}_{(g_{1},g_{2},g_{3})}(\mathbb{F}_q)$ denotes the subfamily of curves $C\in\mathcal{B}_{g}(\mathbb{F}_q)$ such that the three hyperelliptic quotients of $C$ have genera $g_{1},g_{2}$ and $g_{3}$.
The main theorem of the paper is the following:
\bigskip\noindent {\bf Theorem \ref{main}}\, {\em
If the three genera $g_{1},g_{2},g_{3}$ go to infinity, then we have that
$$
\frac{|\{C\in\mathcal{B}_{(g_1,g_2,g_3)}(\mathbb{F}_q):\,\mathrm{Tr}(\mathrm{Frob}_C)=-M\}|}{|\mathcal{B}_{(g_1,g_2,g_3)}(\mathbb{F}_q)|}=
\mathrm{Prob}\left(\sum_{j=1}^{q+1}X_{j}=M\right)
$$
where the $X_{j}$ are i.i.d. (identically independently distributed) random variables such that
$$
X_{i}=\left\{\begin{array}{ccc}-1 & \rm{with\,probability}& \frac{3(q+2)}{4(q+3)}\\
\,&\,&\,\\
1 & \rm{with\,probability}& \frac{6}{4(q+3)}\\
\,&\,&\,\\
3 & \rm{with\,probability}& \frac{q}{4(q+3)}
\end{array}\right..$$}
\subsection*{Outline}
In Section $2$, we introduce the family of biquadratic curves and we give a parametrization of the family in terms of terns of coprime square-free polynomials. In Section $3$, we compute the monodromy group of the family in the sense of Katz and Sarnak (cf. \cite[Ch. 9]{KS99}) and we obtain the corresponding distribution of the Frobenius traces at the $q$-limit. In Section $4$, previous theorem is proven, and in Section $5$ the moments of the Frobenius traces are computed at the $g$-limit, proving that when both $g$ and $q$ go to infinity the normalized trace has a standard complex Gaussian distribution. In last section, Theorem \ref{main} is generalized for an arbitrary cover of the projective line with Galois group isomorphic to $r$ copies of $\mathbb{Z}/2\mathbb{Z}$. The paper concludes with an Appendix, writen by Alina Bucur, giving the heuristic for the distribution of the number of points for the whole family of $r$-quadratic curves.
\subsection*{Notations} We now fix some notations and conventions that will be valid in the sequel.
\begin{itemize}
\item
$p\neq 2$ is a prime integer, and $q$ is a positive power of $p$.
\item
$k=\mathbb{F}_{q}(t)$ is the function field of $\mathbb{P}^{1}_{\mathbb{F}_q}$, and $K/k$ is a finite extension.
\item
$(f,g)$ denotes the greatest common divisor of two polynomials $f,g\in\mathbb{F}_q[t]$.
\item
$\text{deg}(f)$ denotes the degree of a polynomial $f$
\item
$|f|:=q^{\text{deg}(f)}$ denotes the norm of a polynomial $f$.
\item
$\tilde{f}$ is the polynomial obtained inverting the order of the coefficients of $f$.
\item
$g(C)$ denotes the geometric genus of the projective curve $C/\mathbb{F}_q$,
\item
and $\text{Frob}_C$ denotes its geometric Frobenius morphism.
\end{itemize}
\subsection*{Acknowledgements} This work was started in March 2014 at the Arizona Winter School ``Arithmetic Statistic'' and the authors would like to thank the organizers for creating such a stimulating working environment. Moreover, the authors are grateful to Alina Bucur and Chantal David for initiating them into this rich and appealing field of research in number theory and for several discussions and helpful comments during the preparation of the paper. Finally, we would like to thank Patrick Meisner for carefully reading a first draft of the paper and for his useful comments.
\section{The family of biquadratic curves}\label{sec:family}
We first define and give the basic properties of the family of biquadratic curves. We determine its genus in terms of the equations defining the curves, and we study the irreducible components of the coarse moduli space of biquadratic curves.
Recall that if $K/\mathbb{F}_q(t)$ is a finite Galois extension such that $K\cap \bar{\mathbb{F}}_q=\mathbb{F}_q$, then there exists, up to isomorphism, a unique nonsingular projective curve
$C$ with function field $\mathbb{F}_q(C)=K$, together with a regular morphism $\varphi: C\rightarrow\mathbb{P}^{1}_{\mathbb{F}_{q}}$ defined over $\mathbb{F}_{q}$ (cf. \cite[I,Th. 6.6, Th.6.9]{HarBook}).
\begin{defn}\label{def_biq_curve}
We call biquadratic curve a smooth projective curve $C$, together with a regular morphism $\varphi: C\rightarrow\mathbb{P}^{1}_{\mathbb{F}_{q}}$ defined over $\mathbb{F}_{q}$, that induces a field extension with Galois group $\mathrm{Gal}(\mathbb{F}_q(C)/\mathbb{F}_q(t))\simeq \mathbb{Z}/2\mathbb{Z}\times\mathbb{Z}/2\mathbb{Z}$.
\end{defn}
Since $\mathrm{char}(k)\neq 2$, it is clear that every non-cyclic quartic extension of $k$ is of the form $K=k(\sqrt{h_1(t)},\sqrt{h_2(t)})$, for some $h_1(t),h_{2}(t)\in\mathbb{F}_q[t]$ different non-constant polynomials, that we can take to be square-free. Moreover, if the leading coefficient of $h_{i}$ is a square in $\mathbb{F}_{q}$, then we can assume that this is equal to $1$. Therefore, if $C$ is a biquadratic curve, then an affine model of $C$ in $\mathbb{A}^{3}_{\mathbb{F}_{q}}$ is given by
$$
C:\begin{cases}
y_{1}^{2}=h_1(t) \\
y_{2}^{2}=h_2(t)
\end{cases}.
$$
\begin{remark}\label{coprime}
If $K:=k(\sqrt{h_1(t)},\sqrt{h_2(t)})$ is a biquadratic extension of $k$, then there are exactly 3 different quadratic subextensions of $K$, namely $k(\sqrt{h_{1}}),k(\sqrt{h_{2}})$ and $k(\sqrt{h_{1}h_{2}})$.
If we write $h_{i}=f_{i}f$ for $i=1,2$, with $f=(h_1,h_2)$, then clearly we have that $(f_{1},f_{2})=(f_1,f)=(f_2,f)=1$ and these three subextensions are $k(\sqrt{ff_{1}}),k(\sqrt{ff_{2}})$ and $k(\sqrt{f_{1}f_{2}})$.
Two such extensions $k(\sqrt{h_1(t)},\sqrt{h_2(t)})$ and $k(\sqrt{h'_1(t)},\sqrt{h'_2(t)})$ define the same biquadratic extension if and only if we have the equality of sets
$$
\{h_1,h_2,\frac{h_1h_2}{(h_1,h_2)^2}\}=\{h'_1,h'_2,\frac{h'_1h'_2}{(h'_1,h'_2)^2}\}.
$$
\end{remark}
\begin{remark}\label{ram_infty}
Recall that if $\pi:C\rightarrow\mathbb{P}^{1}$ is a degree $2$ regular cover, whose affine plane model is $y^{2}=F(t)$, with $F(t)$ a square-free polynomial over $\mathbb{F}_{q}$, then the point at infinity is ramified in the cover $\pi$ if and only if the degree $d$ of $F$ is odd. Indeed, if we take take $u=\frac{1}{t}$, then the function field of $C$ is
$$
k(C)=k(\sqrt{F(t)})=k(\sqrt{F(1/u)})=k(\sqrt{u^{-d}\tilde{F}(u)})
$$
and then it is clear that $t=\infty$ ramifies if and only if the point $u=0$ ramifies, i.e. if and only if $d$ is odd.
\end{remark}
\begin{proposition} \label{genus}
Let $h_1(t),h_{2}(t)\in\mathbb{F}_q[t]$ be different square-free polynomials, and let $C$ be the curve whose function field is $k(C)=k(\sqrt{h_1(t)},\sqrt{h_2(t)})$. For every $i=1,2$, write $h_i=ff_i$, with $f=(h_{1},h_{2})$, and define $h_{3}:=f_{1}f_{2}$.
If we denote by $C_{i}$ the hyperelliptic curve whose affine plane model is given by the equations $y^{2}=h_{i}(t)$, for $i=1,2,3$, then we have the following formula for the genus of $C$:
$$
g(C)=g(C_{1})+g(C_{2})+g(C_{3}).
$$
Moreover, if we denote by $n:=\operatorname{deg}(f)$ and $n_i:=\operatorname{deg}(f_i)$,
$$
g(C)=g(n_1,n_2,n):=n_1+n_2+n+e_\infty-4,
$$
where $e_{\infty}$ is the ramification index at the point at infinity, that is,
$$
e_{\infty}:=\left\{\begin{array}{ll}2, & \text{if}\;n\equiv n_{1}\equiv n_{2}\equiv 0\,(\mathrm{mod}\,2)\\
1, &otherwise
\end{array} \right..
$$
\end{proposition}
\begin{proof}
Let us denote by $R:=\mathrm{Ram}(\pi)$ the subset of all points of $\mathbb{P}^{1}_{\mathbb{F}_q}$ which are ramified in the cover $\pi: C\longrightarrow\mathbb{P}^{1}_{\mathbb{F}_q}$. Riemann-Hurwitz's formula (cf. \cite[Theorem 7.16]{Rosen02}) implies that $2g(C)-2=4(2\cdot 0-2)+2|R|$. That is, $g(C)=|R|-3$. Again, for the hyperelliptic cover
$\pi_{i}:C_{i}\longrightarrow\mathbb{P}^{1}$ and the ramification sets $R_{i}:=\mathrm{Ram}(\pi_{i})$, we get $g(C_{i})=\frac{|R_{i}|}{2}-1$. Now, the definition of $h_3$ implies that
$$
2|R_{1}\cup R_{2}\cup R_{3}|=|R_{1}|+|R_{2}|+|R_{3}|.
$$
Thus, the formula $g(C)=g(C_{1})+g(C_{2})+g(C_{3})$ holds.
We can also apply Riemann-Hurwitz's formula to the morphism $\pi$, and so we have
$$
2g-2=4(2\cdot 0-2)+2\cdot(n_1+n_2+n_3+e_{\infty}-1).
$$
\end{proof}
Now, we introduce some sets of polynomials that will be useful:
$$
V_d=\{ F\in \mathbb{F}_q[t]:\,F\,\text{monic},\,\text{deg}(F)=d\},
$$
$$
\mathcal{F}_d=\{ F\in \mathbb{F}_q[t]:\,F\,\text{monic},\,\text{square-free},\,\text{deg}(F)=d\},
$$
$$
\widehat{\mathcal{F}}_d=\{ F\in \mathbb{F}_q[t]:\,F\,\text{square-free},\,\text{deg}(F)=d\},
$$
$$
\mathcal{F}_{(n,n_1,n_2)}=\{(f,f_1,f_2)\in \mathcal{F}_n\times \mathcal{F}_{n_1}\times \mathcal{F}_{n_2}:\,(f,f_1)=(f,f_2)=(f_1,f_2)=1\},
$$
$$
\widehat{\mathcal{F}}_{(n,n_1,n_2)}=\{(f,f_1,f_2)\in \mathcal{F}_n\times \widehat{\mathcal{F}}_{n_1}\times \widehat{\mathcal{F}}_{n_2}:\,(f,f_1)=(f,f_2)=(f_1,f_2)=1\},
$$
$$
\mathcal{F}_{[n,n_1,n_2]}=\mathcal{F}_{(n,n_1,n_2)}\cup \mathcal{F}_{(n-1,n_1,n_2)} \cup \mathcal{F}_{(n,n_1-1,n_2)}\cup \mathcal{F}_{(n,n_1,n_2-1)},
$$
$$
\widehat{\mathcal{F}}_{[n,n_1,n_2]}=\widehat{\mathcal{F}}_{(n,n_1,n_2)}\cup \widehat{\mathcal{F}}_{(n-1,n_1,n_2)} \cup \widehat{\mathcal{F}}_{(n,n_1-1,n_2)}\cup \widehat{\mathcal{F}}_{(n,n_1,n_2-1)}.
$$
\begin{defn} We denote by $\mathcal{B}_g(\mathbb{F}_{q})$ the family of biquadratic curves defined over $\mathbb{F}_{q}$ and of fixed genus $g$. It can be written as a disjoint union of subfamilies indexed by unordered $3$-tuples of positive integers $g_1,g_2,g_3$, i.e.
$$
\mathcal{B}_g(\mathbb{F}_{q})=\bigcup_{g_1+g_2+g_3=g }\mathcal{B}_{(g_1,g_2,g_3)}(\mathbb{F}_{q}),
$$
where $\mathcal{B}_{(g_1,g_2,g_3)}(\mathbb{F}_{q})$ denotes the subfamily of biquadratic curves of genus $g=g_1+g_2+g_3$ such that the intermediate curves given by the morphism to $\mathbb{P}^1$ have genus equal to $g_1,g_2$ and $g_3$. This family is in bijection with the family of curves defined by elements in the set of polynomials $\widehat{\mathcal{F}}_{[n,n_{1},n_{2}]}$ such that $g_{i}=\lfloor \frac{n+n_{i}-1}{2}\rfloor$ for $i=1,2$ and $g_{3}=\lfloor \frac{n_1+n_2-1}{2}\rfloor$.
\end{defn}
The family $\mathcal{B}_{g}(\bar{\mathbb{F}}_{q})$ of biquadratic curves defined over $\bar{\mathbb{F}}_{q}$ is a coarse moduli space over $\mathbb{Z}[1/2]$ (cf. \cite[Lemma 3.1]{Pries2005}). A detailed geometric study of this moduli space can be found in \cite{Pries2005} and \cite{Pries2005b}.
\begin{remark} One has the following equalities:
$$
|\mathcal{B}_{(g_1,g_2,g_3)}(\mathbb{F}_{q})|\mbox{'}=\sideset{}{\mbox{'}}\sum_{\scriptscriptstyle C\in \mathcal{B}_{(g_1,g_2,g_3)}(\mathbb{F}_{q})} 1=\sum_{\scriptscriptstyle F\in \widehat{\mathcal{F}}_{[n,n_{1},n_{2}]}}\frac{1}{|\text{Aut}(C)|}=\frac{|\widehat{\mathcal{F}}_{[n,n_{1},n_{2}]}|}{q(q^2-1)},
$$
where the $'$ notation, applied both to cardinality and summation, means that each one of the curves $C$ in the moduli spaces is counted with
the usual weight $\frac{1}{|\text{Aut}(C)|}$.
\end{remark}
\begin{remark}\label{symmetry} Notice that $|\widehat{\mathcal{F}}_{(n,n_1,n_2)}|=(q-1)^2|\mathcal{F}_{(n,n_1,n_2)}|$ and that we can see the set $\widehat{\mathcal{F}}_{(n,n_1,n_2)}$ as the set of the quadratic twists of elements in $\mathcal{F}_{(n,n_1,n_2)}$ given by the equations
$$
C':\begin{cases}y_{1}^{2}=\alpha_{1} ff_1(t)\\y_{2}^{2}=\alpha_{2} ff_2(t)\end{cases}
$$
where $\alpha_{1},\,\alpha_{2}\in\mathbb{F}_{q}^{*}$.
\end{remark}
\section{Monodromy group of the family}
A useful reference for this section is \cite[Ch. 9]{KS99}.
Let $S$ be an open set of $\mathrm{Spec}\,\mathbb{F}_{q}$ and let $\mathcal{C}\rightarrow S$ be a smooth proper morphism of schemes such that the geometric fibers $C_{x}\otimes\overline{\mathbb{F}}_{q}$ are smooth projective curves of genus $g$ over $\overline{\mathbb{F}}_{q}$.
Fix a prime integer $\ell\neq p$. Then, there exists an $\ell$-adic representation
$$
\rho_{\ell}:\pi_{1}(S)\longrightarrow\mathrm{GL}_{2g}(\overline{\mathbb{Q}}_{\ell})
$$
with the following interpolation property:
for every closed point $x:\mathrm{Spec}\,\mathbb{F}_{q}\longrightarrow S$ the induced representation
$$
\mathrm{Gal}(\overline{\mathbb{F}}_{q}/\mathbb{F}_{q})\simeq\pi_{1}(\mathrm{Spec}\,\mathbb{F}_{q})\longrightarrow\pi_{1}(S)
\longrightarrow\mathrm{GL}_{2g}(\overline{\mathbb{Q}}_{\ell})
$$
is isomorphic to the $\ell$-adic representation
$$
\rho_{C_{x},\ell}:\mathrm{Gal}(\overline{\mathbb{F}}_{q}/\mathbb{F}_{q})\longrightarrow\mathrm{Aut}(\mathrm{H}_{et}^{1}(C_{x}\otimes_{\mathbb{F}_{q}}\overline{\mathbb{F}}_{q},\overline{\mathbb{Q}}_{\ell}))\simeq\mathrm{GL}_{2g}(\overline{\mathbb{Q}_{\ell}}).
$$
Once an embedding $\iota:\overline{\mathbb{Q}}_{\ell}\hookrightarrow\mathbb{C}$ is fixed, we have a $2g$-dimensional complex representation $\iota\cdot\rho_{\ell}$. The image of this representation is a subgroup of $\mathrm{GL}_{2g}(\mathbb{C})$ called the monodromy group of the family.
For every integer $d\geq 1$, the set of polynomials $\mathcal{F}_{d}$ defined in Section \ref{sec:family} can be algebraically realized as a Zariski-open subset of $\mathbb{A}^{d}_{\mathbb{F}_{q}}$. This could be done redefining it in the following way:
$$
\mathcal{F}_{d}:=\{(a_{0},\dots,a_{d-1})\in\mathbb{A}^{d}_{\mathbb{F}_{q}}\mid\;D(a_{0},\dots,a_{d-1})\neq 0\},
$$
where $D:\mathbb{A}^{d}_{\mathbb{F}_{q}}\longrightarrow\mathbb{A}^{1}_{\mathbb{F}_{q}}$ is the continuos function such that $D(a_{0},a_{1},\dots,a_{d-1})$ denotes the discriminant of the monic polynomial $a_{0}+a_{1}t+\dots+t^{d}\in\mathbb{F}_{q}[t]$.
Let $\mathcal{H}_{g}$ denote the family of genus $g$ hyperelliptic curves over $\mathcal{F}_{d}$, whose fiber over the polynomial
$F\in\mathcal{F}_{d}$ is given by the curve whose affine plane model is $y^{2}=F(t)$. In \cite[10.1]{KS99}, it is proved that the monodromy group either of the family $\mathcal{H}_{g}$ over $\mathcal{F}_{2g+1}$ and of the family $\mathcal{H}_{g}$ over $\mathcal{F}_{2g+2}$ is $G_{geom}=\mathrm{Sp}_{2g}(\mathbb{C})$.
\begin{proposition}
The monodromy group of the family $\mathcal{B}_{g_1,g_2,g_3}(\mathbb{F}_{q})$ is the biggest possible one, namely it is the symplectic group $\mathrm{Sp}_{2g}(\mathbb{C})$.
\end{proposition}
\begin{proof}
The set of polynomials $\mathcal{F}_{[n,n_1,n_2]}$ defined in Section \ref{sec:family} can be realized as a Zariski-open subset of $\mathbb{A}^{n}_{\mathbb{F}_{q}}\times\mathbb{A}^{n_{1}+1}_{\mathbb{F}_{q}}\times\mathbb{A}^{n_{2}+1}_{\mathbb{F}_{q}}$.
The family of genus $g$ curves over $\mathcal{F}_{[n,n_1,n_2]}$, whose fiber over the $3$-tuple $(f,f_1,f_2)\in \mathcal{F}_{[n,n_1,n_2]}$ is given by the curve whose affine model is $y_{1}^{2}=ff_{1}(t),y_{2}^{2}=ff_{2}(t)$, is exactly the subfamily of genus $g$ biquadratic curves $\mathcal{B}_{g_{1},g_{2},g_{3}}(\mathbb{F}_{q})$ defined in Section \ref{sec:family}.
Let $N:=\mathrm{max}(n,n_2,n_3)$. By the symmetry of the parametrization we can assume for example that $N=n$ and then we fix two square-free polynomials $f_1,f_2$ of degrees $n_1,n_2$ such that $f_1f_2$ is also square-free. Therefore, we can consider the open immersion
$$
\{f\in\mathcal{F}_{n}:\, ff_1,ff_2\;\mathrm{square-free}\}\longrightarrow \mathcal{F}_{[n,n_1,n_2]}:\, f \longmapsto (f,f_1,f_2).
$$
The monodromy group of the family $\mathcal{H}_{g}$ of hyperelliptic curves over this subset of $\mathcal{F}_{n}$ is the same as if we consider the family over all $\mathcal{F}_{n}$. Finally, the monodromy group of the family $\mathcal{B}_{(g_1,g_2,g_3)}(\mathbb{F}_{q})$ can only increase and, after results of \cite[10.1]{KS99}, it is the biggest possible one.
\end{proof}
Applying Deligne's equidistribution theorem (cf. \cite[9.3,9.2]{KS99}) and random matrix theory (cf. \cite[4]{DS94}), we have the following distribution result at the $q$-limit for the family $\mathcal{B}_{(g_1,g_2,g_3)}(\mathbb{F}_{q})$.
\begin{cor}
Let $g\geq 3$ be a fixed integer. When $q$ goes to $\infty$, the classes of the Frobenius automorphisms $\{\mathrm{Frob}_{C}\}_{C\in\mathcal{B}_{(g_1,g_2,g_3)}(\mathbb{F}_{q})}$ acting on the first \'etale cohomology group $\mathrm{H}^{1}_{\text{\'et}}(C,\mathbb{Q}_{\ell})$ are equidistributed with respect to the Haar mesure associated to the maximal compact subgroup of $\mathrm{Sp}_{2g}(\mathbb{C})$, i.e.
$$
\lim_{q\to\infty}\langle\mathrm{Tr}\,\mathrm{Frob}^{m}_{C}\rangle=\left\{\begin{array}{ll} 2g & m=0\\ -\eta_{r} & 1\leq|m|\leq 2g \\ 0 & |m|>2g
\end{array}\right.
$$
where
$$
\eta_{m}:=\left\{\begin{array}{ll} 1 & m\;\mathrm{even}\\ 0 & m\;\mathrm{odd}
\end{array}\right..
$$
\end{cor}
\section{The number of points over $\mathbb{F}_{q}$}\label{distributionFq}
Let $\chi$ denote the quadratic character in $\mathbb{F}_q$. We set, for any element $(f,f_1,f_2)$ in $\widehat{\mathcal{F}}_{(n,n_1,n_2)}$,
$$
S(f,f_1,f_2)=\sum_{x\in\mathbb{F}_q}(\chi(f\cdot f_1(x))+\chi(f\cdot f_2(x))+\chi(f_1\cdot f_2(x))) \text{, and}
$$
$$
\widehat{S}(f,f_1,f_2)=\sum_{x\in\mathbb{P}^{1}(\mathbb{F}_q)}(\chi(f\cdot f_1(x))+\chi(f\cdot f_2(x))+\chi(f_1\cdot f_2(x))),
$$
where for the point at infinity we define
$$
\chi(F(\infty))=\begin{cases}0&\text{deg}(F)\;\text{odd}\\
1&\text{deg}(F)\;\text{even, leading coefficient is a square in }\mathbb{F}_q\\
-1&\text{deg}(F)\;\text{even, leading coefficient is not a square in }\mathbb{F}_q\end{cases}.
$$
Then, for a curve $C\in\mathcal{B}_{(g_1,g_2,g_3)}(\mathbb{F}_q)$ defined by a $3$-tupla $(f,f_1,f_2)$ we have that
$$
\#C(\mathbb{F}_q)=q+1+\widehat{S}(f,f_1,f_2).
$$
Hence, we have the equality
$$
\frac{|\{C\in\mathcal{B}_{(g_1,g_2,g_3)}(\mathbb{F}_q):\,\text{Tr}(\text{Frob}_C)=-M\}|\mbox{'}}{|\mathcal{B}_{(g_1,g_2,g_3)}(\mathbb{F}_q)|\mbox{'}}=
\frac{|\{(f,f_1,f_2)\in\widehat{\mathcal{F}}_{[n,n_1,n_2]}:\,\widehat{S}(f,f_1,f_2)=M\}|}{|\widehat{\mathcal{F}}_{[n,n_1,n_2]}|}.
$$
The goal of this section is to prove the following theorem.
\begin{thm}\label{main} If the three degrees $n,n_{1},n_{2}$ go to infinity, then we have
$$
\frac{|\{(f,f_1,f_2)\in\widehat{\mathcal{F}}_{[n,n_1,n_2]}:\,\widehat{S}(f,f_1,f_2)=M\}|}{|\widehat{\mathcal{F}}_{[n,n_1,n_2]}|}=
\mathrm{Prob}\left(\sum_{j=1}^{q+1}X_{j}=M\right),
$$
where the $X_{j}$ are i.i.d. random variables such that
$$
X_{i}=\left\{\begin{array}{ccc}-1 & \rm{with\,probability}& \frac{3(q+2)}{4(q+3)}\\
\,&\,&\,\\
1 & \rm{with\,probability}& \frac{6}{4(q+3)}\\
\,&\,&\,\\
3 & \rm{with\,probability}& \frac{q}{4(q+3)}
\end{array}\right..$$
More precisely,
$$
\frac{|\{(f,f_1,f_2)\in\widehat{\mathcal{F}}_{[n,n_1,n_2]}:\,\widehat{S}(f,f_1,f_2)=M\}|}{|\widehat{\mathcal{F}}_{[n,n_1,n_2]}|}
=\mathrm{Prob}\left(\sum_{j=1}^{q+1}X_{j}=M\right)
\left(1+ O(q^{-\frac{(1-\epsilon)}{2}\text{min}(n,n_1,n_2)+q})\right).
$$
\end{thm}
The proof of this Theorem runs similarly to the proof of the equivalent statement for hyperelliptic curves (resp. $l-$cyclic covers) in \cite{KR09} (resp. \cite{Betal09}).
\begin{lemma}\label{S} $($\cite[Lemma 4.2]{Betal09}$)$ For $0\leq l\leq q$, let $x_1,...,x_l$ be distinct elements of $\mathbb{F}_q$. Let $U\in\mathbb{F}_q[t]$ be such that $U(x_i)\neq 0$ for $i=0,...,l$. Let $a_1,...,a_l$ be elements of $\mathbb{F}_{q}^{*}$.Then the cardinality of the set
$$
\mathcal{S}^{U}_{d}(a_{1},\dots,a_{l}):=\{F\in\mathcal{F}_d:\,(F,U)=1,\,F(x_i)=a_i,\,1\leq i\leq l\}
$$
is the number
$$
S^{U}_{d}(l)=\frac{q^{d}}{\zeta_q(2)}\left(\frac{q}{q^{2}-1}\right)^{l}\prod_{P|U}(1+|P|^{-1})^{-1}\left(1+O(q^{l-d/2})\right).
$$
\end{lemma}
\begin{lemma}\label{guay} For $0\leq l\leq q$ let $x_1,...,x_l$ be distinct elements of $\mathbb{F}_q$. Let $U\in\mathbb{F}_q[t]$ be such that $U(x_i)\neq 0$ for $i=0,...,l$. Let $a_1,...,a_l,b_1,...,b_l$ be elements of $\mathbb{F}_{q}^{*}$. Then the cardinality of the set
$$
\mathcal{R}^{U}_{n_1,n_2}(a_{1},\dots,a_{l},b_{1},\dots,b_{l}):=\{(f_1,f_2)\in\mathcal{F}_{n_1}\times\mathcal{F}_{n_2}:\,(f_i,U)=(f_1,f_2)=1,\,f_1(x_i)=a_i,\,f_2(x_i)=b_i, \,1\leq i\leq l \}
$$
is the number
$$
R^{U}_{n_1,n_2}(l)=\frac{q^{n_1+n_2}L}{\zeta_{q}^{2}(2)}\left(\frac{q}{(q-1)^{2}(q+2)}\right)^{l}\prod_{P|U}\left(\frac{1}{1+2|P|^{-1}}\right)\left(1+O(q^{l-\frac{\text{min}(n_1,n_2)}{2}})\right),
$$
where $L:= \prod_{P\,\text{prime}}(1-\frac{|P|^{-2}}{(1+|P|^{-1})^2})$.
\end{lemma}
\begin{proof} By inclusion-exclusion principle (same notations as in \cite[Theorem 13.5]{CombBook}), with
$$
f(D)=|\{(f_1,f_2)\in\mathcal{F}_{n_1}\times\mathcal{F}_{n_2}:\,(f_i,U)=1,\,D|(f_1,f_2),\,f_1(x_i)=a_i,\,f_2(x_i)=b_i, \,1\leq i\leq l \}|,
$$
$$
g(D)=|\{(f_1,f_2)\in\mathcal{F}_{n_1}\times\mathcal{F}_{n_2}:\,(f_i,U)=1,\,(f_1,f_2)=D,\,f_1(x_i)=a_i,\,f_2(x_i)=b_i, \,1\leq i\leq l \}|,
$$
where $D$ is a polynomial in $\mathbb{F}_q[x]$, we have
$$
R^{U}_{n_1,n_2}(l)=g(1)=\sum_{\scriptscriptstyle D,\,D(x_i)\neq 0,(D,U)=1}\mu (D) f(D).
$$
But notice that when $(D,U)=1$
$$
f(D)=|\{(f_1,f_2)\in\mathcal{F}_{n_1-\text{deg}(D)}\times\mathcal{F}_{n_2-\text{deg}(D)}:\,(f_i,UD)=1,\,f_1(x_i)=a_i,\,f_2(x_i)=b_i, \,1\leq i\leq l\}|,
$$
hence Lemma \ref{S} implies
\begin{multline*}
f(D)=S^{UD}_{n_1-\text{deg}(D)}(l)\cdot S^{UD}_{n_2-\text{deg}(D)}(l)=\\
=\frac{q^{n_1+n_2-2\text{deg}(D)}}{\zeta_{q}^{2}(2)}\left(\frac{q}{q^{2}-1}\right)^{2l}\prod_{P|UD}(1+|P|^{-1})^{-2}\left(1+O(q^{l+\frac{\text{deg}(D)}{2}-\frac{\text{min}(n_1,n_2)}{2}})\right).
\end{multline*}
So, one has
\begin{multline*}
R^{U}_{n_1,n_2}(l)=\sum_{\scriptscriptstyle D,\,D(x_i)\neq 0,(D,U)=1}\mu (D) f(D)=\\
\frac{q^{n_1+n_2}}{\zeta_{q}^{2}(2)}\left(\frac{q}{q^{2}-1}\right)^{2l}\prod_{P|U}(1+|P|^{-1})^{-2}\sum_{\begin{array}{c}\scriptscriptstyle D(x_i)\neq 0,\,(D,U)=1\\\scriptscriptstyle \text{deg}(D)\leq \text{min}(n_1,n_2)\end{array}}\mu(D)|D|^{-2}\prod_{P|D}(1+|P|^{-1})^{-2}\left(1+O(q^{l-\frac{\text{min}(n_1,n_2)}{2}})\right).
\end{multline*}
Now, we observe that
\begin{multline*}
\sum_{\begin{array}{c}\scriptscriptstyle D(x_i)\neq 0,\, (D,U)=1\\ \scriptscriptstyle\text{deg}(D)\leq \text{min}(n_1,n_2)\end{array}}\mu(D)|D|^{-2}\prod_{P|D}(1+|P|^{-1})^{-2}=\\
\sum_{\scriptscriptstyle D,\,D(x_i)\neq 0,\,(D,U)=1}\mu(D)|D|^{-2}\prod_{P|D}(1+|P|^{-1})^{-2}+O(q^{-2\text{min}(n_1,n_2)}),
\end{multline*}
where we have that
\begin{multline*}
\sum_{\scriptscriptstyle D,\,D(x_i)\neq 0,\,(D,U)=1}\mu(D)|D|^{-2}\prod_{P|D}(1+|P|^{-1})^{-2}=\\
=\left(\frac{(q+1)^{2}}{q(q+2)}\right)^{l}\prod_{P|U}\left(\frac{1+2|P|^{-1}}{(1+|P|^{-1})^2}\right)^{-1}\prod_{P\,\text{prime}}\left(1-\frac{|P|^{-2}}{(1+|P|^{-1})^2}\right)
=\left(\frac{(q+1)^{2}}{q(q+2)}\right)^{l}\prod_{P|U}\left(\frac{1+2|P|^{-1}}{(1+|P|^{-1})^2}\right)^{-1}L.
\end{multline*}
We can prove that $0<L<1$ (see next Remark \ref{non_vanishing_L}).
So, finally
$$
R^{U}_{n_1,n_2}(l)=\frac{q^{n_1+n_2}L}{\zeta_{q}^{2}(2)}\left(\frac{q}{(q-1)^{2}(q+2)}\right)^{l}\prod_{P|U}\left(\frac{1}{1+2|P|^{-1}}\right)\left(1+O(q^{l-\frac{\text{min}(n_1,n_2)}{2}})\right).
$$
\end{proof}
\begin{remark}\label{non_vanishing_L}
We need to prove that the infinite product
$\prod_{P\,\text{prime}}(1-\frac{|P|^{-2}}{(1+|P|^{-1})^2})$
converges to a real number $L$ such that $0<L<1$. The Prime Polynomial Theorem implies that this is equivalent to prove that the infinite product
$$\prod_{\nu\geq 1} \left(1-\frac{1} {(q^\nu + 1)^2} \right)^{\frac{q^{\nu}}{\nu}} $$
converges to a positive real number $\tilde{L}$, in particular, we will see that $0<\tilde{L}<1$ (remember that $q\geq 3$).
Because $\left(1-\frac{1} {(q^{\nu}+1)^2} \right)^{\frac{q^{\nu}}{\nu}}<1$, we have that $\tilde{L}<1$. In order to prove that $0<\tilde{L}$, and since for $z\in(0,1)$ we have $\log(1-z)\geq\frac{z}{z-1}$, it is enough to prove that
$$ \sum_{\nu\geq 1}\frac{q^{\nu}}{\nu}\frac{\frac{1}{(q^{\nu}+1)^2}}{\frac{1}{(q^{\nu}+1)^2}-1}=-\sum_{\nu\geq 1}\frac{1}{\nu}\cdot\frac{1}{q^\nu+2} $$
is convergent. Indeed, we have
$$0\leq \sum_{\nu\geq 1}\frac{1}{\nu}\cdot\frac{1}{q^{\nu}+2}\leq\sum_{\nu\geq 1}\frac{1}{\nu 3^\nu}=\log\frac{3}{2}.$$
Thus,
$$\prod_{\nu\geq 1}\left(1-\frac{1}{(q^{\nu}+1)^2}\right)^{\frac{q^{\nu}}{\nu}}\geq\frac{2}{3}.$$
\end{remark}
\begin{proposition}\label{final} For $0\leq l\leq q$, let $x_1,...,x_l$ be distinct elements of $\mathbb{F}_q$, and $a_1,...,a_l,b_1,...,b_l$ be elements of $\mathbb{F}_{q}^{*}$. Then, for any $1>\epsilon > 0$, we have
$$
|\{(f,f_1,f_2)\in\mathcal{F}_{(n,n_1,n_2)}:\,f(x_i)f_1(x_i)=a_i,\,f(x_i)f_2(x_i)=b_i,\,1\leq i\leq l\}|=
$$
$$
=\frac{KLq^{n_1+n_2+n}}{\zeta_{q}^{3}(2)}\left( \frac{q}{(q-1)^2(q+3)}\right)^{l}(1+O(q^{-(1-\epsilon)n+\epsilon l}+q^{-n-\frac{\text{min}(n_1,n_2)}{2}+l})),
$$
where $K:=\prod_{P}\left(\frac{1+3|P|^{-1}}{(1+|P|^{-1})(1+2|P|^{-1})}\right)$.
\end{proposition}
\begin{proof} First we observe that
$$
|\{(f,f_1,f_2)\in\mathcal{F}_{(n,n_1,n_2)}:\,f(x_i)f_1(x_i)=a_i,\,f(x_i)f_2(x_i)=b_i,\,1\leq i\leq l\}|=
$$
$$=\sum_{ \begin{array}{c}\scriptscriptstyle f\in\mathcal{F}_n\\ \scriptscriptstyle f(x_i)\neq 0\end{array}}\sum_{\begin{array}{c}\scriptscriptstyle f_1\in\mathcal{F}_{n_1}
\\ \scriptscriptstyle f_1(x_i)=a_if(x_i)^{-1}\\ \scriptscriptstyle (f,f_1)=1\end{array}}\sum_{\begin{array}{c} \scriptscriptstyle f_2\in\mathcal{F}_{n_1}\\ \scriptscriptstyle f_2(x_i)=b_if(x_i)^{-1}\\\scriptscriptstyle (ff_1,f_2)=1\end{array}} 1=
$$
$$
=\sum_{\scriptscriptstyle f\in \mathcal{F}_n,\,f(x_i)\neq 0} R^{f}_{n_1,n_2}(l).
$$
Using Lemma \ref{guay} we have that
$$
|\{(f,f_1,f_2)\in\mathcal{F}_{n,n_1,n_2}:\,f(x_i)f_1(x_i)=a_i,\,f(x_i)f_2(x_i)=b_i,\,1\leq i\leq l\}|=
$$
$$
=\frac{q^{n_1+n_2}L}{\zeta_{q}^{2}(2)}\left(\frac{q}{(q-1)^{2}(q+2)}\right)^{l}\sum_{U\in \mathcal{F}_n,U(x_i)\neq 0}\prod_{P|U}\frac{1}{1+2|P|^{-1}}+O(q^{n_1+n_2-\frac{\text{min}(n_1,n_2)}{2}-l})=
$$
$$
=\frac{q^{n_1+n_2}L}{\zeta_{q}^{2}(2)}\left(\frac{q}{(q-1)^{2}(q+2)}\right)^{l}\sum_{\text{deg}(U)=n}c(U)+O(q^{n_1+n_2-\frac{\text{min}(n_1,n_2)}{2}-l}),
$$
where for any polynomial $U$, we define
$$
c(U)=\begin{cases}\mu^2(U)\prod_{P|U}\frac{1}{1+2|P|^{-1}}& U(x_i)\neq 0\\ 0&\text{otherwise}\end{cases}.
$$
In order to evaluate $\sum_{\text{deg}(U)=n}c(U)$, we consider the Dirichlet series
$$
G(w)=\sum_{U}\frac{c(U)}{|U|^w}=\prod_{P,\,P(x_i)\neq 0}(1+\frac{1}{|P|^{w}}\cdot \frac{|P|}{(|P|+2)})=
$$
$$
=\frac{\zeta_q(w)}{\zeta_q(2w)}H(w)(1+\frac{1}{q^{w-1}(q+2)})^{-l},
$$
with
$$
H(w)= \prod_{P}(1-\frac{2}{(1+|P|^w)(|P|+2)}).
$$
Notice that $H(w)$ converges absolutely for $\text{Re}(w)>0$, and $G(w)$ is meromorphic for $\text{Re}(w)>0$ with simple poles at the points $w$ where $\zeta_q(w)=(1-q^{1-w})^{-1}$ has poles, that is, $1+i\frac{2\pi n}{\text{log} q}$. Thus, $G(w)$ has a simple pole at $w=1$ with residue
$$
\frac{K}{\zeta_q(2)\text{log}(q)}\left( \frac{q+2}{q+3}\right)^{l},
$$
where $K=H(1)$.
Using Theorem $17.1$ of \cite{Rosen02}, which is the function field version of the Wiener-Ikehara Tauberian Theorem, we get that
$$
\sum_{\text{deg}(U)=n}c(U)= \frac{K}{\zeta_q(2)}\left( \frac{q+2}{q+3}\right)^{l}q^n+O_q(q^{\epsilon n}),
$$
for all $\epsilon\geq 0$ and where, looking at the proof of the theorem and proceding as in Proposition $4.3$ in \cite{Betal09}, we can exchange $O_q(q^{\epsilon n})$ by $O(q^{\epsilon (n+l)})$.
\end{proof}
\begin{cor}\label{corprefinal} For $0\leq l\leq q$, let $x_1,...,x_l$ be distinct elements of $\mathbb{F}_q$, and let $a_1,...,a_l,b_1,...,b_l$ be elements of $\mathbb{F}_q$ such that $a_1=...=a_{r_{0}}=b_1=...=b_{r_{0}}=0$, $a_{r_{0}+1}=...=a_{r_{0}+r_1}=0=b_{r_{0}+r_1+1}=...=b_{r_{0}+r_1+r_2}$ and $b_{r_{0}+1},...,b_{r_{0}+r_1},a_{r_{0}+r_1+1},...,a_{r_{0}+r_1+r_2},a_j,b_j\neq 0$ if $j>r_{0}+r_1+r_2=m$. Then, for every $\epsilon >0$, the number
$$
\frac{|\{(f,f_1,f_2)\in\mathcal{F}_{(n,n_1,n_2)}:\,f(x_i)f_1(x_i)=a_i,\,f(x_i)f_2(x_i)=b_i,\,f_1(x_i)f_2(x_i)=c_i,\,1\leq i\leq l\}|}{\mid\mathcal{F}_{(n,n_1n_2)}\mid},
$$
where $f(x_i)^2c_i=a_ib_i$, is equal to
$$
\left(\frac{1}{(q-1)(q+3)}\right)^{m}\left(\frac{q}{(q-1)^2(q+3)}\right)^{l-m}\left(1+ O(q^{-\frac{(1-\epsilon)}{2}\text{min}(n,n_1,n_2)+l})\right).
$$
\end{cor}
\begin{proof} Let us write $f=(x-x_1)...(x-x_{r_{0}})f'$, $f_1=(x-x_{r_{0}+1})...(x-x_{r_{0}+r_1})f'_1$, and $f_2=(x-x_{r_{0}+r_1+1})...(x-x_{r_{0}+r_1+r_2})f'_2$. Now, apply Proposition \ref{final} to the $3$-tupla $(f',f'_1,f'_2)$ and sum.
\end{proof}
\begin{cor}\label{corfinal} With notations as in Corollary \ref{corprefinal}, the number
$$
\frac{|\{(f,f_1,f_2)\in\mathcal{F}_{(n,n_1,n_2)}:\,\chi(f(x_i)f_1(x_i))=e^{1}_i,\,\chi(f(x_i)f_2(x_i))=e^{2}_i,\,\chi(f_1(x_i)f_2(x_i))=e_i,\,1\leq i\leq l\}|}{|\mathcal{F}_{(n,n_1,n_2)}|},
$$
where $e^{1}_i,e^{2}_i,e_i\in\{-1,\,0,\,1\}$, $\chi(f(x_i)^2)e_i=e_{i}^{1}e_{i}^{2}$, and exactly $2m$ of them are equal to zero, is equal to
$$
C_{m}^{l}=\left( \frac{q-1}{2}\right)^m
\left( \frac{q-1}{2}\right)^{2(l-m)}
\left(\frac{1}{(q-1)(q+3)}\right)^{m}
\left(\frac{q}{(q-1)^2(q+3)}\right)^{l-m}\left(1+ O(q^{-\frac{(1-\epsilon)}{2}\text{min}(n,n_1,n_2)+l})\right)=
$$
$$
=\left(\frac{1}{2(q+3)}\right)^{m}
\left( \frac{q}{4(q+3)}\right)^{l-m}\left(1+ O(q^{-\frac{(1-\epsilon)}{2}\text{min}(n,n_1,n_2)+l})\right).
$$
\end{cor}
\begin{cor}\label{superfinal} For $0\leq l\leq q$, let $x_1,...,x_l$ be distinct elements of $\mathbb{P}^1(\mathbb{F}_q)$, and let $e^{1}_i,e^{2}_i,e_i\in\{-1,\,0,\,1\}$ be such that $\chi(f(x_i)^2)e_i=e_{i}^{1}e_{i}^{2}$, where exactly $2m$ of them are equal to zero. Then
$$
\frac{|\{(f,f_1,f_2)\in\widehat{\mathcal{F}}_{[n,n_1,n_2]}:\,\chi(f(x_i)f_1(x_i))=e^{1}_i,\,\chi(f(x_i)f_2(x_i))=e^{2}_i,\,\chi(f_1(x_i)f_2(x_i))=e_i\}|}{|\widehat{\mathcal{F}}_{[n,n_1,n_2]}|}
$$
is also equal to the number $C^{l}_{m}$ defined in Corollary \ref{corfinal}.
\end{cor}
\begin{proof}
Distinguish the case in which some $x_j$ is the point at infinity or not. Generalize Corollary \ref{corfinal} for the sets $\widehat{\mathcal{F}}_{(n,n_1,n_2)}$ looking at the symmetry observed in Remark \ref{symmetry}, and add for the different components of $\widehat{\mathcal{F}}_{[n,n_1,n_2]}$.
\end{proof}
\begin{proof} (of Theorem \ref{main}) Apply Corollary \ref{superfinal} in order to compute
$$
\frac{|\{(f,f_1,f_2)\in\widehat{\mathcal{F}}_{[n,n_1,n_2]}:\widehat{S}(f,f_1,f_2)=M|}{|\widehat{\mathcal{F}}_{[n,n_1,n_2]}|}=
$$
$$
=\sum_{\scriptscriptstyle \epsilon_1,...,\epsilon_{q+1}\in \{-1,1,3\}, \sum \epsilon_i =M} \sum_{j=0}^{N_{-1}}\binom{N_{-1}}{j}3^{N_1+N_{-1}} C^{q+1}_{N_1+j}=
$$
$$
=\sum_{\scriptscriptstyle \epsilon_1,...,\epsilon_{q+1}\in \{-1,1,3\},\,\sum \epsilon_i =M} \left(\frac{6}{4}\frac{1}{q+3}\right)^{N_{1}}\left(\frac{3}{4}\frac{q+2}{q+3}\right)^{N_{-1}}\left(\frac{1}{4}\frac{q}{q+3}\right)^{N_{3}}\left(1+ O(q^{-\frac{(1-\epsilon)}{2}\text{min}(n,n_1,n_2)+q})\right)
$$
$$
=\text{Prob}\left(\sum_{1}^{q+1}X_i=M\right)\left(1+ O(q^{-\frac{(1-\epsilon)}{2}\text{min}(n,n_1,n_2)+q})\right),
$$
where we use the notation $N_i$ for the number of elements equal to $i$ in the set $\{\epsilon_1,...,\epsilon_{q+1}\}$.
\end{proof}
\section{Averages and moments sequences}
We want to compute the moments of $\text{Tr}(\text{Frob}_C)/\sqrt{1+q}$. That is, the \textit{k}th moments
$$
M_k(q,g_1,g_2,g_3)=\frac{1}{|\mathcal{B}_{(g_1,g_2,g_3)}(\mathbb{F}_{q})|\mbox{'}}\sum_{\scriptscriptstyle C\in \mathcal{B}_{(g_1,g_2,g_3)}(\mathbb{F}_{q})}\hspace{-1cm}\mbox{'}\hspace{0.8cm}\left(\frac{\text{Tr}(\text{Frob}_C)}{\sqrt{1+q}}\right)^k.
$$
\begin{theorem}\label{moments} With notation in Theorem \ref{main}, we have
$$
M_k(q,g_1,g_2,g_3)=\mathbb{E}\left( \left(\frac{1}{\sqrt{1+q}}\sum_{i=1}^{1+q}X_i\right)^k\right)+O(q^{-\frac{1-\epsilon}{2}min(n,n_{1},n_{2})+k}).
$$
\end{theorem}
\begin{cor} If $g_1,g_2,g_3$ and $q$ tend to infinity, then the moments of $\mathrm{Tr}(\mathrm{Frob}_C)/\sqrt{1+q}$, as $C$ runs over the irreducible component $\mathcal{B}_{(g_1,g_2,g_3)}(\mathbb{F}_{q})$ of the moduli space $\mathcal{B}_{g}(\mathbb{F}_{q})$, are asymptotically Gaussian with mean $0$ and variance $1$.
\end{cor}
\begin{proof} Since the moments of a sum of bounded i.i.d. random variables converge to the Gaussian moments (\cite[Sec. 30]{Bill95}), it follows that, as all $q,g_1,g_2,g_3$ go to $\infty$, $M_k(q,g_1,g_2,g_3)$ agrees with Gaussian moments for all $k$. Then, Theorem $30.2$ in \cite{Bill95} implies the corollary.
\end{proof}
\begin{proof} (of Theorem \ref{moments}) We can write the \textit{k}th moment as
$$
M_k(q,g_1,g_2,g_3)=(-1)^k\frac{q(q^2-1)}{|\widehat{\mathcal{F}}_{[2g_1+2,2g_2+2,2g_3+2]}|}\sum_{\scriptscriptstyle (f_1,f_2,f_3)\in\widehat{\mathcal{F}}_{[2g_1+2,2g_2+2,2g_3+2]}}
(\widehat{S}(f_1,f_2,f_3))^k=
$$
$$
=\frac{(-1)^k q(q^2-1)}{|\widehat{\mathcal{F}}_{[2g_1+2,2g_2+2,2g_3+2]}|}\sum_{\scriptscriptstyle (f_1,f_2,f_3)\in\widehat{\mathcal{F}}_{[2g_1+2,2g_2+2,2g_3+2]}}
\sum_{x\in\mathbb{P}^1(\mathbb{F}_q)}
(\chi(f\cdot f_1(x))+\chi(f\cdot f_2(x))+\chi(f_1\cdot f_2(x)))^k=
$$
$$
=(-1)^k q(q^2-1)\sum_{l=1}^{k}c(k,l)\sum_{(x,b)\in P_{k,l}}\frac{1}{|\widehat{\mathcal{F}}_{[2g_1+2,2g_2+2,2g_3+2]}|}
\sum_{\scriptscriptstyle (f_1,f_2,f_3)\in\widehat{\mathcal{F}}_{[2g_1+2,2g_2+2,2g_3+2]}} B(x,b,f_1,f_2,f_3),
$$
where, borrowing the notation in \cite[Sec. 5]{Betal09}
$$
P_{k,l}=\left\{(x,b):x=(x_1,...,x_l)\in\mathbb{P}^1(\mathbb{F}_q)^l,\,x_i's\,\text{distinct,}\,b=(b_1,...,b_l)\in\mathbb{Z}^{l}_{>0},\,\sum_{i=1}^{l}b_i=k\right\},
$$
$$
B(x,b,f_1,f_2,f_3)=\prod_{i=1}^{l}(\chi(f\cdot f_1(x_i))+\chi(f\cdot f_2(x_i))+\chi(f_1\cdot f_2(x_i)))^{b_i}
$$
and $c(k,l)$ is a certain combinatorial factor. We do not need exact formulas for it, but as it was notice in \cite{Betal09}
\begin{equation}\label{cota}
\sum_{l=1}^{k}c(k,l)\sum_{(x,b)\in P_{k,l}}1=(q+1)^k.
\end{equation}
Fix a vector $(x,b)\in P_{(k,l)}$. Then, the number
$$
\sum_{\scriptscriptstyle (f_1,f_2,f_3)\in\widehat{\mathcal{F}}_{[2g_1+2,2g_2+2,2g_3+2]}} \frac{B(x,b,f_1,f_2,f_3)}{|\widehat{\mathcal{F}}_{[2g_1+2,2g_2+2,2g_3+2]}|}=\sum_{\begin{array}{c}\scriptstyle(\epsilon_1,...\epsilon_l)\\ \scriptstyle\epsilon_i\in\{-1,1,3\}\end{array}}\left(\prod P_{\epsilon_i}\right)\left(\prod \epsilon_{i}^{b_i}\right)=
$$
$$
\sum_{\begin{array}{c}\scriptstyle(\epsilon_1,...\epsilon_l)\\ \scriptstyle\epsilon_i\in\{-1,1,3\}\end{array}}\left(\prod P_{\epsilon_i} \epsilon_{i}^{b_i}\right)=\prod_{i}\left(\frac{3^{b_i}q+6+(-1)^{b_i}3(q+2)}{4(q+3)}\right)
(1+O(q^{-\frac{1-\epsilon}{2}min(n,n_{1},n_{2})+l}))
$$
We obtain that
$$
M_k(q,g_1,g_2,g_3)=(-1)^k q(q^2-1)\sum_{l=1}^{k}c(k,l)\sum_{(x,b)\in P_{k,l}}\prod_{i}\left(\frac{3^{b_i}q+6+(-1)^{b_i}3(q+2)}{4(q+3)}\right)
(1+O(q^{-\frac{1-\epsilon}{2}min(n,n_{1},n_{2})+k})).
$$
where the error term is estimated using \ref{cota}.
On the other hand, the corresponding moment of the normalized sum of our random variables is
$$
\mathbb{E}\left( \left(\frac{1}{\sqrt{1+q}}\sum_{i=1}^{1+q}X_i\right)^k\right)=\frac{1}{(1+q)^{k/2}}\sum_{l=1}^k
\sum_{(i,b)\in A_{k,l}}\mathbb{E}(X_{i_1}^{b_1}...X_{i_l}^{b_l}),$$
where
$$
A_{k,l}=\left\{(i,b):i=(i_1,...,i_l),1\leq i_j\leq q+1,\,i_{j}'s\,\text{distinct,}\,b=(b_1,...,b_l)\in\mathbb{Z}^{l}_{>0},\,\sum_{i=1}^{l}b_i=k\right\}
$$
is clearly isomorphic to $P_{k,l}$.
Since
$$
\mathbb{E}(X_{i}^{b})=\frac{3^bq+6+(-1)^b3(q+2)}{4(q+3)}
$$
and $X_1,...,X_{1+q}$ are independent, we get the equality in the statement of the theorem.
\end{proof}
\section{General case: the family of $r$-quadratic curves}
\begin{defn}
Let $r\geq 1$ be an integer. We call $r$-quadratic curve a non-singular projective curve $C/\mathbb{F}_q$ together with a morpshim $\varphi: C\longrightarrow \mathbb{P}^1_{\mathbb{F}_q}$ defined over $\mathbb{F}_{q}$ such that it induces a function field extension with Galois group
$$
\mathrm{Gal}(\mathbb{F}_q(C)/\mathbb{F}_q(t))\simeq (\mathbb{Z}/2\mathbb{Z})^{r}
$$
Note that when $r=1$ and $r=2$ we find respectively the definition of hyperelliptic curve and that of biquadratic curve given in Section \ref{sec:family}.
\end{defn}
The family of $r$-quadratic curves is studied in \cite{Pries2005b} where it is proved that when we consider the family of this curve defined over $\bar{\mathbb{F}}_{q}$ we obtain a course moduli space over $\mathbb{Z}[1/2]$. In the same paper a formula for the genus of an $r$-quadratic curve is also proved.
\begin{proposition}
Let $r\geq 1$ be an integer and let $C/\mathbb{F}_{q}$ be an $r$-quadratic curve.
\begin{enumerate}
\item
An affine model of $C$ in $\mathbb{A}^{r+1}$ is given by
$$
C:\begin{cases}y_{1}^{2}=h_1(t)\\...\\y_{r}^{2}=h_r(t)\end{cases}
$$
where each $h_{i}$ is square-free and different (up to squares) from $\prod_{j\in J}h_{j}$, for every non-empty subset $J\subseteq\{1,\dots,r\}, J\neq \{i\}$.
\item
For every non-empty $J\subseteq\{1,\dots,r\}$, the affine equation
$$
y^{2}=\prod_{j\in J}h_{j}(t)
$$
defines a quadratic subextension of $\mathbb{F}_q(C)/k$ and every quadratic subextension of $\mathbb{F}_q(C)/k$ is obtained in this way, so there are $2^r-1$ of them.
\item
If we write $C_J$ for the hyperelliptic curve given by the affine equation $y^2=\prod_{J\in J}h_{j}(t)$, then
$$
g(C)=\sum_{J\subseteq\{1,\dots,r\}}g(C_{J}).
$$
\end{enumerate}
\end{proposition}
Next Theorem gives a convenient way of describing the family of $r$-quadratic curves, and it is the key point to compute the distribution of the Frobenius traces.
\begin{theorem}\label{family}
There exists a one-to-one correspondence between the
set of $r$-quadratic extensions of $k$ and the set of unordered $2^{r}-1$-tuples of square-free and pairwise coprime polynomials.
\end{theorem}
\begin{proof}
Let $K=k(\sqrt{h_{1}},\dots,\sqrt{h_{r}})$ be an $r$-quadratic extension. We associate to such an extension, a $2^{r}-1$-tuple $(f_{1},\dots,f_{2^{r}-1})$ of square-free and pairwise coprime polynomials in the following way: for every $i\in\{1,...,2^r-1\}$, we write $B^{i}_{1}\dots B^{i}_{r}$ for the representation of the integer $i$ in base $2$ (so $B^{i}_{j}\in\{1,0\}$ for every $1\leq j\leq r$) and we define $m_{i}$ to be the greatest common divisor of all polynomials $h_{j}$ such that the $B_{j}=1$.
We then define the polynomials $f_{i}$ as the maximum factor in the decomposition of $m_{i}$ which is coprime to all the $h_{j}$ such that $B^{i}_{j}=0$. Notice that, in particular, $f_{2^{r}-1}=(h_{1},\dots,h_{r})$.
Viceversa, given a tuple $(f_{1},\dots,f_{2^{r}-1})$ of square-free and pairwise coprime polynomials, we define the $r$-quadratic extension $k(\sqrt{h_{1}},\dots,\sqrt{h_{r}})$, where $h_{i}$ is defined to be the product of the $f_{j}$ such that the $i$-th digit of $j$ in base $2$ is $1$, i.e. $B_{i}^{j}=1$.
\end{proof}
Notice that, with notations of Theorem \ref{family}, we have that $f_{2^{r}-1}=(h_{1},\dots,h_{r})$.
After Theorem \ref{family}, we are led to define the following sets:
$$
\mathcal{F}_{(n_1,\dots,n_{2^r-1})}:=\{(f_1,\dots,f_{2^r-1})\in\mathcal{F}_{n_1}\times\dots\times\mathcal{F}_{n_{2^r-1}}:\,(f_{i},f_{j})=1,\;\;i,j=1,\dots,2^{r}-1, i\neq j\},
$$
$$
\widehat{\mathcal{F}}_{(n_1,\dots,n_{2^r-1})}:=\{(f_1,\dots,f_{2^r-1})\in\widehat{\mathcal{F}}_{n_1}\times\dots\times\widehat{\mathcal{F}}_{n_{2^r-2}}\times\mathcal{F}_{n_{2^r-1}}:\,(f_{i},f_{j})=1,\;\;i,j=1,\dots,2^{r}-1,i\neq j\},
$$
It is easy to prove that if $C$ is an $r$-quadratic curve whose affine model is given by equations $y_{i}^{2}=h_{i}(t),\;i=1,\dots,r$, then
$$
\#C(\mathbb{F}_{q})=\sum_{x\in\mathbb{P}^{1}_{\mathbb{F}_q}}\prod_{i=1}^{r}(1+\chi(h_{i}(x))).
$$
Now, we express this formula in terms of the polynomials $f_{1},\dots,f_{2^r-1}$ defined in the proof of Theorem \ref{family}.
Let us fix $(f_{1},\dots,f_{2^r-1})\in\widehat{\mathcal{F}}_{(n_1,\dots,n_{2^r-1})}$. For every $i\in\{1,\dots,2^r-1\}$, we define the polynomial $p_{i}$ as the square-free part of the product of the polynomials $f_{j}$ such that the relation between the representations in base $2$ of $i$ and $j$ is the following: $B_{k}^{i}=1\Rightarrow B_{k}^{j}=1$. It is immediate to see that, inside the correspondence of Theorem \ref{family}, the square-free polynomials $p_{1},\dots,p_{2^r-1}$ define all the quadratic subextensions of the $r$-quadratic extension defined by the tuple $(f_{1},\dots,f_{2^r-1})$.
Then we define
$$
\widehat{S}(f_{1},\dots,f_{2^r-1}):=\sum_{x\in\mathbb{P}^{1}}\sum_{i=1}^{2^r-1}\chi(p_{i}(x))
$$
and we can rewrite
$$
\#C(\mathbb{F}_{q})=\sum_{x\in\mathbb{P}^{1}}\prod_{i=1}^{2^r-1}(1+\chi(p_{i}(x)))=q+1+\widehat{S}(f_{1},\dots,f_{2^r-1}).
$$
When $r=2$ we find the formula of Section \ref{distributionFq}.
\begin{lemma}\label{guay_gen}
Let $n_{1},\dots,n_{\beta}$ be positive integers. For $0\leq l\leq q$, let $x_1,...,x_l$ be distinct elements of $\mathbb{F}_q$. Let $U\in\mathbb{F}_q[X]$ be such that $U(x_i)\neq 0$ for $i=0,...,l$. Let be $a^{1}_1,...,a^{1}_l,\dots,a^{\beta}_1,...,a^{\beta}_l\in\mathbb{F}_{q}^{*}$. The number of elements in the set
\begin{multline*}
\mathcal{R}^{U}_{n_{1},\dots,n_{\beta}}((a_{1}^{j},\dots,a_{l}^{j})_{1\leq j\leq\beta}):=\{(f_1,\dots,f_{\beta})\in\mathcal{F}_{n_1}\times\dots\times\mathcal{F}_{n_{\beta}}:\,\\(f_j,U)=1,\,(f_{j},f_{k})_{k\neq j}=1,\,f_j(x_i)=a^{j}_i,1\leq i\leq l,1\leq j,k\leq\beta\}
\end{multline*}
is the number
$$
R^{U}_{n_{1},\dots,n_{\beta}}(l)=\frac{q^{n_{1}+\dots+n_{\beta}}L_{\beta}}{\zeta_{q}^{\beta}(2)}\left(\frac{q}{(q-1)^{\beta}(q+\beta)}\right)^{l}\prod_{P\mid U}\left(\frac{1}{1+\beta|P|^{-1}}\right)(1+O(q^{l-\frac{\mathrm{min}(n_1,\dots,n_{\beta})}{2}})),
$$
where the constant
$$
L_{\beta}:=
\prod_{P\,\text{prime}}\left(\frac{|P|^{\beta-1}(|P|+\beta)}{(|P|+1)^{\beta}}\right).
$$
In a similar way to Remark \ref{non_vanishing_L}, we can see that $L_{\beta}$ is bounded.
\end{lemma}
\begin{proof}
We will prove it by induction on $\beta$. We find Lemma \ref{S} for $\beta=1$, and for $\beta=2$ we find Lemma \ref{guay}. Assume that the equality of the statement is true for $\beta-1$.
By inclusion-exclusion principle, with
\begin{equation*}
f(D)=|\{(f_1,...,f_\beta)\in \mathcal{R}^{U}_{n_{1},\dots,n_{\beta-1}}((a_{1}^{j},\dots,a_{l}^{j})_{1\leq j\leq\beta-1})\times \mathcal{S}_{n_\beta}^{U}(a_{1}^{\beta},\dots,a_{l}^{\beta}):
\,D|(f_1...f_{\beta-1},f_\beta)\}|,
\end{equation*}
\begin{equation*}
g(D)=|\{(f_1,...,f_\beta)\in \mathcal{R}^{U}_{n_{1},\dots,n_{\beta-1}}((a_{1}^{j},\dots,a_{l}^{j})_{1\leq j\leq\beta-1})\times \mathcal{S}_{n_\beta}^{U}(a_{1}^{\beta},\dots,a_{l}^{\beta}):\,D=(f_1...f_{\beta-1},f_\beta)\}|,
\end{equation*}
we have
$$
R^{U}_{n_{1},\dots,n_{\beta}}(l)=g(1)=\sum_{\scriptscriptstyle D,\,D(x_i)\neq 0,(D,U)=1}\mu (D) f(D).
$$
But notice that when $(D,U)=1$ and $D$ is square-free
$$
f(D)=\prod_{P|D}(\beta-1) \cdot R^{UD}_{n_{1},\dots,n_{\beta-1}-\text{deg}(D)}(l)\cdot S^{UD}_{n_{\beta}-\text{deg}(D)}(l)
$$
hence, by induction hypothesis,
\begin{multline*}
f(D)=\frac{q^{n_{1}+\dots+n_{\beta}}L_{\beta-1}}{\zeta_{q}^{\beta}(2)}\left(\frac{q^{2}}{(q^{2}-1)(q-1)^{\beta-1}(q+\beta-1)}\right)^l\\
\prod_{P|U}\frac{1}{(1+(\beta-1)|P|^{-1})(1+|P|^{-1})}\prod_{P|D}\frac{(\beta-1)|P|^{-2}}{(1+(\beta-1)|P|^{-1})(1+|P|^{-1})}\left(1+O(q^{l+\frac{\text{deg}(D)}{2}-\frac{\text{min}(n_1,...,n_\beta)}{2}})\right).
\end{multline*}
So, one has
\begin{multline*}
R^{U}_{n_{1},\dots,n_{\beta}}(l)=\sum_{\scriptscriptstyle D,\,D(x_i)\neq 0,(D,U)=1}\mu (D) f(D)=\\
=\frac{q^{n_{1}+\dots+n_{\beta}}L_{\beta-1}}{\zeta_{q}^{\beta}(2)}\left(\frac{q^{2}}{(q^{2}-1)(q-1)^{\beta-1}(q+\beta-1)}\right)^l
\prod_{P|U}\frac{1}{(1+(\beta-1)|P|^{-1})(1+|P|^{-1})}\cdot\\
\sum_{\begin{array}{c}\scriptscriptstyle D(x_i)\neq 0,\,(D,U)=1\\\scriptscriptstyle \text{deg}(D)\leq \text{min}(n_1,...,n_\beta)\end{array}}\mu(D)\prod_{P|D}\frac{(\beta-1)|P|^{-2}}{(1+(\beta-1)|P|^{-1})(1+|P|^{-1})}\left(1+O(q^{l-\frac{\text{min}(n_1,...,n_\beta)}{2}})\right).
\end{multline*}
Now, we observe that
\begin{multline*}
\sum_{\begin{array}{c}\scriptscriptstyle D\, D(x_i)\neq 0,\,(D,U)=1\\ \scriptscriptstyle\text{deg}(D)\leq \text{min}(n_1,...,n_\beta)\end{array}}\mu(D)\prod_{P|D}\frac{(\beta-1)|P|^{-2}}{(1+(\beta-1)|P|^{-1})(1+|P|^{-1})}=\\
\sum_{\scriptscriptstyle D,\,D(x_i)\neq 0,\,(D,U)=1}\mu(D)\prod_{P|D}\frac{(\beta-1)|P|^{-2}}{(1+(\beta-1)|P|^{-1})(1+|P|^{-1})}+O(q^{-2\text{min}(n_1,...,n_\beta)}),
\end{multline*}
where we have that
$$
\begin{array}{l}
\sum\limits_{\scriptscriptstyle D,\,D(x_i)\neq 0,\,(D,U)=1}\mu(D)\prod\limits_{P|D}\frac{(\beta-1)|P|^{-2}}{(1+(\beta-1)|P|^{-1})(1+|P|^{-1})}=
\\
\,
\\
=\left( 1-\frac{(\beta-1)q^{-2}}{(1+(\beta-1)q^{-1})(1+q^{-1})}\right)^{-l}
\prod\limits_{P|U}\left( 1-\frac{(\beta-1)|P|^{-2}}{(1+(\beta-1)|P|^{-1})(1+|P|^{-1})}\right)^{-1}\prod\limits_{P\,\text{prime}}\left(1-\frac{(\beta-1)|P|^{-2}}{(1+(\beta-1)|P|^{-1})(1+|P|^{-1})}\right)=
\\
\,
\\
=\left( \frac{(q+\beta-1)(q+1)}{q(q+\beta)}\right)^{l}\prod\limits_{P|U}\left( \frac{\beta|P|^{-1}+1}{(1+(\beta-1)|P|^{-1})(1+|P|^{-1})}\right)^{-1}\prod\limits_{P\,\text{prime}} \frac{\beta|P|^{-1}+1}{(1+(\beta-1)|P|^{-1})(1+|P|^{-1})}.
\end{array}
$$
So, the result follows.
\end{proof}
\begin{proposition}\label{final_gen} For $0\leq l\leq q$, let $x_1,...,x_l$ be distinct elements of $\mathbb{F}_q$, and let be $a^{1}_1,...,a^{1}_l,\dots,a^{r}_1,...,a^{r}_l\in\mathbb{F}_{q}^{*}$. Then for any $1>\epsilon > 0$, we have
$$
|\{(f_1,\dots,f_{2^r-1})\in\mathcal{F}_{(n_1,\dots,n_{2^r-1})}:\,h_{j}(x_{i})=a^{j}_{i},\;1\leq i\leq l,\,1\leq j\leq r \}|=
$$
$$
=\frac{q^{n_{1}+\dots+n_{2^{r}-1}}L_{2^{r}-2}}{\zeta_{q}^{2^{r}-1}(2)}\left(\frac{q}{(q-1)^{r}(q+2^{r}-1)}\right)^{l}(1+O(q^{l-\frac{\mathrm{min}(n_1,\dots,n_{\beta})}{2}})).
$$
\end{proposition}
\begin{proof} Apply previous Lemma with $\beta=2^r-1$ and $U(x)=1$. Notice that the value of the polynomials $h_i$ can be fixed by controling the value of $r$ of the $f_j$ polynomials, so we need to multiply the previous number by $(q-1)^{l(2^r-1-r)}$.
\end{proof}
\begin{cor}\label{corofinal_gen}For $0\leq l\leq q$, let $x_1,...,x_l$ be distinct elements of $\mathbb{F}_q$, and let $a^{1}_1,...,a^{1}_l,\dots,a^{r}_1,...,a^{r}_l$ be elements in
$\mathbb{F}_{q}$ such that for exactly $m$ values of $i$ we have $\prod_{i=1}^{r}a_{i}^{j}=0$. Then, for any $1>\epsilon > 0$, we have
$$
|\{(f_1,\dots,f_{2^r-1})\in\mathcal{F}_{(n_1,\dots,n_{2^r-1})}:\,h_{j}(x_{i})=a^{j}_{i},\;1\leq i\leq l,\,1\leq j\leq r \}|=
$$
$$
=\frac{q^{n_{1}+\dots+n_{2^{r}-1}}L_{2^{r}-2}}{\zeta_{q}^{2^{r}-1}(2)}\left(\frac{q}{(q-1)^{r}(q+2^{r}-1)}\right)^{l-m}\left(\frac{1}{(q-1)^{r-1}(q+2^{r}-1)}\right)^{m}
(1+O(q^{l-\frac{\mathrm{min}(n_1,\dots,n_{\beta})}{2}}).
$$
\end{cor}
\begin{proof} Acording to the values of $a^{j}_{i}$ we can decide which of the polynomials $f_k$ should satisfy $f_j(x_i)=0$, see Theorem \ref{family}. Write $f_k(x)=(x-x_i)f'_k(x)$ with $f'_k(x_i)\neq 0$ and apply previous Proposition. Hence for each value of $i$ such that $\prod_{i=1}^{r}a_{i}^{j}=0$, we should multiply the number in Proposition \ref{final_gen} by $(q-1)/q$.
\end{proof}
Finally, previous corollary with $l=q$ together with the fact that there are $\frac{q-1}{2}$ squares in $\mathbb{F}_q$, implies the following generalization of Theorem \ref{main}:
\begin{thm}\label{main_gen}
If the degrees $n_{1},\dots,n_{2^{r}-1}$ go to infinity
$$
\frac{|\{(f_{1},\dots,f_{2^{r}-1})\in\widehat{\mathcal{F}}_{(n_1,\dots,n_{2^r-1})}:\,\widehat{S}(f_{1},\dots,f_{2^r-1})=M\}|}{|\widehat{\mathcal{F}}_{(n_1,\dots,n_{2^r-1})}|}=
\mathrm{Prob}\left(\sum_{j=1}^{q+1}X_{j}=M\right)
$$
where the $X_{j}$ are i.i.d. random variables such that
$$
X_{i}=\left\{\begin{array}{ccc} -1 & \rm{with\,probability}& \frac{(2^{r}-1)(q+2^{r}-2)}{2^{r}(q+2^{r}-1)}\\
\,&\,&\,\\
2^{r-1}-1 & \rm{with\,probability}& \frac{2(2^{r}-1)}{2^{r}(q+2^{r}-1)}\\
\,&\,&\,\\
2^{r}-1 & \rm{with\,probability}& \frac{q}{2^{r}(q+2^{r}-1)}
\end{array}\right..$$
\end{thm}
Observe that Theorem \ref{main_gen} specializes to Theorem $1.1$ of \cite{Betal09} when $r=1$ and to Theorem \ref{main} of the present paper when $r=2$.
\newpage
\section*{Appendix}
\begin{center}by Alina Bucur.
\end{center}
\bigskip\noindent {$1.$ \bf Biquadratic covers.}
Fix a finite field ${\mathbb F}_q$ of characteristic different from $2.$
A biquadratic cover of ${\mathbb P}^1$ over ${\mathbb F}_q$ is a covering map $\pi:C \to {\mathbb P}^1$ such that $\Aut(C/{\mathbb P}^1) \simeq {\mathbb Z}/2{\mathbb Z} \times {\mathbb Z}/2{\mathbb Z}.$ Such a cover has an affine model (here we take ${\mathbb P}^1$ to be marked) given by equations
\[\begin{cases}
y_1^2 =h_1(t) & \\
y_2^2 = h_2(t) &
\end{cases}\]
with $h_1, h_2 \in {\mathbb F}_q[t]$ both square-free. This corresponds to the field extension $k\left(\sqrt{ h_1}, \sqrt{ h_2}\right)$ of the function field $k = {\mathbb F}_q(t)$ of ${\mathbb P}_1.$
Secretly, it also has an implied equation \begin{equation}\label{eq:secretr=2}w^2 = h_1(t) h_2(t),\end{equation} but of course the right hand side is not necessarily square-free anymore and the gcd of $ h_1$ and $h_2$ will appear squared. If we denote $f_3 = gcd( h_1,h_2),$ then we can rewrite our three equations as
\[\begin{cases}
y_1^2 =f_1(t) f_3(t) & \\
y_2^2 = f_2(t) f_3(t) &\\
w^2 = f_1(t) f_2(t)&
\end{cases}\]
with $f_1, f_2, f_3$ all square-free and pairwise coprime.
This corresponds to the field extension $K(\sqrt{f_1f_3}, \sqrt{f_2f_3}, \sqrt{f_1f_2}).$ Note that it is better to think of this as a ``tri-quadratic'' extension, as the roles of $f_1f_3, f_2f_3, f_1f_2$ can be permuted. This also shows that each cover will appear exactly $\#S_3 = 6$ times, which is expected as the automorphism group of the Klein group is indeed $S_3.$
Let us look at geometric points over a point $\alpha \in {\mathbb P}^1$, i.e. we treat our field of definition as algebraically closed.
First, looking at \begin{multline*}(f_1,f_2, f_3)\bmod (t-\alpha)^2 =\\ \left(f_1(\alpha) + f_1'(\alpha)(t-\alpha), h_2(\alpha) + f_2'(\alpha)(t-\alpha), f_3(\alpha) + f_3'(\alpha)(t-\alpha) \right) \bmod (t-\alpha)^2\end{multline*} we see that we have:
\begin{itemize}
\item $(q^2-1)^3$ choices as $f_1,f_2,f_3$ have to be nonzero $\bmod (t-\alpha)^2.$
\item from those we need to exclude the possibility that $(t-\alpha)$ divides any two of $f_1,f_2,f_3$ (as they have to be coprime), i.e. we cannot allow $f_1(\alpha), f_2(\alpha), f_3(\alpha)$ to contain two or three zeros. Let us examine each situation we need to avoid.
\begin{itemize}
\item[\bf Two zeros:] The situation is completely symmetric in $f_1,f_2,f_3$ so it is enough to count one possibility and multiply by $3.$ If $f_1(\alpha) = f_2(\alpha) = 0,$ but $f_3(\alpha) \neq 0,$ then we must have $f_1'(\alpha) \neq 0$ (so $q-1$ choices), $f_2'(\alpha) \neq 0$ ($q-1$ choices) and no restrictions on $f_3'(\alpha)$ ($q$ choices). Taking into account the $q-1$ possibilities for $f_3(\alpha)$, we have to subtract $3 (q-1)^3 q.$
\item[\bf Three zeros:] In this case we must have $f_1'(\alpha), f_2'(\alpha), f_3'(\alpha)$ all nonzero, so there are $(q-1)^3$ such triples.
\end{itemize}
\end{itemize}
Thus we start with \[(q^2-1)^3 - 3q(q-1)^3 - (q-1)^3 = q^2(q-1)^3(q+3)\] triples modulo $(t-\alpha)^2.$
Geometrically, there are two possibilities for the fiber over $\alpha.$
\begin{enumerate}
\item[{\bf I.}] There will be $4$ distinct points when $f_1(\alpha) f_3(\alpha) \neq 0$ and $f_2(\alpha) f_3(\alpha) \neq 0,$ i.e. when $f_1(\alpha), f_2 (\alpha), f_3(\alpha)$ are all nonzero ($(q-1)$ choices each) and $f_1' (\alpha), f_2' (\alpha), f_3'(\alpha)$ have no restrictions ($q$ choices each). Thus there are $q^3(q-1)^3$ possibilities.\\
\item[{\bf II.}] There $2$ distinct points otherwise when either $f_1 (\alpha) f_3(\alpha) =0$ or $f_2(\alpha) f_3(\alpha) =0.$ Since no two of the terms can be zero at the same time, this means that exactly one of $f_1(\alpha), f_2(\alpha), f_3(\alpha)$ is zero. If $f_1(\alpha) =0,$ then $f_1'(\alpha) \neq 0.$ As the situation is again completely symmetric in $f_1, f_2, f_3,$ we have $3q^2(q-1)^3$ such triples.
\end{enumerate}
To count ${\mathbb F}_q$-rational points, we must split each of the two cases above into two further cases.
\begin{enumerate}
\item[{\bf Ia.}] The fiber has $4$ ${\mathbb F}_q$-rational points when $f_1(\alpha)f_3(\alpha)$ and $f_2(\alpha)f_3(\alpha)$ are both quadratic residues (and non-zero, as we are in the first case above). Thus $f_1(\alpha), f_2(\alpha), f_3(\alpha)$ have to be all three either quadratic residues or quadratic non residues. In either case they are all nonzero, and thus $f_1' (\alpha), f_2' (\alpha), f_3'(\alpha)$ have no restrictions imposed on them. Therefore we have
\[ 2 \left(\frac{q-1}{2} \right)^3 q^3 = \frac{q^3(q-1)^3}{4}\] choices, i.e. probability $1/4$ to get this subcase out of case I.
\item[{\bf Ib.}] There are no rational points (but $2$ points of degree $2$) in the fiber when exactly one of $f_1(\alpha), f_2(\alpha), f_3(\alpha)$ is a quadratic residue or exactly one of $f_1(\alpha), f_2(\alpha), f_3(\alpha)$ is a quadratic nonresidue. This situation occurs with probability $3/4$ out of case I. (Note that in this case at least one of the two defining equations has no solution, therefore no rational point.)
\item[{\bf IIa.}] The fiber consists of $2$ ${\mathbb F}_q$-rational points when one of the following three cases occur.
\begin{itemize}
\item $f_1(\alpha)f_3(\alpha) = 0$ and $f_2(\alpha)f_3(\alpha)$ is a nonzero quadratic residue. Thus we need to have $f_1(\alpha)=0,$ $f_1'(\alpha) \neq 0$ and we get \[2 \left( \frac{q-1}{2}\right)^2 (q-1) q^2 = \frac{q^2 (q-1)^3}{2} \textrm{ possibilities.}\]
\item $f_2(\alpha)f_3(\alpha) = 0$ and $f_1(\alpha)f_3(\alpha)$ is a nonzero quadratic residue. As above, there are $\frac{q^2 (q-1)^3}{2}$ such triples.
\item $f_3(\alpha) =0$ and $f_1(\alpha) f_2(\alpha)$ is a nonzero quadratic residue. Similarly, there are $\frac{q^2 (q-1)^3}{2}$ such triples.
\end{itemize}
In conclusion, the probability of getting this subcase out of case II is $1/2.$
\item[{\bf IIb.}] The fiber contains no ${\mathbb F}_q$-rational points (it is one double point of degree $2$) in one of the following three cases.
\begin{itemize}
\item $f_1 (\alpha) f_3(\alpha)= 0$ and $f_2(\alpha)f_3(\alpha)$ is a quadratic nonresidue. There are $\frac{q^2 (q-1)^3}{2}$ such triples.
\item $f_1 (\alpha)f_3(\alpha)$ is a quadratic nonresidue and $f_2(\alpha)h(\alpha) = 0$. As before, there are $\frac{q^2 (q-1)^3}{2}$ such triples.
\item $f_3(\alpha)=0$ and $f_1(\alpha) f_2(\alpha)$ is a quadratic nonresidue. There are also $\frac{q^2 (q-1)^3}{2}$ such triples.
\end{itemize}
In conclusion, the probability of getting this subcase out of case II is also $1/2.$
\end{enumerate}
The upshot is the following ``prediction''.
\bigskip\noindent {\bf Conjecture $1$.}\, {\em
\[\mathrm{Prob}\left( \#C({\mathbb F}_q) = M; \textrm{$C$ biquadratic cover of ${\mathbb P}^1$ defined over ${\mathbb F}_q$}\right) \sim \mathrm{Prob} (X_1+ \dots+X_{q+1} = M)\]
where $X_i$'s are i.i.d. random variables taking values
\[ X_i = \begin{cases}
4 & \textrm{ with probability } \displaystyle \frac{1}{4} \frac{q^3(q-1)^3}{q^2(q-1)^3(q+3)} = \frac{q}{4(q+3)}\\
&\\
2 & \textrm{ with probability } \displaystyle \frac{1}{2} \frac{3q^2(q-1)^3}{q^2(q-1)^3(q+3)} = \frac{3}{2(q+3)}\\
&\\
0 & \textrm{ with probability } \displaystyle \frac{3q}{4(q+3)}+ \frac{3}{2(q+3)}= \frac{3(q+2)}{4(q+3)}
\end{cases}\]}
Note that the expected number of points in a fiber is $1$ and on the whole curve is $q+1.$
\bigskip\noindent {$2.$ \bf The general case: $r$-quadratic covers.}
The argument can be generalized to the case of $r$-quadratic covers of ${\mathbb P}^1$ over ${\mathbb F}_q$, i.e. covers $\pi:C \to {\mathbb P}^1$ with \[\Aut (C/{\mathbb P}^1) \simeq \underbrace{{\mathbb Z}/2{\mathbb Z} \times \dots \times {\mathbb Z}/2{\mathbb Z}}_{r \textrm{ times}}.\]
These are precisely the probabilities that appear in Theorem \ref{main_gen} of the present paper. Our argument works for any $r \in {\mathbb Z}_{>0}.$ For $r=2$ we will recover the predictions from Section $1.$ and for $r=1$ we will recover the results from \cite{KR09}.
An affine model of an $r$-quadratic cover is given by equations
\[\begin{cases}
y_1^2 = h_1(t) &\\
y_2^2 = h_2(t)&\\
\vdots &\\
y_r^2 = h_r(t),&
\end{cases}\]
with $ h_1, \dots, h_r \in {\mathbb F}_q[t]$ square-free polynomials. The cover corresponds to the $r$-quadratic field extension $k\left(\sqrt{ h_1}, \sqrt{ h_2}, \dots, \sqrt{ h_r} \right)$ of $k={\mathbb F}_q(t).$
Together with the ``secret'' equations -- the equivalents of \eqref{eq:secretr=2} -- we get in fact $2^r -1$ equations
\begin{equation}\label{eq:alldef}y_J^2 = \prod_{j \in J} h_j(t)\end{equation}
indexed by the nonempty subsets $J \subseteq \{1, \dots, r\}.$ Again we want to take out gcd's as we did in Section \ref{sec:family}. We obtain $f_1, \dots, f_{2^r-1}$ square-free pairwise coprime polynomials that define this extension. (We can do this by choosing any enumeration of the nonempty subsets $J$. One possibility is the one described in the proof of Theorem 6.3 in the paper.)
Let us examine the fiber above a point $\alpha \in {\mathbb P}^1.$ We first consider the geometric points. In order to ease notation, let $m = 2^r-1.$ We need to look at
\[(f_1, \dots, f_m) \equiv \left(f_1(\alpha) + f_1'(\alpha)(t-\alpha), \dots, f_m(\alpha) + f_m'(\alpha) (t-\alpha) \right) \bmod (t-\alpha)^2.\]
First, since \[(f_1, \dots, f_m) \not \equiv (0,\dots,0) \bmod (t-\alpha)^2,\] we start with at most $(q^2-1)^m$ choices. From these, we need to exclude those that would allow two or more of the $f_i$'s to be divisible by $(t-\alpha),$ since they have to be pairwise coprime.
Fix an integer $k$ with $1 \leq k \leq m.$ Note that if have exactly $k$ zeros among $f_1(\alpha), \dots, f_m(\alpha),$ then the corresponding derivatives of the $k$ polynomials that have a zero at $\alpha$ must be nonzero. Thus, for each choice of $k$ numbers in $\{1, \dots, m\}$ we have \[(q-1)^k (q-1)^{m-k} q^{m-k} =q^{m-k} (q-1)^m \textrm{ choices}.\] Therefore we start with
\[(q^2-1)^m - \sum_{k=2}^m \binom{m}{k} q^{m-k} (q-1)^m
=q^{m-1}(q-1)^m (q+m) \textrm{ tuples modulo }(t-\alpha)^2.\]
Geometrically, we have the following possibilities.
\begin{enumerate}
\item[\bf I.] There will be $2^r$ distinct points when $f_1(\alpha), \dots, f_m(\alpha)$ are all nonzero. Again in this case we have no restrictions on the derivatives. Thus we have $(q-1)^{m}q^m$ possibilities.\\
\item[\bf II.] Otherwise, only one of the $f_1(\alpha), \dots, f_m(\alpha)$ can be zero (as the $f_j$s are pairwise coprime). In this case, the fiber will contain $2^{r-1}$ geometric points. The situation is completely symmetric in $f_1, \dots, f_m.$ Thus, it is enough to count the case when $f_1(\alpha) =0$ and multiply the result by $m.$ Then $f_1'(\alpha) \neq 0$ so the derivative can take $(q-1)$ values. For the other $m-1$ terms, we have $f_j(\alpha) \neq 0$, so it can take $(q-1)$ values, and there are no restrictions on the derivatives $f'_j(\alpha), 2 \leq j \leq m.$ This means that there are $m (q-1) (q-1)^{m-1}q^m$ choices in total that lead to this case. Note that this is equal to \[mq^{m-1}(q-1)^m= q^{m-1}(q-1)^m (q+m) - q^{m}(q-1)^{m}.\]
\end{enumerate}
We now look at the ${\mathbb F}_q$-rational points in the fiber above $\alpha.$ The two cases above split into two cases each.
\begin{enumerate}
\item[{\bf Ia.}] The fiber has $2^r$ ${\mathbb F}_q$-rational points which occurs with probability $1/2^r$ out of case I.
\item[{\bf Ib.}] The fiber has no ${\mathbb F}_q$-rational points which occurs with probability $(2^r-1)/2^r$ out of case I.
\item[{\bf IIa.}] The fiber consists of $2^{r-1}$ ${\mathbb F}_q$-rational points which occurs with probability $1/{2^{r-1}}$ out of case II.
\item[{\bf IIb.}] The fiber contains no ${\mathbb F}_q$-rational points which occurs with probability $(2^{r-1}-1)/{2^{r-1}}$ out of case II.
\end{enumerate}
Note that this already tells us that the expected number of points in a fiber is $1$ and on the whole curve is $q+1.$ Another interesting observation is that the case of biquadratic extensions is a bit different than the general case. For instance, for $r=2$ case II splits into two subcases with probability 50-50; but in general the probabilities for the subcases IIa and IIb are $2^{1-r}$ and $(1 - 2^{1-r}).$ Which means that one has to look at the case $r\geq 3$ in order to get the complete picture.
Since $m = 2^r-1$ we get the following prediction.
\bigskip\noindent {\bf Conjecture $2$.}\, {\em
\[\mathrm{Prob}\left( \#C({\mathbb F}_q) = M; \textrm{$C$ $r$-quadratic cover of ${\mathbb P}^1$ defined over ${\mathbb F}_q$}\right) \sim \mathrm{Prob} (X_1+ \dots+X_{q+1} = M)\]
where $X_i$'s are i.i.d. random variables taking values
\[ X_i = \begin{cases}
2^r & \textrm{ with probability } \displaystyle \frac{1}{2^r} \cdot \frac{q^m(q-1)^m}{q^{m-1}(q-1)^m(q+m)} = \frac{q}{2^r(q+m)}= \frac{q}{2^r(q+2^r -1)}\\
&\\
2^{r-1} & \textrm{ with probability } \displaystyle \frac{1}{2^{r-1}} \cdot \frac{mq^{m-1}(q-1)^m}{q^{m-1}(q-1)^m(q+m)} = \frac{m}{2^{r-1}(q+m)} = \frac{2^r-1}{2^{r-1}(q+2^r -1)}\\
&\\
0 & \textrm{ with probability } \displaystyle \frac{(2^r-1)(q+2^r-2)}{2^r(q+2^r-1)}.
\end{cases}\]}
These are exactly the probabilities that appear in Theorem \ref{main_gen} of the present paper.
The argument works for any $r \in {\mathbb Z}_{>0}.$ For $r=2$ we recover the predictions from previous Section; for $r=1$ we recover the random variables from \cite{KR09}.
|
\section{Introduction}
In marketing literature it has been successively referred the importance of calculating the value of a customer. In fact, such indicative value enables firms to select those customers that can add profit and consequently constitutes an important information to segment the market and efficiently allocate marketing policy resources~\cite{Kumar-2006a,Kumar-2010,Kumar-Petersen-Leone-2007}.
The objective of this work is to establish and study a compartmental model, mathematically translated into a system of ordinary differential equations, for the evolution of the number of customers of some firm, assuming that the customers are divided in two subgroups corresponding to different profitabilities.
Until recently, the value of a customer for a company was based on the present value of future profits generated by a customer over the full course of their dealings with a particular company, this is the customer life-time-value (CLV)~\cite{Kumar-al-2010}. However, other authors refer to the importance of including not only the present and future revenue from the customer purchases, but also the value of the potential to influence other customers under incentives on behalf of the company (customer referral value) or by own initiative (customer influencer value)~\cite{Kumar-2010}. Customer influencing behaviors consists of the intrinsic behaviors motivating the customer to persuade and influence other customers without there being any type of reward on behalf of the company and thus designated the customer influencer value (CIV). In turn, the patterns of customer recommendation are related to the acquisition of new customers due to company initiatives that reward recommendations made to other customers, and thereby establishing the customer referral value (CRV).
According to Kumar et al.~\cite{Kumar-al-2010}, these components are mutually interwoven. Thus, CLV positively correlates with CRV (although only up to a certain point and in an inverted U-shaped relational curve, which means customers reporting average CLV are those most interested in company referral programs) and CLV is positively related with CIV (with an inverted U-shaped relationship in effect between these two concepts).
Much of the literature has focused on the customer referral value through the influence customers might have on the formation of other customers’ attitudes (\cite{Bone-1995}) in the purchasing making decision~\cite{Bansal-Voyer-JSR-2000} and in the reduction of other customers perceived risks~\cite{Godes-Mayzlin-MS-2004}, but little is known about how this processes occur. Since customer referral value and influencer value might have a great impact for companies, these latter try to identify the most influential customers~\cite{Kiss-Bichler-DSS-2008}.
A number of studies allow us to think that the customers of a firm can be classified into several groups according to their influential role over other potential buyers. In imperfect competitive markets information is not purely transparent; some persons are more able than others of influencing people to become a customer of that firm. It is also acceptable to assume that knowing the referrals among each firm’s customers and quantifying their influence constitutes an important asset for the firm competitive advantage, although all customers are important, referrals would be more valuable.
We mostly agree with Marti and Zenou~\cite{Marti-Zenou-2009} when they state that physics/applied mathematics are capable of reproducing many real networks but never reach to explain why they emerge; the economists are very precise to explain why they emerge but their approach does a poor job in matching real world networks. That is why some game theorists are now improving models which take networks as given entities and study the impact of their structure on individuals’ outcomes.
Based on the network theory some models have been tested to study the way influential customers can influence other consumers. For instance Kiss and Bichler~\cite{Kiss-Bichler-DSS-2008} tested real network models, simulated networks and diffusion models to predict influence between customers based on their position within the network. However, as the authors mention this analysis not always is possible if we do not know or do not have information regarding the customer social network. Therefore, other models are needed to try to explain these processes.
In this work we propose a model suitable to describe the dynamics of the number of customers of a given firm. This model is given by a system of ordinary differential equations whose variables correspond to groups of customers and potential customers divided according to their profile and whose parameters reflect the structure of the underlying social network and the marketing policy of the firm. We intend to understand the flows between these groups and its consequences on the raise of customers of the firm. We also want to highlight the usefulness of these models in helping firms deciding their marketing policy.
Specifically, the main objectives of our study is threefold: we intend to obtain theoretical results concerning the long term behavior of the number of customers in various scenarios, we want to present some simulation aimed at illustrating the possibilities of application of our model and, finally, we want to discuss the benefits and limitations of this type of analysis.
As referred, we will consider a compartmental model. As far as we are aware, this is the first time such type of mathematical model is considered in the context of marketing research. We believe that this type of model can be fruitfully explored in this context. This believe is based on the fact that compartmental models have proved to be an important tool not only in the natural sciences, particularly in mathematical epidemiology~\cite{Brauer-Driessche-Jianhong-LNM-2008} and in population biology~\cite{Thieme-PUP-2003, Zhao-SV-2003}, but also, with increasing notoriety in recent years, in the context of economy and other social sciences~\cite{Lin-NAHS-2008, Stiglitz-1993, Tramontana-IRE-2010, Artzrouni-Tramontana-JPM-2014}.
We consider a continuous compartmental model with four compartments, represented by the graph in Figure~\ref{diagram:model} and governed by an autonomous system of four ordinary differential equations.
\begin{figure}[h]
\begin{picture}(360,190)(-30,-20)
\setlength{\unitlength}{.3mm}
\put(5,0){\framebox(60,45){$C$}}
\put(5,100){\framebox(60,45){$R$}}
\put(270,0){\framebox(60,45){$P_C$}}
\put(270,100){\framebox(60,45){$P_R$}}
\put(80,140){\vector(1,0){180}}
\put(160,145){$\beta_2 R$}
\put(260,120){\vector(-1,0){180}}
\put(160,127){$\lambda_3 P_R$}
\put(260,110){\vector(-1,0){180}}
\put(110,92){$\left( \lambda_2 + m_R \lambda_6 \right) R P_R + m \lambda_4 P_R $}
\put(80,40){\vector(1,0){180}}
\put(160,45){$\beta_1 C$}
\put(260,20){\vector(-1,0){180}}
\put(160,27){$\lambda_1 P_C$}
\put(260,10){\vector(-1,0){180}}
\put(130,-7){$\lambda_2 R P_C + m \lambda_4 P_C$}
\put(25,50){\vector(0,1){45}}
\put(0,70){$\lambda_5 C$}
\put(35,95){\vector(0,-1){45}}
\put(40,70){$\lambda_7 R$}
\put(295,50){\vector(0,1){45}}
\put(265,70){$\lambda_5 P_C$}
\put(305,95){\vector(0,-1){45}}
\put(310,70){$\lambda_7 P_R$}
\put(0,25){\vector(-1,0){30}}
\put(-20,10){$\epsilon C$}
\put(0,115){\vector(-1,0){30}}
\put(-20,100){$\epsilon R$}
\put(375,30){\vector(-1,0){40}}
\put(335,37){$(1-\alpha) \gamma$}
\put(375,130){\vector(-1,0){40}}
\put(345,137){$\alpha \gamma$}
\put(335,115){\vector(1,0){40}}
\put(345,105){$\epsilon P_R$}
\put(335,15){\vector(1,0){40}}
\put(345,5){$\epsilon P_C$}
\end{picture}
\caption{The compartmental model}
\label{diagram:model}
\end{figure}
Measuring time in years, we consider the following (pairwise disjoint) compartments: $R(t)$, the referral customers in time t, $C(t)$, the regular customers in time t, $P_R(t)$, the potential referral customers in time t and $P_C(t)$, the potential regular customers in time t. To model transitions between compartments we consider the following parameters: $\lambda_1$, the natural transition rate between $P_C$ and $C$, given by the number of potential regular customers that become regular customers without external influence per year over the number of potential customers (by ``without external influence'' we mean without being influenced by marketing campaigns or referral customers); $\lambda_2$, the referral pull effect, given by the average number of customers that a single referral brings (with no additional incentive) per year over the number of potential customers; $\lambda_3$, the natural $P_R$ to $R$ transition rate, corresponding to the number of potential referral customers that become referral customers without external influence per year over the number of potential referral customers; $m(t)$, the undifferentiated marketing costs, corresponding to marketing costs associated to undifferentiated marketing campaigns per year; $\lambda_4$, the pull effect due to undifferentiated marketing, corresponding to the quotient of the outcome of undifferentiated marketing campaigns per year by the number of potential customers (by ``outcome of undifferentiated marketing campaigns'' it is meant the number of potential customers that become customers in the sequence of undifferentiated marketing campaigns per unitary marketing cost per year); $m_R(t)$, the referral associated marketing costs, corresponding to marketing costs associated to referral directed marketing campaigns per year; $\lambda_6$, the pull effect due to referral directed marketing, given by the referral directed marketing campaigns outcome over the number of potential customers (by ``referral directed marketing campaigns outcome'' it is meant the average number of additional customers that a single referral can bring with incentives per unitary marketing cost per year); $\lambda_5$, the non-central/central transition in the social network equal to the number of individuals non-central in the social network that become central over the total number of individuals in the social network; $\lambda_7$, the central/non-central transition in the social network, given by the number of individuals central in the social network that become non-central over the total number of individuals in the social network; $\beta_1$, the regular customer defection rate, equal to the number of regular customers that cease to be customers over the number of regular customers; $\beta_2$, the referral defection rate, given by the number of referrals that cease to be customers over the number of referrals; $\epsilon$, the customer \& potential customer defection rate, corresponding to the number of individuals that leave the universe of customers and potential customers per year over the number of customers and potential customers (by ``number of individuals that leave the universe of customers and potential customers per year'' it is meant the number of customers and potential customers that cease to be in the set of customers or potential customers per year due to emigration, death, etc.); $\gamma$, the customer and potential customer recruitment rate, given by the number of individuals that enter the universe of customers and potential customers per year over the number of customers and potential customers (by ``number of individuals that leave the universe of customers and potential customers per year'' it is meant the number of customers and potential customers that cease to be in the set of customers or potential customers per year due to immigration, etc.); $\alpha$, the referral recruitment rate, equal to the number of referrals that enter the universe of customers and potential customers per year over number of individuals that enter the universe of customers and potential customers per year.
Our model can be translated into the following system of differential equations~\eqref{eq:modelo} to be studied along this paper.
{\small
\begin{equation}\label{eq:modelo}
\begin{cases}
C'= \lbd_7 R -(\eps+\beta_1+\lbd_5) C + (\lbd_1+m\lbd_4) P_C+\lbd_2RP_C\\
R'= \lbd_5 C-(\eps+\beta_2+\lbd_7)R + (\lbd_3+m\lbd_4) P_R +(\lbd_2+m_R\lbd_6)RP_R\\
P_C'=(1-\alpha)\gamma + \beta_1 C+ \lbd_7 P_R - (\eps+\lbd_5+\lbd_1+m\lbd_4) P_C - \lbd_2RP_C \\
P_R'=\alpha \gamma + \beta_2 R + \lbd_5 P_C - (\eps+\lbd_7+\lbd_3+m\lbd_4) P_R - (\lbd_2+m_R\lbd_6) RP_R
\end{cases}
\end{equation}
}
Notice that $\lambda_2 R P_R$ correspond to the average number of referrals that are brought with no incentive by the referrals per year, that $\lambda_2 R P_C$ is the average number of regular customers that are brought with no incentive by the referrals per year, that $m_R(t) \lambda_6 R P_R$ is the average number of additional customers brought due to incentives per year, that $m(t) \lambda_4 P_R$ is the number of potential referrals that become referrals in the sequence of undifferentiated marketing campaigns per year and that $m(t) \lambda_4 P_C$ is the number of potential regular customers that become regular customers in the sequence of undifferentiated marketing campaigns per year.
This paper is divided in the following way: in section~\ref{section:MR} we state our main results concerning the asymptotic behavior of the number of regular customers and referral customers, in section~\ref{section:S} we present some simulation with the objective of illustrating our theoretical results, in section~\ref{section:P} we prove our results and, finally, in section~\ref{section:Conc} we discuss the results obtained.
\section{Main Results} \label{section:MR}
One of the first natural issues to address when studying a compartmental model is the existence and stability of equilibrium solutions. We obtain several results on the existence and stability of equilibrium solutions in of model~\eqref{eq:modelo} in this section.
We first derive an auxiliary result. Given $\delta \ge 0$ define the sets
$$\Delta_\delta=\{(x,y,z,w) \in (\R_0^+)^4: \gamma/\eps-\delta \le x+y+z+w \le \gamma/\eps+\delta\}.$$
We have the following result that shows that the total population in system~\eqref{eq:modelo} converges to the ratio $\gamma/\eps$, independently of the nonnegative initial conditions considered.
\begin{lemma}[Asymptotic behavior of the total population]\label{teo:general_system}
Let $\eps>0$ and let $(C(t),R(t),P_C(t),P_R(t))$ be some solution of system~\eqref{eq:modelo} with nonnegative initial conditions: $C(t_0) \ge 0$, $R(t_0) \ge 0$, $P(t_0) \ge 0$, $P(t_0) \ge 0$. Then:
\begin{enumerate}[a)]
\item \label{teo:general_system-1} for all $t \ge t_0$, we have $C(t),R(t),P_C(t),P_R(t) \ge 0$;
\item \label{teo:general_system-2} we have $\displaystyle \lim_{t \to +\infty} C(t)+R(t)+P_C(t)+P_R(t) = \dfrac{\gamma}{\eps}$.
In particular, given $\delta>0$ and any solution $(C(t),R(t),P_C(t),P_R(t))$, with nonnegative initial conditions there is $T>0$ such that $(C(t),R(t),P_C(t),P_R(t)) \in \Delta_\delta$ for all $t \ge T$ and any equilibrium solution is in the set
$\Delta_0=\{(x,y,z,w) \in (\R_0^+)^4: x+y+z+w = \gamma/\eps\}$.
\end{enumerate}
\end{lemma}
\begin{remark}
The case $\eps=0$ is not a very interesting case since it corresponds to the situation where there is no customer \& potential customer defection rate which is not a realistic assumption. Nevertheless it is easy to check that, if $\eps=0$, then, given initial conditions $C_0$, $R_0$, $P_{C,0}$ and $P_{R,0}$, we have
$$\displaystyle \lim_{t \to +\infty} C(t)+R(t)+P_C(t)+P_R(t) = \gamma t + C_0+R_0+P_{C,0}+P_{R,0}.$$
In particular if $\eps=\gamma=0$, then the total population remains constant.
\end{remark}
We now obtain a result on the existence of equilibrium solutions. Under the assumption of positivity of the defection rate, the referral pull effect and the non central/central transition in the social network we conclude that there are one, two or three equilibrium solutions, depending on the number of real roots of some third degree polynomial. We need to define the constants
\begin{equation}\label{eq:p_q}
p=\dfrac{\gamma(\alpha\eps+\lbd_5)}{\eps(\eps+\lbd_5+\lbd_7)} \quad \quad \text{and}
\quad \quad q=\dfrac{\gamma((1-\alpha)\eps+\lbd_7)}{\eps(\eps+\lbd_5+\lbd_7)}
\end{equation}
and also
\[
u=\eps+\beta_1+\lambda_5+\lambda_1+m\lambda_4
\quad \text{and} \quad
v=\eps+\beta_2+\lambda_7+\lambda_3+m\lambda_4.
\]
\begin{theorem}[Equilibrium solutions]\label{teo:equilibriums}
Let $\eps,\lambda_5>0$. Then:
\begin{enumerate}[a)]
\item \label{teo:equilibriums-1} system~\eqref{eq:modelo} has up to three equilibrium solutions $(R^*,C^*,P_R^*,P_C^*)$. The first component, $R^*$, is always a nonnegative solution of the cubic equation
\begin{equation}\label{eq:solution-R}
aR^3+bR^2+cR+d=0,
\end{equation}
where $a=-\lambda_2(\lambda_2+m_R\lambda_6)$, $b=\lbd_2(\lbd_2+m_R\lbd_6)p-u(\lambda_2+m_R\lambda_6)-\lambda_2 v$, $c=\lbd_2(\lbd_3+m\lbd_4)p+u(\lbd_2+m_R\lbd_6)p+(\lbd_7+\lbd_2 q)\lbd_5-uv$ and $d=(\lambda_3+m\lambda_4)pu+(\lambda_1+m\lambda_4)q\lambda_5$;
\item \label{teo:equilibriums-2} any equilibrium solution $(R^*,C^*,P_R^*,P_C^*)$ is obtained in the following way: $R^*$ is a nonnegative solution of~\eqref{eq:solution-R} and $P_R^*=p-R^*$, $P_C^*=q-C^*$ and
\begin{equation}\label{eq:C*}
C^*=\dfrac{(\eps+\beta_2+\lambda_7)R^*- (\lambda_3+m\lambda_4+(\lambda_2+m_R\lambda_6)R^*)(p-R^*)}{\lambda_5},
\end{equation}
are nonnegative constants.
\end{enumerate}
\end{theorem}
In the next result we discuss the asymptotic behavior of solutions of~\eqref{eq:modelo} under some assumptions on the parameters that roughly correspond to require that the referral pull effects are bounded by some functions that we can identify with the other ``forces'' in the model such as the natural transition rates, the pull effects due to undifferentiated marketing and the defection rates (see equation~\eqref{eq:cond-equiv}).
In the following theorem we were able to show that, under the mentioned assumptions, the asymptotic behavior of the solutions of~\eqref{eq:modelo} can be obtained by the two dimensional autonomous system~\eqref{eq:sistema_equiv_para_C_and_R_=_0}.
\begin{theorem}[Asymptotic behavior of solutions] \label{teo:asymptotic_behavior}
Let $\eps>0$ and assume that
\begin{equation}\label{eq:cond-equiv}
\small \min\left\{\dfrac{2\eps+2\beta_1+\lbd_5+2\lbd_1+2m\lbd_4}{\lbd_7+q\lbd_2}, \, \frac{2\eps+2\beta_2+\lbd_7+2\lbd_3+2m\lbd_4+2(\lbd_2+m_R\lbd_6)p}{\lbd_5+q\lbd_2} \right\}\!>\!1.
\end{equation}
Consider the system
{\small
\begin{equation}\label{eq:sistema_equiv_para_C_and_R_=_0}
\begin{cases}
C_a'=\lbd_7 R_a -(\eps+\beta_1+\lbd_5) C_a + (\lbd_1+m\lbd_4+\lbd_2 R_a) (q-C_a)\\
R_a'=\lbd_5 C_a-(\eps+\beta_2+\lbd_7)R_a + (\lbd_3+m\lbd_4+(\lbd_2+m_R\lbd_6)R_a) \left(p - R_a\right)
\end{cases}
\end{equation}}
and set $P_{R,a}(t)=p-R_a(t)$ and $P_{C,a}(t)=q-C_a(t)$. Then the asymptotic behavior of $C$, $R$, $P_C$ and $P_R$ in system~\eqref{eq:modelo} is the same as the asymptotic behavior of $C_a$, $R_a$, $P_{C,a}$ and $P_{R,a}$. Namely if $(C(t),R(t),P_C(t),P_R(t))$ is a solution of~\eqref{eq:modelo} with initial condition $(C(t_0),R(t_0),P_C(t_0),P_R(t_0))=(C_0,R_0,P_{C,0},P_{R,0})$ and $(C_a(t),R_a(t))$ is a solution of~\eqref{eq:sistema_equiv_para_C_and_R_=_0} with initial condition $(C_a(t_0),R_a(t_0))=(C_0,R_0)$ then
$$\lim_{t \to +\infty} |C(t)-C_a(t)|+|R(t)-R_a(t)|+|P_C(t)-P_{C,a}(t)|+|P_R(t)-P_{R,a}(t)|=0.$$
\end{theorem}
In the next two theorems we discuss two particular situations where we analyse the existence of equilibriums and their stability. The vector fields plotted with the objective of illustrating the situations correspond to the reduced system~\eqref{eq:sistema_equiv_para_C_and_R_=_0} but we considered situations where Theorem~\ref{teo:asymptotic_behavior} apply so that the asymptotic behavior of referrals and regular clients is the same for both systems.
First we will discuss the situation where there is no transition between referral/potential referral and customer/potential customer and thus we set $\lbd_5=\lbd_7=0$. We consider two cases: the situation where $\lbd_3+m\lbd_4 \ne 0$ and $\lbd_1+m\lbd_4 \ne 0$ (we named it ``static social network'' to reflect the fact that there is no transition between referrals and regular customers) and the situation where $\lbd_1=\lbd_2=\lbd_4=0$, corresponding to the case where all potential customers and potential referrals that become customers are consequence of referral influence (we named it ``word of mouth'' to emphasise that all marketing efforts are related to referrals).
We have the following result in the static social network case.
\begin{theorem}[Static social network]\label{teo:static_social_network}
The following holds for system~\eqref{eq:modelo} with $\eps>0$, $\lbd_2>0$, $\lbd_3+m\lbd_4 \ne 0$, $\lbd_1+m\lbd_4 \ne 0$ and $\lbd_5=\lbd_7=0$: there is a unique equilibrium solution $(R^*,C^*,P_R^*,P_C^*)$ that is locally asymptotically stable and is given by
\begin{equation}\label{eq:static-R*}
R^* = \dfrac{\alpha\gamma}{2\eps}-\theta+\sqrt{\left(\dfrac{\alpha\gamma}{2\eps}+\theta\right)^2 -\frac{\alpha\gamma(\eps+\beta_2)}{\eps(\lbd_2+m_R\lbd_6)}},
\end{equation}
$$C^* = \dfrac{(1-\alpha)\gamma(\lbd_1+m\lbd_4+\lbd_2 R^*)}{\eps(\eps+\beta_1+\lbd_1+m\lbd_4+\lbd_2 R^*)},$$\\[1mm]
$P_R^*=\gamma\alpha/\eps-R^*$ and $P_C^*=\gamma(1-\alpha)/\eps-C^*$ where
\[
\theta=\dfrac{\lbd_3+m\lbd_4+\beta_2+\eps}{2(\lbd_2+m_R \lbd_6)}.
\]
\end{theorem}
Define
\begin{equation}\label{eq:tau}
\tau=\dfrac{\alpha\gamma(\lbd_2+m_R\lbd_6)}{\eps(\eps+\beta_2)}.
\end{equation}
We now consider the word of mouth case. In Figure~\ref{fig1} we show the behavior in the plane $C-R$ in both of the regimes of Theorem~\ref{teo:mri}.
In Figure~\ref{fig1} we used for the plot in the left $m = 40$, $m_R = 0$, $\eps=0.01$ and $\lambda_2 = 10^{-5}$, and, for the plot in the right, $m = 30$, $m_R = 10$, $\eps=0.01$ and $\lambda_2 = 10^{-5}$.
\begin{theorem}[Word of mouth]\label{teo:mri}
The following statements holds for system~\eqref{eq:modelo} with $\eps>0$, $\lbd_2>0$ and $\lbd_1=\lbd_3=\lbd_4=\lbd_5=\lbd_7=0$:
\begin{enumerate}[a)]
\item if $\tau \le 1$ then there a unique locally stable equilibrium given by
\begin{equation*}
(C^*,R^*,P_C^*,P_R^*) =\left(0,0,\frac{\gamma(1-\alpha)}{\eps},\frac{\gamma\alpha}{\eps}\right);
\end{equation*}
\item if $\tau > 1$ then there are two equilibrium solutions. An unstable equilibrium given by
\begin{equation*}
(C_1^*,R_1^*,P_{C,1}^*,P_{R,1}^*) =\left(0,0,\frac{\gamma(1-\alpha)}{\eps},\frac{\gamma\alpha}{\eps}\right);
\end{equation*}
and a locally stable equilibrium given by
\begin{equation*}
\begin{split}
& (C_2^*, R_2^*, P_{C,2}^*,P_{R,2}^*) \\
& = \left(\frac{\alpha(1-\alpha)\lambda_2\gamma^2(1-\frac{1}{\tau})} {\eps^2(\eps+\beta_1+\lbd_2(1-\frac{1}{\tau}))},
\dfrac{\alpha\gamma(1-\frac{1}{\tau})}{\eps},
\frac{(1-\alpha)(\eps+\beta_1)\gamma}{\eps(\eps+\beta_1+\lbd_2(1-\frac{1}{\tau}))}, \frac{\eps+\beta_2}{\lbd_2+m_R\lbd_6}\right).
\end{split}
\end{equation*}
\end{enumerate}
\end{theorem}
\begin{figure}[h!]
\begin{minipage}[b][4cm]{.5\linewidth}
\includegraphics[scale=0.37]{tau_menor_um.eps}
\end{minipage}
\begin{minipage}[b][4cm]{.2\linewidth}
\includegraphics[scale=0.37]{tau_maior_um.eps}
\end{minipage}
\caption{Word of mouth: $\tau<1$ and $\tau>1$}
\label{fig1}
\end{figure}
We next consider a scenario where there is no direct referral influence and thus we set $\lbd_2=\lbd_6=0$. Define
\[
\kappa_1=\dfrac{\lbd_5[\lbd_7p+(\lbd_1+m\lbd_4)q]}{up(\eps+\beta_2+\lbd_7)}
\quad \text{and} \quad \kappa_2=\dfrac{\lbd_7[\lbd_5q+(\lbd_3+m\lbd_4)p]}{vq(\eps+\beta_1+\lbd_5)}.
\]
\begin{theorem}[No referral influence]\label{teo:no_referral_influence}
For system~\eqref{eq:modelo} with $\eps>0$, $\lbd_5>0$ and $\lbd_2=\lbd_6=0$ there is a unique equilibrium solution that is globally asymptotically stable and is given by
$$(R^*, C^*, P_R^*, P_C^*)=(\kappa_1 p, \, \kappa_2 p, \, p-\kappa_1p, \, q-\kappa_2p).$$
\end{theorem}
In Figure~\ref{fig2} we illustrate the behavior of the reduced system in the plane $C-R$ for the setting in Theorem~\ref{teo:no_referral_influence}. In Figure~\ref{fig2} we used $\lbd_2=\lbd_6=0$, $m = 40$, $m_R = 0$, $\lbd_1=\lbd_3=\lbd_4=\lbd_7=0.0002$, $\lbd_5=0.0000018$, $\beta_1=\beta_2=0.18$, $\eps=0.01$ and, for the plot on the left, $\alpha=0$, and, for the plot on the right $\alpha=0.5$. Note that for the figure in the left the number of referrals in the equilibrium point is nonzero, although it seems to be (the number of referrals is very low since the there are no referrals entering the population).
\begin{figure}[h!]
\begin{minipage}[b][4cm]{.5\linewidth}
\includegraphics[scale=0.37]{kapa_alpha0.eps}
\end{minipage}
\begin{minipage}[b][4cm]{.2\linewidth}
\includegraphics[scale=0.37]{kapa_alpha05.eps}
\end{minipage}
\caption{Word of mouth: $\alpha=0$ and $\alpha=0.5$}
\label{fig2}
\end{figure}
\section{Simulation} \label{section:S}
To obtain a better understanding of the behavior of our model, we assume that, for a given corporation, we consider the values for the parameters presented in Table~\ref{tab:ref}.
\begin{table}[htb]
\centering
{\renewcommand{\arraystretch}{1.5}
\begin{tabular}{cc}
\toprule
Parameter & Value \\
\midrule
$\alpha$, $\epsilon$& 0.01\\
$\gamma$ & $\epsilon N_0$\\
$\lambda_1,\ldots,\lambda_4,\lambda_6, \lambda_7$ & 0.0002 \\
$\lambda_5$ & $\lambda_7 R_0/C_0$\\
$\beta_1,~ \beta_2$ & 0.18\\
\bottomrule
\end{tabular}}
\caption{Values of parameters}
\label{tab:ref}
\end{table}
and initial conditions $C_0=2200$, $P_{C,0}=22000$, $R_0=20$ and $P_{R,0}=200$. We write $N_0=C_0+P_{C,0}+R_0+P_{R,0}$. The value considered for $\beta_1$ and $\beta_2$ is based on usual assumptions concerning the defection rate~\cite{Lee-Lee-Feick-JDBCSM-2006}. The assumptions $\lbd_5 C_0=\lbd_7 R_0$ and $\lbd_5 P_{C,0}=\lbd_7 P_{R,0}$ are made to assure that the underlying social network hasn't an initial tendency to ``benefit'' any of the four compartments. We also consider $\gamma=N_0\times \epsilon$ so that the total population converges to an equilibrium where the total population equals $N_0$.
We solved system~\eqref{eq:modelo} (named \emph{initial}) and system~\eqref{eq:sistema_equiv_para_C_and_R_=_0} (named \emph{reduced}) with MATLAB. In the figures we plot the solution for the initial and the reduced systems. We considered two sets of values for $m$ and $m_R$, namely $(m,m_R)=(40,0)$ corresponding to a situation of undifferentiated marketing and $(m,m_R)=(30,10)$ corresponding to a situation where some effort is made for attracting referrals. In both situations we maintain the same total effort, $m+m_R=40$, in order to be able to compare both cases.
In Figures~\ref{fig:undif-costumers} and~\ref{fig:undif-referrals} we consider the evolution of customers and referrals in the case where marketing is used in an undifferentiated way. We can see that the number of customers and referrals decreases in this situation and stabilizes in some lower value for both compartments.
\begin{figure}[!htb]
\centering
\begin{subfigure}[b]{0.49\textwidth}
\includegraphics[scale=0.4]{clients_7.eps}
\caption{$t\in[0;7]$}
\end{subfigure
~
\begin{subfigure}[b]{0.49\textwidth}
\includegraphics[scale=0.4]{clients_inf.eps}
\caption{$t\in[0;500]$}
\end{subfigure}
\caption{Regular customers evolution with $m=40$ and $m_R=0$}
\label{fig:undif-costumers}
\end{figure}
\begin{figure}[!htb]
\centering
\begin{subfigure}[b]{0.49\textwidth}
\includegraphics[scale=0.4]{referral_7.eps}
\caption{$t\in[0;7]$}
\end{subfigure
~
\begin{subfigure}[b]{0.49\textwidth}
\includegraphics[scale=0.4]{referral_inf.eps}
\caption{$t\in[0;500]$}
\end{subfigure}
\caption{Referral evolution with $m=40$ and $m_R=0$}
\label{fig:undif-referrals}
\end{figure}
In Figures~\ref{fig:mark-costumers} and~\ref{fig:mark-referrals} we now consider the evolution of customers and referrals in the case where some marketing effort is used to attract referrals. We can see that there is an initial small decrease in the number of customers that is followed rapidly by an increase that asymptotically doubles its number. There is also an increase in the number of referrals due to the positive value of $m_R$.
\begin{figure}[!htb]
\centering
\begin{subfigure}[b]{0.49\textwidth}
\includegraphics[scale=0.4]{clients_7_II.eps}
\caption{$t\in[0;7]$}
\end{subfigure
~
\begin{subfigure}[b]{0.49\textwidth}
\includegraphics[scale=0.4]{clients_inf_II.eps}
\caption{$t\in[0;500]$}
\end{subfigure}
\caption{Regular customers evolution with $m=30$ and $m_R=10$}
\label{fig:mark-costumers}
\end{figure}
\begin{figure}[!htb]
\centering
\begin{subfigure}[b]{0.49\textwidth}
\includegraphics[scale=0.4]{referral_7_II.eps}
\caption{$t\in[0;7]$}
\end{subfigure
~
\begin{subfigure}[b]{0.49\textwidth}
\includegraphics[scale=0.4]{referral_inf_II.eps}
\caption{$t\in[0;500]$}
\end{subfigure}
\caption{Referral evolution with $m=30$ and $m_R=10$}
\label{fig:mark-referrals}
\end{figure}
In the Figure~\ref{fig:asymptotic} we present the evolution of customers and referrals in the case where $\lbd_2$ is reduced to $10^{-5}$ in order to satisfy the condition~\eqref{eq:cond-equiv} in Theorem~\eqref{teo:asymptotic_behavior}. We can see that, as stated in the theorem, the solutions asymptotically approach the same value. In the previous cases, although condition~\eqref{eq:cond-equiv} is not satisfied, there is computational evidence that the same happens. Thus we conjecture that condition~\eqref{eq:cond-equiv} can be weakened.
\begin{figure}[!htb]
\centering
\begin{subfigure}[b]{0.49\textwidth}
\includegraphics[scale=0.4]{clients_inf_III.eps}
\caption{Regular customers evolution}
\end{subfigure
~
\begin{subfigure}[b]{0.49\textwidth}
\includegraphics[scale=0.4]{referral_inf_III.eps}
\caption{Referrals evolution}
\end{subfigure}
\caption{Asymptotic behavior with $m=30$, $m_R=10$ and $\lambda_2=1\times 10^{-5}$}
\label{fig:asymptotic}
\end{figure}
\vspace{1cm}
\section{Proofs} \label{section:P}
\subsection{Proof of Lemma~\ref{teo:general_system}}
Let $\eps >0$. Analysing the direction of the flow on the boundary of the set $\{(x,y,z,w) \in \R^4: x,y,z,w \ge 0\}$ we immediately obtain~\ref{teo:general_system-1}). Adding the four equations in~\eqref{eq:modelo} and letting $N(t)=C(t)+R(t)+P_C(t)+P_R(t)$, we get
the differential equation
\begin{equation}\label{eq:total pop}
N'=\gamma-\eps N.
\end{equation}
The general solution of~\eqref{eq:total pop}
is $N(t)=\gamma/\eps + C \e^{-\eps t}$ and thus $\displaystyle \lim_{t \to +\infty} N(t) = \gamma/\eps$. In particular, if $N(t)$ corresponds to an equilibrium solution then $N(t)=\gamma/\eps$. This proves~\ref{teo:general_system-2}).
\subsection{Proof of Theorem~\ref{teo:equilibriums}}
Adding equations for $R$ and $P_R$ and for $C$ and $P_C$ in~\eqref{eq:modelo} we get
\[
\begin{cases}
(R+P_R)'=\alpha\gamma-(\eps+\lbd_7)(R+P_R) + \lbd_5 (C+P_C)\\
(C+P_C)'=(1-\alpha)\gamma-(\eps+\lbd_5) (C+P_C) + \lbd_7 (R+ P_R)
\end{cases}
\]
and, setting $X=R+P_R$ and $Y=C+P_C$, we obtain
\begin{equation}\label{eq:sist_X_Y}
\begin{cases}
X'=\alpha\gamma-(\eps+\lbd_7) X + \lbd_5 Y \\
Y'=(1-\alpha)\gamma+\lbd_7 X -(\eps+\lbd_5) Y
\end{cases}.
\end{equation}
The linear system~\eqref{eq:sist_X_Y} has the general solution
\[
\begin{cases}
X(t)=\dfrac{\gamma(\alpha\eps+\lbd_5)}{\eps(\eps+\lbd_5+\lbd_7)}+\dfrac{C_1 \lbd_7 - C_2 \lbd_5}{\lbd_5+\lbd_7}\e^{-(\eps+\lbd_5+\lbd_7)t}+\dfrac{(C_1+C_2)\lbd_5}{\lbd_5+\lbd_7} \e^{-\eps t} \\
Y(t)=\dfrac{\gamma((1-\alpha)\eps+\lbd_7)}{\eps(\eps+\lbd_5+\lbd_7)}+\dfrac{C_2 \lbd_5 - C_1 \lbd_7}{\lbd_5+\lbd_7}\e^{-(\eps+\lbd_5+\lbd_7)t}+\dfrac{(C_1+C_2)\lbd_7}{\lbd_5+\lbd_7} \e^{-\eps t}
\end{cases}.
\]
and thus
\begin{equation}\label{eq:elim_Pr_Pc}
\begin{cases}
P_R(t)=\dfrac{\gamma(\alpha\eps+\lbd_5)}{\eps(\eps+\lbd_5+\lbd_7)}+\theta_1(t) - R(t)\\
P_C(t)=\dfrac{\gamma((1-\alpha)\eps+\lbd_7)}{\eps(\eps+\lbd_5+\lbd_7)}+\theta_2(t) - C(t)
\end{cases}
\end{equation}
with $\theta_i(t) \to 0$ as $t \to \infty$ for $i=1,2$. Replacing $P_R$ and $P_C$ by the expressions in~\eqref{eq:elim_Pr_Pc} in the first two equations in system~\eqref{eq:modelo} we obtain
\begin{equation}\label{eq:C_and_R}
\begin{cases}
C'=\lbd_7 R -(\eps+\beta_1+\lbd_5) C + (\lbd_1+m\lbd_4+\lbd_2R) \left( q+\theta_2(t) - C \right)\\
R'=\lbd_5 C-(\eps+\beta_2+\lbd_7)R + (\lbd_3+m\lbd_4+(\lbd_2+m_R\lbd_6)R) \left(p+\theta_1(t) - R\right)
\end{cases},
\end{equation}
where $p$ and $q$ are given by~\eqref{eq:p_q}.
By Lemma~\ref{teo:general_system} every equilibrium solution must belong to $\Delta_0$. Thus every equilibrium solution must satisfy
\begin{equation}\label{eq:C_and_R_equilibrium}
\begin{cases}
\lbd_7 R -(\eps+\beta_1+\lbd_5) C + (\lbd_1+m\lbd_4+\lbd_2R) \left( q - C \right)=0\\
\lbd_5 C-(\eps+\beta_2+\lbd_7)R + (\lbd_3+m\lbd_4+(\lbd_2+m_R\lbd_6)R) \left(p - R\right)=0
\end{cases}.
\end{equation}
Since $\lambda_5 \ne 0$ we have
\begin{equation}\label{eq:C}
C=\dfrac{(\eps+\beta_2+\lbd_7)R - (\lbd_3+m\lbd_4+(\lbd_2+m_R\lbd_6)R) \left(p - R\right)}{\lbd_5}.
\end{equation}
Substituting~\eqref{eq:C} in the first equation of~\eqref{eq:C_and_R_equilibrium} we obtain equation~\eqref{eq:solution-R}. Each solution of~\eqref{eq:solution-R} uniquely determines $C$, $R$ and $P_R$. Thus, we have at most three equilibrium solutions and~\ref{teo:equilibriums-1}) follows.
By~\eqref{eq:elim_Pr_Pc} for any equilibrium solution $(R^*,C^*,P_R^*,P_C^*)$ we have $P_R^*=p-R^*$ and $P_C^*=q-C^*$. By~\eqref{eq:C} we obtain~\eqref{eq:C*} and~\ref{teo:equilibriums-2}).
\subsection{Proof of Theorem~\ref{teo:asymptotic_behavior}}
Assume that $\eps>0$ and choose $\delta>0$. Additionally, define $V:(\R^+)^2 \to \R$ by $V(x,y)=\frac12 (x^2+y^2)$ and let $(C(t),R(t),P_C(t),P_R(t))$ be a solution of~\eqref{eq:modelo} with initial conditions $C(t_0)=C_0$, $R(t_0)=R_0$, $P_R(0)=P_{R,0}$ and $P_C(0)=P_{C,0}$ and $(C_a(t),R_a(t))$ be a solution of~\eqref{eq:sistema_equiv_para_C_and_R_=_0} with initial conditions $C_a(t_0)=C_0$ and $R_a(t_0)=R_0$. Define $x(t)=C(t)-C_a(t)$ and $y(t)=R(t)-R_a(t)$.
By~\eqref{eq:C_and_R} and~\eqref{eq:sistema_equiv_para_C_and_R_=_0} we have
\begin{equation}\label{eq:estim_x}
\begin{split}
x'
& =C'-C_A'\\
& =\lbd_7 (R-R_a)-(\eps+\beta_1+\lbd_5)(C-C_a)+(\lbd_1+m\lbd_4)(\theta_2(t)-(C-C_a))\\
& \quad + q\lbd_2(R-R_a) +\lbd_2\theta_2(t)R+\lbd_2(R_aC_a-RC)\\
& =\lbd_7 y-(\eps+\beta_1+\lbd_5)x+(\lbd_1+m\lbd_4)(\theta_2(t)-x) + q\lbd_2y +\lbd_2\theta_2(t)R\\
& \quad +\lbd_2(R_a(C_a-C)+(R_a-R)C)\\
& = -(\eps+\beta_1+\lbd_5+\lbd_1+m\lbd_4+\lbd_2R_a)x+ (\lbd_7+q\lbd_2-C\lbd_2)y\\
& \quad + (\lbd_2R+(\lbd_1-m\lbd_4))\theta_2(t)
\end{split}
\end{equation}
and
\begin{equation}\label{eq:estim_y}
\begin{split}
y'
& =R'-R_a'\\
& =\lbd_5 (C-C_a)-(\eps+\beta_2+\lbd_7+\lbd_3+m\lbd_4+(\lbd_2+m_R\lbd_6)p)(R-R_a)\\
& \quad +(\lbd_2+m_R\lbd_6)(R^2-R_a^2)+((\lbd_2+m_R\lbd_6)R+\lbd_3+m\lbd_4)\theta_1(t)\\
& =\lbd_5 x-(\eps+\beta_2+\lbd_7+\lbd_3+m\lbd_4+(\lbd_2+m_R\lbd_6)p)y\\
& \quad +(\lbd_2+m_R\lbd_6)(R+R_a)y+((\lbd_2+m_R\lbd_6)R+\lbd_3+m\lbd_4)\theta_1(t)\\
\end{split}.
\end{equation}
By~\eqref{eq:elim_Pr_Pc} in the proof of Theorem~\ref{teo:general_system}, there is $T_0 \ge 0$ such that $P_R(t),R(t) \le P_R(t)+R(t) < p+\delta$ and
$P_C(t),C(t) \le P_C(t)+C(t) < q+\delta$, for all $t \ge T_0$.
Using this fact, by~\eqref{eq:estim_x} and~\eqref{eq:estim_y} we have
\[
\begin{split}
\frac{d}{dt} V(x,y)
& = xx'+yy'\\
& = -(\eps+\beta_1+\lbd_5+\lbd_1+m\lbd_4+\lbd_2R_a)x^2+
(\lbd_7+q\lbd_2-C\lbd_2)xy\\
& \quad + (\lbd_2R+(\lbd_1-m\lbd_4))x\theta_2(t) + \lbd_5 xy\\
& \quad -(\eps+\beta_2+\lbd_7+\lbd_3+m\lbd_4+(\lbd_2+m_R\lbd_6)p)y^2\\
& \quad +(\lbd_2+m_R\lbd_6)(R+R_a)y^2+
((\lbd_2+m_R\lbd_6)R+\lbd_3+m\lbd_4)y\theta_1(t) \\
& \le -(\eps+\beta_1+\lbd_5+\lbd_1+m\lbd_4)x^2+(\lbd_5+\lbd_7+q\lbd_2-C\lbd_2)xy\\
& \quad -(\eps+\beta_2+\lbd_7+\lbd_3+m\lbd_4+
(\lbd_2+m_R\lbd_6)(p+2\delta))y^2+\Theta(t),
\end{split}
\]
for $t \ge T_0$, where
\small{$$\Theta(t)=(\lbd_2R+(\lbd_1-m\lbd_4))(p+\delta)\theta_2(t)+ ((\lbd_2+m_R\lbd_6)R+\lbd_3+m\lbd_4)(q+\delta)\theta_1(t).$$}
Thus, using the fact that $xy \le 1/2(x^2+y^2)$, we get
\[
\begin{split}
\frac{d}{dt} V(x,y)
& \le -(\eps+\beta_1+\lbd_5+\lbd_1+m\lbd_4)x^2+
\frac12(\lbd_5+\lbd_7+q\lbd_2)(x^2+y^2)\\
& \quad -(\eps+\beta_2+\lbd_7+\lbd_3+m\lbd_4+
(\lbd_2+m_R\lbd_6)(p+2\delta))y^2+\Theta(t)\\
& \le -(\eps+\beta_1+\frac12(\lbd_5-\lbd_7)+\lbd_1+m\lbd_4-\frac12q\lbd_2)x^2\\
& \quad -(\eps+\beta_2+\frac12(\lbd_7-\lbd_5)+\lbd_3+m\lbd_4+
(\lbd_2+m_R\lbd_6)(p+2\delta)\\
& \quad -\frac12q\lbd_2)y^2+\Theta(t),
\end{split}
\]
for $t \ge T_0$. By~\eqref{eq:cond-equiv}, there is $M > 0$ such that
\[
\frac{d}{dt} V(x,y) \le -M V(x,y) +\Theta(t).
\]
Since $\Theta(t)\to 0$ as $t\to 0$, there is $T_1 \ge T_0 > 0$ sufficiently large such that, for $t \ge T_1$, we have $\Theta(t) \le \delta$. Thus
$$ \frac{d}{dt} V(x(t),y(t)) \le -M V(x,y) + \delta$$
for $t\ge T_1$. Therefore, for $t \ge T_1$,
$$ V(x(t),y(t)) \le \frac{\delta}{M} + V(x(T_1),y(T_1)) \e^{-M(t-T_1)}.$$
We conclude that
$$\lim_{t \to +\infty} V(x(t),y(t)) \le \frac{\delta}{M}.$$
Since $\delta>0$ is arbitrary, the theorem follows.
\subsection{Proof of Theorem~\ref{teo:static_social_network}}
Adding equations for $R$ and $P_R$ and for $C$ and $P_C$ in~\eqref{eq:modelo} we get the system
\[
\begin{cases}
(R+P_R)'=\alpha\gamma-\eps(R+P_R)\\
(C+P_C)'=(1-\alpha)\gamma-\eps(C+P_C)
\end{cases}
\]
and thus
\[
\begin{cases}
R(t)+P_R(t)=\frac{\alpha\gamma}{\eps}+C_1\e^{-\eps t}\\[2mm]
C(t)+P_C(t)=\frac{(1-\alpha)\gamma}{\eps}+C_2\e^{-\eps t}
\end{cases}.
\]
Therefore, if $(R^*,C^*,P_R^*,P_C^*)$ is an equilibrium solution, then \begin{equation}\label{eq:PR_PC_equilib_static}
P_R^*=\alpha\gamma/\eps-R^* \quad \quad \text{and} \quad \quad P_C^*=(1-\alpha)\gamma/\eps-R^*.
\end{equation}
Using these expressions and the first two equations in System~\eqref{eq:modelo}, we obtain
\begin{equation}\label{eq_equilibrium_static}
\begin{cases}
-(\eps+\beta_1) C^* + (\lbd_1+m\lbd_4+\lbd_2R^*) \left(\frac{(1-\alpha)\gamma}{\eps}- C^* \right)=0\\
-(\eps+\beta_2)R^* + (\lbd_3+m\lbd_4+(\lbd_2+m_R\lbd_6)R^*) \left(\frac{\alpha\gamma}{\eps}- R^*\right)=0
\end{cases},
\end{equation}
and using the second equation in~\eqref{eq_equilibrium_static} and dividing by $-(\lbd_2+m_R\lbd_6)$ (that is nonzero by assumption), we get
$$(R^*)^2+2\theta R^*-\frac{\alpha\gamma(\lbd_3+m\lbd_4)}{\eps(\lbd_2+m_R\lbd_6)}=0.$$
Thus we have a unique nonnegative root given by~\eqref{eq:static-R*}.
By the first equation in~\eqref{eq_equilibrium_static} we obtain
$$C^* = \frac{(1-\alpha)\gamma}{\eps} \frac{\lbd_1+m\lbd_4+\lbd_2R^* }{\eps+\beta_1+\lbd_1+m\lbd_4+\lbd_2R^*},$$
by~\eqref{eq:PR_PC_equilib_static} we conclude that
$$P_R^*=\dfrac{\alpha\gamma}{2\eps}+\theta-\sqrt{\left(\dfrac{\alpha\gamma}{2\eps}+\theta\right)^2
-\frac{\alpha\gamma(\lbd_3+m\lbd_4)}{\eps(\lbd_2+m_R\lbd_6)}}$$
and it is easy to check that the expression above is nonnegative. Again by~\eqref{eq:PR_PC_equilib_static} we get $P_C^*=\gamma(1-\alpha)/\eps-C^*$ and it is also immediate that this expression is nonnegative.
To study the stability of the equilibrium, we consider the Jacobian matrix $J$ at the equilibrium. Namely
$$
J=\left[
\begin{array}{cccc}
-\eps-\beta_1 & \lbd_2 P_C^* & \lbd_1+m\lbd_4+\lbd_2 R^* & 0 \\
0 & -\eps-A & 0 & B \\
\beta_1 & -\lbd_2 P_C^* & -\eps-\lbd_1-m\lbd_4-\lbd_2 R^* & 0 \\
0 & A & 0 &-\eps-B\\
\end{array}
\right],
$$
where $A=\beta_2-(\lbd_2+m_R\lbd_6)P_R^*$ and $B=\lbd_3+m\lbd_4+(\lbd_2+m_R\lbd_6)R^*$. It is easy check that the eigenvalues of $J$ are $-\eps$, $-(A+B+\eps)$ and $-(\beta_1+\eps+\lbd_1+\lbd_4m-\lbd_2R^*)$. Thus, all eigenvalues have negative real part and we conclude that the equilibrium solution is locally asymptotically stable.
\subsection{Proof of Theorem~\ref{teo:mri}}
By the second equation in~\eqref{eq:modelo} we can see that, if $(R^*,C^*,P_R^*,P_C^*)$ is an equilibrium solution of our system, then
$$((\lbd_2+m_R\lbd_6)P_R^*-\eps-\beta_1)R^*=0$$
and thus $R^*=0$ or $P_R^*=(\eps+\beta)/(\lbd_2+m_R\lbd_6)$. Using the remaining equations and condition~\eqref{eq:tau} it is straightforward to obtain the equilibrium points.
In the equilibrium $(R_1^*,C_1^*,P_{R,1}^*,P_{C,1}^*)$, the Jacobian matrix $J$ is given by
$$
J=\left[
\begin{array}{cccc}
-\eps-\beta_1 & \lbd_2 (1-\alpha) \gamma/ \eps & 0 & 0\\
0 & -\eps-\beta_2+(\lbd_2+m_R\lbd_6) \gamma\alpha / \eps & 0 & 0\\
\beta_1 & -\lbd_2 (1-\alpha)\gamma / \eps & -\eps & 0\\
0 & \beta_2-(\lbd_2+m_R\lbd_6) \gamma\alpha / \eps & 0 & -\eps \\
\end{array}
\right].
$$
We can easily check that the constants $-\eps$, $-b_1-\eps$ and
$\gamma\alpha(\lbd_2+\lbd_6m_R) /\eps - (\beta_2+\eps)$
are the eigenvalues of $J$. Therefore if $\tau>1$ the equilibrium is unstable and if $\tau <1$ it is locally asymptotically stable.
In the equilibrium $(R_2^*,C_2^*,P_{R,2}^*,P_{C,2}^*)$, the Jacobian matrix $J$ has the following form
$$
J=\left[
\begin{array}{cccc}
-\eps-\beta_1 & \lbd_2 P_{C,2}^* & \lbd_2 R_2^* & 0 \\
0 & 0 & 0 & (\lbd_2+m_R\lbd_6)R_2^* \\
\beta_1 & -\lbd_2 P_{C,2}^* & -\eps-\lbd_2 R_2^* & 0\\
0 & -\eps & 0 & -\eps-(\lbd_2+m_R\lbd_6)R_2^* \\
\end{array}
\right].
$$
We can easily check that the negative constants $-\eps$, $-(\beta_1+\eps+\lambda_2R_2^*)$ and $-(\lbd_2+m_R\lbd_6)R_2^*$ are the eigenvalues of $J$. Therefore the equilibrium is locally asymptotically stable.
\subsection{Proof of Theorem~\ref{teo:no_referral_influence}}
In our case, equation~\eqref{eq:solution-R} in Theorem~\ref{teo:equilibriums} has a unique solution which is nonnegative and is given by
$$R^*=\dfrac{(\lbd_3+m\lbd_4)up+(\lbd_1+m\lbd_4)\lbd_5q}{uv-\lbd_5\lbd_7}.$$
Thus, again by Theorem~\ref{teo:equilibriums}, the possible equilibrium solutions are given by $(R^*,C^*,P_R^*,P_C^*)$ where $R^*$ is the constant above,
\[
\begin{split}
C^*
&=\dfrac{(\eps+\beta_2+\lbd_7+\lbd_3+m\lbd_4)R^*-(\lbd_3+m\lbd_4)p}{\lbd_5}\\
&=\dfrac{uvp(\lbd_3+m\lbd_4)+\lbd_5vq(\lbd_1+m\lbd_4) - uvp(\lbd_3+m\lbd_4)+\lbd_5\lbd_7p(\lbd_3+m\lbd_4)}{\lbd_5(uv-\lbd_5\lbd_7)}\\
&=\dfrac{(\lbd_1+m\lbd_4)vq + (\lbd_3+m\lbd_4)\lbd_7p}{uv-\lbd_5\lbd_7}.
\end{split}
\]
\[
P_R^*=p-R^*
=\dfrac{up(\eps+\beta_2+\lbd_7)-\lbd_5(\lbd_7p+(\lbd_1+m\lbd_4)q)}
{uv-\lbd_5\lbd_7}
\]
and
\[
P_C^*=p-C^*
=\dfrac{vq(\eps+\beta_1+\lbd_5)-\lbd_7(\lbd_5q+(\lbd_3+m\lbd_4)p)}
{uv-\lbd_5\lbd_7}.
\]
Since $\max\{\kappa_1,\kappa_2\}\le 1$, is immediate that $P_R^* \ge 0$ and $P_C^* \ge 0$. We conclude that there is a unique equilibrium solution $(R^*,C^*,P_R^*,P_C^*)$.
The Jacobian matrix $J$ at the equilibrium is given by
$$
J=\left[
\begin{array}{cccc}
-\eps-\beta_1-\lbd_5 & \lbd_7 & \lbd_1+m\lbd_4 & 0 \\
\lbd_5 & -\eps-\beta_2-\lbd_7 & 0 & \lbd_3+m\lbd_4 \\
\beta_1 & 0 & -\eps-\lbd_5-\lbd_1-m\lbd_4 & \lbd_7 \\
0 & \beta_2 & \lbd_5 & -\eps-\lbd_7-\lbd_3-m\lbd_4 \\
\end{array}
\right].
$$
We can check that the eigenvalues are all negative and given by $-\eps$, $-\eps-\lbd_5-\lbd_7$ and
$$\frac12 \left( -\sigma -2\lbd_4 m \pm \sqrt{\sigma^2-4[(\eps+\beta_1+\lbd_1+\lbd_5)(\eps+\beta_2+\lbd_3+\lbd_7)-\lbd_7\lbd_5]} \right).$$
Therefore the equilibrium is asymptotically stable. Since the system in this case is linear, the equilibrium is globally asymptotically stable.
\section{Conclusions} \label{section:Conc}
We presented and studied in this paper a compartmental model with four compartments to describe the evolution of the number of customers and potential customers of some corporation based on the marketing policy of the corporation, determined by the effort used in undifferentiated marketing campaigns and in referral directed marketing. Apparently the results obtained are reasonable in the sense that the qualitative behavior obtained is not going against common sense. The results show that the model works in theory in the several scenarios considered, with and without marketing incentives, thus relying on normal marketing policies or on incentives to referrals. The model shows that, in theory, it is possible to predict the influence referrals can have on their peers based on the incentives given to them by the company. The different scenarios also allows to see, in a specific period of time, what happens to the number of current and potential customers based on the zero incentives policy, and therefore, only based on the natural attractiveness power of the referrals (Theorem~\ref{teo:mri} with $\lambda_6=0$); what happens if the company invests in marketing but not on incentives to referrals (Theorem~\ref{teo:no_referral_influence}); and also, what happens if the company only invests on incentives to referrals (Theorem~\ref{teo:mri} with $\lambda_6 \ne 0$). Based on this results, this model allows companies to adjust their marketing policies in order to maximize a specific parameter of the model. For instance, the model allows a company to estimate the amount of investment necessary to transform an $x$ number of potential customers in customers.
This study must be followed by work where the model is used in real world problems. In fact, comparing the results given by the model and real data would of fundamental importance to confirm the usefulness of it. Naturally this is a major task whose feasibility will certainly depend on the accuracy in the estimation of parameters.
We believe that this work opens several possibilities for future studies. For instance, it would be interesting to consider versions of this model with time-varying parameters to model seasonal phenomena that may occur in some economic activities. It can also be of interest to consider age structured populations, to distinguish different consumption habits, or to subdivide the universe of referrals, to reflect different aspects that make those customers important to the corporation.
\bibliographystyle{elsart-num-sort}
|
\section{Notations : groupes et immeubles}
Pour les assertions non prouv\'ees et/ou non r\'ef\'erenc\'ees, on
renvoie le lecteur \`a la monographie de Bushnell et Kutzko
[BK], ou bien au {\S}I de [BS].
\medskip
Si $K$ est un corps localement compact et non archim\'edien, on note
${\mathfrak o}_F$ son anneau d'entiers, ${\mathfrak p}_K$ l'id\'eal maximal de ${\mathfrak o}_K$
et $k_K = {\mathfrak o}_K /{\mathfrak p}_K$ le corps r\'esiduel (fini). On fixe une
fois pour toute un tel corps $F$.
\smallskip
Soit $V$ un $F$-espace vectoriel de dimension finie $N$. On pose
$A={\rm End}_F (V)\simeq {\rm M}(N,F)$. Si $E/F$ est une extension de
corps finie de $F$ plong\'ee dans $A$, le commutant de $E$ dans $A$
est $B = {\rm End}_E (V)\simeq {\rm M}(N/[E:F], E)$. On note $G$ le
groupe ${\rm Aut}_F (V)$ et $G_E$ le groupe ${\rm Aut}_E (V)$,
centralisateur de $E^\times$ dans $G$.
On note ${\rm Her}(A)$ (resp ${\rm Her}(B)$) l'ensemble des
${\mathfrak o}_F$-ordres h\'er\'editaires dans $A$ (resp. des ${\mathfrak o}_E$-ordres
h\'er\'editaires dans $B$). Ce sont des ensembles partiellement
ordonn\'es (EPO) munis des actions par conjugaison de $G$ et $G_E$
respectivement.
On a une injection $G_E$-\'equivariante $j_{E/F}$~: ${\rm
Her}(A)\longrightarrow {\rm Her}(B)$. Elle associe \`a un ordre
h\'er\'editaire $\mathfrak B$ de $B$ attach\'e \`a une ${\mathfrak o}_E$-cha\^ine de
r\'eseaux $({\mathcal L})$ de $V$, l'ordre h\'er\'editaire $\mathfrak A =
\mathfrak A (\mathfrak B )$ associ\'e \`a $({\mathcal L})$ vue comme
${\mathfrak o}_F$-cha\^ine de r\'eseaux.
L'immeuble semisimple $X_F$ de $G$ est naturellement la r\'ealisation
g\'eom\'etrique d'un complexe simplicial de
dimension $N-1$, et de fa\c con abusive, on notera encore $X_F$ ce
complexe simplicial. C'est un espace topologique localement compact sur
lequel $G_F$ agit par automorphismes simpliciaux. Il existe une
bijection d\'ecroissante et $G_F$-\'equivariante entre l'EPO ${\rm
Her}(A)$ et l'EPO des simplexes de $X_F$, que l'on notera $\mathfrak A
\mapsto \sigma_\mathfrak A$. Elle est caract\'eris\'ee comme suit : si $\mathfrak A$
est un ordre h\'er\'editaire de $A$, $\sigma_\mathfrak A$ est l'unique
simplexe de l'immeuble dont le fixateur compact est le sous-groupe
parahorique $U(\mathfrak A
)=\mathfrak A^\times$. On a des notations et faits similaires pour le groupe
$G_E$.
\medskip
Sur les r\'ealisations g\'eom\'etriques $X_F$ et $X_E$, on a
des structures affines : le barycentre de deux points \`a
coefficients positifs est d\'efini. Le fait suivant sera un
ingr\'edient crucial de nos constructions.
\begin{theorem2} [BL] i) Il existe une
unique application $j$~: $X_E \longrightarrow X_F$ qui est $G_E$-\'equivariante
et affine.
ii) De plus les complexes de drapeaux ${\rm Flag} ({\rm Her}(A))$ et
${\rm Flag}({\rm Her}(B))$ correspondent respectivement aux premi\`eres
subdivisions barycentriques de $X_F$ et $X_E$. On montre que $j$
correspond \`a l'application ${\rm Flag}({\rm Her}(B)) \longrightarrow {\rm Flag} ({\rm
Her}(A))$ induite par $j_{E/F}$. En particulier, si $\mathfrak B\in {\rm
Her}(B)$, alors $j$ envoie l'isobarycentre de $\sigma_\mathfrak B$ sur
l'isobarycentre de $\sigma_{\mathfrak A (\mathfrak B )}$.
\end{theorem2}
Dans la suite on identifiera syst\'ematiquement le $G_E$-ensemble
$X_E$ avec son image $j(X_E)$ dans $X_F$. L'inclusion $X_E \subset
X_F$ n'est pas simpliciale en g\'en\'eral. Elle l'est si, et
seulement si, l'extension $E/F$ est non ramifi\'ee. Cependant
l'inclusion $X_E \subset X_F$ est toujours simpliciale apr\`es
passage \`a la premi\`ere subdivision barycentrique.
\medskip
Le sous-$G_F$-ensemble $X(E) = \displaystyle \bigcup_{g\in G} g.X_E$ de $X_F$
est naturellement la r\'ealisation g\'eom\'etrique d'un complexe
simplicial $X[E]$. Si $E/F$ est non ramifi\'ee, alors $X[E]$ est un
sous-complexe simplicial de $X_F$. En g\'en\'eral il faut passer
\`a la premi\`ere subdivision barycentrique pour que l'inclusion
$X(E)\subset X_F$ soit simpliciale.
\medskip
Si $\mathfrak A \in {\rm Her} (A)$, on note $U(\mathfrak A ) = \mathfrak A^\times$ le
sous-groupe parahorique qui fixe le simplexe $\sigma_\mathfrak A$ de
$X_F$. Son sous-groupe pro-unipotent est $U^1 (\mathfrak A ) =
1+\mathfrak P_{\mathfrak A}$, o\`u $\mathfrak P_\mathfrak A$ est le radical de Jacobson de
$\mathfrak A$.
Si $E/F$ est un sous-corps de $A$ et si $\mathfrak B\in {\rm Her}(B)$, $\mathfrak A
= \mathfrak A (\mathfrak B )$, alors on a :
$$
{\mathcal N}(\mathfrak A )\cap G_E = {\mathcal N}(\mathfrak B ), \ U(\mathfrak A )\cap G_E
= U(\mathfrak B ), \ U^1 (\mathfrak A )\cap G_E = U^1 (\mathfrak B ), \ {\mathcal N}(\mathfrak A )
= {\mathcal N}(\mathfrak B ) U(\mathfrak A )\ .
$$
De plus, l'action de $G$ sur $X_E$ poss\`ede la propri\'et\'e suivante :
\medskip
\begin{lemma2} ([BS] Lemma I.3.3) Si deux simplexes de $X_E$
sont conjug\'es sous l'action de $G$, ils le sont sous l'action de
$G_E$.
\end{lemma2}
\noindent En d'autres termes, $G$ n'induit pas plus d'action sur $X_E$ que $G_E$ le fait d\'ej\`a.
\section{Caract\`eres simples}
Les r\'ef\'erences pour cette section sont [BK] et
[BH]. On fixe une fois pour toute une paire simple $[0,\beta ]$
([BH](1.5)), c'est-\`a-dire une extension finie $E/F$ munie
d'un g\'en\'erateur $\beta$ (i.e. $E=F[\beta ]$), satisfaisant les conditions
suivantes :
\smallskip
[PS1] $\beta\not\in {\mathfrak o}_E$,
[PS2] $k_0 (\beta , \mathfrak A (E))<0$ (cf. [BK]{\S}1).
\smallskip
Pour chaque $E$-espace vectoriel $V$ et chaque $\mathfrak B \in {\rm
Her}(B)$, o\`u $B={\rm End}_E (V)$, on a une strate simple $[\mathfrak A (\mathfrak B
), n_\mathfrak B ,0 ,\beta ]$ dans $A={\rm End}_F (V)$, r\'ealisation de
$[0,\beta ]$ dans $A$.
Attach\'ees \`a $[\mathfrak A (\mathfrak B ), n_\mathfrak B ,0 ,\beta ]$ (donc \`a
$[0,\beta ]$, $V$, $\mathfrak B$), on a les donn\'ees suivantes :
\smallskip
$\bullet$ Deux sous-groupes ouverts compacts de $G={\rm Aut}_F (V)$ :
$H^1 (\mathfrak B ) \subset J^1 (\mathfrak B )\subset U^1 (\mathfrak A (\mathfrak B ))$, tous deux
normalis\'es par ${\mathcal N}(\mathfrak B )$.
\smallskip
$\bullet$ Un ensemble fini de caract\`eres simples ${\rm \mathcal
C}(\mathfrak B ) = {\mathcal C}(\mathfrak A (\mathfrak B ), 0,\beta )$ de $H^1 (\mathfrak B )$,
qui ont chacun un $G$-entrelacement donn\'e par $J^1 (\mathfrak B ) G_E J^1
(\mathfrak B )$.
\smallskip
Rappelons que si $\theta \in {\mathcal C}(\mathfrak B )$, il existe \`a
isomorphisme pr\`es une unique rep\'esentation irr\'eductible $\eta =
\eta (\theta )$ de $J^1 (\mathfrak B )$ qui contient $\theta$ par
restriction (la repr\'esentation de Heisenberg de $\theta$).
\smallskip
La paire simple et le $E$-espace vectoriel $V$ \'etant fix\'es, on a
des bijections canoniques :
$$
\tau_{\mathfrak B_1 ,\mathfrak B_2}~: \ {\mathcal C}(\mathfrak B_1 )\longrightarrow {\mathcal C}
(\mathfrak B_2 )\ , \ \mathfrak B_1 ,\ \mathfrak B_2 \in {\rm Her}(B),
$$
\noindent appel\'ees {\it applications de transfert}
([BK](3.6)). Gr\^ace \`a ces applications, si l'on fixe une paire
simple $[0,\beta ]$, un $E$-espace vectoriel $V$, un ordre $\mathfrak B_0$
dans $B={\rm End}_E (V)$ et un caract\`ere simple $\theta_0\in
{\mathcal C}(\mathfrak B_0 )$, on obtient une famille de caract\`eres
simples $(H^1 (\mathfrak B ), \theta (\mathfrak B ))_{\mathfrak B \in {\rm Her}(B)}$ en
posant $\theta (\mathfrak B ) = \tau_{\mathfrak B_0 ,\mathfrak B}(\theta_0 )$. Il lui
est associ\'e une famille de repr\'esentations de Heisenberg
$(H^1 (\mathfrak B ), \theta (\mathfrak B ))_{\mathfrak B \in {\rm Her}(B)}$ d\'efinies
\`a isorphisme pr\`es. Il est facile de v\'erifier que ces deux
familles sont $G_E$-\'equivariantes en un sens \'evident.
Pour chaque paire d'ordres h\'er\'editaires $\mathfrak B_1 \subset \mathfrak B_2$
dans ${\rm Her}(B)$, on peut former le groupe $J^1 (\mathfrak B_1 ,\mathfrak B_2 )
= U^1 (\mathfrak B_1 )J^1 (\mathfrak B_2 )$.
\begin{proposition2} ([BK] (5.1.14-16), (5.1.18), (5.1.19))
Fixons une paire simple $[0,\beta ]$, un $E$-espace vectoriel $V$,
un ordre $\mathfrak B_0\in {\rm Her}(B)$ et un caract\`ere simple
$\theta_0\in {\mathcal C}(\mathfrak B_0 )$. Alors il existe une unique
famille de repr\'esentations $\displaystyle\left( J^1 (\mathfrak B_1 ,\mathfrak B_2 ),\eta
(\mathfrak B_1 ,\mathfrak B_2 )\right)_{\mathfrak B_1 \subset \mathfrak B_2}$ (d\'efinies \`a
isomorphismes pr\`es) qui \'etend la famille $(J^1 (\mathfrak B ), \eta
(\mathfrak B ))_{\mathfrak B}$ au sens suivant :
\smallskip
(i) $\eta (\mathfrak B ,\mathfrak B ) =\eta (\mathfrak B )$, $\mathfrak B\in {\rm Her}(B)$ ;
(ii) $\eta (\mathfrak B_1 ,\mathfrak B_2 )_{\vert J^1 (\mathfrak B_2 )} \simeq \eta (\mathfrak B_1
)$, $\mathfrak B_1 \subset \mathfrak B_2 \in {\rm Her}(B)$ ;
(iii) les induites suivantes sont irr\'eductibles et \'equivalentes :
$$
{\rm Ind}_{J^1 (\mathfrak B_1 )}^{U^1 (\mathfrak A_1 )}\, \eta (\mathfrak B_1 )\simeq {\rm
Ind}_{J^1 (\mathfrak B_1 , \mathfrak B_2 )}^{U^1 (\mathfrak A_1 )}\, \eta (\mathfrak B_1 ,\mathfrak B_2
), \ \mathfrak B_1 \subset \mathfrak B_2 \in {\rm Her}(B)\ .
$$
De plus on a la relation de compatibilit\'e :
$$
\eta (\mathfrak B_1 ,\mathfrak B_2 )_{\vert J^1 (\mathfrak B_2 ,\mathfrak B_3)} \simeq \eta
(\mathfrak B_2 ,\mathfrak B_3 ), \ \mathfrak B_1 \subset \mathfrak B_2 \subset \mathfrak B_3 \in {\rm
Her}(B)\ .
$$
\end{proposition2}
\noindent Il est facile de v\'erifier que la famille $\displaystyle\left( J^1 (\mathfrak B_1 ,\mathfrak B_2 ),\eta
(\mathfrak B_1 ,\mathfrak B_2 )\right)_{\mathfrak B_1 \subset \mathfrak B_2}$ de
repr\'esentations de Heisenberg est $G_E$-\'equivariante en un sens
\'evident.
\section{Repr\'esentations de la s\'erie discr\`ete et types simples}
Sauf si \c ca n'est pas dit express\'ement, on utilisera les
notations de [BK].
\medskip
On fixe une fois pour toute un type simple $(J,\lambda )$ au sens de
[BK](5.5.10). On supposera de plus que l'on est en niveau $>0$.
Concr\`etement cela signifie la chose suivante.
\smallskip
Il existe une paire simple $[0,\beta ]$, un $E$-espace vectoriel $V$
de dimension finie, o\`u $E=F[\beta ]$ et un ordre principal $\mathfrak B_0$
dans $B ={\rm End}_E (V)$. La repr\'esentation $\lambda$ du groupe $J
= J(\mathfrak B_0 ) = J^1 (\mathfrak B_0 )U (\mathfrak B_0 )$ est de la forme $\kappa_0
\otimes \rho$, o\`u $\kappa_0$ est une $\beta$-extension d'un
caract\`ere simple $\theta_0\in {\mathcal C}(\mathfrak B_0 )$
([BK](5.2.1)), et $\rho$ est l'inflation d'un repr\'esentation
irr\'eductible cuspidale de $J/J^1 (\mathfrak B_0 )$ de la forme
suivante. Rappelons que le quotient $J/J^1 (\mathfrak B_0 )$ s'identifie
\`a ${\rm GL}(n/e, k_E )^{\times e}$, o\`u $n:={\rm dim}_E (V)$, $e$
est la p\'eriode de l'ordre $\mathfrak B_0$, et $k_E$ d\'esigne le corps
r\'esiduel de $E$. Alors la condition sur $\rho$ et que, comme
repr\'esentation de ${\rm GL}(n/e, k_E )^{\times e}$, elle est de
la forme $\rho_0^{\otimes e}$, o\`u $\rho_0$ est une
repr\'esentation (irr\'eductible, cuspidale) de ${\rm GL}(n/e,k_E
)$.
\medskip
Les donn\'ees d\'ecrivant le type simple $(J,\lambda )$ de $G = {\rm
Aut}_F (V)$ ne sont pas uniques. C'est pour cette raison que {\it nous
fixons une fois pour toute} :
\smallskip
-- une paire simple $[0,\beta ]$,
-- un $E$-espace vectoriel $V$
-- un ordre principal $\mathfrak B_0$ dans $B= {\rm End}_E (V)$,
-- un caract\`ere simple $\theta_0 \in {\mathcal C}(\mathfrak B_0 )$,
-- une $\beta$-extension $\kappa_0$ de $\theta_0$,
-- une repr\'esentation cuspidale irr\'eductible $\rho_0$ de ${\rm
GL}(n/e,k_E )$.
\smallskip
De plus d'apr\`es la Proposition 2.1, ces donn\'ees donnent lieu \`a
une famille $G_E$-\'equivariante de repr\'esentations de Heisenberg
$\displaystyle\left( J^1 (\mathfrak B_1 ,\mathfrak B_2 ),\eta (\mathfrak B_1 ,\mathfrak B_2
)\right)_{\mathfrak B_1 \subset \mathfrak B_2}$.
\smallskip
Fixons une extension non ramifi\'ee $L/E$ contenue dans $B$, v\'erifiant
$[L:F]=n/e$, telle que le groupe multiplicatif $L^\times$ normalise ${\mathfrak B}_0$.
On pose ${\rm End}_L \, V \simeq {\rm M}(e,L)$ et
$G_L = {\rm Aut}_L \, V$. On a une application canonique ${\rm Her}(C)\longrightarrow
{\rm Her}(B)$ ainsi qu'une inclusion simpliciale et $G_L$-\'equivariante
$X_L\subseteq X_E$. Notons que l'unique ordre ${\mathfrak C }_0 \in {\rm Her}(C)$
v\'erifiant ${\mathfrak B }_0 ={\mathfrak B }({\mathfrak C }_0 )$ est un ordre minimal
(ou {\it ordre d'Iwahori}) ; par contre ${\mathfrak B}_0$ n'est pas minimal en
g\'en\'eral. On note $X[L] = \displaystyle \bigcup_{g\in G} gX_L$ que l'on munit de la structure
simpliciale naturelle $G$-invariant qui prolonge celle de $X_L$.
\medskip
On fixe un ordre maximal ${\mathfrak C }_{\rm max}\supseteq {\mathfrak C}_0$ et on pose
${\mathfrak B}_{\rm max} = {\mathfrak B}( {\mathfrak C}_{\rm max})$ ; cet ordre est
toujours maximal. D'apr\`es [BK](5.2.2-5), à isomorphisme pr\`es, il existe une
unique $\beta$-extension $\kappa_{\rm max}$ de $\eta ({\mathfrak B}_{\rm max} )$ telle
que
$$
{\rm Ind}_{J({\mathfrak B}_0 )}^{U({\mathfrak B}_0 )U^1 ({\mathfrak A}({\mathfrak B}_0 )}
\kappa_0 \simeq {\rm Ind}_{U({\mathfrak B}_0)J^1 ({\mathfrak B}_{\rm
max})}^{U({\mathfrak B}_0 ) U^1 ({\mathfrak A}({\mathfrak B}_0 ))} \kappa_{\rm max}\ .
$$
\noindent On peut alors former la repr\'esentation $\lambda_{\rm max} =
\kappa_{\rm max} \otimes \rho$ de $J_{\rm max} := U(\mathfrak B_0 )J^1 (\mathfrak B_{\rm max})$ ; elle
est irr\'eductible.
\begin{theorem2} a) La paire $(J_{\rm max},\lambda_{\rm max})$ est un type
de $G$ qui d\'efinit la m\^eme composante de Bernstein
${\mathcal R}_{(J,\lambda )}(G)$ que $(J,\lambda )$.
\noindent b) Soit $(\pi ,{\mathcal V})$ un objet de ${\mathcal R}_{(J,\lambda )}(G)$. On a
${\mathcal V}^{\lambda_{\rm max}} = {\mathcal V}^{\eta (\mathfrak B_0 ,\mathfrak B_{\rm max})}$.
\noindent c) Soit $(\pi ,{\mathcal V})$ une repr\'esentation irr\'eductible essentiellement
de carr\'e int\'egrable modulo le centre, objet de ${\mathcal R}_{(J,\lambda )}(G)$.
Alors $(\pi ,{\mathcal V})$ contient la paire $(J_{\rm max} ,\lambda_{\rm max})$
(resp. la paire $(J,\lambda )$) avec multiplicit\'e $1$.
\end{theorem2}
\section{Un syst\`eme de coefficients}
Un syst\`eme de coefficients \'equivariant sur le $G$-complexe simplicial
$X[L]$ est la donn\'ee de $(({\mathcal V}_{\sigma})_\sigma , (r_\tau^\sigma )_{\tau \subset
\sigma}, (\varphi_{g,\sigma})_{g,\sigma})$, o\`u :
\medskip
-- pour chaque simplexe $\sigma$ de $X[L]$, $\mathcal V_\sigma$ est un $\mathbb C$-espace vectoriel,
-- pour $\tau\subseteq \sigma$, $r_\tau^\sigma\in {\rm Hom}_\mathbb C (\mathcal V_\sigma ,
\mathcal V_{\tau})$.
-- pour $\sigma$ simplexe, $g\in G$, $\varphi_{g,\sigma}\in
{\rm Hom}_\mathbb C (\mathcal V_\sigma , \mathcal V_{g\sigma})$.
-- pour chaque simplexe $\sigma$, $r_\sigma^\sigma =\varphi_{1,\sigma}={\rm id}_{\mathcal V_\sigma}$,
-- tous les diagrammes que l'on peut imaginer commutent,
-- pour tout $\sigma$ la repr\'esentation de $G_\sigma$ dans $\mathcal V_\sigma$ induite
par le syst\`eme de coefficients est lisse.
\medskip
Soit \`a pr\'esent $(\pi ,\mathcal V )$ une repr\'esentation lisse de $G$. Soit
$\mathfrak C \in {\rm Her}(C)$ tel que $\mathfrak C_{\rm min}\subseteq \mathfrak C \subseteq \mathfrak C_{\rm max}$, et
soit $\sigma_\mathfrak C$ le simplexe de $X_L$ attach\'e \`a $\mathfrak C$. On pose alors
$$
\mathcal V_{\sigma_\mathfrak C} = \sum_{g\in U(\mathfrak A )/U(\mathfrak B )J^1 (\mathfrak B_{\rm max})}
\pi (g)\mathcal V^{\eta (\mathfrak B ,\mathfrak B_{\rm max})}
$$
\noindent o\`u $\mathfrak B =\mathfrak B (\mathfrak C )$ et $\mathfrak A =\mathfrak A (\mathfrak B )$.
Si $\sigma$ est un simplexe quelconque de $X[L]$, on peut toujours l'\'ecrire
$\sigma = g.\sigma_{\mathfrak C}$, pour un $g\in G$ et un $\mathfrak C\in {\rm Her}(C)$ tel que
$\mathfrak C_{\rm min}\subseteq \mathfrak C \subseteq \mathfrak C_{\rm max}$, et on pose $\mathcal V_\sigma =
\pi (g) \mathcal V_{\sigma_\mathfrak C}$.
\begin{theorem2} a) Si $\sigma$ est un simplexe de $X[L]$, $\mathcal V_\sigma$ est bien d\'efini,
c'est-\`a-dire ne d\'epend d'aucun choix.
\noindent b) Si $\tau\subseteq \sigma$ sont des simplexes de $X[L]$, on a
$\mathcal V_\sigma\subseteq \mathcal V_\tau$.
\noindent c) Si $\sigma$ est un simplexe de $X[L]$ et si $g\in G$, alors $\mathcal V_{g\sigma}
=\pi (g)\mathcal V_\sigma$.
\end{theorem2}
On peut donc d\'efinir un syst\`eme de coefficients ${\mathcal C}(\pi ) =
((\mathcal V_\sigma )_\sigma , (r_\tau^\sigma )_{\tau \subseteq \sigma}, (\varphi_{g,\sigma})_{g,\sigma})$
sur $X[L]$, en d\'efinissant $r_\tau^\sigma$ comme \'etant l'inclusion $\mathcal V_\sigma \subseteq \mathcal V_\tau$, et
$\varphi_{g,\sigma}$ comme \'etant l'application $\mathcal V_\sigma \longrightarrow \mathcal V_{g\sigma}$ induite par $\pi (g)$.
\begin{theorem2} Supposons que $(\pi ,\mathcal V )\in {\mathcal R}_{(J,\lambda )}(G)$. Alors le complexe $X[L]$
et le syst\`eme de coefficients ${\mathcal C}(\pi )$ ne d\'ependent que de l'endo-classe $\Theta$
du caract\`ere simple $\theta_0$, et donc d'aucun autre choix fait dans la construction. On notera
${\mathcal C}_\Theta (\pi )$ ce syst\`eme de coefficients canoniquement attach\'e \`a $\pi$.
\end{theorem2}
Ce syst\`eme de coefficients peut se calculer presque enti\`erement si $(\pi ,\mathcal V)$ est irr\'eductible
et essentiellement de carr\'e int\'egrable, ce que nous supposerons jusqu'\`a la fin de cette section.
\medskip
Soit $\mathfrak C\in {\rm Her}(C)$ tel que $\mathfrak C_{\rm min}\subseteq \mathfrak C \subseteq
\mathfrak C_{\rm max}$. On pose comme d'habitude $\mathfrak B =\mathfrak B (\mathfrak C )$ et
$\mathfrak A =\mathfrak A (\mathfrak B )$. Le quotient ${\mathbb G}_\mathfrak B =U(\mathfrak B )/U^1 (\mathfrak B )$
est un produit de groupes lin\'eaires g\'en\'eraux sur $k_E$, le corps
r\'esiduel de $E$. Il poss\`ede ${\mathbb P}_{\mathfrak B_0}=U(\mathfrak B_0)
/U^1 (\mathfrak B )$ comme sous-groupe parabolique. Ce dernier admet comme
radical unipotent ${\mathbb U}_{\mathfrak B_0}=U^1 (\mathfrak B_0)/U^1 (\mathfrak B )$, et
comme facteur de Levi
$$
{\mathbb L}_{\mathfrak B_0} = {\mathbb P}_{\mathfrak B_0} / {\mathbb U}_{\mathfrak B_0}
= U(\mathfrak B_0)/U^1 (\mathfrak B_0) \simeq {\rm GL}(n/e,k_E )^{\times e}\ .
$$
\noindent Soit ${\rm St} (\mathfrak B ,\rho )$ la repr\'esentation de Steinberg
g\'en\'eralis\'ee de ${\mathbb G}_\mathfrak B$ de support cuspidal
$({\mathbb L}_{\mathfrak B_0}, \rho )$ : c'est l'unique sous-repr\'esentation
irr\'eductible g\'en\'erique de l'induite parabolique
${\rm ind}_{{\mathbb P}_{\mathfrak B_0}}^{{\mathbb G}_\mathfrak B}\, \rho$. On peut alors
former la repr\'esentation $\kappa_{\rm max}\otimes {\rm St}(\mathfrak B ,\rho )$
de $U(\mathfrak B )J^1 (\mathfrak B_{\rm max})$. On montre que cette repr\'esentation est irr\'eductible.
\begin{proposition2} a) On a $\mathcal V^{\eta (\mathfrak B ,\mathfrak B_{\rm max})} \simeq \kappa_{\rm max}\otimes
{\rm St}(\mathfrak B ,\rho )$ comme $U(\mathfrak B )J^1 (\mathfrak B_{\rm max})$-modules.
\noindent b) La repr\'esentation induite :
$$
\lambda (\mathfrak A ):= {\rm ind}_{U(\mathfrak B )J^1 (\mathfrak B_{\rm max})}^{U(\mathfrak A )} \, \left( \kappa_{\rm max}\otimes
{\rm St}(\mathfrak B ,\rho )\right)
$$
\noindent est irr\'eductible.
\noindent c) Soit $(\pi ,\mathcal V )\in {\mathcal R}_{(J,\lambda )}(G)$
une repr\'esentation irr\'eductible essentiellement de carr\'e int\'egrable.
Si ${\mathcal C}_\Theta (\pi ) =
((\mathcal V_\sigma )_\sigma , (r_\tau^\sigma )_{\tau \subseteq \sigma}, (\varphi_{g,\sigma})_{g,\sigma})$
est le syst\`eme de coefficients attach\'e \`a $\pi$, alors
$$
\mathcal V_{\sigma_\mathfrak C} = \mathcal V^{\lambda (\mathfrak A )} \simeq \lambda (\mathfrak A )
$$
\noindent o\`u l'isomorphisme est un isomorphisme de $U(\mathfrak A )$-modules.
\end{proposition2}
\begin{remark2} Il est tr\`es important de noter que la donn\'ee du type $(J,\lambda )$ de $\pi$ ne permet pas
de conna\^itre enti\`erement le syst\`eme de coefficients ${\mathcal C}_\Theta (\pi )$, mais seulement les
espaces $\mathcal V_{\sigma_\mathfrak C }$ comme $U(\mathfrak A )$-modules. Le stabilisateur $G_{\sigma_\mathfrak C}$ de $\sigma_\mathfrak C$ dans
$G$ a la forme d'un produit semi-direct $\langle \Pi_\mathfrak A \rangle \ltimes U(\mathfrak A )$, pour un certain
\'el\'ement $\Pi_\mathfrak A$ de $G$ qui normalise $\mathfrak A$. Il faudrait alors conna\^itre l'action de $\Pi_\mathfrak A$
sur $\mathcal V^{\lambda (\mathfrak A )}$. Ce probl\`eme devrait \^etre r\'esolu en se donnant $\pi$ par un type
simple \'etendu. Jusqu'\`a pr\'esent, les types simples \'etendus ne sont construits que pour les
repr\'esentations supercuspidales (cf. [BK]{\S}6) et les repr\'esentations de la s\'erie discr\`ete de
niveau $0$ ([BH2], [SZ]).
\end{remark2}
\section{R\'esolutions et pseudo-coefficients}
En gardant les notations du {\S}3, on fixe un type simple $(J,\lambda
)$ ainsi qu'une repr\'esentation $(\pi ,\mathcal V )\in {\mathcal
R}_{(J,\lambda )}$. On lui a associ\'e un $G$-complexe simplicial
$X_\pi = X[L]$ muni d'un syst\`eme de coefficients $G$-\'equivariant
${\mathcal C}_\Theta (\pi )$. On peut alors former, pour $q= 0,
...,d-1$, $d={\rm dim}(X_\pi )$, les espaces $C_{q}^{\rm or}(X_\pi
,{\mathcal C}_\Theta (\pi ))$ de cha\^ines orient\'ees
de $X_\pi$ \`a coefficients dans ${\mathcal C}_\Theta (\pi )$ ; ce
sont des $G$-modules lisses. Ils sont munis d'applications bords
$G$-\'equivariantes et forment ainsi un complexe de cha\^ines
naturellement augment\'e sur la repr\'esentation $\mathcal V$ :
\begin{equation}
C_{d}^{\rm or}(X_\pi
,{\mathcal C}_\Theta (\pi ))
\stackrel{\partial}{\longrightarrow} C_{d-1}^{\rm or}(X_\pi
,{\mathcal C}_\Theta (\pi ))
\stackrel{\partial}{\longrightarrow}
\cdots
\stackrel{\partial}{\longrightarrow}
C_{0}^{\rm or}(X_\pi ,{\mathcal C}_\Theta (\pi ))
\stackrel{\epsilon}{\longrightarrow} \mathcal V
\longrightarrow 0
\end{equation}
\begin{proposition2} Le complexe augment\'e (1) est un complexe dans
la cat\'egorie ${\mathcal R}_{(J,\lambda )}(G)$.
\end{proposition2}
\noindent {\bf Conjecture 1} {\it Le complexe augment\'e (1) est une r\'esolution de
$\mathcal V$ dans la cat\'egorie ${\mathcal R}_{(J,\lambda )}(G)$.}
\medskip
Cette conjecture est une cons\'equence de la conjecture suivante :
\medskip
\noindent {\bf Conjecture 2}. {\it Consid\'erons le complexe obtenu de (1) en
appliquant le foncteur (l'\'equivalence de cat\'egories)
${\mathcal R}_{(J,\lambda )}(G) \longrightarrow {\mathcal
H}_{\rm Spher}(G,\lambda_{\max})-{\rm Mod}$, ${\mathcal W}\mapsto
{\mathcal W}^{\lambda_{\rm max}}$ :
\begin{equation}
C_{d}^{\rm or}(X_\pi
,{\mathcal C}_\Theta (\pi ))^{\lambda_{\rm max}}
\stackrel{\partial}{\longrightarrow} C_{d-1}^{\rm or}(X_\pi
,{\mathcal C}_\Theta (\pi ))^{\lambda_{\rm max}}
\stackrel{\partial}{\longrightarrow}
\cdots
\stackrel{\partial}{\longrightarrow}
C_{0}^{\rm or}(X_\pi ,{\mathcal C}_\Theta (\pi ))^{\lambda_{\rm max}}
\stackrel{\epsilon}{\longrightarrow} \mathcal V^{\lambda_{\rm max}}
\longrightarrow 0
\end{equation}
\noindent Alors le complexe de cha\^ines augment\'e (2) est canoniquement
isomorphe au complexe de cha\^ines d'un appartement standard ${\mathcal
A}_L$ de $X_L$ \`a coefficients constants dans $\mathcal V^{\lambda_{\rm
max}}$. Puisque ${\mathcal A}_L$ est contractile, ce dernier complexe est
exact.}
\medskip
Dans [BS], nous d\'emontrons la Conjecture 2, et donc la Conjecture 1 dans le cas
particulier suivant.
\begin{theorem2} {\it Avec les notations pr\'ec\'edentes, supposons que
$(\pi ,\mathcal V )$ est irr\'eductible et essentiellement de carr\'e
int\'egrable. Alors la conjecture 2 est vraie.}
\end{theorem2}
Nous supposons dor\'enavant que la repr\'esentation $(\pi ,\mathcal V )$ est
irr\'eductible et essentiellement de carr\'e int\'egrable.
\medskip
Soit $Z$ le centre de $G$. Fixons une mesure de Haar $\mu_{G/Z}$ sur
$G/Z$.
\smallskip
Pour chaque
simplexe $\sigma$ de $X_\pi$, on note $G_\sigma$ le stabilisateur
global de $\sigma$ dans $G$, et $\lambda_\sigma$ la repr\'esentation
irr\'eductible de $G_\sigma$ dans $\mathcal V_\sigma $ induite par le syst\`eme de
coefficients ${\mathcal C}_\Theta (\pi )$ ; on note $\tau^\mathcal V_\sigma$
le caract\`ere de $\lambda_\sigma$, que l'on \'etend par $0$ \`a $G$ en une
fonction localement constante \`a support compact modulo le centre ;
on note $\epsilon_\sigma~: \ G_\sigma \longrightarrow \{ \pm 1\}$ le caract\`ere
de $G_\sigma$ d\'efini comme suit : si $g\in G_\sigma$,
$\epsilon_\sigma (g)$ est la signature de la permutation des sommets
de $\sigma$ induite par $g$ ; enfin on \'etend la fonction
$\epsilon_\sigma$ \`a $G$ en une
fonction localement constante \`a support compact modulo le centre.
Pour $q=0,...,d={\rm dim}\, X_\pi$, fixons un ensemble ${\mathcal
F}_q$ de repr\'esentants des orbites de $G$ dans les $q$-simplexes
de $X_\pi$. A la suite de Kottwitz [Ko] et Schneider et Stuhler
[SS], on attache \`a $(\pi ,\mathcal V )$ une fonction dite {\it d'Euler-Poincar\'e}
par la formule :
$$
f_{\rm EP}^\mathcal V := \sum_{q=0}^d \sum_{\sigma \in {\mathcal F}_q} (-1)^q
\mu_{G/Z}(G_\sigma /Z )^{-1}\, {\bar
\tau}_\sigma^\mathcal V\ \epsilon_\sigma\ .
$$
En suivant une id\'ee originale de Kottwitz, reprise par Schneider
et Stuhler, on d\'emontre :
\begin{theorem2} La fonction $f_{\rm EP}^\mathcal V$ est un
pseudo-coefficient de $(\pi ,\mathcal V )$.
\end{theorem2}
\section{La formule de caract\`ere}
On fixe une repr\'esentation $(\pi ,\mathcal V )$ de $G$ suppos\'ee
irr\'eductibe et essentiellement de carr\'e int\'egrable. On note
$\chi_\pi$ son caract\`ere d'Harish-Chandra. On garde
les notations des sections pr\'ec\'edentes. Nous aurons besoin du
r\'esultat suivant (d\^u \`a Kazdhan [Ka] dans le cas d'un corps de base
de caract\'eristique $0$ et d'un groupe r\'eductif \`a centre
compact, et \`a Badulescu [Ba] dans le cas de notre groupe).
\begin{theorem2} Soit $f_\pi$ un pseudo-coefficient et $\gamma\in G$ un
\'el\'ement elliptique r\'egulier. Alors $\chi_\pi (\gamma )$ est donn\'e
par l'int\'egrale orbitale convergente :
$$
\chi_\pi (\gamma ) = \int_{G/Z} f_{\pi} (g^{-1}\gamma^{-1}g)d\mu_{G/Z}
(\dot{g})\ .
$$
\end{theorem2}
L'application de ce r\'esultat au pseudo-coefficient $f_{\rm EP}^\mathcal V$
donne une formule pour la valeur de
$\chi_\pi$ en un \'el\'ement elliptique r\'egulier $\gamma$. Elle
s'exprime de fa\c con \'el\'egante dans le cadre g\'eom\'etrique
suivant.
Soit $\vert X_\pi \vert$ la r\'ealisation g\'eom\'etrique standard de
$X_\pi$ ; c'est un espace topologique localement compact muni d'une
action de $G$. L'ensemble de points fixes $\vert X_\pi \vert^\gamma$
est compact. Il est muni de la structure simpliciale naturelle suivante : ses
simplexes sont les intersections non vides $\sigma (\gamma ):=\sigma
\cap \vert X_\pi \vert^\gamma$, pour un simplexe $\sigma$ de $X_\pi$
globalement fixe par $\gamma$. En fait $\sigma (\gamma )$ d\'etermine
enti\`erement $\sigma$.
Avec ces notations, on a alors la formule de caract\`ere suivante.
\begin{theorem2} La valeur du caract\`ere d'Harish-Chandra de $\pi$ en
un \'el\'ement elliptique r\'egulier $\gamma$ est donn\'ee par
$$
\chi_\pi (\gamma )=\sum_{q=0}^{{\rm dim}\vert X_\pi\vert^\gamma}
\sum_{\sigma (\gamma )\in \vert X_\pi\vert^\gamma_q} (-1)^q \ {\rm
Tr}(\gamma ,\lambda_\sigma )\ ,
$$
\noindent o\`u $\vert X_\pi\vert^\gamma_q$ d\'esigne l'ensemble des
$q$-simplexes de $\vert X_\pi\vert^\gamma$.
\end{theorem2}
Cette derni\`ere formule ne peut en aucun cas être consid\'er\'ee
comme {\it effective} en g\'en\'eral. En effet pour un \'el\'ement
$\gamma$ elliptique r\'egulier quelconque, il n'existe aucune
description connue de l'ensemble de points fixes $\vert
X_\pi\vert^\gamma$ (ni de l'ensemble $\vert X \vert^\gamma$
d'ailleurs). Il y a un cas cependant o\`u la formule se simplifie de
fa\c con raisonnable, c'est celui o\`u $\gamma$ est minimal sur $F$
au sens de [BK](1.4.14).
\begin{lemma2} Supposons $\gamma$ elliptique r\'egulier et minimal
sur $F$. Alors $\vert X\vert^\gamma = \vert X\vert^{K^\times}$, o\`u
$K=F[\gamma ]$. En particulier $\vert X\vert^\gamma$ se r\'eduit \`a
un point, image canonique de l'immeuble $X_K$ dans $X$. Dans un
autre langage, $ \vert X\vert^\gamma =\{ x_\gamma \}$, o\`u
$x_\gamma$ est l'isobarycentre du simplexe attach\'e \`a l'unique
ordre h\'er\'editaire $\mathfrak A_\gamma$ de $A$ normalis\'e par
$K^\times$ (ou de fa\c con \'equivalente par $\gamma$).
\end{lemma2}
On en d\'eduit une {\it formule simple de caract\`ere} :
\begin{theorem2} Soit $\gamma$ un \'el\'ement elliptique r\'egulier de
$G$, minimal sur $F$. Soit $\mathfrak A_\gamma$ l'unique ordre
h\'er\'editaire de $A$ normalis\'e par $F[\gamma ]^\times$, et soit
$x_\gamma$ l'isobarycentre du simplexe de $X$ correspondant \`a
$\mathfrak A_\gamma$. Si $x_\gamma \in \vert X_\pi \vert$, soit $\sigma_\gamma$
l'unique simplexe de $X_\pi$ dont l'int\'erieur contient
$x_\gamma$. Alors :
$$
\chi_\pi (\gamma )=\left\{
\begin{array}{ll}
{\rm Tr}(\gamma ,\lambda_{\sigma_\gamma}) & {\rm si}\ f(L/F)\vert
f(F[\gamma ]/F)\ {\rm et}\ e(L/F)\vert
e(F[\gamma ]/F),\\
0 & {\rm sinon.}
\end{array}
\right.
$$
\end{theorem2}
Des cas particuliers de cette derni\`ere formule furent obtenus dans
[Br].
\bigskip
\centerline{\bf R\'ef\'erences}
\bigskip
[Ba] A.I. Badulescu, {\it Un r\'esultat de transfert et un
r\'esultat d'int\'egrabilit\'e locale des caract\`eres en caract\'eristique
non nulle}, J. Reine Angew. Math. 565 (2003), 101--124.
[Br] P. Broussous, {\it Transfert du pseudo-coefficient de
Kottwitz et formules de caractère pour la série discrète de GL(N)
sur un corps local}, Canad. J. Math. 66 (2014), no. 2, 241–283.
[BH] C.J. Bushnell et G. Henniart, {\it Local tame lifting
for GL(N). I. Simple characters}, Inst. Hautes Études
Sci. Publ. Math. No. 83 (1996), 105–233.
[BH2] C.J. Bushnell et G. Henniart, {\it Explicit
functorial correspondences for level zero representations of $p$-adic
linear groups}, J. Number Theory 131 (2011), no. 2, 309--331.
[BK] C.J. Bushnell et P.C. Kutzko, {\it Smooth
representations of reductive $p$-adic groups: structure theory via
types}, Proc. London Math. Soc. (3) 77 (1998), 582--634.
[BL] P. Broussous et B. Lemaire, {\it Building of ${\rm
GL}(m,D)$ and centralizers}, Transform. Groups 7 (2002), no. 1,
15–50.
[BS] P. Broussous et P. Schneider, {\it Simple characters
and coefficient systems on the building}, arXiv:1402.2501, 2014.
[Ka] D. Kzhdan, {\it Cuspidal geometry of p-adic groups},
J. Analyse Math. 47 (1986), 1–36.
[Ko] R.E. Kottwitz, {\it Tamagawa numbers}, Ann. of
Math. (2) 127 (1988), no. 3, 629--646.
[SS] P. Schneider et U. Stuhler, {\it Representation
theory and sheaves on the Bruhat-Tits building},
Inst. Hautes Études Sci. Publ. Math. No. 85 (1997), 97--191.
[SZ] A.J. Silberger et E.-W. Zink,{\it An explicit
matching theorem for level zero discrete series of unit groups of
$p$-adic simple algebras}, J. Reine Angew. Math. 585 (2005), 173--235.
\end{document}
|
\section{Introduction}
\label{sec:intro}
We study experimentally the orientational dynamics of neutrally buoyant non-axi\-sym\-me\-tric particles suspended in a viscous shear flow.
\obs{The rotation of axisymmetric particles in a shear flow has been studied in several experiments (we give a brief account below).
The case of non-axi\-sym\-me\-tric particles, by contrast, has received little attention experimentally. This is surprising because it is known in theory\cite{Hinch1979,Yarin1997} that the orientational dynamics can be very sensitive to small deviations from the axisymmetric limit. How sensitively the dynamics is affected by slight breaking of axisymmetry depends upon the orientational trajectory which in turn is determined by the initial orientation.}
\obs{In order to verify this theoretical prediction experimentally it is necessary to use particles with well-defined shapes, to make sure that inertial effects and thermal noise are negligible, and to compare different orientational trajectories of the same particle.}
\obs{In this paper we describe experimental observations of the orientational motion of micron-sized glass particles suspended in a pressure-driven micro-channel flow. The particles have different shapes: axisymmetric, slightly non-axisymmetric, and substantially non-axisymmetric (the latter are
strongly triaxial particles made for the purpose of this experiment by joining two micron-sized glass rods). We verify that inertial and thermal torques have negligble effects by showing that the orientational dynamics is invariant under reversal of the pressure.
An optical trap allows us to manipulate the same particle into different initial orientations, to study different orientational
trajectories of the same particle. The results of our measurements confirm the predictions of Refs.~\citenum{Hinch1979,Yarin1997}, show
the sensitive dependence of the orientational dynamics upon particle shape and initial orientation, and qualitatively illustrate the different types of orientational dynamics computed in Refs.~\citenum{Hinch1979,Yarin1997}.
We have not attempted to quantititatively compare individual experimental
particle trajectories with theory,
to compare individual trajectories
requires to compute the resistance tensors for the actual shape of the particles used in the experiments
(the theory of Refs. \citenum{Hinch1979,Yarin1997} is formulated for ellipsoidal particles). This is beyond the scope of the present paper. Besides the resulting trajectories are of little general interest, they depend very
sensitively on the particular shape of the particle in question.
}
\obs{The remainder of this Introduction briefly introduces the relevant theory as well as the wider context of this work.}
The orientational dynamics of a particle suspended in a viscous flow is determined by resistance tensors that relate the local flow velocity and its gradients to the torque acting on the particle\cite{Happel1965,Kim1991}.
\obs{A given particle shape corresponds to a set of resistance tensors. Their elements are computed by solving Stokes' equations in appropriate geometries\cite{Happel1965,Kim1991}. The case of an ellipsoid in a viscous shear flow was first solved by Jeffery\cite{Jeffery1922}.}
\obs{In this limit} the equation of motion is given by the condition that the \obs{hydrodynamic} torque vanishes at every instant. For particles that possess three orthogonal mirror planes\cite{Bretherton1962}, particle shape enters the orientational equation of motion through two parameters, $\Lambda$ and $K$. For an ellipsoid with half-axes $a$, $b$ and $c$, for example, $\Lambda=(\lambda^2\!-\!1)/(\lambda^2\!+\!1)$ and $K=(\kappa^2\!-\!1)/(\kappa^2\!+\!1)$ with aspect ratios $\lambda=a/c$ and $\kappa=b/c$.
The case $K=0$ corresponds to an axisymmetric particle. For axisymmetric particles the orientational dynamics is exactly solvable\cite{Jeffery1922}.
When $|\Lambda|<1$ (so that $0 < \lambda < \infty$)
there are infinitely many degenerate periodic orbits, the so-called \lq Jeffery orbits\rq. This is a consequence of the fact that the dynamical system has a conserved quantity: the \lq orbit constant\rq. The value of the orbit constant is determined by the initial orientation of the particle. The orientational motion of an axisymmetric particle in a simple shear is sometimes referred to as \lq tumbling\rq{}, the particle spends a long time aligned with the flow direction, and it periodically changes orientation by $180$ degrees. Different Jeffery orbits differ in the functional form of these periodic \lq flips\rq{}.
The degeneracy for $K=0$ is particular to the simple shear flow, and it means that small perturbations can have a large effect.
Inertial forces, for example, are neglected in Jeffery's theory. These forces induce \lq orbit drift\rq{} into a final stable orbit. This was already suggested by Jeffery\cite{Jeffery1922}, and was discussed in many experimental papers starting with Taylor\cite{Taylor1923}. See also Ref.~\citenum{Karnis1966}.
\obs{The corresponding theory is discussed by \citet{subramanian2005} and by Einarsson {\em et al.}\cite{einarsson2015a,einarsson2015b,candelier2015}.}
Small particles may be affected by thermal noise so that the orbit constant performs a random walk giving rise to a statistical distribution of orientations. This mechanism forms the theoretical basis for understanding the rheology of dilute suspensions\cite{Brenner1974,Hinch1972}.
A third possibility, the topic of this work, is that the particle is not perfectly axisymmetric. This leads to a more complicated orientational equation of motion, also derived by Jeffery\cite{Jeffery1922}. Some numerical examples of its solutions were reported by \citet{Gierszewski1978} who found that the motion of a non-axisymmetric particle in a simple shear flow is qualitatively different from that of an axisymmetric particle. \citet{Hinch1979} analysed the structure of the solutions to the equation of motion.
They found that for short times a nearly axisymmetric ellipsoidal particle approximately follows a Jeffery orbit, but on longer time scales the \lq orbit constant{\rq} does not remain
constant. It oscillates, giving rise to \lq doubly periodic\rq{} tumbling: time series of the components of the unit vector aligned with the major axis of the particle show two distinct periods. Subsequently Yarin {\em et al.}\cite{Yarin1997} inferred from numerical experiments and analytical calculations
that ellipsoidal particles may tumble periodically, quasi-periodically, or in a chaotic fashion -- depending on the particle shape and on its initial orientation. The term \lq quasi-periodic\rq{} refers to doubly periodic motion with incommensurable periods. Our experimental results support these theoretical predictions: our analysis demonstrates that the tumbling may indeed be periodic, doubly periodic, or possibly chaotic, depending on particle shape and initial orientation. Our results are in good qualitative agreement with theoretical predictions.
\obs{This work considers the orientational motion of small neutrally buoyant particles in a time-independent viscous shear flow.} This is a special but important case. It is of theoretical interest because of its degeneracy and sensitivity to small perturbations, and it is of practical interest because it fundamentally relates to theories and experiments concerning the rheology of dilute suspensions. Theories are commonly formulated in terms of Jeffery's equation\cite{Brenner1974,Petrie1999}.
Recently there has been a surge of interest in describing the tumbling of non-spherical particles in turbulent\cite{Par12,Pum11,Gus14,Ni14,Che13,Byron2015} and other complex flows\cite{Wil09,Wil10a,Wil11}.
Since it is difficult to solve the coupled particle-flow problem most theoretical and numerical studies rely
on Jeffery's equation as an approximation to the orientational dynamics.
Some exceptions are described in Refs.~\citenum{Marchioli2010,Einarsson2013a,Challabotla2015}.
This article is organised as follows. In Section~\ref{sec:background} we enumerate previous experimental efforts to validate Jeffery's equations. In Section~\ref{sec:methods} we describe the experimental setup. Section~\ref{sec:results} contains our experimental results. These results are discussed in Section~\ref{sec:discussion}, and we conclude in Section~\ref{sec:conclusions}.
\section{Previous experimental work}\label{sec:background}
In this section we give a brief account of earlier experiments observing the orientational dynamics of single particles in shear flows.
\citet{Taylor1923} immersed millimeter-sized aluminum spheroids in sodium silicate between two concentric rotating cylinders approximately \unit[10]{mm} apart. In his brief report he asserts that the tumbling of the spheroids is in qualitative agreement with Jeffery's predictions, but that the orientational dynamics drifts after many (order of $100$) particle rotations.
In a related study Eirich {\em et al.}\cite{Eirich1936} observed the orientations of glass rods and silk fibres in a Taylor-Couette device.
The particles had diameters between $10$ and \unit[50]{$\upmu$m} and aspect ratios between $5$ and $100$. No quantitative data on the orientational dynamics was measured, but they observed that the particles tend to align with the flow direction or
with the vorticity direction.
\citet{Binder1939} studied fibers of many different aspect ratios suspended in glycerine.
He employed a similar device with two concentric cylinders and found, as Taylor, that the orientational dynamics
slowly drifts.
Mason and co-workers have studied the orientational dynamics of small particles in shear flows during two decades\cite{Trevelyan1951,Mason1956,Bartok1957,Goldsmith1961,Goldsmith1962,Anczurowski1968}. Initially \citet{Trevelyan1951} used a setup of two concentric cylinders rotating in opposite directions, making it possible to study a single particle over an extended period of time. The gap between the cylinders was approximately \unit[10]{mm}, the suspending liquid was white corn syrup, and the particles were \unit[9.5]{$\upmu$m} diameter glass fibers cut to different lengths. By observing the particle orientation in a plane orthogonal to vorticity, \citet{Trevelyan1951}
found fairly good quantitative agreement of their experimental results with Jeffery's theory for one particle rotation (Fig.~7 in Ref.~\citenum{Trevelyan1951}). However, for longer time series (up to 30 revolutions) their results were inconclusive: sometimes orbit drift was observed, sometimes not, and sometimes the change of orbit appeared seemingly erratic.
In order to compare quantitatively with Jeffery's equations, valid for spheroidal particles, Trevelyan and Mason fitted the value of $\Lambda$ to measurements of the tumbling period, yielding in their words an \lq effective aspect ratio\rq{} for cylinders.
\citet{Bretherton1962} later showed that this procedure is consistent.
\citet{Mason1956} extended the experiment to hundreds of particle rotations (Fig.~2 in Ref.~\citenum{Mason1956}), but the observed orbit drift was apparently erratic.
Mason and Manley mention convective currents as a possible cause for the observed drift, but no conclusions could be drawn concerning the single-particle dynamics.
\citet{Bartok1957} used a similar device consisting of concentric cylinders, with a camera-equipped microscope observing along the vorticity axis, allowing
to very precisely measure the tumbling behavior of high aspect ratio ($\lambda\approx45)$ acrylic (\lq Orlon\rq) fibres. The experimental results
were found to agree quantitatively with Jeffery's predictions for one particle rotation (Fig.~5 in Ref.~\citenum{Bartok1957}). However, no data on the orbit drift, if any, was presented.
\citet{Goldsmith1961} performed the first quantitative measurements of the rotations of disks. They used the same coaxial-cylinder setup described above,
with silicone oil for the suspending liquid. The disks were fabricated by heating and compressing
polystyrene spheres. The disk diameters were \unit[400-850]{$\upmu$m},
and their aspect ratios ranged from $\lambda=1/20$ to $\lambda = 1/4$. Goldsmith and Mason showed that Jeffery's theory quantitatively predicts the orientational motion of
axisymmetric disks, and that the orbit remained constant over $120$ particle rotations (Table~II in Ref.~\citenum{Goldsmith1961}).
In a sequel \citet{Goldsmith1962} described measurements of the motion of particles suspended in a liquid flowing through a circular tube. The tube diameters were \unit[2-8]{mm}, the flow was pressure-driven by a syringe pump, and observations were recorded through a microscope traveling along the tube. Measurements were made on an assortment of particles
of sizes $\approx$\unit[0.1]{mm} with aspect ratios ranging from $\lambda=1/20$ (disks) to $\lambda=100$ (fibres). Goldsmith and Mason concluded that the dynamics along
Jeffery orbits compares well with theory for short times (Fig.~5 in Ref.~\citenum{Goldsmith1962}). However, they found measurable orbit drift after a single particle rotation, for both a rod
and a disk,
see Fig.~7 in the same paper. The drift was attributed to particle asymmetry.
\citet{Anczurowski1968} fabricated prolate spheroidal particles by polymerising an electrostatically deformed droplet, and showed that Jeffery's theory holds quantitatively for one particle rotation given the true aspect ratio $\lambda$. They used the same concentric-cylinder device described above.
Harris {\em et al.}\cite{Harris1979} performed experimental measurements on non-axi\-sym\-me\-tric particles ($K\neq0$). They used an apparatus with counter-rotating cylinders with a gap of \unit[27]{mm}, which was filled with a glucose solution. The particles were machined from a composite material (\lq Tufnol\rq) into cuboids of $\unit[1.75]{mm}\!\times\!\unit[1.28]{mm}$ cross-section and \unit[2.5-9.5]{mm} in length. They measured the unknown elements of the resistance tensors by observing simple rotations around each of the principal axes. With these numerical values of the resistance tensors they compared orientational trajectories of the cuboids to numerical solutions of Jeffery's equations and found reasonable quantitative agreement for two particle revolutions (Fig.~9 in their paper).
\citet{Stover1990} investigated the effect of a wall on fibre motion using a pressure-driven flow of corn syrup through a rectangular channel. The fibres had cross-sectional diameters of \unit[50]{$\upmu$m} and lengths \unit[600]{$\upmu$m}, resulting in an aspect ratio of $12$. They found that the orientational dynamics are in quantitative agreement with Jeffery's theory for two particle rotations when the particle is at least one particle length away from the wall.
\citet{Kaya2009} observed \emph{E. coli} bacteria advected in a microfluidic channel of rectangular cross section.
They found that the orientational motion of the bacteria approximately follows Jeffery orbits.
\citet{Einarsson2013a} described examples of orientational trajectories of polymer microrods in a microchannel. They found that the trajectories of some particles were periodic, admitting comparison to Jeffery's theory.
Other trajectories were seen to be aperiodic \obs{(Fig.~8 of Ref.~\citenum{Einarsson2013a}). The authors suggested
that this aperiodic motion may have been caused by lack of axisymmetry
of the particle in question. But it was not possible to draw definite conclusions, for several reasons. First the particles were
produced by shearing polymer microspheres\cite{alargova2004} which does not produce sufficiently well-controlled shapes. This is a problem because the shape of a given particle in the channel cannot be accurately observed. Second it could not be excluded that thermal torques affected the orientational dynamics. Third, and most importantly, the setup did not allow to record different orientational trajectories of the same particle. This motivated us to perform the experiments summarised in the following, they overcome the problems listed above.}
\section{Methods}
\label{sec:methods}
Fig.~\ref{fig:exp_setup}(a) shows a schematic drawing of the experimental setup to observe the orientational motion of small particles advected
in a microchannel flow. A dilute suspension of microrods in a density-matched fluid is introduced into the microchannel. A syringe pump (a standard Harvard Apparatus infuse/withdraw model) is used to drive the flow of the particles. The system is placed under an inverted microscope equipped with a motorised translation stage. A CCD camera is used to register the orientation of a given particle. The initial orientation and position of the particle in the channel is set using an optical trap.
\begin{figure}[t]
\begin{overpic}[width=8cm,clip]{fig1.pdf}
\end{overpic}
\caption{\label{fig:exp_setup} (a) Schematic picture of the experimental setup,
elements are not drawn to scale. $L_{1,2}$: lenses in a Keplerian telescope configuration. $M_{1,2,3}$: mirrors. D$_{1,2}$: diaphragms. DM: dichroic mirror
reflecting the laser beam vertically towards the microfluidic system. The microscopic objective focuses the beam on the sample plane. Illumination is provided from the top.
(b) Coordinate system in the lab frame spanned by orthogonal unit vectors $\hat{\bf x}$, $\hat{\bf y}$, and $\hat{\bf z}$. The $x$-axis (flow direction) is directed along the channel length, the $y$-axis along the depth of the channel, and the $z$-axis along the channel width. The unit vector $\ve n$ points
along the major axis of the particle. The polar angle between $\ve n$ and the ${z}$-axis is denoted by $\theta$.
The particles are kept roughly equally far away from the side walls of the channel. Since the channel is much wider than deep this means that the $\hat{\bf y}$-direction
is the shear direction, so that the local linearisation of the flow-velocity
obeys $u_x = sy$ in the frame co-moving with the centre-of-mass
of the particle, where $s$ is the shear rate.}
\end{figure}
Fig.~\ref{fig:exp_setup}(b) shows the coordinate system that is used in this paper. The $x$-axis is parallel to the channel length along the flow direction. The $y$-axis is directed along the depth of the channel. The $y$-axis is also the optical axis of the microscope objective. The $z$-axis points along the channel width. The orientation of the particle is determined by the unit vector $\ve n$ along the major axis of the particle.
The experiment is performed with cylindrical glass rods with diameters \unit[3]{$\upmu$m} $\pm$ \unit[0.01]{$\upmu$m} (PF-30S, Nippon Electric Glass Co., Ltd).
The microrods were manufactured as spacers in liquid-crystal devices (PF-30S, Nippon Electric Glass Co., Ltd). This requires precise diameters.
The lengths of the rods vary between approximately \unit[10]{$\upmu$m} and \unit[30]{$\upmu$m}.
An electron-microscope image of the particles is shown in Fig.~\ref{fig:particle_image}(a).
This Figure shows that the end surfaces of the cylindrical rods are randomly
inclined and uneven, indicating that the particles were obtained
by breaking longer glass rods. While this is irrelevant for the intended industrial use as spacers,
it is important for our application. Random inclinations of the end surfaces break axisymmetry: sometimes only very slightly
[lower left particle in Fig.~\ref{fig:particle_image}(a)],
sometimes more [c.f. particle in the centre of Fig.~\ref{fig:particle_image}(a)].
We investigate the orientational dynamics of highly asymmetrical particles
by studying aggregates of glass rods. Following \citet{Lewandowski2008}
a dilute suspension of microrods is left to evaporate in order to produce double particles,
Fig.~\ref{fig:particle_image}(b)\obs{--(d)}.
The glass particles have an index of refraction of $1.56$ and a density of \unit[$\rho_{\rm p}$ = 2.56]{g cm$^{-3}$}.
\begin{figure}[t]
\includegraphics[width=8cm]{fig2.pdf}
\caption{\label{fig:particle_image} (a) Electron-microscope image of the glass particles used in the experiments. Taken by S. Gustafsson, Chalmers.
\obs{(b-d) Optical-microscope images of the double particles used
in the experiments. The angular dynamics of the particle in panel (b)
is shown in Fig.~\ref{fig:particle3},
panel (c) corresponds to Fig.~S8, and panel (d) to Fig.~S9.
The last two figures are found in the Supplementary Online Material \cite{supp}.}}
\end{figure}
To achieve neutral buoyancy the fluid must have the
same density as the particles. \obs{To this end we mix} \unit[22.2]{$\%$wt} water, \unit[4.4]{$\%$wt} glycerol, and \unit[73.4]{$\%$wt} sodium metatungstate monohydrate (Alfa Aesar GmbH).
\obs{Density matching is achieved by titration while observing the particle under the microscope until the particle
is buoyant for several minutes. But there are additional sources of error that are difficult to control.
For example, when the fluid is pumped through the microchannel minute changes in temperature are expected
to slightly change the fluid density. Such changes could induce small variations in the $y$-position of the particle
that we sometimes observe during the recordings.}
The suspension is highly diluted in order to avoid particle-particle interactions. The mixture has a viscosity of
\unit[$\mu$ = 25]{mPa s} at \unit[20]{$^o$C}. We estimate the shear Reynolds number ${\rm Re}_s= \rho_{\rm f} s a^2/\mu$ as follows.
The particle length $a$ is of the order of \unit[20]{$\upmu$m}. The shear rate $s$ is determined
by the flow speed $u_x$ which is in turn given by the flow rate $8$ $\mu$l/min and
the cross section of the channel, \unit[2.5]{mm} $\times$ \unit[200]{$\upmu$m}.
This results in $u \approx 0.3$mm s$^{-1}$ and a shear rate of $s \approx 3 {\rm s}^{-1}$ at $60\upmu$m depth assuming
a parabolic profile in the $y$-direction, and that the channel
is much wider than deep.
The density of the fluid $\rho_{\rm f}$ equals the particle density, \unit[$\rho_{\rm p}$ = 2.56]{g cm$^{-3}$}.
This gives ${\rm Re}_s \approx 10^{-4}$. Inertial effects are thus negligible
on the time-scale of the experiment.
The microchannel is produced using standard soft lithography. The process begins by milling a rectangular moulding form in aluminum.
The surfaces of the mould are mechanically polished. A $10$:$1$ mixture of
polydimethylsiloxane (PDMS) and sylgard 184 Dow Corning
(Sigma Aldrich) is poured on the moulding form and allowed to cure for several hours. The PDMS replica obtained after peeling off the mould is sealed to a glass slide (thickness \unit[0.17]{mm}) by oxygen plasma bonding. This results in a rectangular channel that is \unit[40]{mm} long, \unit[2.5]{mm} wide, and \unit[150]{$\upmu$m} deep. For some measurements, a channel with a depth
of \unit[200]{$\upmu$m} was used, see Table \ref{tab:1}.
The suspension is injected into the channel using thin tubing connected to a syringe pump.
PTFE tubes (Cole-Parmer) with outer diameters of \unit[0.76 or 1.07]{mm} are used, the former in connection with a plastic syringe
(\unit[1]{ml}, Terumo), the latter in connection with a glass syringe (\unit[500]{$\upmu$l}, Hamilton).
The optical system is built around a Nikon $X$60 microscope objective (NA 1.0, WD \unit[2]{mm}).
The particle motion is recorded with a CCD video camera (Leica). The channel is mounted on a translation stage that
moves the microchannel over the fixed observation microscopic objective.
The stage is driven by a stepper motor that records the position of the stage. By moving the channel a given particle is kept within the field
of view of the objective, despite the fact that the particle is advected by the fluid through the microchannel.
A single-beam optical trap is used to set the initial orientation and position of a given particle.
The optical trap makes use of the microscope objective (Fig.~\ref{fig:exp_setup}), it provides sufficient magnification to
not only visualise the particle, but also to trap it with a continuous infrared laser of wavelength \unit[1075]{nm} (\unit[10]{W}, IPG-Laser GmbH).
The most efficient way of trapping a glass rod with this setup is to direct the laser beam towards one of the two ends of the particle.
Different orientations can be imposed on the particle by moving the channel sufficiently quickly to cause one end
of the particle to leave the trap, yet sufficiently slowly so that the other end of the particle is kept trapped.
\obs{The experimental procedure described above ensures stable and reproducible experimental conditions. But a number of external factors
were difficult to control and could still influence the experimental results to a small degree: evaporation due to the presence of unsaturated air in the reservoir, the presence of air bubbles in the channel that are difficult to get rid of, and pressure drops ininside the channel
due to the elasticity of the PDMS structure. }
We use the image-analysis algorithm employed by \citet{Einarsson2013a}. Images are acquired
at a rate of $100$ frames per second.
Each frame is stored as $8$-bit gray-scale bitmap with
$692 \times 520$ pixels. The pixel size is $0.21\upmu$m.
For a given frame the image analysis proceeds in three basic steps. First static noise is reduced by
subtracting the time-averaged intensity from each frame. Then the
boundary of the projection of the particle into the image plane is detected, and finally
an ellipse is fitted to the boundary. Details are given in Ref.~\citenum{Einarsson2013a}.
The output defines the position and the orientation of the particle in the image.
The centre-of-mass coordinates of the particle in the laboratory frame are determined
using the output from the stepper motor recording the position of the stage.
\obs{The main sources of uncertainty in the image analysis are the limited resolution of the camera and diffraction. The latter
gives rise to a diffuse particle boundary. The uncertainty in the determination of the particle orientation is largest when the
short end of the rod faces the camera, i.e. for small values of $n_z$ (see Fig.~\ref{fig:exp_setup}).}
A typical experiment starts by capturing a particle with the optical trap.
The particle is brought into the desired location in the $x$-$z$-plane by moving the channel. All particles are
started close to one of the inlets at approximately equal $z$-distances to both side walls. We verified
that the $z$-position remains centred, with an
error typically at most one particle length.
This implies that the shear in the $z$-direction remains very small (the channel is much wider than deep).
The $y$-coordinate thus corresponds to the shear direction, and the $z$-coordinate is the vorticity direction.
\obs{The particle is brought to the desired initial orientation
as described above and the centre-of-mass of the particle is placed at a depth of $60$ mm. The precision in determining the initial depth
is determine by the depth-of-field of the microscope. It is estimated to be of the order of one particle length.}
\obs{Then the particle is} released
to follow the flow in the microchannel.
We then invert the pressure gradient so that the particle is advected back in the opposite $x$-direction.
For each orientational trajectory we record both forward and backward dynamics.
Since Stokes' equation is invariant under simultaneous pressure inversion and time reversal, the backward dynamics must exactly
retrace the forward dynamics unless irreversible effects such as inertia or thermal noise affect the dynamics.
Examples for the resulting video-microscopy recordings of the orientational dynamics in the $x$-$z$-plane can be viewed
via the MULTIMEDIA VIEW links in Figs.~\ref{fig:particle1} and \ref{fig:particle3}.
\begin{figure}[p]
\begin{overpic}[width=14cm]{fig3.pdf}
\end{overpic}
\caption{\label{fig:particle1}
Orientational dynamics of particle $1$.
(a) Dynamics of $n_z$. Here $n_z^{(i)}$ denote the values of $n_z$ at subsequent zero crossings of $n_x$, $i=1,2,3,\ldots$. The data are taken from panels (b-f).
Red $\circ$ data from panel (b); blue $\Box$ data from (c); green $\triangle$ data from (d); black $ + $ data from (e); magenta $\star$ data from (f).
Panels (b-f) show orientational dynamics as a function of c.o.m.-position $x$ in the channel, Eq. (\ref{eq:tx}). Data in different panels correspond to different initial orientations.
Solid blue and dashed red lines represent forward and backward trajectories, respectively. The flow direction is reversed at $x=0$.
The horizontal arrow in panel (b) indicates the period $X_{\rm p}$[mm] of the trajectory. The video-microscopy recording of the orientational dynamics shown in panel (b) can be viewed at MULTIMEDIA VIEW, that of panel (d) at
MULTIMEDIA VIEW.}
\end{figure}
\begin{figure}[p]
\begin{overpic}[width=14cm]{fig4.pdf}
\end{overpic}
\caption{\label{fig:particle2} Orientational dynamics of particle $2$. See caption of Fig.~\ref{fig:particle1} for details. }
\end{figure}
For a given particle this procedure is repeated many times to obtain orientational trajectories with different initial orientations.
We record the length of the projection of the particle into the $x$-$z$-plane as a function of time. We estimate the particle length
using the procedure described in
Ref.~\citenum{Einarsson2013a}.
Once the particle length is known we can extract the components
of the unit vector $\ve n$ determining the orientation of the particle, as a function of time.
We plot the orientation not as a function of time but as a function of distance that the centre-of-mass of the particle
has traveled through the channel, advected by the flow:
\begin{equation}
\label{eq:tx}
x(t) = \int_0^t\!\!{\rm d}t' u_x(t')\,.
\end{equation}
Here $u_x(t)$ is the instantaneous flow velocity at time $t$. This transformation simplifies the analysis because it accounts for
the fact that the shear-rate is time-dependent: the flow velocity changes when the pressure is reversed, and
in order to avoid inertial effects these reversals must be performed slowly. The invariance of Stokes equation under
time and pressure reversal implies $x(t) = -x(-t)$. We overlay forward and backwards dynamics by plotting the orientation as a function of centre-of-mass position.
\section{Experimental results}
\label{sec:results}
Figures~\ref{fig:particle1}, \ref{fig:particle2}, and \ref{fig:particle3} show orientational dynamics of three different particles.
For each particle five different orientational trajectories are shown, corresponding to different initial orientations [panels (b) to (f)].
In all three Figures we show trajectories of the $x$- and $z$-components of the unit vector $\ve n$ that points along the major axis of the particle. Here
$n_x$ is the component in the flow direction, and $n_z$ is the component in the vorticity direction. The third component of $\ve n$ is determined by normalisation,
$|\ve n|=1$.
In panel (a) of each Figure we summarise the orientational dynamics by
recording the values of $n_z$ whenever $n_x=0$. We denote the resulting sequence of consecutive $n_z$-values by $n_z^{(i)}$, $i=1,2,3,\ldots$. For an axisymmetric particle Jeffery's equation predicts that $n_z^{(i+1)}=n_z^{(i)}$, shown as the solid line along the diagonal in panel (a).
As mentioned above, the particle is first advected along a stream line
of the pressure-driven flow in the channel. We then
invert the pressure gradient so that the centre-of-mass of
the particle is advected back to where it came from.
For each orientational trajectory we show the \lq forward dynamics\rq{} (blue solid line), going from right-to-left in the Figure.
After the {reversal} follows the \lq backward dynamics\rq{} (red dashed line). Since Stokes' equation is invariant under simultaneous pressure inversion and time reversal,
the backward orientational dynamics must exactly retrace the forward dynamics unless irreversible effects due to inertia or thermal noise affect the dynamics
noticeably on the time-scale of the experiment.
Consider first the trajectories shown in Fig.~\ref{fig:particle1}, corresponding to particle $1$.
Panels (b) to (f) show orientational trajectories of $n_x$ and $n_z$ for different initial orientations. In all cases the backward dynamics retraces the forward dynamics very well.
This shows that neither inertial forces nor rotational diffusion affect the orientational dynamics.
We attribute the small dephasing visible in each panel
to a small density mismatch \obs{(discussed in Section~\ref{sec:methods})}, causing the particle to sink (or float), changing the shear rate it experiences.
Apart from this slight dephasing all orientational trajectories are fairly
periodic. For a given $n_x$-trajectory the relative variation of the centre-of-mass distance between two consecutive
$n_x\!=\!0$-events (\lq half-period\rq{}) is of the order of $10$\%.
Between different panels we observe variations in the period $X_{\rm p}$, between \unit[2.1]{mm} and \unit[2.6]{mm}, \obs{caused by the uncertainty in the $y$-position mentioned above.} Panel (a) indicates that $n_z^{(i+1)}$ is approximately equal
to $n_z^{(i)}$.
\begin{figure}[t]
\begin{overpic}[width=14cm]{fig5.pdf}
\end{overpic}
\caption{\label{fig:particle3} Orientational dynamics of particle $3$. See caption of Fig.~\ref{fig:particle1} for details. The video-microscopy recording of the orientational dynamics shown in panel (d) can be viewed via MULTIMEDIA VIEW}
\end{figure}
Fig.~\ref{fig:particle2} shows the orientational dynamics of particle $2$
for different initial orientations [panels (b) to (f)]. In all cases the backward dynamics retraces the forward dynamics well, at least for a few millimetres. As for particle $1$ the trajectories show a slight dephasing within each panel.
Different panels show quite different periods $X_{\rm p}$, ranging between $1.7$mm and $5.9$mm. In panels (d) and (f) we see that the amplitude of $n_z$ changes aperiodically. This is also apparent from panel (a) exhibiting a somewhat wider scatter around the diagonal than panel (a) in Fig.~\ref{fig:particle1}.
Fig.~\ref{fig:particle3} shows the orientational dynamics of particle $3$, an asymmetric double particle. For all initial orientations the backward dynamics retraces the forward dynamics fairly well, with the exception of the trajectories shown in panel (e). As for particles $1$ and $2$ we observe a slight dephasing within each panel. The dynamics of $n_x$ is quite periodic, but again the periods $X_{\rm p}$ vary from panel to panel, ranging from \unit[1.4]{mm}
to \unit[4]{mm}. In all cases $n_z$ shows distinct aperiodicity. Panel (a) exhibits a larger scatter around the diagonal than panel (a) in Fig.~\ref{fig:particle2}.
\begin{table}[t]
\begin{tabular}{llllllll}
\hline\hline
Particle&Particle\hspace*{3mm} & Particle & Channel\hspace*{3mm} &Flow rate & Figure \\
number&type$^a$& length($\upmu$m)$^b$\hspace*{3mm} & depth ($\upmu$m)\hspace*{3mm} &($\upmu$l/min)\hspace*{3mm} & & \\\hline
Particle $1$ &single & 20.5 & 200 & 8 & Fig.~\ref{fig:particle1}\\
Particle $2$ &single & 24.3 & 150 & 5 & Fig.~\ref{fig:particle2}\\
Particle $3$ &double$^c$ & \obs{28.7} & 200 & 8 & Fig.~\ref{fig:particle3}\\
Particle $4$ &single & 20.1 & 150 & 7.5 & Fig.~S1$^d$ \\
Particle $5$ &single & 29.5 & 200 & 8 & Fig.~S2\\
Particle $6$ &single & 17 &150 & 5 & Fig.~S3 \\
Particle $7$ &single & 25 & 150 & 5 & Fig.~S4 \\
Particle $8$ &single & 26.5 & 150 & 5 & Fig.~S5 \\
Particle $9$ &single & 22.5 & 150 & 7.5 & Fig.~S6 \\
Particle $10$&single & 25.5 & 150 & 5 & Fig.~S7 \\
\obs{Particle $11$}&double$^e$ & 26 & 200 & 8 & Fig.~S8 \\
\obs{Particle $12$}&double$^f$ & 17 & 200 & 8 & Fig.~S9 \\
\hline\hline
\end{tabular}
\caption{Description of particles and other experimental parameters. \mbox{}$^a$Single or double particle. \mbox{}$^b$Particle length as extracted from image-analysis algorithm, see Section \ref{sec:methods}. We estimate
the error to be of the order of $1\upmu$m. \obs{\mbox{}$^c$This particle is shown in Fig.~\ref{fig:particle_image}(b). The lengths of the two glass rods
are 26.4 and 17.3$\upmu$m.}
\mbox{}$^d$These Figures are in the Supplementary Online Material \cite{supp}.
\obs{\mbox{}$^e$This particle is shown in Fig.~\ref{fig:particle_image}(c). The lengths of the two glass rods are 26 and 16.5$\upmu$m.
\mbox{}$^f$This particle is shown in Fig.~\ref{fig:particle_image}(d).
The lengths of the two glass rods are both 16$\upmu$m. }
}
\label{tab:1}
\end{table}
The orientational dynamics of particles $1$, $2$, and $3$ was obtained under slightly different experimental conditions. These are summarised in Table~\ref{tab:1},
as well as the particle properties. Table~\ref{tab:1} also gives information about other particles for which we obtained precise orientational dynamics. The corresponding Figures are found in the Supplementary Online Material \cite{supp}.
Video-microscopy recordings corresponding to the data shown in
Figs.~\ref{fig:particle1}(b), (d), and \ref{fig:particle3}(d)
can be viewed via the MULTIMEDIA VIEW links in the corresponding figure
captions.
\section{Discussion}
\label{sec:discussion}
The results summarised in Figs.~\ref{fig:particle1} to \ref{fig:particle3} show the orientational dynamics of single and double glass rods.
In general the particles are not perfectly axisymmetric (as seen in Fig.~\ref{fig:particle_image}), and therefore do not follow perfect Jeffery orbits. In this section we relate our experimental results to the theoretical predictions valid for ellipsoidal particles\cite{Jeffery1922,Hinch1979}. The particles in our experiment do not satisfy the mirror symmetries assumed in this theory, but it is plausible that the effects of breaking axisymmetry predicted by this theory apply at least qualitatively to our glass rods.
The equation of motion for the orientational dynamics of an ellipsoidal particle\cite{Jeffery1922,Hinch1979} can be cast in the form\cite{Einarsson2013b}
\begin{subequations}
\label{eq:jeffery}
\begin{align}
\dot{\ve n}&={\ma O}\ve n +\Lambda\, \big({\ma S}\ve n - (\ve n\cdot{\ma S}\ve n)\ve n\big)
+ \frac{K(1-\Lambda^2)}{K\Lambda-1}(\ve n\cdot{\ma S}\ve p)\ve p\,,\\
\dot{\ve p}&={\ma O}\ve p +K \,\big({\ma S}\ve p - (\ve p\cdot{\ma S}\ve p)\ve p\big) + \frac{\Lambda(1-K^2)}{K\Lambda-1}(\ve n\cdot{\ma S}\ve p)\ve n\,.
\end{align}
\end{subequations}
Following the convention outlined
in Section \ref{sec:methods}, $\ve n$ is a unit vector that points along the major axis of the ellipsoidal particle. The unit vector $\ve p$
is orthogonal to $\ve n$, directed along the particle axis corresponding
to the length $b$ used in the definition of the aspect ratio $\kappa$
(defined in the Introduction).
The geometry of the ellipsoid is characterised by the two
shape parameters $\Lambda$ and $K$ that are defined in the Introduction, Section \ref{sec:intro}.
The matrices $\ma S$ and $\ma O$ are the symmetric and anti-symmetric parts of $\ma A$, the matrix of fluid-velocity gradients. In our case this matrix takes the form
\begin{equation}
\ma A=
\left[ {\begin{array}{cccc}
0 & s & 0 \\
0 & 0 & 0 \\
0 & 0 & 0 \\
\end{array} } \right]\,,
\end{equation}
where $s$ is the shear strength (Fig.~\ref{fig:exp_setup}).
Eqs. (\ref{eq:jeffery}) are symmetric under the simultaneous exchange of $\ve n$ and $\ve p$ as well as $\Lambda$ and $K$: describing the motion of the same particle using
a different coordinate system within the particle must result in the
same dynamics.
Note also that the non-linear coupling between $\ve n$ and $\ve p$ involves the strain $\ma S$ only. This coupling maintains
orthogonality of the two vectors $\ve n$ and $\ve p$.
The anti-symmetric part $\ma O$
just causes a solid-body rotation. For axisymmetric particles, $K=0$,
so that the tumbling of $\ve n$ becomes independent of
the dynamics of $\ve p$ (but not vice versa). The resulting equation for $\ve n$ has infinitely many degenerate periodic solutions, the Jeffery orbits\cite{Jeffery1922}. The dynamics of $\ve p$ describes how the particle spins around its symmetry axis.
When $K\neq 0$ (and $\Lambda \neq 0$), no general closed-form solutions
of Eqs.~(\ref{eq:jeffery}) are known.
It is convenient to represent the numerical solutions of Eq.~(\ref{eq:jeffery}) in terms of a Poincar\'{e} surface-of-section\cite{Strogatz}, recording the locations at which the dynamics intersects a surface in the phase space of Eq.~(\ref{eq:jeffery}). This section is constructed as follows.
\citet{Hinch1979} have shown
that the vector $\ve n$ rotates around the vorticity at a positive angular velocity so that one can reduce the dimensionality of the problem by parametrising
the orientational dynamics in terms of the corresponding angle.
A suitable condition\cite{Hinch1979,Yarin1997} defining the surface-of-section is that $\ve n$ is perpendicular to the flow direction, $n_x=0$.
Following Yarin {\em et al.}\cite{Yarin1997} we take the coordinates in the surface-of-section to be $\psi$
and $n_z$, where
$\psi$ is the Euler angle parametrising the spin of the particle around the axis $\ve n$, and $n_z$ is the $z$-component of $\ve n$,
the cosine of the angle $\theta$ between $\ve n$ and vorticity (Fig.~\ref{fig:exp_setup}).
When $n_x=0$ we record
the coordinates $(\psi, n_z)$.
To define
the Euler angles ($\theta,\phi,\psi$) we use the convention of Goldstein\cite{Goldstein} and express $\ve n$ and $\ve p$ as
\begin{equation}
\ve n =
\left [\begin{array}{l}\phantom{-}\sin\theta\sin\phi \\
-\sin\theta\cos\phi \\
\phantom{-}\cos\theta\end{array}\right]\,,\quad\mbox{and}\quad
\ve p = \left [\begin{array}{l}-\sin\psi\cos\phi-\cos\theta\sin\phi\cos\psi \\
-\sin\psi\sin\phi + \cos\theta\cos\phi\cos\psi \\
\phantom{-}\cos\psi\sin\theta\end{array}\right]\,.
\end{equation}
So $\theta$
is the polar angle depicted in Fig.~\ref{fig:exp_setup}. The angle $\phi$ is referred to as the \lq precession angle\rq{}. This angle measures the direction of the projection of $\ve n$ into the flow-shear plane.
Eqs.~(\ref{eq:jeffery}) correspond to Eqs.~(2.1) and (2.2) in Ref.~\citenum{Yarin1997}, setting $s=-1$ and defining the aspect ratios $\lambda$ and $\kappa$
in terms
of the particle axes\cite{Yarin1997} as follows: $a_x=\lambda a_z$ and $a_y = \kappa a_z$.
In the experimental time series shown in Figs.~\ref{fig:particle1} to \ref{fig:particle3}, instances where $n_x\! =\! 0$ correspond to
peaks in the oscillation in $n_z$. The $n_z$-coordinate in the surface-of-section is therefore easily read off from the experimentally observed time series. The angle $\psi$, by contrast, cannot be measured in our experiment because we cannot track how the particles spin around $\ve n$.
Four different surfaces-of-section are shown in Fig.~\ref{fig:poincares}, obtained by numerical integration of Eqs.~(\ref{eq:jeffery}) for a large number of different initial orientations,
and plotting the sequence $[\psi^{(i)},n_z^{(i)}]$ of $(\psi,n_z)$ evaluated at consecutive zero crossings of $n_x$, labeled
by $i = 1,2,3,\ldots$. Similar sections are found in Ref.~\citenum{Yarin1997}. The
map that gives $[\psi^{(i+1)},n_z^{(i+1)}]$ in terms of $[\psi^{(i)},n_z^{(i)}]$ is called the Poincar\'e{} map.
Fig.~\ref{fig:poincares}(a) depicts the orientational dynamics of an axisymmetric particle, $K=0$. The coordinate $n_z$ is a conserved quantity on the surface-of-section, Jeffery orbits appear as horizontal lines in Fig.~\ref{fig:poincares}(a), one-parameter families parametrised by $\psi$. In the literature Jeffery orbits are commonly identified by their orbit constant $C$. It is given by the value of $\tan\theta$ at $n_x=0$ (see for example Eq.~(3) in Ref.~\citenum{Hinch1979}). In this paper we characterise Jeffery orbits by $n_z=\cos\theta=1/(1+C^2)$ on the surface-of-section ($n_x=0$). Fig.~\ref{fig:poincares}(a) illustrates that the orientational dynamics depends on the initial orientation, determining the value of $n_z$.
We remark that the periods of $\ve n$ and $\psi$ are not in general commensurate for $K=0$. But the tumbling of $\ve n$ is independent of that of the angle $\psi$ and thus periodic for axisymmetric particles.
Fig.~ \ref{fig:poincares}(b) shows the orientational dynamics of a weakly asymmetric ellipsoidal
particle ($K \approx 0.095$ and $\Lambda=12/13$). We see that Jeffery orbits
with $n_z \approx \obs{\pm} 1$ remain almost unchanged.
But there are substantial changes for smaller values of
$|n_z|$, compared with the surface-of-section for $K=0$.
We see that $n_z$ ceases to be a constant of motion,
doubly-periodic orientational dynamics results. The most substantial
changes occur near $n_z=0$. The $n_z\!=\!0\,$-Jeffery orbit
is replaced by two fixed points at $(0,0)$ and $(\pm\pi/2,0)$ on the surface-of-section.
This fact and the surface-of-section patterns in the vicinity of these points follow from the time-reversal symmetry
of Eqs.~(\ref{eq:jeffery}). The general principle is explained in Section 6.6 of Ref.~\citenum{Strogatz}. See also Ref. \citenum{Politi1986}.
The invariance of Stokes equation referred to in Section \ref{sec:methods} implies that Eqs. (\ref{eq:jeffery}) are invariant
under
\begin{equation}
\label{eq:symmetry}
t \to -t\,,\quad n_x \to -n_x\,,\quad\mbox{and}\quad p_x\to -p_x\,.
\end{equation}
The fixed point $(0,0)$ is mapped onto itself under this transformation. It follows that
the dynamics in its immediate neighbourhood can neither be expanding nor contracting.
In other words the determinant describing the linearised motion in the vicinity of this fixed point,
\begin{equation}
\det \left [
\begin{array}{ll}
\frac{\partial \psi^{(i+1)}}{\partial \psi^{(i)}} & \frac{\partial \psi^{(i+1)}}{\partial n_z^{(i)} }\\
\frac{\partial n_z^{(i+1)}}{\partial \psi^{(i)}}& \frac{\partial n_z^{(i+1)}}{\partial n_z^{(i)}}\end{array}\right ] \,,
\end{equation}
must be of unit modulus despite the fact that the dynamics (\ref{eq:jeffery}) is dissipative.
We find numerically that the eigenvalues are $\exp(\pm {\rm i}\sigma)$. The point $(0,0)$ is thus an elliptic fixed point, surrounded
by a one-parameter family of closed orbits that appear as concentric closed curves, much like so-called \lq tori\rq{} in so-called \lq Hamiltonian\rq{} systems with area-preserving phase-space dynamics\cite{Lichtenberg}.
For near-axisymmetric particles \citet{Hinch1979} analysed the corresponding orbits by multiple-scales analysis.
As these orbits rotate around the elliptic point, the value of
$n_z$ changes sign. This doubly-periodic motion may be quasi-periodic or periodic, depending on whether the two frequencies are incommensurate
or not (corresponding to irrational or rational winding numbers, respectively).
\begin{figure}[p]
\begin{overpic}[width=14cm]{fig6.pdf}
\end{overpic}
\caption{\label{fig:poincares} Poincar\'e{} surfaces-of-section for $\Lambda=12/13$ and different values of $\kappa$
[$K=(\kappa^2-1)/(\kappa^2+1)$]. The angle $\psi$ is defined up to $(\psi+\pi/2)\mbox{mod}\pi$.}
\end{figure}
\begin{figure}[p]
\begin{overpic}[width=9cm]{fig7.pdf}
\end{overpic}
\caption{\label{fig:max_nz} Range of Poincar\'e{} map for
$\Lambda=12/13$ and different values of $\kappa$ corresponding
to the values used in Fig.~\ref{fig:poincares}, $K=(\kappa^2-1)/(\kappa^2+1)$.
The data shown was obtained by
\obs{by iterating the Poincar\'e{} map once for
randomly chosen initial conditions. The initial conditions were determined by sampling $\psi$ uniformly over the surface-of-section for given values of $n_z$.}
}
\end{figure}
The point ($\pm \pi/2,0$) is a hyperbolic fixed point.
\obs{We find numerically that the eigenvalues are real
and opposite}, $\gamma$ and $\gamma^{-1}$.
The fact that the $n_z\!=\!0\,$-orbit is destroyed upon infinitesimal perturbation and replaced by a discrete set
of fixed points of alternating stability is typical for Hamiltonian systems\cite{Lichtenberg}. The mechanism in our dissipative system is analogous,
a consequence of the symmetry (\ref{eq:symmetry}).
For larger asymmetries chaotic orientational dynamics occurs, seen as a region with a stochastic scatter of points
in Fig.~\ref{fig:poincares}(c). Almost all Jeffery orbits are modified, only those with $n_z$ close to $\pm 1$ remain.
Whether the orientational dynamics is periodic, quasi-periodic, or chaotic depends upon the initial condition on the surface-of-section.
Fig.~\ref{fig:poincares}(d) shows the orientational dynamics for ellipsoidal particles with $K = 3/5$ and $\Lambda=12/13$.
This value of $K$ is similar the corresponding value for the double particle $3$ (Figs.~\ref{fig:particle_image}(b) and \ref{fig:particle3}),
but this particle has $\Lambda \approx 40/41$. We do not seek quantitative correspondence between these parameters because the particles used in our experiment are not ellipsoidal. This prevents us from drawing quantitative conclusions, but still allows for a qualitative comparison between theory and experiment.
The orientational dynamics displayed in Fig.~\ref{fig:poincares}(d) is either on tori or chaotic.
Motion on tori can occur with large amplitudes, so that $n_z$ changes from $n_z \approx -1$ to $n_z \approx 1$.
How are these observations reflected in the experimental time series shown in Figs.~\ref{fig:particle1} to \ref{fig:particle3}?
Since we cannot measure the angle $\psi$ in our experiments, we concentrate on the dynamics of $n_z$.
Fig.~\ref{fig:max_nz} shows the range of changes of $n_z$ in one iteration of the Poincar\'e{} map for particles with different degrees of asymmetry, determined by numerically recording the changes $n_z^{(i+1)}-n_z^{(i)}$ along orientational trajectories. The diagonal corresponds to symmetric particles where $n_z^{(i+1)} = n_z^{(i)}$. The larger the asymmetry, the larger is the range of $n_z^{(i+1)}-n_z^{(i)}$ that may occur, reflecting the properties of the surfaces-of-section shown in Fig.~\ref{fig:poincares}.
Fig.~\ref{fig:max_nz} can be directly compared with Figs.~\ref{fig:particle1}(a) to \ref{fig:particle3}(a). Our experimental results show that the Poincar\'e{} map may scatter significantly around the diagonal. The above discussion explains that this is a consequence of doubly-periodic and possibly
chaotic orientational dynamics of asymmetric particles. The range of scatter differs between different particles, a consequence of different degrees of asymmetry. We see that the scatter is largest for the double particle $3$, with $K \approx 3/5$
[Fig.~\ref{fig:particle3}(a)].
Particle $1$, by contrast, shows only negligible scatter around the diagonal [Fig.~\ref{fig:particle1}(a)]. We infer that this particle is highly symmetric,
$K$ is very small.
The data shown in Fig.~\ref{fig:particle1} are consistent with the conclusion that particle $1$ follows Jeffery orbits, $n_z$ remains approximately constant
in the surface-of-section. But we remark that the shape of the $n_z\approx 0$-orbit does not look like a Jeffery orbit for an axisymmetric particle.
We cannot exclude that this is due
to possible non-ellipsoidal deviations from axisymmetry. The surface-of-section dynamics is most sensitive to such shape perturbations
near $n_z=0$.
A more likely explanation is that the shape of the trajectory is a consequence of systematic (and reproducible) tracking
errors due to diffraction and finite pixel size. \obs{These errors are largest for small values of $n_z$, as discussed
in Section~\ref{sec:methods}.}
Particle $2$ is also a single glass rod, but it shows a somewhat larger range of scatter around the diagonal [Fig. \ref{fig:particle2}(a)].
We attribute this to a more substantial breaking of axisymmetry at the tips of the particle [as seen for instance in Fig.~\ref{fig:particle_image}(a)]. Particle $2$ shows fairly periodic motion for $n_z \approx 1$ and distinctly doubly-periodic orientational dynamics for small values of $n_z$. This confirms that the initial orientation determines whether the particle tumbles
periodically or aperiodically. Note also that $n_z$ changes sign along the trajectories that remain near $n_z=0$ [Fig. \ref{fig:particle2}(b), (d), and (f)].
All of these observations are qualitatively consistent with the surfaces-of-section shown in
Fig. \ref{fig:poincares}(b) and \ref{fig:poincares}(c).
For particle $3$ the value of $n_z$ changes sign for all orientational trajectories shown.
This is consistent with the fact that the Fig. \ref{fig:poincares}(d) shows
predominantly this type of motion. Fig.~\ref{fig:particle3}(e) is consistent with chaotic orientational motion in the stochastic layer around the elliptic island. The surfaces-of-section in Fig.~\ref{fig:poincares} show that there are two types of doubly-periodic motion: either $n_z$ has the same sign, or its sign changes periodically. The trajectory in Fig.~\ref{fig:particle3}(e) exhibits both behaviours, indicating chaotic dynamics.
Now consider the trajectory shown in Fig.~\ref{fig:particle3}(c).
It is not periodic or doubly-periodic, but the reversal works well
(at least initially). But we cannot conclude that the dynamics is consistent with surfaces-of-section discussed above. The $n_z$-values
at $n_x=0$ in Fig.~\ref{fig:particle3}(c) change from approximately $0.25$ to $0.75$ in modulus.
Explaining this behaviour in terms of chaotic dynamics on the surfaces-of-section
requires a large stochastic region, larger than the one shown in
Fig. \ref{fig:poincares}(d).
This might mean that the trajectory shown in panel (d) is a doubly-periodic piece of a chaotic trajectory that may show different
behaviours at larger times. But to determine whether these behaviours can be explained by chaotic dynamics would require to derive and numerically integrate the
orientational equations of motion for the precise shape of the particle, and to experimentally record the angle $\psi$. Since the current experimental
setup does not allow to reliably extract how the angle $\psi$ changes,
we plan to perform experiments with small triangular platelets
that will allow to record the angle $\psi$.
The winding number of the trajectory shown in Fig.~\ref{fig:particle3}(d) is roughly $7$ corresponding to seven $n_x\!=\!0\,$-crossings
while the trajectory winds approximately once round the elliptic fixed point in the centre of the surface-of-section. The winding number
of the trajectory in Fig.~\ref{fig:poincares}(e) is roughly $4$ corresponding to four $n_x=0$-crossings while the surface winds once around the elliptic fixed point.
These observations are in qualitative agreement with the fact that the winding numbers of tori winding around the elliptic fixed points increase as
the distance from that point increases (corresponding to larger maximal values of $n_z$).
Additional results for seven more particles are shown in the Supplementary Online Material \cite{supp}, Figs.~S1 to S7. In general the results shown in these Figures
support the observations and qualitative conclusions summarised above. But the trajectory shown in Fig.~S7 is difficult to reconcile with the surfaces-of-section
shown in Fig.~\ref{fig:poincares}. Fig.~S7(c) shows an orbit near $n_z=0$ where $n_z$ appears not to change sign.
But panels (d) and (f) in Fig.~S7 show distinct sign changes, not consistent with the surfaces-of-section shown in Fig.~\ref{fig:poincares}.
It was pointed out in Section \ref{sec:results} that the periods $X_{\rm p}$ observed in the orientational trajectories can differ substantially between different trajectories
of the same particle. Consider for instance particle $1$ (Fig.~\ref{fig:particle1}).
For an axisymmetric ellipsoidal particle ($K=0$) the Jeffery period (time units) is given by\cite{Jeffery1922}
\begin{equation}
\label{eq:Tp}
T_{\rm p} = \pi \frac{\lambda^2+1}{s\lambda}\,.
\end{equation}
For $K=0$ this is twice the return time to the surface-of-section. For $K\neq 0$ the return
time depends upon the starting position on the surface-of-section, but for small values of $K$ the deviations
from (\ref{eq:Tp}) are small, of the order of $K$. For the nearly axisymmetric particle $1$ we can use Eq. (\ref{eq:Tp}) to estimate the period $X_{\rm p}$
in Fig.~\ref{fig:particle1}. Using the parameters from Table \ref{tab:1} and
assuming that the particle was located at a depth of \unit[60]{$\upmu$m},
Eq. (\ref{eq:Tp}) gives $X_{\rm p} \approx$~\unit[2.3]{mm}.
This is consistent with the range of periods observed in Fig.~\ref{fig:particle1}(b) to (f),
namely \unit[2.1]{mm} to \unit[2.6]{mm}.
We infer that the precision in determining the depth at which particle $1$ moves through the channel is of the order
of one particle length. The variations in periods observed in Figs.~\ref{fig:particle2} and \ref{fig:particle3} indicate that the actual depths vary more
between different panels than in Fig. \ref{fig:particle1}.
\section{Conclusions}
\label{sec:conclusions}
Theory and numerical simulations predict that the orientational dynamics of small neutrally buoyant particles in a shear flow is very
sensitive to breaking of axisymmetry\cite{Gierszewski1978,Hinch1979,Yarin1997}.
Axisymmetric particles tumble periodically on Jeffery orbits, but when the symmetry is broken, the Jeffery orbits are modified. Depending on the initial orientation, periodic, doubly-periodic, or chaotic tumbling may result.
\obs{In order to experimentally verify these predictions it is necessary to use particles of well-defined shape, to ensure that inertial effects and rotational diffusion are negligible, and to compare different orientational trajectories (corresponding
to different initial orientations) of the same particle.}
In this paper we describe measurements of the orientational motion of small glass rods suspended in a micro-channel shear flow.
\obs{The measurements are precise enough to allow for a definite comparison with theory\cite{Hinch1979,Yarin1997}.
First, the glass rods have highly symmetric circular cross sections. Slight imperfections at the ends break the axisymmetry
weakly for some particles [Fig.~\ref{fig:particle_image}(a)]. We also use strongly triaxial particles designed for our purpose
by joining two glass rods. Second, by reverting the pressure-driven flow we observe that the orientation retraces its trajectory for many periods. This means that neither inertial effects nor rotational diffusion matter on this time scale. Third by means of an optical trap we change
the orientation of the particle, making it possible to measure different orientational trajectories for the same particle.}
Our results \obs{confirm the theoretical predictions\cite{Hinch1979,Yarin1997}. We observe periodic and doubly-periodic tumbling,
and how the nature of the orientational dynamics depends upon the initial orientation, and on the degree to which axisymmetry is broken.
Our measurements are consistent with the features of the surfaces-of-section shown in Figs.~\ref{fig:poincares} and \ref{fig:max_nz}.}
In the future we plan to experimentally map out Poincar\'{e} surfaces-of-section (Fig.~\ref{fig:poincares}). With the present particles this is not possible because we cannot resolve how the particle spins around its major axis.
We therefore plan to perform corresponding experiments using micron-sized triangular platelets. We expect that it will be possible to resolve the tumbling and spinning dynamics of these particles using the methods described in this paper.
\begin{acknowledgments}
We thank S. Gustafsson (Chalmers) for taking the electron-microscope image Fig.~\ref{fig:particle_image}(a) and for granting us permission to use the image in this paper. B. Mehlig gratefully acknowledges discussions with S. \"Ostlund (University of Gothenburg) on the implications of time-reversal symmetry for the surface-of-section dynamics shown in Fig.~\ref{fig:poincares}.
This work was supported by grants from the Carl Trygger Foundation for Scientific Research and the Swedish Research Council. Support from the MPNS COST Action MP1305 \lq Flowing matter\rq{} is gratefully acknowledged.
\end{acknowledgments}
\providecommand{\noopsort}[1]{}\providecommand{\singleletter}[1]{#1}%
|
\section{Montecarlo wave function method}
The statistics of the photons emitted by the system under pulsed excitation were first addressed by performing quantum Montecarlo simulations \cite{Dum1992}. This method not only allows to work with larger truncated Hilbert spaces but also provides direct access to individual photon counts, thus embodying the closest theoretical simulation of an actual Hanbury Brown-Twiss experiment. In brief, the algorithm is based on the stochastic evolution of the system wave function through the Schr\"{o}dinger equation
\begin{equation}\label{Sch}
\hat{{\cal{H}}}\left| \psi \right\rangle = i\hbar \frac{\partial }{{\partial t}}\left| \psi \right\rangle \, ,
\end{equation}
written for the non-Hermitian effective Hamiltonian
\begin{equation}\label{Heff}
\hat{{\cal{H}}} = {\hat{{\cal{H}}}_{s}} - \frac{{i\hbar }}{2}\sum\limits_j {\kappa_j\hat a_j^\dag {{\hat a}_j}} \, ,
\end{equation}
The non Hermitian part of \ref{Heff} results in a decay of the norm $\left\langle {\psi }\mathrel{\left | {\vphantom {\psi \psi }}\right. \kern-\nulldelimiterspace}{\psi } \right\rangle$. During the evolution of Eq.~\ref{Sch}, random numbers $0<r<1$ are drawn and the condition $\left\langle {\psi }\mathrel{\left | {\vphantom {\psi \psi }}\right. \kern-\nulldelimiterspace}{\psi } \right\rangle \le r$ decides for the action of a jump operator, $\hat{a}_j\left| \psi \right\rangle$, corresponding to the measurement of a photon. The proper quantum jump operator is chosen such that $j$ is the smallest integer satisfying $\sum\nolimits_j P_j \ge r$, where $P_j$ are the probabilities for the mode $j$ to emit a photon at a given time. Each evolution of Eq.~\ref{Sch} produces a stochastic quantum trajectory associated with the state ${{{\left| \psi{\left( t \right)} \right\rangle }_j}}$, and the procedure can be repeated $N$ times to form an ensemble average of realizations in view of approximating the system density matrix as $\hat \rho \left( t \right) \mathop {\rm{ = }}\limits_{N \to \infty } {\left| {\Psi \left( t \right)} \right\rangle}{\left\langle {\Psi \left( t \right)} \right|}$, where
\begin{equation}
\left| {\Psi \left( t \right)} \right\rangle = \sum_{j = 1}^N {{{\left| {\psi \left( t \right)} \right\rangle }_j}}/N \, ,
\end{equation}
and its potential mixed nature. Any observable or correlation are obtained from $\langle {\hat O( t )}\rangle = {\rm{Tr}}[ {\hat \rho ( t )\hat O} ]$. The full access to photon counts and emission times history allows to mimic the experimental detection scheme. The two-times second-order correlation function, $g^{(2)}(t_1,t_2)$, can be reconstructed from the statistics of photons delays analogously to a Hanbury Brown and Twiss (HBT) setup. Further details on the numerical procedure employed to obtain the results of Fig.~2 in the main text are given in the following.
\section{Two-time correlations under pulsed excitation}
In our Montecarlo simulations we have worked on the basis of trajectories containing single pulses, which makes the data analysis more flexible. We have tracked the quantum jumps performed by the driven cavity from which the photon antibunching is expected. We point out that in reality one should expect a weak mixing between both cavity fields to occur in the guiding channels. It can be accounted for within an input-output treatment. In such a case the laser detuning should be properly adapted according to the prescriptions of Ref. \onlinecite{Flayac2013}. To analyze delays within a given pulse we need to track the trajectories where \emph{at least} two quantum jumps occurred within $\Delta T$ and these events are obviously rare given the relatively small occupation of the cavity 1 as one can see from Fig.~2 of the main text. We have therefore performed a large campaign of massively parallelized simulations on an high end cluster based on $N = 1.5 \times10^{8}$ pulses from which we have recorded the whole quantum jump history. We considered pulses of duration 4 ns separated by 20 ns to avoid any overlap bringing some unwanted pulse to pulse correlations. Our simulation therefore covers not less than 1 seconds of recorded data.
To build the Monte-Carlo curve of Fig. 2g (blue disks), we have worked on quantum jumps that occurred in a time window of width $\Delta T=T_2-T_1=1.57$ ns centered on the $g^{(2)}(t,t)$ minimum (see yellow surface and blue curve of Fig. 2f) mimicking the temporal filtering. Inside this global window we have considered sub-windows of variable duration $\Delta t=t_2-t_1$ ranging from 6.5 ps to $\Delta T$. Each of these sub-windows was gradually displaced by $\Delta t$ within $\Delta T$, starting from the condition $t_1=T_1$ ($t_2=t_1+\Delta t$) and until $t_2=T_2$ ($t_1=t_2-\Delta t$) is fulfilled. For a given value of $\Delta t$, the un-normalized second order correlation $G^{(2)}(\Delta t)$ is obtained from the sum of photon pair counts recorded by slicing the time window, which increases the statistics by $N_w=\Delta T/\Delta t$. Therefore it allows to work with a number of counts that would correspond to $N \times N_w$ trajectories (pulses) reducing by $N_w$ the required computational time. Finally, the $g^{(2)}(\Delta t)$ is obtained by normalizing $G^{(2)}(\Delta t)$ to the number double counts expected from a Poissonian statistics. The errors (magenta curve) are computed from the square root of the number of counts averaged over the sliding windows. Obviously the error is inhomogeneous versus $\Delta t$, given that $N_w$ is variable, and it is small both in the regions of narrow and wide $\Delta t$ where respectively $N_w$ is large and number of counts is important. The previously described procedure is summarized in Fig.~S1 (see captions).
To build the histograms of Fig.~2d-e displaying the pulse-to-pulse statistics, we have performed a Montecarlo rearrangement of our single pulse trajectories to randomize their time ordering, as it would be obtained from many pulse trajectories or in an actual experimental situation.
\begin{figure}[h]
\includegraphics[width=0.5\textwidth,clip]{FigS1.pdf}
\caption{(a) Average occupation of cavity 1 (red) and the equal time second order correlation function (blue line). The yellow region highlights the global time window $\Delta T$ from which the photon counts are extracted. (b) Zoom in between the vertical dashed lines of panel (a) showing an illustrative set of sub-windows for a given value of $\Delta t$. The wavy lines illustrate and example of quantum jump series for a given trajectory. Only the 2 jumps event are kept for the $G^{(2)}(\Delta t)$ statistics. 3 and more jumps events are totally absent in the conditions we consider.}
\label{FigS1}
\end{figure}
\section{Quantum master equation}
The master equation for the density matrix reads
\begin{equation}\label{rho2res}
\dot{\rho}=\frac{1}{i \hbar}[\rho,\hat{{\cal{H}}_{s}}] + \mathcal{L}^{(1,2)} \, ,
\end{equation}
where losses are accounted for through Liouvillian operators in the usual Lindblad form for the two resonators modes,
$\mathcal{L}^{(1,2)}= \sum_{i=1,2} {\kappa_{i}} [\hat{a}_{i}\rho \hat{a}_{i}^{\dag} - 0.5(\hat{a}_{i}^{\dag}\hat{a}_{i}\rho-\rho\hat{a}_{i}^{\dag}\hat{a}_{i})]$.
Further sources of loss, such as nonlinear absorption (e.g. related to the imaginary part of $\chi^{(3)}$) or pure dephasing, could also be added to Eq. \ref{rho2res} (see, e.g. \onlinecite{Ferretti2012,Ferretti2013}), but we neglect them here for simplicity. Moreover, the model can be generalized to include input and output quantum channels \cite{Flayac2013}.
Single-time evolution and steady state numerical results of Eq.~\ref{rho2res} can be straightforwardly performed, as in Refs.~\onlinecite{Liew2010,Bamba2011,Ferretti2013}.
Here, we were additionally able to confirm the Montecarlo results (cyan curve in Fig.~2g of the main text).
The un-normalized and normalized two-times second-order correlation functions of cavity 1 were computed as
\begin{eqnarray}
{G^{(2)}}\left( {{t},{t'}} \right) &=& {\rm{Tr}}\left[ {\hat a_1^\dag {{\hat a}_1}{U_{{t} \to {t'}}}\left( {\hat a_1^\dag {{\hat a}_1}\hat \rho \left( {{t'}} \right)} \right)} \right]\\
{g^{(2)}}\left( {{t},{t'}} \right) &=& \frac{{{G^{(2)}}\left( {{t},{t'}} \right)}}{{{\rm{Tr}}\left[ {\hat a_1^\dag {{\hat a}_1}\hat \rho \left( {{t}} \right)} \right]{\rm{Tr}}\left[ {\hat a_1^\dag {{\hat a}_1}\hat \rho \left( {{t'}} \right)} \right]}} \, ,
\end{eqnarray}
where ${\hat{\cal{U}}}_{{t} \to {t'}}(\hat O)$ is the propagator of the operator $\hat O$ from $t$ to $t'$ associated with Eq. \ref{rho2res}. The photon statistics produced within a time window $\Delta t=t_2-t_1$ is obtained from
\begin{equation}
{g^{(2)}}\left( {\Delta t} \right) = \frac{{\iint_{\Delta t} {{G^{(2)}}\left( {t,t'} \right) dt dt'}}}{{\iint_{\Delta t} {n_1( t )n_1( {t'} ) dt dt'}}}
\end{equation}
where $n_1( t )={\rm{Tr}}[ {\hat a_1^\dag {{\hat a}_1}\hat \rho \left( t \right)}]$. This exact calculation perfectly reproduces the Montecarlo wave function results within the error envelope, as it is reported in Fig.~2g (cyan curve).
\section{Photonic crystal cavities design}
Photonic crystal cavities allow to achieve record figures of merit today, such as ultra-small mode volumes and ultra-high quality factors~\cite{Notomi2010}.
One of the most used photonic crystal cavity designs is realized by removing three air holes in a triangular lattice~\cite{Akahane2003}, which is usually defined a L3 point defect.
Recently, a combination of fast simulation tools~\cite{Andreani2006} and genetic optimization~\cite{Minkov2014} have allowed to show that Q factors largely exceeding $10^6$ can be designed for such cavities, which was shown experimentally~\cite{Lai2014}.
For the photonic crystal cavities design used in this work, we started from a standard SOI photonic crystal membrane, with the silicon layer thickness of 220 nm. We set the lattice constant to $a=400$ nm and the holes radius to $r=112$ nm ($r/a=0.28$) to tune the cavity mode resonant wavelength in the relevant telecom band, i.e. $\lambda = 1.5$ $\mu$m ($\sim 0.825$ eV). The three holes along the cavity axis have been shifted by $s_1=120$ nm ($s_1 /a=0.3$) $s_2=100$ nm ($s_2 /a=0.25$), and $s_3=40$ nm ($s_3 /a=0.1$), to reach a theoretical (unloaded) $Q\sim 1.25 \times 10^6$. Since we aim at coupling these cavities with access and output waveguides, we thus allow the loaded Q-factor to be in the $0.8\times 10^6$ range.
A plot of the normalized electric field intensity, i.e. .the function $|\vec{\alpha} (\mathbf{r})|^2$ with $\int |\vec{\alpha}(\mathbf{r})|^2 \mathrm{d}\mathbf{r} =1$, is shown in Fig.~S2 for our optimized L3 cavity design, which was the building block for the the photonic crystal molecules in Fig.~3 of the main text.
\begin{figure}[t]
\includegraphics[width=0.45\textwidth,clip]{FigS2.pdf}
\caption{Electric field intensity profile at the center of the silicon photonic crystal membrane.}
\label{FigS2}
\end{figure}
\section{Estimating the effective photon-photon interaction}
In this work, we focus on photonic nanostructures in silicon, which is a strongly nonlinear material already at the level of classical electromagnetic response. In particular, bulk silicon is characterized by a relatively large $\chi^{(3)}$ susceptibility, while nominally $\chi^{(2)} = 0$ (neglecting surface contributions) owing to the centrosymmetric nature of the elementary crystalline cell \cite{Boyd2008}. Strongly enhanced nonlinear effects have been already reported in L3 photonic crystal cavities~\cite{Galli2010}.
The photon-photon interaction energy in each resonator is given in terms of the material $\chi^{(3)}$ by the simplified expression \cite{Ferretti2012}
\begin{equation}\label{ham_kerr}
U= \frac{D(\hbar\omega_i)^2 } {8\varepsilon_0}
\int \mathrm{d}\mathbf{r} \,\frac{\chi^{(3)}(\mathbf{r})}{\varepsilon^2(\mathbf{r})} \, |{\alpha} (\mathbf{r}) |^4 \, ,
\end{equation}
where $\vec{\alpha} (\mathbf{r})$ is the three-dimensional cavity field profile, normalized as $\int |\vec{\alpha}(\mathbf{r})|^2 \mathrm{d}\mathbf{r} =1$, and $D$ represents the multiple contributions of the same order of magnitude given by the different elements of the $\chi^{(3)}$ tensor \cite{Boyd2008}.
From the calculated mode profile shown in Fig.~S2, the effective nonlinearity of such a silicon nanocavity can be estimated through Eq.~\ref{ham_kerr}, by using $\chi^{(3)}\sim 0.9\times 10^{-18}$ m$^2/$V$^2$, which is an appropriate order of magnitude for the elements of the bulk silicon third-order susceptibility tensor \cite{Boyd2008}, and D=24 \cite{Sipe1987}. For the cavity mode profile of our optimized photonics crystal structure, see Fig.~S2, a quantitative estimate of this integral results in $U\simeq 0.8\times 10^{-3}$ $\mu$eV, close to what was assumed in the model calculations of the main text and confirming the order of magnitude estimates already given in the literature~\cite{Ferretti2013}.
|
\section*{Summary}{\thispagestyle{empty}}\setcounter{page}{1}\label{summary}
\section*{Summary}\label{summary}
}
\pagestyle{empty}
\noindent This thesis concerns exact and approximate treatments of electronic energy transfer in photosynthetic systems. While the methods used are completely general, their application is focused on the Fenna-Matthews-Olson (FMO) pigment-protein antenna complex, found in certain bacteria. \\[0.5cm]
The FMO complex is a trimer consisting of 24 bacteriochlorophyll (BChl) sites, each of which is coupled to a dissipative environment, which renders the energy transfer irreversible, and means that the dynamics of this transfer must be treated using techniques appropriate for open quantum systems. \\[0.5cm]
Recently, experimental evidence of quantum beating effects in the energy transfer in the FMO complex was found, and theoretical and experimental results suggested that quantum coherence might be observed at room temperature (300 K). One of the aims of this thesis is to ascertain how important coherence effects are for this transfer. \\[0.5cm]
Firstly, the Hierarchical Equations of Motion (HEOM) are introduced, which give numerically exact transfer dynamics, albeit with computational cost rising steeply with the size of the system. The efficient numerical implementation of the HEOM is discussed, and a Taylor-series integration is compared to the traditional Runge-Kutta method, the former proving to be more efficient. \\[0.5cm]
The accuracy of the HEOM can also be improved by representing the bath correlation function as a Pad\'e series instead of the more traditional Matsubara series: the former converges very rapidly compared to the latter, and allows more efficient simulation of the dynamics.\\[0.5cm]
Results are then presented for the FMO complex, both for a 7-site subsystem of the monomer and for the trimer. Numerically exact results are calculated to provide a benchmark for cheaper approximate methods including the Redfield and F\"orster theories. It is found that incoherent F\"orster theory describes the overall features of the dynamics well at 300 K, suggesting that coherence has little if any effect on energy transfer efficiency at room temperature. \\[0.5cm]
Finally, F\"orster theory is used to test the effects of two phenomena on the energy transfer dynamics in the FMO complex: that of including vibrational structure in the environment (traditionally modelled as being unstructured), and that of static disorder due to a slowly fluctuating environment. \\[0.5cm]
It is found that energy transfer in the complex is very robust with respect to changes in the environment at room temperature, and that the results are largely the same if structure is introduced into the environment or disorder is accounted for. Rather than electronic coherence, it is this robustness that is advantageous for the complex's biophysical role.
\newpage
{\renewcommand{\baselinestretch}{1}\normalsize
\section*{Acknowledgements}\label{acknowledgements}
}
\pagestyle{empty}
I would like to thank David Manolopoulos for supervising my Part II project and for his unparalleled guidance and inspiration throughout my degree.
The members of the Manolopoulos group, Nike, Josh, Michele, Jon and Kenji, also helped make this year an enjoyable and productive one. Along with members of the Barford and Wilson groups, they have also provided many interesting conversations over many hours of coffee.
I thank my friends for providing a distraction when I needed one, including Karolis, David R., Emily N., Richard, Emily B., Alan, Emma, Michael, Iain, Nat, Paul and Karen.
The support of my parents has been invaluable both during my degree and otherwise, and I would like to thank them very much for this, along with my brother, Nicholas, and my girlfriend Anne, as well as my grandparents and great-aunt.
Finally, this work is dedicated to the memory of my Nan, Joyce Wilkins, whom I miss very much.
\renewcommand\thepage{}
{\renewcommand{\baselinestretch}{1}\normalsize
\tableofcontents{\thispagestyle{empty}}
}
\renewcommand\thepage{\arabic{page}}
{\renewcommand{\baselinestretch}{1}\normalsize
\chapter{Introduction}\pagestyle{fancy}\setcounter{page}{1}\label{chapter_introduction}
}
\fancyhead{}
\renewcommand{\sectionmark}[1]{\markright{\thesection.\ #1}}
\fancyhead[LO,RE]{\rightmark}
\pagestyle{fancy}
The process of photosynthesis is one whose importance cannot be overestimated: photosynthetic organisms capture energy from sunlight and store it in chemical form. This energy is taken up by organisms higher up the food chain \cite{blankenship,renger_review,fleming_lightharv}.
In an antenna, photons of sunlight are absorbed by pigments, generally based on chlorophyll molecules, and the resulting excitation is transported between these pigments, eventually reaching an energy ``funnel'', in which the excitation moves over time towards pigments of lower energy, meaning that the process is irreversible. Its destination is the reaction centre, in which the energy is converted to chemical form \cite{blankenship}.
This work focuses on the excitation energy transfer that occurs in the energy funnel, although the methods used are suitable for much more general problems.
This transfer is extremely efficient, and it is immediately apparent that it would be very useful if we could design systems that were as effective at transferring energy. For this, it is necessary to understand the features of the system that lead to this efficiency.
As an example, we will be interested in the effect of the environment surrounding the pigments: presumably (as with many natural systems), the environment has been finely tuned to facilitate transfer. If we were to alter this environment, what effect would it have on transfer rates?
We will see in Section \ref{quant_coh} that the importance of quantum-mechanical effects in the transfer is currently under question. By modelling the transfer using methods that both include and ignore these quantum effects, we can gain an idea of how vital they really are at physiological temperatures.
This transfer is of current experimental and theoretical interest \cite{fleming_lightharv,coherence,ishizakifleming,schulten_LH2,2d_fmo}, and we now discuss a complex whose study will allow us to benchmark the techniques used, and to learn about the physics of energy transfer.
\section{Fenna-Matthews-Olson Complex}
The Fenna-Matthews-Olson (FMO) complex is a pigment-protein complex in which bacteriochlorophyll (BChl) pigments are bound to a protein. It is found in green sulfur bacteria such as \emph{C. tepidum} and \emph{P. aestuarii} \cite{fleming_lightharv}, and collects electronic excitation from a light-harvesting chlorosome, funnelling it towards the reaction centre.
The complex itself is a trimer, each monomer of which was thought originally to comprise 7 BChl molecules \cite{fenna_matthews_olson,fenna_matthews}, though it is now accepted that there is an 8th, more weakly bound BChl site in addition \cite{eight_site,schmidt_am_busch}. The currently accepted arrangement of bacteriochlorophylls is shown in Fig. \ref{fmo_figure}.
\begin{figure}[!ht]
\begin{center}
\includegraphics[width=13.5cm, keepaspectratio=true]{fmo.png}
\caption[Bacteriochlorophyll arrangement in FMO monomer.]{\small{Bacteriochlorophyll arrangement in FMO monomer. Figure created with Chem3D 2010, using PDB entry 3ENI \cite{olbrich}.}}
\label{fmo_figure}
\end{center}
\end{figure}
The sites labelled 1, 6 and 8 are those found nearest to the chlorosome, so are the ones that capture the excitation energy, while those labelled 3 and 4 are closest to the reaction centre and lowest in energy, so it is to these sites that the energy is funnelled.
There are several reasons for the popularity of the FMO complex in studies of energy transfer: its structure is well-documented \cite{fenna_matthews,eight_site}, and has been since 1974 \cite{fenna_matthews_olson}, making it ideal for theoretical studies, and its solubility in water means it is convenient for use in experiments \cite{coherence,eight_site}. Because it has been extensively studied in the past, the data required to carry out our simulations is readily available \cite{adolphs_renger}.
Next, we turn to the problem of finding a model Hamiltonian to describe the excitation energy transfer, which will require an introduction to the open quantum system.
\section{Open Quantum Systems}\label{open_qsys}
Consider a system of N ``sites'' (for example, BChl sites) labelled by Latin letters, each of which has a ground and a single excited electronic state. If $\ket{e_{j}}$ refers to site $j$ being in the excited state, and $\ket{g_{j}}$ the ground state, we will focus on the single-excitation manifold, that is, states $\ket{j}$:
\begin{equation}
\ket{j} = \ket{e_{j}}\prod_{k\ne j}^{\text{N}}\ket{g_{k}}.
\end{equation}
Using the states $\lbrace\ket{j}\rbrace$ as a basis, the so-called ``system'' Hamiltonian is given by:
\begin{equation}
\hat{H}_{S} = \sum_{j}^{\text{N}}\hbar\omega_{j}\ket{j}\bra{j} + \sum_{j,k \ne j}^{\text{N}}\hbar J_{jk}(\ket{j}\bra{k} + \ket{k}\bra{j}).
\end{equation}
Here, $\hbar\omega_{j}$ is the energy of site $j$ and $\hbar J_{jk}$ represents the coupling between two sites, due to dipolar interactions \cite{renger_review}. The eigenstates of the Hamiltonian (referred to as excitons) are represented by Greek letters $\ket{\nu}$, such that:
\begin{equation}
\hat{H}_{S}\ket{\nu} = \hbar\omega_{\nu}\ket{\nu}.
\end{equation}
Where:
\begin{equation}
\ket{\nu} = \sum_{j}^{\text{N}}U_{\nu j}\ket{j}.\label{exciton_basis}
\end{equation}
If $\hat{H}_{S}$ were sufficient to describe the entire process, we see that we could use a time-dependent state vector $\ket{\psi(t)} = e^{i\hat{H}_{S}t/\hbar}\ket{\psi(0)}$ to describe the dynamics of the system, and furthermore that these dynamics would be reversible (see Fig. \ref{irreversible} (a)).
Clearly, this is not the full story, as we are not considering a completely isolated system. Rather, the electronic system represented by $\hat{H}_{S}$ is an open system, surrounded by an environment with which it can exchange energy. We take each site to be associated with its own ``bath'', the vibrations of the bacteriochlorophyll and of the surrounding protein, and if each bath is comprised of harmonic oscillators, then the total bath Hamiltonian is:
\begin{equation}
\hat{H}_{B} = \sum_{j}^{\text{N}}\sum_{\alpha}\left( \frac{\hat{p}_{j\alpha}^{2}}{2m_{j\alpha}} + \frac{1}{2}m_{j\alpha}\omega_{j\alpha}^{2}\hat{q}_{j\alpha}^{2} \right).
\end{equation}
Here, $\hat{q}_{j\alpha}$ is the coordinate of the $\alpha^{th}$ oscillator in the $j^{th}$ bath, $\hat{p}_{j\alpha}$ is the conjugate momentum, $m_{j\alpha}$ the mass and $\omega_{j\alpha}$ the frequency. The system-bath interaction is:
\begin{equation}
\hat{H}_{SB} = \sum_{j}^{\text{N}}\ket{j}\bra{j}\sum_{\alpha}g_{j\alpha}\hat{q}_{j\alpha} \equiv \sum_{j}^{\text{N}}\hat{V}_{j}\hat{\xi}_{j},
\end{equation}
\noindent with $\hat{\xi}_{j} = \sum_{j}g_{j\alpha}\hat{q}_{j\alpha}$ the generalized force exerted on the $j^{th}$ site by its bath and $\hat{V}_{j} \equiv \ket{j}\bra{j}$. Before considering the effect that the environment will have on the dynamics of the excitation transfer, we introduce the spectral density, $J_{j}(\omega)$, for the $j^{th}$ site:
\begin{equation}\label{spectral_density}
J_{j}(\omega) = \sum_{\alpha}\frac{g_{j\alpha}^{2}}{2 m_{j\alpha}\omega_{j\alpha}}\delta(\omega - \omega_{j\alpha}).
\end{equation}
This function has peaks at the frequencies of the bath oscillators, weighted by the strength of their coupling to the site $j$, and quantifies the system-bath interaction. In general, it is replaced by a smooth function of $\omega$ (corresponding to an uncountably infinite number of oscillators). Commonly, for reasons that will be discussed later on in this thesis, the spectral density chosen is the Lorentz-Drude function:
\begin{equation}\label{lorentz_drude}
J_{j}(\omega) = \frac{2\hbar}{\pi}\frac{\lambda_{j}\gamma_{j}\omega}{\gamma_{j}^2 + \omega^2}.
\end{equation}
Here $\lambda_{j} = \int_{0}^{\infty}d\omega J_{j}(\omega)/\omega$ is the reorganization energy, a measure of system-bath coupling strength, and $\gamma_{j}$ is the characteristic frequency of the bath, which satisfies $\dfrac{dJ_{j}(\omega)}{d\omega}(\omega=\gamma_{j}) = 0$. Fig. \ref{drude_lorentz} illustrates this spectral density.
\begin{figure}[!ht]
\begin{center}
\includegraphics[width=13.5cm, keepaspectratio=true]{drude_lorentz.png}
\caption[Lorentz-Drude spectral density.]{\small{Lorentz-Drude spectral density of Eqn. \eqref{lorentz_drude}. Parameters used: $\lambda_{j} = 35\text{ cm}^{-1}$, $\gamma_{j} = 106.1\text{ cm}^{-1}$.}}
\label{drude_lorentz}
\end{center}
\end{figure}
The harmonic oscillators of the bath cause a kind of quantum-mechanical friction, one effect of which is exactly the same as classical friction: energy is dissipated from the electronic system, giving irreversible energy transfer.
There is, in addition, the purely quantum-mechanical effect of decoherence. The vibrational states associated with each electronic state may be excited or de-excited (we refer to them as absorbing or emitting a phonon), which alters the site energy \cite{leegwater_klug}. This in turn causes the coherence between two sites to decay, as the phase $e^{i\theta}$ of one state $\ket{k}$ with respect to another, $\ket{j}$, is altered when the energy of a site changes, and when these phases are randomized, their average value is zero. Over time, quantum-mechanical oscillations will be damped.
Both of these phenomena are illustrated in Fig. \ref{irreversible}, which shows the dynamics of a closed system and of a system with exactly the same Hamiltonian $\hat{H}_{S}$, but coupled to a bath of harmonic oscillators.
Since a larger reorganization energy $\lambda_{j}$ gives a stronger interaction between bath and system \cite{spin_boson}, we will expect stronger dissipation and more decoherence. For the same reason, higher temperatures also lead to greater decoherence.
\begin{figure}[!ht]
\begin{center}
\includegraphics[width=13.5cm, keepaspectratio=true]{Irrev.png}
\caption[Comparison of closed and open quantum system dynamics.]{\small{Comparison of the dynamics of (a) a closed and (b) an open quantum system, showing the effects of dissipation and decoherence.}}
\label{irreversible}
\end{center}
\end{figure}
One further property of the open quantum system must be mentioned: that of Markovianity. Due to the statistical nature of the environment, the force acting on a given site is a stochastic function of time.
If this stochastic process were Markovian, then at any given time no memory of its past behaviour would be required to predict its behaviour in the future; only its current value would be needed. There is no reason to assume \emph{a priori} that the behaviour of the function is Markovian, so that an accurate description of the dynamics should allow for non-Markovian effects.
The techniques used to deal with open quantum systems are applicable not only to biological systems such as the FMO complex, but also to many others, such as electron and proton transfer \cite{filtering1,proton_transfer,spin_boson}, as well as damping of Rabi oscillations due to excitation in quantum dots \cite{variational}. The theoretical tools described in this work are thus very versatile, and useful in many fields of current scientific interest.
\section{Quantum Coherence In Energy Transfer}\label{quant_coh}
We might intuitively expect that for the FMO complex at a physiological temperature of around 300 K, there would be a very strong interaction between environment and system (due to thermal excitation of bath oscillators), leading to very fast decoherence.
Fairly recently Engel \emph{et al.} \cite{coherence} observed quantum beating effects in a 2-dimensional electronic spectroscopy study of energy transfer in the FMO complex at a cryogenic temperature of 77 K, and Collini \emph{et al.} \cite{coherence2} observed similar effects at an ambient temperature of 294 K for a system similar to the FMO complex, leading to the suggestion that quantum coherence is important for energy transfer, for example in tunnelling through energy barriers \cite{ishizakifleming}.
This result was followed by a number of theoretical studies by Ishizaki and Fleming \cite{ishizakifleming,heom1,i&f_redfield} with the aim of accurately modelling the energy transfer dynamics in the FMO complex.
Importantly, using the Hierarchical Equations of Motion (which will be discussed in this thesis), the dynamics were simulated at both 77 K and 300 K for a 7-site subsystem of the monomer \cite{ishizakifleming}. At both temperatures oscillations were observed, suggesting that even at room temperature, quantum coherence was seen in the energy transfer, thus sparking renewed interest in the FMO complex.
One of our aims in the following work will be to appraise whether or not quantum coherent effects are necessary for efficient energy transfer at room temperature.
In the literature, it is conventional to discuss excitation transfer dynamics by showing time-dependent site populations \cite{ishizakifleming} (i.e., in the site basis $\lbrace\ket{j}\rbrace$), and we will use this convention in the work to follow.
A number of timescales are important in the study of the FMO complex: electronic coherence lasts for a number of femtoseconds (around 500 fs at 77 K, and around 200 fs at 300 K), so that a numerical simulation up to 1 ps will capture the effects of coherence, if appropriate.
After the coherence has decayed, the population dynamics show a simple decay, like those predicted by a rate equation, and eventually the populations reach a steady state. This occurs at around 10 ps at 77 K, and around 5 ps at 300 K, and will be observed in a numerical simulation up to 15 ps.
A number of experiments have been carried out which show that the fluorescence lifetime of bacteriochlorophylls in the FMO complex is on the order of 1 ns \cite{fluor_lifetime}. This means that we need not consider loss of excitation energy via spontaneous emission on the timescales used in our simulations, as such events are suitably rare.
\section{Summary}
A brief survey of the remainder of this thesis is now appropriate. Firstly, in Chapter \ref{chapter_heom} the Hierarchical Equations Of Motion (HEOM) for calculation of exact energy transfer dynamics will be introduced, and the advantages and disadvantages of using these equations will be described.
Next, in Chapter \ref{approximate} we will introduce several cheaper methods that can be used to obtain approximate results for the dynamics, the Redfield and F\"orster theories.
Chapter \ref{implementation} contains a description of some methods used to implement the HEOM efficiently, while doing so more accurately than in previous work.
Chapter \ref{numerical_results} is dedicated to numerical results for the 7-site and the 24-site FMO complex (the latter of which has not had any exact results published yet) from the HEOM and from the approximate methods described in Chapter \ref{implementation}. It is found that at 300 K, the incoherent F\"orster theory describes the general features of the energy transfer quite well.
Chapter \ref{applications} then covers applications of F\"orster theory: we find the effects of using a structured spectral density instead of the Lorentz-Drude function, and of static disorder due to slow fluctuations in the protein environment.
Finally, Chapter \ref{conclusions} concludes.
{\renewcommand{\baselinestretch}{1}\normalsize
\chapter{Hierarchical Equations Of Motion}\label{chapter_heom}
}
There are several methods for finding a numerically exact description of the dynamics of an open quantum system, many of which have been in use for some decades. The Feynman-Vernon influence functional \cite{influence} is based on the path integral formulation of quantum mechanics \cite{feynman}, wherein the dynamics of the system alone are considered, having integrated out those of the bath.
A more recent method is the self-consistent hybrid method \cite{spin_boson}, in which the spectral density is not considered as a continuum, but is discretized, giving a number of ``bath modes''. The dynamics of a number of these modes is treated exactly along with the system, while the dynamics of the rest of the modes is treated classically: the number of modes treated exactly is increased until the results converge.
In this Chapter we present the Hierarchical Equations of Motion (HEOM), which were formulated originally by Tanimura and Kubo \cite{tanimura1,tanimura2} using the influence functional as a starting point. We work instead with operators \cite{heom1}, deriving the influence operator from a consideration of the statistical properties of a bath of harmonic oscillators.
Section \ref{RDO} introduces the reduced density operator, and gives the general equation of motion for this operator. In Section \ref{infl}, the influence operator is derived, and in Section \ref{EOM}, a hierarchy of auxiliary density operators is introduced, giving exact equations of motion.
\section{Reduced Density Operator}\label{RDO}
An ensemble of quantum systems is most appropriately described not with wavefunctions but rather, using the density operator formalism. If $\hat{H}$ is the Hamiltonian for a system and $\lbrace|\phi_{j}\rangle\rbrace$ the eigenstates of $\hat{H}$, then the density operator is given by:
\begin{equation}\label{densop}
\hat{\rho} = \sum_{n} p_{j} \ket{\phi_{j}}\bra{\phi_{j}},
\end{equation}
\noindent where $p_{j}$ is the probability that a system in this ensemble will be found in the state $\ket{\phi_{j}}$. Particularly important is the situation in which the system is a member of a canonical ensemble, so that $p_{j} = e^{-\beta\epsilon_{j}}/Z_{\beta}$, where $\beta = 1/k_{B}T$, $\epsilon_{j}$ is the energy eigenvalue corresponding to eigenstate $\ket{\phi_{j}}$, and $Z_{\beta} = tr[e^{-\beta\hat{H}}]$ is the canonical partition function. In this case, the canonical density operator is $\hat{\rho}_{\beta} = e^{-\beta\hat{H}}/Z_{\beta}$.
A diagonal matrix element of the density operator, $\bra{\phi_{k}}\hat{\rho}\ket{\phi_{k}}$, gives the population of state $\ket{\phi_{k}}$ \cite{density_matrix}, and by cyclic permutation of operators within a trace, the average value of an operator $\hat{A}$ is given by:
\begin{equation}
\langle A\rangle = \sum_{j,k} p_{j} \bra{\phi_{j}}\hat{A}\ket{\phi_{k}}\inner{\phi_{k}}{\phi_{j}} \equiv tr[\hat{A}\hat{\rho}].
\end{equation}
For a system described by a single state vector, if this state is not an eigenstate of the Hamiltonian then it will evolve in time, and similarly, from the Time-Dependent Schr\"odinger equation, if a density matrix does not commute with the Hamiltonian, we will observe dynamics \cite{density_matrix}, given by $i\hbar d\hat{\rho}(t)/d t = [\hat{H},\hat{\rho}(t)]$.
\subsection{Superoperators And The Interaction Picture}\label{supops}
It proves useful now to introduce some definitions. Firstly, the Liouvillian superoperator $\hat{\mathcal{L}}$ is defined such that:
\begin{equation}
\hat{\mathcal{L}}\hat{O} = [\hat{H},\hat{O}].
\end{equation}
Further, we can define two superoperators $\hat{A}^{\rightarrow}$ and $\hat{A}^{\leftarrow}$ such that $\hat{A}^{\rightarrow}\hat{O} = \hat{A}\hat{O}$ and $\hat{A}^{\leftarrow}\hat{O} = \hat{O}\hat{A}$. Then the commutation and anticommutation superoperators are given by \cite{irreversible}:
\begin{equation}
\commut{A} = \hat{A}^{\rightarrow} - \hat{A}^{\leftarrow} \Rightarrow \hat{A}^{\times}\hat{O} = [\hat{A},\hat{O}],
\end{equation}
\begin{equation}
\anticommut{A} = \hat{A}^{\rightarrow} + \hat{A}^{\leftarrow} \Rightarrow \hat{A}^{\circ}\hat{O} = \lbrace\hat{A},\hat{O}\rbrace.
\end{equation}
A useful relation is the following, which can be verified by differentiating both its left and right hand sides:
\begin{equation}
e^{i\hat{\mathcal{L}}t/\hbar}\hat{O} \equiv e^{i\hat{H}t/\hbar}\hat{O}e^{-i\hat{H}t/\hbar}.
\end{equation}
In order to derive an equation of motion for the density operator, we first split up the total Hamiltonian as $\hat{H} = \hat{H}_{0} + \hat{V}$ (and thus, the Liouvillian as $\hat{\mathcal{L}} = \hat{\mathcal{L}}_{0} + \hat{\mathcal{L}}_{V}$). Then, in the interaction picture with respect to $\hat{H}_{0}$, denoted by a tilde following the notation of \cite{heom1}, the state vectors and operators are given by (pp. 172-173 of \cite{dirac}):
\begin{subequations}\label{interaction}
\begin{equation}\label{int_vec}
\ket{\tilde{\phi}(t)} = e^{i\hat{H}_{0}t/\hbar}\ket{\phi(t)} = e^{i\hat{H}_{0}t/\hbar}e^{-i\hat{H}t/\hbar}\ket{\phi(0)},
\end{equation}
\begin{equation}\label{int_op}
\tilde{O}(t) = e^{i\hat{\mathcal{L}_{0}}t/\hbar}\hat{O}.
\end{equation}
\end{subequations}
Inserting Eqn. \eqref{int_vec} into the time-dependent Schr\"odinger equation gives:
\begin{equation}\label{lipmann-schwinger}
i\hbar\frac{d}{d t}\ket{\tilde{\phi}(t)} = \tilde{V}(t)\ket{\tilde{\phi}(t)},
\end{equation}
\noindent and this gives the expression for the time-derivative of the density operator in the interaction picture:
\begin{equation}\label{von_neumann}
\frac{d}{d t}\tilde{\rho}(t) = -\frac{i}{\hbar}\tilde{\mathcal{L}}_{V}(t)\tilde{\rho}(t).
\end{equation}
Solution of this differential equation will give an expression for the time-dependence of the density operator for the system described by Hamiltonian $\hat{H}$.
\subsection{Time-Evolution And The Partial Trace}\label{partial_trace}
Solution of \eqref{von_neumann} is not as straightforward as it first appears. To understand this, we first integrate the equation to give a recurrence relation:
\begin{equation}\label{recurrence}
\tilde{\rho}(t) = \tilde{\rho}(0) - \frac{i}{\hbar}\int_{0}^{t}dt_{1}\tilde{\mathcal{L}}_{V}(t_1)\tilde{\rho}(t_{1}).
\end{equation}
The density matrix at a given time is determined by an integral over density matrices in the past, and can be expressed as an infinite series by expanding \eqref{recurrence}:
\begin{equation}\label{infinite_series}
\tilde{\rho}(t) = \tilde{\rho}(0) + \sum_{k=1}^{\infty}\left[\left(\frac{-i}{\hbar}\right)^{k}\int_{0}^{t}dt_{1}\dots\int_{0}^{t_{k-1}}dt_{k}\tilde{\mathcal{L}}_{V}(t_{1})\dots\tilde{\mathcal{L}}_{V}(t_{k})\right]\tilde{\rho}(0).
\end{equation}
There is a notable similarity between this and the so-called perturbation expansion \cite{feynman}. By analogy, we can interpret \eqref{infinite_series} as a sum over alternative possibilities.
A general term in the sum represents the effect of $\hat{H}_{SB}$ on the system at times $t_{1},t_{2},\dots,t_{k}$. Liouvillians at two different times do not necessarily commute, and so the order in which they are applied is important.
We introduce the time-ordering operator $\mathcal{T}$. For a product of two operators, the action of $\mathcal{T}$ is:
\begin{equation}\label{time_order}
\mathcal{T}\hat{O}_{1}(t_{1})\hat{O}_{2}(t_{2}) = \hat{O}_{1}(t_{1})\hat{O}_{2}(t_{2})\Theta(t_{1}-t_{2}) + \hat{O}_{2}(t_{2})\hat{O}_{1}(t_{1})\Theta(t_{2}-t_{1}).
\end{equation}
Here, $\Theta(t)$ is the Heaviside step function (equal to 1 when $t>0$ and 0 when $t<0$). When $\mathcal{T}$ acts on a product of $n$ operators, there will be a total of $n!$ terms on the right hand side (this is the number of ways that the operators could be ordered).
Applying $\mathcal{T}$ to \eqref{infinite_series} so that the influences of the $\tilde{\mathcal{L}}_{V}(t)$ are experienced in chronological order (thus preserving causality) gives:
\begin{align}
\tilde{\rho}(t) & = \hat{\rho}_{I}(0) + \mathcal{T}\sum_{k=1}^{\infty}\left[\frac{1}{k!}\left(\frac{-i}{\hbar}\right)^{k}\int_{0}^{t}dt_{1}\dots\int_{0}^{t}dt_{k}\tilde{\mathcal{L}}_{V}(t_{1})\dots\tilde{\mathcal{L}}_{V}(t_{k}) \right]\hat{\rho}_{I}(0)\nonumber \\
& = \mathcal{T}\sum_{k=0}^{\infty}\frac{1}{k!}\left(\frac{-i}{\hbar}\int_{0}^{t}dt_{1}\tilde{\mathcal{L}}_{V}(t_{1})\right)^{k}\tilde{\rho}(0).
\end{align}
\noindent This is the power series for the exponential function, so can be rewritten:
\begin{equation}\label{density_evolution}
\tilde{\rho}(t) = \mathcal{T}\exp\left(-\frac{i}{\hbar}\int_{0}^{t}dt_{1}\tilde{\mathcal{L}}_{V}(t_{1})\right)\tilde{\rho}(0).
\end{equation}
The above discussion is entirely general, and is suitable for any quantum-mechanical system, including the system-plus-environment supersystem we are considering. However, the environment has an infinite number of degrees of freedom, making it impossible to implement \eqref{density_evolution} in the form shown.
A popular method of dealing with this problem is to take a partial trace; the reduced density operator is a trace, over all bath degrees of freedom, of the total density operator:
\begin{equation}\label{reduced_density}
\tilde{\rho}_{S}(t) = \text{tr}_{B}[\tilde{\rho}(t)].
\end{equation}
Finally, we assume that the initial density matrix is given by $\tilde{\rho}(0) = \tilde{\rho}_{S}(0)\hat{\rho}_{B,\beta}$, where $\hat{\rho}_{B,\beta} = e^{-\beta \hat{H}_{B}}/Z_{B}$ is the canonical density matrix for the bath (note that no tilde is used in this case, since $\tilde{\rho}_{B,\beta} = \hat{\rho}_{B,\beta}$ and the canonical density operator is unchanged in the interaction representation).
This is justified in an electronic excitation process by invoking the Franck-Condon principle \cite{heom1}: when electronic excitation occurs, the baths remain in their equilibrium state because nuclear motion is slow on the timescale of electronic reorganization.
With the definition $\left< \tilde{O}(t) \right>_{\beta} = \text{tr}_{B}[\tilde{O}(t)\hat{\rho}_{B,\beta}]$, the time-evolution of the reduced system density matrix in the interaction picture is given by:
\begin{equation}\label{trace_rdo}
\tilde{\rho}_{S}(t) = \mathcal{T}\left< \exp\left(-\frac{i}{\hbar}\int_{0}^{t}dt_{1}\tilde{\mathcal{L}}_{V}(t_{1})\right)\right>_{\beta}\tilde{\rho}_{S}(0).
\end{equation}
This expression is extremely important, and will be revisited in the work of Chapter \ref{approximate}.
\section{Influence Operator}\label{infl}
Evaluating the bath average in \eqref{trace_rdo} will give an expression that depends not on the bath operators but only on terms that give the influence of the bath on the system. In order to achieve this, some observations must be made about the bath. Firstly, we set $\hat{H}_{0} = \hat{H}_{S}+\hat{H}_{B}$ and $\hat{V} = \hat{H}_{SB}$, so that in the remainder of this Chapter the interaction Liouvillian is $\tilde{\mathcal{L}}_{V}(t) = \tilde{\mathcal{L}}_{SB}(t)$.
\subsection{The Force Operator}\label{force_op}
The operator $\tilde{\xi}_{j}(t) = \sum_{\alpha}g_{j\alpha}\tilde{q}_{j\alpha}(t)$ gives the force exerted on the $j^{th}$ site by the bath oscillators $\alpha$ associated with this site. It is our aim here to find an explicit form for $\tilde{\xi}_{j}(t)$, and to show that the force exerted on a site by its bath is a random, Gaussian force, which will allow the bath average to be performed simply.
In the interaction picture, differentiating Eqn. \eqref{int_op} gives:
\begin{equation}
\frac{d}{d t}\tilde{O}(t) = \frac{i}{\hbar}\hat{\mathcal{L}}_{0}\tilde{O}(t).
\end{equation}
Using $\hat{H}_{0} = \hat{H}_{S}+\hat{H}_{B}$ gives $d\tilde{q}_{j\alpha}(t)/d t = \tilde{p}_{j\alpha}(t)/m_{j\alpha}$, and $d\tilde{p}_{j\alpha}(t)/d t = -m_{j\alpha}\omega_{j\alpha}^2\tilde{q}_{j\alpha}(t)$, whose solution is:
\begin{equation}\label{bath_position}
\tilde{q}_{j\alpha}(t) = \tilde{q}_{j\alpha}(0)\cos\left(\omega_{j\alpha}t\right) + \frac{\tilde{p}_{j\alpha}(0)}{m_{j\alpha}\omega_{j\alpha}}\sin\left(\omega_{j\alpha}t\right).
\end{equation}
From this we infer \cite{quantum_langevin} that the force operator at time $t$ is determined by the positions and momenta of the bath oscillators at time $0$, and so we now show that these positions and momenta are Gaussian variables. That is:
\begin{equation}\label{posn_gauss}
\left< f(\hat{q}_{j\alpha})\right>_{\beta} \propto \int_{-\infty}^{\infty}dq_{j\alpha}f(q_{j\alpha})e^{-\kappa q_{j\alpha}^2},
\end{equation}
\noindent for any function $f(\hat{q}_{j\alpha})$ of $\hat{q}_{j\alpha}$ in the Schr\"odinger representation, and that a similar relation will be true for functions of momentum.
Recalling the definition of the bath average, the trace can be evaluated in the position basis. The trace is then given by $\int\prod_{j,\alpha} dq_{j\alpha}\bra{q_{j\alpha}}\dots\ket{q_{j\alpha}}$. Since the canonical density matrix is a product $\hat{\rho}_{B,\beta}= \prod_{j}\hat{\rho}_{j,\beta} = \prod_{j,\alpha}\hat{\rho}_{j\alpha,\beta}$, integration over each coordinate $q_{j^{\prime}\alpha^{\prime}}$ gives the partition function $Z_{j^{\prime}\alpha^{\prime}}$, except for coordinate $q_{j\alpha}$:
\begin{align}
\left<f(\hat{q}_{j\alpha})\right>_{\beta} & = \int_{-\infty}^{\infty}dq_{j\alpha} \langle q_{j\alpha} | f(\hat{q}_{j\alpha})\hat{\rho}_{j\alpha,\beta}|q_{j\alpha}\rangle / Z_{j\alpha} \nonumber\\
& = \frac{1}{Z_{j\alpha}} \int_{-\infty}^{\infty}dq_{j\alpha}f(q_{j\alpha})\langle q_{j\alpha} | \hat{\rho}_{j\alpha,\beta}|q_{j\alpha}\rangle.
\end{align}
The canonical density matrix element $\bra{q_{j\alpha}}\hat{\rho}_{j\alpha,\beta}\ket{q_{j\alpha}}$ is perhaps most easily evaluated by the path integral formulation of quantum mechanics \cite{feynman}, which gives\footnote{This result is obtained by noting that the canonical density matrix $e^{-\beta\hat{H}}$ is analogous to the propagator $e^{-i\hat{H}t/\hbar}$, for ``thermal time'' $t=-i\beta\hbar$, and treating it using the path integral method for evaluating such propagators.}:
\begin{equation}
\left<f(\hat{q}_{j\alpha})\right>_{\beta} = \frac{1}{Z_{j\alpha}}\int_{-\infty}^{\infty}dq_{j\alpha}f(q_{j\alpha})\exp\left( -\frac{m_{j\alpha}\omega_{j\alpha}}{\hbar}\tanh\left(\frac{\beta\hbar\omega_{j\alpha}}{2}\right) q_{j\alpha}^2 \right).
\end{equation}
The positions and momenta of thermal bath oscillators are stochastic variables (due to the statistical nature of the problem) with Gaussian distributions. Since a linear combination of variables with Gaussian distributions is also Gaussian \cite{kubo}, we have the result that the force operator $\tilde{\xi}_{j}(t)$ acting on site $j$ has such a distribution.
\subsection{Evaluation Of Bath Average}\label{bath_average}
We can now eliminate the dependency of Eqn. \eqref{trace_rdo} on the operators of the bath. Using the identity $(\hat{A}\hat{B})^{\times} = \frac{1}{4}\lbrace \hat{A}^{\times},\hat{B}^{\circ}\rbrace + \frac{1}{4}\lbrace \hat{A}^{\circ},\hat{B}^{\times}\rbrace + \frac{1}{2}[\hat{A},\hat{B}]^{\times}$ for the commutator of a product of operators \cite{irreversible} gives:
\begin{align}
\tilde{\mathcal{L}}_{SB}(t) & = \sum_{j}^{\text{N}}\left(\tilde{\xi}_{j}(t)\tilde{V}_{j}(t)\right)^{\times} \nonumber\\
& \equiv \frac{1}{2}\sum_{j}^{\text{N}}\left( \tilde{\xi}_{j}(t)^{\circ}\tilde{V}_{j}(t)^{\times} + \tilde{\xi}_{j}(t)^{\times}\tilde{V}_{j}(t)^{\circ}\right).
\end{align}
We can infer that the interaction Liouvillian also has a Gaussian distribution with respect to a bath average, as it will also be a linear combination of the force operators. This distribution means that the statistical properties of $\tilde{\mathcal{L}}_{SB}(t)$ are fully determined by its autocorrelation function \cite{kubo}:
\begin{align}
\left<\tilde{\mathcal{L}}_{SB}(t)\tilde{\mathcal{L}}_{SB}(s)\right>_{\beta} = \frac{1}{2}\sum_{j}^{\text{N}} & \left( \tilde{V}_{j}(t)^{\times}\left<\tilde{\xi}_{j}(t)^{\circ}\tilde{\xi}_{j}(s)^{\circ}\right>_{\beta}\tilde{V}_{j}(s)^{\times} +\right.\nonumber\\
& \tilde{V}_{j}(t)^{\times}\left<\tilde{\xi}_{j}(t)^{\circ}\tilde{\xi}_{j}(s)^{\times}\right>_{\beta}\tilde{V}_{j}(s)^{\circ} +\nonumber\\
& \tilde{V}_{j}(t)^{\circ}\left<\tilde{\xi}_{j}(t)^{\times}\tilde{\xi}_{j}(s)^{\circ}\right>_{\beta}\tilde{V}_{j}(s)^{\times} +\nonumber\\
& \left.\tilde{V}_{j}(t)^{\circ}\left<\tilde{\xi}_{j}(t)^{\times}\tilde{\xi}_{j}(s)^{\times}\right>_{\beta}\tilde{V}_{j}(s)^{\circ}\right).
\end{align}
Noting that, for example, $\left<\tilde{\xi}_{j}(t)^{\circ}\tilde{\xi}_{j}(s)^{\circ}\right> = \text{tr}_{B,j}[\tilde{\xi}_{j}(t)^{\circ}\tilde{\xi}_{j}(s)^{\circ}\hat{\rho}_{j,\beta}]$, this trace will have four terms, and can be simplified (by cyclic permutation of the operators within the trace) to give $2\left<\lbrace\tilde{\xi}_{j}(t),\tilde{\xi}_{j}(s)\rbrace\right>_{\beta}$. The remaining three bath averages can be carried out (with the third and fourth vanishing) to give:
\begin{multline}\label{autocorr_lvl}
\left<\tilde{\mathcal{L}}_{SB}(t)\tilde{\mathcal{L}}_{SB}(s)\right>_{\beta} =\dots\\
\sum_{j}^{\text{N}} \tilde{V}_{j}(t)^{\times}\left(\alpha_{r,j}(t-s)\tilde{V}_{j}(s)^{\times} + i\alpha_{i,j}(t-s)\tilde{V}_{j}(s)^{\circ}\right).
\end{multline}
Here $\alpha_{r,j}(t) = \frac{1}{2}\left<\lbrace \tilde{\xi}_{j}(t),\tilde{\xi}_{j}(0)\rbrace\right>_{\beta}$ and $i\alpha_{i,j}(t) = \frac{1}{2}\left<\left[\tilde{\xi}_{j}(t),\tilde{\xi}_{j}(0)\right]\right>_{\beta}$, so that $\alpha_{j}(t) = \alpha_{r,j}(t) + i\alpha_{i,j}(t)$ is equal to $\left<\tilde{\xi}_{j}(t)\tilde{\xi}_{j}(0)\right>_{\beta}$, the autocorrelation function of the force exerted on the system by the bath.
Using the identity for a Gaussian process $x(t)$ \cite{kubo},
\begin{equation}\label{gaussian}
\left<\exp\left(\int_{0}^{t}dt_{1} f(t_{1})x(t_{1})\right)\right> = \exp\left(\frac{1}{2}\int_{0}^{t}dt_{1}\int_{0}^{t}dt_{2}f(t_{1})f(t_{2})\left<x(t_{1})x(t_{2})\right>\right),
\end{equation}
we finally obtain \cite{heom1}\footnote{Note that Eqn. \eqref{gaussian} has been adapted here due to the time-ordering; the operator $\mathcal{T}$ allows us to change the integration limits and removes the factor of $\frac{1}{2}$.}:
\begin{multline}\label{influence}
\tilde{\rho}_{S}(t) = \mathcal{T}\prod_{j}^{\text{N}}\exp\left(-\frac{1}{\hbar^2}\int_{0}^{t}dt_{1}\int_{0}^{t_{1}}dt_{2} \tilde{V}_{j}(t_{1})^{\times}\left[\alpha_{r,j}(t_{1}-t_{2})\tilde{V}_{j}(t_{2})^{\times}\right.\right.\\
+\left.\left.i\alpha_{i,j}(t_{1}-t_{2})\tilde{V}_{j}(t_{2})^{\circ}\right]\right)\tilde{\rho}_{S}(0).
\end{multline}
The operator acting on $\tilde{\rho}_{S}(0)$ is the influence operator \cite{heom1}. In the position representation, this becomes the influence functional of Feynman and Vernon \cite{influence,weiss}, which expresses the time-propagation of the reduced density matrix in terms only of statistical quantities of environment variables.
It is shown in Appendix \ref{QFDT} that the bath correlation function $\alpha_{j}(t)$ can be expressed in terms of the spectral density, such that $J_{j}(-\omega)=-J_{j}(\omega)$:
\begin{equation}\label{bath_corrfunc}
\alpha_{j}(t) = \int_{0}^{\infty}J_{j}(\omega)\left[\coth(\beta\hbar\omega/2)\cos(\omega t) - i\sin(\omega t)\right]d\omega.
\end{equation}
Thus, the influence operator encodes the effects of the baths (through the spectral densities) and of the temperature (through the $\coth(\beta\hbar\omega/2)$ term) on the system.
\section{Equations Of Motion}\label{EOM}
With an explicit form for the bath correlation function $\alpha_{j}(t)$, we can now find a differential equation to evolve the reduced density matrix. We take:
\begin{equation}\label{corrfunc_exp}
\alpha_{j}(t) = p_{j0}\delta_{+}(t) + \sum_{k=1}^{K}p_{jk}e^{-\gamma_{jk}t}.
\end{equation}
Here, $\delta_{+}(t)$ is a one-sided delta function, which has all of the properties of the delta function (pp. 58-60 of \cite{dirac}), but the integration limits in its definition are $(0,\infty)$ instead of $(-\infty,\infty)$. For example (and most importantly), $\int_{0}^{\infty}\delta_{+}(t)dt = 1$. We also use the notation $p_{jk} = a_{jk} + ib_{jk}$, where $a_{jk}$ and $b_{jk}$ are real.
The reasoning behind this expression is given in Chapter \ref{implementation}, and the advantage of this sum of exponentials will quickly become apparent in the following work.
\subsection{Auxiliary Density Operators}
Using \eqref{corrfunc_exp}, we can write the time-evolution of the density operator as \cite{tanimura2}
\begin{multline}\label{hierarchy1}
\tilde{\rho}_{S}(t) = \mathcal{T}\prod_{j}\exp\left(\frac{1}{\hbar^2}\int_{0}^{t}dt_{1}\tilde{\phi}_{j}(t_{1})\tilde{\theta}_{j,0}(t_{1})\right)\times\dots\\
\prod_{k}\exp\left(\frac{1}{\hbar^2}\int_{0}^{t}dt_{1}\int_{0}^{t_{1}}dt_{2}\tilde{\phi}_{j}(t_{1})e^{-\gamma_{jk}(t_{1}-t_{2})}\tilde{\theta}_{j,k}(t_{2})\right)\tilde{\rho}_{S}(0),
\end{multline}
\noindent where we have introduced the notation,
\begin{subequations}\label{notation}
\begin{equation}
\tilde{\phi}_{j}(t) = i\tilde{V}_{j}(t)^{\times},
\end{equation}
\begin{equation}
\tilde{\theta}_{j,k}(t) = ia_{jk}\tilde{V}_{j}(t)^{\times} - b_{jk}\tilde{V}_{j}(t)^{\circ}.
\end{equation}
\end{subequations}
We define a set of auxiliary density operators (ADOs) $\hat{\sigma}_{\textbf{n}}(t)$, indexed by the matrix $\textbf{n}$ whose elements are $n_{jk}$:
\begin{multline}\label{ADOs}
\tilde{\sigma}_{\textbf{n}}(t) = \mathcal{T}\prod_{j}\exp\left(\frac{1}{\hbar^2}\int_{0}^{t}dt_{1}\tilde{\phi}_{j}(t_{1})\tilde{\theta}_{j,0}(t_{1})\right)\times\dots\\
\prod_{k}\left(\frac{1}{\hbar^2} \int_{0}^{t}dt_{1}e^{-\gamma_{jk}(t-t_{1})}\tilde{\theta}_{j,k}(t_{1}) \right)^{n_{jk}}\times\dots\\
\exp\left(\frac{1}{\hbar^2}\int_{0}^{t}dt_{1}\int_{0}^{t_{1}}dt_{2}\tilde{\phi}_{j}(t_{1})e^{-\gamma_{jk}(t_{1}-t_{2})}\tilde{\theta}_{j,k}(t_{2})\right)\tilde{\rho}_{S}(0).
\end{multline}
The ADO with $\textbf{n} = \textbf{0}$ is the reduced density operator (RDO). The reason for these ADOs being introduced is made clear when the time-derivative of Eqn. \eqref{ADOs} is taken \cite{heom1}:
\begin{multline}\label{timederiv_ADO}
\frac{d}{d t}\tilde{\sigma}_{\textbf{n}}(t) = \sum_{j,k}\left( \tilde{\phi}_{j}(t)\tilde{\theta}_{j,0}(t) - n_{jk}\gamma_{jk}\right)\tilde{\sigma}_{\textbf{n}}(t) + \dots\\
\sum_{j,k}\left(\tilde{\phi}_{j}(t)\tilde{\sigma}_{\textbf{n}_{jk+}}(t) + n_{jk}\tilde{\theta}_{j,k}(t)\tilde{\sigma}_{\textbf{n}_{jk-}}(t)\right).
\end{multline}
The ADOs $\hat{\sigma}_{\textbf{n}}(t)$ and $\hat{\sigma}_{\textbf{n}_{jk\pm}}(t)$ have matrices $\textbf{n}$ that are identical, except that for the latter the element $n_{jk}$ is replaced by $n_{jk}\pm1$. For each ADO, we define:
\begin{equation}
\mathcal{M}_{\textbf{n}} = \sum_{j}^{\text{N}}\sum_{k}^{K}n_{jk}.
\end{equation}
And now, noting that $d\tilde{O}(t)/d t = e^{i\mathcal{L}_{0}t/\hbar}d\hat{O}/d t + \frac{i}{\hbar}\mathcal{L}_{0}e^{i\mathcal{L}_{0}t/\hbar}\hat{O}$, the time-derivative \eqref{timederiv_ADO} can be rearranged and rewritten in the Schr\"odinger picture:
\begin{multline}\label{heom_eqs}
\frac{d}{d t}\hat{\sigma}_{\textbf{n}}(t) = -\frac{i}{\hbar}\hat{\mathcal{L}}_{S}\hat{\sigma}_{\textbf{n}}(t) - \sum_{j,k}\left(a_{j0}\hat{V}_{j}^{\times}\hat{V}_{j}^{\times} - ib_{j0}\hat{V}_{j}^{\times}\hat{V}_{j}^{\circ} + n_{jk}\gamma_{jk} \right)\hat{\sigma}_{\textbf{n}}(t) + \dots\\
\sum_{j,k}\left(i\hat{V}_{j}^{\times}\hat{\sigma}_{\textbf{n}_{jk+}}(t) + in_{jk}p_{jk}\hat{V}_{j}\hat{\sigma}_{\textbf{n}_{jk-}}(t) + in_{jk}p_{jk}^{\ast}\hat{\sigma}_{\textbf{n}_{jk-}}(t)\hat{V}_{j} \right).
\end{multline}
We see now that these ADOs form a hierarchy: a given ADO's level in the hierarchy is given by its value of $\mathcal{M}_{\textbf{n}}$, with only one operator (the reduced density operator) in the lowest level, $\text{N}\times K$ operators in the first level, and so on. Each ADO is coupled to operators on the levels above and below.
An interesting qualitative picture of the HEOM is given by reference to creation and annihilation operators (pp. 136-139 of \cite{dirac}): to each element of $\textbf{n}$ we can assign a ``bath mode'' labelled by $jk$. Then, a given ADO $\hat{\sigma}_{\textbf{n}}$ has $n_{jk}$ quanta in the bath mode $jk$.
It is then possible to see that a given level of the hierarchy corresponds to a certain total number of quanta, and that each ADO is connected to the ADOs that can be formed from it by creating or annihilating one quantum.
By propagating these ADOs through time (we see in \ref{comments} that a finite number of these operators can be used), the time-evolution of the diagonal elements of the RDO (the site populations) can be found.
\subsection{Comments}\label{comments}
The remainder of this Chapter is given to a discussion of several important points about the HEOM.
\begin{itemize}
\item We have made no approximation in this derivation, so have a method that is numerically exact and able to describe the effects of quantum-mechanical coherence in energy transfer and of non-Markovian system-bath interactions.
\item For all ADOs but the RDO, the initial condition is that all elements of these matrices be equal to zero.
\item Other than the RDO, the ADOs cannot be system density matrices, as their traces are not conserved.
\item In principle, the hierarchy continues to infinity. However, this is not possible in practice. Integrating \eqref{heom_eqs}:
\begin{multline}
\hat{\sigma}_{\textbf{n}}(t) = \int_{0}^{t}dt_{1}\cdot e^{-\left(\frac{i}{\hbar}\hat{\mathcal{L}}_{S}+\sum_{j,k}\left(a_{j0}\hat{V}_{j}^{\times}\hat{V}_{j}^{\times} -ib_{j0}\hat{V}_{j}^{\times}\hat{V}_{j}^{\circ} + n_{jk}\gamma_{jk} \right)\right)(t-t_{1})}\times\\
\sum_{j,k}\left(i\hat{V}_{j}^{\times}\hat{\sigma}_{\textbf{n}_{jk+}}(t_{1}) + in_{jk}p_{jk}\hat{V}_{j}\hat{\sigma}_{\textbf{n}_{jk-}}(t_{1}) + in_{jk}p_{jk}^{\ast}\hat{\sigma}_{\textbf{n}_{jk-}}(t_{1})\hat{V}_{j} \right).
\end{multline}
For ADOs above some level $\mathcal{N}$ of the hierarchy, the values of $n_{jk}$ will be so large that $\exp\left(-\sum{j,k}n_{jk}\gamma_{jk}(t-t_{1})\right)$ will decay very rapidly and thus will be proportional to $\delta_{+}(t-t_{1})$. This gives the result \cite{tanimura2,heom1} that for these ADOs, the time derivative does not depend on ADOs in any other levels, so their elements will remain at their initial value of zero.
\item We obtain converged results by carrying out calculations with an increasing number of levels $\mathcal{N}$, until increasing this number has no further effect on the resulting dynamics.
\item For a calculation using $\mathcal{N}$ levels of the hierarchy, the number of ADOs required is \cite{ishizakifleming}:
\begin{equation}\label{num_ados}
\frac{\left(\mathcal{N}+(\text{N}K)\right)!}{\mathcal{N}!(\text{N}K)!}.
\end{equation}
For some calculations, this number can be extremely large, meaning that the HEOM can be very computationally expensive, requiring a lot of time and memory to run. We will wish to implement the equations as efficiently as possible; this will be discussed further in Chapter \ref{implementation}.
\end{itemize}
{\renewcommand{\baselinestretch}{1}\normalsize
\chapter{Approximate Methods}\label{approximate}
}
Although the HEOM provide a very powerful method for calculating energy transfer dynamics, the large computational cost incurred means that we will need to seek an approximate method with a much lower cost, but with dynamics that are as accurate as possible, since we will be unable to use the HEOM for several of the physical features of the FMO complex that we wish to investigate.
There are a great number of methods that are suitable candidates, both historical and modern. Some of these, such as the F\"orster (1948) \cite{forster1} and Redfield (1957) \cite{redfield1} theories, treat some physical quantity as a perturbation; these two methods will be presented in this Chapter.
Other methods include the Zeroth Order Functional Expansion quantum master equation (ZOFE, 2011) \cite{zofe1}, in which a key operator is treated as being independent of the bath, and which has recently been applied both to the 7-site FMO system \cite{zofe_fmo1} and the 24-site trimer \cite{zofe_fmo2}.
This more modern method will not be considered in this thesis, as preliminary calculations (not reported here) showed the ZOFE to be less accurate than simple F\"orster theory for the FMO complex at 300 K.
The existence of so many methods, all with different conditions for applicability, means that it is necessary for us to decide which we will use in our investigations, and which we need not consider any further. We turn now to the task of introducing two techniques whose validity we will test in Chapter \ref{numerical_results}.
\section{Redfield Theory}\label{redfield_theory}
The perturbative nature of both the Redfield and F\"orster theories \cite{redfield_forster} means that the validity of each depends on some parameter being small enough to justify it being treated as a perturbation. In the case of the Redfield theory, the coupling between the system and bath is assumed to be weak, characterized by reorganization energies $\lambda_{j}$ that are small (compared to dipolar couplings $J_{jk}$). Thus, $\hat{H}_{SB}$ is taken as the perturbation to the Hamiltonian.
In qualitative terms, we might consider an electronic system evolving under its own Hamiltonian $\hat{H}_{S}$, which is then weakly coupled to an environment. The phonons destroy phase information \cite{leegwater_klug}, assisting the relaxation of the system to an equilibrium state.
\subsection{Time-Dependent Redfield Theory}
We take as a starting point Eqn. \eqref{trace_rdo}, with $\tilde{\mathcal{L}}_{V}(t) = \tilde{\mathcal{L}}_{SB}(t)$. Taking the time-derivative gives:
\begin{align}
\frac{d}{d t}\tilde{\rho}_{S}(t) & = \left<-\frac{i}{\hbar}\mathcal{T}\tilde{\mathcal{L}}_{SB}(t)\exp\left(-\frac{i}{\hbar}\int_{0}^{t}dt_{1}\tilde{\mathcal{L}}_{SB}(t_{1})\right)\right>_{\beta}\tilde{\rho}_{S}(0)\nonumber\\
& \approx -\frac{i}{\hbar} \left<\tilde{\mathcal{L}}_{SB}(t)\right>_{\beta}\tilde{\rho}_{S}(0) - \frac{1}{\hbar^2}\int_{0}^{t}dt_{1}\left<\tilde{\mathcal{L}}_{SB}(t)\tilde{\mathcal{L}}_{SB}(t_{1})\right>_{\beta}\tilde{\rho}_{S}(0),\label{redfield_approx}
\end{align}
\noindent with the second line being the series expansion up to second order in the perturbation $\hat{H}_{SB}$. The first term on the right vanishes, since it is an average of a Gaussian variable, which is zero \cite{kubo}.
The correlation function appearing in the integrand is given explicitly by \eqref{autocorr_lvl}. We also replace $\tilde{\rho}_{S}(0)$ on the right hand side by $\tilde{\rho}_{S}(t)$, assuming that the weak perturbation causes only a negligible change in $\tilde{\rho}_{S}(t)$ between time $0$ and time $t$. Then, expanding the commutators and anticommutators \cite{heom1}:
\begin{multline}\label{redfield_timederiv}
\frac{d}{d t}\tilde{\rho}_{S}(t) = -\frac{1}{\hbar^2}\sum_{j}\int_{0}^{t}dt_{1}\left\lbrace \alpha_{j}(t-t_{1})\tilde{V}_{j}(t)\tilde{V}_{j}(t_{1})\tilde{\rho}_{S}(t)\right. +\\
-\alpha_{j}(t-t_{1})\tilde{V}_{j}(t_{1})\tilde{\rho}_{S}(t)\tilde{V}_{j}(t) - \alpha_{j}^{\ast}(t-t_{1})\tilde{V}_{j}(t)\tilde{\rho}_{S}(t)\tilde{V}_{j}(t_{1}) +\\
\left.\alpha_{j}^{\ast}(t-t_{1})\tilde{\rho}_{S}(t)\tilde{V}_{j}(t_{1})\tilde{V}_{j}(t)\right\rbrace.
\end{multline}
We choose to evaluate the density matrix in the exciton basis, where $\hat{H}_{S}\ket{\mu} = \hbar\omega_{\mu}\ket{\mu}$, because this means that the matrix elements of $\tilde{V}_{j}(t)$ have the form:
\begin{equation}
\bra{\mu}e^{i\hat{H}_{S}t/\hbar}\hat{V}e^{-i\hat{H}_{S}t/\hbar}\ket{\nu} = e^{i\omega_{\mu}t}\bra{\mu}\hat{V}\ket{\nu}e^{-i\omega_{\nu}t}.
\end{equation}
The density matrix can be transformed between exciton and site representation by using the matrix $\hat{U}$ of Eqn. \eqref{exciton_basis} that diagonalizes $\hat{H}_{S}$ \cite{leegwater_klug}.
In finding the matrix elements of \eqref{redfield_timederiv}, the completeness relation is used: $\sum_{\mu'}\ket{\mu'}\bra{\mu'}=\hat{1}$. For the first term on the right-hand side, the matrix elements are given by:
\begin{multline}\label{multiline_redfield}
-\frac{1}{\hbar^2}\sum_{j}\int_{0}^{t}dt_{1}\alpha_{j}(t-t_{1})\bra{\mu}\tilde{V}_{j}(t)\tilde{V}_{j}(t_{1})\tilde{\rho}_{S}(t)\ket{\nu} = \\
-\frac{1}{\hbar^2}\sum_{j}\sum_{\kappa}\sum_{\mu'}\int_{0}^{t}dt_{1}\cdot \alpha_{j}(t-t_{1})e^{i(\omega_{\mu}-\omega_{\nu})t}e^{i(\omega_{\mu'}-\omega_{\kappa})(t-t_{1})}\times\\
\bra{\mu}\hat{V}_{j}\ket{\kappa}\bra{\kappa}\hat{V}_{j}\ket{\mu'}\bra{\mu'}\hat{\rho}_{S}(t)\ket{\nu}.
\end{multline}
\noindent Defining a ``partial Fourier transform'',
\begin{equation}\label{partial_fourier}
\alpha_{j}[\omega;t] = \int_{0}^{t}\alpha_{j}(t_{1})e^{i\omega t_{1}}dt_{1},
\end{equation}
\noindent and changing the variable of integration, Eqn. \eqref{multiline_redfield} can be rewritten as:
\begin{equation}
-\frac{e^{i(\omega_{\mu}-\omega_{\nu})t}}{\hbar^2}\sum_{j}\sum_{\mu'}\sum_{\nu'}\delta_{\nu\nu'}\sum_{\kappa}\alpha_{j}[\omega_{\mu'}-\omega_{\kappa};t]\inner{\mu}{j}\inner{j}{\kappa}\inner{\kappa}{j}\inner{j}{\mu'}\rho_{\mu',\nu'}(t).
\end{equation}
\noindent Here, $\rho_{\mu',\nu'}(t) = \bra{\mu'}\hat{\rho}(t)\ket{\nu'}$, and $\inner{j}{\kappa} = U_{\kappa j}$, as in Eqn. \eqref{exciton_basis}.
Treating the other three terms in \eqref{redfield_timederiv} in the same manner gives the Time-Dependent Redfield equation \cite{leegwater_klug,i&f_redfield}:
\begin{equation}\label{tdredf}
\frac{d}{d t}\rho_{\mu,\nu}(t) = -\frac{i}{\hbar}(\omega_{\mu}-\omega_{\nu})\rho_{\mu,\nu}(t) - \sum_{\mu'}\sum_{\nu'}R_{\mu,\nu,\mu',\nu'}(t)\rho_{\mu',\nu'}(t).
\end{equation}
The key quantity in the Redfield theory is the matrix $R_{\mu,\nu,\mu',\nu'}(t)$, which governs the relaxation of the electronic system to equilibrium. It is given by:
\begin{multline}\label{relax_matr}
R_{\mu,\nu,\mu',\nu'}(t) = \Gamma_{\nu',\nu,\mu,\mu'}(t) + \Gamma_{\mu',\mu,\nu,\nu'}^{\ast}(t) \\ - \delta_{\nu\nu'}\sum_{\kappa}\Gamma_{\mu,\kappa,\kappa,\mu'}(t) - \delta_{\mu\mu'}\sum_{\kappa}\Gamma_{\nu,\kappa,\kappa,\nu'}^{\ast}(t).
\end{multline}
\noindent Here $\Gamma_{\mu,\nu,\mu',\nu'}(t)$ is the time-dependent damping matrix, and its elements are \cite{i&f_redfield,tanimura_redfield}:
\begin{equation}
\Gamma_{\mu,\nu,\mu',\nu'}(t) = \frac{1}{\hbar^2}\sum_{j}\inner{\mu}{j}\inner{j}{\nu}\inner{\mu'}{j}\inner{j}{\nu'}\alpha_{j}[\omega_{\nu'} - \omega_{\mu'};t].
\end{equation}
The time-dependence of the relaxation matrix \eqref{relax_matr} highlights the non-Markovian nature of the Time-Dependent Redfield theory: Eqn. \eqref{partial_fourier} is responsible for memory effects \cite{ishizaki_review,tanimura_redfield}, and using Eqn. \eqref{corrfunc_exp}, this function is given by \cite{tanimura_redfield}:
\begin{equation}
\alpha_{j}[\omega;t] = p_{j0} + \sum_{k=1}^{K}p_{jk}\frac{1-e^{(i\omega - \gamma_{jk}) t}}{\gamma_{jk}-i\omega}.
\end{equation}
\subsection{Discussion}
The following points are salient when considering Redfield theory:
\begin{itemize}
\item Due to the assumption of weak system-bath coupling in Eqn. \eqref{redfield_approx}, the full dynamics are not captured: the perturbation $\hat{H}_{SB}$ is proportional to a linear combination of bath coordinates, so can only induce single-phonon transitions in the bath \cite{redfield_forster} (there will only be non-zero matrix elements between bath states whose energy levels are adjacent).
\item In its original form \cite{redfield1}, the Redfield theory was Markovian. This theory can be derived simply from the Time-Dependent Redfield theory by assuming that $e^{-\gamma t}\approx \gamma^{-1}\delta_{+}(t)$ in Eqn. \eqref{partial_fourier}: qualitatively, by time $t$, the integrand has already decayed to zero, so that the upper bound of the integral can be set to infinity without appreciating its value.
\item The only change made to the Time-Dependent theory to derive the Markovian theory is to replace $\alpha_{j}[\omega;t]$ by $\alpha_{j}[\omega] = \alpha_{j}[\omega;t\rightarrow\infty]$ \cite{ishizaki_review}, giving a relaxation matrix $R_{\mu,\nu,\mu',\nu'}$ that is time-independent.
\item When implementing the Time-Dependent Redfield theory to evolve the density matrix through time, the relaxation matrix \eqref{relax_matr} must be calculated at each timestep, whereas it is not dependent on time in the Markovian Redfield case so only needs to be calculated once.
\item More recently, a modified Redfield theory has been developed \cite{redfield_forster,modified_redfield}, which can interpolate between the Redfield and F\"orster theories. However, this theory ignores off-diagonal density matrix elements in the exciton basis, so cannot be used to describe dynamics in the site basis \cite{fleming_lightharv}.
\end{itemize}
\section{F\"orster Theory}\label{forster_theory}
F\"orster theory, or F\"orster-Dexter theory \cite{forster1,forster2,dexter}, was originally formulated to describe resonant energy transfer (RET) between two electronic states (a donor and an acceptor) \cite{RET}.
More recently, Jang \emph{et al.} introduced a generalized F\"orster theory \cite{silbey}, which takes into account situations in which environmental phonons have not relaxed to equilibrium before energy transfer occurs.
In this Section, we derive the original F\"orster theory, and then explain how to calculate the resulting energy transfer rate constants.
\subsection{Resonant Energy Transfer Rate}
For clarity, we begin by considering two sites, labelled 1 and 2, each of whose electronic states are associated with a continuum of vibrational modes, so that when there is no interaction between the sites, the vibronic states are,
\begin{equation}\label{vibronic_states}
\ket{\psi(E_{1},E_{2})},
\end{equation}
\noindent where $E_{j}$ is the vibronic energy of site $j$. $\hat{H}_{0}$ is the unperturbed Hamiltonian:
\begin{equation}
\hat{H}_{0} = \sum_{j=1}^{2}\hbar\omega_{j}\ket{j}\bra{j} + \hat{H}_{B} + \hat{H}_{SB},
\end{equation}
\noindent such that $\ket{\psi(E_{1},E_{2})}$ is an eigenstate of $\hat{H}_{0}$:
\begin{equation}
\bra{\psi(E_{1},E_{2})}\hat{H}_{0}\ket{\psi(E_{1},E_{2})} = E_{1} + E_{2}.
\end{equation}
Now, we add an interaction $\hat{J}$ (which has no diagonal matrix elements between vibronic states) to the Hamiltonian (note that this gives $\hat{H}_{0} + \hat{J} = \hat{H}$, the full Hamiltonian):
\begin{equation}
\hat{J} = \sum_{j,k\ne j}\hbar J_{jk}(\ket{j}\bra{k} + \ket{k}\bra{j}).
\end{equation}
By construction, this interaction is weak (its matrix elements are smaller than the reorganization energies $\lambda_{j}$) and treating it as a perturbation, it will induce vibronic transitions, leading to energy transfer between sites.
We take site 1 to be the ``donor'' and site 2 the ``acceptor'', with the former being electronically excited initially and the latter being unexcited, giving an initial state denoted $\ket{\psi(\epsilon_{1},\epsilon_{2})}$ (where $\epsilon_{1}$ is a vibrational level of the excited electronic state of site 1 and likewise $\epsilon_{2}$ for the ground electronic state of site 2).
Invoking Fermi's Golden Rule (\cite{forster_original}, p. 178 of \cite{dirac}) gives the probability of being in some state $\ket{\psi(E_{1},E_{2})}$ as a function of time. Since this probability is the square modulus of a single probability amplitude, we are ignoring phase information and any theory built on this framework will be incoherent. The probability is given by:
\begin{equation}
P(E_{1},E_{2},t;\epsilon_{1},\epsilon_{2},0) = |\bra{\psi(\epsilon_{1},\epsilon_{2})}\hat{J}\ket{\psi(E_{1},E_{2})}|^{2} \cdot \frac{4\sin^{2}(\Delta E t/2\hbar)}{\Delta E^{2}},
\end{equation}
\noindent where $\Delta E = (\epsilon_{1}+\epsilon_{2}) - (E_{1} + E_{2})$ is the energy difference between the initial and final vibronic states. The total rate of transfer from site 1 to site 2 is given by integrating over all final energies $E_{1}$ and $E_{2}$ (where $E_{1}$ is a vibrational level of the ground electronic state of site 1 and likewise $E_{2}$ for the excited electronic state of site 2):
\begin{multline}
P_{1\rightarrow 2}(t;\epsilon_{1},\epsilon_{2},0) = \\ \int_{0}^{\infty}dE_{1}\int_{0}^{\infty}dE_{2}|\bra{\psi(\epsilon_{1},\epsilon_{2})}\hat{J}\ket{\psi(E_{1},E_{2})}|^{2} \cdot \frac{4\sin^{2}(\Delta E t/2\hbar)}{\Delta E^{2}}.
\end{multline}
The integrand contains a function of the form $\sin^{2}(\kappa x t)/x^{2}$ (where $\kappa$ is a constant), which in the limit as $t\rightarrow \infty$ is proportional to the delta-function. We thus make the approximation \cite{forster_original}:
\begin{equation}
\frac{4\sin^{2}(\Delta E t/2\hbar)}{\Delta E^{2}} \approx \frac{\pi t}{\hbar^{2}}\delta\left(\frac{\Delta E}{2\hbar}\right) = \frac{2\pi t}{\hbar}\delta(\Delta E),
\end{equation}\label{delta_approx}
\noindent and use the property (p. 60 of \cite{dirac}):
\begin{align}
\delta(\Delta E) = \delta(\epsilon_{1}+\epsilon_{2}-E_{1}-E_{2}) & = \int_{-\infty}^{\infty}d(\hbar\omega)\cdot\delta(\hbar\omega+\epsilon_{2}-E_{2})\delta(\hbar\omega-\epsilon_{1}+E_{1}) \nonumber\\
& = \hbar\int_{-\infty}^{\infty}d\omega\cdot\delta(E_{2}-[\epsilon_{2}+\hbar\omega])\delta(E_{1}-[\epsilon_{1}-\hbar\omega]).
\end{align}
This gives:
\begin{multline}\label{delta_fns}
P_{1\rightarrow 2} = 2\pi t\int_{-\infty}^{\infty}d\omega\int_{0}^{\infty}dE_{1}\int_{0}^{\infty}dE_{2} |\bra{\psi(\epsilon_{1},\epsilon_{2})}\hat{J}\ket{\psi(E_{1},E_{2})}|^{2}\times \\
\delta(E_{2}-[\epsilon_{2}+\hbar\omega])\delta(E_{1}-[\epsilon_{1}-\hbar\omega]),
\end{multline}
\noindent and we see, since $\Delta E = 0$ (due to the approximation in Eqn. \eqref{delta_approx}), that energy is conserved, and from Eqn. \eqref{delta_fns}, that $\hbar\omega$ is the energy transferred from donor to acceptor \cite{forster_original}.
Finally, the rate constant for energy transfer is found by using the definition $\int_{-\infty}^{\infty}f(x)\delta(x-a)dx = f(a)$ and taking the time-derivative of the probability:
\begin{equation}\label{forster_rate}
k_{1\rightarrow 2}^{F}(\epsilon_{1},\epsilon_{2}) = 2\pi\int_{-\infty}^{\infty}d\omega|\bra{\psi(\epsilon_{1},\epsilon_{2})}\hat{J}\ket{\psi(\epsilon_{1}-\hbar\omega,\epsilon_{2}+\hbar\omega)}|^{2}.
\end{equation}
Since the probabilities depend linearly on time, the rate constants are time-independent. Next, we will rewrite the integrand in terms of the more familiar functions of open quantum systems.
\subsection{Spectral Overlap}\label{spectral_overlap}
We now assume that the system is initially at equilibrium, so that we do not precisely know the initial vibronic energy; rather, we know that the probability density of being in state $\ket{\psi(\epsilon_{1},\epsilon_{2})}$ is given by $g(\epsilon_{1})g(\epsilon_{2})$, where $g(\epsilon) = e^{-\beta\epsilon}/Z$ and $Z = \int_{0}^{\infty}e^{-\beta\epsilon}d\epsilon$ is the canonical partition function. The total rate of transfer is given by:
\begin{multline}
k_{1\rightarrow 2}^{F} =2\pi\int_{-\infty}^{\infty}d\omega\int_{0}^{\infty}d\epsilon_{1}\int_{0}^{\infty}d\epsilon_{2}\cdot g(\epsilon_{1})g(\epsilon_{2})\times\\
|\bra{\psi(\epsilon_{1},\epsilon_{2})}\hat{H}\ket{\psi(\epsilon_{1}-\hbar\omega,\epsilon_{2}+\hbar\omega)}|^{2}.
\end{multline}
Within the Born-Oppenheimer approximation, the vibronic state is given as a product of electronic and vibrational states:
\begin{equation}
\ket{\psi(E_{1},E_{2})} = \ket{\psi_{\text{elec}}}\ket{\chi_{1}(E_{1})}\ket{\chi_{2}(E_{2})},
\end{equation}
\noindent where $\ket{\chi_{j}(E_{j})}$ is the vibrational state of site $j$. We also take $\ket{1}$ as the initial electronic state, with only the donor excited, and $\ket{2}$ as the final electronic state, with only the acceptor excited. Then, using $\langle 1|\hat{J}|2\rangle = \hbar J_{12}$:
\begin{multline}
k_{1\rightarrow 2}^{F} = 2\pi\hbar^{2}|J_{12}|^{2}\int_{-\infty}^{\infty}d\omega \left(\int_{0}^{\infty}d\epsilon_{1}\cdot g(\epsilon_{1}) |\langle\chi_{1}(\epsilon_{1})|\chi_{1}(\epsilon_{1}-\hbar\omega)\rangle|^{2}\right)\times\\
\left(\int_{0}^{\infty}d\epsilon_{2}\cdot g(\epsilon_{2}) |\langle\chi_{2}(\epsilon_{2})|\chi_{2}(\epsilon_{2}+\hbar\omega)\rangle|^{2}\right).
\end{multline}
The fluorescence and absorption spectral line shapes, $F_{1}[\omega]$ and $A_{2}[\omega]$ respectively, are defined in terms of Franck-Condon factors $S^{2}(\epsilon_{j},\epsilon_{j}\pm\hbar\omega) = |\langle\chi_{j}(\epsilon_{j})|\chi_{j}(\epsilon_{j}\pm\hbar\omega)\rangle|^{2}$ \cite{forster_original}:
\begin{subequations}
\begin{equation}
F_{1}[\omega] = 2\pi\hbar\int_{0}^{\infty}d\epsilon_{1}\cdot g(\epsilon_{1}) S^{2}(\epsilon_{1},\epsilon_{1}-\hbar\omega),
\end{equation}
\begin{equation}
A_{2}[\omega] = 2\pi\hbar\int_{0}^{\infty}d\epsilon_{2}\cdot g(\epsilon_{2}) S^{2}(\epsilon_{2},\epsilon_{2}+\hbar\omega).
\end{equation}
\end{subequations}
\noindent The rate constant is given by:
\begin{equation}\label{overlap_equation}
k_{1\rightarrow 2}^{F} = \frac{1}{2\pi}|J_{12}|^{2}\int_{-\infty}^{\infty}F_{1}[\omega]A_{2}[\omega]d\omega.
\end{equation}
All that is needed now is a method to calculate these spectra using known parameters such as the bath correlation functions $\alpha_{j}(t)$, site energies $\hbar\omega_{j}$ and reorganization energies $\hbar\lambda_{j}$.
This connection is made in \cite{redfield_forster,chromophore_solvent,mukamel}, where the fluorescence and absorption spectra are given by $F_{1}[\omega] = \int_{-\infty}^{\infty}F_{1}(t)e^{i\omega t}dt$ and $A_{2}[\omega] = \int_{-\infty}^{\infty}A_{2}(t)e^{i\omega t}dt$. In the time-domain, the fluorescence and absorption line-shape functions are evaluated using the cumulant expansion method \cite{mukamel,cumulant}, giving:
\begin{subequations}
\begin{equation}
F_{1}(t) = e^{-i(\omega_{1}-\lambda_{1})t-g_{1}^{\ast}(t)},
\end{equation}
\begin{equation}
A_{2}(t) = e^{-i(\omega_{2}+\lambda_{1})t-g_{2}(t)}.
\end{equation}
\noindent The function $g_{j}(t)$ is defined:
\begin{equation}
g_{j}(t) = \frac{1}{\hbar^{2}}\int_{0}^{t}dt_{1}\int_{0}^{t_{1}}dt_{2}\cdot\alpha_{j}(t_{2}).
\end{equation}
\end{subequations}
$\hbar\lambda_{j}$ is the reorganization energy of site $j$, the energy difference between the excited vibrational level reached by a Franck-Condon transition and the ground vibrational level, in a given electronic state \cite{ishizaki_review}.
Parseval's theorem states that the overlap integral in Eqn. \eqref{overlap_equation} can be carried out in either domain, so that with some rearrangement, and assuming that energy transfer in a donor/acceptor pair is not affected by the presence of other sites, the F\"orster rate constant for transfer between two electronic states $\ket{j}$ and $\ket{k}$ is given by the equation ($\Re$ denotes the real part) \cite{redfield_forster}:
\begin{equation}\label{forster_rateconst}
k_{j\rightarrow k}^{F} = 2|J_{jk}|^{2}\Re\left( \int_{0}^{\infty}F_{j}(t)A_{k}(t)dt\right).
\end{equation}
Fig. \ref{overlap} illustrates the spectra $F_{1}[\omega]$ and $A_{2}[\omega]$, showing their overlap at both 77 K and 300 K. The effect of the function $g_{j}(t)$ in $F_{1}(t)$ and $A_{2}(t)$ is to broaden the spectra, from the delta-functions observed if $\alpha_{j}(t) = 0$ (that is, in the absence of baths).
\begin{figure}[!ht]
\begin{center}
\includegraphics[width=13.5cm, keepaspectratio=true]{Spectra.png}
\caption[Overlap of fluorescence spectrum of site 1 and absorption spectrum of site 2.]{\small{Overlap of the fluorescence spectrum of site 1 and the absorption spectrum of site 2 in the FMO complex at (a) 77 K and (b) 300 K.}}
\label{overlap}
\end{center}
\end{figure}
At the higher temperature, the bath correlation function is larger, the spectral lines are broader and there is more overlap between them, thus increasing the rate of transfer from site 1 to site 2.
Chapter \ref{numerical_results} will test the applicability of the F\"orster theory by comparing the predicted dynamics to the exact dynamics of the HEOM. The quality of the predictions of F\"orster theory will allow us to determine whether or not coherence is important in the physics of the FMO.
{\renewcommand{\baselinestretch}{1}\normalsize
\chapter{Implementation Of The HEOM}\label{implementation}
}
The number of ADOs required for a HEOM calculation, Eqn. \eqref{num_ados}, can be very large, and the computational cost of propagating these matrices through time is high. The first Section of this Chapter explains how the HEOM are solved numerically, with an eye towards maximizing efficiency.
The second Section explains why the bath correlation function $\alpha(t)$ can be represented as a sum of exponentials (which is essential for the HEOM), and explores both the Matsubara and the Pad\'e spectral decompositions, highlighting the greater accuracy that can be achieved by using the latter.
\section{Numerical Propagation}
Eqn. \eqref{heom_eqs} provides a set of coupled differential equations for all of the ADOs. If $\underline{\sigma}(t)$ is a vector containing all of the elements of all ADOs, the equations can be rewritten:
\begin{equation}\label{de_a_sigma}
\frac{d}{d t}\underline{\sigma}(t) = f\left(\underline{\sigma}(t),t\right) \equiv \underline{\underline{\mathcal{A}}}\cdot\underline{\sigma}(t).
\end{equation}
The second form follows from the fact that the differential equations do not contain the elements of the ADOs to any order greater than the first: $\underline{\underline{\mathcal{A}}}$ is a linear operator. We now look at the numerical solution of these equations.
Originally, the well-established fourth-order Runge-Kutta (RK4) method was chosen for numerical integration (Sec. 15.1 of \cite{numerical_recipes}). If timestep $\delta t$ is used for the integration, then the error is of the order of $\delta t^{5}$. However, an alternative numerical solution was noted. The formal solution to Eqn. \eqref{de_a_sigma} is $\underline{\sigma}(t+\delta t) = \exp(\underline{\underline{A}}\delta t)\underline{\sigma}(t)$ which can be Taylor-series expanded:
\begin{equation}
\underline{\sigma}(t+\delta t) = \sum_{m=0}^{M}\frac{\delta t^{m}}{m!}\underline{\underline{A}}^{m}\underline{\sigma}(t) + \mathcal{O}(\delta t^{M+1}).
\end{equation}
If $M = 4$ is chosen, the error will again be on the order of $\delta t^{5}$. The Taylor series method has an advantage that is not immediately obvious: it leads to a reduction in the memory required to run a simulation.
The reason for this is that the RK4 method involves calculation of four vectors $\underline{k}_{1}$, $\underline{k}_{2}$, $\underline{k}_{3}$ and $\underline{k}_{4}$. It is possible to use only two of these vectors, with one used for permanent storage of $\underline{k}_{1}$ and the other updated so that it will variously store the remaining three vectors.
On the other hand, the Taylor-series method requires that only one such vector, $\underline{k}_{1}$, is used. The program will loop through four steps, each time calculating $\frac{\delta t}{m}\underline{\underline{A}}\cdot\underline{k}_{1}$ and adding this to the current $\underline{\sigma}(t)$.
The RK4 method requires that the program for implementing the HEOM stores one more vector that contains as many elements as there are in all of the ADOs (that is, $\mathcal{N}\times\text{N}^{2}$, where $\mathcal{N}$ is as in Eqn. \eqref{num_ados}). This can lead to a significant reduction in memory usage.
The time-derivative of each ADO depends on the values of operators on the levels above and below its own. The number of ADOs in a given level is potentially huge, so that an efficient indexing system is vital: before the time-evolution begins, an indexing matrix is set up, each of whose rows corresponds to a single ADO, and whose columns contain references to the operators to which the ADO of this row is coupled, for ease of use at runtime.
For a given problem, we will have a system Hamiltonian $\hat{H}_{S}$ and a bath characterized by $\alpha(t)$, a sum of exponentials. By increasing the number of these exponentials, our $\alpha(t)$ will approach the true function. This number and the number of hierarchical levels may be increased until the dynamics converge and are numerically exact.
Within the program to carry out the HEOM, both the indexing routine and the calculation of the time-derivatives were parallelized: running on multiple CPUs allowed a faster, more efficient program.
The routine calculating the time-derivatives involves (when $\alpha(t)$ contains no $\delta_{+}$-function) the calculation of 6 matrix products for each ADO, or $6 N^3$ scalar multiplications. However, for the commutator and anticommutator of $\hat{V}_{j}$ with a general operator $\hat{O}$, a general matrix element is given by:
\begin{subequations}
\begin{equation}
\bra{m}\hat{V}_{j}^{\times}\hat{O}\ket{n} = \delta_{m j}\bra{n}\hat{O}\ket{n} - \bra{m}\hat{O}\ket{j}\delta_{j n},
\end{equation}
\begin{equation}
\bra{m}\hat{V}_{j}^{\circ}\hat{O}\ket{n} = \delta_{m j}\bra{j}\hat{O}\ket{n} + \bra{m}\hat{O}\ket{j}\delta_{j n}.
\end{equation}
\end{subequations}
Using this formula instead of explicitly calculating the commutators and anticommutators will give only $2 N^{3}$ scalar multiplications (in finding the commutator $\hat{\mathcal{L}}\hat{\sigma}_{\textbf{n}}(t)$), and led to a substantial speeding up of the program, thus increasing efficiency.
\section{Spectral Decomposition}
The Lorentz-Drude spectral density is frequently used for calculations because it allows the bath correlation function to be written analytically as a sum of decaying exponentials, meaning that the HEOM can be used to benchmark our calculations.
Traditionally, the integral in Eqn. \eqref{bath_corrfunc} (or equivalently \eqref{bath_corrfunc_infty}) is carried out as described below to give the Matsubara series \cite{ishizakifleming,tanimura2,filtering1,filtering2}; this series converges very slowly to an exact result, and so we describe an alternative, the Pad\'e series, which is found to converge very much faster.
\subsection{Matsubara Series}
Inserting the definition of $J(\omega)$ from Eqn. \eqref{drude_lorentz}, with subscripts dropped for clarity, into \eqref{bath_corrfunc} and making use of the fact that the integrand is an even function gives the integral:
\begin{equation}
\alpha(t) = \frac{\gamma\lambda\hbar}{\pi}\int_{-\infty}^{\infty}d\omega \left[ \frac{\omega\coth(\beta\hbar\omega/2)e^{-i\omega t}}{\omega^2 + \gamma^2} + \frac{\omega e^{-i\omega t}}{\omega^2 + \gamma^2}\right].
\end{equation}
This can be solved straightforwardly using contour integration: the integrand has poles at $\omega = \pm i\gamma$, and (for the first term) at $\sinh(\beta\hbar\omega/2) = 0 \Rightarrow \omega = \pm i 2\pi j / \beta\hbar$, with $j$ an integer. As we are only interested in $\alpha(t)$ for $t > 0$, we use the contour below, with $R\rightarrow\infty$:
\setlength{\unitlength}{7cm}
\begin{center}
\begin{picture}(1,1)
\put(-0.25,0.5){\vector(1,0){1.5}}
\put(0.5,0){\vector(0,1){1}}
\put(0.55,1){Im$(\omega)$}
\put(1.25,0.43){Re$(\omega)$}
\linethickness{0.5mm}
\put(0.05,0.5){\line(1,0){0.9}}
\put(0.6,0.5){\line(-1,1){0.03}}
\put(0.6,0.5){\line(-1,-1){0.03}}
\put(0.5,0,5){\arc[180,360]{0.45}}
\put(0.94,0.55){$R$}
\put(0.01,0.55){$-R$}
\end{picture}
\end{center}
The integral over the semi-circle vanishes (pp. 113-115 of \cite{complex_analysis}), so that the integral over the real line is given by a sum of residues:
\begin{align}
\alpha(t) & = -2\gamma\lambda\hbar i \left\lbrace\Res_{\omega = -i\gamma} + \sum_{k=1}^{\infty}\Res_{\omega = -i2\pi k / \beta\hbar}\right\rbrace \cdot \left\lbrace \frac{\omega\coth(\beta\hbar\omega/2)e^{-i\omega t}}{\omega^2 + \gamma^2} + \frac{\omega e^{-i\omega t}}{\omega^2 + \gamma^2}\right\rbrace \nonumber\\
& = \gamma\lambda\hbar \left\lbrace \cot\left(\frac{\beta\hbar\gamma}{2}\right) - i\right\rbrace e^{-\gamma t} + \sum_{k=1}^{\infty}\frac{4\gamma\lambda\nu_{k}}{\beta\left(\nu_{k}^2 - \gamma^2\right)}e^{-\nu_{k}t}.
\end{align}
Here, $\nu_{k} = 2\pi k/\beta\hbar$ is called a Matsubara frequency and we have a sum of decaying exponentials, as desired. it is possible to rewrite $\cot(\theta)$ as an infinite sum by applying the residue theorem once again (pp. 131-133 of \cite{complex_analysis}), giving the form used by Ishizaki and Fleming \cite{ishizakifleming}:
\begin{equation}\label{matsubara2}
\alpha(t) = \frac{2\lambda}{\beta}\left\lbrace 1-\sum_{k=1}^{\infty}\frac{2\gamma^{2}}{\nu_{k}^2 - \gamma^2} - i\frac{\beta\hbar\gamma}{2} \right\rbrace e^{-\gamma t} + \frac{2\lambda}{\beta}\sum_{k=1}^{\infty}\frac{2\gamma\nu_{k}}{\nu_{k}^2 - \gamma^2}e^{-\nu_{k}t}.
\end{equation}
It is from this expression that a truncation is suggested: a certain number $K$ of exponential terms are retained, and for $k > K$, a Markovian approximation is used: it is assumed that $\nu_{k}$ is large enough that $\nu_{k}e^{-\nu_{k}t} \approx \delta_{+}(t)$, or equivalently that these terms decay quickly enough that they are essentially zero for $t > 0$, and only contribute at $t = 0$.
In practice, one would successively increase $K$ until converged results were found for the dynamics. Fig. \ref{matsubara_figure} shows $\alpha(t)$ for some values of $K$.
\begin{figure}[!ht]
\begin{center}
\includegraphics[width=14cm, keepaspectratio=true]{matsubara_2.png}
\caption[Convergence of Matsubara series.]{\small{Convergence of the Matsubara series at 77 K with varying number of exponential terms $K$, for the Lorentz-Drude spectral density. \emph{Inset: Magnified bath correlation functions at short times.}}}
\label{matsubara_figure}
\end{center}
\end{figure}
We see that a large value of $K$ would be required to fully converge the HEOM calculations, and so we next turn our attention to a better method of representing $\alpha(t)$.
\subsection{Pad\'e Series}\label{section_pade}
Faster convergence can be obtained by finding the Pad\'e approximant to the Bose-Einstein function, $f_{Bose}(x) = (1-e^{-x})^{-1}$ \cite{pade1,pade2}. That is, we find a rational function that approximates $f_{Bose}(x)$. We will focus on the [$N$-1/$N$] approximant (for reasons that will be discussed at the end of this Section), which uses a fraction of the form:
\begin{equation}\label{n-1/n}
\frac{\sum_{j=0}^{N-1}a_{j}x^{j}}{\sum_{k=0}^{N}b_{k}x^{k}}.
\end{equation}
In Appendix \ref{pade_approximant}, it is shown that the Bose-Einstein function can be approximated \cite{pade2}:
\begin{equation}\label{pade_bose}
f_{Bose}(x) \approx \frac{1}{x} + \frac{1}{2} + 2x\sum_{j=1}^{N}\frac{\eta_{j}}{x^2 + \xi_{j}^2},
\end{equation}
\noindent where the $\eta_{j}$ can be found using:
\begin{equation}
\eta_{j} = \left(N^2 + \frac{3}{2}N\right)\frac{\prod_{k=1}^{N-1}(\zeta_{k}^2 - \xi_{j}^2)}{\prod_{k\neq j}^{N} (\xi_{k}^2 - \xi_{j}^2)},\qquad j = 1,2,\dots,N
\end{equation}
Here, $\xi_{j}=2/c_{j}$ and $\zeta_{j}=2/\tilde{c}_{j}$, where $c_{j}$ are the positive eigenvalues of matrix $\underline{\underline{\Lambda}}$ and $\tilde{c}_{j}$ the positive eigenvalues of matrix $\underline{\underline{\tilde{\Lambda}}}$. $\underline{\underline{\Lambda}}$ is a $2N\times 2N$ matrix, and $\underline{\underline{\tilde{\Lambda}}}$ a $2N-1\times 2N-1$ matrix, with elements \cite{pade2}:
\begin{equation}\label{lambda_matrices}
\Lambda_{mn} = \frac{\delta_{m,n\pm 1}}{\sqrt{(2m+1)(2n+1)}} \hspace{2cm} \tilde{\Lambda}_{mn} = \frac{\delta_{m,n\pm 1}}{\sqrt{(2m+3)(2n+3)}}.
\end{equation}
\noindent There will be $N$ parameters $\xi_{j}$ and $N-1$ parameters $\zeta_{j}$.
Inserting \eqref{pade_bose} into \eqref{bath_corrfunc_infty} gives, with $\nu_{j} = \xi_{j}/\beta\hbar$ (hereafter, the Pad\'e frequencies):
\begin{equation}
\alpha(t) = \frac{2\gamma\lambda\hbar}{\pi}\int_{-\infty}^{\infty}d\omega \frac{\omega e^{-i\omega t}}{(\omega^2 + \gamma^2)}\cdot\left\lbrace \frac{1}{\beta\hbar\omega} + \frac{1}{2} + \frac{2\omega}{\beta\hbar}\sum_{j=1}^{N}\frac{\eta_{j}}{\omega^2 + \nu_{j}^2}\right\rbrace.
\end{equation}
This integral can be evaluated using the same contour used to derive the Matsubara series, to give, by analogy with Eqn. \eqref{matsubara2}:
\begin{equation}\label{pade_alpha}
\alpha(t) = \frac{2\lambda}{\beta}\left\lbrace 1 - \sum_{j=1}^{N}\frac{2\eta_{j}\gamma^2}{\nu_{j}^2 - \gamma^2} - i\frac{\gamma\beta\hbar}{2}\right\rbrace e^{-\gamma t} + \frac{2\lambda}{\beta}\sum_{j=1}^{N}\frac{2\eta_{j}\nu_{j}\gamma}{\nu_{j}^2 - \gamma^2}e^{-\nu_{j}t}.
\end{equation}
Fig. \ref{pade_figure} shows how the Pad\'e series for $\alpha(t)$ changes as the $N$ of [$N$-1/$N$] is varied. The improved convergence is very noticeable: while the Matsubara series required a large number of terms to give a good approximation, the Pad\'e series requires a much smaller number.
\begin{figure}[!ht]
\begin{center}
\includegraphics[width=14cm, keepaspectratio=true]{pade_2.png}
\caption[Convergence of Pad\'e series.]{\small{Convergence of the [$N$-1/$N$] Pad\'e series at 77 K with varying $N$, for the Lorentz-Drude spectral density. \emph{Inset: Magnified bath correlation functions at short times.} The code used to produce this figure was supplied by David Manolopoulos.}}
\label{pade_figure}
\end{center}
\end{figure}
Using the Pad\'e decomposition, we can hope to predict the dynamics of energy transfer with an accuracy that might not be afforded to us by the Matsubara decomposition: HEOM simulations can be run with increasing number of exponentials until convergence is achieved.
On the other hand, because the Matsubara series requires so many more exponential terms to achieve convergence, it is likely that at some point, the computational expense will prohibit more exponentials from being used, before convergence is achieved.
\subsection{Discussion}
The Matsubara and Pad\'e series are not the only methods of representing $\alpha(t)$ as a sum of exponentials: there are several alternatives, including the Meier-Tannor decomposition \cite{meier_tannor}, in which the spectral density is fitted to an expression of the form:
\begin{equation}
J(\omega) = \frac{\pi}{2}\sum_{k=1}^{n}\frac{\omega}{((\omega+\Omega_{k})^{2} + \Gamma_{k}^{2})((\omega-\Omega_{k})^{2}+\Gamma_{k}^{2})}.
\end{equation}
Another possibility is to numerically integrate Eqn. \eqref{bath_corrfunc}, and then to fit the numerical data to an expression of the form \eqref{corrfunc_exp}. It may be appropriate, in this case, to allow both the prefactors and the frequencies in this expression to be complex.
Both of these methods involve numerical fitting, but since the Pad\'e series allows efficient analytical convergence to the exact correlation function (and efficient numerical convergence to the exact energy transfer dynamics) for a Lorentz-Drude spectral density, it is this method that we use.
As well as the [$N$-1/$N$] approximant, the [$N$/$N$] and [$N$+1/$N$] approximants were also considered in the original literature \cite{pade2}. The latter two require that a $\delta_{+}$-function approximation be introduced, whereas the former involves no such approximation.
We also briefly consider the high-temperature limit in order to make the connection with literature in this field \cite{ishizakifleming,tanimura2}. The imaginary part of the bath correlation function is independent of temperature, $\alpha_{i}(t) = -\lambda\gamma\hbar e^{-\gamma t}$. However, for the real part, using $\lim_{\beta\hbar\rightarrow 0}\coth(\beta\hbar\omega/2)=\frac{1}{\beta\hbar\omega}$:
\begin{equation}
\lim_{\beta\rightarrow 0}\alpha_{r}(t) = \frac{2\lambda\gamma}{\pi\beta}\int_{0}^{\infty}d\omega \frac{\cos(\omega t)}{\omega^2 + \gamma^2} = \frac{2\lambda}{\beta}e^{-\gamma t}.
\end{equation}
This high-temperature expression is purely classical, and we might consider further terms (due to the Pad\'e or Matsubara expansion) as ``quantum corrections'' to this \cite{tanimura2}.
Interestingly, even at high temperatures $\alpha_{i}(t)$ contains a factor of $\hbar$, so is inherently quantum-mechanical. This term can be expressed as the Fourier transform of $J(\omega)$, so can be related to quantum dissipation, whereas the $\beta$-dependence of $\alpha_{r}(t)$ relates it to equilibrium fluctuations, which are classical at high temperatures.
{\renewcommand{\baselinestretch}{1}\normalsize
\chapter{Numerical Results}\label{numerical_results}
}
We have now laid the theoretical foundations necessary to carry out calculations for the 7-site FMO system and the 24-site trimer, at a temperature of 300 K. Calculations at a temperature of 77 K have also been carried out, but are not shown here due to the length constraint for this thesis, as well as the fact that this will allow us to focus on results at physiological temperature, which are more relevant to this work.
We will use the same parameters as in the literature \cite{ishizakifleming}, so that $\hbar\lambda_{n} = 35 \text{~cm}^{-1}$ and $\hbar\gamma_{n} = 106.1 \text{~cm}^{-1}$ for each site. These values are found by fluorescent Stokes shift experiments \cite{ishizakifleming,chromophore_solvent}.
For the 7-site FMO, the system Hamiltonian used is that of Adolphs and Renger \cite{adolphs_renger}, given in Appendix \ref{supplementary}, Eqn. \eqref{7site_Hsys}. The couplings $\hbar J_{jk}$ between sites were found using the transition dipole moments of these sites and assuming the protein to provide a dielectric medium, while the site energies $\hbar\omega_{j}$ were calculated using the interaction between the sites and charged amino acid residues \cite{adolphs_renger}.
For the 24-site FMO, we follow the theoretical study of Ritschel \emph{et al.} \cite{zofe_fmo2}, whose ZOFE quantum master equation calculations used two different sets of site energies. One of these was found by Schmidt am Busch \emph{et al.} \cite{schmidt_am_busch} using a method similar to that of Adolphs and Renger \cite{adolphs_renger}, and the other by Olbrich \emph{et al.} \cite{olbrich} using molecular dynamics simulations and electronic structure calculations. The Hamiltonian is given in Appendix \ref{supplementary}, Eqn. \eqref{24site_full}, with Eqn. \eqref{24site_Hsys} giving the intra-monomer couplings, \eqref{24site_sites} giving the site energies for the two different cases \cite{schmidt_am_busch,olbrich} and \eqref{24site_inter} the inter-monomer couplings.
The initial conditions we use will reflect the physics of the complex: bacteriochlorophylls 1, 6 and 8 are closest to the chlorosome, and it is these that will be most likely to be excited initially \cite{schmidt_am_busch,zofe_fmo2}. Thus, for the 7-site system we will carry out simulations with excitation beginning on sites 1 or on 6, and for the trimer, on sites 1, 6 or 8.
The results are colour-coded according to the key in Fig. \ref{key}, with the population of each site denoted by the colours shown.
\begin{figure}[!ht]
\begin{center}
\includegraphics[width=3cm, keepaspectratio=true]{key.png}
\caption[Key for numerical results.]{\small{Key for numerical results.}}
\label{key}
\end{center}
\end{figure}
We set these results out as follows: firstly, those for the 7-site system are presented, both for 1 ps and 15 ps (i.e., on the timescale of the electronic coherence and on the timescale of equilibration), and then those for the trimer.
These results will then allow conclusions to be drawn about which approximate method is most appropriate for our further work.
\section{7-Site Subsystem}
Figs. \ref{300K_7site_e1} and \ref{300K_7site_e6} compare the dynamics of Time-Dependent Redfield and F\"orster theories to those of the HEOM at 300 K, up to 1 ps, with the initial excitation on site 1 and on site 6 respectively. Figs. \ref{SteadyState_300K_e1} and \ref{SteadyState_300K_e6} show the same dynamics up to 15 ps. In each case, the populations of only four sites are shown, as the rest of the populations stay close to zero.
At 300 K, the HEOM dynamics fully converged with 2 Pad\'e exponential terms in $\alpha(t)$ and 4 levels of the hierarchy. At lower temperatures, the weaker system-bath interaction means that fewer levels of the hierarchy are required, but more exponential terms
\begin{figure}[!p]
\begin{center}
\includegraphics[width=13.5cm,keepaspectratio=true]{300K_7site_a.png}
\includegraphics[width=13.5cm,keepaspectratio=true]{300K_7site_c.png}
\caption[Energy transfer dynamics for the 7-site FMO subsystem at 300 K, with initial excitation on site 1.]{\small{Energy transfer dynamics for the 7-site FMO subsystem at 300 K, with initial excitation on site 1. In each case, the solid lines show the HEOM result and the dashed lines show (a) Time-Dependent Redfield, (b) F\"orster results.}}
\label{300K_7site_e1}
\end{center}
\end{figure}
\begin{figure}[!p]
\begin{center}
\includegraphics[width=13.5cm,keepaspectratio=true]{300K_7site_d.png}
\includegraphics[width=13.5cm,keepaspectratio=true]{300K_7site_f.png}
\caption[Energy transfer dynamics for the 7-site FMO subsystem at 300 K, with initial excitation on site 6.]{\small{Energy transfer dynamics for the 7-site FMO subsystem at 300 K, with initial excitation on site 6. In each case, the solid lines show the HEOM result and the dashed lines show (a) Time-Dependent Redfield, (b) F\"orster results.}}
\label{300K_7site_e6}
\end{center}
\end{figure}
\begin{figure}[!p]
\begin{center}
\includegraphics[width=13.5cm,keepaspectratio=true]{SteadyState_300K_a.png}
\includegraphics[width=13.5cm,keepaspectratio=true]{SteadyState_300K_c.png}
\caption[Long-time energy transfer dynamics for the 7-site FMO subsystem at 300 K, with initial excitation on site 1.]{\small{Long-time energy transfer dynamics for the 7-site FMO system at 300 K, with initial excitation on site 1. In each case, the solid lines show the HEOM result and the dashed lines show (a) Time-Dependent Redfield, (b) F\"orster results.}}
\label{SteadyState_300K_e1}
\end{center}
\end{figure}
\begin{figure}[!p]
\begin{center}
\includegraphics[width=13.5cm,keepaspectratio=true]{SteadyState_300K_d.png}
\includegraphics[width=13.5cm,keepaspectratio=true]{SteadyState_300K_f.png}
\caption[Long-time energy transfer dynamics for the 7-site FMO subsystem at 300 K, with initial excitation on site 6.]{\small{Long-time energy transfer dynamics for the 7-site FMO system at 300 K, with initial excitation on site 6. In each case, the solid lines show the HEOM result and the dashed lines show (a) Time-Dependent Redfield, (b) F\"orster results.}}
\label{SteadyState_300K_e6}
\end{center}
\end{figure}
In these results, electronic coherence is observed on short timescales. The formation of a coherent superposition between two states $\ket{j}$ and $\ket{k}$ is expected when the magnitude of $J_{jk}$, the coupling between them, is comparable to $|\omega_{j}-\omega_{k}|$, their energy gap. Consulting Eqn. \eqref{7site_Hsys}, it is for this reason that coherence is observed between states $\ket{1}$ and $\ket{2}$, as well as between $\ket{6}$ and $\ket{5}$.
Whether the first or the sixth site is initially excited, the steady state is the same, but if site 6 is initially excited then this state is reached more quickly. Of the 7 sites, BChl 6 has the highest energy, and we might thus intuitively expect that transfer away from this site is faster, which is borne out by the results.
Calculations were also carried out with the Markovian Redfield theory, and compared to the Time-Dependent Redfield results, the only difference was that the short-time oscillations were less pronounced using the Markovian theory. After the envelope of these oscillations decayed, both theories were quantitatively identical.
F\"orster theory's performance tends to be worse in general than that of Time-Dependent Redfield theory, especially at shorter timescales, as is particularly noticeable in Fig. \ref{300K_7site_e6}.
Both Redfield and F\"orster theories predict the equilibrium populations very well at 300 K. However, as seen in Fig. \ref{SteadyState_300K_e6}, the perturbative methods predict a ``bump'' in the dynamics that is not seen in the HEOM simulation.
Overall, the reasonable accuracy of F\"orster theory's predictions at 300 K suggests that even if coherent effects are important at low temperatures, they do not seem to be very much so at room temperature: energy barriers can be surmounted thermally and therefore tunnelling contributions have little effect.
\section{24-Site Trimer}
In this Section, we present accurate results for the full FMO trimer up to 1 ps, as well as results up to 15 ps that are converged with respect to number of hierarchy levels, but not fully converged with respect to number of exponentials in $\alpha(t)$ (however, it is known that these results are very accurate). These results have not previously been published.
The reason that the 15 ps results are not converged with respect to number of exponentials is one of computation: up to the shorter time, converged results using the HEOM required a large amount of time and memory, and a prohibitively larger time would be required to observe the fully converged steady states.
However, by using the same bath correlation function for all methods, we can carry out a rigorous comparison, which will be illuminating.
Figs. \ref{300K_nmkv_24site_Olb} and \ref{300K_nmkv_24site_SaB} compare the Time-Dependent Redfield and the HEOM results, using respectively the site energies of Olbrich \emph{et al.} (Olb) and of Schmidt am Busch \emph{et al.} (SaB), while Figs. \ref{300K_frst_24site_Olb} and \ref{300K_frst_24site_SaB} repeat this comparison for the F\"orster and the HEOM results.
\begin{figure}[!p]
\begin{center}
\includegraphics[width=13.1cm,keepaspectratio=true]{24site_nmkv_300K_OLB_a.png}
\includegraphics[width=13.1cm,keepaspectratio=true]{24site_nmkv_300K_OLB_b.png}
\includegraphics[width=13.1cm,keepaspectratio=true]{24site_nmkv_300K_OLB_c.png}
\caption[Comparison of energy transfer dynamics to the Redfield approximation for the 24-site FMO trimer at 300 K, using Olb site energies.]{\small{Energy transfer dynamics for the FMO trimer at 300 K using Olb site energies, with HEOM (solid lines) compared to Time-Dependent Redfield (dashed lines). Initial excitation on (a) site 1, (b) site 6, (c) site 8.}}
\label{300K_nmkv_24site_Olb}
\end{center}
\end{figure}
\begin{figure}[!p]
\begin{center}
\includegraphics[width=13.1cm,keepaspectratio=true]{24site_nmkv_300K_SAB_a.png}
\includegraphics[width=13.1cm,keepaspectratio=true]{24site_nmkv_300K_SAB_b.png}
\includegraphics[width=13.1cm,keepaspectratio=true]{24site_nmkv_300K_SAB_c.png}
\caption[Comparison of energy transfer dynamics to the Redfield approximation for the 24-site FMO trimer at 300 K, using SaB site energies.]{\small{Energy transfer dynamics for the FMO trimer at 300 K using SaB site energies, with HEOM (solid lines) compared to Time-Dependent Redfield (dashed lines). Initial excitation on (a) site 1, (b) site 6, (c) site 8.}}
\label{300K_nmkv_24site_SaB}
\end{center}
\end{figure}
\begin{figure}[!p]
\begin{center}
\includegraphics[width=13.1cm,keepaspectratio=true]{24site_frst_300K_OLB_a.png}
\includegraphics[width=13.1cm,keepaspectratio=true]{24site_frst_300K_OLB_b.png}
\includegraphics[width=13.1cm,keepaspectratio=true]{24site_frst_300K_OLB_c.png}
\caption[Comparison of energy transfer dynamics to the F\"orster approximation for the 24-site FMO trimer at 300 K, using Olb site energies.]{\small{Energy transfer dynamics for the FMO trimer at 300 K using Olb site energies, with HEOM (solid lines) compared to F\"orster theory (dashed lines). Initial excitation on (a) site 1, (b) site 6, (c) site 8.}}
\label{300K_frst_24site_Olb}
\end{center}
\end{figure}
\begin{figure}[!p]
\begin{center}
\includegraphics[width=13.1cm,keepaspectratio=true]{24site_frst_300K_SAB_a.png}
\includegraphics[width=13.1cm,keepaspectratio=true]{24site_frst_300K_SAB_b.png}
\includegraphics[width=13.1cm,keepaspectratio=true]{24site_frst_300K_SAB_c.png}
\caption[Comparison of energy transfer dynamics to the F\"orster approximation for the 24-site FMO trimer at 300 K, using SaB site energies.]{\small{Energy transfer dynamics for the FMO trimer at 300 K using SaB site energies, with HEOM (solid lines) compared to F\"orster theory (dashed lines). Initial excitation on (a) site 1, (b) site 6, (c) site 8.}}
\label{300K_frst_24site_SaB}
\end{center}
\end{figure}
Fig. \ref{300K_24site_steady_nmkv} compares the results of the HEOM up to 15 ps to those of the Time-Dependent Redfield, and Fig. \ref{300K_24site_steady_frst} compares the results of the HEOM to those of F\"orster theory on the same timescale. Both of these Figures use the Olb site energies.
\begin{figure}[!p]
\begin{center}
\includegraphics[width=13.1cm,keepaspectratio=true]{ss_nmkv_a.png}
\includegraphics[width=13.1cm,keepaspectratio=true]{ss_nmkv_b.png}
\caption[Long-time energy transfer dynamics for the 24-site FMO trimer at 300 K, compared to Time-Dependent Redfield predictions]{\small{Comparison of long-time dynamics predicted by the HEOM (solid lines) and Time-Dependent Redfield (dashed lines) theories for (a) initial excitation on site 1, (b) initial excitation on site 2. The Olb site energies are used.}}
\label{300K_24site_steady_nmkv}
\end{center}
\end{figure}
\begin{figure}[!p]
\begin{center}
\includegraphics[width=13.1cm,keepaspectratio=true]{ss_frst_a.png}
\includegraphics[width=13.1cm,keepaspectratio=true]{ss_frst_b.png}
\caption[Long-time energy transfer dynamics for the 24-site FMO trimer at 300 K, compared to F\"orster predictions.]{\small{Comparison of long-time dynamics predicted by the HEOM (solid lines) and F\"orster (dashed lines) theories for (a) initial excitation on site 1, (b) initial excitation on site 2. The Olb site energies are used.}}
\label{300K_24site_steady_frst}
\end{center}
\end{figure}
When BChl 8 is initially excited, the decay is largely exponential, with no coherence effects, due to the fact that population transferred to BChl 1 is rapidly transferred onwards to BChl 2, and sufficient population for a noticeable coherent superposition is not built up \cite{zofe_fmo2}.
Considering the Time-Dependent Redfield simulations, we can see that this method performs quite well for both sets of site energies, capturing the features of the transfer dynamics including the oscillations.
At 300 K, F\"orster theory is quite accurate overall, and in cases where it is not as much so, the differences are generally over-predictions of transfer rates rather than under-predictions. The results are somewhat worse than those of Redfield theory: aside from its neglect of the oscillations (whose importance we consider in the next Section), the agreement with exact dynamics tends to be poorer for F\"orster theory. Both theories nevertheless predict the long-time dynamics well, as was seen for the 7-site model.
\newpage \section{Conclusions}
In this Chapter we have presented numerically exact electronic energy transfer dynamics for both the 7-site and the 24-site FMO models, and have used them to benchmark approximate calculations using the Time-Dependent Redfield and F\"orster theories, in order to identify the method most suitable for further use.
Comparing the F\"orster and Redfield theories, although the F\"orster theory predicted the general shape of the dynamics and tended to give reasonably good quantitative agreement, the predictions of the Redfield theory were more accurate. Both methods, however, predicted the long-time decays observed, and the steady state, well.
The fluorescence lifetime of the bacteriochlorophyll sites is on the order of nanoseconds, whereas, as we have seen in this Chapter, steady states are reached in a timescale on the order of tens of picoseconds. Whether or not electronic coherence effects are included in the model of the energy transfer, this steady state is reached in the same amount of time.
The fact that F\"orster theory predicted the dynamics at 300 K so well, in terms of the important features (transfer rates, equilibrium populations and the timescale at which a steady state is reached) is extremely interesting, and allows us to conclude that although quantum coherence effects may be present in the energy transfer at physiological temperature, they appear to be unnecessary for efficient transfer.
With this conclusion, we are prepared to further investigate the physics of the FMO complex, using the F\"orster theory, which has been shown to give fairly good results at temperatures of biophysical interest and is computationally extremely cheap compared to all other methods described in this thesis, including the Time-Dependent Redfield theory. This will be the subject of the following Chapter.
{\renewcommand{\baselinestretch}{1}\normalsize
\chapter{Applications Of F\"orster Theory}\label{applications}
}
There are a number of physical features of the FMO complex which cannot be investigated using the HEOM, due to the computational demand. F\"orster theory, on the other hand, requires very little in terms of computational resources, so can be used for these investigations with ease.
The popularity of the Lorentz-Drude spectral density utilized thus far, within the modelling of open quantum systems, is due in large part to the fact that it allows the bath correlation function to be written as a sum of exponentials.
In reality, there is no reason to expect that this should give us physically correct results: in a pigment-protein complex, the featureless spectral density we have been using does not describe the vibrational environment of a bacteriochlorophyll pigment: for example, vibrations due to certain bonds (for example, O-H) might be expected to couple strongly to a site.
A biological system is generally exquisitely tailored, with its physical parameters striking a balance which, if disturbed past their tolerance, could have drastic effects on its functionality. The spectral density contains all information about the environment, and so altering this function is equivalent to altering the environment that interacts with the bacteriochlorophyll.
In Section \ref{section_specdens}, we explore the effect of using structured spectral densities, and in doing so we hope to learn how robust the energy transfer dynamics are to a change in environment.
The site energies given for the FMO complex are in fact averages over some distribution: due to the slow fluctuation of the protein environment, the actual set of site energies for a given complex will differ from this average, a phenomenon known as static disorder \cite{static_disorder}.
The effect of this disorder, simulated by drawing the site energies from a Gaussian distribution, on the energy transfer dynamics of the complex will be the subject of Section \ref{section_static}.
In this Chapter, we carry out simulations only at a physiological temperature of 300 K, as this is more biologically meaningful, and it is at this temperature that we have shown the F\"orster theory to give reasonably accurate results for the important features of the energy transfer dynamics.
\section{Structured Spectral Density}\label{section_specdens}
The molecular dynamics simulations used by Olbrich \emph{et al.} to find the site energies of the FMO trimer were also used to calculate a spectral density for this complex \cite{olbrich,kleinekathoefer}. Fig. \ref{klein_specdens} shows an average spectral density for the 7-site subsystem, normalized to give a reorganization energy of $\lambda = 35 \text{~cm}^{-1}$, in agreement with experiment.
\begin{figure}[!ht]
\begin{center}
\includegraphics[width=13cm, keepaspectratio=true]{klein_spec.png}
\caption[Structured spectral density for 7-site FMO complex.]{\small{Structured spectral density for 7-site FMO complex, using parameters given in \cite{kleinekathoefer}.}}
\label{klein_specdens}
\end{center}
\end{figure}
The simulations carried out gave the real part of the bath correlation function in the form:
\begin{equation}\label{kl_corrfunc}
\alpha_{r,j}(t) = \sum_{k=1}^{K}c_{jk}\cos(\tilde{\omega}_{jk}t)e^{-\gamma_{jk} t},
\end{equation}
\noindent where $K=15$ and the values of the $c_{jk}$, $\omega_{jk}$ and $\gamma_{jk}$ are given in Appendix \ref{supplementary}, Eqn. \eqref{klein_spectrum_data}. From this, Fourier inversion of the real part of \eqref{bath_corrfunc} gives the spectral density\footnote{A factor of $\frac{2}{\pi}$ is included to give agreement with \cite{kleinekathoefer}, although the scaling to give the correct $\lambda$ means that this factor is unimportant} $J_{j}(\omega) = \frac{2}{\pi}\tanh(\beta\hbar\omega/2)\int_{0}^{\infty}\alpha_{r,j}(t)\cos(\omega t)dt$, or:
\begin{equation}\label{kl_spectral}
J_{j}(\omega) = \frac{\tanh(\beta\hbar\omega / 2)}{\pi} \sum_{k=1}^{K}\left( \frac{c_{jk}\gamma_{jk}}{\gamma_{jk}^{2} + (\omega - \tilde{\omega}_{jk})^{2}} + \frac{c_{jk}\gamma_{jk}}{\gamma_{jk}^{2} + (\omega + \tilde{\omega}_{jk})^{2}} \right).
\end{equation}
In order to use this spectral density in our calculations, we need an expression for the full bath correlation function, so must find $\alpha_{i,j}(t)$:
\begin{equation}
\alpha_{i,j}(t) = -\int_{0}^{\infty}d\omega J_{j}(\omega)\sin(\omega t).
\end{equation}
Using $\tanh(\beta\hbar\omega/2) = \frac{1}{1+e^{-\beta\hbar\omega}}-\frac{1}{1+e^{\beta\hbar\omega}}$ gives an integrand that is an odd function, so that the integral can be rewritten:
\begin{align}
\alpha_{i,j}(t) & = \frac{1}{\pi}\sum_{k=1}^{K}\int_{-\infty}^{\infty}d\omega\cdot \frac{\sin(\omega t)}{1+e^{\beta\hbar\omega}} \left( \frac{c_{jk}\gamma_{jk}}{\gamma_{jk}^{2} + (\omega - \tilde{\omega}_{jk})^{2}} + \frac{c_{jk}\gamma_{jk}}{\gamma_{jk}^{2} + (\omega + \tilde{\omega}_{jk})^{2}} \right) \nonumber\\
& = \frac{2}{\pi}\sum_{k=1}^{K}c_{jk}\gamma_{jk}\int_{-\infty}^{\infty}d\omega \frac{(\omega^{2} + \tilde{\omega}_{jk}^{2} + \gamma_{jk}^{2})\sin(\omega t)f_{Fermi}(\beta\hbar\omega)}{[\gamma_{jk}^{2}+(\omega-\tilde{\omega}_{jk})^{2}][\gamma_{jk}^{2}+(\omega+\tilde{\omega}_{jk})^{2}]}.
\end{align}
Here, $f_{Fermi}(x) = (1+e^{x})^{-1}$ is the Fermi-Dirac function. As with the Bose-Einstein function in \ref{section_pade}, there is an [$N$-1/$N$] Pad\'e approximant that can be used for $f_{Fermi}(x)$ \cite{pade2}:
\begin{equation}
f_{Fermi}(x) \approx \frac{1}{2} - 2x\sum_{l=1}^{N}\frac{\eta_{l}}{x^{2}+\xi_{l}^{2}}.
\end{equation}
For the Fermi-Dirac case, the matrices $\underline{\underline{\Lambda}}$ and $\underline{\underline{\tilde{\Lambda}}}$ are given by:
\begin{equation}
\Lambda_{mn} = \frac{\delta_{m,n\pm 1}}{\sqrt{(2m-1)(2n-1)}} \hspace{2cm} \tilde{\Lambda}_{mn} = \frac{\delta_{m,n\pm 1}}{\sqrt{(2m+1)(2n+1)}}.
\end{equation}
Using the Pad\'e expression gives the following integral, where $\Im$ denotes the imaginary part:
\begin{multline}
\alpha_{i,j}(t) = \frac{4}{\pi\beta\hbar}\sum_{k=1}^{K}\sum_{l=1}^{N}c_{jk}\gamma_{jk}\eta_{l}\Im\left( \int_{-\infty}^{\infty}d\omega\cdot e^{-i\omega t}\times\right.\\
\left. \frac{\omega^{3} + \tilde{\omega}_{jk}^{2}\omega + \gamma_{jk}^{2}\omega}{[\gamma_{jk}^{2}+(\omega-\tilde{\omega}_{jk})^{2}][\gamma_{jk}^{2}+(\omega+\tilde{\omega}_{jk})^{2}][\omega^{2} + \nu_{l}^{2}]} \right).
\end{multline}
This integral can be carried out using the usual semicircular contour. The result is, with $\Omega_{jk} = i\tilde{\omega}_{jk} - \gamma_{jk}$:
\begin{multline}
\alpha_{i,j}(t) = \frac{2}{\beta\hbar}\sum_{k=1}^{K}\sum_{l=1}^{N} \left( \frac{c_{jk}\eta_{l}\Omega_{jk}}{\nu_{l}^{2} - \Omega_{jk}^{2}}e^{\Omega_{jk} t} + \frac{c_{jk}\eta_{l}\Omega_{jk}^{\ast}}{\nu_{l}^{2} - (\Omega_{jk}^{\ast})^{2}}e^{\Omega_{jk}^{\ast} t} \right. \\
\left. -\frac{2\gamma_{jk}c_{jk}\eta_{l}(|\Omega_{jk}|^{2} - \nu_{l}^{2})}{|\nu_{l} - \Omega_{jk}^{2}|^{2}}e^{-\nu_{l} t}\right).
\end{multline}
The real part can also be written in terms of exponentials:
\begin{equation}
\alpha_{r,j}(t) = \sum_{k=1}^{K}\left( \frac{c_{jk}}{2}e^{\Omega_{jk} t} + \frac{c_{jk}}{2}e^{\Omega_{jk}^{\ast} t} \right).
\end{equation}
Since each term in \eqref{kl_spectral} gives rise to a peak in the spectral density, by excluding certain terms we can now find out whether removing the corresponding peaks has an effect on the energy transfer dynamics.
\subsection{Results}
In order to inform the cases we investigate in this Section, we will wish to know the electronic transition frequencies between sites. Fig. \ref{transitions} shows these frequencies along with the structured spectral density.
\begin{figure}[!ht]
\begin{center}
\includegraphics[width=13.5cm, keepaspectratio=true]{transitions.png}
\caption[Inter-site transition energies for the 7-site FMO complex.]{\small{Inter-site transition energies for the 7-site FMO complex.}}
\label{transitions}
\end{center}
\end{figure}
The investigation of Olbrich \emph{et al.} using the ZOFE \cite{zofe_fmo1} led to the conclusion that high-energy vibrational modes do not have any effect on the energy transfer dynamics, since the energy of these modes is far from being in resonance with any transitions.
The spectral density in \ref{klein_specdens} shows several modes above 1300 cm$^{-1}$. By neglecting a number of terms in Eqn. \eqref{kl_spectral}, these modes can be removed from the spectral density with a negligible effect at smaller frequencies. Fig. \ref{highfreq_1ps} compares the dynamics both with and without the high-frequency modes In both cases, the spectral densities have been scaled to give $\lambda = $ 35 cm$^{-1}$.
On this short timescale, there are negligible differences between the predicted dynamics. Another interesting comparison is given by Fig. \ref{1ps_LDD}, which shows the dynamics using the Lorentz-Drude spectral density and the full structured spectral density, with a slightly greater discrepancy than Fig. \ref{highfreq_1ps}, though still a very small difference between the dynamics.
\begin{figure}[!hp]
\begin{center}
\includegraphics[width=13.5cm, keepaspectratio=true]{highfreq_1ps_a.png}
\includegraphics[width=13.5cm, keepaspectratio=true]{highfreq_1ps_b.png}
\caption[Dynamics predicted by F\"orster theory using structured spectral density with and without high-frequency modes.]{\small{(a) Dynamics predicted using structured spectral density with (solid lines) and without (dashed lines) high-frequency modes. (b) Spectral densities used in the calculation of (a).}}
\label{highfreq_1ps}
\end{center}
\end{figure}
\begin{figure}[!hp]
\begin{center}
\includegraphics[width=13.5cm, keepaspectratio=true]{1ps_LDD_a.png}
\includegraphics[width=13.5cm, keepaspectratio=true]{1ps_LDD_b.png}
\caption[Dynamics predicted by F\"orster theory using Lorentz-Drude and structured spectral density.]{\small{(a) Dynamics predicted using Lorentz-Drude (solid lines) and structured (dashed lines) spectral density. (b) Spectral densities used in the calculation of (a).}}
\label{1ps_LDD}
\end{center}
\end{figure}
\newpage It is possible that even a small difference in the short-time dynamics could lead to a larger difference in the longer-time dynamics, so it is important to follow the simulations to this limit.
Fig. \ref{15ps_full_semi_ldd} repeats the two comparisons, but with long-time simulations. As before, there is excellent agreement for the structured spectral density and the same with high-frequency peaks omitted, and good (though less so than the former) agreement for the structured spectral density and the Lorentz-Drude. We now analyse these results.
\begin{figure}[!ht]
\begin{center}
\includegraphics[width=13.5cm, keepaspectratio=true]{15ps_full_semi_ldd_a.png}
\includegraphics[width=13.5cm, keepaspectratio=true]{15ps_full_semi_ldd_b.png}
\caption[Long-time dynamics predicted by F\"orster theory using different spectral densities.]{\small{Long-time dynamics for (a) fully structured (solid lines) vs. omitted high-frequency peaks (dashed lines) spectral densities. (b) Lorentz-Drude (solid lines) vs. fully structured (dashed lines) spectral densities.}}
\label{15ps_full_semi_ldd}
\end{center}
\end{figure}
\subsection{Discussion}
The effect of spectral structure on the dynamics of energy transfer in the FMO complex is very little at room temperature.
We have seen in Fig. \ref{highfreq_1ps} that the high-frequency structure appears to be unnecessary in determining the efficiency of the energy transfer, and in Fig. \ref{1ps_LDD} that the fully structured spectral density of Olbrich \emph{et al.} \cite{kleinekathoefer} can be replaced with the Lorentz-Drude spectral density with only a small effect.
This latter result is remarkable because it suggests that the actual character of the vibrational modes is unimportant, and that it is some other property of the bath that determines the energy transfer. It is interesting to look further into what this property is.
Since we have shown that F\"orster theory provides a reasonable description of the dynamics, an investigation of absorption and fluorescence spectra, as in Section \ref{spectral_overlap}, will be informative.
\newpage Fig. \ref{spectral_comparison} shows the fluorescence spectra of site 1 for both the Lorentz-Drude spectral density and the fully structured spectral density of Fig. \ref{kl_spectral}.
\begin{figure}[!h]
\begin{center}
\includegraphics[width=13.5cm, keepaspectratio=true]{klein_spectra.png}
\caption[Site 1 fluorescence spectra for Lorentz-Drude and structured spectral densities.]{\small{Site 1 fluorescence spectra, $F_{1}[\omega]$, for Lorentz-Drude (solid black line) and fully structured (dashed blue line) spectral densities.}}
\label{spectral_comparison}
\end{center}
\end{figure}
The spectrum is narrowed to a small degree, but the effect of this on the spectral overlap is negligible compared to either of the following effects: the much greater narrowing that comes from decreasing the temperature (see Fig. \ref{overlap}) or the shift that comes about from increasing the reorganization energy.
\newpage Fig. \ref{shifted_spectrum} shows the effect both on the dynamics and on $F_{1}[\omega]$ of increasing the reorganization energy from 35 cm$^{-1}$ to 70 cm$^{-1}$, while using the same Lorentz-Drude spectral density: there is a more significant change in the dynamics, reflecting the fact that the change results in a larger effect on the position and width of the spectral line, and thus a larger effect on its overlap with other lines (and via Eqn. \eqref{overlap_equation}, on the energy transfer rate constants).
\begin{figure}[!h]
\begin{center}
\includegraphics[width=13.5cm, keepaspectratio=true]{shifted_spectrum_a.png}
\includegraphics[width=13.5cm, keepaspectratio=true]{shifted_spectrum_b.png}
\caption[Site 1 fluorescence spectra for two different reorganization energies.]{\small{(a) Energy transfer dynamics for the 7-site FMO complex using a Lorentz-Drude spectral density with $\lambda = 35$ cm$^{-1}$ (solid lines) and $\lambda = 70$ cm$^{-1}$ (dashed lines). (b) Site 1 fluorescence spectra, $F_{1}[\omega]$, for the two reorganization energies.}}
\label{shifted_spectrum}
\end{center}
\end{figure}
These observations lead us to the conclusion that it is the temperature and the environmental reorganization energy that have the greatest effect on the dynamics of excitation energy transfer, rather than the precise shape of the spectral density $J(\omega)$.
\newpage\section{Static Disorder}\label{section_static}
The phenomenon of static disorder causes inhomogeneous line-broadening in electronic spectroscopy. That is, because each FMO complex in a sample is experiencing a different protein environment, giving different electronic transition energies, the spectral properties of each individual complex will differ and the superimposition of different spectral lines gives a broadened line.
In modelling the effects of inhomogeneous broadening, we treat the site energies as random variables with a Gaussian distribution \cite{static_disorder}.
Thus, from the original input data files for our numerical simulations, a large number of other input files are created, in which a random value is added to each site energy, drawn from a Gaussian distribution. The dynamics are found for each input file, and the results are averaged.
In order to carry out this procedure, we require a method of producing Gaussian random numbers. The standard Box-Muller algorithm was chosen for this purpose (Sec. 7.2 of \cite{numerical_recipes}).
The HEOM would be more difficult to use in finding the effect of this disorder, because a converged calculation for a single realization of Gaussian disorder would take a good fraction of an hour, whereas a F\"orster calculation is completed in a matter of seconds. We may require a large number of realizations to find the average effects, so it is preferable to use the faster technique, whose results have been shown to be accurate enough for this purpose at 300 K.
The effect of inhomogeneous broadening on 2-dimensional electronic spectra of the FMO complex has been the subject of a number of studies \cite{static_disorder,2d_fmo}, but in the following, we hope to find the effects, if any, on energy transfer dynamics in the hope once again of understanding the effect of the protein environment.
\subsection{Results}
An implementation of static disorder is characterized by the standard deviation, $\sigma$, of the Gaussian distribution. Often, the full-width at half-maximum (FWHM), equal to $2\sqrt{2\ln 2}\sigma$, is specified instead. Several studies have used a FWHM on the order of 50 cm$^{-1}$ \cite{static_disorder,2d_fmo,read_fleming}, and so this value will be used in this Section.
We show firstly the results for 3 realizations of static disorder in Fig. \ref{realizations}: alongside each set of results is a diagram showing the energy levels of the different sites for this realization, compared to the energy levels in the absence of disorder.
\begin{figure}[!hp]
\begin{center}
\includegraphics[width=12.3cm, keepaspectratio=true]{realizations_a.png}
\includegraphics[width=12.3cm, keepaspectratio=true]{realizations_b.png}
\includegraphics[width=12.3cm, keepaspectratio=true]{realizations_c.png}
\caption[Dynamics for three realizations of static disorder.]{\small{Dynamics for 3 realizations of static disorder. The left-hand column gives the dynamics, and the right-hand column shows the site energies for these realizations. The solid line in each case is the situation in which static disorder is ignored, and the dashed line indicates its presence.}}
\label{realizations}
\end{center}
\end{figure}
Each individual set of site energies gives dynamics that differ a little from the case where disorder is neglected: this is a general observation that can be made for any one realization that is taken, although in some cases (as in Fig. \ref{realizations} c), the differences are slightly more pronounced.
In a real physical application, there will be a great number of FMO complexes, and the dynamics will be an average over many realizations. Fig. \ref{static_disorder} compares the average dynamics for different numbers of realizations to the dynamics in the absence of disorder. A relatively small number of realizations is required for the average dynamics to converge.
\begin{figure}[!ht]
\begin{center}
\includegraphics[width=13.5cm, keepaspectratio=true]{static_disorder_a.png}
\includegraphics[width=13.5cm, keepaspectratio=true]{static_disorder_b.png}
\caption[Average dynamics for numbers of realizations of static disorder.]{\small{Average dynamics (dashed lines) for (a) 5 and (b) 50 realizations of static disorder. The solid lines in each case show the dynamics without disorder.}}
\label{static_disorder}
\end{center}
\end{figure}
\subsection{Discussion}
The results presented in this Section illustrate that inhomogeneous disordering effects on the dynamics of the FMO complex are minimal. This can be explained by examining the energy level diagrams in the right-hand column of Fig. \ref{realizations}.
In the results we have observed, the populations of bacteriochlorophylls 5, 6 and 7 remain close to zero, and only the other four have appreciable populations throughout the simulation.
The electronic energy separations of the four sites involved in the transfer are larger than the differences induced by static disorder, so that the important transition energies are not likely to be very different for two different realizations.
Thus, while there will generally be small differences for the dynamics of these two realizations, the most important features will be the same when comparing the two, and these small differences are not important in the FMO complex because of the effective averaging of the dynamics.\\[4cm]
\section{Conclusions}
In the two subsections of this Chapter, conclusions have been drawn separately about the effects of a structured spectral density and of static disorder due to an inhomogeneous environment. Here, we will make some comments about the implications for the FMO complex.
Firstly, it was found that the structure of the spectral density is largely unimportant at 300 K, and that the temperature and reorganization energy were far more important in determining the rate of energy transfer.
Secondly, the effects of static disorder characterized by a physically reasonable FWHM were found to be very little, in addition to which the average behaviour converged with only a small number of realizations.
This leads to a conclusion which may seem surprising at first: that is, the electronic energy transfer dynamics in the FMO complex are largely insensitive to the fine-structure of the environment. Similarly, the disorder inherent in any experiment has little effect, giving a very robust transfer dynamics.
The energy transfer dynamics of the FMO complex are very stable with respect to changes in its vibrational environment. Perhaps this is not a coincidence, but rather a result of the complex having evolved to perform its job so robustly.
{\renewcommand{\baselinestretch}{1}\normalsize
\chapter{Conclusions And Further Work}\label{conclusions}
}
In this project, we have investigated the physics of the Fenna-Matthews-Olson complex. The Hierarchical Equations of Motion were utilized to find numerically exact excitation energy transfer dynamics for the full 24-site trimer, with a fully converged bath correlation function. These are the first exact results that have been computed for this system.
We have also applied the approximate Redfield and F\"orster methods (the latter of which neglects electronic coherence) to the problem, and found that at room temperature, F\"orster theory described the features of the dynamics well, leading to the conclusion that quantum-mechanical coherence is, for the most part, irrelevant for the efficiency of transfer.
Having shown that F\"orster theory is applicable at this temperature, we then applied it to the problem of altering the effect of the environment on the system. It was found that the effects of adding structure to the spectral density of the bath phonons coupled to the electronic system or of including static disorder were minimal, and that the bath reorganization energies and the temperature were the most important parameters for determining the energy transfer dynamics.
There has been much excitement recently about the importance of coherence in photosynthetic systems \cite{coherence,ishizaki_review,nalbach_thorwart,quant_coherence_assists,shim_rebentrost_spectral}, and the incoherent nature of F\"orster theory has led to its dismissal as an appropriate method for describing transfer \cite{i&f_redfield,fleming_lightharv}. However, the results of this work show that the search for an adequate description of electronic energy transfer need not necessarily require that the effects of coherence be included.
Before moving on to discuss potential further work, it is quite satisfying to note that the energy transfer dynamics in the FMO complex, structurally characterized nearly 40 years ago \cite{fenna_matthews_olson}, can be modelled in a very satisfactory way by a theory that is even more venerable \cite{forster1}.
\section{Further Work}\label{further_work}
The different facets of this work suggest a number of directions for further research:
\begin{itemize}
\item The availability of numerically exact results for larger and more complex systems than previously available allows effective benchmarking of approximate methods: the current glut of such methods, giving different results, means that it is a great advantage to be able to test them against a method (the HEOM) that can be used for many different systems and parameter regimes.
\item Having seen that we need not requite that coherence effects be included in an accurate prediction of energy transfer dynamics, methods can be sought that are more accurate than the F\"orster theory, and with greater ranges of applicability, while still being incoherent.
\item An approximate method that gives accurate results at lower temperatures might allow for a more comprehensive test of the effects of a structured spectral density at these temperatures.
\end{itemize}
With such investigations, perhaps we will be able to gain more insight into how important, if at all, quantum-mechanical interference effects are in biology.
|
\section{Introduction}
\setcounter{equation}{0}
The discovery of the ghost-free massive gravity theory
by de Rham, Gabadadze, and Tolley (dRGT) \cite{deRham:2010kj}
(see \cite{Hinterbichler:2011tt,*deRham:2014zqa} for a review)
opens up the possibility to explain the dark energy and the
cosmic acceleration
\cite{1538-3881-116-3-1009,*0004-637X-517-2-565} by a tiny
mass of the gravitons. The dRGT field equations
admit the de Sitter solution with the cosmological constant mimicked by the
graviton mass. This solution can describe the late time acceleration, but
a special analysis is needed to decide
whether its other properties are physically acceptable.
A closer look reveals that the de Sitter solution in the dRGT theory is actually not unique,
and a number of its versions have been found
\cite{Koyama:2011xz,*Koyama:2011yg},%
\cite{Chamseddine:2011bu},%
\cite{D'Amico:2011jj},%
\cite{Gumrukcuoglu:2011ew},%
\cite{Gratia:2012wt,*Wyman:2012iw}
\cite{Volkov:2012cf,*Volkov:2012zb,*Volkov:2013roa}.
A special attention was received by one solution
whose physical and reference metrics are of the manifestly homogeneous and isotropic
Freedman-Lema$\hat{{\i}}$tre-Robertson-Walker
(FLRW) form \cite{Gumrukcuoglu:2011ew}. However,
a detailed analysis revealed that this solution is unstable
\cite{DeFelice:2012mx,*DeFelice:2013awa}.
For other known solutions only the physical metric is manifestly FLRW
while the reference metric looks inhomogeneous, for which reason
they are considered to be less interesting \cite{D'Amico:2011jj}.
All of this has reduced the interest towards the dRGT theory,
the focus shifting towards its extensions, as for example the bigravity
\cite{Hassan:2011zd},\cite{Volkov:2011an,*vonStrauss:2011mq,*Comelli:2011zm},%
\cite{Akrami:2012vf,*Comelli:2014bqa,*Konnig:2014xva,%
*DeFelice:2014nja,*Solomon:2014iwa} and other generalizations
admitting FLRW solutions
\cite{D'Amico:2012zv,*Huang,*DeFelice:2013dua,*Mukohyama:2014rca,*deRham:2014gla,*Heisenberg:2015voa}.
However, we would like to argue that it may be premature to abandon the dRGT theory
on the basis of negative evidence obtained from just one solution, because
the theory admits infinitely many other solutions that could
be physically interesting. They all have the same physical (de Sitter) metric
but different values of the reference metric depending on the
Stuckelberg field $T(t,r)$ subject to a complicated differential equation
\cite{D'Amico:2011jj},\cite{Gratia:2012wt},%
\cite{Volkov:2012cf,*Volkov:2012zb,*Volkov:2013roa}.
Below we shall describe a simple way to obtain these solutions by applying the
Gordon ansatz \cite{Visser} and using the global embedding coordinates.
The $T$-equation then assumes a simple form,
$(\partial_t{T})^2-(\partial_r T)^2=1$, whose essentially general solution is known.
The simplest solution $T=t$
\cite{Gumrukcuoglu:2011ew} is unstable
\cite{DeFelice:2012mx,*DeFelice:2013awa}
but other solutions could be stable.
One can choose $T$ such that both metrics are invariant under the timelike
isometry, which gives special solutions since only for them
the canonical energy is time independent.
We conjecture that their energy is minimal and hence these solutions are stable.
We also give explicit examples where the reference metric
looks inhomogeneous but shares with the physical metric the same
translational and rotational isometries. Hence, solutions considered to be non-FLRW
can actually be homogeneous and isotropic.
All of this suggests that viable dRGT cosmologies
may exist.
\section{The dRGT massive gravity}
\setcounter{equation}{0}
The theory is defined on a four-dimensional spacetime manifold endowed with
two Lorentzian metrics, the physical metric $g_{\mu\nu}$
and the flat reference metric
$
f_{\mu\nu}=\eta_{AB}\partial_\mu\Phi^A\partial_\nu\Phi^B
$
with $\eta_{AB}={\rm diag}[-1,1,1,1]$.
The fields $\Phi^A(x)$ are sometimes called
Stuckelberg scalars. The theory is defined
by the action
\begin{eqnarray} \label{1}
&&S
=\frac{M_{\rm Pl}^2}{m^2}\int \, \left(
\frac{1}{2}\,R({g})
-{\cal U}\right)\sqrt{-g}\,d^4x
\,.
\end{eqnarray}
The metrics and all coordinates are
assumed to be dimensionless, the length scale
being the inverse graviton mass $1/m$.
The interaction between the two metrics
is expressed by
a scalar function
of the tensor
$
\gamma^\mu_{~\nu}=\sqrt{{{g}}^{\mu\alpha}{{f}}_{\alpha\nu}}
$
where ${{g}}^{\mu\nu}$ is the inverse of ${g}_{\mu\nu}$
and the square root
is understood in the matrix sense, i.e.
$
(\gamma^2)^\mu_{~\nu}\equiv \gamma^\mu_{~\alpha}\gamma^\alpha_{~\nu}={g}^{\mu\alpha}
{f}_{\alpha\nu}.
$
If $\lambda_A$ ($A=0,1,2,3$) are the eigenvalues of $\gamma^\mu_{~\nu}$
then the interaction potential is given by
$
{\cal U}=\sum_{n=0}^4 b_k\,{\cal U}_k
$
where $b_k$ are parameters and
${\cal U}_k$ are defined
by the
relations
\begin{eqnarray} \label{4}
{\cal U}_0&=&1,~~~~~
{\cal U}_1=
\sum_{A}\lambda_A=[\gamma],~~~~~
{\cal U}_2=
\sum_{A<B}\lambda_A\lambda_B
=\frac{1}{2!}([\gamma]^2-[\gamma^2]),\nonumber \\
{\cal U}_3&=&
\sum_{A<B<C}\lambda_A\lambda_B\lambda_C
=
\frac{1}{3!}([\gamma]^3-3[\gamma][\gamma^2]+2[\gamma^3]),\nonumber \\
{\cal U}_4&=&
\lambda_0\lambda_1\lambda_2\lambda_3
=
\frac{1}{4!}([\gamma]^4-6[\gamma]^2[\gamma^2]+8[\gamma][\gamma^3]+3[\gamma^2]^2
-6[\gamma^4])\,. \nonumber
\end{eqnarray}
Here, using the hat to denote matrices, one defines
$[\gamma]\equiv {\rm tr}(\hat{\gamma})= \gamma^\mu_{~\mu}$,
$[\gamma^k]\equiv {\rm tr}(\hat{\gamma}^k)= (\gamma^k)^\mu_{~\mu}$.
The parameters $b_k$ can apriori be arbitrary, but
if one requires the flat space to be a solution of the theory and
$m$ to be the Fierz-Pauli
mass of the gravitons in flat space, then
the five $b_k$ are
expressed in terms of two arbitrary
parameters, sometimes called $c_3$ and $c_4$, as
\begin{equation} \label{bbb}
b_0=4c_3+c_4-6,~~
b_1=3-3c_3-c_4,~~
b_2=2c_3+c_4-1,~~
b_3=-(c_3+c_4),~~
b_4=c_4.
\end{equation}
The metric $g_{\mu\nu}$ and the scalars $\Phi^A$ are the variables of the theory.
Varying the action (\ref{1}) with respect to $g_{\mu\nu}$
gives the Einstein equations
$G_{\mu\nu}=T_{\mu\nu}$
with
\begin{eqnarray} \label{T0}
T^\mu_{~\nu}&=&
\{b_1\,{\cal U}_0+b_2\,{\cal U}_1+b_3\,{\cal U}_2
+b_4\,{\cal U}_3\}\gamma^\mu_{~\nu}
-\{b_2\,{\cal U}_0+b_3\,{\cal U}_1+b_4\,{\cal U}_2\}(\gamma^2)^\mu_{~\nu} \nonumber \\
&+&\{b_3\,{\cal U}_0+b_4\,{\cal U}_1\}(\gamma^3)^\mu_{~\nu}
-b_4\,{\cal U}_0\,(\gamma^4)^\mu_{~\nu}-{\cal U}\,\delta^\mu_\nu \,.
\end{eqnarray}
Varying the action with respect to $\Phi^A$ gives
the conservation conditions
$
{\nabla}_\mu T^\mu_\nu=0\,.
$
These are equations for the Stuckelberg scalars, but they
are actually not independent and
follow from the Bianchi identities for the Einstein equations.
\section{de Sitter space}
\setcounter{equation}{0}
The above field equations admit solutions for which the physical metric is de Sitter.
Specifically,
the de Sitter space can be globally visualized
as the hyperboloid
\begin{equation}
-X_0^2+\sum_i X_i^2+X_4^2=\alpha^2
\end{equation}
is the 5D Minkowski space with the metric
\begin{equation} \label{dSS}
ds^2=-dX_0^2+\sum_i dX_i^2+dX_4^2.
\end{equation}
Rescaling the coordinates,
$X_0=\alpha t$, $X_i=\alpha x_i$, $X_4=\alpha r$
with $x_i\equiv (x,y,z)$,
the metric reads
\begin{eqnarray} \label{dS}
ds_g^2&=&\alpha^2\left\{-dt^2+dr^2+dx^2+dy^2+dz^2
\right\}
\nonumber \\
&=&\alpha^2\left\{-dt^2+dr^2+dR^2+R^2\,d\Omega^2
\right\}
\end{eqnarray}
where $d\Omega^2=d\vartheta^2+\sin^2\vartheta d\varphi^2$ and
\begin{equation} \label{hyper}
R^2\equiv x^2+y^2+z^2=1+t^2-r^2.
\end{equation}
Let us choose the flat
reference metric as
\begin{equation} \label{dSf}
ds_f^2=\alpha^2 u^2\left\{-dT^2+dX^2
+dY^2+dZ^2 \right\},
\end{equation}
where $u$ is a constant and $T,X,Y,Z$ are the Stuckelberg fields.
It turns out that $\alpha$, $u$ and $T,X,Y,Z$
can be chosen such that the two metrics fulfill the field equations.
Specifically, it is sufficient to make sure that
they fulfill the following relation (the Gordon ansatz) \cite{Visser},
\begin{equation} \label{Gordon}
f_{\mu\nu}=\omega^2\left(g_{\mu\nu}+(1-\zeta^2)V_\mu V_\nu\right),
\end{equation}
where $\omega,\zeta$ are functions and
\begin{equation} \label{norm}
g^{\mu\nu}V_\mu V_\nu\equiv V^\mu V_\mu=-1.
\end{equation}
If Eq.\eqref{Gordon} is fulfilled,
then one can see at once that
\begin{equation}
\gamma^\mu_{~\nu}
=
\sqrt{g^{\mu\alpha}f_{\alpha\nu} }
=\omega\left(
\delta^\mu_\nu+(1-\zeta)V^\mu V_\nu
\right),
\end{equation}
since $\gamma^\mu_\alpha\gamma^\alpha_\nu=g^{\mu\alpha}f_{\alpha\nu}$.
One has
$
(\gamma^n)^\mu_{~\nu}=\omega^n\left(
\delta^\mu_\nu+(1-\zeta^n)V^\mu V_\nu
\right)
$
and so
the energy-momentum tensor \eqref{T0} is
\begin{equation} \label{TT}
T^\mu_\nu=-\left\{P_0(\omega)-\zeta\omega P_1(\omega)\right\}\delta^\mu_\nu
+\omega(\zeta-1)P_1(\omega)V^\mu V_\nu\,
\end{equation}
with
\begin{equation}
P_m(\omega)= b_m+2b_{m+1}\,\omega+b_{m+2}\,\omega^2;~~~~~m=0,1,2.
\end{equation}
Let us set $\omega=u$ where $u$ is a constant chosen such that
\begin{equation} \label{u}
P_1(u)=0.
\end{equation}
Then
the energy-momentum tensor \eqref{TT} reduces to
$T^\mu_\nu=-P_0(u)\delta^\mu_\nu$
and the
Einstein equations
become
\begin{equation}
G^\mu_\nu+\Lambda\delta^\mu_\nu=0
\end{equation}
with $\Lambda=P_0(u)$.
The de Sitter metric \eqref{dS} is a solution of these equations provided that
\begin{equation} \label{alpha}
\frac{1}{\alpha^2}=\frac{\Lambda}{3}=\frac{P_0(u)}{3}.
\end{equation}
Therefore, the metrics \eqref{dS} and \eqref{dSf} will indeed fulfill the
field equations if $u$ and $\alpha$ are defined by \eqref{u},\eqref{alpha},
provided that one can adjust
the functions $T,X,Y,Z$ such that the Gordon relation \eqref{Gordon}
is fulfilled.
Let us choose in \eqref{dSf} $T=T(t,r)$, $X=x$, $Y=y$, $Z=z$ so that the f-metric becomes
\begin{equation} \label{dSff}
ds_f^2=\alpha^2 u^2\left\{-dT^2+dx^2
+dy^2+dz^2 \right\}
=
\alpha^2 u^2\left\{-dT^2+dR^2
+R^2 d\Omega^2 \right\}.
\end{equation}
The two metrics
\eqref{dS} and \eqref{dSff} are related to each other as
\begin{equation}
ds_f^2=u^2\left(
ds_g^2+dt^2-dr^2-dT^2
\right).
\end{equation}
This will be compatible with the Gordon relation \eqref{Gordon} if
\begin{equation} \label{grd}
\partial_\mu t\partial_\nu t-
\partial_\mu r\partial_\nu r-
\partial_\mu T\partial_\nu T=(1-\zeta^2)V_\mu V_\nu\,.
\end{equation}
Assuming that the indices $\mu,\nu$ correspond to $(t,r,\vartheta,\varphi)$
yields $V_\vartheta=V_\varphi=0$ and
\begin{eqnarray} \label{TTT}
(\partial_t T)^2-1=(\zeta^2-1)V_t^2\,, \nonumber \\
(\partial_r T)^2+1=(\zeta^2-1)V_r^2\,, \nonumber \\
\partial_t{T}\,\partial_r T=(\zeta^2-1)V_t V_r\,.
\end{eqnarray}
From the first two of these relations one obtains
\begin{equation} \label{VV}
V_t^2=\frac{ (\partial_t {T})^2-1}{\zeta^2-1},~~~~~~
V_r^2=\frac{ (\partial_ r T)^2+1}{\zeta^2-1},
\end{equation}
while the normalization condition \eqref{norm} determines $\zeta$.
Finally, inserting \eqref{VV} to the third relation in \eqref{TTT} yields
\begin{eqnarray}
(\partial_t {T})^2 (\partial_r T)^2 =
( (\partial_t {T})^2-1)( (\partial_r T)^2+1 )
\end{eqnarray}
and therefore
\begin{equation} \label{T}
(\partial_t{T})^2-(\partial_r T)^2=1.
\end{equation}
This completes the procedure, because $V_\mu$ and $\zeta$ are determined by the above
formulas and
the Gordon relation is fulfilled.
Summarizing, the de Sitter solution in the theory is described by
Eqs.\eqref{dS},\eqref{dSff} where $u$,$\alpha$ are defined by \eqref{u},\eqref{alpha}
and $T$ is a solution of the differential equation \eqref{T}.
Since there are infinitely many $T$'s subject to \eqref{T},
there are infinitely many de Sitter solutions. They all
have the same physical metric \eqref{dS} but differ one from the other
by the choice of $T$ in the reference metric \eqref{dSff}.
The physical properties of solutions with different $T$'s,
as for example their stability, can be different.
These solutions were actually obtained previously
\cite{D'Amico:2011jj},\cite{Gratia:2012wt},\cite{Volkov:2012cf,*Volkov:2012zb,*Volkov:2013roa},
but within a different computation scheme yielding the $T$-equation in a
form that gives little hope to solve it (see Eq.\eqref{c0} in the Appendix).
Our procedure yields its equivalent form
\eqref{T}, which is simple
and admits a general solution.
In addition, by slightly modifying our procedure, we can obtain new things.
Specifically,
it was assumed in the above derivation
that both metrics have the same spatial $SO(3)$ symmetry.
However, let us rather choose
\begin{equation} \label{dSfff1}
ds_f^2=\alpha^2 u^2\left\{-dt^2+dx^2
+dy^2+dZ^2 \right\}
\end{equation}
with $Z=Z(r,z)$, so that the two metrics share the same $SO(1,2)$ symmetry in the
$t,x,y$ subspace.
Repeating the above analysis one obtains
\begin{equation} \label{Ta1}
(\partial_r{Z})^2+(\partial_z Z)^2=1,
\end{equation}
and this gives new solutions.
When expressed in the standard spherical coordinates, their f-metric will not look spherically
symmetric, since for generic $Z$ it has no common with the g-metric $SO(3)$ symmetry,
although it has its own $SO(3)$ in the $x,y,Z$ space.
Below
we shall mainly be discussing equation \eqref{T} since the analysis of \eqref{Ta1}
is similar.
\section{The simplest solution}
\setcounter{equation}{0}
Even though there are infinitely many solutions of Eq.\eqref{T},
almost all known dRGT cosmologies reported
in the literature correspond to just one simplest solution,
\begin{equation} \label{sol}
T=t.
\end{equation}
A slightly more general choice is
\begin{equation} \label{sol0}
T=\cosh(\xi)\,t+\sinh(\xi)\, r
\end{equation}
with a constant $\xi$. However, the value of $\xi$ can be changed
by a boost in the $t,r$ plane of the ambient 5D Minkowski space,
which does not affect the g-metric \eqref{dS}, hence one can set
$\xi=0$ without loss of generality.
Rewriting \eqref{sol} in different coordinates gives results which
look very different, and it has not been recognized that they actually
describe
the same solution. Let us see what happens when this solution
is expressed in the standard
spatially flat, closed, and open coordinate systems.
\subsubsection{Flat slicing}
Let us express $t,r$,$R$ in \eqref{dS} in terms of
two new coordinates $\tau$ and $\rho$ as
\begin{equation}
t=\sinh\tau+\frac{\rho^2}{2}e^\tau,~~~
r=\cosh\tau-\frac{\rho^2}{2}e^\tau,~~~
R=e^\tau\rho\,.
\end{equation}
This solves the constraint \eqref{hyper} and transforms the de Sitter metric \eqref{dS}
to the standard FLRW form with flat spatial sections,
\begin{equation} \label{flat}
ds_g^2=\alpha^2\{-d\tau^2+a^2(\tau)(d\rho^2+\rho^2d\Omega^2)\},
\end{equation}
where $a(\tau )=e^\tau$. The function $T=t$ can be represented as
\begin{equation}
T=\frac12\int\frac{d\tau}{\dot{a}(\tau)}+\frac12\left(1+\rho^2 \right)a(\tau).
\end{equation}
This solution was found in Ref.\cite{D'Amico:2011jj}
for $b_k$ given by \eqref{bbb} with $c_3=c_4=0$, and later
for arbitrary $b_k$ \cite{Volkov:2012cf,*Volkov:2012zb,*Volkov:2013roa}
(solution in \cite{D'Amico:2011jj},\cite{Volkov:2012cf,*Volkov:2012zb,*Volkov:2013roa}
contains an integration constant that can be obtained by using \eqref{sol0} instead of \eqref{sol}).
Although the g-metric \eqref{flat}
is manifestly homogeneous and isotropic, the f-metric \eqref{dSff}, when expressed in the
$\tau,\rho$
coordinates, becomes non-diagonal and $\rho$-dependent, which suggests that it is inhomogeneous.
For this reason it is sometimes said that the dRGT theory does not
admit genuinely homogeneous and isotropic cosmologies with flat spatial sections
\cite{D'Amico:2011jj}.
However, we shall shortly comment on this.
\subsubsection{Closed slicing}
If one chooses in \eqref{dS}
\begin{equation}
t=\sinh(\tau),~~~r=\cosh(\tau)\cos(\rho),~~~~R=\cosh(\tau)\sin(\rho),
\end{equation}
this solves the constraint \eqref{hyper} and the de Sitter metric \eqref{dS}
assumes the FLRW form with closed spatial sections,
\begin{equation}
ds_g^2=\alpha^2\{-d\tau^2+a^2(\tau)(d\rho^2+\sin^2(\rho)d\Omega^2)\}
\end{equation}
with $a(\tau)=\cosh(\tau)$. These coordinates
cover the whole of de Sitter space.
The Stuckelberg field
is
$
T=\sinh(\tau),
$
and the f-metric \eqref{dSff} expressed in the $\tau,\rho$ coordinates
is again non-diagonal and $\rho$-dependent, which suggests that there are no
genuinely homogeneous and isotropic cosmologies with closed spatial sections either.
\subsubsection{Open slicing}
For the open slicing one has
\begin{equation}
t=\sinh(\tau)\cosh(\rho),~~~~r=\cosh(\tau),~~~R=\sinh(\tau)\sinh(\rho),
\end{equation}
and the g-metric becomes
\begin{equation}
ds_g^2=\alpha^2\{-d\tau^2+a^2(\tau)(d\rho^2+\sinh^2(\rho)d\Omega^2)\}
\end{equation}
with $a(\tau)=\sinh(\tau)$. The Stuckelberg field is
$
T=\sinh(\tau)\cosh(\rho)
$
and the specialty now is that
the f-metric \eqref{dSff} becomes diagonal in the $\tau,\rho$ coordinates,
\begin{equation} \label{muk}
ds_f^2=\alpha^2 u^2\{-\cosh(\tau)^2d\tau^2+a^2(\tau)(d\rho^2+\sinh^2(\rho)d\Omega^2)\}.
\end{equation}
This solution, discovered in Ref.\cite{Gumrukcuoglu:2011ew},
is broadly viewed as the only homogeneous and
isotropic dRGT cosmology, because both metrics
are manifestly homogeneous and isotropic, so that they share the same
rotational and translational Killing symmetries.
However, this solution is completely equivalent to its
flat and closed versions. Therefore, the latter also have the same
common isometries, hence they are all
homogeneous and isotropic, although their symmetries are not manifest.
The conclusion is that sometimes solutions can be FLRW in a non-manifest way.
At the same time, although homogeneous and isotropic,
the solution $T=t$ is not static whereas the de Sitter space is.
Specifically,
let us consider the
\subsubsection{Static slicing}
Setting
\begin{equation} \label{ss}
t=\sqrt{1-\rho^2}\,\sinh(\tau),~~~r=\sqrt{1-\rho^2}\,\cosh(\tau),~~~R=\rho
\end{equation}
solves the condition \eqref{hyper} and reduces the de Sitter metric \eqref{dS} to
the static form
\begin{equation}
ds_g^2=\alpha^2\left\{
-\Sigma\,d\tau^2+\frac{d\rho^2}{\Sigma}+\rho^2d\Omega^2
\right\}
\end{equation}
with $\Sigma=1-\rho^2$.
The $T=t$ solution then becomes
\begin{equation} \label{sol1}
T(\tau,\rho)=\sqrt{1-\rho^2}\,\sinh(\tau),
\end{equation}
and it is non-static even in static coordinates. Therefore,
the g-metric is invariant
under the action of the (locally) timelike Killing vector $\partial/\partial\tau$,
but the Stuckelberg field $T$ and the f-metric are not invariant.
As a result, the timelike isometry is not shared by both
metrics.
As the solution $T=t$ is not static, it is unlikely to
describe the ``ground state'' of the theory. This is probably the reason
why this solution was found to be unstable \cite{DeFelice:2012mx,*DeFelice:2013awa}.
Therefore, we need to consider other solutions for $T$.
\section{Other solutions}
\setcounter{equation}{0}
Solutions of the $T$-equation
$(\partial_t{T})^2-(\partial_r T)^2=1$ can be constructed in different ways.
A fairly general
solution
containing an arbitrary function $W(\xi)$ is given by \cite{Courant},
\begin{eqnarray}
T&=&\cosh(\xi)\,t+\sinh(\xi)\,r+W(\xi)\,, \nonumber \\
0&=&\sinh(\xi)\,t+\cosh(\xi)\,r+\frac{dW(\xi)}{d\xi}\,,
\end{eqnarray}
where the second line implicitly determines the dependence of
$\xi$ on $t,r$. Together with \eqref{sol0}, this gives if not all but
probably almost all solutions. However, this formula is difficult to use
since one cannot explicitly determine $\xi(t,r)$ for a generic $W(\xi)$.
\begin{figure}[h]
\hbox to \linewidth{ \hss
\resizebox{6cm}{5cm}
{\includegraphics{T.eps}}
\hspace{1mm}
\hss}
\caption{{\protect\small
The $T$-equation in the method of characteristics.
}}%
\label{Fig}
\end{figure}
The $T$-equation can also be integrated by applying the method of
characteristics \cite{Courant}, which has a simple geometric
interpretation. Let us consider the 2D Minkowski space spanned by
$x^a\equiv \{t,r\}$
with the metric $g_{ab}={\rm diag}[1,-1]$. The $T$-equation reads
$g^{ab}\partial_a T\partial_b T\equiv \langle \nabla T,\nabla T\rangle =1$.
Let $\gamma=x^a(s)$ be a spacelike curve
and $T$ is constant along it. At every point of $\gamma$
there is a unit timelike normal $n$ such that $\langle n,n\rangle=1$ and
$\langle n,\partial/\partial s\rangle =0$.
The $T$-equation is equivalent to
$\partial T/\partial n=1$ \cite{Courant}.
This allows one to pass from $\gamma$ where $T=T(\gamma)$ to a neighboring
curve $\tilde{\gamma}$ where $T=T(\tilde{\gamma})$ (see Fig.\ref{Fig}) and so on,
thereby extending $T$
to the whole of the space. The solution is therefore defined, up to an additive constant,
by the choice of the initial curve $\gamma$. For example, the solution \eqref{sol0}
can be obtained by choosing $\gamma$ to be a straight line.
In practice solutions of $(\partial_t{T})^2-(\partial_r T)^2=1$ can be obtained
by changing the variables and then separating them \cite{Khosravi:2013axa}.
Let us illustrate the method
by passing to the static
coordinates \eqref{ss}, in which case the $T$-equation becomes
\begin{equation} \label{sol11}
\frac{1}{\Sigma}\left(\frac{\partial T}{\partial \tau}\right)^2
-
\frac{\Sigma}{1-\Sigma}\left(\frac{\partial T}{\partial \rho}\right)^2=1.
\end{equation}
It is easy to see that $T(\tau,\rho)$ given by \eqref{sol1} fulfills this equation,
but now we can obtain also other solutions, in particular those for which
$dT$ does not depend on time and the f-metric is static.
The most general solution of \eqref{sol11} of this type
is obtained by separating the variables,
\begin{equation} \label{static}
T=\sqrt{1+q^2}\,\tau+\int\frac{\rho \,d\rho}{\Sigma}\,\sqrt{q^2+\rho^2}\,,
\end{equation}
where $q$ is an integration constant. If $q=0$ then the solution becomes
especially simple,
\begin{equation}
T=\tau+\int\frac{d\rho}{\Sigma}-\rho\equiv V-\rho,
\end{equation}
and choosing $V$ and $\rho$ as coordinates, the two metrics become
\begin{eqnarray}
ds_g^2&=&\alpha^2\{-\Sigma\, dV^2+2dV d\rho+\rho^2d\Omega^2\}, \nonumber \\
ds_f^2&=&u^2\alpha^2\{-dV^2+2dV d\rho+\rho^2d\Omega^2\}.
\end{eqnarray}
\section{Energy}
One can compute the canonical energy for systems with non-trivial Stuckelberg fields
in the same way as this is done in the unitary gauge
\cite{Volkov:2014qca,*Volkov:2014ida}.
The computation will be presented separately \cite{MM} but its result is as follows.
For a solution expressed in the static coordinates $\tau,\rho$ the energy
on a hypersurface of constant $\tau$ is
\begin{equation}
E=\int {\cal E}\,d\rho
\end{equation}
with the radial energy density
\begin{equation}
{\cal E}=u^2 P_2(u)\rho^2 \partial_\tau T.
\end{equation}
Applying this formula to the $T=t$ solution \eqref{sol1} gives
the time-dependent value,
\begin{equation}
{\cal E}=u^2 P_2(u)\rho^2\sqrt{1-\rho^2}\cosh(\tau),
\end{equation}
which indicates once again that this solution cannot describe the ground state.
On the other hand, the energy will be time independent
if $\partial_\tau T$ is time independent,
but all such solutions are given by \eqref{static}, in which case
\begin{equation}
{\cal E}=u^2 P_2(u)\sqrt{1+q^2}\rho^2.
\end{equation}
This corresponds to the constant volume energy density
\begin{equation}
\epsilon=u^2 P_2(u)\sqrt{1+q^2},
\end{equation}
and the
total energy is $E=\epsilon V$ where $V$ is the (infinite) volume of the 3-space.
We remember that $u$ is a solution of the algebraic equation $P_1(u)=0$,
therefore, depending on choice of $u$ and also on values of the parameters $b_k$,
the energy can be positive, negative, or zero.
The actual value of the background energy is probably not so important,
but it is important to know if the energy is minimal or not.
We conjecture that the static solutions \eqref{static} correspond to the energy minima and
are therefore stable. Therefore, they are candidates for describing the de Sitter
ground state in the theory. To prove the conjecture will
require to resolve the constraints and to compute the energy for deformations of the background
\cite{Volkov:2014qca,*Volkov:2014ida}. We presently have partial results
supporting our conjecture, but the detailed analysis will be presented elsewhere \cite{MM}.
\section{Conclusions}
We have shown that the de Sitter vacua in the dRGT theory are labeled by solutions of
$(\partial_t{T})^2-(\partial_r T)^2=1$. The simplest solution $T=t$
is manifestly homogeneous and isotropic when written in the open chart, but it is unstable.
Therefore, one should study other solutions. One could worry that other solutions
will not be FLRW because their reference metric is inhomogeneous.
However, as we have seen, this is not
necessarily the case, as the reference metric can look inhomogeneous in some coordinates
while sharing common translational isometries with the physical metric.
The important issue is the number of common isometries of the two metrics. Since each
of them describes a maximal symmetry space, each metric has ten isometries, some of which can be
common, as for example the SO(3) rotational isometries. The number of common isometries depends
on choice of $T$, for example for $T=t$ this number is six, but the same
can be true for other choices of $T$ as well.
Requiring the timelike isometry to be common for both metrics reduces the
set of solutions to a one-parameter family \eqref{static}. These solutions are physically
distinguished since only for them the energy is time-independent. We conjecture that
these solutions are stable and describe therefore the de Sitter ground state of the theory.
The stability will follow if one shows that the energy increases for deformations
of the de Sitter background, but such an analysis goes beyond the scope of the present paper
and will be reported separately \cite{MM}.
\acknowledgments
We are grateful to Gary Gibbons and Kei-ichi Maeda for constructive suggestions.
The work of M.S.V. was partly supported by the Russian Government Program of Competitive Growth
of the Kazan Federal University.
|
\section{Introduction}
The relation between answers of definite logic programs and their least Herbrand
models is not trivial.
In some cases the equivalence
\begin{equation}
\label{eq}
{\ensuremath{\cal M}}\xspace_P\models Q \ \ \mbox{ iff } \ \ P\models Q
\end{equation}
does not hold
(where $P$ is a definite program, ${\ensuremath{\cal M}}\xspace_P$ its least Herbrand model, and $Q$ a
query, i.e.\ a conjunction of atoms%
\footnote
{%
The semantics of non closed formulae is understood as usually
(see e.g.\ \cite{vanDalen,Apt-Prolog}),
so that ${\it IT}\models Q$ iff
${\it IT}\models\forall Q$, where ${\it IT}$ is an interpretation or a
theory, $Q$ a formula, and $\forall Q$ its
universal closure.%
}%
).
So programs with the same least Herbrand model may have different sets of
answers.
(By definition, $Q$ is an answer of $P$ iff $P\models Q$.)
For a simple counterexample \cite[Exercise 4.5]{Doets},
assume that the underlying language has only one function symbol, a constant $a$.
Take a program $P=\{\, p(a)\,\}$.
Now ${\ensuremath{\cal M}}\xspace_P\models p(X)$ but $P\mathrel{\,\not\!\models} p(X)$.
This counterexample can be in a natural way generalized for any finite set
of function symbols,
see the comment following the proof of Prop.\,\ref{prop:counterexample}.
Equivalence (\ref{eq}) holds for ground queries
(\citeNP{Lloyd87}, Th.\,6.6;\,\,\citeNP{Apt-Prolog}, Th.\,4.30).
For a possibly nonground $Q$ (and a finite $P$)
a sufficient condition for (\ref{eq})
is that there are infinitely many constants in the
underlying language
(\citeNP
{DBLP:books/mk/minker88/Maher88}; \citeNP[Corollary 4.39]{Apt-Prolog}).
\citeN{DBLP:books/mk/minker88/Maher88} states without proof
that instead of an infinite supply of constants it is sufficient that there is
a non constant function symbol not occurring in $P,Q$.
The author is not aware of any proof of this property
(except for \cite[Appendix]{drabent.arxiv.coco14}).
This paper presents a more general sufficient condition,
and shows that the condition is also a necessary one.
To obtain the sufficient condition, we show a property of (possibly
nonground) atoms containing symbols not occurring in a program $P$.
Namely, when such atom is true in ${\ensuremath{\cal M}}\xspace_P$ then, under certain conditions, a
certain more general atom is a logical consequence of $P$.
As an initial step, we obtain
a generalization of the theorem on constants \cite{shoenfield67},
for a restricted class of theories, namely definite clause programs.
We also give an alternative proof for the original theorem.
\paragraph{Related problem.}
This paper studies (in)equivalence of
two views at the declarative semantics of definite
clause programs. One of them considers answers true in the least Herbrand
models of programs, the other -- answers that are logical consequences of
programs.
The subject of this paper should be compared with a related issue
(which is outside of the scope of this paper).
There exists (in)equivalence between the declarative
semantics and the operational one, given by SLD-resolution.
As possibly first pointed in
(\citeNP{DM87};\,\,\citeyearNP{DM88}),
two logically equivalent programs
(i.e.\ with the same models, and thus the same logical consequences)
may have different sets of SLD-computed answers for the same query.
For instance take
$P_1= \{\, p(X).\,\}$, and
$P_2= \{\, p(X).\ p(a).\,\}$
Then for a query $p(Y)$ program $P_2$ gives two distinct computed answers,
and $P_1$ one. This phenomenon gave rise to the {\em s-semantics},
see e.g.\ \cite{DBLP:journals/tcs/Bossi09} for overview and references.
\paragraph{Preliminaries.}
We consider definite clause logic programs. A query is a conjunction of atoms.
A query $Q$ is an {\em answer} (or a {\em correct answer}) of a program $P$
iff $P\models Q$.
\citeN{Apt-Prolog} calls it a correct instance (of some query).
We do not need to refer to SLD-computed answers, as
each computed answer is an
answer, and each answer is a computed answer for some query,
by soundness and completeness of SLD-resolution.
Similarly, we do not need to consider to which query $Q_0$ a given query is
an answer.
The Herbrand universe
(for the alphabet of function symbols of the underlying language)
will be denoted by {\ensuremath{\cal H U}}\xspace, and the least Herbrand model of a program $P$ by ${\ensuremath{\cal M}}\xspace_P$.
Remember that ${\ensuremath{\cal M}}\xspace_P$ depends on the underlying language.
We require ${\ensuremath{\cal H U}}\xspace\neq\emptyset$.
Names of variables will begin with an upper-case letter.
Otherwise we use the standard definitions and notation of \cite{Apt-Prolog},
including the list notation of Prolog.
(However in discussing the semantics of first order formulae we use a
standard term ``variable assignment'' instead of ``state''
used in \cite{Apt-Prolog}.)
The paper is organized as follows. The next section presents some
necessary definitions.
Section \ref{sec:lemma}
shows how existence of answers containing symbols not occurring in the program
implies existence of more general answers. The main result of this section
is compared with theorem on constants \cite{shoenfield67}.
Section \ref{sec:main}
contains the central technical lemma of this paper.
Section \ref{sec:H}
studies when the least Herbrand models provide an exact characterization of
program answers. A new sufficient condition for equivalence
(\ref{eq}) is presented, and it is shown in which sense the condition is a
necessary one.
\section{Definitions}
This section introduces three notions needed further on.
Let ${\ensuremath{\mathscr F}}\xspace$ be the set of function symbols of the underlying language;
let $F\subseteq {\ensuremath{\mathscr F}}\xspace$. An {\em alien} w.r.t.\ $F$
is a non-variable term with its main function symbol from ${\ensuremath{\mathscr F}}\xspace\setminus F$.
An alien w.r.t.\ a theory $T$ (for instance a program)
means an alien w.r.t.\ the set of function symbols occurring in $T$.
An occurrence of an alien $t$ (w.r.t.\ $F$, in an atom or substitution)
will be called a {\em maximal alien} if the occurrence is not within
an alien $t'\neq t$.
By a generalization of a query we mean the result of systematic replacement
of maximal aliens in the query by new variables. More formally,
let \P be a theory or a set of function symbols.
Let the maximal aliens of a query $Q$ w.r.t.\ $\P$
be the occurrences in $Q$ of distinct terms $\seq t$.
Let $\seq V$ be distinct variables not occurring in $Q$.
Let a query $Q'$ be
obtained from $Q$ by replacing (each occurrence of)
$t_i$ by $V_i$, for $i=1,\ldots,n$.
(So $Q=Q'\{V_1/t_1,\ldots,V_n/t_n\}$.)
Such $Q'$ will be called
{\em $Q$ generalized} for $\P$.
We will also call it a/the {\em generalization} of $Q$ (for \P).
Note that it is unique up to variable renaming.
\begin{example}
The standard append program APPEND \cite[p.\,127]{Apt-Prolog}
contains two function symbols $[\,]$ and $[\ |\ ]$.
Terms $a,f([a,b])$ are aliens w.r.t.\ APPEND, term $[a,b]$ is not.
Maximal aliens in $A = app( [a], [[\,]\,|\,g(a,X)], [g(a,Y),Z,[a]] )$
are the first and the last occurrences of $a$ and the (single) occurrences of
$g(a,X)$ and $g(a,Y)$.
Atom $app( [V_1], [[\,]|V_2], [V_3,Z,[V_1]] )$
is $A$ generalized for APPEND.
\end{example}
Let $Q'$ be a query not containing aliens w.r.t.\ \P, and
$\theta$ be a substitution such that $Dom(\theta)\subseteq {\it Var}(Q')$.
Then $Q'$ is a generalization of $Q'\theta$ for \P (and for $\P\cup\{Q'\}$)
iff $\theta=\{V_1/t_1,\ldots,V_n/t_n\}$
where $\seq t$ are distinct aliens w.r.t.\ \P.
The correspondence between a ground atom and its generalization
is described, in other terms, in
\cite[Def.\,4]{naish.tplp.floundering14}.
%
It is used in that paper to represent nonground atoms by
ground ones, in analysis of floundering in the context of delays.
\section{On program answers and aliens}
\label{sec:lemma}
Given a query containing aliens which is an answer of a program $P$,
this section shows which more general queries are answers of $P$.
The main result (Lemma \ref{lemma:alien:substitution})
is compared with theorem on constants,
used by \cite{DBLP:books/mk/minker88/Maher88} to prove
equivalence (\ref{eq}) for a case with an infinite alphabet of constants.
It is rather obvious that answers containing aliens can be generalized.
Assume that a query
$Q$ is an answer of $P$, and that $Q$ contains aliens w.r.t.\ $P$.
Then $Q$ is
a proper instance of some computed answer $Q'$.
It is however not obvious which replacements of aliens in $Q$ by variables
result in answers.
\begin{example}
By replacing aliens w.r.t.\ $P$ by variables in an answer $Q$,
we obtain some queries which are answers of $P$,
and some which are not.
Let $P= \{p(X,X,Y)\}$ and $Q=p( f(a),f(a),b )$.
So $P\models Q$.
Now $p(f(V_1),V_2,b)$ and $p(V_1,V_2,b)$ are not answers of $P$,
but $p(f(V),f(V),Z)$, $p(V,V,b)$ and $p(V,V,Z)$ are.
\end{example}
\begin{lemma}
\label{lemma:alien:substitution}
Let $P$ be a program, $Q$ a query, and
$\rho=\{V_1/t_1,\ldots,V_k/t_k\}$ be a substitution where
$\seq[k]t$ are distinct aliens w.r.t.\ $P\cup\{Q\}$. Then
\begin{equation}
\label{th:alien:substitution1}
P\models Q \ \ \mbox{ iff } \ \ P\models Q\rho\,.
\end{equation}
\end{lemma}
Note that terms $\seq[k]t$ may be nonground
(and may contain variables from $\{\seq[k]V\}$), some
$t_i,t_j$ may be unifiable, or contain common variables,
$Q$ may contain variables other than $\seq V$
and may contain aliens w.r.t. $P$.
So $Q$ is not necessarily a generalization of $Q\rho$ for $P$,
but it is one for $P\cup\{Q\}$.
\begin{example}
In the previous example, the cases in which the more general atom is an
answer of $P$ satisfy conditions of Lemma \ref{lemma:alien:substitution},
and the remaining ones do not.
\end{example}
\begin{proof}[Proof (Lemma \ref{lemma:alien:substitution})]
Without loss of generality assume that variables $\seq[k]V$ occur in $Q$.
Let $\seq[l]X$ be the remaining variables of $Q$.
The ``only if'' case is obvious.
Assume $P\models Q\rho$.
By completeness of SLD-resolution,
$Q\rho$ is an instance of some computed answer $Q\varphi$ for $P$ and $Q$:
$Q\rho=Q\varphi\sigma$.
Each function symbol occurring in $\varphi$ occurs in $P$ or $Q$.
Moreover (for $i=1,\ldots,k$) $t_i=V_i\varphi\sigma$
and the main symbol of $t_i$ does not occur in $V_i\varphi$;
hence $V_i\varphi$ is a variable.
As $\seq[k] t$ are distinct, variables $V_1\varphi,\ldots,V_k\varphi$ are
distinct.
Similarly, $X_j=X_j\varphi\sigma$ for $j=1,\ldots,l$, thus
%
$V_1\varphi,\ldots,V_k\varphi,X_1\varphi,\ldots,X_l\varphi$
are distinct variables.
%
Thus $Q\varphi$ is a variant of $Q$ and, by soundness of
SLD-resolution, $P\models Q$.
\end{proof}
\begin{corollary}
\label{th:generalized}
\label{cor:generalized}
Let $P$ be a program, $Q$ a query, and $Q'$ be $Q$ generalized for $P$.
Then $P\models Q$ iff $P\models Q'$.
\end{corollary}
\begin{proof}
$Q=Q'\rho$ for a certain $\rho=\{V_1/t_1,\ldots,V_k/t_k\}$.
The premises of Lemma \ref{lemma:alien:substitution}
are satisfied by $P$, $Q'$, and $\rho$
(as $\seq[k]t$ are aliens w.r.t.\ $P$, but also w.r.t.\ $Q'$).
\end{proof}
\begin{example}
Consider again program APPEND. Assume that the underlying language
has more function symbols than those occurring in the program,
i.e.\ $[\,]$, $[\ |\ ]$.
Assume that we know that the least Herbrand model ${\ensuremath{\cal M}}\xspace_{\rm APPEND}$
contains an atom
$Q =\linebreak[3] app([\seq[m]t], [\SEQ{t}{m+1}k], [\seq[k]t])$,
where $\seq[k]t$ are distinct aliens w.r.t.\ APPEND.
Note that $P\models Q$, as equivalence (\ref{eq}) holds for ground queries.
{\sloppy\par}
By Corollary \ref{cor:generalized},
${\rm APPEND}\models app([\seq[m]V], [\SEQ{V}{m+1}k], [\seq[k]V])$,
where $\seq[k] V$ are distinct variables.
Hence, for any terms $\seq[k] s$,
\mbox{%
${\rm APPEND}\models app([\seq[m]s], [\SEQ{s}{m+1}k], [\seq[k] s])$%
}.
{\sloppy\par}
\end{example}
\begin{example}
Consider the map colouring program \cite[Program 14.4]{Sterling-Shapiro}.
%
%
%
We skip any details, let us only mention that the names of colours and
countries do not occur in the program.
(The function symbols occurring in the program are
$F=\{\, [\,],
[\, |\,],
{\it region} \,\}
$.)
By Corollary \ref{cor:generalized},
for any answer $Q$ of the program,
the generalization $Q'$ of $Q$ w.r.t.\ $F$ is an
answer of the program. So is each instance of $Q'$.
Thus systematic replacing (some) names of colours or countries in $Q$ by
other terms results in a query $Q''$ which is an answer of the program.%
\footnote{%
Thus it is possible that neighbouring countries get the same colour.
This does not mean that the program is incorrect.
Its main predicate {\it color\_map}
describes a correct map colouring provided that its second argument
is a list of distinct colours.
}
\end{example}
The proof of equivalence (\ref{eq})
for an infinite set of constants of
\cite[proof of Prop.\,6]{DBLP:books/mk/minker88/Maher88} employs a so called
theorem on constants \cite{shoenfield67},
see also free constant theorem in \cite[p.\,56]{HandbookLAILP:FOL}.
The theorem states that (\ref{th:alien:substitution1}) holds for an arbitrary
theory $P$ and formula $Q$, when the distinct aliens $\seq[k]t$ are constants.
Its proofs in \cite{shoenfield67,HandbookLAILP:FOL} are syntactical, but a
rather simple semantic proof is possible:
Let $F$ be the set of function and predicate symbols from $P,Q$, let
{\ensuremath{\cal X}}\xspace be the set of the free variables of $Q$.
Notice that
for any interpretation $I$ (for $F$) and any variable assignment $\sigma$
(for {\ensuremath{\cal X}}\xspace)
there exists a variable assignment $\sigma'$
(for ${\ensuremath{\cal X}}\xspace\setminus\{\seq[k]V\}$)
and an interpretation $I'$
(for $F\cup\{\seq[k]t\}$) such that
$\sigma'(X)=\sigma(X)$ for each $X\in{\ensuremath{\cal X}}\xspace\setminus\{\seq[k]V\}$,
$I'(t_i)=\sigma(V_i)$ for each $i$, and all the symbols
of $F$ have the same interpretation in $I$ and $I'$.
Thus $I\models P$ iff $I'\models P$, and
$I\models_\sigma Q$ iff $I'\models_{\sigma'} Q\rho$.
Conversely, for each interpretation $I'$ for $F\cup\{\seq[k]t\}$
and variable assignment $\sigma'$ for ${\ensuremath{\cal X}}\xspace\setminus\{\seq[k]V\}$
there exist $I,\sigma$ as above. (In particular, the two equivalences hold.)
Now the theorem follows:
\smallskip
\begin{tabular}{l}
$P\models Q$ iff \\
for every $I,\sigma$ (as above) $I\models P$ implies $I\models_{\sigma}Q$
iff \\
for every $I',\sigma'$ (as above) $I'\models P$ implies
$I'\models_{\sigma'}Q\rho$ iff \\
$P\models Q\rho$.
\end{tabular}
\smallskip
\citeN[p.\,634]{DBLP:books/mk/minker88/Maher88}
states that ``The same effect
[as adding new constants] could be obtained with one new function symbol (of
arity $>0$) to obtain new ground terms with new outermost function symbol.''
This idea does not apply to the proof of the previous paragraph;
when $\seq[k]t$ are such terms then the proof fails.%
\footnote{%
Informally, this is because such new terms cannot be interpreted
independently, in contrast to $k$ new constants.
%
Sometimes no interpretation for the new symbol $f$ is possible,
such that $\seq[k]t$ are interpreted as a given $k$ values.
For instance take $t_i=f^i(a)$ for $i=1,\ldots,k$.
Then for any interpretation for $f$,
if $t_1,t_2$ have the same value then all $\seq[k]t$ also have the same value.
} %
So do the proofs of \cite{shoenfield67,HandbookLAILP:FOL}.
In the context of \cite{shoenfield67} -- first order logic with equality -- the
generalization of the theorem on constants to terms with a
new outermost symbol does not hold.
For a counterexample, note that
$\{a=b\}\models f(a)=f(b)$ \ but \
$\{a=b\}\mathrel{\,\not\!\models} V_1=V_2$.
The generalization in Lemma \ref{lemma:alien:substitution} is
sound and has a simple proof, due to restriction to
definite programs and queries.
From Lemma \ref{lemma:alien:substitution} it follows that equivalence (\ref{eq})
holds whenever the underlying language has a non constant function symbol $f$
(or a sufficient number of constants) not occurring in $P,Q$.%
\footnote{%
Assume that $\seq[k]V$ are the variables of $Q$, and that there exist
distinct ground terms $\seq[k]t$ with their main symbols
not occurring in $P,Q$.
Let $\rho=\{V_1/t_1,\ldots,V_k/t_k\}$.
Assume ${\ensuremath{\cal M}}\xspace_P\models Q$, so ${\ensuremath{\cal M}}\xspace_P\models Q\rho$,
and $P\models Q\rho$ as $Q\rho$ is ground.
By Lemma \ref{lemma:alien:substitution}, $P\models Q$.
} %
(See also \cite[Appendix]{drabent.arxiv.coco14} for a direct proof.)
We however aim for a more general sufficient condition for (\ref{eq}),
allowing $f$ to occur in $Q$;
in this case Lemma \ref{lemma:alien:substitution} is not applicable.
\section{Least Herbrand models and program answers}
\label{sec:main}
This section shows conditions under which
truth in ${\ensuremath{\cal M}}\xspace_P$ of a query with aliens implies
that a certain more general query is an answer of $P$.
This is a central technical result of this paper (Lemma \ref{lemma:MP}).
From it,
the sufficient conditions for equivalence (\ref{eq}) follow rather
straightforwardly, as shown in the next section.
We begin with proving an auxiliary property, by means the two following lemmas.
\begin{lemma}
\label{lemma:unifier}
Two distinct terms have at most one unifier of the form $\{X/u\}$ where $u$
is not a variable.
\end{lemma}
\begin{proof}
Let $\theta=\{X/u\}$, $\theta'=\{X'/u'\}$ be distinct substitutions,
where neither of $u,u'$ is a variable.
We show that if $s_1\theta=s_2\theta$ then $s_1\theta'\neq s_2\theta'$,
for any distinct terms $s_1,s_2$.
The proof is
by induction on the sum $|s_1|+|s_2|$ of the sizes of $s_1,s_2$.
(Any notion of term size would do, providing that $|t|<|t'|$ whenever $t$ is
a proper subterm of $t'$.)
Assume that the
property holds for each $s_1',s_2'$ such that $|s_1'|+|s_2'|<|s_1|+|s_2|$.
Let $s_1\neq s_2$ and $s_1\theta=s_2\theta$.
Notice that at most one of $s_1,s_2$ is a variable.
(Otherwise $s_1\theta,s_2\theta$ are $s_1,s_2$ -- two distinct variables,
or exactly one of $s_1\theta,s_2\theta$ is a variable, contradiction.)
Assume that exactly one of $s_1,s_2$, say $s_1$, is a variable.
Then $s_1=X$ (as $s_1\theta\neq s_1$), so
$X$~does not occur in $s_2$ (as $X,s_2$ are unifiable),
hence $s_2\theta=s_2=u$.
Now if $X'\neq X$ then $s_1\theta'=X$ which is distinct from any instance of
$s_2$. Otherwise $X'= X$, hence $s_1\theta'=u'\neq u=s_2=s_2\theta'$.
If both $s_1,s_2$ are not variables then
$s_i=f(\seq[l]{{s_i}})$, for $i=1,2$. For some~$j$, ${s_1}_j\neq{s_2}_j$
and $|{s_1}_j|+|{s_2}_j|<|{s_1}|+|{s_2}|$.
By the inductive assumption,
${s_1}_j\theta'\neq{s_2}_j\theta'$;
thus ${s_1}\theta'\neq{s_2}\theta'$.
\end{proof}
\begin{lemma}
\label{lemma:distinct}
let \P be a theory or a set of function symbols.
Let $\seq[m]t$ be a sequence of distinct terms, where $\seq t$
($0\leq n \leq m$) are variables, and $\SEQ t{n+1}m$ are aliens w.r.t.~$\P$.
Assume that if $\SEQ t{n+1}m$ are ground then there exist ground
aliens $\seq u$ w.r.t.\ $\P$, pairwise distinct from $\SEQ t{n+1}m$.
Then the sequence has
a ground instance $(\seq[m]t)\sigma$ consisting of $m$ distinct aliens
w.r.t.\ $\P$.
\end{lemma}
\nopagebreak
\begin{proof}
Consider first the case of $\SEQ t{n+1}m$ ground.
Then $\sigma=\{t_1/u_1,\ldots,\linebreak[3]t_n/u_n\}$
is a substitution providing the required instance.
{\sloppy\par}
Let some $t_j$ ($n<j\leq m$) be nonground.
Its main symbol, say $f$, is a non-constant function symbol not
occurring in $\P$. Thus the set $Al$ of ground aliens w.r.t.\ $\P$ is infinite.
\pagebreak[3]
Let $\seq[l]X$ be the variables occurring in $\seq[m]t$.
For some $s_1\in Al$ substitution $\theta_1=\{X_1/s_1\}$ is not a unifier of any
pair $t_i,t_j$ ($1\leq i<j\leq m$),
as by Lemma \ref{lemma:unifier}
each such pair has at most one unifier of the form $\{X_1/s\}$, $s\in{\ensuremath{\cal H U}}\xspace$.
Thus $(\seq[m] t)\theta_1$ is a sequence of $m$ distinct terms.
Applying this step repetitively we obtain the required sequence
$(\seq[m] t)\theta_1\cdots\theta_l$ of distinct ground terms.
\end{proof}
\begin{lemma}
\label{lemma:MP}
Let $P$ be a program,
$Q$ an atom, and $Q'$ be $Q$ generalized for $P$.
If
\begin{enumerate}[(a)]
\item
\label{lemma:MP:condition1}
the underlying language has a non-constant function symbol not occurring in~P,
or
\item
\label{lemma:MP:condition2}
$Q$ contains exactly $n\geq0$ (distinct) variables, and
the underlying language has (at least) $n$ constants not occurring in $P,Q$,
\nopagebreak
\end{enumerate}
\nopagebreak
then ${\ensuremath{\cal M}}\xspace_P\models Q$ iff $P\models Q'$.
\end{lemma}
\begin{proof}
Note that $Q= Q'\varphi$ where
$\varphi=\{\, X_1/u_1,\ldots X_m/u_m\,\}$, $\seq[m]X$ are the variables
of $Q'$ not occurring in $Q$, and $\seq[m]u$ are the maximal aliens in $Q$
(precisely: the distinct terms whose occurrences in $Q$ are the maximal aliens
w.r.t.\ $P$).
Let $\seq[n] Y$ be the variables occurring in $Q$.
We construct a ground instance $Q\sigma$ of $Q$, such that
$Q'$ is $Q\sigma$ generalized for $P$.
To apply Lemma \ref{lemma:distinct} to terms $\seq Y,\seq[m]u$,
note that if $\seq[m]u$ are ground then there exist $n$ ground aliens
w.r.t.\ $P$ pairwise distinct from $\seq[m]u$.
(They are either the constants from condition \ref{lemma:MP:condition2},
or can be taken from the infinite set of ground aliens w.r.t.\ $P$ with the main
symbol from condition \ref{lemma:MP:condition1}.)
By Lemma \ref{lemma:distinct},
there exists a ground instance $(\seq Y,\seq[m]u)\sigma$,
consisting of $n+m$ distinct aliens w.r.t.\ $P$,
where the domain of $\sigma$ is $\{\seq Y\}$.
Note that
$\varphi\sigma=\sigma \cup \{\, X_1/u_1\sigma,\ldots, X_m/u_m\sigma\,\} $.
The substitution
maps variables $\seq Y,\seq[m]X$ to distinct aliens
$(\seq Y,\seq[m]u)\sigma$ w.r.t.\ $P$.
So $Q'$ is $Q'\varphi\sigma$ generalized for $P$.
Thus $P,\ Q'\varphi\sigma$ and $Q'$ satisfy the conditions of
Corollary~\ref{cor:generalized}.
Now ${\ensuremath{\cal M}}\xspace_P\models Q$ implies ${\ensuremath{\cal M}}\xspace_P\models Q\sigma$ and then
$P\models Q\sigma$
(as equivalence (\ref{eq}) from Introduction holds for ground queries).
As $Q\sigma=Q'\varphi\sigma$, by
Corollary \ref{cor:generalized} $P\models Q'$.
The ``if'' case is obvious, as $Q$ is an instance of $Q'$.
\end{proof}
\begin{remark}
The premises of Lemma \ref{lemma:MP} can be weakened by
stating that $Q,Q'$ are atoms such that $Q=Q'\varphi$ for a substitution
$\varphi=\{\, X_1/u_1,\ldots X_m/u_m\,\}$,
where $\seq[m]u$ are distinct aliens w.r.t.\ $P\cup\{Q'\}$, and variables
$\seq[m]X$ do not occur in $Q$.
\end{remark}
\begin{proof}
Obtained by minor modifications of the proof above.
The first sentence, describing $\varphi$, is to be dropped.
Each ``w.r.t.\ $P$'' is to be changed to ``w.r.t.\ $P\cup\{Q'\}$''$\!$.
In the third paragraph, substitution $\varphi\sigma$ together with $P$ and
$Q'$ satisfy the condition of Lemma \ref{lemma:alien:substitution}.
At the end of the proof,
Lemma \ref{lemma:alien:substitution} should be applied instead of
Corollary \ref{cor:generalized}.
\end{proof}
It remains to generalize Lemma \ref{lemma:MP} to arbitrary queries.
\begin{corollary}
\label{cor:MP}
Lemma \ref{lemma:MP} also holds for non-atomic queries. Moreover, condition
\ref{lemma:MP:condition2} of the lemma can be replaced by:
\begin{enumerate}[(a)]
\setcounter{enumi}2
\item
\label{cor:MP:condition3}
for each atom $A$ of $Q$ with $k\geq0$ (distinct) variables,
the underlying language has (at least) $k$ constants not occurring in $P,A$.
\end{enumerate}
\end{corollary}
\begin{proof}
Note that condition \ref{lemma:MP:condition2} implies condition
\ref{cor:MP:condition3}.
So assume that the latter holds.
Let $Q=\seq[l]A$ generalized for $P$ be $Q'=\seq[l]{A'}$.
Then each $A_i'$ is $A_i$ generalized for $P$.
So Lemma \ref{lemma:MP} applies to each $A_i,A_i'$. Thus ${\ensuremath{\cal M}}\xspace_P\models Q$ implies
$P\models A_i'$, for each $i=1,\ldots,l$.
Hence $P\models Q'$.
\end{proof}
\section{Characterization of program answers by the least Herbrand model}
\label{sec:H}
This section studies when the least Herbrand models exactly characterize the
program answers.
First a sufficient condition is presented for equivalence (\ref{eq}) from
Introduction.
Then we show that the sufficient condition is also necessary.
Conditions \ref{th:MP:condition1}, \ref{th:MP:condition2} below are the same
as conditions \ref{lemma:MP:condition1}, \ref{cor:MP:condition3}
of Lemma \ref{lemma:MP} and Corollary \ref{cor:MP}.
\begin{theorem}
[Characterizing answers by ${\ensuremath{\cal M}}\xspace_P$]
\label{th:MP}
Let $P$ be a program, and $Q$ a query such that
\begin{enumerate}[(a)]
\item
\label{th:MP:condition1}
the underlying language has a non-constant function symbol not occurring in~P,
or
\item
\label{th:MP:condition2}
for each atom $A$ of $Q$ with $k\geq0$ (distinct) variables,
the underlying language has (at least) $k$ constants not occurring in $P,A$.
\nopagebreak
\end{enumerate}
\nopagebreak
Then ${\ensuremath{\cal M}}\xspace_P\models Q$ iff $P\models Q$.
\end{theorem}
Note that condition \ref{th:MP:condition1}
implies that the equivalence holds for every query $Q$,
including queries containing the new symbol.
Also, it holds for every query $Q$ and every finite program $P$ when the
alphabet contains infinitely many function symbols, as then
condition \ref{th:MP:condition1} or \ref{th:MP:condition2} is satisfied.
From the theorem the known sufficient conditions follow:
the alphabet containing infinitely many constants (and $P$ finite),
or $Q$ ground.
Condition \ref{th:MP:condition2}
is implied by its simpler version:
the language has $k\geq0$ constants not occurring in $P,Q$,
and each atom of $Q$ contains no more than $k$ variables.
\begin{proof}[Proof of Th.\,\ref{th:MP}]
Let $Q'$ be $Q$ generalized for $P$.
By Corollary \ref{cor:MP},
${\ensuremath{\cal M}}\xspace_P\models Q$ implies $P\models Q'$, hence $P\models Q$, as $Q$ is an
instance of $Q'$.
The reverse implication is obvious.
\end{proof}
We conclude with showing in which sense
the sufficient condition of Th.\,\ref{th:MP}
is also necessary. As expected, it is strictly speaking not a necessary
condition for (\ref{eq}),
as it is violated for some $P,Q$ for which (\ref{eq}) holds.
\begin{example}
\noindent
Consider program APPEND and assume that the only function symbols of the
underlying language are $[\,]$,
$[\ |\ ]$.
Let $Q=app( [X], [Y], [X,Y] )$.
Then ${\ensuremath{\cal M}}\xspace_{\rm APPEND}\models Q$ and ${\rm APPEND}\models Q$,
but the condition of Th.\,\ref{th:MP} is violated.
On the other hand, consider a program $P$ of three clauses
$ app(\,[\,],L,L\,)$.\,;
{\small
$app(\,[[\,]|K],L,[[\,]|M]\,) \gets \linebreak[3] app(\,K,L,M\,)$.\,;
\hspace{0pt plus .5ex}%
$app(\,[[H|T]|K],L,[[H|T]|M]\,) \gets \linebreak[3] app(\,K,L,M\,)$.
}
Programs \mbox{APPEND} and $P$ have the same least Herbrand model but
different sets of answers,
as e.g.\ $P\mathrel{\,\not\!\models} Q$.
The condition of Th.\,\ref{th:MP} is violated by $P,Q$, and the equivalence
does not hold.
Note that $P$ cannot be used to append lists when new function symbols are
added to the language; $app([a],[b],[a,b])$ is then not an answer of~$P$.
\sloppy
\end{example}
Roughly speaking,
the sufficient conditions of Th.\,\ref{th:MP} and Lemma \ref{lemma:MP}
are also necessary,
when all what is known about a program is the set of function
symbols employed in it:
\begin{proposition}
\label{prop:counterexample}
Let ${\ensuremath{\mathscr F}}\xspace$ be the set of function symbols of the underlying language, and
$F_0\subseteq {\ensuremath{\mathscr F}}\xspace$ be its finite subset.
Let $Q$ be a query,
such that the predicate symbols of the atoms of $Q$ are distinct.
Assume that ${\ensuremath{\cal M}}\xspace_P\models Q$ iff $P\models Q$, for each finite program $P$
such that
$F_0$ is the set of function symbols occurring in $P$. Then the
sufficient condition of Th.\,\ref{th:MP} holds.
\end{proposition}
The proposition also holds when $F_0$ and the considered program $P$ are
infinite.
\begin{proof}
Let $Q$ be a query whose atoms have distinct predicate symbols.
Assume that the sufficient condition of Th.\,\ref{th:MP} does not hold.
We show that for a certain program $P$
(such that $F_0$ is the set of the function symbols occurring in $P$),
\linebreak[3]
\mbox{${\ensuremath{\cal M}}\xspace_P\models Q$} but $P\mathrel{\,\not\!\models} Q$.
As condition \ref{th:MP:condition1} of Th.\,\ref{th:MP} does not hold,
all the non-constant function symbols of ${\ensuremath{\mathscr F}}\xspace$ are in $F_0$.
As condition \ref{th:MP:condition2} does not hold,
there is an atom $A$ in $Q$ with $k$ distinct variables $\seq[k]Y$,
for which the number of constants from ${\ensuremath{\mathscr F}}\xspace\setminus F_0$ not occurring in
$A$ is $l<k$\/; let $\seq[l]a$ be the constants.
The atom can be represented as $A = B[\seq[n]b,\seq[k]Y]$,
where
$\seq[n]b$ are those (distinct) constants of $A$ which are not in $F_0$.%
\footnote{%
Formally, $B[\seq[n+k]t]$ can be defined as the instance
$B\{V_1/t_1,\ldots V_{n+k}/t_{n+k}\}$ of an atom $B$, whose
(distinct) variables are $\seq[n+k]V$, and
whose function symbols are from $F_0$.
}
So ${\ensuremath{\mathscr F}}\xspace\setminus F_0 = \{ \seq[l]a,\seq[n]b \}$.
Let $P_0$ be the set of atoms of $Q$ except for $A$.
Let ${\cal V} = \{\seq[n+k-1]X\}$ be $n+k-1$ distinct variables.
Let $P$ consist of the unary clauses of $P_0$ and the
unary clauses of the form $B[\seq[n+k]t]$ where
(i)~$\{\seq[n+k]t\} =\cal V $
(so a variable occurs twice),
or (ii)~$\{\seq[n+k]t\} ={\cal V} \cup \{f(\vec Z)\} $
where $f\in F_0$, its arity is $m\geq0$,
and $\vec Z$ is a
tuple of $m$ distinct variables pairwise distinct from those in $\cal V$.
Note that $P$ is finite iff $F_0$ is.
Each ground atom $B'=B[\seq[n+k]u]$ (where $\seq[n+k]u\in{\ensuremath{\cal H U}}\xspace$)
is an instance of some clause of $P$,
as if $\seq[n+k]u$ are distinct terms then the main symbol of some of
them is in $F_0$, and $B'$ is an instance of a clause of the form (ii),
otherwise $B$ is an instance of a clause of the form (i).
Thus ${\ensuremath{\cal M}}\xspace_P\models A$, hence ${\ensuremath{\cal M}}\xspace_P\models Q$
(as $P_0\models A'$ for each atom $A'$ of $Q$ distinct from $A$).
To show that $P\mathrel{\,\not\!\models} Q$,
add new constants $a_{l+1},\ldots,a_k$ to the alphabet.
Then $B[\seq[n]b,\seq[k]a]$ is not an instance of any clause of $P$,
so $B[\seq[n]b,\seq[k]a]$ is false in the least Herbrand model of $P$
with the extended alphabet.
\end{proof}
%
The proof provides a family of counterexamples for a claim that
${\ensuremath{\cal M}}\xspace_P\models Q$ and $P\models Q$ are equivalent.
In particular, setting $Q=p(V)$ ($k=1$, $n=0$) results in
$P = \{\, p(f(\vec Z)) \mid f\in F \,\}$,
a generalization
of the counterexample from Introduction to any underlying
finite set $F$ of function symbols.
From the proposition it follows that a more general sufficient condition
(than that of Th.\,\ref{th:MP}) for
the equivalence of ${\ensuremath{\cal M}}\xspace_P\models Q$ and $P\models Q$ is impossible,
unless it uses more information about $P$ than just the set of involved symbols.
\section{Conclusion}
In some cases the least Herbrand model does not characterize the set of
answers of a definite program.
This paper generalizes the sufficient condition for
${\ensuremath{\cal M}}\xspace_P\models Q$ iff $P\models Q$, to
``a non-constant function symbol not in $P$, or $k$ constants not in $P,A$
for each atom $A$ of $Q$''.
It also shows
that the sufficient condition cannot be improved unless more is
known about the program than just which function symbols occur in it.
As a side effect, it is shown
which more general queries are implied to be answers of $P$ by $Q$ being an
answer.
\paragraph{Acknowledgement.}
Comments of anonymous referees helped improving the presentation.
\bibliographystyle{acmtrans}
|
\section{Taxi Price Comparison Experiment}
\paragraph*{\textbf{The New York City Taxi Dataset.}}
The Freedom of Information Law in United States encourages public authorities
to release their data where appropriate to the benefit of the citizens. In 2014
the law was exploited by Chris Whong to acquire and post on the web one of the most
comprehensive taxi mobility datasets available today. The dataset describes
taxi journeys in New York City during the full course of 2013, and informs us not only on the
origin and destination points of taxi trips, noted in the related jargon as pick up and
drop off points respectively, but also on the financial costs incurred to the customer (trip fair)
with unprecedented detail. This rather dense mobility dataset, containing hundreds of millions
of trips is of gigabytes in size and can be downloaded here~\url{http://chriswhong.com/open-data/foil_nyc_taxi/}. A sample of the traces generated
by the data can be seen is in Figure~\ref{taxitraces}, where we have drawn a black point for every pick up
and drop off point of a taxi journey.
\begin{figure}
\centering
\includegraphics[width=0.7\linewidth]{taxitraces.png}
\caption{Marking the traces of new york city yellow taxis. For every pick up and drop off point in a uniform sample of the data we draw a black point.}
\label{taxitraces}
\end{figure}
\paragraph*{\textbf{Comparing Taxi Prices}}
In August 2014, Uber opened up an API with access to valuable information about its services.
The occasion allowed us to perform a first head to head comparative analysis of prices between Uber and Yellow taxis
in New York City. To achieve this we run the following experiment :
\begin{itemize}
\item 1. For every trip in the New York City Yellow Taxi dataset, record the geographic coordinates (latitude and longitude) of the pick up and drop off points.
\item 2. Retrieve the total fare paid by the customer for the trip (including the tip).
\item 3. Query Uber's API and ask how much they would charge for the same trip (same pick up and drop off points),
considering the cheapest version of the service, Uber X.
\item 4. Uber's API returns a value range indicating the minimum and maximum price estimate. We take the mean
of the two values.
\item 5. We then compare the prices from the two services.
\end{itemize}
As can be observed in Figure~\ref{pricecomp} where the distribution of prices for the two services is shown, despite the qualitative similarity of the two distribution, yellow taxi appear on average (median) 1.4 US dollars cheaper than Uber X. In Figure~\ref{ubervsyellow}, we compare Uber and yellow taxis from another perspective: for every observed yellow taxi price, we show the median Uber X price. Uber appears more expensive for prices below 35 dollars and begins to become cheaper only after that threshold. As one would expect, the cheaper journeys are those that are in principle of shorter range. As observed in a variety of empirical data, human mobility tends to be characterised by a vast majority of short trips~\cite{gonzalez2008understanding, brockmann2006scaling}. This observation therefore suggests that Uber's economical model exploits this trend of human mobility in order to maximise revenue. We also confirm the skewed frequency distribution of movement distances in the present context by visualising it in Figure~\ref{tripdistances}, where we note a mean distance for a yellow taxi trip in New York equal to $2.09$.
The above experiment may involve a number of biases which we refer to here.
The NYC Yellow taxi data corresponded to year 2013 whereas Uber to 2014. Although note that the prices for yellow taxis in the city had last changed in 2012 after 8 years~\cite{fares}. So it should offer a good approximation of today’s prices. Further, there was no control for time of the day/week for the API query, an additional dimension which should be incorporated when available. However, we argue that the process of comparing two different companies that provide the same service in the same geographic area is of value to commuters. Just as consumer have open access to airfares for a long time now allowing for transparency in a free, competitive, market we believe that similar approaches could benefit commuters in modern cities.
\begin{figure}
\centering
\includegraphics[width=0.7\linewidth]{uberVsTaxiPrices.pdf}
\caption{Distribution of prices per journey for Uber X and Yellow Taxis in New York City.}
\label{pricecomp}
\end{figure}
\begin{figure}
\centering
\includegraphics[width=0.7\linewidth]{taxiPriceVsUberPrice.pdf}
\caption{Median Uber price for a given Yellow Taxi price.}
\label{ubervsyellow}
\end{figure}
\begin{figure}
\centering
\includegraphics[width=0.7\linewidth]{tripdistances.pdf}
\caption{Distribution of geographic distances between drop off and pick up points for Yellow Taxi journeys.}
\label{tripdistances}
\end{figure}
\begin{figure}
\centering
\includegraphics[width=0.7\linewidth]{TaxiPriceGeoComparison.pdf}
\caption{Geographic comparison between Uber and Yellow Taxi prices. We paint an area black if Uber is cheaper by trip majority and yellow otherwise.}
\label{ios1}
\end{figure}
\section {Helping commuters}
Our observations show that it might be financially advantageous on average for travellers to chose either Yellow Cabs or Uber depending on the duration of their journey. However the specific journey they are willing to take matters.
In order to help users to take the right decision, we have developed a smartphone app, called OpenStreetCab, designed as follows.
One limitation for the design of our service is that only prices for trips with origins and destinations
in the New York City Taxi Dataset can in principle be retrieved. In order to evaluate the price of any trip, as needed for a usable App, we have divided the NY region into a mesh with cells of size around 100m by 100m in order to index trips in the database efficiently. For each user query, we find a set of trips in our dataset with the origin in neighbouring cells of desired origin and, among them, we find the trip whose destination is closest to the desired one. This strategy has the advantage of being sufficiently fast to perform online queries and expected to provide reliable price estimates. For the same trip, Uber price is obtained through their API.
\begin{figure}
\centering
\includegraphics[width=0.7\linewidth]{ios1.png}
\caption{The proof of the concept}
\label{ios1}
\end{figure}
A real-time prototype has been designed and is currently launched on popular mobile platforms. Future improvements include the possibility to change predictions depending on the time of the day, or on the expected traffic on the way, but also to suggest other types of transportations, such as walking when the distance is sufficiently short, or only part of the way, in situations when a small change in the origin point can lead to a significant change in the price quote. In the meanwhile the current version (Fig. \ \ref{ios1}) already provides a fully working solution, including geolocation services and address retrieval. We are planning to launch the application on the related stores very soon.
\small
\bibliographystyle{plain}
|
\section{Introduction}
Given an ideal $\ag$ of the maximal order $\mathcal{O}_K$ of $K = \Q(\zeta_{2^s})$, we want to decide if
$\ag$ is principal, and if so, compute $\alpha\in\mathcal{O}_K$ such that $\ag = (\alpha)\mathcal{O}_K$. This corresponds
to the Principal Ideal Problem (PIP), which is a fundamental problem in computational number theory.
The resolution of the PIP in classes of number fields of large degree
recently received a growing attention due to its connection with cryptosystems based on
the hardness of finding short generators of principal ideals such as the homomorphic encryption
scheme of Vercauteren and Smart~\cite{fre_smart}, and the multilinear maps of Garg, Gentry and Halevi~\cite{mult_maps}.
A generator of a principal ideal of the maximal order $\mathcal{O}_K$ of $K = \Q(\zeta_{p^s})$ can be found in
heuristic subexponential time $L_\Delta(2/3+\varepsilon,c)$ for some $c>0$ where
$$L_{\Delta}(a,b) = e^{(b+o(1))(\ln|\Delta|)^a(\ln\ln|\Delta)^{1-a} }$$
by using an algorithm of Biasse and Fieker~\cite{biasse_subexp,ANTS_XI}. We can also
find one in quantum polynomial time with an algorithm of Biasse and Song~\cite{SODA16} which
relies on the hidden subgroup resolution algorithm for $\R^{O(n)}$ of Eisentr\"{a}ger, Hallgren, Kitaev and Song~\cite{STOC2014}.
To go from an arbitrary generator $\alpha$ to a small one, we need to multiply $\alpha$ by the right unit.
This is an instance of the Bounded Distance Decoding problem in the lattice of the logarithms of the
complex embeddings of the elements of $K$. Campbel, Groves and Shepherd~\cite{GCHQ} observed that
using an LLL reduction of the basis of this lattice consisting of the so-called cyclotomic units~\cite[Chap. 8]{Washington}
and performing a simple round-off reduction yielded the solution to the problem. This fact was corroborated by Schank
in a replication study~\cite{Pari_implementation}. Shortly thereafter, Cramer, Ducas, Peikert and Regev~\cite{CVP_cyclo}
proved that this fact was due to the intrinsic geometric properties of the cyclotomic units and that the LLL reduction
was not necessary. The draft of Campbel et al. also contained elements on ongoing work on a quantum algorithm
for solving the PIP which was interrupted when they conjectured that the methods of Eisentr\"{a}ger et al.~\cite{STOC2014}
would ultimately yield a quantum polynomial time algorithm for solving the PIP.
These recent developments raised the question of the hardness of the Shortest Vector Problem (SVP) in ideal
lattices (and principal ideal lattices in particular). According to Cramer et al.~\cite[Sec. 6]{CVP_cyclo},
a short generator of a principal ideal in a cyclotomic ring is at least within a factor $e^{\tilde{O}(\sqrt{n})}$
of a shortest element. To the best of our knowledge, there is currently no method that can leverage the knowledge of
a short generator to derive an element with length within a better approximation factor to the first minima. However,
a short generator of a principal ideal is also a solution to $\gamma$-SVP for $\gamma\in e^{\tilde{O}(\sqrt{n})}$.
In addition, Cramer, Ducas and Wesolowski~\cite{Stickelberger} showed that given an ideal $I$ in a cyclotomic ring
and under reasonable assumptions on the ideal class group,
there is a heuristic method relying on the Discrete Logarithm Problem (DLP) in $\operatorname{Cl}(\mathcal{O}_K)$ to find an ideal $J\subseteq \mathcal{O}_K$
with $\mathcal{N}(J)\in e^{O(n^{3/2})}$ such that $IJ$ is principal. Then a short generator of the principal ideal $IJ\subseteq I$
yields a solution to $\gamma$-SVP in $I$ for $\gamma\in e^{\tilde{O}(\sqrt{n})}$.
\paragraph{\textbf{Contributions}} We describe an attack running in heuristic complexity $2^{O(n^{1/2+\varepsilon})}$ for
arbitrarily small $\varepsilon>0$
against cryptographic schemes relying on the hardness of finding a short generator of an
ideal in $\Q(\zeta_{p^s})$, including the homomorphic encryption
scheme of Smart and Vercauteren~\cite{fre_smart}, and the multilinear maps of Garg, Gentry and Halevi~\cite{mult_maps}.
We rely
on a different ideal class group computation and PIP algorithms than the subsexponential methods of Biasse and Fieker~\cite{ANTS_XI} which
run in time $2^{O(n^{2/3 + \varepsilon})}$
for arbitrarily small $\varepsilon > 0$. We take
advantage of the small height of the defining polynomial of fields of the form $\Q(\zeta_{p^s})$, and we use a modified
$\qg$-descent method to solve the PIP.
We also describe practical improvements to our algorithm for computing the ideal class group and solving the PIP that
do not improve the theoretical complexity, but that have a significant impact on the performances of our methods.
Finally, we present a PIP resolution method that leverages a precomputation on $\Q(\zeta_{p^s})$ of a higher cost
than $2^{O(n^{1/2+\varepsilon})}$. Using this precomputation, we achieve a better heuristic complexity than $2^{O(n^{1/2})}$ when
solving all subsequent PIP instances for which $\mathcal{N}(I)\leq 2^{n^b}$ in $\Q(\zeta_{p^s})$.
More specifically, by spending a precomputation time in $2^{O(n^{2-3a+\varepsilon})}$ for an arbitrarily small $\varepsilon>0$, we can solve the
PIP on input ideals $I$ with $\mathcal{N}(I)\leq 2^{n^b}$ in time $2^{O\left(n^{a + o(1)}\right)}$ when
\begin{enumerate}
\item $b\leq 7a - 2$.
\item $\frac{2}{5} < a < \frac{1}{2}$.
\end{enumerate}
Combined
with the recent heuristic reduction from $\gamma$-SVP to the short PIP of Cramer et al.~\cite{Stickelberger} (under the same assumptions
on the class group), this yields a
heuristic algorithm for $\gamma$-SVP in ideals $I$ of $\Q(\zeta_{p^s})$ satisfying $\mathcal{N}(I)\leq 2^{n^b}$ with precomputation on $\Q(\zeta_{p^s})$
for $\gamma\in e^{\tilde{O}(\sqrt{n})}$ with a better trade-off approximation/cost than BKZ. For
example:
\begin{itemize}
\item With a precomputation of cost $2^{O(n^{5/7 + \varepsilon})}$ on $\Q(\zeta_{p^s})$, the instances of $\gamma$-SVP in ideals $I$ of
$\Q(\zeta_{p^s})$ satisfying $\mathcal{N}(I)\leq 2^{n+o(1)}$ can be solve in time $2^{O(n^{3/7+o(1)})}$.
\item Given a $2^{O(n^{5/7 + \varepsilon})}$
precomputation,
the private keys of the multilinear maps of Garg, Gentry and Halevi~\cite{mult_maps} can be retrieved
in time $2^{O(n^{3/7+o(1)})}$ from the corresponding public keys.
\end{itemize}
\section{Background}
\paragraph{\textbf{Lattices}} A lattice is a discrete additive subgroup of $\R^n$ for some integer $n$. The first
minima of a lattice $\mathcal{L}$ is defined by $\lambda_1 := \min_{\vec{v}\in\mathcal{L}\setminus \{0\}}\|\vec{v}\|$.
A basis of $\mathcal{L}$ is a set of linearly independent vectors $\vec{b}_1,\cdots,\vec{b}_k$ such that
$\mathcal{L} = \Z \vec{b}_1 + \cdots + \Z \vec{b}_k$. The determinant of $\mathcal{L}$ is $\det(\mathcal{L}) = \sqrt{\det(B\cdot B^T)}$ where
$B = (\vec{b}_i)_{i\leq k}\in\R^{k\times n}$ is the matrix of a basis of $\mathcal{L}$. For a full dimensional lattice
$\mathcal{L}$, the best upper bound we know on $\lambda_1(\mathcal{L})$ is in $O\left( \sqrt{n}\det(\mathcal{L})^{1/n}\right)$.
The problem of finding a shortest vector $v\in\mathcal{L}$ is known as the Shortest Vector Problem (SVP), while
the problem of finding $v\in\mathcal{L}$ such that $\|\vec{v}\|\leq \gamma \lambda_1(\mathcal{L})$ for
some $\gamma\geq 1$ is known has $\gamma$-SVP. A solution $\vec{v}$ to $\gamma$-SVP satisfies $\|\vec{v}\|\in O\left( \gamma\sqrt{n}\det(\mathcal{L})^{1/n}\right)$.
Given the matrix of a basis $A$ as
input, the LLL algorithm~\cite{LLL} returns a basis $(\vec{b}_i)_{i\leq n}$ such that $\frac{\|\vec{b}_1\|}{\det(\mathcal{L})^{1/n}}\in 2^{O(n)}$
in polynomial time in $n$ and $\log(|A|)$. The BKZ algorithm~\cite{Ajtai_BKZ} with block size $k$ returns
a basis $(\vec{b}_i)_{i\leq n}$ such that $\frac{\|\vec{b}_1\|}{\det(\mathcal{L})^{1/n}}\in O(k^{n/k})$ in
time $2^{O(k)}\mathrm{Poly}(n,\log(|A|)$. Finally, the HKZ algorithm returns a basis $(\vec{b}_i)_{i\leq n}$ such
that $\frac{\|\vec{b}_1\|}{\det(\mathcal{L})^{1/n}}\in O(\sqrt{n})$
in time $2^{O(n)}$. The problem of finding $n$ linearly independent vectors $(\vec{v}_i)_{i\leq n}$ such that
$\max_i\|\vec{v}_i\|\leq \max_{(\vec{b}_i)_{i\leq n} \mbox{ basis of }\mathcal{L}}\max_i\| \vec{b}_i\|$ is the Shortest Independent Vectors
Problem (SIVP), and $\sqrt{n}\gamma$-SVIP efficiently reduces to $\gamma$-SVP.
\paragraph{\textbf{Number fields}} A number field $K$ is a finite extension of $\Q$. Its ring of integers $\mathcal{O}_K$ has the structure
of a $\Z$-lattice of degree $n=[K:\Q]$, and the orders $\mathcal{O}\subseteq\mathcal{O}_K$ are the sublattices of $\mathcal{O}_K$
which have degree $n$ and which are equipped with a ring structure. A number field
has $r_1\leq n$ real embeddings $(\sigma_i)_{i\leq r_1}$ and $2r_2$ complex
embeddings $(\sigma_i)_{r_1 < i \leq 2r_2}$ (coming as $r_2$ pairs of conjugates). The field $K$ is isomorphic to
$\mathcal{O}_K\otimes\Q$ where $\mathcal{O}_K$ denotes the ring of integers of $K$. We can embed $K$ in
$K_\R := K\otimes \R \simeq \R^{r_1}\times \C^{r_2}, $
and extend the $\sigma_i$'s to $K_\R$. Let $T_2$ be the Hermitian form on $K_\R$ defined by
$T_2(x,x') := \sum_i \sigma_i(x)\overline{\sigma_i}(x')$,
and let $\| x\| := \sqrt{T_2(x,x)}$ be the corresponding $L_2$-norm. The norm of an element $x\in K$ is defined by $\mathcal{N}(x) = \prod_i\sigma_i(x)$.
Let $(\alpha_i)_{i\leq d}$ such that
$\mathcal{O}_K = \oplus_i \Z\alpha_i$, then the discriminant of $K$ is given by $\Delta = \det^2(T_2(\alpha_i,\alpha_j))$.
The volume of the fundamental domain is $\sqrt{|\Delta|}$, and the size of the input of algorithms
working on an integral basis of $\mathcal{O}_K$ is in $O(\log(|\Delta|))$. In $K = \Q(\zeta_{p^s})$, the degree satisfies
$[K:\Q] = (p-1)p^{s-1}$ and $\Delta = \pm p^{p^{s-1}(ps - s - 1)}$, therefore $\log(|\Delta|)\sim n\log(n)$
and we can express the complexity of our algorithms in terms of $n$ (a choice we made in this paper),
which makes it easier to compare with other lattice reduction result. However, most of the literature on class group
computation presents complexities in terms of $\log(|\Delta|)$, which is in general the right value to measure the
input. For example, it makes no sense to express the complexity with respect to the degree of $K$ in infinite
classes of quadratic number number fields.
\paragraph{\textbf{Cyclotomic fields}} A cyclotomic field is an extension of $\Q$ of the form $K = \Q(\zeta_N)$ where $\zeta_N = e^{2i\pi/N}$ is
a primitive $N$-th root of unity. The ring of integers $\mathcal{O}_K$ of $K$ is
$\Z[X]/(\Phi_N(X))$ where $\Phi_N$ is the $N$-th cyclotomic polynomial. When $N$ is a power of two,
$\Phi_N(X) = X^{N/2} + 1$, and when $N = p^s$ is a power of $p>2$, we have
$\Phi_N(X) = X^{p^{s-1}(p-1)} + X^{p^{s-1}(p-2)} + \cdots + 1$ (which of course generalizes the case $p = 2$).
Elements $a\in \mathcal{O}_K$ are residues of polynomials in $\Z[X]$ modulo $\Phi_N(X)$, and can be identified
with their coefficient vectors $\vec{a}\in\Z^{\phi(N)}$ where $\phi(N) = p^{s-1}(p-1)$ is the
Euler totient of $N$ (and the degree of $\Phi_N(X)$).
\paragraph{\textbf{The ideal class group}} Elements of the form $\frac{I}{d}$ where $I\subseteq \mathcal{O}_K$ is an (integral) ideal of
the ring of integers of $K$ and $d > 0$ are called fractional ideals. Like the orders of $K$, they have the
structure of a $\Z$-lattice of degree $n=[K;\Q]$, and they form a multiplicative group $\mathcal{I}$. Elements of $\mathcal{I}$
admit a unique decomposition as a power product of prime ideals of $\mathcal{O}_K$ (with
possibly negative exponents). The norm of
integral ideals is given by $\mathcal{N}(I) := [\mathcal{O}_K:I]$, which extends to fractional
ideals by $\mathcal{N}(I/J) := \mathcal{N}(I)/\mathcal{N}(J)$. The norm of a principal (fractional) ideal agrees with the norm
of its generator $\mathcal{N}(x\mathcal{O}_K) = |\mathcal{N}(x)|$. The principal fractional ideals $\mathcal{P}$ of $K$ are a subgroup of $\mathcal{P}$
and ideal class group of $\mathcal{O}_K$ is defined by
$\operatorname{Cl}(\mathcal{O}_K) := \mathcal{I}/\mathcal{P}.$ We denote by $[\ag]$ the class of a fractional $\ag$ in $\operatorname{Cl}(\mathcal{O}_K)$ and by $h$ the
cardinality of $\operatorname{Cl}(\mathcal{O}_K)$ which is a finite group. In $\operatorname{Cl}(\mathcal{O}_K)$ we identify two fractional ideals $\ag,\bg$ if there is $\alpha\in K$
such that $\ag = (\alpha)\bg$. This is denoted by $\ag\sim \bg$.
\paragraph{\textbf{Units of $\mathcal{O}_K$}} Elements $u\in\mathcal{O}_K$ that are invertible in $\mathcal{O}_K$ are called units. Equivalently, they
are the elements $u\in\mathcal{O}_K$ such that $(u)\mathcal{O}_K = \mathcal{O}_K$ and also such that $\mathcal{N}(u) = \pm 1$. The unit group of $\mathcal{O}_K$
where $K$ is a cyclotomic field has rank $r = n/2-1$ and has the form
$\mathcal{O}_K^* = \mu\times \langle\epsilon_1\rangle\times \cdots\times \langle\epsilon_r\rangle$ where $\mu$ are roots of
unity (torsion units) and the $\epsilon_i$ are non-torsion units. Such $(\epsilon_i)_{i\leq r}$ are called a system
of fundamental units of $\mathcal{O}_K$. Units generate a lattice $\mathcal{L}$ of rank $r$ in $\R^{r+1}$ via the embedding
$$x\in K\longmapsto \operatorname{Log}(x) := \left( \ln(|\sigma_1(x)|) , \cdots, \ln(|\sigma_{r+1}(x)|)\right)$$
where the complex embeddings $(\sigma_i)_{i\leq n}$ are ordered such that the first $r = n/2$ ones are not
conjugates of each other. The volume $R$ of $\mathcal{L}$ is an invariant of $K$ called the regulator.
The regulator $R$ and the class number $h$ satisfy
$hR = \frac{|\mu|\sqrt{|\Delta|}}{2^{r_1}(2\pi)^{r_2}}\lim_{s\rightarrow 1} \left( (s-1)\zeta_\K (s)\right),$
where $\zeta_\K (s) = \sum_\ag \frac{1}{\mathcal{N}(\ag)^s}$ is the usual $\zeta$-function associated to $K$ and
$|\mu|$ is the cardinality of $\mu$ the group of torsion units. This
allows us to derive a bound $h^*$ in polynomial time under GRH that satisfies $h^* \leq hR < 2h^*$ (\cite{BachEulerProd}).
When $K = \Q(\zeta{p^s})$, logarithm vectors of units of the form $u_j = \frac{\zeta_{p^s}^j - 1}{\zeta_{p^s} - 1}$ for $j\in\Z_{p^s}^*$
together with $\mu$ generate a sublattice of $\mathcal{L}$ of index $h^+(p^s)$ where $h^+(N)$ is the class number of the
maximal real subfield of $\Q(\zeta_N)$~\cite[Lemma 8.1]{Washington}. It is conjectured
that $h^+(2^s)=1$ (Weber class number problem) and that
$h^+(p^s)$ remains bounded for fixed $p$ and increasing $s$ (see~\cite{Pomerance_conjecture}).
\paragraph{\textbf{Notations}} Throughout this paper, $\|A\| = \max_{i,j}|A_{i,j}|$ denotes the infinite norm of a matrix.
We denote by $\ln(x)$ the natural logarithm of $x$ and by $\log(x)$ its base-$2$ logarithm.
\section{High level description of the algorithm}
In cryptosystems based on the hardness of finding a short generator of a principal ideal (in particular~\cite{fre_smart}
and~\cite{mult_maps}), the secret key is a small generator $g\in K = \Q(\zeta_N)$ of a principal ideal $\ag\subseteq\mathcal{O}_K$, and
the public parameters include a $\Z$-basis of $\ag$. We present a subexponential method for retrieving a generator of an
input ideal that is asymptotically faster than the current state of the art~\cite{biasse_subexp,ANTS_XI}. Combined with the
Bounded Distance Decoding method in $\operatorname{Log}(\Z[\zeta_{N}]^*)$ presented in~\cite{CVP_cyclo}, this yield a
better classical attack against the schemes relying the hardness of finding a short generator in a principal ideal.
We also show how to leverage the precomputation of a large set of relations between generators of $\operatorname{Cl}(\mathcal{O}_K)$ to
solve the PIP in $\mathcal{O}_K$ in time $2^{O(n^{a})}$ for $a < 1/2$. Combined with the techniques of~\cite{Stickelberger},
this provides a method for solving $\gamma$-SVP in ideal lattices of $K$ for $\gamma \in e^{\tilde{O}(n^{1/2})}$
with a better trade-off approximation/cost than BKZ.
\paragraph{\textbf{Solving the Principal Ideal Problem}}
All subexponential time algorithms for solving the Principal Ideal Problem follow the same high level strategy. They derive from the
subexponential algorithm for classes of fixed degree number fields~\cite{Buchmann}, which itself is a generalization of the algorithm
of Hafner and McCurley~\cite{hafner} for quadratic number fields. Let $B>0$ be a smoothness bound and
$\mathcal{B}:= \{ \mbox{prime ideals }\pg \mbox{ with } \mathcal{N}(\pg)\leq B\}$. We first need to compute a generating
set of the lattice $\Lambda$ of all the vectors $(e_1,\cdots,e_m)\in\Z^m$ such that
\begin{equation}\label{eq:rel}
\exists \alpha\in K, \ \ (\alpha)\mathcal{O}_K = \pg_1^{e_1}\cdots\pg_m^{e_m},
\end{equation}
where $m = |\mathcal{B}|$. When $B>12\ln^2|\Delta|$, the classes of ideals in $\mathcal{B}$ generate $\operatorname{Cl}(\mathcal{O}_K)$
under the GRH~\cite[Th. 4]{BBounds}.
Therefore, $(\mathcal{B},\Lambda)$ is a presentation of the group $\operatorname{Cl}(\mathcal{O}_K)$ and the search for a generating set of the
relations of the form~\eqref{eq:rel} is equivalent to computing the group structure of $\operatorname{Cl}(\mathcal{O}_K)$. Indeed, the morphism
$$\begin{CD}
\Z^m @>{\varphi}>> \mathcal{I} @>{\pi}>> \operatorname{Cl}(\mathcal{O}_K)\\
(e_1, \ldots, e_m) @>>> \prod_i\p_i^{e_i} @>>> \prod_i[\p_i]^{e_i}
\end{CD},$$
is surjective, and the class group $\operatorname{Cl}(\mathcal{O}_K)$ is isomorphic to $\Z^m/\ker(\pi\circ\varphi) = \Z^m/\Lambda.$
If we only compute a generating set for a sublattice of $\Lambda$ of finite
index, then the cardinality of the tentative class group we obtain is a multiple of $h =|\operatorname{Cl}(\mathcal{O}_K)|$. This can
be tested by using an estimate of $hR$ given by the methods of~\cite{BachEulerProd} in polynomial time under the GRH.
If we are missing relations to generate $\Lambda$, we obtain a multiple of $hR$, and this tells us we need
to collect more relations.
Let $\ag$ be an input ideal. We find an extra relation of the form $\ag = (\alpha)\pg_1^{y_1}\cdots\pg_m^{y_m}.$
The input ideal $\ag$ is principal, if and only if $\pg_1^{y_1}\cdots\pg_m^{y_m}$ is principal too. Then, assuming we have a generating
set for the lattice $\Lambda$ of relations of the form~\eqref{eq:rel}, we can rewrite $\pg_1^{y_1}\cdots\pg_m^{y_m}$ as a power-product
of the relations generating $\Lambda$. More specifically, if the relations $(\alpha_i)\mathcal{O}_K = \pg_1^{e_{i,1}}\cdots\pg_m^{e_{i,m}},\ \ \mbox{for } i\leq m$
generate $\Lambda$, then there is $x\in\Z^m$ such that $xA=y$ for $A = (e_{i,j})_{i,j\leq m}$. Then
$\pg_1^{y_1}\cdots\pg_m^{y_m} = (\alpha_1^{x_1}\cdots\alpha_m^{x_m})\cdot\mathcal{O}_K$, which gives us a generator of $\ag$, $\ag = (\alpha\cdot \alpha_1^{x_1}\cdots\alpha_m^{x_m})\cdot\mathcal{O}_K.$
\begin{algorithm}[ht]
\caption{Principal Ideal Problem}
\begin{algorithmic}[1]\label{alg:PIP}
\REQUIRE Ideal $\ag\subseteq\mathcal{O}_K$, smoothness bound $B$.
\ENSURE $\mathrm{false}$ or $\beta\in K$ such that $\ag = (\beta)\mathcal{O}_K$.
\STATE Compute $\mathcal{B} = \left\lbrace \p_1,\hdots,\p_m\right\rbrace$ of norm less than $B$.
\STATE Compute a lattice $\Lambda\subseteq\Z^m$ of random relations between elements of $\mathcal{B}$.
\STATE $h\leftarrow \left| \Z^m/\Lambda\right|$, $R\leftarrow \mbox{regulator of }K$.
\STATE Check $h\cdot R$ with the estimates of~\cite{ANTS_XI}[Sec. 4.3]. Find more relations if necessary.
\STATE $A \leftarrow (e_{i,j})_{i,j\leq m}$, where $(\alpha_i)=\prod_i\pg_j^{e_{i,j}}$ are a basis of $\Lambda$.
\STATE Find $\vec{y}\in\Z^m$ such that $\ag = (\alpha)\pg_1^{y_1}\cdots\pg_m^{y_m}.$
\STATE Solve $\vec{x}A=\vec{y}$.
\RETURN $\mathrm{false}$ if no solution or $\alpha\cdot\alpha_1^{x_1},\hdots,\alpha_{m}^{x_m}$.
\end{algorithmic}
\end{algorithm}
\begin{theorem}
Algorithm~\ref{alg:PIP} runs in heuristic complexity $2^{O(n^{1/2+\varepsilon})}$ for
arbitrarily small $\varepsilon>0$ when
$B = 2^{n^{1/2}}$ with $K = \Q(\zeta_{p^s})$ and $n = [K:\Q]$.
\end{theorem}
\paragraph{\textbf{PIP with precomputation on $K$}}
Algorithm~\ref{alg:PIP} can be divided into two steps: the computation of a basis of the lattice of
relations between the classes of a generating set for $\operatorname{Cl}(\mathcal{O}_K)$, and the decomposition of the input
ideal $\ag$ with respect to this generating set. The computation of relations between generators of the
ideal class group is independent of the specific instances of the PIP and can be precomputed. Moreover,
the larger the smoothness bound $B$ is, the faster the decomposition of an input ideal over $\mathcal{B}$ is.
Naturally, this implies that the precomputation is more expensive since it requires the calculation of more
relations.
\begin{theorem}
If we precompute all relations between ideals of norm less than $B \in 2^{O(n^a)}$ for some $1/2 < a < 1$,
then the cost of solving the PIP is in $2^{O\left(n^{\frac{2}{3} - \frac{a}{3} + \varepsilon}\right)}$ for arbitrarily small $\varepsilon > 0$
while the cost of the precomputation is in $2^{\tilde{O}(n^a)}$.
\end{theorem}
For example, for $a = 2/3$, a precomputation of all relations between prime ideals of norm less than $B \in 2^{O(n^{2/3})}$
can be done in time $2^{\tilde{O}(n^a)}$. It allows us to solve all instances of the PIP in $\mathcal{O}_K$ in time
$2^{O(n^{4/9 + \varepsilon})}$ for arbitrarily small $\varepsilon > 0$
\paragraph{\textbf{Reducing the short-PIP and $\gamma$-SVP to the PIP}}
Assume that the input ideal $\ag\subseteq\mathcal{O}_K$ is generated by a short element $g$, and that we have computed $\alpha\in\mathcal{O}_K$
such that $\ag = (\alpha)\cdot\mathcal{O}_K$.
Given a generating set $\gamma_1,\cdots,\gamma_r$ of the unit group $\mathcal{O}_K^*$, all
generators $g'$ of $\ag$ are of the form
$$g' = \alpha\cdot \gamma_1^{x_1}\cdots\gamma_r^{x_r}\ \ \mbox{for some } (x_1,\cdots,x_r)\in\Z^r.$$
The problem of finding $g$ (or another short generator, which is equivalent for the sake of a cryptanalysis),
boils down to finding $(x_1,\cdots,x_r)$ such that $\alpha\cdot \gamma_1^{x_1}\cdots\gamma_r^{x_r}$ is short. In the
lattice of logarithm embeddings, this can be done by finding $(x_1,\cdots,x_r)$ such that
$\| \operatorname{Log}(\alpha) - \sum_i x_i \operatorname{Log}(\gamma_i)\|$ is small. To do this, we find the closest vector to $\operatorname{Log}(\alpha)$
in the lattice $\mathcal{L}:= \Z\operatorname{Log}(\gamma_1)+\cdots+\Z\operatorname{Log}(\gamma_r)$.
The closest vector problem is a notoriously hard problem without prior knowledge on the properties of the lattice and the
target vector. In our situation, things are made easier by the knowledge of the distribution of the target vector (given in
the description of the cryptosystems~\cite{fre_smart,mult_maps}) and of a good basis for $\mathcal{L}$. Indeed, the decryption
algorithm of schemes relying on the hardness of the short-PIP works under the assumption that the generator $g$ is small. This
means that the target vector $\operatorname{Log}(\alpha)$ is very close to the lattice $\mathcal{L}$. Our instance of the Closest Vector Problem
therefore turns into an instance of the Bounded Distance Decoding problem (BDD) since we have a bound on
$\operatorname{dist}(\operatorname{Log}(\alpha),\mathcal{L})$. Moreover, we know a very good basis $\operatorname{Log}(\gamma_i)$ for $\mathcal{L}$ with which
Babai's nearest plane algorithm~\cite{barbai} returns the correct value.
\begin{algorithm}[ht]
\caption{short-PIP to PIP reduction}
\begin{algorithmic}[1]\label{alg:CVP}
\REQUIRE A generator $\alpha\in\mathcal{O}_K$ of the ideal $\ag\subseteq\mathcal{O}_K$.
\ENSURE A short generator of $\ag$.
\STATE Let $(\gamma_i)_{i\leq r}$ be the cyclotomic units of $K$.
\STATE $\mathcal{L}\leftarrow \Z\operatorname{Log}(\gamma_1)+\cdots\Z\operatorname{Log}(\gamma_r)$.
\STATE Find the closest vector $\sum_i x_i\operatorname{Log}(\gamma_i)\in\mathcal{L}$ to $\operatorname{Log}(\alpha)$ by using Babai's round-off method.
\RETURN $\frac{\alpha}{\prod_i\gamma_i^{x_i}}$.
\end{algorithmic}
\end{algorithm}
\begin{theorem}[Th. 4.1 of~\cite{CVP_cyclo}]
When $K = \Q(\zeta_{p^s})$, Algorithm~\ref{alg:CVP} runs in polynomial time.
\end{theorem}
A short generator of a principal ideal in $K = \Q(\zeta_{p^s})$ yields a solution to $\gamma$-SVP for $\gamma\in e^{\tilde{O}(\sqrt{n})}$.
Moreover, under reasonable assumptions on the ideal class group,
given an arbitrary input ideal $I\subseteq\mathcal{O}_K$, the heuristic methods of~\cite{Stickelberger} allow us to find an ideal $J$
with $\mathcal{N}(J)\in 2^{O(n^{3/2})}$ such that $IJ$ is principal. Then a short generator of $IJ$ is a solution to $\gamma$-SVP for $I$ with $\gamma\in e^{\tilde{O}(\sqrt{n})}$.
The close principal multiple algorithm of~\cite{Stickelberger} uses the decomposition of an input ideal on a short generating set
of the ideal class group. This can be done in quantum polynomial time using~\cite{SODA16}, or in time $2^{O(n^{\frac{2}{3} - \frac{a}{3} + \varepsilon})}$ for arbitrarily small $\varepsilon > 0$
given a precomputation of cost in $2^{\tilde{O}(n^a)}$.
\begin{theorem}
If we precompute all relations between ideals of norm less than $B \in 2^{O(n^a)}$ for some $1/2 < a < 1$,
then the cost of solving $\gamma$-SVP for $\gamma\in e^{\tilde{O}(\sqrt{n})}$ is in $2^{O\left(n^{\frac{2}{3} - \frac{a}{3} + \varepsilon}\right)}$
for arbitrarily small $\varepsilon > 0$
while the cost of the precomputation is in $2^{\tilde{O}(n^a)}$.
\end{theorem}
\section{Smoothness of ideals}\label{sec:smoothness}
Given $B>0$, the expected time to find a relation of the form $(\alpha) = \pg_1^{e_1}\cdots\pg_m^{e_m}$
where $\mathcal{N}(\pg_i)\leq B$ depends on the probability that a random ideal $\ag$ of bounded norm is
$B$-smooth, that is to say of the form $\ag = \pg_1^{e_1}\cdots\pg_m^{e_m}$. In~\cite{Scourfield},
Scourfield established a result on the smoothness of ideals in a number field
comparable to the ones known on integers. Let
$$\Psi(x,y) := \left| \{ \ag\subseteq \mathcal{O}_K, \mathcal{N}(\ag)\leq x, \ag\ y\mbox{--smooth}\} \right|,$$
and $\varepsilon > 0$, then $\frac{\Psi(x,y)}{x}\sim \lambda_K\rho(u)$, where $u=\frac{\ln(x)}{\ln(y)}$,
$\rho$ is the Dickman function, $\lambda_K$ is the residue of the zeta function $\zeta_K(s)$ at
$s=1$ and
$(\ln\ln(x))^{\frac{5}{3}+\varepsilon}\leq \ln(y)\leq \ln(x),\ x \geq x_0(\varepsilon)$
for some $x_0(\varepsilon)$. In the case where $K$ is normal and $\frac{n}{\ln|\Delta|}\rightarrow 0$,
$\lambda_K$ can be bounded absolutely, but there is no such result in the general case.
During our relation search algorithm, we draw principal ideals at random. There is no known
analogue of Sourfield's result for restricted classes of ideals. This is one of the reasons why the complexity of the number field
sieve~\cite{NF_sieve} is only heuristic.
We therefore
rely on the following heuristic for the smoothness of ideals.
\begin{heuristic}\label{heuristic:smoothness_ideals}
We assume that the probability $P(\iota,\mu)$ that a principal ideal of $\mathcal{O}_K$ of norm bounded by $\iota$
is a power-product of prime ideals of norm bounded by $\mu$ satisfies
\begin{equation}\label{eq:smooth_ideal}
P(\iota,\mu)\geq e^{\left( -u \ln u (1+o(1))\right)},\ \mbox{for }u = \ln(\iota) / \ln(\mu).
\end{equation}
\end{heuristic}
\begin{comment}
Moreover, we need to decompose ideals as power products of primes
of inertia degree 1, that is of the form $p\mathcal{O} + (\theta-v_p)\mathcal{O}$,
where $v_p$ is a root of $T(X)\bmod p$. In general, prime ideals can have inertia degree $f\geq 2$ and
thus be of the form $p\mathcal{O} + T_p(\theta)\mathcal{O}$ where $\deg(T_p) = f$.
In~\cite{ANTS_XI}, Biasse and Fieker argued that the proportion of
inertia degree $f\geq 2$ prime ideals of norm less than $B$ was asymptotically small when $B = L_\Delta(a,c)$
for some $0<a<1$ and $c>0$. This is due to the fact that the
norm of a prime ideal inertia degree $f$ is $p^f$, and therefore for large values of $q$, ideals above $q$ can only be of
norm less than $B$ is $f = 1$. The primes of $\Q(\zeta_{N})$
with inertia degree 1 (totally split primes) are the prime ideals above $p = kN + 1$ (see for example~\cite{Washington}[Th. 2.13]).
The properties of the primes of the form $k2^s+1$ have been extensively studied in the scope of the factorization of Fermat numbers,
but nothing indicates that an odd number of the form $k2^s+1$ has less (or more) chance to be prime than any other
odd number. The proportion of split primes less than $B$ is expected to be around $\pi(B)/N$. Heuristically, we expect to
have to draw $O(\log(|\Delta|))$ primes at random to get one of inertia degree 1. Suppose the norm behaves like a random number.
A result of Erd\"{o}s and Kac~\cite{Erdos_Kac} tells us that almost all numbers $x$ have $\ln\ln(x)(1+o(1))$ different prime divisors.
We therefore expect to have to repeat $\log(|\Delta|)^{\tilde{O}(\log\log|\Delta|)}$ the search for a relation to obtain
one that involves only degree 1 primes. If the run time is subexponential (which is the case in all our algorithms), then
this overhead is asymptotically negligible.
\begin{heuristic}\label{heuristic:smoothness}
We assume that the probability $P(\iota,\mu)$ that a principal ideal of $\mathcal{O}$ of norm
bounded by $\iota$ is is a power-product of degree 1 prime ideals of norm bounded by $\mu$ satisfies
\begin{equation}\label{eq:smooth_principal}
P(\iota,\mu)\geq e^{\left( -u \ln u (1+o(1))\right)},\ \mbox{for }u = \ln(\iota) / \ln(\mu).
\end{equation}
\end{heuristic}
\end{comment}
\section{Computation of $\operatorname{Cl}(\mathcal{O}_K)$}\label{sec:CL_group}
In this section, we show how to compute $\operatorname{Cl}(\mathcal{O}_K)$ where $K = \Q(\zeta_{p^s})$ is
a cyclotomic field of prime power conductor in time $L_\Delta(1/2,c)$ for some constant $c>0$
and $\Delta = \operatorname{disc}(K)$
under Heuristic~\ref{heuristic:smoothness_ideals}. The best known heuristic complexity
for fields of degree $n\in \tilde{\Theta}(\log(|\Delta|))$ is in $2^{O(n^{2/3+\varepsilon})}$
where $\varepsilon>0$ is an
arbitrarily small constant~\cite{ANTS_XI}. Cyclotomic fields of prime power conductor
have a defining polynomial
with height 1, which allows us to use a different technique than the one described
in~\cite{ANTS_XI}. All we have to do is to draw elements $\alpha\in\mathcal{O}_K$ with small
coefficients on the power basis $1,\zeta_{p^s},\cdots,\zeta_{p^s}^{n-1}$ and test them for smoothness
with respect to a factor basis $\mathcal{B} = \{ \pg\ \mid \ \mathcal{N}(\pg)\leq B\}$ for
some smoothness bound $B > 0$. The smoothness test is simply done by checking
if $\mathcal{N}(\alpha)$ is $B$-smooth as an integer using either a factoring algorithm~\cite{NF_sieve,pomerance_qs}
or a dedicated smoothness test algorithm~\cite{bernstein}. Every time we have a relation of the form
$$(\alpha) = \pg_1^{e_1}\cdots\pg_m^{e_m},$$
we store the vector $(e_1,\cdots,e_m)$ in the rows of a matrix $M$. Once enough relations are found, we
complete the computation as in all previous subexponential methods~\cite{BSub,HMSub,ANTS_XI} by processing the matrix
$M$.
\begin{algorithm}[ht]
\caption{Computation of the class group of $\Q(\zeta_l)$}
\begin{algorithmic}[1]\label{alg:class_group}
\REQUIRE A smoothness bound $B > 0$, a constant $A > 0$ and a conductor $N$.
\ENSURE $d_i$ such that $\operatorname{Cl}(\mathcal{O}_K)=\bigoplus_i \Z/d_i$ for $K = \Q(\zeta_{N})$ and $M\in\Z^{k\times m}$,$(\alpha_i)_{i\leq k}\in\mathcal{O}_K^n$
such that for each row $M_i$ of $M$, $\mathcal{B}^{M_i} = (\alpha_i)$, where $\mathcal{B}= \{\pg\ \mid \ \mathcal{N}(\pg)\leq B\}$.
\STATE Compute $\mathcal{B} = \{\pg \ \mid \ \mathcal{N}(\pg)\leq B\}$.
\STATE $m\leftarrow |\mathcal{B}|$, $k\leftarrow m$, $M\in\Z^{0\times m}$.
\WHILE{ The number of relations is less than $k$}
\STATE $(a_i)_{i\leq n}\xleftarrow[]{\mathcal{R}}[-A,A]^n$, $\alpha\leftarrow \sum_i a_i\zeta_N^i$.
\IF{$(\alpha)$ is $\mathcal{B}$-smooth}
\STATE Find $(e_i)_{i\leq m}$ such that $(\alpha)=\prod_i\pg_i^{e_i}$.
\STATE $M\leftarrow \left(\frac{M}{(e_i)}\right)$.
\ENDIF
\ENDWHILE
\STATE \textbf{if} $M$ does not have full rank \textbf{then} $k\leftarrow 2k$ and go to Step~3.
\STATE $H\leftarrow \mathrm{HNF}(M)$. $d\leftarrow \det(H)$. $B\leftarrow \ker(M)$.
\STATE $L\leftarrow \{\operatorname{Log}(\alpha_1),\cdots,\operatorname{Log}(\alpha_k)\}$. $C\leftarrow \left(L^{B_i}\right)$.
\STATE Let $V$ be the volume of the lattice generated by the rows of $C$, and $h^*$ be an approximation of $hR$ given by the
methods of~\cite{BachEulerProd}.
\STATE \textbf{if} $dV > 1.5h^*$ \textbf{then} $k\leftarrow 2k$ and go to Step~3.
\STATE $\operatorname{diag}(d_i,\cdots,d_m)\leftarrow \mathrm{SNF}(H)$.
\RETURN $(d_i)_{i\leq m}$, $M$. $(\alpha_i)_{i\leq k}$.
\end{algorithmic}
\end{algorithm}
The run time of Algorithm~\ref{alg:class_group} depends on the probability of smoothness of
principal ideals, which is ruled by Heuristic~\ref{heuristic:smoothness_ideals}. This gives
us a bound on the average time to find a relation. However, there is no indication that the
relations we find are distributed according to a distribution in $\Lambda$ allowing us to
terminate the computation in subexponential time. Suppose we found a full rank sublattice $\Lambda_0$ of
$\Lambda$, Hafner and McCurley~\cite{hafner}
proved under GRH that their relation search for quadratic fields yielded relation vectors such that
the probability of drawing one in the the coset $\vec{w} + \Lambda_0$ for any $\vec{w}\in \Lambda$
was bounded from below by a large enough bound. This allowed them to justify that the algorithm would
terminate with high enough probability in subexponential time. It it reasonable to assume that
by drawing coefficient vectors uniformly at random in $[-A,A]$, the generators of the principal ideals
of our relations will be well enough distributed to justify that the relations themselves are
equally distributed in $\Lambda$, but proving it remains an open question.
\begin{heuristic}\label{heuristic:relation_lattice}
Let $\Lambda_0$ be a sublattice of $\Lambda$ corresponding to the relations between primes
in $\mathcal{B}$, and $A > 1$ be a constant. If $(a_i)_{i\leq n}$ is
drawn uniformly at random in $[-A,A]^n$, and if $(\alpha)\mathcal{O}_K = \mathcal{B}^{\vec{w}_\alpha}$ is $\mathcal{B}$-smooth,
then for any $\vec{w}\in\Lambda$
$$P\left( \vec{w}_\alpha \in \vec{w} + \Lambda_0\right) \geq \frac{\det(\Lambda)}{\det(\Lambda_0)}\left( 1 + o(1)\right).$$
\end{heuristic}
\begin{proposition}[GRH+Heuristic~\ref{heuristic:smoothness_ideals}+Heuristic~\ref{heuristic:relation_lattice}]\label{prop:class_group}
Algorithm~\ref{alg:class_group} with $B \in 2^{O(n^{1/2})}$ and $N$ of the form $p^s$
is correct and its heuristic complexity is in $2^{O(n^{1/2})}$
\end{proposition}
\begin{proof}
The run time depends on the smoothness probability of $\alpha\in\mathcal{O}_K$ drawn in Step~4. Let $P\in\Z[X]$ such that
$\alpha = P(\zeta_{N})$ for $N = p^s$. The norm of $\alpha$ is given by $\mathrm{Res}(\Phi_N,P)$ where
$\Phi_N$ is the $N$-th cyclotomic polynomial. The first $n$ rows of the resultant have
length less than $\sqrt{n}$ while the last $n$ rows have length bounded by $\sqrt{n}A$.
By Hadamard's bound, the resultant is bounded by $n^nA^n$. This means
that $\log(|\mathcal{N}(\alpha)|) \leq n\log(n)(1+o(1))$ (as $A$ is a constant).
Let $u := \frac{\log(|\mathcal{N}(\alpha)|)}{\log(B)}$, from Heuristic~\ref{heuristic:smoothness_ideals},
the probability of finding a smooth $\alpha$ is at least
$e^{-u\ln(u)(1+o(1))} \in \frac{1}{2^{O\left(n^{1/2}\right)}}$, and therefore the
whole relation search takes time $2^{O(n^{1/2})}$.
The linear algebra phase (HNF and SNF computation) takes time $|\mathcal{B}|^{4+o(1)} \in 2^{O(n^{1/2})}$,
which is asymptotically the same as the relation collection phase.
\end{proof}
\begin{corollary}[GRH+Heuristic~\ref{heuristic:smoothness_ideals}+Heuristic~\ref{heuristic:relation_lattice}]
Algorithm~\ref{alg:class_group} has heuristic complexity $2^{O(n^{a})}$ when $B \in 2^{O(n^{a})}$ for
$\frac{1}{2} \leq a < 1$ and $b > 0$.
\end{corollary}
\begin{comment}
\begin{proposition}\label{prop:class_group}
Algorithm~\ref{alg:class_group} with $B = L_\Delta\left(1/2,1/\sqrt{6}\right)$ and $N$ of the form $p^s$
is correct and its heuristic complexity is in $L_\Delta\left(1/2,2\sqrt{2}/3+o(1)\right).$
\end{proposition}
\begin{proof}
The run time depends on the smoothness probability of $\alpha\in\mathcal{O}_K$ drawn in Step~4. Let $P\in\Z[X]$ such that
$\alpha = P(\zeta_{N})$ for $N = p^s$. The norm of $\alpha$ is given by $\mathrm{Res}(\Phi_N,P)$ where
$\Phi_N$ is the $N$-th cyclotomic polynomial. The first $n$ rows of the resultant have
length less than $\sqrt{n}$ while the last $n$ rows have length bounded by $\sqrt{n}A$.
By Hadamard's bound, the resultant is bounded by $n^nA^n$. This means
that $\log(|\mathcal{N}(\alpha)|) \leq n\log(n)(1+o(1)) = \log(|\Delta|)(1+o(1))$ (as $A$ is a constant).
Let $u := \frac{\log(|\mathcal{N}(\alpha)|)}{\log(B)}$, from Heuristic~\ref{heuristic:smoothness},
the probability of finding a smooth $\alpha$ is at least
$e^{-u\log(u)(1+o(1))} = L_\Delta\left(\frac{1}{2},-\frac{\sqrt{6}}{2} + o(1)\right)$, and therefore the
whole relation search takes time $L_\Delta\left(\frac{1}{2},\frac{\sqrt{6}}{2} + \frac{1}{\sqrt{6}} + o(1)\right)$.
The linear algebra phase (HNF and SNF computation takes time $|\mathcal{B}|^4+o(1) = L_\Delta\left(\frac{1}{2},\frac{2\sqrt{2}}{3}+o(1)\right)$,
which is asymptotically the same as the relation collection phase.
\end{proof}
\begin{corollary}
Algorithm~\ref{alg:class_group} has heuristic complexity $L_\Delta\left(a,\max(4b,\frac{1}{2b} + b) + o(1)\right)$ when $B = L_\Delta(a,b)$ for
$\frac{1}{2} \leq a < 1$ and $b > 0$.
\end{corollary}
\end{comment}
\section{A $\qg$-descent algorithm to solve the PIP}\label{sec:q_desc}
In this section we show how decompose an input ideal $I\subseteq\mathcal{O}_K$ into a
$B$-smooth product for ideals in $\operatorname{Cl}(\mathcal{O}_K)$ for some $B>0$. In other words,
we find $\pg_1,\cdots,\pg_m$ with $\mathcal{N}(\pg_i)\leq B$, $(e_i)_{i\leq m}\in\Z^m$ and
$\alpha\in K$ such that $I = (\alpha)\prod_i\pg_i^{e_i}$. Then $I$ is principal if
and only if $\prod_i\pg_i^{e_i}$ is principal as well, which is checked by solving a linear system
as described in Algorithm~\ref{alg:PIP}. In the case where $I$ is principal, this
process also returns a generator.
To decompose $I$, we execute a $\qg$-descent consisting of searching for small
elements $\alpha$ of $I$ such that $(\alpha)/I$ is smooth with respect to ideals of norm less than
$I$. Then we iterate the process on all divisors of $(\alpha)/I$ that have norm greater
than $B$ until we finally break it down into a product of ideals of norm less than
$B$.
The
idea of the $\qg$-descent derives from the algorithms based on the number
field sieve~\cite{NF_sieve} to solve the discrete logarithm problem in time $L_q(1/3)$ in $\F_q$
(see in particular~\cite{Adleman93,gordon_GFp,NFS_Fre}). This idea was also used to
solve the discrete logarithm problem in the Jacobian of certain classes of $C_{ab}$ curves~\cite{Enge2}.
A $\qg$-descent strategy was used to
derive relations in $\operatorname{Cl}(\mathcal{O}_K)$ in~\cite{biasse_Phd,biasse_subexp,ANTS_XI}, but it was restricted
to classes of fields with degree $n \in O(\log(|\Delta|)^a$ for some $a < 1/2$. In this paper, we
achieve a heuristic $L_\Delta(1/2+\varepsilon,1)$ complexity for any $\varepsilon>0$ in
cyclotomic fields despite the fact that the degree of these fields satisfies $n\in\tilde{\Theta}(\log(|\Delta|))$.
\paragraph{\textbf{Finding short elements in $I$}}
We assume that $I$ is an ideal of $\mathcal{O}_K$. We want to find elements $\alpha\in I$
of small norm. To do this, we restrict the search to the lattice
$$\mathcal{L}_I := \Z v_{1,1} + \Z (v_{2,2}\zeta_N + v_{2,1}) + \cdots + \Z(v_{k,k}\zeta_N^k + v_{k,k-1}\zeta_N^{k-1}+\cdots + v_k)\subseteq I,$$
for some $k > 0$ where $N = q^s$ is the conductor of $K$. The coefficients $v_{i,j}$ are given
by the HNF of the $\Z$-basis of $I$ which has the shape
\[ H:=\left( \begin{array}{ccccc}
v_{1,1} & 0 & \hdots & \hdots & 0 \\
v_{2,1} & v_{2,2} & \ddots & & \vdots \\
\vdots & 0 & \ddots & \ddots & \vdots \\
\vdots & \vdots& \ddots & \ddots& 0 \\
v_{N,1}& 0 & \hdots & 0 & v_{N,N}\\
\end{array} \right).\]
\begin{lemma}\label{lemma:smooth}
We can find a vector
in $\mathcal{L}_I$ of length less than $\sqrt{k}\mathcal{N}(I)^{\frac{1}{k}}$ in time $2^{O(k)}$.
\end{lemma}
\begin{proof}
The determinant of $\mathcal{L}_I$ is that of the upper left $k\times k$
submatrix of $H$ and satisfies $\det(\mathcal{L}_I) \leq \prod_{i\leq N}v_{i,i} = \mathcal{N}(I)$.
\begin{comment}
$\sqrt{\det(A\cdot A^T)}$, and
\[ AA^T:=\left( \begin{array}{ccccc}
1 +v_1^2 & v_1v_2 & \hdots & v_1v_{k-1} & v_1v_k \\
v_1v_2 & 1 + v_2^2 & & & v_2v_k \\
\vdots & & \ddots & & \vdots \\
v_1v_{k-1} & & & \ddots & v_{k-1}v_k \\
v_1v_k & \hdots & \hdots & v_{k-1}v_k & 1+v_{k}^2\\
\end{array} \right).\]
\end{comment}
An HKZ-reduction returns a basis
whose first vector has length less than $\sqrt{k}(\det(\mathcal{L}_I))^{1/k}$ in time $2^{O(k)}$.
\end{proof}
\begin{lemma}\label{lemma:smooth_combinations}
In time $2^{O(k)}$, we can find an element $\alpha\in I$
such that $\mathcal{N}(\alpha)\leq n^{n}\mathcal{N}(I)^{\frac{n}{k}}.$
\end{lemma}
\begin{proof}
Let $\alpha$ be the first vector of an HKZ-reduced basis of $\mathcal{L}_I$. The calculation of this basis
takes time $2^{O(k)}$ and by Lemma~\ref{lemma:smooth}, the length of its first vector $(\alpha_1,\cdots,\alpha_k)$ is bounded
by $\sqrt{k}\mathcal{N}(I)^{\frac{1}{k}}$.
By the same argument as in the proof of Proposition~\ref{prop:class_group},
the algebraic norm of $\alpha := \sum_i \alpha_i \zeta_N^i$ satisfies
$$\mathcal{N}(\alpha) \leq \left.\sqrt{n}\right.^n \left( \|(\alpha_1,\cdots,\alpha_k)\| \right)^n\leq \underbrace{n^{n/2} k^{n/2}}_{\leq n^n}\mathcal{N}(I)^{\frac{n}{k}}.$$
\end{proof}
\paragraph{\textbf{A round of the $\qg$-descent}}
Let $\varepsilon > 0$ be an arbitrarily small constant and $\qg$ be a prime
ideal with $\log(\mathcal{N}(\qg)) \leq n^b$ for some $1/2\leq b \leq 1$.
We would like to decompose the class of $\qg$ in $\operatorname{Cl}(\mathcal{O}_K)$ as a product of primes $\pg$
with $\log(\mathcal{N}(\pg))\leq n^{b-\varepsilon}$ in time $2^{O(n^{1/2 + \varepsilon})}$.
\begin{algorithm}[ht]
\caption{One round of the $\qg$-descent}
\begin{algorithmic}[1]\label{alg:one_round}
\REQUIRE Prime ideal $\qg$ with $\log(\mathcal{N}(\qg))\leq n^b$, a constant $\varepsilon > 0$ and $A\geq 1$.
\ENSURE Prime ideals $\pg_i$, integers $e_i$ and $\alpha\in\qg$ such that $(\alpha)/\qg = \prod_i\pg_i^{e_i}$
and $\log(\mathcal{N}(\pg_i))\leq n^{b-\varepsilon}$.
\STATE $S\leftarrow \left\{\pg_i\ \mbox{such that }\mathcal{N}(\pg_i)\leq 12\log(|\Delta|)^2\right\}$.
\WHILE{no relation has been found}
\STATE $(x_i)\xleftarrow[]{R}[0,A]^{|S|}$, $I\leftarrow \qg\prod_i \pg_i^{x_i}$ where $R$ is the uniform distribution.
\STATE Construct a basis for the lattice $\mathcal{L}_I$ with $k = n^{1/2 + \varepsilon}$.
\STATE HKZ-reduce $\mathcal{L}_I$ and let $\alpha$ be its first vector.
\STATE \textbf{if} $(\alpha)$ is $2^{n^{b-\varepsilon}}$-smooth, find $(\pg_i)$,$(e_i)$ such that $(\alpha) = \prod_i\pg_i^{e_i}$.
\ENDWHILE
\RETURN $\alpha$, $(e_i)$, $(\pg_i)$, $(e_i)$.
\end{algorithmic}
\end{algorithm}
\begin{proposition}[GRH + Heuristic~\ref{heuristic:smoothness_ideals}]\label{prop:run_time_qstep}
Let $\varepsilon > 0$ be a constant, $1/2\leq b\leq 1$,
and $\qg$ be a prime ideal with $\log(\mathcal{N}(\qg))\leq n^b$.
There is a large enough constant $A$ such that Algorithm~\ref{alg:one_round} returns a decomposition of
$\qg$ in $\operatorname{Cl}(\mathcal{O}_K)$ as a product of degree one primes $\pg_i$ with
$\log(\mathcal{N}(\pg))\leq n^{b-\varepsilon}$ in time $2^{O(n^{1/2 + \varepsilon})}$ when
the conductor $N$ is of the form $p^s$.
\end{proposition}
\begin{proof}
The ideal $I$ created in Step~3 of Algorithm~\ref{alg:one_round} satisfies $\mathcal{N}(I) = \mathcal{N}(\qg)^{1+o(1)}$. Therefore,
according to Lemma~\ref{lemma:smooth_combinations}, the $\alpha$ derived in Step~5
satisfies $\log(\mathcal{N}(\alpha)) \leq \frac{n}{k}\log\left(\mathcal{N}(\qg)\right)(1+o(1))$. As $\log\left(\mathcal{N}(\qg)\right) \leq n^{b}$ and $k = n^{1/2 + \varepsilon}$,
this gives us $\log(\mathcal{N}(\alpha)) \leq n^{1/2 + b - \varepsilon}(1+o(1))$. We want $(\alpha)/I$ to be
$B$-smooth for $\log(B) = n^{b - \varepsilon}$.
Let $u = \frac{\log(\mathcal{N}(\alpha))}{\log(B)}$, the probability of finding such an $\alpha$ is
$$P = e^{-u\ln(u)(1+o(1))} = \frac{1}{2^{\tilde{O}(n^{1/2}(1+o(1))}}.$$
For any constant $A\geq 2$, the size $A^{|S|} \geq 2^{n^2}$ of the search space is sufficient to find our $\alpha$. The run time
of this procedure is dominated by the reduction of $\mathcal{L}_q$ which takes $2^{O(k)} = 2^{O(n^{1/2 + \varepsilon})}$.
\end{proof}
\paragraph{\textbf{Full procedure}}
Assume we are given an arbitrary ideal $I\subseteq\mathcal{O}_K$ as input. Before initiating the $\qg$-descent, we need to
break it down into a product of primes $\qg_i$ with $\log(\qg_i)\leq n$. To do this,
we multiply $\ag$ by an random power product of ideals
of the factor base $\mathcal{B}$, and BKZ-reduce it. This is done by finding
a short element $\phi\in\cg$ where $\ag^{-1}=\frac{1}{d}\cg$
with $d\in\Z_{>0}$ and $\cg\subseteq \mathcal{O}$. When using the BKZ algorithm with block size $l = n^{1/2}$, such a short element satisfies
$$\| \phi \| \leq n^{n^{1/2}/4} |\Delta|^{\frac{1}{2n}} \mathcal{N}(\cg)^{\frac{1}{n}}.$$
Then, the ideal $\bg:= \frac{\phi}{l}\ag$ satisfies
$\bg\subseteq\mathcal{O}_K$, $\mathcal{N}(\bg)\leq 2^{\frac{n^{3/2}\log(n)}{2}} \sqrt{|\Delta|}$ and it is
$2^n$-smooth with probability $\frac{1}{2^{\tilde{O}(n^{1/2}(1+o(1))}}$ by following the same argument as in the proof
of Proposition~\ref{prop:run_time_qstep}. Repeating the process $2^{n^{1/2 + o(1)}}$ times yields the
original $2^n$-smooth decomposition.
\begin{algorithm}[ht]
\caption{Initial decomposition}
\begin{algorithmic}[1]\label{alg:LLL_red}
\REQUIRE An ideal $I\subseteq \mathcal{O}_K$, factor base $\mathcal{B}$, and a constant $A\geq 1$.
\ENSURE $\alpha\in K$ such that $\bg:= (\alpha)I$ is $2^{n}$-smooth.
\STATE $\ag\leftarrow \prod_i\pg_i^{e_i}$ for random $e_i\leq A$ (use a square-and-multiply strategy).
\STATE $\cg \leftarrow l\ag^{-1}$ where $l$ is the denominator of $\ag$.
\STATE Find a BKZ reduced $\phi\in\cg$ with block size $l = n^{1/2}$.
\STATE $\alpha\leftarrow \frac{\phi}{l}$.
\STATE \textbf{if} $\bg:= (\alpha)I$ is not $2^{n}$-smooth
for degree 1 primes \textbf{then} go to Step~1.
\RETURN $\alpha$.
\end{algorithmic}
\end{algorithm}
\begin{algorithm
\caption{$\q$-descent}
\begin{algorithmic}[1]\label{alg:full_q_descent}
\REQUIRE $I\subseteq \mathcal{O}_K$ and an arbitrarily small $\varepsilon > 0$.
\ENSURE Prime ideals $(\qg_i)_{i\leq l}\in\mathcal{B}$ with $\mathcal{N}(\q_i)\leq 2^{n^{1/2}}$, integers $(e_i)$ and $(\phi_j)_{j\leq k}\in K$ such that
$I =\prod_{j\leq k}(\phi_j) \cdot\prod_{i\leq l}\mathfrak{q}_i^{e_i}$
\STATE Decompose $I = (\phi_1)\prod_i\qg_i$ using Algorithm~\ref{alg:LLL_red}.
\STATE $\mathrm{genList}\leftarrow \left\lbrace \phi_1\right\rbrace $, $\mathrm{primeList}\leftarrow \left\lbrace\mathfrak{q}_1,\hdots,\mathfrak{q}_l\right\rbrace$, $\mathrm{expList}\leftarrow \{1,\cdots,1\}$.
\STATE $b\leftarrow 1$.
\WHILE {$b > 1/2$}
\FOR {$\q\in\mathrm{primeList}$ with $\log(\mathcal{N}(\qg))> n^{b-\varepsilon}$}
\STATE Find $(\mathfrak{q}_i)_{i\leq l}$, $(e_i)_{i\leq l}$ and $\phi_k$ such that
$\mathfrak{q}=(\phi_k)\prod_{i\leq l}\mathfrak{q}_i^{e_i}$ by a round of the $\qg$-descent.
\STATE $\mathrm{genList}\leftarrow \mathrm{genList}\cup\left\lbrace\phi_k\right\rbrace$, $\mathrm{primeList}\leftarrow\mathrm{primeList}\cup\left\lbrace \mathfrak{q}_1,\hdots,\mathfrak{q}_l\right\rbrace$.
\STATE $\mathrm{expList}\leftarrow \{e_1,\cdots,e_l\}$.
\STATE Remove $\qg$ from $\mathrm{primeList}$ and $\mathrm{expList}$.
\ENDFOR
\STATE $b\leftarrow b-\varepsilon$.
\ENDWHILE
\RETURN $\mathrm{genList}$, $\mathrm{primeList}$, $\mathrm{expList}$.
\end{algorithmic}
\end{algorithm}
\begin{proposition}
Algorithm~\ref{alg:full_q_descent} decomposes $I$ as an $2^{n^{1/2}}$-smooth product in $\operatorname{Cl}(\mathcal{O}_K)$
in time $2^{O(n^{1/2+\varepsilon})}$ for arbitrarily small $\varepsilon > 0$.
\end{proposition}
\begin{proof}
The cost of the initial decomposition is in $2^{O(l)}$ where $l = n^{1/2}$ is the block size of the BKZ reduction.
The depth of the decomposition tree arising from the $\qg$-descent is $\frac{1}{2\varepsilon}$ which is
a constant. The arity of this tree is at most $n^2$, so Algorithm~\ref{alg:one_round} is
called at most polynomially many times.
\end{proof}
\section{Practical improvements to the computation of $\operatorname{Cl}(\mathcal{O}_K)$ and the resolution of the PIP}
The subexponential methods we presented in Section~\ref{sec:CL_group} and Section~\ref{sec:q_desc} have
a heuristic asymptotic run time in $2^{O(n^{1/2+\varepsilon})}$, and depending on the choices made for
implementation, the practical performances of these algorithms can vary significantly. In this section, we
present some practical improvements which do not affect the asymptotic complexity, but which impact
the practical run time. As the computation of the ideal class group and the resolution of the
Principal Ideal Problem are widely studied problems, there are many existing improvements that readily
apply to our algorithm, and many of them are folklore. For example, for each relation $\prod_i\pg_i = (\alpha)$
we find, we immediately get an additional $|\mathrm{Gal}(K/\Q)|-1$ others by including
$\prod_i\pg_i^{\sigma} = (\sigma(\alpha))$ for all $\sigma \in \mathrm{Gal}(K/\Q)$. This
method is present in the class group algorithm available in both PARI~\cite{pari} and Magma~\cite{magma}. The large prime variants are
another folklore practical improvement. It was originally described in the context of integer factorization~\cite{Lenstra:2LP,Lenstra:3LP},
and was successfully adapted to the resolution of the DLP in finite fields, and in the Jacobian of curves.
The single large prime variant for computing $\operatorname{Cl}(\mathcal{O}_K)$ was first presented by Jacobson~\cite{JacobsonPhd} while
the double large prime variant was first successfully used for computing class groups by Biasse~\cite{biasse}. Other improvements
impacting the practical performances of our methods include quadratic~\cite{Jdl} and lattice~\cite{biasse_fieker}
sieving, as well as optimized methods of computing the HNF and the SNF of a large integer matrix. In this section, we
restrict ourselves to improvements that are very specific to the settings of the algorithm. More specifically,
we develop two points:
\begin{itemize}
\item We show how to efficiently reduce the resolution of the PIP in $\operatorname{Cl}(\mathcal{O}_{\Q(\zeta_N)})$ to an instance of
the PIP in $\operatorname{Cl}(\mathcal{O}_{\Q(\zeta_N)^+})$.
\item We show how to enhance the relation search by looking for elements of small norm near the cyclotomic units.
\end{itemize}
\paragraph{\textbf{Solving the PIP in the maximal real subfield}}
Using subfields (and in particular the maximal real subfield $K^+$) is folklore. Several references use a variant
around this idea. We specifically rely on the Gentry-Szydlo method~\cite{Gentry_Szydlo}
and its extension by Howgrave-Graham and Szydlo~\cite{szydlo_norm_eq} for solving norm
equations of the form $\mathcal{N}_{\Q(\zeta_N)/\Q(\zeta_N)^+}(x) = g$. Halving the degree of the
field can have a significant impact on the practical behavior of the PIP algorithm.
The input size which is given by $\log(|\Delta|)$ is halved in the case of a power of
two cyclotomic. In addition, the LLL algorithm has a practical behavior significantly
better than the worst case estimates for lower dimensions. Indeed, it is expected to
return a basis $(b_i)$ such that $\frac{\|b_1\|}{\det(L)^{1/d}}\approx (1.02)^d$ (see~\cite{LLL_average}). This
makes a difference in practice during a relation search based on the search for small elements
in LLL-reduced bases of ideals\footnote{In $K^+$ the height of the defining polynomial is higher, therefore other
relation search methods may be used}.
Given an input ideal $I \subseteq \mathcal{O}_K$, we want to find a generator of the ideal $I'$ generated
by $\mathcal{N}_{K/K^+}(I)$ in $\mathcal{O}_{K^+}$. The ideal $I'$ is principal.
However, the norm map is not surjective. This means that there can be a generator
of $I'$ that is not the norm of a generator of $I$. If we solve the PIP for $I'$ in $\mathcal{O}_{K^+}$ and
find $g\in\mathcal{O}_{K^+}$ such that $I' = (g)$, then
we need to find another generator $g'$ that is the norm of an element in $K$. To proceed,
we find the right unit $u\in\mathcal{O}_{K^+}^*$ such that $g' = ug$ is totally positive. Given a
set of fundamental units $(u_1,\cdots,u_r)$ for $\mathcal{O}_{K^+}^*$, this can be done by solving
a linear system in $\F_2$. For $\alpha\in K^+$, let $\delta(\alpha)\in\F_2^r$ be the signature of
$x$, i.e. $\delta(\alpha)_i = 0$ if $\sigma_i(\alpha)\geq 0$, and $\delta(\alpha)_i = 1$ otherwise. Let
$M = (\delta(u_i)_j)_{i,j\leq r}$, then the right product of the $u_i$ turning $g$ into a totally
positive number is $u = \prod_i u_i^{x_i}$ where $(x_1,\cdots,x_r)M = \delta(g)$.
\begin{proposition}
When $K$ is of the form $K = \Q(\zeta_{2^s})$, there is always a solution to the system $(x_1,\cdots,x_r)M = \delta(g)$.
\end{proposition}
\begin{proof}
According to~\cite[Prop. 1]{signature_unit}, there are units of arbitrary signature in $\Q(\zeta_{2^s})$. Therefore,
there must be a linear combination of the signatures of the units in the generating set that matches $\delta(g)$.
\end{proof}
\begin{proposition}
When $K = \Q(\zeta_{2^s})$, a totally positive generator $g'$ of $I'$ is necessarily the norm of a generator of $I$.
\end{proposition}
\begin{proof}
$I'$ is generated by at least one
totally positive number (i.e. the image $\mathcal{N}_{K/K^+}(g_0)$ of a generator $g_0$ of $I$ by the norm map).
Then from~\cite{squares_unit,weber},
we know that the totally positive units are exactly the squares of units, which are also the norms
of the units of $\mathcal{O}_K$ that are in $\mathcal{O}_{K^+}$. The two totally positive generators $g',\mathcal{N}_{K/K^+}(g_0)$ of $I'$ differ by a
totally positive unit, hence a square, and hence the image of a unit $u_0$ of $\mathcal{O}_K\cap\mathcal{O}_{K^+}$ by the norm map,
i.e. $g' = \mathcal{N}_{K/K^+}(u_0)\mathcal{N}_{K/K^+}(g_0) = \mathcal{N}_{K/K^+}(u_0g_0)$, which is the image of a generator of
$I$.
\end{proof}
Before applying the Howgrave-Graham and Szydlo~\cite{szydlo_norm_eq} norm equation resolution method,
we need to make sure that the input is polynomially bounded, which is not guaranteed if we take
an arbitrary solution to the PIP in $\mathcal{O}_{K^+}$ (even after adjusting the signature). However, we
know the existence of a short (totally positive) generator of $I'$, namely the norm of a short generator
of $I$. We are facing an instance of the Bounded Distance Decoding problem similar to the one solved
by Cramer et al.~\cite{CVP_cyclo}. The logarithms of the norms of the cyclotomic units enjoy similar geometric properties
as the logarithms of the cyclotomic units themselves. Therefore, applying the method of~\cite{CVP_cyclo}
on $g'$ yields a small generator $g''$ of $I'$ on which we can apply the algorithm for the resolution of the
norm equation $\mathcal{N}_{K/K^+}(x) = g''$ given in~\cite{szydlo_norm_eq}.
\begin{algorithm}[ht]
\caption{Reduction from the PIP in $\mathcal{O}_K$ to the PIP in $\mathcal{O}_{K^+}$}
\begin{algorithmic}[1]\label{alg:PIP_subfield}
\REQUIRE $I\subseteq\mathcal{O}_K$, $I'=\mathcal{N}_{K/K^+}(I)\mathcal{O}_{K^+}$, $ \mathcal{O}_{K^+}^*= \langle u_1,\cdots,u_r\rangle$, $g$ with $I'=(g)$.
\ENSURE $g_0\in\mathcal{O}_K$ with $I = (g_0)$.
\STATE $M\leftarrow (\delta(u_i)_j)_{i,j\leq r}$, $y\leftarrow \delta(g)$.
\STATE Solve $(x_1,\cdots,x_r)M = y$. $g'\leftarrow \prod_i u_i^{x_i}$.
\STATE Find a close vector $\operatorname{Log}(u)$ to $\operatorname{Log}(g')$ in $\mathcal{L} = \Z\operatorname{Log}(u'_1)+\cdots+\Z\operatorname{Log}(u'_r)$ where the $u'_i$ are the norms of the
cyclotomic units of $\mathcal{O}_K$ by using the methods of~\cite{CVP_cyclo}.
\STATE $g''\leftarrow g'\cdot u^{-1}$.
\STATE Solve $\mathcal{N}_{K/K^+}(g_0) = g''$.
\RETURN $g_0$.
\end{algorithmic}
\end{algorithm}
\paragraph{\textbf{Search for relations around the cyclotomic units}}
Our relation search method to compute $\operatorname{Cl}(\mathcal{O}_K)$ consists of drawing polynomials $f(\zeta_N)$
with small coefficients at random and to check the norm of the resulting algebraic
integer for smoothness. As we saw, the algebraic norm is bounded from above by a function of
the length of the vector of coefficients of $f(X)$. Therefore, it is natural to search
elements represented by a very small coefficient vector.
We observe that $\mathcal{N}(f(\zeta_N))$ is a polynomial in the coefficients of $f$, and is therefore
a continuous function. In cyclotomic fields, we know a set of minimums for this function before
hand: the cyclotomic units $u_a = \frac{\zeta_N^a - 1}{\zeta_N -1}$. For each $a$, $\mathcal{N}(u_a) = 1$,
and we expect small variations around the $u_a$ to yield algebraic numbers of small norm,
although the coefficient vector of the corresponding polynomials might be long. In Table~\ref{tab:num_ev},
we compared the strategy consisting of drawing algebraic integers that are small variations around
the cyclotomic units to the sampling of random coefficient vectors. We drew coefficient vectors
of the same Hamming weight $w = 10,20,30,50,75,100$ in $K = \Q(\zeta_{2^{512}})$.
For each $w$, we drew 100 random coefficient vectors
of coefficients in $\{0,1\}$ (denoted by ``Random vectors''), and we drew 100 elements that differed
from a cyclotomic unit $u$ of weight $w$ by a term in $\zeta_N^i$ for some $i$.
\begin{table}[ht]
\caption{Average $\log(\mathcal{N}(\alpha))$}
\label{tab:num_ev}
\begin{center}
\begin{tabular}{|r|r|r|}
\hline
\multicolumn{1}{|r|}{Weight} &
\multicolumn{1}{r|}{Random vectors}&
\multicolumn{1}{|r|}{Unit variations} \\
\hline
10 & 301 & 154 \\
20 & 430 & 156 \\
30 & 503 & 158 \\
50 & 586 & 159 \\
75 & 638 & 156 \\
100 & 674 & 154 \\
\hline
\end{tabular}
\end{center}
\end{table}
In Table~\ref{tab:num_ev}, we observe that the size of the elements $\alpha$
increases as the Hamming weight of their coefficient vector over the power
basis gets larger. Meanwhile, at a comparable Hamming weight, the size of a small
variation around a cyclotomic unit does not seem to be affected. In fact, it is
the distance to the unit that seems to rule the size of $\alpha$. For example, we
measured small variations around $u_{101}$ (of Hamming weight 101).
\begin{itemize}
\item $\log(u_{101} + \zeta_N^3) = 151$.
\item $\log(u_{101} + \zeta_N^3 - \zeta_N^{150}) = 204$.
\item $\log(u_{101} + \zeta_N^3 + \zeta_N^{60} - \zeta_N^{150}) = 238$.
\item $\log(u_{101} + \zeta_N^3 + \zeta_N^{60} - \zeta_N^{150} + \zeta_N^{200}) = 283$.
\end{itemize}
If we draw $\alpha'$ and $\alpha$ at random such that $\log(|\mathcal{N}(\alpha')|)\leq \frac{|\log(\mathcal{N}(\alpha))|}{d}$ for some $d > 0$,
then the expected time $T'$ to find a smooth $\alpha'$ satisfies $T'\leq \sqrt[d]{T}$ where $T$ is
the expected time to find a smooth $\alpha$. Our numerical results indicate that the unit variation method
should provide a significant speed-up over the random vectors method even when we allow a Hamming distance larger
than one.
\section{PIP and $\gamma$-SVP in $\Q(\zeta_{p^s})$ with precomputation}\label{sec:short_PIP}
In Section~\ref{sec:q_desc}, we showed how to compute the PIP in heuristic subexponential time $2^{O(n^{1/2+ \varepsilon})}$.
This provides an attack against schemes relying on the hardness of finding a short generator of a principal ideal
such as~\cite{fre_smart,mult_maps}. Also, according to~\cite{CVP_cyclo}, the size of a short generator of $I$ resulting from the BDD algorithm
on the log-unit lattice is within $e^{O(\sqrt{n})}$ of the first minima of the ideal lattice $I$. Moreover, it was recently conjectured~\cite{Stickelberger} that
for most fields $\Q(\zeta)$, any ideal of $\mathcal{O}_{\Q(\zeta)}$ was within a short enough ideal multiple from a principal ideal. Therefore,
solutions to the PIP in $\mathcal{O}_{\Q(\zeta)}$ yields solution to $\gamma$-SVP in ideals of $\mathcal{O}_{\Q(\zeta)}$ for $\gamma \in e^{O(\sqrt{n})}$.
Since the methods of Section~\ref{sec:q_desc} run in heuristic subexponential time $2^{O(n^{1/2+ \varepsilon})}$, this does not
offer a better trade-off than using a BKZ reduction with block size in $O(\sqrt{n})$.
In this section, we show how to leverage a subexponential precomputation to solve all instances of $\gamma$-SVP in ideals of $\mathcal{O}_{\Q(\zeta_N)}$ for
$N = p^s$ and
$\gamma \in e^{O(\sqrt{n})}$ in time better than $2^{O(\sqrt{n})}$, thus achieving a better trade-off than the BKZ reduction. To the
best of our knowledge, this is the first time a method for solving $\gamma$-SVP beats the BKZ time/approximation trade-off even with a precomputation on the
field. The general idea is to use the ideal class group computation given by Algorithm~\ref{alg:class_group} with a larger factor base bound $B$.
This gives us a basis for all relations between the prime ideals of norm less than $B$. Then given an input ideal $I$, we first compute
$\alpha$ and $\qg_1,\cdots,\qg_k$ such that $I = (\alpha)\prod_i\qg_i$ with $\mathcal{N}(\qg_i)\leq B$, and then we solve the PIP problem for $\prod_i\qg_i$.
The larger the smoothness bound $B$ is, the more expensive the precomputation gets. Meanwhile all subsequent PIP resolutions using this precomputation
get faster since their
run time is dominated by the decomposition in $\operatorname{Cl}(\mathcal{O}_K)$ of $I$ as a product of ideals of norm bounded by $B$.
\begin{enumerate}
\item Precomputation: Given $B$, find a basis for the lattice $\mathcal{L}$ of vectors $(e_i)_{i\leq k}$ such that $\prod_i \qg_i^{e_i}$ for the $\mathcal{N}(\qg_i)\leq B$.
\item Decomposition: Given $I$ and $B$, find $(\qg_i)_{i\leq k}$ and $\alpha\in K$ such that $I = (\alpha)\prod_i \qg_i^{e_i}$.
\item Resolution: Given the decomposition of $I$ and $\mathcal{L}$, find a solution of $\gamma$-SVP for $I$.
\end{enumerate}
\paragraph{\textbf{Precomputation step}}
Given $B$, Algorithm~\ref{alg:class_group} returns a basis $(\vec{b}_1,\cdots,\vec{b}_k)$ of the lattice $\mathcal{L}$ of
vectors $\vec{x}$ such that $\mathcal{B}^{\vec{x}}\sim (1)$ for $\mathcal{B} = \{ \pg\ \mid \ \mathcal{N}(\pg)\leq B\}$
together with $\alpha_i\in K$ such that $\mathcal{B}^{\vec{b}_i} = (\alpha_i)$. Our precomputation step additionally
processes this basis and the $(\alpha_i)$ to return an HNF-reduced basis $(\vec{h}_1,\cdots,\vec{h}_m)$ for $\mathcal{L}$.
Using~\cite[Prop. 6.3]{Arne}, we can find a unimodular $U\in\mathrm{Gl}_{k\times k}(\Z)$ such that
\[ UB
=\left( \begin{array}{cccc}
h_{11}& 0 & \hdots & 0 \\
\vdots & h_{22}& \ddots & \vdots \\
\vdots & \vdots & \ddots & 0 \\
* & * & \hdots & h_{\tiny{N}\tiny{N}}\\
0 & \hdots & \hdots & 0 \\
\vdots & & & \vdots \\
0 & \hdots & \hdots & 0
\end{array}
\right)
\]
is the HNF of $B = (\vec{b}_i)_{i\leq k}$ with $\|U\|\leq \left( \sqrt{m}\|B\|\right)^m$ in time
$$O\left( km^{\theta - 1}\log(\beta) + km\log(m)\mathrm{Mult}(\log(\beta))\right)$$
for $\beta = \left( \sqrt{m}\|B\|\right)^m$, $\mathrm{Mult}(x)$ the complexity of $x$-bit
integer multiplication, and $2\leq \theta\leq 3$ the matrix multiplication
exponent. The matrix $H = UB$ has a small essential part. Under GRH, $h_{i,i} = 1$ for $i> 12\log(|\Delta|)^2$.
We leverage this to facilitate the resolution of the linear system giving the solution to the PIP.
However, for this to yield a generator (as opposed to just the answer whether or not $I$ is principal),
we need to compute the $(\beta_i)_{i\leq m}$ such that $\prod_{j\leq k}\alpha_j^{U_{i,j}} = \beta_i$
for $i\leq m$.
As the coefficients of $U$ and the number of terms $m$ in the product are large, we cannot afford to
write down these algebraic numbers on the power basis. However, we know that they are used to compute
an element of $I$ whose length is within $e^{\tilde{O}(\sqrt{n})}$ of the first minima $\lambda_1(I)\leq \sqrt{n}|\Delta|^{1/n}\mathcal{N}(I)^{1/n}$
of the ideal lattice $I$. So we compute $\beta_i\bmod \pg_j$ for a collection of split prime ideals
$\pg_1,\cdots,\pg_l$ such that $\prod_j\mathcal{N}(\pg_j) \leq e^n n^{n/2} |\Delta|\mathcal{N}(I)$.
We also need to keep $\operatorname{Log}(\beta_i)$ as part of the precomputation as they are needed for computing
a short generator of a principal ideal by solving an instance of the BDD in $\operatorname{Log}(\Z[\zeta_N]^*)$. Each
of these values satisfies $\operatorname{Log}(\beta_i) = \sum_{j\leq k} U_{i,j}\operatorname{Log}(\alpha_j)$. The logarithm vectors of
the $\alpha_j$ have polynomial size, but the bit size of the $U_{i,j}$ is in $2^{O(n^a)}$ for $1/2 < a < 1$
depending on the parameters of the precomputation. As we are aiming at lowering down the
cost of subsequent resolutions of the short-PIP which requires the values of $\operatorname{Log}(\beta_i)$, we must
find different generators corresponding to the rows of $UB$. Let $i_0$ such that
$\{\pg_1,\cdots,\pg_{i_0}\} = \{\pg\ \mid \ \mathcal{N}(\pg) < 12\ln(|\Delta|)^2\}$ and $i_1$ such that
$\mathcal{N}(\pg_{i_1}) > 12\ln(|\Delta|)^2$,
then $\beta_{i_1}$ is the generator of the principal ideal $\pg_{i_1}\cdot\left( \prod_{j\leq i_o}\pg_j^{h_{i_0,j}}\right) = (\beta_{i_1})\mathcal{O}_K$.
We use the BBD solution in the log-unit lattice of Cramer et al.~\cite{CVP_cyclo} to find $\beta'_{i_1}$ such that
$\pg_{i_1}\cdot\left( \prod_{j\leq i_o}\pg_j^{h_{i_0,j}}\right) = (\beta'_{i_1})\mathcal{O}_K$ and such that the bit size of the representation of
$\operatorname{Log}(\beta'_{i_1})$ is bounded.
\begin{algorithm}[ht]
\caption{Precomputation step}
\begin{algorithmic}[1]\label{alg:precomputation}
\REQUIRE Split primes $(\pg_i)_{i\leq l}$, $B > 0$, and conductor $N$.
\ENSURE $H$ in HNF form with $(\beta_j\bmod \pg_i)$ such that $\mathcal{B}^{H_i} = (\beta_i)\mathcal{O}_K$ for $\mathcal{B} = \{ \pg\ \mid \ \mathcal{N}(\pg)\leq B\}$,
and $(\operatorname{Log}(\beta_j))_{j\leq m}$.
\STATE Compute a generating set $\vec{b}_1,\cdots,\vec{b}_k$ of the lattice $\mathcal{L}$ of vectors $\vec{x}$ such that $\mathcal{B}^{\vec{x}}\sim (1)$
and $(\alpha_i)_{i\leq k}$ such that $\mathcal{B}^{\vec{b}_i} = (\alpha_i)$ using Algorithm~\ref{alg:class_group}.
\STATE Find $U\in \mathrm{Gl}_{k\times k}(\Z)$ such that $U\cdot B = H$ in HNF form for $B = (\vec{b}_i)_{i\leq k}$ using~\cite[Prop. 6.3]{Arne}.
\STATE \textbf{for} $i\leq k$, $j\leq l$ \textbf{do} Compute $\beta_i \bmod \pg_j:=\alpha_1^{u_{i,1}}\cdots\alpha_k^{u_{i,k}}\bmod \pg_j$.
\STATE Let $u_1,\cdots,u_r$ be the cyclotomic units of $\Q(\zeta_N)$.
\FOR {$1\leq i\leq m$}
\STATE $I_i\leftarrow \prod_{j\leq m}\pg_j^{h_{i,j}}$.
\STATE Let $(x_j)_{j\leq r}$ be the output of Algorithm~\ref{alg:CVP} solving the BDD on input $I_i,(u_j)_{j\leq r}$.
\STATE $\beta_j\bmod\pg_k\leftarrow u_1^{x_1}\cdots u_{r}^{x_r}\beta_j\bmod \pg_k$ for $j\leq m$, $k\leq l$.
\STATE $\operatorname{Log}(\beta_j)\leftarrow x_1\operatorname{Log}(u_1)+\cdots+ x_r\operatorname{Log}(u_{r}) + \operatorname{Log}(\beta_j)$ for $j\leq m$.
\ENDFOR
\RETURN $H , (\beta_i\bmod \pg_j)_{i\leq m,j\leq l}, (\operatorname{Log}(\beta_j))_{j\leq m}$
\end{algorithmic}
\end{algorithm}
\begin{proposition}[GRH + Heuristic~\ref{heuristic:smoothness_ideals}]
Assume that $B \in 2^{O(n^a)} $ for $a \geq 1/2 $, and that $l\in O ( \log(|\Delta|))$,
$\mathcal{N}(\pg_j)\in O(\log(|\Delta|))$, then the heuristic expected run time of
Algorithm~\ref{alg:precomputation} is in $2^{O(n^a)}$,
and the bit size of the representation of the $\operatorname{Log}(\beta_j)$ is polynomial
in $n$.
\end{proposition}
\begin{proof}
The run time of Algorithm~\ref{alg:precomputation} is clearly dominated by the cost of the search for
relations and the computation of the HNF of the relation matrix (together with the premultipliers).
We need to bound the $\operatorname{Log}(\beta_i)$.
The upper bound on a generator $\beta_i$ of the principal
ideal $I_i = \prod_{j\leq m}\pg_j^{h_{i,j}}$ is essentially given by the
norm of the integral ideal $I_i$. When $i\leq i_0$, $I_i$ has the shape $I_i = \prod_{j\leq i_0}\pg_j^{h_{i_0,j}}$
while when $i > i_0$, $I_i$ is of the form $I_{i} := \pg_{i}\cdot\left( \prod_{j\leq i_0}\pg_j^{h_{i_0,j}}\right)$.
For each $j\leq i_0$, $\mathcal{N}(\pg_j)\leq 12\ln(|\Delta|)^2$ while $\mathcal{N}(\pg_i)\in L_\Delta(a,b)$ if $i > i_0$ and
$h_{i,j}\leq |\operatorname{Cl}(\mathcal{O}_K)|\in O(\sqrt{|\Delta|})$ for $i,j\leq i_0$. Therefore, in any case $\mathcal{N}(I_i)\in 2^{\tilde{O}(|\Delta|)}$,
and $\|\beta_i\|\leq 2^{\tilde{O}(n^{1/2})}\mathcal{N}(I_i)^{1/n}\in 2^{\tilde{O}(|\Delta|)}$. For each $\sigma\in\mathrm{Gal}(K/\Q)$,
$\max_\sigma|\sigma(\beta_i)|\leq \|\beta_i\|\in 2^{\tilde{O}(|\Delta|)}$, and
$\min_\sigma|\sigma(\beta_i)|\geq \frac{|\mathcal{N}(\beta_i)|}{\left(\max_\sigma|\sigma(\beta_i)|\right)^{n-1}} \geq \frac{1}{2^{\tilde{O}(n|\Delta|)}}$.
Therefore, for all $\sigma\in \mathrm{Gal}(K/\Q)$, $|\ln(|\sigma(\beta_i))|\in \tilde{O}(|\Delta|)$, and the representation of
the vector $\operatorname{Log}(\beta_i)$ has a polynomial bit size in $n$.
\end{proof}
\paragraph{\textbf{Decomposing $I$}}
\begin{comment}
More specifically, by spending a precomputation time in $2^{O(n^{2-3c+\varepsilon})}$ for an arbitrarily small $\varepsilon>0$, we can solve the
PIP on input ideals $I$ with $\mathcal{N}(I)\leq 2^{n^b}$ in time $2^{O\left(n^{c}\right)}$ when
\begin{enumerate}
\item $b\leq 7c - 2$.
\item $\frac{2}{5} < c < \frac{1}{2}$.
\end{enumerate}
\end{comment}
The second step of our search for a generator of $I$ satisfying $\mathcal{N}(I)\leq 2^{n^b}$ is to break it down as a product of ideals of
norm less than $ B \in 2^{O(n^{2-3a+\varepsilon})}$ for an arbitrarily small $\varepsilon>0$ in time $2^{O(n^{a+o(1)})}$ where
\begin{enumerate}
\item $b\leq 7a - 2$.
\item $\frac{2}{5} < a < \frac{1}{2}$.
\end{enumerate}
We proceed by a $\qg$-descent procedure similar to that of Section~\ref{sec:q_desc}. Given a
prime ideal $\qg$ such that $\log(\mathcal{N}(\qg)) \leq n^b$,
we look for $\alpha\in\qg$ such that
$(\alpha)/\qg = \prod\qg_i$ where the $\qg_i$ are prime ideals satisfying $\log(\mathcal{N}(\qg_i))\leq n^{b - \varepsilon}$.
We look for small vectors of $\mathcal{L}_I$ with the same definition of $\mathcal{L}_I$ as in Section~\ref{sec:q_desc}
by using the BKZ reduction method with block size $l := n^{a}$ and
by parameterizing $\mathcal{L}_q$ with the degree $k := n^{\min\left\{4a-1,b+2a-1-\varepsilon\right\}}$.
\begin{lemma}\label{lemma:smooth_BKZ}
Let $l\leq k\leq n$. By using a BKZ reduction on $\mathcal{L}_I$, we can find a vector
in $\mathcal{L}_I$ of length less than $l^{k/2l}\mathcal{N}(I)^{\frac{1}{k}}$ in time $2^{O(l)}$.
\end{lemma}
\begin{proof}
The determinant of $\mathcal{L}_I$ is that of the upper left $k\times k$
submatrix of $H$ and satisfies $\det(\mathcal{L}_I) \leq \prod_{i\leq N}v_{i,i} = \mathcal{N}(I)$.
An BKZ-reduction with block length $l$ returns a basis
whose first vector has length less than $l^{k/2l}(\det(\mathcal{L}_I))^{1/k}$ in time $2^{O(l)}$.
\end{proof}
\begin{lemma}\label{lemma:smooth_combinations_BKZ}
In time $2^{O(l)}$, we can find an element $\alpha\in I$
such that $\mathcal{N}(\alpha)\leq l^{\frac{kn}{2l}(1+o(1))}\mathcal{N}(I)^{\frac{n}{k}}.$
\end{lemma}
\begin{proof}
Let $\alpha$ be the first vector of an HKZ-reduced basis of $\mathcal{L}_I$. The calculation of this basis
takes time $2^{O(k)}$ and by Lemma~\ref{lemma:smooth}, the length of its first vector $(\alpha_1,\cdots,\alpha_k)$ is bounded
by $l^{k/2l}\mathcal{N}(I)^{\frac{1}{k}}$.
By the same argument as in the proof of Proposition~\ref{prop:class_group},
the algebraic norm of $\alpha := \sum_i \alpha_i \zeta_N^i$ satisfies
$$\mathcal{N}(\alpha) \leq \left.\sqrt{n}\right.^n \left( \|(\alpha_1,\cdots,\alpha_k)\| \right)^n
\leq \underbrace{n^{n/2} l^{kn/2l}}_{l^{\frac{kn}{2l}(1+o(1))}}\mathcal{N}(I)^{\frac{n}{k}}.$$
\end{proof}
\begin{algorithm}[ht]
\caption{One round of the $\qg$-descent}
\begin{algorithmic}[1]\label{alg:one_round_2}
\REQUIRE Prime ideal $\qg$ with $\log(\mathcal{N}(\qg))\leq n^b$, $\varepsilon > 0$, $a > 0$, and a conductor $N$.
\ENSURE Prime ideals $\pg_i$, integers $e_i$ and $\alpha\in\qg$ such that $(\alpha)/\qg = \prod_i\pg_i^{e_i}$
and $\log(\mathcal{N}(\pg_i))\leq n^{b-\varepsilon}$.
\STATE $S\leftarrow \left\{\pg_i\ \mbox{such that }\mathcal{N}(\pg_i)\leq 12\log(|\Delta|)^2\right\}$.
\WHILE{no relation has been found}
\STATE $(x_i)\xleftarrow[]{R}[0,A]^{|S|}$, $I\leftarrow \qg\prod_i \pg_i^{x_i}$ where $R$ is the uniform distribution.
\STATE Construct a basis for the lattice $\mathcal{L}_I$ with $k := n^{\min\left\{4a-1,b+2a-1-\varepsilon\right\}}$.
\STATE BKZ-reduce $\mathcal{L}_I$ with block size $l :=n^{a}$ and let $\alpha$ be its first vector.
\STATE \textbf{if} $(\alpha)$ is $2^{n^{b-\varepsilon}}$-smooth, find $(\pg_i)$,$(e_i)$ such that $(\alpha) = \prod_i\pg_i^{e_i}$.
\ENDWHILE
\RETURN $\alpha$, $(e_i)$, $(\pg_i)$, $(e_i)$.
\end{algorithmic}
\end{algorithm}
\begin{proposition}[GRH + Heuristic~\ref{heuristic:smoothness_ideals}]
Let $\varepsilon > 0$ be an arbitrarly small constant and let $b,c > 0$ be constants satisfying
\begin{enumerate}
\item $2-3a + 2\varepsilon \leq b \leq 7a - 2$.
\item $\frac{2}{5} + \frac{\varepsilon}{5} \leq a \leq \frac{1}{2}$.
\end{enumerate}
Let $\qg$ be a prime with $\log(\mathcal{N}(\qg))\leq n^b$. When $N = q^s$,
there is a large enough constant $A$ such that Algorithm~\ref{alg:one_round_2} returns a decomposition of
$\qg$ in $\operatorname{Cl}(\mathcal{O}_K)$ as a product of primes $\pg_i$ with
$\log(\mathcal{N}(\pg))\leq n^{b-\varepsilon}$ in time $2^{\tilde{O}(n^{a + o(1) })}$.
\end{proposition}
\begin{proof}
According to Lemma~\ref{lemma:smooth_combinations}, any $\alpha$ derived in Step~4 of Algorithm~\ref{alg:one_round_2}
satisfies $\log(\mathcal{N}(\alpha)) \in O\left( \frac{nk\log(l)}{l}(1+o(1)) + \frac{n}{k}\log(q)\right)$.
As $k\leq n^{4a-1}$ and $l = n^a$, we get
$$\frac{nk}{l}\leq n^{1+4c-1-c} n^{3a}.$$
Likewise since $k = \min\{ n^{4a - 1} , n^{b+2a-1-\varepsilon}\}$, we get
\begin{align*}
\frac{n}{k}\log(q) &\leq n^{1 + b - 4a + 1} \leq n^{3a}\\
\frac{n}{k}\log(q) &\leq n^{1+b-(b+2a-1-\varepsilon)} \leq n^{2(1-a) + \varepsilon} \leq n^{3a}
\end{align*}
The latter inequality follows from the fact that by definition $a\geq \frac{2}{5} + \frac{\varepsilon}{5}$.
In each case, $\log(\mathcal{N}(\alpha)) \in O (n^{3a}(1+o(1)))$, and testing the smoothness of $\mathcal{N}(\alpha)$ with the
Number Field Sieve takes time $2^{O(n^{a+o(1)})}$. As $k\leq n^{b+2a-1-\varepsilon}$, we also have
$$\frac{nk}{l}\leq n^{1+b+2a-1-\varepsilon -a} = n^{a +b-\varepsilon}.$$
In addition, we can show that $k\geq n^{1-a+\varepsilon}$. Indeed, from the definition of $a,b$ we get
\begin{align*}
1-a + \varepsilon &\leq 4a-1 \Leftrightarrow \frac{2}{5} + \frac{\varepsilon}{5} \leq a\\
1-a + \varepsilon &\leq b + 2a -1 -\varepsilon \Leftrightarrow 2-3a + 2\varepsilon \leq b
\end{align*}
Therefore, we have the following inequality
$$\frac{n}{k}\log(q) \leq n^{1 + b -(1-a + \varepsilon)} = n^{a + b - \varepsilon}.$$
This means that $\log(\mathcal{N}(\alpha)) \in O (n^{a + b - \varepsilon}(1+o(1)))$, and from
Heuristic~\ref{heuristic:smoothness_ideals}, the number of $\alpha$ we need to test
before obtaining one such that $(\alpha)/I$ is $2^{n^{b-\varepsilon}}$-smooth is bounded by
$2^{O(n^{a + o(1)})}$. For correctness, we need to check that we always have $l\leq k\leq n$.
First, $k\geq n^{1-a + \varepsilon}$, and
$$1-a + \varepsilon \geq a \Leftrightarrow a \leq \frac{1}{2} + \frac{\varepsilon}{2}$$
\end{proof}
Since $a < \frac{1}{2}$, we must have $k \geq n^a = l$. On the other hand, we have
$k = \min\{ n^{4a - 1} , n^{b+2a-1-\varepsilon}\}$, and $4a - 1 \leq 1 \Leftrightarrow a\leq \frac{1}{2}$,
which is satisfied by definition of $a$. Therefore, we always have $k\leq n$.
\begin{algorithm}
\caption{$\q$-descent}
\begin{algorithmic}[1]\label{alg:full_q_descent_2}
\REQUIRE $I\subseteq \mathcal{O}_K$ with $\mathcal{N}(I)\leq 2^{n^{b_0}}$, $\varepsilon > 0$ and $a$ such that
$b_0\leq 7a - 2$ and $\frac{2}{5}+ \frac{\varepsilon}{5}\leq a \leq \frac{1}{2}$.
\ENSURE Prime ideals $(\qg_i)_{i\leq l}\in\mathcal{B}$ with $\mathcal{N}(\q_i)\leq 2^{n^{2-3a+2\varepsilon}}$, integers $(e_i)$ and $(\phi_j)_{j\leq k}\in K$ such that
$I =\prod_{j\leq k}(\phi_j) \cdot\prod_{i\leq l}\mathfrak{q}_i^{e_i}$.
\STATE $\mathrm{genList}\leftarrow \left\lbrace 1\right\rbrace $, $\mathrm{primeList}\leftarrow \left\lbrace I \right\rbrace$,
$\mathrm{expList}\leftarrow \{1\}$.
\STATE $b\leftarrow b_0$.
\WHILE{$b > 2-3a+2\varepsilon$}
\FOR {$\q\in\mathrm{primeList}$ with $\mathcal{N}(\q)> n^{b-\varepsilon/2}$}
\STATE Find $(\mathfrak{q}_i)_{i\leq l}$, $(e_i)_{i\leq l}$ and $\phi_k$ such that
$\mathfrak{q}=(\phi_k)\prod_{i\leq l}\mathfrak{q}_i^{e_i}$ by Algorithm~\ref{alg:one_round_2}.
\STATE $\mathrm{genList}\leftarrow \mathrm{genList}\cup\left\lbrace\phi_k\right\rbrace$, $\mathrm{primeList}\leftarrow\mathrm{primeList}\cup\left\lbrace \mathfrak{q}_1,\hdots,\mathfrak{q}_l\right\rbrace$.
\STATE $\mathrm{expList}\leftarrow \mathrm{expList}\cup\{e_1,\cdots,e_l\}$.
\STATE Remove $\q$ from $\mathrm{primeList},\mathrm{expList}$.
\ENDFOR
\STATE $b\leftarrow b - \varepsilon$.
\ENDWHILE
\RETURN $\mathrm{genList}$, $\mathrm{primeList}$, $\mathrm{expList}$.
\end{algorithmic}
\end{algorithm}
\begin{proposition}[GRH + Heuristic~\ref{heuristic:smoothness_ideals}]
Algorithm~\ref{alg:full_q_descent} decomposes $I$ with $\mathcal{N}(I)\leq 2^{n^b}$ as an $2^{O(n^{2-3a+2\varepsilon})}$-smooth product in $\operatorname{Cl}(\mathcal{O}_K)$
in time $2^{O\left((n^{a + o(1)}\right)}$ for arbitrarily small $\varepsilon > 0$ and
\begin{enumerate}
\item $2-3a + 2\varepsilon \leq b \leq 7a - 2$.
\item $\frac{2}{5} + \frac{\varepsilon}{5} \leq a \leq \frac{1}{2}$.
\end{enumerate}
\end{proposition}
\paragraph{\textbf{Resolution step}}
Given an input ideal $I$ of $\mathcal{O}_K$, we use the method of Cramer, Ducas, and Wesolowski~\cite{Stickelberger}
to compute a an ideal $J$ with $\mathcal{N}(J)\in 2^{\tilde{O}(n^{1+c})}$ for some $c >0$ such that $IJ$ is principal.
Under the heuristic
used in~\cite{Stickelberger} stating that the class group is generated by the prime ideals above a number of
primes in $O(\log(n))$, we have $c = 1/2$. Given a generator $\alpha$ of $IJ$, the short generator recovery
method of Cramer et al.~\cite{CVP_cyclo} returns an element $\beta\in IJ\subseteq I$ approximating the shortest vector of $I$ by a factor
$e^{\tilde{O}(\sqrt{n})}$. Our ideal decomposition technique is used twice in this resolution step:
\begin{enumerate}
\item To decompose the class of $I$ as a product of a short basis of generators of $\operatorname{Cl}(\mathcal{O}_K)$ in~\cite[Alg. 1 Step 1]{Stickelberger}.
\item To decompose the principal ideal $IJ\subseteq I$ given by~\cite[Alg. 1 Step 1]{Stickelberger} on a short basis of generators
of $\operatorname{Cl}(\mathcal{O}_K)$ and find its generator.
\end{enumerate}
In both cases, we use the fact that under the GRH, the essential part of the HNF of the relation matrix
precomputed has less than $12\ln(|\Delta|)^2$ columns (corresponding to the prime ideals of norm less than
$12\ln(|\Delta|)^2$). The HNF has the form $H = \left(\begin{smallmatrix} H_1 & (0) \\ H_2 & I \end{smallmatrix}\right)$
where $I$ is an identity block corresponding to the relations of the form $\mathfrak{P}\sim \prod_i\pg_i^{e_i}$ where $(-e_i)$
is a row vector of $H_2$, $\mathcal{N}(\pg_i)\leq 12\ln(|\Delta|)^2$ and $\mathfrak{P}\in\mathcal{B}$, $\mathcal{N}(\mathfrak{P}) > 12\ln(|\Delta|)^2$.
Given a decomposition of an input ideal over $\mathcal{B}$, it is straightforward to rewrite all large prime ideals as products of
the ideals of norm less than $12\ln(|\Delta|)^2$. We describe this procedure in Algorithm~\ref{alg:decomp}.
\begin{algorithm}[ht
\caption{Decomposition over a small generating set}
\begin{algorithmic}[1]\label{alg:decomp}
\REQUIRE Hermite form $H = \left(\begin{smallmatrix} H_1 & (0) \\ H_2 & I \end{smallmatrix}\right)$ of the
matrix of relations between primes of $\mathcal{B}=\{\pg\ \mid\ \mathcal{N}(\pg)\leq L_\Delta(a,b)\}$ for some $a,b > 0$,
and input ideal $I$.
\ENSURE $(e_i)$ such that $I\sim \prod\pg_i^{e_i}$ for the $\pg_i$ such that $\mathcal{N}(\pg_i)\leq 12\ln(|\Delta|)^2$.
\STATE Use the $\qg$-descent to find $I\sim \prod_i\qg_i$ with $\qg_i\in\mathcal{B}$.
\STATE $i_0\leftarrow |\mathcal{B}_0|$.
\FOR {$\qg_j$ with $\mathcal{N}(\pg_j) > 12\ln(|\Delta|)^2$}
\STATE Use the corresponding row in $H_2$ to find $\qg_j\sim \prod_{i\leq i_0}\pg_j^{f_i}$.
\STATE Update the decomposition of $I$.
\ENDFOR
\RETURN $(e_i)$ where $I\sim \prod_{i\leq i_0}\pg_i^{e_i}$.
\end{algorithmic}
\end{algorithm}
Algorithm~\ref{alg:decomp} can be used in substitution of Step~1 of~\cite[Alg. 1]{Stickelberger} which starts
by decomposing an ideal over the factor base, and then finds a close vector in the relation lattice
to the ideal annihilating the input ideal in the class group. In~\cite{Stickelberger}, the authors only
considered quantum polynomial time methods to perform this decomposition. Indeed, the best available
classical subexponential solutions~\cite{biasse_subexp,ANTS_XI} (and even the methods described in Section~\ref{sec:q_desc})
did not yield a better trade-off approximation/cost than the BKZ reduction. However,
we can leverage an expensive (but still subexponential) precomputation to achieve a better compromise. The
call to~\cite[Alg. 1]{Stickelberger} must be preceded by a random walk in the Cayley graph of the class group
bringing the input ideal $I$ in $\operatorname{Cl}^{-}(\mathcal{O}_K) = \ker(\mathcal{N}_{K/K^+})$ where $ \mathcal{N}_{K/K^+} : \operatorname{Cl}(\mathcal{O}_K)\rightarrow \operatorname{Cl}(\mathcal{O}_{K^+})$. This consists
of multiplying $I$ by random products of primes of norm in $O(n)$ until it equals an ideal in $\operatorname{Cl}^{-}(\mathcal{O}_K)$. The
norm of $I$ gets multiplied by a factor in $2^{\tilde{O}(n)}$.
\begin{heuristic}\label{heuristic:class_group}
When $K = \Q(\zeta_N)$ for a conductor of the form $N = p^s$, the class number of the maximal subfield of $K$ satisfy $h^+(K)\in O(n)$ and the class group of $\mathcal{O}_K$ is generated by
primes above a number polylogarithmic in $n$ of primes,
\end{heuristic}
\begin{proposition}[GRH + Heuristic~\ref{heuristic:smoothness_ideals}]
When using a precomputation corresponding to Algorithm~\ref{alg:class_group} with
smoothness bound $B\in 2^{O(2-3a+2\varepsilon)}$ for $a,b,\varepsilon > 0$ such that
\begin{enumerate}
\item $2-3a + 2\varepsilon \leq b \leq 7a - 2$,
\item $\frac{2}{5} + \frac{\varepsilon}{5} \leq a \leq \frac{1}{2}$,
\end{enumerate}
the run time of~\cite[Alg. 1 Step 1]{Stickelberger}
on the input ideal $I$ satisfying $\log(\mathcal{N}(I)) \leq n^b$
is in $2^{O\left(n^{a+o(1)}\right)}$.
\end{proposition}
\begin{algorithm}[ht
\caption{$\gamma$-SVP with precomputation for $\gamma\in e^{\tilde{O}(\sqrt{n})}$}
\begin{algorithmic}[1]\label{alg:gamma_SVP}
\REQUIRE Hermite form $H = \left(\begin{smallmatrix} H_1 & (0) \\ H_2 & I \end{smallmatrix}\right)$ of the
relation matrix with $(\alpha_i\bmod \pg_j)_{i\leq k,j\leq l}$, $(\operatorname{Log}(\alpha_i))_{i\leq m}$
where $m = |\mathcal{B}|$ for $\mathcal{B} = \{ \pg\ \mid \ \mathcal{N}(\pg)\leq L_\Delta(a,b)\}$ for
some $a,b$ satisfying $\mathcal{B}^{H_i} = (\alpha_i)\mathcal{O}_K$, and input ideal $I$.
\ENSURE $\beta\in I$ with $\|\beta\|\leq e^{\tilde{O}(\sqrt{n})}\mathcal{N}(I)^{1/n}$.
\STATE $i_0\leftarrow \dim(H_1)$. $\mathcal{B}_0 \leftarrow \{ \pg_1,\cdots,\pg_{i_0}\}$.
\STATE $I\leftarrow II'$ where $\mathcal{N}(I')\in 2^{\tilde{O}(n)}$ and $II'\in\operatorname{Cl}^-(\mathcal{O}_K)$.
\STATE Decompose $I$ over $\mathcal{B}_0$ using Algorithm~\ref{alg:decomp}.
\STATE Find $J$ such that $IJ$ principal using~\cite[Alg. 1]{Stickelberger}. $I\leftarrow IJ$.
\STATE Decompose $I = (\delta) \prod_{i \leq i_0}\pg_i^{e_i}\cdot \prod_{i > i_0}\pg_i^{f_i}$
using Algorithm~\ref{alg:full_q_descent_2}.
\STATE Deduce $(\delta'\bmod \pg_i)$ and $\operatorname{Log}(\delta')$ such that $I = (\delta')\mathcal{B_0}^{(e_i) + \sum_j f_jH_j}$.
\STATE $\vec{y}\leftarrow (e_i) + \sum_j f_jH_j$. Solve $\vec{x}H = \vec{y}$.
\STATE Deduce $(\beta_0\bmod \pg_i)$ and $\operatorname{Log}(\beta_0)$ such that $I = (\beta_0)$.
\STATE Use the methods of~\cite{CVP_cyclo} to derive $(x_i)$ such that $\beta_0\prod_i u_i^{x_i}$ is a short generator of $I$ where
the $u_i$ are the cyclotomic units.
\STATE Compute $\beta_0\prod_i u_i^{x_i}\bmod\pg_j$ for $j\leq l$ and reconstruct $\beta = \beta_0\prod_i u_i^{x_i}$ by the
Chinese Remainder Theorem.
\RETURN $\beta$
\end{algorithmic}
\end{algorithm}
\begin{proposition}[GRH + Heuristic~\ref{heuristic:smoothness_ideals} + Heuristic~\ref{heuristic:class_group}]
When using a precomputation corresponding to Algorithm~\ref{alg:class_group} with smoothness bound
$B \in 2^{O(2-3a+2\varepsilon)}$ for $a,b,\varepsilon > 0$ such that
\begin{enumerate}
\item $2-3a + 2\varepsilon \leq b \leq 7a - 2$,
\item $\frac{2}{5} + \frac{\varepsilon}{5} \leq a \leq \frac{1}{2}$,
\end{enumerate}
on the input ideal $I$ satisfying
$\log(\mathcal{N}(I))\leq n^b$, the run time of Algorithm~\ref{alg:gamma_SVP}
is in $2^{O(n^{a +o(1)})}$, and it
returns a solution to $\gamma$-SVP for $\gamma\in e^{\tilde{O}(\sqrt{n})}$.
\end{proposition}
For example, with $c= \frac{3}{7}$ and $b=1+o(1)$, if an entity spends a precomputation cost in $2^{(n^{5/7 + \varepsilon})}$ for
an arbitrarily small $\varepsilon > 0$, then
all subsequent instances of $\gamma$-SVP in ideals $I$ of $\mathcal{O}_K$ satisfying $\log(\mathcal{N}(I))\leq n^{1+o(1)}$ for $\gamma \in e^{\tilde{O}(\sqrt{n})}$ will
take heuristic time in $2^{O(n^{3/7+o(1)})}$. In particular, the public keys of the multilinear maps
of Garg, Gentry and Halevi satisfy the requirement on the norm of the input. This means that with a precomputation of cost $2^{(n^{5/7 + \varepsilon})}$,
our key recovery attack takes heuristic time $2^{O(n^{3/7+o(1)})}$. The attack
with precomputation can have to main scenarios:
\begin{itemize}
\item An entity precomputes the data and the attacker downloads it. In this case the storage required is proportional
to the expected time of the attack, that is $2^{O(n^a)}$.
\item The attacker is allowed to query the entity having done the precomputation. In this case, the entity sends the attacker
the significantly smaller matrix corresponding to the relations between the short ideals before hand, and for each challenge, the
attacker asks for the decomposition of the large primes in the decomposition of the ideal given as input to the $\qg$-descent.
\end{itemize}
\paragraph{\textbf{Remark}}
The precomputation can go on after the computation of the HNF of the large relation matrix. The entity
performing the precomputation can continue creating new relations by performing $\qg$-descents on the large
primes which are not already in the factor basis $\mathcal{B}$. The HNF of the relation matrix can be updated at
a minimal cost.
\begin{comment}
\section{An algorithm to solve the PIP in time $2^{N^{1/2+\varepsilon}}$}\label{sec:PIP}
We solve the Principal Ideal Problem by first finding a generating set for all the relations of the form
$(\alpha_i) = \prod_j\pg_j^{m_{i,j}}$ where $\pg_j\in\mathcal{B}$. Then we find an extra relation
of the form $(\alpha)\ag = \prod_j\pg_j^{y_j}$ (where $\ag$ is the input ideal). If $\ag$ is principal,
this relation must be a product of the relation
previously found. This strategy is very classic, and the ingredient that allows it to run in polynomial
time is the use of Algorithm~\ref{alg:full_q_descent} to derive relations. During the computation of the
lattice of relations, we compute the ideal class group $\operatorname{Cl}(\mathcal{O}_K)$. It is not \textit{sensu stricto}
necessary because all we need to check is that the determinant of the lattice generated by the relations
found at a given time is the class number $|\operatorname{Cl}(\mathcal{O}_K)|$. However, the extra step consisting of deriving the
class group does not increase the overall complexity, and since our algorithm is the first polynomial time
method allowing the computation of $\operatorname{Cl}(\mathcal{O}_K)$ in an infinite class of number fields, we chose to present it
as well.
\paragraph{Ideal class group computation}
Let $\mathcal{B}:= \{ \pg\subseteq\mathcal{O}_K \mbox{ such that }\mathcal{N}(\pg)\leq 2^{N^{1/2}}\}$,
and
$$\begin{CD}
\Z^m @>{\varphi}>> \mathcal{I} @>{\pi}>> \operatorname{Cl}(\mathcal{O})\\
(e_1, \ldots, e_l) @>>> \prod_i\p_i^{e_i} @>>> \prod_i[\p_i]^{e_i}
\end{CD}.$$
Under the generalized Riemann Hypothesis (GRH), the classes of ideals in $\mathcal{B}$ generate
$\operatorname{Cl}(\mathcal{O}_K)$ (see (see~\cite[Th. 4]{BBounds})), which means that $\pi\circ\varphi$ is
surjective and $\operatorname{Cl}(\mathcal{O})\simeq \Z^m/\ker(\pi\circ\varphi).$ Therefore, computing the class group boils down to
computing $\ker(\pi\circ\varphi)$, which is given by the lattice $\mathcal{L}$ of $(e_1,...,e_l)\in\Z^m$ such that
$$\p_1^{e_1},\hdots , \p_m^{e_m} = (\alpha),$$
where $\alpha\in\mathcal{O}_K$. We collect many relations of the form
$\prod_i\p_i^{2^{(j)}_i} = (\alpha_j)$ and put them in the rows of the relation matrix $M := (e^{(j)}_i)$.
To find a generating set for $\mathcal{L}$, we compute the Hermite Normal Form (HNF) of $M$, that is, a unimodular
matrix $U$ such that $UM = \left( \frac{H}{(0)}\right)$, where $H\in\Z^{l\times l}$ is lower triangular. The computation
of the HNF of an integer matrix is a well studied problem (see for example~\cite{Arne}), and has polynomial complexity.
During the collection of random relations, we can check the rank of the current relation matrix, and once it has full
rank, we can check the determinant $d$ of the lattice spanned by its rows by computing its HNF. If $d = |\operatorname{Cl}(\mathcal{O}_K)|$,
we have a generating set for $\mathcal{L}$. To do so, we use an approximation of the Euler product to certify the result:
$hR = \frac{|\mu|\sqrt{|\Delta|}}{2^{r_1}(2\pi)^{r_2}}\lim_{s\rightarrow 1} \left( (s-1)\zeta_\K (s)\right),$
where $h = |\operatorname{Cl}(\mathcal{O}_K)|$, $R$ is the regulator of $K$, $\zeta_\K (s) = \sum_\ag \frac{1}{\mathcal{N}(\ag)^s}$
is the usual $\zeta$-function associated to $K$ and
$|\mu|$ is the cardinality of $\mu$ the group of torsion units. Indeed, it
allows us to derive a bound $h^*$ in polynomial time under GRH that satisfies $h^* \leq hR < 2h^*$ (\cite{BachEulerProd}).
If the tentative class number and regulator do not satisfy this, we collect more relations. Normally, the computation
of the ideal class group is interleaved with the unit group computation (since we need $R$ to certify the result), however,
in the special case of $K = (\Q\zeta_{2^{n}})$, we already know a set of fundamental units $\gamma_i = \frac{\zeta_N^i - 1}{\zeta_N-1}$,
and we can readily calculate a polynomial approximation of the regulator. Eventually, the Smith Normal Form (SNF) of $H$
yields the group structure of $\operatorname{Cl}(\mathcal{O}_K)$. Indeed, there are unimodular matrices $U,V$ such that $UHV = \operatorname{diag}(d_1,\cdots,d_l)$,
and the corresponding diagonal coefficients satisfy
$$\operatorname{Cl}(\mathcal{O}_K) \simeq \Z / d_1\Z \times \cdots \times \Z/d_l\Z.$$
\begin{algorithm}[ht]
\caption{Class group of $\mathcal{O}_K$}
\begin{algorithmic}[1]\label{alg:cl_grp}
\REQUIRE $K$, $\mathcal{B} = \{ \pg_i,\ \mbox{such that }\mathcal{N}(\pg_i)\leq B\}$ that generates $\operatorname{Cl}(\mathcal{O})$ .
\ENSURE $\operatorname{Cl}(\mathcal{O}_K)$ and a generating set of the lattice $\mathcal{L}\subseteq\Z^l$ of relations.
\STATE $M\leftarrow \mbox{empty matrix with }l\mbox{ columns}$, $k\leftarrow 1$. $L\leftarrow \mbox{empty list}$.
\FOR {$i\leq l$}
\STATE Call Algorithm~\ref{alg:full_q_descent} on $\pg_i$, $K'$ and $\mathcal{B}$. Let $\alpha,(e_j)_{j\leq l}$ such
that $(\alpha)\pg_i=\prod_j\pg_j^{e_j}$.
\STATE Append $(e_1,\cdots,e_{i-1},e_i-1,e_{i+1},\cdots,e_l)$ to $M$. $L\leftarrow L\cup\{\alpha\}$. $k\leftarrow k+1$.
\ENDFOR
\STATE If $M$ does not have full rank, \textbf{return} failure.
\STATE $m\leftarrow \frac{2\log|\Delta|}{\log(2/a)}$ where $a = 1 + O(k/|\Delta|)$.
\STATE Derive $m$ random relations $(\alpha_k) = \prod_i\pg_j^{e_j}$ by calling
Algorithm~\ref{alg:full_q_descent} on $\ag = \mathcal{O}_K$, $K'$ and $\mathcal{B}$.
\STATE Compute the regulator $R$ of $K$ by using the cyclotomic units $\gamma_i = \frac{\zeta_N^i-\zeta_N}{\zeta_N-1}$.
\STATE Compute $U$ such that $UM = \frac{(H)}{(0)}$ where $H$ is in Hermite Normal Form. $h\leftarrow \det(H)$.
\STATE If $h R$ is not close enough to
$\frac{|\mu|\sqrt{|\Delta|}}{2^{r_1}(2\pi)^{r_2}}\lim_{s\rightarrow 1} \left( (s-1)\zeta_\K (s)\right)$ \textbf{return} failure.
\STATE $A\leftarrow \mbox{empty matrix with }l\mbox{ columns}$, $L'\leftarrow \mbox{empty list}$.
\FOR {$i\leq l$}
\STATE Append the $i$-row of $H$ to $A$.
\STATE $\gamma_i\leftarrow \prod_j\alpha_j^{u_{i,j}}$, $L'\leftarrow L'\cup\{\gamma_i\}$
\ENDFOR
\STATE Compute the SNF $D = \operatorname{diag}(d_1,\cdots,d_l)$ of $H$.
\RETURN $A$, $L'$, $\prod_i\Z/d_i\Z$.
\end{algorithmic}
\end{algorithm}
\noindent The run time of Algorithm~\ref{alg:cl_grp} depends on our capacity to derive relations in $\operatorname{Cl}(\mathcal{O}_K)$ almost
uniformly at random. Most subexponential algorithms relying on a relation search in an ideal class group
rely on the heuristic that the relations are chosen uniformly at random. Algorithm~\ref{alg:full_q_descent}
use random power-products in $\operatorname{Cl}(\mathcal{O}_K)$, but the fact that we use elements in a subfield makes it complicated
to derive statements on the distribution of the relations in $\operatorname{Cl}(\mathcal{O}_K)$ that we derive using Algorithm~\ref{alg:full_q_descent}.
We therefore rely on the standard heuristic that the randomization we induce is enough for our purposes.
\begin{heuristic}\label{heuristic:random}
The relations obtained by calling Algorithm~\ref{alg:full_q_descent} are chosen uniformly at random in $\operatorname{Cl}(\mathcal{O}_K)$.
\end{heuristic}
\noindent Heuristic~\ref{heuristic:random} relies in fact on a very classical heuristic which consists of
saying that if a power-product $\prod_i\pg_i^{e_i}$ of primes from $\mathcal{B}$ is chosen such that
the exponent vector is larger than the diameter of the relation lattice $\Lambda$, and if the LLL reduced
ideal
$(\alpha)\prod_i\pg_i^{e_i} = \ag = \prod_j\pg^{f_j}$ is smooth too, then the relation $(e_i-f_i)_{i\leq m}$
is a uniform random choice of a relation in $\Lambda$. This heuristic is used to analyze most of the class group
algorithms, and it can be further refined by randomizing the choice of an LLL-reduced element. The main particularity
of our heuristic is that we make that assumption twice. The first time, we reduce a random power product
$\prod_i\pg_i^{e_i}\sim \ag$, but $\mathcal{N}(\ag)\approx 2^{n^2}\sqrt{|\Delta|}$, which according to Section~\ref{sec:smoothness}
leaves us no chance to find a smooth $\ag$ (and thus a relation). Although we may not be able to find it efficiently,
let $(f_i)$ be an exponent vector chosen at random such that $\prod_i\pg_i^{e_i}\sim \prod_j\pg^{f_j} =: \ag$. According
to the usual heuristic, $u:= (e_i-f_i)_{i\leq m}$ is a random choice of a relation in $\Lambda$. Unfortunately,
we cannot compute it. But we can find another relation involving $\ag$ by decomposing it as $\ag = (\alpha)\prod_i\pg_i^{g_i}$
by using Algorithm~\ref{alg:full_q_descent}. As $\alpha$ lies in a subfield $K'$ of $K$. the relation $v:= (f_i-g_i)_{i\leq n}$
arises as the lift of a
relation in $\operatorname{Cl}(\mathcal{O}_{K'})$. As by assumption $v$ is a random choice of a relation in $\operatorname{Cl}(\mathcal{O}_{K'})$, $v$
is a random choice of a relation in
$\Lambda'\subseteq \Lambda$, which we define as the lattice of relations in $\operatorname{Cl}(\mathcal{O}_K)$ arising as the lift of a
relation in $\operatorname{Cl}(\mathcal{O}_{K'})$. The relations found by Algorithm~\ref{alg:full_q_descent} are
of the for $w = u + v$ where $u$ is chosen at random in $\Lambda$ and $v$ is chosen at random in $\Lambda'$.
Note that the presence of subfields is usually bad for the randomness of the relation search. For example, there is
an algorithm of the Pari~\cite{pari}
software that searches for small $\alpha$ in a prime ideal $\pg$ or a short power-product $\prod_i\pg_i$. When
$(\alpha)$ is $\mathcal{B}$-smooth, we obtain a relation of the form $(\alpha) = \prod_i\pg_i^{e_i}$. However,
when $\alpha$ is in a subfield $K'$, this relation is only a relation in $\operatorname{Cl}(\mathcal{O}_{K'})$ (which is valid in $\operatorname{Cl}(\mathcal{O}_K)$).
The difference there is that there is no prior reduction of the already short ideal. Indeed, dealing with an ideal
of norm $2^{n^2}|\Delta|$ slows down the search for short elements by enumeration techniques.
\begin{proposition}[GRH and Heuristic~\ref{heuristic:random}]
Let $K = \Q(\zeta_{2^n})$, $N = 2^n$, and
$$\mathcal{B} := \{ \pg\subseteq\mathcal{O}_K \mbox{ such that } \mathcal{N}(\pg)\leq 2^{N^{1/2}}\},$$
then Algorithm~\ref{alg:cl_grp} runs in time $2^{N^{1/2+\varepsilon}}$, and output a correct or failure with probability
less than $1 - 1/N$.
\end{proposition}
\begin{proof}
The correctness comes from the fact that the relation matrices $M,A$ and the list
$L,L'$ carry valid relations of the form $(\alpha_i) = \prod_j\pg_j^{m_{i,j}}$ and
$(\gamma_i) = \prod_j\pg_j^{a_{i,j}}$, and that Step~11 checks if enough relations were
found to generate $\operatorname{Cl}(\mathcal{O}_K)$. The guarantee on the run time is due to the fact that we call
Algorithm~\ref{alg:full_q_descent} a polynomial number of times. Finally, we follow the
same argument (assuming the randomness of the relations under Heuristic~\ref{heuristic:random})
as in~\cite{hafner}[Sec. 3.2] to
prove the success probability.
\end{proof}
\paragraph{Solving the Principal Ideal Problem}
Given the factor base
$\mathcal{B} := \{ \pg\subseteq\mathcal{O}_K \mbox{ such that } \pg\cap\Q=(p) \mbox{ with } p\leq 12\log^2|\Delta|\},$
of size $l$, Algorithm~\ref{alg:cl_grp} provides us with a matrix $M\in\Z^{l\times l}$ and a list $(\alpha_i)_{i\leq l}$
such that $\forall i\leq l (\alpha_i) = \prod_j\pg_j^{m_{i,j}}$. Moreover, the rows of $M$ generate the lattice
$\mathcal{L}\subseteq \Z^l$ of all the vectors $(e_i)_{i\leq l}$ such that $\prod_i \pg_i^{e_i}$. Given an input
ideal $\ag\subseteq \mathcal{O}_K$, we first need to find an extra relation
$$(\alpha)\ag = \pg_1^{y_1}\cdots\pg_l^{y_l}.$$
It is clear that $\ag$ is principal if and only if $\pg_1^{y_1}\cdots\pg_l^{y_l}$ is principal as well, and in addition,
if $\pg_1^{y_1}\cdots\pg_l^{y_l} = (\beta)$, then $\ag = (\alpha^{-1}\beta)$. Therefore, once $\ag$ has been decomposed
over the factor base, our problem boils down to finding a generator (if it exists) of the ideal $\prod_j\pg_j^{y_j}$. But
the rows of $M$ generate all possible relations between elements in $\mathcal{B}$, and the vector $(y_1,\cdots,y_l)$
must belong to $\mathcal{L}$. In particular, we can solve the linear system $XM = Y$, and the coefficients of $X$
give us
$$\pg_1^{y_1}\cdots\pg_l^{y_l} = (\alpha_1^{x_1}\cdots\alpha_l^{x_l})\mathcal{O}_K.$$
This solves our problem because $\ag = (\alpha^{-1}\prod_i\alpha_i^{x_i})\mathcal{O}_K$.
\begin{algorithm}[ht]
\caption{Principal Ideal Problem}
\begin{algorithmic}[1]\label{alg:PIP2}
\REQUIRE Ideal $\ag\subseteq\mathcal{O}$, factor base $\mathcal{B} =(\pg_j)_{j\leq l}$ and a subfield $K'\subseteq K$.
\ENSURE $\mathrm{false}$ or $\alpha$ such that $\ag = (\alpha)\mathcal{O}_K$.
\STATE Use Algorithm~\ref{alg:cl_grp} with $\mathcal{B}$ and $K'$ to get $L=(\alpha_i)_{i\leq l}$ and $M$ such that
$(\alpha_i)=\prod_j\pg_j^{m_{i,j}}$.
\STATE Use Algorithm~\ref{alg:full_q_descent} with $K'$, $\ag$ and $\mathcal{B}$ to find $Y$ and $\alpha$ such that
$(\alpha)\ag = \prod_j\pg_j^{y_j}$.
\STATE Solve the linear system $XM = Y$. If it has no solution \textbf{return} $\mathrm{false}$
\RETURN $\alpha^{-1}\alpha_1^{x_1}\cdots \alpha_l^{x_l}$.
\end{algorithmic}
\end{algorithm}
\begin{proposition}[GRH and Heuristic~\ref{heuristic:random}]
Let $K = \Q(\zeta_{2^n})$, $N = 2^n$, $\ag\subseteq\mathcal{O}_K$ and
$$\mathcal{B} := \{ \pg\subseteq\mathcal{O}_K \mbox{ such that } \mathcal{N}(\pg)\leq 2^{N^{1/2}}\},$$
then Algorithm~\ref{alg:PIP2} runs in time $N^{1/2+\varepsilon}$ and polynomial in $\log(\mathcal{N}(\ag))$, and output a correct or failure with probability
less than $1 - 1/N$.
\end{proposition}
\section{A reduction from the short-PIP to the PIP}\label{sec:short_PIP}
In Section~\ref{sec:PIP}, we showed how to find a generator $\alpha\in K$ of an ideal $\ag\subseteq \mathcal{O}_K$ in heuristic
polynomial time. Campbel, Groves and Shepherd~\cite{GCHQ}[Sec. 3.1] mentioned that there was a polynomial time
reduction of the problem of finding a short generator of the ideal to the problem of finding an arbitrary one. Their
claim was not supported by any proof, but after exchanging with them, Cramer, Ducas, Peikert and Regev agreed on its
validity and posted a preliminary sketch of the proof online~\cite{CVP_cyclo}. In this section, we give details
on how to reduce the short-PIP to the PIP in polynomial time. Our goal is not to take credit for this method, as most of it
is due to Campbel, Groves and Shepherd for the original idea, and to Cramer, Ducas, Peikert and Regev for its further
dissemination. However, we wish to present a self contained description of our attack, and to this date, there is no proof
of the reduction from short-PIP to the PIP publicly available.
\paragraph{The distribution of the secret key}
An important aspect of the efficient reduction from the short-PIP to the PIP is that the generator $g$ of the ideal $\ag$
is relatively short. This means that given the arbitrary generator $\alpha$ of $\ag$ found by using Algorithm~\ref{alg:PIP2},
$\operatorname{dist}(\operatorname{Log}(\alpha),\operatorname{Log}(\mathcal{O}_K^*))\leq \|\operatorname{Log}(g)\|$ is small. Finding a short generator of $\ag$ consists of finding
$u\in\mathcal{L}=\operatorname{Log}(\mathcal{O}_K^*)$ which minimizes $\|\operatorname{Log}(\alpha)-u\|$. This is an instance of the Closest Vector Problem,
and the knowledge of the proximity of the target vector $\operatorname{Log}(\alpha)$ to the lattice helps its resolution. In~\cite{mult_maps},
the secret $g$ is sampled on the power basis (that is, as vectors of coefficients in $\Z^N$) according to a spherical Gaussian
distribution of average 0 and standard deviation $\sigma_0$. To derive statements on $\operatorname{Log}(g)\in\R^r$, we first observe that
according to~\cite{ducas_durmus}[Th. 5], the distribution of the Minkowski embedding of $g$ is also spherical centered around 0.
Then, we consider each coordinate $x\in R$ independently as a Gaussian centered in 0 with
standard deviation $\sigma$, and study the distribution of $\log|x|$.
\begin{proposition}\label{prop:distribution_g}
Assume that $X$ is a random variable of probability density function $\rho_0 := \frac{1}{\sqrt{2\pi\sigma}}e^{\frac{-x^2}{2\sigma}}$, then
the random variable $\log|X|$ has probability density function
$$\rho(x) := \frac{2}{\sqrt{\pi}} \exp\left[ (x-\log(\sqrt{2\sigma})) - \exp\left( 2(x-\log(\sqrt{2\sigma})\right)\right].$$
In particular, its average is in $O(\log(\sqrt{2\sigma}))$ and its standard deviation is in $O(1)$.
\end{proposition}
\begin{proof}
To find the probability distribution of $\log|X|$, we search for $\rho$ such that
$P(\log|X|\geq a) = \int_a^{\infty}\rho(x)dx$. We already know that
$P(|X|\geq t) = \frac{2}{\sqrt{2\pi\sigma}}\int_t^{\infty}e^{\frac{-x^2}{2\sigma}}dx$. This means that
\begin{align*}
P(\log|X|\geq a) &= \sqrt{\frac{2}{\pi\sigma}}\int_{e^a}^{\infty} e^{\frac{-x^2}{2\sigma}}dx\\
&= \frac{2}{\sqrt{\pi}} \int_{\frac{e^a}{\sqrt{2\sigma}}}^{\infty} e^{-u^2} du\ \mbox{with } x = \sqrt{2\sigma}u\\
&= \frac{2}{\sqrt{\pi}} \int_{a - \log(\sqrt{2\sigma})}^{\infty} e^{v-e^{2v}} dv\ \mbox{with } u = e^v\\
&= \frac{2}{\sqrt{\pi}} \int_{a}^{\infty} e^{(w-\log(\sqrt{2\sigma})-e^{2(w-\log(\sqrt{2\sigma}))}} dw\ \mbox{with } w = v+\log(\sqrt{2\sigma}).
\end{align*}
This gives us the probability density function of $\log|X|$. To compute its average and standard deviation, we rely on a result on a generalized
Gumbel distribution~\cite{Gumbel} which states that for $k>0$
$$ \frac{1}{\Gamma(k)} \int_\R e^{kx - e^{x}}dx = 1.$$
Our probability density function $\rho(x)$ is not a Gumbel distribution per se, but we can easily derive its first and second order
moments from that of the generalized Gumbel distribution for $k = 1/2$. Let us start with the average.
\begin{align*}
\int_\R x\rho(x)dx &= \frac{2}{\sqrt{\pi}}\int_\R (x + \log(\sqrt{2\sigma}))e^{x - e^{2x}}dx\\
&= \frac{1}{\sqrt{\pi}}\int_\R \left(\frac{u}{2}+\log(\sqrt{2\sigma})\right) e^{\frac{u}{2}-e^u}du\ \mbox{with } x = \frac{u}{2}\\
&= \underbrace{\frac{1}{2\sqrt{\pi}}\int_\R ue^{\frac{u}{2}-e^u}du}_{b\in O(1)}
+ \log(\sqrt{2\sigma})\underbrace{\frac{1}{\sqrt{\pi}}\int_\R e^{\frac{u}{2} - e^u}du}_{1}.
\end{align*}
The average of $\log|X|$ is therefore $\mu = \log(\sqrt{2\sigma}) + b\in O(\log(\sqrt{2\sigma}))$. Its standard deviation
is given by
\begin{align*}
\sigma_\rho^2 &= \int_\R (x-\mu)^2\rho(x)dx\\
&= \frac{2}{\sqrt{\pi}} \int_\R (x-a-\log(\sqrt{2\sigma}))^2e^{x - \log(\sqrt{2\sigma})-2^{x-\log(\sqrt{2\sigma})}}dx\\
&= \frac{2}{\sqrt{\pi}} \int_\R (u-a)^2 e^{u - 2^{2u}}du \in O(1).
\end{align*}
\end{proof}
\paragraph{A good basis of $\mathcal{L} = \operatorname{Log}(\mathcal{O}_K^*)$}
Solving the Closest Vector Problem, even with the promise that the target vector $\operatorname{Log}(\alpha)$ is close to the
lattice $\mathcal{L}$ (thus turning it into an instance of the Bounded Distance Decoding problem) can be computationally
difficult. The lattice of the logarithms of the Archimedean embeddings is generated by the vectors corresponding
to
$$\zeta_a := \zeta^{(1-a)/2}\frac{1 - \zeta^a}{1-\zeta} = \pm \frac{\sin(\pi a /2^n)}{\sin(\pi/2^n)},\ \mbox{for } 2\nmid a,\ 1<a<2^{n-1},\ \zeta = e^{2j\pi/2^{n}}.$$
When $N$ is a power of 2, the $\operatorname{Log}(\zeta_a)$ are actually a basis for $\mathcal{L}$, and as originally observed by
Campbel, Groves and Shepherd~\cite{GCHQ}, it is a very good one to solve an instance of the Bounded Distance Decoding
problem. To facilitate our analysis, we use the slightly modified basis given by
\begin{align*}
\beta_a &\leftarrow \zeta_b\cdot \zeta_a^{-1} = -\zeta^a \frac{\sin(3\pi a /2^n)}{\sin(\pi a/2^n)}\ \mbox{where } b = 3a\bmod 2^{n}, 3a<2^{n-1}.\\
\beta_a &\leftarrow \zeta_b\cdot \zeta_a^{-1} = -\zeta^{-a} \frac{\sin(3\pi a /2^n)}{\sin(\pi a/2^n)}\ \mbox{where } b = -3a\bmod 2^{n}, 3a\geq2^{n-1}.\\
\end{align*}
The $\|\operatorname{Log}(\beta_a)\|$ are all close to the $n$-th root of the volume of $\mathcal{L}$, and they
are consequently close to being orthogonal. Their Gram Schmidt Orthogonalization (GSO) inherits their
good properties, thus allowing an efficient resolution of the BDD problem. To prove this, we start by
showing that for $b\neq a$, $\|\operatorname{Log}(\beta_a)\|\in O(\|\operatorname{Log}(\beta_a)\|)$. Then we show that their common
asymptotic value is close to the $n$-th root of the volume of $\mathcal{L}$.
\begin{proposition}\label{prop:size_units}
The Euclidean norm of the $\operatorname{Log}(\beta_a)$ satisfies
$$\forall a,b,\ \|\operatorname{Log}(\beta_b)\| \in O(\|\operatorname{Log}(\beta_a)\|) \in O\left(\sqrt{ \int_{1/n}^{\pi/3 - 1/n} \log^2\left( \left| \frac{\sin(3x)}{\sin(x)}\right|\right)}\right).$$
\end{proposition}
\begin{proof}
The vectors $\operatorname{Log}(\beta_a)$ are of the form $(\log|\sigma_i(\beta_a)|)_{i\in\Z_N^*,i<N/2}$. Each $\sigma_i\in G = \mathrm{Gal}(\Q(\zeta_N)/\Q)$
acts via the operation $\zeta\mapsto\zeta^i$. In particular
$$\sigma_i(\beta_a) = -\zeta^{ia} \frac{\sin(3ai\pi/2^n)}{\sin(ai\pi/2^n)}\mbox{ for } 3b<2^{n-1}\mbox{ and }
\sigma_i(\beta_a) = -\zeta^{-ia} \frac{\sin(3ai\pi /2^n)}{\sin(ai\pi a/2^n)}\mbox{ for } 3b\geq 2^{n-1}.$$
This means that $\|\operatorname{Log}(\beta_a)\|^2 = \sum_i \log^2|\sigma_i(\beta_a)| = \sum_i\log^2\left( \left| \frac{\sin(3ai\pi /2^n)}{\sin(ai\pi/2^n)} \right|\right)$.
As $a$ is invertible modulo $2^{n}$, the $ai\bmod 2^{n}$ take all possible values modulo $2^{n}$. We can therefore rearrange the terms of the sum and get
$$\|\operatorname{Log}(\beta_a)\|^2 \in O\left( \sum_i\log^2\left( \left| \frac{\sin(3i\pi /2^n)}{\sin(i\pi/2^n)} \right|\right)\right) \in
O\left( \int_{1/n}^{\pi/3 - 1/n} \log^2\left( \left| \frac{\sin(3x)}{\sin(x)}\right|\right)\right).$$
\end{proof}
\noindent We just established that the Euclidean norm of the basis vectors of $\mathcal{L}$ was very homogeneous. To make it a very good basis
to solve the BDD problem, we also want their common asymptotic size to be close to the $r$-th root of the volume of $\mathcal{L}$.
For this, we rely on a result of~\cite{Washington} showing how the eigenvalues of the matrix whose rows are the $\operatorname{Log}(\beta_a)$
relate to $L(\chi,1)$ for $\chi$ in the character group of $G = \mathrm{Gal}(\Q(\zeta_{2^n})/\Q)$.
\begin{proposition}\label{prop:size_eigenvalues}
The size of the basis vectors satisfies $\|\operatorname{Log}(\beta_a)\| \in O(\mathrm{Vol}(\mathcal{L})^{1/r})$ where $r$ is the rank of $\mathcal{L}$.
\end{proposition}
\begin{proof}
According to the proof of~\cite{Washington}[Lem. 5.26], the eigenvalues of the matrix whose rows are the $\operatorname{Log}(\beta_a)$ are of the form
$\sum_{\sigma\in G}\chi(\sigma)\log(|1-\zeta^\sigma|) = -\frac{\tau}{2}(\chi)L(\chi,1)$, where $L$ is the Dirichlet $L$-function, $\chi$ is a Dirichlet
character of the character group $\widehat{G}$ of $G = \mathrm{Gal}(\Q(\zeta_{2^n})/\Q)$ and $\tau(\chi)$ is the Gauss sum of $\chi$. All the characters
of $\widehat{G}$ have the same conductor $f_\chi$, and according to~\cite{Washington}[Lem. 4.8], $|\tau(\chi)| = \sqrt{f_\chi}$. Also, the values of the
$L$-function at different $\chi$ are related by $L(\chi,1) = L(\tilde{\chi},1)\prod_{p\mid f_\chi}\left( 1 - \frac{\chi(p)}{p}\right)$,
where $\tilde{\chi}$ is a primitive character. Our
conductor is a power of $2$, therefore for each $\chi_1\neq \chi_2$, $|L(\chi_1,1)|\in O(|L(\chi_2,1))|$. Moreover,
$$\pm \prod_{\chi\neq 1} -\frac{\tau(\chi)}{2}L(\chi,1) = h^+ R^+ \in O(\mathrm{Vol}(\mathcal{L})).$$
Therefore, each eigenvalue $\lambda_i$ of the matrix $M$ whose rows are the $\operatorname{Log}(\beta_a)$ satisfies
$\lambda_i\in O\left( \mathrm{Vol}(\mathcal{L})^{1/r}\right)$. Moreover, the Euclidean norm of the rows of $M$ is related to
the eigenvalues of $M$ by
$$\sum_i |\lambda_i|^2 \leq \|M\|_F^2 = \sum_a \|\operatorname{Log}(\beta_a)\|^2 \leq r \max_i |\lambda_i|^2,$$
where $\|M\|_F$ is the Frobenius norm of $M$.
Therefore, for each $a$, $\|\operatorname{Log}(\beta_a)\|\in O(\mathrm{Vol}(\mathcal{L})^{1/r})$.
\end{proof}
\paragraph{Solving the BDD problem}
The basis we chose for $\mathcal{L}$ has a very regular structure. Each vector satisfies $\|\operatorname{Log}(\beta_a)\|\in O(\mathrm{Vol}(\lambda)^{1/r})$.
We solve the BDD problem by using Babai nearest plane algorithm. It relies heavily on the size of the Gram Schmidt Orthogonalization (GSO) of
our basis for $\mathcal{L}$. Let $(b^*_a)_{a<2{n-1}}$ be the GSO of $(\operatorname{Log}(\beta_a))_{a<2^{n-1}}$. The vectors of the GSO
satisfy $\|b^*_a\|\leq \|\operatorname{Log}(\beta_a)\|\in O(\mathrm{Vol}(\lambda)^{1/r})$ and
$\prod \|b^*_a\| = \mathrm{Vol}(\mathcal{L}$. Therefore, the size of the smallest one satisfies
$\|b^*_{a_0}\|\in O(\mathrm{Vol}(\mathcal{L})^{1/r}) \in O(\sqrt{N})$. Meanwhile, we know that
the lattice point we are looking for is within a ball of small. Therefore, Babai's nearest
plane algorithm return the closest vector in $\mathcal{L}$ to the target vector $\operatorname{Log}(\alpha)$ where
$\alpha$ is the arbitrary generator found by using Algorithm~\ref{alg:PIP2}.
\begin{algorithm}[ht]
\caption{Babai's nearest plane algorithm}
\begin{algorithmic}[1]\label{alg:babai}
\REQUIRE Target vector $\operatorname{Log}(\alpha)$ where $\alpha$ generates $\ag$.
\ENSURE The shortest generator $g$ of $\ag$.
\STATE $\operatorname{Log}(\alpha)\leftarrow \operatorname{Log}(\alpha) - \frac{\log(|\mathcal{N}(\alpha)|)}{r}(1,\cdots,1)$.
\STATE $\beta_a\leftarrow \zeta_b\zeta_a^{-1}$.
\STATE Let $(b^*_a)_{a < 2^{n-1}}$ be the GSO of $(\operatorname{Log}(\beta_a))_{a<2^{n-1}}$.
\STATE Let $(a_1,\cdots,a_r)\in\R^{r}$ such that $\operatorname{Log}(\alpha) = \sum_i a_i b^*_{\sigma_i}$.
\FOR {$i\leq r$}
\STATE $g_i\leftarrow \lceil a_i\rfloor$.
\ENDFOR
\RETURN $\alpha\prod_i \beta_i^{-g_i}$.
\end{algorithmic}
\end{algorithm}
\begin{proposition}
When $g$ is chosen according to $\mathcal{D}_{\Z^N,N}$, algorithm~\ref{alg:babai} is correct and runs in
polynomial time.
\end{proposition}
\begin{proof}
The only thing we need to observe is that Step~1 turns the target vector into a vector with coordinates in $O(1)$ as we observed
that the stardard deviation was constant. The lattice point we are looking is therefore in a ball of constant radius.
\end{proof}
\section{Conclusion and further perspectives}
We described an algorithm that finds the generator of a principal ideal in $\Q(\zeta_{2^n})$ in heuristic complexity $2^{N^{1/2+\varepsilon}}$.
Our algorithm yields a heuristic attack on the homomorphic encryption
scheme of Vercauteren and Smart~\cite{fre_smart}, and the multilinear maps of Garg, Gentry and Halevi~\cite{mult_maps}
that rely on the hardness of finding a short generator of a principal ideal. Indeed, the ideals used in these
schemes have a very small generator, and under this assumption, we can use the reduction from short-PIP to PIP to
derive in polynomial time the secret small generator from an arbitrary generator also obtained in polynomial
time by calling Algorithm~\ref{alg:PIP2}.
Our attack relies on the Generalized Riemann Hypothesis, which is widely accepted as true, and also on Heuristic~\ref{heuristic:random}
which states that Algorithm~\ref{alg:full_q_descent} returns random relations. The latter allows us to carry on a complexity analysis,
but it would be probably very difficult to make a rigorous statement on the probability distribution of the relations found by
Algorithm~\ref{alg:full_q_descent}. Therefore, it will be important to design an efficient implementation our method. To properly
compare it with the existing methods, it will have to incorporate many classical practical improvements such as the use of
a small factor base following the bounds described in~\cite{belabasBd}, fast modular linear algebra to test the rank of the
relation matrix, optimized methods to intersect an ideal with a subfield, batch smoothness testing~\cite{bernstein}, etc ...
Besides the cryptographic applications, it will also be interesting to see how large the degree of the field can be pushed
in the ideal class group computation. Indeed, the methods implemented in the main computer algebra softwares behave extremely
badly when the degree of the number field exceeds 100.
Also, besides the straightforward generalization to power-of-$p$ cyclotomics
(for a fixed $p$), it will be interesting to see how this method generalizes to wider classes of number fields. It seems
unlikely that it will be applicable to many other well identified infinite classes of $K$ because it relies both on the existence
of subfields of prescribed degree.
\end{comment}
|
\section{Introduction}
With the discovery of the Higgs boson in the range of 125 -- 126~GeV the standard model (SM) has been completed \cite{Chatrchyan:2012ufa,Aad:2012tfa}. The uncertainty in the Higgs mass measurement has continuously decreased and is well below 0.5~GeV today \cite{Khachatryan:2014ira}. This small uncertainty is currently much better than the theoretical prediction in any scenario beyond the Standard Model (BSM). Therefore, more precise calculations are necessary to better confront BSM models with the Higgs mass measurement. This has two motivations of particular weight: (1) In the minimal supersymmetric standard model (MSSM) and many extensions thereof, radiative corrections are required to be at least as significant as the tree-level contribution, so higher-order corrections are especially important. (2) In the Standard Model and non-supersymmetric extensions thereof a precise calculation is required to extract the parameters of the model, which when run to high energies gives information about the stability or lifetime of the potential -- which may point the way to new physics if, as appears currently, the potential is metastable. Beyond these motivations, for a generic model of new physics with boundary conditions fixed from the top down (such as supersymmetric models) it is important to know what regions of parameter space are allowed, compatible with the measured Higgs mass. For example, a one-loop calculation may naively lead to excluding certain constrained scenarios, whereas with a two-loop calculation the Higgs mass may be large enough; this is related to the difficulty in estimating the error in the Higgs mass calculation, since at two-loop order there are new contributions from particles that have no direct couplings to the Higgs, and a simple variation of the renormalisation scale as an estimate of the error is not sufficient.
In general, there are three approaches to tackle the problem of finding the Higgs mass precisely: (i) effective potential calculations, (ii) diagrammatic calculations, (iii) renormalisation group equation methods. We shall concentrate in the following on the first two options. Calculations from the effective potential suffer from a larger uncertainty compared to diagrammatic calculations because of the missing momentum contributions. However, these are only really pronounced at one-loop level, and it is already possible to calculate the full one-loop Higgs mass inlcuding momentum dependence for generic models using {\tt SARAH}\xspace \cite{Staub:2008uz,Staub:2009bi,Staub:2010jh,Staub:2012pb,Staub:2013tta} to produce {\tt SPheno}\xspace \cite{Porod:2003um,Porod:2011nf} output or {\tt SOFTSUSY} \cite{Allanach:2001kg} output via {\tt FlexibleSUSY} \cite{Athron:2014yba}; explicit results have been known for some time for specific models such as the MSSM with real parameters \cite{Brignole:1991wp,Chankowski:1991md,Dabelstein:1994hb,Pierce:1996zz} and complex \cite{Pilaftsis:1998pe,Pilaftsis:1998dd,Frank:2006yh}; and for the NMSSM with real parameters \cite{Degrassi:2009yq,Staub:2010ty,Ender:2011qh} and complex \cite{Graf:2012hh}. At two loops the momentum effects are expected to be small: according to recent calculations for specific models they are comparable to the experimental uncertainty \cite{Martin:2004kr,Borowka:2014wla,Degrassi:2014pfa} and since the momentum-dependent corrections due to new physics scale at best as $m_H^2/M_{\rm New\ Physics}^2$ relative to the effective potential contribution, we expect this to be a general result. Hence effective potential calculations, with their concomitant great simplification over the diagrammatic approach, should be useful at two loop order and beyond (even if the inclusion of the momentum dependence will ultimately be necessary to reach the experimental accuracy).
In general even a two-loop calculation of the dominant contributions at zero external momentum is available for just two supersymmetric models: the MSSM \cite{Hempfling:1993qq,Carena:1995wu,Heinemeyer:1998jw,Zhang:1998bm,Heinemeyer:1998np,Heinemeyer:1999be,Espinosa:1999zm,Espinosa:2000df,Degrassi:2001yf,Brignole:2001jy,Brignole:2002bz,Martin:2002iu,Martin:2002wn,Dedes:2002dy,Dedes:2003km,Heinemeyer:2007aq,Hollik:2014bua} and partially for the next-to-minimal supersymmetric standard model (NMSSM) with real \cite{Degrassi:2009yq} and complex \cite{Muhlleitner:2014vsa} parameters. There continues to be much work in this direction and there are now some calculations of the strong (i.e. proportional to $\alpha_s$) momentum-dependent contributions for the MSSM \cite{Martin:2004kr,Borowka:2014wla,Degrassi:2014pfa}. These results have variously been made available to the community in model-specific public codes: {\tt FeynHiggs}~\cite{Heinemeyer:1998yj}, {\tt SoftSUSY}~\cite{Allanach:2001kg}, {\tt SuSpect}~\cite{Djouadi:2002ze} and {\tt SPheno}~\cite{Porod:2003um,Porod:2011nf} for the MSSM and {\tt NMSPEC} \cite{Ellwanger:2006rn}, {\tt Next-to-Minimal SOFTSUSY} \cite{Allanach:2013kza,Allanach:2014nba}, and {\tt NMSSMCALC} \cite{Baglio:2013iia} for the NMSSM. There are also some three-loop results, in the Standard Model \cite{Martin:2013gka,Martin:2014cxa} and the MSSM \cite{Martin:2007pg,Kant:2010tf} with the code {\tt H3m} based on \cite{Kant:2010tf}.
The state of the art in these calculations is however somewhat suprising given that the \emph{complete generic} expressions for the two-loop effective potential, valid for a general renormalisable quantum field theory, have been available for more than ten years by the work of Martin \cite{Martin:2001vx}. These were applied to a \emph{complete} two-loop calculation of the light Higgs mass in the MSSM in the effective potential approach in Ref.~\cite{Martin:2002wn}.
Furthermore, generic results for the diagrammatic calculation including the momentum dependence up to leading order in gauge couplings have been available in the literature for almost as long \cite{Martin:2003it,Martin:2003qz,Martin:2005eg}. Unfortunately the results of Ref.~\cite{Martin:2002wn} suffered the so-called ``Goldstone boson catastrophe'' (recently re-explored in \cite{Martin:2014bca,Elias-Miro:2014pca}) due to the presence of tachyons in the tree-level spectrum so were numerically unstable. Perhaps due to this no public code was made available to exploit these prior to Ref.~\cite{Goodsell:2014bna} where an implementation in {\tt SARAH}\xspace/{\tt SPheno}\xspace was presented.
Currently, the only generic two-loop results relevant for the Higgs mass calculation still not present in the literature are the all-electroweak loops and the corrections to the $Z$-boson mass relevant for determining the electroweak expectation value. These will be the subject of future work. Here we shall instead continue the process started in Ref.~\cite{Goodsell:2014bna} of making the pioneering generic results of Martin available in a public code -- which entails performing some new calculations.
As we stated above, calculating the two-loop corrections to the Higgs mass in the effective potential approach is expected to be a good approximation. However, there is more than one way to actually perform even this calculation: either we can calculate the potential and numerically take the derivatives, as done in Refs.~\cite{Martin:2002iu} and \cite{Goodsell:2014bna}, or we can perform the calculation diagrammatically and set the external momentum to zero. In this work we shall exploit this equivalence: we shall \emph{analytically} take the derivatives of the effective potential, producing expressions equivalent to the diagrammatic calculation and having the same structure, but with much simpler loop functions. The advantages of this over the first method are that the results are numerically stable\footnote{i.e. not subject to potential errors from ill-judged step-sizes in the numerical derivation or from parameters being too small.}; it is in principle a faster computation for more complicated models where the numerical method must make several passes to ensure stability; it can later be extended to a full diagrammatic calculation by simply changing the loop functions -- but at zero momentum the loop functions are much simpler and therefore significantly faster to evaluate. We shall therefore compute the analytic expressions for the first and second derivatives of the two-loop effective potential and implement them in {\tt SARAH}\xspace. As in Refs.~\cite{Goodsell:2014bna,Goodsell:2014pla} we shall ignore broken gauge groups, and adopt the same ans\"atze regarding the contribution of the electroweak gauge couplings to the tree-level Higgs mass matrix, to which references we refer the reader; the reasons for restricting to the so-called ``gaugeless limit'' are (a) partial circumvention of the Goldstone boson catastrophe (complete evasion in the case of the MSSM or any theory where the electroweak gauge couplings entirely determine the Higgs quartic potential); (b) significant simplification in the expressions and therefore speed in calculation; (c) the electroweak contributions are expected to be small, of the same magnitude as the momentum-dependent contributions. On the other hand, in the appendix we provide just the tadpole contributions in the case of broken gauge groups, and will return to the full expressions in future work.
The layout of the paper is as follows. In Sec.~\ref{SEC:derivatives} we explain our procedure to take the derivatives of the effective potential to extract the two-loop tadpole functions for a general theory with massless gauge bosons in a form convenient for automation. In subsection \ref{sec:Tadpoles} we summarize our results for the tadpole diagrams; we present the full set of second derivatives in appendix \ref{APP:SecondDerivatives}. The implementation of these results in {\tt SARAH}\xspace, including some technical details of the translation of the generic results into an an algorithm, is explained in Sec.~\ref{sec:sarah} before we conclude in Sec.~\ref{sec:summary}. Impatient readers interested in using our implementation of the results might want to jump directly to subsection~\ref{sec:manual}.
\section{Derivatives of the effective potential with massless vectors}
\label{SEC:derivatives}
In this section we derive the expressions for the two-loop tadpoles in a general quantum field theory with massless gauge bosons in Landau gauge. To do this, we analytically take the dervatives of the expressions in \cite{Martin:2001vx}. Writing the couplings in the notation of that paper, the theory is defined by real scalars $\phi_i$, Weyl fermions $\psi_I$ and massless gauge bosons $A_\mu^a$ where the gauge covariant derivative for the fermions and scalars are
\begin{align}
D_\mu \psi_I \equiv& \partial_\mu \psi_I+ ig A_\mu^a (T^a)_I^J \psi_J \nonumber\\
D_\mu \phi_i \equiv& \partial_\mu \phi_i + ig A_\mu^a \theta^a_{ij} \phi_j.
\end{align}
The structure constants $\theta^a_{ij}$ are imaginary antisymmetric matrices that obey the gauge algebra but, since we are writing in terms of real bosonic fields, for complex representations they will have twice the dimension of the equivalent generators $T^a$ (so e.g. a $U(1)$ generator is two-dimensional). We define as usual $\mathrm{tr}(\theta^a \theta^{a}) = \sum_i d(i) C(i)$ where $d(i)$ is the dimension of the representation of field $i$ and $C(i)$ the quadratic casimir, and similarly for $T^a$.
The lagrangian is then composed of the normal kinetic terms of the scalars and fermions using the above covariant derivatives supplemented by purely scalar and scalar-fermion interactions
\begin{align}
{\cal L}_S =& -\frac{1}{6} \lambda^{ijk} \phi_i \phi_j \phi_k - \frac{1}{24} \lambda^{ijkl} \phi_i \phi_j \phi_k \phi_l \nonumber\\
{\cal L}_{SF} =& - \frac{1}{2} y^{IJk} \psi_I \psi_J \phi_k + {\mathrm c.c.}
\label{EQ:couplingdefinitions}\end{align}
$y$ is in general a dimensionless complex tensor with $y^{IJk}=y^{JIk}$, while $\lambda^{ijk},\lambda^{ijkl}$ are real, symmetric tensors.
\subsection{Effective potential}
\begin{figure}[hbt]
\includegraphics[width=0.5\linewidth]{EffPotDiagrams.pdf}
\label{fig:EffPot}
\caption{Two-loop diagrams contributing to the effective potential in the gaugeless limit.}
\end{figure}
We can simplify the expression for the effective potential in the Landau gauge given in Ref.~\cite{Martin:2001vx} for our case; with all gauge groups unbroken some diagrams do not contribute. The non-vanishing diagrams are shown in Fig.~\ref{fig:EffPot}. \\
The contribution of each diagram to the effective potential is given by:
\begin{align}
\label{EQ:Potentialfirst}
V^{(2)}_{SSS} =& \frac{1}{12} (\lambda^{ijk})^2 f_{SSS} (m_i^2, m_j^2,m_k^2) \\
V^{(2)}_{SS} =& \frac{1}{8} \lambda^{iijj} f_{SS} (m_i^2, m_j^2) \\
V^{(2)}_{FFS} =& {1\over 2} |y^{IJk}|^2 f_{FFS} (m^2_I, m^2_J, m^2_k) \\
V^{(2)}_{\overline{FF}S} =& {1\over 4} y^{IJk} y^{I'J'k} M^*_{II'} M^*_{JJ'}
f_{\overline{FF}S} (m^2_I, m^2_J, m^2_k) + {\rm c.c.}\\
V^{(2)}_{SSV} =& {g^2\over 4} d(i) C(i) f_{SSV} (m^2_i,m^2_i, 0)\\
V^{(2)}_{FFV} =& {g^2 \over 2} d(I) C(I) f_{FFV} (m^2_I, m^2_I, 0) \\
V^{(2)}_{\overline{FF}V} =& -{g^2 \over 2} d(I) C(I) m_I^2 f_{\overline{FF}V} (m^2_I, m^2_I, 0)
\label{EQ:Potential}\end{align}
Here, $y$ and $\lambda$ are the trilinear and quartic couplings of eq.~(\ref{EQ:couplingdefinitions}), $g$ is a gauge coupling, and $M$ are fermion masses.
The loop functions, given in terms of standard basis functions given in appendix \ref{APP:loopfunctions}, are the same for $\overline{MS}$ and $\overline{DR}^\prime$ in the case of
\begin{align}
f_{SSS}= & - I(x,y,z) \nonumber \\
f_{SS}= & J(x,y) \nonumber\\
f_{FFS}= & J(x,y) - J(x,z) - J(y,z) + (x + y - z) I(x,y,z) \nonumber\\
f_{\overline{FF}S} =& 2 I(x,y,z) \nonumber\\
f_{SSV} = & (x+y)^2 + 3 (x+y)I(x,y,0) + 3J(x,y) - 2xJ(x) - 2yJ(y)
\end{align}
but differ for those with vectors and fermions:
\begin{equation}
\begin{array}{|c|c|c|} \hline
& \overline{MS} & \overline{DR}^\prime \\ \hline
f_{FFV} & 0 & - (x+y)^2 + 2x J(x) + 2y J(y)\\
f_{\overline{FF}V} & 6I (x,y,0) + 2(x+y) - 4 J(x) - 4 J(y) & 6I (x,y,0)\\ \hline
\end{array}
\end{equation}
All of these functions are symmetric on the substitution of their first two indices, but may not be so with the third. In fact, we can then combine the vector-fermion diagrams to give
\begin{align}
V^{(2)}_{FFV}+V^{(2)}_{\overline{FF}V} \equiv {g^2 \over 2} d(I) C(I) F_{FV} (m_I^2)
\end{align}
where
\begin{align}
F_{FV} (x) \equiv& -4 x^2 + 4 x J(x) - 6 x I(x,x,0) + \delta_{\rm{MS}} 4x J(x)
\end{align}
where $ \delta_{\rm{MS}} $ is zero for $\overline{\rm{DR}}^\prime$ and one for $\overline{\rm{MS}}$.
\subsection{Derivatives of the potential}
Here we shall analytically take the derivatives of the potential. In a generic model, we may want the derivatives of the Higgs potential in terms of some unrotated fields, i.e. in a basis where the mass matrix is not diagonal; let us say that we start with such a case. We write the tree-level potential in terms of some expectation values $\hat{v}_i$ and the associated real fluctuations $S_i^0$ as
\begin{align}
V^{\mathrm{scalar\ tree}} = V_0 (\hat{v}_i) + \frac{1}{2} \hat{m}_{ij}^2 S_i^0 S_j^0 + \frac{1}{6} \hat{\lambda}_0^{ijk} S_i^0 S_j^0 S_k^0 + \frac{1}{24} \hat{\lambda}^{ijkl}_0 S_i^0 S_j^0 S_k^0 S_l^0.
\end{align}
Then there is a tree-level rotation
\begin{equation}
S_i^0 = R^0_{ij} S_j
\end{equation}
to diagonalise the mass matrix; we then obtain
\begin{align}
V^{\mathrm{scalar\ tree}} = V_0 + \frac{1}{2} m_{i}^2 S_i S_i + \frac{1}{6} \lambda^{ijk} S_i S_j S_k + \frac{1}{24} \lambda^{ijkl} S_i S_j S_k S_l.
\end{align}
We can write $\phi_i = v_i + S_i$ and work with the couplings of eq.~(\ref{EQ:couplingdefinitions}). We then need the quantities which enter in the effective potential calculation, which are masses and couplings depending on $\{S_i\}$.\\
In general, we have
\begin{align}
m_{ij}^2 (S) =& \frac{\partial^2}{\partial S_i \partial S_j} V \nonumber\\
=& m_{ i}^2 \delta_{ij}+ \lambda^{ijk} S_k + \frac{1}{2} \lambda^{ijkl} S_k S_l\nonumber\\
\lambda^{ijk} (S) =& \frac{\partial^3}{\partial S_i \partial S_j \partial S_k} V \nonumber\\
&= \lambda^{ijk} + \lambda^{ijkl} S_l \nonumber\\
\lambda^{ijkl} (S) =& \lambda^{ijkl}
\end{align}
with the shorthand notation $S\equiv\{S_i\}$.
Hence we can write
\begin{align}
\frac{\partial}{\partial S_r} \lambda^{ijkl} (S) =& 0 \nonumber\\
\frac{\partial}{\partial S_r} \lambda^{ijk} (S) =& \lambda^{ijkr} (S) \nonumber\\
\frac{\partial}{\partial S_r} m_{ij}^2 (S) =& \lambda^{ijr} + \lambda^{ijkr} S_k .
\end{align}
Similarly for the fermions we have
\begin{align}
m_I^2 \delta_J^I =& M^{I I'}M_{JI'}^* \nonumber\\
\rightarrow \frac{\partial}{\partial S_r} M^{I I'}M_{JI'}^* =& y^{II'r} M_{JI'}^* + M^{I I'} y_{JI'r} \nonumber\\
\frac{\partial}{\partial S_r} y^{IJs} =& 0.
\end{align}
However, for the purposes of the effective potential, we then require a further diagonalisation for $m_{ij}^2 (S) $: we put
\begin{equation}
S_i = R_{ij} (S)\, S_j^\prime.
\end{equation}
We name the couplings in this basis as $\tilde{m}_i(S), \tilde{\lambda}^{ijk}_S(S),\tilde{\lambda}^{ijkl}(S) $.
We then express the effective potential by inserting the couplings and masses in the basis $\{S_i^\prime\}$ into the formulae of eqs.~(\ref{EQ:Potentialfirst})--(\ref{EQ:Potential}). However, to take the derivatives we rewrite the expressions in terms of the basis $\{S_i\}$, and use the trick (with ${\bf m^2}\equiv(m_{ij}^2)$)
\begin{align}
\frac{\partial}{\partial S_r} \left( \frac{1}{q^2 + {\bf m^2}} \right)_{ij} =& -\left(\frac{1}{q^2 + {\bf m^2}}\right)_{ik} \frac{\partial m_{k k'}^2}{\partial S_r} \left(\frac{1}{q^2 + {\bf m^2}}\right)_{k'j} \,.
\end{align}
For similar expressions we write by abuse of notation (using $C\equiv 16\pi^2 \mu^{2\epsilon}(2\pi)^{-d}$ \cite{Martin:2003qz})
\begin{align}
\mathbf{J}(m^2_{ik},m_{jl}^2) &\equiv C^2\int d^dq d^dk \left( \frac{1}{q^2 + {\bf m^2}} \right)_{ik} \left( \frac{1}{k^2 + {\bf m^2}} \right)_{jl}.
\end{align}
For fermion propagators we can write
\begin{align}
\frac{M_{II'}}{q^2 + m_{I}^2} \rightarrow M_{IJ} \frac{1}{q^2 + M^{JK}M_{KI'}} = \frac{1}{q^2 + M_{IJ}M^{JK}} M_{KI'}.
\end{align}
Let us demonstrate our method on a brief example:
\begin{align}
\frac{\partial}{\partial S_p^0 }\frac{1}{8} \tilde{\lambda}^{iijj} f_{SS} (\tilde{m}_i^2, \tilde{m}_j^2) =& R_{rp}^0\frac{\partial}{\partial S_r} \frac{1}{8} \tilde{\lambda}^{iijj} f_{SS} (\tilde{m}_i^2, \tilde{m}_j^2) \nonumber\\
=& R_{rp}^0\frac{\partial}{\partial S_r} \frac{1}{8} \lambda^{ik jl} f_{SS} (m_{ik}^2, m_{jl}^2) \nonumber\\
=& R_{rp}^0 \frac{1}{4}\lambda^{ik jl} (S) f^{(1,0)}_{SS} (m_{im}^2,m_{nk}^2; m_{jl}^2) \frac{\partial}{\partial S_r}( m_{mn}^2) \nonumber\\
=& R_{rp}^0\frac{1}{4}\lambda^{ik jl} (S) f^{(1,0)}_{SS} (m_{im}^2,m_{nk}^2; m_{jl}^2) \lambda^{mn r} (S) \nonumber\\
\underset{S\rightarrow 0}{\longrightarrow}& R_{rp}^0\frac{1}{4}\lambda^{ik jj} \lambda^{i k r} f^{(1,0)}_{SS} (m_i^2, m_k^2;m_j^2 ).
\end{align}
Recall that $f_{SS} (x,y) \equiv J(x,y)$ where $J$ is the finite loop function and ${\bf J}(x)=J(x)-\frac x\epsilon$, see Eq.~(\ref{eq:oneloopfunctionsAJ}); here we have defined
\begin{align}
f^{(1,0)}_{SS} (x, y; z)\equiv& -C^2 \int d^d k d^d q \frac{1}{k^2 + x} \frac{1}{k^2 + y} \frac{1}{q^2 + z} + \frac{C}{\epsilon}\bigg({\bf J}(z) - C z \int d^d k \frac{1}{k^2 + x} \frac{1}{k^2 + y} \bigg) + \frac{z}{\epsilon^2} \nonumber\\
=& -C^2 \bigg( \int d^d k \frac{1}{k^2 + x} \frac{1}{k^2 + y} \bigg) \bigg(\int d^d q \frac{1}{q^2 + z} +\frac{z}{\epsilon} \bigg) + \frac{C}{\epsilon}{\bf J}(z) + \frac{z}{\epsilon^2}\nonumber\\
=& \frac{1}{x-y} \bigg( {\bf J}(x) - {\bf J}(y) \bigg) J(z) + \frac{1}{\epsilon} J(z) \nonumber\\
=& \frac{1}{x-y} \bigg( J(x) - J(y) \bigg) J(z) \nonumber\\
=& - B_0 (x,y) J(z).
\end{align}
Note that the $R_{ra}^0$ are not functions of $S_i$ and so do not present complications if we want to take futher derivatives. \\
We can similarly take the derivatives of all the remaining loop functions; these are derived from the basis:
\begin{align}
\frac{\partial}{\partial S_r} J(m_i^2,m_j^2) \rightarrow & -B_0 (m_i^2,m_k^2) J(m_j^2) \frac{\partial m_{ik}^2}{\partial S_r} - J(m_i^2) B_0 (m_j^2,m_k^2) \frac{\partial m_{jk}^2}{\partial S_r} \nonumber\\
\frac{\partial}{\partial S_r} B_0 (m_i^2,m_j^2) \rightarrow& - C_0 (m_i^2,m_k^2,m_j^2) ( \frac{\partial m_{ik}^2}{\partial S_r} + \frac{\partial m_{jk}^2}{\partial S_r}) \nonumber\\
\frac{\partial}{\partial S_r} I(m_i^2,m_j^2,m_k^2) \rightarrow& - U_0 (m_i^2,m_l^2,m_j^2,m_k^2) \frac{\partial m_{il}^2}{\partial S_r} - U_0 (m_j^2,m_l^2,m_i^2,m_k^2) \frac{\partial m_{jl}^2}{\partial S_r} \nonumber\\
&- U_0 (m_k^2,m_l^2,m_i^2,m_j^2) \frac{\partial m_{kl}^2}{\partial S_r}
\end{align}
and
\begin{align}
\frac{\partial}{\partial S_r} m_i^2 I(m_i^2,m_j^2,m_k^2) \rightarrow& \frac{\partial m_{il}^2}{\partial S_r} I(m_l^2,m_j^2,m_k^2) - m_i^2 \frac{\partial m_{il}^2}{\partial S_r} U_0 ( m_l^2,m_i^2,m_j^2,m_k^2) \nonumber\\
&-\frac{\partial m_{jl}^2}{\partial S_r}m_i^2 U_0 (m_l^2,m_j^2,m_k^2,m_i^2) - \frac{\partial m_{kl}^2}{\partial S_r}m_i^2 U_0 (m_l^2,m_k^2,m_i^2,m_j^2) .
\end{align}
The derivative of a typical term in the effective potential will have the form
\begin{align}
\frac{\partial}{\partial S_r} A^{ijk} A^{ijk} f_\alpha(x_i,y_j,z_k) =& 2 f_\alpha(x_i,y_j,z_k) A^{ijk} \frac{\partial}{\partial S_r} (A^{ijk}) \nonumber\\
&+ \bigg\{ A^{ijk} A^{i'jk} \frac{\partial m^2_{ii'}}{\partial S_r} f_\alpha^{(1,0,0)}(x_i,x_{i'};y_j,z_k) + (x\leftrightarrow y) + (x\leftrightarrow z)\bigg\}.
\end{align}
where, generalising the above, it is straightforward to show that for a generic function appearing in the effective potential $f_\alpha$ composed of polynomials (even containing monomials with negative exponents) multiplying the loop functions above, we can write
\begin{align}
f_\alpha^{(1,0,0)} (x,u; y,z) \equiv& \frac{f_\alpha (x,y,z) - f_\alpha(u,y,z)}{x-u}
\end{align}
and similarly for permutation of the indices. On the other hand, this explicit expression is often inconvenient in practice due to the need to carefully take the smooth limit when $x=u$; it is instead more practical to rewrite the right hand side in terms of our basis of loop functions multiplied by suitable polynomials. In the following we present explicit expressions for the first derivatives (and, in the appendix, the second derivatives) which have been appropriately simplified to remove the apparent singularities.
\subsection{First derivatives}
\label{sec:Tadpoles}
\begin{figure}[hbt]
\includegraphics[width=0.4\linewidth]{1pointDiagrams}
\label{fig:Tad2L}
\caption{Tadpole diagrams at the two-loop level which don't vanish in the gaugeless limit. }
\end{figure}
Here we gather the set of two-loop tadpole diagrams. There are only three topologies, two of which only apply to the all-scalar case. All generic possible diagrams are given in Fig.~\ref{fig:Tad2L}. Here we present our results for the first derivatives in the basis $\{S_i\}$, so without the rotation matrices $R$; the full tadpole in the original basis is then given by
\begin{align}
\frac{\partial V^{(2)}}{\partial S_p^0} = R_{rp}^0 [T_{S} + T_{SSFF}+ T_{FFFS}+T_{SV}+T_{FV}]
\label{EQ:FullTadpoleGaugeless}\end{align}
with the tadpoles on the right-hand side to be defined below.
\subsubsection{All scalars}
We start with the purely scalar diagrams which are in the first row of Fig.~\ref{fig:Tad2L}. The entire contribution is given by
\begin{equation}
T_{S} = T_{SS} + T_{SSS} + T_{SSSS}
\end{equation}
with
\begin{align}
T_{SS} =& \frac{1}{4}\lambda^{ik jj} \lambda^{i k r} f^{(1,0)}_{SS} (m_i^2, m_k^2;m_j^2 ) \\
T_{SSS} =&\frac{1}{6} \lambda^{rijk} \lambda^{i jk} f_{SSS} (m_i^2, m_j^2,m_k^2) \\
T_{SSSS} = & \frac{1}{4} \lambda^{rii'} \lambda^{ijk} \lambda^{i'jk} f_{SSS}^{(1,0,0)} (m_{i}^2,m_{i'}^2; m_{j}^2,m_{k}^2) .
\end{align}
The new loop functions are defined as
\begin{align}
f^{(1,0)}_{SS} (x,y;z ) \equiv&- B_0 (x,y) J(z) \nonumber\\
f_{SSS}^{(1,0,0)} (x,y;u,v ) \equiv& U_0 (x,y,u,v).
\end{align}
Note that $f^{(1,0)}_{SS}$ corresponds to $X_{SSS}$, $f_{SSS}$ corresponds to $S_{SSS}$, and $f^{(1,0,0)}_{SSS}$ to $W_{SSSS}$ of Ref.~\cite{Martin:2003it} in the limit of zero external momentum.
\subsubsection{Scalars and fermions}
We have, first, the diagrams with two scalar propagators:
\begin{align}
T_{SSFF}=& \frac{1}{2}y^{IJk} y_{IJl} f_{FFS}^{0,0,1} (m^2_I, m^2_J; m_k^2,m_l^2) \lambda^{klr} \nonumber\\
& - \bigg[\frac{1}{2} y^{IJk} y^{I'J'k} M^*_{II'} M^*_{JJ'} \lambda^{klr} U_0(m_l^2,m^2_k,m^2_I, m^2_J) + c.c.\bigg].
\end{align}
Then we turn to one scalar and three fermion propagators:
\begin{align}
T_{FFFS}=& 2\mathrm{Re} \big[ y^{IJr} y_{IKm} y^{KLm} M^*_{JL} \big] T_{F\overline{F}FS} (m_I^2, m_J^2,m_K^2,m_m^2) \nonumber\\
& + 2\mathrm{Re} \big[ y_{IJr} y^{IKm} y^{JLm} M^*_{KL} \big] T_{FF\overline{F}S} (m_I^2, m_J^2,m_K^2,m_m^2) \nonumber\\
&- 2\mathrm{Re} \big[ y^{IJr} y^{KLm} y^{MNm} M^*_{IK}M^*_{JM} M^*_{LN}\big] T_{\overline{F}\overline{F}\overline{F}S} (m_I^2, m_J^2,m_L^2,m_m^2).
\end{align}
Here we have defined
\begin{align}
f_{FFS}^{1,0,0} (m_{I'}^2,m^2_I, m^2_J; m_k^2) \equiv & - B_0 (m_{I'}^2,m^2_I) J(m_J^2) + B_0 (m_{I'}^2,m^2_I) J(m_k^2) + I( m_{I'}^2, m_J^2,m_k^2)\nonumber\\
&- (m_I^2 + m_J^2 - m_k^2) U_0(m_{I'}^2,m^2_I, m^2_J, m_k^2), \nonumber\\
f_{FFS}^{0,0,1} (m^2_I, m^2_J; m_k^2, m_l^2) \equiv & B_0 (m_l^2,m^2_k) J(m_I^2) + B_0 (m_l^2,m^2_k) J(m_J^2) - I( m_{I}^2, m_J^2,m_l^2)\nonumber\\
&- (m_I^2 + m_J^2 - m_k^2) U_0(m_l^2,m_k^2,m^2_I, m^2_J) \nonumber\\
T_{\overline{F}\overline{F}\overline{F}S} (m_I^2, m_J^2,m_L^2,m_m^2) \equiv& U_0( m_I^2, m_J^2,m_L^2, m_m^2). \nonumber\\
T_{F\overline{F}FS} (m_I^2, m_J^2,m_K^2,m_m^2) \equiv& f_{FFS}^{1,0,0} (m_I^2, m_J^2,m_K^2;m_m^2) \nonumber\\
T_{FF\overline{F}S} (m_I^2, m_J^2,m_K^2,m_m^2) \equiv& I(m_I^2,m_K^2,m_m^2) - m_I^2 U_0(m_I^2, m_J^2,m_K^2,m_m^2).
\end{align}
\subsubsection{Diagrams with vectors}
Finally, the two generic diagrams involving vectors are given by
\begin{align}
T_{SV}=& \frac{g^2}{2 } d(i) C(i) \lambda^{iir}\bigg(3I(0,m_i^2,m_i^2) - J(m_i^2) +2 m_i^2 \bigg) \\
=& \frac{g^2}{2 } d(i) C(i) \lambda^{iir} m_i^2 \bigg[ -12 + 11 \log m_i^2/Q^2 - 3 \log^2 m_i^2/Q^2) \bigg] .\label{EQ:TadpoleScalarVector}\nonumber\\
T_{FV}=& g^2 d(I) C(I) \mathrm{Re}( M_{I I'} y^{II'r} ) 4 \bigg(-3I(0,m_I^2,m_I^2) + 5 J(m_I^2) -4 m_I^2 + \delta_{\overline{\mathrm{MS}}} \big[ 2 J(m_I^2) + m_I^2\big] \bigg)\\
=& g^2 d(I) C(I) \mathrm{Re}( M_{I I'} y^{II'r} ) 4m_I^2 \bigg[ 6 - 7 \log m_I^2/Q^2 +3 \log^2 m_I^2/Q^2 + \delta_{\overline{\mathrm{MS}}} \big[ 2 \log m_I^2/Q^2 - 1\big]\bigg]. \nonumber
\end{align}
\subsection{Second derivatives}
To find the second derivatives of the potential we can apply the same technique. However, in principle, we can simply use the results of \cite{Martin:2003it}, which computed (diagrammatically) the two-loop scalar self-energies at leading order in gauge couplings. Since we want the self-energies for neutral scalars, this comprises all of the contributions, and if we want the results in the effective potential approach we can simply set the external momenta to zero (and, for this work, the masses of the gauge bosons to zero). In fact, for the majority of the diagrams, these yield the same result. However, in a few cases we find that by taking the derivatives of the potential we find simpler results (which are of course entirely equivalent).
The full result is given by
\begin{align}
\frac{\partial V^{(2)}}{\partial S_p^0\partial S_q^0} =& R_{ip}^0 R_{jq}^0 \Pi_{ij} (0) \nonumber\\
\equiv& R_{ip}^0 R_{jq}^0 \bigg[ \Pi^{S}_{ij} + \Pi^{SF(W)}_{ij} + \Pi^{SF_4 (M)}_{ij} + \Pi^{S_2 F_3 (M)}_{ij} + \Pi^{S_3F_2 (V)}_{ij} + \Pi^{SF_4 (V)}_{ij} + \Pi^{SV}_{ij} + \Pi_{ij}^{FV}\bigg]
\end{align}
where the superscripts correspond to the numbers of scalars, fermions and vectors with types of topology listed for diagrams $(M)$ and $(V)$.
We give the complete set of relevant expressions in appendix \ref{APP:SecondDerivatives}; the corresponding diagrams are shown in Figs.~\ref{fig:SelfScalars} and \ref{fig:SelfMixed}. The expressions for $ \Pi^{SF_4 (V)}_{ij}, \Pi^{SV}_{ij}$ and $\Pi_{ij}^{FV} $ exhibit particular simplifications in our approach.
\label{sec:selfenergies}
\begin{figure}[hbt]
\includegraphics[width=0.66\linewidth]{2pointDiagrams_Scalar}
\caption{Two-loop self-energy diagrams involving only scalars. }
\label{fig:SelfScalars}\end{figure}
\begin{figure}[hbt]
\includegraphics[width=0.5\linewidth]{2pointDiagrams_Mixed}
\caption{Remaining two-loop self-energy diagrams which do not vanish in the gaugeless limit. }
\label{fig:SelfMixed}\end{figure}
\section{Implementation in {\tt SARAH}\xspace}
\label{sec:sarah}
We have implemented the new routines in {\tt SARAH}\xspace. By including the first and second derivatives of the effective potential using the analytic expressions here, rather than numerically taking the derivatives of the couplings and masses as performed in the previous version \cite{Goodsell:2014bna}, we can guarantee greater numerical stability, accuracy and speed improvements. Moreover, this approach allows a straightforward upgrade to the \emph{pole} mass calculation by simply changing the loop functions called to those defined in Ref.~\cite{Martin:2003it} based on loop functions implemented in {\tt TSIL} \cite{Martin:2005qm}, which will be made possible in a future version.
\subsection{Method}
For any given supersymmetric\footnote{A two-loop Higgs mass calculation in non-SUSY models will become available in future releases} model {\tt \$MODEL}, once the user has specified the particle content and their symmetries, {\tt SARAH}\xspace calculates all of the vertices and masses. It then writes a \texttt{Fortran}\xspace code (placed in the suggestively named {\tt 2LPole\_\$MODEL.f90})
which implements our expressions, linking to a static \texttt{Fortran}\xspace code (named {\tt 2LPoleFunctions.f90}) of the basis functions for the generic first and second derivatives of the effective potential defined in section \ref{SEC:derivatives} and the appendix. These two pieces of \texttt{Fortran}\xspace code are called by {\tt SPheno}\xspace during the calculation of the loop corrections to the Higgs mass. Here we shall give a few details of how {\tt SARAH}\xspace writes {\tt 2LPole\_\$MODEL.f90}.
The overall algorithm is to
\begin{enumerate}
\item Generate masses and couplings for all relevant particles in the gaugeless limit.
\item Populate and classify all tadpole topologies according to particle content.
\item For each tadpole topology, pass the set of diagrams along with information specifying the symmetries to a generic writer function.
\item Rotate the total tadpole vector by the Higgs rotation matrix to the non-diagonal basis, c.f. eq.~(\ref{EQ:FullTadpoleGaugeless}).
\item Populate and classify all second derivative topologies according to particle content.
\item For each mass topology, pass the set of diagrams along with information specifying the symmetries to a generic writer function.
\item Rotate the mass matrix to the non-diagonal basis (as with the tadpoles).
\end{enumerate}
The writer function is actually identical for tadpoles and mass diagrams with a switch to adjust the number of Higgses. It cycles through the list of diagrams and applies the following process for each:
\begin{enumerate}[(a)]
\item Determine symmetry factor of diagram due to permutations.
\item Determine the colour factor; for diagrams with a gluon propagator this is simply $d(I) C(I)$ whereas otherwise we must trace over colour indices of the vertices. In principle, for four-point vertices there can be two colour structures for the vertex which superficially leads to more than one colour factor for such diagrams. However, as we can simply see by inspecting the expressions in the appendix, or by considering that the colour factors have to be inherited from a corresponding vacuum diagram (since differentiating with respect to neutral Higgs fields cannot introduce any additional colour factor), for diagrams with a four-point vertex consisting of four coloured fields the colour factor is given by a trace over the indices in pairs. Hence such four-point vertices are saved with the colour factor of the pairs of indices traced over. To be more explicit, such vertices can only contribute if they come from differentiating $V_{SS}^{(2)}$ which contains the coupling $\lambda^{iijj}$. We then must simply take care that the indices of the vertices correspond correctly to the indices that are traced over.
\item Write a nested set of loops to sum the diagram over the generations of all particles and, for the inner loop, the external Higgs legs, since the most computationally demanding aspect is evaluating the loop functions and this can be evaluated before calling the Higgs loop -- indeed, we also check that the coupling multiplying it is non-zero first too.
\end{enumerate}
There are subtleties in translating our results into a form usable by {\tt SARAH}\xspace, both stemming from the fact that in the {\tt SPheno}\xspace code the couplings are stored in terms of either real or complex scalars, and four-component spinor fermions, while, since our results are based on those of Refs.~\cite{Martin:2001vx} and \cite{Martin:2003it} and for economy we use real scalars and two-component fermions. The translation between the two bases as required by {\tt SARAH}\xspace and {\tt SPheno}\xspace is largely as described in Ref.~\cite{Goodsell:2014bna}, however here we have the additional complication for fermions of translating chains of couplings and masses such as
$$C_1=\mathrm{Re}( y^{IJr} y^{KLm} y^{MNm} M^*_{IK}M^*_{JM} M^*_{LN}). $$
In {\tt SARAH}\xspace, the interactions of Weyl spinors $\psi$ are derived from the corresponding Dirac spinor interactions $\Psi$ as
\begin{align}
\mathrm{Vertex} = i \frac{\delta {\cal L}}{\delta \phi_r \delta \bar \Psi_I \delta \Psi_J} = i \left(\frac{\delta {\cal L}}{\delta \phi_r \delta \psi^R_I \delta \psi^L_J} P_L + \frac{\delta {\cal L}}{\delta \phi_r \delta \psi^{L*}_I \delta \psi^{R*}_J} P_R\right) \equiv i (c_L P_L + c_R P_R) ,
\end{align}
with $\psi^Y_X = P_Y \Psi_X$ ($Y=L,R$; $X=I,J$) and polarization operators $P_L$, $P_R$;
and so $c_{L,R} \leftrightarrow -y^{IJr}, c_{L,R}^* \leftrightarrow -y_{IJr}$. When we have complex scalars $\Phi$, we should write
\begin{align}
\mathcal{L} \supset& \Phi^m (\overline{\Psi}_I (c_L (I,J,m) P_L + c_R (I,J,m) P_R) \Psi_J + \Phi_m (\overline{\Psi}_I (c_L (I,J,\overline{m}) P_L + c_R (I,J,\overline{m})P_R) \Psi_J \nonumber\\
=& \Phi^m (\overline{\Psi}_I (c_L (I,J,m) P_L + c_R (I,J,m) P_R) \Psi_J + \Phi_m (\overline{\Psi}_J (c_R^* (I,J,m) P_L + c_L^* (I,J,m) P_R) \Psi_J
\end{align}
and so $c_L(I,J,\overline{m}) = c_R^*(J,I,m)$. For a given topology, {\tt SARAH}\xspace populates the diagrams using Dirac propagators (i.e. links fermions with its conjugate) and so for the coupling $C_1$ above we will find sets of particles
\begin{align}
\{\overline{\Psi}_I,\Psi_J,\phi^r\},\{\overline{\Psi}_N,\Psi_I,\phi^m\},\{\overline{\Psi}_J,\Psi_N,\overline{\phi}_m\}.
\end{align}
Suppose each of the fermions is a Dirac spinor with Weyl spinors $\psi^I_{L,R}$ etc, then to construct coupling $C$ above we must sum over the left and right-handed Weyl fermions (which have opposite representations of all gauge groups) and thus (noting that {\tt SPheno}\xspace always internally stores the fermion masses as real positive definite)
\begin{align}
C_1 = \mathrm{Re} \bigg[&y^{\psi^I_L,\psi^J_R,r}y^{\psi^N_L,\psi^I_R,m}y^{\psi^J_L,\psi^N_R,\overline{m}} + y^{\psi^I_R,\psi^J_L,r}y^{\psi^N_R,\psi^I_L,\overline{m}}y^{\psi^J_R,\psi^N_L,m}\bigg] M_I M_J M_L \nonumber\\
=& -\mathrm{Re} \bigg[c_R^* (I,J,r) c_R^* (N,I,\overline{m}) c_R^*(J,N,m) + c_L(I,J,r) c_L(N,I,\overline{m}) c_L(J,N,m) \bigg] M_I M_J M_N \nonumber\\
=& -\mathrm{Re} \bigg[c_R (I,J,r) c_R (N,I,\overline{m}) c_R(J,N,m) + c_L(I,J,r) c_L(N,I,\overline{m}) c_L(J,N,m) \bigg] M_I M_J M_L.
\end{align}
On the other hand, if the fermions are all Majorana then $\psi^I_L = \psi_R^I$ and we therefore only have half of this sum, so we include an extra factor of $1/2$.
If we consider another example
\begin{align}
C_2 =& \mathrm{Re}( y^{IJr} y_{IKm} y^{KNm} M_{JN}^* )
\end{align}
in {\tt SARAH}\xspace we would generate the set of particles
\begin{align}
\{\overline{\Psi}_I,\Psi_J,\phi^r\},\{\overline{\Psi}_K,\Psi_I,\overline{\phi}_m\},\{\overline{\Psi}_J,\Psi_K,\phi^m\}.
\end{align}
and thus
\begin{align}
C_2 \rightarrow& \mathrm{Re} \bigg[y^{\psi^I_L,\psi^J_R,r}y_{\psi^I_L,\psi^K_R,\overline{m}}y^{\psi^K_R,\psi^N_L,m}+y^{\psi^I_R,\psi^J_L,r}y_{\psi^I_R,\psi^K_L}^my^{\psi^K_L,\psi^N_R}_{\overline{m}} \bigg] M_J \nonumber\\
\rightarrow& -\mathrm{Re} \bigg[c_R^* (I,J,r) c_L^* (K,I,m) c_R^*(J,K,m) + c_L (I,J,r) c_R (K,I,m) c_L(J,K,m)\bigg] M_J \nonumber\\
=&-\mathrm{Re} \bigg[c_R (I,J,r) c_L (K,I,m) c_R(J,K,m) + c_L (I,J,r) c_R (K,I,m) c_L(J,K,m)\bigg] M_J .
\end{align}
The above show that it is straightforward to translate the two-component results into expressions in {\tt SARAH}\xspace
\begin{align}
\mathrm{Re}\bigg[& \big[ \prod_{i=1}^m y^{I_i J_i s_i} \prod_{j=1}^n y_{I_j J_j s_j} \big] \big[ \prod_{k=1}^p M_{I_k J_k} \big] \bigg] \rightarrow\nonumber\\
& \left(\frac{1}{2} \right)^M (-1)^{m+n} \mathrm{Re}\bigg[ \prod_{i=1}^mc_L(I_i,J_i,s_i) \prod_{j=1}^nc_R(I_j,J_j,s_j) + \prod_{i=1}^mc_R(I_i,J_i,s_i)\prod_{j=1}^nc_L(I_j,J_j,s_j)\bigg] \big[ \prod_{k=1}^p M_{I_k} \big]
\end{align}
where $M=1$ for Majorana fermions and zero otherwise.
A final point regarding the translation into a basis of real and complex scalars is that the new routines assume that there is a unique way of constructing a gauge- and global symmetry-invariant coupling $\lambda^{ijk}$ from complex scalars other than the complex conjugate of the whole coupling; i.e. if $\lambda^{ijk}$ is permitted for given complex $i,j,k$ then $\lambda^{ij}_k$ is not. This is evidently true -- if both are permitted then we can generate a holomorphic mass term at one loop which violates the premise. However, it is important in that we cannot write gauge singlets as complex scalars if they have couplings violating the above condition, no matter how small the couplings -- for example for sneutrinos in see-saw models.
\subsection{How to use the new routines}
\label{sec:manual}
To study a model with {\tt SARAH}\xspace the general procedure is as follows: the user should download and run {\tt SARAH}\xspace with the demanded model. {\tt SARAH}\xspace derives all analytical expressions for mass matrices, vertices, renormalisation group equations as well as loop corrections and exports this information into \texttt{Fortran}\xspace source code. The \texttt{Fortran}\xspace source code is compiled together with {\tt SPheno}\xspace and all numerical
calculations are then performed by the new {\tt SPheno}\xspace module. This includes a calculation of the entire mass spectrum, branching ratios as well as flavour and other precision observables \cite{Porod:2014xia}. For the mass spectrum all one-loop corrections to any particle are included in a diagrammatic way \cite{Staub:2010ty}. For a supersymmetric model there are now in addition three options to get two-loop corrections in the Higgs sector. The first two are based on the effective potential approach presented in Ref.~\cite{Goodsell:2014bna} while the new routines are called by the (now default) third option.
A step-by-step description to obtain a spectrum generator for an arbitrary model {\tt \$MODEL} implemented in {\tt SARAH}\xspace reads as follows:
\begin{enumerate}
\item Download the most recent {\tt SARAH}\xspace and {\tt SPheno}\xspace versions into a directory {\tt \$PATH}. Both packages are located at {\tt HepForge}:
\begin{lstlisting}
http://sarah.hepforge.org/
http://spheno.hepforge.org/
\end{lstlisting}
\item Enter the directory and extract both codes
\begin{lstlisting}
> cd $PATH
> tar -xf SARAH-4.5.0.tar.gz
> tar -xf SPheno-3.3.3.tar.gz
\end{lstlisting}
\item Start \Mathematica, load {\tt SARAH}\xspace, run {\tt \$MODEL}, and generated the {\tt SPheno}\xspace output
\begin{lstlisting}
<< $PATH/SARAH-4.5.0/SARAH.m;
Start["$MODEL"];
MakeSPheno[];
\end{lstlisting}
\item Leave \Mathematica, enter the {\tt SPheno}\xspace directory and create a new sub-directory for your model
\begin{lstlisting}
> cd $PATH/SPheno-3.3.0
> mkdir $MODEL
> cp $PATH/SARAH-4.5.0/Output/$MODEL/EWSB/SPheno/* $MODEL
\end{lstlisting}
\item Compile {\tt SPheno}\xspace together with the new module
\begin{lstlisting}
> make Model=$MODEL
\end{lstlisting}
\end{enumerate}
After these steps a new binary {\tt bin/SPheno\$MODEL} is available. To run it an input file in the Les Houches format is needed.
{\tt SARAH}\xspace writes a template for that file which has to be filled with numbers. To enable the new functions for a calculation of the two-loop Higgs masses based on our new loop functions the following flags have to be set:
\begin{lstlisting}
Block SPhenoInput #
...
7 0 # Skip two loop masses: True/False
8 3 # Choose two-loop method
9 1 # Gaugeless limit: True/False
\end{lstlisting}
{\tt 8 -> 3} chooses the new approach to calculate the loop corrections. The other options for flag {\tt 8} would correspond the effective potential calculations based on {\tt SARAH}\xspace ({\tt 8->1} for a fully numerical derivation, {\tt 8->2} for a semi-analytical derivation). Also some hard-coded corrections are available which are based on results in literature: {\tt 8->8} uses the known $\alpha_S(\alpha_b + \alpha_t)$ corrections for the MSSM, NMSSM, TMSSM or any variant thereof with up to four neutral CP-even Higgs fields and including models with Dirac gauginos \cite{Goodsell:DG}. {\tt 8->9} uses the corrections of option 8 and adds the two-loop MSSM $(\alpha_t + \alpha_b + \alpha_\tau)^2$ results based on Refs.~\cite{Brignole:2001jy,Degrassi:2001yf,Brignole:2002bz,Dedes:2002dy,Dedes:2003km}. Note that the last two options are not included by default in the {\tt SPheno}\xspace output of {\tt SARAH}\xspace. To include them, the user must make sure to include in the {\tt SPheno.m} of the considered model
\begin{lstlisting}
Use2LoopFromLiterature = True;
\end{lstlisting}
Finally, {\tt SPheno}\xspace is executed by
\begin{lstlisting}
> ./bin/SPheno$MODEL $MODEL/LesHouches.in.$MODEL
\end{lstlisting}
and the output is written to
\begin{lstlisting}
SPheno.spc.$MODEL
\end{lstlisting}
\subsection{Validation}
We have intensively used the {\tt SPheno}\xspace output to validate our new two-loop functions, in particular:
\begin{itemize}
\item We found a numerical agreement of more than $10$ digits between our code and using public routines for the MSSM based on Refs.~\cite{Brignole:2001jy,Degrassi:2001yf,Brignole:2002bz,Dedes:2002dy,Dedes:2003km} for the self-energies. In order to perform this validation, it is necessary to use the same assumptions: turn off the first and second generation Yukawa couplings; take the Goldstone boson and light Higgs masses in the loops to be zero, and set the tree-level mixing angle of the neutral CP-even scalars to $\alpha=\beta-\pi/2 $. \\
The excellent agreement between all four possibilites to calculate the two-loop Higgs masses in the CMSSM is also shown in Fig.~\ref{fig:CMSSM}.The parameter point used here is the same as in \cite{Goodsell:2014bna},
\begin{equation}
\label{eq:CMSSMpoint}
M_0 = M_{1/2} = 1~\text{TeV},\quad A_0=-2~\text{TeV},\quad \tan\beta=10,\quad \text{sgn}(\mu)=1
\end{equation}
\begin{figure}[hbt]
\includegraphics[width=0.75\linewidth]{mh2L}
\caption{Comparison between the diagrammatic calculation of the two-loop Higgs masses with {\tt SARAH}\xspace/{\tt SPheno}\xspace presented here ({\tt diag}: diagrammatical) with the effective potential calculation of Ref.~\cite{Goodsell:2014bna} ({\tt p-num}: purely numerical, {\tt semi}: semi-analytical): , and with the routines based on Refs.~\cite{Brignole:2001jy,Degrassi:2001yf,Brignole:2002bz,Dedes:2002dy,Dedes:2003km} ({\tt ref}: reference). The fixed parameters are those in \qref{eq:CMSSMpoint}.}
\label{fig:CMSSM}
\end{figure}
\item Similarly, we found full agreement with available results for the $\alpha_s(\alpha_b+\alpha_t)$ corrections in the NMSSM \cite{Degrassi:2009yq} and for Dirac gauginos \cite{Goodsell:DG}.
\item We compared the full two-loop corrections, i.e. also including corrections not involving the strong interaction, for the NMSSM, models with Dirac gauginos as well as for the B-L-SSM \cite{O'Leary:2011yq} against the results using the other two options based on a completely independent implementation in {\tt SARAH}\xspace presented in Ref.~\cite{Goodsell:2014bna}. We usually found very good agreement. Tiny differences were based on numerical artefacts in the routines using the effective potential ansatz. Similarly, we could reproduce the results of Ref.~\cite{Dreiner:2014lqa} for the two-loop contributions to the Higgs mass stemming from $R$-parity violating couplings.
\end{itemize}
The new routines of course provide better stability. For example, in the routines based on numerically taking the derivatives of the potential, it is necessary to take care with the initial step size; if there are neutral scalars which have small expectation values then the results from those methods could become inaccurate -- this problem occurs in general for any neutral scalar having expectation value $v_i \ll M_{SUSY}$. Less significantly, the numerical method can suffer (small) errors when there are small couplings present, such that they do not induce a sufficient shift in particle masses or couplings upon variation of the Higgs vevs to accurately take the derivative. Hence, it is very important to have two independent implementations of generic two-loop Higgs mass calculations in {\tt SARAH}\xspace/{\tt SPheno}\xspace: this is the only possibility to cross check results for models beyond the (N)MSSM at the moment. Thus, we highly motivate users to test all options for the model under consideration and to compare the results.
\section{Summary}
\label{sec:summary}
We have presented the derivation of a new set of expressions for calculation of the tadpoles and self-energies at the two-loop level. These expressions include all generic diagrams which do not vanish in the gaugeless limit and are valid in the limit of zero external momenta. This set of loop functions is simpler than the set of expressions obtained by taking the limit $p^2 \to 0$ in the pole mass functions available in literature so far. This allows for a rapid numerical evaluation of the Higgs mass. We have implemented these functions in \texttt{Fortran}\xspace and included them in the new version of {\tt SARAH}\xspace\ {\tt 4.5.0}. This provides the possibility to automatically calculate the Higgs mass in a wide range of supersymmetric models with a guaranteed numerical accuracy and stability. The obtained precision for the Higgs mass is comparable with the one dedicated spectrum generators provide so far for the MSSM, and can now be applied to the study of a wide variety of models.
Aside from accuracy and stability, one of the principal advantages of this approach is that it is readily extendable. It would be straightforward to extend the calculation to non-zero momentum by changing the functions in the code and linking with the library {\tt TSIL}. On the other hand, including the electroweak contributions should be possible by applying these techniques to the full effective potential; we presented the expressions in this case for the tadpoles in the appendix, but the second derivatives are currently unknown -- as are the full set of equivalent expressions in the diagrammatic approach. Furthermore, to truly reach the full two-loop precision we would require the two-loop shift in the Z-mass that determines the electroweak expectation value. We hope to return to these issues in future.
\section*{Acknowledgement}
We are indebted to Pietro Slavich for a large number of fruitful discussions, and comments on the draft.
|
\section{Introduction}
To model a compact object it is generally assumed that the
underlying matter distribution is homogeneous i.e.
perfect fluid obeying Tolman-Oppenheimer-Volkoff (TOV) equation.
The nuclear matter of density $\rho \sim 10^{15}$~gm/cc, which is
expected at the core of the compact terrestrial object, becomes
anisotropic in nature as was first argued by Ruderman
\cite{Ruderman1972}. In case of anisotropy the pressure inside
the fluid sphere can specifically be decomposed into two parts:
radial pressure, $p_r$ and the transverse pressure, $p_t$, where $p_t$
is in the perpendicular direction to $p_r$. Their difference
$\Delta=p_t-p_r$ is defined as the anisotropic factor. Now, the
anisotropic force ($\frac{2\Delta}{r}$) will be repulsive in nature
if $\Delta>0$ or equivalently $p_t>p_r$ and attractive if $p_t<p_r$.
So it is reasonable to consider pressure anisotropy to develop
our model under investigation. It has been shown that in case of
anisotropic fluid the existence of repulsive force helps to construct
compact objects \cite{gokhroo1994}.
Anisotropy may occur for different reasons in any stellar distribution.
It could be introduced by the existence of the solid core or
for the presence of type $3A$ superfluid \cite{Kippenhahn1990}.
Different kinds of phase transitions \cite{sokolov1980}, pion condensation
~\cite{sawyer1972} etc. are also reasonable for anisotropy. It may also
occur by the effects of slow rotation in a star. Bowers and Liang
\cite{Bowers1974} showed that anisotropy might have
non-negligible effects on such parameters like equilibrium mass and
surface redshift. Very recently other theoretical advances also indicate
that the pressure inside a compact object is not essentially isotropic in
nature \cite{Varela2010,Rahaman2010a,Rahaman,Rahaman2012a,Kalam2012,Hossein2012,Kalam2013}.
In recent years the extension of General Relativity to higher
dimensions has become a topic of great interest. As a special
mention in this line of thinking we note that `Whether the usual
solar system tests are compatible with the existence of higher
spatial dimensions' has been investigated by Rahaman et
al.~\cite{Rahaman2}. Some other studies in higher dimension are
done by Liu and Overduin \cite{Liu} for the motion of test
particle whereas Rahaman et al.~\cite{Rahaman3} have investigated
higher dimensional gravastars.
One of the most interesting outcomes of string theory is that
the target spacetime coordinates become noncommuting operators on
$D$-brane~\cite{Witten1996,Seiberg1999}. Now the noncommutativity of
a spacetime can be encoded in the commutator
$\left[x^{\mu},x^{\nu}\right]=i \theta^{\mu \nu}$, where
$\theta^{\mu \nu}$ is an anti-symmetric matrix and is of dimension
$(length)^{2}$ which determines the fundamental cell
discretization of spacetime. It is similar to the way the
Planck constant $\hbar$ discretizes phase space~\cite{Smailagic2003}.
In the literature many studies are available on noncommutative
geometry, for example, Nazari and Mehdipour~\cite{Nozari2009} used Lorentzian
distribution to analyze `Parikh-Wilczek Tunneling' from
noncommutative higher dimensional black holes. Besides this
investigation some other noteworthy works are on galactic rotation
curves inspired by a noncommutative-geometry
background~\cite{Rahaman2012b}, stability of a particular class of
thin-shell wormholes in noncommutative geometry~\cite{Peter2012},
higher-dimensional wormholes with noncommutative geometry
\cite{Rahaman2012c}, noncommutative BTZ black hole
\cite{Rahaman2013}, noncommutative wormholes \cite{Rahaman2014a}
and noncommutative wormholes in $f(R)$ gravity with Lorentzian
distribution~\cite{Rahaman2014b}.
It is familiar to search for the natural relationship between
geometry and matter through the Einstein field equations where it
is very convenient to use inheritance symmetry. The well known
inheritance symmetry is the symmetry under conformal Killing
vectors (CKV) i.e.
\begin{equation}
L_{\xi} g_{ik}=\psi g_{ik},
\end{equation}
where $L$ is the Lie derivative of the metric tensor, which
describes the interior gravitational field of a stellar
configuration with respect to the vector field $\xi$, and $\psi$
is the conformal factor. It is supposed that the vector $\xi$
generates the conformal symmetry and the metric $g$ is conformally
mapped onto itself along $\xi$. It is to note that neither $\xi$
nor $\psi$ need to be static even though one considers a static
metric~\cite{Harko1,Harko2}. We also note that (i) if $\psi=0$
then Eq. (1) gives the Killing vector, (ii) if $\psi=$ constant it
gives homothetic vector and (iii) if $\psi=\psi(\textbf{x},t)$
then it yields conformal vectors. Moreover it is to be mentioned
that for $\psi=0$ the underlying spacetime becomes asymptotically
flat which further implies that the Weyl tensor will also vanish.
So CKV provides a deeper insight of the underlying spacetime
geometry.
A large number of works on conformal motion have been done
by several authors. A class of solutions for anisotropic stars admitting
conformal motion have been studied by Rahaman et
al.~\cite{Rahaman7}. In a very recent work Rahaman et
al.~\cite{Rahaman8} have also described conformal motion in higher
dimensional spacetimes. Charged gravastar admitting conformal
motion has been studied by Usmani et al.~\cite{Usmani2011}. Contrary to
this work Bhar \cite{piyali} has studied higher dimensional charged
gravastar admitting conformal motion whereas
relativistic stars admitting conformal motion has been analyzed by
Rahaman et al.~\cite{Rahaman2010b}. Inspired by these earlier works
on conformal motion we are looking forward for a new class of
solutions of anisotropic stars under the framework of
General Relativity inspired by noncommutative geometry
in four and higher dimensional spacetimes.
In the presence of noncommutative geometry there are two different
distributions available in the literature: (a) Gaussian and (b)
Lorentzian \cite{Mehdipour2012}. Though these two mass distributions
represent similar quantitative aspects, for the present investigation
we are exploiting a particular Lorentzian-type energy
density of the static spherically symmetric smeared and
particle-like gravitational source in the multi-dimensional
general form~\cite{Nozari2009,Mehdipour2012}
\begin{equation}
\rho=\frac{M \sqrt\phi}{\pi^{2}(r^{2}+\phi)^{\frac{n+2}{2}}},
\end{equation}
where $M$ is the total smeared mass of the source, $\phi$ is the
noncommutative parameter which bears a minimal width $\sqrt\phi$
and $n$ is positive integer greater than 1. In this approach,
generally known as the {\it noncommutative geometry inspired
model}, via a minimal length caused by averaging noncommutative
coordinate fluctuations cures the curvature singularity in black
holes
\cite{Smailagic2003,Spallucci2006,Banerjee2009,Modesto2010,Nicolini2011}.
It has been argued that it is not required to consider the length
scale of the coordinate noncommutativity to be the same as the
Planck length as the noncommutativity influences appears on a
length scale which can behave as an adjustable parameter
corresponding to that pertinent scale \cite{Mehdipour2012}.
It is interesting to note that Rahaman et al. \cite{Rahaman7}
have found out a new class of interior solutions for anisotropic
compact stars admitting conformal motion under $4D$ framework of GR.
On the other hand, Rahaman et al. \cite{Rahaman8} have studied
different dimensional fluids, higher as well as lower, inspired by
noncommutative geometry with Gaussian distribution of energy density
and have shown that at $4D$ only one can get a stable configuration
for any spherically symmetric stellar system. However, in the present
work we have extended the work of Rahaman et al. \cite{Rahaman7} to
higher dimensions and that of Rahaman et al. \cite{Rahaman8} to
higher dimensions with energy density in the form of Lorentzian
distribution. In this approach we are able to generalize both
the above mentioned works to show that compacts stars
may exist even in higher dimensions.
In this paper, therefore we use noncommutative geometry inspired model
to combine the microscopic structure of spacetime with the relativistic
description of gravity. The plan of the present investigation is as follows:
in Section 2 we formulate the Einstein field equations for the interior
spacetime of the anisotropic star. In Section 3 we solve the
Einstein field Equations by using the density function of
Lorentzian distribution type in higher dimensional spacetime as
given by Nozari and Mehdipour~\cite{Nozari2009}. We consider
the cases $n=~2,~3,~4$ and $9$ i.e. $4D,~5D,~6D$ and $11D$
spacetimes in Section 4 to examine expressions for physical
parameters whereas the matching conditions are provided in Section 5.
Various physical properties are explored in Section 6 with
interesting features of the model and present them with
graphical plots for comparative studies among the results of
different dimensional spacetimes. Finally we complete the paper
with some concluding remarks in Section 7.
\section{The interior spacetime and the Einstein field equations}
To describe the static spherically symmetric spacetime (in
geometrical units $G=1=c$ here and onwards) in higher dimension the
line element can be given in the standard form
\begin{equation}
ds^{2}=-e^{\nu(r)}dt^{2}+e^{\lambda(r)}dr^{2}+r^{2}d\Omega_n^{2},
\end{equation}
where
\begin{equation}
d\Omega_n^{2}=d\theta_1^{2}+\sin^{2}\theta_1d\theta_2^{2}+
\sin^{2}\theta_1\sin^{2}\theta_2d\theta_3^{2}+...+\prod_{j=1}^{n-1}\sin^{2}\theta_jd\theta_n^{2},
\end{equation}
where $\lambda$, $\nu$ are functions of the radial coordinate $r$.
Here we have used the notation $D=n+2$,~$D$ is the dimension of
the spacetime.
The energy momentum tensor for the matter distribution can be
taken in its usual form \cite{Lobo}
\begin{equation}
T_{\nu}^{\mu}=(\rho+p_r)u^{\mu}u_{\nu}-p_rg^{\mu}_{\nu}+(p_t-p_r)\eta^{\mu}\eta_{\nu},
\end{equation}
with $u^{\mu}u_{\mu}=-\eta^{\mu}\eta_{\mu}=1$ and
$u^{\mu}\eta_{\nu}= 0$. Here the vector $u^{\mu}$ is the fluid
$(n+2)$-velocity and $\eta^{\mu}$ is the unit space-like vector
which is orthogonal to $ u^{\mu}$, where $\rho$ is the matter
density, $p_r$ is the radial pressure in the direction of
$\eta^{\mu}$ and $p_t$ is the transverse pressure in the
orthogonal direction to $p_r$. Since the pressure is anisotropic
in nature so for our model $p_r \neq p_t$. Here $p_t-p_r = \Delta$
is the measure of anisotropy, as defined earlier.
Now for higher ( $n \geq 2 )$ dimensional spacetime the Einstein
equations can be written as \cite{Rahaman2012b}
\begin{equation}
e^{-\lambda}\left[\frac{n\lambda'}{2r}-\frac{n(n-1)}{2r^{2}}\right]+\frac{n(n-1)}{2r^{2}}=8\pi
\rho=8\pi T_0^{0},
\end{equation}
\begin{equation}
e^{-\lambda}\left[\frac{n(n-1)}{2r^{2}}+\frac{n\nu'}{2r}\right]-\frac{n(n-1)}{2r^{2}}=8\pi
p_r=-8\pi T_1^{1},
\end{equation}
\[\frac{1}{2}e^{-\lambda}\left[\frac{1}{2}(\nu')^{2}+\nu''-\frac{1}{2}\lambda'\nu'+\frac{(n-1)}{r}(\nu'-\lambda')
+\frac{(n-1)(n-2)}{r^{2}}\right]\]
\begin{equation}
-\frac{(n-1)(n-2)}{2r^{2}}=8\pi p_t=-8\pi T_2^{2}=-8\pi T_3^{3},
\end{equation}
where $\prime$ denotes differentiation with respect to the radial
coordinate $r$ i.e. $\prime \equiv \frac{d}{dr}$.
\section{The solution under conformal Killing vector}
Mathematically, conformal motions or conformal Killing vectors
(CKV) are motions along which the metric tensor of a spacetime
remains invariant up to a scale factor. A conformal vector field
can be defined as a global smooth vector field $x$ on a manifold,
$\ss$, such that for the metric $g_{ab}$ in any coordinate system
on $\ss$ $x_{a;b} = \psi g_{ab} + F_{ab}$, where $\psi : \ss
\rightarrow real ~number$ is the smooth conformal function of
$x$, $F_{ab}$ is the conformal bivector of $x$. This is
equivalent to $L_x g_{ik} = \psi g_{ik}$, (as considered in Eq.
(1) in the usual form) where $L_x$ signifies the Lie derivatives
along $x_a$.
To search the natural relation between geometry and matter through
the Einstein equations, it is useful to use inheritance symmetry.
The well known inheritance symmetry is the symmetry under
conformal Killing vectors (CKV). These provide a deeper insight into
the spacetime geometry. The CKV facilitate generation
of exact solutions to the Einstein's field equations. The study of
conformal motions in spacetime is physically very important
because it can lead to the discovery of conservation laws and devise
spacetime classification schemes. Einstein's field equations being
highly non linear partial differential equations, one can reduce
the partial differential equations to ordinary differential equations
by use of CKV. It is still a challenging problem to the theoretical
physicists to know the exact nature and characteristics of compact stars
and elementary particle like electron.
Let us therefore assume that our static spherically symmetry spacetime
admits an one parameter group of conformal motion. The conformal Killing vector,
as given in Eq. (1), can be written in a more convenient form:
\begin{equation}
L_\xi g_{ik}=\xi_{i;k}+\xi_{k;i}=\psi g_{ik},
\end{equation}
where both $i$ and $k$ take the values $ ~1,~2...,~n+2$. Here
$\psi$ is an arbitrary function of the radial coordinate $r$ and
$\xi$ is the orbit of the group. The metric $g_{ij}$ is
conformally mapped onto itself along $\xi_i$.
Let us further assume that the orbit of the group to be orthogonal
to the velocity vector field of the fluid,
\begin{equation}
\xi^{\mu}u_{\mu}=0.
\end{equation}
As a consequence of the spherically symmetry from Eq. (10) we have
\[\xi^{1}=\xi^{3}=...=\xi^{n+1}=0.\] Now the conformal Killing equation
for the line element (3) gives the following equations:
\begin{equation}
\xi^{2}\nu'=\psi,
\end{equation}
\begin{equation}
\xi^{n+2}=C_1,
\end{equation}
\begin{equation}
\xi^{2}=\frac{\psi r}{2},
\end{equation}
\begin{equation}
\xi^{2}\lambda'+2{\xi^{2}}^\prime =\psi,
\end{equation}
where $2$ stands for the spatial coordinates $r$,`$\prime$' and
`$,$' denotes the partial derivative with respect to $r$ and $C_1$
is a constant.
The above set of equations consequently gives
\begin{equation}
e^{\nu}=C_2^{2}r^{2},
\end{equation}
\begin{equation}
e^{\lambda}=\left(\frac{C_3}{\psi}\right)^{2},
\end{equation}
\begin{equation}
\xi^{i}=C_1\delta_{n+2}^{i}+\left( \frac{\psi
r}{2}\right)\delta_2^{i},
\end{equation}
where $\delta$ stands for the `Kronecker delta'
and $C_2$,~$C_3$ are constants of integrations.
Using Eqs. (15)-(17) in the Einstein field Eqs. (6)-(8), we get
\begin{equation}
\frac{n(n-1)}{2r^{2}}\left(1-\frac{\psi^{2}}{C_3^{2}}\right)-\frac{n\psi
\psi'}{rC_3^{2}}=8\pi \rho,
\end{equation}
\begin{equation}
\frac{n}{2r^{2}}\left[(n+1)\frac{\psi^{2}}{C_3^{2}}-(n-1)\right]=8\pi
p_r,
\end{equation}
\begin{equation}
\frac{n\psi\psi'}{rC_3^{2}}+n(n-1)\frac{\psi^{2}}{2r^{2}C_3^{2}}-\frac{(n-1)(n-2)}{2r^{2}}=8\pi
p_t.
\end{equation}
We thus have three independent Eqs. (18)-(20)
with four unknowns $\rho$,~$p_r$,~$p_t$,~$\psi$. So we
are free to choose any physically reasonable {\it ansatz} for any
one of these four unknowns. Hence we choose density profile $\rho$ in
the form given in Eq. (2) in connection to higher dimensional
static and spherically symmetric Lorentzian distribution of
smeared matter as provided by Nozari and Mehdipour~\cite{Nozari2009}.
This density profile will be employed as a key tool in our present
study.
Therefore, substituting Eq. (2) into (18) and solving, we
obtain
\begin{equation}
\psi^{2}=C_3^{2}-\frac{16MC_3^{2}\sqrt{\phi}}{n\pi}\frac{1}{r^{n-1}}\int\frac{r^{n}}{(r^{2}+\phi)^{\frac{n+2}{2}}}dr+\frac{A}{r^{n-1}},
\end{equation}
where $A$ is a constant of integration which is determined by
invoking suitable boundary conditions.
Now the equation (21) gives the expression of the conformal factor $\psi$.
Assigning $n=~2,~3,~4$ and $9$ i.e. $4D,~5D,~6D$ and $11D$ spacetimes
respectively if we perform the above integral then the conformal factor
$\psi$ can be obtained for different dimension that is necessary to
find out the other physical parameter namely $p_r$ and $p_t$ for these dimension.
Here $A,~-\infty <A <\infty$, is a constant of integration that can be later on found
out from the boundary condition $p_r(R)=0$,~$R$ being the radius of the star.
\section{Exact solutions of the models in different dimensions}
The above set of equations are associated with dimensional
parameter $n$ and hence to get a clear picture of the physical
system under different spacetimes we are interested for studying
several cases starting from standard $4D$ to higher $5D$, $6D$
and $11D$ spacetimes as shown below.
\subsection{Four dimensional spacetime ($n=2$)}
The conformal parameter $\psi(r)$ and the metric potential
$e^{\lambda}$ are given as,
\begin{equation}
\psi=\sqrt{C_3^{2}-\frac{4MC_3^{2}\sqrt{\phi}}{\pi
r}\left[\frac{1}{\sqrt{\phi}}\arctan\left(\frac{r}{\sqrt{\phi}}\right)
-\frac{r}{r^{2}+\phi} \right]+\frac{A}{r}},
\end{equation}
\begin{equation}
e^{-\lambda}=1+\frac{A}{C_3^{2}r}-\frac{4M\sqrt{\phi}}{\pi
r}\left[\frac{1}{\sqrt{\phi}}\arctan\left(\frac{r}{\sqrt{\phi}}\right)
-\frac{r}{r^{2}+\phi}\right].
\end{equation}
The radial and transverse pressures are obtained as
\begin{equation}
p_r=\frac{1}{8\pi
r^{2}}\left[2+\frac{3A}{C_3^{2}r}-\frac{12M\sqrt{\phi}}{\pi
r}\left\{\frac{1}{\sqrt{\phi}}\arctan\left(\frac{r}{\sqrt{\phi}}\right)
-\frac{r}{r^{2}+\phi} \right\} \right],
\end{equation}
\begin{equation}
p_t=\frac{1}{8\pi}\left[\frac{1}{r^{2}}-\frac{8M\sqrt{\phi}}{\pi(r^{2}+\phi)^{2}}\right].
\end{equation}
To find the above constant of integration we impose the boundary
condition $p_r(r=R)=0$, where $R$ is the radius of the fluid
sphere as mentioned earlier, which gives
\begin{equation}
A=\frac{4MC_3^{2}\sqrt{\phi}}{\pi}\left\{
\frac{1}{\sqrt{\phi}}\arctan\left(\frac{R}{\sqrt{\phi}}\right)
-\frac{R}{R^{2}+\phi}\right\}-\frac{2}{3}C_3^{2}R.
\end{equation}
\subsection{Five dimensional spacetime ($n=3$)}
In this case the solution set can be obtained as follows:
\begin{equation}
\psi=\sqrt{C_3^{2}+\frac{16MC_3^{3}\sqrt{\phi}}{9\pi
r^{2}}\frac{3r^{2}+2\phi}{(r^{2}+\phi)^{\frac{3}{2}}}+\frac{A}{r^{2}}},
\end{equation}
\begin{equation}
e^{-\lambda}=1+\frac{16M\sqrt{\phi}}{9\pi
r^{2}}\frac{3r^{2}+2\phi}{(r^{2}+\phi)^{\frac{3}{2}}}+\frac{A}{C_3^{2}r^{2}},
\end{equation}
\begin{equation}
p_r=\frac{3}{8\pi r^{2}}\left[1+\frac{32M\sqrt{\phi}}{9\pi
r^{2}}\frac{3r^{2}+2\phi}{(r^{2}+\phi)^{\frac{3}{2}}}+\frac{2A}{C_3^{2}r^{2}}\right],
\end{equation}
\begin{equation}
p_t=\frac{1}{8\pi}\left[\frac{2}{r^{2}}-\frac{8M\sqrt{\phi}}{\pi(r^{2}+\phi)^{\frac{5}{2}}}\right],
\end{equation}
with
\begin{equation}
A=-\frac{C_3^{2}R^{2}}{2}\left[1+\frac{32M\sqrt{\phi}}{9\pi
R^{2}}\frac{3R^{2}+2\phi}{(R^{2}+\phi)^{\frac{3}{2}}}\right].
\end{equation}
\subsection{Six dimensional spacetime ($n=4$)}
Here the solutions are as follows:
\begin{equation}
\psi=\sqrt{C_3^{2}-\frac{MC_3^{2}\sqrt{\phi}}{2\pi
r^{3}}\left[\frac{3}{\sqrt{\phi}}\tan^{-1}\left(\frac{r}{\sqrt{\phi}}\right)
-\frac{5r^{3}+3r\phi}{(r^{2}+\phi)^{2}}
\right]+\frac{A}{r^{3}}},
\end{equation}
\begin{equation}
e^{-\lambda}=1-\frac{M \sqrt{\phi}}{2\pi
r^{3}}\left[\frac{3}{\sqrt{\phi}}\tan^{-1}\left(\frac{r}{\sqrt{\phi}}\right)
-\frac{5r^{3}+3r\phi}{(r^{2}+\phi)^{2}}
\right]+\frac{A}{C_3^{2}r^{3}},
\end{equation}
\begin{equation}
p_r=\frac{1}{4 \pi r^{2}}\left[2-\frac{5M\sqrt{\phi}}{2\pi
r^{3}}\left\{\frac{3}{\sqrt{\phi}}\tan^{-1}\left(\frac{r}{\sqrt{\phi}}\right)
-\frac{5r^{3}+3r\phi}{(r^{2}+\phi)^{2}}
\right\}+\frac{5A}{r^{3}C_3^{2}}\right],
\end{equation}
\begin{equation}
p_t=\frac{1}{8\pi}\left[\frac{3}{r^{2}}-\frac{8M\sqrt{\phi}}{\pi(r^{2}+\phi)^{3}}\right],
\end{equation}
and
\begin{equation}
A=\frac{MC_3^{2}\sqrt{\phi}}{2\pi}\left\{\frac{3}{\sqrt{\phi}}\tan^{-1}\left(\frac{R}{\sqrt{\phi}}\right)
-\frac{5R^{3}+3r\phi}{(R^{2}+\phi)^{2}}
\right\}-\frac{2C_3^{2}R^{3}}{5}.
\end{equation}
\subsection{Eleven dimensional spacetime ($n=9$)}
For this arbitrarily chosen higher dimension the solutions can be
obtained as
\begin{equation}
\psi=\sqrt{C_3^{2}+C_3^{2}\frac{16M\sqrt{\phi}}{2835\pi
r^{8}}\frac{315r^{8}+840r^{6}\phi+1008r^{4}\phi^{2}+576r^{2}\phi^{3}+128\phi^{4}}{(r^{2}
+\phi)^{\frac{9}{2}}}+\frac{A}{r^{8}}},
\end{equation}
\begin{equation}
e^{-\lambda}=1+\frac{16M\sqrt{\phi}}{2835\pi
r^{8}}\frac{315r^{8}+840r^{6}\phi+1008r^{4}\phi^{2}+576r^{2}\phi^{3}+128\phi^{4}}{(r^{2}
+\phi)^{\frac{9}{2}}}+\frac{A}{C_3^{2}r^{8}},
\end{equation}
\begin{equation}
p_r=\frac{9}{8\pi r^{2}}\left[1+\frac{16M\sqrt{\phi}}{567\pi
r^{8}}\frac{315r^{8}+840r^{6}\phi+1008r^{4}\phi^{2}+576r^{2}\phi^{3}+128\phi^{4}}{(r^{2}
+\phi)^{\frac{9}{2}}}+\frac{5A}{C_3^{2}r^{8}}\right],
\end{equation}
\begin{equation}
p_t=\frac{1}{\pi}\left[\frac{1}{r^{2}}-\frac{M\sqrt{\phi}}{\pi(r^{2}+\phi)^{\frac{11}{2}}}\right],
\end{equation}
and
\begin{equation}
A=-\frac{C_3^{2}}{5}\left[R^{8}+\frac{16M\sqrt{\phi}}{567\pi}\frac{315R^{8}+840R^{6}\phi+1008R^{4}\phi^{2}
+576R^{2}\phi^{3}+128\phi^{4}}{(R^{2}
+\phi)^{\frac{9}{2}}}\right].
\end{equation}
\begin{figure}[h]
\centering
\includegraphics[width=0.4\textwidth]{pr4.eps}
\caption{The graphical plot for radial pressure vs radius which
has a definite cut-off at $4.17$~km}
\end{figure}
Let us now turn our attention to the physical analysis of the stellar
model under consideration, i.e. whether it is a normal star or
something else. To do so, primarily we try to
figure out the radius of the stellar configuration. It is to note
that in Eqs. (29), (34) and (39) the radius of the star $R$ has
been mentioned under the boundary condition, $p_r(R) =0$, i.e. we
get analytical results in the respective cases. So it seems that
we can proceed on without further plot descriptions to get $p_r(R)
=0$ for all dimensions. However, for $4D$ case it reveals that the
radius of the star is very small with a numerical value of
$4.17$~km (Fig. 1). This obviously then indicates that the star is
nothing but a compact object (see Table 1 of all the Refs.
\cite{Rahaman,Kalam2012,Hossein2012,Kalam2013}
for comparison with the radius of some of the real compact stars).
\section{Matching conditions}
Now, we match our interior solutions with the exterior vacuum
solutions. The generalization of Schwarzschild solution, which as
obtained by Tangherlini \cite{Tangherlini1963} reads as
\begin{equation}
ds ^2 = - \left(1 - \frac{\mu_{n}}{ r^{n-1}}\right) dt^2 +
\left(1 - \frac{\mu_{n}}{ r^{n-1}}\right)^{-1} dr^2 + d
\Omega_{n} ^2.\label{eq16}
\end{equation}
Here, \[ {\Omega}_{ n }=\frac{2 \pi^{\frac{n+1}{2}}}{\Gamma\left(
\frac{n+1}{2}\right)},\] the area of a unit $n$-sphere and
\[\mu_{n}=16\pi GM/n c^2 {\Omega}_{n},\]
is the constant of integration with $M$, the mass of the black hole with $n=2,~3,~4,~9$.
\subsection{Four dimensional spacetime ($n=2$)}
For $4D$ case, our interior solution should match to the
exterior Schwarzschild spacetime at the boundary $r=a_4$ given by
\begin{equation}
ds^{2}=-\left(1-\frac{\mu_2}{r}\right)dt^{2}+\left(1-\frac{\mu_2}
{r}\right)^{-1}dr^{2}+r^{2}d\Omega_{2}.
\end{equation}
Now using the matching conditions at the boundary $r=a_4$, we have
\begin{equation}
1-\frac{\mu_2}{a_4}=C_2^{2}a_4^{2},
\end{equation}
and
\begin{equation}
1-\frac{\mu_2}{a_4}=1+\frac{A}{C_3^{2}a_4}-\frac{4M\sqrt{\phi}}{\pi
a_4}\left[\frac{1}{\sqrt{\phi}}\arctan\left(\frac{a_4}{\sqrt{\phi}}\right)
-\frac{a_4}{a_4^{2}+\phi}\right].
\end{equation}
Solving the above two equations, we obtain
\begin{equation}
C_2^{2}=\frac{1}{a_4^{2}}\left(1-\frac{\mu_2}{a_4}\right),
\end{equation}
\begin{equation}
\frac{A}{C_3^{2}}=\frac{4M\sqrt{\phi}}{\pi
}\left[\frac{1}{\sqrt{\phi}}\arctan\left(\frac{a_4}{\sqrt{\phi}}\right)-\frac{a_4}{a_4^{2}+\phi}\right]-\mu_2.
\end{equation}
\subsection{Five dimensional spacetime ($n=3$)}
For $5D$ case, our interior solution should match to the exterior $5D$
Schwarzschild spacetime at the boundary $r=a_5$, given by
\begin{equation}
ds^{2}=-\left(1-\frac{\mu_3}{r^{2}}\right)dt^{2}+\left(1-\frac{\mu_3}{r^{2}}\right)^{-1}dr^{2}
+r^{2}d\Omega_{3}.
\end{equation}
Now using the matching conditions at the boundary $r=a_5$, we have
\begin{equation}
1-\frac{\mu_3}{a_5^{2}}=C_2^{2}a_5^{2},
\end{equation}
and
\begin{equation}
1-\frac{\mu_3}{a_5^{2}}=1+\frac{A}{C_3^{2}a_5^{2}}+\frac{16M\sqrt{\phi}}{9\pi
a_5^{2}}\frac{3a_5^{2}+2\phi}{(a_5^{2}+\phi)^{\frac{3}{2}}}.
\end{equation}
Solving the above two equations, we obtain
\begin{equation}
C_3^{2}=\frac{1}{a_5^{2}}\left(1-\frac{\mu_3}{a_5^{2}}\right),
\end{equation}
and
\begin{equation}
\frac{A}{C_3^{2}}=-\mu_3-\frac{16M\sqrt{\phi}}{9\pi}\frac{3a_5^{2}+2\phi}{(a_5^{2}
+\phi)^{\frac{3}{2}}}.
\end{equation}
\subsection{Six dimensional spacetime ($n=4$)}
For $6D$ case, our interior solution should match to the exterior $6D$
Schwarzschild spacetime at the boundary $r=a_6$, given by
\begin{equation}
ds^{2}=-\left(1-\frac{\mu_4}{r^{3}}\right)dt^{2}+\left(1-\frac{\mu_4}{r^{3}}\right)^{-1}dr^
{2}+r^{2}d\Omega_{4}.
\end{equation}
Now using the matching conditions at the boundary $r=a_6$, we have
\begin{equation}
1-\frac{2\mu_4}{a_6^{3}}=C_2^{2}a_6^{2},
\end{equation}
and
\begin{equation}
1-\frac{\mu_4}{a_6^{3}}=1-\frac{M \sqrt{\phi}}{2\pi
a_6^{3}}\left[\frac{3}{\sqrt{\phi}}\tan^{-1}\left(\frac{a_6}{\sqrt{\phi}}\right)
-\frac{5a_6^{3}+3a_6\phi}{(a_6^{2}+\phi)^{2}}
\right]+\frac{A}{C_3^{2}a_6^{3}}.
\end{equation}
Solving the above two equations, we obtain
\begin{equation}
C_2^{2}=\frac{1}{a_6^{2}}\left(1-\frac{\mu_4}{a_6^{3}}\right),
\end{equation}
and
\begin{equation}
\frac{A}{C_3^{2}}=\frac{M
\sqrt{\phi}}{2\pi}\left[\frac{3}{\sqrt{\phi}}\tan^{-1}\left(\frac{a_6}{\sqrt{\phi}}\right)
-\frac{5a_6^{3}+3a_6\phi}{(a_6^{2}+\phi)^{2}} \right]-\mu_4.
\end{equation}
\subsection{Eleven dimensional spacetime ($n=9$)}
For $11D$ case, our interior solution should match to the exterior
11D Schwarzschild spacetime at the boundary $r=a_{11}$, given by
\begin{equation}
ds^{2}=-\left(1-\frac{\mu_{9}}{r^{8}}\right)dt^{2}+\left(1-\frac{\mu_{9}}{r^{8}}\right)^{-1}
dr^{2}+r^{2}d\Omega_{9}.
\end{equation}
Now using the matching conditions at the boundary $r=a_{11}$, we have
\begin{equation}
1-\frac{\mu_{9}}{a_{11}^{8}}=C_2^{2}a_{11}^{2},
\end{equation}
and
\begin{equation}
1-\frac{\mu_9}{a_{11}^{8}}=1+\frac{16M\sqrt{\phi}}{2835\pi
a_{11}^{8}}\frac{315a_{11}^{8}+840a_{11}^{6}\phi+1008a_{11}^{4}\phi^{2}
+576a_{11}^{2}\phi^{3}+128\phi^{4}}{(a_{11}^{2}
+\phi)^{\frac{9}{2}}}+\frac{A}{C_3^{2}a_{11}^{8}}.
\end{equation}
Solving the above two equations, we obtain
\begin{equation}
C_2^{2}=\frac{1}{a_{11}^{2}}\left(1-\frac{\mu_{9}}{a_{11}^{8}}\right),
\end{equation}
and
\begin{equation}
\frac{A}{C_3^{2}}=\frac{M
\sqrt{\phi}}{2\pi}\left[\frac{3}{\sqrt{\phi}}\tan^{-1}\left(\frac{a_{11}}{\sqrt{\phi}}\right)
-\frac{5a_{11}^{3}+3a_{11}\phi}{(a_{11}^{2}+\phi)^{2}}
\right]-\mu_{9},
\end{equation}
where $a_j$ (j=4,~5,~6~and~11) are the radii of the fluid spheres
in different dimensions.
\section{A comparative study of the physical features of the model}
Let us now carry out a comparative study of the physical features
based on the solutions set obtained in the previous Section 4.
This can be done in different ways. However, in the present
investigation the best method we may adopt for comparative study,
firstly, in connection to stability of the models for different
dimensions which may be considered as most crucial one and
secondly, for other physical parameters viz., density, pressure,
pressure anisotropy, pressure gradient, conformal parameter and
metric potential.
\subsection{Stability of the stellar configuration}
The Generalized Tolman-Oppenheimer-Volkoff (TOV) equation can be
written in the form
\begin{equation}
-\frac{M_G(r)(\rho+p_r)}{r^2}e^{\frac{\nu-\mu}{2}}-\frac{dp_r}{dr}+\frac{2}{r}(p_t-p_r)=0,
\end{equation}
where $M_G(r) $ is the gravitational mass within the sphere of radius $r$
and is given by
\begin{equation}
M_G(r)=\frac{1}{2}r\nu'e^{\frac{\mu-\nu}{2}}.
\end{equation}
Substituting (64) into (63), we obtain
\begin{equation}
-\frac{\nu'}{2}(\rho+p_r)-\frac{dp_r}{dr}+\frac{2}{r}(p_t-p_r)=0.
\end{equation}
The above TOV equation describe the equilibrium of the stellar
configuration under gravitational force $F_g$, hydrostatic force
$F_h$ and anisotropic stress $F_a$ so that we can write it in the
following form:
\begin{equation}
F_g+F_h+F_a=0,
\end{equation}
where
\[F_g=-\frac{\nu'}{2}(\rho+p_r),\]
\[F_h=-\frac{dp_r}{dr},\]
\[F_a=\frac{2}{r}(p_t-p_r).\]
\begin{equation}
\end{equation}
\begin{figure}[h]
\centering
\includegraphics[width=0.4\textwidth]{newtov2.eps}
\includegraphics[width=0.4\textwidth]{newtov3.eps}
\includegraphics[width=0.4\textwidth]{newtov4.eps}
\includegraphics[width=0.4\textwidth]{newtov9.eps}
\caption{The three different forces namely gravitational forces
$(F_g)$, hydrostatic forces $(F_h)$ and anisotropic forces $(F_a)$
are plotted against $r$~(km) for $4D$ spacetime (top left), $5D$
spacetime (top right), $6D$ spacetime (bottom left) and $11D$
spacetime (bottom right). From the figure balancing status of the
forces can be clearly observed for $4D$,~$5D$,~$6D$,~$11D$
spacetimes}
\end{figure}
We have shown the plots of TOV equations for $4D$, $5D$, $6D$ and
$11D$ spacetime in Fig. 2. From the plots it is overall clear that
the system is in static equilibrium under three different forces, viz.
gravitational, hydrostatic and anisotropic, for example, in the $4D$ case
to attain equilibrium, the hydrostatic force is counter
balanced jointly by gravitational and anisotropic forces. In $5D$
also the situation is exactly same, the only difference being in
the radial distances. In $4D$ it is closer to 5 whereas in
$5D$ it is closer to 8. This distance factor can
also be observed in the higher dimensional spacetimes
though the balancing features between the three forces are clearly
different in the respective cases.
\subsection{Energy conditions}
Now we check whether all the energy conditions are satisfied or
not. For this purpose, we shall consider the following inequalities:
\[
(i)~NEC: \rho+p_r\geq 0,~\rho+p_t\geq 0,
\]
\[
(ii)~WEC: \rho+p_r\geq 0,~\rho\geq 0,~\rho+p_t\geq 0,
\]
\[
(iii)~SEC: \rho+p_r\geq 0,~\rho+p_r+2p_t\geq 0.
\]
Fig. 3 indicates that in our model all the energy conditions are
satisfied through out the interior region.
\begin{figure}[h]
\centering
\includegraphics[width=0.4\textwidth]{ec2.eps}
\includegraphics[width=0.4\textwidth]{ec3.eps}
\includegraphics[width=0.4\textwidth]{ec4.eps}
\includegraphics[width=0.4\textwidth]{ec9.eps}
\caption{The energy conditions in the interior have been
plotted against $r$ for $4D$ spacetime (top left), $5D$ spacetime
(top right), $6D$ spacetime (bottom left) and $11D$ spacetime
(bottom right) }
\end{figure}
\subsection{Anisotropy of the models}
We have shown the possible variation of radial and transverse pressures in Fig. 4
(top left and right of the panel respectively). Hence the measure of anisotropy
$\Delta=(p_t-p_r)$ in $4$, $5$, $6$ and $11$
dimensional cases are respectively given as
\begin{equation}
\Delta_4=\frac{1}{8\pi}\left[\frac{12M\sqrt{\phi}}{\pi
r^{3}}\left\{\frac{1}{\sqrt{\phi}}\arctan\left(\frac{r}{\sqrt{\phi}}\right)-\frac{r}{r^{2}+\phi}\right\}-\frac{1}{r^{2}}
-\frac{8M\sqrt{\phi}}{\pi(r^{2}+\phi)^{2}}-\frac{3A}{C_3^{2}r^{3}}\right],
\end{equation}
\begin{equation}
\Delta_5=\frac{1}{8\pi}\left[\frac{32M\sqrt{\phi}}{3\pi
r^{4}}\left\{\frac{3r^{2}
+2\phi}{(r^{2}+\phi)^{\frac{3}{2}}}\right\}-\frac{1}{r^{2}}
-\frac{8M\sqrt{\phi}}{\pi(r^{2}+\phi)^{\frac{5}{2}}}+\frac{6A}{C_3^{2}r^{4}}\right],
\end{equation}
\begin{equation}
\Delta_6=\frac{1}{8\pi}\left[\frac{5M\sqrt{\phi}}{\pi
r^{5}}\left\{\frac{3}{\sqrt{\phi}}\tan^{-1}\left(\frac{r}{\sqrt{\phi}}\right)
-\frac{5r^{3}+3r\phi}{(r^{2}+\phi)^{2}}
\right\}-\frac{1}{r^{2}}-\frac{8M\sqrt{\phi}}{\pi(r^{2}+\phi)^{3}}-\frac{10A}{r^{5}C_3^{2}}\right],
\end{equation}
\begin{eqnarray}
\Delta_{11}= \frac{1}{8\pi}\left[\frac{M\sqrt{\phi}}{4\pi
r^{10}}\left\{\frac{315r^{8}+840r^{6}\phi+1008r^{4}\phi^{2}+576r^{2}\phi^{3}+128\phi^{4}}{(r^{2}
+\phi)^{\frac{9}{2}}}\right\} \right. \nonumber\\
- \left. \frac{1}{r^{2}}-\frac{8M\sqrt{\phi}}{\pi(r^{2}+\phi)^{\frac{11}{2}}}+\frac{45A}{C_3^{2}r^{10}}\right],
\end{eqnarray}
All these are plotted in Fig. 4 (bottom left of the panel). From all the plots
we see that $\Delta<0$ ~i.e., $p_t~<p_r$ and hence the anisotropic
force is attractive in nature. A detailed study shows that
firstly, in every case of different dimensions the measure of
anisotropy is a decreasing function of $r$. Secondly, from $4D$
onward measure of anisotropy is increasing gradually and
is attaining maximum at $5D$. Surprisingly, it is very high
compared to $4D$ and $11D$ spacetimes. This observation therefore
dictates that $4D$ configuration represents almost a spherical
object as departure from isotropy is very less than the higher
dimensional spacetimes.
Moreover, in all the above cases of different dimension one can
note that the pressure gradient $\frac{dp_r}{dr}$ is a decreasing
function of $r$ (bottom right panel of Fig. 4).
\begin{figure}[h]
\centering
\includegraphics[width=0.4\textwidth]{pr.eps}
\includegraphics[width=0.4\textwidth]{pt.eps}
\includegraphics[width=0.4\textwidth]{delta.eps}
\includegraphics[width=0.4\textwidth]{de.eps}
\caption{Variation of the radial pressure $p_r$ in the interior of
the compact star in $4D$,~$5D$,~$6D$,~$11D$ have been plotted
against $r$ (Left). The figures show that profiles of $p_r$ are
monotonic decreasing function of $r$. The anisotropic factor
$\Delta$ for $4D$,~$5D$,~$6D$,~$11D$ cases are shown against $r$
(right). In both the cases the suffixes in the legends indicate
the dimension of the spacetime}
\end{figure}
\subsection{Compactness and redshift of the star}
At the end of previous Section we did a primary test to get a preliminary
idea about the structure of the star under consideration and we have seen
that the star actually represents a compact object with a radius
$4.17$~km. However, for further test of confirmation one can
perform some specific calculations for `compactness factor'
\cite{Rahaman,Kalam2012,Hossein2012,Kalam2013}.
To do so we first define gravitational mass of the system of
matter distribution as follows:
\begin{equation}
m(r) = \int_0^{ r}~ \left[ \frac{2 \pi^{\frac{n+1}{2}} }{\Gamma
\left(\frac{n+1}{2}\right)}\right]r^n \rho dr.
\end{equation}
Therefore, the compactness factor and surface redshift of the star
can be respectively given by
\begin{equation}
u(r)= \frac{m(r) }{r},
\end{equation}
\begin{equation}
z_s= [1-2u]^{-1/2}-1.
\end{equation}
Hence for different dimensions we can calculate the expressions for
the above parameters as follows:
\textbf{For n=2:}
\begin{equation}
m(r)=\frac{2M}{\pi}\left[tan^{-1}\left(\frac{r}{\sqrt{\phi}}\right)-\frac{r\sqrt{\phi}}{r^{2}+\phi}\right],
\end{equation}
\begin{equation}
u(r)=\frac{2M}{\pi
r}\left[tan^{-1}\left(\frac{r}{\sqrt{\phi}}\right)-\frac{r\sqrt{\phi}}{r^{2}+\phi}\right],
\end{equation}
\begin{equation}
z_s=\left[1-\frac{4M}{\pi
r}\left\{tan^{-1}\left(\frac{r}{\sqrt{\phi}}\right)-\frac{r\sqrt{\phi}}{r^{2}+\phi}\right\}\right]^{-\frac{1}{2}}-1.
\end{equation}
\textbf{For n=3:}
\begin{equation}
m(r)=\frac{2M}{3}\left[2-\frac{(3r^{2}+2\phi)\sqrt{\phi}}{(r^{2}+\phi)^{\frac{3}{2}}}\right],
\end{equation}
\begin{equation}
u(r)=\frac{2M}{3r}\left[2-\frac{(3r^{2}+2\phi)\sqrt{\phi}}{(r^{2}+\phi)^{\frac{3}{2}}}\right],
\end{equation}
\begin{equation}
z_s=\left[1-\frac{4M}{3r}\left\{2-\frac{(3r^{2}+2\phi)\sqrt{\phi}}{(r^{2}+\phi)^{\frac{3}{2}}}\right\}\right]^{-\frac{1}{2}}-1.
\end{equation}
\textbf{For n=4:}
\begin{equation}
m(r)=M\left[tan^{-1}\left(\frac{r}{\sqrt{\phi}}\right)-\frac{r\sqrt{\phi}(5r^{2}+3\phi)}{3(r^{2}+\phi)^{2}}\right],
\end{equation}
\begin{equation}
u(r)=M\left[\frac{1}{r}tan^{-1}\left(\frac{r}{\sqrt{\phi}}\right)-\frac{\sqrt{\phi}(5r^{2}+3\phi)}{3(r^{2}+\phi)^{2}}\right],
\end{equation}
\begin{equation}
z_s=\left[1-2M\left\{\frac{1}{r}tan^{-1}\left(\frac{r}{\sqrt{\phi}}\right)-\frac{\sqrt{\phi}(5r^{2}+3\phi)}{3(r^{2}+\phi)^{2}}\right\}\right]^{-\frac{1}{2}}-1.
\end{equation}
\textbf{For n=9:}
\begin{equation}
m(r)=\frac{M\pi^{3}}{3780}\left[128-\frac{128\phi^{4}+576\phi^{3}r^{2}+1008\phi^{2}r^{4}
+840\phi
r^{6}+315r^{8}}{(r^{2}+\phi)^{\frac{9}{2}}}\sqrt{\phi}\right],
\end{equation}
\begin{equation}
u(r)=\frac{M\pi^{3}}{3780r}\left[128-\frac{128\phi^{4}+576\phi^{3}r^{2}+1008\phi^{2}r^{4}
+840\phi
r^{6}+315r^{8}}{(r^{2}+\phi)^{\frac{9}{2}}}\sqrt{\phi}\right],
\end{equation}
\begin{equation}
z_s=\left[1-\frac{M\pi^{3}}{1890r}\left\{128-\frac{128\phi^{4}+576\phi^{3}r^{2}+1008\phi^{2}r^{4}
+840\phi
r^{6}+315r^{8}}{(r^{2}+\phi)^{\frac{9}{2}}}\sqrt{\phi}\right\}\right]^{-\frac{1}{2}}-1.
\end{equation}
\begin{figure}[h]
\centering
\includegraphics[width=0.4\textwidth]{u.eps}
\includegraphics[width=0.4\textwidth]{z.eps}
\caption{The compactness factor (left) and surface redshift (right)
of compact star in $4D$,~$5D$,~$6D$,~$11D$ dimensions are plotted against $r$~(km) for the specified range}
\end{figure}
The nature of variation of the above expressions for compactness
factor and surface redshift of the star can be seen in the Fig. 5
in the left and right panel respectively for all the values of
$n$. It is observed from Fig. 5 (left panel) that compactness
factors for different dimensions are gradually increasing with
decreasing $n$ and maximum for $4D$ spacetime. Thus, very
interestingly, at the center the star is most dense for
$4$-dimension with a very small yet definite core whereas in the
$11D$ case here seem to be no core.
We note that in connection with the isotropic case and in the absence of
the cosmological constant it has been shown for the surface
redshift analysis that $z_s \leq
2$~\cite{Buchdahl1959,Straumann1984,Boehmer2006}. On the other
hand, B{\"o}hmer and Harko \cite{Boehmer2006} argued that for an
anisotropic star in the presence of a cosmological constant the
surface redshift must obey the general restriction $z_s \leq 5$,
which is consistent with the bound $z_s \leq 5.211$ as obtained by
Ivanov~\cite{Ivanov2002}. Therefore, for an anisotropic star
without cosmological constant the above value $z_s \leq 1$ is
quite reasonable as can be seen in the $4D$ case (Fig. 5, right
panel) \cite{Rahaman}. In the other cases of higher
dimension the surface redshift values are increasing and seem to
be within the upper bound~\cite{Ivanov2002}.
We note that integration of $m(r)$ from $0$ to $R$, where $R$
is the radius of the fluid distribution, gives $M$ (total mass of
the source) i.e.
\[M=\int_0^{ R}~ \left[ \frac{2 \pi^{\frac{n+1}{2}} }{\Gamma
\left(\frac{n+1}{2}\right)}\right]r^n \rho dr.\] This equation
gives the radius $R$ of the fluid distribution. Thus solutions of
the following equations provide the corresponding radius of
different dimensional situations.
\textbf{For $n=2$:}
\begin{equation}
M=\frac{2M}{\pi}\left[tan^{-1}\left(\frac{R}{\sqrt{\phi}}\right)
-\frac{R\sqrt{\phi}}{R^{2}+\phi}\right],
\end{equation}
\textbf{For $n=3$:}
\begin{equation}
M=\frac{2M}{3}\left[2-\frac{(3R^{2}+2\phi)\sqrt{\phi}}{(R^{2}+\phi)^{\frac{3}{2}}}\right],
\end{equation}
\textbf{For $n=4$:}
\begin{equation}
M=M\left[tan^{-1}\left(\frac{R}{\sqrt{\phi}}\right)-\frac{R\sqrt{\phi}(5R^{2}+3\phi)}
{3(R^{2}+\phi)^{2}}\right],
\end{equation}
\textbf{For $n=9$:}
\begin{equation}
M=\frac{M\pi^{3}}{3780}\left[128-\frac{128\phi^{4}+576\phi^{3}R^{2}+1008\phi^{2}R^{4}
+840\phi
R^{6}+315R^{8}}{(R^{2}+\phi)^{\frac{9}{2}}}\sqrt{\phi}\right],
\end{equation}
\subsection{Some other physical parameters}
In this subsection we have shown the panel of the plots for the
conformal parameter $\psi(r)$ (top left), the metric potential
$e^{\lambda}$ (top right) and the density $\rho$ (bottom) for $4$
and extra dimensional spacetimes (Fig. 6). It is observed that for
all the physical parameters the features are as usual for $4D$,
however for extra dimension they take different shapes. A special
mention can be done for density where central densities are
abruptly decreasing as one goes to higher dimensions. Thus, from
the plot it reveals that the central density is maximum for $4D$
whereas it is minimum for $11D$ spacetime showing most compactness
of the star for standard $4$-dimension. Note that this same result
was observed in Fig 5 (left panel).
\begin{figure}[h]
\centering
\includegraphics[width=0.4\textwidth]{psi.eps}
\includegraphics[width=0.4\textwidth]{elambda.eps}
\includegraphics[width=0.4\textwidth]{rho.eps}
\caption{The conformal factor $\psi(r)$ (top left),
the metric potential $e^{\lambda}$ (top right) and matter density (bottom)
are plotted against $r$~(km) for $4D$,~$5D$,~$6D$,~$11D$ dimensional spacetimes.
The suffixes in the left panel indicates the dimension of the spacetime
and for right panel $D=n+2$ stands for the dimension of the spacetime}
\end{figure}
\section{Conclusion}
In the present paper we have studied thoroughly a set of new
interior solutions for anisotropic stars admitting conformal motion
in higher dimensional noncommutative spacetime. Under this
spacetime geometry the Einstein field equations are solved by
choosing a particular Lorentzian type density distribution
function as proposed by Nozari and Mehdipour \cite{Nozari2009}. The
studies are conducted not only with standard $4D$ dimensional
spacetime but also for three special cases with higher dimension,
such as $5D$, $6D$ and $11D$. In general it is noted that the
model parameters e.g. matter-energy density, radial as well as
transverse pressures, anisotropy and others show physical
behaviours which are mostly regular throughout the stellar
configuration.
Also it is specially observed that the solutions represent a star
with mass $2.27$~$M_{\odot}$ and radius $4.17$~km
which falls within the range ($0 < z_s \leq 1$) of a compact
star \cite{Rahaman,Kalam2012,Hossein2012,Kalam2013}.
However, it has been shown that for a strange star of radius
$6.88$~km surface redshift turns out to be $z_s = 0.5303334$
\cite{Rahaman} whereas the maximum surface redshift for a
strange star $Her X-1$ of radius $7.7$~km is $0.022$
\cite{Kalam2012} and that for a compact star $4U~1820-30$ of
radius $10$~km turns out to be again $0.022$
\cite{Hossein2012}. Therefore it seems that our compact star
may be a strange quark star (see Table 1).
\begin{table*}
\centering
\begin{minipage}{140mm}
\caption{Values of the model parameter $z_s$ for
different Strange Stars}\label{tbl-1}
\begin{tabular}{@{}lrrrrrr@{}}
\hline Strange Star candidates & $M$($M_{\odot}$) & $R$(km) &
$M/R$ & $z_s$ \\ \hline Her X-1 & 0.88 &
7.7 & 0.168 & 0.0220 [Ref. 13]\\
& & & & 0.2285 [Ref. 15]\\
4U 1820-30 & 2.25 & 10.0 & 0.332 &
0.0220 [Ref. 14]\\
& & & & 0.7246 [Ref. 15]\\
SAX J 1808.4-3658(SS1) & 1.435 & 7.07 & 0.299 &
0.5787 [Ref. 15]\\ SAX J 1808.4-3658(SS2) & 1.323 & 6.35 &
0.308 & 0.6108 [Ref. 15]\\ Rahaman model [Ref. 11] &
1.46 & 6.88 & 0.313 & 0.5303334
\\ Our proposed model & 2.27 & 4.17 & 0.804
& $0 < z_s \leq 1$\\ \hline
\end{tabular}
\end{minipage}
\end{table*}
However, through several mathematical case studies we have given
emphasis on the acceptability of the model from physical point of
view for various structural aspects. As a consequence it is
observed that for higher dimensions, i.e. beyond $4D$ spacetime,
the solutions exhibit several interesting yet bizarre features.
These features seem physically not very unrealistic.
Thus, as a primary stage, the investigation indicates that compact
stars may exist even in higher dimensions. But before placing a
demand in favour of this highly intrigued issue of compact stars
with extra dimensions we need to perform more specific studies and
to look at the diversified technical aspects related to higher
dimensional spacetimes of a compact star. Basically our approach,
dependent on a particular energy density distribution of
Lorenztian type, which gives higher dimensional existence of
compact stars may not be only the way to have sufficient evidence
in favour of it. We further need to employ other type of density
distributions as well. Moreover, one may also think for other than
higher dimensional embedding of GTR and thus opt for alternative
theories of gravity to find conclusive proof for higher
dimensional compact stars.
However, in the literature there are some evidences available
in favour of `Extra Dimensions in Compact Stars'
\cite{Liddle1990,BLL2003,Paul2004,BLL2006,BLL2010,CP2010}.
\section*{Acknowledgments}
FR and SR wish to thank the authorities of the Inter-University
Centre for Astronomy and Astrophysics (IUCAA), Pune, India for
providing the Visiting Associateship under which a part of this
work was carried out. We all express our grateful thanks to both
the referees for their several suggestions which have enabled us
to improve the manuscript substantially.
|
\section{Introduction}\let\thefootnote\relax\footnote{\emph{Abbreviations:} MFA, metabolic flux analysis; EFMs, elementary flux modes; CHO, Chinese hamster ovary; Lac, lactate; Glc, glucose}
In previous work we presented a column generation based algorithm for solving the EFMs-based metabolic flux analysis (MFA) problem \citep{Oddsdottir2014}.
In this work we present a more refined model where the column generation algorithm is combined with robustness. For the sake of completeness a short description of the background follows. A more detailed background can be found in \emph{e.g.,\ } \citet{Oddsdottir2014}.
A metabolic reaction network is represented by the stoichiometric matrix $A$, which together with the flux vector ($v$) gives the overall change in concentration of each metabolite ($C$).
The rows of the stoichiometric matrix ($A$) refer to either external metabolites ($A_x$) or internal ($A_i$). The flux space is given by a set of vectors $v$ that satisfy the pseudo-steady state assumption and flow direction assumption,
\begin{equation}\label{eq:Cone}
\left\{v: \begin{bmatrix}
\phantom{-} A_{i}\\
-A_{i}\\
-I_j
\end{bmatrix}v \leq \begin{bmatrix} 0\\0\\0
\end{bmatrix}, \; j\in J_{\text{irrev}} \right\},
\end{equation}
where $I_j$ is a reduced identity matrix with ones only when $j\in J_{\text{irrev}}$ and $J_{\text{irrev}}$ is the set of irreversible reactions.
When all reactions in the network are irreversible \eqref{eq:Cone} is a cone where any ray can be written as a non-negative linear combination of the extreme rays \citep[Part I.4 Theorem 4.8]{Nemhauser1999}.
EFMs contain information how extracellular metabolites are connected by detailing which reactions are required for their uptake or production \citep{Llaneras2010}. They are vectors in the flux space, each EFM includes only a minimial set of reactions and is nondecomposable \citep{Klamt2002a}. Further, any vector in the flux space can be denoted as a non-negative linear combination of the EFMs \citep{Schilling1999,Papin2003},
\begin{equation}
v= \sum_{l=1}^L w_l e_l =Ew, \qquad \gamma \geq 0,
\label{eq:combin}
\end{equation}
where $e$ denotes a single EFM and the matrix $E$ contains the EFMs as columns. In this sense the EFMs generate the flux space and
are related to the definition of extreme rays in the cone \eqref{eq:Cone} with only irreversible reactions. In fact when a metabolic network only has irreversible reactions the EFMs and the extreme rays of the cone \eqref{eq:Cone} are equal \citep{Gagneur2004}. We assume, without loss of generality, that the metabolic network has only irreversible reactions, \emph{i.e.,\ } $v_j\geq 0 \; \forall j$.
When the network includes reversible reactions finding all the EFMs is equivalent to finding all the extreme rays of a cone in an extended space where all reactions are irreversible \citep{Gagneur2004,Urbanczik2005}.
For modest-sized networks enumeration of EFMs is possible and computer programs exist for that purpose, \emph{e.g.,\ } Metatool \citep{VonKamp2006}. However, with increased network size enumeration of EFMs becomes prohibitive \citep{Klamt2002a}. Thus focus has shifted to identify only a subset of the EFMs \citep{DeFigueiredo2009,Kaleta2009,Tabe-Bordbar2013}.
This work considers the solution of the EFMs-based metabolic flux analysis (MFA) problem \citep[Chapter 5.2]{Provost2006} when the network is large and there are known bounds on measurement errors. EFMs-based MFA uses the decomposition of $v$ given by \eqref{eq:combin} to create a macroscopic network ($A_xE$). The macroscopic fluxes ($w$) are then adjusted so that the flux in the network fit the cell specific external flux measurements ($Q$), \emph{i.e.,\ }
\begin{equation}\label{eq:EFMsbMFA}
\begin{aligned}
\underset{w}{\text{minimize}} \quad & \frac{1}{2} \|Q-\mathcal{I}A_xE w \|_2^2\\
\text{subject to } \quad
& w \geq 0.
\end{aligned}
\end{equation}
The formulation given by \eqref{eq:EFMsbMFA} includes multiple repetitions of the same experiments, \emph{i.e.,\ } if $q_{k}$ are results from one repetition, $k$, then $Q^T= [q_{1}^T, \ldots q_{d}^T]$, where $d$ denotes the number of repetitions.
$\mathcal{I}$ is a stacked identity matrix consisting of $d$ identity matrices of size $M_{ext}$ (number of external metabolites) or $\mathcal{I}=[I_{M_{ext}}, \ldots, I_{M_{ext}}]^T$, where $I_{M_{ext}}$ is repeated $d$ times.
EFMs-based MFA as given by \eqref{eq:EFMsbMFA} requires the whole set of EFMs, limiting the application to simplified networks. Methods that can solve the EFMs-based MFA problem without enumerating EFMs exist. One method identifies EFMs beforehand through a series of linear programming (LP) problems \citep{Jungers2011}.
This method is based on the existence of a feasible flux vector $v$, an assumption we will examine in Section \ref{sec:Feasab}.
In our previous work we introduced a more integrated approach that enables identification of EFMs in conjunction with solving the EFMs-based MFA problem \citep{Oddsdottir2014}. The approach was based on an optimization technique named column generation \citep{Lubbecke2005}, in which large networks can be handled by relying on a master problem and a subproblem that are solved iteratively.
The subproblem gives the master problem a new column every iteration until the solution of the subproblem indicates that the solution of the master problem is optimal to the full optimization problem.
The experimental measurements used to calculate the fluxes in $Q$ in the EFMs-based MFA problem \eqref{eq:EFMsbMFA} are prone to errors, which have been stated to reach at least 20\% \citep{Goudar2009}. For this reason we wanted to consider the sensitivity of the solution with respect to these errors.
Additionally, in some cases certain metabolites included in the network, are difficult to measure and thus remain unmeasured in the data set. Even though those metabolites are unmeasured in this specific experimental setup some information on their fluxes can be available, and a bound can be added.
We therefore present an extension to our previous column generation algorithm given by \citet{Oddsdottir2014}. This extension includes both a robust formulation and a version that deals with unmeasured metabolites, while still having the benefit of working with larger networks. In the robust formulation the error on each measurement is assumed bounded, while unmeasured metabolites are given a feasible interval.
In the robust formulation the aim is to minimize the objective function when the assumed errors are such that the objective is as disadvantageous as it can be. For more information on robust optimization please see \citet{Mulvey1995} or \citet{Ben-Tal2009}.
Previous work on robust least-squares mainly focus on errors in both the measurements and the model, in general those formulations are difficult to solve (NP complete) \citep{Ghaoui1997}. However, we show that for this special case, where the errors are only in measurements and bounded by an interval, the robust problem can be formulated as a convex quadratic programming (QP) problem. Furthermore, column generation can be applied to this QP, allowing the problem to be solved without previous enumeration of EFMs.
The paper is outlined as follows. In Section \ref{sec:Feasab} it is shown how the stacked least-squares can be written as least-squares of averages along with an example, showing that metabolic reaction networks do not necessarily have a feasible flux vector for a given set of external measurements. Then we present the main results of this paper in Section \ref{sec:Robustv}; a robust version of the EFMs-based MFA, where column generation can also be applied, along with a version in which intervals for unmeasured metabolites are included. Finally in Section \ref{sec:CS} we present some results comparing the solutions of the robust problem to the EFM-based MFA.
\section{On the Feasibility of the EFMs-based MFA}\label{sec:Feasab}
In this section we examine the uniqueness of the stacked EFMs-based MFA and if there always exists a flux vector that fits the network and measurements exactly. These observations support our main results shown in Section \ref{sec:Robustv}.
To simplify the discussion we consider a problem equivalent to the EFMs-based MFA where a flux vector $v$ is sought,
\begin{equation}\label{eq:EFMMFA}
\begin{aligned}
\underset{v}{\text{minimize}} \quad & \frac{1}{2} \norm{\mathcal{I}A_xv-Q}_2^2\\
\text{subject to } \quad & A_iv=0,\\
& v \geq 0.
\end{aligned}
\end{equation}
Problem \eqref{eq:EFMMFA} is equivalent to \eqref{eq:EFMsbMFA}, by using the decomposition of $v$ given by \eqref{eq:combin}, thus removing the equality constraint.
With the stacking of multiple measurements the objective function of \eqref{eq:EFMMFA} seems to represent an overdetermined problem. However, problem \eqref{eq:EFMMFA} can be represented as if it only has one measurement, or as an underdetermined problem by,
\begin{equation*}
\norm{\mathcal{I}A_{x}v-Q}_2^2
\sum_{k=1}^d \norm{ A_{x}v-q_k}_2^2
=d v^TA_{x}^TA_{x}v -2 \sum_{k=1}^d q_k^TA_{x}v + \sum_{k=1}^d q_k^Tq_k.
\end{equation*}
Thus, the solution $v$ of \eqref{eq:EFMMFA} is equal to the solution of
\begin{equation}\label{eq:EFMMFAav}
\begin{aligned}
\underset{v}{\text{minimize}} \quad & \norm{ A_{x}v-\frac{1}{d}\sum q_k}_2^2\\
\text{subject to } \quad & A_iv=0,\\
& v \geq 0.
\end{aligned}
\end{equation}
Consequently, for a given experimental condition, stacking repetitions is equal to using the average value of the flux measurements.
In light of that the data fitting can equivalently use the average, \emph{i.e.,\ } only one measurement, it becomes important to consider if there always exists a solution to \eqref{eq:EFMMFAav} with zero residual.
That is, if
\begin{equation} \label{eq:exv}
\exists v: \; A_xv=q, \; A_iv=0, \; v \geq 0, \; \text{for any }q.
\end{equation}
For robustness the existence of a solution is especially relevant, because when there is only one measurement that fits the network exactly robust optimization will not give a different solution from the non-robust solution.
Although, it should be noted that when there are repetitions, or multiple measurements, the solution of the robust optimization can differ from the non-robust solution.
Previous analysis of calculability in networks have considered when there exists a unique $v$ that satisfies \eqref{eq:exv} without the positivity constraint. Hence examining if a network is underdetermined or determined.
In general a full rank matrix has the whole of $\mathcal{R}$ as its range, indicating that there always exists a $v$ such that $A_xv = q$ and $A_iv=0$. When the network is underdetermined this $v$ would not be unique \citep{Klamt2002}.
However, this assumes that $v$ can be negative in all values. In metabolic networks reactions are often restricted to only one direction. Hence, an underdetermined network may not have a solution for all sets of measurements. A small example of how this can happen follows.
\begin{figure}[H]
\centering
\begin{tikzpicture}[scale =0.5, state/.style ={ scale=0.6,circle ,top color =white , bottom color = processblue!20 ,draw,processblue , text=black , minimum width =1 cm},interm/.style={fill,circle,inner sep=0pt,scale=0.01}]
\node[interm](cc3){};
\node[state] (c3) [below left =of cc3]{\ce{C3}};
\node[state] (c5) [right =of cc3]{\ce{C5}};
\node[state] (c1) [left =of c3]{\ce{C1}};
\node[state] (c2) [below =of c1]{\ce{C2}};
\node[state] (c4) [right =of c2]{\ce{C4}};
\node[interm] (c5c6) [below right =of c5]{};
\node[state] (c7) [right = of c5c6]{\ce{C7}};
\node[state] (c8) [below = of c7]{\ce{C8}};
\node[interm] (c3c4) [above right =of c4]{};
\node[state] (c6) [right = of c3c4]{\ce{C6}};
\draw[->] (c1) to node[above] {$v_1$} (c3);
\draw[->] (c2) to node[above] {$v_2$} (c4);
\draw[-] (c3) to node {} (cc3);
\draw[->] (cc3) to node[above] {$v_4$} (c5);
\draw[-](c6) to node{} (c5c6);
\draw[-](c3) to node{} (c3c4);
\draw[-](c4) to node{} (c3c4);
\draw[->] (c3c4) to node[above] {$v_6$} (c6);
\draw[->] (c5c6) to node[above] {$v_8$} (c7);
\draw[->] (c4) to node[left]{$v_3$}(c3);
\draw[->] (c4) to node[above]{$v_9$}(c8);
\draw[->] (c5) to node [right] {$v_{7}$}(c6);
\draw[->] (c4) to node [left] {$v_{5}$}(c5);
\draw[line width =0.5mm, color =redcol](-3.5,-6.5) rectangle (3.7,3.7);
\node[redcol] (Int) [above =of c5] {Internal};
\node[kthblue] (Ext) [right =of Int] {External};
\end{tikzpicture}
\caption{A reaction network with underdetermined stoichiometry.}\label{fig:unddRN}
\end{figure}
Consider the network shown in Figure \ref{fig:unddRN}. If one external metabolite is not measured then the network has underdetermined stoichiometry, and thus, there exists a $v$ such that $A_xv=q$ and $A_iv=0$ for any $q$. However depending on which external metabolite is not measured $v$ might not be positive.
\begin{itemize}
\item If \ce{C1} is not measured, then, depending on what the measurements are, there might not exist a $v \geq 0$ that satisfies the stoichiometry. With no measurements on \ce{C1}, $v_1$ is free, however $v_2,v_8$ and $v_9$ are fixed from measurements. Flow balance requires that $v_2 \leq v_8+v_9$, additionally if $v_2 \leq v_9$ then the flow through \ce{C4} cannot be fulfilled. Thus, if $v_2$ is too high flow balance can not be fulfilled and no feasible $v$ exists.
\item If \ce{C2} is not measured, then $v_2$ is free, and can be chosen so that the flow to \ce{C7} and \ce{C8} is satisfied, note that any lack of flow from $v_1$ can be compensated by sending through $v_3$.
\end{itemize}
Hence, errors in measurements can lead to nonexistance of a flux vector for the given network that fits the measurements exactly.
\section{The Robust Variant of the EFMs-based MFA}\label{sec:Robustv}
This section contains the main results of this work, here we present an extension of the EFMs-based MFA problem, where errors in $Q$ are taken more directly into consideration. For this purpose we make use of a technique named robust optimization \citep{Mulvey1995,Ben-Tal2009}.
The robust optimization problem is to minimize the residual when the errors in the data give a worst-case scenario outcome, \emph{i.e.,\ } the errors in the data are such that the residual is maximized. Inherent in least-squares is the assumption that the errors are bounded by the two-norm, \emph{i.e.,\ } $\norm{\Delta Q} \leq \beta$. In fact, when the errors are assumed bounded by the two norm, the least squares problem gives the same solution as its robust variant. However, in this work we assume that the errors in $Q$ are bounded by an interval, a more restrictive assumption that might cause the solution to change. The interval is such that $Q_{real}=Q+\Delta Q$ where $\Delta Q_i =[\Delta q_1^T, \ldots \Delta q_d^T]^T$ and $\abs{\Delta q_{ki}} \leq \theta_{ki} \abs{q_{ki}}$, $k$ refers to a specific repetition and $i$ to the metabolite.
In order to simplify notation $\theta$ is stacked in the same way as $Q$ and $\Delta Q$, the subindex $s$ then refers to a specific element in those vectors. Note that in general the percentage of error on each metabolite is the same for all repetitions, \emph{i.e.,\ } $\theta_{k_1i}=\theta_{k_2i}$ for all $k_1$ and $k_2$. The robust problem is then given by
\begin{equation}\label{eq:RobOr}
\underset{w \geq 0}{\text{minimize}} \underset{|\Delta Q_{s}| \leq \theta_s |Q_{s}|}{\text{maximize}} \quad \frac{1}{2} \norm{\mathcal{I} A_{x}E w -Q+\Delta Q}.
\end{equation}
As shown in Appendix \ref{sec:DerRQP}, problem \eqref{eq:RobOr} can equivalently be formulated as a quadratic programming problem in the form
\begin{equation}\label{eq:Robdr}
\begin{aligned}
\underset{w,t}{\text{minimize}} \quad & \frac{1}{2} \norm{\mathcal{I}A_{x}E w-Q}^2+\mathbf{1}^T t\\
\text{subject to}\quad & t_s - \left(\mathcal{I}A_{x}E w-Q\right)_s \theta_s Q_s\geq0, \quad \forall s\\
& t_s + \left(\mathcal{I}A_{x}E w-Q\right)_s \theta_s Q_s \geq 0, \quad \forall s\\
&w \geq 0.
\end{aligned}
\end{equation}
When $\theta =0$ the above formulation is equivalent to the EFMs-based MFA \eqref{eq:EFMsbMFA}.
To make the notation more compact we define $\Theta$ and $\tilde{Q}$ as diagonal matrices with $\theta$ and $Q$ on the diagonal, respectively. Further,
the objective function of \eqref{eq:Robdr} can be stated as minimizing the average measure of $q$ over all measurements,
\begin{equation}\label{eq:Robdrav}
\begin{aligned}
\underset{w,t}{\text{minimize}} \quad & \frac{1}{2} \norm{A_{x}E w-\frac{1}{d}\sum_{k=1}^dq_{k}}^2+\mathbf{1}^T t\\
\text{subject to}\quad & t-\Theta \tilde{Q}\left(\mathcal{I}A_{x}E w-Q\right) \geq 0,\\
& t+\Theta \tilde{Q}\left(\mathcal{I}A_{x}E w-Q\right) \geq 0,\\
& w\geq0,
\end{aligned}
\end{equation}
The formulation in \eqref{eq:Robdrav} shows that even when the average value gives a zero norm solution of the EFMs-based MFA, the robust solution might be different.
The reason for this difference can be seen when the constraints in \eqref{eq:Robdrav} are examined. For multiple measurements of the same metabolites $\left(\mathcal{I}A_{x}E w-Q\right)_s$ for each specific measurement will in general not be equal to zero for all $s$, forcing $t$ to increase from zero.
With enough increase in $t$ the robust solution might deviate from the non-robust solution, \emph{i.e.,\ } increasing $\norm{\mathcal{I}A_{x}E w-Q}$ while decreasing $t$.
Thereby, giving a non-zero value of $\norm{\mathcal{I}A_{x}E w-Q}$ the robust solution.
The change in the solution depends on two factors, how far the measurement is from the best least-squares calculated flux and how high the error on that measurement is assumed to be.
No change in the optimal solution is expected when either the measurements are good or the assumed interval is tight, since then $t$ can remain close to zero.
\subsection{Column Generation of the Robust Variant of EFMs-based MFA}
For large networks enumerating all EFMs beforehand is prohibitive. For that reason, we present two problems: A master problem and subproblem that can be solved iteratively to identify the necessary EFMs along with solving problem \eqref{eq:Robdr}. Their derivation can be seen in appendix \ref{sc:MPSP}. The master problem is given by
\begin{subequations} \label{eq:RobMP}
\begin{align}
\underset{w,t}{\text{minimize}} \quad & \frac{1}{2} \norm{\mathcal{I}A_{x}E_B w_B-Q}^2+\mathbf{1}^Tt\\
\text{subject to}\quad & t-\Theta\tilde{Q}\left(\mathcal{I}A_{x}E_Bw_B -Q\right) \geq 0, \label{eq:l1} \\
& t+\Theta \tilde{Q}\left(\mathcal{I}A_{x}E_Bw_B -Q\right) \geq 0, \label{eq:l2}\\
&w_B \geq 0,
\end{align}
\end{subequations}
where the index $B$ indicates that only the known columns of $E$ are used.
The corresponding subproblem requires information from the solution of the master problem. More specifically the macroscopic fluxes, $w_B$ and the dual solutions, $\lambda_m$ and $\lambda_p$ corresponding to the constraints \eqref{eq:l1} and \eqref{eq:l2} respectively, are required. The subproblem is given by
\begin{equation} \label{eq:RobSubp}
\begin{aligned}
\underset{e}{\text{minimize}}\quad &\left(\mathcal{I}A_{x}E_Bw_B-Q +\Theta \tilde{Q} \lambda_{m}-\Theta \tilde{Q}\lambda_{p}\right)^T\mathcal{I}A_{x}e \\
\text{subject to}\quad & A_ie=0,\\
& \mathbf{1}^Te=1, \\
& e_j \geq 0 \quad \forall j.
\end{aligned}
\end{equation}
The subproblem \eqref{eq:RobSubp} identifies EFMs \citep{Oddsdottir2014} until the objective function value is non-negative. At that stage the optimal solution of the master problem is also the optimal solution of the full problem.
\subsection{Inclusion of Intervals on Unmeasured Metabolites in the Robust Variant}
In this section a further extension of the EFMs-based MFA is introduced, where unmeasured metabolites are taken into consideration. Unmeasured metabolites are external metabolites that are a part of the network used but have no measurement data. An example is the metabolite \ce{CO2} a gas that is difficult to measure without special experimental setup.
Intervals are estimated on those metabolites and modelled with a penalty function. In this way the intervals are allowed to be infeasible for the first few iterations of the column generation. A robust optimization problem that considers feasible intervals on unmeasured metabolites can be stated as,
\begin{equation}\label{eq:Robint}
\begin{aligned}
\underset{w \geq 0}{\text{minimize}} \;& \left(M_u^T\max(\mathbf{0},(-Q_n^u+A_{x,n}Ew)) \right. \\&+M_l^T\max(\mathbf{0},(Q_n^l-A_{x,n}Ew)) +\\& \; \left. \underset{\abs{\Delta Q}_k\leq \theta_k \abs{Q_k} }{\text{maximize}} \; \frac{1}{2} \norm{\mathcal{I}A_{x}E w -Q+\Delta Q} \right),
\end{aligned}
\end{equation}
where $A_{x,n}$ are the rows from the stoichiometric matrix that correspond to the unmeasured metabolites, $Q_n^u$ and $Q_n^l$ are the upper and lower bounds on the given interval respectively. The quantities $M_u$ and $M_l$ indicate how large the penalty is for not satisfying the specific interval constraint. In general $M_u$ and $M_l$ will be set to a sufficiently large number by the user.
The inner maximization problem is unchanged from \eqref{eq:RobOr} and hence, \eqref{eq:Robint} can be represented as a convex quadratic programming problem,
\begin{subequations}\label{eq:Robintder}
\begin{align}
\underset{w,t,z^u,z^l}{\text{minimize}} \quad & \frac{1}{2} \norm{\mathcal{I}A_{x}E w-Q}^2+\mathbf{1}^T t +M_u^Tz^u + M_l^Tz^l\\
\text{subject to}\quad & t-\Theta \tilde{Q}\left(\mathcal{I}A_{x}Ew -Q\right) \geq 0, \\
& t+\Theta \tilde{Q}\left(\mathcal{I}A_{x}Ew -Q\right) \geq 0, \\
& z^u-A_{x,n}Ew\geq -Q_n^u, \label{eq:lu}\\
& z^l+A_{x,n}Ew \geq Q_n^l, \label{eq:ll}\\
& z^u \geq 0,\\
& z^l \geq 0,\\
&w \geq 0.
\end{align}
\end{subequations}
The formulation from \eqref{eq:Robintder} can be solved using column generation where the subproblem generates columns of $E$ by,
\begin{equation}\label{eq:SPint}
\begin{aligned}
\underset{e}{\text{minimize}}\quad &\left((\mathcal{I}A_{x}E_Bw_B-Q +\Theta\tilde{Q}\lambda_{m}-\Theta\tilde{Q}\lambda_{p})^T\mathcal{I}A_{x}\right.\\ &\left.+(\lambda_u- \lambda_l)^TA_{x,n} \right)e\\
\text{subject to } \quad
& A_{i}e=0, \\
&\mathbf{1}^Te \leq 1, \\
& e_j \geq 0\; \forall j,
\end{aligned}
\end{equation}
where $\lambda_u$ and $\lambda_l$ are the dual variables corresponding to constraints \eqref{eq:lu} and \eqref{eq:ll}.
\section{Case-Study: Cultivation of CHO Cells} \label{sec:CS}
\subsection{Particulars of the Data}
Data were obtained from the same experimental setup as described in \citet{Oddsdottir2014}. A Chinese hamster ovary (CHO) cell line producing a monoclonal antibody (mAb) was cultivated during 11 days according to a pseudo-perfusion protocol (daily sample collection and medium exchange) to imitate steady-state conditions. The cultivation was carried out in parallel cultures using different medium compositions. Cell-specific metabolic rates (external fluxes) were calculated for the last seven days of culture. Two different media were selected for the present work in order to show extreme situations of our findings, the resulting fluxes are presented in Tables \ref{tb:datam1} (Medium 1) and \ref{tb:datam11} (Medium 5). In addition to the measured data, an interval on \ce{CO2} flux was estimated as $4.95-7.09$, based on the intervals given by \citet{Goudar2011} and \citet{Aunins1993}.
\begin{table*}[tbh]
\centering
\scriptsize
\rowcolors{1}{white}{kthgrey}
\begin{tabular}{l|rrrrrrr}
\bf{Metabolite} & \bf{$q_{6,1}$} & \bf{$q_{7,1}$} & \bf{$q_{8,1}$} & \bf{$q_{9,1}$} & \bf{$q_{10,1}$} & \bf{$q_{11,1}$} & \bf{$q_{12,1}$}\\ \hline
Ala & 0.45 & 0.51 & 0.44 & 0.40 & 0.40 & 0.41 & 0.46\\
Arg &-0.27 &-0.26 &-0.14 &-0.19 &-0.11 &-0.47 &-0.22\\
Asn &-0.17 &-0.22 &-0.17 &-0.18 &-0.20 &-0.18 &-0.15\\
Asp &0.04 &0.07 &0.07 &0.06 &0.06 &0.07 &0.07\\
Biomass &0.61 &0.60 &0.50 &0.70 &0.53 &0.56 &0.65\\
Cys &-0.09 &-0.11 &-0.07 &-0.09 &-0.09 &-0.07 &-0.05\\
Glucose (Glc) &-3.52 &-4.06 &-2.64 &-3.26 &-3.96 &-2.92 &-3.43\\
Gln &-1.60 &-1.97 &-1.61 &-2.38 &-2.31 &-1.90 &-1.71\\
Glu &0.22 &0.32 &0.25 &0.27 &0.30 &0.33 &0.28\\
Gly &0.04 &0.07 &0.05 &0.03 &0.03 &0.05 &0.03\\
His &-0.02 &-0.05 &-0.02 &-0.01 &-0.01 &-0.01 &-0.02\\
Ile &-0.10 &-0.13 &-0.10 &-0.10 &-0.11 &-0.12 &-0.11\\
Lactate (Lac) &5.48 &7.40 &5.89 &6.20 &6.78 &7.02 &6.00\\
Leu &-0.19 &-0.22 &-0.17 &-0.17 &-0.18 &-0.20 &-0.19\\
Lys &-0.05 &-0.05 &-0.06 &-0.05 &-0.08 &-0.07 &-0.04\\
Met &-0.05 &-0.07 &-0.04 &-0.05 &-0.06 &-0.04 &-0.03\\
\ce{NH4+} &1.17 &-- &1.17 &1.15 &1.23 &1.24 &1.18\\
Phe &-0.10 &-0.12 &-0.10 &-0.09 &-0.09 &-0.12 &-0.12\\
Pro &-0.10 &-0.14 &-0.09 &-0.11 &-0.11 &-0.13 &-0.10\\
Ser &-0.00 &0.01 &-0.01 &0.00 &0.00 &-0.03 &0.01\\
Thr &-0.11 &-0.10 &-0.11 &-0.10 &-0.12 &-0.10 &-0.07\\
Trp &-0.03 &-0.07 &-0.02 &-0.03 &-0.03 &-0.02 &-0.02\\
Tyr &-0.09 &-0.12 &-0.08 &-0.08 &-0.08 &-0.12 &-0.10\\
Val &-0.15 &-0.17 &-0.13 &-0.14 &-0.15 &-0.16 &-0.13\\
mAb &2.1e-04 &2.3e-04 &1.8e-04 &1.8e-04 &1.7e-04 &2.2e-04 &1.9e-04\\
\end{tabular}
\caption{External fluxes obtained from a CHO cell cultivation, given for each metabolite from the final seven days of the cultivation for medium 1.
The unit is $\text{pmol}\cdot \text{cell}^{-1} \cdot \text{day}^{-1}$, except for Biomass which has the unit $\text{day}^{-1}$.}
\label{tb:datam1}
\end{table*}
\begin{table*}[tbh]
\centering
\scriptsize
\rowcolors{1}{white}{kthgrey}
\begin{tabular}{l|rrrrrrr}
\bf{Metabolite} & \bf{$q_{6,5}$} & \bf{$q_{7,5}$} & \bf{$q_{8,5}$} & \bf{$q_{9,5}$} & \bf{$q_{10,5}$} & \bf{$q_{11,5}$} & \bf{$q_{12,5}$}\\ \hline
Ala &0.43 &0.45 &0.46 &0.50 &0.44 &0.40 &0.43\\
Arg &-0.45 &-0.57 &-0.22 &-0.17 &-0.11 &-0.23 &-0.51\\
Asn &-0.18 &-0.22 &-0.20 &-0.22 &-0.21 &-0.19 &-0.17\\
Asp &0.07 &0.08 &0.08 &0.08 &0.09 &0.09 &0.08\\
Biomass &1.11 &0.57 &0.55 &0.59 &0.52 &0.55 &0.56\\
Cys &-0.11 &-0.14 &-0.12 &-0.13 &-0.11 &-0.09 &-0.12\\
Glc &-3.56 &-3.22 &-3.14 &-2.79 &-3.23 &-3.18 &-3.19\\
Gln &-1.79 &-1.71 &-1.60 &-2.41 &-1.81 &-1.85 &-1.77\\
Glu &0.25 &0.31 &0.22 &0.24 &0.22 &0.23 &0.21\\
Gly &-0.00 &-0.06 &0.01 &0.04 &0.03 &0.03 &0.04\\
His &-0.03 &-0.03 &-0.04 &-0.03 &-0.03 &-0.03 &-0.03\\
Ile &-0.13 &-0.20 &-0.11 &-0.11 &-0.11 &-0.11 &-0.13\\
Lac &6.70 &6.02 &5.41 &6.24 &5.61 &5.82 &5.84\\
Leu &-0.21 &-0.30 &-0.19 &-0.19 &-0.18 &-0.19 &-0.22\\
Lys &-0.09 &-0.15 &-0.04 &-0.04 &-0.06 &-0.09 &-0.07\\
Met &-0.05 &-0.07 &-0.06 &-0.05 &-0.06 &-0.04 &-0.04\\
\ce{NH4+} &1.37 &1.45 &1.26 &1.33 &1.31 &1.11 &1.25\\
Phe &-0.12 &-0.11 &-0.10 &-0.11 &-0.09 &-0.07 &-0.13\\
Pro &-0.17 &-0.23 &-0.12 &-0.11 &-0.12 &-0.12 &-0.15\\
Ser &-0.09 &-0.06 &-0.01 &0.03 &0.00 &0.01 &0.01\\
Thr &-0.14 &-0.23 &-0.13 &-0.11 &-0.13 &-0.11 &-0.10\\
Trp &-0.03 &-0.03 &-0.03 &-0.03 &-0.03 &-0.02 &-0.02\\
Tyr &-0.12 &-0.12 &-0.10 &-0.11 &-0.08 &-0.07 &-0.06\\
Val &-0.18 &-0.24 &-0.15 &-0.15 &-0.15 &-0.15 &-0.18\\
mAb &2.3e-04 &2.3e-04 &1.8e-04 &2.2e-04 &1.6e-04 &2.0e-04 &2.1e-04\\
\end{tabular}
\caption{External fluxes obtained from a CHO cell cultivation, given for each metabolite from the final seven days of the cultivation for medium 5.
The unit is $\text{pmol}\cdot \text{cell}^{-1} \cdot \text{day}^{-1}$, except for Biomass which has the unit $\text{day}^{-1}$.}
\label{tb:datam11}
\end{table*}
\subsubsection{Particulars of the Error on the Data}
The errors on the measurements ($\Delta Q$) are assumed bounded by an error parameter $\theta_s$ that varies for each metabolite but remains constant between repetitions, \emph{i.e.,\ } $\abs{\Delta Q}_s \leq \theta_s \abs{Q}_s$. The estimation of the error parameter was mostly based on the estimated errors of experimental measurements, along with the evaluated variance in the data set. Finally, consistency with the analysis given by \citet{Goudar2009} was ensured.
The values of $\theta$ for each metabolite are given in Table \ref{tb:alpha}.
\begin{table}[tbh]
\begin{minipage}[c]{0.48\textwidth}
\centering
\scriptsize
\rowcolors{1}{white}{kthgrey}
\begin{tabular}{l|c}
\bf{Metabolite} & \bf{Error ($\theta$ [\%])}\\ \hline
Ala & 13.04 \\
Arg & 17.25 \\
Asn & 20.36 \\
Asp & 13.72 \\
Biomass & 17.42 \\
Cys & 17.61 \\
Glc & 14.73 \\
Gln & 15.39 \\
Glu & 13.73 \\
Gly & 15.47 \\
His & 17.10 \\
Ile & 15.31 \\
Lac & 17.52 \\
Leu & 15.49 \\
Lys & 14.55 \\
Met & 13.78 \\
\ce{NH4+} & 13.96 \\
Phe & 16.05 \\
Pro & 15.29 \\
Ser & 15.94 \\
Thr & 15.71 \\
Trp & 15.01 \\
Tyr & 13.58 \\
Val & 23.05 \\
mAb & 18.57 \\
\end{tabular}
\end{minipage}\begin{minipage}[c]{0.5\textwidth}
\caption{The percentage error on each metabolite, $\theta_i$ (\%)}
\label{tb:alpha}\end{minipage}
\end{table}
\subsection{Description of the Metabolic Network}
The network used in this study is based on a network available in the literature \citep[Section 2.2]{ZamoranoRiveros2012}. The network was extended in several ways to better fit this study. More reactions were made reversible and some transport reactions were added. The final network consists of 101 reactions, whereof 29 are reversible, and 100 metabolites, whereof 28 are external. Metabolites that are included in the network but not measured are \ce{CO2}, Choline, and Ethanolamine.
In some experiments an external metabolite concentration is set to zero in the medium. This does not exclude the optimal solution from using that metabolite in the optimal solution. Hence, in order to get a solution that fits better with the experimental set-up, columns of $A_{x}$ corresponding to reactions from those metabolites are removed, thus blocking the optimal EFMs from using those reactions. For media 1 and 5 those metabolites are mAb.
\subsection{Technicalities on Normalization}
The results are presented based on a normalized version of the EFMs-based MFA.
The normalized version aims at fitting the network with the measurements divided by the average value for each specific metabolite in the medium considered. The network is normed similarly by dividing each row of $A_x$ with the average of the measurement for the corresponding metabolite in the medium considered.
Thus if the average value is defined as,
\begin{equation*}
\bar{q}_{i,g}=\sum_{k=1}^d \frac{q_{i,k,g}}{d},
\end{equation*}
then the external network ($A_x$) and measurements ($Q$) are redefined as follows,
\begin{equation*}
\begin{aligned}
a_{ij}&=\frac{a_{ij}}{\bar{q}_{i,g}} \quad \forall \; j \in J_{ext},\\
q_{i,k,g}&=\frac{q_{i,k,g}}{\bar{q}_{i,g}}.
\end{aligned}
\end{equation*}
Where, $a_{ij}$ is an element from $A_x$ and $J_{ext}$ represents the set of all measured external metabolites in the network. When $\abs{\bar{q}_{i,g}}<0.02$ the value is replaced with $\abs{\bar{q}_{i,g}}=0.02$, in order to avoid dividing by too small values. This minimum is chosen to affect only a few metabolites. For metabolite 1 this affects mAb and Ser, for metabolite 5 this affects those same metabolites along with Gly.
\subsection{Results and Discussion}\label{sec:Res}
In this section the results for two experimental conditions using two different media are given. In order to demonstrate the difference of the EFMs-based MFA with and without robustness the flux over each EFM and the flux to each external metabolite for three levels of error are shown. The levels of error are 0\%, 5\%, and 100\% of the $\theta$ error given in Table \ref{tb:alpha}. Additionally, the effects of adding an interval are examined by considering the results with a given interval on \ce{CO2} for $\theta$ equal either to zero or 100\%. The 0\% error interval is equivalent to the EFMs-based MFA without robustness. The solution with 100 \% of error interval is referred to as the robust solution. The results are shown in the following tables and figures, where Tables \ref{tb:M1EFMsw} and \ref{tb:M5EFMsw} show the flux over each EFM for medium 1 and 5 respectively. Furthermore, the value of the objective functions for the EFMs-based MFA with and without robustness is shown.
Figures \ref{fig:M1EFMsw} and \ref{fig:M5EFMsw}, show in the same manner, the flux over each EFM where the flux has been normed with respect to the flux given by the robust solution, this gives an overview of how different the fluxes are for each error interval.
Tables \ref{tb:M1metw} and \ref{tb:M5metw}
are similarly constructed but show the flux to each external metabolite.
\begin{table*}[p]
\centering
\scriptsize
\rowcolors{1}{white}{kthgrey}
\begin{tabular}{c|p{0.5\linewidth}|rrrrr}
EFM & \bf{Macroscopic Reaction} & $w_0$ & $w_{0.05}$ & $w_1$ & $w_{0,inv}$ & $w_{1,inv}$ \\ \hline
1 &0.5 Glu $\Rightarrow$ 0.5 Ala + 1 \ce{CO2} & 5.71 & 5.84 & 5.90 & 4.36 & 4.23\\
2 &0.5 Glc $\Rightarrow$ 1 Lac & 4.96 & 4.86 & 4.81 & 4.45 & 4.20\\
3 &1 Asn $\Rightarrow$ 1 Lac & 1.58 & 1.58 & 1.58 & 1.58 & 1.58\\
4 &0.5 Glc + 0.5 Asn + 0.5 Ala $\Rightarrow$ 1 Ser + 0.5 Glu + 1 \ce{CO2} & 1.58 & 1.72 & 1.77 & 0.25 & 0.13\\
5 &1 Gln + 1 Asp $\Rightarrow$ 1 Asn + 1 Glu & 1.55 & 1.52 & 1.58 & 1.39 & 1.36\\
6 &1 Ser $\Leftrightarrow$ 1 Gly & 1.47 & 1.67 & 1.64 & 0.33 & 0.27\\
7 &1 Ala + 1 \ce{CO2} $\Leftrightarrow$ 1 Asp & 0.88 & 0.95 & 0.98 & 0.35 & 0.33\\
8 &1 Asp + 1 Gly $\Rightarrow$ 1 Asn + 1 \ce{CO2} & 0.59 & 0.67 & 0.61 & 0.09 & 0.04\\
9 &0.16667 Tyr + 0.83333 Ala $\Rightarrow$ 1 Asp & 0.57 & 0.58 & 0.61 & 0.57 & 0.61\\
10 &1 Lac + 1 Gly $\Rightarrow$ 1 Ala + 1 \ce{CO2} & 0.54 & 0.57 & 0.66 & 0.08 & 0.06\\
11 &0.16667 Leu + 0.16667 Lac + 0.83333 Ala $\Rightarrow$ 1 Asp & 0.50 & 0.49 & 0.46 & 0.38 & 0.30\\
12 &1 Gln $\Rightarrow$ 1 Glu + 1 \ce{NH4+} & 0.34 & 0.34 & 0.28 & 0.49 & 0.50\\
13 &0.05195 Glc + 0.0656 Gln + 0.0278 Ser + 0.046 Arg + 0.0552 Thr + 0.0552 Lys + 0.0644 Val + 0.046 Ile + 0.2596 Leu + 0.0644 Phe + 0.0184 Met + 0.0736 Lac + 0.0744 Gly + 0.2032 Glu + 0.0184 Cys + 0.0184 His + 0.046 Pro + 0.0092 Trp + 0.006 Ethanolamine + 0.0171 Choline $\Rightarrow$ 0.084 Asp + 0.218 \ce{CO2} + 1 Biomass & 0.31 & 0.31 & 0.34 & 0.43 & 0.48\\
14 &1 Ala $\Rightarrow$ 1 Lac + 1 \ce{NH4+} & 0.29 & 0.26 & 0.34 & 0.47 & 0.52\\
15 &1 Ser $\Rightarrow$ 1 Lac + 1 \ce{NH4+} & 0.27 & 0.24 & 0.32 & 0.07 & 0.05\\
16 &1 Gly $\Rightarrow$ 1 \ce{NH4+} + 1 \ce{CO2} & 0.23 & 0.29 & 0.21 & 0.08 & 0.05\\
17 &0.5 Val $\Rightarrow$ 0.5 Ala + 1 \ce{CO2} & 0.22 & 0.22 & 0.22 & 0.22 & 0.22\\
18 &0.5 Glc + 1 Arg $\Rightarrow$ 1 Ser + 1 Glu & 0.21 & 0.23 & 0.23 & 0.21 & 0.23\\
19 &0.03155 Glc + 0.0756 Gln + 0.0368 Asn + 0.046 Arg + 0.1982 Thr + 0.0552 Lys + 0.0644 Val + 0.046 Ile + 0.0828 Leu + 0.0644 Phe + 0.0184 Met + 0.3718 Lac + 0.1452 Ala + 0.0164 Glu + 0.0184 Cys + 0.0184 His + 0.046 Pro + 0.0092 Trp + 0.006 Ethanolamine + 0.0171 Choline $\Rightarrow$ 0.3726 \ce{CO2} + 1 Biomass & 0.17 & 0.17 & 0.18 & 0.06 & 0.04\\
20 &0.5 Asp + 0.25 Phe $\Rightarrow$ 0.25 Ala + 0.5 Glu + 1 \ce{CO2} & 0.14 & 0.13 & 0.17 & 0.16 & 0.17\\
21 &0.16667 Lys + 0.33333 Lac + 0.66667 Ala $\Rightarrow$ 1 Asp & 0.14 & 0.14 & 0.14 & 0.14 & 0.14\\
22 &1 Ala + 0.2 Trp $\Rightarrow$ 0.8 Asp + 0.4 Glu & 0.12 & 0.11 & 0.11 & 0.12 & 0.11\\
23 &0.03155 Glc + 0.0912 Gln + 0.0686 Ser + 0.0368 Asn + 0.046 Arg + 0.0552 Thr + 0.0552 Lys + 0.0644 Val + 0.046 Ile + 0.0828 Leu + 0.3296 Phe + 0.0184 Met + 0.0736 Ala + 0.0744 Gly + 0.0008 Glu + 0.0184 Cys + 0.0184 His + 0.046 Pro + 0.0092 Trp + 0.006 Ethanolamine + 0.0171 Choline $\Rightarrow$ 0.2092 Asp + 0.3064 \ce{CO2} + 1 Biomass & 0.11 & 0.12 & 0.08 & 0.10 & 0.07\\
24 &1 Pro $\Rightarrow$ 1 Glu & 8.52e-2 & 8.19e-2 & 8.15e-2 & 8.52e-2 & 8.15e-2\\
25 &1 Ile $\Rightarrow$ 1 Glu + 1 \ce{CO2} & 8.25e-2 & 7.99e-2 & 7.99e-2 & 8.25e-2 & 7.99e-2\\
26 &1 Cys $\Rightarrow$ 1 Lac + 1 \ce{NH4+} & 5.68e-2 & 4.13e-2 & 2.29e-2 & 7.20e-2 & 5.79e-2\\
27 &1 Gly + 1 Cys $\Rightarrow$ 1 Asn & 5.23e-2 & 7.53e-2 & 9.46e-2 & 3.71e-2 & 5.97e-2\\
28 &1 Thr $\Rightarrow$ 0.5 Gly + 0.5 Glu + 0.5 \ce{CO2} & 4.43e-2 & 4.39e-2 & 4.19e-2 & 5.92e-2 & 5.22e-2\\
29 &1 Ser + 1 Met $\Rightarrow$ 1 Asp + 1 Cys & 3.78e-2 & 3.85e-2 & 3.92e-2 & 3.78e-2 & 3.92e-2\\
30 &1 Asp + 1 His $\Rightarrow$ 1 Asn + 1 Glu & 0.56e-2 & 0.61e-2 & 0.93e-2 & 0.56e-2 & 0.89e-2\\
31 &1 His $\Rightarrow$ 1 Glu + 1 \ce{NH4+} & 0.56e-2 & 0.65e-2 & 0.33e-2 & 0.56e-2 & 0.37e-2\\
32 & AAs $\Rightarrow$ 1 mAb & 0.02e-2 & 0.02e-2 & 0.02e-2 & 0.02e-2 & 0.02e-2\\
33 &0.5 Glc $\Rightarrow$ 1 \ce{CO2} & - & - & - & 1.84 & 2.24\\
Norm & $\norm{A_{x}E w-\frac{1}{d}\sum_{k=1}^dq_{k}}^2$ & 0 & 0.22 & 0.42 & 0 & 0.42 \\
Rob. N.& $\norm{A_{x}E w-\frac{1}{d}\sum_{k=1}^dq_{k}}^2+\mathbf{1}^T t_\theta$ & 463.38 & 444.79 & 442.41 & 463.80 & 442.41
\end{tabular}
\caption{Fluxes of each EFM for medium 1, solved with varied assumed error. $w_0$, $w_{0.05}$, and $w_1$ indicate the results with $\theta$ at $0$, $5$, and $100$ of the values given in Table \ref{tb:alpha} \% respectively. $w_{0,inv}$ and $w_{1,inv}$ indicate the results with an interval given for \ce{CO2} with $\theta$ at $0$ and $100$ \% respectively. $t_\theta=\abs{\tilde{Q}_{\theta}\left(\mathcal{I}A_{x}Ew -Q\right)}$ where $\theta$ is given by Table \ref{tb:alpha}.
The unit of the fluxes is $\text{pmol}\cdot \text{cell}^{-1} \cdot \text{day}^{-1}$, except for Biomass for which the unit is $\text{day}^{-1}$. AAs denotes amino acids.}
\label{tb:M1EFMsw}
\end{table*}
\begin{figure}[htb]
\centering
\includegraphics[trim=100 0 90 0,clip,width=1\linewidth]{FluxEFMsn1}
\caption{Flux over each EFM, normed with respect to the full error robust solution, for medium 1.}
\label{fig:M1EFMsw}
\end{figure}
\begin{table*}[tbh]
\begin{minipage}[c]{0.7\textwidth}
\centering
\scriptsize
\rowcolors{1}{white}{kthgrey}
\begin{tabular}{c|rrrrr}
\bf{Metabolite} & $q_0$ & $q_{0.05}$ & $q_1$ & $q_{0,inv}$ & $q_{1,inv}$ \\ \hline
Ala & 0.44 & 0.44 & 0.44 & 0.44 & 0.44\\
Arg & -0.24 & -0.26 & -0.26 & -0.24 & -0.26\\
Asn & -0.18 & -0.18 & -0.18 & -0.18 & -0.18\\
Asp & 0.06 & 0.07 & 0.07 & 0.06 & 0.07\\
Biomass & 0.59 & 0.60 & 0.60 & 0.59 & 0.60\\
\ce{CO2} & 8.39 & 8.76 & 8.83 & 6.98 & 7.07\\
Choline & -0.01 & -0.01 & -0.01 & -0.01 & -0.01\\
Cys & -0.08 & -0.09 & -0.09 & -0.08 & -0.09\\
Ethanolamine & -36e-4 & -36e-4 & -36e-4 & -36e-4 & -36e-4\\
Glc & -3.40 & -3.43 & -3.43 & -3.40 & -3.43\\
Gln & -1.93 & -1.91 & -1.90 & -1.93 & -1.90\\
Glu & 0.28 & 0.28 & 0.28 & 0.28 & 0.28\\
Gly & 0.04 & 0.05 & 0.05 & 0.04 & 0.05\\
His & -0.02 & -0.02 & -0.02 & -0.02 & -0.02\\
Ile & -0.11 & -0.11 & -0.11 & -0.11 & -0.11\\
Lac & 6.40 & 6.20 & 6.20 & 6.40 & 6.20\\
Leu & -0.19 & -0.19 & -0.19 & -0.19 & -0.19\\
Lys & -0.06 & -0.06 & -0.06 & -0.06 & -0.06\\
Met & -0.05 & -0.05 & -0.05 & -0.05 & -0.05\\
NH4 & 1.19 & 1.18 & 1.18 & 1.19 & 1.18\\
Phe & -0.10 & -0.10 & -0.10 & -0.10 & -0.10\\
Pro & -0.11 & -0.11 & -0.11 & -0.11 & -0.11\\
Ser & -0.00 & -0.00 & -0.01 & -0.00 & -0.01\\
Thr & -0.10 & -0.10 & -0.10 & -0.10 & -0.10\\
Trp & -0.03 & -0.03 & -0.03 & -0.03 & -0.03\\
Tyr & -0.10 & -0.10 & -0.10 & -0.10 & -0.10\\
Val & -0.15 & -0.15 & -0.15 & -0.15 & -0.15\\
mAb & 2e-4 & 2e-4 & 2e-4 & 2e-4 & 2e-4
\end{tabular}
\end{minipage}\begin{minipage}[c]{0.28\textwidth}
\caption{Fluxes of each metabolite for medium 1, solved with varied assumed error. $w_0$, $w_{0.05}$, and $w_1$ indicate the results with $\theta$ at $0$, $5$, and $100$ \% of its given value in Table \ref{tb:alpha} respectively. $w_{0,inv}$ and $w_{1,inv}$ indicate the results with an interval given for \ce{CO2} with $\theta$ at $0$ and $100$ \% respectively.
The unit of the fluxes is $\text{pmol}\cdot \text{cell}^{-1} \cdot \text{day}^{-1}$, except for Biomass for which the unit is $\text{day}^{-1}$.}
\label{tb:M1metw}\end{minipage}
\end{table*}
Considering the results for medium one shown in Table \ref{tb:M1EFMsw} it can be noted that introducing robustness has no impact on which EFMs are required. Further, the change in the norm and robust norm is low, indicating that the impact of robustness on the solution is low. When the flux to each metabolite shown in Table \ref{tb:M1metw} is considered, there are some minor adjustments in the solution when robustness is added, however the larges change comes when an interval on \ce{CO2} is enforced, resulting in a lower flow to \ce{CO2}.
Considering the results from medium five shown in Table \ref{tb:M5EFMsw} it can be noted that for all levels of robustness one more EFM is required in the solution in order to reach optimality. When the flux to \ce{CO2} is bounded a few EFMs are dropped and new ones are introduced.
In general introducing a new metabolite in the model by giving an interval of its value, \emph{i.e.,\ } without experimental measurements, has a large impact on the solution than introducing robustness. Furthermore the intervals can be satisfied without a large change in the norm of the data fitting. This indicates that giving realistic intervals on unmeasured metabolites can help guide the solution. These results show that, here, larger improvements of the residual error are brought by improving the model structure, i.e. introducing new metabolites, compared to taking into account the robust solution, which address the parameter estimation.
\begin{table*}[p]
\centering
\tiny
\rowcolors{1}{white}{kthgrey}
\begin{tabular}{c|p{0.6\linewidth}|rrrrr}
EFM & \bf{Macroscopic Reaction} & $w_0$ & $w_{0.05}$ & $w_1$ & $w_{0,inv}$ & $w_{1,inv}$ \\ \hline
1 &0.5 Glc $\Rightarrow$ 1 Lac & 3.20 & 3.82 & 2.30 & 2.69 & 2.02\\
2 &0.5 Glu $\Rightarrow$ 0.5 Ala + 1 \ce{CO2} & 2.64 & 2.72 & 3.13 & 2.53 & 2.48\\
3 &0.5 Glc + 1 Ala $\Rightarrow$ 1 Ser + 1 Lac & 2.62 & 1.93 & 3.44 & 3.22 & 3.72\\
4 &0.5 Gln + 0.5 Lac $\Rightarrow$ 1 Asp & 2.05 & 2.27 & 2.20 & 1.49 & 1.62\\
5 &1 Ser $\Leftrightarrow$ 1 Gly & 1.94 & 2.15 & 2.33 & 2.45 & 2.32\\
6 &1 Asp + 1 Gly $\Rightarrow$ 1 Asn + 1 \ce{CO2} & 1.89 & 2.09 & 2.26 & 1.12 & 1.29\\
7 &1 Asn $\Rightarrow$ 1 Lac & 1.60 & 1.70 & 1.74 & 1.60 & 1.69\\
8 &0.5 Glu $\Rightarrow$ 0.5 Lac + 0.5 \ce{NH4+} + 1 \ce{CO2} & 1.43 & 1.62 & 1.09 & 1.13 & 1.27\\
9 &0.5 Asn + 1 Lac $\Rightarrow$ 0.5 Ala + 0.5 Glu + 1 \ce{CO2} & 1.24 & 1.32 & 1.11 & 3e-7 & --\\
10 &1 Ser $\Rightarrow$ 1 Ala & 1.01 & 0.22 & 1.60 & 1.07 & 1.89\\
11 &1 Gln + 1 Asp $\Rightarrow$ 1 Asn + 1 Glu & 0.54 & 0.35 & 0.37 & 0.73 & 0.69\\
12 &0.5 Glc + 1 Arg $\Rightarrow$ 1 Ser + 1 Glu & 0.36 & 0.46 & 0.49 & 0.35 & 0.50\\
13 &0.25 Tyr $\Rightarrow$ 0.25 Glu + 1 \ce{CO2} & 0.34 & 0.26 & 0.28 & 0.26 & 0.29\\
14 &1 Asn $\Rightarrow$ 1 Asp + 1 \ce{NH4+} & 0.32 & 0.20 & 0.46 & 0.38 & 0.42\\
15 &0.25465 Glc + 0.0656 Gln + 0.0468 Asn + 0.0616 Asp + 0.046 Arg + 0.0552 Thr + 0.0552 Lys + 0.0644 Val + 0.046 Ile + 0.5446 Leu + 0.0644 Phe + 0.0184 Met + 0.0736 Ala + 0.0744 Gly + 0.0184 Cys + 0.0184 His + 0.046 Pro + 0.0092 Trp + 0.006 Ethanolamine + 0.0171 Choline $\Rightarrow$ 0.3668 Glu + 0.9804 \ce{CO2} + 1 Biomass & 0.30 & 0.27 & 0.23 & 2e-5 & --\\
16 &0.027502 Glc + 0.063938 Gln + 0.066862 Ser + 0.035867 Asn + 0.044834 Arg + 0.053801 Thr + 0.053801 Lys + 0.32125 Val + 0.044834 Ile + 0.080702 Leu + 0.32125 Phe + 0.017934 Met + 0.072515 Gly + 0.025731 Glu + 0.017934 Cys + 0.017934 His + 0.044834 Pro + 0.0089669 Trp + 0.005848 Ethanolamine + 0.016667 Choline $\Rightarrow$ 0.024951 Lac + 0.36569 Ala + 1 \ce{CO2} + 0.97466 Biomass & 0.26 & 0.08 & 0.09 & 0.04 & 0.03\\
17 &1 Gln $\Rightarrow$ 1 Glu + 1 \ce{NH4+} & 0.25 & 0.27 & 0.28 & 0.32 & 0.23\\
18 &0.52941 Val + 0.17647 Leu + 0.058824 Pro $\Rightarrow$ 0.058824 Arg + 0.52941 Glu + 1 \ce{CO2} & 0.12 & 0.22 & 0.22 & 0.23 & 0.20\\
19 &1 Thr $\Rightarrow$ 1 Asp & 0.10 & 0.10 & 0.10 & 0.10 & 0.10\\
20 &0.33333 Asn + 0.33333 Asp + 0.33333 Lys + 0.33333 Pro $\Rightarrow$ 0.33333 Arg + 0.66667 Glu + 1 \ce{CO2} & 0.10 & 0.12 & 0.12 & 0.09 & 0.12\\
21 &1 Ile + 1 Cys $\Rightarrow$ 1 Asp + 1 Glu & 0.09 & 0.08 & 0.07 & 0.09 & 0.08\\
22 &0.16667 Glc + 0.33333 Asn + 0.33333 Met + 0.33333 Pro $\Rightarrow$ 0.33333 Arg + 1 Lac + 0.33333 \ce{CO2} & 0.08 & 0.05 & 0.05 & 0.06 & 0.05\\
23 &0.25 Ala + 0.25 Trp $\Rightarrow$ 0.5 Glu + 1 \ce{CO2} & 0.07 & 0.07 & 0.06 & 0.06 & 0.08\\
24 &1 Pro $\Rightarrow$ 1 Glu & 0.05 & 0.05 & 0.05 & 0.05 & 0.05\\
25 &0.028217 Glc + 0.0756 Gln + 0.0686 Ser + 0.0368 Asn + 0.0616 Arg + 0.2044 Tyr + 0.0552 Thr + 0.0552 Lys + 0.0644 Val + 0.2228 Ile + 0.0828 Leu + 0.0368 Phe + 0.0184 Met + 0.0744 Gly + 0.0008 Glu + 0.0184 Cys + 0.0184 His + 0.046 Pro + 0.0092 Trp + 0.006 Ethanolamine + 0.0171 Choline $\Rightarrow$ 0.224 Ala + 0.4956 \ce{CO2} + 1 Biomass & 0.04 & 0.15 & 0.17 & 0.07 & 0.05\\
26 &0.028217 Glc + 0.0656 Gln + 0.5834 Ser + 0.046 Arg + 0.0276 Tyr + 0.0552 Thr + 0.3204 Lys + 0.0644 Val + 0.046 Ile + 0.0828 Leu + 0.0368 Phe + 0.5332 Met + 0.0744 Gly + 0.0264 Glu + 0.0184 His + 0.046 Pro + 0.0092 Trp + 0.006 Ethanolamine + 0.0171 Choline $\Rightarrow$ 0.3484 Ala + 0.468 \ce{NH4+} + 0.9736 \ce{CO2} + 0.4964 Cys + 1 Biomass & 0.03 & 0.05 & 0.06 & 0.04 & 0.05\\
27 &1 Cys $\Rightarrow$ 1 Ala & 0.03 & 0.06 & 0.07 & 0.04 & 0.06\\
28 &0.024997 Glc + 0.058115 Gln + 0.060773 Ser + 0.032601 Asn + 0.06343 Asp + 0.040751 Arg + 0.024451 Tyr + 0.048901 Thr + 0.048901 Lys + 0.057052 Val + 0.040751 Ile + 0.073352 Leu + 0.032601 Phe + 0.0163 Met + 0.065911 Gly + 0.038979 Cys + 0.0163 His + 0.050319 Pro + 0.25 Trp + 0.0053154 Ethanolamine + 0.015149 Choline $\Rightarrow$ 0.022679 Lac + 0.17665 Ala + 1 \ce{CO2} + 0.8859 Biomass & 0.02 & 0.03 & 0.04 & 0.03 & 0.02\\
29 &1 Lac + 1 His $\Rightarrow$ 1 Ala + 1 Glu & 99e-4 & 118e-4 & 175e-4 & 100e-4 & 151e-4\\
30 &1 His $\Rightarrow$ 1 Glu + 1 \ce{NH4+} & 99e-4 & 92e-4 & 35e-4 & 99e-4 & 59e-4\\
31 & AAs $\Rightarrow$ 1 mAb & 2e-4 & 2e-4 & 2e-4 & 2e-4 & 2e-4\\
32 &0.33333 Asp + 0.33333 Phe $\Rightarrow$ 0.66667 Glu + 1 \ce{CO2} & - & 0.18 & 0.18 & 0.21 & 0.24\\
33 &1 Lac + 1 Gly $\Rightarrow$ 1 Ala + 1 \ce{CO2} & - & - & - & 1.28 & 0.95\\
34 &0.028217 Glc + 0.0656 Gln + 0.0686 Ser + 0.0368 Asn + 0.0716 Asp + 0.046 Arg + 0.0276 Tyr + 0.0552 Thr + 0.0552 Lys + 0.0644 Val + 0.046 Ile + 0.343 Leu + 0.0368 Phe + 0.0184 Met + 0.2346 Lac + 0.0736 Ala + 0.0744 Gly + 0.0184 Cys + 0.0184 His + 0.046 Pro + 0.0092 Trp + 0.006 Ethanolamine + 0.0171 Choline $\Rightarrow$ 0.2238 Glu + 0.0368 \ce{CO2} + 1 Biomass & - & - & - & 0.45 & 0.42\\
& $\norm{A_{x}E w-\frac{1}{d}\sum_{k=1}^dq_{k}}^2$ & 0.00 & 0.69 & 2.27 & 0.01 & 2.27 \\
& $\norm{A_{x}E w-\frac{1}{d}\sum_{k=1}^dq_{k}}^2+\mathbf{1}^T t_\theta$ &1151.23 & 1073.57 & 941.05 & 1154.05 & 945.09
\end{tabular}
\caption{Fluxes of each EFM for medium 5, solved with varied assumed error. $w_0$, $w_{0.05}$, and $w_1$ indicate the results with $\theta$ at $0$, $5$, and $100$ of the values given in Table \ref{tb:alpha} \% respectively. $w_{0,inv}$ and $w_{1,inv}$ indicate the results with an interval given for \ce{CO2} with $\theta$ at $0$ and $100$ \% respectively. $t_\theta=\abs{\tilde{Q}_{\theta}\left(\mathcal{I}A_{x}Ew -Q\right)}$ where $\theta$ is given by Table \ref{tb:alpha}.
The unit of the fluxes is $\text{pmol}\cdot \text{cell}^{-1} \cdot \text{day}^{-1}$, except for Biomass for which the unit is $\text{day}^{-1}$. AAs denotes amino acids.}
\label{tb:M5EFMsw}
\end{table*}
\begin{figure}[htb]
\centering
\includegraphics[trim=100 0 90 0,clip,width=1\linewidth]{FluxEFMsn5}
\caption{Flux over each EFM, normed with respect to the full error robust solution, for medium 5.}
\label{fig:M5EFMsw}
\end{figure}
\begin{table*}[tbh]
\begin{minipage}[c]{0.7\textwidth}
\centering
\scriptsize
\rowcolors{1}{white}{kthgrey}
\begin{tabular}{c|rrrrr}
\bf{Metabolite} & $q_0$ & $q_{0.05}$ & $q_1$ & $q_{0,inv}$ & $q_{1,inv}$\\ \hline
Ala & 0.44 & 0.44 & 0.44 & 0.45 & 0.44\\
Arg & -0.32 & -0.42 & -0.45 & -0.32 & -0.45\\
Asn & -0.20 & -0.20 & -0.20 & -0.20 & -0.20\\
Asp & 0.08 & 0.08 & 0.08 & 0.08 & 0.08\\
Biomass & 0.64 & 0.57 & 0.57 & 0.64 & 0.57\\
\ce{CO2} & 8.47 & 9.11 & 8.95 & 7.09 & 7.09\\
Choline & -0.01 & -0.01 & -0.01 & -0.01 & -0.01\\
Cys & -0.12 & -0.12 & -0.12 & -0.12 & -0.12\\
Ethanolamine & -38e-4 & -34e-4 & -34e-4 & -38e-4 & -34e-4\\
Glc & -3.19 & -3.19 & -3.19 & -3.16 & -3.14\\
Gln & -1.85 & -1.79 & -1.79 & -1.84 & -1.77\\
Glu & 0.24 & 0.23 & 0.23 & 0.24 & 0.24\\
Gly & 0.01 & 0.02 & 0.03 & 0.01 & 0.03\\
His & -0.03 & -0.03 & -0.03 & -0.03 & -0.03\\
Ile & -0.13 & -0.13 & -0.13 & -0.13 & -0.11\\
Lac & 5.95 & 5.84 & 5.84 & 6.00 & 6.24\\
Leu & -0.21 & -0.21 & -0.19 & -0.21 & -0.19\\
Lys & -0.07 & -0.09 & -0.09 & -0.07 & -0.09\\
Met & -0.05 & -0.05 & -0.05 & -0.05 & -0.05\\
\ce{NH4+} & 1.30 & 1.31 & 1.31 & 1.30 & 1.31\\
Phe & -0.11 & -0.11 & -0.11 & -0.11 & -0.11\\
Pro & -0.14 & -0.15 & -0.15 & -0.14 & -0.15\\
Ser & -0.01 & -0.02 & -0.06 & -0.01 & -0.06\\
Thr & -0.13 & -0.13 & -0.13 & -0.13 & -0.13\\
Trp & -0.03 & -0.03 & -0.03 & -0.03 & -0.03\\
Tyr & -0.09 & -0.10 & -0.11 & -0.09 & -0.10\\
Val & -0.17 & -0.17 & -0.18 & -0.17 & -0.15\\
mAb & 2e-4 & 2e-4 & 2e-4 & 2e-4 & 2e-4
\end{tabular}
\end{minipage}\begin{minipage}[c]{0.28\textwidth}
\caption{Fluxes of each metabolite for medium 5, solved with varied assumed error. $w_0$, $w_{0.05}$, and $w_1$ indicate the results with $\theta$ at $0$, $5$, and $100$ \% of its given value in Table \ref{tb:alpha}, respectively. $w_{0,inv}$ and $w_{1,inv}$ indicate the results with an interval given for \ce{CO2} with $\theta$ at $0$ and $100$ \% respectively.
The unit of the fluxes is $\text{pmol}\cdot \text{cell}^{-1} \cdot \text{day}^{-1}$, except for Biomass for which the unit is $\text{day}^{-1}$.}
\label{tb:M5metw}
\end{minipage}
\end{table*}
\section{Conclusion} \label{sec:conc}
In this work we have examined the effect of errors in measurement on the solution of the EFMs-based MFA.
The approach has been to derive a robust form of the EFMs-based MFA, a form that considers measurement errors more directly, in the sense that each value based on measurements is given an error interval and the aim is to minimize the maximum possible error on the given interval with respect to the least-squares measure.
These types of robust optimization problems are in general not easily solved \citep{Ghaoui1997}. However, for this special case we showed that the robust form can be stated as a convex quadratic optimization problem where column generation can be applied to achieve an optimal solution. Furthermore, it was demonstrated that an external metabolite could be taken into account in the model when no measurements exists this metabolite, by considering an interval of its value.
By considering the worst-case scenario, we see how unfavourable the fit could be within the given error intervals. The case-study compares the solution of the non-robust and robust EFMs-based MFA. Those results demonstrate that the optimal solution to the robust problem is similar to the optimal solution to the non-robust formulation, \emph{i.e.,\ } the EFMs-based MFA is rather robust to those assumed errors. This indicates that the errors on measurement do not induce a large change in the solution.
Measurement errors were known to be around 20\%, therefore it was important to consider the effects of those errors, especially with respect to which EFMs are used in the optimal solution. Our second contribution, the addition of intervals, is especially relevant for metabolites for which known intervals are available in the literature but are problematic to measure in the experimental setup. In fact, the addition of intervals had a larger effect on the solution than the addition of around 20\% errors in measurements in the presented cases.
The robust solution addressed the parameter estimation, while adding a metabolite by considering its interval, addressed the model structure.
We showed a way to include the knowledge of unmeasured metabolites in the column generation method. This important result allowed to achieve improvements of the model structure. This approach can be a general strategy to improve a model structure by introducing a new metabolite in a model. It can also be an approach to identify which variables should be measured when designing a new experiment.
\renewcommand{\abstractname}{Acknowledgements}
\begin{abstract}
The work of the authors from the Department of Mathematics was supported by the Swedish Research Council. The work of the authors from the Division of Industrial Biotechnology was supported by KTH and the Swedish Governmental Agency for Innovation Systems (VINNOVA). The CHO cell line was kindly provided by Selexis (Switzerland). Culture media were kindly provided by Irvine Scientific (CA, USA).
Finally, we thank the editor and the two anonymous referees for their valuable comments and suggestions.
\end{abstract}
|
\section{Introduction}\label{sec:intro}
In the standard $\Lambda$CDM scenario, the first stars, the so-called
Population III (PopIII) stars, are predicted to
form in dark matter minihalos of typical mass $\sim 10^6\;{\rm M}_{\odot}$ at
$z\sim 20-30$ out of a gas of pristine composition \citep[see][for recent reviews]{Ciardi2005,BrommYoshida2011}. Thanks to their peculiar chemical
composition, they are expected to be more massive than the subsequent
stellar populations with typical mass of $\sim 40\;{\rm M}_{\odot}$ \citep{Hosokawa2011},
possibly extending up to hundred solar masses \citep{Hirano2014}. The formation of
these stars mark the fundamental transition from a simple, very
homogeneous Universe to the complex and structured one we see already
in place a billion years after the Big Bang. During this period of
time two fundamental transitions are
expected to occur: (i) a change in the SFR mode, i.e. from massive
PopIII to solar-size PopII/I stars and (ii) the cosmic
reionization, i.e. the change of the inter-galactic
medium (IGM) from a neutral
to a fully ionized state. There is a general consensus that the first
transition is driven by the so-called chemical feedback
\citep{Schneider2002}, i.e. the enrichment of
star forming clouds by the first supernova explosions above a critical threshold of $Z_{\rm crit}=10^{5\pm
1}\;{\rm Z}_{\odot}$ \citep{Schneider2003}. As the chemical feedback is
essentially a local effect, on cosmological scale the two populations
should be coeval for a long period of time \citep{Schneider2006,Tornatore2007,Maio2010}. Understanding how and when the change in the SFR mode happened
is one of the main goals for current studies of galaxy formation in
the early Universe and the detection of PopIII stars one of the main
challenges of the next generation of space and ground facilities
\citep[see][for a review]{BrommYoshida2011}. The cosmic reionization of the IGM has been
extensively studied in the
last years both from a theoretical and observational point of view
\citep[see][for a recent review]{Loeb2013}. While many steps towards our understanding
of reionization process have been done, still many fundamental details are missed: In
which way does it proceed? How gradual and how prolonged was the process? Was
radiation from early stars sufficient to sustain this phase
transition? Do PopIII stars or quasars have a major role in driving the
process? Is some other more exotic process at work? etc.
Historically, the exploration of the distant Universe has been
carried out following two main pathways: the observations of bright
quasars detected in wide shallow surveys \citep{Fan2012}, and of distant galaxies identified through the drop-out
technique in small fields
\citep{Bouwens2014}. Thanks to their extreme brightness Gamma-Ray Bursts (GRBs)
represent an alternative way to access those early epochs.
As demonstrated by GRB~090423 at $z=8.2$
\citep{Salvaterra2009,Tanvir2009}, they can be detected even at
distances much larger than any other cosmic object. In principle, their afterglow emission can be
observed up to $z\sim 20$ \citep{Ciardi2000,Guo2004} providing useful
information about the ionization and metal enrichment history of the
early Universe. Here, I will review the present observational status of
this research field and discuss the possible role of GRBs in
the exploration of the Universe during and before the reionization epoch.
This paper is organized as follows. In Section~2, a brief summary of the
observations of the GRBs at $z>6$ and of their host
galaxies is given. Section~3 presents the expected rate of high-$z$ GRBs
both from PopII and PopIII stars and the expected properties
of the relative host galaxies. In Section~4, the use of GRBs as a
probe of the early
Universe is reviewed. Finally in Section~5 I present some ideas for the future of
this research field. The conclusions are drawn in Section~6.
\section{Observations}
\begin{table*}
\begin{center}
\begin{tabular}{lcccccccccc}
\hline
\hline
GRB & $z$ & $E_{\rm p}$ & $E_{\rm iso}$ & $\log(N_{\rm HI})$ & $\log(N_{\rm H,X})$ & $Z$ &
$A_V$ & $M_{\rm UV,host}$ & SFR$_{\rm host}$ \\
& & [erg] & [erg] & [cm$^{-2}$] & [$10^{21}$ cm$^{-2}$] &
[${\rm Z}_{\odot}$] & & [AB] & [${\rm M}_{\odot} {\rm yr}^{-1}$] \\
\hline
050904 & 6.3 & 3178 & $1.24\times 10^{54}$ & 21.6
& $63^{+34}_{-29}$& $-1.6\pm 0.3$ & $0.15\pm0.07$& $>-19.95$ &
$<4.1$ \\
080913 & 6.7 & 1008 & $7\times 10^{52}$ & 19.84 & $95^{+89}_{-77}$ & -
& $0.12\pm 0.03$ & $>-19.00$ & $<1.3$\\
090423 & 8.2 & 746 & $1.88\times 10^{53}$ & - &
$102^{+49}_{-54}$& - & $<0.1$ & $>-16.95$ & $<0.38$ \\
130606A & 5.9 & 2028 & $2.7\times 10^{53}$& 19.93 & $<30$ &$-1.35\pm
0.15$ & $<0.05$ & - & -\\
140515A & 6.3 & 376 & $5.1\times 10^{52}$& 18.62
&$<226$ &$<-0.8$ & $0.11\pm 0.02$ & - & -\\
\hline
090429B & 9.4 & 437 & $4.31\times 10^{52}$& - &$140\pm10$ &- & $0.10\pm 0.02$ &$>-19.65$ & $<2.4$\\
120521C & 6.0 & - & $1.9\times 10^{53}$ & - & $<60$ & - & $<0.05$ & - & -\\
120923A & 8.5 & 376 & $5.1\times 10^{52}$& - &$<720$ &- & - & - & - \\
\hline
\hline
\end{tabular}
\end{center}
\caption{List of the GRBs at $z>6$ detected by {\it Swift} and of
their observed properties: peak energy ($E_{\rm p}$), isotropic equivalent
energy in the 1-10000 keV range ($E_{\rm iso}$), hydrogen column
density in the host ($N_{\rm HI}$), X-ray equivalent hydrogen column
density ($N_{\rm H,X}$), host metallicity ($Z$) and dust extinction ($A_V$). The last two columns report
the limits on the host galaxy luminosity ($M_{\rm
UV,host}$) and SFR. See text for references. }
\end{table*}
\subsection{The Swift high-$z$ GRB sample}
In ten years of operations {\it Swift} has detected a handful of bursts
with spectroscopically confirmed redshift larger than 6. In
addition, other three GRBs have well constrained photometric redshift above
this limit. The
observed high-$z$ sample represents $\sim 1$\% of all {\it
Swift} bursts, $\sim 2.5$\% of those with known $z$. The main properties of the $z>6$ GRB sample
are given in Table~1.
\begin{itemize}
\item{{\bf GRB~050904} at z=6.3 \citet{Kawai2006}}
This burst was firstly imagined by the 25-cm telescope TAROT \citep{Klotz2005}. Its high-$z$ nature was recognized by multi-wavelength
photometric data \citep{Haislip2006,Tagliaferri2005} and firmly confirmed
spectrocopically three days after the {\it Swift} trigger by the
Subaru telescope \citep{Kawai2006}. The afterglow spectrum provided
a upper limit on the neutral hydrogen fraction at the GRB
redshift of $x_{\rm HI}<0.17$ at 1$\sigma$ confidence \citep{Totani2006} and a measure of the metallicity at the
level of $\sim 0.1\;{\rm Z}_{\odot}$ \citep{Kawai2006}. Recently, \citet{Thoene2013}
revised this value, inferring a slightly lower metallicity from the S
II $\lambda1243$ equivalent width, resulting in $\log(Z/{\rm Z}_{\odot})=-1.6\pm 0.3$. The afterglow spectral energy
distribution (SED) requires the presence of SMC or supernova (SN) type
dust at a level of $A_V=0.15\pm 0.07$ (\citealt{Stratta2011} but see
\citealt{Zafar2011a}). The modeling of afterglow data from X-ray to
radio suggests GRB~050904 to be an energetic burst blowing up in a dense
medium with $n\simeq 680$ cm$^{-3}$ \citep{Frail2006,Laskar2014}.
\item{{\bf GRB~080913} at $z=6.7$ \citep{Greiner2009}}
The high-$z$ nature of this burst was recognized via the detection of
a spectral break between the i$^\prime$ and z$^\prime$ bands of the
GROND instrument and then confirmed spectroscopically by VLT observations. The analysis of the red damping wing constrained $x_{\rm HI}<0.73$ at
90\% confidence level \citep{Patel2010}. In spite of the rapid
follow-up campaign, the faintness
of the afterglow prevented the detection of any metal absorption line, but
the S$_{\rm II}$+Si$_{\rm II}$ at 2.9$\sigma$ level \citep{Patel2010}.
Dust absorption of $A_V=0.12\pm 0.03$ is found from SED fitting \citep{Zafar2011b}.
\item{{\bf GRB~090423} at $z=8.2$ \citep{Salvaterra2009,Tanvir2009}}
The spectroscopic redshift of this burst was secured by TNG \citep{Salvaterra2009} and by VLT
\citep{Tanvir2009}, and still represents the distance record for a
cosmic object. Radio observations by
VLA were reported by \citet{Chandra2010}. The analysis of the
multi-wavelength dataset show that the afterglow is reminiscent of many
other lower redshift bursts, suggesting that in spite of its extreme
redshift, its progenitor and the medium in which it blowed up were not
peculiar. Indeed, its detection is consistent with the
high-$z$ tail of PopII/I GRB redshift distribution
\citep{Salvaterra2009}. From the X-ray to optical SED no absorption by dust is
evident with $A_V<0.1$ (\citealt{Salvaterra2009,Tanvir2009}, but see \citealt{Laskar2014}).
\item{{\bf GRB 130606A} at $z=5.9$
\citep{Chornock2013,CastroTirado2013,Totani2014,Hartoog2014}}
The redshift of this burst was obtained by Gemini-North
\citep{Chornock2013}, GTC \citep{CastroTirado2013}, Subaru \citep{Totani2014} and VLT \citep{Hartoog2014}. In particular, the
superior resolution and wavelength coverage of the VLT/X-shooter
instrument showed the potentiality of GRBs as tool to study in great
details the metal enrichment of star forming region inside high$-z$ galaxies
\citep{Hartoog2014}. Precise column densities of H, Al, Si and Fe are
reported together with limit on C, O, S and Ni. The host metallicity
is constrained to be in the range of 0.03-0.06 solar and the high [Si/Fe] in
the host suggests the presence of dust depletion (though $A_V<0.05$
from SED fitting). The best fit of
the Ly$\alpha$ absorption line is obtained for $\log(N_{\rm
HI})=19.94\pm 0.01$ and negligible neutral hydrogen in the external
medium, with $x_{\rm HI}<0.03$ at 3$\sigma$ significance.
\item{{\bf GRB 140515A} \citep{Chornock2014,Melandri2015}}
The redshift of this burst has been secured by Gemini-North
\citep{Chornock2014}, GTC and VLT
\citep{Melandri2015}. \citet{Chornock2014} analysed the Gemini-North
spectra finding no evidence
of narrow absorption lines, indicating a host metallicity
$Z<0.15\;{\rm Z}_{\odot}$. However, \citet{Melandri2015} by modeling the X-ray to optical SED found evidence for
dust absorption to the level of $A_V\sim 0.1$ indicating some metal enrichment.
The red damping wing of Lyman-$\alpha$ can be fitted
equally well by a single host galaxy absorber with $\log(N_{\rm
HI})=18.62\pm 0.08$ or a pure IGM absorption with neutral hydrogen
fraction $x_{\rm HI}\sim 0.06$ \citep{Chornock2014}.
\end{itemize}
Other three GRBs have accurate photometric redshift measurement that place
in the $z>6$ sample:
\begin{itemize}
\item{{\bf GRB 090429B} at $z\sim 9.4$ \citep{Cucchiara2011}}
\citet{Cucchiara2011} collected the afterglow data obtained with
Gemini-North, VLT and GROND. In the best
fit model these data are all consistent with a photometric redshift of $z=9.4$ and low
extinction $A_V=0.10\pm 0.02$. A secondary solution at very low
redshift is still allowed by the SED fitting, but it seems
unlikely due to the lack of any
detection of the GRB host galaxy (see next Section).
\item{{\bf GRB 120521C} at $z\sim 6$ \citep{Laskar2014}}
The large multi-wavelength dataset of the afterglows lead to
$z\simeq 6.0$ also supported by the analysis of the very low
signal-to-noise Gemini-North spectrum. All afterglow
data after $\sim 0.25$ days can be well fitted assuming
constant-density medium with $n_0\sim 0.05$ cm$^{-3}$ and no dust
extinction, $A_V<0.05$. The radio observations revealed the existence
of a jet break corresponding to an jet angle $\theta_{\rm jet}\simeq 3$
degrees. An isotropic energy $E_{\rm iso}=1.9\pm0.8 \times 10^{53}$
erg is obtained by extrapolating the BAT fluence.
\item{{\bf GRB 120923A} at $z\sim 8.5$ \citep{Tanvir2013}}
The optical-NIR afterglow of this burst was imagined by Gemini-North
telescope $\sim 1.4$ hours after the {\it Swift} trigger. The blue IR color, H-K$\sim 0.1$ mag together
with the break between the Y- and J-band is suggestive of a strong
high-$z$ candidate \citep{Levan2012}. The fit to all photometric data gives $z\sim
8.5$ \citep{Tanvir2013}. No radio or mm
detection is reported.
\end{itemize}
\subsection{General properties of high-$z$ GRBs}\label{sec:obsprop}
In spite of the burst to burst differences and of the small size of
the sample, some interesting clues about the nature
of high-$z$ bursts can be gathered by comparing them with
bursts at lower redshifts. Thus, in the following, I will compare the
properties of $z>6$ bursts with those of a well selected,
complete sample of bright {\it Swift} bursts \citep[BAT6,][]{Salvaterra2012}. The BAT6 sample has been selected only on the basis
of the $\gamma$-ray peak flux as seen by {\it Swift}/BAT and is
characterized by a very high completeness in redshift \citep[95\%,][]{Salvaterra2012,Covino2013}. Thus, it represents a perfect sub-sample to
compare the results of the high-$z$ burst population with.
The top-left panel of Fig.~1 shows the position of $z>6$ bursts in the
$E_{\rm p}-E_{\rm iso}$ plane. The black points represents the BAT6 sample
\citep{Nava2012} and the shaded area shows the 3$\sigma$ scatter around the
$E_{\rm p}-E_{\rm iso}$ correlation.
High-$z$ bursts nicely follow the Amati correlation suggesting that their
prompt emission properties do not differ significantly from bright
lower-$z$ bursts. This consistency calls for a
common central engine (and progenitor star).
\begin{figure*}\label{fig1}
\centering
{\includegraphics[scale=0.4]{review_salvaterra_fig1a.ps}}
{\includegraphics[scale=0.4]{review_salvaterra_fig1b.ps}}
{\includegraphics[scale=0.4]{review_salvaterra_fig1c.ps}}
{\includegraphics[scale=0.4]{review_salvaterra_fig1d.ps}}
\caption{Comparison of the properties of $z>6$ bursts (red squares and
lines) with those of a
well selected complete sub-sample of bright {\it Swift} GRB, the BAT6
sample (black points and lines; \citealt{Salvaterra2012}). {\it
Top-left panel}: $E_{\rm p}-E_{\rm iso}$ correlation where BAT6 bursts are
from \citet{Nava2012}. The shaded regions represents the $3\sigma$ scatter around the
best-fitting relation. {\it Top-right panel}:
rest-frame 2-10 keV light curves normalized to their isotropic energy where BAT6
data are from \citet{DAvanzo2012}. The shaded region
represents the 2$\sigma$ scatter around the mean value of
the $L_{\rm X}/E_{\rm iso}$ distributions at the given rest-frame time. {\it Bottom-left panel}: dust extinction as a
function of redshift, where BAT6 data are from \citet{Covino2013}. {\it Bottom-right panel}:
intrinsic X-ray equivalent hydrogen column densities $N_{\rm H,X}$ as
a function of redshift, where BAT6 data are from \citet{Campana2012}. The shaded regions represent the effect of intervening
material along the line-of-sight (see \citealt{Campana2015} for the
details).}
\end{figure*}
In the top-right panel of Fig.~1 I have then compared the rest-frame
2-10 keV light
curves normalized for the corresponding $E_{\rm iso}$ of $z>6$
bursts with those obtained from the BAT6 sample \citep{DAvanzo2012}.
Indeed, the X-ray light curves of low-$z$ GRBs tend to cluster when such a
normalization is performed and strong correlation between $L_{X}$ and
$E_{\rm iso}$ is found at early times. This effect is seen also in high-$z$
GRBs and their X-ray light curves can not be distinguish from
those of low- and intermediate-$z$ bursts. This conclusion is also supported by more extensive,
although model dependent, analysis of afterglow emission. It has been shown
\citep{Chandra2010,Laskar2014} that the total energy,
microphysical parameters and medium properties of the high-$z$
GRBs are in line with those derived for lower redshift bursts.
In the bottom-left panel of Fig.~1, the available host dust extinction
estimates for $z>6$ bursts are plotted against those obtained for the
BAT6 sample \citep{Covino2013}. Being selected only on the
basis of the GRB $\gamma-$ray peak flux, the BAT6 sample is representative of the
extinction distribution in GRB selected galaxies:
50\% of bursts suffer less than 0.3-0.4 mag extinction and
only 13\% of GRBs have $A_V>2$
mag. The high-$z$ sample is characterized by little or no
absorption\footnote{However, the lack of highly
extincted bursts may reflect a bias in the sample
as even a small amount of absorption at $z>6$ will prevent the
detection of the optical afterglow and thus of the redshift measure.},
suggesting a decrease in dust content in
star-forming environments at high redshifts \citep{Zafar2011b}. Similarly, the metallicities inferred for $z\sim 6$ bursts are in line,
though at the lower end, with the distribution of $Z$ measured in
lower redshift GRB afterglows \citep{Savaglio2009,Sparre2014}. The observed $Z$ and $A_V$ values (or limits) are indeed in agreement
with those expected for high-$z$ galaxies populating the
faint end of the luminosity function \citep[see e.g.][]{Salvaterra2011}.
At variance with lower redshift bursts, all high-$z$ GRBs have measured (or limit consistent with) very high
X-ray equivalent hydrogen column densities, $N_{\rm H,X}$ (bottom-right
panel of Fig.~1). At lower
redshift, a high $N_{\rm H,X}$ is usually connected to highly extincted afterglows \citep{Campana2012,Covino2013}.
However, at higher redshifts this interpretation seems to be unrealistic
given our knowledge of galaxy evolution at early times. Moreover, little or no dust absorption is
found in high-$z$ bursts, calling for unusual metal-to-dust ratios.
In recent years, a well defined trend of increasing $N_{\rm
H,X}$ with redshift has been found using both GRBs and AGNs
\citep{Campana2010,Behar2011,Starling2013,Campana2015}. High-$z$ GRBs extend this
relation to higher redshifts quite nicely with no $N_{\rm H,X}$ value
or limit below the observed lower envelope.
\citet{Campana2015} have shown, by means of dedicated numerical simulations, that this effect can be naturally
explained by the
absorption of intervening metals along the line-of-sight (LOS), being
the observed lower
envelope due to metals present in the IGM. An additional
contribution in most of the LOSs comes from metals in the diffuse
gas in galaxy groups at $z\sim 0-2$.
In conclusion, high-$z$ GRBs do not show any peculiar feature with
respect to low- and intermediate-$z$ events. Both the energetics and
the afterglow properties are found to be very similar. The medium
in which they blow up does not differ too much in terms of density, metallicity and
dust content. Finally, also the duration of their prompt emission is
consistent with those of the low-$z$ sample taking into account
cosmological and instrumental effects \citep{Littlejohns2013}. All these findings support the idea that the GRBs
detected so far at $z>6$ represent the high redshift tail of the PopII/I GRB redshift distribution.
\subsection{High-$z$ GRB host galaxies}\label{sec:obshost}
In the last few years, deep searches with both ground- and space-based
facilities \citep{Tanvir2012,Basa2012,Walter2012,Berger2014} have been
carried out searching for the host galaxy of
high-$z$ GRBs. Indeed, the a-priori, precise knowledge of the position and
distance of the galaxy provided by the GRB detection allows to
optimize the telescope setup, pushing the instruments to their
limit.
\citet{Tanvir2012} reported the observation of the field of four
GRBs at $z>6$ with HST with filters choosen to cover the rest-frame wavelength range
1200-1500\AA, i.e. above the Lyman-$\alpha$. In spite of the very deep
limits reached (see Table~1), none of the targets
have been identified. Similar results were obtained by \citet{Basa2012} using slightly shallower VLT observations.
More recently, \citet{Berger2014} imagined the field
of GRB~090423 with {\it Spitzer} at 3.6 $\mu$m and ALMA at 222 GHz,
but, also in this case, no detection is reported. These results imply that all the
hosts lie below the characteristic luminosity value at their respective redshifts, with star
formation rates SFR$<4\;{\rm M}_{\odot}$ yr$^{-1}$ in all cases. In particular, GRB~090423
was possible to derive a strong limit on the unobscured
SFR$<0.38\;{\rm M}_{\odot}$ yr$^{-1}$ \citep{Tanvir2012}, while the ALMA non
detection required the obscured SFR to be $<5\;{\rm M}_{\odot}$ yr$^{-1}$ \citep{Berger2014}.
\citet{Tanvir2012} also stacked all images deriving a limit on the mean
SFR per galaxy of $<0.17\;{\rm M}_{\odot}$ yr$^{-1}$, consistent with the idea
that the bulk of star formation activity is missed in current deep HST
surveys \citep{Salvaterra2011}. Moreover, these findings offer independent
evidence that the galaxy
luminosity function is evolving rapidly with redshift with a
steeper faint-end slope or with a decreasing characteristic
luminosity at high redshift \citep{Tanvir2012}.
\section{Theory}
\subsection{The high-$z$ GRB population}
As described in previous Section, {\it Swift} has detected almost one
GRB at $z>6$ per year. Considering the 1.4 sr field of view of the BAT
instrument, this corresponds to a rate\footnote{While this represents a lower limit, given the fact that
some high-$z$ burst can be among GRBs without redshift measurement, it
should be noted that there is a possible positive bias at work here. Indeed,
most of the observational programs at large ground- and
space-facilities have as their primary goal the detection of high-$z$
events.} of $\sim 0.6$ bursts yr$^{-1}$
sr$^{-1}$. It is worth to ask whether such a rate is expected on the
basis of our knowledge of the intrinsic GRB redshift distribution and luminosity
function \citep{Bromm2002,Daigne2006,SalvaterraChincarini2007,Salvaterra2009b,Butler2010,Wanderman2010,Robertson2012,Salvaterra2012,Ghirlanda2015}.
\begin{figure}\label{fig:ngt6}
\centering
{\includegraphics[scale=0.38]{review_salvaterra_fig2.ps}}
\caption{Peak flux distribution of bursts at $z>6$ as measured in the
15-150 keV {\it Swift}/BAT band. The solid (dotted) histogram shows the distribution of
bursts with spectroscopic (spectroscopic and photometric) redshift. The
shaded regions report model predictions \citep{Salvaterra2012}, taking into account the uncertainties in the
determination of the GRB luminosity function and evolution. Dark (light) color refers to the density
(luminosity) evolution models. The dashed line is obtained assuming
no evolution and clearly
underestimates the number of high-$z$ detections at all peak fluxes.}
\end{figure}
Figure~2 shows the distribution of peak
fluxes of bursts at $z>6$ as measured in the 15-150 keV {\it
Swift}/BAT band. The histogram is compared with the model results under
different assumptions on the evolution of the GRB luminosity function
\citep{Salvaterra2012}. The models have
been calibrated by jointly fitting the differential peak
flux distribution of BATSE long GRBs and the redshift distribution of
the BAT6 sample. The models reproduce the rate and peak flux
distribution of $z>6$ burst remarkably well\footnote{Since the BAT6 sample is limited to
$z<5.4$ there is no assumption in the model about the peak flux distribution of
$z>6$ GRBs}
with a small excess in the predicted population at the faintest
fluxes. However, since, in general, to faint
$\gamma$-ray fluxes correspond faint optical afterglows
\citep[e.g.][]{Melandri2014}, some high-$z$ faint bursts can be missed
being the detection of its afterglow difficult even with 8-m class telescopes.
By extrapolating model results to
fainter fluxes \citep{Salvaterra2008,Ghirlanda2015}, it can be shown that bursts $z>6$ represent $\sim 10$\% of the whole population,
suggesting that GRBs are quite efficient in selecting high-$z$ objects. The detection of
these events is one of the main goals of any future GRB
mission.
\subsection{GRBs from PopIII stars}\label{sec:popIII}
Different authors \citep{Meszaros2010,Komissarov2010,Suwa2011,Toma2011,Nagakura2012,Nakauchi2012,Piro2014} have proposed
that the conditions for jet breakout could be met also during the
final phases of the collapse of a PopIII massive star. Although under
different frameworks, all models consistently predict PopIII GRBs to be
very energetic events, with total energies
exceeding by orders of magnitude those expected in PopII events. In particular
\citet{Toma2011} suggested that PopIII bursts could
have an equivalent isotropic energy $\sim 10^{56-57}$ erg making their
detection possible even at the highest redshifts. A much longer
prompt emission, with typical duration of $10^4$ s, is also foreseen
in most of the models.
However, all these
characteristics are shared observationally with the population of ultra-long GRBs
recently detected at much lower redshift \citep{Levan2014} and likely
associated with PopII
blue supergiant stars \citep{Nakauchi2013}. Thus, they do not represent a unique feature to
firmly indentify a PopIII event. Strong evidence
could be provided by the absence of metal absorption lines in the
afterglow spectrum down to a level of the critical
metallicity. However, the measure of such low metallicity (or a limit
consistent with it) requires extremely high signal-to-noise spectra
that can be even beyond the capabilities of
30-m class telescopes. In the absence of any other observational tool
to identify PopIII events, radio observations can provide an important
clue. Indeed extremely powerful radio afterglows
\citep{Ciardi2000,Toma2011,Ghirlanda2013b} are expected given the peculiar energetics of PopIII
events. \citet{Ghirlanda2013b} showed that PopIII radio
afterglows should reside in a very distinct position in observed peak
time-peak flux plane, reaching much larger peak fluxes at later times
than any PopII GRBs.
\smallskip
We have seen in Section~\ref{sec:obsprop} that it is unlikely that the observed $z>6$
bursts are associated with PopIII progenitors. This piece of
information has been used to put a limit on the expected rate of
PopIII events \citep{Bromm2006,Campisi2011,Toma2011,deSouza2011,MaioBarkov2014,Mesler2014}. By means of dedicated numerical simulations including the
effect of the chemical feedback, \citet{Campisi2011} computed the
expected redshift distribution of PopIII GRBs. The PopIII GRB rate is
found to
increase from $z=20$ to $z=8-10$ and then it decreases rapidly becoming
negligible at $z\sim
5$. Under the assumption that all PopIII GRBs can be detected by {\it
Swift}\footnote{This may be not the case since the {\it Swift} trigger algorithm \citep{Lien2014} is not very sensitive to extremely long GRBs.}, the maximum allowed
rate is $\sim 0.03$ Gpc$^{-3}$ yr$^{-1}$. Assuming a typical jet angle
of $\sim 7$ degree and a typical PopIII mass of $\sim 100\;{\rm M}_{\odot}$,
this corresponds to $<1$ PopIII GRB every 500 PopIII stars, comparable
with the PopII/SN type Ib/c ratio \citep{Ghirlanda2013a}.
\medskip
In conclusion, even if GRBs can really result from the
collapse of a massive PopIII star, they will be extremely
rare with $<0.05$ event yr$^{-1}$ sr$^{-1}$ at $z>6$. PopIII GRBs
become the dominant observed population only at $z>10-15$
\citep{Campisi2011}. On the other hand, even a single detection of a
GRB powered by PopIII stars, i.e. a extremely high-$z$,
very energetic, long event with a bright radio afterglow peaking a
late times, would represent a breakthrough in the study of the first
generation of stars.
\begin{figure}
\centering
{\includegraphics[scale=0.35]{review_salvaterra_fig3a.ps}}
{\includegraphics[scale=0.35]{review_salvaterra_fig3b.ps}}
{\includegraphics[scale=0.35]{review_salvaterra_fig3c.ps}}
\caption{
Properties of simulated GRB hosts in the redshift range
$z=6-10$. Panels from top to bottom show the distribution of UV
absolute magnitude, of the SFR (in
$\rm {\rm M}_{\odot}~yr^{-1}$), and of gas phase metallicities (in solar units). The dotted line shows the normalized cumulative
distributions (see right y-axis). The arrows in the first two panels
report the limits on $M_{\rm UV}$ and SFR of GRB hosts \citep{Tanvir2012}.
The shaded areas in the bottom panel correspond to the metallicity measured
in the GRB~050904 \citep{Thoene2013} and GRB~130606A
\citep{Hartoog2014} afterglow spectrum. The
arrow shows the limit inferred for GRB~140515A \citep{Chornock2014}.
}
\end{figure}
\subsection{High-$z$ GRB host galaxies}
In absence
of any detection (see Section~\ref{sec:obshost}), the
relation of GRB selected galaxies with the typical high-$z$ galaxy
population has been studied by means of dedicated
semi-analytical models \citep{Trenti2012} or numerical
simulations \citep{Salvaterra2013,Elliott2015}.
\citet{Salvaterra2013} derived the expected properties
of the GRB hosts at $z=6-10$ by means of cosmological numerical simulations including all
relevant feedback effects at these redshifts. GRBs are found to
explode in bursty galaxies with typical star formation rates
SFR$\simeq 0.03-0.3$ M$_\odot$ yr$^{-1}$, stellar masses
$M_\star\simeq 10^6-10^8$ M$_\odot$, specific star formation rates
sSFR$\simeq 3-10$ Gyr$^{-1}$. The distribution of their UV
luminosities (see top panel of Fig.~3) places them in the faint end of the galaxy
luminosity function, below the capabilities of current
instruments. This is consistent with the lack of detection and suggests that deep observations with the
James Webb Space Telescope (JWST) are required to pinpoint them. Finally, the bulk of the GRBs
are found to explode in galaxies already enriched with metal to
$Z\simeq 0.03-0.1$ Z$_\odot$ (bottom panel of Fig.~3), consistently with the
metallicity inferred in $z\sim 6$ GRB afterglow spectra
(\citealt{Salvaterra2013}, but see \citealt{Cen2014}).
The same simulations allow to predict the properties of host galaxies
of PopIII GRB events \citep{Campisi2011}. These bursts are
found to reside typically in objects at the lower end of the stellar
mass with $M_\star< 10^7\;{\rm M}_{\odot}$ as they have the
higher probability of having still a pristine composition. While a
metallicity below $Z_{\rm crit}$ is obviously required in the star
forming cloud from which the PopIII GRB arises, some of the PopIII
GRB hosts are found in galaxies with $Z\sim 10^{-3}\;{\rm Z}_{\odot}$. This may
reflect the fact that some PopIII event can be associated with
pockets of metal-free gas in the outskirt of a enriched galaxy
\citep{Tornatore2007}.
\section{High-$z$ GRBs as a tool}
\subsection{Independent measure of the SFR}
It is well established that long GRBs are associated to the death of
massive stars \citep[see][and references
therein]{Hjorth2012}. Therefore, they should trace to some extent the star
formation activity through cosmic times and they can be use to measure
in an independent way the global SFR
\citep{Kistler2009,Ishida2011,Robertson2012}. In this respect, GRBs
have many advantages compared with usual probes: (i) they are detected at higher
redshifts; (ii) they are independent on the galaxy brightness; (iii)
they do not suffer of usual biases affecting optical/NIR surveys.
Indeed, although hampered by small statistics, the
detection of the few high-$z$ GRBs already suggests that the global
SFR at $z\simeq 8-9$ is 3-5
times higher than deduced from high-$z$ galaxy searches through
the drop-out technique \citep{Ishida2011}.
An important caveat to be mention here is that GRBs may be not perfect
tracers of the SFR. Indeed, in order to reproduce the redshift
distribution of bursts detected by {\it Swift}
\citep[e.g.][]{SalvaterraChincarini2007,Wanderman2010,Robertson2012,Salvaterra2012}, some kind of evolution with redshift is required. Thus, a precise
knowledge of the nature and the value of this bias
is needed to properly use GRBs as SFR probe in the early
Universe. This can be achieved by measuring the GRB LF and
its evolution at low and intermediate redshifts
\citep{Salvaterra2012,Wanderman2010,Pescalli2015} and by studying well selected sample of low$-z$ GRB host
galaxies
\citep{Boissier2013,Perley2013,Hunt2014,Trenti2014,Vergani2014}. In
particular, \citet{Vergani2014} have shown that the stellar mass
distribution of a small, but complete, sample of GRB selected galaxies at $z<1$ is consistent
with the existence of a mild metallicity bias in the GRB hosts, with
bursts forming more efficiently in environments with
$Z<0.3-0.5\;{\rm Z}_{\odot}$ \citep{Vergani2014}. If these findings are
confirmed by larger samples, GRBs can be considered good tracers of
the SFR at least at $z>3-4$.
\subsection{The reionization history}
Similar to quasars, GRBs have been used to measure the neutral hydrogen
fraction in the IGM by fitting the red damping wing of the Ly$\alpha$
and therefore to place strong constraints on the reionization redshift
\citep{Totani2006,Chornock2013,Totani2014,Hartoog2014,Chornock2014}.
With respect to quasars, GRBs have many advantages: (i) they are already
detected at much larger redshift than quasars; (ii) they reside in
average cosmic regions, less affected by local ionization effects or
clustering, and (iii) their spectrum follows a power-law making
continuum determination much easier. The main limitation of this
method is represented by the presence of the host galaxy absorption
\citep{McQuinn2008}, although this can be in principle separated being
less extended. \citet{Chen2007} found that $\sim 20-30$\% of GRBs have small enough hydrogen column
density to allow a direct measure of absorption from a partially
neutral IGM. Moreover, the typical $N_{\rm HI}$ value for GRB hosts is
expected to decrease with increasing redshift \citep{Nagamine2008}.
This idea is now supported by the fact that three out of the four $z>6$ GRBs for
which the intrinsic HI column density has been measured, showed
$N_{\rm HI}<10^{20}$ cm$^{-2}$ (see Table~1). In particular, a very
low $N_{\rm HI}$ is found in the case of GRB~140515A
\citep{Chornock2014} suggesting that a statistical sample of bursts
with low intrinsic
column densities can be gathered in future allowing to firmly
constrain the reionization process.
Another effective method to constrain the reionization history is
based on the statistics of peaks/gaps in the afterglow spectrum between Ly$\alpha$
and Ly$\beta$, i.e. corresponding to transmission/absorption regions along
the LOS \citep{Paschos2005}. \citet{Gallerani2008} have shown that
the distribution of the dimension of the largest dark gap is sensitive
to the assumed reionization history
provided that a statistical sample of $\sim 20$ LOSs is gathered.
Finally, the study of the HI forest over-imposed on the GRB radio
afterglow with SKA of a very high redshift burst could offer, in principle, a powerful tool to
study the reionization process \citep{Xu2011,Ciardi2015}. However, it requires
the detection of an extremely bright radio afterglow at the level of
$\sim 10$ mJy at $z>7-8$, that may be possible in the case of PopIII GRBs.
Beside constraining the reionization history and its variance along
different LOSs, GRBs can provide an estimate of the typical escape
fraction of UV photons from early galaxies. The measure of this
crucial quantity from the observation of the ecaping
Lyman continuum radiation is already very difficult at $z\sim 3$
\citep{Vanzella2012}, and may be
impossible for the small, faint galaxies responsible for the bulk of star
formation at $z>6$. A statistical sample of GRBs provide an
alternative way to infer it from
the distribution of $N_{\rm HI}$ over many LOSs. Useful
constrains have so far obtained at $z=2-4$ with $f_{\rm esc}<7.5$\% \citep{Fynbo2009}, being higher redshift studies prevented by the small
statistic of the sample. The measure of the escape fraction from
typically star-forming galaxies during the
epoch of reionization will provide the missing piece of the puzzle in
our understanding of how galaxies reionize the IGM.
\subsection{Studying high-$z$ galaxies}
Theoretical models and observational evidence suggests that
most of the UV photons that reionize the Universe are emitted by low
mass, faint galaxies missed even in the deepest observations by
HST. The detection of these objects is one of the main scientific
goal of the next generation of space telescopes. However,
little or no information about the physical properties (e.g. metal and
dust content) of these objects will be accessible even with JWST.
High-$z$ GRBs can provide
a useful, complementary tool to investigate the building up of metals
at these early stages of galaxy formation and, in particular, in those
objects that provide the bulk of the ionizing photons
\citep{Salvaterra2013}. Indeed, GRB events are produce by the same
massive stars that emit the
FUV photons able to ionize hydrogen. Metal absorption line observations of
a large sample of good signal-to-noise GRB spectra will allow to
recover the cosmic metal enrichment history, to extend the
mass-metallicity and FMR to higher redshift \citep{Salvaterra2013,Laskar2011}, to
perform stellar population studies \citep{Grieco2014,Ma2015}, and to search for peculiar
nucleo-synthesis pattern (see next Section).
Metal absorption lines over-imposed to the
afterglow radiation have been already
detected even $z>6$ showing that high-$z$ galaxies are already
enriched to a few percent solar. In particular, the spectrum of
GRB~130606A obtained with VLT/X-shooter \citep{Hartoog2014} shows a great
variety of metal elements allowing accurate abundance and gas
kinematic studies of the inner region of a $z\sim 6$ galaxy. Even better
and richer datasets will be gathered when the 30-m
class telescope like E-ELT will become operative. These studies can be complemented by X-ray spectroscopic
observations of bright GRBs with the next generation of X-ray
instruments \citep{Campana2011}. In particular, with the Athena
satellite \citep{Nandra2013} it will become possible to directly
measure the abundance patterns in X-ray afterglows allowing, in principle, to discriminate between different nucleo-synthesis
sources \citep{Jonker2013}.
Beside metal absorption lines, the observation of H$_2$
molecular absorption and of the local dust law can provide
further details about the host enrichment. As discussed in
Section~\ref{sec:obsprop}, evidence of dust extinction in four $z>6$ GRBs
has been derived by studying the X-ray-to-optical SED. Furthermore, in
the case of GRB~140515A dust deplition is suggested by the peculiar
metal abundance ratio \citep{Hartoog2014}. These observations show
that GRBs can provide fundamental clues about the presence of dust in
high-$z$ galaxies. Moreover, they can be use to constrain the dust
formation channels at work at those early epoch. It is interesting to
note that for GRB~071025 at $z\sim 5$ \citep{Perley2010} and tentatively
for GRB~050904 at $z=6.3$ \citep{Stratta2011} evidences for SN
synthesized dust has been reported.
Tracking the dust enrichment history of first galaxies is also very
important for better understand the PopIII/PopII transition. Indeed,
it has been shown that even a little amount of dust in
the medium will induce the formation of low mass stars \citep{Schneider2003,Schneider2012}.
\subsection{Indirect search for PopIII stars}
PopIII GRBs (if any) are expected to be extremely rare and difficult
to distinguish from other GRB populations (see Section~\ref{sec:popIII}).
An indirect, but fruitful, way to search for the elusive PopIII stars
is represented by the study of the metal and dust composition of the
ISM of distant galaxies enlightened by PopII GRBs.
Indeed, peculiar metal abundance ratios \citep{HegerWoosley2002} and dust composition \citep{Schneider2004} are foreseen as the result of the enrichment of
first massive SN explosions. The detection of such a signature
in the optical-NIR afterglow of
a very high-$z$ GRBs will provide a strong evidence for the existence
of very massive ($>100\;{\rm M}_{\odot}$) PopIII stars.
\citet{Wang2012} found that PopIII enriched gas will result in much larger
equivalent widths of metal absorption lines than typical PopII
enrichment. Moreover, \citet{Ma2015} showed that in principle the [C/O] and
[Si/O] alone could be
enough to distinguish a PopIII enriched envirorment,
although, in practice, the detection of more elements
(such as S and Fe) is needed. PopII GRBs exploding in a PopIII enrichment medium are
expected to be rare at $z=6$ ($<1$\% of the GRB population at
that redshift) but the probability to find them increases
with $z$ being $\sim 10$\% at $z=10$ \citep{Ma2015}. They
should reside in small host galaxies with $M_\star\sim
10^{5-6}\;{\rm M}_{\odot}$ and $Z<10^{-2.8}\;{\rm Z}_{\odot}$. Thus, the
detection of a $z>10$ burst with a metallicity below this threshold
will be a strong candidate to be either a PopIII event or a PopII GRBs
blowing up in a PopIII enriched medium.
\subsection{Others}
For the sake of completeness \citep[see][for a more complete list]{McQuinn2009,Amati2013} I would like to briefly mention here some
other possible uses of GRBs in the study of early Universe:
{\it - constrain the dark matter particle} \citep{Mesinger2005,deSouza2013}. The redshift distribution of GRBs can set limits
on the dark matter particle mass, m$_{\rm x}$. Using a sub-sample of
$z>4$ GRBs m$_{\rm x}>1.6-1.8$ keV at
95\% confidence level is found.
{\it - non-Gaussianity} \citep{Maio2012}. Deviations from Gaussianity of
the primordial density field will translate in different rates of GRBs
$z\gg 6$. A single GRB
detection at $z>15$ will favor
non-Gaussian scenarios with a positive non-linear parameter.
{\it - radiation field} \citep{Inoue2010,Kakuwa2012}. The ultraviolet
intergalactic radiation field below the Lyman edge energy can cause
attenuation in the $\gamma$-ray spectra of GRBs. This may
be observable at high-$z$ by the Cherenkov Telescope Array (CTA).
{\it - magnetic fields} \citep{Takahashi2011}. Pair echoes from
luminous high-$z$ GRBs may be detectable with CTA providing a unique
probe of weak intergalactic magnetic fields at early stages of
structure formation.
\section{Present and future of the fields}\label{sec:future}
The detection of high-$z$ GRBs is one of the main goals driving the
design of the next GRB missions \citep{Amati2013}. As already discussed, GRBs are
very effective in selecting high-$z$ objects and sufficiently bright to be
detectable up to beginning of the star formation activity. While the forthcoming SVOM
satellite \citep{Godet2012} or the proposed LOFT mission \citep{Amati2015} will
surely increase the $z>6$ GRB sample, it is also clear that
in order to full exploit the potentiality of GRBs as a probe of the early Univese, a much larger sample of
well studied, high-$z$ GRBs should be collected.
\begin{figure}\label{fig:flim}
\centering
{\includegraphics[scale=0.35]{review_salvaterra_fig4.ps}}
\caption{Required sensitivity, in terms of minimum peak flux
$P_{\Delta E}$ that
can be detected in the given energy band $\Delta E$, and field-of-view to
detect 10 GRB yr$^{-1}$ at $z>8$. Different lines correspond to
different energy bands as labeled in the plot, i.e. to different
mission concept. See \citet{Ghirlanda2015} for the details of
the calculation.}
\end{figure}
Due to redshift scaling, in the observer frame the prompt emission of
high-$z$ GRBs is expected to peak at softer energies than lower
redshift bursts. However, as the sensitivity of current GRB satellites
allow us to detect only the bright-end of the GRB LF at
high-$z$, the mean observed peak energy is $<E_{\rm p,
obs}>\sim 150$ keV similar to low-$z$ bursts. Therefore, while a facility operating in the soft X-rays
will be, in general, more efficient in selecting high-$z$ GRBs, a much
better sensitivity than current instruments should also be foreseen.
The scientific goals outlined in the previous section call for a large
statistical sample of well studied high-$z$ GRBs. To be competitive
with other probes, at
least a hundred of GRBs at $z>6$ with several tens of them lying at
$z>8$ should be detected. \citet{Ghirlanda2015} computed the expected detection rate of
high-$z$ GRBs by a generic detector with defined energy band and
sensitivity by means of a observational tested population
synthesis model of long GRBs. Following the results of this work, Figure~4
shows the sensitivity and
field-of-view of a mission able to detect 10 GRBs yr$^{-1}$ at
$z>8$ and operating in different energy bands. For instance, by
adopting the {\it Swift} sensitivity of $\sim 0.4$ ph s$^{-1}$
cm$^{-2}$ and energy band, we can see that the goal is never reached. On the other hand, limiting ourselves to the {\it Swift} FOV
of 1.4 sr, a hundred times better sensitivity than {\it Swift} is
needed. A X-ray instrument
will instead require a sensitivity of 0.1 ph s$^{-1}$ cm$^{-2}$
in the 2-5 keV band for the same FOV. It is worth to note that the same mission will also detect
thousands of bursts at lower redshifts (plus other
transients). In order to effectively select high-$z$
candidates, the presence of
a 0.5-1 m infrared telescope on-board should be foreseen (together
with fast repointing capabilities). This instrument will allow not
only to promptly measure the
redshift but also to perform low-resolution (R$\sim 1000$) spectroscopic studies
(e.g. to identify metal absorption lines) when the afterglow is at the
maximum of its brightness. A rapid dissemination of best high-$z$
candidates is also fundamental in order to trigger follow-up
observations by future facilities operating at different wavelengths:
optical-NIR (E-ELT, JWST if still operating), radio (SKA), X-ray
(ATHENA), TeV (CTA).
In conclusion, the study of the high-$z$ Universe with GRBs requires a
X-ray detector with unprecedented combination of
sensitivy and FOV coupled with an infrared telescope to select
reliable high-$z$ candidates. The THESEUS mission
recently proposed for the M4 ESA call matches all these
characteristics.
\section{Conclusions}
In this paper, I have briefly reviewed the status and the prospect for
the exploration of the high redshift Universe with GRBs.
A few solid considerations can be drawn:
(i) GRBs do exist at very high redshift and can be detected
and studied with present-day facilities;
(ii) high-$z$ GRBs are very similar to low- and intermediate-$z$ ones;
(iii) GRBs are an efficient way to select high-$z$ objects;
(iv) high-$z$ GRBs are hosted in faint galaxies that
are missed even in the deepest surveys;
(v) high-$z$ GRBs have proved to be an independent and powerful tool
to study the early Universe.
Although the current sample is limited to a few bursts,
future dedicated missions can provide a sufficiently large number of
GRBs at $z>6$ to study the first bound structures in a complementary
way with respect to galaxy and quasar surveys.
\section*{Acknowledgments}
This research was supported by the ASI-INAF contract (1/005/11/1). The author would like to thank A.~Melandri for providing the X-ray light
curve for the $z>6$ sample, S.~Campana for providing the measure of
the N$_{\rm H,X}$ for some of the high-$z$ bursts, G.~Ghirlanda for
provinding the $E_{\rm iso}$ values and the expected rate of GRBs at
$z>8$.
\section*{References}
|
\section{Introduction}
Spiral structure is a commonplace and visually striking feature of many galaxies
and yet there
is still disagreement as to the correct theory that explains its origin after decades of debate. The
first well known theory \citep{Lin:Shu:1964} is that density waves propagating
through the disk of the galaxy are the responsible
agent. The density wave theory for spiral modes, described in detail by \citet{Bertin:Lin:1996}, calls for a long-lived, quasi-steady global spiral pattern. Others call for more transient spiral patterns, whether from swing-amplified noise \citep{Goldreich:Lynden-Bell:1965,Julian:Toomre:1966}, recurrent cycles of groove modes \citep{Sellwood:2000}, or superposed transient instabilities \citep{Sellwood:Carlberg:2014}. Other theories have also been proposed, with
one in particular, the manifold theory \citep{Athanassoula:2009a,Athanassoula:2009b,Athanassoula:2010}, rejecting the density wave concept
altogether in favor of an explanation involving stars in chaotic highly
eccentric orbits.
The density wave theory, as originally articulated by \citet{Lin:Shu:1966}, had a
very specific prediction for the pitch angle of the spiral pattern produced
by the waves.
They calculated the pitch angle to be a ratio of the density of material
in the galaxy's disk to a certain quantity made up
of the frequencies of orbital motions in the disks,
\begin{equation}
\tan|P| = \frac
{2\pi m G(\sigma_o+F \sigma_*)}
{R(\kappa^2-(\omega-m\Omega)^2)},
\end{equation}
where $P$ is the logarithmic spiral arm pitch angle, $m$ is the number
of spiral arms in the pattern, $G$ is the gravitational
constant, $\Omega$ is the angular frequency of orbits of particles in the
disk, $R$ is the galactocentric radius, $\kappa$ is the epicyclic frequency of the same particles, and $\omega$
is the frequency associated with the perturbation that excites the density
waves. Note that as long as this perturbation is some form of self-excitation
arising from within the disk itself, it follows that all of the terms in
the denominator should depend on
the mass of the central bulge of the galaxy. The simplest case of a dominant bulge (approximated as a point mass) would imply, for instance,
$\Omega \propto \sqrt{M}$ and $\kappa \propto \sqrt{M}$, with $M$ being the central mass. The numerator depends on the density of gas in the disk, $\sigma_o$, and the stellar disk density, $\sigma_*$, with a factor
$F$, called the reduction factor, which underweights the stellar density (compared
to the gas density), since it is primarily within the gas that the
density wave propagates. In this Letter we will present some evidence that $F << 1$.
Focusing on the masses and densities involved in this relation, we find
that
\begin{equation}
\tan|P| \propto \frac{\sigma_o+F \sigma_*}{M_o},
\label{DWT}
\end{equation}
where $M_o$ is the mass of the galactic bulge, or else the total mass
interior to the radius in question.
This formula is known to work very well in the case of spiral density
waves in Saturn's rings \citep{Shu:1984}.
Bulge-dominated galaxies are not too distant from the Saturnian
situation of a
small dense core with negligible mass in the disk, though disk-dominated
galaxies are obviously far more complex.
Generally speaking, the density wave theory predicts
that the pitch angle of the spiral arms in galaxies does
depend on the radial distribution of matter in the galaxy,
and experimental studies concur \citep{Seigar:2006,Seigar:2014}. Results of
this type broadly agree with \citet{Lin:Shu:1966} that a thin (dense)
disk and massive (small) central bulge
should result in a tight (loose) spiral. This is not surprising, since we
would expect a standing
wave pattern (such as in a vibrating string) to depend on the ratio of a
restoring force or tension (in this case
the central mass, or at least the mass inside a given radius $R$)
to the density of the medium (in this case the density of
gas in the disk at radius $R$).
Although the precise nature of the relation between these
three quantities can be expected to vary between galaxies of different
types (bulge-dominated versus disk-dominated, for instance), nevertheless
we show in this Letter that the three quantities, spiral arm pitch angle,
central bulge mass, and gas density in the disk, do strongly correlate to
form a fundamental plane that may play a similar role in tying together
gross features of disk galaxies to that played by the fundamental plane
of elliptical galaxies \citep{Djorgovski:Davis:1987,Dressler:1987}.
We take as our sample disk galaxies from the DiskMass Survey \citep[DMS;][]{DMSI}, which
is ideal for our purposes since it deals with the disk densities of
a sample of face-on galaxies and includes measurements of the central bulge mass.
Using the technique of \citet{Me:2012}, we measure the
pitch angle for these galaxies and find that our sample of
24 galaxies, when plotted in a volume defined by these three
quantities, delineates a plane with very low
scatter. There is only a 0.0047\% chance that this plane could
have been formed by statistical accident.
The plane satisfies a number of requirements, which one would expect of a
useful fundamental plane. The plane is steeply
inclined across the volume formed by the three related quantities. In other
words, it is not merely a relation between two of the three quantities,
with the third essentially irrelevant. The galaxies are distributed quite
widely and fairly uniformly across the plane. There is no particular
evidence of a favored curvilinear relation on the face of the plane.
Finally, and most importantly,
the plane is oriented as one would expect on the basis of the density wave
theory. A large bulge mass and a rarified disk produces the tightest
spirals. A small bulge and a dense disk produce the loosest spirals.
We submit that any successful theory of galactic spiral structure must
be able to explain this result.
One final point is worthy of note. The DMS measured not only the
density of atomic hydrogen in the disk of each galaxy (the quantity
used in our relation), but also the density of molecular hydrogen and
the dynamical disk mass density (the total density in the disk). Our
results suggest that it is the gas density, not the total density in
the disk that matters for spiral density waves. This suggests an apparent decoupling between the stars and gas. The fact that the
density of atomic hydrogen fits noticeably better than that for molecular
hydrogen may simply be due to the fact
that it is a much more reliable measurement, since molecular hydrogen
is estimated indirectly from observations of other molecules, not
hydrogen itself \citep{Westfall:2011,DMSVII}.
\section{Data and Analysis}
The DMS PPak Sample \citep{DMSVI} consists of 30 nearly face-on galaxies whose disk densities have previously
been closely studied. However, four of these galaxies do not have
central stellar bulge masses available and three provide them only as upper limits, so they are excluded from our sample. In addition to the 23 remaining DMS galaxies, we also include our own Galaxy, the Milky Way, in our sample of galaxies.
In addition to the method described in \citet{Me:2012}, which utilizes a two-dimensional fast Fourier transform software called \textit{2DFFT}, we also measured pitch angles for all of the sample galaxies using new software called \textit{Spirality}. \textit{Spirality} (Shields et al. 2015, in preparation) is a novel method for measuring spiral arm pitch angle by fitting galaxy images to spiral coordinate systems (templates) of known pitch. For a given pitch angle template, the mean pixel value is found along each of typically 1000 spiral axes. The fitting function, which shows a local maximum at the best-fit pitch angle, is the variance of these means. In other words, we choose the pitch angle that exhibits the greatest contrast between the mean luminosity along the spiral axes. The presumption is that where the pitch angle of the spiral axes is equal to the pitch angle of the galaxy's spiral arms there will be some axes that fall precisely along the true spiral arms (and thus are much brighter in the mean) and some that never coincide with the true spiral arms (and thus are, on average, dim). Where the pitch of the axes is not equal to the pitch of the spiral arms, each axis will cross the true spiral arms a roughly equal number of times, making the mean brightness along each axis roughly equal. Error bars are found by varying the inner radius of the measurement annulus and finding the standard deviation of the best-fit pitch angles.
The two techniques yield measurements that agree within
the error bars in almost all cases. As a final and important test, we
visually inspected each galaxy, comparing them to overlays
of synthetic spirals on transparency paper, in order to confirm
the measured pitch angle. Our overlay transparencies showed
spirals of different sizes and different pitch angles in steps of
$5^{\circ}$ from $5^{\circ}$ to $85^{\circ}$. We were therefore
able to visually confirm the pitch angle to within $5^{\circ}$.
We were satisfied in all cases that the measured pitch angle of
\textit{2DFFT} was reliable and strongly supported by the combination
of \textit{Spirality} and
visual inspection.
For the sake of consistency, we chose to use
only the results of \textit{2DFFT} in this Letter. The pitch angles, $P$,
given in Table \ref{Sample} are the results of the \textit{2DFFT} routine.
The images used were obtained from the NASA/IPAC Extragalactic Database,\footnote{\url{http://ned.ipac.caltech.edu/}} and/or from the pODI (partial One Degree Imager) camera on the WIYN 3.5 $m$ telescope. The WIYN images were all acquired as 120 $s$ exposures, calibrated using QuickReduce1.0 from the ODI Pipeline, Portal, and Archive,\footnote{\url{http://portal.odi.iu.edu}} and processed using a five-point dither pattern for each galaxy and subsequently stacking the images using \textit{SWarp} \citep{Bertin:2002}. Additionally, KPNO 2.1 $m$ imaging for UGC 463, 1529, 1908, 4036, and 11318 were measured to confirm previous pitch angle measurements. Unless otherwise specified (Milky Way data have been determined in very different ways than other galaxies), all data for stellar galactic bulge
masses $(M_{\star}^{bulge})$ and maximum neutral atomic hydrogen
($\rm{H_{I}}$) gas mass surface densities ($\Sigma_{H_{I}}^{max}$) come
from \citet{Martinsson:2011} and \citet{DMSVII}; see Table \ref{Sample}. For the determination of the stellar bulge masses, the $K$-band light profile was decomposed into a central S\'ersic component (convolved with a seeing disk) and a number of exponential disks \citep{DMSVI}. The bulge masses were determined using the integral of the light from the central S\'ersic component and the mass-to-light ratio ($M/L$) derived from the disk using vertical velocity dispersions \citep{DMSVII}. Gas densities were determined from 21 cm line measurements \citep{Martinsson:2011}.
\begin{deluxetable*}{llccclllc}
\tablecolumns{9}
\tablecaption{Sample\label{Sample}}
\tablehead{
\colhead{Galaxy Name} & \colhead{Type} & \colhead{Band} & \colhead{Image Source} & \colhead{$m$} & \colhead{$\tan|P|$} & \colhead{$\log(M_{\star}^{bulge}/M_{\sun})$} & \colhead{$\Sigma_{H_{I}}^{max}/(M_{\sun}pc^{-2})$} & \colhead{Excluded} \\
\colhead{(1)} & \colhead{(2)} & \colhead{(3)} & \colhead{(4)} & \colhead{(5)} & \colhead{(6)} & \colhead{(7)} & \colhead{(8)} & \colhead{(9)}
}
\startdata
Milky Way & SBc & 21 cm & 1 & 4 & $0.414\pm0.051$\tablenotemark{a} & $9.95\pm0.03$\tablenotemark{b} & $4.98\pm0.53$\tablenotemark{c,d} & \\
UGC 448 & SABc & $r$ & 2 & 4 & $0.327\pm0.033$ & $9.76_{-0.51}^{+0.23}$ & $4.58\pm0.46$ & \\
UGC 463 & SABc & $B$ & 3 & 3 & $0.412\pm0.066$ & $9.35_{-0.41}^{+0.21}$ & $6.18\pm0.66$ & \\
UGC 1081 & SBc & $r$ & 2 & 2 & $0.452\pm0.064$ & $8.81_{-0.24}^{+0.16}$ & $6.25\pm0.62$ & \\
UGC 1087 & Sc & $r$ & 2 & 2 & $0.188\pm0.039$ & $8.64_{-0.25}^{+0.16}$ & $4.42\pm0.46$ & \\
UGC 1529 & Sc & 645.0 nm\tablenotemark{e} & 4 & 3 & $0.490\pm0.096$ & $8.98_{-0.39}^{+0.20}$ & $6.59\pm0.66$ & \\
UGC 1635 & Sbc & $r$ & 2 & 3 & $0.209\pm0.014$ & $8.74_{-0.29}^{+0.17}$ & $2.60\pm0.32$ & \\
UGC 1862 & SABcd$^1$ & $r$ & 2 & 2 & $0.444\pm0.074$ & \nodata & $9.14\pm0.91$ & \checkmark \\
UGC 1908 & SBc$^2$ & 645.0 nm\tablenotemark{e} & 4 & 3 & $0.376\pm0.069$ & $9.68_{-0.54}^{+0.23}$ & $4.62\pm0.46$ & \\
UGC 3091 & SABd & $i$ & 2 & 2 & $0.555\pm0.092$ & \nodata & $5.59\pm0.56$ & \checkmark \\
UGC 3140 & Sc & $r$ & 2 & 3 & $0.290\pm0.090$ & $9.65_{-0.24}^{+0.15}$ & $4.87\pm0.54$ & \\
UGC 3701 & Scd & $r$ & 2 & 2 & $0.276\pm0.090$ & $8.69_{-0.31}^{+0.18}$ & $5.55\pm0.57$ & \\
UGC 3997 & Im & $g$ & 5 & 2 & $0.185\pm0.048$ & $8.53_{-0.27}^{+0.17}$ & $5.01\pm0.54$ & \\
UGC 4036 & SABbc & 645.0 nm\tablenotemark{e} & 4 & 2 & $0.268\pm0.021$ & $8.92_{-0.23}^{+0.15}$ & $5.20\pm0.56$ & \\
UGC 4107 & Sc & $g$ & 5 & 2 & $0.371\pm0.041$ & $8.65_{-0.31}^{+0.18}$ & $5.42\pm0.54$ & \\
UGC 4256 & SABc & $g$ & 5 & 2 & $0.555\pm0.099$ & $9.29_{-9.29}^{+0.36}$ & $9.75\pm0.98$ & \checkmark \\
UGC 4368 & Scd & $g$ & 5 & 2 & $0.439\pm0.043$ & $9.21_{-0.41}^{+0.21}$ & $5.95\pm0.66$ & \\
UGC 4380 & Scd & $g$ & 5 & 3 & $0.430\pm0.095$ & $8.86_{-0.20}^{+0.13}$ & $4.08\pm0.41$ & \\
UGC 4458 & Sa & $g$ & 5 & 1 & $0.243\pm0.056$ & $10.67_{-0.39}^{+0.20}$ & $3.28\pm0.53$ & \\
UGC 4555 & SABbc & $g$ & 5 & 2 & $0.213\pm0.017$ & $8.96_{-0.39}^{+0.20}$ & $4.58\pm0.47$ & \\
UGC 4622 & Scd & $g$ & 5 & 4 & $0.401\pm0.099$ & $9.89_{-0.41}^{+0.21}$ & $3.50\pm0.38$ & \\
UGC 6903 & SBcd & $g$ & 5 & 2 & $0.283\pm0.041$ & $8.03_{-0.62}^{+0.25}$ & $4.94\pm0.59$ & \\
UGC 6918 & SABb$^3$ & F606W & 6 & 3 & $0.306\pm0.044$ & $8.04_{-8.04}^{+0.70}$ & $7.04\pm0.72$ & \checkmark \\
UGC 7244 & SBcd & $g$ & 5 & 2 & $0.627\pm0.105$ & \nodata & $5.53\pm0.60$ & \checkmark \\
UGC 7917 & SBbc & $g$ & 5 & 3 & $0.278\pm0.025$ & $10.01_{-10.01}^{+00.34}$ & $2.70\pm0.28$ & \checkmark \\
UGC 8196 & Sb & $g$ & 5 & 5 & $0.144\pm0.009$ & $10.73_{-0.26}^{+0.16}$ & $2.74\pm0.28$ & \\
UGC 9177 & Scd & $g$ & 5 & 2 & $0.256\pm0.035$ & $9.55_{-0.58}^{+0.24}$ & $3.92\pm0.42$ & \\
UGC 9837 & SABc & $g$ & 5 & 6 & $0.482\pm0.061$ & $8.35_{-0.29}^{+0.17}$ & $7.95\pm0.80$ & \\
UGC 9965 & Sc & $g$ & 5 & 3 & $0.237\pm0.037$ & \nodata & $5.63\pm0.58$ & \checkmark \\
UGC 11318 & SBbc & 645.0 nm\tablenotemark{e} & 4 & 3 & $0.569\pm0.101$ & $9.69_{-0.50}^{+0.23}$ & $6.51\pm0.67$ & \\
UGC 12391 & SABc & $r$ & 2 & 4 & $0.235\pm0.091$ & $8.98_{-0.28}^{+0.17}$ & $4.90\pm0.49$ & \\
\enddata
\tablecomments{\emph{Columns:}
(1) Galaxy name.
(2) Hubble morphological type from either the UGC \citep{Nilson:1973}
or RC3 \citep{RC3}
catalogs. Notes on morphologies: 1 = peculiar, 2 = starburst, and 3 = AGN.
(3) Filter waveband/wavelength used for pitch angle calculation.
(4) Telescope/literature source of imaging used for pitch angle calculation.
(5) Harmonic mode (number of spiral arms).
(6) Tangent of the pitch angle of the galactic logarithmic spiral arms.
(7) Base 10 logarithm of the stellar bulge mass of the galaxy, in solar masses.
(8) Maximum surface density in the galactic $\rm{H_{I}}$ gas, in solar masses per square pc.
(9) Indication of galaxies that are excluded in fittings due to missing measurements or measurements that are merely upper limits.
\emph{Image Sources:}
(1) \citet{Levine:2006};
(2) WIYN 3.5 $m$ pODI;
(3) JKT 1.0 $m$;
(4) Palomar 48 inch Schmidt;
(5) SDSS;
(6) HST.
}
\tablenotetext{a}{\citet{Levine:2006}.}
\tablenotetext{b}{\citet{McMillan:2011}.}
\tablenotetext{c}{No error estimates were provided by its reference so we have assigned the mean error of the included sample, $\pm0.53$ $M_{\sun}pc^{-2}$.}
\tablenotetext{d}{Calculated using Equation (2) from \citet{Ferriere:2001}.}
\tablenotetext{e}{IIIaJ emulsion.}
\end{deluxetable*}
\vspace{9mm}
\section{Results}
We find a best linear fit for Equation (\ref{DWT}) from the included data sample of 24 galaxies of
\begin{equation}
\tan|P| = (0.375\pm0.092)\frac{\Sigma_{H_{I}}^{max}/(M_{\sun}pc^{-2})}{\log\left(M_{\star}^{bulge}/M_{\sun}\right)} + (0.127\pm0.049).
\label{propto}
\end{equation}
The root mean squared error ($E_{rms}$) is equal to 0.0909 (a residual scatter of $31.2\%$ per galaxy on average), with $R^2 = 0.344$, and a $p$-value equal to $2.59$ $\times$ $10^{-3}$ for Equation (\ref{propto}). A plot of this linear fit, along with the included data sample, is given in Figure \ref{2DPlot}.
\begin{figure}
\includegraphics[width=8.6cm]{f1.eps}
\caption{Two-dimensional plot of the linear fit defined by the multivariate normally distributed sampling fit of Equation (\ref{propto}) depicted by the solid black line, along with the plotted points of the 24 galaxy member data set. The Milky Way is depicted distinctly in \textcolor{green}{green}. The axes $[x, y]$ depict [$\tan|P|$, $(\Sigma_{H_{I}}^{max}/(M_{\sun}pc^{-2}))/\log(M_{\star}^{bulge}/M_{\sun})$], respectively.
\label{2DPlot}}
\end{figure}
The formula describing the fundamental plane for spiral galaxies from the sample is as follows:
\begin{eqnarray}
\frac{\Sigma_{H_{I}}^{max}}{M_{\sun}pc^{-2}} = (5.70\pm1.40)\tan|P| - \nonumber \\
(0.677\pm0.199)\log\left(M_{\star}^{bulge}/M_{\sun}\right) + (9.29\pm1.96).
\label{Plane}
\end{eqnarray}
$E_{rms} = 0.770$ $M_{\sun}pc^{-2}$ (a residual scatter of $16.7\%$ per galaxy on average\footnote{Compare this to the fundamental plane for elliptical galaxies, which has a residual scatter of $\sim20\%$ per galaxy on average \citep{Kormendy:Djorgovski:1989}.}), with $R^2 = 0.613$, and a $p$-value $= 4.71$ $\times$ $10^{-5}$ for Equation (\ref{Plane}). It is interesting to note that the addition of the extra dimension cuts the residual scatter approximately in half. A three-dimensional plot of this plane, along with the included data sample, is given in Figure \ref{3DPlot}\footnote{A 3D animated gif of this figure can be accessed at \url{http://dafix.uark.edu/~ben/movie.gif}.}.
\begin{figure*}
\includegraphics[width=5.97cm]{f2a.eps}
\includegraphics[width=5.97cm]{f2b.eps}
\includegraphics[width=5.97cm]{f2c.eps}
\caption{Three-dimensional plot of the plane defined by the fit of Equation (\ref{Plane}) with the multivariate normally distributed sampling depicted by a translucent \textcolor{blue}{blue} meshed surface, along with the plotted points of the 24 galaxy member included data set (depicted by \textcolor{red}{red} spheres with the Milky Way in \textcolor{green}{green}). Note that the points will appear slightly darker when they are projected behind and partially obscured by the plane. The axes $[x, y, z]$ depict [$\tan|P|$, $\log(M_{\star}^{bulge}/M_{\sun})$, $\Sigma_{H_{I}}^{max}$], respectively. Left: the view has been oriented parallel to the plane. Middle: the view has been oriented at an orientation sufficient to view the face of the plane. Right: the view has been projected along an orthogonal vector above the plane.\label{3DPlot}}
\end{figure*}
The errors presented in Equations (\ref{propto}) and (\ref{Plane}) have been determined by sampling individual data points from multivariate normal distributions using the following algorithm.
\begin{enumerate}
\item For each measurement, draw a new measurement based on multivariate normal distributions with the mean and variance of each variable for all 24 galaxies.
\item Fit linear (or planar) best-fit coefficients to the 24 points.
\item Repeat steps 1 and 2 $10^6$ times, saving the fitted coefficients after each step.
\item Use the distribution of the coefficients based on these $10^6$ fits to define the best-fitting (median) coefficients and their error (the $1\sigma$ confidence interval of the distribution).
\end{enumerate}
The orientation of the fundamental plane illustrated in Figure \ref{3DPlot} is exactly as one would expect on the basis of the spiral density wave theory. According to Equation (\ref{DWT}), the pitch angle is minimized (tightest winding) when the $\rm{H_{I}}$ mass surface density is low and the central mass is high.
Alternatively, the pitch angle is maximized (loosest winding) when the $\rm{H_{I}}$ mass surface density is high and the central mass is low. This behavior is illustrated in both figures. Note in the middle panel of Figure \ref{3DPlot} how the plane slopes from the lower left front (low pitch angle, low $\rm{H_{I}}$ mass surface density, and high central mass) corner of the cube to the upper right back (high pitch angle, high $\rm{H_{I}}$ mass surface density, and low central mass) corner of the cube.
Furthermore, this indicates that the shape of the plane is strongly correlated to all three variables (the individual variable $p$-values of the intercept, $\tan|P|$, and $\log(M_{\star}^{bulge}/M_{\sun})$ are $9.97$ $\times$ $10^{-5}$, $8.03$ $\times$ $10^{-5}$, and $9.41$ $\times$ $10^{-4}$, respectively for Equation (\ref{Plane})).
\section{Discussion}
If one favors the standing wave picture of spiral structure, our result
is not unexpected.
In analogy with standing waves on a string, we would expect the
wavelength of the resulting pattern to be strongly determined by the
tension or restoring force (in this case the central gravitating mass)
and the density of the medium (in this case the gas in the disk).
It is worth noting that our case is probably analogous to a string with
non-uniform density, since the gas density generally falls off with increasing radius in a
galactic disk\footnote{This is true for the total gas density; however, the
atomic gas density generally has a peak value at some radius and decreases
toward the center due to conversion of the atomic to molecular gas.}. Additionally, it seems reasonable that
the gravitational restoring force increases
with increasing radius, since there will be more mass inside the given
radius. Both effects would tend to cause the pitch angle to tighten with
increasing radius, and this effect is often seen in spiral galaxies \citep[see, e.g.,][]{Me:2012}.
It has been often proposed in the past that different mechanisms may
explain spiral structure in different galaxies; for instance, the
mechanism that produces grand design spirals may differ from the one
that produces flocculent spirals \citep[e.g.,][]{D'Onghia:2013}. Furthermore, it has been demonstrated that galaxies that appear to have grand design structure in infrared light can appear flocculent in blue images, suggesting that stellar and gaseous disks are decoupled \citep{Grosbol:Patsis:1998}. It is striking that the sample used
in this study contains quite a few flocculent or multi-armed patterns,
which are not grand design. The existence of a very low
scatter planar correlation for all of these galaxies is thus very
significant and implies that different galactic morphologies all adhere to the same imposed mechanism of density wave theory. It is true that four of the galaxies have a noticeably
greater scatter than the others, and further study with larger
samples might yet support the existence of two kinds of spiral
structure. It is worth noting that two of these four galaxies represent
the extremes of gas density for the sample, one having clearly the
highest gas density in its disk, another clearly the lowest.
In recent years, there has been some discussion that spiral arms may be quite transient, persisting for only one or two revolutions of the disk galaxy \citep{Toomre:Kalnais:1991}. In recent years, there have been attempts to show theoretically
that more long-lasting spiral patterns are possible \citep{Sellwood:Carlberg:1984,D'Onghia:2013}. This Letter
suggests that even if spiral patterns are transient,
some resonant mechanism compels the pattern, when it reforms, to resume something close to its previous pitch angle.
The relation discovered here might be useful as a tool in the study
of disk galaxies. One of the three quantities, disk density, is
relatively difficult to measure. The relation found here could be
used to measure it indirectly from the other two quantities
(pitch angle and central bulge mass), which would be easier to
measure. In addition, the existence of the three-way correlation
may enable more careful studies of the important relation between
pitch angle and central mass, which is itself a very useful marker
for quantities such as the central black hole mass \citep{Seigar:2008,Berrier:2013}.
It has long been known that pitch angle does depend on the distribution
of mass \citep[e.g.,][]{Seigar:2006,Seigar:2014}
and on the size of the central bulge
(for instance, the observed correlation between pitch angle and sigma
reported in \citet{Seigar:2008} and \citet{Berrier:2013}, as well as the qualitative relation of
pitch angle to bulge size featured in the Hubble classification).
In addition, it has been reported that pitch angle varies with the
total mass of gas in galactic disks \citep{Roberts:1975}.
This Letter demonstrates the fundamental way in which we can understand
how the spiral structure depends on the distribution of mass in
disk galaxies. Furthermore, it illustrates how the qualitative Hubble morphological types can exhibit varying pitch angles for galaxies that have similarly sized bulges and are thus categorized as the same type. These galaxies likely have different gas densities in their disks. Although the density wave theory provided the
inspiration for this study, other theories may also be able to
explain this result. Certainly, any successful theory of spiral
structure must be able to do so.
\acknowledgments
The authors thank Marc Seigar and Matthew Bershady, who provided valuable comments on drafts of the Letter. We thank the National Optical Astronomy Observatory for observing time on the WIYN 3.5 $m$ telescope at Kitt Peak National Observatory. We acknowledge the use of data obtained with PPak and SparsePak IFUs on the Calar Alto and WIYN 3.5 $m$ telescopes, made available by the DiskMass Survey team. This research has made use of the
NASA/IPAC Extragalactic Database, which is operated by
the Jet Propulsion Laboratory, California Institute of Technology,
under contract with the National Aeronautics and Space Administration.
Plots were generated using \textit{Mathematica}. Numerical computations and statistical analyses were performed with \textit{Matlab}. We also thank
the National Science Foundation (NSF) for REU Site grant \#0851150, which
contributed to a significant part of the data collection for this Letter. This research has made use of NASA's Astrophysics
Data System. K. B. W. acknowledges support from NSF (USA) grant OISE-0754437 and NWO (NL) grant 614.000.807.
|
\section{MAMBO observations} \label{sect:obs}
We mapped the field around $\epsilon$~Eridani with the
117-channel Max-Planck Bolometer array {\it MAMBO-2} \citep{Krey98}
of the \textit{IRAM} 30-meter radiotelescope on Pico Veleta, Spain (2900~m).
\textit{MAMBO-2} has a half-power spectral bandwidth from 210 to
290 GHz, with an effective frequency centered on 250~GHz (1.20~mm) for thermal emission spectra.
The 117-channel array is $240'' \times 240''$ in size, and
the bolometers are arranged in a hexagonal pattern
with a beam separation of $22''$. This is twice the effective FWHM beam ($10.7''$) and provides an undersampled image.
We used the standard on-the-fly mapping technique, where one map consists
of 39 azimuthal subscans of 74 sec in duration each with a scanning velocity of $\rm 5''/sec$ over $370''$ in azimuth, and
with an incremental step of $8''$ over $312''$ in elevation, while wobbling the secondary mirror at 2 Hz by $60''$
or $46''$ in azimuth (the observations are split between the two throws).
This scanning pattern produces time streams of data that are converted
into a fully sampled spatial map with $3.5''$ pixels.
We conducted 14 forty-eight-minute observations of the field around $\epsilon$~Eridani
on nine different days between November 16 and December 4, 2007, totaling 11.2 hours of on-source observations.
Atmospheric conditions were good during these winter observations,
with typical zenith opacities between 0.1 and 0.3 at 250 GHz and low sky noise, that is,\textit{} low atmospheric fluctuations.
The telescope pointing was checked before and after each forty-eight-minute map by using
the same nearest bright point source (J0339-018); it
was found to be stable to greater accuracy than $3''$.
The absolute flux calibration is based on observations of several standard calibration sources, including
planets, and on a tipping curve (sky dip) of the atmospheric opacity once every few hours.
The resulting absolute flux calibration uncertainty is estimated to be about 10\% (rms).
The data were analyzed using the mopsic software package written by R. Zylka at \textit{IRAM}. The chopped observations
acquired with the telescope pointing alternately on and off source at the wobbler rate produce
double-beam maps that were combined to a single map using the shift-and-add procedure
\footnote{http://www.iram.es/IRAMES/mainWiki/CookbookMopsic} \citep{Emer79}. Compared to a complete image
restoration, this produces maps with a sensitivity higher
by about a factor 2, at the expense of no sensitivity to emission
structures in scan direction that are larger that the wobbler throw of $60''$ or $46''$.
The intensity map of the entire field observed around $\epsilon$~Eridani is presented in Fig~\ref{fig:allfield_raw}.
The total integration time of $11.2$ hours yielded an rms noise of $\rm \sim 0.81~mJy/11''$~beam in the central part
of the co-added map ($r < 50''$) where the structure is located. As seen in Fig~\ref{fig:allfield_raw},
noise is not uniform across the map because the scanned field is about twice
as large as the bolometer array size, so that more data are taken in the central region of the map, where the source of
interest is, than near the edges.
\begin{figure}[h!]
\resizebox{8.5cm}{!}{\includegraphics[angle=-90] {Fig1.ps}}
\caption{Intensity map of the entire field around $\epsilon$~Eridani observed at $\lambda= 1.2$~mm with \textit{MAMBO}
in November 2007 at Pico Veleta (\textit{IRAM} 30-meter radiotelescope).
A ring is clearly apparent in the central region of the map centered on the position of $\epsilon$~Eridani on 15 November 2007
($\alpha= \rm 3^h~32^m 55.32^s$ and $\delta= -9^{\circ}~27'~29.7''$). The point source interior to the ring
is within one pixel from the map center coincidental with the photosphere of $\epsilon$~Eridani.
Within $50''$ from the map center, the noise rms is 0.81~mJy/$11''$beam, and the peak brightness is 3.1~mJy/$11''$beam.
Over the whole map, the noise increases with distance from the center because of the scanning law of the telescope. The color code
is blue, purple, red, orange, yellow, and white for increasing positive
intensity and black for all negative intensities. The pixel size is $3.5''$, and the noise is uncorrelated across pixels.
}
\label{fig:allfield_raw}
\end{figure}
\begin{figure}[h!]
\resizebox{8.5cm}{!}{\includegraphics[angle=-90] {Fig2.ps}}
\caption{Synthesized beam resulting from the shift-and-add procedure used to restore the \textit{MAMBO} map from
the chopped observations
of the $\epsilon$~Eridani field conducted on nine days between November 16 and December 4, 2007.
Each observation is 48 minutes long.
The hour angles of the observations comprise only $\pm 2$ hours because $\epsilon$~Eridani
culminates at the relatively low elevation of $43^{\circ}$ at Pico Veleta. Dashed white contours are -7\%, -4\%, -2\%,
orange contours are +25\%, +50\%, +90\%, of the peak at the center.
}
\label{fig:dirty}
\end{figure}
\null
\null
In the intensity map of Fig~\ref{fig:allfield_raw}, a ring is clearly apparent in the central region of the map, and
a point source is located within only one pixel ($3.5''$) from the map center.
Moreover, there are negative (black) side lobes
east and west of the ring, which have slightly distorted the outer edges
of the structure. This is expected given that the observations were taken over
a limited range of hour angles ($\pm 2$~hours), producing the point spread function shown in Fig.~\ref{fig:dirty}
for the shift-and-add procedure of image restoration. To overcome these slight systematics in the map, we restored the map with
the CLEAN algorithm \citep{Hogb74}, which is widely used
for image restoration in radio interferometry with complicated point spread functions because of limited $u-v$ coverage
\citep{Clar80}. With this algorithm, the central region of the map of Fig.~\ref{fig:allfield_raw} was
first deconvolved iteratively with the synthesized beam of Fig.~\ref{fig:dirty}
by adopting a gain of 0.1 (fraction of the synthesized
beam scaled to the map maximum and removed
at each iteration), and in running five hundred iterations that produced 46 CLEAN $\delta-$functions.
Then, these resulting CLEAN spikes were convolved with a 2D Gaussian with a FWHM
of $10.7''$ for the beam of the telescope. The final CLEAN map is shown in Fig.~\ref{fig:clean}, from which the photosphere
has been subtracted and the residual map not added back for clarity, as is standard.
\section {Photometric excess of $\epsilon$~Eridani at long wavelengths} \label{sect:photo}
A ring is clearly apparent in the central region of Fig~\ref{fig:allfield_raw}. Additionally, a
point source is coincidental with the position
of $\epsilon$~Eridani at the date of observation (map center) within astrometric uncertainty ($3.5''$).
The measured flux density of this point source is $1.2 \pm 0.3$~mJy. This is somewhat in excess of the flux density
of $0.53\pm 0.25$ mJy for the photosphere of $\epsilon$~Eridani we predicted at $\lambda=1.2$mm with the NexGen stellar atmospheric
model \citep{Haus99}, $T_{eff}=5034~\pm~228$~K \citep{Pale15}, and the UBVRIJHK magnitudes in \citet{Duca02}.
\citet{Grea14} have detected a warm component associated with $\epsilon$~Eridani in their {\it Herschel} PACS images
that they modeled with an inner belt of 14~AU in radius ($4''$).
The flux density of their model (their Fig.~3) extrapolated at 1.2~mm is higher than 5~mJy, meaning that {\it }
it is well in excess of our MAMBO measurement ($1.2 \pm 0.3$~mJy), but it may contribute to the slight excess above the
photospheric level just mentioned. Their model is based on a modified blackbody law; their emission spectrum scales
as $\lambda^{-(2+\beta)}$ with $\beta=0.4$ longward of 150~$\mu$m. A steeper spectrum is required with our MAMBO measurement,
which indicates a population of smaller dust grains in this warm belt.
Aperture photometry made in our MAMBO image of Fig~\ref{fig:allfield_raw} has yielded a total flux density of $17.3~\pm~3.5$~mJy
at 1.2~mm within a radius of $30''$, including the contribution
of the point source. The uncertainty includes statistical error
and a 10\% error in absolute flux calibration.
A first map of the structure around $\epsilon$~Eridani at 1.2~mm was made with the \textit{SIMBA}
bolometer array at the \textit{SEST} radiotelescope
in Chile with a beam FWHM of $24''$.
This map showed an extended structure that can be fit with a circular Gaussian intensity distribution with a FWHM of $36.4''$
and a total flux density of $21.4 \pm 5.1$~mJy \citep{Schu04}, which is consistent with our measurement at the same wavelength.
An earlier millimeter observation of $\epsilon$~Eridani was conducted
with an \textit{MPIfR} bolometer system also at {\it SEST} and yielded $24.2\pm3.4$~mJy at 1300$\mu$m \citep{Chin91},
somewhat in excess of our measurement, even though we observed at a shorter wavelength.
Properties of emitting particles in the cold belt can be tested with the modified blackbody law. From the total flux density
of $37.0\pm2.5$~mJy at 850~$\mu$m \citep{Grea05} and our measurement at 1.2~mm, one finds
$\beta=0.4^{+0.6}_{-0.4}$. Notwithstanding the large uncertainty, this value is lower than $\beta =1.0\pm 0.15$
calculated between 450 and 850~$\mu$m ($S_{450}=250 \pm 20$ mJy in \citet{Grea05}) and might indicate a change in
slope for the emission of medium-sized particles (pebbles).
The compound of these slopes and uncertainties, however, is consistent with the model developed by \citet{Gasp12b} for
a collisional cascade that reaches
equilibrium with a mass distribution that is steeper than the traditional solution by Dohnanyi.
Their model implies a slope of $-2.65$ ($\beta =0.65$)
for the long-wavelength Rayleigh-Jeans tail of the SED, consistent with our values of $\beta$ above.
The slope of their model does not depend on material properties of the medium-sized particles since they
predominantly contribute to the emission at long-wavelength in fully absorbing the incident light
of the star ($Q_{abs}=1$), as also discussed in \citet{Wyat02}.
\begin{figure}[h!]
\resizebox{12.5cm}{!}{\includegraphics[angle=-90]{Fig3.ps} }
\caption{\textit{MAMBO} intensity map of the structure around $\epsilon$~Eridani at $\lambda= 1.2$mm restored with the CLEAN algorithm.
The star $\epsilon$~Eridani is at the map center (0,0), which is at $\alpha= \rm 3^h~32^m 55.32^s$ and $\delta= -9^{\circ}~27'~29.7''$ (15 November 2007).
The telescope beam is shown as a hatched circle (FWHM = $10.7''$).
Contour levels are 30\%, 60\%, and 90\% of the northwest brightness peak. The first contour (30\%) is at the level
of the CLEAN residuals, which were not added back to the map for clarity, as is standard.
The point source (1.2 mJy) close to the map center has been subtracted (see text of Sect. \ref{sect:photo}).} The pixel size is $3.5''$, and the map is unsmoothed.
\label{fig:clean}
\resizebox{6.5cm}{!}{\includegraphics[scale=0.5, angle=0]{Fig4.ps} }
\caption{\textit{SCUBA} intensity image of the structure around $\epsilon$~Eridani observed at $\lambda=850~\mu$m in \citet{Grea05}. The full field is
$70''$ wide in RA and DEC coordinates; north is up and east to the left.}
\label{fig:scubaimag}
\end{figure}
\section{Observed azimuthal structure around $\epsilon$~Eridani } \label{sect:azi}
The final \textit{MAMBO} image of the structure around $\epsilon$~Eridani at 1.2~mm is presented in Fig.~\ref{fig:clean}
with the photosphere subtracted. The ring-like appearance is striking and was originally discovered in \textit{SCUBA}
images at 450~$\mu$m and 850~$\mu$m \citep{Grea98, Grea05}, the latter is
reproduced in Fig.~\ref{fig:scubaimag}.
The main features in the \textit{MAMBO} and \textit{SCUBA} images are similar.
There are four emission clumps
in the southeast, northeast, northwest, and southwest sectors, and there are
two deep hollows to the east and west (east is toward positive RA).
These features are similarly located in the two images except for the southeast clump,
as reported in Table \ref{tab:location}. This positional coincidence with the exception of the southeast clump
is also apparent by comparing the \textit{MAMBO} and \textit{SCUBA} azimuthal profiles shown in Fig.~\ref{fig:azim}.
Hence, the most noticeable difference is the southeast clump, which appears
as a prominent arc-like feature in the \textit{SCUBA} image while it is
clearly more compact and fainter in the \textit{MAMBO} image.
However, in the most recent image of $\epsilon$~Eridani with \textit{SCUBA2} (Greaves and Holland, private communication),
this clump is actually point-like and positioned
consistently with its counterpart in the \textit{MAMBO} image.
In fact, it is possible that the arc-like feature in the earlier \textit{SCUBA} image is an artifact that is due to
incomplete removal of atmospheric fluctuations.
The MAMBO structure we observed in November 2007 (Fig.~\ref{fig:clean}) is slightly elongated in the north-south direction, while the \textit{SCUBA}
structure resulting from the integration of data from 1997 to 2002 (Fig.~\ref{fig:scubaimag}) is oriented southeast to northwest.
This appears inconsistent, but the most recent image of $\epsilon$~Eridani made in March 2011 with \textit{Herschel/PACS}
shows a structure that is slightly elongated in the north-south direction \citep{Grea14}, thus similar as with \textit{MAMBO} in November 2007.
Greaves and colleagues argued that the eastwest extension seen earlier was caused by the disk moving against
a background of contaminant sources, with the significant proper motion of the star ($1''$/yr) almost entirely westward (see also \citet{Poul06}).
Hence, these contaminants contribute to the brightness distribution of the disk, and snapshots will eventually disentangle the
two emissions. It is already clear from the available data, however, that the disk emission dominates.
The slight elongation of the structure in the north-south direction is indicative of a moderately inclined structure
($32^{\circ}$) \citep{Grea14}.
The overall agreement between the two images made with \textit{MAMBO} and \textit{SCUBA} provides tentative evidence
of the presence of four emission clumps, but to characterize
them fully, additional observations are required.
\begin{table}
\caption{Coordinates of the four emission clumps in the images.}
\label{tab:location}
\centering
\begin{tabular}{l| c c | c c}
\hline\hline
& \multicolumn{2}{c}{MAMBO} & \multicolumn{2}{c}{SCUBA} \\
Sector & $\Delta \alpha cos\delta$ ($''$) & $\Delta \delta$ ($''$) & $\Delta \alpha cos\delta$ ($''$) & $\Delta \delta$ ($''$) \\
\hline
NW & $-9.5$ & $+18$ & $-11$ & $+16$ \\
SW & $-9.5$ & $-9$ & $-12$ & $-11$ \\
NE & $+16$ & $+9$ & $+16$ & $+11$ \\
SE & $+6$ & $-15$ & $+18$ & $-13$ \\
\hline
\end{tabular}
\end{table}
\begin{figure}[t!]
\resizebox{8.5cm}{!}{\includegraphics[angle=-90] {Fig5.ps}}
\caption{Azimuthal profiles of the dust emission around $\epsilon$~Eridani computed for \textit{MAMBO} (black) and \textit{SCUBA} (blue).
The \textit{MAMBO} profile has been computed from the unsmoothed image with the photosphere subtracted.
Each \textit{MAMBO} point represents the mean brightness over a small sector with an opening angle of
$10^{\circ}$ and ranging radially from $10''$ to $30''$. Points are independent
and uncertainties are the noise rms of the map divided by the square root of the number of pixels
in each sector ($\sim$ 6 pixels). The \textit{SCUBA} profile is
from \citet{Grea05} and has been divided by 2, for clarity.
The SE counterpart is inconsistent between \textit{MAMBO} and \textit{SCUBA;} this is discussed in Sect. \ref{sect:azi}.
The angle increases counterclockwise from zero at north,
and clumps are labeled.}
\label{fig:azim}
\end{figure}
\section{Observed radial structure around $\epsilon$~Eridani } \label{sect:prof}
The radial brightness profile of the structure in Fig.~\ref{fig:prof} was
calculated by averaging intensities over $4''$ wide, elliptical annuli of increasing radii. To maximize the peak of the
profile, the structure must be projected on the sky
with an inclination of $\sim 25^{\circ}$ and PA of $\sim 0^{\circ}$, and
the center of the ellipses must be displaced by $2''$ westward from the map center. This optimal geometry
is consistent with the slightly inclined structure ($32^{\circ}$) along the north-south direction seen
in the 160~$\mu$m image provided with {\it Herschel/PACS} \citep{Grea14}.
We have fit a Gaussian and a constant level
($ c + a \exp\big(-0.5\times \big({(r-r_0)/\sigma_0}\big)^2\big) $) to this
measured radial profile and found a constant level $c=-0.08\pm 0.03$~mJy/$11''$beam, an amplitude
$a=1.65\pm 0.13$~mJy/$11''$beam, a peak radius
$r_0=17.7\pm0.4''$, and an
FWHM $= 2 \sigma_0 \times \sqrt{2 ln2} = 12\pm 1''$. The fit is characterized
by a normalized $\chi^2_{\nu}$ of 0.95 for the number of degrees of freedom of 26.
The peak radius ($17.7'' \pm 0.4''$) is comparable to $\sim 18''$ measured in \citet{Grea05}.
Hence, at the distance of the star of 3.2~pc,
this peak radius corresponds to $R=57 \pm 1.3$~AU.
The fit FWHM of the profile is consistent with the
\textit{IRAM} 30-meter radiotelescope beam of $10.7''$ \citep{Kram13}. Thus,
we estimated limits for the belt width $\Delta R$ in deconvolving the
profile with the measured lower and upper limits of its FWHM,
$11'' \leq \sqrt{10.7''^2 + (\Delta R)^2} \leq 13''$, and found $2.5'' \leq \Delta R \leq 7''$, or $8 \leq \Delta R \leq 22$~AU at 3.2~{\rm pc}.
Hence, the fractional belt width is $0.1 \leq \Delta R / R \leq 0.4$,
smaller than the fractional width of the Kuiper Belt
$\Delta R / R$ of 0.5 in \citet{Ster97}, and possibly as narrow as the belt around Fomalhaut
at millimeter wavelength (Boyles et al 2012).
Our upper limit of the width of the belt around $\epsilon$~Eridani
is more stringent than the estimate made from the radial profile at 850~$\mu$m in \citet{Grea05}
because the beam of the JCMT is broader $(14'')$. Moreover, \citet{Grea14} modeled their {\it Herschel} PACS images and
determined a mean radius of 61~AU and a width of $\sim 15$~AU ($\Delta R / R = 0.2-0.3$),
indicating that the belt is quite radially confined according to the model.
\begin{figure}[h!]
\resizebox{8.5cm}{!}{\includegraphics[angle=-90] {Fig6.ps}}
\caption{Radial profile of the dust emission around $\epsilon$~Eridani computed from the photosphere-subtracted
\textit{MAMBO} image.
Points are independent and represent the mean brightness calculated over $4''$ wide, elliptical
annuli of increasing semi-major axis (structure is inclined by 25$^{\circ}$ relative
to the plane of the sky along the north-south direction (PA=0$^{\circ}$)). The uncertainty of each brightness point
is the noise rms of the map divided by the square root of the number of pixels in each annulus. The Gaussian in red is based on
the four-parameter fit described in Sect. \ref{sect:prof}.}
\label{fig:prof}
\end{figure}
\begin{table*}
\caption{Sizes of debris disks estimated from images at long
wavelength. }
\label{tab:radii}
\centering
\begin{tabular}{l|l|c c c c c |c|c|c}
\hline\hline
& & $R_{in}$ & $R_{out}$ & $R_{mean}$ & $\Delta R$ & $\Delta R/R_{mean}$ & $\lambda$ & & ref. \\
& & (AU) & (AU) & (AU) & (AU) & & ($\mu$m) & & \\
\hline
Vega & A0V & $< 85$ & $> 105$ & & & $>0.2$ & 1300 & IRAM-PdB & 1 \\
HD21997 & A3IV/V & $55\pm16$ & & & $< 108$ \tablefootmark{a} & $< 2$ & 886 & ALMA & 2 \\
Fomalhaut & A4V & & & 141.5 & $16\pm3$ & 0.1 & 857 & ALMA & 3 \\
HD107146 & G2V & 50 & 170 & & & 1.1 & 880 & ALMA & 4 \\
$\eta$~Corvi & F2V & & & $164\pm 4$ & $< 60$ & $< 0.3$ & 850 & SCUBA2 & 5 \\
Sun & G2V & 30 & 50 & & & 0.5 & & $-$ & 6 \\
$\epsilon$ Eridani & K2V & & & $57\pm 1.3$ & $< 22$ & $<0.4$ & 1200 & IRAM-30m & 7 \\
AU Mic & M1V & $8^{+11}_{-1}$ & $40\pm0.4$ & & & 1.3 & 1300 & ALMA & 8 \\
\hline
\end{tabular}
\tablebib{
(1)~\citet{Piet11}; (2)~\citet{Moor13}; (3)~\citet{Bole12}; (4)~\citet{Hugh11}; (5)~\citet{Duch14}; (6)~\citet{Ster97}; (7)~this work; (8)~\citet{MacG13}.
}
\tablefoot{\tablefootmark{a} angular resolution is $1.5''$, the belt is smaller than beam, hence $\Delta R <1.5''\times 72~$pc = 108~AU.
}
\end{table*}
\section{Discussion} \label{sect:discu}
\subsection{Robustness of the clumps in the MAMBO and SCUBA images}
Asymmetries and clumps observed in
images of debris disks can be related, in theory, to dust and planetesimals trapped in resonances
with planets just interior to the inner edge of the belt \citep[e.g.][]{Liou96, Ozer00, Wyat03, Kriv07}.
However, reliable identification of predicted structures in real images at long wavelength has been difficult.
For example, \citet{Piet11} and \citet{Hugh12} could not recover the clumpy structure around Vega
reported earlier. Similarly, \citet{Hugh12} could not confirm the asymmetries found above $3 \sigma$
in the structure around HD 107146 in previous observations.
In contrast, for $\epsilon$~Eridani, the similarities between the \textit{MAMBO}
and \textit{SCUBA} images discussed here provide tentative evidence that the observed structure is robust.
The four clumps may result from mean-motion resonances with an unseen planet in the system that
is not detectable with current radial velocity measurements because
its orbital period is as long as tens or hundreds of years.
Inferring the orbital parameters of a planet from its imprint on the structure of a disk is an area of active research.
The ring-like structure of the solar system between 30 and 50~AU has been described as particles trapped in
mean-motion resonances with Neptune, and the clearing within 10~AU described as particules ejected because
of the giant planets Jupiter and Saturn \citep{Liou96,Liou99, Moro02}.
Clumps due to mean-motion resonances must be significantly enhanced to be observable, and two mechanisms have been studied.
Dust grains can be trapped into resonances during their transport inward
by Poynting-Robertson drag \citep{Roqu94, Kuch03, Dell05}, or resonant planetesimals and grains can be swept along
during planet migration \citep{Wyat03, Rech06}.
First-order mean-motion resonances ($p+1:p$, $p>0$) with low $p$ are the strongest, but planet mass and orbit eccentricity
have a significant impact on the final structure \citep{Rech06}.
Predicting the number of observable clumps and their relative intensities requires additional constraints. \citet{Mare09} determined
an upper limit of four Jupiter masses for any planet (1 Gyr models) inside the inner rim of the submillimeter ring
of $\epsilon$~Eridani with \textit{Spitzer/IRAC} data. We note also that the two hollows east and west
seen consistently in both the \textit{SCUBA} and \textit{MAMBO} images require a planet of sub-Jupiter mass
if arising from first-order mean-motion resonances according to several simulations \citep{Ozer00, Quil02, Rech06}.
\subsection{Narrow belt of planetesimals around $\epsilon$~Eridani}
The \textit{MAMBO} image in Fig.~\ref{fig:clean} and the radial brightness profile in Fig.~\ref{fig:prof} currently provide the most stringent upper limit on the width of the planetesimal belt around $\epsilon$~Eridani determined
directly from observations. In Table~\ref{tab:radii}, we compare this relative width ($0.1 \leq \Delta R / R \leq 0.4$)
to the sizes of debris disks determined from millimeter or submillimeter observations that trace parent planetesimals.
Given its mean radius of $57 \pm 1.3$~AU, the belt around $\epsilon$~Eridani has a cavity of a few tens of AU,
which can host a large planetary system.
The inner edge of a planetesimal belt results from the planetary formation phase when the largest planets eject
smaller bodies into the Oort cloud \citep{Gold14}, and from outward migration
\citep{Levi03}. The outer edge is set by the competition between viscous spreading in the protoplanetary disk
and disruptive effect of stellar flybys in star forming region \citep{Roso14}. Additionally, at a later stage
of evolution, the outer edge of the disk can be truncated further and its mass depleted as a result of other destructive
stellar flybys for a period as long as $\sim 100$~Myr, when the central star is still in the expanding open cluster
of its birth \citep{Lest11}.
The belt around $\epsilon$~Eridani ($57$~AU, $0.1 \leq \Delta R / R \leq 0.4$) is narrower and at a larger distance than the present-day
Kuiper Belt ($R_{mean}=$ 40~AU, $\Delta R / R =0.5$) \citep[][inner edge $\sim$ 30~AU and
outer edge $\sim$50~AU]{Ster97}. Furthermore, planetesimals most probably formed within $\sim 30$ AU
in the solar system and were subsequently pushed outward by Neptune's 2:1 mean-motion resonance during
its final phase of migration. This means that the entire Kuiper Belt formed closer to the Sun and was transported
outward during the final stages of planet formation \citep{Levi03}. Hence, the original
Kuiper Belt around the younger Sun was more compact than it is today, contrasting even more with the relatively large belt
around $\epsilon$~Eridani, which is only 850~Myr in age \citep{Difo04}.
It is useful to compare the properties of the disk around $\epsilon$~Eridani also with other disks that are spatially resolved
and are listed in Table~\ref{tab:radii}. The broadest belt is
around the solar-type star HD107146 with $\Delta R/R \sim 1.1$ \citep[][inner radius $\sim$ 50~AU, outer radius $\sim$ 170~AU]{Hugh11},
and the narrowest belt is around the A-type star Fomalhaut~A
with $\Delta R/R \sim 0.1$ \citep[][mean radius 135~AU, width $16 \pm 3$~AU]{Bole12}. These authors also showed that
two shepherding planets on both sides of the belt of Fomalhaut~A can be responsible for its narrow width.
This might be of interest for the $\epsilon$~Eridani belt.
Finally, our characterization of the belt around $\epsilon$~Eridani
with a mean radius of $57\pm 1.3$~AU and a width smaller than $<22$~AU
is not consistent with model fits to the SED of this star. \citet{Back09} found a broader submillimeter ring between 35 and 90 AU,
and \citet{Reid11} found a birth ring between 55 to 90~AU. In this latter model dust grains are transported inward by
Poynting-Robertson and stellar wind drags and are constrained by the $850~\mu$m flux density from JCMT/SCUBA.
Unfortuntately, the authors of these two studies
did not predict the flux density of the disk at 1.2~mm for comparison with our MAMBO measurement.
\null
\null
\section{Conclusion} \label{sect:concl}
The \textit{MAMBO} image of the debris disk around $\epsilon$~Eridani at 1.2~mm we presented shows a structure similar
to the one identified earlier with \textit{SCUBA} at 850~$\mu$m : a ring-like structure broken into four
emission clumps in the northeast, northwest, southwest, southeast sectors, and two deep hollows east and west.
In theory, this structure can be the imprint
of resonant clumps related to gravitational perturbations from undetected long-period planets in the system.
However, the reliability of images made with a single radiotelescope and bolometer camera has been debated before because these images can be distorted if the signal of the large atmospheric fluctuations
inherent in this type of observation is not completely removed from the data.
We argued that the identification of three of the four emission clumps (NE, NW, SW)
at the same locations within astrometric uncertainty in two images made
with different radiotelescopes, namely the \textit{IRAM} 30-meter and \textit{JCMT},
and different bolometer cameras, namely \textit{MAMBO} and \textit{SCUBA}, provides tentative
evidence that the observed structure is robust. The southeast clump is the most discrepant
in brightness and extension in the two images, but it is possible that this is an artifact.
Overall, this provides new impetus for future studies
to relate this structure to the presence of undetected planets in the system.
Additionally, we provided the most stringent upper limit on the width of the belt of planetesimals
around $\epsilon$~Eridani with $8 \leq \Delta R \leq 22$~AU. This corresponds to a relative width of $0.1 \leq \Delta R / R \leq 0.4,$ which
is narrower than for the Kuiper Belt.
Future observations at long wavelengths on finer scales with interferometers will provide additional information
on the emission clumps in the belt around $\epsilon$~Eridani, for instance,\textit{} their true size, to investigate
the interplay between debris disks and associated planetary systems.
\null
\null
\begin{acknowledgements}
We are grateful to the staff of the IRAM 30-m telescope, especially St\'ephane Leon now at ALMA,
for his dedication in managing the MAMBO pool, to Patrick Charlot
at Observatoire de Bordeaux for useful
discussions on the CLEAN algorithm, and to Sarah Maddison at Swinburne University for her benevolent assistance.
We are very grateful to our referee for constructive suggestions that have improved the paper. This work was funded in part by
CNES and CNRS PNP.
\end{acknowledgements}
\null
\null
\bibliographystyle{aa}
|
\section{#1} \setcounter{equation}{0} }
\documentclass[a4paper,11pt]{article}%
\usepackage{geometry}
\usepackage{amsmath}
\usepackage{amsfonts}
\usepackage{amssymb}
\usepackage{graphicx}
\usepackage{indentfirst}
\usepackage[small,bf]{caption}
\usepackage{slashed}
\setcounter{MaxMatrixCols}{30}
\providecommand{\U}[1]{\protect\rule{.1in}{.1in}}
\newtheorem{theorem}{Theorem}
\newtheorem{acknowledgement}[theorem]{Acknowledgement}
\newtheorem{algorithm}[theorem]{Algorithm}
\newtheorem{axiom}[theorem]{Axiom}
\newtheorem{case}[theorem]{Case}
\newtheorem{claim}[theorem]{Claim}
\newtheorem{conclusion}[theorem]{Conclusion}
\newtheorem{condition}[theorem]{Condition}
\newtheorem{conjecture}[theorem]{Conjecture}
\newtheorem{corollary}[theorem]{Corollary}
\newtheorem{criterion}[theorem]{Criterion}
\newtheorem{definition}[theorem]{Definition}
\newtheorem{example}[theorem]{Example}
\newtheorem{exercise}[theorem]{Exercise}
\newtheorem{lemma}[theorem]{Lemma}
\newtheorem{notation}[theorem]{Notation}
\newtheorem{problem}[theorem]{Problem}
\newtheorem{proposition}[theorem]{Proposition}
\newtheorem{remark}[theorem]{Remark}
\newtheorem{solution}[theorem]{Solution}
\newtheorem{summary}[theorem]{Summary}
\newenvironment{proof}[1][Proof]{\noindent\textbf{#1.} }{\ \rule{0.5em}{0.5em}}
\geometry{left=1.5cm, right=1.5cm, top=1.5cm, bottom=1.5cm}
\setlength{\topmargin}{-1cm} \setlength{\evensidemargin}{-0.75cm}
\setlength{\oddsidemargin}{-0.75cm} \setlength{\textwidth}{17.5cm}
\setlength{\textheight}{24.0cm} \setlength{\parskip}{10pt}
\newcommand{pdf}{pdf}
\DeclareMathOperator{\Tr}{Tr}
\renewcommand{\refname}{References}
\newcommand{\mbox{$\frac{1}{2}$}}{\mbox{$\frac{1}{2}$}}
\hyphenation{coun-ter-term}
\begin{document}
\date{}
\title{\textbf{Dimension two condensates in the Gribov-Zwanziger theory in the Coulomb gauge}}
\author{\textbf{M.~S.~Guimaraes}\thanks{<EMAIL>}\,\,,
\textbf{B.~W.~Mintz}\thanks{<EMAIL>}\,\,,
\textbf{S.~P.~Sorella}\thanks{<EMAIL>}\,\,\,,\\[2mm]
{\small \textnormal{ \it Departamento de F\'{\i }sica Te\'{o}rica, Instituto de F\'{\i }sica, UERJ - Universidade do Estado do Rio de Janeiro,}}
\\ \small \textnormal{ \it Rua S\~{a}o Francisco Xavier 524, 20550-013 Maracan\~{a}, Rio de Janeiro, Brasil}\normalsize}
\maketitle
\begin{abstract}
We investigate the dimension two condensate $\langle {\bar \phi}^{ab}_i \phi^{ab}_i -{\bar \omega}^{ab}_i \omega^{ab}_i \rangle$
within the Gribov-Zwanziger approach to Euclidean Yang-Mills theories in the Coulomb gauge, in both $3$ and $4$ dimensions.
An explicit calculation shows that, at the first order, the condensate
$\langle {\bar \phi}^{ab}_i \phi^{ab}_i -{\bar \omega}^{ab}_i \omega^{ab}_i \rangle$ is plagued by a non-integrable IR divergence
in $3D$, while in $4D$ it exhibits a logarithmic UV divergence, being proportional to the Gribov parameter $\gamma^2$. These results
indicate that in 3D the transverse spatial Coulomb gluon two-point correlation function exhibits a scaling behaviour, in agreement with Gribov's
expression. In 4D, however, they suggest that, next to the scaling behaviour, a decoupling solution might emerge too.
\end{abstract}
\section{Introduction}
Th Coulomb gauge, $\partial_i A^a_i=0$, $i=1,...., D-1$, is largely employed in analytic
\cite{Cucchieri:1996ja,Cucchieri:2000hv,Zwanziger:2002sh,Zwanziger:2003de,Greensite:2004ke,Zwanziger:2004np,Greensite:2004ur,Zwanziger:2006sc,Schleifenbaum:2006bq,Watson:2006yq,Epple:2006hv,Watson:2007vc,Campagnari:2009wj,Burgio:2009xp,Watson:2010cn,Leder:2011yc,Weber:2011zzd,Szczepaniak:2011bs,Watson:2011kv,Campagnari:2011bk,Watson:2012ht,Huber:2014isa} as well as in lattice numerical investigations \cite{Cucchieri:2000gu,Cucchieri:2000kw,Burgio:2008jr,Nakagawa:2009zf,Burgio:2012bk} of non-perturbative aspects of Euclidean yang-Mills theories. An impressive number of results are nowadays available in this gauge, providing a consistent scenario for confinement.
\\\\In this letter we focus on aspects of the Gribov-Zwanziger formulation of the Coulomb gauge \cite{Gribov:1977wm,Zwanziger:2002sh,Zwanziger:2003de,Zwanziger:2004np,Zwanziger:2006sc},
which implements the restriction of the domain of integration in the
functional Euclidean integral to the Gribov Region $\Omega$, defined as the set of all field configurations fulfilling the Coulomb condition and for which the Faddeev-Popov operator,
${\cal M}^{ab}= -\delta^{ab}\partial_i \partial_i -g f^{acb}A^c_i\partial_i$, is strictly positive, namely
\begin{equation}
\Omega = \{ A^a_i \;, \;\; \partial_iA^a_i =0\;, \;\; {\cal M}^{ab} >0 \;\; \} \;. \label{gr}
\end{equation}
The restriction to the region $\Omega$ accounts for the existence of the Gribov copies, which affect the Coulomb gauge. The so called Gribov-Zwanziger action
\cite{Gribov:1977wm,Zwanziger:2002sh,Zwanziger:2003de,Zwanziger:2004np,Zwanziger:2006sc} is the final result of the restriction to the region $\Omega$. Besides
the Coulomb gauge, the restriction to the Gribov region has been effectively implemented in the Landau \cite{Gribov:1977wm,Zwanziger:1988jt,Zwanziger:1989mf}
and maximal Abelian gauges \cite{Capri:2006vv,Capri:2006cz}, where the corresponding Gribov-Zwanziger actions have been worked out. A feature of the Gribov-Zwanziger
set up in the Landau and maximal Abelian gauges is that the two-point gluon correlation function is strongly suppressed in the infrared region in momentum space $k$,
attaining a vanishing value when $k=0$. This kind of behaviour is usually referred to as the scaling solution, also observed in the study of the Dyson-Schwinger
equations in these gauges, see \cite{Alkofer:2000wg,Huber:2009wh}. Such a behaviour is also found for the spatial gluon correlation function in the Coulomb gauge.
\\\\Nevertheless, next to the scaling solution, another type of solution, called decoupling solution, was found in both Landau
\cite{Cornwall:1981zr,Aguilar:2004sw,Dudal:2007cw,Aguilar:2008xm,Dudal:2008sp,Dudal:2008rm,Fischer:2008uz} and maximal Abelian gauges
\cite{Capri:2008ak,Capri:2008vk,Capri:2010an}, in both $3$ and $4$ dimensions. Similarly to the scaling solution, the decoupling
solution is also strongly suppressed in the infrared region, displaying a violation of positivity. However, it does not attain a vanishing value at $k=0$.
\\\\Within the Gribov-Zwanziger approach, the decoupling solution arises as the consequence of the existence of dimension two condensates
\cite{Dudal:2007cw,Dudal:2008sp,Dudal:2008rm,Capri:2008ak,Capri:2008vk,Capri:2010an}. The resulting action accounting for the inclusion of
these condensates is known as the Refined-Gribov-Zwanziger action \cite{Dudal:2007cw,Dudal:2008sp,Dudal:2008rm,Capri:2008ak,Capri:2008vk,Capri:2010an}.
\\\\In the present work, we investigate the condensate $\langle {\bar \phi}^{ab}_i \phi^{ab}_i -{\bar \omega}^{ab}_i \omega^{ab}_i \rangle$ in Coulomb gauge,
in both $3$ and $4$ dimensions. In $3D$ we find, by an explicit first order calculation, that the integral defining
$\langle {\bar \phi}^{ab}_i \phi^{ab}_i -{\bar \omega}^{ab}_i \omega^{ab}_i \rangle$ exhibits a non-integrable IR divergence, showing that the condensate cannot
be safely introduced in $3D$. As a consequence, in $3D$, the transverse equal-time gluon propagator exhibits the scaling type behaviour given by Gribov's expression.
\\\\Moreover, in $4D$, we find that the condensate is safe in the IR, being plagued by a mild UV logarithmic divergence, precisely as the gap equation defining the
Gribov parameter. This suggests that, apart from a UV proper renormalization, a non-vanishing dimension two condensate
$\langle {\bar \phi}^{ab}_i \phi^{ab}_i -{\bar \omega}^{ab}_i \omega^{ab}_i \rangle$ might show up also in the Coulomb gauge.
As a consequence, the spatial transverse two-point gluon correlation function exhibits a decoupling solution, next to the well known scaling one. To some extent,
this suggests a kind of common feature of the Landau, Coulomb and maximal Abelian gauge in $4D$ Euclidean space-time. \\\\The present letter is organised as follows.
In Sect.2 we give a short overview of the Gribov-Zwanziger action in the Coulomb gauge. Sect.3 is devoted to the evaluation of the condensate
$\langle {\bar \phi}^{ab}_i \phi^{ab}_i -{\bar \omega}^{ab}_i \omega^{ab}_i \rangle$ at the first order, in both $3D$ and $4D$.
\section{The Gribov-Zwanziger action in the Coulomb gauge}
The Gribov-Zwanziger action implementing the restriction to the Gribov region $\Omega$, eq.\eqref{gr}, in the Coulomb gauge, $\partial_i A^a_i=0$, $i=1,...,(D-1)$, reads \cite{Gribov:1977wm,Zwanziger:2002sh,Zwanziger:2003de,Zwanziger:2004np,Zwanziger:2006sc}
\begin{eqnarray}
S_{GZ} & = & \int d^Dx \left( \frac{1}{4} F^{a}_{\mu\nu} F^{a}_{\mu\nu} + b^a \partial_i A^a_i + {\bar c}^a \partial_i D_{i}^{ab} c^b \right) \nonumber \\
&+& \int d^Dx \left( - {\bar \phi}^{ab}_{\mu} \partial_i D_{i}^{ab} \phi^{cb}_\mu + {\bar \omega}^{ab}_{\mu} \partial_i D_{i}^{ab} \omega^{cb}_\mu
- g f^{acm} (\partial_i {\bar \omega}^{ab}_{\mu}) (D_{i}^{cp} c^p) \phi^{mb}_\mu \right) \nonumber \\
&+& \int d^Dx \left( \gamma^2 g f^{abc} A^a_i (\phi^{bc}_i - {\bar \phi}^{bc}_i ) - (D-1) (N^2-1) \gamma^4 \right) \;. \label{gza}
\end{eqnarray}
The field $b^a$ is the Lagrange multiplier enforcing the Coulomb condition, $\partial_i A^a_i=0$, while the fields $({\bar c}^a, c^a)$ are the Faddeev-Popov ghosts.
The fields $({\bar \phi}^{ab}_{\mu}, { \phi}^{ab}_{\mu})$, $\mu=1,...,D$, are a set of commuting fields while $({\bar \omega}^{ab}_{\mu}, { \omega}^{ab}_{\mu})$ are
anti-commuting. These fields are introduced in order to implement the restriction to the region $\Omega$ through a local action, eq.\eqref{gza}. Finally, the parameter
$\gamma^2$ appearing in expression \eqref{gza} is the Gribov parameter. It has the dimension of mass squared and has a dynamical origin, being determined
in a self consistent way through the gap equation
\begin{equation}
\frac{\partial {\cal E}_v}{\partial \gamma^2}= 0 \;, \label{ev}
\end{equation}
where ${\cal E}_v$ is the vacuum energy, namely
\begin{equation}
e^{-V {\cal E}_v} = \int [{\cal D}\Phi] \; e^{-S_{GZ}} \;. \label{z}
\end{equation}
To the first order
\begin{equation}
{\cal E}_v = -(D-1) (N^2-1) \gamma^4 + \frac{(D-2)}{2} (N^2-1) \int \frac{d^Dk}{(2\pi)^D} \; \log \left(k^2_D +{\vec k}^2 + \frac{2Ng^2\gamma^4}{{\vec k}^2} \right) \; + \; \;\;{\rm terms \; independent\; from}\; \; \gamma^2 \;,
\end{equation}
so that the gap equation, eq.\eqref{ev}, takes the form
\begin{equation}
\int \frac{d^Dk}{(2\pi)^D} \frac{1}{k^2_D\; {\vec k}^2 + {\vec k}^4 +{2Ng^2\gamma^4} }= \frac{(D-1)}{(D-2)} \frac{1}{2Ng^2} \;. \label{gp1}
\end{equation}
From the Gribov-Zwanziger action, it follows that the tree level two-point gluon spatial correlation function is given by
\begin{equation}
\langle A^a_i(k) A^b_j(-k) \rangle = \frac{\delta^{ab}}{ k^2_D +{\vec k}^2 + \frac{2Ng^2\gamma^4}{{\vec k}^2}} \left( \delta_{ij} -\frac{k_i k_j} {{\overrightarrow k}^2} \right) \;, \label{gzgluon}
\end{equation}
leading to an equal-time transverse form factor \cite{Burgio:2012bk}
\begin{equation}
D^{tr}_{GZ} ({\overrightarrow k}) = \frac{|{\overrightarrow k}|}{\sqrt{{\overrightarrow k}^4 + 2Ng^2\gamma^4}} \;, \label{trgz}
\end{equation}
which exhibits the scaling behaviour, $D^{tr}_{GZ} (0)=0$. Before going any further, it is worth reminding briefly how expression \eqref{gzgluon} is derived. A simple calculation shows in fact that the quadratic part of the Gribov-Zwanziger action in the gluon sector takes the following form
\begin{equation}
S_{GZ}^{\rm quadr-gluon} = \int d^Dx \;\left( \frac{1}{2} A^a_D (-{\vec{\partial}}^2\;) A^a_D + \frac{1}{2} A^a_i \left(-\partial_D^2 -{\vec{\partial}}^2+ \frac{2Ng^2\gamma^4}{-{\vec{\partial}}^2}\right) A^a_i +b^a \partial_i A^a_i \right) \;, \label{quadr}
\end{equation}
from which one immediately derives the tree level expression for the transverse spatial gluon propagator given in eq.\eqref{gzgluon}. Moreover, one has to observe that, unlike expression \eqref{gzgluon}, the tree level temporal correlator $\langle A^a_D(k) A_D^b(-k)\rangle$ does not display an energy resolution, due to the well known existence of residual temporal gauge transformations which affect the Coulomb condition. The resolution of the energy behaviour of this correlation function is a quite delicate point which requires a detailed mastering of the renormalization procedure of the Coulomb gauge. Here, we remind the reader to the large literature existing on this subject, both in the continuum as well as in the lattice formulation \cite{Cucchieri:1996ja,Cucchieri:2000hv,Zwanziger:2002sh,Zwanziger:2003de,Greensite:2004ke,Zwanziger:2004np,Greensite:2004ur,Zwanziger:2006sc,Schleifenbaum:2006bq,Watson:2006yq,Epple:2006hv,Watson:2007vc,Campagnari:2009wj,Burgio:2009xp,Watson:2010cn,Leder:2011yc,Weber:2011zzd,Szczepaniak:2011bs,Watson:2011kv,Campagnari:2011bk,Watson:2012ht,Huber:2014isa,Cucchieri:2000gu,Cucchieri:2000kw,Burgio:2008jr,Nakagawa:2009zf,Burgio:2012bk}.
\section{The dimension two condensate $\langle {\bar \phi}^{ab}_i \phi^{ab}_i -{\bar \omega}^{ab}_i \omega^{ab}_i \rangle$}
In order to evaluate the condensate $\langle {\bar \phi}^{ab}_i \phi^{ab}_i -{\bar \omega}^{ab}_i \omega^{ab}_i \rangle$, we couple the operator $ \left({\bar \phi}^{ab}_i(x) \phi^{ab}_i(x) -{\bar \omega}^{ab}_i(x) \omega^{ab}_i(x) \right)$ to a constant source $J$, {\it i.e.}
\begin{equation}
S_{GZ} \rightarrow S_{GZ} + \int d^Dx J \left({\bar \phi}^{ab}_i(x) \phi^{ab}_i(x) -{\bar \omega}^{ab}_i(x) \omega^{ab}_i(x) \right) \;, \label{jact}
\end{equation}
and we evaluate
\begin{equation}
e^{-V {\cal E}_v(J) } = \int [{\cal D}\Phi] \; e^{-\left(S_{GZ}+ \int d^Dx J \left({\bar \phi}^{ab}_i(x) \phi^{ab}_i(x) -{\bar \omega}^{ab}_i(x) \omega^{ab}_i(x) \right)\right)} \;. \label{zj}
\end{equation}
The condensate is thus obtained by differentiating ${\cal E}_v(J)$ with respect to $J$ and setting $J=0$ at the end, namely
\begin{equation}
\frac{\partial {\cal E}_v(J)}{\partial J}\Bigl|_{J= 0} = \frac{ \int [{\cal D}\Phi] \; \left({\bar \phi}^{ab}_i(x) \phi^{ab}_i(x) -{\bar \omega}^{ab}_i(x) \omega^{ab}_i(x) \right) \;e^{-S_{GZ}}}{\int [{\cal D}\Phi] \;e^{-S_{GZ}}} = \langle {\bar \phi}^{ab}_i \phi^{ab}_i -{\bar \omega}^{ab}_i \omega^{ab}_i \rangle \;. \label{cond}
\end{equation}
At the first order, we get
\begin{equation}
{\cal E}_v(J) = \frac{(D-2)}{2} (N^2-1) \int \frac{d^Dk}{(2\pi)^D} \; \log \left(k^2_D +{\vec k}^2 + \frac{2Ng^2\gamma^4}{{\vec k}^2+J} \right) \; + \; \;\;{\rm terms \; independent\; from}\; \; J \;. \label{[j1}
\end{equation}
Taking the derivative with respect to $J$ and setting $J=0$, it turns out that
\begin{equation}
\langle {\bar \phi}^{ab}_i \phi^{ab}_i -{\bar \omega}^{ab}_i \omega^{ab}_i \rangle = - Ng^2\gamma^4 (N^2-1) (D-2)
\int \frac{d^Dk}{(2\pi)^D} \frac{1}{k^2_D\;{\vec k}^2 + {\vec k}^4 + {2Ng^2\gamma^4}} \;\frac{1}{ {\vec k}^2} \;. \label{j2}
\end{equation}
\subsection{The $3D$ case}
Let us start by considering first expression \eqref{j2} in $3D$. Here, we have
\begin{equation}
\langle {\bar \phi}^{ab}_i \phi^{ab}_i -{\bar \omega}^{ab}_i \omega^{ab}_i \rangle = - Ng^2\gamma^4 (N^2-1)
\int \frac{dk_3\; d^2{\vec k} }{(2\pi)^3} \frac{1}{k^2_3\;{\vec k}^2 + {\vec k}^4 + {2Ng^2\gamma^4}} \;\frac{1}{ {\vec k}^2} \;, \label{cd31}
\end{equation}
while for the gap equation
\begin{equation}
\int \frac{dk_3 \;d^2{\vec k}}{(2\pi)^3} \frac{1}{k^2_3\; {\vec k}^2 + {\vec k}^4 +{2Ng^2\gamma^4} }= \frac{1}{Ng^2} \;. \label{gp31}
\end{equation}
We see that, while the integral defining the gap equation, eq.\eqref{gp31}, is convergent in both IR and UV regions, the expression for the condensate, eq.\eqref{cd31}, exhibits a non-integrable singularity in the IR, due to the presence of the term $\frac{1}{ {\vec k}^2}$ which is non-integrable in two-dimensional space. This indicates that the condensate $\langle {\bar \phi}^{ab}_i \phi^{ab}_i -{\bar \omega}^{ab}_i \omega^{ab}_i \rangle$ cannot be introduced in $3D$, due to the existence of infrared divergences. A similar phenomenon occurs in the Landau gauge in $2D$, where the analogous condensate cannot be introduced due to infrared divergences \cite{Dudal:2008xd}.
\subsection{The $4D$ case}
Let us turn now to the $4D$ case, where for the condensate and for the gap equation we get
\begin{equation}
\langle {\bar \phi}^{ab}_i \phi^{ab}_i -{\bar \omega}^{ab}_i \omega^{ab}_i \rangle = - 2Ng^2\gamma^4 (N^2-1)
\int \frac{dk_4 \;d^3{\vec k}}{(2\pi)^4} \frac{1}{k^2_4\;{\vec k}^2 + {\vec k}^4 + {2Ng^2\gamma^4}} \;\frac{1}{ {\vec k}^2} \;, \label{cd41}
\end{equation}
and
\begin{equation}
\int \frac{dk_4 \; d^3{\vec k} }{(2\pi)^4} \frac{1}{k^2_4\; {\vec k}^2 + {\vec k}^4 +{2Ng^2\gamma^4} }= \frac{3}{2} \frac{1}{2Ng^2} \;. \label{gp1}
\end{equation}
In this case, both expressions are safe in the IR, while they suffer from UV divergences which should be handled by a suitable renormalization procedure in the Coulomb gauge. In fact, taking a closer look at the expression \eqref{cd41}, we may write
\begin{equation}
\langle {\bar \phi}^{ab}_i \phi^{ab}_i -{\bar \omega}^{ab}_i \omega^{ab}_i \rangle = - Ng^2\gamma^4 (N^2-1) \frac{1}{\pi^3} \int_0^\infty d\rho \int_0^\infty dr \frac{1}{\rho^2 r^2 + r^4 + 2Ng^2\gamma^4} \;. \label{j3}
\end{equation}
where use has been made of three dimensional polar coordinates. Making the change of variables
\begin{equation}
\rho \rightarrow (2Ng^2\gamma^4)^{1/4} \;\rho \;, \qquad r \rightarrow (2Ng^2\gamma^4)^{1/4} \; r \;, \label{ch}
\end{equation}
we get
\begin{equation}
\langle {\bar \phi}^{ab}_i \phi^{ab}_i -{\bar \omega}^{ab}_i \omega^{ab}_i \rangle = - (N^2-1){\sqrt{2g^2N} }\;\gamma^2 \frac{1}{2\pi^3} \int_0^\infty d\rho \int_0^\infty dr \frac{1}{\rho^2 r^2 + r^4 + 1} \;. \label{cd4}
\end{equation}
To evaluate the integral
\begin{equation}
I = \int_0^\infty d\rho \int_0^\infty dr \frac{1}{\rho^2 r^2 + r^4 + 1} \;, \label{it1}
\end{equation}
we adopt two-dimensional polar coordinates, $(\rho = R \cos\theta, \;r= R \sin\theta)$, obtaining
\begin{equation}
I = \int_0^{\sqrt{2} \Lambda} dR R\int_0^{\frac{\pi}{2}} d\theta\; \frac{1}{R^4 \sin^2\theta + 1} \;, \label{it2}
\end{equation}
where $\Lambda$ is a cutoff.
From
\begin{equation}
\int_0^{\phi} d\theta \; \frac{1}{R^4 \sin^2\theta + 1} = \frac{\arctan(\sqrt{R^4+1} \tan\phi)} {\sqrt{R^4+1}} \;, \label{it3}
\end{equation}
we finally get
\begin{equation}
I = \frac{\pi}{2} \int_0^{\sqrt{2} \Lambda} dR \frac{R}{\sqrt{R^4+1}} = \frac{\pi}{4} {\rm arcsinh}(2\Lambda^2) \;, \label{it4}
\end{equation}
which diverges logarithmically as $\Lambda \rightarrow \infty$. \\\\We see therefore that, apart from a UV renormalization, a non-vanishing two-dimensional condensate $\langle {\bar \phi}^{ab}_i \phi^{ab}_i -{\bar \omega}^{ab}_i \omega^{ab}_i \rangle$ might emerge in $4D$. The effect of this condensate on the dynamics of the theory can be taken into account by introducing the Refined Gribov-Zwanziger action in the Coulomb gauge
\begin{equation}
S_{RGZ} = S_{GZ} + \int d^4x \;\mu^2 \left({\bar \phi}^{ab}_i(x) \phi^{ab}_i(x) -{\bar \omega}^{ab}_i(x) \omega^{ab}_i(x) \right) \;, \label{rgz}
\end{equation}
where the parameter $\mu^2$ can be obtained order by order by evaluating the effective potential for the operator $ \left({\bar \phi}^{ab}_i(x) \phi^{ab}_i(x) -{\bar \omega}^{ab}_i(x) \omega^{ab}_i(x) \right)$, as done in \cite{Dudal:2011gd} in the case of the Landau gauge. Evaluating now the spatial two-point gluon correlation function with the refined action \eqref{rgz}, one gets the decoupling solution
\begin{equation}
\langle A^a_i(k) A^b_j(-k) \rangle_{RGZ} = \frac{\delta^{ab}}{ k^2_4 +{\overrightarrow k}^2 + \frac{2Ng^2\gamma^4}{{\overrightarrow k}^2+\mu^2}} \left( \delta_{ij} -\frac{k_i k_j} {{\overrightarrow k}^2} \right) \;, \label{rgzgluon}
\end{equation}
leading to an equal time transverse form factor of the decoupling type, namely
\begin{equation}
D^{tr}_{RGZ} ({\overrightarrow k}) = \frac{\sqrt{{\overrightarrow k}^2+\mu^2}}{\sqrt{{\overrightarrow k}^2 \left({\overrightarrow k}^2+\mu^2\right) + 2Ng^2\gamma^4}} \;. \label{rtrgz}
\end{equation}
Even if being well beyond the aim of the present letter, one might expect that
the existence of a decoupling type behavior for the transverse gluon propagator
should entail modifications on the infrared behavior of the ghost, perhaps resulting in a milder behavior of the ghost form factor in the deep infrared region, similarly to what happens in the Landau gauge. It is worth in fact to point out that,
within the Schwinger-Dyson approach, a decoupling type solution for the transverse gluon and its consequences on
the ghost form factor as well as on the Coulomb potential have been already analysed by the authors of \cite{Epple:2007ut}. We hope to report soon on this important topic, which deserves a more complete and detailed analysis.
\section*{Acknowledgments}
The Conselho Nacional de Desenvolvimento Cient\'{\i}fico e
Tecnol\'{o}gico (CNPq-Brazil), the Faperj, Funda{\c{c}}{\~{a}}o de
Amparo {\`{a}} Pesquisa do Estado do Rio de Janeiro, the
Coordena{\c{c}}{\~{a}}o de Aperfei{\c{c}}oamento de Pessoal de
N{\'{\i}}vel Superior (CAPES), are gratefully acknowledged.
|
\section{Introduction}
\label{sec:intro}
Herbig Ae/Be stars are intermediate mass (2-8M$_{\sun}$) pre-main
sequence and zero age-main sequence stars whose spectral energy
distribution reveals an infrared excess (Herbig 1960; Waters \& Waelkens
1998; Vieira et al. 2003). They are the precursors to main sequence
A-stars many of which have debris disks (Su et al. 2005; 2006). As
the higher mass analogs to classical T Tauri stars, Herbig Ae/Be stars
provide an important link between their lower mass solar analogs and high
mass stars.
A great deal of progress has been made in the study of the circumstellar
environment of Herbig Ae/Be stars since the last international conference
on Herbig Ae/Be stars (The et al. 1994a). Over the past two decades,
advances in coronagraphic imaging (Grady et al. 1999, 2005, 2014;
Weinberger, et al. 1999),
the (sub)millimeter (Mannings \& Sargent 1997; Dent et al. 2005) and infrared observations
(e.g. Meeus et al. 2001, 2012), and SED modeling (Dullemond et al. 2007) have
advanced our understanding of this environment. While the the geometry
of the dust giving rise to the infrared excess was controversial two decades
ago (The et al. 1994b), it is now clear that Herbig Ae/Be stars possess flared
disks in a Keplerian orbit - usually with a puffed up inner rim. Additionally,
the study of post-main sequence stars that have evolved from A stars indicate that these disks form giant planets
perhaps even more prodigiously than their lower mass counterparts (Johnson
et al. 2010).
In addition to these tools, near infrared spectroscopy provides valuable
insight to the gas in the inner region of the disk. For example, while this
region ($\lesssim$50~AU) remains difficult to study by direct imaging, the
kinematic profile of resolved emission lines can serve as a surrogate for
high resolution imaging by providing spatial information about the gas
on milliarcsecond scales assuming the gas is in a Keplerian orbit and that
the stellar mass and disk inclination are known. Ground based high resolution
(R$\gtrsim$20,000) spectroscopy of ro-vibrational CO lines have been
used to measure the distribution of molecular gas relative to dust in the
inner disk. This technique was pioneered in the study of classical T Tauri
stars and embedded YSOs (e.g. Najita et al. 2000, 2007) and has been
applied extensively to gas in disks around Herbig Ae/Be Stars (e.g.
Brittain \& Rettig 2002; Brittain et al. 2003, 2007; Blake \& Boogert 2004;
Goto et al. 2006; van der Plas et al. 2009, 2010; Salyk et al. 2009;
Brown, J. et al. 2013). The study of ro-vibrational CO lines has also complemented
FIR studies of disks around Herbig Ae/Be stars by adding parameter space to
constrain thermochemical models of disks (e.g. Carmona et al. 2014; Thi et al. 2014).
Spectroscopy of the infrared \ion{H}{1} recombination lines has also
advanced our understanding of Herbig Ae/Be stars. Moderate resolution
spectroscopy combined with interferometry has constrained the size of
the emitting region of \ion{H}{1} to typically within $\sim$0.1~AU (Eisner
et al. 2009; Kraus et al. 2008). A great deal of work has also been done
to calibrate these lines against the accretion luminosity so that they can
be used to measure the accretion rate of Herbig Ae/Bes (van den Ancker
et al. 2005; Garcia Lopez et al. 2006, Donehew \& Brittain 2011; Mendigut{\'{\i}}a
et al. 2011; Pogodin et al. 2012).
A particularly exciting development in the use of NIR spectroscopy to
study gas in disks is the application of spectro-astrometry. Spectro-astrometry is
the measurement of the spatial center of the point spread function (PSF) of
a spectrum as a function of wavelength (Bailey et al. 1998a).
Because variations in the center of the PSF can be measured on much
smaller angular scales than the diffraction limit of an instrument
spectro-astrometry can provide spatial
information on sub-milliarcsecond scales. This technique was initially
developed with the advent of CCDs in the 1980s (see Bailey 1998b for
a review) and has since been applied extensively to extract additional
spatial information from spectra of a number of astrophysical phenomena
such as binaries (Bailey et al. 1998a; Porter et al. 2004), YSO jets (Takami
et al. 2003, Whelan et al. 2004; Davies et al. 2010), and circumstellar disks
(Acke \& van den Ancker 2006; Pontoppidan et al. 2008; Brittain et al. 2013;
Brown, L. et al. 2013).
In this review we discuss how modeling line profiles spectrally and
spatially provide information about the distribution of gas in disks
around Herbig Ae/Be stars. We then highlight two example applications:
identification of signposts of planet formation and the study of the origin
of \ion{H}{1} emission lines in Herbig Ae/Be stars. We follow with a
discussion of sources of potential artifacts that may complicate the
interpretation of spectro-astrometry measurements. We show the scope of these artifacts
and how to acquire the data so as to mitigate their effect. We conclude
with a brief discussion of future prospects for ground based study of gas in disks.
\section{Technique}
\subsection{Line Profile Modeling}
Planet forming disks begin
their lives
with high optical depths at most wavelengths. For the minimum
mass solar nebula, the surface density at 1~AU is 1700~g~cm$^{-2}$ or
about N(H)=10$^{27}$~cm$^{-2}$. For a gas to dust ratio and grain size
distribution representative of the local interstellar medium, only the top
10$^{-5}$ of the disk is optically thin at 5$\mu$m. Grain growth and settling
increase the depth of the optically thin portion of the disk atmosphere,
but even so NIR observations only probe a thin layer of the disk. The line
emission observed in
the
disk arises from this thin warm layer of the atmosphere.
Thus observing H$_2$ can be very challenging because of the weak
oscillator strengths of the quadrupole transitions (Bitner et al. 2007;
Carmona et al. 2011)
and large H$_2$ column densities are needed for detectable emission.
CO has proven to be the most commonly observed
species in disks despite the fact that its abundance is 10$^{-4}$ -- 10$^{-6}$
that of H$_2$. The ro-vibrational transitions of CO are a useful tracer of gas
for a number of reasons. First, the relatively low sublimation temperature and
high dissociation temperature makes it abundant throughout a circumstellar
disk. Second, even relatively small columns of CO are self shielding.
Ro-vibrational transitions are also readily observable from the ground
(moderate Doppler shifts are necessary for the J$\lesssim$11 v=1--0 transitions).
We can trace the distribution of this warm layer of gas in the
disk by modeling the kinematic profile of the lines. If we assume
the gas follows a circular Keplerian orbit and know the disk inclination
and stellar mass, then the line profile can be readily calculated (e.g.
Smak 1981; Figure 1). If the line is resolved, the inner and outer edge
of the gas emission can be calculated. The half width at zero intensity
(HWZI) of the emission line corresponds to the projected velocity of
the gas at the inner edge of the detected region. The half-width of the peak
separation corresponds to the velocity of outer extent of the emitting region.
\begin{figure}[h]
\includegraphics[width=\columnwidth]{figure1.eps}
\caption{Schematic of a rotating disk, emergent spectral line and spectro-
astrometric signal. In the top panel the projected isovelocity contours are
colored. The corresponding parts of the emission line are given the same
color. The half width at zero flux corresponds to the velocity of the gas at the
inner edge of the disk. The half-width of the peak separation corresponds to
the velocity of outer extent of the emitting region. When the slit of the
spectrograph is aligned with the semi-major axis of the disk (the rectangle
superimposed on the schematic), the blue shifted gas is spatially offset
toward one end of the disk and the red-shifted gas is spatially offset toward
the other end of the slit. The magnitude of the offset depends on the spatial
extent of the gas, the radial distribution of the disk flux, and the relative
brightness of the continuum. }
\label{fig:1}
\end{figure}
Generally the data are modeled by assuming the gas distribution is axisymmetric and follows circular
Keplerian orbits, however, this may not accurately describe gas in circumstellar
disks. For example, the gas in orbit may be elliptical giving rise to asymmetric
line profiles (Reg{\'a}ly et al. 2010) or the gas may arise in a non-Keplerian orbit such
as a disk wind (Carr 2007; Pontoppidan et al. 2011).
\subsection{Spectroastrometry}
\label{sec:Observations}
Using a relatively new technique,
spectroastrometry, one can test the assumption that the gas arises in
a Keplerian orbit.
Spectro-astrometry has been described in a number of papers (e.g. Bailey 1998ab;
Porter et al. 2004; Brannigan et al. 2006; Whelan \& Garcia 2008),
so we limit our discussion to a brief overview of this technique - particularly as it
applies to circumstellar disks.
While the spatial resolution of a PSF is limited to $\sim$1.2D/$\lambda$,
the centroid of the PSF can be measured to a small fraction of that. For a well sampled Gaussian
PSF dominated by photon noise, the center of the Gaussian can be measured to an accuracy of,
\begin{equation}
\delta b \sim 0.4 \frac{FWHM}{SNR}
\end{equation}
\noindent where $b$ is the center of the Gaussian, $FWHM$
is the full width at half maximum power of the Gaussian, and $SNR$ is the
signal to noise ratio of the PSF (Brannigan et al. 2006). With a SNR of 100 and an AO corrected beam of 0$\arcsec$.12,
the centroid of the PSF can be measured to an accuracy of 0.5 milliarcseconds
assuming the PSF is adequately sampled. The spectro-astrometric
signal is not a direct measure of the extent of the gas, rather it is the projection of the center of
light of a given velocity channel along the axis of the slit.
Typically, the
continuum is much brighter
than the emission lines, so the continuum dominates the center of light.
The extent of the offset depends on the angular extent
of the emission, the asymmetry of the distribution of the emission
along the axis of the slit, the brightness of the emission relative to the continuum,
and the relative position
angle of the disk and the slit. For example, if the slit is aligned with the semi-major axis of an
inclined axisymmetric disk, gas in a Keplerian orbit will show emission lines where the red shifted
side of the line is spatially offset toward one end of the slit while the blue shifted side of the emission line
is offset toward the opposite end of the slit (figure 1). If the slit is rotated so that it is aligned with the semi-minor axis
of the disk then the center of light of the red and blue shifted gas will be aligned with the center of the PSF and
no offset will be measured.
This technique has provided
valuable information about gas in disks. For example, high resolution spectra of the
[\ion{O}{1}]$\lambda$~6300~\AA\ lines have been
analyzed spectro-astrometrically (Acke \& van den Ancker 2006). These authors
show that the line is consistent with gas in Keplerian orbit
about Herbig Ae/Be stars rather than arising in an outflow as
in the case of T Tauri stars. In contrast, spectro-astrometric
analysis of high resolution spectra of ro-vibrational CO emission
from strongly accreting YSOs shows that the gas is less extended
than the narrow profiles indicate if the gas were in a Keplerian
orbit (Pontoppidan et al. 2011; Bast et al. 2012).
In the section that follows we describe two applications of spectral
profile fitting and SA to elucidate aspects of disks around Herbig Ae/Be stars.
\section{Applications}
\subsection{Signposts of Planet Formation}
The identification of gas giant planets forming in
circumstellar disks will represent a major
milestone in the study of planet formation.
By identifying forming planets we can connect the initial conditions
in disks to the kinds of planets that form. We can also determine
the disk radii at which planets form for comparison with
the orbital radii of planets observed in more mature systems.
Such observations can also test theoretical predictions that
planets clear gaps in disks and accrete mass through circumplanetary
disks, processes that affect the evolution of planetary masses.
This approach is complementary to the direct detection of
forming gas giant planets.
Direct detection is challenging due to the
small angular separation between the forming gas giant planet and
the host star as well as the high
contrast between the forming planet and the disk.
However, several exciting candidate forming objects have been identified
by direct imaging (e.g., T~Cha Hu\'{e}lamo et al. 2011; LkCa~15 Kraus \& Ireland 2012;
HD~100546 Quanz et al. 2013).
\begin{figure}[h]
\includegraphics[width=\columnwidth]{figure2.eps}
\caption{Synthetic line profile from gas arising from an eccentric annulus. The
line profile of a line arising from an annulus is plotted. The asymmetry of the
line changes with the eccentricity of the disk. The asymmetry arises from two
effects. First, the gas near periastron is moving faster than gas near apastron.
Second, the gas at periastron is closer to the star and thus brighter than the more
distant gas. Even if the line is not resolved, this effect can be inferred from the
shift of the centroid of the line relative to the systemic velocity of the system. }
\label{fig:2}
\end{figure}
Another approach to identifying forming gas giant
planets is to identify dynamical markers of tidal interactions between the disk and
forming object. Hydrodynamic modeling of planet disk interaction indicates that a
planet with a mass greater than $\sim$3M$_{\rm Jupiter}$ on a circular orbit can induce an
eccentricity in the disk in the vicinity of its orbit (e.g. Kley \& Dirksen 2006).
Simulations show that the eccentricity can be as high as 0.25 and falls off
as approximately r$^{-2}$. The profile of a line arising from an eccentric annulus will generally not be
symmetric. The gas at periastron will be faster than the gas at apastron and
generally the temperatures will differ as well. This results in a distinctive asymmetric
line profile (figure 2; see Reg{\'a}ly et al. 2010 and Liskowsky et al. 2012). Interestingly,
the semi-major axis of the disk precesses very slowly ($\sim$10${\degr}$
per 1000 orbits), so the line profile
should remain constant over decadal timescales. Such a line profile
has been observed in OH spectra from the Herbig Ae binary V380~Ori (Fedele et al. 2011)
and the Herbig Be star HD~100546 (Liskowsky et al. 2012).
In the case of V380~Ori, it is likely a stellar companion that is driving the
eccentricity whereas in the case of HD~100546 it is likely that a
substellar object is driving the eccentricity. In principle, an asymmetric line profile could
also arise from the uneven sampling of a resolved axisymmetric disk
(e.g. Hein Bertelsen et al. 2014) which we will discuss in section 3.
\begin{figure*}[!t]
\begin{center}
\includegraphics[width=6.5in]{figure3.eps}
\end{center}
\caption{Evolution of CO emission line profile and spectro-astrometric signal (Brittain et al. 2014).
In the first column the v=1--0 P26 CO line observed in 2003 is plotted (black)
and compared to the same line observed in 2006 (red), 2010 (green), and
2013 (cyan). The Doppler shift of the excess emission can be seen in this column.
In the second column, the spectro-astrometric signal is plotted
for each epoch along with the model (dot-dashed line) and the residual.
The spectro-astrometric signal is modeled by adding an
extra source of emission in the inner edge of the disk at the position
corresponding to the Doppler shift of the excess (illustrated in the third
column). While the inner edge of the disk is roughly 130$mas$ from the
star, the center of light of the CO emission is offset from the center of the
PSF $\pm$15$mas$ along the axis of the slit. The extra source of emission
in orbit around the star causes the center of light to shift. This source of emission
was hidden by the circumstellar disk in 2003. }
\label{fig:3}
\end{figure*}
Another way to identify ongoing planet formation is by detecting
emission arising from a circumplanetary disk feeding the planet.
Circumplanetary disks can be discovered
by their line emission (Brittain et al.\ 2013; 2014)
and also potentially by their dust continuum emission
(Zhu 2014; Isella et al.\ 2014).
Simulations of gas giant formation
indicate that the disk feeding the gas
giant planet will have a radius roughly one-third of the Hill Sphere
(Quillen \& Trilling 1998; Ayliffe \& Bate 2009ab; Martin \& Lubow 2011),
though the full accretion envelope is likely intrinsically three dimensional with complicated flow patterns
(e.g. Tanigawa et al. 2012; Ayliffe \& Bate 2012; Gressel et al. 2013). Thus a 5M$\rm_{Jupiter}$
planet forming at an orbital radius of 10~AU should have a
circumplanetary disk with a radius of 0.3~AU - comparable to the size of the
emitting region of ro-vibrational CO emission observed toward classical T Tauri stars (Najita et al. 2003).
In the case of Herbig Ae/Be
stars with transition disks, hot bands of ro-vibrational
CO emission are observed from R$\gtrsim$10~AU (e.g. HD~179213, van der Plas 2010;
HD~100546, Brittain et al. 2009; HD~141569 Brittain et al.
2003; HD~97048, van der Plas et al. 2009; and Oph~IRS~48,
Brown et al. 2012) due to the UV fluorescence of the gas. To detect the
IR ro-vibrational emission lines that result from this process, a large
area is needed because these transitions are exceeding optically thin.
For example, the line flux of CO emission arising from HD~100546 is
comparable to that observed from classical T Tauri stars in Taurus
(compare for example, Brittain et al. 2009 and Najita et al. 2003).
However, the emitting region of the CO observed toward HD~100546
extends over 50~AU while the emission from the classical T Tauri
stars typically originates within 1~AU of the star. Thus UV fluoresced gas
originating from a disk with a radius less than 1~AU will reveal
ro-vibrational CO lines with luminosities 3-4 orders of magnitude fainter
than what is observed from the transition disks mentioned above.
The contrast between the gas conditions
which produce UV fluorescence (broadly distributed, perhaps tenuous, gas) and collisionally
excited emission (warm, dense material) provides a way to identify
circumplanetary disks. The surface density and temperature of the circumplanetary disk is expected to be
enhanced relative to the feeding material (Klahr \& Kley 2006; Matin \& Lubow 2011). If the gas
is of order $10^3$~K and optically thick, then the flux of the emission from the circumplanetary disk will be
similar to what is observed from classical T Tauri stars.
A Herbig Ae/Be star with a transition disk created by a
forming gas giant planet should then give rise to
v=1--0 lines that are a blend of emission from the circumstellar disk and
circumplanetary disk while the hotband transitions will be overwhelmingly
dominated by the UV fluoresced gas in the circumstellar disk. Thus as
the circumplanetary disk orbits the star, the shape of the v=1--0
lines will change with the motion of the circumplanetary disk. The hotband
lines provide a constant fiducial profile against which to
measure the evolution of the v=1--0 lines. This behavior, variable v=1--0
profiles with constant hotband profiles, is observed in
HD~100546 (Brittain et al. 2013; 2014).
In January 2006 the red side of the v=1--0 P26 line at 0--10~km~s$^{-1}$
brightened relative to what was observed in January 2003 (second row of figure 3). The Doppler shift
of this excess emission was 6$\pm$1~km~s$^{-1}$. In December 2010, the v=1--0 P26 was still
brighter than it was in January 2003, however, the Doppler shift was --1$\pm$1~km~s$^{-1}$ (third row of figure 3).
In March 2013 the excess emission was Doppler shifted --6$\pm$1~km~s$^{-1}$. Assuming the source
of the excess emission is in a Keplerian orbit, the projected Doppler shifts and epochs of the observations
place the emission near the inner edge of the outer disk (R$\sim$13~AU). One way
to test whether the shift of the excess v=1--0 line emission is due to gas in a
Keplerian orbit and determine the radial position of the emission is to use spectro-astrometry.
Consider an inclined axisymmetric disk of gas
in a circular Keplerian orbit. If the slit is aligned
near the semi-major axis of the inclined disk, then the blue side of the
emission lines formed in the disk will be offset in one direction along
the slit axis and vice versa for the red side of the line, and the offsets
would be symmetric (figure 1). This was observed for the v=1--0 P26 line in 2003
(top row of figure 3). But what happens
if we add an extra localized source of emission at a given velocity
to our disk? In this case, the center of light at the velocity of the extra source
will be shifted closer to the added source of emission.
The extent and direction of the
shift will depend on the location of the center of light along the slit axis.
In the case of HD~100546, the first three observations were taken with
the slit in its default position angle of 90$\degr$ (Brittain et al. 2013).
The position angle of the disk is 140$\degr$ (Ardila et al. 2007). In 2006
not only was the red side of the line brighter
relative to its brightness in 2003, the center of the PSF of the red side
of the line was less extended in 2006 than it was in 2003 (second row of figure 3).
As it turns out, for the combination of the slit PA and disk PA,
an excess source of emission that orbits close to the inner disk edge of the outer disk (R$\sim$13~AU)
with a projected velocity of 6~km~s$^{-1}$ will be located near
the center of the slit, just as required to explain the spectro-astrometric signal
(second row of figure 3). In 2010, the
spectral profile of the extended emission
was broader and shifted to $-1\pm$1~km~s$^{-1}$. Placing this emission
in the disk near the inner rim at the orbital position corresponding to this shift
reproduces the spectro-astrometric signal (third row of figure 3). The same is
observed for the observations acquired in 2013. From 2006 to 2010 to 2013, the Doppler
shift of the emission, location of the emission projected along the slit axis,
and period of the emission are all consistent with a source of thermal CO
emission in a circular Keplerian orbit near 13~AU (Brittain et al. 2014).
\subsection{The Origin of \ion{H}{1} lines}
A second important application of spectro-astrometry to the study of Herbig
Ae/Be stars is the study of the origin of the \ion{H}{1} recombination lines.
The origin of these lines remains uncertain (Kraus et al. 2008):
do they form in the stellar wind, disk wind, decretion disk, accretion flow,
or some combination thereof? If the \ion{H}{1} lines are connected
to accretion, then they may
serve as a useful proxy of the stellar accretion rate (e.g. Garcia Lopez et al.
2006). Several studies have
attempted to calibrate various emission line diagnostics to
the accretion luminosity (e.g. Donehew \& Brittain 2011; Mendigut{\'{\i}}a et al.
2011, 2012, 2013; Oudmaijer et al. 2011; Pogodin et al. 2012; see also
Salyk et al. 2013). Typically, these studies convert the veiling of
the Balmer discontinuity to an accretion luminosity using a magnetospheric
accretion model (e.g.
Muzerolle et al. 2004) and compare this to the luminosity of various emission
lines.
For example, Donehew \& Brittain (2011) apply this method and
find that the relationship between the luminosity of Br$\gamma$ and the
accretion luminosity determined for Herbig Ae stars is consistent with the
relationship found for classical T Tauri
stars (Muzerolle et al. 1998) and intermediate mass T Tauri stars (Calvet et
al. 2004); however, the Herbig Be stars
do not follow the same trend.
\begin{figure*}[!ht]
\includegraphics[width=6.5in]{figure4.eps}
\caption{Spectro-astrometric measurement of Pa$\beta$ (left column) and
Br$\gamma$ (right column) with NIFS on Gemini North. The spectral profile
of the line is plotted (black) above the spectro-astrometric signals. The
red lines correspond to the centroid measurements made along the slitlet (the x-direction) and the
blue lines refer to the centroid measurements made perpendicular to the slitlet (the y-direction). No significant
signal is detected for the HAe stars, but a signal is detected for the HBe stars.
Modeling the line profiles and signals is an ongoing project (Adams et al. 2015 - in prep.).}
\label{fig:4}
\end{figure*}
Whether the correlation between the luminosity of
the
\ion{H}{1} emission lines
and
the inferred accretion luminosity is physical or fortuitous
remains to be seen.
One way to
investigate this issue
is to measure the variability of the veiling
of the Balmer discontinuity and emission line luminosities
simultaneously (Mendigut{\'{\i}}a et al. 2013). With the advent of broad coverage
spectrographs
such as X-Shooter on the VLT (covering $\sim0.3-2.4\micron$), it is now
possible to simultaneously observe the
veiling of the Balmer discontinuity and the emission line
diagnostics that fall within this range (Oudmaijer et al. 2011; Mendigutia et al. 2014).
Ongoing studies along these lines will be
able to determine whether or not the veiling and line emission are
connected. However, they will not be able to establish that Herbig Ae/Be
stars accrete magnetospherically - as assumed in the calculation of
the accretion luminosity - or identify where the \ion{H}{1} emission lines originate.
At first blush the assumption that Herbig Ae stars accrete
magnetospherically would seem unjustified. Herbig Ae/Be
stars are not fully convective thus one would not expect
that they would generate the strong kilogauss magnetic fields
characteristic of their lower mass analogs (Johns-Krull 2007).
Indeed, studies of the magnetic properties of Herbig~Ae
stars indicate that they do not possess strong well ordered
surface fields (Alecian et al. 2014). On the other
hand red shifted hydrogen lines are observed in Herbig~Ae
stars with velocities of several hundred km s$^{-1}$
indicative of material in free fall onto the star (the tell tale
signature of magnetospheric accretion;
Guimar{\~a}es et al. 2006). Thus the assumption that at least
some Herbig Ae/Be stars accrete magnetospherically is
not without warrant.
\begin{figure*}[!b]
\includegraphics[width=6.5in]{figure5.eps}
\caption{Illustration of artifacts resulting from pointing errors. In row A, the
isovelocity contours are plotted in red. Also plotted are the same contours
convolved with two dimensional Gaussian profile with a full width at half max
equal to 0$\arcsec$.6. The portion of the disk sampled by the slit on the
Phoenix spectrograph is marked with green dot-dashed lines. In row B, the
resultant line profile arising from the disk is plotted. In the first panel, the PSF
of the disk is centered in the slit. In the next the center of the PSF is offset by
one pixel, and so on. In row C, the spectro-astrometric measurement of the
PSF is plotted. Each panel corresponds to the offset noted in the row above.
When the disk emission is resolved and the PSF is not properly centered in the
slit, the profile is asymmetric. This has a minor effect on the line profile and
spectro-astrometric signal unless the offset is substantial. Misaligning the center
of the PSF and slit results in about a 20\% error when offset is half a slit width
and 50\% error when the PSF is offset a full slit width. With care the center of the
PSF can be measured and placed in the center of the slit with an accuracy of about
0.5 pix. }
\label{fig:5}
\end{figure*}
Before the recombination \ion{H}{1} lines can be used reliably as a proxy for
the stellar accretion rate, it is crucial that we understand the
physical origin of these lines. One promising development is the use of
spectro-astrometric
measurements of \ion{H}{1} emission using integral field spectrographs. This
method has recently been applied to massive YSOs. Davies et al. (2010)
measured
Br$\gamma$ emission from the massive YSO W33A with NIFS on Gemini
North. They
were able to measure the centroid of the PSF with an accuracy of 0.1 ~$milli
$arcseconds and show that the line originates in an outflow.
Application of the same method to Herbig Ae/Be stars is beginning to show
promising results. For example
(Adams et al. 2015), show they can achieve a root
mean square fidelity of 100$\mu$arcseconds
with their observation of Herbig Ae/Be stars. This is within a factor of a few
of being able to distinguish emission from a magnetosphere, outflow, and
gas in a Keplerian
orbit in the inner 0.1AU. A study of four Herbig Ae/Be stars finds SA signatures
from \ion{H}{1}
from the HBe stars but not the HAe stars in the sample (Figure 4).
By combining these data we are
also able to show that our fidelity is limited by photon noise. Increased
integration time should
provide a means to improve the sensitivity of these measurements by a factor
of a few and
thus discriminate between a wind, disk, or funnel flow origin for the nearest
Herbig Ae/Be stars.
Making such high precision measurements requires attention
to potential artifacts in the data.
\section{Pitfalls}
Several authors have summarized best practices
for using SA to study
phenomena such as binaries and outflows
(Bailey et al. 1998; Baines et al. 2004; Porter et al. 2004; Brannigan et al.
2006; Whelan \& Garcia 2008). Because SA involves making high
precision measurements of the center of the PSF, even small artifacts
can lead to spurious signals. We summarize the principal best practices
extracted from the work of these authors:
\begin{enumerate}
\item Observe the source twice at each desired PA - once at the original
position and then with the slit rotated 180${\degr}$. Instrumental
artifacts should rotate with the instrument. If they are reproducible,
then they should cancel when the signals are combined.
\item Observe a standard star as a secondary check for instrumental artifacts
\item Acquire very high signal to noise ratio flats every time the grating is
moved. If the grating in the spectrograph moves when the telescope slews,
it may be necessary to retake flats with each observation. Small errors in
the flat-fielding can shift the center of the PSF and result in spurious signals.
\item The slit width should be narrow compared to the PSF. If the PSF
does not fill the slit and the PSF is not perfectly symmetric, it can result
in a spurious signal that will not be removed by combining with the SA
signal acquired with the instrument rotated 180${\degr}$.
\item Acquire spectra over a wide spectral range. It is crucial that the
continuum be well sampled to ensure a baseline measurement is
available for comparison to the region covered by the spectral line.
This can prove to be challenging when observing strong \ion{H}{1}
lines from Herbig Ae/Be stars with high resolution spectrographs
as the the full width at zero intensity of the line can be comparable
to the spectral grasp of the order.
\end{enumerate}
\begin{figure}
\includegraphics[width=2.75in]{figure6.eps}
\caption{Illustration of the effect of an asymmetric PSF. In panel A, the PSF
of a star acquired with PHOENIX on Gemini South is plotted (red contours).
A synthetic, elongated profile is also plotted (black contours). An emission line was
synthesized assuming it arose from gas in a disk in a Keplerian circular orbit. In panel B the
resultant line profiles that resulted from the PSF plotted in panel A are plotted. In panel C
the spectro-astrometric signal
is similarly plotted. While the focus of the PSF has only a minor effect on the spectral line
profile, the effect on the measurement of the spectro-astrometric signal is significant. Proper
focusing of the source is important for acquiring reliable signals. }
\label{fig:6}
\end{figure}
In what follows, we highlight two additional issues that may complicate
the interpretation of spectro-astrometric measurements. First we
discuss the challenges of analyzing spectro-astrometric signals
from resolved sources and second we explore the challenges of
making SA measurements from spectra acquired with an IFS.
When observing molecular emission from nearby disks, it is possible that the
emission will be spatially resolved. For example, the molecular NIR emission
observed from HD~100546 extends from $\sim$10$\pm2$~AU to beyond
50~AU (van der Plas et al. 2009; Brittain et al. 2009; Carmona et al. 2011;
Liskowsky et al. 2012). In addition there is a
compact ring of solids that extends from 0.25~AU to less than 0.7~AU that
dominates
the continuum in the NIR (Benisty et al. 2010; Mulders et al. 2011, 2013; Pani{\'c} et al.
2014).
At the distance to HD~100546 (97$\pm$4~AU; van Leeuwan 2007), the inner
hole
of the disk subtends 0$ \arcsec .28\pm0\arcsec.03$ along the semimajor axis.
Adopting
an inclination of 42$\degr$ (Ardila et al. 2007; Pineda et al. 2014) indicates
that the
inner hole subtends 0$\arcsec.19\pm0\arcsec.03$ along the minor axis of
the disk.
Spectroscopic observations of
molecular lines from the disk using adaptive optics and $0\arcsec.2$ slit will
result in the occultation of a significant fraction of the disk. Small changes in the
positioning of the slit can give rise to varying line shapes (Hein Bertelsen et al. 2014).
Reconstruction of the line profile representing
the underlying structure of the disk thus requires precise knowledge of the slit position. Even
pointing uncertainties of order $0\arcsec$.05 can undermine the interpretation
of the spectrum. In addition, since the continuum is unresolved, the PSF of the
continuum and the PSF of the emission lines is not identical. This can give
rise to artifacts as well since the PSF does not fill the slit.
In such cases it can be advantageous to work at lower spatial resolution and
wider slit widths.
Here we compare the effect of slight pointing offsets on the shape of the
spectral lines to the effects
described by Hein Bertelsen et al. (2014). We consider conditions with
seeing of 0$\arcsec$.6 and
a four pixel slit with a width of 0$\arcsec$.34 (the slit width of Phoenix on
Gemini South; \citealt{2003SPIE.4834..353H,2000SPIE.4008..720H,1998SPIE.3354..810H}).
If the slit is offset from the center of the PSF, an asymmetric PSF and spectro-astrometric signal does result
(figure 5); however, to achieve
a deviation of 20\% from the original measurement, the PSF of the continuum
must be shifted
as much as
2 pixels from the center of the slit.
With care the star can be centered on the slit to within $\pm$0.5pixels
(about $\pm$0$\arcsec$.05 for Phoenix on Gemini South and CRIRES on the VLT).
While such an offset can be significant when using adaptive optics (a PSF
FWHM$\sim 0\arcsec.15$) and $0\arcsec.2$ slit (Hein Bertelsen et al. 2014)
it is much less significant when the seeing is $0\arcsec.6$.
An additional concern arises from making measurements of
spectro-astrometric signals from an uncorrected PSF. For example, an
asymmetric PSF can give rise to a spurious spectro-astrometric signal. We modeled
the effect of a poorly focused beam convolving our disk model with an observed
PSF and an asymmetric PSF (figure 6a). The line profile shows only a modest effect
due to the occultation of part of the disk (figure 6b). The spectro-astrometric signal, however,
shows a significant asymmetry (figure 6c).
Perhaps of greater concern is that rotating the instrument will not remove artifacts
arising from this effect if it is not steady (say due to variable seeing).
Imaging the PSF throughout the observations is a helpful
way to identify problems with a variable PSF.
\section{Conclusions}
\label{sec:conclusions}
Spectroscopy has been used to elucidate many aspects of Herbig Ae/Be
stars since the previous conference on Herbig Ae/Be stars held in 1994; however, there is much exciting work
left to be done. Over the past two decades, we have moved from detecting and characterizing line emission arising
from circumstellar material to using these features to elucidate the star-disk connection and identify signposts
of forming gas giant planets. The application of spectro-astrometric measurement to the study of disks is still in its
infancy. While this technique shows remarkable promise for revealing sub-$milli$arcsecond spatial information
from spectra, attention is needed to avoid several pitfalls that can complicate the interpretation of such data. The
detection of companions with orbits of order $\sim$10~AU by measuring the Doppler shift of emission arising
from the circumplanetary disk requires observations on decadal timescales. Archival data is thus a crucial resource
facilitating the long term monitoring of these systems.
\acknowledgments
S.D.B. acknowledges support for this work from the National Science Foundation under grant number AST-0954811. Basic research in infrared astronomy at the Naval Research Laboratory is supported by 6.1 base funding.
|
\section{Introduction}
Water is special fluid for its biological relevance and technological applications but most intriguing is that it presents thermodynamic and dynamic properties properties with anomalous (or unusual) behavior~\cite{Franks:water:matrix}. The origin of its anomalous properties is actively discussed in the literature, with different thermodynamic scenarios competing to describe its behavior on regular and metastable regimes~\cite{debenedetti03:review, malenkov2009:jpcm:review}. Among alternative views on water thermodynamics it should be relevant to mention the second critical point hypothesis~\cite{poole92:nat} and the singularity free scenario~\cite{sastry96:pre,sastry98:jcp}, which will be relevant in the context of the current work.
The second critical point hypothesis was first proposed by Poole {\it et al.} based on computer simulations of atomistically detailed water models, and relates the observed divergence on water's thermodynamic response functions to the critical end point of a deeply metastable liquid-liquid phase transition~\cite{poole92:nat}. Singularity free scenario was proposed by Sastry~\textit{et al.} using lattice models for liquid water which supported the idea that waterlike anomalies indeed exist, at least in lattice systems, without a liquid-liquid coexistence~\cite{sastry96:pre}. In some simplified models of fluid it was possible to obtain both second critical point and singularity free scenario by adjusting some physical parameters of the system~\cite{giancarlo03:pre,heckmann2013:jcp}
For instance, in the model proposed by Franzese~\textit{et al.}~\cite{giancarlo03:pre} it is possible to observe the liquid-liquid coexistence to shrink and the temperature of second critical point decrease to $T_c \rightarrow 0$ by decreasing the correlation between hydrogen bonds inside within individual molecules.
Lattice models of fluid have been extensively used to investigate the above mentioned anomalous properties of water due to possibility of obtaining analytical or numerical results, while exploring a wide range of physical parameters. In this direction, both thermodynamics~\cite{barbosa08,marcia05:jcp, giancarlo03:pre, sastry96:pre,sastry98:jcp, heckmann2013:jcp} and kinetics~\cite{szortyka10:jcp, giancarlo08:prl} where investigated in lattice models with waterlike behavior.
Nevertheless, approximations employed in two and three dimensions(3D), and even some exact solutions in one dimension (1D), tend to generate complex sets of equations whose analyses is often performed numerically. Thus, it should desired to design models for which one could obtain simple analytical expressions connecting thermodynamic anomalous behavior to phase transitions and critical behavior.
To achieved this goal we previously investigate 1D lattice models with pair interaction between the first neighboring molecules, with interactions spanning two~\cite{barbosa11:jcp} and three lattice sites~\cite{barbosa13:pre}. While in Ref.~\cite{barbosa11:jcp} both van der Waals and hydrogen bond like interactions were used, resulting in a line of temperature of maximum density associated to a ground state phase transition (GSPT), in Ref.~\cite{barbosa13:pre} it was proposed a core-softened fluid with pair interactions up to three sites, resulting in two temperature of maximum density lines associated to two GSPT. Besides obtaining exact results, in the latter work we used an analytical approximation in the neighborhood of the critical point to obtain a simple expression for Gibbs free energy, and used it to mathematically study the relation between anomalous density behavior and GSPT.
In this work we proceed on this direction by investigating the repulsive 1D lattice gas, which is even simpler than our previous models and presents waterlike anomalies in density, thermodynamic response functions and self diffusion constant. The model was studied through transfer matrix technique, the Takahashi method (within a two state approximation, as will be discussed latter), and Monte Carlo simulations. With the results obtained from these techniques a connection between temperature of maximum density and GSTP was found as in a previous work with more complex models~\cite{barbosa13:pre}. In addition, it was also found that GSPT does present a residual entropy, due to phase mixing, and it is shown that this property is fundamental in determining waterlike anomalies for the model considered here. Finally, a comparison between regions with density and diffusion anomaly indicated that this model presents so called \textit{hierarchy of anomalies}~\cite{debenedetti01:nat}.
\section{\label{thermo}Thermodynamics}
\subsection{Model and Ground State}
By defining the presence of a particle on site $k$ with an occupation variable $\eta_k = 1$ (and absence with $\eta_k=0$) the effective Hamiltonian of the 1D repulsive lattice gas in the grand canonical ensemble becomes:
\begin{equation}
\label{eq:H}
\mathcal{H} = \sum_i \epsilon{}{\eta{}}_i{\eta{}}_{i+1}-\sum_i \mu{}{\eta{}}_i,
\end{equation}
where $\epsilon>0$ is the strength of the (repulsive) interaction and $\mu$ is the chemical potential. Note that this Hamiltonian can be seen as a special case from our previous model for liquid water~\cite{barbosa11:jcp} by setting $\epsilon_{hb}=0$ and $\epsilon_{vdw}=-\epsilon$.
The ground state of the unidimensional repulsive lattice gas is composed by three phases: a gas in which the lattice is empty, a dense fluid (DF), presenting a completely filled lattice, and a half-filled lattice with each particle separated by holes, here denominated as
softened fluid (SF)\footnote{Notation for this phase has changed from bonded fluid to softened fluid, since our previous work, because there are no~\textit{hydrogen bonded} in the current context.}. In a lattice with $L$ sites separated by a distance $l$, periodic boundary conditions, and $N$ particles, the enthalpy per particle of these phases are $h_\texttt{gas}=0$, $h_{\texttt{DF}}=\epsilon + Pl$ and $h_{\texttt{SF}}=2Pl$, with $P$ meaning pressure. Considering this, the SF is stable in the ground state for pressures $0 < P \leq P_c$ with $P_c l = \epsilon$.
Two ground state phase transitions are present at null temperature and were calculated by equating the enthalpy per particle at different phases: a G-SF at $P=0$ and a SF-DF transition at $P_c$, which will be called ``second critical point'' as in previous works~\cite{stanley99:pre,barbosa11:jcp,barbosa13:pre}.
\subsection{Transfer Matrix Technique}
Thermodynamics of the model proposed here is obtained using transfer matrix technique and will be described shortly. Other derivations using this technique can be found elsewhere~\cite{barbosa10:jcp}. We start from the grand canonical partition function as a trace of a matrix
\begin{equation}
\Xi{}\left(T,L,\mu\right)= \sum_{\vec \eta} e^{-\beta \mathcal{H} } = Tr \left \{ \mathcal{P}^L \right \},
\label{equ:traco}
\end{equation}
where $\beta= 1/k_BT$, $\vec \eta = \{ \eta_1, \ldots, \eta_N \}$, and the elements of $\mathcal{P}$ are given by
\begin{equation}
\mathcal{P}_{\eta \eta' }=e^{-\beta{}\left(\epsilon{}{\eta{}}{\eta{}}'-\
\mu{}{\eta{}}\right)}.
\label{eq:matriz}
\end{equation}
In the thermodynamic limit only the largest eigenvalue of $\mathcal{P}$ will contribute to the partition function:
\begin{equation}
\Xi{}\left(T,L,\mu\right)={\lambda{}}^L=e^{\beta{}P l L}.
\end{equation}
The eigenvalue $\lambda$ is related to the fugacity $z=e^{\beta \mu}$ via characteristic function of $\mathcal{P}$,
\begin{equation}
\left(az-\lambda{}\right)\left(1-\lambda{}\right)-z=0, \label{eq:z}
\end{equation}
with $a = e^{-\beta \epsilon}$.
Next we introduce reduced variables $t=k_BT/\epsilon$, $p=Pl/\epsilon$, $g = \mu / \epsilon$, $v = V/Nl$ and $s = S/{Nk_B}$. Since the equation of state~(\ref{eq:z}) connects the Gibbs free energy $g$ to thermodynamic variables $t$ and $p$, it is possible to obtain the `volume' $v$ and entropy $s$ per particle as
\begin{equation}
v = \left(\frac{\partial g}{\partial p}\right)_p, \hspace{0.4cm} s = -{\left(\frac{\partial g}{\partial t}\right)}_p. \nonumber
\end{equation}
With these definitions, density can be calculated as
\begin{equation}
\label{eq:density}
\rho = \frac{1}{v} = \frac{\left(1-\lambda{}\right)[a\left(1-\lambda{}\right)-1]}{a{(1-\lambda{})}^2-1+2\lambda{}},
\end{equation}
and entropy becomes
\begin{equation}
\label{eq:entropy}
s = -\ln{\left(z\right)}+\frac{p}{t} + \frac{za[\left(1-\lambda\right)+p\lambda]-p \lambda^2}{zt \left [a\left(1-\lambda\right)-1 \right ]} .
\end{equation}
Thermodynamic response functions, such as thermal expansion coefficient $\alpha$, isothermal compressibility $k_T$, and isobaric heat capacity $c_P$ are calculated from standard definitions~\cite{salinas:introduction} which, within the convention adopted here, reads
\begin{subequations}
\begin{eqnarray}
\alpha v & = & \left ( \frac{\partial v}{\partial t} \right )_p,\label{eq:alphav}\\
k_t v & = & -\left ( \frac{\partial v}{\partial p} \right )_t\\
c_p & = & t\left ( \frac{\partial s}{\partial t} \right )_p. \label{eq:cp}
\end{eqnarray}
\end{subequations}
Explicit (and exact) expressions for these functions are too lengthy to be reproduced here. The temperature of maximum density line was calculated from~(\ref{eq:alphav}) by numerically solving $\alpha v=0$.
\begin{figure}
\begin{center}
\includegraphics[scale=0.7]{tmd-pxt.eps}
\caption{Location of temperature of maximum density line ($T_{MD}$) in the $p$ vs. $t$ phase diagram. The ground state phase transition between the softened fluid and dense fluid is indicated with a triangle, while its pressure value is marked with a dotted line. An example of the mixed SF/DF state is shown along the dotted line, used to indicate the critical pressure $p=p_c=1$.}
\label{fig:diagram}
\end{center}
\end{figure}
\begin{figure}
\begin{center}
\includegraphics[scale=0.7]{mult.eps}
\caption{The temperature dependence of (a) density and (b) entropy for pressures equal, slightly below and above the ground state phase transition at $p=p_c=1$.}
\label{fig:dens}
\end{center}
\end{figure}
\subsection{\label{two-state}Two states approximation}
To obtain further theoretical insight about this system, it is interesting to investigate its entropy in the neighborhood of the second critical point using a two state approximation based on a lattice version of the Takahashi method, as employed in Ref.~\cite{barbosa13:pre}. In that work the Gibbs free energy per particle was approximated near the GSPT by realizing that on this region the probability of finding local states different from those of the coexisting structures was vanishing small. In the current model these states would correspond to softened and dense fluids, whose ground state local molecular spacing are $r_S=2r_D=2l$. Thus, the Gibbs free energy (non-reduced units) becomes approximately:
\begin{equation}
G/N \approx h_c -k_B T \ln \left [ e^{-\beta (P-P_c)r_S} + e^{-\beta (P-P_c)r_D}\right ],
\label{eq:g-pc}
\end{equation}
where $h_c$ is a constant identical to the local microscopic enthalpy at the GSPT (see Ref.~\cite{barbosa13:pre}). Within the notation adopted here, the Gibbs free energy on Eq.~(\ref{eq:g-pc}) reads
\begin{equation}
g = \frac{h_c}{\epsilon} + \frac{3}{2}\Delta p - t \ln \left [ 2 \cosh \left ( \frac{\Delta p}{2t} \right ) \right ],
\label{eq:g-two-states}
\end{equation}
with $\Delta p = p-p_c$. From the latter, molecular volume and entropy becomes
\begin{equation}
v = \frac{3}{2}-\frac{1}{2}\tanh \left( \frac{\Delta p}{2t} \right),
\label{eq:v-2s}
\end{equation}
and
\begin{equation}
s = \ln \left [ 2 \cosh \left( \frac{\Delta p}{2t} \right) \right ] - \left ( \frac{\Delta p}{2t} \right) \tanh \left ( \frac{\Delta p}{2t} \right).
\label{eq:entropy-2s}
\end{equation}
It is possible to see that by approaching the GSPT at constant pressure, i.e., $(t \rightarrow 0, p = p_c = 1 )$, one obtains the residual entropy $s \rightarrow \ln 2$, discussed before. For any other route used to approach the ground state, entropy results in a null value, indicating that in this model residual entropy only occurs exactly at the GSPT. Molecular volume also depends on the path along which criticality is approached, resulting in
\begin{subequations}
\begin{eqnarray}
\lim_{p \rightarrow p_c } \lim_{t \rightarrow 0} v & = & \left \{
\begin{array}{lc}
2, & \qquad p \rightarrow p_c^- \\
1, & \qquad p \rightarrow p_c^+ \label{eq:v-lowt-1}
\end{array}
\right . \\
\lim_{t \rightarrow 0 } \lim_{p \rightarrow p_c} v & = & \frac{3}{2}. \label{eq:v-lowt-2}
\end{eqnarray}
\end{subequations}
It is important to investigate the critical behavior of the system near the GSPT, by calculating analytical expressions for thermodynamic response functions~(\ref{eq:alphav})-(\ref{eq:cp}), within the two-states approximation adopted here. It is convenient to define $f(x)=1-\tanh^2x$, with $x=\Delta p/2t$, such that
\begin{subequations}
\begin{eqnarray}
\alpha v & = & \Delta p^{-1} x^2 f(x) ,\label{eq:alphav-2s}\\
k_t v & = & \frac{f(x)}{4t},\label{eq:ktv-2s}\\
c_p & = & x^2f \left (x \right) . \label{eq:cp-2s}
\end{eqnarray}
\end{subequations}
Since $0 \leq f(x) \leq 1$, thermodynamic response functions diverge toward criticality as $k_t \sim t^{-1}$, $\alpha~\sim {\Delta p}^{-1} x^2$ and $c_p \sim x^2$. The low temperature behavior on response functions is consistent with waterlike behavior and the critical exponents for $k_t$ and $c_p$ are identical to those found in a continuous one dimensional core softened model proposed by Sadr-Lahijany {\emph et al}~\cite{stanley99:pre} in the same context.
\subsection{Fluid Structure}
\begin{figure}
\begin{center}
\includegraphics[scale=0.7]{entrot.eps}
\caption{Entropy as a function of pressure at a fixed temperatures from exact (continuous and dashed lines) and two state approximation (circles). Note that entropy presents a maximum value at low temperatures.}
\label{fig:entro}
\end{center}
\end{figure}
\begin{figure}
\begin{center}
\includegraphics[scale=0.7]{alpha.eps}
\caption{Thermal expansion coefficient, $\alpha$, as a function of pressure at fixed temperatures. At low temperatures, an oscillatory behavior is observed in the neighborhood of the ground state critical pressure $p_c=1$.}
\label{fig:alpha}
\end{center}
\end{figure}
We start discussing our model's liquid structure from its exact pressure vs. temperature phase diagram, shown in Fig.~\ref{fig:diagram}.
The model presents a GSPT between softened and dense phases at $p=p_c=1$, from which a temperature of maximum density line, $T_{MD}$, emerges separating a region with density anomaly ($\alpha < 0$) from a region with normal density behavior ($\alpha>0$). This line was obtained by numerically solving the equation $\alpha=0$, and was found to extend up to a temperature $t=0.279(1)$, and proceed retracing to lower temperatures ending exactly at pressure $p_{low}=1/2$. An explanation for this precise number can be obtained is as follows: at null temperature, the entropy gain for contracting or expanding the softened fluid by a single site will be same (a lattice defect), thus the system contracts or expands through a competition between fluctuations in energy and volume, which mathematically corresponds to free energy fluctuations of $\delta g_{con} = \epsilon -Pl$ (contraction) or $\delta g_{exp} = +Pl$ (expansion). By equating both variations it follows that $p_{low}= 1/2$.
Another interesting feature of this phase diagram is the mixed SF/DF state, which occurs near the GSPT. This can be observed in Fig.~\ref{fig:dens}~(a), where density is shown as a function of temperature for the critical pressure $p_c=1$ and values slightly below and above $p_c$. Density anomaly is evident for $p<p_c$ while density behaves normally for $p>p_c$. Exactly at $p=p_c=1$ density is kept constant up to $t\approx 0.15$. Density behaves this way because $\alpha \propto (\partial \rho / \partial T )_P \approx 0$ near the $T_{MD}$ line.
The entropy increase with temperature is depicted in Fig.~\ref{fig:dens}~(b), for the pressures investigated in Fig.~\ref{fig:dens}~(a). From this data it is evident that a residual entropy occurs exactly at $p=p_c=1$, while for other pressures ground state entropy start to increase from a null value. It is well known that, due to the Maxwell relation
\begin{equation}
- \left ( \frac{\partial S}{\partial P} \right)_{T,N} = \left ( \frac{\partial V}{\partial T} \right)_{P,N},
\label{eq:Maxwell}
\end{equation}
a density maximum at fixed pressure is mathematically equivalent to an entropy maximum as function of pressure at fixed temperature. Thus, in Fig.~\ref{fig:entro}, entropy is investigated as a function of pressure (at fixed temperatures) using exact and approximate expressions obtained previously, Eqs.~(\ref{eq:entropy}) and~(\ref{eq:entropy-2s}). It is important to note accuracy of the two states approximation near the transition. The relevance of a residual entropy for the appearance of a $T_{MD}$ line is evident since a region of anomalous entropy increase appears at low temperature in the neighborhood of the SF/DF transition. What is interesting in this behavior is that entropy anomaly appears naturally due to residual entropy, whose origin is the two phase mixture in the GSPT. In this way, anomalous entropy behavior (and anomalous density) is intrinsically associated to phase transition in our 1D model since mixing and critical behavior cannot be dissociated when continuous and discontinuous transitions are collapsed in the same point.
Residual entropy also explains the oscillatory behavior observed on thermal expansion coefficient ($\alpha$) as a function of pressure, near the critical transition, Fig.~\ref{fig:alpha}. This behavior was observed previously~\cite{barbosa11:jcp,barbosa13:pre} and its relation with entropy anomaly is simple: below the critical pressure $\alpha$ must be negative due to the fast increase in entropy towards its residual value $\ln2$, and above $p_c$ entropy must fast decrease to a low value, since it is null at $t=0$.
It's worth mentioning that two states approximation also predicts that isothermal compressibility and constant pressure heat capacity increase while lowering temperature, at pressures near the critical one, through Eqs.~(\ref{eq:ktv-2s}) and~(\ref{eq:cp-2s}). Both response functions must present maximum values as a function of temperature, for pressures near the critical one. These results are not shown here but they can be rationalized as follows: the increase in isothermal compressibility should be expected near the GSPT because while increasing pressure in that region the molecular volume is abruptly changing from typically softened fluid to typically dense fluid. On the other hand, the existence of a maximum as a function of temperature in $c_p$, at pressures near the critical value, can be predicted from Eq.~(\ref{eq:cp}) and Fig.~\ref{fig:dens} (b). Its explanation proceeds as follows: let us consider a system in the softened fluid, at $t=0$ and $p<p_c$, where all pair of particles are separated by a hole. In this state, while increasing the temperature entropy will preferentially be increased by allowing neighboring particles to contract. As temperature increases further, an even higher number of particles pairs are allowed to fluctuate between softened and dense fluid states, but the nearest the system is to the critical pressure, the fastest is the approach to an entropic value of $\ln 2$, since the enthalpy difference between the two states is proportional to $\Delta p$. At higher temperatures this mixing mechanism will compete with translational diffusion, i.e., volume expansion, there must be a maximum in $(\partial s/\partial t)_p$ between those temperatures were each mechanism is dominating the entropic behavior of the system.
\section{\label{MC}Monte Carlo}
\begin{figure}
\begin{center}
\includegraphics[scale=0.7]{difusao.eps}
\caption{Self-diffusion constant as a function of density at fixed temperature.}
\label{fig:difu}
\end{center}
\end{figure}
\begin{figure}
\begin{center}
\includegraphics[scale=0.4]{hierarchy.eps}
\caption{The location of the maximum (circles) and minimum (square) values of the self-diffusion constant are
compared to the line of temperature of maximum density (line) in the $t$ vs. $\rho$ phase diagram.}
\label{fig:hierarchy}
\end{center}
\end{figure}
Standard NVT Monte Carlo simulations with Metropolis algorithm were used to calculate the self diffusion constant of our model as a function of density and temperature~\cite{binder,comment2}. The system dynamics proceeds as follows: particles seated on site $k$ were allowed to jump to nearest neighbor ($k\pm 1$) and next nearest neighbor sites ($k\pm 2$). Jumping to next nearest neighbor sites is necessary for particles to diffuse in a 1D lattice, because there is no way for them to turn around their neighboring particles (the same strategy was used in our previous paper~\cite{barbosa11:jcp}. System size was equal to $L=10^3$ and simulation times were equal to or higher than $10^7$ Monte Carlo steps after an equilibrium time of $10^6$ Monte Carlo steps.
As in Refs.~\cite{barbosa11:jcp,marcia07:pa,marcia07:pa2}, self diffusion constant was obtained by linearly adjusting the mean square displacement of particles to the Einstein-Smoluchowski equation for the Brownian motion in one dimension,
\begin{equation}
\left \langle \frac {1}{N} \sum_{k=1}^{N} \left [ X_k(\tau) - X_k(0) \right ]^2 \right \rangle_\mathcal{R} = 2 D (t,\rho) \tau
\end{equation}
where $X_k(\tau)$ is the position of particle $k$ at time $\tau$ and $\mathcal{R} \geq 5$ is the number of different initial conditions used to average square displacements.
Fig.~\ref{fig:difu} exhibits $D(\rho,t)$ for different isotherms. At higher temperatures, as illustrated for $t=0.7$, the diffusion constant is `normal' and monotonically decreases with density. For $t < 0.55$ minimum and maximum diffusion appears, delimiting a region where diffusion anomalously increase with density, as observed experimentally~\cite{angell:jcp76}.
In liquid water, density anomaly is associated to diffusion anomaly through a hierarchy of anomalies, with the former enclosing the latter in the phase diagram~\cite{debenedetti01:nat}. To investigate this hierarchy diffusion isotherms were adjusted to a fourth degree polynomial in the neighborhood of extremum temperatures. At low temperatures, the location of maximum and minimum diffusion were estimated using six points around each simulated extremum, and the best estimate for their values were taken from the analytically adjusted function. On intermediate temperatures, where maximum and minimum where too close, a single function and a higher number of points were used to adjust diffusion isotherms. The result is shown in Fig.~\ref{fig:hierarchy}, where is possible to see that an strict hierarchy of anomalies is found, with the same order observed in atomically detailed models of liquid water~\cite{debenedetti01:nat}.
\section{\label{final}Final discussions and conclusions}
We investigated the exact thermodynamics of the one dimensional repulsive lattice gas model using transfer matrix technique. In the ground state, the model presents a phase transition between a softened fluid and a dense fluid, which is characterized by a mixture between two states. The main consequence of this mixed state is a residual entropy in a single point in the pressure vs. temperature phase diagram. Using thermodynamic relations, it was argued that the presence of a residual entropy in a single point is consistent with a temperature of maximum density line ($\alpha=0$) emanating from the ground state phase transition. Considering this it is possible to state that ground state phase transition, residual entropy and density anomalies are thermodynamic properties of our model which are intrinsically related and cannot be dissociated from each other.
Further discussion about ground state phase transitions in one dimensional models is relevant at this point. First of all, the two transitions observed in our model can be seen as remanent from phase transitions on higher dimensional systems collapsing to $T \rightarrow 0$ as dimensionality is reduced to $d \rightarrow 1$. This collapse is also evident on the analytical solution of a lattice gas inside the Bethe lattice: in this case, while continuously reducing the coordination number the liquid-gas phase transition continuously shrinks to $T\rightarrow 0$~\cite{hu98:pre,barbosa10:jcp}. This is a relevant issue for one dimensional systems because the observed ground state phase transitions incorporate elements from both continuous and discontinuous phase transitions~\footnote{While crossing the neighborhood of the ground state phase transitions, in the $P$ vs. $T$ phase diagram, it is possible to observe the discontinuity in extensive variables (\textit{e.g.} volume) and the diverging behavior on response functions (\textit{e.g.} thermal expansion coefficient)}. While usually a discontinuous transition allows for any mixtures between two coexisting phases, when a continuous transition is collapsed on it, any mixture except for that maximizing the entropy is forbidden. Thus, for fluid-fluid transition entropy is maximized for equal distribution of randomly mixed interactions compatible with softened and dense fluids, resulting in an entropy per particle $s = S/N = k_b \ln 2$ (as in a coin toss game).
The viability of using a simple two states approximation to accurately describe the system in the neighborhood of the ground state phase transition is another issue deserving attention, since there is a long tradition on using two state models to describe liquid water~\cite{malenkov2009:jpcm:review,sanches87,tanaka2014:nature}. Recently, a modified regular solution model within a two solvent approach was adjusted to computer simulation data from ST2 model, being accurately used to describe computational data above the second critical point, in a region called Widom line~\cite{poole11:prl}. This raises the possibility of connecting the current approach to more complex, higher dimensional and off-lattice models. Nevertheless, it should be emphasized that the two states character of the approximation employed here is conceptually distinct from the ones implemented in Ref.~\cite{poole11:prl}, where two states refers to two fluids with distinct free energies. Our approach resembles more the approximation implemented in Ref.~\cite{heckmann2013:jcp}, in the sense that two states are localized microscopic states with well defined local energies and entropies, which are useful for truncating the partition function in some kind of expansion. Certainly it should be interesting to link these different views, but this is outside the scope of the current work.
To finish, it should be mentioned that in a previous work on lattice models with waterlike behavior~\cite{barbosa11:jcp} we found an oscillatory behavior on thermal expansion coefficient $\alpha$ associated to density anomaly and ground state phase transition. More recently this effect was shown to be general, being a feature of any ground state phase transition between different fluid structures in one dimensional lattice models with interactions restricted to first neighboring particles and, inspired by these models, it was possible to design a three dimensional spherically symmetric pair potentials with two liquid-liquid phase transitions, both of them associated to temperature of maximum density lines~\cite{barbosa13:pre}. The oscillatory behavior on $\alpha$, as well as the anomalous behavior on isothermal compressibility and constant pressure heat capacity, can be explained in terms of a residual entropy using thermodynamical and statistical arguments, while for the current model the the relation between GSPT, residual entropies and waterlike behavior can be mathematically described within the two states approximation.
Obviously we are tempted to ask whether such entropic effects couldn't play an important rule in more complex systems, such as higher dimensional and off-lattice models. Current work is being done on this direction through molecular dynamics simulations of the 3D core softened models, as in Ref.~\cite{barbosa13:pre}.
This work has been supported by CNPq and FAP-DF. MAAB acknowledges Marcia Barbosa and Evy Salcedo for useful discussions.
|
\subsubsection{Open-Loop Control}\label{sec:OL_construction}
For $t_i\ge 0$, consider a bounded or unbounded time interval in one of the following forms: $\mathcal T = [t_i, t_{i+1})$, $\mathcal T = [t_i, t_{i+1}]$ or $\mathcal T=[t_i,\infty)$.
\begin{enumerate}
\item At the initial time $t_i$ measure $x(t_i)$ and construct a smooth path $c(t_i)\in\Gamma(x^\star(t_i),x(t_i))$.
\item For each $t\in\mathcal T$, apply the control signal $u(t) = k_p(c(t),u^\star(t),t,1)$, where $c(t)$ is the forward image of the path $c(t_i)$ with the path of controls \eqref{eq:k_p}, i.e. for each $s\in[0,1]$ and $t\in\mathcal T$, $c(t,s)$ satisfies \begin{equation}\label{eq:forward_image}
\ddt c(t,s) = f(c(t,s),t)+B(c(t,s),t)k_p(c(t),u^\star(t),t,s).\end{equation}
\end{enumerate}
When this strategy is applied on an interval $\mathcal T = [t_i,\infty)$ the length of the curve $c(t)$ shrinks exponentially, and allows us to establish Theorem \ref{thm:CCM} claim 1.
The bound $R=\sqrt{\frac{\alpha_2}{\alpha_1}}$ given in the theorem is achieved if the initial path $c_0$ is a minimal geodesic joining $x^\star(t_i)$ to $x(t_i)$, existence of which is established in Lemma \ref{lem:hopf_rinow} in the appendix.
With any other initial path, exponential stability is still achieved with the same rate but perhaps with larger overshoot. Note that when $c(t_i)$ is a geodesic it is in general {\em not} the case that $c(t)$ is a geodesic for $t>t_i$, see Fig. \ref{fig:Gamma_x}.
\begin{figure}
\begin{center}
\includegraphics[width=0.65\columnwidth]{Figures/Gamma_Revised}
\caption{An illustration of the geometry of solutions using the open-loop or sampled-data CCM-based control over an interval $[t_i, t_{i+1}]$. The target trajectory $x^\star(t)$ is shown in red, system trajectory $x(t)$ in black. Paths joining $x^\star(t)$ to $x(t)$ are shown in blue. }
\label{fig:Gamma_x}
\end{center}
\end{figure}
\subsubsection{Sampled-Data Feedback Controller}
The open-loop controller can be extended to sampled-data feedback by recomputing geodesics at the sampling instants. To be precise:
\begin{enumerate}
\item At each sample time $t_i$, measure the state $x(t_i)$ and compute a minimal geodesic $\gamma_i \in \arg\min_{c\in\Gamma_i} E(c,t_i)$ where $\Gamma_i:=\Gamma(x^\star(t_i),x(t_i))$.
\item On the interval $\mathcal T = [t_i, t_{i+1})$ apply the open-loop control described above with $c(t_i)=\gamma_i$.
\end{enumerate}
Note that this is stabilizing with {\em any} choice of sample times, including uniform sampling: $t_i=it_s$ for some fixed $t_s>0$.
\subsubsection{Smooth Feedback, Uniquely Defined Almost Everywhere and in a Neighbourhood of $x^\star$}
By taking the limit as sampling interval goes to zero, one can obtain a continuous-time controller which does away with the need to solve \eqref{eq:forward_image} over the inter-sample intervals.
Specifically:
\begin{enumerate}
\item Measure the state $x(t)$ and a minimal geodesic $\gamma = \arg\min_{c\in\Gamma(x^\star(t),x(t))} E(c,t_i)$.
\item Apply the control signal $u(t) = k_p(\gamma,u^\star(t),t,1)$.
\end{enumerate}
This defines a mapping $(x(t), x^\star(t),u^\star(t),t\mapsto u(t)$, however a difficulty is that it may be multiply-defined or non-smooth at some states $x(t)$, specifically points on the {\em cut locus}, denoted by $\mathfrak C(x^\star,t)$, which is the set of points for which non-unique minimizing geodesics exist from $x^\star$ (cut points) and/or the first-order minimality condition fails (conjugate points).
This set is known to have zero Lebesgue measure. Let use define $\mathfrak D(x^\star,t):=\R^n/(\mathfrak C(x^\star,t) \cup x^\star )$ which is diffeomorphic to punctured open ball. For every $x\in \mathfrak D(x^\star,t)$ there is a unique minimal geodesic $\gamma$ joining $x$ and $x^\star$ \cite[Ch 13]{docarmo1992riemannian}, and we show in the appendix that the above controller is smooth on $\mathfrak D(x^\star,t)$ and continuous at $x=x^\star(t)$. This controller is universally exponentially stabilizing under the technical assumption that the set of times at which $x(t)\in \mathfrak C(x^\star,t)$ has zero measure.
\subsection{Stronger Conditions Giving Simpler Controllers}\label{sec:strong}
Since the differential dynamics are linear, it is tempting to look for an admissible differential feedback controller of the form $\delta_u= K(x,t)\d_x$ satisfying \eqref{eq:ccm_explicit}. We will show that this is possible under the following slightly stronger conditions:
\begin{enumerate}[C1:]
\item \label{C1} if $\delta_x\ne 0$ satisfies $\delta_x'MB=0$, then
\[
\delta_x'\left(\pder[M]{t}+\partial_f M+\pder[f]{x}'M+M\pder[f]{x}+2\lambda M\right)\delta_x <0,
\]
\item \label{C2} for each $i=1, 2, ..., m$,
$
\partial_{b_i} M + \pder[b_i]{x}'M+M\pder[b_i]{x} = 0.
$
\end{enumerate}
These stronger conditions also hold under the assumptions of Proposition \ref{prop:contracting}, as is clear from the proof in the appendix. Condition C\ref{C1} says that the uncontrolled system is contracting in directions orthogonal to the span of the control inputs. Condition C\ref{C2} ensures that large $u$ of unknown sign cannot cause expansion of $\|\d_x\|$. Formally it states that the vector fields $b_i$ are Killing fields for the metric $M$.
In particular, if $B$ is of the form $[0, I]'$, with $0$ and $I$ the zero and identity matrices of appropriate dimension, then Condition C\ref{C2} says that $M$ must not depend on the last $m$ state variables.
By applying Finsler's theorem (see, e.g., \cite{uhlig1979recurring}) pointwise in $x$ and $t$ to condition C\ref{C1}, we immediately obtain the following:
\begin{prop} \label{prop:rho} Condition C\ref{C1} is equivalent to the existence of a scalar multiplier $\rho(x,t)$ such that for all $x,t$:
\begin{equation}\label{eq:CCM_rhoform}
\pder[M]{t}+\partial_f M+\pder[f]{x}'M+M\pder[f]{x}-\rho MBB'M+2\lambda M < 0.
\end{equation}
\end{prop}
One can then construct the differential feedback gain $K(x,t) = -\frac{1}{2}\rho(x,t)B(x,t)'M(x,t)$ which satisfies \eqref{eq:ccm_explicit} and is always path integrable since it is independent of $u$.
\begin{remark}\label{rem:rho_gm}
If \eqref{eq:CCM_rhoform} holds for some multiplier $\rho(x,t)=\rho_0(x,t)$, then it clearly holds for any $\rho(x,t)\ge \rho_0(x,t)$ for all $x,t$ since $MBB'M\ge 0$. This can be interpreted as the differential feedback having infinite up-side gain margin, and also implies that one can construct a smooth $\rho(x,t)$.
\end{remark}
\subsection{Riemannian Energy as a CLF}
The proof of Theorem \ref{thm:CCM} uses an explicit construction of a particular stabilizing controller, but in doing so we have actually shown that the Riemannian energy $E(x, x^\star,t)$ can always be decreased, and hence be used as control Lyapunov function (CLF) for {\em any} target trajectory of the system.
The formula for first variation of energy \cite[p. 195]{docarmo1992riemannian} gives a particularly convenient expression for the time derivative of the energy functional as an affine function of $u$:
\begin{align}
\frac{1}{2}\frac{d}{dt}E(x, x^\star, t) =& \langle \gamma_s(t,0), \dot x^\star \rangle_{x^\star,t} - \langle \gamma_s(t,1), f(x,t)\rangle_{x,t}\notag\\
&-\langle \gamma_s(t,1),B(x,t)u\rangle_{x,t}+\frac{1}{2}\pder[E]{t},\label{eq:Edot}
\end{align}
When $x(t)\in \mathfrak C(x^\star,t)$ the above formula still holds with $=$ replaced by $\le$ and $\frac{d}{dt}$ replaced by the Dini derivative.
In proving Theorem 1, we have also proven that for any $x^\star, u^\star, t$, the convex set (either a half-space or all of $\R^m$):
\[
\mathcal U =\left\{u\in\ R^m : \ddt E(x, x^\star, t) \le - 2\lambda E(x,x^\star, t) \right\},
\]
where $ \ddt E(x, x^\star, t)$ is given by \eqref{eq:Edot}, is always non-empty.
This opens up the possibility of using many other particular controllers based on CLFs that may have further desirable properties. For example, pointwise min-norm control \cite{freeman2008robust}:
$
u(t) = \arg\min_{\tilde u\in \mathcal U} \|\tilde u\|^2
$
would have reduced control magnitude, and can be generalized to provide approximate optimality with guaranteed stability \cite{primbs2000receding}.
\subsection{Invariance Under Coordinate Change and Feedback}
Metrics and dual metrics are tensors: geometrical objects that are `intrinsic'' and have coordinate representations that transform appropriately under smooth coordinate changes. In the following theorem we establish that the CCM criteria are invariant under such coordinate changes and, additionally, under affine feedback laws.
\begin{thm}\label{thm:coord}
If the CCM condition \eqref{eq:CCM} (or equivalently \eqref{eq:weak_kernel}) is satisfied for system \eqref{eq:sys}, then \eqref{eq:CCM} and \eqref{eq:weak_kernel} still hold under:
\begin{enumerate}
\item affine feedback transformations $u(x,v) = \alpha(x)+\beta(x)v$ with $\beta$ a smooth non-singular $n\times n$ matrix function.
\item differential coordinate changes $\delta_\xi = \Phi(x)\delta_x$, in which $\Phi(x)$ is a non-singular matrix for all $x$, with the new CCM
$
M_\xi(x,t) :=\Psi'(x) M(x,t)\Psi(x)$ and dual CCM $
W_\xi(x,t):= \Phi(x) W(x,t) \Phi(x)'
$, where $\Psi(x) = \Phi(x)^{-1}$;
\item coordinate changes $\xi = \phi(x)$, $\phi$ a smooth diffeomorphism, with the new CCM and dual CCM $M_\xi, W_\xi$ as above with $\Phi(x) = \pder[\phi]{x}$ evaluated at $x=\phi^{-1}(\xi)$.
\end{enumerate}
\end{thm}
\begin{remark} If $\Phi(x)$ has bounded singular values over all $x,t$ within $[\sigma_{\min }, \sigma_{\max }]$, then the uniform bounds on $M$ and $W$ are also preserved under coordinate change.
\end{remark}
\subsection{Necessity for Feedback Linearizable Systems}
A corollary of Theorem \ref{thm:coord} is that for feedback linearizable systems, existence of a CCM is guaranteed. A system of the form \eqref{eq:sys}
is feedback linearizable if there exists a change of variables and feedback transformation such that the transformed system is linear time-invariant:
$
\dot \xi = G\xi+Hv,
$
where the pair of constant matrices $(G, H)$ is controllable \cite{khalil2002nonlinear}.
\begin{cor} For any feedback linearizable system there is a control contraction metric that verifies universal stabilizability, given by $W(x,t) = \Phi(x,t)P\Phi(x,t)'$ where $P$ is any constant symmetric positive definite matrix satisfying
$
H_\perp(GP+PG')H_\perp'<0.
$
\end{cor}
The proof is immediate from Theorem \ref{thm:coord}, and we note that such a $P$ is guaranteed to exist if $(G, H)$ is stabilizable \cite{dullerud2000course}.
The converse is not true: the necessary and sufficient conditions for feedback-linearizability consist of a controllability condition and an involutivity (complete integrability) condition to find an appropriate coordinate change \cite{khalil2002nonlinear}. In contrast, existence of a CCM depends only on stabilizability, and does not require complete integrability of the differential control. This is because a metric corresponds to a {\em differential} change of coordinates $\delta_z = \Theta(x,t)\delta_x$, i.e. $M = \Theta'\Theta$, but there is no requirement that this should be integrable to an explicit change of coordinates $z=\theta(x)$.
\begin{ex}
The following system
\[
\ddt \bm{x_1\\x_2} = \bm{-x_1-x_1^3 +x_2^2\\0}+\bm{0\\1}u=:f(x)+Bu
\]
is not feedback linearizable in the sense of \cite{isidori1995nonlinear}, since the vector fields $B$ and
$ ad_fB := \pder[f]{x}B-\pder[B]{x}f =[2x_2, 0]'
$
are not linearly independent when $x_2=0$. However, it is universally stabilizable as verified by \eqref{eq:CCM_rhoform_dual} with $W=I$ and multiplier $\rho(x) =1+2x_2^2$. Additionally, since we can take $B_\perp=[1 \ 0]'$, condition \eqref{eq:weak_kernel} reduces to the fact that $\pder[f_1]{x_2}=-1-3x_1^2<0 \ \forall x$.
\end{ex}
\section{Introduction}
\input{ccmIntro}
\section{Preliminaries and Problem Setup}\label{sec:setup}
\input{ccmSetup}
\section{Control Contraction Metrics}\label{sec:ccm}
\input{ccmCCM2}
\subsection{Dual Metrics and Convexity of Synthesis}\label{sec:dual}
\input{ccmDual}
\section{Properties of Control Contraction Metrics}\label{sec:properties}\input{ccmCLF}
\input{ccmConverse}
\input{ccmSynchr}\section{Illustrative Example}\label{sec:examples}
\input{ccmAndrieu}
\section{Stability and Stabilization of Submanifolds}\label{sec:flow_inv}
Convergence of a nonlinear system to a submanifold of state space is a requirement that appears in many applications, including coordination of multi-agent systems \cite{tanner2007flocking}, synchronization of oscillators \cite{dorfler2014synchronization}, computational neuroscience \cite{burak2009accurate}, and nonlinear control design
\cite{astolfi2007nonlinear}. Design of controllers to stabilize submanifolds has been investigated recently using transverse feedback linearization \cite{nielsen2008local}, reduction and backstepping
\cite{el2013reduction}, and modifying controllers for drift-free systems \cite{montenbruck2015compensating}.
Suppose a submanifold is defined by a level set of some smooth funciton $
Z(t) = \{x:z(x,t) = c\}
$
where $z:\R^n\times\R^+\to \R^q$ and $\pder[z]{x}$ has rank $q$ for all $x$. The sets $Z(t)$ are called
{\em controlled invariant} if there exists a smooth mapping $u^\star:Z(t)\times \R^+\to \R^m$ such that
\[\pder[z(x,t)]{t}+\pder[z(x,t)]{x}(f(x,t)+B(x,t)u^\star(x,t) )=0
\]
for all $x\in Z(t)$ for all $t\in\R^+$. In the case of an uncontrolled system, this reduces to the condition for a manifold to be {\em flow invariant}: $\pder[z]{t}+\pder[z]{x}f(x,t)=0$.
The objective is to design a controller guaranteeing exponential convergence to $Z(t)$, i.e. for each $x(0)$ one can construct a control signal $u(t)$ such that the solution $x(t)$ of \eqref{eq:sys} satisfies
$
\inf_{y\in Z(t)} |x(t)-y|\le e^{-\lambda t} \bar R
$ for some $\bar R>0$.
Assume we can construct a smooth matrix function $G(x,t)$ with columns that form a basis for the null space of $\pder[z(x,t)]{x}$. In order to study stability and stabilization of $Z(t)$, we construct a ``virtual control system'':
\begin{equation}\label{eq:virt}
\dot x = f(x,t)+\bar B(x,t)\bar u
\end{equation}
where $\bar B(x,t) = [B(x,t) \ G(x,t)]$ and $\bar u = [u' \ v']'$, with $u$ the actual control input and $v$ a newly introduced ``virtual control''.
\begin{thm}\label{thm:manifold_stab}
If there exists a CCM for the virtual control system \eqref{eq:virt} satisfying the strong conditions C\ref{C1}, C\ref{C2}, then any time-varying submanifold of the form $Z(t) = \{x:z(x,t) = c\}$ can be exponentially stabilized (open-loop, sampled-data, or continuously almost everywhere) with rate $\lambda$.
\end{thm}
The proof uses the concept of a ``shadow state'' $\bar x(t)$, which has the property that $\bar x(t) \in Z(t)\, \forall t$ and can be thought of as a generalised projection of $x(t)$ onto $Z(t)$. The virtual control system \eqref{eq:virt} is constructed so that it can represent dynamics of the real system when $v=0$, but can also represent $\bar x(t)\in Z(t)$ when $u=x^\star(\bar x,t)$ and $v$ is arbitrary. The idea is to ensure $x(t)$ converges to $\bar x(t)$, and therefore to $Z(t)$.
The following corollary gives simple convex criteria for an uncontrolled system to converge to a submanifold.
\begin{cor}\label{thm:synch} Consider an uncontrolled system of the form \eqref{eq:sys} with $B=0 \ \forall x,t$.
Suppose there exists a uniformly bounded dual metric $W$, invariant on level sets of $z(x,t)$, satisfying:
\begin{equation}\label{eq:CCM_Z}
\pder[z]{x}\left(-\pder[W]{t}-\partial_fW + \pder[f]{x}W+W\pder[f]{x}'+2\lambda W\right)\pder[z]{x}'< 0,
\end{equation}
Then all solutions of the system converge exponentially with rate $\lambda$ to the set $Z(t)$.
\end{cor}
Note that this corollary is based solely on the {\em existence} of a universally stabilizing controller. Actual computation of a control signal is not required.
Corollary \ref{thm:synch} generalizes results on {\em partial contraction} in \cite{wang2005partial, Pham2007}. The latter showed that convergence to a {\em linear} manifold defined by $z(x)=Vx = 0$, with $V$ a constant matrix, is guaranteed by the condition that $V(\pder[f]{x}+\pder[f]{x}')V'$ is uniformly negative definite.
A similar notion of {\em horizontal contraction} was studied in \cite{forni2014differential}.
|
\section{Introduction}
Almost a hundred years after Einstein wrote down the equations of General Relativity \cite{Einstein1916}, solutions of the Einstein equation remain notoriously difficult to find beyond those which exhibit significant symmetries. Even for these highly symmetric solutions, basic questions remain unanswered. A famous example is the question of the non-perturbative stability of the Kerr solution -- more than 50 years after its discovery, it is not known whether the exterior Kerr solution is stable. The main difficulty of solving the Einstein equation is its non-linearity, which defies perturbative approaches.
One of the main approaches in our hunt for solutions is the use of numerical methods. Numerical methods have been used to solve the Einstein equation for many decades, but the past decade has seen tremendous advances. A particular watershed moment was the breakthrough in evolving the inspiral mergers of two black holes \cite{Pretorius:2005gq,Baker:2005vv,Campanelli:2005dd} in 2005, a crucial milestone in the growth of numerical relativity as a discipline and as a tool. The other driver of this development is an explosion in the availability of large and powerful supercomputing clusters and the maturity of parallel processing technology such as the Message Passing Interface (MPI) and OpenMP, which open up new computational approaches to solving the Einstein equation.
We anticipate that this development will continue to accelerate, partly driven by upcoming observational projects. The gravitational wave detector LIGO is expected to start Advanced LIGO science runs in late 2015, and there are hopes that the sensitivity might be good enough to achieve a first detection of gravitational waves from binaries. In the longer term, the European Space Agency (ESA) has designated the space-based eLISA detector an L3 launch slot (expected launch date around 2034), and the LISA Pathfinder spacecraft has a firm launch date of late 2015.
Beyond searching for gravitational waves and black holes, numerical relativity is now beginning to find uses in the investigation of other areas of fundamental physics. For example, standard GR codes are now being adapted to study modified gravity \cite{Berti:2015itd}, cosmology \cite{Wainwright:2014pta,Johnson:2011wt} and even string theory motivated scenarios \cite{Cardoso:2012qm,Chesler:2013lia,Cardoso:2014uka,Choptuik:2015mma}. In particular, there is an increasing focus on solving GR coupled to matter equations in the strong-field regime: cosmic string evolution with GR, realistic black hole systems with accretion disks, non-perturbative systems in the early universe, etc. This nascent, but growing, interest in using numerical relativity as a mature scientific tool to explore other broad areas of physics is one key motivation of this work. Since it is often difficult to have an intuitive picture of the entire evolution ahead of time, the code must be able to automatically adapt to ensure that all regions of interest always remain adequately resolved.
In the numerical GR community, this requirement is largely met through a moving-box mesh refinement scheme. This type of setup consists of hierarchies of boxes nested around some specified centres, and the workflow typically requires the user to specify the exact size of these boxes beforehand. These boxes are then moved around, either along a prespecified trajectory guided by prior estimates, or by automatically tracking certain quantities or features in the solution as it evolves. Boxes which come within a certain distances of each other may also be allowed to merge. A number of moving-box mesh refinement codes have been made public over the recent years, many of which are built on top of the well-known $\mathtt{CACTUS}$ framework \cite{Goodale2002a,Loffler:2011ay}. One such implementation is the McLachlan/Kranc code \cite{Brown:2008sb,Kranc:web}, which uses finite difference discretisation and the Baumgarte-Shapiro-Shibata-Nakamura (BSSN) evolution scheme \cite{Baumgarte:1998te,Shibata:1995we}. Similarly, the \texttt{LEAN} code \cite{Sperhake:2006cy,Zilhao:2010sr}, which uses the \texttt{CACTUS} framework, and \texttt{BAM} and AMSS-NCKU \cite{Marronetti:2007ya, PhysRevD.82.024005} also implement the BSSN formulation of the Einstein equations. There is also $\mathtt{GRHydro}$ which implements general-relativistic magnetohydrodynamics (MHD) for the Einstein Toolkit \cite{EinsteinToolkit:web}, building yet another layer of physics on top of evolution codes such as McLachlan/Kranc. There are also non-$\mathtt{CACTUS}$ codes such as $\mathtt{SPeC}$ \cite{Pfeiffer:2002wt} and \texttt{bamps} \cite{Hilditch:2015aba}, which implement the generalised harmonic formulation of the Einstein equations using a pseudospectral method. In addition to these public codes, there is a plethora of closed-source codes.
The moving-box mesh refinement technique has found great success in astrophysically motivated problems such as two-body collision/inspiral. Outside of this realm, however, the setup can quickly become impractical, especially where one expects new length scales of interest to emerge dynamically over the course of the evolution. This can occur generically in highly nonlinear regimes, either by interaction between GR and various matter models, or by gravitational self-interaction itself which can exhibit complicated unstable behaviour in higher dimensions. In such situations, it is necessary to develop a code which has the flexibility to create refinement regions of arbitrary shapes and sizes, anywhere in the computational domain as may be required. This can be achieved by using a fully adaptive mesh refinement (AMR) technique, whose feature is generally characterised by the ability to monitor a chosen quantity at each time step and insert higher resolution sub-regions where this quantity fails to lie within some chosen bounds. Of course, the efficacy of such codes depend crucially on a sensible choice of these criteria, however when implemented correctly they can be an extremely powerful tool. The advantage here is twofold: AMR ensures that small emergent features remain well-resolved at all times, but also that only those regions which require this extra resolution gets refined, thus allowing more problems to fit within a given memory footprint. To the best of our knowledge, \texttt{PAMR/AMRD} \cite{PAMR} and \texttt{HAD} \cite{Neilsen:2007ua} are the only two codes with full adaptive mesh refinement (AMR) capabilities in numerical GR.
In this work, we introduce $\mathtt{GRChombo}$, a new multi-purpose numerical relativity code. $\mathtt{GRChombo}$ is built on top of the $\mathtt{Chombo}$ \cite{Chombo} framework. $\mathtt{Chombo}$ is a set of tools developed by Lawrence Berkeley National Laboratory for implementing block-structured AMR for solving partial differential equations. $\mathtt{GRChombo}$ features \footnote{Since the \texttt{Chombo} core is dimension-independent for up to six spatial dimensions, $\mathtt{GRChombo}$ could potentially be extended to simulate fully higher dimensional spacetimes without any symmetry assumptions, subject to computational resource availability.} include the following.
\begin{itemize}
\item{\emph{BSSN formalism with moving puncture}: $\mathtt{GRChombo}$ evolves the Einstein equation in the BSSN formalism. An option to turn on the CCZ4 constraint damping modification\cite{Alic:2011gg,Bona:2003fj} is also available. Singularities of black holes are managed using the moving puncture gauge conditions \cite{Campanelli:2005dd, Baker:2005vv}.}
\item{\emph{Adaptive Mesh Refinement}: $\mathtt{Chombo}$ provides full adaptive mesh refinement with non-trivial nesting topologies via the Berger-Rigoutsos block-structured adaptive mesh algorithm \cite{BergerRigoutsis91}. The user only needs to specify regridding criteria, and $\mathtt{Chombo}$ does the rest. Kreiss-Oliger dissipation is used to control errors, from both truncation and the interpolation associated with regridding. }
\item{\emph{MPI scalability}: $\mathtt{GRChombo}$ inherits the parallel infrastructure of $\mathtt{Chombo}$, with ability to scale efficiently to many thousands of CPU-cores per run.}
\item{\emph{Standardized Output and Visualization}: $\mathtt{GRChombo}$ uses \texttt{Chombo}'s $\mathtt{HDF5}$ output format, which is supported by many popular visualization tools such as $\mathtt{VisIt}$. In particular, the output files can be used as input files if one chooses to continue a previously stopped run -- i.e. the output files are also checkpoint files.}
\end{itemize}
In this paper, we will detail these capabilities of $\mathtt{GRChombo}$ and illustrate how they expand the current field in numerical GR to permit new physics to be explored. The design methodology, scaling properties and performance of $\mathtt{GRChombo}$ in a number of standard simulations are included.
The paper is organized as follows:
In Sec. \ref{sec-GRChombo} we describe the implementation of the code. In particular, in Sec. \ref{sec-eqns} we establish the exact formulation of the equations which were used and our notation conventions, and in Sec. \ref{sec-code} we detail the AMR methodology and other ``coding'' aspects, such as finite differencing, dissipation and load balancing.
In Sec. \ref{sec-newphys} we give examples of several areas of physics which the code is well adapted to explore, and in which it offers advantages over existing codes.
In Sec. \ref{sec-tests}, we present the results of standard tests, including the Apples with Apples tests \cite{Babiuc:2007}, black holes and black hole mergers, and critical collapse. We test the AMR capabilities of the code, its robustness to regridding errors, and its scaling and convergence properties.
We discuss our results and future directions in Sec. \ref{sec-dis}.
Videos of several tests conducted in this paper, and examples of some new problems being tackled using the code, can be viewed via our website at http://grchombo.github.io.
\section{$\mathtt{GRChombo}$} \label{sec-GRChombo}
In this section, we will describe our numerical implementation of the Einstein equation in $\mathtt{GRChombo}$.
\subsection{$\mathtt{GRChombo}$ equations and notation conventions} \label{sec-eqns}
The purpose of this subsection is to clearly state the equations of motion that we have implemented and our conventions for completeness. Since these are standard in the field, the experienced reader may want to skip this subsection.
Many numerical relativity codes implement the so called BSSN formulation of the Einstein equation \cite{Nakamura:1987zz,Shibata:1995we,Baumgarte:1998te}. This formulation expresses the Einstein equation in a strongly hyperbolic form, and together with the $``1+\log"$ slicing \cite{Bona:1994dr} and the ``gamma-driver" gauge conditions \cite{Alcubierre:2002kk}, has allowed the stable simulation of dynamical spacetimes of interest, including black hole binaries.
More recently, other refined formulations of the Einstein equation based on the Z4 system \cite{Bona:2003fj,Gundlach:2005eh} have been proposed, most notably the Z4c formulation \cite{Bernuzzi:2009ex} and the CCZ4 formulation \cite{Alic:2011gg}.\footnote{Both the BSSN and CCZ4 equations have been written in a fully covariant form \cite{Brown:2009dd,Baumgarte:2012xy,Sanchis-Gual:2014nha}. These covariant formulations can be advantageous in certain cases, and we plan to implement them in the future.} In the Z4 system, both the Hamiltonian and the momentum constraint are promoted to dynamical variables and hence constraint violating modes can propagate and eventually exit the computational domain. This may potentially result in a more stable evolution. In addition, the Z4 system can be augmented with damping terms so that constraint violating modes can be exponentially suppressed. In practical terms, the changes required between the CCZ4 equations and the standard BSSN equations are minimal and in \texttt{GRChombo} we have implemented both.
In this work, we follow the indexing convention of \cite{ShapiroBook}. The signature is $(-+++)$, and low-counting Latin indices $a,b,\dots$ are abstract tensor indices while Greek indices $\mu,\nu,\dots$ denote spacetime component indices and run from $0,1,2,3$. Spatial component indices are labeled by high-counting Latin indices $i,j,\dots$ which runs from $1,2,3$. Unless otherwise stated, we set $G=1$ and $c=1$.
\subsubsection{Evolution equations}
The Z4 system with constraint damping is \cite{Gundlach:2005eh}
\begin{equation}
R_{ab} +\nabla_a\,Z_b + \nabla_b\,Z_a-\kappa_1\left[n_a\,Z_b + n_b\,Z_a-(1+\kappa_2)\,g_{ab}\,n^c\,Z_c\right]=8\,\pi\left(T_{ab} -\frac{1}{2}\,g_{ab}\,T\right)
\label{eq:Z4eqs}
\end{equation}
where $R_{ab}$ is the Ricci tensor associated with the metric $g$ on the spacetime manifold ${\mathcal M}$, and $\nabla$ is the corresponding metric compatible covariant derivative. $T_{ab}$ is the stress-energy tensor of the matter and $T\equiv g_{ab}\,T^{ab}$ is its trace. If we set $Z^a=0$, the Z4 equations \eqn{eq:Z4eqs} reduce to the standard (trace-reversed) Einstein equation. Here $\kappa_1$ and $\kappa_2$ are parameters which control the damping.
In the \texttt{GRChombo} code we use the standard $3+1$ ADM decomposition of the spacetime metric,
\begin{equation}
ds^2=-\alpha^2\,dt^2+\gamma_{ij}(dx^i + \beta^i\,dt)(dx^j + \beta^j\,dt)\,,
\end{equation}
so that $\gamma_{ij}$ is the induced metric on the spatial slices and
\begin{equation}
n^\mu = \frac{1}{\alpha}\left(\partial_t^\mu - \beta^i\,\partial_i^\mu\right)\,,
\end{equation}
is the corresponding timelike unit normal. The extrinsic curvature is defined as
\begin{equation}
K_{ij} = -\frac{1}{2}\,(\pounds_n\gamma)_{ij}\,,
\end{equation}
where $\pounds$ denotes the Lie derivative. As is customary, we decompose the induced metric as $\gamma_{ij}=\frac{1}{\chi^2}\,\tilde\gamma_{ij}$ so that $\det\tilde\gamma_{ij}=1$ and $\chi = \left(\det\gamma_{ij}\right)^{-\frac{1}{6}}$. Similarly, the extrinsic curvature is decomposed into its trace, $K=\gamma^{ij}\,K_{ij}$, and its traceless part so that
\begin{equation}
K_{ij}=\frac{1}{\chi^2}\left(\tilde A_{ij} + \frac{1}{3}\,K\,\tilde\gamma_{ij}\right)\,,
\end{equation}
with $\tilde\gamma^{ij}\,\tilde A_{ij}=0$. In the Z4 system, one further defines $\Theta$ as the projection of the Z4 four-vector along the normal timelike direction, $\Theta \equiv -n_\mu\,Z^\mu$. Finally, the spacelike components of the four-vector, $Z_i$, are included in a variable $\hat\Gamma^i$ defined as
\begin{equation}
\hat\Gamma^i \equiv \tilde\Gamma^i + 2\,\tilde\gamma^{ij}\,Z_j\,,
\end{equation}
where $\tilde\Gamma^i=\tilde\gamma^{jk}\,\tilde\Gamma^i_{~jk}$ and $\tilde\Gamma^i_{~jk}$ are the Christoffel symbols associated to the conformal metric $\tilde\gamma_{ij}$,
\begin{equation}
\tilde \Gamma^i_{~jk} = \frac{1}{2}\,\tilde\gamma^{il}\left(\partial_j\tilde\gamma_{kl} + \partial_k\tilde\gamma_{jl} - \partial_l\tilde\gamma_{jk}\right)\,.
\end{equation}
Summarizing, the dynamical variables for the Z4 system are
\begin{equation}
\{\chi,\,\tilde\gamma_{ij},\,K,\,\tilde A_{ij},\,\Theta,\,\hat\Gamma^i\}\,.
\end{equation}
Setting to zero the Z4 four-vector, $Z^\mu=0$, this system reduces to the standard BSSN system.
Finally, we recall the various components of the matter stress tensor in the standard $3+1$ decomposition:
\begin{equation}
\rho = n_a\,n_b\,T^{ab}\,,\quad S_i = -\gamma_{ia}\,n_b\,T^{ab}\,,\quad S_{ij} = \gamma_{ia}\,\gamma_{jb}\,T^{ab}\,,\quad S = \gamma^{ij}\,S_{ij}\,.
\label{eq:Mattereqns}
\end{equation}
We are now ready to write down the evolution equations for CCZ4 system in the standard $3+1$ decomposition \cite{Alic:2011gg}:
\begin{align}
&\partial_t\chi=\frac{1}{3}\,\alpha\,\chi\, K - \frac{1}{3}\,\chi \,\partial_k \beta^k + \beta^k\,\partial_k \chi\,, \label{eqn:dtchi}\\
&\partial_t\tilde\gamma_{ij} =-2\,\alpha\, \tilde{A}_{ij}+\tilde{\gamma}_{ik}\,\partial_j\beta^k+\tilde{\gamma}_{jk}\,\partial_i\beta^k-\frac{2}{3}\,\tilde{\gamma}_{ij}\,\partial_k\beta^k +\beta^k\,\partial_k \tilde{\gamma}_{ij}\,, \label{eqn:dttgamma} \\
&\partial_t K = -\gamma^{ij}D_i D_j \alpha + \alpha\left(R + 2\,D_iZ^i + K^2 - 2\,K\,\Theta\right) + \beta^i\partial_iK \nonumber\\
&\hspace{1.2cm}-3\,\alpha\,\kappa_1(1+\kappa_2)\Theta+4\pi\,\alpha(S-3\,\rho), \label{eqn:dtK} \\
&\partial_t\tilde A_{ij} = \chi^2\left[-D_iD_j \alpha + \alpha\left( R_{ij} + D_iZ_j + D_jZ_i-8\pi\,\alpha \,S_{ij}\right)\right]^\textrm{TF} \nonumber \\
&\hspace{1.3cm}+ \alpha \tilde{A}_{ij}(K-2\,\Theta)-2\,\alpha\,\tilde{A}_{il}\,\tilde{A}^l{}_j+\tilde{A}_{ik}\,\partial_j\beta^k + \tilde{A}_{jk}\,\partial_i\beta^k \nonumber \\
&\hspace{1.3cm}-\frac{2}{3}\,\tilde{A}_{ij}\,\partial_k\beta^k+\beta^k\,\partial_k \tilde{A}_{ij}\,, \label{eqn:dtAij}\\
&\partial_t\Theta =\frac{1}{2}\,\alpha\left(R + 2\,D_iZ^i-\tilde{A}_{ij}\,\tilde{A}^{ij} + \frac{2}{3}\,K^2 - 2\,\Theta\,K\right) \nonumber \\
&\hspace{1.2cm} -Z^i\,\partial_i\alpha + \beta^k\,\partial_k\Theta - \alpha\,\kappa_1\,(2+\kappa_2)\Theta - 8\pi\,\alpha\,\rho\,,\\
&\partial_t\hat\Gamma^i=-2\,\tilde{A}^{ij}\,\partial_j \alpha +2\,\alpha\left(\tilde\Gamma^i_{jk}\,\tilde{A}^{jk}-\frac{2}{3}\,\tilde\gamma^{ij}\partial_j K - 3\,\tilde{A}^{ij}\frac{\partial_j \chi}{\chi}\right) \nonumber \\
&\hspace{1.3cm} +\beta^k\partial_k \hat\Gamma^{i} +\tilde\gamma^{jk}\partial_j\partial_k \beta^i +\frac{1}{3}\,\tilde\gamma^{ij}\partial_j \partial_k\beta^k \nonumber \\
&\hspace{1.3cm} + \frac{2}{3}\,\tilde\Gamma^i\,\partial_k \beta^k -\tilde\Gamma^k\partial_k \beta^i
+2\,\kappa_3\left(\frac{2}{3}\,\tilde\gamma_{ij}\,Z_j\,\partial_k\beta^k - \tilde\gamma^{jk}\,Z_j\,\partial_k\beta^i\right) \nonumber \\
&\hspace{1.3cm} + 2\,\tilde\gamma^{ij}\left(\alpha\,\partial_j\Theta - \Theta\,\partial_j\alpha - \frac{2}{3}\,\alpha\,K\, Z_j\right)
-2\,\alpha\,\kappa_1\,\tilde\gamma^{ij}\,Z_j - 16\pi\,\alpha\,\tilde\gamma^{ij}\,S_j\,.\label{eqn:dtgamma}
\end{align}
Here $D_i$ is the metric compatible covariant derivative with respect to the physical metric $\gamma_{ij}$ and $[\ldots]^\textrm{TF}$ denotes the trace free part of the expression inside the parenthesis. The three-dimensional Ricci tensor, $R_{ij}$, is split as
\begin{equation}
R_{ij} = \tilde R_{ij} + R^\chi_{ij}\,,
\end{equation}
where
\begin{equation}
\tilde R_{ij} = -\frac{1}{2}\tilde{\gamma}^{lm}\partial_m\partial_l\tilde{\gamma}_{ij}+\tilde{\Gamma}^k\tilde{\Gamma}_{(ij)k}+\tilde{\gamma}^{lm}(2\tilde{\Gamma}^k_{l(i}\tilde{\Gamma}_{j)km}+\tilde{\Gamma}^k_{im}\tilde{\Gamma}_{klj}) \label{eqn:conformalR}
\end{equation}
and
\begin{equation}
R^{\chi}_{ij}=\frac{1}{\chi}(\tilde{D}_i\tilde{D}_j\chi + \tilde{\gamma}_{ij}\tilde{D}^l\tilde{D}_l \chi)-\frac{2}{\chi^2}\tilde{\gamma}_{ij}\tilde{D}^l\chi \tilde{D}_l\chi. \label{eqn:Rchi}
\end{equation}
where $\tilde D_i$ is the metric compatible covariant derivative with respect to the conformal metric $\tilde\gamma_{ij}$. Note that the three-dimensional Ricci Scalar is then $R = \gamma^{ij}R_{ij}$.
Equations \eqn{eqn:dtchi}--\eqn{eqn:dtgamma} are the CCZ4 evolution equations as originally presented in \cite{Alic:2011gg}, including the extra damping parameter $\kappa_3$. This parameter controls the coupling of some quadratic terms in the evolution equation for $\hat \Gamma^i$. The choice $\kappa_3=1$ corresponds to the fully covariant CCZ4 system, but as discussed in \cite{Alic:2011gg}, it leads to instabilities in the evolution of spacetimes containing black holes. More recently, \cite{Alic:2013xsa} showed that replacing $\kappa_1\to \kappa_1/\alpha$ in \eqn{eqn:dtchi}--\eqn{eqn:dtgamma} allows to stably evolve black hole spacetimes whilst retaining the full covariance of the CCZ4 system. In $\mathtt{GRChombo}$ we have included a parameter that allows us to switch from the original formulation of the CCZ4 system to the more recent one proposed in \cite{Alic:2013xsa}, with the aforementioned redefinition of $\kappa_1$.
Note that in the actual evolution, the values of the three-vector $Z_i$ are computed from the knowledge of the evolved variable $\hat\Gamma^i$ and $\tilde\Gamma^i$, which is computed from the conformal metric, $\tilde\gamma_{ij}$. Finally, we note that the evolution equations \eqn{eqn:dtchi}--\eqn{eqn:dtgamma} reduce to the standard BSSN equations upon setting $\Theta=0$ and $Z^i=0$, and using the Hamiltonian constraint, \eqn{eqn:HamiltonianConst}, in the evolution equation for $K$, \eqn{eqn:dtK}, to eliminate the Ricci scalar $R$.
\subsubsection{Gauge conditions}
To complete the set of evolution equations, we need to choose slicing conditions -- we specify the gauge via driving conditions for the lapse $\alpha$ and shift $\beta^i$ \cite{Alcubierre:2002kk}. The optimal gauge conditions are in general physics dependent, and $\mathtt{GRChombo}$ allows the user to code in whichever gauge conditions are well adapted to the application at hand. However, the most commonly used conditions which have been implemented in $\mathtt{GRChombo}$ are detailed below.
The \emph{alpha-driver} condition is usually written as a first order differential equation
\begin{equation}
\partial_t \alpha = -\mu_{\alpha_1}\alpha^{\mu_{\alpha_2}}K + \mu_{\alpha_3}\beta^i\partial_i \alpha. \label{eqn:alphadriver}\,
\end{equation}
The commonly used $1+\log$ slicing applicable for black hole inspirals corresponds to $\mu_{\alpha_1}=2$, $\mu_{\alpha_2}=1$ and $\mu_{\alpha_3}=1$. On the other hand, the \emph{maximal slicing} condition, which preserves $K=0$ and $\partial_t K=0$ at all slices, is a second order differential equation
\begin{equation}
D^2\alpha = \alpha[K_{ij}K^{ij}+4\pi(\rho+S)], \label{eqn:maximalslicing}
\end{equation}
which is useful for spherically symmetric collapse problems such as the critical scalar collapse scenarios.
We specify the evolution equation for $\beta^i$ using the \emph{gamma-driver} conditions \cite{Alcubierre:2002kk},
\begin{eqnarray}
\partial_t \beta^i& =& \eta_1 B^i\, \label{eqn:betadriver}\\
\partial_t B^i &=& \mu_{\beta_1}\alpha^{\mu_{\beta_2}}\partial_t \hat\Gamma^i-\eta_2 B^i\, ,\label{eqn:gammadriver}
\end{eqnarray}
where $B^i$ is an auxiliary vector field, while $\eta_1$, $\eta_2$, $\mu_{\beta_1}$ and $\mu_{\beta_2}$ are input parameters. The usual hyperbolic gamma-driver condition uses the parameters $\eta_1=3/4$, $\mu_{\beta_1}=1$, $\mu_{\beta_2}=0$ and $\eta_2=1$. We have also included parameters that allow us to turn on standard advection terms in \eqn{eqn:betadriver}--\eqn{eqn:gammadriver}.
In our tests in Sec. \ref{sec-vacblack} and Sec. \ref{sec-choptuik}, where black holes are present, we manage the singularities with the so-called \emph{moving punctures method} \cite{Campanelli:2005dd,Baker:2005vv}, which is a combination of the $1+\log$ slicing for $\alpha$ and gamma-driver for $\beta^i$. In addition, we hard code the condition $\alpha>0$ as is usual practice.
\subsubsection{Constraint equations}
$\mathtt{GRChombo}$ computes both the Hamiltonian constraint,
\begin{equation}
H = R + K^2-K_{ij}K^{ij}-16\pi \rho , \label{eqn:HamiltonianConst}
\end{equation}
and the momentum constraint,
\begin{equation}
M_i = \gamma^{jk}(\partial_l K_{ij}-\partial_iK_{jl}-\Gamma^m_{~jl}K_{mi}+\Gamma^m_{~ij}K_{lm})-8\pi S_i , \label{eqn:MomentumConst}
\end{equation}
in order to monitor the accuracy of the calculation.
For the algebraic constraints of BSSN, we do not enforce (by hand) the condition that the conformal metric has a determinant of one, but we do enforce after each timestep that $\tilde{A}_{ij}$ is traceless.
\subsubsection{Scalar matter evolution equations} \label{eqn:scalarmatterevolve}
We have included a single minimally coupled scalar field $\phi$ as matter content
\begin{equation}
{L}_{\phi} = \frac{1}{2}\nabla_\mu \phi\nabla^{\mu} \phi + V(\phi) \label{eqn:scalaraction},
\end{equation}
with the equation of motion
\begin{equation}
\nabla_{\mu}\nabla^{\mu}\phi - \frac{dV}{d\phi}=0 \label{eqn:scalarEOM}.
\end{equation}
As is usual, we decompose the second order \eqn{eqn:scalarEOM} into two first order variables $\phi$ and $\Pi_M$
\begin{equation}
\Pi_M \equiv \frac{1}{\alpha}(\partial_t \phi -\beta^i\partial_i \phi) \label{eqn:phiM}.
\end{equation}
We note that our $\Pi_M$ is negative of $\Pi$ in some references, e.g. \cite{ShapiroBook}. \eqn{eqn:scalarEOM} is then decomposed into the following equations
\begin{equation}
\partial_t \phi = \alpha \Pi_M +\beta^i\partial_i \phi \label{eqn:dtphi}
\end{equation}
and
\begin{equation}
\partial_t \Pi_M=\beta^i\partial_i \Pi_M +\gamma^{ij}(\alpha\partial_j\partial_i \phi + \partial_j \phi\partial_i \alpha)+\alpha\left(K\Pi_M-\gamma^{ij}\Gamma^k_{ij}\partial_k \phi+\frac{dV}{d\phi}\right). \label{eqn:dtphiM}
\end{equation}
We also use the energy momentum tensor of the scalar field
\begin{align}
T_{\mu \nu} = \nabla_\mu \phi \nabla_\nu \phi - \tfrac{1}{2} g_{\mu \nu} (\nabla_\lambda \phi \, \nabla^\lambda \phi + 2V)
\end{align}
to calculate the matter components of the BSSN/CCZ4 system via \eqn{eq:Mattereqns}.
\subsection{$\mathtt{GRChombo}$ code implementation} \label{sec-code}
$\mathtt{GRChombo}$ is a physics engine built around the publicly-available adaptive-mesh framework $\mathtt{Chombo}$ \cite{Chombo}. $\mathtt{GRChombo}$ solves the system of hyperbolic partial differential equations of the Einstein equation and scalar matter content (see section \ref{sec-eqns}) using a finite difference scheme.
A key feature of $\mathtt{GRChombo}$ is its highly flexible adaptive mesh refinement capability -- to be precise, $\mathtt{GRChombo}$
uses Berger-Oliger style \cite{bergeroliger,BergerColella} adaptive mesh refinement with Berger-Rigoutsos \cite{BergerRigoutsis91} block-structured grid generation. $\mathtt{GRChombo}$ supports full non-trivial mesh topology -- i.e. many-boxes-in-many-boxes. Morton ordering is used to map grid responsibility to neighbouring processors in order to optimize processor number scaling.
\subsubsection{Discretization and Time-stepping} \label{sect:discretization}
We would like to evolve a set of fields in space (the state-vector $\Phi({\bf x},t) = \{\phi_1,\phi_2,\phi_3,\dots\}$) through time $t$ via the equations of motion
\begin{equation}
\frac{\partial \Phi}{\partial t} = {\cal F}(\Phi) ,
\end{equation}
where ${\cal F}$ is some operator on $\Phi$ which, in the case of the Einstein equation, is non-linear.
In $\mathtt{GRChombo}$, both the space and time coordinates are discretized. Evolution in time is achieved through time-stepping $t\rightarrow t+\Delta t$, where at each time step we compute the fluxes for each grid point individually. Time stepping is implemented using the standard 4th Order Runge-Kutta method, and hence, as usual, we only need to store the values of the state-vector at each time step.
$\Phi$ itself is discretized into a cell-centered grid. Spatial derivatives across grid points are computed using standard 4th order stencils for all spatial derivatives, except for advection terms which are implemented using an upwind stencil. The form of the stencils used exactly follows equations (2.2) through (2.6) of \cite{Zlochower:2005bj}.
\subsubsection{Berger-Rigoutsos Block-structured AMR} \label{sect:BRAMR}
$\mathtt{GRChombo}$ uses {\tt Chombo}'s implementation of the Berger-Rigoutsos adaptive-mesh-refinement algorithm \cite{BergerRigoutsis91}, which is one of the standard block-structured AMR schemes. Block-structured AMR regrids by overlaying variable size boxes, instead of remeshing on a cell-by-cell basis (the ``bottom-up'' approach). The main challenge is to find an efficient algorithm to \emph{partition} the cells which need regridding into rectangular ``blocks''. In this section, we will briefly discuss the algorithm.
For a given grid at some refinement level $l$ where $l=0$ is the base level and $l_{max}$ is some preset maximum refinement level, we first ``tag'' cells for which refining is required. The refinement condition used by $\mathtt{GRChombo}$ is discussed later in this section. The primary problem of AMR is to efficiently partition this grid into regions which require adaptive remeshing. In \emph{block-structured} AMR these regions are boxes in 3D or rectangles in 2D. \emph{Efficiency} is measured by the ratio of tagged over untagged cell points in the final partitions.
In each partition, we compute the \emph{signatures} or traces of the tagging function $f(x,y,z)$ of any given box
\begin{eqnarray}
X(x) &=& \int f(x,y,z) dy dz, \\
Y(y) &=& \int f(x,y,z) dx dz, \\
Z(z) &=& \int f(x,y,z) dy dx,
\end{eqnarray}
where $f(x,y,z) =1$ if it is tagged for refinement and $0$ otherwise. Given these traces, we can further compute the Laplacian of the traces $\partial^2_x X(x)$, $\partial^2_y Y(y)$ and $\partial^2_z Z(z)$. Given the Laplacians, the algorithm can search for all (if any) inflection points \emph{individually for each direction} -- i.e. the locations of zero crossings of the Laplacian, and then pick the one whose $\delta (\partial^2_i X_i)$ is the greatest (corresponding to the line -- or plane in 3D -- separating the largest change in the Laplacian). This point then becomes the line of partition for this particular dimension. Roughly speaking, this line corresponds to an edge between tagged and untagged cells in the orthogonal directions of the signature. Furthermore, if there exists a point $x_i$ with zero signature $X_i(x_i)=0$ (i.e. no cells tagged along the plane orthogonal to the direction), then this ``hole'' is chosen to be the line of partition instead.
After a partitioning, we check whether or not each partition is \emph{efficient}, specifically whether it passes a user-specified threshold or \emph{fill factor}, $\epsilon <1.0$,
\begin{equation}
\frac{\mathrm{Tagged~Cells}}{\mathrm{Total~Cells}} > \epsilon
\end{equation}
If this is true, then we check if this box is \emph{properly nested}\footnote{Properly nested means that (1) a $l+1$ level cell must be separated from an $l-1$ cell by at least a single $l$ level cell and (2) the physical region corresponding to a $l-1$ level cell must be completely filled by $l$ cells if it is refined, or it is completely unrefined (i.e. there cannot be ``half-refined'' coarse cells).} \cite{BergerColella, bergeroliger} and if so we accept this partition and the partitioning for this particular box stops. If not, then we continue to partition this box recursively until either all boxes are accepted or partitioning no longer can be achieved (either by the lack of any tagged cells or reaching a preset limit on the number of partitions). Furthermore, $\mathtt{GRChombo}$ allows one to set the maximum partition size, which if exceeded will force a partitioning of the box.
Note that a higher value of $\epsilon$ means that the partitioning will be more aggressive which will lead to a higher efficiency in terms of final ratio of tagged to untagged cells -- generating more boxes in the process. However, this is not necessarily always computationally better as partitioning requires computational overhead, which depends on the number and topology of the processors. The ideal fill ratio is often a function of available processors, their topology and of course the physical problem in question.
A partitioned box is then \emph{refined}, i.e. its grids split into a finer mesh using the (user definable) refinement ratio $n^l = \delta x^{l+1}/\delta x^l$, and this process continues recursively until we either have no more tagged cells, or when we reached a preset number of refinement levels $l_{max}$.
Finally we need to specify a prescription for tagging which cells are required to be refined. $\mathtt{GRChombo}$ tags a cell when any (set of) user selected fields $\phi \in \Phi$ pass a chosen threshold $\sigma(\phi)$, which sets a limit on the $L^2$ norm of the change in the value of the field across that cell, i.e.
\begin{equation}
f(x,y,z)=\left\{
\begin{array}{cc}
1 & \mathrm{if}~\sqrt{\sum\limits_{i=1}^{3} (\phi({\bf x}+\delta x \, \hat{\bf x}_i) - \phi({\bf x}-\delta x \, \hat{\bf x}_i))^2} > \sigma(\phi) \\
\\
0 & \mathrm{otherwise}.
\end{array}
\right. \label{eqn:tagging}
\end{equation}
This condition can be augmented, for example by using estimated truncation errors as tagging conditions instead.
Partitioning can be done at every time-step for each refinement level and this is a user preset choice per refinement level. However, the user may wish to select a lower frequency because it might be useful to not partition at every timestep for a given refinement level. One consideration is that it is important to let numerical errors dissipate (e.g. via Kreiss-Oliger dissipation, see Sec. \ref{sect:kreissoliger}) before remeshing. Once a new hierarchy of partitions is determined, we interpolate via linear interpolation from coarse to fine mesh, and average from fine to coarse mesh.
Since the finer mesh has a smaller Courant number, each mesh level's timestep is appropriately reduced via
\begin{equation}
\Delta t^{l+1} = \frac{\Delta t^l}{n^l}.
\end{equation}
$\mathtt{GRChombo}$ follows standard Berger-Collela AMR evolution algorithm \cite{BergerColella}. Starting from the coarsest mesh, it advances the coarse mesh 1 time step i.e. $t \rightarrow t+\Delta t^l$. Then it advances the next finest mesh $n^l$ times until the fine mesh ``catches up'' with the coarse mesh time. Once both coarse and fine mesh are at the same time $t$, $\mathtt{GRChombo}$ synchronizes them by averaging over the fine cells to the coarse cell values. We add that in a conservative system, this simple synchronization is not conservative and requires proper \emph{refluxing} -- the coarse fluxes are replaced with a time-averaged fine mesh fluxes. This step incurs additional overhead, and is at the moment not implemented by $\mathtt{GRChombo}$ as GR equations are not conservative. Nevertheless, we intend to implement conservative refluxing as an option in a future version of $\mathtt{GRChombo}$.
\subsubsection{Load Balancing} \label{sect:loadbalance}
$\mathtt{GRChombo}$'s efficiency when running on a large number of distributed-memory nodes is highly dependent on efficient load balancing of the available computational work across those nodes. Load balancing seeks to avoid the situation where most of the nodes are waiting for some small subset of nodes to finish their computational work, and it does this by seeking to distribute the amount of work to be done per time step evenly among all of the nodes. This can be non-trivial when AMR boxes at many different refinement levels are simultaneously being evolved across the system. In addition, even within a single node, multiple OpenMP threads might be running, and the per-node workload needs to be balanced amongst those threads.
For the inter-node load balancing, $\mathtt{GRChombo}$ leverages {\tt Chombo}'s load balancing capabilities to distribute the AMR boxes among the available nodes. It does this by building a graph of the boxes to be distributed, adding edges between neighbouring and overlapping boxes. A bin packing / knapsack algorithm is used to balance the computational work among nodes, where the work is assumed to be proportional to the number of grid points, and then an exchange phase is used to minimise the communication cost. Because this load balancing procedure can be costly, we normally run it only every few time steps. In between runs of the load balancing procedure, new boxes generated by AMR refinement stay on the node which holds the parent box.
Within each node, the computational work is divided amongst the available OpenMP threads by iterating over the boxes to process using OpenMP's dynamic scheduling capability. This allows each thread to take the next available box from the queue of unprocessed boxes, instead of deciding ahead of time which boxes each thread will process. This is important because the boxes are varying in size. We generally divide even the coarsest level into multiple boxes so that it can be processed in parallel by multiple threads.
\subsubsection{Kreiss-Oliger Dissipation} \label{sect:kreissoliger}
In a finite difference scheme, instabilities can arise from the appearance of high frequency spurious modes. Furthermore, regridding generates errors an order higher than the typical error of the evolution operator, hence it is doubly crucial that we control these errors. The standard prescription to deal with this is to implement some form of numerical dissipation to damp out these modes. $\mathtt{GRChombo}$ implements $N=3$ Kreiss-Oliger \cite{TUS:TUS1547} dissipation. In this scheme, for all evolution variables $u\in \{\tilde{A}_{ij},\tilde{\gamma}_{ij},K,\chi,\Theta,\tilde{\Gamma}^{i}\}$, the evolution equations are modified as follows
\begin{equation}
\partial_t u_m \rightarrow \partial_t u_m + \frac{\sigma}{64 \Delta x}(u_{m+3}-6u_{m+2}+15u_{m+1}-20u_{m}+15u_{m-1}-6u_{m-2}+u_{m-3}),
\end{equation}
where $m\pm n$ labels the grid point $m$, $n$ the total offset from $m$ and $\sigma$ is an adjustable dissipation parameter usually of the order ${\cal O}(10^{-2})$. This 3rd order scheme is accurate as long as the integration order of the finite difference scheme is 5 or less (which it is in our implementation using 4th order Runge-Kutta).
\subsubsection{Boundary Conditions} \label{sect:BC}
$\mathtt{GRChombo}$ supports both periodic (in any direction) boundary conditions, as well as any particular boundary conditions the user may want to specify (such as Neumann or Dirichlet types). A particular popular type of boundary condition is the so-called Sommerfield \cite{Alcubierre:2002kk} boundary condition, where out-going radiation is dissipated away. For any field $f$, we impose the condition at the boundary
\begin{equation}
\frac{\partial f}{\partial t} = -\frac{v x_i}{r} \frac{\partial f}{\partial x_i} -v\frac{f-f_0}{r}
\end{equation}
where $r= \sqrt{x_1^2+x_2^2+x_3^2}$ is the radial distance from the center of the grid, $f_0$ is the desired space-time at the boundary (typically Minkowski space for asymptotically flat spacetimes) and $v$ the velocity of the ``radiation'', which is typically chosen to be 1.
\subsubsection{Initial conditions} \label{sect:IC}
$\mathtt{GRChombo}$ supports several ways of entering initial conditions.
\begin{itemize}
\item{Direct equations -- Initial conditions which are described by known analytic equations, such as the Schwarzchild solution, can be entered directly in equations form.}
\item{Checkpointing -- The {\tt HDF5} format output files from $\mathtt{GRChombo}$ doubles as checkpointing files. A run can simply be continued from any previous state as long as its {\tt HDF5} output file is available. }
\item{Entering from data file -- $\mathtt{GRChombo}$ allows one to insert data from a file.}
\item{Relaxation -- $\mathtt{GRChombo}$ has a rudimentary capability to solve for the initial metric given some initial mass distribution, and assuming a moment of time symmetry and conformal flatness. Given a guess for $\chi$, $\mathtt{GRChombo}$ relaxes it to the correct initial metric using a dissipation term which is proportional to a user chosen dissipation coefficient times the Hamiltonian constraint.}
\end{itemize}
The initial conditions used in this paper are mostly analytic or approximate analytic solutions, and so are entered directly into the code. In the critical collapse, a Mathematica numerical solution as a function of the radius is interpolated onto the initial grid.
\section{Using $\mathtt{GRChombo}$ for new physics} \label{sec-newphys}
The primary advantage of $\mathtt{GRChombo}$ over existing publicly available codes is its robust AMR ability. In this section we discuss several physical systems which can be studied with \texttt{GRChombo} but that would be hard (if not impossible) to simulate using codes based on moving-box mesh refinement. Most of the examples discussed in this section are still ``work in progress" by the authors and the results will be presented in separate publications.
\subsection{Asymmetric scalar field bubbles}
\label{sec-newbubs}
One of the most fascinating and as yet not fully understood aspects of general relativity is the appearance of critical phenomenon in gravitational collapse as first discovered by Choptuik \cite{Choptuik:1992jv}. A comprehensive review can be found in \cite{Gundlach:2007gc}.
Briefly, if we have an initial configuration, such as a Gaussian shaped bubble of scalar field, and allow this to evolve under the action of gravity, the result will be either the formation of a black hole, or dispersal of the field to infinity depending on the ``strength" of the initial data. Varying any one initial parameter $p$ of the configuration (such as the height of the bubble), one finds that there is a critical point $p^*$ at which the transition between the two end states occurs, and that the mass of the black hole created on the supercritical side follows the scaling relation,
\begin{equation}
M \propto (p - p^*)^{\gamma},
\end{equation}
where the scaling constant $\gamma$ is universal in the sense that it does not depend on the choice of family of initial data. For a massless scalar in a spherically symmetric collapse, $\gamma$ has been numerically determined to be around 0.37.
The other key phenomenon which is observed is that of self-similarity in the solutions, or ``scale-echoing''. Close to the critical point, and in the strong field region, the value of any gauge independent field $\phi$ at a point $x$ and time $T$ exhibits the following scaling relation,
\begin{equation}
\phi (x,t) = \phi(e^{\Delta} x, e^{\Delta} T) ,
\end{equation}
where $\Delta$ is a dimensionless constant with another numerically determined value of 3.44 for a massless scalar field in the spherical case. The time $T$ here is measured ``backwards" - it is the difference between the critical time at which the formation of the black hole occurs and the current time, with time being the proper time measured by a central observer.
What one sees is therefore that, as the time nears the critical time by a factor of $e^{\Delta}$, the same field profile is seen but on a scale $e^{\Delta}$ smaller. This scale-echoing may be either continuous or discrete. In Choptuik's seminal paper \cite{Choptuik:1992jv}, a 1+1 adaptive mesh code was used to study such behaviour near the critical point. Since then there has been some progress in studying the phenomenon in non spherically symmetric cases, see \cite{Abrahams:1993wa,Choptuik:2003ac,Sorkin:2010tm,Healy:2013xia,Hilditch:2013cba,Hilditch:2015aba}, but progress in making firm conclusions has been slower than expected, due to the extremely high refinements required to study the stages of the collapse, which are magnified three-fold in full $3+1$ codes.
We are currently investigating the problem of asymmetric bubble collapse with $\mathtt{GRChombo}$. A snapshot of the evolution of one such example is shown in Figure \ref{fig-BubCol}. The mesh refines so that the field profile is consistently resolved as the critical time is approached. Note that since the profile is highly irregular, with a range of length scales represented in different (disconnected) regions, $\mathtt{GRChombo}$ provides a significant advantage over moving-box mesh refinement in terms of computational cost. In addition to the adaptation of the mesh to the local curvature, $\mathtt{GRChombo}$ can automatically add in new levels of refinement during the evolution, which allows the scale-echoing behaviour to be probed consistently as the profile shrinks even in the absence of spherical symmetry. The results will be discussed in detail in a separate publication.
\begin{figure}[t]
\subfigure{
\includegraphics[width=.425\textwidth]{SFMesh.png}
}
\subfigure{
\includegraphics[width=.45\textwidth]{SFMeshelev.png}
}
\caption{An asymmetric scalar field bubble collapse: In the left image, a slice through the centre of the field bubble is shown at a particular time during the collapse, with the instantaneous mesh superimposed. This same slice is shown in 3D in the right image to better illustrate how the mesh is adapted to the curvature. The AMR efficiently tracks the scales in the profile. In addition, new levels are added as the profile shrinks, leading to a more efficient computation in terms of memory and computational resources, and requiring less ``human input'' ahead of and during the run.
\label{fig-BubCol}}
\end{figure}
\subsection{Higher dimensional black holes/anti-de Sitter}
\label{sec-newother}
In recent years it has been realized that the dynamics of general relativity, even in vacuum, beyond the traditional asymptotically flat four-dimensional setting is much richer than previously anticipated. Some of these new directions involve considering more than four spacetime dimensions and/or new boundary conditions, such anti-de Sitter (AdS) or Kaluza-Klein (KK) asymptotics. Black holes, as primary objects in any theory of gravity, in these new set ups exhibit two important new physical phenomena: Firstly, black holes with topologically non-spherical horizons are possible. The black ring of \cite{Emparan:2001wn} is the first example of an asymptotically flat vacuum black hole with a non-spherical horizon. By now it is clear that this solution is the tip of the iceberg, and many more new types of black holes are known to exist. Secondly, vacuum black holes can be dynamically unstable under gravitational perturbations, as Gregory and Laflamme first showed in the case of black strings in asymptotically KK spacetimes \cite{Gregory:1993vy} . Rapidly spinning asymptotically flat vacuum (and AdS) black holes suffer from these Gregory-Lafllamme-type-of instabilities \cite{Emparan:2003sy,Dias:2009iu,Dias:2010eu}, and new types of non-axisymmetric instabilities \cite{Shibata:2010wz}. In fact, anti-de Sitter itself is non-linearly unstable to the formation of black holes \cite{Bizon:2011gg}, and the process is turbulent. There is a lot of interest in studying the dynamics of gravity, and black holes, in AdS motivated by the gauge/gravity correspondence \cite{Maldacena:1997re}. See \cite{Emparan:2008eg,Horowitz:2012nnc} for some (relatively recent) reviews with references.
In a remarkable paper, \cite{Lehner:2010pn} studied the endpoint of the Gregory-Laflamme instability of black strings in five-dimensions. This paper gave convincing evidence that the black string would pinch off in finite asymptotic time, thus providing a potential counter-example of the weak cosmic censorship conjecture in non-asymptotically flat vacuum spacetimes. The evolution the Gregory-Laflamme instability for black strings showed that the horizon develops a fractal structure, with thin necks connecting bulges at different scales. Moreover, the non-linear instability of AdS is of a turbulent nature \cite{Bizon:2011gg}, and in fact AdS black holes also suffer from turbulent type-of-instabilities \cite{Holzegel:2011uu,Carrasco:2012nf,Adams:2013vsa,Yang:2014tla}. A common feature in all these instabilities is that in the fully non-linear regime, new length scales are dynamically generated and a priori one does not know where they will appear. Therefore, if one wants to use numerical GR to determine the endpoints of these instabilities, one needs a numerical method that can automatically resolve these newly generated length scales. Therefore, it seems that full AMR is not an option but a necessity.\footnote{The AMR capabilities of the \texttt{PAMR/AMRD} code were essential in \cite{Lehner:2010pn}.} In addition, with \texttt{GRChombo} one should be able to simulate higher dimensional spacetimes with no symmetry assumptions using the fact that the \texttt{Chombo} core is dimension independent up to six spatial dimensions. This should find applications to the AdS/CFT correspondence, where one may be interested in simulating $(4+1)$-dimensional asymptotically AdS spacetimes in full generality.
Some of the aforementioned instabilities are currently being investigated using \texttt{GRChombo} and will be discussed elsewhere. In Fig. \ref{fig-Ring} we display snapshots of the meshes that are dynamically generated by \texttt{GRChombo} during evolution for the case of an unstable black ring (\textit{left}) and a higher dimensional black hole (\textit{right}). \texttt{GRChombo} not only can adapt the mesh to the non-trivial topology of the horizon but it also adds new levels in regions where new structures appear during the evolution. This essential capability ensures that all relevant length scales are correctly resolved whilst keeping the computational cost of the simulation under control. For these simulations, moving-box mesh refinement would be prohibitively expensive and hence it is not a realistic option. Previous works on higher dimensional black hole physics in asymptotically flat spaces using moving boxes include \cite{Zilhao:2010sr,Yoshino:2009xp}. In AdS, the code of \cite{Bantilan:2012vu} is based on \texttt{PAMR/AMRD}, whilst \cite{Chesler:2013lia} uses a fixed domain decomposition and pseudospectral discretisation.
\begin{figure}
\begin{center}
\subfigure{
\includegraphics[scale=0.215]{ring_mesh.png}
}
\subfigure{
\includegraphics[scale=0.215]{mp6_mesh.png}
}
\caption{Snapshots of the rotation plane of dynamically generated meshes by \texttt{GRChombo} during the evolution of the instabilities of a black ring (\textit{left}) and of a rapidly spinning spherical black hole (\textit{right}). AMR ensures that the mesh is adapted to the non-trivial topology of the horizon and new levels are added where new structure appears. This is essential in order to have enough resolution where it is needed while keeping the computational cost of the simulation under control.
\label{fig-Ring}}
\end{center}
\end{figure}
\subsection{$N$-body problems}
\label{sec-newBHs}
Simulations of single black holes and binary mergers often make use of symmetries in the problem or Newtonian approximations to predict the levels of resolution required at each point in space, for each time-step. In three (or more) body problems, the trajectories of the objects are generally not known ahead of time and must be calculated numerically. The shift vector can be used to predict the movement of black hole centres locally, but this tracking must be added in for each black hole and the boxes adjusted accordingly. For $\mathtt{GRChombo}$, it is trivial to add multiple black holes to a spacetime, without actively tracking their central points, and so these many-body systems are as easy to set up as binary ones (although they clearly require greater computational resource to run). For example, the $\mathtt{GRChombo}$ mesh can be seen adapting at each timestep in the triple black hole merger shown in Figure \ref{fig-TriBH}.
\begin{figure}
\begin{center}
\subfigure{
\includegraphics[width=.3\textwidth]{BH1a.png}
}
\subfigure{
\includegraphics[width=.3\textwidth]{BH2a.png}
}
\subfigure{
\includegraphics[width=.3\textwidth]{BH3a.png}
}
\caption{Triple black hole merger: Three black holes are evolved with $\mathtt{GRChombo}$. The mesh is shown, which has adapted to the local curvature in $\chi$, the variable plotted.
\label{fig-TriBH}}
\end{center}
\end{figure}
Another (albeit rather contrived) example is the ring of black holes shown in Figure \ref{fig-RingBH}. The set up of this ring was no more difficult than that of the binary or triple black hole spacetime -- $\mathtt{GRChombo}$ automatically remeshes the grid given a set of analytic initial conditions without further user intervention, and there is no need to try and predict the paths of the black holes individually. It is clear that other fields of research, such as magnetohydrodynamics, would benefit greatly from this ability to maintain a consistent level of resolution throughout a simulation, and make this resolution follow the inherently unpredictable movement in a many body system.
\begin{figure}
\begin{center}
\subfigure{
\includegraphics[width=.3\textwidth]{BHRing1.png}
}
\subfigure{
\includegraphics[width=.3\textwidth]{BHRing2.png}
}
\subfigure{
\includegraphics[width=.3\textwidth]{BHRing3.png}
}
\caption{Multiple Black Hole merger: A ring of black holes are evolved with $\mathtt{GRChombo}$. The mesh is shown, which has adapted to the local curvature in $\chi$, the variable plotted.
\label{fig-RingBH}}
\end{center}
\end{figure}
\section{Testing $\mathtt{GRChombo}$}
\label{sec-tests}
We detail the results of the standard Apples with Apples tests \cite{Babiuc:2007} in Sec. \ref{sec-apples} when turning off AMR and using fixed resolution grids. In Sec. \ref{sec-vacblack} we turn on the AMR abilities of the code and demonstrate that it can stably evolve spacetimes containing black-hole-type singularities. In \ref{sec-headon} we study convergence of our code in a head on collision of two black holes. In Sec. \ref{sec-choptuik} we demonstrate the ability of the code to evolve matter content by considering scalar fields with gravity, by recreating the results of the sub-critical and critical cases of Choptuik scalar field collapse detailed in \cite{AlcubierreBook}. In Sec. \ref{sec-scaling} we discuss the weak scaling properties of the code on the Mira supercomputer.
The test figures referred to in this section can be found in Appendix \ref{sec-appendix}.
\subsection{Apples with Apples tests}
\label{sec-apples}
In this section we describe the results of applying the code to the standard Apples with Apples tests in \cite{Babiuc:2007}. Here we give a brief description of the key features of the tests, but the reader should refer to the paper for full specifications. Where we do not specify details, our treatment can be assumed to follow that of the standard tests. The AMR capabilities of the code are not utilised in these tests, which were designed for a uniform resolution, in order to make our results comparable to other codes. (We consider the effects of regridding on code performance in Section \ref{sec-vacblack}.)
\subsubsection{Robust stability test}
\label{subsubsec:rob}
The robust stability test introduces small amounts of random noise, scaled with the grid spacing, to all of the evolution variables, in order to test the code's robustness against numerical noise. The test was conducted at resolutions of $\rho = 4$, $\rho=2$ and $\rho=1$, corresponding to grid spacings of 0.005, 0.01 and 0.02 respectively. No dissipation was added in the test.
As shown in Figure \ref{fig-robust}, the error growth in the evolution variables did not increase with increasing grid resolution, and the Hamiltonian constraint $H$ did not grow more for higher resolutions. Therefore, we conclude that the test is passed.
\subsubsection{Linear wave test}
A wave of fixed amplitude is propagated across the grid in the $x$-direction with periodic boundary conditions. The amplitude is small enough that the non-linear terms are below numerical precision, such that the behaviour under the Einstein equation is approximately linear. The test measures the errors in magnitude and phase introduced by the code after 1000 crossing times.
As can be seen from Figure \ref{fig-linear}, this error is 12 orders of magnitude smaller than the signal and therefore negligible.
\subsubsection{Gauge wave tests}
The BSSN formulation is known to produce unsatisfactory results for the gauge wave tests. $\mathtt{GRChombo}$ is no different in this respect. As can be seen in Figure \ref{fig-gauge}, it becomes unstable after around 50 crossing times, with the Hamiltonian constraint increasing exponentially, even for a relatively small initial amplitude of the gauge wave of $A = 0.1$.
As was shown in \cite{Alic:2011gg} stability can be achieved by adding in the CCZ4 constraint damping terms. $\mathtt{GRChombo}$ shows exactly this behaviour (figure \ref{fig-gauge}).
\subsubsection{Gowdy wave test}
The Gowdy wave evolves a strongly curved spacetime: an expanding vacuum universe containing a plane polarised gravitational wave propagating around a 3-torus.
In the expanding direction we use the analytic gauge, $\partial_t \alpha = -\partial_t \sqrt{\gamma_{zz}}$. The collapsing direction is evolved starting at $t=t_0$ with harmonic slicing for the lapse and zero shift. A Kreiss-Oliger dissipation coeffecient of $\sigma=0.05$ was used in both directions.
The results for both the BSSN and CCZ4 codes in the collapsing direction are shown in Figure \ref{fig-gowdy-collapsing}, and in the expanding direction in Figure \ref{fig-gowdy-expanding}.
As is found in the Apples with Apples tests \cite{Babiuc:2007} for other simple BSSN codes, $\mathtt{GRChombo}$ with BSSN and CCZ4 gives a less than satisfactory performance in this test in the expanding direction. The evolution is stable for approximately the first 30 crossing times, after which high frequency instabilities develop and cause code crash, due to the exponentially growing $\gamma_{zz}$ component. In \cite{Babiuc:2007} it was found that this behaviour of BSSN could be controlled with dissipation, but that long term accuracy was not achievable.
In the contracting direction the evolution is stable for the full 1000 crossing times and we were able to confirm the convergence of our code. As shown in Figure \ref{fig-gowdy-convergence}, both BSSN and CCZ4 exhibit 4\textsuperscript{th} order convergence initially. While convergence is never lost, the order is reduced at later times. This is similar to the behaviour found in \cite{Babiuc:2007} and \cite{Cao:2011fu}.
\subsection{Vacuum black hole spacetimes}
\label{sec-vacblack}
In this subsection we show that our code can stably evolve spacetimes containing black holes.
All the simulations presented here used the BSSN formulation of the Einstein equations, along with the gamma-driver and alpha-driver gauge conditions. Adding CCZ4 constraint damping gives better performance for the Hamiltonian constraint, as would be expected, but the results are broadly similar and so are not presented here. Unless otherwise stated, we perform the simulations with up to 8 levels of refinement and we based our tagging/regridding criterion, \eqn{eqn:tagging}, on the value of $\chi$. We emphasise that the purpose of this subsection is to demonstrate that we can stably evolve black hole spacetimes, but we are not interested in extracting gravitational wave data or in studying convergence; this will be done in the next subsection.
Where we refer to taking an $L^2$ norm of the Hamiltonian constraint $H$ in a test, this is calculated as follows (using the weighted variable sum function in VisIt):
\begin{equation}
|| H ||_{2} = \sqrt{\sum_i m_i H_i^2} , \label{eqn:L2norm}
\end{equation}
where $m_i$ is $V_i/V_{tot}$, the fraction of the total grid volume $V$ occupied by the $i$th box. Where the grid contains a black hole, we excise the interior by setting $H$ to zero within the region in which the lapse $\alpha$ is less than 0.3 (which is an approximate rule of thumb for the location of an event horizon for a black hole in the moving puncture gauge). The difference in the results is small, since the error norm is dominated by regridding errors at the boundaries between meshes. We also exclude the values on the outer boundaries of the grid, which can distort the results in cases where periodic boundaries are used.
\subsubsection{Schwarzschild black hole}
First we evolve a standard Schwarzschild black hole in isotropic gauge, with a conformally flat metric, the lapse initially set to one everywhere, and the conformal factor $\chi$:
\begin{equation}
\chi = \left( 1+\frac{M}{2r} \right)^{-2} \label{eqn:SCchi}
\end{equation}
In this simulation, we chose the outer boundary of the domain to be at $600M$ and the spatial resolution in the coarsest mesh is $10M$. We impose Sommerfeld boundary conditions. The initial value of $\chi$ through a slice is shown in Figure \ref{fig-profile}. We see the expected ``collapse of the lapse" at the singularity and the solution quickly stabilises into the ``trumpet" puncture solution described in \cite{Hannam:2008sg}. We find an apparent horizon and are able to evolve the black hole stably and without code crash for well over $t = 10000M$ time steps as shown in Figure \ref{fig-SC} (left). In this figure we show the $L^2$ norm of the Hamiltonian constraint across the whole grid, and it remains bounded throughout the evolution.
We monitor the ADM mass of the black hole by integrating over a surface near the asymptotically flat boundary, as seen in Figure \ref{fig-SC} (right). We also monitor the angular momentum and linear momentum of the black hole, and find that these remain zero as expected, as shown in Figure \ref{fig-SC} (right). These simple ADM measures rely on asymptotic flatness at the surface over which they are integrated, and so are sensitive to errors introduced by reflections at the boundaries, initial transients from approximate gauge choice or if the black hole is moving nearer the boundary (as in the boosted case). They are therefore less reliable as the simulation progresses, and we use them simply to confirm that we are evolving the correct spacetime initially.
\begin{figure}
\begin{center}
\includegraphics[width=.6\textwidth]{BHchi.png}
\caption{The profile for $\chi$ through a slice perpendicular to the z axis is shown.
\label{fig-profile}}
\end{center}
\end{figure}
\subsubsection{Kerr black hole}
In this sub-subsection we present the results of a simulation of the Kerr black hole spacetime in quasi isotropic gauge as in \cite{Brandt:1996si} with the angular momentum parameter $a = J/M$ set to 0.2. The domain size was chosen to be $(320M)^3$ and the grid spacing in the coarsest level was $4M$. We impose periodic boundary conditions for simplicity, which limits the duration of the simulation due to boundary effects.
In Figure \ref{fig-KE} (left) we show the $L^2$ norm of the Hamiltonian constraint throughout the evolution. This plot shows that the amount of constraint violation remains stable during the simulation. In the right panel of Figure \ref{fig-KE} we display the ADM measures for the three components of the angular momenta and the mass. This Figure shows that these quantities remain (approximately) constant during the simulation.
\subsubsection{Boosted black hole}
In this sub-subsection we evolve a boosted black hole using the perturbative approximation from \cite{ShapiroBook}, with initial momenta set to $P_x = 0.02$, $P_y = 0.02$ and $P_z = 0.0$. The domain size was chosen to be $(640)^3$, with spatial resolution in the coarsest grid of $4M$. We imposed periodic boundary conditions at the outer boundaries of the domain. The black hole moves across the grid diagonally as expected, as is seen in Figure \ref{fig-Boostmove}.
\begin{figure}
\begin{center}
\subfigure[Initial position]{
\includegraphics[width=.3\textwidth]{Boost1.pdf}
}
\subfigure[Position at $t=100$]{
\includegraphics[width=.3\textwidth]{Boost2.pdf}
}
\subfigure[Grid at $t=100$]{
\includegraphics[width=.3\textwidth]{Bgrid2.pdf}
}
\caption{Boosted black hole, movement: The boosted black hole moves across the grid diagonally with initial momenta of $P_x = 0.02$, $P_y = 0.02$ and $P_z = 0.0$, as expected, and the grid adapts to this movement, with the high resolution grids following the movement.
\label{fig-Boostmove}}
\end{center}
\end{figure}
In the left panel of Figure \ref{fig-BoostBH} we show the $L^2$ norm of the Hamiltonian constraint across the domain as a function of time. This plot shows that the constraints remain bounded throughout the simulation. In the right panel of Figure \ref{fig-BoostBH} we display the components of the ADM linear momentum during the simulation. In the continuum limit they should be constant and in our simulation they are indeed approximately constant.
\subsubsection{Binary inspiral}
\label{subsubsec:binary}
In this sub-subsection we superpose the initial perturbative solution for two boosted black holes in \cite{ShapiroBook}, sufficiently separated, to simulate a binary inspiral merger. The domain size was $(200M)^3$ with a grid spacing in the coarsest level of $5M$. As in some of the previous tests, for simplicity we imposed periodic boundary conditions at the outer boundaries of the domain.
We are able to evolve the merger stably such that the two black holes merge to form one with a mass approximately equal to the sum of the two. The progression of the merger is shown in Figure \ref{fig-BiBH}. The time evolution of the $L^2$ norm of the Hamiltonian constraint across the grid is shown in Figure \ref{fig-Binary}. Again this remains stable throughout the simulation.
\begin{figure}
\begin{center}
\subfigure{
\includegraphics[width=.4\textwidth]{movie0005.png}
}
\subfigure{
\includegraphics[width=.4\textwidth]{movie0015.png}
}
\subfigure{
\includegraphics[width=.4\textwidth]{movie0028.png}
}
\subfigure{
\includegraphics[width=.4\textwidth]{movie0032.png}
}
\caption{Binary Black Hole merger: Two black holes are evolved with $\mathtt{GRChombo}$. The final stages of the merger are shown.
\label{fig-BiBH}}
\end{center}
\end{figure}
\subsection{Convergence test: head on collision of two black holes}
\label{sec-headon}
In this subsection we simulate the head on collision of two black holes and analyse the convergence of the code. We set up Brill-Lindquist initial data consisting of two static black holes of mass $0.5M$ with a separation of $10M$, located at the centre of the computational domain. We extract the gravitational wave signal (see below). An initial burst of radiation is seen, which is a property of the superimposed initial data, prior to the main signal. Even though this set up could be simulated in axisymmetry, we have evolved the system without imposing any symmetry assumptions. So the results below correspond to a full $3+1$ simulation with \texttt{GRChombo}.
We performed runs at three different resolutions with 7 levels of refinement, each level having half the grid spacing as
the previous one. The grid spacings were \linebreak $0.03125M/4M$ for the low resolution run, $0.02083M/2.66667M$ for the medium resolution run and $0.01563M/2M$ for the high resolution run. Here the numbers refer to the resolution on the finest/coarsest grids respectively. The outer boundary of the domain is located at $200M$ and we impose periodic boundary conditions for simplicity. This puts an upper bound on the time up to which we can evolve the system before boundary effects influence physical observables.
\begin{figure}
\begin{center}
\subfigure{
\includegraphics[width=0.8\textwidth]{psi4.pdf}
}
\subfigure{
\includegraphics[width=0.75\textwidth]{psi4err.pdf}
}
\caption{Convergence test of head on collision. \textit{Top}: The real part of the $\ell=2$, $m=0$ mode of $r\Psi_4$ on the sphere of radius $R=60M$. \textit{Bottom}: Differences in the real part of the $\ell=2$, $m=0$ mode of $r\Psi_4$ between three different resolutions. We also show the data rescaled by a factor consistent with either third ($\times 4.11$) or fourth ($\times 5.64$) order convergence.
\label{fig-psi4}}
\end{center}
\end{figure}
In Figure \ref{fig-psi4} (top) we display the real part of the $\ell=2$, $m=0$ mode of $r\Psi_4$ extracted on a sphere of radius $R=60M$ using 4th order interpolation. We use $320$ grid points in both the polar and azimuthal directions on the extraction sphere. Following \cite{Loffler:2011ay}, we test convergence by comparing a physical quantity $\Psi$ at different resolutions. The convergence is of order $Q$ if for a set of grid spacings $h_1$, $h_2$, $h_3$, the differences between the numerically computed physical quantity $\Psi$ at successive resolutions satisfy
\begin{equation}
\frac{\Psi_{h_1} - \Psi_{h_2}}{\Psi_{h_2} - \Psi_{h_3}} = \frac{h_1^Q - h_2^Q}{h_2^Q - h_3^Q}\,.
\end{equation}
With the resolutions used in these runs, assuming $4^\textrm{th}$ order convergence the above factor is $\approx 5.953$, whilst assuming $3^\textrm{rd}$ order convergence the factor is $\approx 4.115$.
The gravitational wave content of the superimposed initial data is reflected in the non-zero initial signal. The collision of the two black holes takes place at $t\sim50$ on this plot, so the signal before this collision time should be regarded as unphysical. As can be seen in this plot, the results for the two higher resolutions are indistinguishable on the scale employed here, whilst the lowest resolution shows a very slight drift towards later times, but is still in very good agreement. The bottom plot in Figure \ref{fig-psi4} shows the absolute value of the difference between $r\Psi_4$ computed low and medium resolution (solid blue), medium and high resolution (solid black), and this latter curve scaled up by the convergence factor assuming $3^\textrm{rd}$ (dotted orange) and $4^\textrm{th}$ (dotted red) order convergence. This plot shows that in the highly dynamical stages of the evolution, when there is a lot of regridding and the boxes move around the domain, the convergence is closer to $3^\textrm{rd}$ order. On the other hand, when the system has nearly settled, and hence the boxes do not move much, the convergence order is closer to $4$. We can explain this loss of convergence due to regridding because in the interpolation used in \texttt{GRChombo} only the values of the functions are matched across levels, but not their derivatives. We hope to improve this aspect of the code in the future.
\subsection{Choptuik scalar field collapse}
\label{sec-choptuik}
We now test the scalar field part of the code, by simulating the Choptuik scalar field collapse as described in \cite{AlcubierreBook} and illustrated in Figure \ref{fig-chop3D}. The referenced description is for a 1+1 simulation which is evolved using a constrained evolution, such that the lapse $\alpha$ and the single degree of freedom for the metric, $A$, are both solved for on each slice using ODEs obtained from the constraint equations. The only degrees of freedom which are truly evolved are those of the field, $\phi$, $\Psi$ and $\Pi$.
Our evolution is carried out using the full $3+1$ BSSN equations, without assuming or adapting coordinates to spherical symmetry. We are able to replicate the results obtained in \cite{AlcubierreBook}, subject to some minor differences due to the fact that we evolve with the puncture gauge rather than according to the maximal slicing constraint equation, see Figures \ref{fig-chopsnaps} and \ref{fig-chopalp}, which can be found in Appendix \ref{sec-appendix}. Videos of the results can be viewed via our website at http://grchombo.github.io.
We see that $\mathtt{GRChombo}$ can accurately evolve the field profile in the presence of gravity, and copes with the collapse of the supercritical case into a singularity, without code crash. For the subcritical cases we see that the field disperses as expected.
\begin{figure}
\begin{center}
\subfigure{
\includegraphics[width=.3\textwidth]{Choptuik1}
}
\subfigure{
\includegraphics[width=.3\textwidth]{Choptuik2}
}
\subfigure{
\includegraphics[width=.3\textwidth]{Chopsup.png}
}
\caption{In Choptuik scalar field collapse, the initial specially symmetric configuration in the first figure (which shows the values on a slice perpendicular to the z axis) collapses, splitting into an ingoing and an outgoing wave as seen in the second image. If the amplitude of the initial perturbation is greater than a certain critical value, the ingoing wave will result in the formation of a black hole, as seen from the output of the apparent horizon finder in the third figure, which shows that an apparent horizon with a mass of about 0.25 has formed by $t=15$.
\label{fig-chop3D}}
\end{center}
\end{figure}
\subsection{MPI Scaling properties}
\label{sec-scaling}
We now turn to the performance aspects of $\mathtt{GRChombo}$. Here we perform a number of scaling tests to show that our code can exploit the parallelism offered in modern supercomputers to a reasonable extent. Whilst {\tt Chombo} does have the capability to partially utilise threads through hybrid OpenMP routines, we will limit our attention to pure MPI mode in these tests, as we have found that this gives significantly smaller run-to-run performance variations.
Our strong scaling test is performed using a head on binary black hole system. We set up Brill-Lindquist initial data for two static black holes of mass $0.5M$, with a separation of $6M$. Our overall computational domain is a box of size $160M$, and at the coarsest level, we fix the total number of grid points to $320$ in each direction, giving a grid spacing of $0.5M$. The centre of mass of the system is at the centre of the domain. For the mesh refinement, we fix the total number of levels to six. The simulation is allowed to run up to the time of $2M$. The bulk of this test was performed on the SuperMike-II cluster at the Louisiana State University. Each compute node consists of two 2.6GHz 8-core Sandy Bridge Xeon processors, connected via a InfiniBand QDR fabric. We fix the computational load across all jobs and vary the number of core count from 16 to 2048. Our data in Figure \ref{fig-StrongScaling} shows excellent strong scaling up to 200 cores on this cluster. We continue to see a reasonable speedup up to around 1000 cores for this particular problem.
\begin{figure}[htp]
\begin{center}
\includegraphics[height=8cm]{strongscaling.pdf}
\caption{
Strong scaling behaviour of GRChombo on the SuperMike-II cluster at the Louisiana State University. The code achieves excellent strong scaling up to 200 cores, and a useful scaling up to around 1000 cores.
\label{fig-StrongScaling}}
\end{center}
\end{figure}
Of course, in a production environment, it is often desirable to use additional cores to be able to run a larger simulation, rather than to speed up a problem of fixed size. In this scenario, weak scaling behaviour is of interest. We begin at 1024 cores with an identical setup to that in the strong scaling test. We then scale up the number of grid points at the coarsest level proportional to the increase in core count up to 10240, whilst adjusting the tagging threshold in order to maintain the shape and size of the refined regions. We also adjusted the time step size (i.e. the Courant factor) so that each simulation would reach the target stop time in the same number of steps. We use the Mira Blue Gene/Q cluster at the Argonne National Laboratory for this due to the larger number of cores available. Figure \ref{fig-WeakScaling} shows a less-than-perfect scaling behaviour in this setup, with the main bottleneck appearing in the regridding and box generation stages. We are working together with the developers of {\tt Chombo} to improve this aspect of the code performance. It is worth noting, however, that even in its current state the code still shows a useful level of scalability: the wallclock time increases by less than 2x over the 10x increase in core count.
\begin{figure}[htp]
\begin{center}
\includegraphics[height=8cm]{weakscaling.pdf}
\caption{
Weak scaling behaviour of GRChombo on the Mira Blue Gene/Q cluster at the Argonne National Laboratory over a 10x increase in core count.
\label{fig-WeakScaling}}
\end{center}
\end{figure}
\subsection{Performance comparison}
\label{sec-performance}
Lastly, we demonstrate that $\mathtt{GRChombo}$'s performance on standard $3+1$ black hole problems is comparable to that of an existing numerical relativity code.
Our comparison target is the $\texttt{Lean}$ code~\cite{Sperhake:2006cy,Zilhao:2010sr}, a $3+1$ numerical relativity code designed to evolve four and higher dimensional vacuum spacetimes. $\texttt{Lean}$ is based on the {\texttt{Cactus}} computational toolkit~\cite{Cactuscode:web} and realises moving-box mesh refinement via the \texttt{Carpet} package~\cite{Schnetter:2003rb,CarpetCode:web}, both of which are part of the open-source {\texttt{Einstein Toolkit}}~\cite{Loffler:2011ay,EinsteinToolkit:web}. Initial data is constructed either analytically or numerically by employing the {\texttt{TwoPunctures}} spectral solver~\cite{Ansorg:2004ds}. In order to track apparent horizons, $\texttt{Lean}$ makes use of~\texttt{AHFinderDirect}~\cite{Thornburg:2003sf,Thornburg:1995cp}.
The $\mathtt{GRChombo}$ setup is identical to that in the strong scaling test as detailed in Sec. \ref{sec-scaling}. The $\texttt{Lean}$ code is subject to the limitation of Carpet, where successive levels may only occur a collection of nested-box hierarchies, whose sizes are typically related by a power of two. In this case, we first fix boxes of side lengths $160$, $80$, $40$ and $20M$ at the centre of the domain, encompassing both black holes, then fix further boxes of side lengths 5 and $2.5M$ centred at each of the black holes. During the evolution, $\texttt{Lean}$ has the capability to track the black holes and move or merge the finer boxes as appropriate, however the shape and size of the boxes remain unchanged. The $\mathtt{GRChombo}$ code is not subject to this box structure limitation, and therefore we simply tune the regridding threshold so that the size of the finest level matches that of the $\texttt{Lean}$ setup. We make no attempt to match the sizes of the intermediate levels as this would defeat the spirit of fully-flexible AMR.
\begin{figure}[htp]
\begin{center}
\includegraphics[height=8cm]{LeanComparison.pdf}
\caption{
Runtime and scaling comparison between $\mathtt{GRChombo}$ (orange) and $\texttt{Lean}$ (blue). The leftmost data points show disproportionately large wallclock times as the machine becomes memory-limited at this core count.
\label{fig-LeanComparison}}
\end{center}
\end{figure}
Our comparison tests were performed on the COSMOS VIII shared memory facility. Both codes were executed on the same SGI UV1000 machine, utilising up to 60 Nehalem EX 2.67GHz CPUs with 6 cores per CPU, giving up to 360 cores in total. In all of these runs, we pin one MPI rank to each core and disabled all checkpointing activity since we wish to exclude I/O bottleneck. We allowed the simulation to run up to coordinate time $t = 2$, and measured the wall-clock time taken to execute the time evolution portion of the code (i.e. we excluded the time spent during initial setup).
Within the range of 150-360 cores, both $\mathtt{GRChombo}$ and $\texttt{Lean}$ exhibit similar performance and strong scaling characteristics (figure \ref{fig-LeanComparison}). Below 150 cores, we cannot meaningfully test the strong scaling behaviour as the machine becomes memory-limited. We have not performed this comparison on a larger cluster due to the lack of resource availability, but we have no reason to expect any significant difference provided that the problem size is also scaled up appropriately. Having said this, we believe that a framework like Cactus probably remains the better choice when it comes to these standard problems, owing to the wealth of existing tools and resources and a more mature community of users. Instead, we intend for $\mathtt{GRChombo}$ to be complementary to existing numerical relativity codes in order to open up new avenues of research by enabling a wider range of problems to be tackled at a feasible level of resources (see section \ref{sec-newphys} for examples of such problems).
\section{Discussion}
\label{sec-dis}
In this paper, we introduced and described $\mathtt{GRChombo}$, a new multi-purpose numerical relativity code built using the {\tt Chombo} framework. It is a $3+1$D finite difference code based on the BSSN/CCZ4 evolution scheme. It supports Berger-Collela type AMR evolution with Berger-Rigoutsos block structured grid generation, and is fully parallelized via the Message Passing Interface, and time evolution is via standard 4-th order Runge-Kutta time-stepping.
We illustrated some areas of physics that can potentially benefit from this code, such as multiple black hole mergers and scalar field collapse. Such fields require a code which adapts to changes in the range and location of scales at different points in space and time in the simulation. We emphasise that setting the initial conditions for these mergers are trivial -- $\mathtt{GRChombo}$ automatically remeshes the grid given a set of analytic initial conditions without further user intervention.
We showed that $\mathtt{GRChombo}$ successfully passes the standard ``Apples with Apples'' tests\footnote{We note that we perform as well as any other BSSN code in the Gowdy wave test as we expected in a pre-determined gauge. Although, \cite{Cao:2011fu} managed to achieve long-term evolution by considering different gauge conditions.}. In addition to these tests, we evolved standard single black hole spacetimes (Schwarzschild and Kerr) and showed that it is stable to more than $T =10000M$. Using the moving puncture gauge, we also show that $\mathtt{GRChombo}$ stably evolves the merger of two and three black holes in inspiral and head-on collisions. We simulated the supercritical collapse of a scalar field configuration, and found that it forms a black hole as expected, to show that the code supports non vacuum spacetimes. Finally we tested the MPI scaling properties of the code, both strongly and weakly, and compared this with an alternative numerical relativity code based on the popular Cactus framework.
Nevertheless, despite its power, the AMR capability of $\mathtt{GRChombo}$ has to be treated with care. As we mentioned earlier, coarse-fine boundaries could be a significant source of inaccuracy, even though the Hamiltonian constraint may still be kept under control. A way to reduce coarse-fine boundary errors is to introduce conservative refluxing during interlevel operations. Although refluxing requires significant overhead, we intend to implement it in the next iteration of $\mathtt{GRChombo}$. The litmus test for accuracy of $\mathtt{GRChombo}$ is its ability to make accurate predictions of outgoing gravitational wave-forms. We leave this, and the introduction of a set of ``best practices'' for the use of AMR in general relativistic systems, for a follow-up work.
\section*{Acknowledgements}
We would first like to thank the $\texttt{Lean}$ collaboration for allowing us to use their code as a basis for comparison, and especially Helvi Witek for helping with the setting up and running of the $\texttt{Lean}$ simulation. We would like to thank Erik Schnetter, Ulrich Sperhake, Helvi Witek, Luis Lehner, Carlos Palenzuela and Tom Giblin for many useful conversations, and members of the $\mathtt{Chombo}$ collaboration, Daniel Martin and Brian Van Straalen. We would especially like to thank Juha J{\"a}ykk{\"a} and James Briggs for their amazing technical support. This work was undertaken on the COSMOS Shared Memory system at DAMTP, University of Cambridge operated on behalf of the STFC DiRAC HPC Facility. This equipment is funded by BIS National E-infrastructure capital grant ST/J005673/1 and STFC grants ST/H008586/1, ST/K00333X/1. EAL acknowledges support from an STFC AGP grant ST/L000717/1. PF and ST are supported by the European Research Council grant ERC-2011-StG279363HiDGR. PF is also supported by the Stephen Hawking Advanced Research Fellowship from the Centre for Theoretical Cosmology, University of Cambridge. MK is supported by an STFC studentship. He started his work on this project as a summer student funded by the Bridgwater Summer Undergraduate Research Programme at the Centre for Mathematical Sciences, University of Cambridge, and by King's College, Cambridge. HF is supported by the US Department of Energy (DOE), and this research used resources of the Argonne Leadership Computing Facility, which is a DOE Office of Science User Facility, both supported under Contract DE-AC02-06CH11357. Part of the performance test for this work was performed on Louisiana State University's High Performance Computing facility.
|
\section{Introduction}
The Standard Model (SM), although highly successful, leaves many unanswered questions in its wake such as: what (if anything) stabilises the recently discovered Higgs boson mass?
Are the three known gauge forces unified into a simple gauge group
which also explains charge quantisation?
What is the origin of the three families of quarks and leptons and their pattern of masses, mixing and CP violation?
Why is CP so accurately conserved by the strong interactions?
The answers to such questions may help resolve longstanding cosmological puzzles such as the nature of dark matter and the origin of matter-antimatter asymmetry, both of which are unexplained within the SM.
In this paper we propose a realistic and fairly complete model capable of addressing all the above questions unanswered by the SM. The basic ingredients of our model are Supersymmetry (SUSY) together with an $SU(5)$ Grand Unified Theory (GUT), flavoured by an $A_4$ family symmetry (for a review see e.g. \cite{King:2013eh}). The model is minimal in the sense that $SU(5)$ is the smallest GUT group and $A_4$ is the smallest family symmetry group that admits triplet representations. Also, below the GUT scale, the model yields the minimal supersymmetric standard model (MSSM) supplemented by a minimal two right-handed neutrino seesaw mechanism. The model is realistic in the sense that it provides a successful (and natural) description of the fermion mass and mixing spectrum, including spontaneous CP violation, while resolving the strong CP problem. It is fairly complete in the sense that GUT and flavour symmetry breaking are addressed, including doublet-triplet splitting, Higgs mixing and the origin of the MSSM $\mu$ term.
The model also allows a WIMP dark matter candidate due to the conserved MSSM R-parity, and permits matter-antimatter asymmetry via leptogenesis involving the two right-handed neutrinos. We shall show that the leptogenesis phase is equal to the single phase appearing in the neutrino mass matrix, providing a direct link between neutrino oscillations and matter-antimatter asymmetry, although we shall not discuss cosmological aspects any further in this paper.
We emphasise the predictive nature of the model in the lepton sector, where the entire PMNS matrix is predicted without any free parameters, up to a discrete choice of a single phase. Large lepton mixing is accounted for by the seesaw mechanism \cite{Minkowski:1977sc} with constrained sequential dominance (CSD) \cite{King:1998jw,King:2005bj,Antusch:2011ic}). With a diagonal two right-handed neutrino mass matrix $M_R$, the dominant right-handed neutrino $\nu_R^{\rm atm}$ mainly responsible for the atmospheric neutrino mass $m_3$ has couplings to $(\nu_e, \nu_{\mu}, \nu_{\tau})$ proportional to $(0,1,1)$, while the subdominant right-handed neutrino $\nu_R^{\rm sol}$ giving the solar neutrino mass $m_2$ has couplings to $(\nu_e, \nu_{\mu}, \nu_{\tau})$ proportional to $(1,3,1)$. These couplings, corresponding to the so called CSD3 scheme \cite{King:2013iva,Bjorkeroth:2014vha}, originate from $A_4$ vacuum alignment.%
\footnote{CSD4 models have been discussed in \cite{King:2013hoa}}
The model consequently predicts a normal neutrino mass hierarchy, $m_3>m_2\gg m_1=0$.
As mentioned above, the lepton sector is controlled by a relative phase which is selected to be $2\pi/3$, chosen from the nine complex roots of unity arising from spontaneous CP violation of a $\mathbb{Z}_9 \times \mathbb{Z}_{6}$ discrete symmetry, by a mechanism proposed in \cite{Antusch:2011sx}. Such a spontaneous CP violating scenario had been proposed previously in order to account for the smallness of CP violation in the soft SUSY sector \cite{Ross:2004qn}. We also employ a $\mathbb{Z}_4^R$ discrete R-symmetry (as the origin of MSSM R-parity, as in \cite{Lee:2011dya}) and a missing partner (MP) mechanism \cite{Masiero:1982fe} for doublet-triplet splitting as recently advocated for flavoured GUTs in \cite{Antusch:2014poa}. The model predicts very sparse lepton and down-type quark Yukawa matrices, with five texture zeroes, and Yukawa elements involving simple $SU(5)$ Clebsch-Gordan (CG) ratios of $4/9$ and $9/2$ for the first and second families, with $m_{\tau}/m_b=1$ for the third family, all in excellent agreement with their experimental values run up to the GUT scale \cite{Antusch:2009gu}.
Quark mixing originates predominantly from a non-diagonal and naturally hierarchical up-type Yukawa matrix, provided by the broken $\mathbb{Z}_9$ discrete family symmetry. Quark CP violation, however, comes exclusively from a single off-diagonal element in the down Yukawa matrix. By contrast, to excellent approximation, all lepton mixing and CP violation originates from the neutrino mass matrix, whose structure is also controlled by the $A_4$ family symmetry and the $\mathbb{Z}_{6} $ symmetry via the CSD3 type vacuum alignment as described above \cite{King:2013iva}.
Although there have been many attempts in the literature based on $A_4$ flavoured $SU(5)$ SUSY GUTs (for an incomplete list see e.g. \cite{deMedeirosVarzielas:2005qg}), we would argue that none are as successful or complete as the present one. For example, many of the previous models predicted mixing very close to tri-bimaximal and are by now excluded. Indeed the present model is the first one based on CSD3 capable of predicting all the lepton mixing parameters consistent with current data on lepton mixing \cite{King:2013iva,Bjorkeroth:2014vha} (see also \cite{King:2013hoa}). The full literature on flavoured SUSY GUTs, i.e. which involve
a (discrete) family symmetry, is quite extensive (for an incomplete list see e.g. \cite{huge}). The goal of all these models is clear: to address the questions left unanswered by the SM. It will take some time and (experimental) effort to resolve all these models. However the most promising models are those that make testable predictions while being theoretically complete and consistent.
While there are many different chiral superfields in this model, indeed almost exactly a hundred, it is important to note that we are explicitly presenting a renormalisable model. Any ``non-renormalisable terms'' generated below the Planck scale are required to have a specific well defined realization through multiple renormalisable terms involving heavy messenger fields that can be integrated out around the GUT scale. The respective effective theory after they are integrated out is actually more predictive than otherwise, with a normal neutrino mass hierarchy, a zero lightest neutrino mass, and all lepton mixing angles and CP phases predicted. The model presented here is amongst the most predictive and complete SUSY GUTs of flavour, consistent with current data.
The layout of the remainder of the paper is as follows: in Section \ref{sec:SM} we describe the superfields directly related to the SM fermions and neutrinos, as well as their Yukawa structures as imposed by the GUT and family symmetries when certain $A_4$ breaking vacuum expectation values (VEVs) are applied; we also perform a global fit to the parameters of the model and present our predictions for the lepton sector. In Section \ref{sec:A4} we describe the superfields that are responsible for breaking the family symmetries and how the required $A_4$ breaking VEVs arise. In Section \ref{sec:GUT} several aspects related to the GUT are discussed, particularly how to break $SU(5)$ and $\mathbb{Z}_4^R$ down to the MSSM with R-parity in a viable way (i.e. addressing doublet-triplet splitting, the origin of the $\mu$ term and proton decay). We also discuss the resolution to the strong CP problem. In Section \ref{sec:link} we discuss the link between leptogenesis and the oscillation phase in this model. Finally in Section \ref{sec:con} we summarise our main results and conclude. Appendix~\ref{A4} summarises the $A_4$ conventions used in this paper, in the basis of \cite{Ma:2001dn}.
\section{The Yukawa sector of the model \label{sec:SM}}
\input{table_1}
The model involves an $A_4\times SU(5)$ CP invariant superpotential at the GUT scale,
where all symmetries, including CP, are spontaneously broken along supersymmetric flat directions,
as discussed in Sections \ref{sec:A4} and \ref{sec:GUT}.
As already noted, it involves a further
$\mathbb{Z}_9 \times \mathbb{Z}_{6}$ discrete family symmetry as well as
a $\mathbb{Z}_4^R$ discrete R-symmetry.
The purpose of this section is to describe those aspects of the model pertaining to the Yukawa sector,
i.e. relevant for understanding quark and lepton masses, mixing and CP violation.
The flavour sector of the model is very important in our approach, since we
make a serious attempt to understand and, where possible, predict the experimentally observable
fermion masses and mixing matrices.
In Table \ref{ta:SMF} we show the matter superfields $F$, $T_i$ that contain the quarks and leptons,
as well as the right-handed neutrino superfields $N_i^c$ and double seesaw superfield
$\Gamma$, all of which carry unit $\mathbb{Z}_4^R$ charge. Apart from the $A_4\times SU(5)$ assignments of
$F\sim (3,\overbar{5})$, $T_i\sim (1,{10})$, $N^c_i \sim (1,{1})$,
under $\mathbb{Z}_9$ they transform as $F\sim0$, $T_i\sim (5,7,0)$, $N^c_i\sim (7,8)$.
Unlike the rest of the quarks and leptons, the right-handed neutrinos are further charged under $\mathbb{Z}_{6}$ (as are some of the symmetry breaking scalars).
In Table \ref{ta:SMF} we also display the six Higgs superfields, generally denoted $H$ (but also $\Lambda$) which serve to break the $SU(5)$ gauge symmetry.
The two light MSSM Higgs doublet superfields $H_u$ and $H_d$ will emerge from $H_5$ and a mixture of $H_{\bar{5}}$ and $H_{\overbar{45}}$ by a mechanism discussed later. The superfield $\xi$ which breaks $\mathbb{Z}_{9}$ is particularly central to this theory, as it is responsible for both right-handed neutrino masses and the up-type quark mass hierarchy. Finally we have the $\theta_i$ superfields which break $\mathbb{Z}_{6}$ and help to control Dirac neutrino masses, and nine $A_4$ breaking triplet flavons generally denoted $\phi$, with various vacuum alignments, responsible for large lepton mixing.
With these assignments, only the top quark gets a mass from a renormalisable Yukawa coupling $H_5T_3T_3$
(which has $\mathbb{Z}_4^R$ charge 2 as required for an allowed superpotential term).
All the other quark and lepton Yukawa couplings must arise through higher order terms.
This provides the basic reason why most of the SM (or strictly MSSM) Yukawa couplings appear to be so small.
The observed hierarchy of Yukawa couplings between the three families will be explained via
a discrete $\mathbb{Z}_{9}$ version of the Froggatt-Nielsen mechanism
\cite{Froggatt:1978nt}, with powers of the low VEV of $\xi$ controlling the hierarchy in the up-type quark sector, and also, in part,
the smallness of the down quark and electron.
In order to enhance predictivity we need the messengers listed in Table \ref{ta:Mess}, which is the price we pay for having a renormalisable theory at the GUT scale. We denote these superfields either as fermion messengers, $X_i$, or scalar messengers, $\Sigma_i$, depending on
whether they carry similar quantum numbers to, respectively, the quarks and leptons
(with odd $\mathbb{Z}_4^R$ charge) or the symmetry breaking scalars
(with even $\mathbb{Z}_4^R$ charge). The fermion messengers $X_i$ carry similar quantum numbers to down-type quarks and charged leptons (and neutrinos). Scalar messengers $\Sigma_i$ have quantum numbers similar to $H_5$ (the superfield that gives the top quark a renormalisable mass term). The $\Sigma_i$ messengers do not get VEVs, which means we need not consider the effect of diagrams with $\Sigma_i$ superfields in external legs to the masses of SM fermions.
The messengers group themselves in pairs of two superfields with a renormalisable bare mass coupling which respects all the symmetries. Their masses are therefore expected to be at or around the GUT scale.
Although there will be in general distinct masses for different pairs, for simplicity and because they are all expected to be at a similar mass scale, we take masses of all such pairs to be $M$ and set it equal to the GUT scale
in our numerical estimates. We emphasise that the successful predictions of the model
in the lepton sector (namely predicting the PMNS matrix) is independent of the specific values of these
mass parameters.
\subsection{Up quarks \label{sec:u}}
Apart from the top quark mass, which originates from a lowest order Yukawa coupling,
the remaining up-type quark Yukawa couplings appear from higher order terms that result from combining several renormalisable terms involving $\Sigma_i$ messengers and the GUT singlet superfield $\xi$.
To be precise, the up-type quark Yukawa couplings arise from $\Sigma_i$ messenger tower diagrams shown in Fig.~\ref{fig:upmass}.
For example, the most suppressed coupling arises from the first diagram in Fig.~\ref{fig:upmass}. Other less suppressed couplings arise from the diagrams where at the base one has the respective $T_i T_j$, with a shorter tower leading up to $H_5$. The least suppressed coupling, the renormalisable $H_5 T_3 T_3$ operator responsible for the top quark mass, is the last diagram in Fig.~\ref{fig:upmass}.
The effective superpotential responsible for the up-type Yukawa couplings is
\begin{equation}
W_{\mathrm{up}} = u_{ij} H_5 T_i T_j \left( \frac{\xi}{M} \right)^{n_{ij}}.
\end{equation}
The resulting symmetric Yukawa matrix for up-type quarks is
\begin{equation}
Y_{ij}^u = u_{ij} \left(\frac{\braket{\xi}}{M}\right)^{n_{ij}} \sim \pmatr{\tilde{\xi}^4 & \tilde{\xi}^3 & \tilde{\xi}^2 \\ &\tilde{\xi}^2 & \tilde{\xi} \\& & 1},
\label{upYuk}
\end{equation}
where $ \tilde{\xi} = \braket{\xi}/M \sim 0.1 $. The explicit form of $ Y^u $ is given in Eq.~\ref{eq:yufull} and includes the coefficients $ u_{ij} $, which are $ \mathcal{O}(1) $ and, by enforcing CP conservation at the GUT scale, necessarily real. Thus, the hierarchy of the up quark masses as well as the CKM mixing angles are given by powers of $\tilde{\xi}$. Due to the structure of this matrix, any phase introduced by $\braket{\xi}$ can be reabsorbed by appropriate redefinition of the three $T_i$ fields, so $Y^u$ does not contain a source of CP violation.
\begin{figure}[ht]
\begin{subfigure}[b]{0.17\textwidth}
\centering
\includegraphics[scale=0.4]{images/u11.pdf}
\end{subfigure}%
\begin{subfigure}[b]{0.17\textwidth}
\centering
\includegraphics[scale=0.4]{images/u12.pdf}
\end{subfigure}%
\begin{subfigure}[b]{0.17\textwidth}
\centering
\includegraphics[scale=0.4]{images/u13.pdf}
\end{subfigure}%
\begin{subfigure}[b]{0.17\textwidth}
\centering
\includegraphics[scale=0.4]{images/u22.pdf}
\end{subfigure}%
\begin{subfigure}[b]{0.17\textwidth}
\centering
\includegraphics[scale=0.4]{images/u23.pdf}
\end{subfigure}%
\begin{subfigure}[b]{0.17\textwidth}
\centering
\includegraphics[scale=0.4]{images/u33.pdf}
\end{subfigure}%
\caption{Diagrams responsible for the masses and mixings of the up-type quarks.}
\label{fig:upmass}
\end{figure}
\subsection{Down quarks, charged leptons and flavons \label{sec:dcl}}
When considering the Yukawa structures of down quarks and charged leptons we must inevitably discuss $A_4$ triplet flavons.%
\footnote{As a point of terminology, we refer to as ``flavons'' any superfields that are GUT singlets transforming non-trivially under the family symmetry and that get VEVs. In particular not only $A_4$ but strictly speaking also $\mathbb{Z}_9$ and $\mathbb{Z}_{6}$ are family symmetries, so we also refer to $\xi$ as a ``flavon''.}
The assignments of all the flavons under the family symmetries appear in Table \ref{ta:SMF}. Indeed, since the three families of $F$ transform as a triplet of $A_4$ (see Table \ref{ta:SMF}), all $T_i H_{\bar 5}F$ terms require a contraction with at least one $A_4$ triplet flavon to be invariant.
\begin{figure}[ht]
\centering
\begin{subfigure}{0.5\textwidth}
\centering
\includegraphics[scale=0.45]{images/t3.pdf}
\caption{}
\end{subfigure}%
\begin{subfigure}{0.5\textwidth}
\centering
\includegraphics[scale=0.45]{images/t2.pdf}
\caption{}
\end{subfigure}
\begin{subfigure}{0.5\textwidth}
\centering
\includegraphics[scale=0.45]{images/t1.pdf}
\caption{}
\end{subfigure}%
\begin{subfigure}{0.5\textwidth}
\centering
\includegraphics[scale=0.45]{images/t23.pdf}
\caption{}
\label{fig:ydoffdiag}
\end{subfigure}
\caption{Diagrams responsible for the masses of the down-type quarks and charged leptons.}
\label{fig:charmass}
\end{figure}
From the diagrams shown in Fig.~\ref{fig:charmass}, integrating out the fermion messengers $X$,
which acquire large masses as a result of either explicit mass terms or GUT scale Higgs VEVs,
we obtain effective operators of the form
\begin{equation}
W_{\mathrm{down}} = d_{33} T_3 \frac{H_{\bar 5}\phi_\tau}{M}F + d_{22} T_2\frac{H_{\overbar{45}}H_{24}\phi_\mu}{M^2}F + d_{11} T_1\frac{H_{\bar 5}\xi\phi_e}{\braket{\Lambda_{24}}^2}F + d_{12} T_1\frac{H_{\bar 5}\xi\phi_\mu}{\braket{\Lambda_{24}}\braket{H_{24}}}F ,
\label{eq:dcl}
\end{equation}
where $ d_{ij} $ are $ \mathcal{O}(1) $ couplings.
The light MSSM doublet $H_d $ is a combination of the doublets inside $ H_{\bar{5}} $ and $ H_{\overbar{45}} $, as discussed in Section~\ref{sec:splitting}, hence the $d_{22}$ term also leads to a relevant Yukawa coupling.
The alignment of the respective flavon VEVs of $\phi_{e,\mu,\tau}$ (discussed in Section \ref{sec:A4}) is
\begin{equation}
\braket{\phi_e} = v_{e} \pmatr{1\\0\\0} \qquad \braket{\phi_\mu} = v_{\mu} \pmatr{0\\1\\0} \qquad \braket{\phi_\tau} = v_{\tau} \pmatr{0\\0\\1},
\end{equation}
such that, apart from $d_{12}$, the contraction appearing with $T_{1,2,3}$ isolates the respective $F_{1,2,3}$ family. This would lead to diagonal Yukawa structures if not for the additional term connecting $T_1 (\phi_\mu F)$ (see Fig.~\ref{fig:ydoffdiag}).
The resulting effective Yukawa matrices are, schematically:
\begin{equation}
Y^d_{LR} \sim Y^e_{RL} \sim \pmatr{ \dfrac{\braket{\xi} v_{e}}{v_{\Lambda_{24}}^2} & \dfrac{\braket{\xi} v_{\mu}}{{v_{\Lambda_{24}}} {v_{H_{24}}}} & 0 \\[2ex] 0&\dfrac{{v_{H_{24}}} v_{\mu}}{M^2} & 0 \\[2ex] 0& 0& \dfrac{v_{\tau}}{M}},
\label{downYuk}
\end{equation}
where $ v_{\Lambda_{24}} $ and $ v_{H_{24}} $ are the respective VEVs of $ \Lambda_{24} $ and $ H_{24} $ (given in Eq.~\ref{eq:su5vevs}), and we include the subscripts $LR$ to emphasise the role of the off-diagonal term to left-handed mixing from $Y^d$. The off-diagonal term in $Y^e$ also provides a tiny contribution to left-handed charged lepton mixing $ \theta_{12}^e \sim m_e/m_\mu $ which may safely be neglected. It also introduces CP violation to the CKM matrix via the phase of $\braket{\xi}$.
Furthermore, because the underlying renormalisable theory is known,
the diagrams in Fig.~\ref{fig:charmass} are the only contributions for each family. The $SU(5)$ contractions and associated CG coefficients appearing for each family are unique \cite{Antusch:2009gu,Antusch:2014poa}.
With the GUT scale symmetry breaking as discussed in Section~\ref{sec:GUT},
each of the scalars here get a VEV with the group structure:
\begin{align}
\begin{split}
\braket{H_{\bar 5}}^a &=\delta^a_5~ v_d/\sqrt{2}\\
\braket{H_{\overbar{45}}}^{ab}_c &=(\delta^{[a}_c-\delta^{[a|}_5\delta^5_c-4\delta^{[a|}_4\delta^4_c)\delta^{b]}_5~v_d/\sqrt{2} \\
\braket{H_{24}}^a_b &=\mathrm{diag}(2,2,2,-3,-3)~v_{H_{24}} \\
\braket{\Lambda_{24}}^a_b &=\mathrm{diag}(2,2,2,-3,-3)~v_{\Lambda_{24}},
\end{split}
\label{eq:su5vevs}
\end{align}
where the indices run $a,b,c=1,...,5$. This leads to the GUT scale prediction:
\begin{equation}
\dfrac{Y^e_{33}}{Y^d_{33}} = 1, \qquad \dfrac{Y^e_{22}}{Y^d_{22}}=\dfrac{9}{2}, \qquad \dfrac{Y^e_{11}}{Y^d_{11}} = \dfrac{Y^e_{21}}{Y^d_{12}} = \dfrac{4}{9}.
\end{equation}
The explicit forms of $ Y^d$ and $ Y^e$, including CG and $ d_{ij} $ coefficients, are given
later in Eq.~\ref{eq:ydfull} and Eq.~\ref{eq:yefull}, respectively.
\subsection{Neutrinos and CSD3 \label{sec:neutrinos}}
In order to obtain the CSD3 vacuum alignment in this model we couple the neutrinos to a set of flavons distinguished by the $\mathbb{Z}_{6}$ symmetry. Of the superfields in Table~\ref{ta:SMF}, only the right-handed neutrinos and some of the flavons are charged under this symmetry. For clarity, we relabel two of the flavon fields
as $ \phi_{\mathrm{atm}} \equiv \phi_3 $ and $ \phi_{\mathrm{sol}} \equiv \phi_4$, to highlight their role in producing neutrino mixing. We also write
$N_\mathrm{atm}^c \equiv N_1^c$ to denote the right-handed neutrino that dominantly leads to the atmospheric neutrino mass, and $N_\mathrm{sol}^c \equiv N_2^c$ as that which contributes mainly to the solar neutrino mass.
The relevant terms in the superpotential giving neutrino masses are thus
\begin{equation}
W_\nu = y_1 H_{5} F\frac{\phi_\mathrm{atm}}{\braket{\theta_2}} N_\mathrm{atm}^c
+ y_2 H_{5} F\frac{\phi_\mathrm{sol}}{\braket{\theta_2}} N_\mathrm{sol}^c
+ y_3 \frac{\xi^2}{M} N_\mathrm{atm}^c N_\mathrm{atm}^c
+ y_4 \xi N_\mathrm{sol}^c N_\mathrm{sol}^c.
\label{eq:neutrinomassW}
\end{equation}
The flavons $\phi_\mathrm{atm}$ and $ \phi_\mathrm{atm} $ gain VEVs according to the CSD3 alignment:
\begin{equation}
\braket{\phi_{\mathrm{atm}}} = v_{\mathrm{atm}} \pmatr{0\\1\\1} \qquad\qquad \braket{\phi_{\mathrm{sol}}} = v_{\mathrm{sol}} \pmatr{1\\3\\1},
\end{equation}
where $v_{\mathrm{atm}}$ and $v_{\mathrm{sol}}$ are generally complex. Denoting the phases of VEVs as $\rho_i = \arg (v_i)$, only the relative phase $ \rho_\mathrm{atm}\!-\!\rho_\mathrm{sol} $ between the VEVs is physically relevant, and is constrained to a discrete set of values, as discussed in Section \ref{sec:phasefixing}. The flavon $\xi$ (already responsible for the up quark masses) is also acting as a Majoron by generating hierarchical right-handed neutrino masses. At the effective level, the Dirac terms result from coupling the neutrinos (and $H_5$) to $\phi_\mathrm{atm}$ and $\phi_\mathrm{sol}$ via the flavon $\theta_2$ (an $A_4$ singlet carrying $\mathbb{Z}_{6}$ charge). The corresponding diagrams with associated messengers appear in Fig~\ref{fig:Dirac}.
\begin{figure}[ht]
\centering
\begin{subfigure}{0.5\textwidth}
\centering
\includegraphics[scale=0.6]{images/nu1.pdf}
\caption{}
\end{subfigure}%
\begin{subfigure}{0.5\textwidth}
\centering
\includegraphics[scale=0.6]{images/nu2.pdf}
\caption{}
\end{subfigure}
\begin{subfigure}{0.5\textwidth}
\centering
\includegraphics[scale=0.6]{images/n1.pdf}
\caption{}
\end{subfigure}%
\begin{subfigure}{0.5\textwidth}
\centering
\includegraphics[scale=0.6]{images/n2.pdf}
\caption{}
\end{subfigure}
\caption{Renormalisable diagrams leading to neutrino effective terms. Diagrams (a) and (b) are responsible for neutrino Yukawa terms (leading to Dirac masses) while (c) and (d) give right-handed neutrino Majorana mass terms.}
\label{fig:Dirac}
\end{figure}
In turn, the Majorana mass term for $N_\mathrm{atm}^c $ is also non-renormalisable and we refer to the superfield $\Gamma$ as the respective messenger. It couples only to $N_\mathrm{atm}^c $ and simply provides the non-renormalisable mass term for $N_\mathrm{atm}^c $, suppressed relative to the mass of $N_\mathrm{sol}^c $. As $\Gamma$ has the quantum numbers of a third right-handed neutrino, one can also consider this field as mediating a double seesaw mechanism, responsible for the $N_\mathrm{atm}^c $ mass. The mixing term $\frac{\xi^6}{M^5} N_{\mathrm{atm}}^c N_{\mathrm{sol}}^c$, though allowed by the symmetries, is absent as there is no combination of messengers able to produce it.
We write $\braket{\xi} = |v_{\xi} | e^{i\rho_{\xi}}$, where $ \rho_{\xi} $ is chosen from a discrete set of available phases, discussed in Section~\ref{sec:GUTbreaking} (see Eq.~\ref{eq:GUTVEVs}).
This phase originates from the spontaneous CP violation of a discrete Abelian symmetry \cite{Antusch:2011sx,Ross:2004qn}, in our case the $\mathbb{Z}_9$.
We will now show that $ \rho_{\xi} $ and $ \rho_\mathrm{atm}\!-\!\rho_\mathrm{sol} $ fix the relative phases within the effective neutrino mass matrix and consequently the leptonic mixing angles.
In a Supersymmetric (SUSY) model, the relevant terms in the superpotential giving neutrino masses, in the diagonal charged lepton basis, are
\begin{equation}
W_\nu = y_{\mathrm{atm}}^i H L_i N_\mathrm{atm}^c
+ y_{sol}^i H L_i N_\mathrm{sol}^c
+ M_{\rm atm} N_\mathrm{atm}^c N_\mathrm{atm}^c
+ M_{\rm sol} N_\mathrm{sol}^c N_\mathrm{sol}^c,
\label{eq:neutrinomassW2}
\end{equation}
where $L_i$ are three families of lepton doublets and the (CP conjugated) right-handed neutrinos $N_{\rm atm}^c$ and $N_{\rm sol}^c$ with real positive masses $M_{\rm atm}$ and $M_{\rm sol}$ do not mix.
The structure of $\lambda^\nu$ is determined by the vacuum alignments of $\phi_\mathrm{atm}$ and $\phi_\mathrm{sol}$. The Dirac and Majorana matrices, derived by comparing the superpotential terms in Eqs.~\ref{eq:neutrinomassW} and \ref{eq:neutrinomassW2} are
\begin{equation}
\lambda^\nu = \pmatr{0&b\\a&3b\\a&b} \qquad\qquad M^c = \pmatr{\dfrac{y_3\braket{\xi}^2}{M}&0\\[2ex] 0&y_4\braket{\xi}},
\label{seesaw}
\end{equation}
where $a = y_1 v_{\textrm{atm}}/\braket{\theta_2}$ and $b = y_2 v_{\textrm{sol}}/\braket{\theta_2}$.
For the see-saw mechanism we shall introduce a different convention for Yukawa and Majorana masses.
The Yukawa matrices $Y^{e}$, $Y^{\nu}$ are defined in a LR convention by
\begin{equation}
\mathcal{L}^{LR} = -H^dY^e_{ij}\overline{L}^i_L e^j_R - H^u Y^{\nu}_{ij} \overline{L}^i_L \nu^{i}_R + \mathrm{h.c.},
\end{equation}
where $i,j=1,2,3$ label the three families of lepton doublets $L_i$, right-handed charged leptons $e^j_R$ and right-handed neutrinos $\nu_R^j$ below the GUT scale; $H^u, H^d$ are the Higgs doublets which develop VEVs $v_u,\,v_d$. The physical effective neutrino Majorana mass matrix $m^{\nu}$ is determined by the seesaw mechanism,
\begin{equation}
m^{\nu} = v_u^2 Y^{\nu} M^{-1}_{R} Y^{\nu \mathrm{T}},
\label{eq:seesaw}
\end{equation}
where the light Majorana neutrino mass matrix $m^\nu$ is defined%
\footnote{The conventions for $ Y^{\nu,e} $ and $m^{\nu}$ differ, respectively, by overall Hermitian and overall complex conjugation compared to those used in the Mixing Parameter Tools package \cite{Antusch:2005gp}, which was used when performing global fits.}
by \( \mathcal{L}^{LL}_\nu = -\tfrac{1}{2} m^\nu \overline{\nu}_{L} \nu^{c}_L + \mathrm{h.c.} \), while the heavy right-handed Majorana neutrino mass matrix $M_R$ is defined by \(\mathcal{L}^{RR}_\nu = -\tfrac{1}{2} M_{R} \overline{\nu^c}_{R} \nu_{R} + \mathrm{h.c.} \).
There is a simple dictionary between the seesaw basis and the SUSY basis, as follows:
compared to the SUSY basis in Eq.~\ref{eq:neutrinomassW} used in leptogenesis calculations we see that $Y^{\nu}=(\lambda^{\nu})^*$, while $M_R=(M^c)^*$.
Hence the neutrino matrices become, in the seesaw basis,
\begin{equation}
Y^{\nu} = (\lambda^{\nu})^*=\pmatr{0&b^*\\a^*&nb^*\\a^*&(n-2)b^*} \qquad
M_R =(M^c)^*= \pmatr{\left(\dfrac{y_3\braket{\xi}^2}{M}\right)^*&0\\[2ex] 0&(y_4\braket{\xi})^*}.
\label{seesawy}
\end{equation}
Seesaw produces the effective neutrino mass matrix
\begin{equation}
m^\nu = m_a \pmatr{0&0&0\\0&1&1\\0&1&1}
+ m_b e^{i\eta} \pmatr{1&3&1\\3&9&3\\1&3&1} ,
\label{eq:mnu}
\end{equation}
where $m_a = v_u^2 |a|^2 / (y_3|v_{\xi}|^2 /M)$ and $m_b = v_u^2 |b|^2 / (y_4|v_{\xi}|)$ and we have multiplied throughout by an overall phase which we subsequently drop, keeping only the (physical) relative phase
\begin{equation}
\eta\equiv -\rho_{\xi}+2(\rho_\mathrm{atm}\!-\!\rho_\mathrm{sol}) ,
\label{eta}
\end{equation}
where we recall the above definitions of phases,
\begin{equation}
\rho_{\xi} \equiv \mathrm{arg}(\braket{\xi}) , \qquad \rho_\mathrm{atm}\!-\!\rho_\mathrm{sol} \equiv \mathrm{arg}(v_{\mathrm{atm}} v_{\mathrm{sol}}^\ast) ,
\label{varphi}
\end{equation}
and that CP conservation at high energies ensures that $y_i$ and $M$ are real. By arguments given in Section \ref{sec:phasefixing} and Section \ref{sec:GUTbreaking}, we can restrict the physical phase $ \eta $ to a discrete choice, namely one of the nine complex roots of unity. The values $\eta= \pm 2\pi/3$ are preferred by CSD3 \cite{King:2013iva,Bjorkeroth:2014vha}. Note that the model predicts a normal neutrino mass hierarchy, namely $m_3>m_2\gg m_1=0$, which will be tested in the near future.
The sign of $ \eta $ has phenomenological significance, as it fixes the leptonic Dirac phase $ \delta^l $. Specifically, a positive $ \eta $ uniquely leads to negative $ \delta^l $, and \emph{vice versa} \cite{Bjorkeroth:2014vha}. As experimental data hints at $ \delta^l \sim -\pi/2 $, the \emph{a posteriori} preferred solution has positive $ \eta =+2\pi/3$. The sign of $ \eta $ also has cosmological significance, as discussed in Section \ref{sec:link}. For example a positive $ \eta =+2\pi/3$, together with the requirement that baryon asymmetry is positive, implies that the lightest right-handed neutrino should be $N_1^c=N_\mathrm{atm}^c $, while $N_2^c=N_\mathrm{sol}^c $ should be somewhat heavier, which is the natural ordering in our model.
\subsection{Full parameter fit \label{sec:fit}}
The structure of the Yukawa matrices and neutrino mass matrix is set by the theory, up to $\mathcal{O}(1)$ coefficients. The VEVs of the fields $ \xi $, $\Lambda_{24}$ and $H_{24}$ are at or near the GUT scale, but otherwise undetermined. This freedom coincides with the choice of coefficients in the Yukawa matrices, providing no extra degrees of freedom in the determination of the Yukawas other than to provide the appropriate scale. The same is true for the flavon fields $\phi_e$, $\phi_\mu$ and $\phi_\tau$, which provide the necessary hierarchy in the down-quark and charged lepton Yukawa sector.
The neutrino matrix $ m^\nu $ is given in Eq.~\ref{eq:mnu}. Letting $ v_f $ represent the VEV of a field $ f $, the Yukawa matrices are as follows:
\begingroup
\begin{align}
Y^u &= \pmatr{ u_{11} |\tilde{\xi}^4| & u_{12} |\tilde{\xi}^3| & u_{13} |\tilde{\xi}^2| \\[2ex]
u_{12} |\tilde{\xi}^3| & u_{22} |\tilde{\xi}^2| & u_{23} |\tilde{\xi}| \\[2ex]
u_{13} |\tilde{\xi}^2| & u_{23} |\tilde{\xi}| & u_{33}
} \label{eq:yufull} \\ \nonumber \\
Y^d &=\frac{1}{\sqrt{2}} \pmatr{\dfrac{1}{4} d_{11} \dfrac{|v_{\xi} v_e|}{|v_{\Lambda_{24}}|^2}
& d_{12} \dfrac{ | v_{\xi} v_\mu |}{|v_{\Lambda_{24}} v_{H_{24}}|} e^{i \zeta} & 0 \\[2.5ex]
0 & 2 d_{22} \dfrac{| v_{H_{24}} v_\mu | }{M^2} & 0 \\[2.5ex]
0 & 0 & d_{33} \dfrac{|v_\tau |}{M}
} \label{eq:ydfull}\\ \nonumber \\
Y^e &=\frac{1}{\sqrt{2}}\pmatr{\dfrac{1}{9} d_{11} \dfrac{|v_{\xi} v_e|}{|v_{\Lambda_{24}}|^2} & 0 & 0 \\[2.5ex]
d_{12} \dfrac{|v_{\xi} v_\mu|}{|v_{\Lambda_{24}} v_{H_{24}}|} e^{i \zeta} & 9 d_{22} \dfrac{|v_{H_{24}} v_\mu|}{M^2} & 0 \\[2.5ex]
0 & 0 & d_{33} \dfrac{|v_\tau|}{M} .
} \label{eq:yefull}
\end{align}
\endgroup
As already remarked, the phases in $Y^u$ from powers of $\braket{\xi} = |v_{\xi} |e^{i\rho_{\xi}}$
can be removed by field redefinition.
Without loss of generality we have rephased fields such that the only phase appearing in
$ Y^d $ and $ Y^e$ is the phase $ \zeta $ as shown in Eqs.~\ref{eq:ydfull}, \ref{eq:yefull}, so all quark CP violation
originates from the single phase $ \zeta $ appearing in $Y^d_{12}$.
In turn, $ \zeta $ is determined by a combination of phases coming from various field VEVs; more precisely, $\zeta=\rho_\xi-2\rho_{H_{24}}-\rho_{\Lambda_{24}}$.
As long as it is reasonably far from zero, it can produce the necessary CP violation. Different choices of
$ \zeta $ do not affect the goodness-of-fit, corresponding simply to different but equally valid choices of $ \mathcal{O}(1) $ coefficients. For our fit we choose $ \zeta = \pi/3$.
Note that the corresponding phase in $Y^e_{21}$ does not contribute to leptonic CP violation, since this term does not affect left-handed mixing, to an accuracy of $\mathcal{O}(m_e/m_{\mu})$.
To fit the real coefficients $u_{ij}$, $d_{ij}$, $m_a$ and $m_b$, we propose a function $\chi^2$ that relates the $N$ physical predictions $P_i(\{x\})$ for a given set of input parameters $\{x\}$ to their current best-fit values $\mu_i$ and their associated $1\sigma$ errors, denoted $ \sigma_i $, by
\begin{equation}
\chi^2 = \sum_{i=1}^N \left(\frac{P_i(\{x\})-\mu_i}{\sigma_i} \right)^2.
\label{eq:chisq}
\end{equation}
The errors $ \sigma_i $ are equivalent to the standard deviation of the experimental fits to a Gaussian distribution. For most parameters, this is essentially the case, with the exception of the (lepton) atmospheric angle $ \theta^l_{23} $. For a normal hierarchy, the distribution is roughly centered on $ \theta^l_{23} = 45^\circ $, while the best fit value is given by $ \theta^l_{23} = 42.3^\circ $. So as to not overstate the error for $ \theta^l_{23} $, we approximate its distribution by a Gaussian about $ 42.3^\circ $, setting $ \sigma_{\theta^l_{23}} = 1.6^\circ $.
We now wish to minimise $ \chi^2 $. In this analysis, $N = 18$, corresponding to six mixing angles $\theta^l_{ij}$ (neutrinos) and $\theta^q_{ij}$ (quarks), the CKM phase $\delta^q$, nine Yukawa eigenvalues for the quarks and charged leptons, and two neutrino mass-squared differences $\Delta m^2_{21}$ and $\Delta m^2_{31}$.
In the lepton sector, we use the PDG parametrisation of the PMNS matrix \cite{Beringer:1900zz}
\(U_{\mathrm{PMNS}} = R^l_{23} U^l_{13} R^l_{12} P_\textrm{PDG} \) in terms of \(s_{ij}=\sin \theta^l_{ij}\), \(c_{ij}=\cos\theta^l_{ij}\), the Dirac CP violating phase \(\delta^l\) and
further Majorana phases contained in
\(P_\textrm{PDG} = \textrm{diag}(1,e^{i\frac{\alpha_{21}}{2}},e^{i\frac{\alpha_{31}}{2}})\).
Experimentally, the leptonic phase $ \delta^l $ is poorly constrained at $ 1\sigma $ (and completely unconstrained at $ 3\sigma $), so is not fit, and left as a pure prediction of the model, as are the (completely unconstrained) Majorana phases $ \alpha_{21} $ and $ \alpha_{31} $. As this model predicts only two massive left-handed neutrinos, i.e. $ m_1 = 0 $, one Majorana phase is zero, which we take to be
$ \alpha_{31} =0$.
\begingroup
\newcommand{\exto}[1]{{\smaller$ \times 10^{#1} $}}
\newcommand{$^\circ$}{$^\circ$}
\begin{table}[!ht]
\centering
\footnotesize
\begin{tabular}{| c | c c | c c |}
\hline
\multirow{2}{*}{\makecell{Parameter \\ {\scriptsize (from \cite{Antusch:2013jca})}}} & \multicolumn{2}{c|}{\rule{0pt}{3ex}$ \tan \beta = 5 $} & \multicolumn{2}{c|}{$ \tan \beta = 10 $} \\[0.5ex]
\cline{2-5}
& $ \mu_i $ & $ \sigma_i $ & $ \mu_i $ & $ \sigma_i $ \\
\hline \hline
\rule{0pt}{3ex}%
$\theta^q_{12}$ & 13.027$^\circ$ & 0.0814$^\circ$ & 13.027$^\circ$ & 0.0814$^\circ$ \\
$\theta^q_{13}$ & 0.1802$^\circ$ & 0.0281$^\circ$ & 0.1802$^\circ$ & 0.0281$^\circ$ \\
$\theta^q_{23}$ & 2.054$^\circ$ & 0.384$^\circ$ & 2.054$^\circ$ & 0.384$^\circ$ \\
$\delta^q$ & 69.21$^\circ$ & 6.19$^\circ$ & 69.21$^\circ$ & 6.19$^\circ$ \\
\rule{0pt}{3ex}%
$y_u$ & 2.92 \exto{-6} & 1.81 \exto{-6} & 2.88 \exto{-6} & 1.79 \exto{-6} \\
$y_c$ & 1.43 \exto{-3} & 1.00 \exto{-4} & 1.41 \exto{-3} & 9.87 \exto{-5} \\
$y_t$ & 5.34 \exto{-1} & 3.41 \exto{-2} & 5.20 \exto{-1} & 3.15 \exto{-2} \\
\rule{0pt}{3ex}%
$y_d$ & 4.81 \exto{-6} & 1.06 \exto{-6} & 4.84 \exto{-6} & 1.07 \exto{-6} \\
$y_s$ & 9.52 \exto{-5} & 1.03 \exto{-5} & 9.59 \exto{-5} & 1.04 \exto{-5} \\
$y_b$ & 6.95 \exto{-3} & 1.75 \exto{-4} & 7.01 \exto{-3} & 1.78 \exto{-4} \\
\rule{0pt}{3ex}%
$y_e$ & 1.97 \exto{-6} & 2.36 \exto{-8} & 1.98 \exto{-6} & 2.38 \exto{-8} \\
$y_\mu$ & 4.16 \exto{-4} & 4.97 \exto{-6} & 4.19 \exto{-4} & 5.02 \exto{-6} \\
$y_\tau$ & 7.07 \exto{-3} & 7.27 \exto{-5} & 7.15 \exto{-3} & 7.42 \exto{-5} \\
\hline
\end{tabular}
\caption{Best fit values for quark and charged lepton parameters when run to the GUT scale as calculated in \cite{Antusch:2013jca}, with the SUSY breaking scale set at 1 TeV. We have included an overall contribution from threshold corrections corresponding to $ \bar{\eta}_b = -0.24375 $ which affects primarily the $ b $ quark Yukawa coupling $ y_b $. $ \mu_i $ represents the best-fit value and $ \sigma_i $ the error, as defined in Eq.~\ref{eq:chisq}.}
\label{tab:data}
\end{table}
\endgroup
The running of best-fit and error values to the GUT scale are generally dependent on SUSY parameters, notably $\tan \beta$, as well as contributions from SUSY threshold corrections. We extract the GUT scale CKM parameters and all Yukawa couplings (with associated errors) from \cite{Antusch:2013jca} for judicious choices of $ \tan \beta $. In further reference to \cite{Antusch:2013jca}, we choose for the parameter $ \bar{\eta}_b $ parametrising the threshold corrections a value $ \bar{\eta}_b = -0.24375 $; a non-zero value is required primarily to produce a necessary (small) difference in $ b $ and $ \tau $ Yukawa couplings. Experimental neutrino parameters are extracted from \cite{Gonzalez-Garcia:2014bfa}. All data is reproduced in Tables \ref{tab:data} and \ref{tab:lepdata}.
\begingroup
\newcommand{\exto}[1]{{\smaller$ \times 10^{#1} $}}
\newcommand{$^\circ$}{$^\circ$}
\begin{table}[ht]
\centering
\footnotesize
\renewcommand{\arraystretch}{1.5}
\begin{tabular}{| c | r r | }
\hline
\rule{0pt}{4ex}\makecell{Parameter \\ {\scriptsize (from \cite{Gonzalez-Garcia:2014bfa})}}
& \(\mu_i \pm1\sigma\) & 3\(\sigma\) range \\[1.5ex] \hline\hline
\( \theta^l_{12} \,(^\circ)\) & 33.48 \(^{+0.78}_{-0.75}\) & 31.29 \(\rightarrow\) 35.91 \\
\( \theta^l_{23} \,(^\circ)\) & 42.3 \(^{+3.0}_{-1.6}\) & 38.2 \(\rightarrow\) 53.3 \\
\( \theta^l_{13} \,(^\circ)\) & 8.5 \(^{+0.20}_{-0.21}\) & 7.85 \(\rightarrow\) 9.10 \\
\( \delta^l \,(^\circ)\) & 306 \(^{+39}_{-70}\) & 0 \(\rightarrow\) 360 \\ [5pt]
\(\dfrac{\Delta m^2_{21}}{10^{-5}}\) eV\(^2\) & 7.50 \(^{+0.19}_{-0.17}\) & 7.02 \(\rightarrow\) 8.09 \\ [8pt]
\(\dfrac{\Delta m^2_{31}}{10^{-3}}\) eV\(^2\) & +2.457 \(^{+0.047}_{-0.047}\) & +2.317 \(\rightarrow\) +2.607\\ [5pt]
\hline
\end{tabular}
\caption{Table of current best fits to experimental data for lepton mixing angles and neutrino masses case of normal mass squared ordering taken from \cite{Gonzalez-Garcia:2014bfa}, with 1\(\sigma\) and 3\(\sigma\) uncertainty ranges.}
\label{tab:lepdata}
\end{table}
\endgroup
Minimisation by differential evolution was performed in Mathematica, producing the set of $ \mathcal{O}(1) $ input coefficients and the corresponding physical parameters seen in Table \ref{tab:fit}, with an associated $ \chi^2 = 7.98 $ (for $ \tan \beta = 5 $) and $ \chi^2 = 7.84 $ (for $ \tan \beta = 10 $).
In this fit, the VEVs of $ \xi $, $ \Lambda_{24} $, $ H_{24} $ and the three $ \phi_{e,\mu,\tau} $ are fixed by hand in terms of the scale $ M $, which is taken to be the GUT scale, i.e. $ M \approx 3 \times 10^{16} $ GeV. Similarly, the Higgs doublet VEV enters only implicitly through $ m_a $ and $ m_b $, but is understood to take (at the GUT scale) the value $ v_H = 174 $ GeV. We set
\begin{equation}
\begin{split} v_{\xi} &= 6 \times 10^{-2} M \\ v_{\Lambda_{24}} &= M \\ v_{H_{24}} &= 3 \times 10^{-1}M \end{split} \qquad \qquad
\begin{split} v_e &= 10^{-3} M \\ v_\mu &= 10^{-3} M \\ v_\tau &= 5 \times 10^{-2} M \end{split}.
\label{eq:vevfit}
\end{equation}
The value of $ v_{\xi} $ is chosen to accommodate not only the fit to $ Y^u $ parameters but also to control the $ \mu $-term, as discussed in Section \ref{sec:splitting}. Meanwhile the factor $ \sim$3 split between $v_{\Lambda_{24}}$ and $ v_{H_{24}} $ assists in establishing a hierarchy between the $ e $ and $ \mu $ families.
With the above numerical values for the VEVs, we get the following Yukawa matrices in terms only of $ \mathcal{O}(1) $ coefficients and the complex phase $ \zeta $:
\begingroup
\newcommand{\pmatrs}[1]{\begin{pmatrix*}[r] #1 \end{pmatrix*}}
\newcommand{\extom}[1]{{{\scriptstyle\times 10^{#1}}}}
\begin{align}
Y^u &= \phantom{\frac{1}{\sqrt{2}}} \pmatrs{1.296 \extom{-5} \cdot u_{11} & 2.16 \extom{-4} \cdot u_{12} & 3.6\extom{-3} \cdot u_{13} \\ 2.16 \extom{-4} \cdot u_{21} & 3.6\extom{-3} \cdot u_{22} & 6 \extom{-2} \cdot u_{23} \\ 3.6\extom{-3} \cdot u_{31} & 6\extom{-2} \cdot u_{32} & u_{33}} \\[2ex]
Y^d &= \frac{1}{\sqrt{2}} \pmatr{ 1.5\extom{-5} \cdot d_{11} \phantom{e^{i \zeta}} & 2 \extom{-4} \cdot d_{12} e^{i \zeta} & 0 \\ 0 & 6 \extom{-4} \cdot d_{22} \phantom{e^{i \zeta}} & 0 \\ 0 & 0 & 5 \extom{-2} \cdot d_{33}} \\[2ex]
Y^e &= \frac{1}{\sqrt{2}} \pmatr{6.67 \extom{-6} \cdot d_{11} \phantom{e^{i \zeta}} & 0 & 0 \\ \phantom{6.6}2 \extom{-4} \cdot d_{12} e^{i \zeta} & 2.7 \extom{-3} \cdot d_{22} & 0 \\ 0 & 0 & 5 \extom{-2} \cdot d_{33}}.
\end{align}
\endgroup
\input{table_fit_2}
It is worth reiterating that the neutrino mass matrix phase $ \eta $ can be forced to admit only phases coming from the nine complex roots of unity, essentially due to spontaneous CP violation with the $\mathbb{Z}_9$ symmetry, where we select $ \eta = \pm 2\pi/3 $, both of which yield equally good $ \chi^2 $ fits, with only minor adjustments to $ \mathcal{O}(1) $ coefficients. The primary effect is in the prediction of $ \delta^l $; as previously stated, positive $ \eta $ corresponds to negative $ \delta^l $. As this is preferred by experiment, the results presented in Table \ref{tab:fit} are for $ \eta =+2\pi/3$. The fit also predicts the Majorana phases $ \alpha_{21} = 72^\circ $ and $ \alpha_{31} = 0 $.
In order to understand the significance of the $ \chi^2 $ fit, and assess the strength of the model overall, it is prudent to enumerate the parameters and predictions of the model. The nominal parameter count at the GUT scale is very large, owing to the diverse field content. However, at the scale where we are able to make predictions, many of these parameters combine to give a constrained set of free parameters that need to be determined. Notably, the VEVs of Higgs and flavon fields such as those given in Eq.~\ref{eq:vevfit} do not constitute true degrees of freedom, as they can be absorbed by redefining other parameters.
Relevant parameters that require consideration include: six $ u_{ij} $, four $ d_{ij} $, masses $ m_a $ and $ m_b $, phases $ \eta $ and $ \zeta $, the threshold factor $ \bar{\eta}_b $, and $ \tan \beta $, for a total of $N_I=16$ input parameters.
However three of these parameters, namely
$\tan \beta $, $\eta$ and $ \zeta $, are fixed prior to the fit, with the latter two phases restricted to discrete choices, as discussed previously.
Finally, the factor $ \bar{\eta}_b $ affects only the coupling $ y_b $ and is fitted by hand. As mentioned earlier, the model fits $ N = 18 $ observables, including nine Yukawa eigenvalues, two neutrino mass squared differences, six mixing angles and the quark CP phase. In addition the model predicts
the leptonic CP phase $ \delta^l $, two Majorana phases (one of which is zero) and a massless physical neutrino.
\section{\secheadmath{A_4} symmetry breaking and the flavon vacuum\label{sec:A4}}
In order to address $A_4$ symmetry breaking we need to address three aspects of the flavon vacuum: what drives some flavons to have VEVs a few orders of magnitude below the GUT scale, what determines their vacuum alignment, and what fixes the relative vacuum phase $\rho_\mathrm{atm}\!-\!\rho_\mathrm{sol} \equiv \mathrm{arg}(v_{\mathrm{atm}} v_{\mathrm{sol}}^{\ast})$ (and consequently the physical phase $ \eta $). In this section we consider each of these issues in turn.
\subsection{Driving the flavon vacuum expectation values \label{sec:radiative}}
The flavon $\phi$ VEVs are driven by radiative breaking \cite{Ibanez:1982fr} (see e.g. \cite{Ibanez:2007pf} for a recent review). The soft squared mass terms appear as $m_i^2 \phi_i \phi_i^\dagger \mathrm{ln}(\phi_i \phi_i^\dagger/\Lambda_i^2)$ and become negative at the required scales $\Lambda_i < M$, lifting the respective flat direction to a few orders of magnitude below the GUT scale for each flavon: $\langle \phi_i \phi_i^\dagger \rangle \sim \Lambda_i^2$. Hierarchies of VEVs are thus naturally expected due to the logarithmic nature of this mechanism. The precise symmetry breaking scale $\Lambda_i$ for each field $ \phi_i$ depends on otherwise undetermined parameters in the model which are different for each flavon, such as superpotential terms involving messengers.%
\footnote{Examples of such a Yukawa coupling for the neutrino flavons are $\phi_{\mathrm{atm}} F X_{14}$ and $\phi_{\mathrm{sol}} F X_{12}$.}
Therefore a hierarchy for such flavon VEVs is generated and remains stable due to radiative breaking \cite{Greene:1986jb}. On the other hand those for $v_e, v_{\mu}, v_{\tau}$ in Eq.~\ref{eq:vevfit} arise from $F$-terms as discussed later.
\subsection{Flavon vacuum alignment}
\input{table_2}
Thus far we have assumed that the $A_4$ triplet VEVs are aligned in special directions. In this section we describe how these directions are obtained by the superpotential terms allowed by the symmetries.
In doing this, the role of $\mathbb{Z}_{6}$ becomes clearer. The driving sector, a set of superfields $ A_i $ and $ O_{ij} $ with $\mathbb{Z}_4^R$ charge 2, is listed fully in Table \ref{ta:driving}.%
\footnote{Note that the $O$ (and $P$) fields that carry no $\mathbb{Z}_{6}$ charge couple to $H_{5} H_{\bar{5}} \xi^n$ (with some power of $\xi$), e.g. $P_{22} H_{5} H_{\bar{5}}$. We do not discuss these further as the respective $F$-terms do not affect the alignment nor the origin of the $\mu$-term.}
The inclusion of the $ \mathbb{Z}_{6} $ symmetry is necessary because the driving superpotential, responsible for aligning the flavons, needs to have each driving field isolated,
as shown below (see also \cite{King:2013iva}):
\begin{align}
\begin{split}
W_{\mathrm{align}} &\sim A_\mu\phi_\mu\phi_\mu+A_\tau\phi_\tau\phi_\tau + A_2(\phi_2\phi_2+\phi_2\theta_1) \\
&\qquad+ O_{e\mu}\phi_e\phi_\mu+O_{e\tau}\phi_e\phi_\tau+O_{\mu\tau}\phi_\mu\phi_\tau \\
&\qquad+ O_{e3}\phi_e\phi_3+O_{23}\phi_2\phi_3+O_{12}\phi_1\phi_2+O_{13}\phi_1\phi_3 \\
&\qquad+ O_{\mu 5}\phi_\mu\phi_5+O_{25}\phi_2\phi_5+O_{\mu 6}\phi_\mu\phi_6+O_{56}\phi_5\phi_6 \\
&\qquad+ O_{64}\phi_6\phi_4+O_{14}\phi_1\phi_4.
\label{flavon}
\end{split}
\end{align}
The additional $\mathbb{Z}_{6}$ charges ensure each term is separated from all others, leading to an array of vanishing $F$-term conditions that force mutual orthogonality conditions between many of the vacuum alignments. Since this was fully
discussed in \cite{King:2013iva}, we need only state the results here, namely that Eq.~\ref{flavon} leads to the following vacuum alignment patterns:
\begin{align}
\begin{array}{rlrlrl}
\braket{\phi_e} & \sim \pmatr{1\\0\\0} \hspace{5ex} & \braket{\phi_\mu} & \sim \pmatr{0\\1\\0} \hspace{5ex} & \braket{\phi_\tau} & \sim \pmatr{0\\0\\1} \\ [5ex]
\braket{\phi_1} & \sim \pmatr{2\\-1\\1} & \braket{\phi_2 } & \sim \pmatr{1\\1\\-1} & \braket{\phi_3} & \sim \pmatr{0\\1\\1} \\ [5ex]
\braket{\phi_4} & \sim \pmatr{1\\3\\1} & \braket{\phi_5} & \sim \pmatr{1\\0\\1} & \braket{\phi_6 } & \sim \pmatr{1\\0\\-1}.
\end{array}
\end{align}
The role of the VEVs (containing two zero entries) of the superfields $\phi_{e,\mu,\tau}$ appearing in Eq.~\ref{eq:dcl} was already discussed in Section \ref{sec:dcl}. Meanwhile, the role of the VEVs of the flavons $\phi_{3,4}$ (redubbed $ \phi_{\mathrm{atm,sol}} $) was described in Section \ref{sec:neutrinos}. It is the special structure of these vacuum alignments,
combined with the phase of $ \eta $ in the neutrino mass matrix, that leads to the very successful prediction of the leptonic mixing angles (as described in Section \ref{sec:fit}). The remaining VEVs are not directly relevant to the masses and mixings of SM fermions, but help shape the VEVs of $ \phi_{\mathrm{atm}} $ and $ \phi_{\mathrm{sol}} $.
\subsection{Flavon vacuum phases}
\label{sec:phasefixing}
With the direction of the $A_4$ triplet flavons $\phi$ fixed, we turn now to a discussion of how to fix the relative phase $\rho_\mathrm{atm}\!-\!\rho_\mathrm{sol} \equiv \mathrm{arg}(v_{\mathrm{atm}} v_{\mathrm{sol}}^{\ast})$ to a discrete choice. We present a mechanism which does this by adding a number of fields $ P $ that are $ A_4 $ and $ SU(5) $ singlets, also given in Table \ref{ta:driving}, and which resemble the $ O $ fields except they do not force orthogonality between the flavons $ \phi $.
These fields and their respective charge assignments result in the following invariant superpotential terms:
\begin{align}
\begin{split}
W_{\mathrm{phase}}
&\sim P_{ee}(\phi_e\phi_e+ M^2+P_{ee}^2)+P_{\mu\mu}(\phi_\mu\phi_\mu+Z_2 Z_3+P_{\mu\mu}^2) \\
&\qquad+P_{e4}(\phi_e\phi_4+\theta_1\theta_2)+P_{22}(\phi_2\phi_2+\theta_1\theta_1) \\
&\qquad+ P_{1e}(\phi_1\phi_e+\phi_6\phi_\tau)+P_{44}(\phi_4\phi_4+\phi_5\phi_\tau)+P_{34}(\phi_3\phi_4+\phi_6\phi_e)\\
&\qquad+ P_{33}^{1,2}(\phi_3\phi_3+\phi_1\phi_\mu+\phi_5\phi_e)+ P_{2\tau}(\phi_2\phi_\tau+\phi_3\phi_6+\phi_4\phi_5),
\label{eq:fixing}
\end{split}
\end{align}
where each term technically has an associated real coupling $ \lambda $ which is $ \mathcal{O}(1) $ and may be made positive
by field redefinitions. We omit these for simplicity as they have no effect on the general argument presented here, with one caveat: the two superfields $ P_{33}^{1,2}$ have exactly the same quantum numbers but different $ \lambda $ couplings to flavons. Due to this duplication there are two independent relations between the flavon VEVs
involving different $\lambda$ couplings which leads to an additional constraint on the phases of the respective VEVs. Exact values of these $ \lambda $ are not specified; it suffices that they are not equal.
Furthermore, the primary role of the SU(5) adjoint fields $ Z_2 $ and $ Z_3 $ which couple to $ P_{\mu\mu} $ is in the GUT breaking mechanism (as discussed in Section \ref{sec:GUTbreaking}). Their phases are fixed separately by other superpotential terms.
We begin the analysis of these terms by noting they do not affect the alignments of the flavons $ \phi$.
The corresponding $ F $-terms for each field $ P_{ij} $ produces a set of coupled equations that admit a solution where none of the $A$, $O$, and $P$ fields but all the flavons obtain a VEV. Omitting the (real, positive, $\mathcal{O}(1)$) $\lambda$ coefficients, these VEVs have the structure:
\begin{equation}
\begin{aligned}
v_e &\sim M &
v_\mu &\sim (v_{Z_2} v_{Z_3})^\frac{1}{2} \\
v_\tau &\sim (v_{Z_2}~v_{Z_3})^{-\frac{1}{3}} M^\frac{5}{3} &
v_1 &\sim (v_{Z_2} v_{Z_3})^{-\frac{1}{2}}~v_3^2 \\
v_2 &\sim (v_{Z_2} v_{Z_3})^{-\frac{1}{6}} M^{-\frac{7}{3}}~v_3^3 \qquad&
v_4 &\sim (v_{Z_2} v_{Z_3})^{-\frac{1}{6}} M^\frac{1}{3}~v_3 \\
v_5 &\sim M^{-1}~v_3^2 &
v_6 &\sim (v_{Z_2} v_{Z_3})^{-\frac{1}{6}} M^{-\frac{2}{3}}~v_3^2 \\
v_{\theta_1} &\sim (v_{Z_2} v_{Z_3})^{\frac{1}{6}}M^{-\frac{7}{3}}~v_3^3 &
v_{\theta_2} &\sim (v_{Z_2} v_{Z_3})^{-\frac{1}{6}} M^\frac{11}{3}~v_3^{-2} \\[1ex]
v_{O}&=~v_{P}=~v_A=~0.
\end{aligned}
\end{equation}
Regarding the magnitudes of the VEVs, two comments are in order. We assumed above that
$M$ sets the scale of the VEV of $ \phi_e $, which is in contradiction with our previous assumption that
it be $ \mathcal{O}(10^{-3}) M$. This violates our simplifying assumption that all mass scales are equal,
and demonstrates that some spectrum of mass scales is in fact required in this model.
As for the VEV $v_3$, as discussed earlier, it is driven to a specific scale $\Lambda_3$ radiatively \cite{Greene:1986jb}.
Writing $ \rho_i \equiv \mathrm{arg}(v_i)$, this VEV structure gives (up to multiples of $\pi$) the phase relation
\begin{equation}
\rho_4= \frac{2 \pi n}{3} - \frac{1}{6}(\rho_{Z_2} + \rho_{Z_3}) + \rho_3 ,
\label{fix}
\end{equation}
where $n$ is an integer, and similar relations for the other flavons as linear combinations of $ \rho_3 $, $(\rho_{Z_2} + \rho_{Z_3})$ and multiples of $ 2\pi/3 $. This is an important equation since it fixes the relative phase $\rho_3-\rho_4 = \rho_\mathrm{atm}\!-\!\rho_\mathrm{sol}$ in terms of $\frac{1}{6}(\rho_{Z_{2}}+\rho_{Z_{3}})$. As discussed in Section \ref{sec:GUTbreaking}, $\rho_{Z_{2}}+\rho_{Z_{3}} = \frac{2 \pi k'}{3}$, where we also establish that $\rho_{\xi}= \frac{2 \pi k}{9}$, for integers $k$, $k'$. From Eq.~\ref{eta}, $\eta\equiv -\rho_{\xi}+2(\rho_\mathrm{atm}\!-\!\rho_\mathrm{sol} )$, so we conclude that $\eta$ is one of the nine complex roots of unity.
\section{GUT scale symmetry breaking, proton decay and the strong CP problem \label{sec:GUT}}
In this section we discuss the aspects of the model related to grand unification, starting with how the R-symmetry and the GUT gauge group are spontaneously broken. We refer to the superfields involved as the scalar sector; they are listed in Tables \ref{ta:SMF} and \ref{ta:RMPM}. We then describe the details of the MP mechanism, and finish this section with an analysis that justifies the absence of dangerous proton decay operators in the model.
\subsection{\secheadmath{SU(5)} and \secheadmath{\mathbb{Z}_4^R} breaking \label{sec:GUTbreaking}}
As previously discussed, the $\Upsilon$ messengers form pairs; their mass scale, unprotected by any symmetry, is near the highest scale of the theory, which we represent generically as $M$. The GUT breaking superpotential with non-renormalisable terms is then%
\footnote{A renormalisable term of the form $ Z_{2}H_{24}\Pi_2$, allowed by the symmetries, has been dropped to make the discussion more transparent. This term mixes the VEVs of the GUT breaking scalars with the ones in the MP mechanism so they should be naturally around the same scale ($M \sim\!M_{\mathrm{GUT}}$). Beyond this, its practical effect is minimal: the fields obtain VEVs with or without this term. Since the VEVs get very complicated when this ``mixing'' term is included, we ignore it for simplicity, simply bearing in mind that VEVs from both sets of fields are related.}%
\begin{align}
\label{eq:GUTpot}
\begin{split}
W_{\mathrm{GUT}} & = Z_1\left(M\Lambda_{24}+\frac{\lambda_1}{M^2}H_{24}\xi^3+\lambda_2 Z_1^2\right)+Z_2\left(\frac{\lambda_3}{M^2}\Lambda_{24}\xi^3+\lambda_4Z_2^2\right) \\
& \qquad +Z_3\left(\lambda_5H_{24}^2+\lambda_6 Z_3^2\right).
\end{split}
\end{align}
We have five GUT adjoint superfields, three of which (the $Z$ fields) are charged by $2$ and two ($ \Lambda_{24} $ and $ H_{24} $) by $0$ under the R-symmetry. Also appearing in $W_{\mathrm{GUT}}$ is the Majoron $\xi$, the GUT singlet field which we have seen is involved in giving mass to several SM fermions and whose VEV breaks lepton number by giving the right-handed neutrinos their Majorana mass.
The supressions of the non-renormalisable terms in Eq.~\ref{eq:GUTpot} come precisely from the mass of the $\Upsilon$ messengers, as displayed in Fig.~\ref{fig:GUTpotcomplete}.
\begin{figure}[ht]
\centering
\begin{subfigure}{0.5\textwidth}
\centering
\includegraphics[scale=0.45]{images/pot1.pdf}
\caption{}
\end{subfigure}%
\begin{subfigure}{0.5\textwidth}
\centering
\includegraphics[scale=0.45]{images/pot2.pdf}
\caption{}
\end{subfigure}
\caption{Diagrams for non-renormalisable potential terms.}
\label{fig:GUTpotcomplete}
\end{figure}
$W_{\mathrm{GUT}}$ has a non-trivial minimum:
\begin{equation}
\label{eq:GUTVEVs}
\begin{split}
v_{Z_1} &=-\frac{i 2^{2/3} \lambda _1 \lambda _4^{1/3} \lambda _6^{1/6}}{3^{1/2} \lambda _2 \lambda _3 \lambda _5^{1/2}}M\\
v_{Z_2} &=\frac{i 2^{1/3} \lambda _1 \lambda _6^{1/6}}{3^{1/2} \lambda _2^{2/3} \lambda _3 \lambda _5^{1/2}}M \\
v_{Z_3} &= \frac{i \lambda _1 \lambda _4^{1/3}}{3^{1/2} \lambda _2^{2/3} \lambda _3 \lambda _5^{1/2} \lambda _6^{1/6}}M
\end{split}\qquad
\begin{split}
v_{H_{24}} &=\frac{ \lambda _1 \lambda _4^{1/3} \lambda _6^{1/3}}{\lambda _2^{2/3} \lambda _3 \lambda _5}M \\
v_{\Lambda_{24}} &=\frac{2^{1/3} \lambda _1^2 \lambda _4^{2/3} \lambda _6^{1/3}}{\lambda _2 \lambda _3^2 \lambda _5} M\\
\langle \xi\rangle^3 &=\frac{2^{1/3} \lambda _4^{1/3}}{\lambda _2^{1/3} \lambda _3}M^3,
\end{split}
\end{equation}
where all the adjoint scalars get a VEV of the form $\braket{\Phi_{24}}=v_{\Phi_{24}}~\mathrm{diag}(2,2,2,-3,-3)$.
By themselves, the $F$-terms associated with $W_{\mathrm{GUT}}$ also allow a trivial minimum where the magnitude of each VEV vanishes. But after SUSY is broken and we consider the effects of the small contribution from radiative breaking \cite{Ibanez:1982fr} to the scalar components of the GUT breaking superfields (as we did in Section \ref{sec:A4} for the $A_4$ breaking flavons), the stationary point with vanishing magnitudes is no longer a minimum due to the radiatively induced negative squared mass term. To a very good approximation the true minima are given by the magnitudes in Eq.~\ref{eq:GUTVEVs}, which are now a lower energy state than the trivial $ F $-term solution.
We conclude that Eq.~\ref{eq:GUTpot} can generate GUT and R-symmetry breaking at high scale,
with $\mathbb{Z}_4^R$ broken to $\mathbb{Z}_2^R$ (standard R-parity preserved) by the $Z_i$ VEVs. Because $\mathbb{Z}_4^R$ is broken at a high scale, the no-go theorem from \cite{Fallbacher:2011xg} does not apply to our model and we verified that all the components of the $SU(5)$ adjoints acquire GUT scale masses.
A slightly unappealing issue with $W_{\mathrm{GUT}}$
is that the minimum requires some non-${\cal O}(1)$ choice of $\lambda$ parameters if we are to obtain a hierarchy between the VEVs of $H_{24}$ and $\Lambda_{24}$, and an appropriate value for $ \braket{\xi}\!/M $ as shown in Eq.~\ref{eq:vevfit}. These requirements come from the successful fit to up and down quark and charged lepton masses, as discussed in Sections \ref{sec:u}, \ref{sec:dcl} and \ref{sec:fit}, and partly also for the $\mu$ term, as will be discussed shortly.
However, since the messengers will in general have different masses (recall we set them all equal to $M$ only for simplicity), the $\lambda$ parameters need not be as hierarchical as Eq.~\ref{eq:GUTVEVs} appears to indicate. For example, if the masses of messengers $\Sigma$ are slighly larger than the GUT scale masses of messengers $\Upsilon$, this would allow all $\lambda$ to be $\mathcal{O}(1)$.
We note also that, although we are considering a situation where the superpotential parameters (the $M$ and $\lambda$ couplings) are real due to CP conservation, the VEVs of the GUT breaking scalars may be complex, since they depend on $n^\mathrm{th}$ order roots of real numbers. As has been shown in Sections \ref{sec:neutrinos} and \ref{sec:phasefixing}, the phases of the fields $ \xi $, $ Z_2 $ and $ Z_3 $ are relevant for establishing the physical phase $ \eta $ in the neutrino mass matrix, which controls neutrino masses and mixing. We see immediately that $ \rho_{\xi} = \frac{2\pi k}{9} $, for integer $ k $, i.e. one of nine roots of unity. While $ \rho_{Z_2} $ and $ \rho_{Z_3} $ individually can be any of six roots, originating in the factor $ \lambda_5^{1/6} $, their product $ Z_2 Z_3 $ cancels this factor such that the largest root is a third, giving $ \rho_{Z_2} + \rho_{Z_3} = \frac{2\pi k'}{3} $, for integer $ k' $.
\subsection{Doublet-triplet splitting, Higgs mixing and the \secheadmath{\mu} term \label{sec:splitting}}
Given that we have a number of GUT representations containing (SM gauge group) $SU(2)$ doublets and triplets we turn now to a brief discussion of how doublet-triplet splitting is achieved in this model. Although one could alternatively introduce the double MP mechanism \cite{Hisano:1994fn} as demonstrated in \cite{Antusch:2014poa},
here we limit ourselves to describing how the MP mechanism is implemented with the fields listed in Table \ref{ta:RMPM}.
We have a superpotential
\begin{equation}
W_\Pi = \Pi_1\left(\lambda_7\Pi_1^2+ M\Pi_2+\frac{\lambda_8}{M^2}\Pi_2^4\right),
\label{eq:W75}
\end{equation}
which gives $\Pi$, the \textbf{75}s of $SU(5)$, their VEVs
\begin{equation}
\label{eq:75VEV}
v_{\Pi_1} =-\frac{ 1}{ {16}^{1/3} \lambda _7^{1/2} \lambda_8^{1/6}}~M, \qquad
v_{\Pi_2} =-\frac{ 1}{4 \lambda_8^{1/3}}M,
\end{equation}
which are aligned with the SM singlet inside the $SU(5)~\textbf{75}$. The non-renormalisable term in $W_\Pi$ comes from the diagram in Fig.~\ref{fig:W75}.
\begin{figure}[!ht]
\centering
\includegraphics[scale=.6]{images/pot6.pdf}
\caption{Diagram for non-renormalisable term in the 75 potential.}
\label{fig:W75}
\end{figure}
With Eq.~\ref{eq:75VEV}, the MP mechanism proceeds through the superpotential:
\begin{align}
\begin{split}
W_{\mathrm{MP}} &\sim H_{\bar 5} \Omega_1\Pi_2+H_5\Omega_2\Pi_1+\xi\Omega_1\Omega_2 \\
&\qquad + H_{\bar{45}}\Omega_3\Pi_2+M H_{\bar{45}}H_{45}+M\Omega_3\Omega_4 \\
&\qquad + H_{\bar 5} H_{45}\Pi_2+H_5 H_{\bar{45}}\Pi_1\left(\frac{\xi}{M}\right)^8.
\end{split}
\end{align}
The very high order non-renormalisable term at the end arises through the $\Sigma$ messengers already listed in Table \ref{ta:Mess} (half of which participate in the up quark Yukawa terms, as discussed in Section \ref{sec:u} and illustrated in Fig.~\ref{fig:upmass}).
Strictly speaking this term does not participate in splitting the masses of $SU(2)$ doublets and triplets, rather it is the source of the $\mu$ term in our model, as shown below.
The terms in $W_{\mathrm{MP}}$ generate mixing between the \textbf{45}s and \textbf{5}s of $SU(5)$.
The mass matrix for the $SU(2)$ triplets contained in the \textbf{45}s and \textbf{5}s is:
\begin{align}
\begin{split}
M_{\mathbf{3}} &= \overline{\textbf{3}}^T \pmatr{0 &v_{\Pi_2} &v_{\Pi_2} & 0 \\ v_{\Pi_1}\tilde{\xi}^8 & M & 0 & v_{\Pi_2} \\ v_{\Pi_1} & 0 & \braket{\xi} &0 \\0 &0 & 0 & M} \textbf{3} \\[2ex]
\overline{\textbf{3}}^T &=\pmatr{\textbf{3}(H_{\bar 5})&\textbf{3}(H_{\bar{45}})& \textbf{3}(\Omega_2)& \textbf{3}(\Omega_4)} \\
\textbf{3}^T &=\pmatr{\textbf{3}(H_5) &\textbf{3}(H_{45}) & \textbf{3}(\Omega_1) & \textbf{3}(\Omega_3)},
\end{split}
\end{align}
where once again $ \tilde{\xi} = v_{\xi}/M $. Taking $\braket{\Pi_{1,2}}\sim M$, the eigenvalues of this mass matrix are all of order $M$ (i.e. at the GUT scale), leading us to conclude there are no light $SU(2)$ triplets. Conversely, for the doublets we have the matrix:
\begin{align}
\pmatr{\mathbf{2}(H_{\bar 5}) & \mathbf{2}(H_{\overbar{45}})} \pmatr{0 & v_{\Pi_2} \\ v_{\Pi_1} \tilde{\xi}^8 & M} \pmatr{\mathbf{2}(H_{5}) \\ \mathbf{2}(H_{45})}.
\end{align}
It is clear that were it not for $\tilde{\xi}^8$, the determinant of this mass matrix would vanish. We may rotate to the basis of the MSSM Higgs doublets $H_{u,d}$ and a pair of very heavy doublets $H_{u,d}^H$:
\begin{align}
\begin{split}
\pmatr{\mathbf{2}(H_{\bar{5}}) \\ \mathbf{2}(H_{\overbar{45}})} &\approx \frac{1}{\sqrt{2}}\pmatr{1 & -1 \\ 1 & 1} \pmatr{H_d^H \\ H_d} \\
\pmatr{\mathbf{2}(H_{5}) \\ \mathbf{2}(H_{45})} &\approx \pmatr{\tilde{\xi}^8 & 1 \\ -1 & \tilde{\xi}^8} \pmatr{H_u^H \\ H_u}.
\end{split}
\end{align}
The usual MSSM term $\mu H_d H_u$ comes from this mechanism with:
\begin{equation}
\mu\sim\frac{v_{\Pi_1}v_{\Pi_2}}{M}\tilde\xi^8,
\end{equation}
where $v_{\Pi_1}$ provides the necessary $\mathbb{Z}_4^R$ breaking. Using the fit from Eq.~\ref{eq:vevfit} we see that $\tilde{\xi}^8 \sim 1.6\times 10^{-10} M_{\mathrm{GUT}}$. If we choose the couplings at the vertices of the tower that generates the $\xi^8$ term to be $\sim 0.5$ we may get a term $\mu\sim \mathcal{O}(10^2-10^3)$ GeV without any fine-tuning.
\subsection{Proton decay}
A classic problem in GUT theories, and in particular those based on $SU(5)$, is the prediction of excessively fast proton decay. The most dangerous processes come from the ``dimension 5'' operators $TTTF$ (for a discussion of dimension 6 operators we refer the reader to \cite{Antusch:2014poa}).
The ``dimension 5'' operators are forbidden by the symmetries of the model, but related higher order
operators of the following form are allowed:
\begin{equation}
T_i T_j T_k F \frac{Z \phi}{M^3}\left(\frac{\xi}{M}\right)^{n_{ijk}},
\label{proton}
\end{equation}
where the extra superfields shown are needed for such terms to be invariant under the symmetries. Since we are working with the renormalisable theory, in order for this type of effective term to be present at the GUT scale, there must be messengers allowing them. In this case, an analysis of the $SU(5)$ index structure revels there should either be messengers that are $SU(5)$ $\overbar{\textbf{10}}$, or $SU(5)$ $\textbf{5}$ that are also charged under $\mathbb{Z}_4^R$. As one can confirm from Table \ref{ta:Mess}, our model has neither: $\overbar{\textbf{10}}$ messengers were not used, and the $\textbf{5}$ messengers are all neutral under $\mathbb{Z}_4^R$. We conclude therefore that our symmetry content, together with the existing set of messengers, do not allow any such GUT scale suppressed operators that would lead to excessively fast proton decay to be generated. The operators in Eq.~\ref{proton} may in principle be generated by physics at the Planck scale, with the scale $M$ replaced by the Planck mass, leading to highly suppressed proton decay.
\subsection{Strong CP problem and the Nelson-Barr resolution}
We first recall the strong CP problem, namely that the physical angle $\overline{\theta}=\theta_{\mathrm{QCD}}-\theta_q$, where $\theta_{\mathrm{QCD}}$ multiplies the topological gluon term $(g_s^2/32\pi^2)G\tilde{G}$ and $\theta_q= \arg \det (Y^uY^d)$, is limited to be $\overline{\theta} <10^{-10}$ by the non-observation of the neutron EDM \cite{Beringer:1900zz, Baker:1997ed}. The origin of such a small number, $\overline{\theta} <10^{-10}$, is commonly called the strong CP problem. It is interesting to compare this with the CP violation related to the weak interaction in the quark sector; the relevant quantity is the Jarlskog invariant $J^q \sim \det[Y^u Y^{u\dagger},Y^dY^{d\dagger}]$, which, when compared to data, is required to be non-vanishing, and indeed in the standard parameterisation, requires a large phase angle $\delta^q \sim 1$.
It turns out that our model resolves the strong CP problem without relying on the introduction of axions. Unlike the axion solution, which requires a global $U(1)$ symmetry with a colour anomaly, we shall rely on the fact that the high energy theory conserves CP, ensuring that $\theta_{\mathrm{QCD}}=0$. CP is then spontaneously broken in such a way as to yield $\delta^q\sim 1$ while maintaining $\overline{\theta} <10^{-10}$ and in particular $ \theta_q <10^{-10}$. How it achieves this feat can be seen from Eq.~\ref{eq:yufull} where $Y^u$ is real, while the structure of $Y^d$ in Eq.~\ref{eq:ydfull} gives it a real determinant. This is due to the lack of a Yukawa term $ Y^d_{21} $, meaning the coupling $ Y^d_{12} $ (which is the only complex Yukawa coupling) does not appear in the determinant of $ Y^d $. Therefore there are no contributions to $ \theta_q$ even after spontaneous CP breaking. This is similar to the Nelson-Barr mechanism \cite{Nelson:1983zb, Barr:1984qx}, where the triangular form of Yukawa matrices was proposed, although in our model $\theta_q$ vanishes due to the triangular form of $Y^d$ only, with $Y^u$ being non-triangular and real.
For a successful resolution of the strong CP problem, we must ensure that no higher order corrections to the Yukawa matrices arise which would violate the bound $ \theta_q <10^{-10}$. The main focus is on the Yukawa coupling $Y^d_{21}$ which is zero at leading order but which may in principle receive higher order corrections, violating the triangular structure. However in our model such higher order corrections are absent at the field theory level with the specified messenger sector. This entry in the Yukawa matrix would arise from the coupling of the bilinears $T_2 H_{\bar 5, \overbar{45}}$ to the bilinear $\phi_e F$. Since these terms are non-renormalisable, we require messengers to form them. The messengers that could produce such terms are the $ X_i $ fields in Table \ref{ta:Mess}. With these messengers, the only allowed connection to $\phi_e F$ is $T_1 H_{\bar 5}$ (contributing to $ Y^d_{11} $), thus forbidding the $ Y^d_{21} $ term, even when allowing for all higher-order corrections. Therefore the specified model has no strong CP violation arising from $Y^d_{21}$ since the required operators are not generated at the field theory level.
It is also important to consider the effect of higher order corrections arising from the Planck scale, since such operators only have to respect the symmetries of the model, and do not require the specified messenger sector to generate them. The biggest contribution would come from the term%
\footnote{This term would also give a contribution to lepton angles of $\mathcal{O}(10^{-3})$ which is negligible.}
\begin{equation}
T_2 H_{\bar 5}\frac{\phi_e}{M_P}F.
\end{equation}
With a general choice of phase, such a term would lead to $\theta_q\sim 10^{-4}$ which is far too big. However the contribution to $\theta_q$ may be avoided by a judicious choice of GUT breaking phases. As stated in section \ref{sec:fit}, the physical phase in the down quark Yukawa matrix is $\zeta=\rho_\xi-2\rho_{H_{24}}-\rho_{\Lambda_{24}}$. The new Planck suppressed term has a phase $\zeta'=-\rho_\xi+2\rho_{\Lambda_{24}}$. Choosing a relation between phases $2\rho_{H_{24}}=\rho_{\Lambda_{24}}$, then $\zeta=-\zeta'$ and this way the contribution to $\theta_{q}$ vanishes. This happens for one in three cases. The next biggest contribution comes from a term
\begin{equation}
T_2 H_{\bar{45}}\frac{\xi^2\phi_e}{M_P^3}F,
\end{equation}
giving $\theta_q\sim 10^{-14}$ which is several orders of magnitude below the current experimental bound. Any other Planck suppressed terms allowed by the symmetries are further suppressed so we need not consider them. Therefore the model may resolve the strong CP problem even in the presence of Planck scale operators controlled only by symmetry. Finally, extra contributions may come from SUSY breaking terms. If we assume that there is no extra CP violation in this sector, which is controlled by the spontaneously CP violating flavons, the SUSY flavour problem is under control and such contributions to $\overline{\theta}$ are also expected to be negligible \cite{Antusch:2013rla}.
\section{The leptogenesis link}
\label{sec:link}
The link between leptogenesis and the PMNS matrix was first studied for sequential dominance in \cite{King:2002qh}. In the original form of CSD, the columns of the Dirac mass matrix in the flavour basis were orthogonal to each other and consequently the CP asymmetries for cosmological leptogenesis \cite{DiBari:2012fz} vanished \cite{Antusch:2006cw,King:2006hn,Bertuzzo:2009im, Choubey:2010vs}. In this model, leptogenesis does not vanish since the columns of the Dirac mass matrix in the flavour basis are not orthogonal.
Interestingly, since the seesaw mechanism in this model with two right-handed neutrinos only involves a single phase $\eta=2\pi /3$, both the leptogenesis asymmetries and the neutrino oscillation phase must necessarily originate from this phase, providing a direct link between the two CP violating phenomena in this model.
Following the arguments in \cite{Antusch:2006cw}, the produced baryon asymmetry $Y_B$ from leptogenesis in the seesaw model in Eq.~\ref{seesaw} satisfies
\begin{equation}\label{eq:BAU}
Y_B \propto \pm\sin \eta,
\end{equation}
where the ``$+$'' sign applies to the case $M_{\rm atm} \ll M_{\rm sol}$ and the ``$-$'' sign holds for the case $M_{\rm sol} \ll M_{\rm atm}$. Since the observed baryon asymmetry $Y_B$ is positive, it follows that, for $M_{\rm atm} \ll M_{\rm sol}$, we must have $\sin \eta$ to be positive, while for $M_{\rm sol} \ll M_{\rm atm}$ we must have $\sin \eta$ to be negative. We have seen that positive $\eta$ is associated with negative $\delta^l$ and {\it vice versa}. Although the global fits do not distinguish the sign of $\eta$, the present hint that $\delta^l \sim -\pi/2$ would require positive $\eta$, then in order to achieve positive $Y_B$ we require $M_{\rm atm} \ll M_{\rm sol}$, which is natural in our model, corresponding to ``light sequential dominance'', where successful leptogenesis may be achieved in the two right-handed neutrino model as discussed in \cite{Antusch:2011nz}.
\section{Conclusion \label{sec:con}}
We have presented here a fairly complete realisation of an $SU(5)$ GUT flavoured with
$A_4$, which leads to the MSSM plus two right-handed neutrinos below the GUT scale.
The $A_4$ family symmetry unifies the three families of 5-plets $F$ and its vacuum alignment
determines the Yukawa matrices.
In addition a $\mathbb{Z}_9\times \mathbb{Z}_{6}$ symmetry provides the mass hierarchy and controls
spontaneous CP violation in both the quark and lepton sectors while a $\mathbb{Z}_4^R$ symmetry
is broken to $\mathbb{Z}_2^R$, identified as the usual R-parity.
Proton decay is under control in this model, with the symmetries forbidding dangerous dimension-5 operators, and similar (but higher order) operators being very suppressed.
The strong CP problem is resolved in a similar way to the Nelson-Barr mechanism.
The model is highly predictive and satisfies many distinct (and non-trivial) phenomenological requirements.
Imposing CP at the high scale is an important feature of the model. If we do not impose CP then
all couplings become complex, leading to all VEVs having undetermined phases.
In particular the phase $ \eta $, present in both neutrino mixing and leptogenesis, is no longer restricted to a discrete choice. However the link between leptogenesis and low energy phenomenology remains.
On the other hand we would no longer solve the strong CP problem.
We highlight the ubiquitous nature of the flavon field $\xi$ across all the sectors of the model: it triggers
spontaneous CP violation in both the quark and lepton sectors, generates up-type quark mass hierarchies and CKM mixing, explains the smallness of down quark and electron masses and breaks lepton number, providing
the hierarchy between solar and atmospheric right-handed neutrino masses.
In addition, the $\xi$ field is responsible for generating the small (complex) $\mu$ term.
The phase of the $\xi$ field VEV also contributes to the relative phase $\eta$
appearing in the neutrino mass matrix.
This phase, the only one
appearing in the neutrino mass matrix and in the formula
for the baryon asymmetry of the universe,
provides a direct link between the PMNS phase $\delta^l$ and leptogenesis.
We emphasise that the entire PMNS matrix is predicted without any free parameters, up to nine choices for $ \eta $, where we select $ \eta = +2\pi/3$, since it is preferred by comparing CSD3 to data.
The required vacuum alignments are provided from $A_4$.
The model effectively serves to
yield the CSD3 scheme (with two right-handed neutrinos) within a fully working and viable SUSY GUT of flavour
in which all quark and lepton masses and mixings are successfully described.
Indeed the model provides an excellent fit (better than one sigma) to all quark and lepton (including neutrino) masses, mixing and CP violation. All fermion mass hierarchies are understood in the sense that purportedly
$\mathcal{O}(1)$ couplings indeed contain no strong hierarchies. However the most immediate
predictions of the model are those of CSD3 with two right-handed neutrinos
and $ \eta = +2\pi/3$,
namely a normal neutrino mass hierarchy with $m_1=0$, a reactor angle of $\theta^l_{13}\approx 8.7^{\circ}$,
a solar angle $\theta^l_{12}\approx 34^{\circ}$, close to maximal
atmospheric mixing $\theta^l_{23}\approx 46^{\circ}$ and almost maximal leptonic CP violation, with
an oscillation phase $\delta^l\approx -87^{\circ}$ consistent with the current
hint $\delta^l\approx -\pi/2$.
The reason why the field content is so large is that the model is fairly complete.
In particular it is {\em renormalisable at the GUT scale}, which requires a large explicit field content including many heavy messenger superfields. It also addresses many aspects relevant both to a GUT and to family symmetry models (stopping short of discussing the details of SUSY breaking and its string theory completion).
In particular, the $A_4$, $SU(5)$ and $R$-symmetry symmetry breaking sectors all require large field content.
For example, the GUT symmetry is
broken by an explicit superpotential at the GUT scale, including doublet-triplet splitting via
a missing partner mechanism (leaving no light exotic degrees of freedom at the low scale),
Higgs mixing and the origin of the MSSM $\mu$ term of the right order of magnitude.
Despite the many successes of the model, there are inevitably several important issues
that lie beyond the scope of this paper.
To take one example, we do not discuss GUT scale threshold corrections, which will
be important in maintaining successful gauge coupling unification in the presence of many fields, including
colour triplets, at the GUT scale.
In fact all the additional superfields in non-trivial representations of
the gauge group may have masses at or above the GUT scale.
Another important issue is that of the low energy superpartner spectrum in this model. While we expect
SUSY induced flavour changing to be under control for the 5-plets, which are unified into an $A_4$
triplet, this is not the case for the 10-plets $T_i$ which are singlets of $A_4$,
leading to flavour violation in the super-CKM basis.
It would be interesting to study this in the future.
\section*{Acknowledgements}
This project has received funding from the European Union's Seventh Framework Programme for research, technological development and demonstration under grant agreement no PIEF-GA-2012-327195 SIFT.
The authors also acknowledge partial support from the European Union FP7 ITN-INVISIBLES (Marie Curie Actions, PITN- GA-2011- 289442) and CONACyT.
|
\section{INTRODUCTION}
The parametrization of longitudinal profiles for particle showers produced by the primary
nuclei in atmosphere, is an essential tool for the primary nuclei identification
and the evaluation of primary energy. Experiments that sample the shower
longitudinal development using Cherenkov light images \cite{HiRes} or air fluorescence
\cite{Auger,TArray} from different traversed atmosphere depths, extract the position of shower maximum,
which is sensitive to the incident primary nucleus.
The integral of shower profile strongly correlates with primary energy \cite{Matthews}.
The shower longitudinal profile is a dependence of the shower particle number
($N$) on a given traversed atmospheric depth, $T$. The parametrization of shower profile commonly
used in cosmic-ray experiments is Gaisser-Hillas formula \cite{GH}:
\begin{equation}\label{one}
N(X)=N_{max}\left(\frac{X}{X_{max}}\right)^{X_{max}}\exp{(X_{max}-X)}\;,
\end{equation}
where $X=(T-X_0)/\lambda$ and $X_{max}=(T_{max}-X_0)/\lambda$.
The maximum number of shower particles $N_{max}$ at the traversed atmosphere depth $T_{max}$ along with $X_0$ and $\lambda$
in expression (1) are free parameters that depend on the primary nucleus and energy.
The standard primary nuclei composition consists of
the first 28 nuclei of the periodic table with mass (nucleon) numbers $A=1,\dots56$
usually divided into four-six groups (species) H, He, CNO-like, Si-like, Fe-like.
The large number of nuclei species (more than four) increases the uncertainties of the
inverse problem ($E$ and $A$ reconstruction) falsely improving the agreement of experiment
with the theory \cite{samo}.
The primary energy region responsible for particle shower
detection at the observation level begins at about $E>1$ PeV and ends at GZK cutoff
energies \cite{Auger}.
The efficiency of four-parametric parametrization (1) is in its
applicability to a wide range of energies and primary nuclei.
However, the observed correlations between parameters
result in a loss of the physical meaning of $X_0$ and $\lambda$ \cite{Mont} and reduce the range
of effective atmosphere depths for Eq.~(1).
\section{Parametrization}
Here, an alternative parametrization $N(T,E,\boldsymbol{\varepsilon})$ for particle shower longitudinal profile is proposed
using three non-correlating parameters that depend on the primary particle energy and
nucleon energy, $\boldsymbol{\varepsilon}$:
\begin{equation}\label{two}
N(x)=N_{\max}\exp{(-\frac{1}{2}\left(\frac{\ln{x}}{\delta(x)}\right)^2)}\;,
\end{equation}
where
\begin{equation}\label{three}
\delta(x)=\alpha-\beta(\tanh{x})^{\frac{1}{4}}
\end{equation}
is the profile shape function of variable
\[
x=\frac{T}{T_{max}}\;.
\]
The shower maximum position, $T_{max}(\boldsymbol{\varepsilon})$, and
shape function, $\delta(x,\boldsymbol{\varepsilon})$, turned out to be
dependent on the primary particle energy per nucleon,
\[
{\boldsymbol{\varepsilon}}=\frac{E}{A}\;,\;\;(\text{PeV/n}).
\]
The maximum number of shower particles, $N_{max}(E,\boldsymbol{\varepsilon})$
is factored into the primary energy and a function of nucleon energy only. The corresponding approximations
for the parameters of shower longitudinal profile (2,3) are:
\begin{subequations}
\begin{align}
&\alpha =0.707+0.209{\boldsymbol{\varepsilon}}^{-0.084}\;,\;\;\;\;\; \beta =\sqrt{\alpha/2.59}\;,\\
&T_{max} =433.5+38.9(\ln({\boldsymbol{\varepsilon}}A_{Fe}))^{0.857},\;\text{(g/cm$^2$)},\\
&N_{max} =0.653(E/1\text{GeV})(1-e^{-2.5{\boldsymbol{\varepsilon}}^{0.12}})\;,
\end{align}
\end{subequations}
where $A_{Fe}=56$ and $\boldsymbol{\varepsilon}$ is in the units of PeV/n. The goodness-of-fit tests for (4a-c) were $\chi^2<1$ at
negligible correlations between the $\alpha$, $T_{max}$ and $N_{max}$ parameters.
\section{Shower Profiles}
The values of
free parameters in expressions (1) and (2) were obtained from simulated shower profiles
(training sample) using {\footnotesize{CORSIKA}} \cite{CORSIKA} ({\footnotesize{SIBYLL}} \cite{SIBYLL}) code for four primary nuclei $A\equiv1, 4, 16, 56$
at six energies $E\equiv1, 10, 100, 500, 2500, 10^4$ PeV. Shower profiles were studied for
10 atmosphere depths $T\equiv 100, 200,\dots1000$ g/cm$^2$ at two zenith angles, $\cos\theta=0.7$ and $1$.
The shower particle energy threshold was $E_e>1$ MeV. Simulation statistics were provided for less than 2-3\%
statistical errors in the whole measurement range.
The averaged shower profiles were approximated by expressions (1) and (2) using 13 reference depths.
The results are presented in Fig.~\ref{curve1}. It is seen that the parametrization (1) (dashed lines) underestimates the shower sizes
at large atmosphere depths.
\begin{figure}
\includegraphics[scale=0.35]{curve1.eps}
\caption{\label{curve1} Average longitudinal shower profiles for 13 traversed atmosphere depths
produced by H, He, O and Fe primary nuclei with 6 energies (from 1 PeV to 10 EeV).
The symbols are {\footnotesize{CORSIKA}} shower simulated data (training sample).
The dashed lines are results from the 4-parametric approximation (1). The solid lines
are parametrizations (2-4) computed for corresponding primary nuclei and energies.}
\end{figure}
The parametrization errors of expressions (1) and (2) and corresponding $\chi^2_{/\text{d.o.f.}}$ are presented in Fig.~\ref{err}
for different primary energies and nuclei.
The upper and middle panels show the errors of the four-parametric approximations of {\footnotesize{CORSIKA}}
simulated shower profiles using the $N_{max}, T_{max}, X_0$ and $\lambda$ parameters of expression (1)
and the $N_{max}, T_{max}, \alpha$ and $\beta$ of expressions (2,3). The lower panel of Fig.~\ref{err} shows
the errors of shower profiles $N(T,E,A)$ from expressions (2-4).
\begin{figure}
\includegraphics[scale=0.35]{err.eps}
\caption{\label{err} Parameterization errors $(N_{fit}-N_{corsika})/\Delta N_{corsika}$ for the 4-parametric expressions (1) and (2,3)
are shown in the upper
and middle panels respectively. The lower panel corresponds to the errors of the shower profiles $N(T,A,E)$ from
expressions (2-4) for different primary energies and nuclei.
}
\end{figure}
The normalized simulated (symbols) and parametrized (lines) shower profiles
are presented in Fig.~\ref{curve2}. It is seen that the parametrization (2)
effectively describe the shower profiles in the regions of both the maximum ($x\simeq1$, inset figure, solid line)
and asymptotic depths ($x\simeq3$). Equation (1) is systematically biased about $-2\%$
(inset figure, dashed line) at $x\simeq1$.
\begin{figure}
\includegraphics[scale=0.35]{curve2.eps}
\caption{\label{curve2}Normalized particle shower profiles. The symbols are {\footnotesize{CORSIKA}}
simulated data (training sample). The lines are the results of parametrization (2). Inset figure
is a zoom of selected rectangular region for $0.8<x<1.2$. The dashed lines correspond to
the parametrization (1).
}
\end{figure}
\section{Parameters}
The study of $T_{max}(\boldsymbol{\varepsilon})$ and $N_{max}(\boldsymbol{\varepsilon})$ dependence
on nucleon energy ($\boldsymbol{\varepsilon}$) are presented in the upper and lower panel of Fig.~\ref{Nmax} respectively.
The approximations of shower profiles using
parametrizations (1) and (2) were trailed for different lower ($T_{low}$) and upper ($T_{up}$) limits of traversed
atmosphere depth.
The estimated values for $T_{max}$ (Fig.~\ref{Nmax}, upper panel) were unbiased for all trails.
The line in Fig.~\ref{Nmax} corresponds to the expression (4b). The asterisk and cross symbols in Fig.~\ref{Nmax}
are correspondingly renormalized {\footnotesize{CORSIKA}} simulated data from \cite{Swordy}.
Estimations of $N_{max}$ (Fig.~\ref{Nmax}, lower panel) using expression (1) for approximations of
shower profile turned out to be dependent on boundary conditions for atmosphere depth (hollow and bold star symbols),
whereas the expression (2) remained practically unbiased (hollow and bold circle symbols) for different boundaries.
\begin{figure}
\includegraphics[scale=0.4]{tmax_nmax.eps}
\caption{\label{Nmax}Parameters $T_{max}$ (upper panel) and $N_{max}/E$ in units of particle/GeV (lower panel)
derived from expressions (1) and (2) for the different boundaries of traversed depths.
The lines correspond to the expressions (4b) and (4c) for $T_{max}$ and $N_{max}$ respectively.
}
\end{figure}
The shower profile shape functions $\delta(\boldsymbol{\varepsilon}|T)$ and $\delta(x|\boldsymbol{\varepsilon})$ are presented in Fig.~\ref{delta},
where the symbols (left panel) are the data extracted from {\footnotesize{CORSIKA}} simulated training sample. The solid lines in both panels correspond to the expressions (3,4a).
The dashed lines in the right panel of Fig.~\ref{delta} are the $0.5\%$ accuracy logarithmic simplifications of shape function (3),
\begin{equation}
\delta(x)\simeq
\begin{cases}
a-b\ln{x}, &\text{if $0.07\lesssim x\leqq1$;}\\
a-b\ln{x}/(1+1.59\ln{x}), &\text{if $x\geqq1$,}
\end{cases}
\end{equation}
where
\begin{equation}
\begin{split}
a& =0.215+0.145{\boldsymbol{\varepsilon}}^{-0.084}\;,\\
b& =0.086+0.011{\boldsymbol{\varepsilon}}^{-0.084}\;,
\end{split}
\end{equation}
at $\chi^2/450=0.7$.
The approximation (5) provides the analytic solution of the inverse function $N_0^{-1}(x)$ for $x<1$ and $N_1^{-1}(x)$ for $x>1$.
\begin{figure}
\includegraphics[scale=0.4]{sigma.eps}
\caption{\label{delta}Shower profile shape function (expression (3)) depending on the nucleon energy (left panel)
and normalized atmosphere depth (right panel). The solid lines correspond to the expressions (3,4a). The dashed lines in the right panel
are the logarithmic simplifications of the shape function according to (5,6).
}
\end{figure}
\section{Verification}
The verification of the universality of approximations (1) and (2) was performed by extrapolating the shower
profiles from $100-1428$ g/cm$^2$ interval to the $T=10$ g/cm$^2$ observation level,
corresponding to the earlier stage of shower development. The results are presented in Fig.~\ref{zero}.
\begin{figure}
\includegraphics[scale=0.4]{zero.eps}
\caption{\label{zero} Extrapolations of the parametrization (1) (left panel) and (2) (right panel) to the earliest
stage of shower development for different primary nuclei and energies (lines). The symbols are
the {\footnotesize{CORSIKA}} simulated data.
}
\end{figure}
The symbols at $T=10$ g/cm$^2$ in Fig.~\ref{zero} are the corresponding data from {\footnotesize{CORSIKA}}
simulated control sample, whereas the symbols at $T=100$ g/cm$^2$ are the representatives of the training sample (Section III).
It is seen that parametrization (1) being trained in the $100-1428$ g/cm$^2$ depth interval can not be extrapolated
to the region less than about 50 g/cm$^2$ (lines, left panel), whereas the parametrization (2) works correctly up to the beginning of
the atmosphere (lines, right panel).
The verification of the shower profiles (2-4) by the control samples of different nuclei and energies
are shown in Fig.~\ref{verify}.
The shower profile for primary Fe nucleus with energy $E=500$ PeV and corresponding $\boldsymbol{\varepsilon}\simeq8.93$ PeV/n
from training sample (Section III) are compared to the control sample of shower profiles produced by primary H, He, C, O and Si nuclei
with the same energy per nucleon (symbols). The lines in Fig.~\ref{verify} are the corresponding congruent predictions
from the parametrizations (2-4).
The results in Fig.~\ref{verify} confirm the $\boldsymbol{\varepsilon}$-dependence of the shower longitudinal profile shape
(expressions 4a,b). The shower profile amplitude ($N_{max}$) also depends linearly on the primary energy, $E$ (expression 4c).
\begin{figure}
\includegraphics[scale=0.4]{check.eps}
\caption{\label{verify}Control samples of shower profiles (symbols) produced by the
different primary nuclei with the same nucleon energy $\boldsymbol{\varepsilon}\simeq8.93$ PeV/n.
The lines are the predictions from (2-4).
}
\end{figure}
The good agreement in Fig.~\ref{verify} between predictions (lines) and simulated data
indicates the correctness of expressions (2-4) for shower profile description at least
with the accuracies of about $2-3\%$ in the whole measurement range.
\section{\label{sec2} INTEGRAL}
The right hand side of parametrization (2) at the corresponding normalization can be considered
as a probability density function and be used for primary energy evaluation \cite{Auger,Matthews}.
Unfortunately this function was missed by mathematicians and
by using numerical technique the required normalization
\begin{equation}\label{int}
\int_0^\infty{f(x,\boldsymbol{\varepsilon})\mathrm{d}x}\simeq1\pm10^{-4}
\end{equation}
was provided for probability density function
\begin{equation}
f(x)=\frac{1}{\sqrt{2\pi}\delta_0}\exp{(-\frac{1}{2}\left(\frac{\ln{x}}{\delta(x)}\right)^2)}
\end{equation}
with additional parameter
\[
\delta_0=0.226+0.148{\boldsymbol{\varepsilon}}^{-0.092}.
\]
The goodness-of-fit test for $\delta_0(\boldsymbol{\varepsilon})$ was $\chi^2=0.01$ in the $10^{-2}\leq\boldsymbol{\varepsilon}\leq10^4$ (PeV/nucleon) interval
and the upper limit of integral (7), $x_{max}=3$.
It is interesting to note the relation between parameters $\delta_0$ and
shape function $\delta(x)$ from expression (3):
\begin{equation}
\delta_0(\boldsymbol{\varepsilon})\simeq\frac{1}{x_{max}}\int_0^{x_{max}}\delta(x)\mathrm{d}x\; \pm1\%\;.
\end{equation}
The statistical parameters, the average ($\bar{x}$) and standard deviation ($\sigma_x$) of distribution (8), are well approximated
($0.1\%$ errors) by the following expressions that depend on the nucleon energy:
\[
\bar{x}=1.036+0.094{\boldsymbol{\varepsilon}}^{-0.12}
\]
at $\chi^2=0.1$, and
\[
\sigma_x=0.226+0.176{\boldsymbol{\varepsilon}}^{-0.092}
\]
at $\chi^2=1.1$.
\section{Fluctuations}
The main source of shower profile fluctuations is the depth of the first interaction of primary
particles in the atmosphere \cite{Stanev}. The exponentially distributed uncertainty of the
first interaction point results in the corresponding fluctuations of the shower profile (2) depending on the rate
of change ($\mathrm{d} N/\mathrm{d} x$) of the profile with respect to the depth, $x$. Thus, the fluctuations
should be maximal at the beginning of shower development ($x\simeq0$, Fig.~\ref{curve2}),
and minimal in the region of shower maximum, $x=1$. The dependence of the interaction length, $\lambda(A,E)$,
on the primary particle also results in the mass ($A$) and energy ($E$) dependences of shower profile fluctuations.
The statistical measure of fluctuations is the standard deviation of shower particle number,
$\sigma_N$. The corresponding values of $\sigma_N(x,A,E)/N(x)$ obtained from the
shower simulated dataset (Section III) are presented in Fig.~\ref{sigma} (symbols).
The inset panel shows the region of minimal fluctuations in detail.
\begin{figure}
\includegraphics[scale=0.4]{fluct.eps}
\caption{\label{sigma} Normalized standard deviations ($\sigma_N/N$) of shower particles for different primary nuclei and primary energies (symbols).
The lines are represent the parametrization (10) for energies of 1 PeV (dotted lines) and 10 EeV (solid lines). The dashed lines describe the fluctuations for intermediate
energies. The inset panel zooms in on the region of minimal fluctuations at $x\simeq1$.
}
\end{figure}
The lines in Fig.~\ref{sigma} correspond to the parametrizations
\begin{equation}
\frac{\sigma_N}{N}\simeq
\begin{cases}
a_1-a_2\ln{x}, &\text{if $x\leqq1$;}\\
a_1+a_3(\ln^{\eta}x)/x, &\text{if $x\geqq1$,}
\end{cases}
\end{equation}
where
\[
\begin{split}
a_1& =0.165A^{-0.32}E^{-0.13}\;,\\
a_2& =0.68A^{-0.185}E^{-0.009}\;,\\
a_3&=3.77A^{-0.386}E^{-0.035}\;,\\
\eta&=2.67A^{-0.080}E^{-0.027}
\end{split}
\]
at $\chi^2/470\simeq1.7$.
\section{Summary}
The standard inverse problem of cosmic ray physics in PeV-EeV energy region is the identification of
a primary nucleus (or elemental composition) and the estimation of its energy (or energy spectrum) by the detected
shower response at the observation level in the frames of a given interaction model.
The efficiencies of primary particle and primary energy estimators depend on both the accuracy (Section III) and
universality (Section V) of shower longitudinal profile description.
Historically, the conventional shower longitudinal profiles were proposed in 1960 (Greisen function) \cite{Greisen},
1977 (Gaisser-Hillas function) \cite{GH} and 2001 (Gaussian-In-Age approach) \cite{HiRes1}.
The efficiencies and accuracies of listed profile parametrizations are compared in Refs. \cite{Matthews,Song,Auger15} in detail.
The last Gaussian-In-Age approach \cite{Matthews} reduced the number of parameters to 3, and
decreased the intercorrelations between parameters of profile function in return for the narrow range of applicability in the vicinity of
shower maximum: $0.75 \lesssim s \lesssim1.25$ \cite{Song}, where $s=3/(1+2/x)$ is the shower age parameters.\\
The alternative shower longitudinal profile description (expressions 2-4), as opposed to the parametrizations
\cite{Matthews,GH,Greisen} , represents the first complete
formula for shower profile, $N(T,A,E)$, depending on the
atmosphere depth ($T$), primary nucleus ($A$) and primary energy $E$.
Expressions (2-4) provide the accuracies of about $2-3\%$ for the region of $0<T\le 1450$ g/cm$^2$, $A\le56$, $1 \text{ PeV}\le E\le10\text{ EeV}$.
The results are obtained in the frames of {\footnotesize{SIBYLL}} \cite{SIBYLL} interaction model (Section III).
The position of the shower maximum, $T_{max}(\boldsymbol{\varepsilon})$ from expression (4b) and profile shape function,
$\delta(x,\boldsymbol{\varepsilon})$ from expression (4a), depend only on the primary nucleon energy $\boldsymbol{\varepsilon}=E/A$,
which is in agreement with the prediction of superposition model \cite{Sommers}.
The amplitude of the profile $N_{max}(E,\boldsymbol{\varepsilon})$ from expression (4c) depends on both the primary energy ($E$)
and nucleon energy ($\boldsymbol{\varepsilon}$).
The intercorrelations between the
$N_{max}(E,\boldsymbol{\varepsilon})$, $T_{max}(\boldsymbol{\varepsilon})$ and $\delta(x,\boldsymbol{\varepsilon})$
shower profile parameters are negligible.
The profile shape function, $\delta(x,\boldsymbol{\varepsilon}$), from (3) has the simple logarithmic
representation (5) which provides an analytic solution for the corresponding inverse profile function, which can
be used in the Constant-Intensity-Cut method \cite{GH}.
The fluctuations of particle shower longitudinal profile, $\sigma_N/N$, from parametrization (10) depend on
the energy ($E$) and mass number ($A$) of the primary nuclei.
\section{Acknowledgments}
I wish to thank James Matthews for useful correspondence.
|
\section{Introduction}
At the Dec 2014 Tokyo meeting, Edo Ibar presented the results of deep ALMA continuum surveys, Nick Scoville discussed ALMA's role in measuring dust in galaxies, and Linda Tacconi was expected to discuss the progress and future role of ALMA in understanding gas in the interstellar medium (ISM) of galaxies, since Linda could not attend, I presented on this area, and
am contributing this part of the proceedings.
The ISM has always been a factor in observational astronomy. Absorption is clear to the naked eye in the Milky Way, and it provides some of the most visually striking features of nearby galaxies. Spectroscopic signatures of diffuse material in emission and absorption have been seen whenever new spectroscopic instruments have been fielded, and in almost every new class of galaxy found. The importance of understanding the effects and details of the ISM has been highlighted in the past year, as subtle features of polarized dust emission have very-publicly confounded early claims of the detection of an unambiguous signal of cosmic inflation.
ALMA is the first telescope with the sheer size to be able to image the ISM quickly in almost any type of high-redshift galaxy. Previous interferometers could generally detect the most luminous high-redshift galaxies, but in observing campaigns that would span several days of integration time, from the earliest (e.g. Solomon et al.\ 1992; Frayer\ et al.\ 1998, 1999; Downes\ et al.\ 1999), to the most recent (e.g. Walter\ et al.\ 2012; Huynh\ et al.\ 2014). Before ALMA, the detections have all been at a relatively modest significance, and while they provided important positional information, redshifts, total luminosities and a view of galaxy-wide ISM properties, they have not typically revealed very detailed astrophysical information about the galaxies involved.
The technology available for receivers and correlators has advanced tremendously, in part to deliver the ambitious requirements of ALMA. The first interferometers sensitive enough to detect distant galaxies' ISM could do so over bandwidths of only a fraction of a GHz, and were thus barely able to span the whole width of a molecular or atomic line from a galaxy's ISM. Advances made leading to ALMA have boosted the performance of existing interferometers, for example enabling the NOEMA project to expand the IRAM PdB interferometer, whose technical specifications exceed those of ALMA at present. Tremendous increases in bandwidth have provided spectrographs that have made existing mm-wave facilities more competitive: for example, the EMIR wideband multi-band spectrograph at the IRAM 30M telescope has provided a great deal of insight into the ISM of nearby galaxies (Saintonge et al. 2011), and helped to expand the line spectral energy distributions (LSED) for a range of galaxies (e.g.\ Wei\ss\ et al.\ 2005; Carilli \& Walter\ 2013).
\section{ALMA and the ISM in distant galaxies}
There are multiple components to the ISM, and it is dynamic, with the production of energy in stars and from AGN affecting the physical state and motions of gas within and around the host galaxy. These energy sources are themselves controlled by the supply and processing of ISM gas. In relatively quiescent systems, and in undisturbed parts of interacting and active galaxies, dense molecular gas should trace the potential of the galaxy/halo. However, feedback from active regions can drive winds, at speeds exceeding 1000\,km/s from AGNs, to escape the galaxy halo. Mechanical energy from radio jets can also impact the ISM, with consequences for fueling AGN and star-formation, while the ongoing rain of gas from the intergalactic medium into the galaxy, and right down onto an AGN might potentially be visible to molecular line observations (e.g. Combes et al.\ 2013, 2014).
Understanding the state of the ISM requires at least several CO lines to be observed to build up an LSED (e.g. Wei\ss\ et al. 2005). Probing high-density tracers (e.g. Imanishi \& Nakanishi 2014; Viti et al.\ 2014) can better resolve the distribution of gas within the galaxy. Furthermore, optical depth effects mean that observations of isotopes of CO might be required to sample the ISM in full. Hence, deep observations with multiple tunings are required, to compile broadband (sub)mm-wave spectra. Atmospheric transmission windows also make it impossible to observe all galaxies over the same range of wavelengths, and so multi-line studies are constrained to specific redshift ranges; see
the recent review by Carilli \& Walter\ (2013).
The most complete LSED of the ISM in distant galaxies comes from a composite spectra obtained from bright gravitationally-lensed galaxies (e.g Spilker et al. 2014). While there are effects to consider about the uniformity of these samples, and how differential magnification affects the SED/LSED (Blain 1999; Serjeant 2012), the picture of a rich range of lines being probed is clear.
The ISM of distant galaxies has also been probed by the {\it Spitzer} and {\it Herschel} space missions, and at longer wavelengths from the ground using SPT, ACT, LMT, the IRAM 30M, JCMT, APEX, CSO and ASTE. There are now wide fields with spectral coverage providing accurate SEDs for distant dusty galaxies, the wavelenght range spanning from starlight to longwards of the peak of the thermal dust emission in the far-infrared(IR). The total luminosity radiated by the ISM, powered by absorbed optical and ultraviolet radiation is now well known across a substantial part of the Universe's history (e.g. Casey et al.\ 2014). Large surveys comparing {\it Spitzer} stellar masses, optical spectroscopic properties, and total luminosities from {\it Herschel} have revealed a perhaps surprising statistical uniformity of galaxies' properties at $z \sim 1$-2 - the "Main Sequence (MS)" (Elbaz et al. 2011; Magdis et al.\ 2012; Saintonge et al.\ 2012; Tacconi et al.\ 2013). The way in which this is established, regulated and maintained is still not known in detail. Determining the key details of the gas fueling star formation within carefully-selected subsamples of MS galaxies with ALMA, in conjunction with resolved images and spectra in the near-IR, should reveal the processes in galaxies that lie on the relation. Furthermore, there is a comparable correlation in the range of metallicities of galaxies - the "Fundamental Metallicity Relation (FMR)" - that applies to bulk galaxy properties, and is inevitably important for understanding the net result of the processes at work in typical galaxies over the bulk of cosmic time (Mannucci et al.\ 2010; Maier et al.\ 2014); ALMA can observe typical distant galaxies on these relations, and the results can be compared with existing large-scale surveys of the cold gas component of nearby galaxies (e.g. Santoinge et al. 2011).
Programs are underway to image galaxies out to large cosmic distances over contiguous areas that are large enough to overwhelm ALMA's mapping speed. The area of sky with excellent imaging quality in depth, bands and resolution, ranges from the CANDELS coverage of exquisite angular resolution over a fraction of a square degree using {\it HST} to the many square degrees of sky that will soon be covered to the maximum depth that can be reached by the 1.5-deg$^2$-field Subaru Telescope's Hyper Suprime-Cam instrument. Euclid and WFIRST plan very wide high-resolution surveys. There are degree-sized deep surveys by ground-based (sub)mm telescopes IRAM 30M, ASTE and JCMT, and 1000-deg$^2$ fields surveyed by {\it Spitzer} and {\it Herschel}. The SPT and ACT ground-based CMB telescopes probe thousands of square degrees of the mm sky, and over the whole sky there are mid-IR and mm-wave surveys from WISE and Planck Surveyor. New wide/deep radio imaging capabilities are arriving from SKA precursors.
\section{The unique capabilities of ALMA}
ALMA's excellent spatial resolution allows high-redshift galaxies to be investigated, even though their ISM continuum emission is too faint to be detected above confusion noise using lower-resolution telescopes. While spectral-line observations reduce confusion, serendipitous discoveries of line-emitting galaxies are rare, and no single-dish telescopes can resolve individual high-redshift galaxies.
ALMA can reveal serendiptious redshifts for galaxies whose location might not be known sufficiently accurately for conventional optical/near-IR spectroscopy. A prototype is the case of the brightest submillimeter galaxy (SMG) found in the Hubble Deep Field (Hughes et al. 1998), whose identification was unclear. An early IRAM PdBI image found an accurate position, but no
convincing counterpart (Downes et al. 1999). A redshift was finally determined using the updated PdBI by Walter et al.\ (2012).
One-to-one matching and submm detection of substantial numbers of galaxies identified in deep multiwaveband surveys, like significant stellar-mass-selected samples from CANDELS, demands ALMA observations, since only a handful of the galaxies within these fields are sufficiently bright to be probed using previous interferometers (e.g. Walter et al. 2014) in a reasonable time. The resolution is limited to match the requirement to obtain significant detections in each synthesized beam. ALMA's spectral coverage in all the submm windows from 3-0.3\,mm, resolving power, and sensitivity puts it in a strong position to resolve typical high-redshift galaxies, even the faintest objects in CANDELS, given enough time.
ALMA could resolve the bulk of the ISM gas in a high-redshift galaxy, whether associated with a cool ordered disk, with turbulent active regions, with infall, or in the process expulsion by winds. It will reveal the location, relative excitation and temperature of different components of the ISM, provide dynamical evidence to distinguish merging subunits from systematic halo-wide rotation, and provide an accurate measurement of the dynamical mass under testable assumptions about whether the gas is virialized, has enhanced motions from turbulence or is supported by rotation. Observations of redshifted CO(1-0) from the Jansky VLA (JVLA) can be combined to provide a consistent high-resolution picture of the coolest molecular gas (e.g. Wagg et al. 2014).
At present, this sort of investigation has been demonstrated for galaxies that have been made brighter and physically larger on the sky by gravitational lensing, most dramatically the "Eyelash" (Swinbank et al. 2010); see also Bothwell et al. (2013) and Spilker et al. (2014). ALMA's power will allow similar studies in galaxies that are not boosted by gravitational lensing, while large samples of bright lensed galaxies can be despatched very quickly using ALMA (Vieira et al. 2013; Messias et al. 2014), or even with smaller interferometers (Bussmann et al. 2013; Rawle et al. 2014).
The physical conditions in even the most heavily obscured regions can be probed using ALMA alone, by measuring molecular and atomic fine-structure line ratios (e.g. Meijerink et al.\ 2011; Tomassetti et al. 2014), and by observing rotational--vibrational and high-rotational-level molecular emission that can only occur in energetic regions (Imanishi \& Nakanishi 2013). By comparing the ALMA images of galaxies with those obtained from ground-based imaging near-IR spectrographs, {\it HST} in the optical, and radio interferometers, it should be possible to obtain an excellent picture of the motion of molecular, atomic and ionized gas in specific powerful galaxies, providing insight into the fuel available for, and both the duration and current efficiency of star formation and AGN fueling. The independent requirements of a nearby adaptive-optics tip-tilt star, lines accessible in ALMA atmospheric windows, and relative brightness means that gathering large samples of these galaxies would be challenging; however, studying individual examples remains important.
In deep images, ALMA can provide very high fidelity maps of the dynamics of gas in distant galaxies, probing molecular gas that is likely to be amongst the coldest component of the ISM, and thus to be the most accurate tracer of the gravitational field. It is possible that very precise measurements could perhaps probe the gravitational lensing shear field on small scales (Blain 2002a), or exploit negative lensing magnification in the cores of dense regions to increase the effective primary beam area for serendiptious surveys in cases where sensitivity is abundant (Blain 2002b).
Any deep ALMA observation of a distant galaxy conducts an effective deep pencil-beam serendipitous spectroscopic survey throughout the primary beam. ALMA's survey performance in going deeper than previous facilities (in both line and continuum observations) has been highlighted by Hatsukade et al. (2013); also see Ibar (this work) and Carniani et al. (2015). ALMA is an almost ideal spectrograph for resolving the ISM in individual high-redshift galaxies and their immediate surroundings, especially those with redshifts that are already known, and in which line-of-sight velocity information is guaranteed. However, ALMA does not survey for potential companions on scales larger than the primary beam, and is thus not perfectly matched to compliment the arcminute-scale patrol radius of ambitious new optical/near-IR imaging spectrographs like MUSE and KMOS. On the other hand, ALMA is well matched to future ELT instruments, with excellent adaptive-optics correction over relatively small fields.
\section{ALMA imaging results for infrared-bright and normal galaxies}
There are now huge numbers of galaxies selected by their ISM dust emission: the SMGs (Smail et al. 1997; Blain et al.\ 2002; Wei\ss\ et al. 2009; Karim et al. 2013), and their cousins selected at shorter wavelengths by Herschel, Spitzer and WISE: "distant dusty galaxies" (DDGs e.g. Casey et al.\ 2014). These galaxies are up to several hundred times more luminous than typical high-redshift galaxies, and are much rarer, by a factor of about 100 for the SMGs, and by up to 10$^4$ in the case of WISE-selected "HotDOGs" (Eisenhardt et al. 2012; Wu et al. 2014; Assef et al.\ 2015; Tsai et al.\ 2015). Nevertheless, these rare (and presumably short-lived) galaxies are bursting with an extreme rate of activity and must be associated with more mundane less-active hosts, that themselves probably lie on the MS and FMR relations. ALMA can discriminate the ordinary and remarkable parts of these galaxies, and provide more complete size, dynamical, and excitation information than accessible at other wavelengths due to substantial extinction. ALMA's ability to resolve the structures, dynamics and physical conditions in the ISM should soon allow us to understand SMGs/DDGs in detail.
Imaging high-redshift SMGs with ALMA (e.g. Karim et al. 2013) has revealed that an interesting number break up into multiple components. Perhaps this shows a limit to the surface brightness of galaxies, with all very luminous objects breaking up into interacting subunits, or perhaps many are even accidental superpositions.
The most impressive haul of ALMA spectra so far has been for galaxies identified as gravitational lenses from wide-area multicolor mm-wave continuum surveys primarily aimed at measuring the abundance of Sunyaev--Zeldovich effect clusters (e.g. Hezaveh et al. 2013; Vieira et al. 2013; Wei\ss et al. 2013; Spilker et al.\ 2014); comparable samples are available from the Planck, Herschel and ACT telescopes.
The first distant galaxies imaged by ALMA can be divided into several catagories: bright lensed objects (see above); SMGs and dusty AGNs at various resolutions (e.g. Swinbank et al.\ 2012; Nagao et al.\ 2012; Wagg et al.\ 2012; Carniani et al.\ 2013, Carilli et al.\ 2013; Hodge et al.\ 2013; Huynh et al.\ 2013; Karim et al.\ 2013; Wang et al.\ 2013b; De Breuck et al.\ 2014; Simpson et al.\ 2014, 2015; Thomson et al.\ 2014); more typical star-forming galaxies (e.g. Decarli et al.\ 2014; Ono et al.\ 2014; Ota et al.\ 2014; Riechers et al.\ 2014); high-redshift QSOs (e.g. Wang et al.\ 2013a; Gilli et al.\ 2014); and more unusual objects: gamma-ray burst (GRB) host galaxies (Berger et al.\ 2014; Hatsukade et al.\ 2014; Michalowski et al.\ 2014), Lyman-$\alpha$ emitters (Ouchi et al.\ 2012; Williams et al.\ 2014), and low-metallicity dwarves (Hunt et al.\ 2014).
\section{High-speed molecular outflows}
Molecular material ejected by powerful low-redshift galaxies has been detected from the ground (Iono et al.\ 2007; Nesvadba et al.\ 2008; Feruglio et al.\ 2010; Alatalo et al.\ 2011; Aalto et al.\ 2012; Cicone et al.\ 2014), and space (Fischer et al.\ 2010; Sturm et al.\ 2011), in the broad wings of molecular line emission and in P-Cygni profiles of OH lines respectively.
Powerful AGN at high redshifts also show broad molecular line wings in sensitive, broad-band mm-wave spectral images of redshifted relatively-bright fine-structure lines (Cicone et al.\ 2015).
The detection of these broad lines requires substantial S/N detections of the narrower, bound ISM emission in order to be reach the necessary sensitivity. While providing otherwise unavailable insight into the mechanical energy output by luminous galaxies, they can thus also detail the less-dramatic but comparably interesting dynamics of gas in the host galaxy and any nearby companions. The first such ALMA results are appearing (Garcia-Burillo et al. 2014; Sun et al. 2014).
\section{Large-scale structure and probing galaxy mass scales}
There are various indications that, on a variety of spatial scales, luminous dusty galaxies are strongly clustered, whether from spectroscopy of counterparts (Blain et al. 2004), from correlation functions using {\it Herschel} surveys (Viero et al. 2013), from the overdensity of ultraluminous SMG companions in the fields of WISE-selected galaxies from ground-based images (Jones et al. 2014, 2015), or within the ALMA primary beam: both in follow-up images of SMG surveys (e.g. Karim et al. 2013), or around WISE--radio-selected AGNs (Lonsdale et al.\ 2015; Silva \& Sajina 2015; this volume).
The modest field of view of ALMA is always likely to make it more powerful for the study of individual galaxies than for conducting large surveys. However, because of its sensitivity and spatial resolution, and the absence of confusion noise, discovering serendipitous galaxies in ALMA fields will be a regular occurence (Tamura et al.\ 2014).
ALMA is not ideally suited to studying galaxy clustering on scales larger than the primary beam; however, it does offer the opportunity to test whether the detailed astrophysics of galaxies depend on their environments, which is likely to be important for understanding the development of the MS and FMR relations, and potentially to reveal the astrophysics responsible for the downsizing process. It can also provide high-resolution extinction-free images for modelling gravitational lenses, and reveal structure otherwise too small to probe (e.g. Vieira et al.\ 2013).
\section{Future capabilities?}
Owing to the modest 10-km/s resolution necessary to study galaxies, there are few demands from this science area on the spectral resoution of the ALMA correlator. However, outflows that are several 1000 km/s in width do challenge the available bandwidth coverage at the very highest frequencies. There are closely separated lines from different species that could be imaged simultaneously if the bandwidth could be increased to span a larger fraction of an atmospheric window, which would also increase the chances of success in serendipitous line (and continuum) surveys in parallel to deep observations.
Absorption line studies against powerful continuum point sources would also potentially benefit from an increase in bandwidth (e.g.\ Wiklind \& Combes 1998). Measurements of very precise frequency offsets in absorbers can be used to limit changes in fundamental constants, and species that are far too rare to be detected in emission can be measured in absorption (Muller et al.\ 2007, 2014; Bagdonaite et al. 2012). The precision of ALMA spectroscopy and frequency accuracy also potentially allows some such tests using emission lines (e.g. Lentati et al. 2013).
\section{Summary}
ALMA's first three years of operations, during which the time available on the array and its capabilities have been much less than will be ultimately delivered
has shown its remarkable capabilities.
ALMA will provide a vital resource for studying galaxies in the years ahead, providing
data of exceptional and quite unprecedented quality.
\acknowledgements I thank Linda Tacconi, Caitlin Casey, Ian Smail, Mark Swinbank and Fabian Walter for providing material for this presentation, and the organizers for such a wide-ranging program of highlights in such an excellent venue. AWB is supported by a Royal Society Wolfson Research Merit award.
|
\section{Introduction}
\label{sec:intro}
Cellular Automata (CAs) are often used for constructing models in a variety of fields of application, including chemistry, biology, medicine, physics, ecology and the study of socioeconomic interactions. In many of these settings, Stochastic CAs (SCAs) are considered due to the stochastic nature of the phenomenon under study or due to a lack of understanding of the exact rules driving the phenomenon \cite{Ganguly03asurvey,das2012survey}. Therefore, a better understanding of the dynamics of SCAs is crucial. Only a few methods for dealing with the analysis of models based on SCAs have been developed. In many practical applications, especially in cases where the averaged behavior of the system is of concern, sampling methods, relying on extensive computer simulations, are sufficient \cite{aids,2136045,4167259}. Techniques built on the mean--field theory can be used to study the long-term behavior of SCAs \cite{RSA:RSA20126}. The theory of Markov chains can be applied to provide analytical tools for analyzing the model's behavior~\cite{Mairesse201442}, although in practice, due to the theoretical and computational complexity of such tools, the application scope is limited.
This paper is devoted to providing effective analytical tools based on deterministic CAs for the analysis of multi-state SCAs. Although the theoretical foundations of the presented results are already (at least partially) known in the literature \cite{Mairesse201442,lebowitz,computing-pca}, so far the applications are limited. Therefore, the main aim of this paper is to provide a complete, formal description of the discussed properties in a form that is suitable for applications and that does not require a strong mathematical background, as well as to present examples that can motivate further applications in the domain of systems modeling. The methods presented here are developed in the context of 1D SCAs on finite lattices, but can be easily generalized to the case of higher dimensions and infinite spaces.
Two main results are presented in this paper. The first one involves constructing images that show the cell-wise probability distribution over the state set, at any time step. The method is based on associating an SCA with a deterministic, Continuous CA (CCA). The second result shows the equivalence of SCAs and stochastic mixtures of deterministic CAs. Based on this finding, any SCA can be decomposed into a finite set of deterministic CAs, each of them contributing to the behavior of the stochastic system. An effective method for finding a decomposition is presented. It allows to uncover the deterministic component in the mixture with the highest impact on the behavior of the SCA.
This paper is organized as follows. We start with some preliminaries and definitions in Section~\ref{sec:preliminary}. In Section~\ref{sec:cca}, we introduce the concept of CCAs and the formalism enabling the analysis of multi-state CAs. Section \ref{sec:multi-sca} contains the definition of multi-state SCAs and holds the main results of this paper. The paper is concluded with Section \ref{sec:experiments}, discussing the experimental results that illustrate our results. A summary is presented in Section~\ref{sec:summary}.
\section{Preliminaries}
\label{sec:preliminary}
Informally, a CA is a discrete dynamical system in which the space is subdivided into discrete elements, referred to as cells. At every consecutive, discrete time step, each cell is assigned one of $N$ states using a deterministic rule, which depends only on the previous state of the considered cell and the states of its neighboring cells \cite{neumann1948}.
Formally speaking, let the state set $\mathcal{S}$ be a finite set of $N>1$ elements. Elements of the set $\mathcal{C} = \{ c_i \mid i=1,\dotsc,M \}$ will be referred to as cells. Every cell $c_i$ is assigned a state $s(c_i,t)\in \mathcal{S}$ at each time step $t\in\mathbb{N}$, according to a local, deterministic rule. The vector $s(\cdot, t) \in \mathcal{S}^M$ will be referred to as the configuration, and as the initial configuration when $t=0$. The sequence $(s(\cdot,0),s(\cdot,1),\ldots)$ will be referred to as the space-time diagram of the CA. For technical reasons, we impose periodic boundary conditions, but our results do not depend on this assumption.
The function $A\colon \mathcal{S}^M \to \mathcal{S}^M$ satisfying, for every $t\in\mathbb{N}$:
\begin{equation}
s(\cdot, t+1) = A(s(\cdot, t))\,,
\end{equation}
will be referred to as a global CA rule or simply a CA, if there exists a radius $r\in\mathbb{N}$ and a function $f\colon\mathcal{S}^{2\,r+1}\to\mathcal{S}$ satisfying:
\begin{equation}
s(c_i, t+1) = f(s(c_{i-r}, t), \dotsc, s(c_{i+r},t))\,,
\label{eq:local-fun}
\end{equation}
for every $i$ and $t\in\mathbb{N}$. Such a function $f$ will be referred to as a local rule. Note that a local rule uniquely defines the global rule, while for a given global rule, multiple local rules exist. Additionally, it is assumed that $r\ll M$. The vector $(c_{i-r}, c_{i-r+1},\dotsc,c_{i+r-1},c_{i+r})$ will be referred to as the neighborhood of cell $c_i$ and $R=2\,r+1$ will denote the neighborhood size. For the sake of simplicity, $s(c_{i-r},\dotsc,c_{i+r}, t)\in\mathcal{S}^R$ will denote the state of the neighborhood of cell $c_i$ at time step $t$, and will be referred to as the neighborhood configuration of $c_i$ at time step $t$.
\section{Continuous CAs}
\label{sec:cca}
There exist multiple ways of extending the definition of CAs to cover infinite state sets. Examples of such approaches include Coupled Map Lattices (CMLs)~\cite{kaneko93} and so-called fuzzy CAs \cite{flocchini2000convergence,Betel:2009:ABF:1621142.1621283,Betel20095}. In this section, we present Continuous CAs (CCAs), which can be seen as a generalization of the ideas presented in \cite{flocchini2000convergence}. Our formalism is based on a polynomial representation of discrete CA rules. We start with formulating the continuous counterparts of binary CAs. After that we present a generalization to cover multi-state CAs.
\subsection{Binary CAs}
\label{sec:binary-cca}
Binary CAs are widely studied \cite{RevModPhys.55.601,wolfram-class}, because they allow to evolve complex patterns and exhibit complex behavior despite their intrinsic simplicity. The state set of such a CA $A$ is $\mathcal{S} = \{0,1\}$. We will now formally define and characterize its local rule $f\colon \mathcal{S}^R\to\mathcal{S}$. Let $l = (l_i)_{i=1}^{2^R}$ be a binary vector. We consider a system of equations:
\begin{equation}
f(s_{i,1},\dotsc,s_{i,R}) = l_i\,,
\end{equation}
where $(s_{i,1},\dotsc,s_{i,R})$ is a binary vector such that $i = 1 + \sum_{j=0}^{R-1} s_{i,R-j}\,2^{j}$. As can be seen, such a system of equations is uniquely defined by the vector $l$. The vector $l$ will be referred to as the lookup table (LUT) of the local rule $f$. It is not difficult to check that such a system uniquely defines the function $f$, since it lists all of the possible input configurations, and maps them to corresponding values by components of the vector $l$.
Following \cite{flocchini2000convergence}, we know that the function $f$ can be expressed as a polynomial, which is of interest for our purposes. In order to define it, we introduce two auxiliary functions. We start with the function $\ind\colon \{1,\dotsc,2^R\}\to \{1,2\}^R$. It is defined in such a way that $\ind(i)[m]$ is the $m$--th digit, incremented by one, read from left to right, of the binary representation of the integer $i-1$, padded with ones on the left, so that it always has length $R$. Consequently, it holds that:
\begin{equation}
i = 1+\sum_{m=1}^{R} (\ind(i)[R-(m-1)]-1)\,2^{m-1}.
\label{eq:bin-def}
\end{equation}
The values of $\ind(i)$ for $R=3$ and $i\in\{1,\dotsc,8\}$ are shown below:
\begin{gather*}
\ind(1) = (1,1,1),\ \ind(2) = (1,1,2),\ \ind(3) = (1,2,1),\ \ind(4) = (1,2,2),\\
\ind(5) = (2,1,1),\ \ind(6) = (2,1,2),\ \ind(7) = (2,2,1),\ \ind(8) = (2,2,2).
\end{gather*}
The function $\ind$ is related to the binary representation of integers. In \cite{flocchini2000convergence} a simpler formulation using the function $\bin$, which yields the binary representation of an integer, is used. The construction presented here, although a bit more complicated in the binary setting, allows for a smoother generalization to the multi-state case.
Using the function $\ind$, we now define the function $n\colon \mathcal{S} \times \mathbb{N} \times \mathbb{N} \to \mathcal{S}$, which for $s\in\mathcal{S}$ and $m,i\in\mathbb{N}$, is given by:
\[
n(s, m, i) =
\begin{cases}
s &, \textrm{if}\ \ind(i)[m] = 2,\\
1- s &, \textrm{if}\ \ind(i)[m] = 1.
\end{cases}
\]
Note that we will use vectors of states of the form $(s_1,\dotsc,s_R)\in\mathcal{S}^R$ and for simplicity, for any $m\in\{1,\dotsc,R\}$, we will write $n(s_m,i)$ instead of $n(s_m, m, i)$. Using the functions $\ind$ and $n$, we can write the polynomial representation of the local rule $f$ as:
\begin{equation}
f(s_1,\ldots,s_{R}) = \sum_{i=1}^{2^R} l_i \, \left(\prod_{m=1}^{R} n(s_m,i) \right)\,.
\label{eq:fca-gen}
\end{equation}
The following example shows the explicit form of this polynomial for a member of the family of Elementary CAs (ECAs).
\begin{example}[Elementary CAs]
\label{exa:one}
Binary, 1D CAs with neighborhood radius $r=1$ are commonly referred to as ECAs~\cite{RevModPhys.55.601}. There are 256 such ECAs. Treating the LUT entries $l_i$ as digits of an integer written in base 2, we can enumerate the local rules of ECAs. By convention, the binary vectors are read in the reverse order, {\it i.e.}\ $(l_i)_{i=1}^8$ is interpreted as $(l_8, l_7, \dotsc, l_1)_2$. For example, given the LUT $l = (0,1,1,0,1,0,0,1)$ of ECA 150, and denoting the Boolean complement as $\bar{s} = 1-s$, its local rule can be written, according to Eq.~$(\ref{eq:fca-gen})$, as:
\[
f_{150}(s_1,s_2,s_3) = \bar{s}_1\,\bar{s}_2\,s_3
+ \bar{s}_1\,s_2\,\bar{s}_3
+ s_1\,\bar{s}_2\,\bar{s}_3
+ s_1\,s_2\,s_3\,.\tag*{\qed}\]
\end{example}
Using the above notation, a CCA can be defined analogously to a binary CA, with two notable differences. Firstly, the state set of a CCA is the unit interval, {\it i.e.}\ $\mathcal{S}=[0,1]$, and, secondly, the local rule $f\colon [0,1]^R\to[0,1]$ is given by Eq.~(\ref{eq:fca-gen}) with coefficients $l_i\in [0,1]$. We will refer to such a vector $(l_i)_{i=1}^{2^R}$ as a generalized~LUT. It is easy to check that this definition of a CCA is formally correct. Indeed, the values of the function $f$ in Eq.\ (\ref{eq:fca-gen}) are guaranteed to belong to the unit interval if $l_i\in[0,1]$ for all $i$ and $s_m\in[0,1]$ for all $m~\in~\{1,\dotsc,R\}$. Note that this construction is directly related to the one presented in~\cite{flocchini2000convergence}, where fuzzy CAs are constructed as polynomials representing fuzzified logical functions. Following the same line of reasoning, an alternative polynomial representation for the local rules of binary CAs is presented in \cite{schule2008global}, which is consistent with a logical representation of the local rules.
In order to introduce the formalism that is needed in the multi-state setting, we present a slightly modified way of representing binary CAs compared to the one obtained through Eq.~$(\ref{eq:fca-gen})$. Let us assume that the state set is given by $\mathcal{S}_2 = \{ (1,0), (0,1) \}\subset \mathbb{R}^2$. Then the local rule is a function $f\colon \mathcal{S}_2^R\to \mathcal{S}_2$ and can be represented as a vector function $f = (f_1, f_2)$, where $f_j\colon \mathcal{S}_2^R\to \{0,1\}$. Note that any $s = (s_1, s_2)\in \mathcal{S}_2$ satisfies $s_2 = 1-s_1$. Similarly, $f_2 = 1 - f_1$, so that the local rule $f$ can be defined using $f_2$ only. \hl{Let $l\in\mathcal{S}_2^{2^R}$ be given, {\it i.e.} for every $i=1,\dotsc,2^R$, it holds that $l_i = (l_{i,1}, l_{i,2})\in \mathcal{S}_2$. Let $(s_1,\dotsc,s_R) \in \mathcal{S}_2^R$.} We can define $f_2$ with a formula similar to Eq.\ (\ref{eq:fca-gen}) as:
\begin{equation}
f_2(s_1,\dotsc,s_{R}) = \sum_{i=1}^{2^R} l_{i,2} \, \left(\prod_{m=1}^{R} n(s_{m,2},i) \right).
\end{equation}
Since $s_{m,1} = 1 - s_{m,2}$, we can rewrite the formula for the function $n$ as:
\[
n(s_{m,2}, i) =
\begin{cases}
s_{m,2} &, \textrm{if}\ \ind(i)[m] = 2,\\
s_{m,1} &, \textrm{if}\ \ind(i)[m] = 1,
\end{cases}
\]
which can be simplified to:
\begin{equation}
n(s_{m,2}, i) = s_{m,\ind(i)[m]}\,,
\end{equation}
and thus $f_2$ can be written as:
\begin{equation}
f_2(s_1,\ldots,s_{R}) = \sum_{i=1}^{2^R} l_{i,2} \, \left(\prod_{m=1}^{R} s_{m,\ind(i)[m]} \right).
\end{equation}
Having defined the function $f_2$, we can prove the following fact, which gives a direct formula for $f_1$.
\begin{lemma}
The function $f_1$ can be expressed as:
\begin{equation}
f_1(s_1,\ldots,s_{R}) = \sum_{i=1}^{2^R} l_{i,1} \, \left(\prod_{m=1}^{R} s_{m,\ind(i)[m]} \right).
\end{equation}
\end{lemma}
\noindent Due to the above, we can write a general expression for $f_j$, $j=1,2$, as:
\begin{equation}
f_j(s_1,\ldots,s_{R}) = \sum_{i=1}^{2^R} l_{i,j} \, \left(\prod_{m=1}^{R} s_{m,\ind(i)[m]} \right).
\label{eq:final-cca-2}
\end{equation}
Resorting to this vector notation, a CCA can be defined with a state set:
\begin{equation}
S_c = \{(x_1, x_2) \in [0,1]^2 \mid x_1+x_2 = 1\}\subset \mathbb{R}^2\,,
\label{eq:sc-2}\end{equation}
and a local rule $f\colon S_c^R \to S_c$ given by Eq.\ (\ref{eq:final-cca-2}) with $(l_j)_{j=1}^{2^R} \in S_c^{2^R}$.
\begin{example}
The local rule $f_{150}$ of ECA 150 presented in Example \ref{exa:one} can be rewritten in the context of the state set $\mathcal{S}_2$, as $f\colon \mathcal{S}^3_2 \to \mathcal{S}_2$, $f=(f_1, f_2)$. Note that for $s_i\in\mathcal{S}_2$, it holds that $s_i=(s_{i,1}, s_{i,2})$. The local rule $f$ can be written as:
\begin{align*}
f_1(s_1, s_2, s_3) & = s_{1,2}\,s_{2,2}\,s_{3,1} + s_{1,2}\,s_{2,1}\,s_{3,2} + s_{1,1}\,s_{2,2}\,s_{3,2} + s_{1,1}\,s_{2,1}\,s_{3,1} \\
f_2(s_1, s_2, s_3) & = s_{1,2}\,s_{2,2}\,s_{3,2} + s_{1,2}\,s_{2,1}\,s_{3,1} + s_{1,1}\,s_{2,2}\,s_{3,1} + s_{1,1}\,s_{2,1}\,s_{3,2}\,.\\[-\normalbaselineskip]\tag*{\qed}
\end{align*}
\end{example}
The representation given by Eq.~(\ref{eq:final-cca-2}) is equivalent to that given by Eq.~(\ref{eq:fca-gen}), but allows for an easier generalization to multi-state CAs. Therefore, we will use it throughout the remainder of this paper.
\subsection{Multi-state CAs}
\label{sec:multi-ca}
We start with generalizing the state set $\mathcal{S}_2$ defined in Section \ref{sec:binary-cca} to the case of $N$ states. Let $\mathcal{S}_N\subset \mathbb{R}^N$ be the set of all base vectors of the $N$--dimensional Euclidean space, {\it i.e.}
\begin{equation}
\mathcal{S}_N =\Big\{ e \in \{0,1\}^N \mid \norm{e} = 1\Big\}\,.
\label{eq:state-set}
\end{equation}
Let the $i$--th state in a multi-state CA be represented by the base vector $e_i$ which has zeros everywhere, except at position $i$, where it has a one. For instance, if $N=3$, then $\mathcal{S}_3 = \{ (1,0,0), (0,1,0), (0,0,1)\}$. For the sake of simplicity, in the remainder of this paper, we will write $\mathcal{S}$ instead of $\mathcal{S}_N$, assuming that $N$ is fixed.
We now generalize the function $\ind \colon \{1,\dotsc,N^R\} \to \{1,\dotsc,N\}^R$, defined above for the binary case, to enumerate the $N^R$ neighborhood configurations. Let $i\in\{1,\dotsc,N^R\}$ be the index identifying a neighborhood configuration. We assume that such a configuration consists of the following states:
\begin{equation}
(e_{\ind(i)[1]}, e_{\ind(i)[2]},\dotsc,e_{\ind(i)[R]})\,.
\end{equation}
We assume that $\ind$ satisfies the following equation:
\begin{equation}
i = 1 + \sum_{m=1}^{R}\Big(\ind(i)[R-(m-1)]-1\Big)\,N^{m-1}\,.
\end{equation}
A direct formula for calculating $\ind$ is given by:
\begin{equation}
\ind(i)[m] = \left\lfloor\frac{(i-1)\ \textrm{mod}\ N^{R-(m-1)}}{N^{R-m}}\right\rfloor + 1\,.
\end{equation}
Essentially, $\ind(i)$ yields the positional representation of $i-1$ in base $N$, padded with zeros on the left, so that it always has length $R$, where each of the digits is incremented by one. Hence, one easily finds:
\begin{align*}
\ind(1) &= (1,\dotsc,1,1)\,, \\
\ind(2) &= (1,\dotsc,1,2)\,, \\
\vdots \\
\ind(N) &= (1,\dotsc,1,N)\,, \\
\ind(N+1) &= (1,\dotsc,2,1)\,, \\
\vdots\\
\ind(N^R) &= (N,\dotsc,N,N)\,.
\end{align*}
Any local rule $f$ of an $N$--state, deterministic CA can be uniquely defined by writing down its outputs for all of the possible neighborhood configurations. Formally, let us consider the system of equations:
\begin{equation}
f(e_{\ind(i)[1]},\dotsc,e_{\ind(i)[R]}) = l_{i},
\end{equation}
where $i\in\{1,\dotsc,N^R\}$ and $l_i \in \mathcal{S}$. Similarly to the binary case, the matrix $L= (l_i)_{i=1}^{N^R}\in\mathcal{S}^{N^R}$ will be called the LUT of a multi-state CA. Basically, $L$ is a matrix of $N^{R}$ columns and $N$ rows, containing only zeros and ones, in such a way that every column contains exactly one non-zero entry. Each of the matrices $L\in\mathcal{S}^{N^{R}}$ uniquely defines an $N$--state CA in terms of its local rule.
In order to define a CCA in the context of multiple states, we need to formally define its state set. Let $\mathcal{S}_c$ be defined as:
\begin{equation}
\mathcal{S}_c = \left\{ (x_1, \dotsc, x_N) \in [0,1]^N \mid \sum_{i=1}^{N} x_i = 1\right\}\subset \mathbb{R}^N\,,
\end{equation}
which is consistent with the definition for $N=2$ given by Eq.~$(\ref{eq:sc-2})$. Note that any $s\in\mathcal{S}_c$ satisfies $s_i = 1 - \sum_{j\neq i} s_j$. The set $\mathcal{S}_c$ is commonly referred to as a standard $(N-1)$--simplex \cite{rudin-principles}.
\begin{example}
Let $N=3$. The set $\mathcal{S}_c$ is a 2D triangle placed in the 3D Euclidean space, with vertices $(1,0,0)$, $(0,1,0)$ and $(0,0,1)$. \qed
\end{example}
In the case of a CCA, the local rule is a vector function $f\colon \mathcal{S}_c^{R}\to\mathcal{S}_c$, given by $f=(f_1,\dotsc,f_N)$, for which there exists a matrix $P\in\mathcal{S}_c^{N^R}$, referred to as the generalized LUT, such that for $j\in\{1,\dotsc,N\}$ it holds that $f_j\colon\mathcal{S}_c^R\to[0,1]$ and:
\begin{equation}
f_j(s_1,\dotsc,s_R) = \sum_{i=1}^{N^R} P_{ij}\left(\prod_{m=1}^{R}s_{m,\ind(i)[m]}\right)\,.
\label{eq:def-f}
\end{equation}
We will show that the definition given by Eq.\ $(\ref{eq:def-f})$ is formally correct, meaning that the function $f$ with components $f_j$ satisfies $f\colon \mathcal{S}_c^{R}\to\mathcal{S}_c$, which is equivalent to showing that $f_j(s_1,\dotsc,s_R) \in [0,1]$ and $\sum_{j=1}^N f_j(s_1,\dotsc,s_R) = 1$ for any $(s_1,\dotsc,s_R)\in \mathcal{S}_c^R$.
\begin{lemma}
For any $(s_1,\dotsc,s_R)\in\mathcal{S}_c^R$, it holds that
\begin{equation}
\sum_{i=1}^{N^R}\left[\prod_{m=1}^R s_{m,\ind(i)[m]}\right] = 1\,.
\label{eq:lem1}
\end{equation}
\end{lemma}
\begin{proof}
Let
\[\Theta = \sum_{i=1}^{N^R}\left[\prod_{m=1}^R s_{m,\ind(i)[m]}\right].\]
Note that for every $j\in\{1,\dotsc,N\}^R$, there exists exactly one $i\in\{1,\dotsc,N^R\}$ such that $j = \ind(i)$. Due to this, it holds that:
\[ \Theta = \sum_{j\in\{1,\dotsc,N\}^R}\left[\prod_{m=1}^{R} s_{m,j_m}\right].\]
By regrouping the elements of the above sum, we may write:
\[ \Theta = \sum_{j\in\{1,\dotsc,N\}^{R-1}}\left[\prod_{m=1}^{R-1} s_{m,j_m}\left(\sum_{i=1}^{N} s_{R,i}\right)\right].\]
Since $s_R \in \mathcal{S}_c$, we know that $\sum_{i=1}^{N} s_{R,i} = 1$, and thus:
\[ \Theta = \sum_{j\in\{1,\dotsc,N\}^{R-1}}\left[\prod_{m=1}^{R-1} s_{m,j_m}\right].\]
By repeating this regrouping procedure $R-1$ times, we finally get:
\[ \Theta = \sum_{j=1}^N s_{1,j} = 1\,.\qedhere\]
\end{proof}
\begin{proposition}
For any $j\in\{1,\dotsc,N\}$ and any $(s_1,\dotsc,s_R)\in\mathcal{S}_c^R$, it holds that
$f_j(s_1,\dotsc,s_R) \in [0,1]$.
\end{proposition}
\begin{proof}
Firstly, $f_j(s_1,\dotsc,s_R)\geq 0$ is fulfilled since all of the elements in the sum in Eq.\ (\ref{eq:def-f}) are non-negative. Therefore, showing that $f_j(s_1,\dotsc,s_R)\leq 1$ is sufficient to prove the proposition. Since $P_{ij}\leq 1$, we can write:
\[ f_j(s_1,\dotsc,s_R) = \sum_{i=1}^{N^R}P_{ij}\left(\prod_{m=1}^R s_{m,\ind(i)[m]}\right)\leq
\sum_{i=1}^{N^R}\prod_{m=1}^R s_{m,\ind(i)[m]} = 1\,. \]
\end{proof}
\begin{proposition}
For any $(s_1,\dotsc,s_R)\in\mathcal{S}_c^R$, it holds that:
\[\sum_{j=1}^N f_j(s_1,\dotsc,s_R) = 1\,.\]
\end{proposition}
\begin{proof}
Let $j\in\{1,\dotsc,N\}$, then it suffices to show that:
\begin{equation}
1 - f_j(s_1,\dotsc,s_R) = \sum_{k\neq j} f_k(s_1,\dotsc,s_R)\,.
\end{equation}
This identity is proven by the following chain of equalities:
\begin{align*}
1 - f_j(s_1,\dotsc,s_R) &= 1 - \sum_{i=1}^{N^R}P_{ij}\left(\prod_{m=1}^R s_{m,\ind(i)[m]}\right) \\
&=1 - \sum_{i=1}^{N^R}\left(1-\sum_{k\neq j}P_{ik}\right)\prod_{m=1}^R s_{m,\ind(i)[m]} \\
&=1 - \sum_{i=1}^{N^R}\prod_{m=1}^R s_{m,\ind(i)[m]} + \sum_{i=1}^{N^R}\sum_{k\neq j}P_{ik}\left(\prod_{m=1}^R s_{m,\ind(i)[m]}\right) \\
&= \sum_{k\neq j}\sum_{i=1}^{N^R}P_{ik}\left(\prod_{m=1}^R s_{m,\ind(i)[m]}\right)
= \sum_{k\neq j} f_k(s_1,\dotsc,s_R)\,.\qedhere
\end{align*}
\end{proof}
We have shown the correctness of the CCA definition. We conclude this section with the observation that the local rule $f\colon \mathcal{S}^R\to\mathcal{S}$ of any $N$--state, deterministic, 1D CA can always be written in the form of Eq.~$(\ref{eq:def-f})$, where $(s_1,\dotsc,s_R)\in\mathcal{S}^R$, $f_j\colon \mathcal{S}^R\to\{0,1\}$ and $\sum_{j=1}^{N} f_j(s_1,\dotsc,s_R) = 1$. This shows that CCAs are a generalization of $N$--state, deterministic CAs.
\section{Properties of SCAs}
\label{sec:multi-sca}
In this section, we provide a general definition of an $N$--state SCA, which is formulated using the notation introduced in Section \ref{sec:cca} for multi-state CAs and CCAs. Subsequently, we show the nature of the relation between SCAs and CCAs. Finally, we present a method for decomposing an SCA into a set of deterministic CAs.
\subsection{General construction of multi-state SCAs}
We define SCAs, the stochastic counterpart of multi-state CAs. We assume that the state set $\mathcal{S}$ is given by Eq.\ (\ref{eq:state-set}) and contains all the standard base vectors $e_i$. In such a setting, the states are assigned according to a probability distribution, which depends on the neighborhood configuration at the previous time step. Therefore, SCAs are not defined through a local rule, but rather by a set of conditional expressions, \hl{which for $j \in\{1,\dotsc,N\}$ and $k \in \{ 1,\dotsc,N^R\}$ can be written as}:
\begin{equation}
\mathbb{P}\Big(s(c_i, t+1) = e_j \mid s(c_{i-r},\dotsc,c_{i+r},t) = (e_{\ind(k)[1]},\dotsc,e_{\ind(k)[R]})\Big) = p_{kj}\,.
\label{eq:pkj}
\end{equation}
\hl{The conditional probabilities $p_{kj}$ in Eq.~(\ref{eq:pkj}) do not depend on the choice of $i$ and $t$. Obviously, it holds that $p_k = (p_{k1}, \dotsc, p_{kN}) \in \mathcal{S}_c$ for all $k$. Furthermore, the matrix $P = (p_{kj})$ can be considered an element of $\mathcal{S}_c^{N^R}$. Consequently, every element in $\mathcal{S}_c^{N^{R}}$}
uniquely defines the probability distribution of some local rule. The matrix $P$ will be referred to as the probabilistic lookup table (pLUT). Such matrices are commonly known as stochastic matrices \cite{stoma} and are typically used in the analysis of Markov chains.
Any SCA is uniquely defined by a pLUT belonging to $\mathcal{S}_c^{N^{R}}$. Since we know from Section \ref{sec:multi-ca} that every element of $\mathcal{S}_c^{N^{R}}$ uniquely defines a CCA, we elaborate on the relationship between an SCA defined by a pLUT $P\in\mathcal{S}_c^{N^R}$ and a CCA defined by the same $P$ considered as generalized LUT.
\subsection{Relationship between CCAs and SCAs}
In contrast to deterministic CAs, only the initial state $s(\cdot, 0)$ is known with certainty in the case of SCAs, while for $s(\cdot, t)$, $t>0$, only a probability distribution can be calculated. Let $\pi(c_i, t\mid I_0)$ denote the probability distribution over the set of states $\mathcal{S}$ for the $i$--th cell at time step $t$, given that the evolution started from the initial condition $I_0$, {\it i.e.}\ $s(\cdot, 0) = I_0$. Whenever it will not bring confusion, we will omit $I_0$ and write $\pi(c_i, t)$. Since there are $N$ states, we need to specify $N$ probabilities summing up to 1, to define $\pi(c_i, t)$. For this reason, let us represent $\pi(c_i, t)$ as a vector of probabilities, {\it i.e.}\ $\pi(c_i, t)\in\mathcal{S}_c$, such that:
\begin{equation}
\pi(c_i, t \mid I_0)[k]= \mathbb{P}\Big(s(c_i, t) = e_j \mid s(\cdot, 0) = I_0\Big),
\end{equation}
where $j=1,\dotsc,N$.
When using SCAs in practical problem settings, one might be interested in computing the values of $\pi(c_i, t)$ for any $t$. Normally, this is achieved by means of sampling methods \cite{aids,2136045,4167259}, {\it i.e.}\ multiple space-time diagrams are generated from the same initial condition and probability distributions are estimated, based on the frequencies of reaching any of the states in a given cell. Unfortunately, such an approach has serious drawbacks, especially in terms of the computational burden it brings along. Yet, one can calculate $\pi(c_i, t)$ much faster by relying on CCAs.
\begin{proposition}
Let $P\in\mathcal{S}_c^{N^R}$ be the pLUT of an SCA, and let $f\colon\mathcal{S}_c^{R} \to \mathcal{S}_c$ be a function defined according to Eq.\ $(\ref{eq:def-f})$ using $P$ as generalized LUT, then:
\begin{equation}
\pi(c_i, t+1) = f(\pi(c_{i-r}, t), \dotsc, \pi(c_{i+r}, t))\,,
\end{equation}
for any $i$ and $t$.
\label{prop:cca-mstate}
\end{proposition}
\begin{proof}
Let $k\in\{1,\dotsc,N^R\}$ and $j\in\{1,\dotsc,N\}$, then $p_{kj}$ is the probability of reaching state $e_j$ if the neighborhood configuration was known to be $(e_{\ind(k)[1]},\dotsc,e_{\ind(k)[R]})$, which will be referred to as the $k$--th neighborhood. Assuming that $\pi(\cdot, t)$ is known, the probability of cells $(c_{i-r},\dotsc,c_{i+r})$ being in the $k$--th neighborhood can be calculated with the following formula:
\begin{multline}
\mathbb{P}\Big(s(c_{i-r},\dotsc,c_{i+r}, t)=(e_{\ind(k)[1]},\dotsc,e_{\ind(k)[R]})\Big)\\ = \prod_{m=1}^{R} \pi(c_{i-r+m-1}, t)\big[\ind(k)[m]\big].
\label{eq:nei-prob}
\end{multline}
Due to the Total Probability Theorem, the probability of reaching state $e_j$ at time step $t+1$ is expressed as:
\begin{multline}
\mathbb{P}\Big(s(c_i, t+1) = e_j\Big) = \\
\sum_{k=1}^{N^R} \mathbb{P}\Big(s(c_i, t+1) = e_j \mid s(c_{i-r},\dotsc, c_{i+r}, t)=(e_{\ind(k)[1]},\dotsc,e_{\ind(k)[R]})\Big) \times\\
\mathbb{P}\Big(s(c_{i-r},\dotsc, c_{i+r}, t)=(e_{\ind(k)[1]},\dotsc,e_{\ind(k)[R]})\Big)\,.
\label{eq:total-prob}
\end{multline}
Substituting Eqs.~$(\ref{eq:pkj})$ and $(\ref{eq:nei-prob})$ in Eq.\ $(\ref{eq:total-prob})$, we get:
\begin{equation}
\mathbb{P}\Big(s(c_i, t+1) = e_j\Big) = \sum_{k=1}^{N^R}
p_{kj}\left(\prod_{m=1}^{R} \pi(c_{i-r+m-1}, t)\big[\ind(k)[m]\big]\right).
\label{eq:pre-f}
\end{equation}
Taking into account the definition of the function $f$ in Eq.\ $(\ref{eq:def-f})$, we can rewrite Eq.\ $(\ref{eq:pre-f})$ as:
\[
\mathbb{P}\Big(s(c_i, t+1) = e_j\Big) = f_j(\pi(c_{i-r},t),\dotsc,\pi(c_{i+r},t))\,.
\]
Since $\pi(c_i, t+1)[j] = \mathbb{P}\big(s(c_i, t+1) = e_j\big)$, the following is proven:
\[\pi(c_i, t+1) = f(\pi(c_{i-r},t),\dotsc,\pi(c_{i+r},t))\,.\qedhere\]
\end{proof}
As shown above, the construction of CCAs is directly connected to the Total Probability Theorem, and thus it enables us to better understand the evolution of cell-wise probability distributions of SCAs. It is important to highlight that the evolution of those distributions is deterministic in SCAs, although the dynamical system itself is stochastic. The finding above agrees with \cite{Betel20095}, where a similar result in the case of binary CAs was shown. Yet, this paper was confined to study the asymptotic behavior of deterministic CAs. \hl{Interestingly, in~\cite{busic2013} an SCA is defined as a discrete-time deterministic dynamical system acting on the set of probability measures on the set of all configurations. Although the definition presented there is much more abstract, and suits mostly theoretical applications, it reassembles the main ideas of CCAs.}
\subsection{Decomposition of SCAs}
In the previous subsection, we have shown that CCAs are directly related to SCAs, since the evolution of a CCA is equivalent to the evolution of the cell-wise probability distributions of an SCA. Here, we uncover the relationship between SCAs and deterministic, $N$--state CAs, and discuss the possibility of decomposing an SCA into deterministic CAs. Let us start with an introductory example motivating our general construction.
\begin{example}[$\alpha$--asynchronous CAs]
Classically, states in deterministic CAs are updated synchronously, {\it i.e.}\ a new state is assigned to all cells simultaneously at every time step according to the local rule. Yet, different approaches of breaking the synchronicity of CAs have been proposed \cite{schonfisch1999synchronous}. Interestingly, the choice of the update scheme, which defines the order or timing of cell state updates, has very important repercussions on the dynamical properties of CAs~\cite{Baetens2012383}. Here we focus on so-called $\alpha$-asynchronous CAs ($\alpha$-ACAs)~\cite{Fates05anexperimental}.
Any $\alpha$-ACA is defined by its deterministic counterpart $A$ and a probability~$\alpha$, called the synchrony rate, which controls whether or not its cells are updated. More precisely, $\alpha$ is the probability of applying the local rule $f$ of $A$, {\it i.e.}:
\[
s(c_i, t+1) =
\begin{cases}
f(s(c_{i-r},\dotsc,c_{i+r},t)) &, \textrm{with probability $\alpha$}, \\
s(c_i, t) &, \textrm{otherwise}.
\end{cases}
\]
Note that if $\alpha=0$, such a system remains at its initial configuration, whereas the system is equivalent to CA $A$ if $\alpha=1$.\qed
\end{example}
The essential property of the construction presented above is that an $\alpha$--ACA can also be seen as a synchronous, stochastic system in which the local rule of $A$ is selected with probability $\alpha$, while the identity rule is selected with probability~$1-\alpha$. Hence, we may say that the rule of a CA $A$ is stochastically mixed with the identity rule.
The idea behind $\alpha$--ACA can be easily extended to the mixing of $q\geq 2$ deterministic rules. Consider synchronous, deterministic CAs $A_1, \dotsc, A_q$ and probabilities $\alpha_1, \dotsc, \alpha_q$ that satisfy $\sum_{i=1}^{q} \alpha_i = 1$. Then, we define an SCA $\widetilde{A}$ in which the state of every cell at every consecutive time step is obtained by evaluating one of the local rules, corresponding to one of the $A_i$. The rule is chosen randomly and independently for every cell at each time step with the selection probability for CA $A_i$ being $\alpha_i$. Such systems, referred to as stochastic mixtures of deterministic CAs in the remainder of this paper, hold very interesting properties that are currently also investigated by others \cite{IOPORT.05887733}. For technical reasons, we assume that the local rules of $A_1, \dotsc, A_q$ are defined with a common radius $r\geq 0$. Since each local rule with radius $r'<r$ can be represented equivalently as a local rule with radius $r$, we are not loosing generality with this assumption.
The first observation is a direct consequence of the definition presented above.
\begin{fact}
Assume that $r\geq 0$ is the radius of the neighborhood of the local rules of automata $A_1, \dotsc, A_q$. Let $R=2\,r+1$ and suppose that $L_i \in \mathcal{S}^{N^R}$ is the LUT of CA $A_i$, for $i=1,\dotsc,q$. Then the pLUT $P\in\mathcal{S}_c^{N^R}$ of the stochastic mixture~$\widetilde{A}$ satisfies $P = \sum_{i=1}^{q} \alpha_i L_i$.
\label{fac:seven}
\end{fact}
Sums of the form $\sum_{i=1}^{q} \alpha_i L_i$, where $\alpha_i \in [0,1]$ and $\sum_{i=1}^{q} \alpha_i=1$ are commonly referred to as convex combinations, which allows us to rephrase Fact \ref{fac:seven} as: the pLUT of a stochastic mixture of deterministic CAs is the corresponding convex combination of the LUTs of the mixed CAs.
Having defined the concept of a stochastic mixture, we are interested in characterizing this class of SCAs. Interestingly, we are able to show that in general there are no characteristics that distinguish stochastic mixtures from other SCAs, since any SCA can be represented as a stochastic mixture. This fact is not surprising if we envisage a stochastic mixture of deterministic CAs as a convex combination of LUTs of deterministic CAs. We can rely on the classical Krein--Milman theorem from convex set theory \cite{Krein1940}. It states that a compact convex set is the convex hull of its extreme points. In our context, the set $\mathcal{S}^{N^R}$ containing the LUTs of deterministic CAs is the set of extreme points of $\mathcal{S}_c^{N^R}$, which is convex and compact. The convex hull of $\mathcal{S}^{N^R}$ is the set of all possible convex combinations of elements from $\mathcal{S}^{N^R}$, which is precisely the set of all stochastic mixtures of deterministic CAs. Due to this, we can state the following theorem.
\begin{thm}
Any SCA can be represented as a stochastic mixture of a finite number of deterministic CAs.
\end{thm}
The decomposition of a stochastic mixture into a convex combination described above is not unique, therefore many methods for constructing the decomposition can be formulated. In the proof of Proposition \ref{prop:stoch-mix-decom}, we introduce one possible algorithm of which it will be shown in Proposition \ref{prop:motiv} that it uncovers the most influential component of the stochastic mixture.
\begin{proposition}
Let $P\in\mathcal{S}_c^{N^R}$. Then there exists an integer $n^\ast$ such that for $i=1,2,\dotsc,n^\ast$, we can define coefficients $\alpha_i\in [0,1]$, $\sum_{i=1}^{n^\ast} \alpha_i = 1$
and matrices $L^i \in \mathcal{S}^{N^R}$, that satisfy:
\begin{equation}
P = \sum_{i=1}^{n^\ast} \alpha_i\,L^i\,.
\end{equation}
\label{prop:stoch-mix-decom}
\end{proposition}
\begin{proof}
We start by defining auxiliary matrices $P^m$, for $m\geq 0$. Let:
\begin{equation}
P^m = \begin{cases}
P &, \textrm{if }m=0, \\
P^{m-1} - \alpha_m\,L^m &, \text{if }m>0,
\end{cases}
\label{eq:pm-mat}
\end{equation}
where, for $m>0$:
\begin{equation}
\alpha_m = \min_i\max_j P^{m-1}_{ij}\,,\label{eq:alpha-def}
\end{equation}
and $L^m = (L^m_{ij})$ such that:
\begin{equation}
L^m_{ij} =
\begin{cases}
1, & \textrm{for }j = \min \Big\{j \mid P^{m-1}_{il} \leq P^{m-1}_{ij}, \textrm{for any $l\in\{1,\dotsc,N\}$}\Big\}\,, \\
0, & \textrm{otherwise}.
\end{cases}
\label{eq:lm-mat}
\end{equation}
Note that, for any $m$ and $i$, it holds that $L^m_{ij} = 1$ for a single value of $j$. Consequently $L^m \in \mathcal{S}^{N^R}$.
Firstly, note that $\alpha_1$ is one of the elements of $P^0=P$, so $\alpha_1\in [0,1]$. Additionally, it holds that $\alpha_1>0$, since if $\alpha_1=0$, then for some $i$, we would have $\max_j P_{ij}=0$, which contradicts the fact that $\sum_j P_{ij} = 1$ for any~$i$.
Secondly, note that for any $i$ and $j$, it holds that $P^0_{ij}\geq 0$. We will show next that the assumption $P^{m-1}_{ij}\geq 0$ leads to $P^{m}_{ij}\geq 0$. Indeed, we know that $P^m_{ij} \in \{ P^{m-1}_{ij}, P^{m-1}_{ij} - \alpha_m \}$. If $P^{m}_{ij} = P^{m-1}_{ij}$, then obviously $P^{m}_{ij}\geq 0$, while if $P^{m}_{ij} = P^{m-1}_{ij}-\alpha_m$, it holds that $P^{m-1}_{ij}\geq P^{m-1}_{il}$ for any $l$, meaning that $P^{m-1}_{ij} = \max_l P^{m-1}_{il}$. On the other hand, since $\alpha_m = \min_i\max_l P^{m-1}_{il}$, it holds that $P^{m-1}_{ij}\geq \alpha_m$. Since it was assumed that $P^{m-1}_{ij}\geq 0$, we see that $P^m_{ij}\geq 0$ as well.
Thirdly, following the argument above, we can easily show that for every $i$, $j$ and $m$, it holds that $P^{m}_{ij} \leq P^{m-1}_{ij}$. This means that for every $i$, it holds that $\max_j P^{m}_{ij} \leq \max_j P^{m-1}_{ij}$, and thus $\alpha_{m+1} \leq \alpha_{m}$. Since $\alpha_1\in\, ]0,1]$, and $\alpha_m\geq 0$ for any $m$, we see that $\alpha_m\in[0,1]$.
Note that for any $i$, it holds that $\sum_j P^m_{ij} = \left(\sum_j P^{m-1}_{ij}\right) - \alpha_m$. Expanding this recursively, yields $\sum_j P^m_{ij} = (\sum_j P^{0}_{ij}) - \sum_{l=1}^m\alpha_l$. From the definition of a stochastic matrix, we know that $\sum_j P^{0}_{ij} =1$, such that $\sum_j P^m_{ij} = 1 - \sum_{l=1}^m\alpha_l$, which means that the column sums in matrix $P^m$ are equal. Since for all $m$ we have shown that $P^m_{ij}\geq 0$, we know that for any $m$, it holds that $1 - \sum_{l=1}^{m}\alpha_l \geq 0$, therefore $\sum_{l=1}^{m}\alpha_l \leq 1$.
We further note that $\alpha_m = 0$, for some $m$, if and only if $P^{m-1}$ contains one zero column. Since we have shown that the column sums in $P^m$ are equal for any $m$, we know that $\alpha_m=0$ if and only if $P^{m-1} = \mathbf{0}$. Therefore, $\alpha_m > 0$ if and only if $P^{m-1}\neq \mathbf{0}$, which implies that in each of the columns of $P^{m-1}$, there is at least one non-zero entry.
Let $z(m) = \#\{(i,j) \mid P^m_{ij} = 0\}$ denote the number of zeros in matrix $P^m$. Note that if $P^m\neq \mathbf{0}$, then $z(m) < z(m+1)$, which follows from the fact that if $P^m\neq \mathbf{0}$, there exists $P^m_{ij}=\alpha_{m+1}>0$, and thus for at least one couple $(i,j)$, we know that $P^{m+1}_{ij} = P^m_{ij} - \alpha_{m+1} = 0$. Since $z(m)$ cannot be greater than the total number of elements in $P^m$, it is not possible that $z(m) < z(m+1)$ for all~$m$. As such, we know that there exist $n^\ast$ for which $P^{n^\ast} = \mathbf{0}$ and $P^{n^\ast -1}\neq \mathbf{0}$, and thus $\sum_{l=1}^{n^\ast}\alpha_l = 1$. Using the definition of $P^{n^\ast}$ multiple times, we get:
\[ \mathbf{0} = P^{n^\ast} = P^{n^\ast -1} - \alpha_{n^\ast}\,L^{n^\ast} = \ldots = P^0 - \alpha_1\,L^1 - \ldots - \alpha_{n^\ast}\,L^{n\ast} = P - \sum_{i=1}^{n^\ast}\alpha_i\,L^i,\]
such that $P = \sum_{i=1}^{n^\ast}\alpha_i\,L^i$.
\end{proof}
To illustrate the construction presented in the proof of Proposition \ref{prop:stoch-mix-decom}, every step of the algorithm is visualized in the following example.
\begin{example}
Although Proposition \ref{prop:stoch-mix-decom} deals with stochastic matrices with dimensions corresponding to pLUTs, it is obvious that the same method can be used to decompose a stochastic matrix of arbitrary size. For simplicity, we apply the method to an exemplary $4\times 3$ stochastic matrix $P$ given by:
\[
P = \begin{bmatrix}
0.6 & 0 & 0.2 & 0.4 \\
0.3 & 1 & 0.3 & 0.4 \\
0.1 & 0 & 0.5 & 0.2
\end{bmatrix}.
\]
According to Eqs.~(\ref{eq:pm-mat}) and (\ref{eq:lm-mat}), we build matrices $P^0$ and $L^1$:
\[
P^0 = P = \begin{bmatrix}
{\bf 0.6} & 0 & 0.2 & {\bf 0.4} \\
0.3 & {\bf 1} & 0.3 & 0.4 \\
0.1 & 0 & {\bf 0.5} & 0.2
\end{bmatrix},
\quad
L^1 = \begin{bmatrix}
1 & 0 & 0 & 1 \\
0 & 1 & 0 & 0 \\
0 & 0 & 1 & 0
\end{bmatrix}.
\]
Note that we highlighted the selected maximal elements in each of the columns using boldface font. The boldface entries in $P^{0}$ correspond to the positions of the ones in matrix $L^1$. Note that in the last column of $P^0$, there is no unique maximal element. Yet, as given by Eq.\ (\ref{eq:lm-mat}), we pick the first of the entries from top to bottom. Picking the other possibility would affect the subsequent matrices $L^m$ and result in a different decomposition, but the coefficients $\alpha_m$ would not change.
Following Eq.\ (\ref{eq:alpha-def}), we find that $\alpha_1 = 0.4$. We proceed and calculate $P^1$ and~$L^2$:
\[
P^1 = P^0 - \alpha_1\,L^1 = \begin{bmatrix}
0.2 & 0 & 0.2 & 0 \\
{\bf 0.3} & {\bf 0.6} & {\bf 0.3} & {\bf 0.4} \\
0.1 & 0 & 0.1 & 0.2
\end{bmatrix},
\quad
L^2 = \begin{bmatrix}
0 & 0 & 0 & 0 \\
1 & 1 & 1 & 1 \\
0 & 0 & 0 & 0
\end{bmatrix},
\]
and from this we see that $\alpha_2 = 0.3$. We continue the procedure:
\[
P^2 = P^1 - \alpha_2\,L^2 = \begin{bmatrix}
{\bf 0.2} & 0 & {\bf 0.2} & 0 \\
0 & {\bf 0.3} & 0 & 0.1 \\
0.1 & 0 & 0.1 & {\bf 0.2}
\end{bmatrix},
\quad
L^3 = \begin{bmatrix}
1 & 0 & 1 & 0 \\
0 & 1 & 0 & 0 \\
0 & 0 & 0 & 1
\end{bmatrix},
\]
and $\alpha_3 = 0.2$. We proceed one more step and get:
\[
P^3 = P^2 - \alpha_3\,L^3 = \begin{bmatrix}
0 & 0 & 0 & 0 \\
0 & {\bf 0.1} & 0 & {\bf 0.1} \\
{\bf 0.1} & 0 & {\bf 0.1} & 0
\end{bmatrix},
\quad
L^4 = \begin{bmatrix}
0 & 0 & 0 & 0 \\
0 & 1 & 0 & 1 \\
1 & 0 & 1 & 0
\end{bmatrix},
\]
and $\alpha_4 = 0.1$. This is the final step, since: $P^4 = P^3 - \alpha_4\,L^4 = {\bf 0}$. Therefore, the decomposition can be written as:
\begin{align*}
P = 0.4 &\times \begin{bmatrix}
1 & 0 & 0 & 1 \\
0 & 1 & 0 & 0 \\
0 & 0 & 1 & 0
\end{bmatrix} + 0.3 \times \begin{bmatrix}
0 & 0 & 0 & 0 \\
1 & 1 & 1 & 1 \\
0 & 0 & 0 & 0
\end{bmatrix} \\ + 0.2 &\times \begin{bmatrix}
1 & 0 & 1 & 0 \\
0 & 1 & 0 & 0 \\
0 & 0 & 0 & 1
\end{bmatrix} + 0.1 \times \begin{bmatrix}
0 & 0 & 0 & 0 \\
0 & 1 & 0 & 1 \\
1 & 0 & 1 & 0
\end{bmatrix}.\\[-\normalbaselineskip]\tag*{\qed}
\end{align*}
\end{example}
As mentioned earlier, the construction presented in the proof of Proposition~\ref{prop:stoch-mix-decom} is one of many possibilities of decomposing a pLUT, but in the following proposition we will show that it enables us to capture the element of the mixture with the highest possible coefficient.
\begin{proposition}
Let $P\in\mathcal{S}_c^{N^R}$, $p=1,\dotsc, p_{max}$, $q=1,\dotsc, q_{max}$, and let $\alpha_p, \beta_q \in [0,1]$, $L^p, M^q \in \mathcal{S}^{N^R}$ be such that $\sum_p \alpha_p = \sum_q \beta_q = 1$. Moreover, let $P=\sum_p \alpha_p\,L^p = \sum_q \beta_q\,M^q$, and let $\alpha_p$, $L^p$ be defined as in the proof of Proposition \ref{prop:stoch-mix-decom}.
It then holds that $\max_q \beta_q \leq \alpha_1 = \max_p \alpha_p$.
\label{prop:motiv}
\end{proposition}
\begin{proof}
Let $i$ and $j$ be such that $P_{ij} = \alpha_1$. Then for any $k$, it holds that $P_{ik}\leq \alpha_1$. Let $q\in\{1,\dotsc,q_{max}\}$ and let $l$ be such that $M^q_{il} = 1$. Since $P~=~ \sum_q \beta_q\,M^q$, we know that $\beta_q \leq P_{il}$, and thus $\beta_q \leq P_{il} \leq \alpha_1$. Therefore, for any $q~\in~\{1,\dotsc,q_{max}\}$, it holds that $\beta_q \leq \alpha_1$ and thus $\max_q \beta_q \leq \alpha_1$.
\end{proof}
The informal meaning of the above proposition is that the presented approach to decompose an SCA uncovers the deterministic rule that has the highest probability of being executed, and therefore is likely to have the highest influence on the behavior of the system. The exact relations between the dynamical behavior of SCAs and the dynamical characteristics of the components of stochastic mixtures are still under investigation.
\hl{The experiments in the following section underline that in some cases those CAs that have the highest probability of application indeed greatly influence the behavior of the SCA. Yet, in some cases, those with a very small probability of application can also play an important role. Therefore, at this stage we do not claim that the overall dynamics of the SCA is always determined by the component of the mixture with the highest probability of application.
The decomposition of an SCA as a stochastic mixture might also find its use in SCA-based modeling. Indeed, in some cases it might be easier to express the model in terms of deterministic components having a meaning in the language relevant to the modeling task. Due to the presented equivalence, we are guaranteed that such a practice will not limit the modeling potential.}
\section{Experiments}
\label{sec:experiments}
\subsection{Analysis of $\alpha$--asynchronous ECAs}
\label{sec:aaca}
Through the following experiment, we will analyze the behavior of $\alpha$-asyn\-chro\-nous ECAs. For a given rule, the cell-wise distance between the space-time diagram of the deterministic CA and the space--time diagram of a CCA representation of an $\alpha$--ACA variant, for $\alpha$ ranging from $0$ to $1$, was measured. More formally, if $A$ is an ECA rule, and $A_\alpha$ is the CCA representation of the $\alpha$--asynchronous variant of $A$ for $\alpha\in[0,1]$, we measured the distance $D(\alpha)$ between space-time diagrams.
For a random initial condition $I_0\in \{0,1\}^M$, where $M>0$ denotes the number of cells and $T$ the number of time steps, the distance $D$ is defined as:
\begin{equation}
D(\alpha) = \frac{1}{M\,T} \sum_{t=1}^{T} \norm{A^t(I_0) - A^t_\alpha(I_0)}\,,
\end{equation}
where $A^t(I_0)$ denotes the result of applying the global rule $A$ $t$ times to input~$I_0$. In this experiment, we choose $M=T=69$.
Unsurprisingly, it turned out that the most pronounced discrepancies between the parallel evolutions are observed when $\alpha$ approaches 1. For that reason, we restrict the further discussion to $\alpha\in[0.9,1]$. We classified the ECAs according to the behavior of the function $D$ (the contents of the classes is shown in Table~\ref{tab:ex1-classes}):
\begin{itemize}
\item Class I: $D(\alpha)$ is almost 0, for all $\alpha\in[0.9,1]$ (Figure~\ref{fig:c1}),
\item Class II: $D(\alpha)$ smoothly decreases towards 0, as $\alpha$ increases (Figure~\ref{fig:c2}),
\item Class IIIa: there is a very sudden drop of $D(\alpha)$ as $\alpha$ approaches 1 (Figure~\ref{fig:c3a}),
\item Class IIIb: there is a very sudden drop of $D(\alpha)$ as $\alpha$ approaches 1, and the behavior is not monotonic (Figure~\ref{fig:c3b}).
\end{itemize}
\begin{figure}
\centering
\subfloat[Class I (ECA 40)]{\label{fig:c1}\includegraphics[width=0.4\textwidth]{40.pdf}}\
\subfloat[Class II (ECA 42)]{\label{fig:c2}\includegraphics[width=0.4\textwidth]{42.pdf}}\
\subfloat[Class IIIa (ECA 38)]{\label{fig:c3a}\includegraphics[width=0.4\textwidth]{38.pdf}}\
\subfloat[Class IIIb (ECA 43)]{\label{fig:c3b}\includegraphics[width=0.4\textwidth]{43.pdf}}\
\caption{Plots of $D(\alpha)$ for representatives of each of the classes defined in Table \ref{tab:ex1-classes}.}\label{pic:ex1}
\end{figure}
\begin{table}
\centering
\begin{tabular}{r|l}
\hline
{\bf Class I} &
0, 8, 12, 32, 40, 64, 68, 72, 76, 77, 93, 96, 128, 132,\\ & 136, 140, 160, 168, 192, 196, 200, 205--207, 220, 221, \\ & 224, 232, 233, 235--239, 249--255 \\ \hline
{\bf Class II} &
1--5, 7, 10, 13, 15--17, 19, 21, 23, 24, 29, 31, 34, 36,\\ & 42, 44, 48, 50, 51, 55, 56, 63, 66, 69, 71, 79, 80, 85, \\ & 87, 92, 95, 100, 104, 108, 112, 119, 127, 130, 138,\\ & 141, 144, 152, 162, 164, 170--172, 174--176, 178, 179, \\ & 186--191, 194, 197, 201--203, 208, 216--219, 222, 223, \\ & 226, 228, 230, 231, 234, 240--248 \\ \hline
{\bf Class IIIa} &
6, 9, 18, 20, 22, 25--28, 30, 33, 35, 37--39, 41, 45, 46, \\ & 49, 52--54, 57--62, 65, 67, 70, 73--75, 78, 82, 83, 86,\\ & 88--91, 94, 97--99, 101--103, 105--107, 109--111, 114--116, \\ &118, 120--126, 129, 131, 133--135, 137, 139, 145--151, \\ & 153--159, 161, 163, 165--167, 169, 173, 177, 180--185, \\ &193, 195, 198, 199,204, 209--211, 214, 215, 225, 227, 229 \\ \hline
{\bf Class IIIb} &
11, 14, 43, 47, 81, 84, 113, 117, 142, 143, 212, 213 \\ \hline
\end{tabular}
\caption{Classes of $\alpha$--Asynchronous ECAs.}
\label{tab:ex1-classes}
\end{table}
The interpretation of the classes is as follows. Rules belonging to Class~I are resistant to $\alpha$--asynchronicity, which means that introducing the asynchronicity aspect does not affect the behavior, or that its impact is negligible. Rules belonging to Class II are affected by the asynchronicity, but the impact on the final behavior is proportional to the synchrony rate $\alpha$, {\it i.e.}\ the behavior steadily approaches the deterministic case as $\alpha$ approaches one. Rules belonging to Classes~IIIa and IIIb are sensitive to $\alpha$--asynchronicity, meaning that the behavior of the system changes drastically as soon as $\alpha$ is smaller than 1. Class~IIIb can be distinguished from Class IIIa, by the noisy behavior of $D(\alpha)$ for $\alpha$ close to one. The cause of the differences is not yet uncovered.
\hl{Classes IIIa and IIIb illustrate the case where the decomposition as a stochastic mixture does not suffice to understand the dynamics of the SCA. Indeed, for large $\alpha$, the behavior of the SCA and that of the deterministic CA with the highest probability of application is quite different.}
\subsection{Stochastic density classifiers}
We consider the SCA defined by the local rule $\mathbf{C}_3$ introduced in \cite{fates:inria-00608485,nazim-automata2013}. This rule is defined by the pLUT shown in Table \ref{tab:plut-nazim}. Note that this pLUT uses the standard notation for its state set $\mathcal{S}=\{0,1\}$, and its entries constitute the probabilities of reaching state $1$. It was shown both analytically and experimentally that this rule is a stochastic solution for the classical density classification problem (DCP) \cite{fates:inria-00608485} with arbitrary precision, {\it i.e.}\ by varying the parameter $\eta$, we can achieve any probability $p<1$ of successful classification. The DCP was introduced in \cite{gacs1978one,packard1988adaptation}.
\renewcommand{\tabcolsep}{.1cm}
\begin{table}[ht]
\centering
\begin{tabular}{>{$}c<{$}|>{$}c<{$}|>{$}c<{$}|>{$}c<{$}|>{$}c<{$}|>{$}c<{$}|>{$}c<{$}|>{$}c<{$}}
\hline
(1,1,1) & (1,1,0) & (1,0,1) & (1,0,0) & (0,1,1) & (0,1,0) & (0,0,1) & (0,0,0) \\
\hline
1 & \eta & 1 & 1-\eta & 1 & 0 & 0 & 0 \\
\hline
\end{tabular}
\caption{pLUT of local rule $\mathbf{C}_3$ \cite{fates:inria-00608485,nazim-automata2013}.}\label{tab:plut-nazim}
\end{table}
Using the CCA representation, we can visualize the average behavior of~$\mathbf{C}_3$. In Figure \ref{pic:c3}, the first 80 rows of the averaged space-time diagram (time goes from top to bottom) obtained using the CCA representation of $\mathbf{C}_3$ with $\eta=0.1$ are shown, together with three typical space-time diagrams obtained by direct evaluation of the SCA rule $\mathbf{C}_3$. All images were obtained using the same initial condition involving 29 cells out of which 16 were black (state 1). As can be inferred from Figure \ref{pic:c3}, two samples lead to behavior that is similar to the one displayed by the average space-time diagram, while the one depicted in Figure \ref{fig:fail} behaves differently, and leads to a wrong classification. Since the probability of obtaining a correct classification for the initial condition at stake was estimated to be 90.1\% on the basis of 10000 repetitions, the impact of the erroneous cases on the averaged image is hardly visible.
\begin{figure}
\centering
\subfloat[Averaged]{\includegraphics[width=0.19\textwidth,fbox]{avg-image.pdf}}\quad
\subfloat[Sample \#1]{\includegraphics[width=0.19\textwidth,fbox]{stochastic-image1.pdf}}\quad
\subfloat[Sample \#2]{\includegraphics[width=0.19\textwidth,fbox]{stochastic-image2.pdf}}\quad
\subfloat[Failing sample]{\label{fig:fail}\includegraphics[width=0.19\textwidth,fbox]{stochastic-image4-fail.pdf}}\quad
\caption{Space time diagrams of $\mathbf{C}_3$ with $\eta = 0.1$.}\label{pic:c3}
\end{figure}
One of the most interesting questions within the analysis of DCP solutions relates to the expected time of convergence towards the outcome of the classification and how this relates to the number of cells. Figure \ref{pic:c3-conv-time} depicts the average convergence time calculated over an ensemble of 5000 random initial conditions for different numbers of cells and values of $\eta$, both for the CCA and the SCA representation. The initial conditions were generated as follows. For every initial condition independently, a probability $p$ was selected randomly, and then the initial states were selected with $p$ being the probability of selecting state 0 at each cell independently. Such a selection procedure assures that each of the possible densities has the same probability in the ensemble of initial conditions, and thus we can evaluate the classifier across a diverse set of densities.
In the case of the CCA representation, we cannot expect reaching a truly homogeneous, global state. Therefore we evolved it until the maximum, absolute difference between the states of any two cells was lower than 0.001. Then, we verified whether the average of the states was closer to 1 or 0. In the case of the SCA, for each of the 5000 initial conditions, 100 simulations were performed, and the results were averaged.
\begin{figure}
\centering
\subfloat[CCA]{\includegraphics[width=0.45\textwidth]{cca-n.pdf}}\quad
\subfloat[SCA]{\includegraphics[width=0.45\textwidth]{sca-n.pdf}}
\caption{Expected convergence time of rule $\mathbf{C}_3$ for different values of $\eta$.}\label{pic:c3-conv-time}
\end{figure}
The results for small grid sizes (smaller than 200) are similar for both representations and agree with the findings presented in \cite{fates:inria-00608485,nazim-automata2013}. Yet, the computing time when using the CCA representation is significantly lower than in the case of SCAs. This indicates that the CCA representation of SCAs proposed in this paper might enable a substantial reduction of the required computing time. Moreover, using the CCA representation, we are able to get more insight into the behavior of the dynamical system by plotting the evolution of the density over time. Figure \ref{pic:c3-conv-time2} shows the results of such an experiment. For the sake of clarity, these results are based on an ensemble of 400 initial conditions (69 cells) out of which 200 had a density greater than 0.5 (green), while the other 200 had a density smaller than 0.5 (red). The ensemble of initial configurations was the same for all values of $\eta$. Note that the range of the horizontal axis differs across plots. From Figure \ref{pic:c3-conv-time2}, we can infer that the time and quality of classification increases with decreasing $\eta$. Due to the computational costs involved in evolving and averaging SCAs directly, drawing and analyzing such plots has become possible only due to the introduction of CCAs.
\begin{figure}
\centering
\subfloat[$\eta=0.1$]{\includegraphics[width=0.48\textwidth]{dcp-plot-1.png}}
\subfloat[$\eta=0.05$]{\includegraphics[width=0.48\textwidth]{dcp-plot-05.png}}\\
\subfloat[$\eta=0.01$]{\includegraphics[width=0.48\textwidth]{dcp-plot-01.png}}
\subfloat[$\eta=0.005$]{\includegraphics[width=0.48\textwidth]{dcp-plot-005.png}} \\
\subfloat[$\eta=0.001$]{\includegraphics[width=0.48\textwidth]{dcp-plot-001.png}}
\caption{Density evolution over time for the CCA representation of SCA $\mathbf{C}_3$ for 200 initial conditions with density greater than 0.5 (green) and 200 other initial conditions with density smaller than 0.5 (red) and different values $\eta$.}\label{pic:c3-conv-time2}
\end{figure}
\subsection{Totalistic SCAs}
\hl{In this section, we analyze a specific class of totalistic SCAs. Although most likely being of no practical use, it holds many similarities with more complex CAs and SCAs that are often resorted to in modelling. Totalistic CAs are generally known for their ability to mimic real-world phenomena \cite{RevModPhys.55.601,wolfram-class,total}. Given their importance and straightforward formulation, we have chosen this class to serve as an example for our analysis. The focus is on the illustration of the tools introduced in this paper, rather than to unveil new properties of some complex CA models.}
We consider the class of 1D, binary SCAs with a unit neighborhood radius, that satisfy the following conditions:
\begin{align*}
\mathbb{P}\big(s(c_i,t+1) = e_2 \mid &\ s(c_{i-1}, c_i, c_{i+1}, t) = (e_2,e_1,e_1)\big) = \\
\mathbb{P}\big(s(c_i,t+1) = e_2 \mid &\ s(c_{i-1}, c_i, c_{i+1}, t) = (e_1,e_2,e_1)\big) = \\
\mathbb{P}\big(s(c_i,t+1) = e_2 \mid &\ s(c_{i-1}, c_i, c_{i+1}, t) = (e_1,e_1,e_2)\big) = p_1,
\end{align*}
\begin{align*}
\mathbb{P}\big(s(c_i,t+1) = e_2 \mid &\ s(c_{i-1}, c_i, c_{i+1}, t) = (e_1,e_2,e_2)\big) = \\
\mathbb{P}\big(s(c_i,t+1) = e_2 \mid &\ s(c_{i-1}, c_i, c_{i+1}, t) = (e_2,e_1,e_2)\big) = \\
\mathbb{P}\big(s(c_i,t+1) = e_2 \mid &\ s(c_{i-1}, c_i, c_{i+1}, t) = (e_2,e_2,e_1)\big) = p_2,
\end{align*}
where $e_1$ and $e_2$ are the base vectors of the Euclidean space $\mathbb{R}^2$.
Such SCAs are commonly referred to as totalistic SCAs. We consider the subclass of such SCAs satisfying:
\begin{align*}\mathbb{P}\big(s(c_i,t+1) = e_2 \mid s(c_{i-1}, c_i, c_{i+1}, t) = (e_2,e_2,e_2)\big) = 1,\\
\mathbb{P}\big(s(c_i,t+1) = e_1 \mid s(c_{i-1}, c_i, c_{i+1}, t) = (e_1,e_1,e_1)\big) = 1.
\end{align*}
In this paper, we will refer to this subclass as totalistic SCAs. By applying the decomposition algorithm given by Proposition \ref{prop:stoch-mix-decom}, we can prove the following fact, revealing their structure.
\begin{fact}
Any totalistic SCA can be written as a stochastic mixture of the following ECAs: 128, 150, 232, 254, where up to three of those rules are applied with non-zero probability.
\end{fact}
We can determine the regions in the $(p_1,p_2)$--plane where a given deterministic CA has the highest probability of application, {\it i.e.} where it is dominant. These regions are given by:
\begin{itemize}
\item ECA 128 is dominant, if $p_1 < 0.5$ and $p_2 < 0.5$,
\item ECA 150 is dominant, if $p_1 < 0.5$ and $p_2 > 0.5$,
\item ECA 232 is dominant, if $p_1 > 0.5$ and $p_2 < 0.5$,
\item ECA 254 is dominant, if $p_1 > 0.5$ and $p_2 > 0.5$.
\end{itemize}
Note that if $p_1=0.5$ or $p_2=0.5$, there is no unique, dominant rule, therefore we have omitted those cases in the description above.
Let $\alpha_{128}$, $\alpha_{150}$, $\alpha_{232}$, $\alpha_{254}$ denote the probabilities of applying ECAs 128, 150, 232, 254, respectively. Figure \ref{fig:regions} depicts the dependence of $\alpha_{150}$ and $\alpha_{232}$ on $p_1$ and $p_2$. We have omitted the remaining two images for $\alpha_{128}$ and $\alpha_{254}$, since they do not yield additional information.
\begin{figure}
\centering
\subfloat[$\alpha_{150}$]{\includegraphics[width=0.48\textwidth]{alpha150.pdf}}
\subfloat[$\alpha_{232}$]{\includegraphics[width=0.48\textwidth]{alpha232.pdf}}
\caption{Dependence of the application probabilities: $\alpha_{150}$ of ECA 150 and $\alpha_{232}$ of ECA 232 in the totalistic SCA, on the values of the parameters $p_1$ and $p_2$.}\label{fig:regions}
\end{figure}
The dynamical characteristics of the four ECAs that compose the totalistic SCAs differ significantly. Rules 128 and 254 are simple -- they belong to Class~I according to Wolfram's classification \cite{wolfram-class}, and their maximum Lyapunov exponents (MLEs) \cite{Bagnoli199234} equals $-\infty$, which means that any changes to their initial configuration do not influence their long-term behavior. Rule 232 belongs to complexity Class II, and its MLE is positive, but very close to zero. In contrast, rule 150 is a Class III rule, and its MLE is the highest among all ECAs, which highlights the fact that this rule is sensitive to the smallest perturbation of its initial configuration. Since rules 150 and 232 are relatively more complex than rules 128 and 254, we might expect a distinct behavior of the stochastic mixture in regions where the former are dominant.
In order to verify this, we set up an experiment involving a random initial condition of $M=49$ cells which was evolved 100 times for $T=49$ time steps. Such a procedure, for the same initial condition, was repeated for multiple different choices of $(p_1,p_2)$ using a $101\times 101$ regular grid, resulting in a set of space-time diagrams. The set of such space-time diagrams is denoted as $\mathcal{I}(p_1,p_2)$. Let $\Delta(p_1, p_2) = \{ \dist(I,J) \mid I\neq J; I,J\in \mathcal{I}(p_1,p_2) \}$ denote the set of all pair-wise Hamming distances between space-time diagrams, and $\Delta_T(p_1, p_2) = \{ \dist(I[T],J[T]) \mid I\neq J; I,J\in \mathcal{I}(p_1,p_2) \}$ denote the set of all pair-wise Hamming distances between the final configurations in the space-time diagrams. The results for the minimum, mean and maximum of $\Delta(p_1,p_2)$ and $\Delta_T(p_1,p_2)$ are shown in Figure \ref{fig:diff}.
\begin{figure}
\centering
\subfloat[$\max \Delta(p_1,p_2)$]{\includegraphics[width=0.48\textwidth]{dmax-plot.pdf}}
\subfloat[$\max \Delta_T(p_1,p_2)$]{\includegraphics[width=0.48\textwidth]{dfmax-plot.pdf}}\\
\subfloat[$\avg \Delta(p_1,p_2)$]{\includegraphics[width=0.48\textwidth]{d-plot.pdf}}
\subfloat[$\avg \Delta_T(p_1,p_2)$]{\includegraphics[width=0.48\textwidth]{df-plot.pdf}}\\
\subfloat[$\min \Delta(p_1,p_2)$]{\includegraphics[width=0.48\textwidth]{dmin-plot.pdf}}
\subfloat[$\min \Delta_T(p_1,p_2)$]{\includegraphics[width=0.48\textwidth]{dfmin-plot.pdf}}
\caption{Impact of the probabilities $p_1$ and $p_2$ on the distances $\Delta(p_1,p_2)$ between the space-time diagrams (a, c, e) and the distances $\Delta_T(p_1,p_2)$ between the final configurations at $T=49$ (b, d, f), obtained obtained from the same initial condition.}\label{fig:diff}
\end{figure}
As can be seen from the charts, the distance is especially high in those regions where the application probability $\alpha_{150}$ is high, {\it i.e.}\ where $p_1>0.5$ and $p_2<0.5$, and near the line $p_2 = -p_1$ in the quarter $p_1<0.5$ and $p_2>0.5$. The first region is the one where ECA 150 is dominant, while ECA 232 is dominant in the second one, but with ECA 150 having a substantially high application probability as well. Although the Hamming distance is just a simple indication of complexity, we already see a strong influence of the ECA 150 component.
\hl{The findings presented above suggest that analyzing the components of a stochastic mixture decomposition of an SCA unveils information on the dynamics of the SCA. In contrast to the $\alpha$-ACAs in Classes IIIa and IIIb ({\it cf.} Table~\ref{tab:ex1-classes}), here we encounter a class of SCA for which uncovering the components of the stochastic mixture with relatively high probability of application gives additional insight into the dynamical properties of the SCA.}
\section{Summary}
\label{sec:summary}
In this paper, two basic, yet important properties of SCAs were discussed. We have shown, both theoretically and practically, that SCAs can be effectively analyzed in the context of deterministic CAs and CCAs. The stochastic mixture representation of SCAs allows to understand the underlying dynamics, while the CCA representation allows to quickly uncover the average behavior of the system. Further research is undertaken to expand the application scope of the presented methods.
\section*{Acknowledgments}
Witold Bołt is supported by the Foundation for Polish Science under International PhD Projects in Intelligent Computing. This project is financed by the European Union within the Innovative Economy Operational Program 2007--2013 and the European Regional Development Fund.
The authors gratefully acknowledge the financial support of the Research Foundation Flanders (FWO project 3G.0838.12.N).
\section*{Bibliography}
\bibliographystyle{elsarticle-num}
|
\section{Introduction}
\setcounter{equation}{0}
Consider a fluid interface in contact with a solid substrate. This scenario
describes a container filled with liquid, a drop sitting on a leaf, or a
vapour bubble \new{inside a liquid filled bottle.} Imagine observing a point
in the vapour phase. As the liquid phase is approached, a \new{rapid, yet}
smooth transition in the density occurs at the liquid-vapour interface.
Staying \new{on} this interface and approaching the substrate, would reveal a
variety of physical effects that \new{become significant.} First, the fluid
feels an attractive force of the wall particles. At the same time, the nature
of the solid substrate forces the fluid particles to \new{`jam'} and restrict
their mobility as the wall is approached.
In this work, we are interested in the effect \newB{of} the wall attractive
forces on the density profile in the vicinity of a three-phase contact line
for a wide range of contact angles. Developing a fundamental understanding of
these small-scale phenomena at equilibrium is important to predict the
dynamic nanoscale behaviour of the moving contact line, which is still a
controversial problem with a wide range of physical explanations being
offered (for a review, see Bonn~\etal~\cite{Bonn.20090527} or Snoeijer and
Andreotti~\cite{SnoeijerAndreotti:2013}). In this context, our intention is
twofold: First, to illustrate and give a general understanding for the
density structure of the fluid as well as its
\new{form} and scale of variations in the vicinity of the contact line;
\new{and second}, to illustrate the impact of the contact line on the normal
pressure distribution acting on the substrate. The latter point is directly
connected with the definiton of the disjoining pressure. The uniqueness of
disjoining pressure definitions was recently discussed critically in several
papers~\cite{Herring:2010vn,Henderson:2011:EPJST:DisjoiningPressure,MacDowell:2011:ResponseEPJST,Henderson:EPJST:ResponseMacDowell,Henderson:NoteContinuingContactLine,Nold:2014:FluidStructure}.
To describe the interaction between a solid substrate and a fluid interface,
we choose to model a simple fluid, \new{i.e.\ }a system of identical
particles in contact with a homogeneous, perfectly flat\new{,} hard wall. The
particles of the fluid are modelled as hard spheres interacting with a
Lennard-Jones type potential decaying with $r^{-6}$, where $r$ is the
interparticle distance.
\newB{The wall and fluid particles are assumed to interact via a similar
Lennard-Jones type potential.}
Contact line models, including nonlocal contributions to the free energy beyond those of the disjoining pressure, have previously been studied
analytically~\cite{Merchant:1992kx,Pismen:2001fk,Snoeijer:2008fk,Getta:1998ly,SnoeijerAndreotti:2013}. However,
for the sake of analytical attainability, only simple models of the free energy model can be considered and
restrictive assumptions on the nature of the density profile at the contact line have to be made.
In contrast, we consider the density structure at the contact line numerically employing
classical density functional theory (DFT), an approach derived from
the statistical mechanics of fluids~\cite{Evans}. DFT has
proven to be a numerically efficient way to model equilibrium properties of
inhomogeneous fluid systems. It can be viewed as middle ground between
continuum hydrodynamics, which is inapplicable at small fluid volumes,
and particle-based Monte-Carlo (MC) or Molecular Dynamics (MD) simulations,
which despite dramatic improvements in
computational power are still restricted to small fluid volumes.
In fact, compared to MC or MD simulations, for which the numerical complexity scales
with the number of particles modelled, DFT gives the ability to solve
directly for the density distribution,
with the advantage that its computational complexity is formally independent of the number of particles.
Thus, modelling larger systems, such as contact lines, becomes feasible.
The predictive qualities of the DFT results depend on the accuracy of the
free-energy model employed. Here, we model the hard-sphere free energy with a
fundamental measure theory (FMT)~\cite{Rosenfeld:1989qc}, while the
attractive forces are included as a Barker-Henderson
perturbation~\cite{Barker:1967rq} in a mean-field manner. DFT-FMT has been
applied successfully in studies of critical point wedge
filling~\cite{Malijevsky:2013:CriticalPointWedgeFilling}, phase transitions
in nanocapillaries~\cite{PeterPRE}, thin films on planar
substrates~\cite{Peter2012} and density computations in the vicinity of
liquid wedges~\cite{Merath:2008}. A previous study by~\cite{Antonio2010} on
equilibrium contact lines utilised a DFT local-density approximation (LDA)
which is not appropriate to describe structuring in the fluid and fails to
describe the oscillatory behavior of the density in the immediate vicinity of
a wall.
The present work parallels our previous study
in~\cite{Nold:2014:FluidStructure} where DFT-FMT was used to analyse the
fluid structure in the immediate vicinity of a contact line for $\thYoung <
90^\circ$. Here we investigate a wide spectrum of contact angles $40^\circ <
\thYoung < 135^\circ$ and we shed further light on the density structure in
the vicinity of the contact line and its dependency \new{on} the wall
strength. A discussion of the special case of a $90^\circ$ contact angle is also included.
We present density profiles slice by slice as we sweep through the
contact line region and we contrast the profiles with that of a planar liquid
film on a substrate with the same film thickness, but at an off-saturation chemical potential.
Interestingly, the two are not that different, which
suggests that results of the planar film case may be transferable to the contact line.
In particular, as
in~\cite{Nold:2014:FluidStructure} we shall scrutinize the ability of
Derjaguin-Frumkin theory~\cite{Derjaguin:1987:50YearsOfSurfaFceScience} for
planar liquid films on a substrate to predict the height profile at the
contact line. We offer a unified Derjaguin-Frumkin treatment of the contact
line for $\thYoung < 90^\circ$ and $\thYoung > 90^\circ$ by appropriately
extending the boundary conditions for the disjoining pressure equation to
account for the case $\thYoung > 90^\circ$. We further study the connection
between the Derjaguin-Frumkin disjoining pressure and the normal pressure
distribution acting on the substrate for non-planar liquid films, such as
given by the contact line, for $40^\circ < \thYoung < 135^\circ$.
In section 2 we give an overview of the DFT model employed to solve for the
equilibrium density profile. The numerical scheme to compute the contact
angles is introduced in section 3. A description of the density structure in
the vicinity of the contact line is given in section 4, before discussing
coarse-grained Hamiltonian approaches in section 5. Finally, a general
discussion of the results and concluding remarks are in section 6.
\section{Statistical mechanics framework \label{sec:DFTmodel}}
\vspace*{0.5cm} \setcounter{equation}{0}
As done for contact angles less than $90^\circ$ in
Ref.~\cite{Nold:2014:FluidStructure}, we employ classical DFT to investigate
the density distribution in the vicinity of an equilibrium contact line at
contact angles both greater and less than $90^\circ$. It is based on a
statistical mechanics description and has been successfully applied in the
study of inhomogeneous fluids. It is based on the theorem of Mermin
\cite{Mermin:1965lo}, which allows the Helmholtz free energy $\FE$ to be
written as a unique functional of the number density $\nDensity({\bf
r})$~\cite{Wu-DFT}. The equilibrium density distribution minimizes the grand
potential \cite{Evans}
\begin{align}
\GrandPotential[\nDensity] = \FE[\nDensity] + \int \nDensity(\pos) \klammCurl{\Vext({\bf r}) - \chemPotN} \dI\pos, \label{eq:GrandPotential}
\end{align}
where $\chemPotN$ is the chemical potential and $\Vext$ is the external
potential, dependent on the position vector $\pos$. We then minimize
Eq.~(\ref{eq:GrandPotential}) by solving the Euler-Lagrange equation
\begin{align}
\frac{\delta \Omega[\nDensity]}{\delta \nDensity({\pos})} = 0,\label{eq:EulerLagrangeEquation}
\end{align}
where for a simple fluid of particles interacting with a Lennard-Jones
potential, the free energy is usually separated into a repulsive hard-sphere
part and an attractive contribution
\begin{align}
\FE[\nDensity] = \FEhs[\nDensity] + \FEattr[\nDensity].
\end{align}
To accurately model both the structure and thermodynamics of hard-sphere
fluids, we use the Rosenfeld FMT approach~\cite{Rosenfeld:1989qc} for the
hard-sphere contribution~\cite{Roth:2010fk}. The attractive interactions are
\new{modelled} with a mean-field Barker-Henderson approach
\cite{Barker:1967rq}
\begin{align}
\FEattr[\nDensity] &= \frac{1}{2 } \iint \phi_{\text{attr}}({|\pos - \pos'|}) \nDensity(\pos)\nDensity(\pos') \dI\pos' \dI\pos, \label{eq:FEattr}
\end{align}
where the attractive interaction potential is given by
\begin{align}
\BHattr\klamm{r} =
\depthLJ \left\{ \begin{array}{ll}
0 & \text{for } r \leq \LJdiam\\
4 \klamm{
\klamm{\frac{\LJdiam}{r}}^{12} -
\klamm{\frac{\LJdiam}{r}}^{6}
} & \text{for } r > \LJdiam
\end{array} \right. .\label{eq:pattr}
\end{align}
Here, \new{$\depthLJ$ is the depth of the Lennard-Jones potential, $\LJdiam$ is the distance from the center of the particle at which the
Lennard-Jones potential is zero, and $r$ is a (scalar) radial distance}. The
simple fluid described by (\ref{eq:GrandPotential})--(\ref{eq:pattr}) has a critical point at \newB{$k_B T_c/\depthLJ = 1.0$}, where $k_B$ is the Boltzmann constant. \new{Computations} in
this work are performed at $T = 0.75 T_c$, at which the liquid and vapour
number densities are well-separated ($\nDensityL \LJdiam^3= 0.622$,
$\nDensityV \LJdiam^3= 0.003$) and at which the liquid-vapour surface tension \newB{resulting from planar DFT computations} is $\surfaceTensionLV = 0.3463 \depthLJ/\LJdiam^2$. All two-dimensional (2D) computations are
performed at the saturation chemical potential, at which the bulk \new{vapour} and
bulk liquid are equally stable.
The wall-fluid particle interaction is modelled
analogously to the fluid-fluid interaction as
\begin{align}
\BHattrWF\klamm{r} =
\depthLJW \left\{ \begin{array}{ll}
\infty & \text{for } r \leq \LJdiam\\
4 \klamm{
\klamm{\frac{\LJdiam}{r}}^{12} -
\klamm{\frac{\LJdiam}{r}}^{6}
} & \text{for } r > \LJdiam
\end{array} \right. , \label{eq:WallFluidInteraction}
\end{align}
where $\depthLJW$ is the depth of the wall-fluid interactions. Let us take a
Cartesian coordinate system with the $x$-$z$ plane parallel to the wall and the $y$-coordinate \new{in the} direction normal to
the wall. The external potential can then be obtained analytically from the
integration of the interactions over the uniform density distribution of wall
particles $\nDensityW$ for $y \leq -\LJdiam$, giving
\begin{align}
\Vext\klamm{y}
&=
\left\{
\begin{array}{ll}
\infty & y \leq 0\\
\frac{2}{3}\pi \alpha_{\text{w}} \LJdiam^3 \klammSquare{\frac{2}{15} \klamm{\frac{\LJdiam}{y + \LJdiam} }^9 - \klamm{\frac{\LJdiam }{y+ \LJdiam}}^3} & y > 0
\end{array}
\right.,
\end{align}
where $\alpha_{\text{w}} = \nDensityW \depthLJW$ is the strength of the wall
potential.
\begin{figure}[ht]
\centering
\includegraphics[width=9cm]{Figure1.eps}
\caption{Plot of the Young contact angle $\theta_{\text{Y}}$ dependence on the strength of the wall attraction $\alpha_{\text{w}}$.
Computations for \new{$\surfaceTensionLV$, $\surfaceTensionWV$ and $\surfaceTensionWL$} are done in a planar geometry, which are then inserted in (\ref{YoungEquation}).
In the bottom left inset $\theta_{\text{Y}}$ is compared \new{to} 2D contact angle measurements $\theta$
which are solved for $y < y_{\text{max}}= 15 \LJdiam$ and with $\theta_{\text{n}} = 90^\circ$ ($\bullet)$, $\theta_{\text{n}} = 120^\circ$ ($+$) and $\theta_{\text{n}} = 40^\circ$ ($\vardiamond$).
The top right inset depicts the contour lines of the density profile for a Cartesian grid ($\theta_{\text{n}} = 90^\circ$) and $\alpha_{\text{w}} \sigma^3 /\depthLJ = 0.55$ \new{(giving $\theta_{\text{Y}}=134.2^\circ$)} for $y_{\text{max}}=15 \sigma$.}
\label{fig:ContactAngleMeasurements90}
\end{figure}
\section{Numerical Method \label{sec:NumericalMethod}}
To solve (\ref{eq:EulerLagrangeEquation}) numerically in a 2D domain, we
employ a spectral collocation method~\cite{Trefethen_2000}. We have used this
method successfully in our previous studies with both DFT-LDA and DFT-FMT
(e.g.~\cite{PeterPRE,Peter2012,Nold:2014:FluidStructure}). It should be
emphasized that because the equations we wish to solve are non-local, the
resulting matrices following discretization are dense, however the advantage
with the spectral collocation method is that through a convenient choice of
collocation points their number may be kept relatively low, leading to
significant reduction in the size of the matrices. The reduction in the
number of points becomes increasingly important when going to higher
dimensions (as the number of points in a product grid scales exponentially
with the dimension).
Consider the tensor product of two one-dimensional (1D) Chebychev grids on
the box $(\xi,\eta) \in [-1,1]\times [-1,1]$. This computational domain is
mapped onto the half space $[-\infty,\infty]\times[0,\infty]$ by
\begin{align}
x' = L_1 \frac{\xi}{\sqrt{1-\xi^2}}
,\qquad
y' = L_2 \frac{1+\eta}{1-\eta}. \label{eq:CartesianMap}
\end{align}
Here, $L_1$ and $L_2$ are numerical parameters determining the spatial resolution of the collocation points close to $x'=0$ and in the vicinity of the wall, respectively.
This Cartesian grid in the physical half-space is then skewed by an angle $\theta_n$ using the map
\begin{align}
x = \frac{x'}{\sin \theta_n} + y' \cot \theta_n
, \qquad
y = y'. \label{eq:SkewedMap}
\end{align} The skewed grid allows us to have more discretization points near the
fluid-fluid and fluid-solid interface where higher density gradients are
expected. In our computations, we assume that the liquid-vapour interface is
at an angle of $\theta_n$ for values $y \geq y_{\text{max}}$, and only solve
for collocation points located at $y < y_{\text{max}}$, such that the resulting density profiles may only be interpreted for $y < y_{\text{max}}$.
In order to minimize the numerical inaccuracy caused by this cut-off, we iteratively adapt $\theta_n$ and increase
$y_{\text{max}}$ to obtain a final result which is fully physically interpretable.
Physically, the contact angle of a liquid wedge is uniquely defined through
the surface tensions of the liquid-\new{vapour} phase\new{,
$\surfaceTensionLV$,} and the wall-fluid pair \new{($\surfaceTensionWV$ and
$\surfaceTensionWL$ being wall-vapour and wall-liquid surface tensions,
respectively)}, given by
\new{the Young} equation
\begin{align}
\surfaceTensionLV \cos \theta_{\text{Y}} = \surfaceTensionWV - \surfaceTensionWL, \label{YoungEquation}
\end{align}
where the surface tensions are quantities that can be extracted from
planar/\new{(1D)} DFT computations and $\theta_{\text{Y}}$ is defined as the
Young contact angle. Given that we restrict our attention to systems at
temperature $T/T_c = 0.75$, the only parameter on which
\new{$\theta_{\text{Y}}$} depends is the strength of the wall attraction
$\alpha_{\text{w}}$. In figure \ref{fig:ContactAngleMeasurements90}, we plot
the dependence of
\new{$\theta_{\text{Y}}$} on the wall attraction. As expected
intuitively, the contact angle decreases with increasing wall-fluid
attraction and reaches complete wetting at the critical value of
$\alpha_{\text{w,crit}} \sigma^3/\depthLJ = 1.50$. In 2D computations, the
contact angle of the liquid-vapour interface has to converge to
\new{$\theta_{\text{Y}}$} at large distances from the wall.
To check this, we have performed computations on a Cartesian grid, employing
(\ref{eq:CartesianMap}) \newB{and (\ref{eq:SkewedMap}) with $\theta_n =
90^\circ$}, and assuming that above a limiting value $y_{\text{max}}$, the
density at the collocation points corresponds to an equilibrium liquid-vapour
interface with a $90^\circ$ contact angle. The result of the density profile
for such a computation is depicted in the top right inset of figure
\ref{fig:ContactAngleMeasurements90}. By measuring the slope of the
isodensity line for $\nDensity = (\nDensityL + \nDensityV)/2$ in the interval
$y \in [10\sigma,14\sigma]$, we obtain an estimate for the contact angle in a
2D setting. The deviations to the
\new{$\theta_{\text{Y}}$} are shown in the bottom left inset of figure
\ref{fig:ContactAngleMeasurements90}, showing very good agreement.
We have also performed computations on skewed grids, to increase the number
of collocation points in the vicinity of the contact line and the
liquid-vapour interface, by assuming that the liquid-vapour interface is at
an angle of $\theta_n$ for values $y > y_{\text{max}}$. This allowed us to
increase the value of $y_{\text{max}}$ to higher values. The corresponding
behaviour is shown in figure \ref{fig:CA}, where for a wall attraction of
$\alpha_w \sigma^3 / \depthLJ=0.55$ \newB{corresponding to $\thYoung =
134.14^\circ$}, the numerical parameters $y_{\text{max}}$ and $\theta_n$ are varied.
It is seen that for all values of $y_{\max}$ and $\theta_n$ the contact angle approaches
$\theta_{\text{Y}}$ for increasing $y$, before converging to $\theta =
\theta_n$ near $y=y_{\max}$ due to the imposed boundary condition.
For reference, the principal results presented in figure~\ref{fig:DensitySlices135} were computed on
a grid with $45 \times 75$ collocation points and parameters $y_{\text{max}} = 35\sigma$ and
$\theta_n = \{135^\circ,120^\circ,90^\circ,60^\circ,40^\circ\}$ for the different rows, respectively.
\begin{figure}[ht]
\centering
\includegraphics[width=9cm]{Figure2.eps}
\caption{Slope of the isodensity line for $\nDensity = \klamm{\nDensityV + \nDensityL}/2$ for $y_{\max} = \{20,25,35\}$,
represented by the dotted, dash-dotted and solid lines, respectively.
Computations are done on a skewed grid with $\theta_n = 135^\circ$,
$134^\circ$ and $134.2^\circ$, represented by horizontal dashed lines, and results for which are
drawn with magenta, blue and black lines,
respectively. The substrate strength is
$\alpha_w \sigma^3/\depthLJ = 0.55$, leading to \new{$\theta_{\text{Y}} =
134.14^\circ$}, depicted by the red horizontal line. The inset
shows a typical contour plot for the density,
where the contour lines correspond to number densities $\klamm{\nDensity -
\nDensityV}/\klamm{\nDensityL - \nDensityV} = \{0.05,0.5,0.95\}$ from left to
right, respectively.}
\label{fig:CA}
\end{figure}
\section{Fluid structure in the vicinity of the contact line \label{sec:DensityStructure}}
Figure \ref{fig:DensitySlices135} reveals the density structure for a fluid
in the vicinity of the contact line for different wall strengths. It can be
seen that depending on the wall strength parameter $\alpha_{\text{w}}$, the
contact density at the wall for the wall-liquid interface changes
significantly. In particular, we have checked the consistency of the observed
behaviour with the wall-fluid virial equation \cite{Herring:2010vn}
\begin{align}
p
= - \int_{-\infty}^\infty \nDensity(y) V'_{\text{ext}}(y) \dI y
= \nDensity(0) - \int_{0}^\infty \nDensity(y) V'_{\text{ext}}(y) \dI y,
\end{align}
where $\nDensity(0)$ stems from the delta-function contribution to $V'_{\text{ext}}$ at $y=0$.
\begin{figure}
\centering
\includegraphics[width=15cm]{Figure3a.eps}
\caption*{For caption, see next page.}
\notag
\end{figure}
\begin{figure}
\centering
\includegraphics[width=15cm]{Figure3b.eps}
\caption{Contour lines for the number density (left column, \new{subfigures I}) and density profiles as \new{a} function of the distance to the
substrate at various positions $x$ along the substrate (right column,
\new{subfigures II}).
In the left column, the contour lines correspond to number densities $\klamm{\nDensity - \nDensityV}/\klamm{\nDensityL - \nDensityV} = \{0.05,0.5,0.95\}$ from left to right.
The height profiles $h_{\text{I,II,III}}$, defined through equation (\ref{eq:MinimizingHamiltonian:Condition}) with boundary conditions (\ref{eq:h:BC1})-(\ref{eq:h:BC2}) and equation (\ref{eq:Def_HIII}),
are depicted by black dash-dotted, dashed and solid lines, respectively, $h_{\text{I}}$ being virtually indistinguishable from $h_{\text{III}}$.
The solid lines in the right column represent the 2D density profile, plotted along the dashed vertical lines of corresponding colour in the left column figures.
These density profiles are compared to the equivalent planar off-saturation liquid or vapour film of the same adsorption film thickness, drawn with dashed lines.
\label{fig:DensitySlices135}
}
\end{figure}
The density plots at different positions in $x$ across the contact line in
the right column of figure \ref{fig:DensitySlices135} provide an insight as
to how the transition between a wall-vapour and a wall-liquid interface leads
to a quasi step-like increase of the density in the contour plots. We note
that this transition is accompanied by a gradual increase of the distance
between the liquid-vapour interface \newB{and} the wall. A similar transition
can be observed when gradually varying the chemical potential for a fluid
film in contact with a planar wall. A typical example \newB{of} the
bifurcation diagram, also widely denoted as the adsorption isotherm,
representing this transition, is shown in figure
\ref{fig:AdsorptionIsotherm41}, where the film thickness $\filmThickness$ of
the liquid or vapour film, defined by
\begin{align}
\filmThickness \defi& \frac{1}{\Delta \nDensity} \int_0^\infty |\nDensity(\infty) - \nDensity(y)| \dI y \label{eq:AdsorptionFilmThickness:Def}\\
\qquad\text{with} \qquad \Delta n =& \nDensityL - \nDensityV,
\end{align}
is plotted versus the deviation of the chemical potential from its saturation
value $\Delta \chemPot$. In particular, figure \ref{fig:AdsorptionIsotherm41} shows the
behaviour for a dewetting scenario of a growing vapour film. Each point on the adsorption isotherm represents a density
profile which satisfies the Euler-Lagrange equation
(\ref{eq:EulerLagrangeEquation}) for a planar configuration.
As saturation is approached, the adsorption isotherm satisfies the expected inverse cubic decay of $\Delta \chemPot$ with $\filmThickness$
for systems with dispersion forces~\cite{DietrichNap:1991}, such as shown in the inset of figure \ref{fig:AdsorptionIsotherm41}.
These density profiles are compared in the right column of figure
\ref{fig:DensitySlices135} with density profiles across the contact line
which have the same adsorption (\ref{eq:AdsorptionFilmThickness:Def}). We
note that the contact line is computed at saturation chemical potential,
whereas the chemical potential for the density profiles of the adsorption
isotherm is naturally off-saturation. Nevertheless, the result is unexpected
and shows a surprisingly good agreement, where for large film thicknesses,
the density profiles at the liquid-vapour interfaces differ because for a
contact line the liquid-vapour interface is at an angle to the wall, while
the dashed lines always describe planar films.
\begin{figure}[ht]
\centering
\includegraphics[width=8cm]{Figure4.eps}
\caption{Plot of the adsorption isotherm for a dewetting scenario with wall attraction \new{of} $\alpha_{\text{w}} \LJdiam^3/\depthLJ = 0.7$,
corresponding to $\theta_{\text{Y}} = 119.9^\circ$.
The inset shows the asymptotic behaviour
for \newB{large} $\filmThickness$, as $\Delta \mu \sim \filmThickness^{-3}$, \new{where the} dashed line is a fit
for $\filmThickness \in [10 \LJdiam,15 \LJdiam]$ to $\Delta \mu = a
\filmThickness^{-3}$, \new{with computed} coefficient $a = -1.21 \depthLJ \LJdiam^3$. \new{In the inset} individual
DFT computations of the equilibrium density \new{are marked with circles and are connected by the solid line \newB{in the main plot} for clarity.}}
\label{fig:AdsorptionIsotherm41}
\end{figure}
\section{Hamiltonian approaches, Derjaguin-Frumkin route and disjoining pressure \label{sec:HamiltonianApproaches}}
In a coarse-grained description of the contact line, the two-dimensional density profile is reduced to a height profile $h(x)$
representing the liquid-vapour interface~\cite{Lipowsky:1987,Mikheev:1991pi}. At equilibrium, this height profile minimizes the Hamiltonian \cite{Herring:2010vn}
\begin{align}
H[h] = \int_{-\infty}^\infty \klammCurl{ \surfaceTensionLV \klamm{ \sqrt{ 1+ (h')^2 } -1 } + V(h) } \dI x, \label{eq:Hamiltonian_h}
\end{align}
where $h' = {\dI h}/{\dI x}$ is the slope of the interface and $V(h)$ is the effective interface potential.
The first term in (\ref{eq:Hamiltonian_h}) accounts for the excess energy stored through the surface tension due to the curvature of the liquid-vapour interface, while
the second term accounts for corrections due to the presence of the substrate. This not only includes direct attractive forces between fluid and wall particles, but also
corrections due to the distorted fluid density profile caused by the presence of the wall.
The effective interface potential $V$ is linked to the disjoining potential $\DisjoiningPressure$ by
\begin{align}
\DisjoiningPressure\klamm{h} \defi - \frac{\dI V}{\dI h}. \label{eq:DisjoiningPressureVh}
\end{align}
Usually, (\ref{eq:Hamiltonian_h}) is only applied in the lubrication approximation. For larger slopes, both
the separate inclusion of the effective surface potential and surface energy~\cite{Pismen:2001fk}
as well as the functional dependence of $V$ on $h$ alone, as opposed to a functional dependence on $h(x)$, were put into question \cite{Henderson:2011:EPJST:DisjoiningPressure,MacDowell:2011:ResponseEPJST,Henderson:EPJST:ResponseMacDowell,Henderson:NoteContinuingContactLine}. Here, we test
for different disjoining pressure definitions whether (\ref{eq:Hamiltonian_h}) may be used to to define height profiles for a large range of contact angles.
In \cite{Nold:2014:FluidStructure} we have compared height profiles resulting
from minimizing (\ref{eq:Hamiltonian_h}) with two different definitions of
the disjoining pressure for contact angles $\theta < 90^\circ$. We note that
these disjoining pressure definitions are different from phenomenological
analytical models such as used e.g. in \cite{Schwartz:1998,Sibley:JengMath:2014} in that they are
obtained directly from DFT computations, and therefore include the full
information of hard-sphere as well as the attractive particle interactions.
The first disjoining pressure definition we consider is based on the
celebrated Derjaguin and Frumkin theory \cite{Derjaguin:1936,Frumkin1938I}:
\begin{align}
\DisjoiningPressure_{\text{I}}\klamm{\filmThickness} \defi - \Delta \chemPot \Delta \nDensity \times
\left\{
\begin{array}{ll}
1 & \text{for } \nDensity|_{y=\infty} = \nDensityV\\
-1 & \text{for } \nDensity|_{y=\infty} = \nDensityL
\end{array}
\right. , \label{eq:DisjoiningPressureAdsorptionIsotherm}
\end{align}
for a system at saturation chemical potential $\chemPot_{\text{sat}}$. $\chemPotN_{\text{eq}}$ is the chemical potential at which a film of thickness $\filmThickness$ is at equilibrium, such as depicted in the adsorption isotherm in figure \ref{fig:AdsorptionIsotherm41}, and where
\begin{align}
\Delta \chemPot &= \chemPotN_{\text{eq}}\klamm{\filmThickness} - \chemPotN_{\text{sat}}.
\end{align}
The first case of (\ref{eq:DisjoiningPressureAdsorptionIsotherm}), \newB{$\nDensity|_{y=\infty} = \nDensityV$}, describes a
wetting scenario where the density at infinite distance from the wall
corresponds to the equilibrium vapour density. In this case, a liquid film
will slowly build as the chemical potential reaches its saturation value. In
contrast, the dewetting case \newB{$\nDensity|_{y=\infty} = \nDensityL$}
describes a vapour film in a bulk \new{liquid} environment, as in
figure \ref{fig:AdsorptionIsotherm41}. The sign switch in
(\ref{eq:DisjoiningPressureAdsorptionIsotherm}) originates from the sign
difference between the density in the film vs. the bulk density. We note that
contact lines with contact angle $\thYoung > 90^\circ$ are described by a
vapour film of varying height, whereas contact lines with contact angle
\newB{$\thYoung < 90^\circ$} are described by a liquid film of varying
height.
\new{As an alternative} to the Derjaguin and Frumkin definition of the disjoining pressure (\ref{eq:DisjoiningPressureAdsorptionIsotherm}), one can define the disjoining pressure based on the normal force balance at the substrate. The disjoining pressure is then defined as the excess pressure acting on the substrate due to the deviation from the equilibrium density profile, caused e.g.\ by the boundary conditions imposed on the system~\cite{Herring:2010vn,Henderson:2011:EPJST:DisjoiningPressure}
\begin{align}
\DisjoiningPressure_{\text{II}}\klamm{x} \defi - \int_{-\infty}^\infty \klamm{ \nDensity(x,y) - \nDensity(\infty,y) } V_{\text{ext}}'(y) \dI y. \label{eq:SumRuleDisjoiningPressue2D}
\end{align}
Note that $\nDensity(x,y) V_{\text{ext}}'(y)$ is the force acting through the
external potential---representing the wall---on the fluid element at point
$(x,y)$. In our case, $\nDensity(x,y)$ is the density profile originating
from a 2D DFT computation of the contact line, and hence
$\DisjoiningPressure_{\text{II}}$ is a quantity containing information of the
full 2D equilibrium density profile; in contrast
(\ref{eq:DisjoiningPressureAdsorptionIsotherm}) is derived from planar 1D
computations.
The equilibrium height profiles $h_{\text{I}}$ and $h_{\text{II}}$
corresponding to the disjoining pressures $\DisjoiningPressure_{\text{I}}$
and $\DisjoiningPressure_{\text{II}}$, respectively, are obtained by
minimizing the Hamiltonian (\ref{eq:Hamiltonian_h}), leading to the defining
equation for $h_{\text{I/II}}$
\begin{align}
- \DisjoiningPressure_{\text{I}/\text{II}} = \surfaceTensionLV \frac{\dI}{\dI x} \klamm{ \frac{h_{\text{I}/\text{II}}'}{\sqrt{1+ (h_{\text{I}/\text{II}}')^2}} }, \label{eq:MinimizingHamiltonian:Condition}
\end{align}
with boundary conditions
\begin{align}
\lim_{x\to -\infty } h_{\text{I}} = h_0 \qquad \text{for } \thYoung < 90^\circ, \label{eq:h:BC1}
\end{align}
and
\begin{align}
\lim_{x\to \infty } h_{\text{I}} = h_0 \qquad \text{for } \thYoung > 90^\circ, \label{eq:h:BC2}
\end{align}
where $h_0$ is the film thickness representing the wall-vapour interface in
the wetting case and the wall-liquid interface in the drying case. We note
that $h_0$ corresponds to the (finite) value at $\Delta \chemPot = 0$ of the
adsorption isotherm in \new{figure \ref{fig:AdsorptionIsotherm41}}. Given
that $\DisjoiningPressure_{\text{I}}$ is a function of \new{$h$, and not of
$x$ directly}, (\ref{eq:MinimizingHamiltonian:Condition}) for $h_{\text{I}}$
is an autonomous ordinary differential equation. This means that with
(\ref{eq:h:BC1}),
\newB{(\ref{eq:h:BC2})} $h_{\text{I}}$ is translationally invariant in $x$.
For simplicity, in figures \ref{fig:DensitySlices135} and
\ref{fig:DisjoiningPressures}, we depict one plot for $h_{\text{I}}$ or
$\DisjoiningPressure_{\text{I}}(h_{\text{I}})$.
The ordinary differential equation (\ref{eq:MinimizingHamiltonian:Condition})
defining the film heights $h_{\text{I/II}}$ can also be interpreted as a form
of the Young-Laplace equation for a pressure jump across a fluid interface,
where the left hand side describes the difference between the pressure acting
on the substrate and the fluid pressure at $y=\infty$, while the right hand
side represents the product of the surface tension with the curvature of the
interface.
Integrating (\ref{eq:MinimizingHamiltonian:Condition}) with respect to $x$ and $h$, respectively, leads to the normal-force balance of Young's equation
\begin{align}
-\int_{-\infty}^\infty \DisjoiningPressure_{\text{I}/\text{II}}(x) \dI x = \surfaceTensionLV \sin \theta_{\text{Y},\text{I}/\text{II}}, \label{eq:SumRuleNormalForceBalance}
\end{align}
and the important expression of Derjaguin-Frumkin theory~\cite{derjaguin1986properties,Schwartz:1998}
\begin{align}
- \int_{h_0}^\infty \DisjoiningPressure_{\text{I}/\text{II}}\klamm{h} \dI h = \surfaceTensionLV \klamm{1 - |\cos \theta_{\text{Y},\text{I}/\text{II}}|}, \label{eq:FrumkinResultDisjoiningPressure1}
\end{align}
where $\theta_{\text{Y},\text{I}/\text{II}} \in [0,180^\circ]$ corresponds to the limiting slope of the height profiles $h_{\text{I,II}}$, respectively, at distances far away from the wall:
\begin{align}
\theta_{\text{Y},\text{I}/\text{II}} = \lim_{h_{\text{I/II}} \to \infty} \tan^{-1} \klamm{ h'_{\text{I/II}}(x)}. \label{eq:definethetaI_II}
\end{align}
Equation (\ref{eq:FrumkinResultDisjoiningPressure1}) can be interpreted as a
force balance in direction parallel to the substrate. For $\thYoung<
90^\circ$, the right hand side of the equation represents the forces of the
liquid-vapour interface acting in the negative $x$-direction. For $\thYoung >
90^\circ$, the height profile decreases from $\infty$ to $h_0$ as $x$ increases. Due to this
inversion of the height profile, (\ref{eq:FrumkinResultDisjoiningPressure1})
represents the force balance in the positive $x$-direction. The force of the
liquid-vapour interface acting in the positive $x$-direction is
$\surfaceTensionLV$, whereas the force acting in the negative direction is
$\surfaceTensionLV |\cos \thYoung|$. We note that here, the modulus accounts
for the fact that for $\thYoung > 90^\circ$, $\cos \thYoung < 0$, given that
we have defined $\theta_{\text{Y},\text{I}/\text{II}} \in [0,180^\circ]$, as
opposed to allowing for negative values of $\theta_{\text{Y,I/II}}$ in
(\ref{eq:definethetaI_II}).
Since both sum rules are derived from (\ref{eq:MinimizingHamiltonian:Condition}),
$\theta_{\text{Y},\text{I}/\text{II}}$ in equations
(\ref{eq:SumRuleNormalForceBalance}) and
(\ref{eq:FrumkinResultDisjoiningPressure1}) are equivalent and ultimately,
both height profiles converge to the slope dictated by the Young contact
angle. Thus $\theta_{\text{Y},\text{I}/\text{II}}$ both correspond to $\thYoung$
defined in the Young equation (\ref{YoungEquation}). We will exploit this
property to estimate the accuracy of our numerical method.
\begin{table}[ht]
\centering
\begin{tabular}{llllll}
\toprule
$\alpha_w \sigma^3/\varepsilon$ & $\theta_{\text{Y}}$ & $-\int_{h_0}^\infty \DisjoiningPressure_{\text{I}}\klamm{h} \dI h$ & $\theta_{\text{Y,I}}$ &
$-\int_{-\infty}^\infty \DisjoiningPressure_{\text{II}}\klamm{x} \dI x$
&$\theta_{\text{Y,II}}$\\\midrule
$0.55$ & $134.2^\circ \pm 0.1^\circ$ & $0.103\pm 0.002$ & $134.5^\circ \pm 0.4^\circ$ & $-0.244 \pm 0.005$ & $135.2^\circ \pm 1.1^\circ$\\
$0.7$ & $119.9^\circ\pm 0.05^\circ$ & $0.172\pm 0.003$ & $120.3^\circ \pm 0.5^\circ$ & $-0.298 \pm 0.002$ & $120.5^\circ \pm 0.7^\circ$ \\
$1.0$ & $89.6^\circ \pm 0.1^\circ$ & $0.345 \pm 0.001$ & $89.8^\circ \pm 0.2^\circ$ & $-0.3463 \pm 10^{-4}$ & ($\star$) \\
$1.25$ & $59.9^\circ \pm 0.1^\circ$ & $0.173 \pm 0.001$ & $60.0^\circ \pm 0.2^\circ$ & $-0.297 \pm 0.003$ & $59.1^\circ \pm 0.8^\circ$\\
$1.375$ & $41.0^\circ\pm 0.1^\circ$ & $0.085 \pm 0.001$ & $41.1^\circ \pm 0.2^\circ$ & $-0.234 \pm 0.007$ & $42.5^\circ \pm 1.6^\circ$\\\bottomrule
\end{tabular}
\caption{Comparison of $\theta_{\text{Y}}$ as defined in (\ref{YoungEquation}), the contact angles $\theta_{\text{I,II}}$ defined through (\ref{eq:definethetaI_II}) as well as the absolute errors of the integrals on the left hand sides of equations (\ref{eq:SumRuleNormalForceBalance}) and (\ref{eq:FrumkinResultDisjoiningPressure1}), respectively. ($\star$): Here, the integral expression gives $\sin \theta_{\text{Y,II}} = 1.0001 \pm 0.0001$, such that an estimate for $\theta_{\text{Y,II}}$ cannot formally be given.}
\label{tab:thetaY_I_II_Comparison}
\end{table}
In table \ref{tab:thetaY_I_II_Comparison}, numerical values for the integrals
of the disjoining pressures are given. Error bounds $\Delta$ are estimated by
comparing the integral expressions with $\surfaceTensionLV \sin \thYoung$ and
$\surfaceTensionLV \klamm{1 - |\cos \thYoung|}$, respectively. These error
bounds are then used to estimate error bounds of $\theta_{\text{Y,I/II}}$ by
\begin{align}
\Delta \theta_{\text{Y,II}} = \left| \frac{\Delta \klammCurl{-\int_{-\infty}^\infty \DisjoiningPressure_{\text{II}}(x) \dI x}}{\surfaceTensionLV \cos \theta_{\text{Y},\text{II}}} \right|
\qquad \text{and} \qquad
\Delta \theta_{\text{Y,I}} = \left|\frac{\Delta \klammCurl{- \int_{h_0}^\infty \DisjoiningPressure_{\text{I}}\klamm{h} \dI h}}{ \surfaceTensionLV \sin \theta_{\text{Y},\text{I}} }\right|.
\end{align}
The above formulations can be derived from
(\ref{eq:SumRuleNormalForceBalance}) and
(\ref{eq:FrumkinResultDisjoiningPressure1}) by using
$\theta_{\text{Y},\text{I/II}} + \Delta \theta_{\text{Y},\text{I/II}}$ and
linearly expanding to first order in $\Delta \theta_{\text{Y},\text{I/II}}$
the right hand side of the respective equation around
$\theta_{\text{Y},\text{I/II}}$. Finally, we compare the film height profiles
$h_{\text{I}}$ and $h_{\text{II}}$ with the adsorption film thickness
\begin{align}
h_{\text{III}}(x) \defi \frac{1}{\Delta \nDensity} \int_0^\infty |\klamm{\nDensity(x,y) - \nDensity(x,\infty)}| \dI y, \label{eq:Def_HIII}
\end{align}
\new{which is the 2D generalisation of (\ref{eq:AdsorptionFilmThickness:Def}).}
This allows us to define a disjoining pressure suggested by the adsorption film height, obtained by inserting $h_{\text{III}}$ into (\ref{eq:MinimizingHamiltonian:Condition}), giving the rescaled curvature
\begin{align}
- \DisjoiningPressure_{\text{III}}(h) \defi \surfaceTensionLV \frac{\dI}{\dI x} \klamm{ \frac{h_{\text{III}}'}{\sqrt{1+ (h_{\text{III}}')^2}} }.
\end{align}
\section{Discussion and conclusion \label{sec:Conclusion}}
We have scrutinized the fluid structure and its properties in the vicinity of
a three-phase contact line by employing a DFT-FMT model. In particular, we
presented density profiles slice by slice as we sweep through the contact
line region and we contrast the density profiles with the profile of a planar
liquid film on a substrate, but with the same film thickness, demonstrating
that the two are quite similar. We also scrutinized the ability of
Derjaguin-Frumkin theory~\cite{Derjaguin:1987:50YearsOfSurfaFceScience} for
planar liquid films on a substrate to predict the height profile at the
contact line and we offered a unified Derjaguin-Frumkin treatment of the
contact line for $\thYoung < 90^\circ$ and $\thYoung > 90^\circ$ by
appropriately extending the boundary conditions for the disjoining pressure
equation to account for the case $\thYoung > 90^\circ$.
In figure \ref{fig:DensitySlices135} we plot the height profiles
$h_{\text{I/II/III}}$ for contact angles in the region $40^\circ < \thYoung <
135^\circ$ and compare them with the contour lines of the density. The figure summarizes
some of the main results of our study as far as the behaviour close to the
contact line is concerned. Additional information on this can be extracted
from figure \ref{fig:DisjoiningPressures} where we compare the disjoining
pressure profiles $\DisjoiningPressure_{\text{I/II/III}}$. An observation we
made in our previous study in~\cite{Nold:2014:FluidStructure} for contact
angles $\thYoung <90^\circ$, was that the location of maximal curvature for
the height profile $h_{\text{II}}$ is shifted towards the fluid phase if
compared with the adsorption height profile $h_{\text{III}}$. This
observation can also be made in figures \ref{fig:DensitySlices135}
\new{(g,i)} and in figure \ref{fig:DisjoiningPressures} \new{(b)}. However,
this does not occur to the same extent in cases where $\thYoung >90^\circ$---such as
observed in figures \ref{fig:DensitySlices135} (a,c) and in figure
\ref{fig:DisjoiningPressures}
\new{(a)}.
\begin{figure}[ht]
\centering
\includegraphics[width=14cm]{Figure5.eps}
\caption{Plots of different disjoining pressure definitions for different wall attractions. \new{Dash-dotted, dashed and solid} lines depict disjoining pressures
\new{$\DisjoiningPressure_{\text{I}}$, $\DisjoiningPressure_{\text{II}}$ and $\DisjoiningPressure_{\text{III}}$}, respectively.
In \new{subfigure (a)}, the black and green lines show data for $\alpha_w \sigma^3/\varepsilon = 0.55$ and $0.7$, respectively, \new{whilst in (b)},
the black, green and magenta lines show data for $\alpha_w \sigma^3/\varepsilon = 1.375, 1.25$ and $1.0$, respectively.}
\label{fig:DisjoiningPressures}
\end{figure}
Furthermore, the maximal absolute curvature of the height profile
$h_{\text{II}}$ \newB{(see dashed lines in figures \ref{fig:DensitySlices135}
and \ref{fig:DisjoiningPressures})} is lower than the maximal absolute
curvature of the adsorption film height $h_{\text{III}}$ \newB{(see solid
lines in figures \ref{fig:DensitySlices135} and
\ref{fig:DisjoiningPressures})}. This can best be seen in figure
\ref{fig:DisjoiningPressures} \newB{(we note that the disjoining pressure
corresponds to the rescaled curvature of the corresponding height profile)}.
While the difference is less pronounced for large contact angles $\thYoung >
90^\circ$, it is still observable. In contrast, the film thickness
$h_{\text{I}}$ \newB{(see dash-dotted lines in figures
\ref{fig:DensitySlices135} and \ref{fig:DisjoiningPressures})} based on the
adsorption isotherm, agrees very well with $h_{\text{III}}$, often to the
point of being virtually indistinguishable (compare the left column of figure
\ref{fig:DensitySlices135}).
It should be noted that for a varying height profile, here enforced by the
boundary conditions, there exist conflicting definitions of the disjoining
pressure---one based on the adsorption isotherm, the other based on the
normal force balance. These two definitions lead to distinct height profiles,
which suggest that the use of the disjoining pressure based on the adsorption
isotherm is more appropriate, given the good agreement of the corresponding
height profile with the adsorption height profile. This is \new{somewhat}
surprising, given that the disjoining pressure based on the normal force
balance $\DisjoiningPressure_{\text{II}}$ contains information from the full
equilibrium 2D density profile, whereas $\DisjoiningPressure_{\text{I}}$ is
derived from purely 1D computations.
At the same time the behaviour of $\DisjoiningPressure_{\text{II}}$ is such
that the maximum absolute normal pressure acting on the substrate is lower
than the curvature of the adsorption height profile would suggest. Also, for
$\thYoung < 90^\circ$, the maximal normal pressure does {\it not} act in the
vicinity of the contact line, but instead at a slightly shifted position
towards the liquid phase. This interpretation could be of interest
for the nanoscale behaviour of contact lines at soft substrates, such as
considered e.g.\ by Lubbers \etal~\cite{DubbersSnoeijer:2014:SoftSolids}.
The special case of $\thYoung$ being very close to $90^\circ$, such as
depicted in figure \ref{fig:DensitySlices135} \new{(e,f)} for $\alpha_w
\sigma^3/\varepsilon = 1.0$, as well as the magenta lines in figure
\ref{fig:DisjoiningPressures} \new{(b)}, deserves a comment. In this case,
the density at very large distances from the wall $\nDensity|_{y \to \infty}$
depends on the position $x$, and hence does not allow for the definition of
an adsorption height profile $h_{\text{III}}$ through (\ref{eq:Def_HIII}).
While the disjoining pressure $\DisjoiningPressure_{\text{I}}$ based on the
adsorption isotherm has a very high absolute maximum, the absolute maximum of
$\DisjoiningPressure_{\text{II}}$ is less pronounced. Also, \new{the} width
of $\DisjoiningPressure_{\text{II}}$ corresponds roughly to the width of the
interface and is slightly shifted towards the fluid phase.
An important observation, therefore, is that the maximal normal pressure
acting on the substrate does not correspond with the maximal curvature of the
adsorption film thickness or the maximal value of the Derjaguin-Frumkin
disjoining pressure $\DisjoiningPressure_{\text{I}}$. One reason for the
softening of the normal pressure profile could be the width of the fluid
interface. In particular, one can observe in figure
\ref{fig:DisjoiningPressures} (b), that the width of
$\DisjoiningPressure_{\text{II}}$ for $\thYoung \approx 90^\circ$, denoted by
the dashed magenta line, corresponds approximately with the width of the
liquid-vapour interface.
It is noteworthy that the main limitation of the model is that its mean-field
nature does not include the description of thermal
fluctuations~\cite{ArcherEvans:2013,EvansHendersonHoyle:1993,MacDowellBenet:2014}.
Inclusion of thermal fluctuations, which become more pronounced with increasing
film thicknesses $\filmThickness$, lead to a broadening of the liquid-vapour interface
and a renormalization of the dependence of $\filmThickness$ on the chemical potential deviation from saturation $\Delta \chemPot$~\cite{ArcherEvans:2013} is needed.
A detailed recent study based on molecular simulations and experiments
has found that thermal fluctuations lead to an effective film-height dependent surface tension $\surfaceTensionLV(\filmThickness)$ in (\ref{eq:Hamiltonian_h})~\cite{MacDowellBenet:2014}.
A final conclusion about the effect on thermal fluctuations for the results presented here could be reached by a molecular simulations study in the spirit of Herring and Henderson's analysis~\cite{Herring:2010vn},
but including dispersion forces and a comparison with the corresponding Derjaguin-Frumkin disjoining pressure. This, however, is beyond the scope of the present study.
The important observation made here is that in a mean-field model, disjoining pressures
obtained from planar films via the Derjaguin-Frumkin route do allow us to
predict with good accuracy the structure of the contact line, hence implying
a small contribution of non-locality. It would be interesting to see if this
holds for other settings, e.g. spherical droplets.
Of particular interest would also be to investigate very large contact angles
close to $180^\circ$, given interesting recent results in this
case~\cite{benilov2013contact} as well as the influence of surface roughness
and chemical heterogeneities which are known to influence wetting phenomena
substantially
(e.g.~\cite{PhysFluids_21_2009,Savva2011,PhysRevLett_104_2010,Raj2011}. We
shall address these and related issues in future studies.
\vspace*{0.5cm}
\section*{Acknowledgements}
We acknowledge financial support from ERC Advanced Grant No. 247031 and Imperial College through a DTG International Studentship.
|
\section{Introduction}
The interaction between photons and acoustic phonons has been investigated in bulk materials since the 1920s \cite{Boyd2008,Agrawal2013}. In case the phonons are generated by optical forces, the interaction is often called stimulated Brillouin scattering (SBS) -- a feedback loop in which energy flows from the optical waves to the mechanical oscillator. Its signature is narrowband amplification of an optical probe that is red-detuned by the phonon resonance frequency from a strong optical pump.
Although the mechanical linewidth does not exceed 100 MHz typically, there is no such inherent optical bandwidth restriction. Compared to cavity-based optomechanics \cite{Carmon2005,Aspelmeyer2014,Cohen2015}, a circuit-oriented approach is intrinsically less power-efficient as the optical field is not resonantly enhanced. Nevertheless, the removal of the bandwidth restriction and the accompanying optical versatility has motivated a great deal of SBS work in small-core waveguides, from photonic crystal \cite{Dainese2006a,Kang2009,Kang2011}, dual-web \cite{Butsch2014} and subwavelength \cite{Beugnot2014} fibres to chalcogenide \cite{Eggleton2013,Kabakova2013,Marpaung2015} and silicon waveguides \cite{Shin2013b,Shin2015,VanLaer2015}. It may provide new integrated signal processing capabilities such as tunable RF notch filters \cite{Casas-bedoya2015} and true time delays \cite{Zhu2007b}. The prospect is especially appealing in silicon photonic wires, whose strong confinement enhances the light-matter coupling. Mass-manufacturable silicon-on-insulator chips are therefore an exciting platform for high-density optomechanical circuitry, perhaps even at the quantum level \cite{Habraken2012,Safavi-naeini2014,Safavi-Naeini2011e}.
Recent work on this front has demonstrated promising photon-phonon coupling efficiencies in all-silicon waveguides \cite{VanLaer2015}. The coupling was sufficiently strong to bring the waveguides into transparency, but phonon leakage and free-carrier absorption precluded actual amplification above the optical propagation loss. Here, we eliminate the phonon clamping loss -- observing an increase of the phonon quality factor from 300 up to 1000 at room temperature -- by fully suspending the silicon nanowires. Thus, we achieve a modest amount of gain exceeding the optical losses. The waveguides consist of a series of suspended beams, supported by silicon dioxide anchors (fig.\ref{fig:structure_and_modes}).
Finally, we observed a strong dependence of the phonon quality factor on the number of and distance between the suspensions. This indicates the presence of geometric disorder that broadens and splits the phonon dispersion relation in some cases, similar to Doppler broadening in gas lasers \cite{Saleh1991}. From a wider perspective and not limited to our system, such geometric disorder may hinder development of nanoscale phonon circuits quite generally \cite{Habraken2012,Safavi-naeini2014,Safavi-Naeini2011e,Hatanaka2014}.
\begin{figure*}
\centering
\subfigure{
\begin{tikzpicture}[>=stealth']
\node[anchor=south west,inner sep=0] (image) at (0,0) {\raisebox{0pt}{\includegraphics[scale=0.25]{suspended_wire.png}}};
\begin{scope}[x={(image.south east)},y={(image.north west)}]
\node at (0.60,0.69) {\textcolor{black}{\large{\textbf{Si wire}}}};
\draw[white,->,very thick,decorate,decoration={snake,amplitude=.5mm,segment length=3mm,post length=2mm}] (0.2493,0.857) -- (0.372,0.75);
\draw[white,<->,thin,decorate,decoration={snake,amplitude=.1mm,segment length=0.4mm,pre length = 0.5mm,post length=0.5mm}] (0.457,0.65) -- (0.486,0.67);
\draw[white,<->,thin,decorate,decoration={snake,amplitude=.1mm,segment length=0.4mm,pre length = 0.5mm,post length=0.5mm}] (0.477,0.63) -- (0.508,0.65);
\draw[white,<->,thin,decorate,decoration={snake,amplitude=.1mm,segment length=0.4mm,pre length = 0.5mm,post length=0.5mm}] (0.437,0.67) -- (0.465,0.688);
\node at (-0.06,0.9) {\large{\textbf{a}}};
\node at (0.07,0.32) {\textcolor{black}{\large{\textbf{$\text{SiO}_{2}$}}}};
\draw[white,*->,thick] (0.24,0.04) -- (0.4,0.04);
\begin{scope}[shift={(0.6,0.085)}]
\draw[black,*->,thick,rotate around={21:(0.25,0.36)}] (0.25,0.36) -- (0.35,0.36);
\node at (0.33,0.422) {\textcolor{black}{\large{{${x}$}}}};
\draw[black,->,thick,rotate around={81:(0.263,0.37)}] (0.263,0.37) -- (0.348,0.37);
\node at (0.254,0.45) {\textcolor{black}{\large{{${y}$}}}};
\end{scope}
\node at (-0.06,0.20) {\large{\textbf{b}}};
\node at (0.38,0.067) {\textcolor{white}{\large{{${z}$}}}};
\draw[white,<->,densely dashed,thick,rotate around={-0.22:(0.520,0.14)}] (0.19,0.14) -- (0.85,0.14);
\node at (0.520,0.172) {\textcolor{white}{\normalsize{{$L_{\text{s}}$}}}};
\draw[white,<->,densely dashed,thick,rotate around={-0.22:(0.1365,0.142)}] (0.085,0.142) -- (0.188,0.142);
\node at (0.1365,0.174) {\textcolor{white}{\normalsize{{$L_{\text{a}}$}}}};
\end{scope}
\end{tikzpicture}
}
\subfigure{
\begin{tikzpicture}[>=stealth']
\node[anchor=south west,inner sep=0] (image) at (0.0,0) {\raisebox{0pt}{\includegraphics[scale=0.45]{mechanical_and_optical_mode.png}}};
\begin{scope}[x={(image.south east)},y={(image.north west)}]
\node at (-0.06,0.98) {\large{\textbf{c}}};
\node at (0.528,0.853) {\scriptsize{\textcolor{white}{{$\bf{|E|}^2$}}}};
\node at (0.51,0.66) {\scriptsize{\textcolor{black}{{$u_{x}$}}}};
\draw[green!50!black,<-,solid,thick,rotate around={0:(0.0,0.0)}] (0.1,0.0) -- (0.1,0.3);
\node at (0.05,0.04) {\scriptsize{\textcolor{black}{{$k_{\text{p}}$}}}};
\draw[red!50!black,<-,solid,thick,rotate around={0:(0.0,0.0)}] (0.15,0.07) -- (0.15,0.3);
\node at (0.22,0.11) {\scriptsize{\textcolor{black}{{$k_{\text{pr}}$}}}};
\draw[black,<-,solid,thick,rotate around={0:(0.0,0.0)}] (0.15,0.00) -- (0.15,0.068);
\node at (0.20,0.04) {\scriptsize{\textcolor{black}{{$K$}}}};
\node at (-0.06,0.33) {\large{\textbf{d}}};
\draw[black,-,solid,thick,rotate around={0:(0.0,0.0)}] (0.4,0.00) -- (0.8,0.0);
\draw[black,-,solid,thick,rotate around={0:(0.0,0.0)}] (0.55,0.05) -- (0.8,0.05);
\draw[black,-,densely dashed,thick,rotate around={0:(0.0,0.0)}] (0.4,0.13) -- (0.8,0.13);
\draw[green!50!black,->,solid,thick,rotate around={0:(0.0,0.0)}] (0.467,0.0) -- (0.467,0.127);
\node at (0.418,0.06) {\scriptsize{\textcolor{black}{{$\omega_{\text{p}}$}}}};
\draw[red!50!black,->,solid,thick,rotate around={0:(0.0,0.0)}] (0.670,0.13) -- (0.670,0.052);
\node at (0.75,0.09) {\scriptsize{\textcolor{black}{{$\omega_{\text{pr}}$}}}};
\draw[black,->,solid,thick,rotate around={0:(0.0,0.0)}] (0.670,0.05) -- (0.670,0.002);
\node at (0.73,0.025) {\scriptsize{\textcolor{black}{{$\Omega$}}}};
\node at (0.395,0.27) {\tiny{\textcolor{black}{{pump}}}};
\begin{scope}[shift = {(-0.03,-0.01)}]
\node[rotate around = {25:(0.37,0.355)}] at (0.37,0.355) {\tiny{\textcolor{black}{{probe}}}};
\end{scope}
\begin{scope}[shift = {(0.55,-0.27)}]
\node[rotate around = {-25:(0.37,0.355)}] at (0.37,0.355) {\tiny{\textcolor{black}{{phonon}}}};
\end{scope}
\draw[black,->,green!50!black,very thick,decorate,decoration={snake,amplitude=.3mm,segment length=1.4mm,pre length = 0mm,post length=2.0mm}] (0.30,0.25) -- (0.55,0.25);
\begin{scope}[shift={(0.28,0.0)}]
\begin{scope}[rotate around={25:(0.30,0.25)}]
\draw[black,->,red!50!black,very thick,decorate,decoration={snake,amplitude=.3mm,segment length=1.4mm,pre length = 0mm,post length=2.0mm}] (0.30,0.25) -- (0.55,0.25);
\end{scope}
\begin{scope}[rotate around={-25:(0.30,0.25)}]
\draw[black,->,very thick,decorate,decoration={snake,amplitude=.3mm,segment length=1.4mm,pre length = 0mm,post length=2.0mm}] (0.30,0.25) -- (0.55,0.25);
\end{scope}
\end{scope}
\end{scope}
\filldraw[black](1.98,1.625) circle (0.7mm);
\end{tikzpicture}
}
\caption{\textbf{A series of suspended silicon nanowires.} \textbf{a}, Impression of a silicon-on-insulator waveguide that consists of a series of suspensions and anchors. The photons propagate along the wire while the phonons are localized at their $z$-point of creation. \textbf{b}, Scanning electron micrograph of an actual suspension of length $L_{\text{s}} = 25.4 \, \mu \text{m}$ held by $L_{\text{a}} = 4.6 \, \mu\text{m}$ long anchors. \textbf{c}, Photonic (top) and phononic (bottom) traveling modes. \textbf{d}, The Brillouin process converts incoming pump photons with energy-momentum $(\hbar \omega_{\text{p}},\hbar k_{\text{p}})$ into redshifted probe (Stokes) photons $(\hbar \omega_{\text{pr}},\hbar k_{\text{pr}})$ and phonons $(\hbar \Omega, \hbar K)$.}
\label{fig:structure_and_modes}
\end{figure*}
\section{Results and discussion}
\label{sec:results}
\subsection{Theoretical background}
\label{subsec:theory}
The following discussion is concerned with forward intra-modal scattering, in which co-propagating pump and probe waves generate low-wavevector, low-group-velocity acoustic phonons (fig.\ref{fig:structure_and_modes}d).
\subsubsection{Brillouin gain.}
\label{subsub:Brillouingain}
First, we briefly treat the small-signal Brillouin gain in a waveguide consisting of suspensions of length $L_{\text{s}}$ and anchors of length $L_{\text{a}}$. The section length is $L_{\text{sec}} = L_{\text{s}}+L_{\text{a}}$ and there are $N = \frac{L}{L_{\text{sec}}}$ such sections with $L$ the total waveguide length (fig.\ref{fig:structure_and_modes}). We denote the input pump power $P_{\text{p}}$ and the red-detuned probe (Stokes) power $P_{\text{pr}}$. As previously shown \cite{VanLaer2015}, the power $P_{\text{pr}}$ of the probe obeys
\begin{align}
\label{eq:gainevolution}
\frac{\text{d}P_{\text{pr}}}{\text{d}z} &= -\left(\tilde{\mathcal{G}} P_{\text{p}}e^{-\alpha z} \Im{\mathcal{L}} + \alpha \right) P_{\text{pr}} &\text{suspensions}\\
\frac{\text{d}P_{\text{pr}}}{\text{d}z} &= - \alpha P_{\text{pr}} & \text{anchors} \notag
\end{align}
in the low-cascading regime and with $\tilde{\mathcal{G}}$ the Brillouin gain coefficient, $\mathcal{L}(\Delta) =\frac{1}{-2\Delta + i}$ the complex Lorentzian, $\Delta = \frac{\Omega - \Omega_{\text{m}}}{\Gamma_{\text{m}}}$ the normalized detuning, $\Gamma_{\text{m}}$ the phonon linewidth and $\Im \mathcal{L} = -\frac{1}{4 \Delta^{2} + 1}$. To derive \eqref{eq:gainevolution}, we assumed that the phonon propagation loss far exceeds the photon propagation loss and that the photon-phonon coupling is weak relative to the spatial phonon decay \cite{VanLaer2015b}. In particular, in this work the photon decay length $\alpha^{-1}$ is about a centimeter, while the phonons spatially decay over a couple of nanometers in the $z$-direction \cite{VanLaer2015}. Indeed, the flat dispersion of these Raman-like \cite{Kang2009,Rakich2012} phonons yields an exceedingly low group velocity \cite{VanLaer2015}. Therefore, each suspension consists of a series of independent mechanical oscillators -- whose frequency depends on the local width \cite{VanLaer2015}. This phonon locality also implies that the anchor does not contribute to the Brillouin gain.
We treat the optical loss $\alpha$ as distributed, although it may in fact be partially localized at the interfaces or be unequal in the suspensions and anchors. This is a good approximation as the SBS strength in the next section only depends on the remaining pump power, not on how some of it was lost in the previous sections. The ansatz $P_{\text{pr}} = g(z) e^{-\alpha z}$ and piecewise integration of \eqref{eq:gainevolution} results in
\begin{equation*}
\ln \frac{g(L)}{g(0)} = - \tilde{\mathcal{G}} P_{\text{p}} L_{\text{s,eff}} \Im{\mathcal{L}} \sum_{k=0}^{N-1} e^{- \alpha k L_{\text{sec}}} = - \tilde{\mathcal{G}} P_{\text{p}} L_{\text{s,eff}} \Im{\mathcal{L}} \frac{1 - e^{-\alpha L}}{1 - e^{-\alpha L_{\text{sec}}}} \approx - \tilde{\mathcal{G}} P_{\text{p}} f_{\text{s}} L_{\text{eff}} \Im{\mathcal{L}}
\end{equation*}
with the effective suspension length $L_{\text{s,eff}} = \frac{1 - e^{-\alpha L_{\text{s}}}}{\alpha} \approx L_{\text{s}}$ since the sections are much smaller than the optical decay length ($L_{\text{sec}} \ll \alpha^{-1}$) and $f_{\text{s}} = \frac{L_{\text{s}}}{L_{\text{sec}}}$. Therefore we obtain
\begin{equation}
\label{eq:Brillouingain_homo}
\ln \frac{P_{\text{pr}}(L)}{P_{\text{pr}}(0)e^{-\alpha L}} = \frac{\tilde{\mathcal{G}} P_{\text{p}} f_{\text{s}} L_{\text{eff}} }{4 \Delta^{2} + 1}
\end{equation}
These experiments are the circuit analog of cavity-based optomechanically induced transparency \cite{Weis2010,Safavi-Naeini2011b}. However, our system features spatially stronger mechanical than optical damping, such that it is the optical response that is modified here \cite{VanLaer2015b}.
\subsubsection{Cross-phase modulation.}
\label{subsubsec:XPMtheory}
Gain measurements provide access to all relevant optomechanical parameters, but require careful calibration of the on-chip pump power $P_{\text{p}}$. In contrast, a cross-phase modulation (XPM) measurement \cite{Shin2013b,VanLaer2015} is, in absence of free-carriers, intrinsically calibrated: it provides access to the ratio of the photon-phonon coupling and the electronic Kerr effect independent of pump power. These experiments are the circuit analog of cavity-based coherent wavelength conversion \cite{Hill2012}, although the conversion need not take place between two optical resonances in our case.
We assume weak XPM and denote the envelopes of the pump and its red- and blue-detuned sidebands $a_{\text{p}}$, $a_{\text{p}-}$ and $a_{\text{p}+}$ and similarly for the injected probe $a_{\text{pr}}$ and the XPM-imprinted blue-shifted sideband $a_{\text{pr}+}$. The imprinted sideband grows as \cite{VanLaer2015}
\begin{align*}
\frac{\text{d}a_{\text{pr}+}}{\text{d}z} &= -\frac{i}{2} \left(4 \gamma_{\text{Ks}} + \tilde{\mathcal{G}} \mathcal{L}\right) \left( a_{\text{p}}a^{\star}_{\text{p}-} + a_{\text{p}+} a^{\star}_{\text{p}} \right) a_{\text{pr}} - \frac{\alpha}{2} a_{\text{pr}+} & \text{suspensions}\\
\frac{\text{d}a_{\text{pr}+}}{\text{d}z} &= -i 2\gamma_{\text{Ka}} \left( a_{\text{p}}a^{\star}_{\text{p}-} + a_{\text{p}+} a^{\star}_{\text{p}} \right) a_{\text{pr}} - \frac{\alpha}{2} a_{\text{pr}+} & \text{anchors}
\end{align*}
with $\gamma_{\text{K}}$ the Kerr parameter. Note that the XPM can also be seen as a four-wave mixing process with photon creations ($a^{\star}$) and annihilations ($a$) given by $a^{\star}_{\text{pr+}}\left( a_{\text{p}}a^{\star}_{\text{p}-} + a_{\text{p}+} a^{\star}_{\text{p}} \right) a_{\text{pr}}$. We assume that the pump and probe remain undepleted by the XPM, but include their absorptive decay. Then we get
\begin{align*}
\frac{\text{d}a_{\text{pr}+}}{\text{d}z} &= C_{\text{s}}e^{-\frac{3\alpha}{2} z} - \frac{\alpha}{2} a_{\text{pr}+} & \text{suspensions}\\
\frac{\text{d}a_{\text{pr}+}}{\text{d}z} &= C_{\text{a}}e^{-\frac{3\alpha}{2} z} - \frac{\alpha}{2} a_{\text{pr}+} & \text{anchors}
\end{align*}
with $C_{\text{s}} = \frac{1}{2} \left(4 \gamma_{\text{Ks}} + \tilde{\mathcal{G}} \mathcal{L}\right) C$, $C_{\text{a}} = 2 \gamma_{\text{Ka}} C$ and $C = -i\left( a_{\text{p}}a^{\star}_{\text{p}-} + a_{\text{p}+} a^{\star}_{\text{p}} \right) a_{\text{pr}}|_{z=0}$. Inserting the ansatz $a_{\text{pr}+}(z) = g(z) e^{-\frac{\alpha}{2}z}$ and piecewise integrating ($g(0)=0$) yields
\begin{equation*}
g(L) \approx \left( C_{\text{s}} L_{\text{s}} + C_{\text{a}} L_{\text{a}} \right) \sum_{k=0}^{N-1} e^{- \alpha k L_{\text{sec}}} = \left( C_{\text{s}} L_{\text{s}} + C_{\text{a}} L_{\text{a}} \right) \frac{1 - e^{-\alpha L}}{1 - e^{-\alpha L_{\text{sec}}}} \approx \left( C_{\text{s}} f_{\text{s}} + C_{\text{a}} f_{\text{a}} \right) L_{\text{eff}}
\end{equation*}
where we used $L_{\text{a,eff}} \approx L_{\text{a}}$, $L_{\text{eff}} = \frac{1 - e^{-\alpha L}}{\alpha}$ and $f_{\text{a}} = 1 - f_{\text{s}}$. Therefore,
\begin{align*}
\left|a_{\text{pr}+}(L)\right|^{2} &= 4 \overline{\gamma}^{2}_{\text{K}} \mathcal{F} |C|^{2} L^{2}_{\text{tot,eff}} e^{-\alpha L} \propto \mathcal{F}
\end{align*}
with the averaged Kerr parameter $\overline{\gamma}_{\text{K}} = \gamma_{\text{Ks}}f_{\text{s}} + \gamma_{\text{Ka}}f_{a}$, the normalized Fano function
\begin{equation}
\label{eq:Fano_homo}
\mathcal{F}(\Delta) = \left| 1 + f_{\text{s}}\frac{\tilde{\mathcal{G}}}{4 \overline{\gamma}_{\text{K}}} \mathcal{L}(\Delta) \right|^{2} = \left| 1 + r \mathcal{L}(\Delta) \right|^{2}
\end{equation}
and $r = f_{\text{s}}\frac{\tilde{\mathcal{G}}}{4 \overline{\gamma}_{\text{K}}}$ the ratio between the mechanically- and Kerr-driven XPM.
Including two-photon absorption, free-carrier index changes and free-carrier absorption, \eqref{eq:Fano_homo} no longer holds. For instance, in case of two-photon absorption the Kerr parameter should be replaced by $\overline{\gamma}_{\text{K}} \rightarrow \overline{\gamma}_{\text{K}} - i \overline{\gamma}_{\text{TPA}}$ and \eqref{eq:Fano_homo} becomes
\begin{equation}
\label{eq:Fano_TPA}
\mathcal{F}(\Delta) = \left| 1 + e^{i \phi} f_{\text{s}}\frac{\tilde{\mathcal{G}}}{4 \overline{\gamma}_{\text{tot}}} \mathcal{L}(\Delta) \right|^{2} = \left| 1 + e^{i \phi} r \mathcal{L}(\Delta) \right|^{2}
\end{equation}
with $\overline{\gamma}_{\text{tot}} = \sqrt{\overline{\gamma}_{\text{K}}^{2} + \overline{\gamma}^{2}_{\text{TPA}}}$ and $\tan{\phi} = \frac{\overline{\gamma}_{\text{TPA}}}{\overline{\gamma}_{\text{K}}}$. In our case, the two-photon $\phi$ is small and positive as $\frac{\overline{\gamma}_{\text{TPA}}}{\overline{\gamma}_{\text{K}}} \approx 0.1$. The free-carrier nonlinearity $\overline{\gamma}_{\text{FC}}$, however, can give rise to negative $\phi$ (see Appendix).
\subsection{Fabrication and passive characterization.}
\label{subsec:fabrication}
We started from air-cladded 220 nm thick, 450 nm wide silicon-on-insulator wires fabricated by 193 nm UV lithography (\textit{www.ePIXfab.eu}) at imec. Next, we patterned an array of apertures in a resist spinned atop the wires. Then we immersed the chip in buffered hydrofluoric acid, which selectively etches the silicon dioxide substrate, until the wires were released. The end result was a series of suspended beams, each typically $25 \, \mu\text{m}$ long and held by $5 \, \mu \text{m}$ silicon dioxide anchors (fig.\ref{fig:structure_and_modes}). Simulations and measurements show that the reflections caused by these anchors are negligibly small (see Appendix). We found optical losses $\alpha \approx 5.5 \, \text{dB/cm}$ by the cut-back method, which are a factor 2 larger than before the etch. This is likely related to a deterioration of the wires' surface state and consistent with both (1) the measured drop in free-carrier lifetime (see Appendix) and (2) the decrease in free-carrier absorption found in the gain experiment (fig.\ref{fig:netgain}b).
\subsection{Optomechanical experiments}
\label{subsec:experiments}
In this section, we discuss the guided-wave optomechanical characterization of a series of suspended silicon nanobeams. We used the experimental set-ups presented in \cite{VanLaer2015}. Our device is characterized by the suspension length $L_{\text{s}}$, the anchor length $L_{\text{a}}$, the section length $L_{\text{sec}} = L_{\text{s}}+L_{\text{a}}$, the number of suspensions $N$, the total length $L = N L_{\text{sec}}$, the total effective length $L_{\text{eff}} = \frac{1 - e^{-\alpha L}}{\alpha}$, the suspended fraction $f_{\text{s}} = \frac{L_{\text{s}}}{L_{\text{sec}}}$ and the waveguide width $w$. Unless stated otherwise, these parameters have values $L_{\text{s}} = 25.4 \, \mu\text{m}$, $L_{\text{a}} = 4.6 \, \mu\text{m}$, $L_{\text{sec}} = 30 \, \mu\text{m}$, $N = 85$, $L = 2535 \, \mu\text{m}$, $L_{\text{eff}} = 2168 \, \mu\text{m}$, $f_{\text{s}} = 0.85$ and $w = 450 \, \text{nm}$. In some cases, our waveguides have a non-suspended input/output section before/after the cascade of suspended nanobeams. We take this into account when calculating input pump powers (gain experiment) or suspended fractions (XPM experiment).
\subsubsection{Brillouin gain.}
\label{subsubsec:Brillouingain}
\begin{figure*}
\centering
\subfigure{
\begin{tikzpicture}
\begin{axis}[scale = 0.34*\textwidth/(4.5in),
width=4.52083333333333in,
height=3.565625in,
xmin=9.05,
xmax=9.15,
xlabel={Frequency $\frac{\Omega}{2\pi}$ (GHz)},
xlabel style = {yshift=0ex},
ymin=-0.5,
ymax=1.5,
ylabel= {Probe power (dB)},
ylabel style = {yshift=-1.5ex},
axis x line*=bottom,
axis y line*=left,
legend style={fill=none,draw=none,legend cell align=left,at={(0.5,1.2)},anchor=north, legend columns=-1},
xtick = {9.06,9.1,9.14},
ytick = {0,0.7,1.4},
minor xtick={9.08,9.10,...,9.14},
minor ytick={0.35,0.7,...,1.55}
]
\addplot [
color=blue,
line width=0.5pt,
mark size=1.35pt,
only marks,
mark=o,
mark options={solid}
]
table[row sep=crcr]{
9.05299999999997 -0.0294301172271991\\
9.05399999999997 -0.033356045028289\\
9.05499999999997 -0.0260693728293786\\
9.05599999999997 -0.0174739006304678\\
9.05699999999997 -0.0271506284315572\\
9.05799999999997 -0.0167717562326474\\
9.05899999999997 -0.0214197840337364\\
9.05999999999997 -0.0153390118348264\\
9.06099999999997 -0.0230048396359157\\
9.06199999999997 -0.018504567437005\\
9.06299999999997 -0.0260875952380947\\
9.06399999999996 -0.0162670230391848\\
9.06499999999996 0.00247514915972595\\
9.06599999999996 0.00355652135863509\\
9.06699999999996 0.0133016935575468\\
9.06799999999996 0.0234206657564569\\
9.06899999999996 0.0280605379553665\\
9.06999999999996 0.0388913101542779\\
9.07099999999996 0.0518111823531866\\
9.07199999999996 0.0548904545520986\\
9.07299999999996 0.0544671267510078\\
9.07399999999996 0.0485941989499184\\
9.07499999999996 0.0583061711488293\\
9.07599999999996 0.0701023433477391\\
9.07699999999996 0.0702270155466503\\
9.07799999999996 0.0547118877455591\\
9.07899999999996 0.0446188599444702\\
9.07999999999996 0.0622483321433807\\
9.08099999999996 0.0716796043422905\\
9.08199999999995 0.0702181765412009\\
9.08299999999995 0.0877327487401117\\
9.08399999999995 0.112145220939023\\
9.08499999999995 0.136966193137933\\
9.08599999999995 0.162299965336843\\
9.08699999999995 0.198013937535753\\
9.08799999999995 0.239822409734663\\
9.08899999999995 0.270530481933574\\
9.08999999999995 0.300394554132484\\
9.09099999999995 0.334990426331394\\
9.09199999999995 0.379654098530305\\
9.09299999999995 0.442526170729216\\
9.09399999999995 0.522194342928125\\
9.09499999999995 0.622596215127035\\
9.09599999999995 0.743523587325947\\
9.09699999999995 0.881778059524856\\
9.09799999999995 0.981303631723766\\
9.09899999999995 1.07182050392268\\
9.09999999999994 1.16879397612159\\
9.10099999999994 1.2471926483205\\
9.10199999999994 1.29024112051941\\
9.10299999999994 1.30630319271832\\
9.10399999999994 1.27848386491723\\
9.10499999999994 1.17391543711614\\
9.10599999999994 1.00648110931505\\
9.10699999999994 0.80903138151396\\
9.10799999999994 0.64646465371287\\
9.10899999999994 0.496576625911781\\
9.10999999999994 0.37776719811069\\
9.11099999999994 0.288051570309601\\
9.11199999999994 0.224498842508511\\
9.11299999999994 0.173136314707421\\
9.11399999999994 0.133040686906331\\
9.11499999999994 0.104412559105243\\
9.11599999999994 0.0811954313041535\\
9.11699999999994 0.0665297035030639\\
9.11799999999993 0.051578475701974\\
9.11899999999993 0.0319026479008846\\
9.11999999999993 0.020657720099794\\
9.12099999999993 0.0191751922987047\\
9.12199999999993 0.0128986644976151\\
9.12299999999993 0.0072534366965256\\
9.12399999999993 0.00301950889543527\\
9.12499999999993 0.0142361810943461\\
9.12599999999993 0.00822615329325686\\
9.12699999999993 0.00227972549216653\\
9.12799999999993 0.00350189769107616\\
9.12899999999993 -0.00907813011001309\\
9.12999999999993 -0.0169500579111021\\
9.13099999999993 -0.0128191857121917\\
9.13199999999993 -0.00802651351328172\\
9.13299999999993 -0.0116719413143711\\
9.13399999999993 -0.00713756911546155\\
9.13499999999993 -0.00887879691655065\\
9.13599999999992 -0.0189765247176412\\
9.13699999999992 -0.02501965251873\\
9.13799999999992 -0.0216248803198189\\
9.13899999999992 -0.00360050812091024\\
9.13999999999992 0.00158436407800182\\
9.14099999999992 0.00565993627691079\\
9.14199999999992 -0.000745191524178915\\
9.14299999999992 -0.00523061932526843\\
9.14399999999992 -0.0113499471263578\\
9.14499999999992 -0.00947417492744669\\
9.14599999999992 -0.00547540272853637\\
9.14699999999992 7.41694703725732e-05\\
9.14799999999992 -0.00235585833071546\\
9.14899999999992 -0.0131025861318052\\
9.14999999999992 -0.025963113932895\\
9.15099999999992 -0.0285313417339852\\
9.15199999999992 -0.0252977695350749\\
9.15299999999992 -0.0118403973361642\\
};
\addlegendentry{Experiment \, \,};
\addplot [
color=red,
solid,
line width=2.0pt,
]
table[row sep=crcr]{
9.05299999999997 -0.0352280491359782\\
9.05324999999997 -0.0349981036681845\\
9.05349999999997 -0.034764650633173\\
9.05374999999997 -0.0345276187390312\\
9.05399999999997 -0.03428693488412\\
9.05424999999997 -0.0340425241020164\\
9.05449999999997 -0.0337943095045019\\
9.05474999999997 -0.0335422122225021\\
9.05499999999997 -0.0332861513449331\\
9.05524999999997 -0.033026043855342\\
9.05549999999997 -0.0327618045662479\\
9.05574999999997 -0.0324933460511105\\
9.05599999999997 -0.0322205785738196\\
9.05624999999997 -0.0319434100155744\\
9.05649999999997 -0.0316617457990958\\
9.05674999999997 -0.0313754888100206\\
9.05699999999997 -0.0310845393153574\\
9.05724999999997 -0.0307887948789115\\
9.05749999999997 -0.0304881502735218\\
9.05774999999997 -0.0301824973899721\\
9.05799999999997 -0.0298717251424512\\
9.05824999999997 -0.0295557193704127\\
9.05849999999997 -0.0292343627366431\\
9.05874999999997 -0.028907534621418\\
9.05899999999997 -0.028575111012559\\
9.05924999999997 -0.0282369643911793\\
9.05949999999997 -0.0278929636129804\\
9.05974999999997 -0.0275429737848666\\
9.05999999999997 -0.0271868561366586\\
9.06024999999997 -0.0268244678877215\\
9.06049999999997 -0.0264556621082556\\
9.06074999999997 -0.0260802875749886\\
9.06099999999997 -0.0256981886210524\\
9.06124999999997 -0.0253092049797442\\
9.06149999999997 -0.0249131716219201\\
9.06174999999997 -0.0245099185866711\\
9.06199999999997 -0.0240992708050348\\
9.06224999999997 -0.0236810479163804\\
9.06249999999997 -0.0232550640770981\\
9.06274999999997 -0.022821127761276\\
9.06299999999997 -0.0223790415529629\\
9.06324999999997 -0.0219286019295805\\
9.06349999999997 -0.0214695990361084\\
9.06374999999997 -0.021001816449579\\
9.06399999999997 -0.0205250309333675\\
9.06424999999997 -0.0200390121808388\\
9.06449999999997 -0.0195435225478037\\
9.06474999999997 -0.0190383167731864\\
9.06499999999997 -0.0185231416873869\\
9.06524999999997 -0.0179977359076831\\
9.06549999999997 -0.0174618295199932\\
9.06574999999997 -0.0169151437463657\\
9.06599999999997 -0.0163573905974382\\
9.06624999999997 -0.0157882725090669\\
9.06649999999997 -0.0152074819623613\\
9.06674999999997 -0.0146147010862506\\
9.06699999999997 -0.0140096012416219\\
9.06724999999997 -0.0133918425861353\\
9.06749999999997 -0.0127610736186716\\
9.06774999999997 -0.012116930702273\\
9.06799999999997 -0.011459037564534\\
9.06824999999997 -0.0107870047741748\\
9.06849999999997 -0.0101004291924691\\
9.06874999999997 -0.00939889339826138\\
9.06899999999997 -0.00868196508507567\\
9.06924999999997 -0.00794919642874724\\
9.06949999999997 -0.00720012342403587\\
9.06974999999997 -0.00643426518845469\\
9.06999999999997 -0.00565112323143823\\
9.07024999999997 -0.0048501806869989\\
9.07049999999997 -0.00403090150775587\\
9.07074999999997 -0.00319272961810523\\
9.07099999999997 -0.00233508802427995\\
9.07124999999997 -0.00145737787879163\\
9.07149999999997 -0.000558977496550664\\
9.07174999999997 0.000360758680031284\\
9.07199999999997 0.00130250116998918\\
9.07224999999997 0.00226694660004194\\
9.07249999999997 0.00325481891138672\\
9.07274999999997 0.00426687063071018\\
9.07299999999997 0.0053038842093133\\
9.07324999999997 0.00636667343432921\\
9.07349999999997 0.00745608491644312\\
9.07374999999997 0.00857299965884131\\
9.07399999999997 0.00971833471221129\\
9.07424999999997 0.010893044921161\\
9.07449999999997 0.0120981247677658\\
9.07474999999997 0.0133346103181432\\
9.07499999999997 0.0146035812785559\\
9.07524999999997 0.0159061631680032\\
9.07549999999997 0.0172435296144989\\
9.07574999999997 0.0186169047829534\\
9.07599999999997 0.0200275659431497\\
9.07624999999997 0.0214768461866101\\
9.07649999999997 0.0229661373020331\\
9.07674999999997 0.0244968928196559\\
9.07699999999997 0.0260706312353268\\
9.07724999999997 0.0276889394261954\\
9.07749999999997 0.029353476270452\\
9.07774999999997 0.0310659764847551\\
9.07799999999997 0.032828254693515\\
9.07824999999997 0.0346422097455364\\
9.07849999999997 0.0365098292946329\\
9.07874999999997 0.0384331946616351\\
9.07899999999997 0.0404144859967828\\
9.07924999999997 0.0424559877628667\\
9.07949999999997 0.0445600945605018\\
9.07974999999997 0.0467293173188064\\
9.07999999999997 0.0489662898764031\\
9.08024999999997 0.0512737759789695\\
9.08049999999997 0.0536546767217858\\
9.08074999999997 0.0561120384677185\\
9.08099999999997 0.0586490612726908\\
9.08124999999997 0.0612691078532854\\
9.08149999999997 0.0639757131335168\\
9.08174999999997 0.066772594409689\\
9.08199999999997 0.0696636621753213\\
9.08224999999997 0.0726530316508059\\
9.08249999999997 0.0757450350646947\\
9.08274999999997 0.0789442347369091\\
9.08299999999997 0.0822554370171394\\
9.08324999999997 0.0856837071341198\\
9.08349999999997 0.0892343850150543\\
9.08374999999997 0.0929131021375307\\
9.08399999999997 0.0967257994785334\\
9.08424999999997 0.10067874662862\\
9.08449999999997 0.104778562141936\\
9.08474999999997 0.109032235194313\\
9.08499999999997 0.113447148624147\\
9.08524999999997 0.118031103432041\\
9.08549999999997 0.122792344814721\\
9.08574999999997 0.127739589808784\\
9.08599999999997 0.132882056617863\\
9.08624999999997 0.138229495692\\
9.08649999999997 0.143792222622903\\
9.08674999999997 0.149581152910033\\
9.08699999999997 0.155607838639291\\
9.08724999999997 0.161884507100591\\
9.08749999999997 0.168424101349197\\
9.08774999999997 0.175240322686513\\
9.08799999999997 0.182347675001457\\
9.08824999999997 0.189761510868547\\
9.08849999999997 0.197498079240879\\
9.08874999999997 0.205574574507015\\
9.08899999999997 0.214009186593863\\
9.08924999999997 0.222821151688913\\
9.08949999999997 0.232030803025432\\
9.08974999999997 0.241659621014521\\
9.08999999999997 0.251730281812113\\
9.09024999999997 0.262266703176048\\
9.09049999999997 0.273294086186183\\
9.09074999999997 0.284838951059848\\
9.09099999999997 0.296929164892449\\
9.09124999999997 0.309593958672596\\
9.09149999999997 0.322863930350064\\
9.09174999999997 0.336771030067608\\
9.09199999999997 0.351348522884804\\
9.09224999999997 0.366630923409073\\
9.09249999999997 0.382653895702877\\
9.09274999999997 0.39945411063418\\
9.09299999999997 0.417069051485469\\
9.09324999999997 0.435536757122589\\
9.09349999999997 0.454895490373013\\
9.09374999999997 0.475183317490259\\
9.09399999999997 0.496437582732841\\
9.09424999999997 0.518694260250197\\
9.09449999999997 0.541987163751161\\
9.09474999999997 0.566346992998435\\
9.09499999999997 0.591800195271731\\
9.09524999999997 0.618367619874299\\
9.09549999999997 0.646062944929072\\
9.09574999999997 0.674890858651081\\
9.09599999999997 0.704844982605754\\
9.09624999999997 0.735905532897521\\
9.09649999999997 0.768036727605146\\
9.09674999999997 0.801183965902213\\
9.09699999999997 0.835270826918172\\
9.09724999999997 0.870195965038187\\
9.09749999999997 0.905830013051409\\
9.09774999999997 0.942012644661895\\
9.09799999999997 0.978549991656529\\
9.09824999999997 1.01521265531586\\
9.09849999999997 1.05173459162051\\
9.09874999999997 1.08781317883023\\
9.09899999999997 1.12311078582793\\
9.09924999999997 1.15725814103673\\
9.09949999999997 1.18985974587418\\
9.09974999999997 1.22050147687858\\
9.09999999999997 1.24876037463675\\
9.10024999999997 1.27421643020311\\
9.10049999999997 1.29646596428336\\
9.10074999999997 1.31513597375993\\
9.10099999999997 1.32989862445301\\
9.10124999999997 1.34048493206456\\
9.10149999999997 1.34669662601707\\
9.10174999999997 1.34841525434868\\
9.10199999999997 1.34560776657625\\
9.10224999999997 1.33832809037785\\
9.10249999999997 1.32671456320773\\
9.10274999999997 1.31098344460635\\
9.10299999999997 1.29141906818166\\
9.10324999999997 1.26836144948748\\
9.10349999999997 1.2421923176094\\
9.10374999999997 1.21332057392547\\
9.10399999999997 1.18216811024357\\
9.10424999999997 1.14915676436198\\
9.10449999999997 1.11469698634731\\
9.10474999999997 1.07917856686733\\
9.10499999999997 1.04296356850235\\
9.10524999999997 1.00638142262528\\
9.10549999999997 0.969726019425726\\
9.10574999999997 0.933254529832933\\
9.10599999999997 0.897187651960915\\
9.10624999999997 0.861710963805824\\
9.10649999999997 0.826977079188476\\
9.10674999999997 0.793108336299354\\
9.10699999999997 0.76019978974967\\
9.10724999999997 0.728322321573529\\
9.10749999999997 0.697525729752872\\
9.10774999999997 0.667841691753126\\
9.10799999999997 0.639286533834171\\
9.10824999999997 0.611863764068148\\
9.10849999999997 0.585566348228317\\
9.10874999999997 0.560378723616062\\
9.10899999999997 0.536278557242797\\
9.10924999999997 0.513238262395392\\
9.10949999999997 0.491226292298523\\
9.10974999999997 0.47020823205071\\
9.10999999999997 0.450147710839476\\
9.11024999999997 0.431007156138202\\
9.11049999999997 0.412748410534894\\
9.11074999999997 0.395333230322718\\
9.11099999999997 0.378723683222674\\
9.11124999999997 0.362882460762572\\
9.11149999999997 0.347773118999602\\
9.11174999999997 0.333360259526078\\
9.11199999999997 0.319609661079336\\
9.11224999999997 0.306488370600351\\
9.11249999999997 0.293964761268975\\
9.11274999999997 0.282008563883872\\
9.11299999999997 0.270590876937597\\
9.11324999999997 0.2596841598591\\
9.11349999999997 0.249262213143639\\
9.11374999999997 0.23930014844369\\
9.11399999999997 0.22977435114832\\
9.11424999999997 0.220662437519213\\
9.11449999999997 0.211943208061296\\
9.11474999999997 0.203596598483748\\
9.11499999999997 0.195603629335114\\
9.11524999999997 0.187946355174323\\
9.11549999999997 0.180607813952279\\
9.11574999999997 0.173571977126381\\
9.11599999999997 0.166823700906733\\
9.11624999999997 0.160348678929309\\
9.11649999999997 0.154133396568649\\
9.11674999999997 0.148165087037116\\
9.11699999999997 0.142431689362743\\
9.11724999999997 0.136921808295161\\
9.11749999999997 0.131624676156488\\
9.11774999999997 0.126530116626628\\
9.11799999999997 0.121628510432254\\
9.11824999999997 0.11691076289445\\
9.11849999999997 0.112368273277645\\
9.11874999999997 0.1079929058745\\
9.11899999999997 0.10377696275684\\
9.11924999999997 0.0997131581183947\\
9.11949999999997 0.0957945941335256\\
9.11974999999997 0.0920147382563958\\
9.11999999999997 0.0883674018848554\\
9.12024999999997 0.0848467203146956\\
9.12049999999997 0.0814471339125535\\
9.12074999999997 0.0781633704373882\\
9.12099999999997 0.0749904284431129\\
9.12124999999997 0.0719235616984912\\
9.12149999999997 0.0689582645626575\\
9.12174999999997 0.0660902582577292\\
9.12199999999997 0.0633154779835593\\
9.12224999999997 0.0606300608220557\\
9.12249999999997 0.0580303343814943\\
9.12274999999997 0.0555128061345934\\
9.12299999999997 0.0530741534063418\\
9.12324999999997 0.0507112139702268\\
9.12349999999997 0.0484209772145348\\
9.12374999999997 0.0462005758422574\\
9.12399999999997 0.0440472780705031\\
9.12424999999997 0.0419584802978602\\
9.12449999999997 0.0399317002097572\\
9.12474999999997 0.0379645702938343\\
9.12499999999997 0.0360548317395027\\
9.12524999999997 0.0342003286971936\\
9.12549999999997 0.0323990028743913\\
9.12574999999997 0.0306488884474127\\
9.12599999999997 0.0289481072689037\\
9.12624999999997 0.0272948643523686\\
9.12649999999997 0.0256874436165914\\
9.12674999999997 0.0241242038736054\\
9.12699999999997 0.0226035750449743\\
9.12724999999997 0.0211240545924085\\
9.12749999999997 0.0196842041494043\\
9.12774999999997 0.01828264634143\\
9.12799999999997 0.016918061783315\\
9.12824999999997 0.0155891862429269\\
9.12849999999997 0.0142948079609639\\
9.12874999999997 0.0130337651176\\
9.12899999999997 0.0118049434370532\\
9.12924999999997 0.0106072739217474\\
9.12949999999997 0.00943973070851089\\
9.12974999999997 0.00830132903942845\\
9.12999999999997 0.00719112334069383\\
9.13024999999997 0.00610820540297879\\
9.13049999999997 0.00505170265754556\\
9.13074999999997 0.00402077654246784\\
9.13099999999997 0.00301462095367711\\
9.13124999999997 0.00203246077608717\\
9.13149999999997 0.00107355049015675\\
9.13174999999997 0.000137172849516143\\
9.13199999999997 -0.000777362374223718\\
9.13224999999997 -0.0016707195833432\\
9.13249999999997 -0.00254353848615722\\
9.13274999999997 -0.00339643517266109\\
9.13299999999997 -0.00423000313685418\\
9.13324999999997 -0.00504481424874242\\
9.13349999999997 -0.00584141967866755\\
9.13374999999997 -0.00662035077662828\\
9.13399999999997 -0.00738211990907378\\
9.13424999999997 -0.00812722125536414\\
9.13449999999997 -0.00885613156612985\\
9.13474999999997 -0.00956931088558457\\
9.13499999999997 -0.0102672032396368\\
9.13524999999997 -0.010950237291638\\
9.13549999999997 -0.011618826967513\\
9.13574999999997 -0.0122733720517941\\
9.13599999999997 -0.0129142587560968\\
9.13624999999997 -0.013541860261525\\
9.13649999999997 -0.0141565372362434\\
9.13674999999997 -0.0147586383295354\\
9.13699999999997 -0.01534850064359\\
9.13724999999997 -0.0159264501840453\\
9.13749999999997 -0.016492802290423\\
9.13774999999997 -0.0170478620474683\\
9.13799999999997 -0.0175919246782991\\
9.13824999999997 -0.0181252759202811\\
9.13849999999997 -0.0186481923845229\\
9.13874999999997 -0.0191609418997289\\
9.13899999999997 -0.0196637838411919\\
9.13924999999997 -0.0201569694456966\\
9.13949999999997 -0.0206407421129307\\
9.13974999999997 -0.0211153376941004\\
9.13999999999997 -0.0215809847683732\\
9.14024999999997 -0.0220379049076782\\
9.14049999999997 -0.0224863129304395\\
9.14074999999997 -0.0229264171447895\\
9.14099999999997 -0.0233584195816986\\
9.14124999999997 -0.0237825162185196\\
9.14149999999997 -0.0241988971934085\\
9.14174999999997 -0.0246077470110052\\
9.14199999999997 -0.0250092447397779\\
9.14224999999997 -0.0254035642014558\\
9.14249999999997 -0.025790874152849\\
9.14274999999997 -0.0261713384604259\\
9.14299999999997 -0.0265451162679886\\
9.14324999999997 -0.0269123621577271\\
9.14349999999997 -0.0272732263049436\\
9.14374999999997 -0.027627854626773\\
9.14399999999997 -0.0279763889251014\\
9.14424999999997 -0.0283189670239644\\
9.14449999999997 -0.0286557229016964\\
9.14474999999997 -0.0289867868179945\\
9.14499999999997 -0.0293122854361564\\
9.14524999999997 -0.0296323419407157\\
9.14549999999997 -0.0299470761506151\\
9.14574999999997 -0.0302566046281646\\
9.14599999999997 -0.030561040783913\\
9.14624999999997 -0.0308604949776613\\
9.14649999999997 -0.0311550746157199\\
9.14674999999997 -0.0314448842445962\\
9.14699999999997 -0.0317300256412788\\
9.14724999999997 -0.0320105979002171\\
9.14749999999997 -0.0322866975171493\\
9.14774999999997 -0.0325584184699342\\
9.14799999999997 -0.0328258522964658\\
9.14824999999997 -0.0330890881698062\\
9.14849999999997 -0.0333482129706672\\
9.14874999999997 -0.0336033113573191\\
9.14899999999997 -0.0338544658330336\\
9.14924999999997 -0.0341017568111908\\
9.14949999999997 -0.0343452626780891\\
9.14974999999997 -0.0345850598535925\\
9.14999999999997 -0.0348212228496882\\
9.15024999999997 -0.0350538243270193\\
9.15049999999997 -0.0352829351494932\\
9.15074999999997 -0.0355086244370327\\
9.15099999999997 -0.0357309596165331\\
9.15124999999997 -0.0359500064711014\\
9.15149999999997 -0.0361658291876535\\
9.15174999999997 -0.0363784904029056\\
9.15199999999997 -0.036588051247838\\
9.15224999999997 -0.0367945713907007\\
9.15249999999997 -0.0369981090785868\\
9.15274999999997 -0.0371987211776354\\
};
\addlegendentry{Fit};
\addplot [
name path = toperror,
color=black,
line width=0.1pt
]
table[x expr=\thisrow{X},y expr=\thisrow{Y} + 0.11,row sep=crcr]{
X Y \\
9.05299999999997 -0.0294301172271991\\
9.05399999999997 -0.033356045028289\\
9.05499999999997 -0.0260693728293786\\
9.05599999999997 -0.0174739006304678\\
9.05699999999997 -0.0271506284315572\\
9.05799999999997 -0.0167717562326474\\
9.05899999999997 -0.0214197840337364\\
9.05999999999997 -0.0153390118348264\\
9.06099999999997 -0.0230048396359157\\
9.06199999999997 -0.018504567437005\\
9.06299999999997 -0.0260875952380947\\
9.06399999999996 -0.0162670230391848\\
9.06499999999996 0.00247514915972595\\
9.06599999999996 0.00355652135863509\\
9.06699999999996 0.0133016935575468\\
9.06799999999996 0.0234206657564569\\
9.06899999999996 0.0280605379553665\\
9.06999999999996 0.0388913101542779\\
9.07099999999996 0.0518111823531866\\
9.07199999999996 0.0548904545520986\\
9.07299999999996 0.0544671267510078\\
9.07399999999996 0.0485941989499184\\
9.07499999999996 0.0583061711488293\\
9.07599999999996 0.0701023433477391\\
9.07699999999996 0.0702270155466503\\
9.07799999999996 0.0547118877455591\\
9.07899999999996 0.0446188599444702\\
9.07999999999996 0.0622483321433807\\
9.08099999999996 0.0716796043422905\\
9.08199999999995 0.0702181765412009\\
9.08299999999995 0.0877327487401117\\
9.08399999999995 0.112145220939023\\
9.08499999999995 0.136966193137933\\
9.08599999999995 0.162299965336843\\
9.08699999999995 0.198013937535753\\
9.08799999999995 0.239822409734663\\
9.08899999999995 0.270530481933574\\
9.08999999999995 0.300394554132484\\
9.09099999999995 0.334990426331394\\
9.09199999999995 0.379654098530305\\
9.09299999999995 0.442526170729216\\
9.09399999999995 0.522194342928125\\
9.09499999999995 0.622596215127035\\
9.09599999999995 0.743523587325947\\
9.09699999999995 0.881778059524856\\
9.09799999999995 0.981303631723766\\
9.09899999999995 1.07182050392268\\
9.09999999999994 1.16879397612159\\
9.10099999999994 1.2471926483205\\
9.10199999999994 1.29024112051941\\
9.10299999999994 1.30630319271832\\
9.10399999999994 1.27848386491723\\
9.10499999999994 1.17391543711614\\
9.10599999999994 1.00648110931505\\
9.10699999999994 0.80903138151396\\
9.10799999999994 0.64646465371287\\
9.10899999999994 0.496576625911781\\
9.10999999999994 0.37776719811069\\
9.11099999999994 0.288051570309601\\
9.11199999999994 0.224498842508511\\
9.11299999999994 0.173136314707421\\
9.11399999999994 0.133040686906331\\
9.11499999999994 0.104412559105243\\
9.11599999999994 0.0811954313041535\\
9.11699999999994 0.0665297035030639\\
9.11799999999993 0.051578475701974\\
9.11899999999993 0.0319026479008846\\
9.11999999999993 0.020657720099794\\
9.12099999999993 0.0191751922987047\\
9.12199999999993 0.0128986644976151\\
9.12299999999993 0.0072534366965256\\
9.12399999999993 0.00301950889543527\\
9.12499999999993 0.0142361810943461\\
9.12599999999993 0.00822615329325686\\
9.12699999999993 0.00227972549216653\\
9.12799999999993 0.00350189769107616\\
9.12899999999993 -0.00907813011001309\\
9.12999999999993 -0.0169500579111021\\
9.13099999999993 -0.0128191857121917\\
9.13199999999993 -0.00802651351328172\\
9.13299999999993 -0.0116719413143711\\
9.13399999999993 -0.00713756911546155\\
9.13499999999993 -0.00887879691655065\\
9.13599999999992 -0.0189765247176412\\
9.13699999999992 -0.02501965251873\\
9.13799999999992 -0.0216248803198189\\
9.13899999999992 -0.00360050812091024\\
9.13999999999992 0.00158436407800182\\
9.14099999999992 0.00565993627691079\\
9.14199999999992 -0.000745191524178915\\
9.14299999999992 -0.00523061932526843\\
9.14399999999992 -0.0113499471263578\\
9.14499999999992 -0.00947417492744669\\
9.14599999999992 -0.00547540272853637\\
9.14699999999992 7.41694703725732e-05\\
9.14799999999992 -0.00235585833071546\\
9.14899999999992 -0.0131025861318052\\
9.14999999999992 -0.025963113932895\\
9.15099999999992 -0.0285313417339852\\
9.15199999999992 -0.0252977695350749\\
9.15299999999992 -0.0118403973361642\\
};
\addplot [
name path = boterror,
color=black,
line width=0.1pt
]
table[x expr=\thisrow{X},y expr=\thisrow{Y} - 0.11,row sep=crcr]{
X Y \\
9.05299999999997 -0.0294301172271991\\
9.05399999999997 -0.033356045028289\\
9.05499999999997 -0.0260693728293786\\
9.05599999999997 -0.0174739006304678\\
9.05699999999997 -0.0271506284315572\\
9.05799999999997 -0.0167717562326474\\
9.05899999999997 -0.0214197840337364\\
9.05999999999997 -0.0153390118348264\\
9.06099999999997 -0.0230048396359157\\
9.06199999999997 -0.018504567437005\\
9.06299999999997 -0.0260875952380947\\
9.06399999999996 -0.0162670230391848\\
9.06499999999996 0.00247514915972595\\
9.06599999999996 0.00355652135863509\\
9.06699999999996 0.0133016935575468\\
9.06799999999996 0.0234206657564569\\
9.06899999999996 0.0280605379553665\\
9.06999999999996 0.0388913101542779\\
9.07099999999996 0.0518111823531866\\
9.07199999999996 0.0548904545520986\\
9.07299999999996 0.0544671267510078\\
9.07399999999996 0.0485941989499184\\
9.07499999999996 0.0583061711488293\\
9.07599999999996 0.0701023433477391\\
9.07699999999996 0.0702270155466503\\
9.07799999999996 0.0547118877455591\\
9.07899999999996 0.0446188599444702\\
9.07999999999996 0.0622483321433807\\
9.08099999999996 0.0716796043422905\\
9.08199999999995 0.0702181765412009\\
9.08299999999995 0.0877327487401117\\
9.08399999999995 0.112145220939023\\
9.08499999999995 0.136966193137933\\
9.08599999999995 0.162299965336843\\
9.08699999999995 0.198013937535753\\
9.08799999999995 0.239822409734663\\
9.08899999999995 0.270530481933574\\
9.08999999999995 0.300394554132484\\
9.09099999999995 0.334990426331394\\
9.09199999999995 0.379654098530305\\
9.09299999999995 0.442526170729216\\
9.09399999999995 0.522194342928125\\
9.09499999999995 0.622596215127035\\
9.09599999999995 0.743523587325947\\
9.09699999999995 0.881778059524856\\
9.09799999999995 0.981303631723766\\
9.09899999999995 1.07182050392268\\
9.09999999999994 1.16879397612159\\
9.10099999999994 1.2471926483205\\
9.10199999999994 1.29024112051941\\
9.10299999999994 1.30630319271832\\
9.10399999999994 1.27848386491723\\
9.10499999999994 1.17391543711614\\
9.10599999999994 1.00648110931505\\
9.10699999999994 0.80903138151396\\
9.10799999999994 0.64646465371287\\
9.10899999999994 0.496576625911781\\
9.10999999999994 0.37776719811069\\
9.11099999999994 0.288051570309601\\
9.11199999999994 0.224498842508511\\
9.11299999999994 0.173136314707421\\
9.11399999999994 0.133040686906331\\
9.11499999999994 0.104412559105243\\
9.11599999999994 0.0811954313041535\\
9.11699999999994 0.0665297035030639\\
9.11799999999993 0.051578475701974\\
9.11899999999993 0.0319026479008846\\
9.11999999999993 0.020657720099794\\
9.12099999999993 0.0191751922987047\\
9.12199999999993 0.0128986644976151\\
9.12299999999993 0.0072534366965256\\
9.12399999999993 0.00301950889543527\\
9.12499999999993 0.0142361810943461\\
9.12599999999993 0.00822615329325686\\
9.12699999999993 0.00227972549216653\\
9.12799999999993 0.00350189769107616\\
9.12899999999993 -0.00907813011001309\\
9.12999999999993 -0.0169500579111021\\
9.13099999999993 -0.0128191857121917\\
9.13199999999993 -0.00802651351328172\\
9.13299999999993 -0.0116719413143711\\
9.13399999999993 -0.00713756911546155\\
9.13499999999993 -0.00887879691655065\\
9.13599999999992 -0.0189765247176412\\
9.13699999999992 -0.02501965251873\\
9.13799999999992 -0.0216248803198189\\
9.13899999999992 -0.00360050812091024\\
9.13999999999992 0.00158436407800182\\
9.14099999999992 0.00565993627691079\\
9.14199999999992 -0.000745191524178915\\
9.14299999999992 -0.00523061932526843\\
9.14399999999992 -0.0113499471263578\\
9.14499999999992 -0.00947417492744669\\
9.14599999999992 -0.00547540272853637\\
9.14699999999992 7.41694703725732e-05\\
9.14799999999992 -0.00235585833071546\\
9.14899999999992 -0.0131025861318052\\
9.14999999999992 -0.025963113932895\\
9.15099999999992 -0.0285313417339852\\
9.15199999999992 -0.0252977695350749\\
9.15299999999992 -0.0118403973361642\\
};
\addplot[gray!15!white] fill between[
of = toperror and boterror,
];
\end{axis}
\node[draw=none] at (1.23,0.3) {\footnotesize $f_{\text{s}}L_{\text{eff}}=1.5 \,\text{mm}$};
\node[draw=none] at (2.4,0.7) {\footnotesize $N=66$};
\node at (-0.9,-0.9) {\large{\textbf{a}}};
\node[draw=none] at (3.7,0.29) {\footnotesize $P_{\text{p}} = 26 \, \text{mW}$};
\begin{scope}[shift={(0.35,-0.07)}]
\draw[black,>=stealth',<->,densely dotted,thick] (1.81,2.13) -- (2.27,2.13);
\begin{scope}[shift={(0,0.00)}]
\draw[black,>=stealth',-,densely dotted,thick] (2.04,2.22) -- (2.53,2.55);
\node[draw=none,text width=1.6cm] at (3.33,2.63) {\footnotesize 12.5 MHz};
\end{scope}
\end{scope}
\end{tikzpicture}%
}
\subfigure{
\begin{tikzpicture}
\begin{axis}[scale = 0.34*\textwidth/(4.5in),
width=4.52083333333333in,
height=3.565625in,
xmin=0,
xmax=50,
xlabel={Pump power $P_{\text{p}}$ (mW)},
ymin=0,
ymax=2.15,
ylabel={On/off gain (dB)},
ylabel style = {yshift=-1.5ex},
axis x line*=bottom,
axis y line*=left,
legend style={fill=none,draw=none,legend cell align=left,at={(0.5,1.2)},anchor=north, legend columns=-1},
xtick = {15,30,45},
ytick = {0.6,1.2,1.8},
minor xtick={7.5,15,...,45},
minor ytick={0.3,0.6,...,2.1}
]
\addplot [
color=black,
solid,
line width=1.0pt,
forget plot
]
table[row sep=crcr]{
0 1.5\\
45 1.5\\
};
\addplot [
color=blue,
line width=0.5pt,
mark size=2.0pt,
only marks,
mark=o,
mark options={solid},
]
table[row sep=crcr]{
1.95937460044138 0.116545752394263\\
2.40145778356686 0.186289631689896\\
3.2465593313025 0.153364552567161\\
3.42591346853364 0.198590142671552\\
4.07689329399471 0.246301302370326\\
4.39282895742181 0.258896609949585\\
5.10165383086894 0.39276985034977\\
7.06236520406498 0.401191951417197\\
7.90698468640645 0.450306078636981\\
9.58734059857148 0.544239845171001\\
11.3228352694945 0.64937218611419\\
15.4523822908621 0.806476841093112\\
18.9477097488396 1.02468375694258\\
21.5947320021443 1.21778098634685\\
25.5865745910813 1.40660312388493\\
28.8939963799571 1.50663072644452\\
31.4462565464831 1.70555897901218\\
35.6832835553455 1.86590477652449\\
38.4633809706373 2.06748822923273\\
38.6405997691 1.92805832080523\\
};
\addlegendentry{Experiment \, \,};
\addplot [
color=red,
solid,
line width=2.0pt,
]
table[row sep=crcr]{
1.95937460044138 0.146723410143881\\
2.05937460044138 0.151865546337258\\
2.15937460044138 0.157007682530636\\
2.25937460044138 0.162149818724013\\
2.35937460044138 0.167291954917391\\
2.45937460044138 0.172434091110768\\
2.55937460044138 0.177576227304146\\
2.65937460044138 0.182718363497523\\
2.75937460044138 0.1878604996909\\
2.85937460044138 0.193002635884278\\
2.95937460044138 0.198144772077655\\
3.05937460044138 0.203286908271033\\
3.15937460044138 0.20842904446441\\
3.25937460044138 0.213571180657788\\
3.35937460044138 0.218713316851165\\
3.45937460044138 0.223855453044543\\
3.55937460044138 0.22899758923792\\
3.65937460044138 0.234139725431298\\
3.75937460044138 0.239281861624675\\
3.85937460044138 0.244423997818052\\
3.95937460044138 0.24956613401143\\
4.05937460044138 0.254708270204807\\
4.15937460044138 0.259850406398185\\
4.25937460044138 0.264992542591562\\
4.35937460044138 0.27013467878494\\
4.45937460044138 0.275276814978317\\
4.55937460044138 0.280418951171695\\
4.65937460044138 0.285561087365072\\
4.75937460044138 0.29070322355845\\
4.85937460044138 0.295845359751827\\
4.95937460044138 0.300987495945204\\
5.05937460044138 0.306129632138582\\
5.15937460044138 0.311271768331959\\
5.25937460044138 0.316413904525337\\
5.35937460044138 0.321556040718714\\
5.45937460044138 0.326698176912092\\
5.55937460044138 0.331840313105469\\
5.65937460044138 0.336982449298847\\
5.75937460044138 0.342124585492224\\
5.85937460044138 0.347266721685602\\
5.95937460044138 0.352408857878979\\
6.05937460044138 0.357550994072356\\
6.15937460044138 0.362693130265734\\
6.25937460044138 0.367835266459111\\
6.35937460044138 0.372977402652489\\
6.45937460044138 0.378119538845866\\
6.55937460044138 0.383261675039244\\
6.65937460044138 0.388403811232621\\
6.75937460044138 0.393545947425999\\
6.85937460044138 0.398688083619376\\
6.95937460044138 0.403830219812753\\
7.05937460044138 0.408972356006131\\
7.15937460044138 0.414114492199508\\
7.25937460044138 0.419256628392886\\
7.35937460044138 0.424398764586263\\
7.45937460044138 0.429540900779641\\
7.55937460044138 0.434683036973018\\
7.65937460044138 0.439825173166396\\
7.75937460044138 0.444967309359773\\
7.85937460044138 0.45010944555315\\
7.95937460044138 0.455251581746528\\
8.05937460044138 0.460393717939906\\
8.15937460044138 0.465535854133283\\
8.25937460044138 0.47067799032666\\
8.35937460044138 0.475820126520038\\
8.45937460044138 0.480962262713415\\
8.55937460044138 0.486104398906793\\
8.65937460044138 0.49124653510017\\
8.75937460044138 0.496388671293548\\
8.85937460044138 0.501530807486925\\
8.95937460044138 0.506672943680303\\
9.05937460044138 0.51181507987368\\
9.15937460044138 0.516957216067058\\
9.25937460044138 0.522099352260435\\
9.35937460044138 0.527241488453812\\
9.45937460044138 0.53238362464719\\
9.55937460044138 0.537525760840567\\
9.65937460044138 0.542667897033945\\
9.75937460044138 0.547810033227322\\
9.85937460044138 0.5529521694207\\
9.95937460044138 0.558094305614077\\
10.0593746004414 0.563236441807455\\
10.1593746004414 0.568378578000832\\
10.2593746004414 0.573520714194209\\
10.3593746004414 0.578662850387587\\
10.4593746004414 0.583804986580964\\
10.5593746004414 0.588947122774342\\
10.6593746004414 0.594089258967719\\
10.7593746004414 0.599231395161097\\
10.8593746004414 0.604373531354474\\
10.9593746004414 0.609515667547852\\
11.0593746004414 0.614657803741229\\
11.1593746004414 0.619799939934607\\
11.2593746004414 0.624942076127984\\
11.3593746004414 0.630084212321361\\
11.4593746004414 0.635226348514739\\
11.5593746004414 0.640368484708116\\
11.6593746004414 0.645510620901494\\
11.7593746004414 0.650652757094871\\
11.8593746004414 0.655794893288249\\
11.9593746004414 0.660937029481626\\
12.0593746004414 0.666079165675004\\
12.1593746004414 0.671221301868381\\
12.2593746004414 0.676363438061759\\
12.3593746004414 0.681505574255136\\
12.4593746004414 0.686647710448513\\
12.5593746004414 0.691789846641891\\
12.6593746004414 0.696931982835268\\
12.7593746004414 0.702074119028646\\
12.8593746004414 0.707216255222023\\
12.9593746004414 0.712358391415401\\
13.0593746004414 0.717500527608778\\
13.1593746004414 0.722642663802156\\
13.2593746004414 0.727784799995533\\
13.3593746004414 0.732926936188911\\
13.4593746004414 0.738069072382288\\
13.5593746004414 0.743211208575665\\
13.6593746004414 0.748353344769043\\
13.7593746004414 0.75349548096242\\
13.8593746004414 0.758637617155798\\
13.9593746004414 0.763779753349175\\
14.0593746004414 0.768921889542553\\
14.1593746004414 0.77406402573593\\
14.2593746004414 0.779206161929308\\
14.3593746004414 0.784348298122685\\
14.4593746004414 0.789490434316063\\
14.5593746004414 0.79463257050944\\
14.6593746004414 0.799774706702817\\
14.7593746004414 0.804916842896195\\
14.8593746004414 0.810058979089572\\
14.9593746004414 0.81520111528295\\
15.0593746004414 0.820343251476327\\
15.1593746004414 0.825485387669705\\
15.2593746004414 0.830627523863082\\
15.3593746004414 0.83576966005646\\
15.4593746004414 0.840911796249837\\
15.5593746004414 0.846053932443215\\
15.6593746004414 0.851196068636592\\
15.7593746004414 0.856338204829969\\
15.8593746004414 0.861480341023347\\
15.9593746004414 0.866622477216724\\
16.0593746004414 0.871764613410102\\
16.1593746004414 0.876906749603479\\
16.2593746004414 0.882048885796857\\
16.3593746004414 0.887191021990234\\
16.4593746004414 0.892333158183612\\
16.5593746004414 0.897475294376989\\
16.6593746004414 0.902617430570367\\
16.7593746004414 0.907759566763744\\
16.8593746004414 0.912901702957121\\
16.9593746004414 0.918043839150499\\
17.0593746004414 0.923185975343876\\
17.1593746004414 0.928328111537254\\
17.2593746004414 0.933470247730631\\
17.3593746004414 0.938612383924009\\
17.4593746004414 0.943754520117386\\
17.5593746004414 0.948896656310764\\
17.6593746004414 0.954038792504141\\
17.7593746004414 0.959180928697519\\
17.8593746004414 0.964323064890896\\
17.9593746004414 0.969465201084273\\
18.0593746004414 0.974607337277651\\
18.1593746004414 0.979749473471028\\
18.2593746004414 0.984891609664406\\
18.3593746004414 0.990033745857783\\
18.4593746004414 0.995175882051161\\
18.5593746004414 1.00031801824454\\
18.6593746004414 1.00546015443792\\
18.7593746004414 1.01060229063129\\
18.8593746004414 1.01574442682467\\
18.9593746004414 1.02088656301805\\
19.0593746004414 1.02602869921143\\
19.1593746004414 1.0311708354048\\
19.2593746004414 1.03631297159818\\
19.3593746004414 1.04145510779156\\
19.4593746004414 1.04659724398494\\
19.5593746004414 1.05173938017831\\
19.6593746004414 1.05688151637169\\
19.7593746004414 1.06202365256507\\
19.8593746004414 1.06716578875844\\
19.9593746004414 1.07230792495182\\
20.0593746004414 1.0774500611452\\
20.1593746004414 1.08259219733858\\
20.2593746004414 1.08773433353195\\
20.3593746004414 1.09287646972533\\
20.4593746004414 1.09801860591871\\
20.5593746004414 1.10316074211209\\
20.6593746004414 1.10830287830546\\
20.7593746004414 1.11344501449884\\
20.8593746004414 1.11858715069222\\
20.9593746004414 1.1237292868856\\
21.0593746004414 1.12887142307897\\
21.1593746004414 1.13401355927235\\
21.2593746004414 1.13915569546573\\
21.3593746004414 1.14429783165911\\
21.4593746004414 1.14943996785248\\
21.5593746004414 1.15458210404586\\
21.6593746004414 1.15972424023924\\
21.7593746004414 1.16486637643262\\
21.8593746004414 1.17000851262599\\
21.9593746004414 1.17515064881937\\
22.0593746004414 1.18029278501275\\
22.1593746004414 1.18543492120613\\
22.2593746004414 1.1905770573995\\
22.3593746004414 1.19571919359288\\
22.4593746004414 1.20086132978626\\
22.5593746004414 1.20600346597964\\
22.6593746004414 1.21114560217301\\
22.7593746004414 1.21628773836639\\
22.8593746004414 1.22142987455977\\
22.9593746004414 1.22657201075315\\
23.0593746004414 1.23171414694652\\
23.1593746004414 1.2368562831399\\
23.2593746004414 1.24199841933328\\
23.3593746004414 1.24714055552666\\
23.4593746004414 1.25228269172003\\
23.5593746004414 1.25742482791341\\
23.6593746004414 1.26256696410679\\
23.7593746004414 1.26770910030017\\
23.8593746004414 1.27285123649354\\
23.9593746004414 1.27799337268692\\
24.0593746004414 1.2831355088803\\
24.1593746004414 1.28827764507368\\
24.2593746004414 1.29341978126705\\
24.3593746004414 1.29856191746043\\
24.4593746004414 1.30370405365381\\
24.5593746004414 1.30884618984719\\
24.6593746004414 1.31398832604056\\
24.7593746004414 1.31913046223394\\
24.8593746004414 1.32427259842732\\
24.9593746004414 1.3294147346207\\
25.0593746004414 1.33455687081407\\
25.1593746004414 1.33969900700745\\
25.2593746004414 1.34484114320083\\
25.3593746004414 1.34998327939421\\
25.4593746004414 1.35512541558758\\
25.5593746004414 1.36026755178096\\
25.6593746004414 1.36540968797434\\
25.7593746004414 1.37055182416771\\
25.8593746004414 1.37569396036109\\
25.9593746004414 1.38083609655447\\
26.0593746004414 1.38597823274785\\
26.1593746004414 1.39112036894122\\
26.2593746004414 1.3962625051346\\
26.3593746004414 1.40140464132798\\
26.4593746004414 1.40654677752136\\
26.5593746004414 1.41168891371473\\
26.6593746004414 1.41683104990811\\
26.7593746004414 1.42197318610149\\
26.8593746004414 1.42711532229487\\
26.9593746004414 1.43225745848824\\
27.0593746004414 1.43739959468162\\
27.1593746004414 1.442541730875\\
27.2593746004414 1.44768386706838\\
27.3593746004414 1.45282600326175\\
27.4593746004414 1.45796813945513\\
27.5593746004414 1.46311027564851\\
27.6593746004414 1.46825241184189\\
27.7593746004414 1.47339454803526\\
27.8593746004414 1.47853668422864\\
27.9593746004414 1.48367882042202\\
28.0593746004414 1.4888209566154\\
28.1593746004414 1.49396309280877\\
28.2593746004414 1.49910522900215\\
28.3593746004414 1.50424736519553\\
28.4593746004414 1.50938950138891\\
28.5593746004414 1.51453163758228\\
28.6593746004414 1.51967377377566\\
28.7593746004414 1.52481590996904\\
28.8593746004414 1.52995804616242\\
28.9593746004414 1.53510018235579\\
29.0593746004414 1.54024231854917\\
29.1593746004414 1.54538445474255\\
29.2593746004414 1.55052659093593\\
29.3593746004414 1.5556687271293\\
29.4593746004414 1.56081086332268\\
29.5593746004414 1.56595299951606\\
29.6593746004414 1.57109513570944\\
29.7593746004414 1.57623727190281\\
29.8593746004414 1.58137940809619\\
29.9593746004414 1.58652154428957\\
30.0593746004414 1.59166368048295\\
30.1593746004414 1.59680581667632\\
30.2593746004414 1.6019479528697\\
30.3593746004414 1.60709008906308\\
30.4593746004414 1.61223222525646\\
30.5593746004414 1.61737436144983\\
30.6593746004414 1.62251649764321\\
30.7593746004414 1.62765863383659\\
30.8593746004414 1.63280077002996\\
30.9593746004414 1.63794290622334\\
31.0593746004414 1.64308504241672\\
31.1593746004414 1.6482271786101\\
31.2593746004414 1.65336931480347\\
31.3593746004414 1.65851145099685\\
31.4593746004414 1.66365358719023\\
31.5593746004414 1.66879572338361\\
31.6593746004414 1.67393785957698\\
31.7593746004414 1.67907999577036\\
31.8593746004414 1.68422213196374\\
31.9593746004414 1.68936426815712\\
32.0593746004414 1.69450640435049\\
32.1593746004414 1.69964854054387\\
32.2593746004414 1.70479067673725\\
32.3593746004414 1.70993281293063\\
32.4593746004414 1.715074949124\\
32.5593746004414 1.72021708531738\\
32.6593746004414 1.72535922151076\\
32.7593746004414 1.73050135770414\\
32.8593746004414 1.73564349389751\\
32.9593746004414 1.74078563009089\\
33.0593746004414 1.74592776628427\\
33.1593746004414 1.75106990247765\\
33.2593746004414 1.75621203867102\\
33.3593746004414 1.7613541748644\\
33.4593746004414 1.76649631105778\\
33.5593746004414 1.77163844725116\\
33.6593746004414 1.77678058344453\\
33.7593746004414 1.78192271963791\\
33.8593746004414 1.78706485583129\\
33.9593746004414 1.79220699202467\\
34.0593746004414 1.79734912821804\\
34.1593746004414 1.80249126441142\\
34.2593746004414 1.8076334006048\\
34.3593746004414 1.81277553679818\\
34.4593746004414 1.81791767299155\\
34.5593746004414 1.82305980918493\\
34.6593746004414 1.82820194537831\\
34.7593746004414 1.83334408157169\\
34.8593746004414 1.83848621776506\\
34.9593746004414 1.84362835395844\\
35.0593746004414 1.84877049015182\\
35.1593746004414 1.8539126263452\\
35.2593746004414 1.85905476253857\\
35.3593746004414 1.86419689873195\\
35.4593746004414 1.86933903492533\\
35.5593746004414 1.87448117111871\\
35.6593746004414 1.87962330731208\\
35.7593746004414 1.88476544350546\\
35.8593746004414 1.88990757969884\\
35.9593746004414 1.89504971589222\\
36.0593746004414 1.90019185208559\\
36.1593746004414 1.90533398827897\\
36.2593746004414 1.91047612447235\\
36.3593746004414 1.91561826066572\\
36.4593746004414 1.9207603968591\\
36.5593746004414 1.92590253305248\\
36.6593746004414 1.93104466924586\\
36.7593746004414 1.93618680543923\\
36.8593746004414 1.94132894163261\\
36.9593746004414 1.94647107782599\\
37.0593746004414 1.95161321401937\\
37.1593746004414 1.95675535021274\\
37.2593746004414 1.96189748640612\\
37.3593746004414 1.9670396225995\\
37.4593746004414 1.97218175879288\\
37.5593746004414 1.97732389498625\\
37.6593746004414 1.98246603117963\\
37.7593746004414 1.98760816737301\\
37.8593746004414 1.99275030356639\\
37.9593746004414 1.99789243975976\\
38.0593746004414 2.00303457595314\\
38.1593746004414 2.00817671214652\\
38.2593746004414 2.0133188483399\\
38.3593746004414 2.01846098453327\\
38.4593746004414 2.02360312072665\\
38.5593746004414 2.02874525692003\\
};
\addlegendentry{Fit};
\end{axis}
\node[draw=none] at (3.3,0.35) {\footnotesize $f_{\text{s}}L_{\text{eff}}=1.8 \,\text{mm}$};
\node[draw=none] at (3.3,0.75) {\footnotesize $N = 85$};
\node[draw=none] at (3.3,1.15) {\footnotesize $\tilde{\mathcal{G}} = 6561 \, \text{W}^{-1}\text{m}^{-1}$};
\node[draw=none] at (3.3,1.55) {\footnotesize $Q= 464$};
\node[draw=none] at (0.75,3.3) {\footnotesize Net gain};
\node[draw=none] at (0.3,2.63) {\footnotesize $\alpha L$};
\node at (-0.7,-0.7) {\large{\textbf{b}}};
\draw[black!20,fill,opacity=0.2] (0.0,2.45) rectangle (4.16,3.49);
\end{tikzpicture}%
}
\caption{\textbf{Brillouin gain exceeding the optical losses.} \textbf{a}, An example of a Brillouin gain resonance, in this case with an on/off gain of $1.4 \, \text{dB}$, quality factor of $Q_{\text{m}} = 728$ and an on-chip input pump power of $26 \, \text{mW}$. The shaded black area indicates uncertainty in the probe power. \textbf{b}, Scan of the on/off gain with pump power. At a pump power of $30 \, \text{mW}$ the transparency point is reached. For $P_{\text{p}} > 30 \, \text{mW}$, more probe photons leave than enter the waveguide. The slope yields the Brillouin gain coefficient $\tilde{\mathcal{G}} = 6561 \, \text{W}^{-1}\text{m}^{-1}$ with a quality factor of $464$ in this particular waveguide. Notably, the on/off gain scales linearly with pump power across the entire sweep -- indicating the absence of free-carrier absorption in this range.}
\label{fig:netgain}
\end{figure*}
First, we measured the amplitude response of our system. We injected a weak probe red-detuned from a strong pump and retrieved the probe power as a function of detuning (fig.\ref{fig:netgain}a). As before \cite{VanLaer2015}, we find gain resonances around $9.1 \, \text{GHz}$. The on/off gain increases with pump power (fig.\ref{fig:netgain}b) and reaches the transparency point $\tilde{\mathcal{G}} P_{\text{p}} = \alpha$ around $P_{\text{p}} = 30 \, \text{mW}$. Beyond this pump power, the ouput exceeds the input probe photon flux. At the maximum pump power of $39 \, \text{mW}$, we obtain guided-wave cooperativities \cite{VanLaer2015b} of $\tilde{\mathcal{C}} = \frac{(f_{\text{s}}\tilde{\mathcal{G}}) P_{\text{p}}}{\alpha} = 1.7$. This modest net gain of $0.5 \, \text{dB}$ (fig.\ref{fig:netgain}b) is a step towards selective on-chip amplifiers that could be used for homodyne detection, in order to eliminate the requirement of a phase-stabilized local oscillator \cite{Atkins1986}.
Notably, the linear scaling between on/off gain and pump power (fig.\ref{fig:netgain}b) indicates the absence of free-carrier absorption up to $40 \, \text{mW}$ \cite{Wolff2015a,Wolff2015b}. In contrast, we previously measured increased nonlinear absorption already at $25 \, \text{mW}$ in silicon wires on a pillar \cite{VanLaer2015}. Both this finding and the higher propagation losses ($5.5 \, \text{dB/cm}$ instead of $2.6 \, \text{dB/cm}$ \cite{VanLaer2015}) likely originate in a deterioration of the wires' surface state during the fabrication of the suspensions. In agreement with this hypothesis, we measured a drop in the free-carrier lifetime (see Appendix).
In case this structure were to be placed in a cavity, such as a silicon microring, it would also have to overcome coupling losses to achieve the photon/phonon lasing threshold ($\tilde{\mathcal{C}} > \mathcal{C} =1$ \cite{VanLaer2015b}). We note that the acoustic linewidth ($\sim 10 \, \text{MHz}$) is a factor $10^{2}$ smaller than typical optical linewidths of silicon microcavities ($\sim 1 \, \text{GHz}$). Therefore, this would produce stimulated emission of phonons, not photons \cite{Grudinin2010,Shen1965,Nunnenkamp2014,VanLaer2015b,Aspelmeyer2014,Cohen2015}. Such a device would not benefit from the spectral purification associated with Brillouin lasers \cite{Debut2000,Debut2001}. The origin of this reversal of the damping hierarchy (going from waveguides to cavities) \cite{VanLaer2015b,Nunnenkamp2014} lies in the exceedingly low group velocity of these Raman-like \cite{Kang2009} acoustic phonons; indeed, despite enormously higher propagation losses they usually still have lower linewidths than photons \cite{Aspelmeyer2014,Nunnenkamp2014,VanLaer2015b}. Only uniquely high-quality optical cavities, to date realized only using silica \cite{Lee2012a,Li2014,Stokes1982,Li2012c,Kabakova2013} or crystalline \cite{Grudinin2009} materials, can produce lower photonic than phononic damping rates.
\subsubsection{Geometric disorder.}
\label{subsubsec:disorder}
Next, we study the quality factor extracted from the gain resonances (fig.\ref{fig:netgain}a). We find that it strongly decreases with the number of suspensions $N$ (fig.\ref{fig:qualityfactor1}a); from $Q_{\text{m}} \approx 10^{3}$ at $N = 6$ to $Q_{\text{m}} \approx 400$ at $N = 85$. For larger numbers of suspensions in a spiral configuration ($N = 1332$, not shown), the quality factor levels off around $Q_{\text{m}} \approx 340$. Notably, this relation changes from sample to sample -- even if they originate from the same wafer (fig.\ref{fig:qualityfactor1}a). We attribute such variations to inhomogeneous broadening by geometric disorder, presumably in the width of the nanowires \cite{VanLaer2015}. Indeed, the sensitivity of the resonance frequency $\frac{\Omega_{\text{m}}}{2\pi}$ to width variations is $\frac{1}{2\pi} \frac{\text{d}\Omega_{\text{m}}}{\text{d}w} \approx 20 \, \text{MHz/nm}$ \cite{VanLaer2015}. Therefore, realistic width variations $\delta w$ of about $0.5 \, \text{nm}$ \cite{Selvaraja2009} yield inhomogeneous linewidths of about $\frac{\Gamma_{\text{inh}}}{2\pi} \approx 10 \, \text{MHz}$ -- comparable to those measured (fig.\ref{fig:netgain}a). Similar disorder has been studied in snowflake crystals \cite{Safavi-naeini2014}.
\begin{figure*}
\centering
\subfigure{
\begin{tikzpicture}
\begin{axis}[scale = 0.34*\textwidth/(4.5in),
width=4.52083333333333in,
height=3.565625in,
xmin=0,
xmax=100,
xlabel={Number of suspensions $N$ (-)},
ymin=255,
ymax=1050,
ylabel={Quality factor $Q_{\text{m}}$ (-)},
ylabel style = {yshift=0.75ex},
axis x line*=bottom,
axis y line*=left,
legend style={fill=none,draw=none,legend cell align=left,at={(0.45,1.08)},anchor=north},
xtick = {30,60,...,90},
ytick = {250,500,750,1000},
minor xtick={15,30,...,90},
minor ytick={125,250,...,1000}
]
\addplot [
color=blue,
line width=0.75pt,
mark size=3.0pt,
dotted,
mark=x,
mark options={solid}
]
table[row sep=crcr]{
6 1010 \\
19 776 \\
66 502 \\
85 456 \\
293 300 \\
};
\addlegendentry{A};
\addplot [
color=green!50!black,
line width=0.75pt,
mark size=3.0pt,
dotted,
mark=o,
mark options={solid}
]
table[row sep=crcr]{
19 740 \\
66 731 \\
85 591 \\
293 417 \\
};
\addlegendentry{B};
\addplot [
color=purple!50!black,
line width=0.75pt,
mark size=3.0pt,
dotted,
mark=triangle,
mark options={solid}
]
table[row sep=crcr]{
6 919 \\
19 681 \\
66 488 \\
85 421 \\
283 296 \\
283 320 \\
};
\addlegendentry{C};
\end{axis}
\node at (-1.1,-0.9) {\large{\textbf{a}}};
\node[draw=none] at (3.5,0.29) {\footnotesize $w = 450 \, \text{nm}$};
\end{tikzpicture}%
}
\subfigure{
\begin{tikzpicture}
\begin{axis}[scale = 0.34*\textwidth/(4.5in),
width=4.52083333333333in,
height=3.565625in,
xmin=330,
xmax=520,
xlabel={Wire width $w$ (nm)},
ymin=255,
ymax=1050,
ylabel={Quality factor $Q_{\text{m}}$ (-)},
ylabel style = {yshift=0.75ex},
axis x line*=bottom,
axis y line*=left,
legend style={fill=none,draw=none,legend cell align=left,at={(0.45,1.08)},anchor=north},
xtick = {350,400,...,500},
ytick = {250,500,750,1000},
minor xtick={325,350,...,525},
minor ytick={125,250,...,1000}
]
\addplot [
color=blue,
line width=0.75pt,
mark size=3.0pt,
dotted,
mark=x,
mark options={solid}
]
table[row sep=crcr]{
350 417 \\
400 405 \\
450 502 \\
};
\addlegendentry{A};
\addplot [
color=green!50!black,
line width=0.75pt,
mark size=3.0pt,
dotted,
mark=o,
mark options={solid}
]
table[row sep=crcr]{
350 602 \\
400 517 \\
450 731 \\
500 822 \\
};
\addlegendentry{B};
\addplot [
color=purple!50!black,
line width=0.75pt,
mark size=3.0pt,
dotted,
mark=triangle,
mark options={solid}
]
table[row sep=crcr]{
450 488 \\
500 654 \\
};
\addlegendentry{C};
\end{axis}
\node at (-0.9,-0.9) {\large{\textbf{b}}};
\node[draw=none] at (3.5,0.29) {\footnotesize $N = 66$};
\end{tikzpicture}%
}
\caption{\textbf{The quality factor decreases with the number of suspensions.} We study the phonon quality factor for three samples (A, B and C) from the same wafer. The samples were designed to be identical. \textbf{a}, The quality factor increases up to 1010 when there are only 6 suspensions. For larger $N$, the quality factors approach $\approx 400$. Unless stated otherwise, all resonances are still well-fit by a Lorentzian function (fig.\ref{fig:netgain}a). \textbf{b}, In general, wider waveguides exhibit slightly larger $Q_{\text{m}}$. However, this pattern is neither linear (samples A/C) nor monotonic (sample B). Some waveguides were defective, possibly because of a collapsed beam, and were excluded from the study.}
\label{fig:qualityfactor1}
\end{figure*}
Further, we investigated the influence of the width $w$ on the quality factor $Q_{\text{m}}$. Since the resonance frequency scales inversely with width ($\frac{\Omega_{\text{m}}}{2\pi} \propto w^{-1}$) \cite{VanLaer2015}, its sensitivity scales inverse quadratically with width ($\frac{\text{d}\Omega_{\text{m}}}{\text{d}w} \propto w^{-2}$). Subsequently, the inhomogeneously broadened linewidth scales similarly ($\Gamma_{\text{inh}} \propto w^{-2}$) in case the size of the width variations $\delta w$ does not depend on $w$. Then the quality factor scales linearly with width ($Q_{\text{m}} = \frac{\Omega_{\text{m}}}{\Gamma_{\text{m}}} \propto w$). We indeed observe overall larger quality factors for wider wires, although this pattern is neither linear nor monotonic (fig.\ref{fig:qualityfactor1}b). We note that, unless stated otherwise, all resonances were still well-fit by a Lorentzian function (fig.\ref{fig:netgain}a). In case of sufficiently sampled geometric disorder, the gain curves would become convolutions of a Lorentzian and a probability function describing the geometric disorder (e.g. distribution of the width $w$). The largest deviations of such Voigt curves with respect to a Lorentzian occur in the tails (large $\Delta = \frac{\Omega - \Omega_{\text{m}}}{\Gamma_{\text{m}}}$), precisely where the relative uncertainty in the measured probe power is highest (fig.\ref{fig:netgain}a). Given this uncertainty, both Lorentzian and e.g. Gaussian-shaped curves produce good fits to the gain resonances. A low-temperature characterization would yield more information regarding the nature of the acoustic broadening mechanisms.
\begin{figure*}
\centering
\subfigure{
\begin{tikzpicture}
\begin{axis}[scale = 0.4*\textwidth/(4.5in),
width=4.52083333333333in,
height=3.565625in,
view={45}{30},
scale only axis,
xmin=9.07,
xmax=9.14,
xlabel={Frequency $\frac{\Omega}{2\pi}$ (GHz)},
xlabel style = {yshift=1.3ex,rotate=-22},
xmajorgrids,
ymin=0,
ymax=100,
ylabel={Anchor $L_{\text{a}}$ ($\mu\text{m}$)},
ylabel style = {yshift=1.3ex,rotate=22},
ymajorgrids,
zmin=-0.1,
zmax=0.6,
zlabel={Probe power (dB)},
zmajorgrids,
xtick = {9.08,9.10,9.12},
ytick = {0,25,50,75,100},
ztick = {0,0.2,0.4,0.6},
minor xtick={9.08,9.09,...,9.13},
minor ytick={12.5,25,...,100},
minor ztick = {0.1,0.2,...,0.6}
]
\addplot3 [
color=blue,
solid,
fill opacity=0.15,
draw=blue!80!white,thin,
fill=blue,
mark=o,
mark size=0.5pt]
table[row sep=crcr] {
9.07999999999996 94 -0.0786609311438378\\
9.08099999999996 94 -0.0786744504580251\\
9.08199999999995 94 -0.0746949497722147\\
9.08299999999995 94 -0.0730380790863992\\
9.08399999999995 94 -0.0724968184005901\\
9.08499999999995 94 -0.0716914177147752\\
9.08599999999995 94 -0.0657319670289658\\
9.08699999999995 94 -0.0591371063431519\\
9.08799999999995 94 -0.0482120356573388\\
9.08899999999995 94 -0.0369092449715263\\
9.08999999999995 94 -0.0240059342857136\\
9.09099999999995 94 -0.00934282359990228\\
9.09199999999995 94 0.00560354708591056\\
9.09299999999995 94 0.0228707177717231\\
9.09399999999995 94 0.0352767584575355\\
9.09499999999995 94 0.0451413291433473\\
9.09599999999995 94 0.0593782198291608\\
9.09699999999995 94 0.0768910105149709\\
9.09799999999995 94 0.105044661200785\\
9.09899999999995 94 0.144359681886597\\
9.09999999999994 94 0.19802983257241\\
9.10099999999994 94 0.262851033258221\\
9.10199999999994 94 0.330976943944035\\
9.10299999999994 94 0.382383504629847\\
9.10399999999994 94 0.404545715315659\\
9.10499999999994 94 0.407804746001473\\
9.10599999999994 94 0.405419506687283\\
9.10699999999994 94 0.379123767373097\\
9.10799999999994 94 0.322377308058908\\
9.10899999999994 94 0.257535298744721\\
9.10999999999994 94 0.208342459430534\\
9.11099999999994 94 0.180532960116346\\
9.11199999999994 94 0.159902270802159\\
9.11299999999994 94 0.12532030148797\\
9.11399999999994 94 0.081263712173782\\
9.11499999999994 94 0.0363905328595946\\
9.11599999999994 94 0.00199230354540965\\
9.11699999999994 94 -0.0217956957687813\\
9.11799999999993 94 -0.0358683850829662\\
9.11899999999993 94 -0.0448201043971564\\
9.11999999999993 94 -0.0528515637113421\\
9.12099999999993 94 -0.0601249230255308\\
9.12199999999993 94 -0.0670953423397167\\
9.12299999999993 94 -0.0703667516539044\\
9.12399999999993 94 -0.0699679609680937\\
9.12499999999993 94 -0.0700548702822817\\
9.12599999999993 94 -0.070705579596468\\
9.12699999999993 94 -0.0727631889106571\\
9.12799999999993 94 -0.0751595982248441\\
9.12899999999993 94 -0.0740346075390308\\
9.12999999999993 94 -0.0783090168532193\\
};
\addplot3 [
color=red,
solid,
fill opacity=0.15,
draw=blue!80!white,thin,
fill=red,
mark=o,
mark size=0.5pt]
table[row sep=crcr] {
9.07999999999996 19 -0.113023371801669\\
9.08099999999996 19 -0.118905594764357\\
9.08199999999995 19 -0.0914400177270475\\
9.08299999999995 19 -0.107962640689739\\
9.08399999999995 19 -0.0892998636524289\\
9.08499999999995 19 -0.0770502866151196\\
9.08599999999995 19 -0.0801867095778099\\
9.08699999999995 19 -0.0594588325405005\\
9.08799999999995 19 -0.0499225555031912\\
9.08899999999995 19 -0.0558140784658813\\
9.08999999999995 19 -0.0287112014285717\\
9.09099999999995 19 -0.0117763243912605\\
9.09199999999995 19 0.0294069626460474\\
9.09299999999995 19 0.0540369596833579\\
9.09399999999995 19 0.0716688167206663\\
9.09499999999995 19 0.0957029737579761\\
9.09599999999995 19 0.118419770795288\\
9.09699999999995 19 0.134096157832596\\
9.09799999999995 19 0.128469074869906\\
9.09899999999995 19 0.114085581907214\\
9.09999999999994 19 0.137665178944525\\
9.10099999999994 19 0.158755305981836\\
9.10199999999994 19 0.197571933019144\\
9.10299999999994 19 0.258871500056453\\
9.10399999999994 19 0.319981277093765\\
9.10499999999994 19 0.355030554131074\\
9.10599999999994 19 0.347007221168384\\
9.10699999999994 19 0.317654308205692\\
9.10799999999994 19 0.254175405243001\\
9.10899999999994 19 0.202018372280313\\
9.10999999999994 19 0.174958759317622\\
9.11099999999994 19 0.147535056354932\\
9.11199999999994 19 0.12143019339224\\
9.11299999999994 19 0.0822457904295491\\
9.11399999999994 19 0.0405296474668611\\
9.11499999999994 19 0.00610006450417029\\
9.11599999999994 19 -0.0258022984585213\\
9.11699999999994 19 -0.0573945214212112\\
9.11799999999993 19 -0.0763923443839018\\
9.11899999999993 19 -0.0752874673465917\\
9.11999999999993 19 -0.0768932903092813\\
9.12099999999993 19 -0.0587937132719715\\
9.12199999999993 19 -0.058210436234664\\
9.12299999999993 19 -0.0502086591973547\\
9.12399999999993 19 -0.0539866821600429\\
9.12499999999993 19 -0.056291405122734\\
9.12599999999993 19 -0.0598518280854249\\
9.12699999999993 19 -0.0650635510481147\\
9.12799999999993 19 -0.0640867740108059\\
9.12899999999993 19 -0.0713871969734955\\
9.12999999999993 19 -0.0767040199361861\\
};
\addplot3 [
color=blue,
solid,
fill opacity=0.15,
draw=blue!80!white,thin,
fill=red,
mark=o,
mark size=0.5pt]
table[row sep=crcr] {
9.09099999999995 44 -0.0900561832253353\\
9.09199999999995 44 -0.0854510968075596\\
9.09299999999995 44 -0.0940109703897901\\
9.09399999999995 44 -0.112443143972015\\
9.09499999999995 44 -0.115533267554246\\
9.09599999999995 44 -0.0989217711364775\\
9.09699999999995 44 -0.0773434147187005\\
9.09799999999995 44 -0.0441155283009328\\
9.09899999999995 44 -0.0133052318831635\\
9.09999999999994 44 0.0123270445346121\\
9.10099999999994 44 0.0184664109523804\\
9.10199999999994 44 0.03728182737015\\
9.10299999999994 44 0.0396830437879251\\
9.10399999999994 44 0.0437240802056943\\
9.10499999999994 44 0.0495511866234632\\
9.10599999999994 44 0.0450039630412379\\
9.10699999999994 44 0.0260672194590074\\
9.10799999999994 44 0.0125038358767823\\
9.10899999999994 44 0.0755943722945519\\
9.10999999999994 44 0.15046394871232\\
9.11099999999994 44 0.220513675130095\\
9.11199999999994 44 0.209566401547865\\
9.11299999999994 44 0.213923387965633\\
9.11399999999994 44 0.33043026438341\\
9.11499999999994 44 0.395639580801178\\
9.11599999999994 44 0.421532377218946\\
9.11699999999994 44 0.426101223636723\\
9.11799999999993 44 0.393637670054491\\
9.11899999999993 44 0.35396375647226\\
9.11999999999993 44 0.299354752890035\\
9.12099999999993 44 0.241393259307805\\
9.12199999999993 44 0.123203955725579\\
9.12299999999993 44 0.0115733621433491\\
9.12399999999993 44 -0.0116723214388828\\
9.12499999999993 44 0.0061661649788937\\
9.12599999999993 44 0.022224231396662\\
9.12699999999993 44 -0.00659356218556913\\
9.12799999999993 44 -0.0253956757677945\\
9.12899999999993 44 -0.0344057893500243\\
9.12999999999993 44 -0.113762302932256\\
9.13099999999993 44 -0.0834270165144802\\
9.13199999999993 44 -0.061919430096711\\
9.13299999999993 44 -0.0788278436789434\\
9.13399999999993 44 -0.0704869572611682\\
9.13499999999993 44 -0.053120770843399\\
9.13599999999992 44 -0.0486149844256237\\
9.13699999999992 44 -0.0359440980078543\\
9.13799999999992 44 -0.0434066115900856\\
9.13899999999992 44 -0.0232174251723099\\
9.13999999999992 44 -0.0281348387545404\\
9.14099999999992 44 -0.00757165233677289\\
};
\addplot3 [
color=blue,
solid,
fill opacity=0.15,
draw=blue!80!white,thin,
fill=blue,
mark=o,
mark size=0.5pt]
table[row sep=crcr] {
9.08999999999995 69 -0.0773910948208139\\
9.09099999999995 69 -0.0763617761958768\\
9.09199999999995 69 -0.0776454575709389\\
9.09299999999995 69 -0.0853154389460001\\
9.09399999999995 69 -0.0877803203210617\\
9.09499999999995 69 -0.080633601696122\\
9.09599999999995 69 -0.0699862830711835\\
9.09699999999995 69 -0.0593626644462431\\
9.09799999999995 69 -0.0461357458213057\\
9.09899999999995 69 -0.0343074271963664\\
9.09999999999994 69 -0.0212750085714282\\
9.10099999999994 69 -0.00246988994649038\\
9.10199999999994 69 0.0116468286784483\\
9.10299999999994 69 0.0191361473033877\\
9.10399999999994 69 0.0260098659283256\\
9.10499999999994 69 0.0392652845532657\\
9.10599999999994 69 0.0618625231782035\\
9.10699999999994 69 0.0839697918031427\\
9.10799999999994 69 0.116363470428081\\
9.10899999999994 69 0.155558339053019\\
9.10999999999994 69 0.204852457677958\\
9.11099999999994 69 0.258489996302898\\
9.11199999999994 69 0.321300204927836\\
9.11299999999994 69 0.377807373552776\\
9.11399999999994 69 0.426986712177714\\
9.11499999999994 69 0.469048520802652\\
9.11599999999994 69 0.455023179427591\\
9.11699999999994 69 0.360812068052529\\
9.11799999999993 69 0.256048876677467\\
9.11899999999993 69 0.177369385302406\\
9.11999999999993 69 0.133661673927345\\
9.12099999999993 69 0.101552312552285\\
9.12199999999993 69 0.0644312011772223\\
9.12299999999993 69 0.0318148198021619\\
9.12399999999993 69 0.00968793842709999\\
9.12499999999993 69 -0.00672124294796199\\
9.12599999999993 69 -0.0206025243230235\\
9.12699999999993 69 -0.0274312056980835\\
9.12799999999993 69 -0.0330104870731463\\
9.12899999999993 69 -0.0398876684482064\\
9.12999999999993 69 -0.0429247498232668\\
9.13099999999993 69 -0.0510874311983285\\
9.13199999999993 69 -0.0516247125733918\\
9.13299999999993 69 -0.0520886939484529\\
9.13399999999993 69 -0.0530199753235144\\
9.13499999999993 69 -0.0511426566985742\\
9.13599999999992 69 -0.0567129380736367\\
9.13699999999992 69 -0.0657816194486954\\
9.13799999999992 69 -0.0746452008237588\\
9.13899999999992 69 -0.0695852821988199\\
9.13999999999992 69 -0.0657768635738814\\
};
\addplot3 [
color=blue,
solid,
fill opacity=0.15,
draw=blue!80!white,thin,
fill=blue,
mark=o,
mark size=0.5pt]
table[row sep=crcr] {
9.07499999999996 4 -0.0883470481688769\\
9.07599999999996 4 -0.0888383567805592\\
9.07699999999996 4 -0.0890412653922426\\
9.07799999999996 4 -0.0877920740039263\\
9.07899999999996 4 -0.0869354826156113\\
9.07999999999996 4 -0.0845144912272951\\
9.08099999999996 4 -0.0811535998389784\\
9.08199999999995 4 -0.0738554084506629\\
9.08299999999995 4 -0.0677924170623457\\
9.08399999999995 4 -0.0538242256740305\\
9.08499999999995 4 -0.0445967342857146\\
9.08599999999995 4 -0.0275446428973976\\
9.08699999999995 4 -0.00611181150908221\\
9.08799999999995 4 0.00830329987923287\\
9.08899999999995 4 0.0270409312675503\\
9.08999999999995 4 0.0467288526558664\\
9.09099999999995 4 0.0786127540441838\\
9.09199999999995 4 0.111053905432498\\
9.09299999999995 4 0.150769076820814\\
9.09399999999995 4 0.204481228209132\\
9.09499999999995 4 0.253357509597447\\
9.09599999999995 4 0.299895640985762\\
9.09699999999995 4 0.34303987237408\\
9.09799999999995 4 0.402820603762394\\
9.09899999999995 4 0.45315578515071\\
9.09999999999994 4 0.471736316539028\\
9.10099999999994 4 0.455103497927342\\
9.10199999999994 4 0.40816671931566\\
9.10299999999994 4 0.348061430703976\\
9.10399999999994 4 0.292916092092292\\
9.10499999999994 4 0.239066903480608\\
9.10599999999994 4 0.190037644868924\\
9.10699999999994 4 0.145431496257241\\
9.10799999999994 4 0.0893221276455562\\
9.10899999999994 4 0.0422155390338729\\
9.10999999999994 4 0.0104115604221887\\
9.11099999999994 4 -0.0150027581894954\\
9.11199999999994 4 -0.0299826668011794\\
9.11299999999994 4 -0.0404760754128629\\
9.11399999999994 4 -0.042507084024549\\
9.11499999999994 4 -0.0517887926362318\\
9.11599999999994 4 -0.0580488012479149\\
9.11699999999994 4 -0.0685982098596008\\
9.11799999999993 4 -0.0769078184712833\\
9.11899999999993 4 -0.0829425270829661\\
9.11999999999993 4 -0.0887848356946493\\
9.12099999999993 4 -0.0873864443063355\\
9.12199999999993 4 -0.0860610529180187\\
9.12299999999993 4 -0.0854752615297011\\
9.12399999999993 4 -0.0855353701413873\\
9.12499999999993 4 -0.083892178753071\\
};
\end{axis}
\node at (-1.1,-0.7) {\large{\textbf{a}}};
\end{tikzpicture}%
}
\subfigure{
\begin{tikzpicture}
\begin{axis}[scale = 0.33*\textwidth/(4.5in),
width=4.52083333333333in,
height=3.565625in,
xmin=0,
xmax=105,
xlabel={Anchor length $L_{\text{a}}$ ($\mu$m)},
ymin=255,
ymax=1050,
ylabel={Quality factor $Q_{\text{m}}$ (-)},
ylabel style = {yshift=0.75ex},
axis x line*=bottom,
axis y line*=left,
legend style={fill=none,draw=none,legend cell align=left,at={(0.40,0.88)},anchor=north},
xtick = {25,50,...,100},
ytick = {250,500,750,1000},
minor xtick={12.5,25,...,100},
minor ytick={125,250,...,1000}
]
\addplot [
color=green!50!black,
line width=0.75pt,
mark size=3.0pt,
dotted,
mark=o,
mark options={solid}
]
table[row sep=crcr]{
2 740 \\
4 731 \\
19 654 \\
44 548 \\
69 897 \\
94 738 \\
};
\addlegendentry{B};
\end{axis}
\node at (-1.1,-0.9) {\large{\textbf{b}}};
\node[draw=none] at (3.5,0.29) {\footnotesize $L_{\text{s}} = 25 \, \mu \text{m}$};
\node[draw=none] at (3.5,0.7) {\footnotesize $N = 19$};
\begin{scope}[shift={(-0.05,-0.1)}]
\begin{scope}[rotate around={-25:(0.965,1.4)},shift={(0.33,0.28)}]
\draw[black] (0.965,1.4) ellipse (8.7mm and 4.0mm);
\draw[black!20,fill,opacity=0.3] (0.965,1.4) ellipse (8.7mm and 4.0mm);
\end{scope}
\draw[black] (2.16,1.13) -- (2.6,1.2);
\node[draw=none] at (3.05,1.25) {split};
\end{scope}
\end{tikzpicture}%
}
\caption{\textbf{The phonon resonance splits at certain anchor lengths.} \textbf{a}, As we sweep the anchor length, the initially clean curve splits at $L_{\text{a}} = 19 \, \mu\text{m}$ and $44 \, \mu\text{m}$ but recombines at $L_{\text{a}} = 69 \, \mu\text{m}$. The pump power was $P_{\text{p}} = 26 \, \text{mW}$ and the position of the first suspension was fixed in this sweep. \textbf{b}, A Lorentzian fit to the gain curves of (\textbf{a}) yields high quality factors at short and long anchors. We suspect that (\textbf{a}) and (\textbf{b}) arise from a nanometer-scale ($\delta w$) width fluctuation in this straight silicon wire.}
\label{fig:qualityfactor2}
\end{figure*}
There are two types of potential width fluctuations: (1) fast sidewall roughness with a coherence length $L_{\text{coh}}$ of only $\approx 50 \, \text{nm}$ \cite{Lee2000} and (2) slow variations in the average waveguide width $w$. We suspect that mechanism (2) is at play here, since even an individual section ($L_{\text{sec}} \approx 30 \, \mu\text{m}$) is much larger than the coherence length ($\approx 50 \, \text{nm}$ \cite{Lee2000}) of the surface roughness. Therefore, sidewall roughness cannot explain the significant changes of $Q_{\text{m}}$ with $N$ (fig.\ref{fig:qualityfactor1}a): even a single section samples it fully ($L_{\text{sec}}/L_{\text{coh}} \approx 600$). In contrast, slow excursions of the waveguide width are consistent with such behavior. We confirm this by scanning the anchor length $L_{\text{a}}$ while keeping the number of suspensions $N$ and the suspension length $L_{\text{s}}$ constant (fig.\ref{fig:qualityfactor2}). In this sweep, the position of the first suspension is fixed. As $L_{\text{a}}$ increases, the initially clean resonance first splits at $L_{\text{a}} = 19 \, \mu\text{m}$ and then recombines at $L_{\text{a}} = 69 \, \mu\text{m}$ (fig.\ref{fig:qualityfactor2}a). Remarkably, the $L_{\text{a}} = 69 \, \mu\text{m}$ wire even produces the highest quality factor (fig.\ref{fig:qualityfactor2}b). In light of the above discussion, this behavior likely stems from a nanometer-scale ($\delta w$) width excursion: short and long anchor waveguides avoid the width fluctuations and thus yield clean profiles. Both fig.\ref{fig:qualityfactor1}$\&$\ref{fig:qualityfactor2} are fingerprints of geometric disorder that hinders the development of integrated Brillouin-based technologies.
\begin{figure*}
\centering
\subfigure{
\begin{tikzpicture}
\begin{axis}[scale = 0.34*\textwidth/(4.5in),
width=4.52083333333333in,
height=3.565625in,
xmin=8.98,
xmax=9.22,
xlabel={Frequency $\frac{\Omega}{2\pi}$ (GHz)},
xlabel style = {yshift=0ex},
ymin=-6,
ymax=5,
ylabel= {Sideband power (dB)},
ylabel style = {yshift=-1.5ex},
axis x line*=bottom,
axis y line*=left,
legend style={fill=none,draw=none,legend cell align=left,at={(0.5,1.2)},anchor=north, legend columns=-1},
xtick = {9,9.1,9.2},
ytick = {-4,-2,0,2,4},
minor xtick={9,9.05,...,9.25},
minor ytick={-5,-4,...,5}
]
\addplot [
color=blue,
line width=0.5pt,
mark size=2.0pt,
only marks,
mark=o,
mark options={solid}
]
table[row sep=crcr]{
9.00000000000002 0.230000000000002\\
9.00200000000002 0.209999999999999\\
9.00400000000002 0.230000000000002\\
9.00600000000002 0.260000000000003\\
9.00800000000002 0.219999999999999\\
9.01000000000002 0.209999999999999\\
9.01200000000002 0.190000000000002\\
9.01400000000002 0.219999999999999\\
9.01600000000002 0.179999999999998\\
9.01800000000002 0.209999999999999\\
9.02000000000002 0.2\\
9.02200000000002 0.219999999999999\\
9.02400000000002 0.249999999999999\\
9.02600000000002 0.240000000000001\\
9.02800000000003 0.270000000000002\\
9.03000000000003 0.300000000000003\\
9.03200000000003 0.270000000000002\\
9.03400000000003 0.350000000000002\\
9.03600000000003 0.37\\
9.03800000000003 0.360000000000001\\
9.04000000000003 0.41\\
9.04200000000003 0.390000000000002\\
9.04400000000003 0.41\\
9.04600000000003 0.379999999999999\\
9.04800000000003 0.350000000000002\\
9.05000000000003 0.440000000000001\\
9.05200000000003 0.480000000000001\\
9.05400000000003 0.520000000000001\\
9.05600000000003 0.61\\
9.05800000000004 0.620000000000004\\
9.06000000000004 0.810000000000004\\
9.06200000000004 0.920000000000001\\
9.06400000000004 0.910000000000002\\
9.06600000000004 1.02\\
9.06800000000004 1.01\\
9.07000000000004 1.05\\
9.07200000000004 1.02\\
9.07400000000004 1.09\\
9.07600000000004 1.13\\
9.07800000000004 1.33\\
9.08000000000004 1.56\\
9.08200000000004 1.76\\
9.08400000000004 1.92\\
9.08600000000004 2.21\\
9.08800000000005 2.42\\
9.09000000000005 2.52\\
9.09200000000005 2.87\\
9.09400000000005 3.4\\
9.09600000000005 3.95\\
9.09800000000005 4.02\\
9.10000000000005 4.25\\
9.10200000000005 4.39\\
9.10400000000005 4.18\\
9.10600000000005 3.44\\
9.10800000000005 2.25\\
9.11000000000005 0.93\\
9.11200000000005 -0.189999999999996\\
9.11400000000005 -1.42\\
9.11600000000005 -2.56\\
9.11800000000006 -3.38\\
9.12000000000006 -3.94\\
9.12200000000006 -4.19\\
9.12400000000006 -4.31\\
9.12600000000006 -4.24\\
9.12800000000006 -4.17\\
9.13000000000006 -4.05\\
9.13200000000006 -3.93\\
9.13400000000006 -3.8\\
9.13600000000006 -3.65\\
9.13800000000006 -3.52\\
9.14000000000006 -3.37\\
9.14200000000006 -3.26\\
9.14400000000006 -3.1\\
9.14600000000006 -3.01\\
9.14800000000007 -2.84\\
9.15000000000007 -2.8\\
9.15200000000007 -2.78\\
9.15400000000007 -2.68\\
9.15600000000007 -2.61\\
9.15800000000007 -2.53\\
9.16000000000007 -2.43\\
9.16200000000007 -2.34\\
9.16400000000007 -2.31\\
9.16600000000007 -2.2\\
9.16800000000007 -2.15\\
9.17000000000007 -2.11\\
9.17200000000007 -2.09\\
9.17400000000007 -2\\
9.17600000000007 -1.94\\
9.17800000000008 -1.93\\
9.18000000000008 -1.89\\
9.18200000000008 -1.82\\
9.18400000000008 -1.84\\
9.18600000000008 -1.74\\
9.18800000000008 -1.76\\
9.19000000000008 -1.66\\
9.19200000000008 -1.61\\
9.19400000000008 -1.59\\
9.19600000000008 -1.51\\
9.19800000000008 -1.52\\
9.20000000000008 -1.5\\
};
\addlegendentry{Experiment \, \,};
\addplot [
color=red,
solid,
line width=2.0pt,
]
table[row sep=crcr]{
9.00000000000001 -0.00965633775813968\\
9.00100000000001 -0.00286348938947101\\
9.00200000000001 0.00405333214052084\\
9.00300000000001 0.0110975300623399\\
9.00400000000001 0.0182726324886406\\
9.00500000000001 0.0255822981552231\\
9.00600000000001 0.0330303224791592\\
9.00700000000001 0.0406206439547232\\
9.00800000000001 0.0483573509089268\\
9.00900000000001 0.0562446886405361\\
9.01000000000001 0.064287066967686\\
9.01100000000001 0.0724890682115114\\
9.01200000000001 0.080855455645363\\
9.01300000000001 0.0893911824409785\\
9.01400000000001 0.0981014011458737\\
9.01500000000001 0.106991473728888\\
9.01600000000001 0.116066982233323\\
9.01700000000001 0.125333740080654\\
9.01800000000001 0.134797804071294\\
9.01900000000001 0.14446548713223\\
9.02000000000001 0.154343371865793\\
9.02100000000001 0.164438324958433\\
9.02200000000001 0.174757512512687\\
9.02300000000001 0.185308416371533\\
9.02400000000001 0.196098851509659\\
9.02500000000001 0.207136984573068\\
9.02600000000001 0.218431353654891\\
9.02700000000001 0.229990889403375\\
9.02800000000001 0.241824937566549\\
9.02900000000001 0.25394328308682\\
9.03000000000001 0.266356175869359\\
9.03100000000001 0.2790743583595\\
9.03200000000001 0.292109095075999\\
9.03300000000001 0.305472204261205\\
9.03400000000001 0.319176091824179\\
9.03500000000001 0.333233787768602\\
9.03600000000001 0.347658985316224\\
9.03700000000001 0.362466082956644\\
9.03800000000001 0.377670229675733\\
9.03900000000001 0.393287373640508\\
9.04000000000001 0.409334314644606\\
9.04100000000001 0.425828760649379\\
9.04200000000001 0.442789388788019\\
9.04300000000001 0.460235911237617\\
9.04400000000001 0.478189146405184\\
9.04500000000001 0.496671095918011\\
9.04600000000001 0.515705027959914\\
9.04700000000001 0.535315567550627\\
9.04800000000001 0.555528794426624\\
9.04900000000001 0.576372349250943\\
9.05000000000001 0.597875548955738\\
9.05100000000001 0.620069512104326\\
9.05200000000001 0.642987295253607\\
9.05300000000001 0.666664041400678\\
9.05400000000001 0.691137141709852\\
9.05500000000001 0.716446411842613\\
9.05600000000001 0.742634284349061\\
9.05700000000001 0.769746018730538\\
9.05800000000001 0.797829930944716\\
9.05900000000001 0.826937644302022\\
9.06000000000001 0.857124363891355\\
9.06100000000001 0.888449176872102\\
9.06200000000001 0.920975381179718\\
9.06300000000001 0.954770845405797\\
9.06400000000001 0.989908402823001\\
9.06500000000001 1.02646628272486\\
9.06600000000001 1.06452858242093\\
9.06700000000001 1.1041857833458\\
9.06800000000001 1.14553531477991\\
9.06900000000001 1.18868216859018\\
9.07000000000001 1.23373956811389\\
9.07100000000001 1.28082969374573\\
9.07200000000001 1.33008446680785\\
9.07300000000001 1.38164639171372\\
9.07400000000001 1.43566945401516\\
9.07500000000001 1.49232006829961\\
9.07600000000001 1.55177806456974\\
9.07700000000001 1.61423769397575\\
9.07800000000001 1.67990862358316\\
9.07900000000001 1.74901687380874\\
9.08000000000001 1.82180562921543\\
9.08100000000001 1.89853582064415\\
9.08200000000001 1.97948633004833\\
9.08300000000001 2.06495360299581\\
9.08400000000001 2.15525035920197\\
9.08500000000001 2.25070295660051\\
9.08600000000001 2.35164677213546\\
9.08700000000001 2.45841868821164\\
9.08800000000001 2.57134538300099\\
9.08900000000001 2.69072556746844\\
9.09000000000001 2.81680352650598\\
9.09100000000001 2.94973022017093\\
9.09200000000001 3.08950667911802\\
9.09300000000001 3.23590237642213\\
9.09400000000001 3.38833860551549\\
9.09500000000001 3.54572371827777\\
9.09600000000001 3.70622384528948\\
9.09700000000001 3.86695077402733\\
9.09800000000001 4.02355109465503\\
9.09900000000001 4.16969372017642\\
9.10000000000001 4.29648709107705\\
9.10100000000001 4.39192732857564\\
9.10200000000001 4.44059507802094\\
9.10300000000001 4.42396553691233\\
9.10400000000001 4.32178916894802\\
9.10500000000001 4.11486484229777\\
9.10600000000001 3.78898913773196\\
9.10700000000001 3.33902726875425\\
9.10800000000001 2.77149922982219\\
9.10900000000001 2.10449434177954\\
9.11000000000001 1.36502382949771\\
9.11100000000001 0.585072144215706\\
9.11200000000001 -0.202245137742081\\
9.11300000000001 -0.965686230052573\\
9.11400000000001 -1.67804649674745\\
9.11500000000001 -2.31775042184074\\
9.11600000000001 -2.86999429786538\\
9.11700000000001 -3.3272992132184\\
9.11800000000001 -3.68924614891339\\
9.11900000000001 -3.96134078391074\\
9.12000000000001 -4.15325453246287\\
9.12100000000001 -4.27688436715068\\
9.12200000000001 -4.34465245536184\\
9.12300000000001 -4.36828670315523\\
9.12400000000001 -4.35812030154881\\
9.12500000000001 -4.32281575640989\\
9.12600000000001 -4.26937320351122\\
9.12700000000001 -4.2032952075654\\
9.12800000000001 -4.12881573863316\\
9.12900000000001 -4.04913702766214\\
9.13000000000001 -3.96664531998945\\
9.13100000000001 -3.88309404995305\\
9.13200000000001 -3.79975275889764\\
9.13300000000001 -3.71752480321722\\
9.13400000000001 -3.63703866265922\\
9.13500000000001 -3.55871787593577\\
9.13600000000001 -3.48283414459476\\
9.13700000000001 -3.40954741664993\\
9.13800000000001 -3.33893601589411\\
9.13900000000001 -3.2710192188576\\
9.14000000000001 -3.20577412968968\\
9.14100000000001 -3.14314826285466\\
9.14200000000001 -3.0830689005782\\
9.14300000000001 -3.02545002905685\\
9.14400000000001 -2.97019745789794\\
9.14500000000001 -2.91721257675847\\
9.14600000000001 -2.86639509005589\\
9.14700000000001 -2.81764498580411\\
9.14800000000001 -2.77086393105557\\
9.14900000000001 -2.72595623876646\\
9.15000000000001 -2.68282951514073\\
9.15100000000001 -2.64139506964522\\
9.15200000000001 -2.60156814967266\\
9.15300000000001 -2.56326804660013\\
9.15400000000001 -2.52641810849244\\
9.15500000000001 -2.49094568601129\\
9.15600000000001 -2.45678203151081\\
9.15700000000001 -2.42386216631317\\
9.15800000000001 -2.39212472737735\\
9.15900000000001 -2.36151180170049\\
9.16000000000001 -2.33196875461166\\
9.16100000000001 -2.3034440564642\\
9.16200000000001 -2.27588911097711\\
9.16300000000001 -2.24925808752581\\
9.16400000000001 -2.22350775896688\\
9.16500000000001 -2.1985973460406\\
9.16600000000001 -2.17448836899105\\
9.16700000000001 -2.15114450674493\\
9.16800000000001 -2.1285314637673\\
9.16900000000001 -2.10661684455348\\
9.17000000000001 -2.08537003560003\\
9.17100000000001 -2.0647620946197\\
9.17200000000001 -2.04476564671053\\
9.17300000000001 -2.02535478715586\\
9.17400000000001 -2.00650499051433\\
9.17500000000001 -1.98819302565013\\
9.17600000000001 -1.97039687635447\\
9.17700000000001 -1.95309566721587\\
9.17800000000001 -1.93626959440657\\
9.17900000000001 -1.91989986106505\\
9.18000000000001 -1.90396861697049\\
9.18100000000001 -1.88845890221958\\
9.18200000000001 -1.87335459463277\\
9.18300000000001 -1.85864036063371\\
9.18400000000001 -1.8443016093608\\
9.18500000000001 -1.83032444978651\\
9.18600000000001 -1.81669565063363\\
9.18700000000001 -1.80340260289369\\
9.18800000000001 -1.79043328476497\\
9.18900000000001 -1.77777622884082\\
9.19000000000001 -1.76542049139145\\
9.19100000000001 -1.75335562359301\\
9.19200000000001 -1.74157164456853\\
9.19300000000001 -1.73005901611546\\
9.19400000000001 -1.71880861900331\\
9.19500000000001 -1.70781173073314\\
9.19600000000001 -1.69706000465945\\
9.19700000000001 -1.6865454503814\\
9.19800000000001 -1.67626041531724\\
9.19900000000001 -1.6661975673827\\
9.20000000000001 -1.65634987869905\\
};
\addlegendentry{Fit};
\end{axis}
\node[draw=none] at (1.32,0.28) {\footnotesize $f_{\text{s}} = 0.6$};
\node[draw=none] at (1.32,0.64) {\footnotesize $N=66$};
\node[draw=none] at (1.32,1.00) {\footnotesize $Q_{\text{m}}=515$};
\node[draw=none] at (1.32,1.36) {\footnotesize $r = 1.1$};
\node[draw=none] at (1.32,1.70) {\footnotesize $\phi = 0$};
\node at (-0.9,-0.9) {\large{\textbf{a}}};
\end{tikzpicture}%
}
\subfigure{
\begin{tikzpicture}
\begin{axis}[scale = 0.34*\textwidth/(4.5in),
width=4.52083333333333in,
height=3.565625in,
xmin=110,
xmax=1100,
xlabel={Quality factor $Q_{\text{m}}$ (-)},
ymin=1100,
ymax=11000,
ylabel={$\tilde{\mathcal{G}}$ (W$^{-1}$ m$^{-1}$)},
ylabel style = {yshift=-0.2ex},
axis x line*=bottom,
axis y line*=left,
legend style={fill=none,draw=none,legend cell align=left,at={(0.55,1.05)},anchor=north, legend columns=-1},
xtick = {250,500,750,1000},
ytick = {2500,5000,7500,10000},
minor xtick={125,250,...,1000},
minor ytick={1250,2500,...,10000}
]
\addplot [
color=black,
solid,
line width=1.0pt,
forget plot
]
table[row sep=crcr]{
0 1.5\\
45 1.5\\
};
\addplot [
color=blue,
line width=0.5pt,
mark size=2.0pt,
only marks,
mark=o,
mark options={solid},
]
table[row sep=crcr]{
306 3218 \\
464 6561 \\
494 6199 \\
776 8432 \\
1010 10360 \\
614 5915 \\
731 6018 \\
603 6485 \\
674 7316 \\
715 7999 \\
678 8558 \\
740 8095 \\
897 8661 \\
738 7738 \\
427 6956 \\
490 5823 \\
681 7312 \\
};
\addlegendentry{Gain \, \,};
\addplot [
color=green!50!black,
line width=0.5pt,
mark size=2.0pt,
only marks,
mark=triangle,
mark options={solid},
]
table[row sep=crcr]{
249 3055 \\
400 3846 \\
461 4255 \\
966 8566 \\
472 3322 \\
515 4489 \\
482 4273 \\
1040 9863 \\
649 7611 \\
561 5443 \\
};
\addlegendentry{XPM \, \,};
\addplot [
color=red!50!black,
line width=0.5pt,
dashed
]
table[row sep=crcr]{
250 2583 \\
1010 10440.0 \\
};
\end{axis}
\node[draw=none] at (3.55,0.35) {\footnotesize $w=450 \,\text{nm}$};
\node[draw=none] at (3.55,0.75) {\footnotesize $\frac{\tilde{\mathcal{G}}}{Q_{\text{m}}} = 10.3 \, \text{W}^{-1}\text{m}^{-1}$};
\node at (-0.7,-0.7) {\large{\textbf{b}}};
\end{tikzpicture}%
}
\caption{\textbf{The efficiency $\tilde{\mathcal{G}}$ reaches up to $10^{4} \, \text{W}^{-1} \text{m}^{-1}$.} \textbf{a}, Example of a Fano resonance obtained from the XPM experiment, used to determine the quality factor $Q_{\text{m}}$ and gain coefficient $\tilde{\mathcal{G}}$ given $\overline{\gamma}_{\text{tot}} = 610 \, \text{W}^{-1} \text{m}^{-1}$ independently from the gain resonances. \textbf{b}, Plot of $(\tilde{\mathcal{G}},Q_{\text{m}})$-pairs for a large set of waveguides obtained from both the gain (fig.\ref{fig:netgain}a) and the XPM experiment (fig.\ref{fig:photonphononoverlap}a). A linear fit without offset yields $\tilde{\mathcal{G}}/Q_{\text{m}} = 10.3 \, \text{W}^{-1} \text{m}^{-1}$. Most variation results from uncertainty in the coupling efficiency ($\approx 25\%$). The two points at $Q_{\text{m}} < 375$ concern a silicon wire on a pillar \cite{VanLaer2015}.}
\label{fig:photonphononoverlap}
\end{figure*}
\subsubsection{Photon-phonon overlap.}
\label{subsubsec:xpm}
Finally, we measure the XPM resonances (fig.\ref{fig:photonphononoverlap}a) for a subset of waveguides to obtain an independent estimate of the photon-phonon interaction efficiency. Combined with the gain data, we obtain the $(\tilde{\mathcal{G}},Q_{\text{m}})$-pairs for a large set of waveguides (fig.\ref{fig:photonphononoverlap}b) with fixed waveguide width $w = 450 \, \text{nm}$. A fit (fig.\ref{fig:photonphononoverlap}b) to this dataset yields a non-resonant nonlinearity of $\frac{\tilde{\mathcal{G}}}{Q_{\text{m}}} = 10.3 \, \text{W}^{-1} \text{m}^{-1}$ -- in good agreement with earlier experiments \cite{VanLaer2015} and predictions \cite{Rakich2012,Qiu2013}. The efficiencies reach up to $\tilde{\mathcal{G}} = 10360 \, \text{W}^{-1}\text{m}^{-1}$, the highest value obtained thus far in the gigahertz range.
\section{Conclusion}
Through a novel opto-acoustic nanodevice, a series of suspended silicon wires, we demonstrate modest ($0.5 \, \text{dB}$) net Brillouin gain with high efficiencies (up to $10^{4} \, \text{W}^{-1}\text{m}^{-1}$). This device is a step towards integrated selective amplifiers. We find that fabrication disorder, likely in the waveguide width, broadens and splits the phonon resonances in some cases. In particular, the phonon quality factor strongly decreases as the number of suspended silicon beams increases. Such disorder is expected to hinder development of nanoscale phonon-based technologies quite generally -- new techniques or better fabrication tools must be developed to address this issue.
\section*{Appendices}
\subsection*{Influence of two-photon absorption and free-carriers on XPM}
In this section, we describe the influence of two-photon absorption (TPA) and free-carriers (FCs) on the Fano resonances. First, we recall that (see $\eqref{eq:Fano_homo}$), in absence of TPA and FCs, the sideband power is proportional to $\mathcal{F}$ with
\begin{equation}
\label{eq:Fano_homo2}
\mathcal{F}(\Delta) = \left| 1 + r \frac{1}{-2\Delta + i} \right|^{2}
\end{equation}
with $r = f_{\text{s}} \frac{ \tilde{\mathcal{G}}}{4 \overline{\gamma}_{\text{K}}}$. One can show that this Fano function $\mathcal{F}$ has one maximum $\mathcal{F}_{\text{max}}$ and one minimum $\mathcal{F}_{\text{min}}$ given by
\begin{equation}
\label{eq:Fanosymm}
\mathcal{F}_{\text{max}} = \frac{r^{2} + 4 + r \sqrt{ r^{2} + 4}}{r^2 + 4 - r \sqrt{r^{2} + 4}} = \frac{1}{\mathcal{F}_{\text{min}}}
\end{equation}
that are fully determined by $r$. This implies $\mathcal{F}_{\text{max}} \, [\text{dB}] = - \mathcal{F}_{\text{min}} \, [\text{dB}]$, as evident in fig.\ref{fig:photonphononoverlap}a. Inverting this for $r$ yields
\begin{equation}
\label{eq:Fanoratio}
r = \frac{\mathcal{F}_{\text{max}} - 1}{\sqrt{\mathcal{F}_{\text{max}}}}
\end{equation}
Applied to fig.\ref{fig:photonphononoverlap}a, we have $\mathcal{F}_{\text{max}} = 4.4 \,\text{dB} = 2.75$ and thus $r=1.1$ through \eqref{eq:Fanoratio} -- in agreement with the $r$ obtained from a least-square fit (fig.\ref{fig:photonphononoverlap}a). The extrema are reached at a detuning of $\Delta_{\text{max/min}} = \frac{1}{4} \left( r \mp \sqrt{r^{2} + 4} \right)$. In the large $r$ limit, we get $\Delta_{\text{max}} \rightarrow -\frac{1}{2r}$ and $\Delta_{\text{min}} \rightarrow \frac{r}{2}$. In the small $r$ limit, we have $\Delta_{\text{max/min}} \rightarrow \frac{r \mp 2}{4}$. Therefore, the maximum XPM is always reached at a negative detuning between $-\frac{1}{2}$ and $0$, typically ($r > 1$) close to the phonon resonance ($\Delta = 0$).
\begin{figure*}
\centering
\subfigure{
\begin{tikzpicture}
\begin{axis}[scale = 0.34*\textwidth/(4.5in),
width=4.52083333333333in,
height=3.565625in,
xmin=9.23,
xmax=9.49,
xlabel={Frequency $\frac{\Omega}{2\pi}$ (GHz)},
xlabel style = {yshift=0ex},
ymin=-6,
ymax=2,
ylabel= {Sideband power (dB)},
ylabel style = {yshift=-1.5ex},
axis x line*=bottom,
axis y line*=left,
legend style={fill=none,draw=none,legend cell align=left,at={(0.5,1.2)},anchor=north, legend columns=-1},
xtick = {9.25,9.35,9.45},
ytick = {-6,-4,...,2},
minor xtick={9,9.05,...,9.25},
minor ytick={-5,-4,...,5}
]
\addplot [
color=blue,
line width=0.5pt,
mark size=2.0pt,
only marks,
mark=o,
mark options={solid}
]
table[row sep=crcr]{
9.25000000000001 0.00999999999999881\\
9.25200000000001 -0.0500000000000033\\
9.25400000000001 0.00999999999999881\\
9.25600000000001 0\\
9.25800000000001 0.0899999999999943\\
9.26000000000001 0.0199999999999935\\
9.26200000000001 0.0700000000000015\\
9.26400000000001 0.0299999999999924\\
9.26600000000001 0.0799999999999959\\
9.26800000000001 0.0700000000000015\\
9.27000000000002 0.0999999999999933\\
9.27200000000002 0.0399999999999995\\
9.27400000000002 0.00999999999999881\\
9.27600000000002 0.0399999999999995\\
9.27800000000002 0.0299999999999924\\
9.28000000000002 0.0199999999999935\\
9.28200000000002 0.00999999999999881\\
9.28400000000002 0.119999999999995\\
9.28600000000002 0\\
9.28800000000002 0.0799999999999959\\
9.29000000000002 0.0499999999999946\\
9.29200000000002 0.0700000000000015\\
9.29400000000002 0.00999999999999881\\
9.29600000000002 0.0700000000000015\\
9.29800000000002 0.0799999999999959\\
9.30000000000003 0.0399999999999995\\
9.30200000000003 0.0899999999999943\\
9.30400000000003 0.0799999999999959\\
9.30600000000003 0.119999999999995\\
9.30800000000003 0.149999999999997\\
9.31000000000003 0.129999999999994\\
9.31200000000003 0.0799999999999959\\
9.31400000000003 0.0299999999999924\\
9.31600000000003 0.0799999999999959\\
9.31800000000003 0.0599999999999927\\
9.32000000000003 0.139999999999998\\
9.32200000000003 0.0499999999999946\\
9.32400000000003 0.209999999999995\\
9.32600000000003 0.159999999999996\\
9.32800000000003 0.109999999999996\\
9.33000000000004 0.0599999999999927\\
9.33200000000004 0.0799999999999959\\
9.33400000000004 0.149999999999997\\
9.33600000000004 0.239999999999995\\
9.33800000000004 0.229999999999997\\
9.34000000000004 0.239999999999995\\
9.34200000000004 0.229999999999997\\
9.34400000000004 0.36\\
9.34600000000004 0.299999999999998\\
9.34800000000004 0.419999999999993\\
9.35000000000004 0.459999999999994\\
9.35200000000004 0.569999999999999\\
9.35400000000004 0.789999999999997\\
9.35600000000004 0.779999999999994\\
9.35800000000004 0.93\\
9.36000000000005 0.919999999999992\\
9.36200000000005 1.05\\
9.36400000000005 1.11\\
9.36600000000005 1.29\\
9.36800000000005 1.19\\
9.37000000000005 0.379999999999993\\
9.37200000000005 -1.35\\
9.37400000000005 -3.62000000000001\\
9.37600000000005 -5.04000000000001\\
9.37800000000005 -4.82\\
9.38000000000005 -4.29000000000001\\
9.38200000000005 -2.87\\
9.38400000000005 -2.33000000000001\\
9.38600000000005 -1.87000000000001\\
9.38800000000006 -1.54000000000001\\
9.39000000000006 -1.49\\
9.39200000000006 -1.29000000000001\\
9.39400000000006 -1.12000000000001\\
9.39600000000006 -1.13\\
9.39800000000006 -0.970000000000008\\
9.40000000000006 -0.990000000000003\\
9.40200000000006 -0.89\\
9.40400000000006 -0.900000000000008\\
9.40600000000006 -0.770000000000004\\
9.40800000000006 -0.780000000000003\\
9.41000000000006 -0.770000000000004\\
9.41200000000006 -0.680000000000001\\
9.41400000000006 -0.760000000000005\\
9.41600000000006 -0.660000000000003\\
9.41800000000007 -0.690000000000004\\
9.42000000000007 -0.569999999999999\\
9.42200000000007 -0.670000000000002\\
9.42400000000007 -0.600000000000001\\
9.42600000000007 -0.700000000000003\\
9.42800000000007 -0.64\\
9.43000000000007 -0.630000000000001\\
9.43200000000007 -0.550000000000007\\
9.43400000000007 -0.560000000000005\\
9.43600000000007 -0.530000000000004\\
9.43800000000007 -0.620000000000007\\
9.44000000000007 -0.510000000000006\\
9.44200000000007 -0.609999999999999\\
9.44400000000007 -0.530000000000004\\
9.44600000000007 -0.560000000000005\\
9.44800000000008 -0.480000000000005\\
9.45000000000008 -0.520000000000005\\
9.45200000000008 -0.590000000000006\\
9.45400000000008 -0.380000000000002\\
9.45600000000008 -0.460000000000002\\
9.45800000000008 -0.390000000000002\\
9.46000000000008 -0.400000000000005\\
9.46200000000008 -0.430000000000002\\
9.46400000000008 -0.470000000000006\\
9.46600000000008 -0.380000000000002\\
9.46800000000008 -0.340000000000003\\
9.47000000000008 -0.360000000000001\\
};
\addlegendentry{Experiment \, \,};
\addplot [
smooth,
color=red,
solid,
line width=2.0pt,
]
table[row sep=crcr]{
9.25000000000001 -0.0190477942171496\\
9.25100000000001 -0.0177929477578657\\
9.25200000000001 -0.0165181239296369\\
9.25300000000001 -0.0152228428165208\\
9.25400000000001 -0.0139066090406187\\
9.25500000000001 -0.0125689111358639\\
9.25600000000001 -0.0112092208911619\\
9.25700000000001 -0.0098269926612041\\
9.25800000000001 -0.00842166264301975\\
9.25900000000001 -0.0069926481163429\\
9.26000000000001 -0.00553934664561301\\
9.26100000000001 -0.00406113524134055\\
9.26200000000001 -0.00255736947847429\\
9.26300000000001 -0.00102738256907636\\
9.26400000000001 0.000529515613367643\\
9.26500000000001 0.00211403956100778\\
9.26600000000001 0.00372692921326065\\
9.26700000000001 0.0053689510963478\\
9.26800000000001 0.00704089952440742\\
9.26900000000001 0.00874359786612655\\
9.27000000000001 0.0104778998809764\\
9.27100000000001 0.0122446911296664\\
9.27200000000001 0.0140448904635227\\
9.27300000000001 0.0158794515980979\\
9.27400000000001 0.0177493647764794\\
9.27500000000001 0.0196556585283722\\
9.27600000000001 0.0215994015313924\\
9.27700000000001 0.0235817045815413\\
9.27800000000001 0.025603722680328\\
9.27900000000001 0.0276666572467441\\
9.28000000000001 0.029771758462704\\
9.28100000000001 0.031920327761492\\
9.28200000000001 0.0341137204694213\\
9.28300000000001 0.0363533486117078\\
9.28400000000001 0.0386406838945786\\
9.28500000000001 0.04097726087659\\
9.28600000000001 0.0433646803431565\\
9.28700000000001 0.0458046128996293\\
9.28800000000001 0.0482988027994228\\
9.28900000000001 0.0508490720253126\\
9.29000000000001 0.0534573246433805\\
9.29100000000001 0.0561255514510303\\
9.29200000000001 0.0588558349423257\\
9.29300000000001 0.061650354615957\\
9.29400000000001 0.0645113926535492\\
9.29500000000001 0.0674413399985701\\
9.29600000000001 0.070442702868862\\
9.29700000000001 0.0735181097389817\\
9.29800000000001 0.0766703188320619\\
9.29900000000001 0.0799022261645413\\
9.30000000000001 0.0832168741915406\\
9.30100000000001 0.0866174611052858\\
9.30200000000001 0.0901073508441519\\
9.30300000000001 0.0936900838757737\\
9.30400000000001 0.0973693888240666\\
9.30500000000001 0.101149195017284\\
9.30600000000001 0.105033646042095\\
9.30700000000001 0.10902711439771\\
9.30800000000001 0.113134217354159\\
9.30900000000001 0.117359834129717\\
9.31000000000001 0.121709124515223\\
9.31100000000001 0.12618754908706\\
9.31200000000001 0.130800891165933\\
9.31300000000001 0.135555280696581\\
9.31400000000001 0.140457220243201\\
9.31500000000001 0.145513613317493\\
9.31600000000001 0.150731795281371\\
9.31700000000001 0.15611956709465\\
9.31800000000001 0.161685232209459\\
9.31900000000001 0.167437636949079\\
9.32000000000001 0.173386214749112\\
9.32100000000001 0.179541034684282\\
9.32200000000001 0.185912854755212\\
9.32300000000001 0.192513180467148\\
9.32400000000001 0.199354329297568\\
9.32500000000001 0.206449501722029\\
9.32600000000001 0.21381285954962\\
9.32700000000001 0.221459612410674\\
9.32800000000001 0.229406113340919\\
9.32900000000001 0.237669964519342\\
9.33000000000001 0.24627013434141\\
9.33100000000001 0.255227087144357\\
9.33200000000001 0.264562927047509\\
9.33300000000001 0.274301557524784\\
9.33400000000001 0.284468858484127\\
9.33500000000001 0.295092882784758\\
9.33600000000001 0.306204074263842\\
9.33700000000001 0.317835509454242\\
9.33800000000001 0.33002316522435\\
9.33900000000001 0.342806214520038\\
9.34000000000001 0.356227352172166\\
9.34100000000001 0.370333152254236\\
9.34200000000001 0.385174457590783\\
9.34300000000001 0.400806800507551\\
9.34400000000001 0.417290851449787\\
9.34500000000001 0.434692888177623\\
9.34600000000001 0.453085272116678\\
9.34700000000001 0.472546908935881\\
9.34800000000001 0.49316365575851\\
9.34900000000001 0.51502861481268\\
9.35000000000001 0.538242218459803\\
9.35100000000001 0.562911956567849\\
9.35200000000001 0.589151513215197\\
9.35300000000001 0.617078948160657\\
9.35400000000001 0.646813350764618\\
9.35500000000001 0.678469062874168\\
9.35600000000001 0.712146033682708\\
9.35700000000001 0.747914000185215\\
9.35800000000001 0.785786752561768\\
9.35900000000001 0.825680347001831\\
9.36000000000001 0.867345071078559\\
9.36100000000001 0.910254013376588\\
9.36200000000001 0.953419047826769\\
9.36300000000001 0.995084058263549\\
9.36400000000001 1.03220872188287\\
9.36500000000001 1.05959380282022\\
9.36600000000001 1.06839792753768\\
9.36700000000001 1.0436546034233\\
9.36800000000001 0.960285646839786\\
9.36900000000001 0.777360423095474\\
9.37000000000001 0.432065628447391\\
9.37100000000001 -0.159571519528843\\
9.37200000000001 -1.07707466919832\\
9.37300000000001 -2.32034126966899\\
9.37400000000001 -3.69478919515575\\
9.37500000000001 -4.78678389765659\\
9.37600000000001 -5.24158062809951\\
9.37700000000001 -5.10556242008983\\
9.37800000000001 -4.67347785067247\\
9.37900000000001 -4.17399498011159\\
9.38000000000001 -3.7083742923693\\
9.38100000000001 -3.30562999062018\\
9.38200000000001 -2.96624351545181\\
9.38300000000001 -2.68198972098679\\
9.38400000000001 -2.44330093405181\\
9.38500000000001 -2.2416045017747\\
9.38600000000001 -2.0698371916132\\
9.38700000000001 -1.92236116695202\\
9.38800000000001 -1.79472412238057\\
9.38900000000001 -1.68341326882704\\
9.39000000000001 -1.58564637138851\\
9.39100000000001 -1.4992060475242\\
9.39200000000001 -1.42231206860138\\
9.39300000000001 -1.35352402759114\\
9.39400000000001 -1.29166732401177\\
9.39500000000001 -1.23577675086605\\
9.39600000000001 -1.18505329012049\\
9.39700000000001 -1.13883081915315\\
9.39800000000001 -1.0965502777807\\
9.39900000000001 -1.05773947987339\\
9.40000000000001 -1.02199722177096\\
9.40100000000001 -0.988980683363452\\
9.40200000000001 -0.958395369805343\\
9.40300000000001 -0.929987027241254\\
9.40400000000001 -0.903535102856089\\
9.40500000000001 -0.878847421236911\\
9.40600000000001 -0.855755824976529\\
9.40700000000001 -0.834112584523386\\
9.40800000000001 -0.813787425455438\\
9.40900000000001 -0.794665054222898\\
9.41000000000001 -0.776643088583023\\
9.41100000000001 -0.759630318361768\\
9.41200000000001 -0.743545237233877\\
9.41300000000001 -0.728314797959437\\
9.41400000000001 -0.713873352733426\\
9.41500000000001 -0.700161747579876\\
9.41600000000001 -0.687126545492349\\
9.41700000000001 -0.674719357625426\\
9.41800000000001 -0.662896265529662\\
9.41900000000001 -0.651617320393786\\
9.42000000000001 -0.640846107660453\\
9.42100000000001 -0.630549367335068\\
9.42200000000001 -0.620696661901843\\
9.42300000000001 -0.611260085067217\\
9.42400000000001 -0.602214005625905\\
9.42500000000001 -0.593534841633441\\
9.42600000000001 -0.585200860804882\\
9.42700000000001 -0.577192003672158\\
9.42800000000001 -0.569489726544538\\
9.42900000000001 -0.562076861744944\\
9.43000000000001 -0.554937492955083\\
9.43100000000001 -0.548056843806577\\
9.43200000000001 -0.541421178111605\\
9.43300000000001 -0.535017710344884\\
9.43400000000001 -0.52883452517373\\
9.43500000000001 -0.522860504991471\\
9.43600000000001 -0.517085264544243\\
9.43700000000001 -0.511499091857152\\
9.43800000000001 -0.506092894765631\\
9.43900000000001 -0.500858152443342\\
9.44000000000001 -0.495786871392069\\
9.44100000000001 -0.490871545423494\\
9.44200000000001 -0.486105119218031\\
9.44300000000001 -0.481480955094335\\
9.44400000000001 -0.476992802665429\\
9.44500000000001 -0.472634771093936\\
9.44600000000001 -0.468401303691052\\
9.44700000000001 -0.464287154632311\\
9.44800000000001 -0.460287367587654\\
9.44900000000001 -0.456397256085264\\
9.45000000000001 -0.452612385447595\\
9.45100000000001 -0.448928556155213\\
9.45200000000001 -0.445341788508756\\
9.45300000000001 -0.441848308472643\\
9.45400000000001 -0.438444534596001\\
9.45500000000001 -0.435127065916577\\
9.45600000000001 -0.431892670762681\\
9.45700000000001 -0.42873827637657\\
9.45800000000001 -0.425660959289994\\
9.45900000000001 -0.422657936389114\\
9.46000000000001 -0.419726556612099\\
9.46100000000001 -0.416864293227815\\
9.46200000000001 -0.41406873664877\\
9.46300000000001 -0.41133758773584\\
9.46400000000001 -0.408668651556004\\
9.46500000000001 -0.406059831557797\\
9.46600000000001 -0.403509124132298\\
9.46700000000001 -0.401014613530277\\
9.46800000000001 -0.398574467108638\\
9.46900000000001 -0.396186930881544\\
9.47000000000001 -0.39385032535378\\
};
\addlegendentry{Fit};
\end{axis}
\node[draw=none] at (1.32,0.34) {\footnotesize $f_{\text{s}} = 0.17$};
\node[draw=none] at (1.32,0.7) {\footnotesize $N=19$};
\node[draw=none] at (1.32,1.06) {\footnotesize $Q_{\text{m}}=966$};
\node[draw=none] at (1.32,1.42) {\footnotesize $r = 0.6$};
\node[draw=none] at (1.32,1.76) {\footnotesize $\phi = -0.2 \pi$};
\node at (-0.9,-0.9) {\large{\textbf{a}}};
\end{tikzpicture}%
}
\subfigure{
\begin{tikzpicture}
\begin{axis}[scale = 0.4*\textwidth/(4.5in),
width=4.52083333333333in,
height=3.565625in,
view={45}{30},
scale only axis,
xmin=-10,
xmax=10.5,
xlabel={Detuning $\Delta$ (-)},
xlabel style = {yshift=1.3ex,rotate=-22},
xmajorgrids,
ymin=-0.5,
ymax=0.4,
ylabel={Phase $\frac{\phi}{\pi}$ (-)},
ylabel style = {yshift=1.3ex,rotate=22},
ymajorgrids,
zmin=-4.5,
zmax=4,
zlabel={Relative power (dB)},
zmajorgrids,
xtick = {-10,-5,0,5,10},
ytick = {-0.4,-0.2,0,0.2,0.4},
ztick = {-4,-2,...,4},
minor xtick={-10,-7.5,...,10},
minor ytick={-0.4,-0.3,...,0.4},
minor ztick = {-5,-4,...,5}
]
\addplot3 [
color=blue,
solid,
fill opacity=0.15,
draw=blue!80!white,thin,
fill=blue]
table[row sep=crcr] {
-10 -0.25 0.145739839195374\\
-9.8 -0.25 0.148549142884006\\
-9.6 -0.25 0.15146839353116\\
-9.4 -0.25 0.154504140716882\\
-9.2 -0.25 0.157663461580377\\
-9 -0.25 0.160954014639648\\
-8.8 -0.25 0.164384100278172\\
-8.6 -0.25 0.167962728871753\\
-8.4 -0.25 0.171699697692247\\
-8.2 -0.25 0.175605677919306\\
-8 -0.25 0.179692313323297\\
-7.8 -0.25 0.183972332460015\\
-7.6 -0.25 0.188459676550408\\
-7.4 -0.25 0.193169645618771\\
-7.2 -0.25 0.198119065944836\\
-7 -0.25 0.203326482468038\\
-6.8 -0.25 0.208812380487849\\
-6.6 -0.25 0.214599441860245\\
-6.4 -0.25 0.220712841931162\\
-6.2 -0.25 0.227180594713669\\
-6 -0.25 0.234033955356221\\
-5.8 -0.25 0.241307890822109\\
-5.6 -0.25 0.249041631971454\\
-5.4 -0.25 0.257279322977729\\
-5.2 -0.25 0.266070787289666\\
-5 -0.25 0.275472433216265\\
-4.8 -0.25 0.285548326663998\\
-4.6 -0.25 0.296371463467031\\
-4.4 -0.25 0.308025278749684\\
-4.2 -0.25 0.320605434972358\\
-4 -0.25 0.334221931889646\\
-3.8 -0.25 0.349001576846804\\
-3.6 -0.25 0.365090835222134\\
-3.4 -0.25 0.382659033709824\\
-3.2 -0.25 0.401901784465276\\
-3 -0.25 0.423044277762052\\
-2.8 -0.25 0.446343636583303\\
-2.6 -0.25 0.472088593331328\\
-2.4 -0.25 0.500592814722983\\
-2.2 -0.25 0.532174120933994\\
-2 -0.25 0.567103017352471\\
-1.8 -0.25 0.6054842295362\\
-1.6 -0.25 0.646989094158309\\
-1.4 -0.25 0.690245247911574\\
-1.2 -0.25 0.731405492136206\\
-1 -0.25 0.760653805646482\\
-0.799999999999999 -0.25 0.753279412108143\\
-0.6 -0.25 0.646146771034244\\
-0.399999999999999 -0.25 0.278413217883554\\
-0.199999999999999 -0.25 -0.706974018154745\\
0 -0.25 -2.65285583008699\\
0.199999999999999 -0.25 -4.41156476673046\\
0.399999999999999 -0.25 -4.24412446517168\\
0.6 -0.25 -3.32636976970489\\
0.799999999999999 -0.25 -2.56648289025602\\
1 -0.25 -2.03608906054768\\
1.2 -0.25 -1.66638136952892\\
1.4 -0.25 -1.40069154448113\\
1.6 -0.25 -1.20318993075773\\
1.8 -0.25 -1.05181516874745\\
2 -0.25 -0.932700162487936\\
2.2 -0.25 -0.8368508874845\\
2.4 -0.25 -0.758243734442576\\
2.6 -0.25 -0.692722212543145\\
2.8 -0.25 -0.637340018918836\\
3 -0.25 -0.589958062964941\\
3.2 -0.25 -0.548989929422142\\
3.4 -0.25 -0.513236630860396\\
3.6 -0.25 -0.481776589942528\\
3.8 -0.25 -0.453890691057152\\
4 -0.25 -0.429010146108851\\
4.2 -0.25 -0.40667953578277\\
4.4 -0.25 -0.386530152998506\\
4.6 -0.25 -0.368260471955391\\
4.8 -0.25 -0.351621630686726\\
5 -0.25 -0.336406496932896\\
5.2 -0.25 -0.322441332425758\\
5.4 -0.25 -0.309579366681349\\
5.6 -0.25 -0.297695791449509\\
5.8 -0.25 -0.286683824263115\\
6 -0.25 -0.276451585108043\\
6.2 -0.25 -0.266919597663013\\
6.4 -0.25 -0.258018774720856\\
6.6 -0.25 -0.249688782204808\\
6.8 -0.25 -0.241876701616199\\
7 -0.25 -0.234535929510553\\
7.2 -0.25 -0.227625266576992\\
7.4 -0.25 -0.221108159403908\\
7.6 -0.25 -0.214952065981149\\
7.8 -0.25 -0.209127922077919\\
8 -0.25 -0.203609690324866\\
8.2 -0.25 -0.198373977465624\\
8.4 -0.25 -0.193399708083261\\
8.6 -0.25 -0.18866784533922\\
8.8 -0.25 -0.184161151027455\\
9 -0.25 -0.179863978650488\\
9.2 -0.25 -0.175762094347011\\
9.4 -0.25 -0.171842521403593\\
9.6 -0.25 -0.168093404812703\\
9.8 -0.25 -0.164503892931813\\
10 -0.25 -0.161064033781724\\
};
\addplot3 [
color=blue,
solid,
fill opacity=0.15,
draw=blue!80!white,thin,
fill=blue]
table[row sep=crcr] {
-10 -0.125 0.194270004263653\\
-9.8 -0.125 0.198097098833885\\
-9.6 -0.125 0.202077586311008\\
-9.4 -0.125 0.20622084889909\\
-9.2 -0.125 0.210537047240605\\
-9 -0.125 0.21503720254908\\
-8.8 -0.125 0.219733289319498\\
-8.6 -0.125 0.22463834023279\\
-8.4 -0.125 0.229766565157543\\
-8.2 -0.125 0.235133486497379\\
-8 -0.125 0.240756093548855\\
-7.8 -0.125 0.246653019039907\\
-7.6 -0.125 0.252844741632808\\
-7.4 -0.125 0.259353818926016\\
-7.2 -0.125 0.266205156409214\\
-7 -0.125 0.27342631895919\\
-6.8 -0.125 0.281047892866361\\
-6.6 -0.125 0.289103908124131\\
-6.4 -0.125 0.297632332889435\\
-6.2 -0.125 0.306675654752672\\
-6 -0.125 0.3162815668977\\
-5.8 -0.125 0.326503781594434\\
-5.6 -0.125 0.337402999022097\\
-5.4 -0.125 0.349048066532244\\
-5.2 -0.125 0.361517372608301\\
-5 -0.125 0.374900531602477\\
-4.8 -0.125 0.389300430681964\\
-4.6 -0.125 0.404835730423691\\
-4.4 -0.125 0.42164393664781\\
-4.2 -0.125 0.439885195307552\\
-4 -0.125 0.459747007000926\\
-3.8 -0.125 0.481450115868572\\
-3.6 -0.125 0.505255902419955\\
-3.4 -0.125 0.531475703546196\\
-3.2 -0.125 0.560482594606926\\
-3 -0.125 0.592726287053813\\
-2.8 -0.125 0.628751883698982\\
-2.6 -0.125 0.669223190982539\\
-2.4 -0.125 0.714950852799604\\
-2.2 -0.125 0.766924088023579\\
-2 -0.125 0.826340629089705\\
-1.8 -0.125 0.894618315877379\\
-1.6 -0.125 0.973342419456045\\
-1.4 -0.125 1.06402411466579\\
-1.2 -0.125 1.16732848291014\\
-1 -0.125 1.28080904204206\\
-0.799999999999999 -0.125 1.392353908763\\
-0.6 -0.125 1.46124476718074\\
-0.399999999999999 -0.125 1.366471464034\\
-0.199999999999999 -0.125 0.806974975923902\\
0 -0.125 -0.618223575602194\\
0.199999999999999 -0.125 -2.46387390496249\\
0.399999999999999 -0.125 -3.29362228856442\\
0.6 -0.125 -3.08693404516726\\
0.799999999999999 -0.125 -2.6161380080608\\
1 -0.125 -2.19026699873569\\
1.2 -0.125 -1.85507546164571\\
1.4 -0.125 -1.59668635548019\\
1.6 -0.125 -1.39560636288858\\
1.8 -0.125 -1.23640274580922\\
2 -0.125 -1.1080384511067\\
2.2 -0.125 -1.00276044154122\\
2.4 -0.125 -0.915084059650522\\
2.6 -0.125 -0.841069430783957\\
2.8 -0.125 -0.777835899465458\\
3 -0.125 -0.7232396288733\\
3.2 -0.125 -0.675658053731669\\
3.4 -0.125 -0.633843790584834\\
3.6 -0.125 -0.596824018589612\\
3.8 -0.125 -0.563830018738721\\
4 -0.125 -0.534247025073882\\
4.2 -0.125 -0.507577976160993\\
4.4 -0.125 -0.483416929928801\\
4.6 -0.125 -0.461429298865068\\
4.8 -0.125 -0.441336968615863\\
5 -0.125 -0.422906960727534\\
5.2 -0.125 -0.405942700395123\\
5.4 -0.125 -0.390277221809596\\
5.6 -0.125 -0.375767830788343\\
5.8 -0.125 -0.362291874892864\\
6 -0.125 -0.349743363431373\\
6.2 -0.125 -0.338030245637872\\
6.4 -0.125 -0.327072202942577\\
6.6 -0.125 -0.31679884603125\\
6.8 -0.125 -0.307148233048263\\
7 -0.125 -0.298065644402879\\
7.2 -0.125 -0.289502563989982\\
7.4 -0.125 -0.28141582750892\\
7.6 -0.125 -0.273766906866023\\
7.8 -0.125 -0.266521306033659\\
8 -0.125 -0.259648048687854\\
8.2 -0.125 -0.253119241807593\\
8.4 -0.125 -0.246909702450557\\
8.6 -0.125 -0.240996637315008\\
8.8 -0.125 -0.235359366600673\\
9 -0.125 -0.22997908520218\\
9.2 -0.125 -0.224838655490223\\
9.4 -0.125 -0.219922426921944\\
9.6 -0.125 -0.215216078522252\\
9.8 -0.125 -0.210706480930061\\
10 -0.125 -0.206381575237497\\
};
\addplot3 [
color=blue,
solid,
fill opacity=0.15,
draw=blue!80!white,thin,
fill=blue]
table[row sep=crcr] {
-10 0 0.213955412592128\\
-9.8 0 0.21824568369072\\
-9.6 0 0.222711075748448\\
-9.4 0 0.227362504139032\\
-9.2 0 0.232211807912982\\
-9 0 0.237271849299183\\
-8.8 0 0.242556626307403\\
-8.6 0 0.248081400481692\\
-8.4 0 0.253862842228103\\
-8.2 0 0.259919196591555\\
-8 0 0.266270472904458\\
-7.8 0 0.272938662397201\\
-7.6 0 0.279947988677345\\
-7.4 0 0.287325196988255\\
-7.2 0 0.295099889397276\\
-7 0 0.303304914601369\\
-6.8 0 0.311976822955324\\
-6.6 0 0.321156399730569\\
-6.4 0 0.330889292640614\\
-6.2 0 0.341226753506061\\
-6 0 0.352226518823318\\
-5.8 0 0.363953860273485\\
-5.6 0 0.376482844303973\\
-5.4 0 0.389897850433149\\
-5.2 0 0.404295411687141\\
-5 0 0.419786458703266\\
-4.8 0 0.436499073086466\\
-4.6 0 0.454581887758888\\
-4.4 0 0.474208315350672\\
-4.2 0 0.495581844447431\\
-4 0 0.518942723832916\\
-3.8 0 0.544576465386402\\
-3.6 0 0.572824749228629\\
-3.4 0 0.604099527142027\\
-3.2 0 0.638901415530217\\
-3 0 0.677843877338381\\
-2.8 0 0.72168524748174\\
-2.6 0 0.771371382452457\\
-2.4 0 0.828092579163572\\
-2.2 0 0.893359180657813\\
-2 0 0.969100130080565\\
-1.8 0 1.05778501176101\\
-1.6 0 1.16255484557214\\
-1.4 0 1.28729589605618\\
-1.2 0 1.43642927485621\\
-1 0 1.61368002234975\\
-0.799999999999999 0 1.81747267133694\\
-0.6 0 2.02559774986978\\
-0.399999999999999 0 2.1490843195471\\
-0.199999999999999 0 1.93220585642266\\
0 0 0.969100130080565\\
0.199999999999999 0 -0.601366154442754\\
0.399999999999999 0 -1.77417350107074\\
0.6 0 -2.14203557926455\\
0.799999999999999 0 -2.07057724287765\\
1 0 -1.87086643357144\\
1.2 0 -1.66245770551988\\
1.4 0 -1.47801619567804\\
1.6 0 -1.32211780682769\\
1.8 0 -1.19170034385802\\
2 0 -1.08233043105447\\
2.2 0 -0.989947588162061\\
2.4 0 -0.911226242130804\\
2.6 0 -0.843540763774447\\
2.8 0 -0.784839287527838\\
3 0 -0.733517023869009\\
3.2 0 -0.688311585096173\\
3.4 0 -0.648221990881918\\
3.6 0 -0.612447512478302\\
3.8 0 -0.580341832511851\\
4 0 -0.5513786563093\\
4.2 0 -0.525125787083739\\
4.4 0 -0.50122544521922\\
4.6 0 -0.479379208527689\\
4.8 0 -0.459336392071392\\
5 0 -0.440885006541303\\
5.2 0 -0.423844664862369\\
5.4 0 -0.408060972681948\\
5.6 0 -0.393401058210773\\
5.8 0 -0.379749983828173\\
6 0 -0.36700784536365\\
6.2 0 -0.355087411676595\\
6.4 0 -0.343912191770721\\
6.6 0 -0.333414842524152\\
6.8 0 -0.323535849556457\\
7 0 -0.314222428484274\\
7.2 0 -0.305427605058384\\
7.4 0 -0.297109441312861\\
7.6 0 -0.289230381538702\\
7.8 0 -0.281756697096138\\
8 0 -0.274658013154643\\
8.2 0 -0.267906903660521\\
8.4 0 -0.26147854337652\\
8.6 0 -0.255350407865584\\
8.8 0 -0.249502013914878\\
9 0 -0.243914694203827\\
9.2 0 -0.238571401077574\\
9.4 0 -0.233456535146865\\
9.6 0 -0.22855579513716\\
9.8 0 -0.223856045985004\\
10 0 -0.219345202653401\\
};
\addplot3 [
color=blue,
solid,
fill opacity=0.15,
draw=blue!80!white,thin,
fill=blue]
table[row sep=crcr] {
-10 0.125 0.20218930014727\\
-9.8 0.125 0.206334517577139\\
-9.6 0.125 0.210652671091931\\
-9.4 0.125 0.215154775536591\\
-9.2 0.125 0.219852798140095\\
-9 0.125 0.224759763346741\\
-8.8 0.125 0.229889871753767\\
-8.6 0.125 0.235258635411854\\
-8.4 0.125 0.240883032165623\\
-8.2 0.125 0.246781682221704\\
-8 0.125 0.252975050753828\\
-7.8 0.125 0.25948568111584\\
-7.6 0.125 0.266338464169418\\
-7.4 0.125 0.273560950389436\\
-7.2 0.125 0.281183712844771\\
-7 0.125 0.289240770942426\\
-6.8 0.125 0.297770087067755\\
-6.6 0.125 0.306814151084528\\
-6.4 0.125 0.316420671249645\\
-6.2 0.125 0.326643394679579\\
-6 0.125 0.33754308639102\\
-5.8 0.125 0.349188703546582\\
-5.6 0.125 0.361658811442264\\
-5.4 0.125 0.375043300763893\\
-5.2 0.125 0.389445482807695\\
-5 0.125 0.404984662230401\\
-4.8 0.125 0.421799317618655\\
-4.6 0.125 0.440051061805804\\
-4.4 0.125 0.459929610810905\\
-4.2 0.125 0.481659068898634\\
-4 0.125 0.505505946874136\\
-3.8 0.125 0.531789485075964\\
-3.6 0.125 0.560895072043694\\
-3.4 0.125 0.593291865071257\\
-3.2 0.125 0.629556175637019\\
-3 0.125 0.670402849530941\\
-2.8 0.125 0.716727849326238\\
-2.6 0.125 0.769666678936066\\
-2.4 0.125 0.830675361316787\\
-2.2 0.125 0.901643568242794\\
-2 0.125 0.985053150158203\\
-1.8 0.125 1.08419862235622\\
-1.6 0.125 1.20348422185186\\
-1.4 0.125 1.34878697864559\\
-1.2 0.125 1.52776434602454\\
-1 0.125 1.74957284502999\\
-0.799999999999999 0.125 2.02201277663192\\
-0.6 0.125 2.33925539491057\\
-0.399999999999999 0.125 2.63980414752451\\
-0.199999999999999 0.125 2.70444953030535\\
0 0.125 2.12901985774785\\
0.199999999999999 0.125 0.887870787551974\\
0.399999999999999 0.125 -0.291374667782859\\
0.6 0.125 -0.942421528310437\\
0.799999999999999 0.125 -1.17763962918451\\
1 0.125 -1.20912798749327\\
1.2 0.125 -1.16005077567375\\
1.4 0.125 -1.08496552491505\\
1.6 0.125 -1.00583152240462\\
1.8 0.125 -0.930880280967849\\
2 0.125 -0.862744703385691\\
2.2 0.125 -0.801820395061009\\
2.4 0.125 -0.747655482258536\\
2.6 0.125 -0.699527029122603\\
2.8 0.125 -0.656676533330413\\
3 0.125 -0.618400428160223\\
3.2 0.125 -0.584078761563551\\
3.4 0.125 -0.553178086358229\\
3.6 0.125 -0.525244353585323\\
3.8 0.125 -0.499892703729284\\
4 0.125 -0.476797075817995\\
4.2 0.125 -0.455680768772261\\
4.4 0.125 -0.436308295558384\\
4.6 0.125 -0.418478531190433\\
4.8 0.125 -0.40201902421809\\
5 0.125 -0.386781303664963\\
5.2 0.125 -0.372637016017696\\
5.4 0.125 -0.359474744964964\\
5.6 0.125 -0.347197388852686\\
5.8 0.125 -0.335719992392682\\
6 0.125 -0.324967948195766\\
6.2 0.125 -0.314875499739201\\
6.4 0.125 -0.305384490559637\\
6.6 0.125 -0.296443315144584\\
6.8 0.125 -0.288006035585334\\
7 0.125 -0.280031634935665\\
7.2 0.125 -0.272483383725911\\
7.4 0.125 -0.265328300487863\\
7.6 0.125 -0.25853669067679\\
7.8 0.125 -0.252081751212747\\
8 0.125 -0.245939230146961\\
8.2 0.125 -0.240087132803677\\
8.4 0.125 -0.234505467242519\\
8.6 0.125 -0.229176023101652\\
8.8 0.125 -0.224082178873527\\
9 0.125 -0.219208733476635\\
9.2 0.125 -0.214541758653522\\
9.4 0.125 -0.210068469274977\\
9.6 0.125 -0.20577710908495\\
9.8 0.125 -0.201656849798117\\
10 0.125 -0.197697701776103\\
};
\addplot3 [
color=blue,
solid,
fill opacity=0.15,
draw=blue!80!white,thin,
fill=blue]
table[row sep=crcr] {
-10 0.25 0.160525524501902\\
-9.8 0.25 0.16393188829912\\
-9.6 0.25 0.167485070927296\\
-9.4 0.25 0.171194718498734\\
-9.2 0.25 0.17507133634322\\
-9 0.25 0.179126386336549\\
-8.8 0.25 0.183372397695509\\
-8.6 0.25 0.187823093452992\\
-8.4 0.25 0.192493535250624\\
-8.2 0.25 0.197400289602924\\
-8 0.25 0.202561619419869\\
-7.8 0.25 0.207997705353123\\
-7.6 0.25 0.213730902493405\\
-7.4 0.25 0.219786039141378\\
-7.2 0.25 0.226190765866279\\
-7 0.25 0.232975964938696\\
-6.8 0.25 0.24017623258663\\
-6.6 0.25 0.247830449523399\\
-6.4 0.25 0.255982459026809\\
-6.2 0.25 0.264681876773422\\
-6 0.25 0.273985063004615\\
-5.8 0.25 0.283956295907262\\
-5.6 0.25 0.294669195998479\\
-5.4 0.25 0.30620846573799\\
-5.2 0.25 0.318672027852047\\
-5 0.25 0.332173671777955\\
-4.8 0.25 0.346846352853978\\
-4.6 0.25 0.362846337179142\\
-4.4 0.25 0.380358451985637\\
-4.2 0.25 0.399602795073322\\
-4 0.25 0.420843389546864\\
-3.8 0.25 0.444399460191926\\
-3.6 0.25 0.47066028347448\\
-3.4 0.25 0.50010496784554\\
-3.2 0.25 0.533329122704344\\
-3 0.25 0.571081279977728\\
-2.8 0.25 0.614313311046352\\
-2.6 0.25 0.664251200798822\\
-2.4 0.25 0.722495814424997\\
-2.2 0.25 0.791168333376414\\
-2 0.25 0.873122630772593\\
-1.8 0.25 0.97225755813437\\
-1.6 0.25 1.09397430949707\\
-1.4 0.25 1.24582622763415\\
-1.2 0.25 1.43835020200461\\
-1 0.25 1.68575280885866\\
-0.799999999999999 0.25 2.00481415243357\\
-0.6 0.25 2.40552820191741\\
-0.399999999999999 0.25 2.8522681187529\\
-0.199999999999999 0.25 3.15744112961746\\
0 0.25 2.91614521731048\\
0.199999999999999 0.25 1.99003869549211\\
0.399999999999999 0.25 0.929561658097673\\
0.6 0.25 0.189082132046746\\
0.799999999999999 0.25 -0.217969154511875\\
1 0.25 -0.416374755392451\\
1.2 0.25 -0.503465610738041\\
1.4 0.25 -0.534006086617912\\
1.6 0.25 -0.536261952437033\\
1.8 0.25 -0.524622735700298\\
2 0.25 -0.506478944088704\\
2.2 0.25 -0.485671249022539\\
2.4 0.25 -0.464205136569904\\
2.6 0.25 -0.443117090679497\\
2.8 0.25 -0.422922717792307\\
3 0.25 -0.40385437534283\\
3.2 0.25 -0.385990029664604\\
3.4 0.25 -0.369324433781651\\
3.6 0.25 -0.353808977360724\\
3.8 0.25 -0.339374160253151\\
4 0.25 -0.325942257151149\\
4.2 0.25 -0.313434368857685\\
4.4 0.25 -0.301774231240547\\
4.6 0.25 -0.290890143930557\\
4.8 0.25 -0.280715811668389\\
5 0.25 -0.271190564514213\\
5.2 0.25 -0.262259232766183\\
5.4 0.25 -0.253871840087616\\
5.6 0.25 -0.245983211366272\\
5.8 0.25 -0.23855255160591\\
6 0.25 -0.23154302789323\\
6.2 0.25 -0.224921371850191\\
6.4 0.25 -0.218657511195516\\
6.6 0.25 -0.212724233824427\\
6.8 0.25 -0.207096884788555\\
7 0.25 -0.201753094869696\\
7.2 0.25 -0.196672538568379\\
7.4 0.25 -0.191836718946746\\
7.6 0.25 -0.187228776673618\\
7.8 0.25 -0.182833320693535\\
8 0.25 -0.17863627810374\\
8.2 0.25 -0.17462476102698\\
8.4 0.25 -0.170786948485278\\
8.6 0.25 -0.167111981494316\\
8.8 0.25 -0.163589869800717\\
9 0.25 -0.160211408870876\\
9.2 0.25 -0.156968105908344\\
9.4 0.25 -0.153852113827058\\
9.6 0.25 -0.150856172240681\\
9.8 0.25 -0.147973554645375\\
10 0.25 -0.145198021075886\\
};
\end{axis}
\node at (-1.1,-0.7) {\large{\textbf{b}}};
\end{tikzpicture}%
}
\caption{\textbf{Influence of phase $\phi$ on the Fano resonance.} \textbf{a}, In some cases, particularly for small $N$, we observe asymmetric ($\mathcal{F}_{\text{max}} \, [\text{dB}] \ll |\mathcal{F}_{\text{min}} \, [\text{dB}]|$) Fano resonances. The data is well-fit by including a phase shift $\phi < 0$ -- physically linked to free-carrier generation (see \eqref{eq:Fano_phi}). \textbf{b}, Plot of the Fano function $\mathcal{F}(\Delta)$ as the phase shift $\phi$ is scanned ($r = 0.5$). The Fano resonance is symmetric ($\mathcal{F}_{\text{max}} \, [\text{dB}] = |\mathcal{F}_{\text{min}} \, [\text{dB}]|$) at $\phi = 0$; while for $\phi < 0$, the resonance becomes significantly deeper as in (\textbf{a}).}
\label{fig:Fano_phi}
\end{figure*}
In some cases, we observe $\mathcal{F}_{\text{max}} \, [\text{dB}] \ll |\mathcal{F}_{\text{min}} \, [\text{dB}]|$ -- a clear indication that \eqref{eq:Fano_homo2} is too simplistic. It turns out that the Fano function $\eqref{eq:Fano_homo2}$ must be replaced by \cite{Shin2013b}
\begin{equation}
\label{eq:Fano_phi}
\mathcal{F}(\Delta) = \left| 1 + e^{i \phi} r \frac{1}{-2\Delta + i} \right|^{2}
\end{equation}
with $r = f_{\text{s}} \frac{ \tilde{\mathcal{G}}}{4 \overline{\gamma}_{\text{tot}}}$ and $\overline{\gamma}_{\text{tot}} = |\overline{\gamma}_{\text{K}} - i \overline{\gamma}_{\text{TPA}} + \overline{\gamma}_{\text{FC}} \overline{P}|$ the \textit{total} nonlinearity, including two-photon absorption and free-carrier effects, $\phi = - \angle \left(\overline{\gamma}_{\text{K}} - i \overline{\gamma}_{\text{TPA}} + \overline{\gamma}_{\text{FC}} \overline{P}\right)$ and $\overline{P}$ the average power in the waveguide. The free-carrier nonlinearity $\overline{\gamma}_{\text{FC}}$ is complex as free carriers modulate both the index and the absorption -- both effects create an imprinted sideband on the probe; in addition, $\overline{\gamma}_{\text{FC}}(\Omega)$ depends on the modulation frequency since free carriers do not respond instantaneously. This is a slow dependency, so we take $\overline{\gamma}_{\text{FC}}$ constant in the range of our sweep \cite{Shin2013b}. This can be shown by solving for the carrier dynamics \cite{Dekker2007}
\begin{equation}
\label{eq:carrier}
\partial_{t} N_{\text{c}} = \frac{\beta_{\text{TPA}}}{2 \hbar \omega} P^{2} - \kappa_{\text{c}}N_{\text{c}}
\end{equation}
in frequency-domain and using the proportionality $\Delta n \propto - N_{\text{c}}$ and $\Delta \alpha \propto N_{\text{c}}$ between both the index and absorption and the carrier concentration $N_{\text{c}}$ \cite{Dekker2007}. Here we denote $\beta_{\text{TPA}}$ the two-photon absorption coefficient and $\kappa_{\text{c}}$ the free-carrier recombination rate.
Notably, $\phi > 0$ in absence of free carriers ($\overline{\gamma}_{\text{FC}}=0$). The observed $\phi < 0$ (fig.\ref{fig:Fano_phi}a) is thus linked to the presence of free carriers; in addition, we still use $\overline{\gamma}_{\text{tot}} \approx \overline{\gamma}_{\text{K}} \approx 610 \, \text{W}^{-1}\text{m}^{-1}$ on the assumption that the Kerr effect remains the dominant background nonlinearity. This is consistent with the observations that (1) in most cases $\mathcal{F}_{\text{max}} \, [\text{dB}] \approx - \mathcal{F}_{\text{min}} \, [\text{dB}]$ and thus $\phi \approx 0$, (2) the background is flat (fig.\ref{fig:photonphononoverlap}$\&$\ref{fig:Fano_phi}a) and (3) the Brillouin efficiencies deduced from the XPM experiment are in reasonable agreement with those inferred from the gain experiment (fig.\ref{fig:photonphononoverlap}b).
\subsection*{Drop in free-carrier lifetime}
Using the set-up presented in \cite{VanLaer2015}, we measured a significant drop in the free-carrier lifetime in the suspended beams with respect to the regular non-suspended waveguide. Both this finding and the higher propagation losses likely originate in a deterioration of the silicon wires' surface state during the fabrication of the suspended beams.
\begin{figure*}
\centering
\subfigure{
\begin{tikzpicture}
\begin{axis}[scale = 0.34*\textwidth/(4.5in),
width=4.52083333333333in,
height=3.565625in,
xmin=-1,
xmax=6.5,
xlabel={Time (ns)},
ymin=0.7,
ymax=1.05,
ylabel={Probe transmission (-)},
ylabel style = {yshift=-1.5ex},
axis x line*=bottom,
axis y line*=left,
legend style={fill=none,draw=none,legend cell align=left,at={(0.70,0.37)},anchor=north},
xtick = {0,2,4,6},
ytick = {0.8,0.9,1.0},
minor xtick={1,2,...,7},
minor ytick={0.75,0.8,...,1},
clip mode=individual
]
\addplot [
color=blue,
line width=0.2pt,
mark size=1.25pt,
only marks,
mark=o,
mark options={solid}
]
table[row sep=crcr]{
-0.999400000000009 0.997494424611349\\
-0.989999999999995 0.994862233812396\\
-0.980599999999995 1.00012582519991\\
-0.971199999999996 1.00069477667852\\
-0.9619 1.00455021315649\\
-0.952500000000001 0.99539088456127\\
-0.943100000000001 0.99900530687121\\
-0.933700000000002 1.00405000998154\\
-0.924400000000006 1.00112544133942\\
-0.914999999999992 1.00009737762598\\
-0.905600000000007 0.999356160283019\\
-0.896199999999993 1.00409584218399\\
-0.886899999999997 0.996718438011357\\
-0.877499999999998 1.00224042819541\\
-0.868099999999998 0.996603857505248\\
-0.858699999999999 0.999901405450019\\
-0.849399999999989 1.00178605722291\\
-0.840000000000003 1.00316497434815\\
-0.83059999999999 0.99889625783781\\
-0.821200000000005 0.998637859041275\\
-0.811900000000009 1.00233762407301\\
-0.802499999999995 0.998608621256958\\
-0.793099999999995 0.998878873209297\\
-0.783699999999996 1.00183267963574\\
-0.7744 1.0024529947895\\
-0.765000000000001 0.999844510302158\\
-0.755600000000001 0.993937687659656\\
-0.746200000000002 0.999913258605824\\
-0.736899999999991 1.00174575649318\\
-0.727500000000006 0.994937303799157\\
-0.718100000000007 1.00436609413632\\
-0.708699999999993 1.00204524622983\\
-0.699399999999997 1.00132694498809\\
-0.689999999999998 1.00208475674918\\
-0.680599999999998 1.00017718887507\\
-0.671199999999999 0.999913258605824\\
-0.661900000000003 1.00365964605039\\
-0.652500000000003 0.999545810775889\\
-0.64309999999999 1.00471694754813\\
-0.633700000000005 1.00211952600621\\
-0.624400000000009 1.00006813984167\\
-0.614999999999995 1.00009105594289\\
-0.60560000000001 1.00022934276061\\
-0.596199999999996 0.994937303799157\\
-0.5869 0.996890703875714\\
-0.577500000000001 0.999149125161636\\
-0.568100000000001 0.999585321295236\\
-0.558700000000002 1.0006600074215\\
-0.549399999999991 1.00284968040375\\
-0.540000000000006 1.00248144236343\\
-0.530600000000007 0.998246704899732\\
-0.521199999999993 0.998764292703188\\
-0.511899999999997 1.00375684192798\\
-0.502499999999998 0.99673582263987\\
-0.493099999999998 1.00028702811885\\
-0.483699999999999 0.99956872687711\\
-0.474400000000003 1.00117759522495\\
-0.465000000000003 1.00134986108931\\
-0.45559999999999 0.997861872441284\\
-0.446200000000005 1.00157349062882\\
-0.436900000000009 1.00074060888097\\
-0.427499999999995 1.00360749216485\\
-0.41810000000001 1.00175760964898\\
-0.408699999999996 1.00213691063472\\
-0.3994 1.00108514060968\\
-0.390000000000001 0.995356905514631\\
-0.380600000000001 0.995552087480209\\
-0.371200000000002 1.00187851183819\\
-0.361899999999991 0.99349516984296\\
-0.352500000000006 0.997258941916035\\
-0.343099999999993 0.999246321039232\\
-0.333699999999993 1.00450438095404\\
-0.324399999999997 1.0003265386382\\
-0.314999999999998 1.00282597409215\\
-0.305599999999998 0.996103654330305\\
-0.296199999999999 0.993156169586956\\
-0.286900000000003 1.00101639230601\\
-0.277500000000003 1.00247591089072\\
-0.26809999999999 0.996712906538649\\
-0.258700000000005 1.00103377693453\\
-0.249400000000009 0.998694754189136\\
-0.239999999999995 0.998987922242697\\
-0.23060000000001 1.0015624276834\\
-0.221199999999996 0.998591236628445\\
-0.2119 0.891740568524923\\
-0.202499999999986 0.736963482410395\\
-0.193100000000001 0.778443127021491\\
-0.183700000000002 0.816986665918811\\
-0.174399999999991 0.800886919494954\\
-0.165000000000006 0.807977477297117\\
-0.155599999999993 0.799271729464014\\
-0.146200000000007 0.798726484297014\\
-0.136900000000011 0.813338264562232\\
-0.127499999999998 0.8100746956641\\
-0.118099999999998 0.809632177847404\\
-0.108699999999999 0.813768929223123\\
-0.0994000000000028 0.822939320763757\\
-0.0900000000000034 0.82819659046818\\
-0.080600000000004 0.834586231657113\\
-0.0712000000000046 0.836936317347922\\
-0.0618999999999943 0.840222802347275\\
-0.0524999999999949 0.864401659767379\\
-0.0431000000000097 0.860477474985752\\
-0.0336999999999961 0.865780576892619\\
-0.0244 0.870728084125355\\
-0.0150000000000006 0.870182048747968\\
-0.00560000000000116 0.880950245691028\\
0.00379999999999825 0.881731763763728\\
0.0131000000000085 0.884357632879586\\
0.0224999999999937 0.886891047380169\\
0.0319000000000074 0.891057036540205\\
0.0412999999999926 0.890545770419844\\
0.0505999999999887 0.898613028260286\\
0.0600000000000023 0.897963475322207\\
0.0694000000000017 0.904312815781405\\
0.0788000000000011 0.905800781940045\\
0.0880999999999972 0.90546178168404\\
0.0974999999999966 0.912730927033656\\
0.106899999999996 0.910587086253842\\
0.116299999999995 0.914143823205534\\
0.125600000000006 0.917143461834422\\
0.135000000000005 0.914609257123452\\
0.144400000000005 0.921183017332545\\
0.153800000000004 0.920062499003841\\
0.1631 0.920452862934997\\
0.172499999999999 0.919947128287345\\
0.181899999999999 0.924567488419881\\
0.191300000000012 0.924314621096055\\
0.200600000000009 0.929789988867279\\
0.209999999999994 0.925096139168755\\
0.219400000000007 0.927681707554878\\
0.228799999999993 0.933990747284341\\
0.238099999999989 0.933191844583128\\
0.247500000000002 0.93008315692084\\
0.256900000000002 0.932577060902075\\
0.266300000000001 0.935863545901429\\
0.275599999999997 0.932519375543827\\
0.284999999999997 0.936415112751524\\
0.294399999999996 0.933375963603288\\
0.303799999999995 0.939230632360251\\
0.313100000000006 0.940788137032943\\
0.322500000000005 0.937386281317094\\
0.331900000000005 0.942574802717852\\
0.341300000000004 0.944672021084836\\
0.3506 0.944298251571805\\
0.359999999999999 0.941925249779774\\
0.369399999999999 0.946068322838588\\
0.378800000000012 0.947659806557919\\
0.388100000000009 0.95022245884282\\
0.397499999999994 0.949079024412894\\
0.406900000000007 0.948200310462598\\
0.416299999999993 0.948510072934285\\
0.425600000000003 0.945901588446941\\
0.435000000000002 0.944803986219457\\
0.444399999999987 0.953302698931178\\
0.453800000000001 0.945022874496645\\
0.463099999999997 0.947775177274415\\
0.472500000000011 0.951567396921421\\
0.481899999999996 0.949084555885602\\
0.491299999999995 0.951911928650134\\
0.500600000000006 0.948774793413915\\
0.509999999999991 0.954629462170878\\
0.519400000000005 0.951418047158286\\
0.528800000000004 0.952928929418147\\
0.5381 0.959387318910745\\
0.547499999999999 0.95826680058204\\
0.556899999999999 0.957433918834188\\
0.566299999999998 0.956525967099574\\
0.575599999999994 0.951228396665416\\
0.584999999999994 0.954555182394504\\
0.594400000000007 0.956307869032774\\
0.603799999999993 0.956238330518722\\
0.613100000000003 0.954675294373322\\
0.622500000000002 0.960852368968163\\
0.631900000000002 0.953411747964578\\
0.641300000000001 0.957973632528479\\
0.650599999999997 0.957330401273496\\
0.660000000000011 0.956336316606705\\
0.669399999999996 0.956273099775748\\
0.67880000000001 0.961277502156346\\
0.688100000000006 0.956680848335418\\
0.697499999999991 0.960232053814402\\
0.706900000000005 0.954962930954174\\
0.716300000000004 0.958295248155971\\
0.7256 0.955491581703048\\
0.734999999999999 0.963513007341046\\
0.744399999999999 0.957117834679405\\
0.753799999999998 0.963794322238803\\
0.763099999999994 0.958261269109332\\
0.772499999999994 0.962668272437389\\
0.781900000000007 0.960030550165728\\
0.791299999999993 0.959990249435993\\
0.800600000000003 0.963098937098281\\
0.810000000000002 0.960806536765719\\
0.819400000000002 0.95920952157368\\
0.828800000000001 0.960232053814402\\
0.838099999999997 0.958013933258214\\
0.847500000000011 0.962668272437389\\
0.856899999999996 0.957146282253335\\
0.86630000000001 0.962737020741055\\
0.875600000000006 0.961008040414393\\
0.884999999999991 0.96344979051009\\
0.894400000000005 0.964282672257942\\
0.903800000000004 0.962450174370589\\
0.9131 0.961972887296867\\
0.922499999999999 0.961898607520494\\
0.931899999999999 0.961386551189746\\
0.941299999999998 0.963983972731673\\
0.950599999999994 0.967121107967891\\
0.960000000000008 0.968149961891709\\
0.969400000000007 0.967121107967891\\
0.978799999999993 0.96344979051009\\
0.988100000000003 0.961173984595654\\
0.997500000000002 0.964834239108038\\
1.0069 0.962593202450629\\
1.0163 0.96261611855185\\
1.0256 0.959743703795262\\
1.035 0.960151452354932\\
1.0444 0.96542057521516\\
1.05380000000001 0.966345911578286\\
1.06310000000001 0.967724828703526\\
1.07249999999999 0.96407563713656\\
1.0819 0.970424977595757\\
1.09129999999999 0.966230540861791\\
1.1006 0.968149961891709\\
1.11 0.964714127129221\\
1.1194 0.967477492852408\\
1.1288 0.965397659113938\\
1.13809999999999 0.963271202962637\\
1.14750000000001 0.969724060982527\\
1.15689999999999 0.964495238852034\\
1.16630000000001 0.964891924466286\\
1.1756 0.964191007853055\\
1.185 0.963673420049598\\
1.1944 0.970229795630179\\
1.2038 0.970034613664601\\
1.2131 0.967569157257295\\
1.2225 0.968132577263196\\
1.2319 0.97241314692934\\
1.24130000000001 0.970161047326514\\
1.25060000000001 0.966017974267699\\
1.25999999999999 0.963483769556729\\
1.2694 0.96877027704547\\
1.27879999999999 0.969293396321635\\
1.2881 0.960226522341693\\
1.2975 0.964955141297242\\
1.3069 0.962835006829037\\
1.3163 0.964816854479525\\
1.32559999999999 0.961570670209907\\
1.33500000000001 0.969867879272953\\
1.34439999999999 0.974958414585728\\
1.35380000000001 0.979359886441077\\
1.3631 0.961214285325389\\
1.3725 0.966759191610665\\
1.3819 0.976670800494264\\
1.3913 0.968183940938348\\
1.4006 0.972528517645836\\
1.41 0.970258243204109\\
1.4194 0.971792831775579\\
1.42880000000001 0.969465662185992\\
1.43810000000001 0.968925158281313\\
1.44750000000001 0.969367676098009\\
1.4569 0.972465300814879\\
1.46629999999999 0.97114881031021\\
1.4756 0.972458979131784\\
1.485 0.970804278581497\\
1.49440000000001 0.972551433747058\\
1.5038 0.969511494388435\\
1.51309999999999 0.974803533349885\\
1.52250000000001 0.972591734476793\\
1.53189999999999 0.973435679170062\\
1.54130000000001 0.972809832543593\\
1.5506 0.971769915674358\\
1.56 0.973786532581872\\
1.5694 0.974044931378406\\
1.5788 0.973803917210384\\
1.5881 0.968718913370318\\
1.5975 0.970591711987405\\
1.6069 0.973970651602032\\
1.6163 0.971310013229149\\
1.62559999999999 0.97305716839471\\
1.63500000000001 0.971562880552975\\
1.6444 0.969114808774183\\
1.65379999999999 0.977630906114416\\
1.6631 0.975234198010776\\
1.6725 0.978711123713386\\
1.6819 0.97379206405458\\
1.6913 0.964955141297242\\
1.70059999999999 0.971160663466014\\
1.71000000000001 0.970247180258692\\
1.71939999999999 0.97152257982324\\
1.72880000000001 0.972091531301849\\
1.7381 0.973143301326889\\
1.74749999999999 0.970953628344631\\
1.7569 0.97163162885664\\
1.7663 0.974280414073719\\
1.77560000000001 0.974395784790215\\
1.785 0.971701167370692\\
1.7944 0.974067847479628\\
1.8038 0.971896349336271\\
1.81309999999999 0.973505217684115\\
1.82250000000001 0.969804662441996\\
1.8319 0.976182450475124\\
1.8413 0.970770299534857\\
1.8506 0.971396146161327\\
1.86 0.972804301070884\\
1.8694 0.975981737036837\\
1.8788 0.970528495156449\\
1.88809999999999 0.970879348568257\\
1.89750000000001 0.97360241356171\\
1.90689999999999 0.971143278837501\\
1.91630000000001 0.968115192634682\\
1.9256 0.972338867152967\\
1.935 0.97301054598188\\
1.9444 0.970993929074366\\
1.9538 0.971999866896962\\
1.96310000000001 0.97737803879059\\
1.9725 0.971132215892084\\
1.98190000000001 0.968822430931009\\
1.9913 0.969976928306353\\
2.00059999999999 0.971057145905323\\
2.01000000000001 0.969799130969287\\
2.0194 0.976257520461885\\
2.0288 0.973935882345006\\
2.0381 0.970948096871923\\
2.0475 0.975297414841733\\
2.0569 0.9676900594465\\
2.0663 0.975389869457007\\
2.07559999999999 0.977257136601385\\
2.08500000000001 0.974941820167602\\
2.09439999999999 0.973033462083102\\
2.10380000000001 0.973596882089002\\
2.1131 0.973212049630554\\
2.1225 0.974631267485528\\
2.1319 0.972338867152967\\
2.1413 0.974176896513028\\
2.15060000000001 0.971999866896962\\
2.16 0.971223880296971\\
2.16940000000001 0.974074169162724\\
2.1788 0.97330924550815\\
2.18809999999999 0.969833110015927\\
2.19750000000001 0.978716655186095\\
2.20689999999999 0.978986907138434\\
2.2163 0.974596498228502\\
2.2256 0.974929967011798\\
2.235 0.972309629368649\\
2.2444 0.978538067638642\\
2.2538 0.976963968547825\\
2.26310000000001 0.972298566423232\\
2.27249999999999 0.97811293445046\\
2.28189999999999 0.972694461827097\\
2.29130000000001 0.977302968803829\\
2.3006 0.975751785814233\\
2.31 0.975211281909554\\
2.3194 0.977412017837229\\
2.3288 0.975148065078598\\
2.3381 0.974056784534211\\
2.3475 0.975671184354763\\
2.35690000000001 0.974120001365167\\
2.3663 0.974998715315463\\
2.37559999999999 0.97535510019998\\
2.38500000000001 0.977619052958612\\
2.39439999999999 0.977929605640686\\
2.4038 0.970793215636079\\
2.4131 0.973447532325867\\
2.4225 0.976998737804851\\
2.4319 0.975613498996515\\
2.4413 0.978372123457382\\
2.45060000000001 0.973649035974541\\
2.45999999999999 0.97673401732522\\
2.46940000000001 0.978222773694247\\
2.47880000000001 0.975383547773911\\
2.4881 0.975458617760672\\
2.4975 0.982158021421291\\
2.5069 0.975746254341524\\
2.5163 0.979434956427839\\
2.5256 0.969476725131409\\
2.535 0.979934369392395\\
2.54440000000001 0.973389846967619\\
2.5538 0.976935520973894\\
2.56309999999999 0.979612753764904\\
2.57250000000001 0.976670800494264\\
2.58189999999999 0.978383186402799\\
2.5913 0.978906305678965\\
2.6006 0.974194281141541\\
2.61 0.97662496829182\\
2.6194 0.974102616736654\\
2.6288 0.978538067638642\\
2.63810000000001 0.980882621856743\\
2.64749999999999 0.978274137369399\\
2.65690000000001 0.981244538213969\\
2.66630000000001 0.979400187170812\\
2.6756 0.974418700891437\\
2.685 0.979767635000747\\
2.6944 0.977980969315838\\
2.7038 0.976584667562085\\
2.7131 0.97733141637776\\
2.7225 0.970591711987405\\
2.73190000000001 0.97740648636452\\
2.7413 0.982652693123526\\
2.75060000000001 0.97390111308798\\
2.76000000000001 0.975177302862915\\
2.76939999999999 0.97897505398263\\
2.7788 0.974050462851115\\
2.7881 0.977429402465742\\
2.7975 0.979584306190973\\
2.8069 0.977935137113395\\
2.8163 0.974838302606911\\
2.82560000000001 0.972349930098384\\
2.83499999999999 0.976590199034794\\
2.84440000000001 0.977527388553725\\
2.85379999999999 0.979526620832725\\
2.86309999999999 0.97964752302193\\
2.8725 0.975550282165559\\
2.8819 0.978228305166955\\
2.8913 0.979078571543321\\
2.9006 0.980928454059187\\
2.91 0.972488216916101\\
2.9194 0.974378400161702\\
2.9288 0.977216835871651\\
2.93810000000001 0.977148087567986\\
2.94750000000001 0.979141788374278\\
2.95689999999999 0.976791702683468\\
2.9663 0.977171003669207\\
2.9756 0.980067124737404\\
2.985 0.977584283701586\\
2.9944 0.972660482780458\\
3.00380000000001 0.980722209148191\\
3.01310000000001 0.974590966755793\\
3.02249999999999 0.981796105064065\\
3.03190000000001 0.980388740364895\\
3.04129999999999 0.977527388553725\\
3.05059999999999 0.97908410301603\\
3.06 0.976677122177359\\
3.0694 0.982847875089105\\
3.0788 0.979136256901569\\
3.0881 0.980692971363874\\
3.0975 0.979796872785065\\
3.1069 0.977791318822968\\
3.1163 0.984611624672792\\
3.12560000000001 0.974286735756815\\
3.13500000000001 0.977148087567986\\
3.1444 0.973803917210384\\
3.1538 0.975602436051098\\
3.1631 0.981962839455713\\
3.1725 0.977664885161055\\
3.1819 0.980750656722121\\
3.19130000000001 0.974240113343985\\
3.20060000000001 0.978400571031312\\
3.20999999999999 0.979354354968369\\
3.21940000000001 0.978929221780186\\
3.22879999999999 0.98388779195834\\
3.2381 0.973642714291445\\
3.2475 0.981841937266509\\
3.2569 0.976952905602407\\
3.2663 0.975889282421563\\
3.2756 0.978372123457382\\
3.28500000000001 0.978124787606264\\
3.2944 0.983077826311709\\
3.3038 0.976142149745389\\
3.31310000000001 0.97651591925842\\
3.32250000000001 0.985071527118001\\
3.3319 0.978251221268177\\
3.3413 0.977073017581225\\
3.3506 0.979905921818465\\
3.36 0.980572069174669\\
3.3694 0.97881385106369\\
3.3788 0.974768764092859\\
3.38809999999999 0.980348439635161\\
3.39749999999999 0.98008371915553\\
3.40690000000001 0.979928837919687\\
3.41629999999999 0.977826088079994\\
3.4256 0.97956692156246\\
3.435 0.977682269789568\\
3.44439999999999 0.979819788886286\\
3.4538 0.984508107112101\\
3.4631 0.979716271325595\\
3.47250000000001 0.979503704731504\\
3.4819 0.976837534885911\\
3.4913 0.981423125761422\\
3.50060000000001 0.976412401697729\\
3.50999999999999 0.977555836127655\\
3.5194 0.976148471428485\\
3.5288 0.97845825638956\\
3.5381 0.979630138393417\\
3.5475 0.982382441171187\\
3.5569 0.980336586479356\\
3.5663 0.979331438867147\\
3.57559999999999 0.979779488156551\\
3.58499999999999 0.977958053214616\\
3.59440000000001 0.984089295607014\\
3.60379999999999 0.984813128321466\\
3.6131 0.974234581871276\\
3.6225 0.97621721973215\\
3.6319 0.976791702683468\\
3.6413 0.980917391113769\\
3.6506 0.98258947629257\\
3.66000000000001 0.977992822471642\\
3.6694 0.982204643834122\\
3.67880000000001 0.981468957963865\\
3.68810000000001 0.979222389833747\\
3.69749999999999 0.98280757435937\\
3.7069 0.982692993853261\\
3.7163 0.979992054750643\\
3.7256 0.975826065590607\\
3.735 0.981193174538817\\
3.7444 0.979986523277934\\
3.7538 0.978331822727647\\
3.76309999999999 0.981250859897065\\
3.77250000000001 0.978044186146794\\
3.78190000000001 0.983819043654675\\
3.79129999999999 0.98056653770196\\
3.8006 0.977366185634786\\
3.81 0.983939155633492\\
3.8194 0.978033123201377\\
3.8288 0.986766528398023\\
3.8381 0.981750272861622\\
3.84750000000001 0.983439742668935\\
3.8569 0.978216452011151\\
3.86630000000001 0.978837557375299\\
3.87560000000001 0.980652670634139\\
3.88499999999999 0.975866366320341\\
3.8944 0.976739548797929\\
3.90379999999999 0.98202605628667\\
3.9131 0.983520344128405\\
3.9225 0.982893707291548\\
3.9319 0.977429402465742\\
3.9413 0.98642199666931\\
3.95059999999999 0.976440849271659\\
3.96000000000001 0.978314438099134\\
3.96939999999999 0.983836428283188\\
3.97879999999999 0.978653438355138\\
3.9881 0.9822907767663\\
3.9975 0.975630883625028\\
4.0069 0.984904792726353\\
4.0163 0.984123274653653\\
4.0256 0.979716271325595\\
4.035 0.979159173002791\\
4.0444 0.980434572567339\\
4.05380000000001 0.982456720947561\\
4.06310000000001 0.97767673831686\\
4.07249999999999 0.985800891305162\\
4.0819 0.977826088079994\\
4.09129999999999 0.981302223572217\\
4.1006 0.983859344384409\\
4.11 0.978463787862269\\
4.1194 0.983606477060583\\
4.1288 0.9817273567604\\
4.13809999999999 0.986961710363602\\
4.14750000000001 0.982876322663035\\
4.15689999999999 0.982566560191348\\
4.16630000000001 0.97881385106369\\
4.1756 0.98158986015307\\
4.185 0.978509620064712\\
4.1944 0.981543237740239\\
4.2038 0.975309267997537\\
4.2131 0.983973924890518\\
4.2225 0.982795721203566\\
4.2319 0.980882621856743\\
4.24130000000001 0.983077826311709\\
4.25060000000001 0.985301478340606\\
4.25999999999999 0.980692971363874\\
4.2694 0.98202605628667\\
4.27879999999999 0.979302991293217\\
4.2881 0.984146190754875\\
4.2975 0.987295179146898\\
4.30690000000001 0.978228305166955\\
4.3163 0.979980991805226\\
4.32559999999999 0.97953847398853\\
4.33500000000001 0.981336992829243\\
4.34439999999999 0.98866303332672\\
4.35380000000001 0.978544389321738\\
4.3631 0.981595391625778\\
4.3725 0.981773188962843\\
4.3819 0.977555836127655\\
4.3913 0.979791341312356\\
4.4006 0.974499302350906\\
4.41 0.98337099436527\\
4.4194 0.97900350155656\\
4.42880000000001 0.984881876625132\\
4.43810000000001 0.984623477828597\\
4.44750000000001 0.980313670378134\\
4.4569 0.980503320871004\\
4.46629999999999 0.98466931003104\\
4.4756 0.984899261253644\\
4.485 0.983083357784418\\
4.4944 0.981095188450835\\
4.5038 0.984686694659553\\
4.51309999999999 0.98008371915553\\
4.52250000000001 0.980917391113769\\
4.53189999999999 0.984335841247744\\
4.54130000000001 0.986358779838354\\
4.5506 0.980078187682821\\
4.55999999999999 0.983462658770157\\
4.5694 0.981031971619878\\
4.5788 0.984238645370149\\
4.58810000000001 0.979509236204212\\
4.5975 0.982433804846339\\
4.6069 0.980773572823343\\
4.6163 0.986691458411262\\
4.62559999999999 0.98191700725327\\
4.63500000000001 0.985284093712093\\
4.6444 0.984387995133283\\
4.6538 0.983543260229627\\
4.6631 0.982543644090127\\
4.6725 0.980463020141269\\
4.6819 0.9809458386877\\
4.6913 0.982284455083205\\
4.70059999999999 0.982250476036565\\
4.71000000000001 0.978262284213595\\
4.71939999999999 0.981871175050826\\
4.72880000000001 0.984358757348966\\
4.7381 0.98366969389154\\
4.74749999999999 0.982353993597257\\
4.7569 0.9826179238665\\
4.7663 0.985031226388266\\
4.77560000000001 0.982681140697457\\
4.785 0.97781423492419\\
4.7944 0.983864875857118\\
4.8038 0.984681163186844\\
4.81309999999999 0.981865643578117\\
4.82250000000001 0.983141043142666\\
4.8319 0.9805096425541\\
4.8413 0.982842343616396\\
4.8506 0.984600561727375\\
4.86 0.987283325991093\\
4.8694 0.981124426235152\\
4.8788 0.987657095504124\\
4.88809999999999 0.988220515510024\\
4.89750000000001 0.98485896052391\\
4.90689999999999 0.980199089872026\\
4.91630000000001 0.981135489180569\\
4.9256 0.983663372208444\\
4.935 0.979911453291173\\
4.9444 0.980290754276912\\
4.9538 0.9808367896543\\
4.96310000000001 0.983169490716596\\
4.9725 0.986496276445684\\
4.98190000000001 0.982204643834122\\
4.9913 0.983525875601114\\
5.00059999999999 0.979653054494639\\
5.01000000000001 0.980440104040048\\
5.0194 0.987243815471745\\
5.0288 0.986139891561167\\
5.0381 0.984221260741636\\
5.0475 0.983939155633492\\
5.0569 0.985508513461988\\
5.0663 0.981939923354492\\
5.07560000000001 0.98574952763001\\
5.08499999999999 0.98164675530093\\
5.09439999999999 0.985106296375027\\
5.10380000000001 0.98499645713124\\
5.1131 0.987151360856471\\
5.1225 0.985134743948958\\
5.1319 0.981175789910304\\
5.1413 0.983192406817818\\
5.1506 0.98593917812288\\
5.16 0.979101487644543\\
5.16940000000001 0.980440104040048\\
5.1788 0.988496298935072\\
5.18809999999999 0.986542108648128\\
5.19750000000001 0.982756210684218\\
5.20689999999999 0.978285200314816\\
5.2163 0.98369814146547\\
5.2256 0.979911453291173\\
5.235 0.983439742668935\\
5.2444 0.985364695171562\\
5.2538 0.982658224596235\\
5.26310000000001 0.982037909442474\\
5.27249999999999 0.981095188450835\\
5.28189999999999 0.984221260741636\\
5.29130000000001 0.982531790934322\\
5.3006 0.987030458667267\\
5.31 0.984767296119023\\
5.3194 0.981543237740239\\
5.3288 0.982180937522513\\
5.3381 0.9825088748331\\
5.3475 0.983238239020262\\
5.35690000000001 0.984365079032062\\
5.3663 0.982261538981983\\
5.37559999999999 0.981502937010504\\
5.38500000000001 0.98391623953227\\
5.39439999999999 0.976992416121755\\
5.4038 0.983393910466492\\
5.4131 0.982405357272409\\
5.4225 0.980199089872026\\
5.4319 0.983382057310688\\
5.4413 0.982790189730857\\
5.45060000000001 0.98399684099174\\
5.45999999999999 0.987651564031415\\
5.46940000000001 0.988157298679067\\
5.47880000000001 0.984439358808436\\
5.4881 0.983819043654675\\
5.4975 0.982658224596235\\
5.5069 0.98585857666341\\
5.5163 0.984410911234505\\
5.5256 0.986030842527767\\
5.535 0.984387995133283\\
5.54440000000001 0.989432698243616\\
5.5538 0.979291138137412\\
5.56309999999999 0.984950624928797\\
5.57250000000001 0.980595775486278\\
5.58189999999999 0.982830490460592\\
5.5913 0.982686672170166\\
5.6006 0.984112211708236\\
5.61000000000001 0.988806061406759\\
5.6194 0.982445658002144\\
5.6288 0.983962861945101\\
5.63810000000001 0.978527004693225\\
5.64749999999999 0.986703311567067\\
5.65690000000001 0.985737674474206\\
5.66629999999999 0.981802426747161\\
5.67559999999999 0.9802733696484\\
5.685 0.986881108904132\\
5.6944 0.985807212988258\\
5.7038 0.983111805358348\\
5.7131 0.984519960267905\\
5.7225 0.981785042118648\\
5.7319 0.987249346944454\\
5.7413 0.985853045190701\\
5.75060000000001 0.985106296375027\\
5.76000000000001 0.987231962315941\\
5.76939999999999 0.981934391881783\\
5.7788 0.98866303332672\\
5.7881 0.984065589295405\\
5.7975 0.982882644346131\\
5.8069 0.984801275165662\\
5.8163 0.984330309775036\\
5.82560000000001 0.981555090896043\\
5.83499999999999 0.983755826823718\\
5.84440000000001 0.984421974179923\\
5.85379999999999 0.987300710619606\\
5.86309999999999 0.985743996157301\\
5.8725 0.988111466476624\\
5.8819 0.983009078008044\\
5.8913 0.986019779582349\\
5.9006 0.990242663890246\\
5.91 0.982756210684218\\
5.9194 0.98682974522898\\
5.9288 0.981118104552056\\
5.93810000000001 0.983629393161805\\
5.94750000000001 0.981170258437596\\
5.9569 0.988622732596985\\
5.9663 0.98839278137438\\
5.9756 0.988645648698207\\
5.985 0.984904792726353\\
5.9944 0.985226408353845\\
6.00380000000001 0.987748759909011\\
6.01310000000001 0.981307755044926\\
6.02249999999999 0.985910730548949\\
6.03190000000001 0.987180598640789\\
6.04129999999999 0.989932901418559\\
6.05059999999999 0.983439742668935\\
6.06 0.98310627388564\\
6.0694 0.988702543846067\\
6.0788 0.987381312079076\\
6.0881 0.983652309263027\\
6.0975 0.987576494044654\\
6.1069 0.982882644346131\\
6.1163 0.989036012629363\\
6.12560000000001 0.985329925914536\\
6.13500000000001 0.98415251243797\\
6.1444 0.986956178890893\\
6.1538 0.983186875345109\\
6.1631 0.985824597616771\\
6.1725 0.982112189218848\\
6.1819 0.988248963083954\\
6.19130000000001 0.988064844063793\\
6.20060000000001 0.986743612296802\\
6.20999999999999 0.986870045958715\\
6.21940000000001 0.984421974179923\\
6.22879999999999 0.983491106344088\\
6.2381 0.979848236460217\\
6.2475 0.982003140185448\\
6.25689999999999 0.982606860921083\\
6.2663 0.983118127041444\\
6.2756 0.991277049286773\\
6.28500000000001 0.982210175306831\\
6.2944 0.985197960779914\\
6.3038 0.988346949171937\\
6.31310000000001 0.98380719049887\\
6.32249999999999 0.989294411425898\\
6.3319 0.981532174794822\\
6.3413 0.982905560447353\\
6.3506 0.99066226560572\\
6.36 0.98488740809784\\
6.3694 0.986789444499245\\
6.3788 0.985427912002519\\
6.38809999999999 0.988656711643624\\
6.39749999999999 0.982962455595214\\
6.40690000000001 0.989547278749724\\
6.41629999999999 0.986100381041819\\
6.4256 0.984123274653653\\
6.435 0.987283325991093\\
6.4444 0.985519576407406\\
6.4538 0.982853406561813\\
6.4631 0.988277410657885\\
6.47250000000001 0.985789828359745\\
6.4819 0.98572108005608\\
6.49130000000001 0.989490383601864\\
6.50060000000001 0.985312541286023\\
6.50999999999999 0.98574952763001\\
6.5194 0.982956924122505\\
6.5288 0.983111805358348\\
6.5381 0.984841575895397\\
6.5475 0.988162830151776\\
6.5569 0.985651541542028\\
6.5663 0.989294411425898\\
6.57559999999999 0.981721825287691\\
6.58499999999999 0.986381695939575\\
6.59440000000001 0.991345797590438\\
6.60379999999999 0.989093697987611\\
6.6131 0.987151360856471\\
6.6225 0.982704056798679\\
6.6319 0.985755059102719\\
6.6413 0.985680779326345\\
6.6506 0.982572091664057\\
6.66000000000001 0.988501830407781\\
6.6694 0.98585857666341\\
6.67880000000001 0.98566892617054\\
6.68810000000001 0.986990948147919\\
6.69749999999999 0.982422741900922\\
6.7069 0.988685949427941\\
6.7163 0.984548407841836\\
6.7256 0.983468190242866\\
6.735 0.988352480644646\\
6.7444 0.985680779326345\\
6.7538 0.981423125761422\\
6.76309999999999 0.987191661586206\\
6.77250000000001 0.988599816495763\\
6.78190000000001 0.984922177354866\\
6.79129999999999 0.986726227668289\\
6.8006 0.983118127041444\\
6.81 0.989104760933029\\
6.8194 0.985122890793153\\
6.8288 0.987714780862372\\
6.8381 0.983577239276266\\
6.8475 0.986628241580306\\
6.8569 0.986232346176441\\
6.86630000000001 0.987231962315941\\
6.87560000000001 0.988041927962572\\
6.88499999999999 0.986053758628988\\
6.8944 0.982422741900922\\
6.90379999999999 0.983450805614353\\
6.9131 0.986358779838354\\
6.9225 0.984531023213323\\
6.9319 0.98761126330168\\
6.9413 0.990282964619981\\
6.95059999999999 0.985209023725332\\
6.96000000000001 0.986668542310041\\
6.96939999999999 0.988053781118376\\
6.97880000000001 0.984916645882158\\
6.9881 0.987904431355241\\
6.9975 0.989518831175794\\
7.0069 0.98780091379455\\
7.0163 0.988542131137515\\
7.0256 0.990202363160511\\
7.035 0.982767273629635\\
7.0444 0.984651925402527\\
7.05380000000001 0.985525107880114\\
7.06310000000001 0.981514790166309\\
7.07249999999999 0.990576132673542\\
7.0819 0.985921793494366\\
7.09129999999999 0.989185362392498\\
7.1006 0.993270750093065\\
7.11 0.987771676010233\\
7.1194 0.985335457387245\\
7.1288 0.983450805614353\\
7.13809999999999 0.985445296631032\\
7.14750000000001 0.987967648186198\\
7.15689999999999 0.988173893097194\\
7.16630000000001 0.987484829639767\\
7.1756 0.991478552935447\\
7.185 0.983973924890518\\
7.1944 0.987283325991093\\
7.2038 0.99260460273686\\
7.2131 0.987674480132637\\
7.2225 0.988157298679067\\
7.2319 0.990242663890246\\
7.24130000000001 0.983928092688075\\
7.25060000000001 0.988749166258898\\
7.26000000000001 0.985404995901297\\
7.2694 0.987329948403924\\
7.27879999999999 0.98996134899249\\
7.2881 0.981526643322113\\
7.2975 0.987605731828971\\
7.30690000000001 0.992185001021386\\
7.3163 0.988001627232837\\
7.32559999999999 0.982491490204587\\
7.33500000000001 0.989133998717346\\
7.34439999999999 0.983537728756918\\
7.35380000000001 0.98680129765505\\
7.3631 0.983743973667914\\
7.3725 0.986536577175419\\
7.3819 0.991754336360495\\
7.3913 0.983732910722496\\
7.4006 0.986806829127758\\
7.41 0.989288879953189\\
7.4194 0.987507745740989\\
7.4288 0.986794975971954\\
7.43809999999999 0.98418649148461\\
7.44750000000001 0.987973179658906\\
7.4569 0.98585857666341\\
7.46629999999999 0.989237516278037\\
7.4756 0.985709226900275\\
7.485 0.987352864505145\\
7.4944 0.981957307983005\\
7.5038 0.985094443219223\\
7.51309999999999 0.985031226388266\\
7.52250000000001 0.989174299447081\\
7.53189999999999 0.989013096528142\\
7.54130000000001 0.982180937522513\\
7.5506 0.98682974522898\\
7.55999999999999 0.989093697987611\\
7.5694 0.981423125761422\\
7.5788 0.983416826567714\\
7.58810000000001 0.99215023176436\\
7.5975 0.982514406305809\\
7.6069 0.985737674474206\\
7.6163 0.986139891561167\\
7.62559999999999 0.987174276957693\\
7.63500000000001 0.986060080312084\\
7.6444 0.987266731572967\\
7.6538 0.986306625952815\\
7.6631 0.991926602224851\\
7.6725 0.986255262277662\\
7.6819 0.986944325735089\\
7.6913 0.991811231508355\\
7.70059999999999 0.987996095760128\\
7.71000000000001 0.98925490090655\\
7.71939999999999 0.987386843551785\\
7.72880000000001 0.986473360344463\\
7.7381 0.984629009301305\\
7.7475 0.985054142489488\\
7.7569 0.983709994621275\\
7.7663 0.986944325735089\\
7.77560000000001 0.985634947123901\\
7.785 0.986915878161158\\
7.79440000000001 0.992213448595316\\
7.8038 0.988150976995972\\
7.81309999999999 0.988731781630385\\
7.82250000000001 0.989875216060311\\
7.8319 0.988628264069694\\
7.8413 0.991616049542777\\
7.8506 0.988415697475602\\
7.86 0.989696628512859\\
7.8694 0.983376525837979\\
7.8788 0.989443761189033\\
7.88809999999999 0.989093697987611\\
7.89750000000001 0.986007926426545\\
7.90689999999999 0.987409759653006\\
7.91630000000001 0.990846384625881\\
7.9256 0.982491490204587\\
7.935 0.991650818799803\\
7.9444 0.992380182986965\\
7.9538 0.985732143001497\\
7.96310000000001 0.988289263813689\\
7.9725 0.991357650746242\\
7.98190000000001 0.992443399817921\\
7.9913 0.986301094480106\\
8.00059999999999 0.985427912002519\\
8.01000000000001 0.987260409889871\\
8.01939999999999 0.987070759397002\\
8.0288 0.989179830919789\\
8.0381 0.989380544358077\\
8.0475 0.986927731316963\\
8.0569 0.986685926938554\\
8.0663 0.986192045446706\\
8.07560000000001 0.984623477828597\\
8.08499999999999 0.991104783422416\\
8.09439999999999 0.987289647674189\\
8.10380000000001 0.984611624672792\\
8.1131 0.993966135233587\\
8.1225 0.987307032302702\\
8.1319 0.989455614344837\\
8.1413 0.98839278137438\\
8.1506 0.986778381553828\\
8.16 0.987484829639767\\
8.16940000000001 0.985312541286023\\
8.1788 0.987834892841189\\
8.18809999999999 0.981653076984026\\
8.19750000000001 0.986548430331224\\
8.20689999999999 0.989111082616124\\
8.2163 0.990460761957046\\
8.2256 0.988530277981711\\
8.235 0.987685543078054\\
8.2444 0.986944325735089\\
8.2538 0.990737335592481\\
8.26310000000001 0.987479298167059\\
8.27249999999999 0.98693879426238\\
8.28190000000001 0.985611240812293\\
8.29130000000001 0.983773211452231\\
8.3006 0.99017944705929\\
8.31 0.991047098064168\\
8.3194 0.985990541798032\\
8.3288 0.984387995133283\\
8.3381 0.985352842015758\\
8.3475 0.984278946099884\\
8.35690000000001 0.984175428539192\\
8.3663 0.98685819280291\\
8.37559999999999 0.984491512693975\\
8.38500000000001 0.987479298167059\\
8.39439999999999 0.984600561727375\\
8.4038 0.98790996282795\\
8.4131 0.985359163698853\\
8.4225 0.989736929242594\\
8.4319 0.990639349504499\\
8.4413 0.992144700291651\\
8.45060000000001 0.990254517046051\\
8.45999999999999 0.982692993853261\\
8.46940000000001 0.989708481668663\\
8.47880000000001 0.989806467756646\\
8.4881 0.989375012885368\\
8.4975 0.991759867833203\\
8.5069 0.982600539237987\\
8.5163 0.995413800662492\\
8.5256 0.992805316175147\\
8.535 0.99290330226313\\
8.54440000000001 0.98931732752712\\
8.55380000000001 0.987266731572967\\
8.56310000000001 0.986680395465845\\
8.57250000000001 0.986134360088458\\
8.5819 0.982267860665079\\
8.59129999999999 0.986105912514528\\
8.6006 0.992828232276369\\
8.61000000000001 0.98996134899249\\
8.61940000000001 0.983497428027183\\
8.6288 0.984485191010879\\
8.63810000000001 0.987416081336102\\
8.64750000000001 0.984209407585831\\
8.65689999999999 0.991185384881886\\
8.66629999999999 0.98758281572775\\
8.6756 0.989415313615103\\
8.68500000000002 0.987961326503102\\
8.69439999999999 0.988059312591085\\
8.7038 0.991173531726081\\
8.71310000000001 0.986841598384784\\
8.72249999999998 0.987697396233858\\
8.7319 0.989202747021011\\
8.7413 0.989943964363977\\
8.75060000000002 0.985031226388266\\
8.75999999999999 0.992162084920164\\
8.76939999999999 0.990019034350737\\
8.7788 0.993420099856199\\
8.78809999999999 0.98974325092569\\
8.7975 0.990421251437699\\
8.8069 0.988754697731607\\
8.81630000000001 0.989495915074572\\
8.82559999999999 0.98955360043282\\
8.83499999999999 0.990862979044007\\
8.84440000000001 0.990783167794925\\
8.85380000000001 0.982819427515174\\
8.86309999999999 0.987140297911054\\
8.8725 0.986852661330202\\
8.8819 0.987898109672146\\
8.89130000000002 0.98831771138762\\
8.9006 0.990995734389016\\
8.91 0.985795359832454\\
8.91940000000001 0.985985010325323\\
8.92879999999998 0.985623093968097\\
8.93810000000001 0.990776846111829\\
8.94750000000001 0.989536215804307\\
8.9569 0.987887046726728\\
8.96629999999999 0.98742161280881\\
8.9756 0.99009884559982\\
8.98500000000001 0.987105528654028\\
8.99439999999998 0.98931732752712\\
9.0038 0.986984626464824\\
9.01310000000001 0.990472615112851\\
9.02250000000001 0.991328412961925\\
9.03189999999999 0.990943580503477\\
9.04129999999999 0.986817892073176\\
9.0506 0.98739316523488\\
9.05999999999999 0.986163597872775\\
9.06939999999999 0.984674841503749\\
9.0788 0.98410589002514\\
9.08810000000001 0.988697012373359\\
9.0975 0.994621219644374\\
9.1069 0.988243431611246\\
9.1163 0.983951008789297\\
9.12559999999999 0.990604580247473\\
9.13499999999999 0.98585857666341\\
9.1444 0.990903279773742\\
9.1538 0.991857063710799\\
9.16309999999999 0.98996134899249\\
9.1725 0.989650796310416\\
9.1819 0.990690713179651\\
9.19130000000001 0.989007565055433\\
9.20059999999999 0.990783167794925\\
9.20999999999999 0.986203108392123\\
9.21940000000001 0.985359163698853\\
9.22880000000001 0.986100381041819\\
9.2381 0.994034883537252\\
9.2475 0.993609750349069\\
9.2569 0.98809961332082\\
9.26630000000002 0.989156914818568\\
9.2756 0.988605347968472\\
9.28500000000001 0.988461529678046\\
9.29440000000001 0.994563534286126\\
9.3038 0.990834531470077\\
9.31310000000001 0.992018266629738\\
9.32250000000001 0.987145829383763\\
9.33190000000002 0.985243792982358\\
9.34129999999999 0.986772059870732\\
9.3506 0.989369481412659\\
9.36000000000001 0.991770930778621\\
9.36939999999998 0.986576877905154\\
9.3788 0.992983903722599\\
9.38810000000001 0.991891832967825\\
9.39750000000001 0.98858243186725\\
9.40689999999999 0.984169106856096\\
9.41629999999999 0.989782761445038\\
9.42560000000002 0.984715142233484\\
9.43499999999999 0.990421251437699\\
9.4444 0.988030865017154\\
9.4538 0.990047481924668\\
9.46309999999998 0.983824575127383\\
9.4725 0.989518831175794\\
9.4819 0.992334350784521\\
9.49130000000001 0.988926963595963\\
9.50059999999999 0.988685949427941\\
9.50999999999999 0.986622710107597\\
9.5194 0.988823446035272\\
9.5288 0.984904792726353\\
9.53809999999999 0.990857447571299\\
9.5475 0.992661497884721\\
9.5569 0.986708843039776\\
9.56630000000001 0.98685819280291\\
9.57559999999999 0.993368736181047\\
9.58500000000001 0.993276281565773\\
9.59440000000001 0.986094059358723\\
9.60380000000001 0.991288902442577\\
9.6131 0.98712844475525\\
9.6225 0.990644880977207\\
9.63190000000002 0.99017944705929\\
9.64129999999999 0.988955411169894\\
9.6506 0.992512148121586\\
9.66000000000001 0.988565047238737\\
9.66940000000001 0.992207917122608\\
9.6788 0.989197215548302\\
9.68810000000001 0.991225685611621\\
9.69750000000001 0.990484468268655\\
9.70689999999999 0.989943964363977\\
9.71629999999999 0.989518831175794\\
9.7256 0.984267092944079\\
9.73500000000001 0.993293666194287\\
9.74439999999998 0.984054526349988\\
9.7538 0.985416058846714\\
9.76310000000001 0.985657863225123\\
9.77249999999999 0.991058951219973\\
9.78189999999999 0.990507384369877\\
9.79129999999999 0.991955049798782\\
9.80060000000002 0.994224534030122\\
9.80999999999999 0.989806467756646\\
9.8194 0.985094443219223\\
9.8288 0.992736567871482\\
9.83809999999998 0.987059696451585\\
9.8475 0.992058567359473\\
9.8569 0.997867403913992\\
9.86630000000001 0.989875216060311\\
9.87559999999999 0.990880363672521\\
9.88499999999999 0.994046736693056\\
9.8944 0.990599048774764\\
9.9038 0.991127699523638\\
9.9131 0.994443422307309\\
9.9225 0.9935291488896\\
9.9319 0.994190554983483\\
9.94130000000001 0.984571323943057\\
9.95059999999999 0.990604580247473\\
9.96000000000001 0.994276687915661\\
9.96940000000001 0.989886279005729\\
9.97880000000001 0.984623477828597\\
9.9881 0.997631921218679\\
9.9975 0.988478914306559\\
10.0069 0.990300349248494\\
10.0163 0.989564663378237\\
10.0256 0.983014609480753\\
10.035 0.990518447315294\\
10.0444 0.991765399305912\\
10.0538 0.983623861689096\\
10.0631 0.991937665170269\\
10.0725 0.989633411681903\\
10.0819 0.9947650379348\\
10.0913 0.986881108904132\\
10.1006 0.99230590321059\\
10.11 0.98761126330168\\
10.1194 0.993702204964343\\
10.1288 0.996195318735192\\
10.1381 0.987283325991093\\
10.1475 0.994661520374109\\
10.1569 0.986824213756271\\
10.1663 0.99036909755216\\
10.1756 0.990478146585559\\
10.185 0.991638965643999\\
10.1944 0.989547278749724\\
10.2038 0.99163343417129\\
10.2131 0.989731397769885\\
10.2225 0.990593517302055\\
10.2319 0.991363182218951\\
10.2413 0.991863385393895\\
10.2506 0.993070036654778\\
10.26 0.990949902186573\\
10.2694 0.992225301751121\\
10.2788 0.993500701315669\\
10.2881 0.987697396233858\\
10.2975 0.991340266117729\\
10.3069 0.98823710992815\\
10.3163 0.991386098320173\\
10.3256 0.983801659026162\\
10.335 0.991231217084329\\
10.3444 0.989404250669685\\
10.3538 0.990219747789025\\
10.3631 0.995367968460048\\
10.3725 0.990799762213051\\
10.3819 0.988455998205337\\
10.3913 0.989243047750746\\
10.4006 0.989265963851968\\
10.41 0.985990541798032\\
10.4194 0.991092930266612\\
10.4288 0.988231578455441\\
10.4381 0.985634947123901\\
10.4475 0.994821933082661\\
10.4569 0.98704784329578\\
10.4663 0.988220515510024\\
10.4756 0.994127338152526\\
10.485 0.991201979300012\\
10.4944 0.988180214780289\\
10.5038 0.99152438513789\\
10.5131 0.988760229204315\\
10.5225 0.98593917812288\\
10.5319 0.992403099088186\\
10.5413 0.992661497884721\\
10.5506 0.987007542566045\\
10.56 0.990650412449916\\
10.5694 0.993690351808539\\
10.5788 0.987622326247098\\
10.5881 0.990523978788003\\
10.5975 0.988697012373359\\
10.6069 0.990184978531998\\
10.6163 0.987887046726728\\
10.6256 0.991770930778621\\
10.635 0.987042311823071\\
10.6444 0.992891449107326\\
10.6538 0.994914387697935\\
10.6631 0.989237516278037\\
10.6725 0.994598303543152\\
10.6819 0.992098868089208\\
10.6913 0.993443015957421\\
10.7006 0.988358012117355\\
10.71 0.992598281053765\\
10.7194 0.991650818799803\\
10.7288 0.989288879953189\\
10.7381 0.991029713435655\\
10.7475 0.986812360600467\\
10.7569 0.989938432891268\\
10.7663 0.994322520118104\\
10.7756 0.98801901186135\\
10.785 0.993443015957421\\
10.7944 0.993880002301408\\
10.8038 0.990432314383116\\
10.8131 0.992144700291651\\
10.8225 0.993081099600195\\
10.8319 0.990082251181694\\
10.8413 0.985870429819214\\
10.8506 0.993546533518113\\
10.86 0.99225928079776\\
10.8694 0.985910730548949\\
10.8788 0.988260816239759\\
10.8881 0.994489254509752\\
10.8975 0.992000882001225\\
10.9069 0.990610901930568\\
10.9163 0.992828232276369\\
10.9256 0.991271517814064\\
10.935 0.987490361112476\\
10.9444 0.988260816239759\\
10.9538 0.988496298935072\\
10.9631 0.99225928079776\\
10.9725 0.986708843039776\\
10.9819 0.98572108005608\\
10.9913 0.993558386673917\\
11.0006 0.992385714459673\\
11.01 0.989593901162555\\
11.0194 0.996075206756374\\
11.0288 0.989357628256855\\
11.0381 0.989151383345859\\
11.0475 0.992759483972704\\
11.0569 0.993368736181047\\
11.0663 0.989736929242594\\
11.0756 0.98869148090065\\
11.085 0.99360421887636\\
11.0944 0.992506616648878\\
11.1038 0.98772031233508\\
11.1131 0.98985229995909\\
11.1225 0.990570601200833\\
11.1319 0.986628241580306\\
11.1413 0.99279425322973\\
11.1506 0.991587601968847\\
11.16 0.988375396745867\\
11.1694 0.995822339432548\\
11.1788 0.994965751373087\\
11.1881 0.99074286706519\\
11.1975 0.990817146841564\\
11.2069 0.987363927450563\\
11.2163 0.992144700291651\\
11.2256 0.985048611016779\\
11.235 0.990587195618959\\
11.2444 0.990518447315294\\
11.2538 0.991800168562938\\
11.2631 0.987432675754228\\
11.2725 0.988536599664807\\
11.2819 0.990191300215094\\
11.2913 0.988806061406759\\
11.3006 0.98901862800085\\
11.31 0.992512148121586\\
11.3194 0.989145061662763\\
11.3288 0.994908856225226\\
11.3381 0.994299604016882\\
11.3475 0.987996095760128\\
11.3569 0.990530300471099\\
11.3663 0.989541747277016\\
11.3756 0.99117906319879\\
11.385 0.990869300727103\\
11.3944 0.987088144025515\\
11.4038 0.991104783422416\\
11.4131 0.989582048006751\\
11.4225 0.992460784446434\\
11.4319 0.989892600688824\\
11.4413 0.992478169074947\\
11.4506 0.990753930010607\\
11.46 0.990328796822425\\
11.4694 0.991581280285751\\
11.4788 0.989789083128133\\
11.4881 0.98758281572775\\
11.4975 0.989921048262755\\
11.5069 0.990570601200833\\
11.5163 0.994241918658635\\
11.5256 0.989996118249516\\
11.535 0.993851554727478\\
11.5444 0.991036035118751\\
11.5538 0.99096096513199\\
11.5631 0.992000882001225\\
11.5725 0.988553194082933\\
11.5819 0.993063714971682\\
11.5913 0.985479275677671\\
11.6006 0.992724714715677\\
11.61 0.990489999741364\\
11.6194 0.989570194850946\\
11.6288 0.987927347456463\\
11.6381 0.988892194338937\\
11.6475 0.985628625440806\\
11.6569 0.989145061662763\\
11.6663 0.995638220412387\\
11.6756 0.989840446803285\\
11.685 0.988869278237716\\
11.6944 0.989789083128133\\
11.7038 0.990886685355616\\
11.7131 0.993173554215469\\
11.7225 0.993828638626256\\
11.7319 0.994247450131343\\
11.7413 0.987593878673167\\
11.7506 0.993759100112204\\
11.76 0.986485213500267\\
11.7694 0.9935291488896\\
11.7788 0.994862233812396\\
11.7881 0.994908856225226\\
11.7975 0.991242280029747\\
11.8069 0.993586834247847\\
11.8163 0.99495468842767\\
11.8256 0.989122145561542\\
11.835 0.993202001789399\\
11.8444 0.997729907306662\\
11.8538 0.990460761957046\\
11.8631 0.991391629792882\\
11.8725 0.985152128577471\\
11.8819 0.991541769766403\\
11.8913 0.991489615880864\\
11.9006 0.987685543078054\\
11.91 0.990558748045029\\
11.9194 0.993420099856199\\
11.9288 0.988191277725706\\
11.9381 0.988731781630385\\
11.9475 0.992064098832182\\
11.9569 0.989599432635264\\
11.9663 0.987771676010233\\
11.9756 0.992437868345212\\
11.985 0.987427144281519\\
11.9944 0.994178701827678\\
12.0038 0.992299581527495\\
12.0131 0.988651180170915\\
12.0225 0.991454846623838\\
12.0319 0.991156147097568\\
12.0413 0.993725121065565\\
12.0506 0.989145061662763\\
12.06 0.989800146073551\\
12.0694 0.989840446803285\\
12.0788 0.988048249645667\\
12.0881 0.988726250157676\\
12.0975 0.990202363160511\\
12.1069 0.98799056428742\\
12.1163 0.990030097296155\\
12.1256 0.988024543334059\\
12.135 0.991541769766403\\
12.1444 0.988708865529163\\
12.1538 0.994638604272887\\
12.1631 0.99185153223809\\
12.1725 0.995195702595691\\
12.1819 0.994638604272887\\
12.1913 0.989921048262755\\
12.2006 0.987834892841189\\
12.21 0.993345820079826\\
12.2194 0.991748014677399\\
12.2288 0.993363204708339\\
12.2381 0.995506255277765\\
12.2475 0.994178701827678\\
12.2569 0.992345413729938\\
12.2663 0.992736567871482\\
12.2756 0.988944348224476\\
12.285 0.988111466476624\\
12.2944 0.995971689195683\\
12.3038 0.988363543590063\\
12.3131 0.991857063710799\\
12.3225 0.993276281565773\\
12.3319 0.993748037166787\\
12.3413 0.987657095504124\\
12.3506 0.993196470316691\\
12.36 0.992156553447456\\
12.3694 0.99439126842177\\
12.3788 0.992471847391851\\
12.3881 0.994828254765756\\
12.3975 0.993914771558435\\
12.4069 0.991673734901025\\
12.4163 0.990862979044007\\
12.4256 0.989748782398398\\
12.435 0.987646032558706\\
12.4444 0.995701437243344\\
12.4538 0.988719928474581\\
12.4631 0.987709249389663\\
12.4725 0.991575748813042\\
12.4819 0.988168361624485\\
12.4913 0.989806467756646\\
12.5006 0.991454846623838\\
12.51 0.991621581015486\\
12.5194 0.991909217596338\\
12.5288 0.993552064990821\\
12.5381 0.992173147865582\\
12.5475 0.996614920450666\\
12.5569 0.991454846623838\\
12.5663 0.989036012629363\\
12.5756 0.990840062942786\\
12.585 0.991426399049908\\
12.5944 0.991345797590438\\
12.6038 0.987921815983754\\
12.6131 0.995643751885096\\
12.6225 0.991375035374755\\
12.6319 0.990294817775785\\
12.6413 0.992455252973725\\
12.6506 0.995000520630113\\
12.66 0.990282964619981\\
12.6694 0.99174248320469\\
12.6788 0.992380182986965\\
12.6881 0.991575748813042\\
12.6975 0.995367968460048\\
12.7069 0.991949518326073\\
12.7163 0.985651541542028\\
12.7256 0.989271495324676\\
12.735 0.992087014933404\\
12.7444 0.988312179914911\\
12.7538 0.994908856225226\\
12.7631 0.993667435707317\\
12.7725 0.989898132161533\\
12.7819 0.989616817263776\\
12.7913 0.988168361624485\\
12.8006 0.991765399305912\\
12.81 0.993202001789399\\
12.8194 0.991891832967825\\
12.8288 0.9949831360016\\
12.8381 0.993437484484713\\
12.8475 0.989564663378237\\
12.8569 0.989811999229355\\
12.8663 0.986950647418184\\
12.8756 0.991484084408156\\
12.885 0.992696267141747\\
12.8944 0.989484061918768\\
12.9038 0.988024543334059\\
12.9131 0.990834531470077\\
12.9225 0.991770930778621\\
12.9319 0.998585705155736\\
12.9413 0.986037164210863\\
12.9506 0.991299965387994\\
12.96 0.987088144025515\\
12.9694 0.991013119017529\\
12.9788 0.990788699267633\\
12.9881 0.991064482692681\\
12.9975 0.992471847391851\\
13.0069 0.985582793238362\\
13.0163 0.990317733877007\\
13.0256 0.993552064990821\\
13.035 0.988168361624485\\
13.0444 0.994799016981439\\
13.0538 0.991133230996347\\
13.0631 0.995126954292026\\
13.0725 0.991162468780664\\
13.0819 0.991352119273534\\
13.0913 0.99304079887046\\
13.1006 0.994477401353948\\
13.11 0.994305135489591\\
13.1194 0.988001627232837\\
13.1288 0.987151360856471\\
13.1381 0.99368482033583\\
13.1475 0.99722417265901\\
13.1569 0.989288879953189\\
13.1663 0.994569855969222\\
13.1756 0.992787931546634\\
13.185 0.991535448083308\\
13.1944 0.991139552679442\\
13.2038 0.991547301239112\\
13.2131 0.990966496604699\\
13.2225 0.992455252973725\\
13.2319 0.992023798102447\\
13.2413 0.993897386929921\\
13.2506 0.988421228948311\\
13.26 0.989260432379259\\
13.2694 0.988863746765007\\
13.2788 0.992408630560895\\
13.2881 0.990966496604699\\
13.2975 0.988892194338937\\
13.3069 0.995230471852718\\
13.3163 0.996908088504227\\
13.3256 0.994719205732357\\
13.335 0.989524362648503\\
13.3444 0.991047098064168\\
13.3538 0.990053013397377\\
13.3631 0.99104156659146\\
13.3725 0.988340627488841\\
13.3819 0.991058951219973\\
13.3913 0.991322881489216\\
13.4006 0.998844103952271\\
13.41 0.988806061406759\\
13.4194 0.991288902442577\\
13.4288 0.988817914562563\\
13.4381 0.995092185035\\
13.4475 0.991943196642977\\
13.4569 0.997132508254123\\
13.4663 0.997006074592209\\
13.4756 0.98907078188639\\
13.485 0.991254133185551\\
13.4944 0.993288134721578\\
13.5038 0.992541385905903\\
13.5131 0.989582048006751\\
13.5225 0.991955049798782\\
13.5319 0.990184978531998\\
13.5413 0.994041205220347\\
13.5506 0.993500701315669\\
13.56 0.992132847135847\\
13.5694 0.993270750093065\\
13.5788 0.995460423075322\\
13.5881 0.990518447315294\\
13.5975 0.996574619720931\\
13.6069 0.992357266885743\\
13.6163 0.991765399305912\\
13.6256 0.993207533262108\\
13.635 0.992357266885743\\
13.6444 0.991754336360495\\
13.6538 0.993512554471474\\
13.6631 0.992392036142769\\
13.6725 0.991512531982086\\
13.6819 0.992920686891643\\
13.6913 0.988346949171937\\
13.7006 0.994098100368208\\
13.71 0.99357577130243\\
13.7194 0.992098868089208\\
13.7288 0.990173915586581\\
13.7381 0.985698163954858\\
13.7475 0.992156553447456\\
13.7569 0.991006797334433\\
13.7663 0.988955411169894\\
13.7756 0.992282987109369\\
13.785 0.989334712155633\\
13.7944 0.98604822715628\\
13.8038 0.990254517046051\\
13.8131 0.999769440315397\\
13.8225 0.993655582551513\\
13.8319 0.994552471340708\\
13.8413 0.995678521142122\\
13.8506 0.993465932058643\\
13.86 0.991081867321195\\
13.8694 0.990806083896146\\
13.8788 0.992891449107326\\
13.8881 0.993311050822799\\
13.8975 0.995385353088561\\
13.9069 0.992426015189408\\
13.9163 0.99201273515703\\
13.9256 0.988955411169894\\
13.935 0.995701437243344\\
13.9444 0.990570601200833\\
13.9538 0.99141533610449\\
13.9631 0.997178340456566\\
13.9725 0.994753184778996\\
13.9819 0.990460761957046\\
13.9913 0.997723585623566\\
};
\addlegendentry{Suspended \, \,};
\addplot [
color=green!50!black,
line width=0.2pt,
mark size=1.25pt,
only marks,
mark=triangle,
mark options={solid}
]
table[row sep=crcr]{
-0.770599999999988 1.00208254826818\\
-0.764299999999992 1.00520334860245\\
-0.758099999999985 0.997169112143552\\
-0.751799999999989 1.00426066943683\\
-0.745599999999982 1.0021323437068\\
-0.739299999999986 1.00164555036718\\
-0.733099999999979 0.995140377290766\\
-0.726799999999983 1.00017229221763\\
-0.720600000000005 0.999741304110943\\
-0.714299999999994 0.996582727926687\\
-0.708100000000002 0.996926144744763\\
-0.701799999999992 0.99866726801241\\
-0.695599999999999 0.993972760109307\\
-0.689299999999989 1.00163868203081\\
-0.683099999999996 0.999579039664402\\
-0.6768 0.994378850496682\\
-0.670599999999993 1.00597775352721\\
-0.664299999999997 1.0022388029204\\
-0.65809999999999 1.00043414754141\\
-0.651799999999994 1.0047921069628\\
-0.645599999999988 1.0021761293511\\
-0.639299999999992 1.00109007366394\\
-0.633099999999985 1.00768796928123\\
-0.626799999999989 0.995951699523472\\
-0.620599999999996 0.996781909681171\\
-0.6143 1.00414219063459\\
-0.608099999999993 1.0030252274338\\
-0.601799999999997 0.999991139846094\\
-0.59559999999999 1.00402371183235\\
-0.589299999999994 0.993798476074133\\
-0.583099999999988 0.993073866587992\\
-0.576799999999992 0.99963484489734\\
-0.570599999999985 1.00827521204014\\
-0.564299999999989 0.997693681333164\\
-0.558099999999982 1.0024628823942\\
-0.551799999999986 0.992218758710982\\
-0.545599999999993 0.992799133133531\\
-0.539299999999983 0.999397887292867\\
-0.53309999999999 1.00834389540375\\
-0.52679999999998 0.996663430878935\\
-0.520600000000002 1.00083336959243\\
-0.514299999999992 0.997163102349236\\
-0.508099999999999 1.00106517594463\\
-0.501800000000003 0.998424300613621\\
-0.495599999999996 1.0025882295328\\
-0.4893 1.00509774793089\\
-0.483099999999993 1.0037300904529\\
-0.476799999999997 1.00039637169142\\
-0.47059999999999 1.0033557661212\\
-0.464299999999994 0.996045280606397\\
-0.458099999999988 0.994172800405837\\
-0.451799999999992 0.999079368194101\\
-0.445599999999985 0.99740692829007\\
-0.439299999999989 0.997849935985388\\
-0.433099999999996 1.00697022813145\\
-0.426799999999986 1.00207567993182\\
-0.420599999999993 1.00022809745057\\
-0.414299999999997 0.993698026654846\\
-0.40809999999999 1.00081534020948\\
-0.401799999999994 0.999703528260955\\
-0.395599999999988 1.00758837840399\\
-0.389299999999992 0.998498993771552\\
-0.383099999999985 0.998998665241853\\
-0.376799999999989 0.997905741218326\\
-0.370599999999996 1.00301835909744\\
-0.364299999999986 0.999779079960932\\
-0.358099999999993 1.00122057205481\\
-0.351799999999983 0.996151739820001\\
-0.34559999999999 0.992543287604064\\
-0.33929999999998 1.00629627262597\\
-0.333100000000002 0.999647723028018\\
-0.326799999999992 0.995165275010077\\
-0.320599999999999 0.997375162234398\\
-0.314300000000003 1.00082135000379\\
-0.308099999999996 0.997456723728691\\
-0.3018 1.00441091429474\\
-0.295599999999993 1.00001002777109\\
-0.289299999999997 0.999566161533724\\
-0.28309999999999 1.00309391079741\\
-0.276799999999994 1.00388634510512\\
-0.270599999999988 1.00791290729707\\
-0.264299999999992 1.00364938750065\\
-0.258099999999985 0.994491319504602\\
-0.251799999999989 1.00491659555935\\
-0.245599999999982 1.0045603006106\\
-0.239299999999986 1.00583437700566\\
-0.233099999999993 1.02158433082469\\
-0.226799999999997 1.11159302030043\\
-0.22059999999999 1.21231974866834\\
-0.214299999999994 1.15326321700581\\
-0.208099999999988 0.911212741121104\\
-0.201799999999992 0.711288090205158\\
-0.195599999999999 0.709714897761551\\
-0.189299999999989 0.743556993953082\\
-0.183099999999996 0.740760207386669\\
-0.176799999999986 0.722930864734194\\
-0.170599999999993 0.7409849736941\\
-0.164299999999983 0.747827038669043\\
-0.15809999999999 0.73577224981252\\
-0.151799999999994 0.71381641067404\\
-0.145599999999988 0.719491116030137\\
-0.139299999999992 0.738712584608889\\
-0.133099999999999 0.736870926067751\\
-0.126800000000003 0.731377373083189\\
-0.120599999999996 0.726738927975637\\
-0.1143 0.729048749494018\\
-0.108099999999993 0.73366839253078\\
-0.101799999999997 0.735029353380817\\
-0.0955999999999904 0.736483895313779\\
-0.0892999999999944 0.735790966029106\\
-0.0830999999999875 0.742707981724593\\
-0.0767999999999915 0.737651340786829\\
-0.0705999999999847 0.724603905617657\\
-0.0642999999999887 0.727793990294972\\
-0.0580999999999818 0.734161624935743\\
-0.0517999999999859 0.748207802066085\\
-0.0455999999999932 0.743588159029323\\
-0.0392999999999972 0.739405513893563\\
-0.0330999999999904 0.743775492903583\\
-0.0267999999999944 0.748482535520546\\
-0.0206000000000017 0.742495664276818\\
-0.0142999999999915 0.742414532053547\\
-0.00809999999999889 0.741453136671343\\
-0.0017999999999887 0.738569036378934\\
0.00440000000000396 0.748557400386887\\
0.0107000000000141 0.750910921695368\\
0.0169000000000068 0.744412187684297\\
0.023200000000017 0.756136094436603\\
0.0294000000000096 0.748931982281203\\
0.0357000000000056 0.747901903535384\\
0.0419000000000125 0.752964811830078\\
0.0482000000000085 0.75981314416195\\
0.0544000000000011 0.749537511985676\\
0.0606999999999971 0.756335877170519\\
0.066900000000004 0.75558679923609\\
0.0731999999999999 0.750698690101797\\
0.0794000000000068 0.756298444737349\\
0.0857000000000028 0.746303813372465\\
0.0919000000000096 0.759538410707489\\
0.0982000000000056 0.75544943250886\\
0.104400000000012 0.756822842218551\\
0.110700000000008 0.75758445486684\\
0.116900000000015 0.756729175281421\\
0.123200000000011 0.754431888476899\\
0.129400000000018 0.764401536268268\\
0.135700000000014 0.754057306582583\\
0.141900000000021 0.754144706162783\\
0.148200000000017 0.758583282682215\\
0.154399999999995 0.76065588903351\\
0.160700000000006 0.760393690292908\\
0.166899999999998 0.766842457156949\\
0.173200000000008 0.76177336735953\\
0.179400000000001 0.770126208771394\\
0.185700000000011 0.759451011127289\\
0.191900000000004 0.764382820051682\\
0.198200000000014 0.767685287882713\\
0.204400000000007 0.76180453243577\\
0.210700000000003 0.764089456234841\\
0.21690000000001 0.770325905651106\\
0.223200000000006 0.770388407511996\\
0.229400000000012 0.772467195366016\\
0.235700000000008 0.770506972168437\\
0.241900000000015 0.771855398304613\\
0.248200000000011 0.773778189069022\\
0.254400000000004 0.762154130756572\\
0.2607 0.774620933940581\\
0.266900000000007 0.769564293002817\\
0.273200000000003 0.778216765588454\\
0.27940000000001 0.771087518299395\\
0.285700000000006 0.767304438631466\\
0.291900000000012 0.766436710186392\\
0.298200000000008 0.773328742308365\\
0.304400000000015 0.778878529796887\\
0.310700000000011 0.785514630535188\\
0.316900000000018 0.78063270290362\\
0.323200000000014 0.775694712476501\\
0.329400000000007 0.767959935482969\\
0.335700000000017 0.775120347848269\\
0.34190000000001 0.775981894790618\\
0.348200000000006 0.778223032945384\\
0.354399999999998 0.777711170178041\\
0.360700000000008 0.773241342728164\\
0.366900000000001 0.777711170178041\\
0.373200000000011 0.786251259609962\\
0.379400000000004 0.782967507995516\\
0.3857 0.795877576437458\\
0.391900000000007 0.783229706736118\\
0.398200000000003 0.784515716865609\\
0.40440000000001 0.778316699882514\\
0.410700000000006 0.77953402664839\\
0.416900000000012 0.78282395976556\\
0.423200000000008 0.788211482807541\\
0.429400000000015 0.78887316116177\\
0.435700000000011 0.783061174932647\\
0.441900000000004 0.784815433893585\\
0.4482 0.792288012146515\\
0.454400000000007 0.786806822167405\\
0.460700000000003 0.787206387635236\\
0.46690000000001 0.784141220825496\\
0.473200000000006 0.789391377140247\\
0.479400000000012 0.786426058770363\\
0.485700000000008 0.791270382260351\\
0.491900000000015 0.788935663022659\\
0.498200000000011 0.790733535919493\\
0.504400000000004 0.789928223481104\\
0.510700000000014 0.786788105950819\\
0.516900000000007 0.793605187352247\\
0.523200000000017 0.794398050930981\\
0.52940000000001 0.794497899370836\\
0.53570000000002 0.788498665121657\\
0.541899999999998 0.796370722988216\\
0.548200000000008 0.799023961324673\\
0.554400000000001 0.80054718662125\\
0.560699999999997 0.796027392024344\\
0.566900000000004 0.792500243740087\\
0.5732 0.800771867074477\\
0.579400000000007 0.795016115349313\\
0.585700000000003 0.794604015167622\\
0.59190000000001 0.794004752820079\\
0.598200000000006 0.794341816427021\\
0.604400000000012 0.788529830197898\\
0.610700000000008 0.799292341567999\\
0.616900000000015 0.803768436374806\\
0.623200000000011 0.802769608559431\\
0.629400000000004 0.794941164628768\\
0.635700000000014 0.799504659015775\\
0.641900000000007 0.802938140362902\\
0.648200000000003 0.803269022467118\\
0.65440000000001 0.80366232057802\\
0.660700000000006 0.805316645244898\\
0.666900000000012 0.802226494861643\\
0.673200000000008 0.798330946185795\\
0.679400000000001 0.807158217931832\\
0.685700000000011 0.805460193474854\\
0.691900000000004 0.804829680196866\\
0.698200000000014 0.799872973553162\\
0.704400000000007 0.804979495783752\\
0.710700000000017 0.802132827924513\\
0.71690000000001 0.806071990536257\\
0.72320000000002 0.800622051487591\\
0.729399999999998 0.805017014071126\\
0.735700000000008 0.803287738683703\\
0.741900000000001 0.801458614856425\\
0.748199999999997 0.811028783034165\\
0.754400000000004 0.81408150098425\\
0.7607 0.810267170385877\\
0.766900000000007 0.808125880670966\\
0.773200000000003 0.803106672166373\\
0.77940000000001 0.811422081145067\\
0.785700000000006 0.814930513212739\\
0.791900000000012 0.813900434466919\\
0.798200000000008 0.809955004498245\\
0.804400000000015 0.816971868633589\\
0.810700000000011 0.812333509380242\\
0.816900000000018 0.808369363194983\\
0.823200000000014 0.804548765239679\\
0.829400000000007 0.806646355164489\\
0.835700000000003 0.811915227695825\\
0.84190000000001 0.812552008330743\\
0.848200000000006 0.810454418405933\\
0.854399999999998 0.812239842443111\\
0.860700000000008 0.812764239924314\\
0.866900000000001 0.808388079411568\\
0.873200000000011 0.817065535570719\\
0.879400000000004 0.80895617668287\\
0.885700000000014 0.804093051122092\\
0.891900000000007 0.814493515311737\\
0.898200000000017 0.817134218934334\\
0.90440000000001 0.813157538035216\\
0.910700000000006 0.816840769263288\\
0.916900000000012 0.816559854306102\\
0.923200000000008 0.814730730478823\\
0.929400000000001 0.81318878896566\\
0.935699999999997 0.809542990170758\\
0.941900000000004 0.820243085534175\\
0.9482 0.815517326700627\\
0.954400000000007 0.81719663494102\\
0.960700000000003 0.81933792465593\\
0.96690000000001 0.823582985798375\\
0.973200000000006 0.81412520077435\\
0.979400000000012 0.813444720349332\\
0.985700000000008 0.8162539557754\\
0.991900000000015 0.808369363194983\\
0.998200000000011 0.810672917356434\\
1.00440000000002 0.821447963440395\\
1.01070000000001 0.815735825651128\\
1.01690000000001 0.820829899022062\\
1.02320000000002 0.815511059343697\\
1.0294 0.814749446695408\\
1.03570000000001 0.818932091831168\\
1.0419 0.822109641794624\\
1.04820000000001 0.817677332632122\\
1.0544 0.823245836337229\\
1.06070000000001 0.815660874930583\\
1.0669 0.823083571890688\\
1.07320000000001 0.822315691885469\\
1.07940000000001 0.822065942004523\\
1.0857 0.823308252343914\\
1.09190000000001 0.821841261551297\\
1.09820000000001 0.828514708868564\\
1.10440000000001 0.826810502908861\\
1.11070000000001 0.832441422620653\\
1.1169 0.824925144577622\\
1.12320000000001 0.834882343509335\\
1.1294 0.824313347516219\\
1.1357 0.820661367218591\\
1.14190000000001 0.817396417674936\\
1.1482 0.81938162444603\\
1.15440000000001 0.821560346594111\\
1.16070000000001 0.824644229620436\\
1.16690000000001 0.82659200395836\\
1.17320000000001 0.828614643162624\\
1.17940000000002 0.824350865803594\\
1.18570000000001 0.828982957700011\\
1.19190000000002 0.832354023040453\\
1.19820000000001 0.821529095663666\\
1.20440000000002 0.825355875121694\\
1.21070000000002 0.819387805948755\\
1.2169 0.820673816078247\\
1.22320000000001 0.824388298236764\\
1.2294 0.825880272602897\\
1.23570000000001 0.827709396430175\\
1.2419 0.831854609132765\\
1.24820000000001 0.824051148775618\\
1.2544 0.823882616972147\\
1.2607 0.821522828306736\\
1.26690000000001 0.826348521434344\\
1.2732 0.830225268039402\\
1.27940000000001 0.829988052872316\\
1.28570000000001 0.825661773652396\\
1.29190000000001 0.814949229429324\\
1.29820000000001 0.82988193707553\\
1.30440000000002 0.826648152608115\\
1.31070000000001 0.832747321151354\\
1.3169 0.820998430825533\\
1.3232 0.814911796996154\\
1.32940000000001 0.825493241848925\\
1.3357 0.840138767181015\\
1.34190000000001 0.842236357105825\\
1.34820000000001 0.84759262921736\\
1.35440000000001 0.849090870940423\\
1.36070000000001 0.821760043473822\\
1.36690000000002 0.820474033344331\\
1.37320000000001 0.835668939731139\\
1.37940000000002 0.845051859888823\\
1.38570000000001 0.829900653292116\\
1.39190000000001 0.822827554652812\\
1.39820000000002 0.839708036636942\\
1.40440000000001 0.838297194494081\\
1.41070000000002 0.831704793545879\\
1.4169 0.825231043108323\\
1.42320000000001 0.82975083770523\\
1.4294 0.836386852589327\\
1.4357 0.837504330915347\\
1.4419 0.830949448254521\\
1.4482 0.830618566150304\\
1.45440000000001 0.836848834063844\\
1.4607 0.831242812071362\\
1.46690000000001 0.834314246238032\\
1.47320000000001 0.835431724564052\\
1.47940000000001 0.839633085916397\\
1.48570000000001 0.84229877311251\\
1.49190000000002 0.837841480376493\\
1.49820000000001 0.828127678114592\\
1.5044 0.838471993654482\\
1.51070000000001 0.847317981617104\\
1.51690000000001 0.836929966287114\\
1.5232 0.839970235377544\\
1.52940000000001 0.833103186829086\\
1.53570000000001 0.838403310290866\\
1.54190000000001 0.836605351539828\\
1.54820000000001 0.835281908977166\\
1.55440000000002 0.836080954058625\\
1.56070000000001 0.831486294595378\\
1.5669 0.840744296885488\\
1.57320000000001 0.840182466971115\\
1.57940000000001 0.841793091847893\\
1.58570000000002 0.84118756214342\\
1.59190000000001 0.837916345242834\\
1.59820000000002 0.841406061093921\\
1.6044 0.846468969388615\\
1.61070000000001 0.840563316222362\\
1.6169 0.83383354854693\\
1.6232 0.843110352907829\\
1.6294 0.834320513594962\\
1.6357 0.845751056530426\\
1.64190000000001 0.840981512052574\\
1.6482 0.84340371672467\\
1.65440000000001 0.845975736983653\\
1.66070000000001 0.846668752122531\\
1.66690000000001 0.842330024042955\\
1.67320000000001 0.847817395524791\\
1.67940000000002 0.839352170959211\\
1.68570000000001 0.843428700298186\\
1.69190000000002 0.84434639589029\\
1.69820000000001 0.847224314679974\\
1.70440000000001 0.84132492887065\\
1.7107 0.839526970119612\\
1.71690000000001 0.85279908574201\\
1.72320000000001 0.844908225804663\\
1.72940000000001 0.849964866742427\\
1.73570000000001 0.848067059551533\\
1.7419 0.847399113840374\\
1.74820000000001 0.849440469261224\\
1.7544 0.850139665902828\\
1.76070000000001 0.840582032438947\\
1.76690000000001 0.844134164296719\\
1.77320000000002 0.845607422446266\\
1.77940000000001 0.844171596729889\\
1.78570000000001 0.852156123604367\\
1.79190000000001 0.850395683140703\\
1.79820000000001 0.849871285659501\\
1.8044 0.836162172136101\\
1.8107 0.851893924863766\\
1.8169 0.844184131443749\\
1.8232 0.847873544174547\\
1.82940000000001 0.850439297076599\\
1.8357 0.848416657872334\\
1.84190000000001 0.850383148426844\\
1.84820000000001 0.847399113840374\\
1.85440000000001 0.849259488598098\\
1.86070000000001 0.855820638615855\\
1.86690000000002 0.849403036828054\\
1.87320000000001 0.856438617179983\\
1.87940000000002 0.854721876506419\\
1.88570000000001 0.849003471360222\\
1.89190000000001 0.85592048705571\\
1.8982 0.857287629408472\\
1.90440000000001 0.844939476735108\\
1.91070000000001 0.855314957351237\\
1.9169 0.855296241134652\\
1.92320000000001 0.857243929618372\\
1.9294 0.850283299986988\\
1.93570000000001 0.858867432625827\\
1.9419 0.856738248353754\\
1.94820000000001 0.860234231561771\\
1.95440000000001 0.856157702222796\\
1.96070000000002 0.855564621377979\\
1.96690000000001 0.857162797395101\\
1.97320000000001 0.859585173775607\\
1.97940000000001 0.857936858903045\\
1.98570000000001 0.857337596555502\\
1.9919 0.858255292147606\\
1.9982 0.860115752759534\\
2.0044 0.855427340504953\\
2.0107 0.862213170975935\\
2.01690000000001 0.85800554226666\\
2.0232 0.851144761075132\\
2.02940000000001 0.850901278551116\\
2.03570000000001 0.856713350634444\\
2.04190000000001 0.852861501748696\\
2.04820000000001 0.860471189166244\\
2.05440000000002 0.853342199439798\\
2.06070000000001 0.850857578761016\\
2.06690000000002 0.857899426469875\\
2.07320000000001 0.860090855040224\\
2.07940000000002 0.857587346436448\\
2.08570000000002 0.857268913191887\\
2.0919 0.857287629408472\\
2.09820000000001 0.861333165379615\\
2.1044 0.858261559504536\\
2.11070000000001 0.861020656075165\\
2.1169 0.857624778869618\\
2.12320000000001 0.868936413731824\\
2.1294 0.865740061697579\\
2.13570000000001 0.85806177677062\\
2.14190000000001 0.853623114396984\\
2.1482 0.864479721975239\\
2.15440000000001 0.868162867349107\\
2.16070000000001 0.861020656075165\\
2.16690000000001 0.86615216187927\\
2.17320000000001 0.859835009510757\\
2.17940000000002 0.861451644181851\\
2.18570000000001 0.866638955218894\\
2.1919 0.86626463088719\\
2.1982 0.858705168179286\\
2.20440000000001 0.868037520210509\\
2.2107 0.86671450691887\\
2.21690000000001 0.862026008810084\\
2.22320000000001 0.864928739464874\\
2.22940000000001 0.868949291862502\\
2.23570000000001 0.862113580098693\\
2.24190000000002 0.869230035111279\\
2.24820000000001 0.862930912125715\\
2.25440000000002 0.86140184874323\\
2.26070000000001 0.870603702383584\\
2.26690000000001 0.87177732935936\\
2.27320000000002 0.864048733868553\\
2.27940000000001 0.866682740863198\\
2.28570000000001 0.864716679579712\\
2.2919 0.869348513913515\\
2.29820000000001 0.868968179787496\\
2.3044 0.86864966068873\\
2.31070000000001 0.865440430523807\\
2.3169 0.86734467678004\\
2.3232 0.859766326147142\\
2.32940000000001 0.867531838945892\\
2.3357 0.871027822153908\\
2.34190000000001 0.867407350349339\\
2.34820000000001 0.875385781575296\\
2.35440000000001 0.872582641797749\\
2.36070000000001 0.867588502720875\\
2.36690000000002 0.865147667686397\\
2.37320000000001 0.868661680277363\\
2.3794 0.868630772763736\\
2.3857 0.866508456828025\\
2.39190000000001 0.872513958434133\\
2.3982 0.866895659290405\\
2.40440000000001 0.87013493842691\\
2.41070000000001 0.872507090097772\\
2.41690000000001 0.868386946822902\\
2.42320000000001 0.873163016220298\\
2.42940000000002 0.870422550012049\\
2.43570000000001 0.871533503418526\\
2.4419 0.86842472267289\\
2.44820000000001 0.875803891551304\\
2.45440000000001 0.875129077503784\\
2.46070000000002 0.874155490824538\\
2.46690000000001 0.872651325161364\\
2.47320000000002 0.877433404353076\\
2.4794 0.873363056516827\\
2.48570000000001 0.870128928632594\\
2.4919 0.876515622906767\\
2.4982 0.878013778775625\\
2.5044 0.871876920236602\\
2.5107 0.875872574914919\\
2.51690000000001 0.870316090798445\\
2.5232 0.877439414147392\\
2.52940000000001 0.877845504534768\\
2.53570000000001 0.883494711192123\\
2.54190000000001 0.877664352163233\\
2.54820000000001 0.877170690487248\\
2.55440000000002 0.877046201890695\\
2.56070000000001 0.885611017333518\\
2.5669 0.87384984985645\\
2.57320000000001 0.876765458641918\\
2.57940000000001 0.878437898545949\\
2.5857 0.881958779473276\\
2.59190000000001 0.880473501735097\\
2.59820000000001 0.881391283181405\\
2.60440000000001 0.870665517410838\\
2.61070000000001 0.872151653691063\\
2.6169 0.87999271818979\\
2.62320000000001 0.876989538115713\\
2.6294 0.883738537132957\\
2.63570000000001 0.877145792767937\\
2.64190000000001 0.881109681390583\\
2.64820000000002 0.879842473331881\\
2.65440000000001 0.87938744604793\\
2.66070000000002 0.877252251981541\\
2.6669 0.885923526637967\\
2.67320000000001 0.879074936743481\\
2.6794 0.877095997329316\\
2.6857 0.880954285280403\\
2.6919 0.883357344464892\\
2.6982 0.884805704895129\\
2.70440000000001 0.887652630316981\\
2.7107 0.873581126196306\\
2.71690000000001 0.88328265130696\\
2.72320000000001 0.882103014536868\\
2.72940000000001 0.889862517541302\\
2.73570000000001 0.894438546642168\\
2.74190000000002 0.889063214897229\\
2.74820000000001 0.886248055531049\\
2.75440000000002 0.886078922748146\\
2.76070000000001 0.886216289475377\\
2.76690000000001 0.885779291574375\\
2.7732 0.884468297871369\\
2.77940000000001 0.8821777076948\\
2.78570000000001 0.883950597018119\\
2.7919 0.888195228889542\\
2.79820000000001 0.894475463950111\\
2.8044 0.881034988232651\\
2.81070000000001 0.88051041904304\\
2.8169 0.887864690202143\\
2.82320000000001 0.885685710491449\\
2.82940000000001 0.886454105621895\\
2.83570000000002 0.889181693699466\\
2.84190000000001 0.889469305284605\\
2.84820000000001 0.893333603030008\\
2.85440000000001 0.892833931559707\\
2.86070000000001 0.888301688103145\\
2.8669 0.894356985147875\\
2.8732 0.892827921765391\\
2.8794 0.89172297815323\\
2.8857 0.891972813888381\\
2.89190000000001 0.888089628217983\\
2.8982 0.894144925262713\\
2.90440000000001 0.893339612824324\\
2.91070000000001 0.888700910154159\\
2.91690000000001 0.897104319692486\\
2.92320000000001 0.894375873072869\\
2.92940000000002 0.886447237285533\\
2.93570000000001 0.888395269186071\\
2.94190000000002 0.891460264287402\\
2.94820000000001 0.891916150113398\\
2.95440000000001 0.885392947654039\\
2.96070000000002 0.887808884969206\\
2.9669 0.890255729797999\\
2.97320000000001 0.892509402666625\\
2.9794 0.889194571830144\\
2.98570000000001 0.890149270584396\\
2.9919 0.881828281082407\\
2.99820000000001 0.900094621635885\\
3.0044 0.894507230005784\\
3.01070000000001 0.88823300473953\\
3.01690000000001 0.891747875872541\\
3.0232 0.892228659417847\\
3.02940000000001 0.887097153613743\\
3.03570000000001 0.897241686419716\\
3.04190000000001 0.891467132623763\\
3.04820000000001 0.896274109534786\\
3.0544 0.889906303185607\\
3.06070000000001 0.898552680122722\\
3.0669 0.898108813885359\\
3.0732 0.897285472064021\\
3.07940000000001 0.901505206216133\\
3.0857 0.894856656618176\\
3.09190000000001 0.889456427153927\\
3.09820000000001 0.89269055503816\\
3.10440000000001 0.891916150113398\\
3.11070000000001 0.891261082532918\\
3.11690000000002 0.900674996058434\\
3.12320000000001 0.89086787027622\\
3.12940000000002 0.901486318291139\\
3.13570000000001 0.900631210414129\\
3.14190000000002 0.899594950165584\\
3.14820000000002 0.894675504246641\\
3.1544 0.889331938557374\\
3.16070000000001 0.894650606527331\\
3.1669 0.899163962058898\\
3.17320000000001 0.900943719718578\\
3.1794 0.896667321791483\\
3.18570000000001 0.902091590432998\\
3.1919 0.890030791782159\\
3.1982 0.895425011452093\\
3.20440000000001 0.903827562448374\\
3.2107 0.908852609038875\\
3.21690000000001 0.898957911968052\\
3.22320000000001 0.895886907072405\\
3.22940000000001 0.899188859778209\\
3.23570000000001 0.900119519355195\\
3.24190000000002 0.897234818083354\\
3.24820000000001 0.904270570143692\\
3.2544 0.905232137234306\\
3.2607 0.902459904970385\\
3.26690000000001 0.903715093440454\\
3.2732 0.902222947365912\\
3.27940000000001 0.902516568745368\\
3.28570000000001 0.90275352634984\\
3.29190000000001 0.910213398180502\\
3.29820000000001 0.903459247910987\\
3.30440000000002 0.906118152624943\\
3.31070000000001 0.902984474159997\\
3.31690000000002 0.89678580059372\\
3.32320000000001 0.903396574341688\\
3.32940000000001 0.902254713421584\\
3.33570000000002 0.903165626531532\\
3.34190000000001 0.905850287506843\\
3.34820000000002 0.899770092742803\\
3.3544 0.904121183827829\\
3.36070000000001 0.897834080430898\\
3.3669 0.90221092777728\\
3.3732 0.901342941769592\\
3.3794 0.905281932672927\\
3.3857 0.913447526064736\\
3.39190000000001 0.910993812899581\\
3.3982 0.906792966672463\\
3.40440000000001 0.913815840602123\\
3.41070000000001 0.904976291704839\\
3.41690000000001 0.902659945266915\\
3.42320000000001 0.908134867889095\\
3.42940000000002 0.918092238529217\\
3.43570000000001 0.903564848582545\\
3.4419 0.906949221324687\\
3.44820000000001 0.89991346926435\\
3.45440000000001 0.912854273511509\\
3.4607 0.905512880483083\\
3.46690000000001 0.904420815001601\\
3.47320000000001 0.908971087841111\\
3.47940000000001 0.913485301914724\\
3.48570000000001 0.908440508857183\\
3.49190000000002 0.910825538658723\\
3.49820000000001 0.907266881881408\\
3.5044 0.908709232517328\\
3.51070000000001 0.909527423086395\\
3.51690000000001 0.913247485768206\\
3.52320000000002 0.906755190822474\\
3.52940000000001 0.918710388801755\\
3.53570000000002 0.916181982478668\\
3.5419 0.908665446873024\\
3.54820000000001 0.912479949179806\\
3.5544 0.917037090355678\\
3.5607 0.908915282608174\\
3.5669 0.906262387688535\\
3.5732 0.905450206913784\\
3.57940000000001 0.913160773021642\\
3.5857 0.91202406335381\\
3.59190000000001 0.907766553351709\\
3.59820000000001 0.907460912383621\\
3.60440000000001 0.910488131634963\\
3.61070000000001 0.911069364599557\\
3.61690000000002 0.912554642337737\\
3.62320000000001 0.900818372579981\\
3.62940000000002 0.912098756511741\\
3.63570000000001 0.907710748118771\\
3.64190000000001 0.913185670740952\\
3.6482 0.922275055373387\\
3.65440000000001 0.913416618551109\\
3.66070000000001 0.913178802404591\\
3.66690000000001 0.92264937970509\\
3.67320000000001 0.9132663736932\\
3.6794 0.920970071464697\\
3.68570000000001 0.9065809067873\\
3.6919 0.909833064054483\\
3.69820000000001 0.916569184941049\\
3.70440000000001 0.915470251123205\\
3.71070000000002 0.916687663743285\\
3.71690000000001 0.911680646535733\\
3.72320000000002 0.916157084759357\\
3.72940000000001 0.91713668123292\\
3.73570000000001 0.912760692428583\\
3.7419 0.917124661644287\\
3.7482 0.917111783513609\\
3.7544 0.914327531661056\\
3.7607 0.921225916994163\\
3.76690000000001 0.916861947778459\\
3.7732 0.916231777917289\\
3.77940000000001 0.917174457082908\\
3.78570000000001 0.914077695925906\\
3.79190000000001 0.914090574056583\\
3.79820000000001 0.916244656047967\\
3.80440000000002 0.918048452884913\\
3.81070000000001 0.910625498362194\\
3.81690000000002 0.914852100850668\\
3.82320000000001 0.916338237130893\\
3.82940000000001 0.915233293518732\\
3.8357 0.913004518369417\\
3.84190000000001 0.912117644436735\\
3.84820000000001 0.914714734123437\\
3.8544 0.91639404236383\\
3.86070000000001 0.913703371594202\\
3.8669 0.918616807718829\\
3.87320000000001 0.916113299115053\\
3.8794 0.917817505074756\\
3.88570000000001 0.922355758325634\\
3.89190000000001 0.918972244125538\\
3.89820000000002 0.920957193334019\\
3.90440000000001 0.923904568175158\\
3.91070000000001 0.929772703054037\\
3.91690000000001 0.917855280924745\\
3.92320000000001 0.919191172347061\\
3.9294 0.921407069365698\\
3.9357 0.921250814713474\\
3.9419 0.915071029072191\\
3.9482 0.927000470790116\\
3.95440000000001 0.926595238944786\\
3.9607 0.916899723628447\\
3.96690000000001 0.926557463094798\\
3.97320000000001 0.929242124070109\\
3.97940000000001 0.923467570274156\\
3.98570000000001 0.918179809817827\\
3.99190000000002 0.916113299115053\\
3.99820000000001 0.923355101266236\\
4.00440000000002 0.925777048375719\\
4.01070000000001 0.933799265245981\\
4.01690000000002 0.927118949592352\\
4.02320000000002 0.923161070764023\\
4.0294 0.920788919093161\\
4.03570000000001 0.923885680250164\\
4.0419 0.922636501574412\\
4.04820000000001 0.920745133448857\\
4.0544 0.924809471490789\\
4.06070000000001 0.922630491780095\\
4.0669 0.925164907897498\\
4.07320000000001 0.930814973096898\\
4.07940000000001 0.917305814015823\\
4.0857 0.926675941897034\\
4.09190000000001 0.924603421399944\\
4.09820000000001 0.92179427182808\\
4.10440000000001 0.924147535573947\\
4.11070000000001 0.919765536975294\\
4.11690000000002 0.926582360814108\\
4.12320000000001 0.922668267630084\\
4.1294 0.925876639252961\\
4.1357 0.924179301629619\\
4.14190000000001 0.915332884395974\\
4.1482 0.923386008779863\\
4.15440000000001 0.928736442805492\\
4.16070000000001 0.91744919053737\\
4.16690000000001 0.925103092870245\\
4.17320000000001 0.92751216184905\\
4.17940000000002 0.921338386002083\\
4.18570000000001 0.926132484782428\\
4.19190000000002 0.918660593363134\\
4.19820000000001 0.93349963407221\\
4.20440000000001 0.921862955191695\\
4.21070000000002 0.925521202846253\\
4.21690000000001 0.933948651561844\\
4.22320000000001 0.92327954956626\\
4.2294 0.929154552781499\\
4.23570000000001 0.923273539771943\\
4.2419 0.93014702738574\\
4.24820000000001 0.9283243426238\\
4.2544 0.923073499475414\\
4.2607 0.928411055370364\\
4.26690000000001 0.925258488980424\\
4.2732 0.93457967996506\\
4.27940000000001 0.93088966625483\\
4.28570000000001 0.925383836119022\\
4.29190000000001 0.924272882712545\\
4.29820000000001 0.924059964285338\\
4.30440000000002 0.93147690901374\\
4.31070000000001 0.923542263432088\\
4.3169 0.92627671984602\\
4.3232 0.938280854721877\\
4.32940000000001 0.929011176259953\\
4.3357 0.924441156953402\\
4.34190000000001 0.927194501292329\\
4.34820000000001 0.928342372006749\\
4.35440000000001 0.925221571672481\\
4.36070000000001 0.939604726555561\\
4.36690000000002 0.930946330029813\\
4.37320000000001 0.925933303027944\\
4.3794 0.931189297428601\\
4.38570000000001 0.92773709986489\\
4.39190000000001 0.934160711447007\\
4.39820000000002 0.934753964000233\\
4.40440000000001 0.932275353115768\\
4.41070000000002 0.931700988487535\\
4.4169 0.933568317435825\\
4.42320000000001 0.929279041378052\\
4.4294 0.928280556979495\\
4.4357 0.934442313237829\\
4.4419 0.936932943710927\\
4.4482 0.9304844344095\\
4.45440000000001 0.933643010593757\\
4.4607 0.931202175559279\\
4.46690000000001 0.933774367526671\\
4.47320000000001 0.932088190949916\\
4.47940000000001 0.937688460710695\\
4.48570000000001 0.93753220605847\\
4.49190000000002 0.931089706551359\\
4.49820000000001 0.934329844229909\\
4.5044 0.932063293230606\\
4.51070000000001 0.937363931817613\\
4.51690000000001 0.935940469106687\\
4.5232 0.929910069781267\\
4.52940000000001 0.92505930722594\\
4.53570000000001 0.936139650861171\\
4.54190000000001 0.930845880610525\\
4.54820000000001 0.933980417617516\\
4.5544 0.934167579783368\\
4.56070000000001 0.939099045290943\\
4.5669 0.940241764753092\\
4.57320000000001 0.928317474287438\\
4.57940000000001 0.936458169959937\\
4.58570000000002 0.93316222704845\\
4.59190000000001 0.938668057184258\\
4.59820000000002 0.931832345420449\\
4.6044 0.935053595174005\\
4.61070000000001 0.937963194165156\\
4.6169 0.93415470165269\\
4.6232 0.935253635470534\\
4.6294 0.93475997379455\\
4.6357 0.93271235101677\\
4.64190000000001 0.931332673950148\\
4.6482 0.935078492893315\\
4.65440000000001 0.935734419015841\\
4.66070000000001 0.937400849125556\\
4.66690000000001 0.933898856123223\\
4.67320000000001 0.94214515246728\\
4.67940000000002 0.934841535288843\\
4.68570000000001 0.939873450215705\\
4.69190000000002 0.941059096780114\\
4.69820000000001 0.937881632670863\\
4.70440000000001 0.941346708365253\\
4.7107 0.939036371721644\\
4.71690000000001 0.934292068379921\\
4.72320000000001 0.941927082787802\\
4.7294 0.940329336041702\\
4.73570000000001 0.941477206756122\\
4.7419 0.93253205718728\\
4.74820000000001 0.939299085587473\\
4.7544 0.942738405020507\\
4.76070000000001 0.93861225195132\\
4.76690000000001 0.933349389214301\\
4.77320000000002 0.939766991002102\\
4.77940000000001 0.945417056201502\\
4.78570000000001 0.941327820440258\\
4.79190000000001 0.941190453713028\\
4.79820000000001 0.941465187167489\\
4.8044 0.937738256149316\\
4.8107 0.936165407122527\\
4.8169 0.941745930416266\\
4.8232 0.940079500306551\\
4.82940000000001 0.939261309737484\\
4.8357 0.93910505508526\\
4.84190000000001 0.942738405020507\\
4.84820000000001 0.944106062498496\\
4.85440000000001 0.933492765735848\\
4.86070000000001 0.940690782242727\\
4.86690000000002 0.940465844226887\\
4.87320000000001 0.93705743230748\\
4.87940000000002 0.938156366125324\\
4.88570000000001 0.941982888020739\\
4.89190000000001 0.937175911109716\\
4.89820000000002 0.936801586778013\\
4.9044 0.938911883125092\\
4.91070000000001 0.941945970712796\\
4.9169 0.946440438319369\\
4.92320000000001 0.941614573483352\\
4.9294 0.939948143373637\\
4.93570000000001 0.945654013805974\\
4.9419 0.943581493308884\\
4.94820000000001 0.947695626789438\\
4.95440000000001 0.946115909426287\\
4.9607 0.941047077191481\\
4.96690000000001 0.951022477214552\\
4.97320000000001 0.942413876127425\\
4.97940000000001 0.949068435519698\\
4.98570000000001 0.939186616579553\\
4.9919 0.949037528006071\\
4.99820000000001 0.943955817640587\\
5.0044 0.937076320232474\\
5.0107 0.948232215567682\\
5.01690000000001 0.942176918522952\\
5.0232 0.942026673665044\\
5.02940000000001 0.941933092582118\\
5.03570000000001 0.94563512588098\\
5.04190000000001 0.946771835548813\\
5.04820000000001 0.944848701367586\\
5.05440000000002 0.943743757755425\\
5.06070000000001 0.944142979806439\\
5.06690000000002 0.942745273356868\\
5.07320000000001 0.95009267617961\\
5.07940000000002 0.945785370738889\\
5.08570000000002 0.941665227464018\\
5.0919 0.941658359127657\\
5.09820000000001 0.94691521207036\\
5.1044 0.945997430624051\\
5.11070000000001 0.944929404319833\\
5.1169 0.945778502402527\\
5.12320000000001 0.944255448814359\\
5.1294 0.944973189964138\\
5.1357 0.941658359127657\\
5.14190000000001 0.94779521766668\\
5.1482 0.94765785093945\\
5.15440000000001 0.949230699966239\\
5.16070000000001 0.952121411032396\\
5.16690000000001 0.942613916423954\\
5.17320000000001 0.94237009048312\\
5.17940000000002 0.943968695771265\\
5.18570000000001 0.945560432723049\\
5.1919 0.949281353946905\\
5.1982 0.946515131477301\\
5.20440000000001 0.946965007508981\\
5.2107 0.948375592089229\\
5.21690000000001 0.940946627772194\\
5.22320000000001 0.953538863949006\\
5.22940000000001 0.948525836947137\\
5.23570000000001 0.943512809945269\\
5.24190000000002 0.947739412433743\\
5.24820000000001 0.953762943422801\\
5.25440000000002 0.948163532204067\\
5.26070000000001 0.951484372834864\\
5.26690000000001 0.957134438034264\\
5.27320000000002 0.951953136791538\\
5.27940000000001 0.947714514714432\\
5.28570000000002 0.950791529404396\\
5.2919 0.944380795952957\\
5.29820000000001 0.948094848840452\\
5.3044 0.949287363741221\\
5.3107 0.947913696468916\\
5.3169 0.947707646378071\\
5.3232 0.953869402636404\\
5.32940000000001 0.948619418030063\\
5.3357 0.946434428525053\\
5.34190000000001 0.94716504780551\\
5.34820000000001 0.94193996091848\\
5.35440000000001 0.949087323444692\\
5.36070000000001 0.954306400537406\\
5.36690000000002 0.953239232775234\\
5.37320000000001 0.941308932515264\\
5.3794 0.946047226062672\\
5.38570000000001 0.943887134276972\\
5.39190000000001 0.943850216969029\\
5.3982 0.948251103492676\\
5.40440000000001 0.94572269716959\\
5.41070000000001 0.950199135393214\\
5.41690000000001 0.949793045005839\\
5.42320000000001 0.946671386129525\\
5.42940000000002 0.949493413832067\\
5.43570000000001 0.946808752856756\\
5.4419 0.948369582294913\\
5.44820000000001 0.948613408235747\\
5.45440000000001 0.9608055638195\\
5.46070000000002 0.952321451328925\\
5.46690000000001 0.955286855553014\\
5.47320000000002 0.954806072007707\\
5.4794 0.952627092297013\\
5.48570000000001 0.945373270557197\\
5.4919 0.953094997711642\\
5.4982 0.950966671981614\\
5.5044 0.945604218367353\\
5.5107 0.95563027237109\\
5.51690000000001 0.953251252363867\\
5.5232 0.953120753972998\\
5.52940000000001 0.950873090898689\\
5.53570000000001 0.953094997711642\\
5.54190000000001 0.946472204375041\\
5.54820000000001 0.942026673665044\\
5.55440000000002 0.951153834147466\\
5.56070000000001 0.949661688072925\\
5.56690000000002 0.951166712278144\\
5.57320000000001 0.955905005825551\\
5.57940000000001 0.954468664983947\\
5.5857 0.954281502818096\\
5.59190000000001 0.953113885636636\\
5.59820000000001 0.950130452029599\\
5.60440000000001 0.947108384030527\\
5.61070000000001 0.946983895433975\\
5.6169 0.951434577396243\\
5.62320000000001 0.958226503515747\\
5.6294 0.954905662884949\\
5.63570000000001 0.945598208573037\\
5.64190000000001 0.950335643578399\\
5.64820000000002 0.957296702480805\\
5.65440000000001 0.955848342050569\\
5.66070000000002 0.951222517511081\\
5.66690000000001 0.955905005825551\\
5.67320000000001 0.944810925517597\\
5.6794 0.958788848555347\\
5.6857 0.948669213468684\\
5.6919 0.955304884935963\\
5.6982 0.94597253290474\\
5.70440000000001 0.954993234173559\\
5.7107 0.95247770598115\\
5.71690000000001 0.956779001627555\\
5.72320000000001 0.954731378849776\\
5.72940000000001 0.952221001909638\\
5.73570000000001 0.951903341352918\\
5.74190000000002 0.957128428239948\\
5.74820000000001 0.948600530105069\\
5.75440000000002 0.955367558505262\\
5.76070000000001 0.959357203389263\\
5.76690000000001 0.962796522822297\\
5.7732 0.952321451328925\\
5.77940000000001 0.952907835545791\\
5.78570000000001 0.951834657989302\\
5.7919 0.961897629300982\\
5.79820000000001 0.956697440133262\\
5.8044 0.954693602999787\\
5.81070000000001 0.951340996313317\\
5.8169 0.945772492608211\\
5.82320000000001 0.961597998127211\\
5.82940000000001 0.957046866745655\\
5.83570000000002 0.954774305952035\\
5.84190000000001 0.954138126296549\\
5.84820000000001 0.950823295460068\\
5.85440000000001 0.952352358842552\\
5.86070000000001 0.951079140989534\\
5.8669 0.958813746274658\\
5.8732 0.9471401500862\\
5.8794 0.959987373250433\\
5.8857 0.946858548295377\\
5.89190000000001 0.958707287061054\\
5.8982 0.950079798048932\\
5.90440000000001 0.958714155397415\\
5.91070000000001 0.950710826452148\\
5.91690000000001 0.956048382347098\\
5.92320000000001 0.959687742076662\\
5.92940000000002 0.948082829251819\\
5.93570000000001 0.95344528286608\\
5.94190000000002 0.957683904943186\\
5.94820000000001 0.956466492323106\\
5.95440000000002 0.962665165889383\\
5.96070000000002 0.952146308751706\\
5.9669 0.956529165892405\\
5.97320000000001 0.952876928032164\\
5.9794 0.962128577111139\\
5.98570000000001 0.961198776076197\\
5.9919 0.959874904242513\\
5.99820000000001 0.956310237670881\\
6.0044 0.957471845058024\\
6.01070000000001 0.956079289860725\\
6.01690000000001 0.956098177785719\\
6.0232 0.956391799165174\\
6.02940000000001 0.960555728084349\\
6.03570000000001 0.953338823652476\\
6.04190000000001 0.955236201572348\\
6.04820000000001 0.964307556821833\\
6.05440000000002 0.949999095096685\\
6.06070000000001 0.969076757882867\\
6.0669 0.963246398853977\\
6.0732 0.962216148399748\\
6.07940000000001 0.957246907042184\\
6.0857 0.957508762365967\\
6.09190000000001 0.960449268870746\\
6.09820000000001 0.950891978823683\\
6.10440000000001 0.958714155397415\\
6.11070000000001 0.96285318659728\\
6.11690000000002 0.963521132308438\\
6.12320000000001 0.955767639098321\\
6.12940000000002 0.950361399839755\\
6.13570000000001 0.950798397740757\\
6.14190000000001 0.961229683589824\\
6.14820000000002 0.961367050317055\\
6.15440000000001 0.95990581175614\\
6.16070000000001 0.9559230352085\\
6.1669 0.966448760682538\\
6.17320000000001 0.958563910539507\\
6.1794 0.959182060812044\\
6.18570000000001 0.95560537465178\\
6.1919 0.959082469934802\\
6.1982 0.956285339951571\\
6.20440000000001 0.962047874158891\\
6.2107 0.960574616009344\\
6.21690000000001 0.956153983018657\\
6.22320000000001 0.959107367654113\\
6.22940000000001 0.965294021631757\\
6.23570000000001 0.964800359955772\\
6.24190000000002 0.957040856951339\\
6.24820000000001 0.960124739977664\\
6.2544 0.960080954333359\\
6.2607 0.96481323808645\\
6.26690000000001 0.963002572913143\\
6.2732 0.957403161694409\\
6.27940000000001 0.966136251378089\\
6.28570000000001 0.952445939925478\\
6.29190000000001 0.955224181983715\\
6.29820000000001 0.96013074977198\\
6.30440000000002 0.955748751173327\\
6.31070000000001 0.956615878638969\\
6.3169 0.960174535416285\\
6.32320000000001 0.960093832464037\\
6.32940000000001 0.96362072318568\\
6.33570000000002 0.963589815672053\\
6.34190000000001 0.968826922147716\\
6.34820000000002 0.952895815957158\\
6.3544 0.961136102506898\\
6.36070000000001 0.963414673094834\\
6.3669 0.96134816239206\\
6.3732 0.959357203389263\\
6.3794 0.959899801961824\\
6.3857 0.965599662599845\\
6.39190000000001 0.965500071722603\\
6.3982 0.961061409348967\\
6.40440000000001 0.966317403749624\\
6.41070000000001 0.962259934044053\\
6.41690000000001 0.960536840159355\\
6.42320000000001 0.9658932839793\\
6.42940000000002 0.968540169104623\\
6.43570000000001 0.961741374648758\\
6.4419 0.967428357156101\\
6.44820000000001 0.96071799253089\\
6.45440000000001 0.962022117897535\\
6.4607 0.965237357856774\\
6.46690000000001 0.963564059410697\\
6.47320000000001 0.961529314763596\\
6.47940000000001 0.964875053113704\\
6.48570000000001 0.95950057991081\\
6.4919 0.952227870246\\
6.49820000000001 0.969744703594026\\
6.5044 0.966991359255099\\
6.51070000000001 0.955080805462168\\
6.51690000000001 0.963626732979996\\
6.52320000000002 0.967416337567468\\
6.52940000000001 0.965069083615917\\
6.53570000000002 0.962777634897303\\
6.5419 0.960523962028677\\
6.54820000000001 0.964019945236694\\
6.5544 0.953907178486393\\
6.5607 0.967210287476622\\
6.5669 0.953875412430721\\
6.5732 0.967128725982329\\
6.57940000000001 0.959563253480109\\
6.5857 0.965905303567933\\
6.59190000000001 0.96771596874124\\
6.59820000000001 0.960043178483371\\
6.60440000000001 0.965174684287475\\
6.61070000000001 0.965006410046618\\
6.61690000000002 0.96177314070443\\
6.62320000000001 0.967697080816246\\
6.62940000000002 0.963683396754979\\
6.63570000000001 0.965724151196397\\
6.64190000000001 0.960973838060357\\
6.6482 0.963664508829985\\
6.65440000000001 0.959800211084581\\
6.66070000000001 0.970187711289344\\
6.6669 0.964613197789921\\
6.67320000000001 0.965731019532759\\
6.6794 0.966504565915476\\
6.68570000000001 0.957977526322642\\
6.6919 0.960724002325207\\
6.69820000000001 0.958114893049872\\
6.70440000000001 0.969189226890787\\
6.71070000000002 0.969220134404414\\
6.71690000000001 0.959649966226673\\
6.72320000000001 0.962334627201984\\
6.72940000000001 0.962640268170072\\
6.73570000000001 0.959588151199419\\
6.7419 0.963546030027749\\
6.7482 0.969994539329176\\
6.7544 0.964176199888919\\
6.7607 0.963464468533455\\
6.76690000000001 0.959300539614281\\
6.7732 0.958720165191732\\
6.77940000000001 0.963945252078762\\
6.78570000000001 0.966386087113239\\
6.79190000000001 0.969257910254402\\
6.79820000000001 0.963995047517383\\
6.80440000000002 0.963046358557448\\
6.81070000000001 0.965718141402081\\
6.81690000000002 0.9672231656073\\
6.82320000000001 0.969513755783869\\
6.82940000000001 0.97260364860451\\
6.83570000000002 0.964513606912679\\
6.84190000000001 0.965387602714683\\
6.84820000000001 0.969626224791789\\
6.8544 0.9586454720338\\
6.86070000000001 0.965549867161224\\
6.8669 0.96339578516984\\
6.87320000000001 0.962484872059893\\
6.8794 0.970574913751725\\
6.88570000000001 0.963920354359452\\
6.89190000000001 0.959812230673214\\
6.8982 0.969570419558852\\
6.90440000000001 0.969588448941801\\
6.91070000000001 0.969114533732855\\
6.91690000000001 0.970381741791557\\
6.92320000000001 0.971848990146788\\
6.9294 0.967834447543476\\
6.93570000000001 0.96123569338414\\
6.9419 0.97357122548944\\
6.9482 0.960574616009344\\
6.95440000000001 0.976736670010058\\
6.9607 0.970537996443782\\
6.96690000000001 0.965162664698843\\
6.97320000000001 0.976830251092984\\
6.97940000000001 0.973465624817882\\
6.98570000000001 0.963832783070842\\
6.99190000000002 0.96699822759146\\
6.99820000000001 0.96632341354394\\
7.00440000000002 0.971542490636655\\
7.01070000000001 0.970032315179164\\
7.01690000000002 0.964781472030778\\
7.02320000000002 0.966611025129079\\
7.0294 0.964120394655981\\
7.03570000000001 0.966136251378089\\
7.0419 0.965930201287243\\
7.04820000000001 0.972853484339661\\
7.0544 0.967909140701408\\
7.06070000000001 0.972679200304487\\
7.0669 0.970537996443782\\
7.0732 0.967166501832318\\
7.07940000000001 0.956691430338946\\
7.0857 0.969482848270242\\
7.09190000000001 0.967497040519716\\
7.09820000000001 0.966392096907556\\
7.10440000000001 0.972410476644343\\
7.11070000000001 0.966504565915476\\
7.11690000000002 0.96807827348431\\
7.12320000000001 0.973615011133745\\
7.1294 0.972397598513665\\
7.1357 0.970119027925729\\
7.14190000000001 0.967897121112775\\
7.1482 0.96603065070653\\
7.15440000000001 0.969601327072479\\
7.16070000000001 0.970943228289112\\
7.16690000000001 0.972685210098804\\
7.17320000000001 0.977710256689305\\
7.17940000000002 0.968771116914779\\
7.18570000000001 0.974919995042435\\
7.19190000000002 0.965512091311235\\
7.19820000000001 0.977467289290516\\
7.20440000000001 0.965731019532759\\
7.21070000000002 0.966517444046154\\
7.21690000000001 0.963383765581208\\
7.22320000000002 0.973553196106491\\
7.2294 0.971898785585409\\
7.23570000000001 0.959581282863058\\
7.2419 0.973215789082731\\
7.2482 0.974370528133513\\
7.2544 0.967172511626634\\
7.2607 0.96823366959449\\
7.26690000000001 0.968334119013777\\
7.2732 0.966254730180325\\
7.27940000000001 0.967691071021929\\
7.28570000000001 0.965998884650858\\
7.29190000000001 0.966529463634786\\
7.29820000000001 0.972429364569337\\
7.30440000000002 0.970368863660879\\
7.31070000000001 0.969089636013545\\
7.3169 0.967753744591228\\
7.32320000000001 0.965699253477087\\
7.32940000000001 0.96467587135922\\
7.3357 0.964619207584237\\
7.34190000000001 0.970063222692791\\
7.34820000000001 0.967359673792486\\
7.35440000000001 0.971748540727501\\
7.36070000000001 0.976580415357834\\
7.36690000000002 0.967897121112775\\
7.37320000000001 0.964981512327307\\
7.3794 0.97075606612326\\
7.38570000000001 0.970000549123492\\
7.39190000000001 0.975919337983037\\
7.39820000000002 0.972885250395333\\
7.40440000000001 0.969963631815549\\
7.41070000000002 0.966786167706298\\
7.4169 0.966174027228077\\
7.42320000000001 0.970481332668799\\
7.4294 0.965124888848854\\
7.4357 0.974152458454035\\
7.4419 0.971080595016342\\
7.4482 0.974227151611966\\
7.45440000000001 0.967815559618482\\
7.4607 0.976030948448912\\
7.46690000000001 0.973153115513433\\
7.47320000000001 0.972004386256967\\
7.47940000000001 0.973509410462187\\
7.48570000000001 0.975356992943437\\
7.49190000000002 0.969245032123724\\
7.49820000000001 0.972616526735188\\
7.50440000000002 0.972347803075044\\
7.51070000000001 0.973427848967893\\
7.51690000000001 0.975793990844439\\
7.5232 0.971193064024262\\
7.52940000000001 0.969851162807629\\
7.53570000000001 0.967229175401617\\
7.54190000000001 0.973790153710964\\
7.54820000000001 0.966280486441681\\
7.5544 0.971299523237866\\
7.56070000000001 0.97271611761243\\
7.5669 0.961179888151203\\
7.57320000000001 0.976761567729369\\
7.57940000000001 0.978921659515069\\
7.58570000000002 0.973515420256503\\
7.59190000000001 0.973246696596358\\
7.59820000000002 0.967897121112775\\
7.60440000000001 0.968302352958105\\
7.61070000000001 0.970750056328944\\
7.6169 0.970343965941568\\
7.6232 0.976792475242996\\
7.6294 0.972635414660183\\
7.6357 0.968496383460318\\
7.64190000000001 0.972166650703508\\
7.6482 0.966416994626866\\
7.65440000000001 0.972984841272575\\
7.66070000000001 0.96688575858354\\
7.66690000000001 0.973852827280263\\
7.67320000000001 0.967141604113007\\
7.67940000000002 0.975469461951357\\
7.68570000000001 0.976904944250916\\
7.69190000000002 0.971574256692327\\
7.69820000000001 0.973671674908728\\
7.70440000000001 0.972541833577257\\
7.7107 0.972135743189882\\
7.71690000000001 0.979633390870532\\
7.72320000000001 0.975569052828599\\
7.7294 0.967141604113007\\
7.73570000000001 0.972816567031718\\
7.7419 0.975987162804607\\
7.74820000000001 0.96978848923833\\
7.7544 0.973022617122564\\
7.76070000000001 0.974177356173345\\
7.76690000000001 0.974495875272111\\
7.77320000000002 0.975494359670668\\
7.77940000000001 0.978291489653899\\
7.78570000000001 0.979633390870532\\
7.79190000000001 0.974770608726572\\
7.79820000000001 0.973465624817882\\
7.8044 0.975045342181033\\
7.8107 0.966236700797376\\
7.8169 0.981499861276776\\
7.8232 0.976992515539525\\
7.82940000000001 0.971736521138868\\
7.8357 0.974171346379029\\
7.84190000000001 0.972410476644343\\
7.84820000000001 0.974064887165425\\
7.85440000000001 0.967160492038001\\
7.86070000000001 0.976761567729369\\
7.86690000000002 0.975538145314972\\
7.87320000000001 0.97368369449736\\
7.87940000000002 0.976755557935053\\
7.88570000000001 0.974476987347117\\
7.89190000000002 0.972460272082964\\
7.89820000000002 0.979683186309153\\
7.9044 0.973827929560952\\
7.91070000000001 0.968508403048951\\
7.9169 0.971567388355965\\
7.92320000000001 0.977566880167758\\
7.9294 0.976511731994218\\
7.93570000000001 0.981087761095085\\
7.9419 0.971280635312872\\
7.94820000000001 0.96998852953486\\
7.95440000000001 0.973153115513433\\
7.9607 0.972971963141897\\
7.96690000000001 0.978409968456135\\
7.97320000000001 0.981524758996087\\
7.97940000000001 0.973528298387181\\
7.98570000000001 0.969526633914547\\
7.99190000000002 0.977049179314508\\
7.99820000000001 0.974214273481288\\
8.0044 0.973190891363421\\
8.0107 0.981480973351782\\
8.01690000000001 0.977341942151918\\
8.0232 0.974677027643646\\
8.02940000000001 0.972073069620583\\
8.03570000000001 0.973122207999806\\
8.04190000000001 0.96600575298722\\
8.04820000000001 0.979927012249987\\
8.05440000000002 0.977679349175678\\
8.06070000000001 0.972298007636423\\
8.06690000000002 0.973209779288415\\
8.07320000000001 0.972273109917112\\
8.07940000000001 0.973384063323589\\
8.08570000000002 0.975637736192214\\
8.09190000000001 0.978360173017514\\
8.09820000000001 0.97780383777223\\
8.1044 0.971124380660647\\
8.11070000000001 0.977317044432607\\
8.1169 0.969151451040799\\
8.12320000000001 0.975463452157041\\
8.1294 0.974177356173345\\
8.1357 0.965399622303315\\
8.14190000000001 0.983016905070628\\
8.1482 0.975319217093449\\
8.15440000000001 0.973584103620118\\
8.16070000000001 0.975444564232047\\
8.16690000000001 0.973265584521353\\
8.17320000000001 0.976599303282828\\
8.17940000000002 0.973839949149585\\
8.18570000000001 0.978671823779918\\
8.1919 0.976299672109056\\
8.1982 0.971736521138868\\
8.20440000000001 0.974008223390442\\
8.2107 0.973521430050819\\
8.21690000000001 0.97561884826722\\
8.22320000000001 0.973759246197337\\
8.22940000000001 0.972997719403253\\
8.23570000000001 0.975026454256039\\
8.24190000000002 0.972660312379493\\
8.24820000000001 0.971960600612663\\
8.2544 0.978846966357137\\
8.26070000000001 0.967977824065023\\
8.26690000000001 0.971960600612663\\
8.27320000000002 0.974314722900575\\
8.27940000000001 0.981893073533474\\
8.28570000000002 0.975425676307052\\
8.2919 0.965611682188477\\
8.29820000000001 0.978147254590307\\
8.3044 0.974177356173345\\
8.3107 0.977698237100672\\
8.3169 0.981175332383694\\
8.3232 0.971929693099036\\
8.32940000000001 0.982791967054788\\
8.3357 0.979371535546749\\
8.34190000000001 0.978965445159374\\
8.34820000000001 0.979551829376239\\
8.35440000000001 0.978877873870764\\
8.36070000000001 0.975919337983037\\
8.36690000000002 0.974701925362957\\
8.37320000000001 0.97368369449736\\
8.3794 0.983772422070396\\
8.38570000000001 0.9744709775528\\
8.39190000000001 0.977754042333609\\
8.3982 0.977217453555365\\
8.40440000000001 0.979152607325225\\
8.41070000000001 0.98309159822856\\
8.41690000000001 0.971336440545809\\
8.42320000000001 0.980656772988399\\
8.4294 0.97737370820759\\
8.43570000000001 0.967416337567468\\
8.4419 0.980357141814628\\
8.44820000000001 0.96933861320665\\
8.45440000000001 0.970550016032414\\
8.46070000000002 0.981450065838155\\
8.46690000000001 0.975744195405818\\
8.47320000000002 0.977292146713297\\
8.4794 0.982992007351318\\
8.48570000000001 0.979414462649008\\
8.4919 0.975219626216207\\
8.4982 0.983148262003543\\
8.5044 0.981213108233683\\
8.5107 0.974402294189185\\
8.51690000000001 0.97640613132266\\
8.5232 0.977535972654131\\
8.52940000000001 0.983816207714701\\
8.53570000000001 0.978397090325457\\
8.54190000000001 0.971080595016342\\
8.54820000000001 0.972654302585177\\
8.55440000000002 0.977935194705145\\
8.56070000000001 0.983503698410252\\
8.56690000000002 0.977754042333609\\
8.5732 0.972404466850026\\
8.57940000000001 0.974420323572134\\
8.5856 0.969769601313336\\
8.59190000000001 0.97780383777223\\
8.5981 0.979739850084136\\
8.60440000000001 0.972616526735188\\
8.61060000000001 0.974301844769898\\
8.61690000000002 0.978215937953922\\
8.62310000000001 0.981037965656464\\
8.62940000000002 0.978471783483389\\
8.6357 0.970912320775485\\
8.64190000000002 0.978222806290284\\
8.6482 0.97967116672052\\
8.65440000000002 0.98450905114517\\
8.66070000000001 0.980837925359935\\
8.66690000000001 0.970007417459854\\
8.67320000000001 0.978852976151454\\
8.67940000000002 0.979982817482924\\
8.6857 0.978085439563053\\
8.69189999999999 0.975350983149121\\
8.6982 0.979801665111389\\
8.70439999999999 0.981037965656464\\
8.7107 0.978778282993522\\
8.7169 0.974458099422122\\
8.72320000000001 0.987823881981651\\
8.7294 0.979146597530909\\
8.73570000000001 0.977822725697225\\
8.7419 0.985295475658565\\
8.74820000000001 0.979121699811598\\
8.7544 0.980425825178243\\
8.76070000000001 0.972185538628503\\
8.76690000000001 0.977385727796223\\
8.77320000000002 0.984465265500865\\
8.7794 0.977280127124664\\
8.78570000000001 0.973815051430274\\
8.7919 0.985976299500401\\
8.79820000000001 0.978335275298204\\
8.8044 0.97922129068884\\
8.81070000000001 0.979471126423991\\
8.8169 0.977092964958813\\
8.82320000000001 0.978927669309385\\
8.82940000000001 0.980869691415607\\
8.83570000000002 0.984390572342934\\
8.84190000000001 0.98124401574731\\
8.84820000000002 0.980425825178243\\
8.85440000000001 0.976593293488512\\
8.86070000000002 0.979583595431911\\
8.8669 0.979845450755694\\
8.87320000000001 0.980588089624784\\
8.8794 0.97669889416007\\
8.88570000000001 0.976118519737521\\
8.89190000000001 0.976081602429578\\
8.89820000000002 0.975419666512736\\
8.90440000000001 0.978179020645979\\
8.91070000000002 0.983010895276312\\
8.91690000000001 0.977410625515533\\
8.92310000000001 0.976330579622683\\
8.92940000000002 0.976936710306588\\
8.9357 0.983984481955558\\
8.94190000000002 0.982180685118613\\
8.94810000000001 0.98453394886448\\
8.95440000000002 0.979858328886372\\
8.9606 0.979240178613835\\
8.96690000000001 0.985014732409787\\
8.9731 0.976218969156808\\
8.97940000000001 0.974507894860743\\
8.98569999999999 0.978009887863076\\
8.99190000000002 0.981518749201771\\
8.9982 0.980825905771302\\
9.00440000000002 0.980982160423526\\
9.0107 0.972741015331741\\
9.01690000000002 0.980801008051991\\
9.0232 0.977635563531373\\
9.02940000000002 0.973809041635958\\
9.03570000000001 0.973896612924568\\
9.04190000000003 0.982473447956023\\
9.04820000000001 0.980894589134917\\
9.0544 0.984770906468953\\
9.06070000000001 0.978385070736825\\
9.06689999999999 0.983472790896625\\
9.0732 0.981587432565386\\
9.07939999999999 0.980282448656696\\
9.0857 0.986774743602428\\
9.0919 0.973396941454267\\
9.09820000000001 0.979053016447983\\
9.1044 0.985314363583559\\
9.11070000000001 0.972291997842106\\
9.1169 0.974158468248351\\
9.12320000000001 0.976137407662515\\
9.1294 0.974264927461955\\
9.13570000000001 0.975975143215974\\
9.14190000000001 0.983422995458004\\
9.14820000000002 0.982860650418404\\
9.15440000000001 0.982186694912929\\
9.16070000000002 0.98295423150133\\
9.1669 0.976006050729601\\
9.17320000000001 0.985782268998188\\
9.1794 0.978721619218539\\
9.18570000000001 0.978622028341297\\
9.1919 0.974564558635726\\
9.19820000000001 0.985301485452881\\
9.20440000000001 0.977797827977914\\
9.21070000000002 0.981780604525554\\
9.21690000000001 0.978053673507381\\
9.22320000000002 0.985526423468721\\
};
\addlegendentry{Not suspended \, \,};
\end{axis}
\end{tikzpicture}%
}
\caption{\textbf{Drop in free-carrier lifetime.} We measured an increase in the free-carrier recombination rate after the suspension of the silicon beams. Both this finding and the higher propagation losses likely originate in a deterioration of the silicon wires' surface state during the fabrication of the suspended beams.}
\label{fig:freecarrier}
\end{figure*}
\subsection*{The interface reflections are negligible}
Our device has discontinuities between the suspended nanobeams and the beams fixed at the anchors (fig.\ref{fig:structure_and_modes}). Since the beams are spaced periodically, optical reflections may build up. However, we simulated an upper bound for the Fresnel reflection at the discontinuity of less than $10^{-4}$ -- indicating that reflections are negligibly small. Empirically, there are indeed no notable differences in the transmission spectrum of a regular waveguide versus that of a suspended waveguide (fig.\ref{fig:passive}). Therefore, our device can be treated as a single-pass structure.
\begin{figure*}
\centering
\subfigure{
\begin{tikzpicture}
\begin{axis}[scale = 0.34*\textwidth/(4.5in),
width=4.52083333333333in,
height=3.565625in,
scale only axis,
xmin=1.51,
xmax=1.555,
xlabel={Wavelength ($\mu$m)},
ymin=-13.5,
ymax=-9,
ylabel={$T(\lambda)$ (dB)},
ylabel style = {yshift=-0.8ex},
legend style={fill=none,draw=none,legend cell align=left,at={(0.70,0.4)},anchor=north},
name=plottrans,
axis x line*=bottom,
axis y line*=left,
minor xtick={1.515,1.52,...,1.55},
minor ytick={-13,-12,...,-9}
]
\addplot [
color=red!40!black,
solid,
line width=0.3pt,
mark size=0.5pt,
mark=o,
mark options={solid}
]
table[row sep=crcr]{
1.51250713403235 -11.1524429955758\\
1.51258908514639 -11.0294546351029\\
1.51266433441909 -11.0435194194671\\
1.51274348712671 -11.1235512874546\\
1.51282527507164 -11.2325952873335\\
1.51290702363979 -10.9595347130265\\
1.51298437256528 -10.917164573171\\
1.51306069186976 -10.8375992202962\\
1.51314200946727 -10.6658612548656\\
1.51322357973674 -10.7357865445643\\
1.51330220089341 -10.7753717636093\\
1.5133808890692 -10.6122233465519\\
1.5134618536293 -10.4647223119887\\
1.51354410491041 -10.4617162597592\\
1.51362451098899 -10.3763662129535\\
1.51370058167335 -10.4907140118935\\
1.51377923591441 -10.5375738302345\\
1.5138633884396 -10.2296030586404\\
1.51394388213522 -10.238903717105\\
1.51402075333272 -10.4543645023799\\
1.51409762367891 -10.5045650323\\
1.5141788825819 -10.4220360556863\\
1.51426298331771 -10.4505689256626\\
1.51434354794663 -10.5444739650905\\
1.51442136357481 -10.4557997363853\\
1.51450070469459 -10.5132263275946\\
1.51458447745613 -10.8706362899125\\
1.51466529416334 -10.9423063800342\\
1.51474292965811 -10.7278339713052\\
1.51482261640458 -10.7799004868411\\
1.51490505135946 -10.7877636398193\\
1.51498795585392 -10.744544633292\\
1.51506552930128 -10.7541468347173\\
1.51514115082533 -10.6794717876737\\
1.51522356899523 -10.7505008417876\\
1.51530875800436 -10.8179592180258\\
1.51538929322073 -10.7394866875222\\
1.5154655647786 -10.6550313740605\\
1.51554329448631 -10.7865530275377\\
1.51562717999784 -10.669609696972\\
1.51570919654206 -10.2035671584135\\
1.51578630280328 -10.2490350212173\\
1.51586295066901 -10.3640318861419\\
1.51594379961506 -10.1241778876816\\
1.516026840494 -10.1925567487413\\
1.51610399515316 -10.2803445697958\\
1.51618023308189 -10.2562780734948\\
1.51626153225376 -10.4512112329367\\
1.51634516645054 -10.2336030712912\\
1.51642627983677 -10.244622114125\\
1.51650087647015 -10.4571174567486\\
1.51657683274718 -10.2919744006543\\
1.51666042070471 -10.1357405908937\\
1.51674172443492 -10.0648804333561\\
1.51681751533438 -10.6672670654391\\
1.51689226029197 -10.4911891204569\\
1.51697220760651 -10.0340970220122\\
1.51705483224357 -10.41111271313\\
1.51713354889984 -10.2393118605058\\
1.51720993016004 -10.3716208096109\\
1.51728705926458 -10.3602726690212\\
1.51736949333751 -10.1057389614069\\
1.51745000472658 -10.4041569456616\\
1.51752484153171 -10.2569305025287\\
1.51760147653325 -10.3329605258883\\
1.51768216051749 -10.7204254029764\\
1.51776340777418 -10.5472947668737\\
1.5178414668377 -10.4021411554871\\
1.51791570588118 -10.4892626922359\\
1.51799268381901 -10.3889923385561\\
1.51807587197371 -10.3982089416675\\
1.51815664140838 -10.478001194379\\
1.51823273955479 -10.4915706581785\\
1.51830924489123 -10.665267822958\\
1.51839008637104 -10.4431897896159\\
1.51847163310779 -10.2630065832906\\
1.51854981259093 -10.2975616342951\\
1.51862642323342 -10.0549149092526\\
1.51870525343363 -10.0228883688934\\
1.51878777370683 -10.2170948798723\\
1.51886726591389 -10.0907989200568\\
1.51894171221596 -9.84825885504345\\
1.51901971852349 -10.095035215485\\
1.51910254995387 -10.1099983215741\\
1.51918348575588 -9.81916355401504\\
1.51926124399411 -9.77620948858443\\
1.51933653995066 -9.77738741542122\\
1.51941583807745 -10.0770805966896\\
1.51949800203748 -10.2130425897703\\
1.51957625384856 -9.91754287321176\\
1.5196520196476 -10.0297141257484\\
1.51972977725248 -10.5391544131007\\
1.51981166011632 -10.4635329661133\\
1.51989111772227 -10.08824873357\\
1.51996746528556 -10.2963215412496\\
1.52004584474244 -10.3118782881087\\
1.52012733454152 -10.1404066181903\\
1.52020821118117 -10.3291757929082\\
1.52028556411803 -10.1960513726109\\
1.52036132936544 -10.0365053326481\\
1.52044123402303 -10.2051693125065\\
1.52052361259095 -10.2703998839215\\
1.5206012461382 -10.4174871557094\\
1.52067647134684 -10.7273591815933\\
1.52075401509362 -10.5695323740899\\
1.52083475631007 -10.0498431870471\\
1.52091459195541 -9.88302158339903\\
1.52099197573843 -10.047920817852\\
1.52106919444702 -9.97819178018165\\
1.52114869670026 -9.87796751202484\\
1.52122976471523 -10.1087433731307\\
1.52130729035563 -10.2133352231127\\
1.52138240176429 -10.0496579963435\\
1.52146256893115 -10.0681215355147\\
1.52154455840901 -10.3236646885014\\
1.52162242905391 -10.1610287789941\\
1.52169853461083 -10.167357046466\\
1.52177483888748 -10.0692777191971\\
1.52185620665802 -9.8648175675658\\
1.52193755498156 -10.4076399997722\\
1.52201450087401 -9.99987652294559\\
1.52209145966279 -9.49085369198044\\
1.52217164143812 -9.69523220610307\\
1.5222540832198 -9.7073252783239\\
1.52233202354533 -9.75646014152442\\
1.52240872870107 -9.61859933751534\\
1.52248868217036 -9.83364500473674\\
1.52257032411434 -9.71520768801383\\
1.5226507360812 -9.55136215584676\\
1.52272643395722 -10.0432656476909\\
1.52280338088146 -10.0306499497965\\
1.52288687292363 -9.9388854068788\\
1.52296845082005 -9.93009135946364\\
1.52304461808183 -10.1328222757588\\
1.52312068867223 -10.4108816474164\\
1.52320137471515 -10.2398186127827\\
1.52328373571871 -10.3270987408925\\
1.5233619001587 -10.3298443782895\\
1.52343729370277 -10.1215660591063\\
1.52351481956497 -10.1911045217381\\
1.52359753724672 -10.222872607163\\
1.5236773530373 -10.1790598829305\\
1.52375164661731 -10.1495615493399\\
1.52382867695956 -10.1398137946671\\
1.52391096009958 -10.4349859836559\\
1.52399184595087 -10.1281225790264\\
1.52406779806629 -9.7414246615623\\
1.52414248584572 -10.2302693260269\\
1.52422093123999 -10.0208347305488\\
1.52430397141495 -9.97374685243051\\
1.52438306935002 -9.75548041504729\\
1.52445617463005 -9.56722385189688\\
1.52453218245132 -10.1540295890223\\
1.52461349214496 -9.57380286251348\\
1.52469323446517 -9.59814869491358\\
1.52476888914088 -9.91282489999423\\
1.52484393624905 -9.95463024483946\\
1.52492487418863 -10.144743698359\\
1.52500699476992 -9.79729113255637\\
1.52508363716399 -10.1058961662164\\
1.52515779825998 -9.99396513251402\\
1.5252348220549 -9.78366042676661\\
1.52531945720942 -9.82606999308067\\
1.52539995276754 -9.85753237629939\\
1.52547416144997 -10.2722317678114\\
1.52555129782219 -9.93774310593208\\
1.52563122385779 -10.0393267946413\\
1.52571349927552 -10.2639354521251\\
1.52579129800535 -10.070683482953\\
1.52586547986176 -10.1865343271203\\
1.52594542501362 -10.0086114703249\\
1.52602731646485 -9.8434445234888\\
1.52610643297648 -9.92095559943792\\
1.52618151118749 -10.2670393499387\\
1.52625798622615 -10.1211846578217\\
1.52634138894263 -9.68888082369929\\
1.52642250995054 -10.0103707516107\\
1.52649957738704 -10.1237143480798\\
1.5265746235495 -10.1114255164152\\
1.52665388146566 -10.0798255438413\\
1.52673814973708 -9.84512735170351\\
1.52681617303349 -10.4159080962626\\
1.52689135504628 -10.4357263206797\\
1.52696942343934 -9.67818467648899\\
1.52705104053115 -9.66965203222325\\
1.52713235434245 -9.53387029837004\\
1.52720906721994 -9.49647549181901\\
1.52728576601992 -9.79141754849423\\
1.52736686369732 -9.58651382004431\\
1.52744918669608 -9.63735787047803\\
1.52752640705658 -9.79716534594641\\
1.52760115938527 -9.43120492915264\\
1.52768172626255 -9.49448905402995\\
1.52776544823751 -9.92044324750548\\
1.52784507665395 -9.68371638439841\\
1.52791991971293 -9.41984630635277\\
1.52799696669274 -9.39163483980623\\
1.52807896471668 -9.55627831261169\\
1.52816024594347 -9.94498601536553\\
1.52823777018888 -9.79449230865123\\
1.52831314356963 -9.83694441786923\\
1.5283945431729 -10.0345149843819\\
1.52847620955535 -10.0651450743169\\
1.52855245470413 -10.0730625163905\\
1.52862948049431 -9.95279564846662\\
1.52870860645192 -10.2606573555891\\
1.52879107348886 -10.2836934489972\\
1.52887036067333 -10.0332252183778\\
1.52894514991526 -10.2207999359883\\
1.52902273064072 -10.3648867312631\\
1.5291035587042 -10.1559473442066\\
1.52918468974216 -9.98833389977078\\
1.52926161965482 -9.92105496717386\\
1.52933773068491 -9.82944686534661\\
1.52941965692254 -9.93745142519317\\
1.52950035439433 -9.67768315541752\\
1.52957798707563 -9.94405785067751\\
1.5296555113118 -10.4880962026366\\
1.52973451396026 -9.56644535880993\\
1.52981736062965 -9.4409616144557\\
1.52989653476883 -10.3112826305346\\
1.52997216943298 -9.84011671892611\\
1.53005123363067 -9.30443586478235\\
1.53013354192199 -9.48245897722355\\
1.53021399795331 -9.68575268657511\\
1.53029029825282 -9.91967854739026\\
1.53036806773137 -9.73654424488123\\
1.53044993103576 -9.59232214983738\\
1.5305315553467 -9.79950673098684\\
1.53061125714171 -9.64643949628173\\
1.5306889394802 -9.63992936955152\\
1.53076916946338 -9.7689972747315\\
1.53085288853793 -9.78470392909645\\
1.53093143188476 -10.1515508455247\\
1.53100783823399 -10.0354346706477\\
1.53108861267256 -9.68325401422723\\
1.53117056202959 -9.72243599217525\\
1.53125071848785 -9.70919932019285\\
1.5313274342198 -9.77442407429158\\
1.53140422017624 -9.94241842405449\\
1.53148678644332 -10.0571426987102\\
1.53156939678186 -9.87227977543671\\
1.53164775580655 -10.4343288651099\\
1.53172460741519 -10.9551608194958\\
1.53180533577886 -9.92696190234597\\
1.53188797590639 -9.83516942625129\\
1.53196541873174 -10.4130722281533\\
1.53204245410438 -10.3231786254431\\
1.53212292166053 -10.6294622219133\\
1.53220530486436 -10.4323782614717\\
1.5322863224311 -9.93415299059864\\
1.5323621734874 -10.0166197570033\\
1.53243932094986 -9.98726736818413\\
1.53252340798878 -9.76876593124776\\
1.53260590226175 -9.63376026329248\\
1.532683234602 -10.0010796962761\\
1.53276093539658 -9.98442638306848\\
1.53284311875153 -9.58244298384687\\
1.53292647654784 -9.55734308254336\\
1.5330058109214 -9.37200595244119\\
1.53308279121061 -9.47761137469466\\
1.53316254654244 -9.46083119796428\\
1.53324705891383 -9.48136979135123\\
1.53332831198032 -9.79490313175403\\
1.53340338006265 -9.48487142110891\\
1.53348335275651 -9.63349104518808\\
1.53356878228113 -10.1346479583032\\
1.53364986365331 -10.0240307167179\\
1.5337275652596 -9.81780787152983\\
1.53380474472117 -9.78364050612976\\
1.53388665681948 -10.0022037370145\\
1.53397101453533 -9.98066382221774\\
1.53404856235008 -9.91175656264821\\
1.53412382891265 -10.1962829455712\\
1.53420402150042 -10.1959071914378\\
1.53428734539044 -10.170224445802\\
1.53436678633986 -10.2850992279737\\
1.5344426552778 -10.3254832451514\\
1.53452270740639 -10.3269193148916\\
1.53460620240305 -10.1659191344179\\
1.53468685745681 -10.2715407363133\\
1.53476282916098 -10.2480495451899\\
1.53483751193158 -10.1552143469023\\
1.53491903142472 -10.2414075023994\\
1.53500207210941 -9.90302509961472\\
1.53507871865871 -10.0425206183562\\
1.53515442970513 -10.1684513262411\\
1.53523400486122 -9.78126399200135\\
1.53531540233374 -9.80611990574759\\
1.53539430873628 -9.86719435022288\\
1.53547006393812 -9.8741889420563\\
1.53554758682009 -9.56766189652571\\
1.53562983841795 -9.56795393956862\\
1.53571045146687 -10.0305798743319\\
1.53578626940988 -9.9188267033806\\
1.53586181387929 -9.69470119205326\\
1.53594285211966 -10.0111319420677\\
1.53602515376334 -10.3966114075501\\
1.53610233002106 -9.95625150035631\\
1.53617886278182 -9.95010366601079\\
1.53625846886223 -10.1801368683834\\
1.53634000235908 -9.99413176400614\\
1.53642072987203 -10.4465795047285\\
1.53649671245436 -10.491298840966\\
1.53657420043654 -10.362204225272\\
1.53665742858384 -10.4366049878212\\
1.53673753849466 -10.1351160947425\\
1.53681459151359 -10.2554924856924\\
1.53689260962164 -10.3830095061539\\
1.53697363801963 -10.2906043678729\\
1.53705455316058 -10.2622277190934\\
1.53713174876957 -10.0740909230004\\
1.53720892665228 -10.3246470066043\\
1.53728870520647 -10.4612500639533\\
1.53737110323347 -10.1573443490375\\
1.53745132769613 -10.4737535347861\\
1.53752710555684 -10.7786700213987\\
1.53760607478286 -10.4343300631954\\
1.53768773448398 -10.2012815382345\\
1.53776592344321 -10.6921903901381\\
1.53784313443304 -10.7105825887771\\
1.53792121239865 -10.186147205685\\
1.53800174057303 -10.2923675118694\\
1.53808207236106 -10.494233978563\\
1.53815841730888 -10.4498284094805\\
1.53823403844179 -10.0250518890297\\
1.53831420608977 -10.3539233338543\\
1.53839608691068 -10.6618772323169\\
1.53847382311385 -9.950607965074\\
1.53855010487226 -10.0740256341395\\
1.53862950032548 -10.3380369853611\\
1.53870991512205 -9.91936855023029\\
1.53878726466054 -10.1475449591544\\
1.53886340688954 -10.5758304037404\\
1.53894136343944 -10.1970505184264\\
1.53902210219172 -9.91121302863003\\
1.53910199696149 -9.96671684060185\\
1.53917712406253 -10.1928154482844\\
1.53925230160621 -10.4051791610197\\
1.53933362816233 -10.4001901386456\\
1.53941441118939 -10.6095818264862\\
1.53949071735543 -10.6032191009889\\
1.53956699996353 -10.8821404498046\\
1.53964601410927 -11.3359199072909\\
1.53972754049163 -10.8148452590903\\
1.53980551951906 -10.7661664285029\\
1.53988086098683 -11.1386957135923\\
1.53995871949584 -10.8285395769884\\
1.54003976258069 -10.7375975746508\\
1.54011961011523 -10.6714798527129\\
1.54019439586913 -10.5286132933918\\
1.5402709716587 -10.5682615544163\\
1.54035400160236 -10.5382540796734\\
1.54043463220075 -10.4965271619919\\
1.54050986904621 -10.4837147174915\\
1.54058567753974 -10.7588770015524\\
1.54066513721262 -10.5990147046242\\
1.54074695736497 -10.3684730256138\\
1.54082552018863 -10.9402204982737\\
1.54089972749023 -10.8612762482161\\
1.54097764671259 -10.4358047916699\\
1.54105978917352 -10.7759805907303\\
1.541137466673 -10.5949495096465\\
1.54121200515779 -10.3768296676015\\
1.54128946923849 -10.8141583123864\\
1.54137273029533 -10.641836870493\\
1.54145449213128 -10.557007738608\\
1.54153013493602 -10.7096146368218\\
1.54160562661147 -10.4205951285992\\
1.54168461437116 -10.4272787986897\\
1.54176783559976 -10.4307854117047\\
1.54184703145867 -10.2638525908787\\
1.54192038157216 -10.858533827457\\
1.54199968990534 -11.0353576533913\\
1.54208233551803 -10.6799778596265\\
1.54216018799047 -10.7872697237481\\
1.54223482145858 -10.5129319127428\\
1.54231153053786 -10.2787343148688\\
1.5423948284758 -10.5195760400877\\
1.54247604976632 -10.8363728169391\\
1.54255151171666 -11.2465383410169\\
1.5426268307564 -11.5104706653216\\
1.54270650857082 -11.0794242996172\\
1.54278998758802 -10.892114079225\\
1.54286709458681 -11.5228113661455\\
1.54294023346521 -11.399033566121\\
1.543019691007 -11.2434838253494\\
1.54310195117297 -11.4723217772483\\
1.54318053376475 -11.1795476391233\\
1.54325564114201 -11.4407854827961\\
1.54333228872701 -11.3850914088279\\
1.54341415745659 -10.8206018947652\\
1.54349513469384 -10.9598966642714\\
1.54357044711164 -11.0228751230082\\
1.54364511735673 -10.8234030337792\\
1.54372682024689 -10.7636611341231\\
1.54381014725921 -10.9305643492513\\
1.54388703604744 -10.7572064640751\\
1.54396165384142 -10.7178698956376\\
1.54404016558981 -10.9576197355614\\
1.54412315615352 -10.5969939201519\\
1.54420352216167 -10.7510302479778\\
1.54428010050865 -10.810105632022\\
1.54435926261131 -10.2975623185984\\
1.54444209577813 -10.5720719375505\\
1.54452351454265 -10.8240389634815\\
1.54459904248964 -10.5978367421241\\
1.54467563596204 -10.5977513521754\\
1.54475952056771 -10.9755140308132\\
1.54484164780129 -11.158465447533\\
1.54491948764278 -10.9912987143079\\
1.54499593595539 -11.3068978356742\\
1.54507493235525 -11.2183028285437\\
1.54515842360422 -11.1541208141039\\
1.54523766989822 -12.4271526507658\\
1.54531417617973 -12.2712102343327\\
1.54539442070308 -11.2779347593199\\
1.54547684188667 -11.4018889414289\\
1.54555751003495 -11.3563850329044\\
1.54563297147942 -11.4589768456075\\
1.54571027847715 -11.5490482729676\\
1.54579280279789 -11.4702680534068\\
1.54587260019884 -11.5433581344131\\
1.54594983597541 -11.5051929099048\\
1.54602590349641 -11.5886793876995\\
1.54610538241269 -11.4242563269557\\
1.54618800114211 -11.0617435969436\\
1.54626542972332 -11.3905988589864\\
1.54634143228539 -11.4209708352293\\
1.54642153397629 -11.0698867655869\\
1.54650290112763 -11.3262423958561\\
1.54658212736784 -11.2489518301656\\
1.54665940559872 -11.1951608476005\\
1.5467391011567 -11.4873278625143\\
1.54682163406938 -11.1415451795247\\
1.54690100730305 -11.072991192303\\
1.54697817631443 -11.44552404999\\
1.54705716830551 -11.3697052806631\\
1.54713949009534 -11.2418708375352\\
1.54722186957979 -11.0924932852039\\
1.54729945750421 -11.1802690200509\\
1.54737693725571 -11.2600514426965\\
1.54745864632161 -11.3170760464854\\
1.54753931372673 -11.5495094508141\\
1.54761728133901 -11.6002970382496\\
1.54769639531783 -11.9032157097278\\
1.54777848112188 -11.6317900900746\\
1.54786086887374 -11.2720905669841\\
1.54793878552756 -11.5488901868866\\
1.54801489216618 -11.4954540943844\\
1.54809581374818 -11.6904793167766\\
1.54817927078676 -11.8532080651642\\
1.54825933258866 -11.5787012210557\\
1.54833672499463 -11.7975698345954\\
1.54841634259548 -12.1205775823569\\
1.54849922211781 -12.0810563022291\\
1.54857882750589 -11.9639709647873\\
1.54865541850007 -11.9855926185265\\
1.54873498524788 -12.2039012449991\\
1.54881737004898 -12.1969046560478\\
1.5488986730345 -12.4479468485345\\
1.54897539760858 -12.335599455794\\
1.54905211152696 -11.9537153771094\\
1.54913502277148 -12.5894867496931\\
1.54921696914033 -12.3807911809539\\
1.54929394713464 -11.7511957887346\\
1.54937092834609 -11.8404361807478\\
1.54945188493016 -12.0880076905739\\
1.54953511379755 -12.2670865834918\\
1.54961387608227 -12.0588425218222\\
1.54969054636478 -12.4684694561136\\
1.54977091813663 -12.3237885722692\\
1.54985364872683 -11.5409405172177\\
1.5499340821206 -11.9012658608552\\
1.55001076211086 -12.0445500324672\\
1.55008980027361 -11.7411866375451\\
1.55017425628137 -11.9938226929917\\
1.55025546371313 -12.3566865478848\\
1.55033105909396 -12.1603669186503\\
1.55040826207272 -12.2263191613465\\
1.55048970076172 -12.7208984706317\\
1.55057228466469 -12.3790608877063\\
1.5506505768807 -12.2067334283698\\
1.5507266150328 -12.3478038492656\\
1.55080888915723 -12.2882820343169\\
1.55089080083657 -12.5962855578755\\
1.5509676786546 -12.6037065775571\\
1.55104330079145 -12.5212067131591\\
1.55112269527804 -13.0792642443836\\
1.55120808792474 -13.341297428308\\
1.55128689666229 -12.7879011287088\\
1.55136269197742 -12.5503454495989\\
1.55143994979278 -12.7174254178609\\
1.55152147823642 -12.8726525564597\\
1.55160357041098 -13.2521320547884\\
1.55167727944487 -13.1201601868678\\
1.55175824648806 -12.7556345302458\\
1.55183481936391 -12.8445599326149\\
1.55191987931061 -12.7844996494383\\
1.55200704713484 -12.68574558355\\
};
\end{axis}
\node at (-0.9,-0.9) {\large{\textbf{a}}};
\begin{axis}[scale = 0.34*\textwidth/(4.5in),
width=4.52083333333333in,
height=3.565625in,
scale only axis,
xmin=-0.2,
xmax=3.5,
xlabel={Distance (mm)},
ymin=0,
ymax=120,
ylabel={$|T(z)|$ (a.u.)},
ylabel style = {yshift=-1.5ex},
legend style={fill=none,draw=none,legend cell align=left,at={(0.85,1.2)},anchor=north},
name=plot1,
at=(plottrans.right of south east),
anchor=left of south west,
axis x line*=bottom,
axis y line*=left,
minor xtick={0.5,1,...,3},
minor ytick={25,50,...,100}
]
\node[right, inner sep=0mm, text=black]
at (axis cs:2,45) {$L = 2.7 \, \text{mm}$};
\addplot [
color=blue,
densely dotted,
line width=0.6pt,
mark size=0.5pt,
mark=o,
mark options={solid}
]
table[row sep=crcr,x expr=\thisrow{x}*10,y expr=\thisrow{y}/4]{
x y \\
0 0.278598865904589\\
0.000646203403479147 143.110610121414\\
0.0012924068069583 6.49188612177484\\
0.00193861021043744 152.573282893612\\
0.00258481361391659 111.332921384056\\
0.00323101701739574 81.5191124754069\\
0.00387722042087489 60.4056016685656\\
0.00452342382435403 32.7774907005184\\
0.00516962722783318 104.23843524198\\
0.00581583063131233 12.7833493421319\\
0.00646203403479147 36.7633575215851\\
0.00710823743827062 52.9787491874336\\
0.00775444084174977 50.5578375639892\\
0.00840064424522892 67.2962500063884\\
0.00904684764870806 416.177434529021\\
0.00969305105218721 51.6040513193207\\
0.0103392544556664 80.0141804885212\\
0.0109854578591455 33.24721676615\\
0.0116316612626247 22.1082073640792\\
0.0122778646661038 37.8090981403369\\
0.0129240680695829 23.7104025595442\\
0.0135702714730621 44.9516152700414\\
0.0142164748765412 195.415815200243\\
0.0148626782800204 55.6858975290417\\
0.0155088816834995 90.5863998778204\\
0.0161550850869787 45.7212829153688\\
0.0168012884904578 18.1448759570638\\
0.017447491893937 83.2490716897523\\
0.0180936952974161 26.5929751105273\\
0.0187398987008953 39.9787503189324\\
0.0193861021043744 22.629613475177\\
0.0200323055078536 28.6163400298128\\
0.0206785089113327 43.7726754835412\\
0.0213247123148119 32.9196280503952\\
0.021970915718291 24.3154671913803\\
0.0226171191217702 51.464811029108\\
0.0232633225252493 62.3588629793612\\
0.0239095259287285 21.5446455917886\\
0.0245557293322076 18.8064389386513\\
0.0252019327356868 29.2483264273311\\
0.0258481361391659 27.4377451214832\\
0.026494339542645 41.379595296015\\
0.0271405429461242 39.7085775493202\\
0.0277867463496033 61.0604909734442\\
0.0284329497530825 20.6619068361219\\
0.0290791531565616 32.768717881618\\
0.0297253565600408 32.6093262402953\\
0.0303715599635199 53.7949257183655\\
0.0310177633669991 48.4805557136165\\
0.0316639667704782 24.9569701440117\\
0.0323101701739574 33.7486580005104\\
0.0329563735774365 46.9684947096118\\
0.0336025769809157 55.8880329127117\\
0.0342487803843948 42.2010143182938\\
0.034894983787874 48.1867556989334\\
0.0355411871913531 35.1000416217374\\
0.0361873905948323 37.928507339866\\
0.0368335939983114 10.750093550454\\
0.0374797974017906 43.8357344893571\\
0.0381260008052697 25.8573582921433\\
0.0387722042087488 19.3188980753955\\
0.039418407612228 7.54031198900055\\
0.0400646110157071 18.7692122340426\\
0.0407108144191863 34.2258078595339\\
0.0413570178226654 36.0723631980864\\
0.0420032212261446 34.0823325165271\\
0.0426494246296237 24.9373046796961\\
0.0432956280331029 75.8748520828429\\
0.043941831436582 25.6161708781118\\
0.0445880348400612 32.3619482742086\\
0.0452342382435403 26.158341793214\\
0.0458804416470195 49.2980534457589\\
0.0465266450504986 20.797536235112\\
0.0471728484539778 63.9441791712905\\
0.0478190518574569 26.2689588935145\\
0.048465255260936 12.6497642920191\\
0.0491114586644152 52.6815707238086\\
0.0497576620678943 53.5305302282998\\
0.0504038654713735 40.1505689407507\\
0.0510500688748527 54.8667479807753\\
0.0516962722783318 29.5806048406265\\
0.0523424756818109 24.2065425054431\\
0.0529886790852901 30.4596752025734\\
0.0536348824887692 33.0450950418801\\
0.0542810858922484 37.5778625365168\\
0.0549272892957275 51.3206935542944\\
0.0555734926992067 93.8246322950565\\
0.0562196961026858 168.484258263152\\
0.056865899506165 41.8544229239614\\
0.0575121029096441 29.119765578187\\
0.0581583063131233 11.8903565928866\\
0.0588045097166024 75.2699376448681\\
0.0594507131200816 42.8432817965068\\
0.0600969165235607 33.114457045127\\
0.0607431199270399 48.9898376615231\\
0.061389323330519 45.8926843867369\\
0.0620355267339982 13.3646083117934\\
0.0626817301374773 18.8622804933334\\
0.0633279335409564 26.8037408426105\\
0.0639741369444356 25.6072179239625\\
0.0646203403479148 32.1249590344095\\
0.0652665437513939 15.6979269467836\\
0.065912747154873 49.3097413313821\\
0.0665589505583522 64.0977576437649\\
0.0672051539618313 50.9241243920733\\
0.0678513573653105 13.6772814369456\\
0.0684975607687896 27.5385693762197\\
0.0691437641722688 19.812293469785\\
0.0697899675757479 66.3049590588749\\
0.0704361709792271 70.018304394827\\
0.0710823743827062 30.7855069226134\\
0.0717285777861854 88.1000999831344\\
0.0723747811896645 21.8020700471247\\
0.0730209845931437 34.2361644553309\\
0.0736671879966228 48.4913242924838\\
0.074313391400102 55.0280057913448\\
0.0749595948035811 41.7615107591242\\
0.0756057982070603 35.3627895345405\\
0.0762520016105394 77.0405265527258\\
0.0768982050140185 48.7340904160149\\
0.0775444084174977 12.9232189522213\\
0.0781906118209768 66.7511263658726\\
0.078836815224456 33.5472507546466\\
0.0794830186279351 64.7148642366718\\
0.0801292220314143 41.3634265244727\\
0.0807754254348934 57.7025876404335\\
0.0814216288383726 33.5366412930769\\
0.0820678322418517 17.1275563899525\\
0.0827140356453309 36.2753557981019\\
0.08336023904881 41.570391648658\\
0.0840064424522892 11.8968204446317\\
0.0846526458557683 32.8136439356188\\
0.0852988492592475 29.2560309082797\\
0.0859450526627266 33.5453606265289\\
0.0865912560662058 93.8590313176646\\
0.0872374594696849 36.8911570007644\\
0.0878836628731641 1.58133324198356\\
0.0885298662766432 6.40350398367065\\
0.0891760696801223 39.8961920861412\\
0.0898222730836015 23.747682556327\\
0.0904684764870807 37.6489141177659\\
0.0911146798905598 48.1184469939348\\
0.0917608832940389 31.5766517509756\\
0.0924070866975181 27.9198483415807\\
0.0930532901009972 27.7580788396174\\
0.0936994935044764 59.9606726928961\\
0.0943456969079555 11.9786417418556\\
0.0949919003114347 103.907663524653\\
0.0956381037149138 68.0790810973014\\
0.096284307118393 48.1799443965238\\
0.0969305105218721 42.4034147396202\\
0.0975767139253513 35.7879104234499\\
0.0982229173288304 38.5973195197655\\
0.0988691207323096 21.0567934827231\\
0.0995153241357887 40.5086480584076\\
0.100161527539268 49.288361865756\\
0.100807730942747 25.7416653780076\\
0.101453934346226 7.95996769767753\\
0.102100137749705 22.4085502384315\\
0.102746341153184 46.4738062571742\\
0.103392544556664 66.6161467914305\\
0.104038747960143 18.5683279371259\\
0.104684951363622 23.4738944974646\\
0.105331154767101 51.6655582980453\\
0.10597735817058 14.6627042554675\\
0.106623561574059 15.866252438779\\
0.107269764977538 66.6064206980169\\
0.107915968381018 17.2943348120497\\
0.108562171784497 20.3819534732317\\
0.109208375187976 7.97777806060541\\
0.109854578591455 48.6130059022771\\
0.110500781994934 20.9297096392071\\
0.111146985398413 28.6106603646791\\
0.111793188801893 66.070273197664\\
0.112439392205372 35.9032839556027\\
0.113085595608851 37.105888620155\\
0.11373179901233 5.82694990053627\\
0.114378002415809 57.920386475502\\
0.115024205819288 66.6353762943846\\
0.115670409222767 23.3098798882148\\
0.116316612626247 20.3671529697224\\
0.116962816029726 52.9030676431532\\
0.117609019433205 14.9781430512254\\
0.118255222836684 28.4465868899329\\
0.118901426240163 36.9393873263666\\
0.119547629643642 51.1571194050021\\
0.120193833047121 60.2070041433063\\
0.120840036450601 42.6888950457503\\
0.12148623985408 23.7149734242054\\
0.122132443257559 12.914182383764\\
0.122778646661038 31.2493661360453\\
0.123424850064517 20.0877680055491\\
0.124071053467996 44.0081332921524\\
0.124717256871475 27.2384507271193\\
0.125363460274955 56.6384977212482\\
0.126009663678434 55.5373688904979\\
0.126655867081913 40.0487760961293\\
0.127302070485392 66.3734867824896\\
0.127948273888871 14.8244378187121\\
0.12859447729235 48.6436071644256\\
0.12924068069583 4.58365542342929\\
0.129886884099309 20.6213558997877\\
0.130533087502788 13.8337736266652\\
0.131179290906267 21.5003072073098\\
0.131825494309746 6.99334030453145\\
0.132471697713225 12.9488494461057\\
0.133117901116704 64.5462020824978\\
0.133764104520184 50.6378750197139\\
0.134410307923663 32.3817740075442\\
0.135056511327142 5.29329586469523\\
0.135702714730621 24.3084035408061\\
0.1363489181341 17.9469009677872\\
0.136995121537579 38.4719842300806\\
0.137641324941058 49.9364309590665\\
0.138287528344538 15.315686908334\\
0.138933731748017 39.0778900916068\\
0.139579935151496 69.8401519431895\\
0.140226138554975 2.94690014357764\\
0.140872341958454 19.6450124164664\\
0.141518545361933 42.2196520595858\\
0.142164748765412 21.2748753278463\\
0.142810952168892 22.4034509685466\\
0.143457155572371 15.6491496296777\\
0.14410335897585 24.6032681193573\\
0.144749562379329 3.04297401620423\\
0.145395765782808 25.4003535368433\\
0.146041969186287 22.6610102216373\\
0.146688172589766 55.6084938798781\\
0.147334375993246 29.751779273606\\
0.147980579396725 23.8277667747971\\
0.148626782800204 12.0491148777312\\
0.149272986203683 23.9479029307347\\
0.149919189607162 66.8126358786934\\
0.150565393010641 55.2221227798004\\
0.151211596414121 21.8785728799815\\
0.1518577998176 34.3302749152252\\
0.152504003221079 8.53294843424021\\
0.153150206624558 23.1893764584444\\
0.153796410028037 15.3219976144395\\
0.154442613431516 35.7059034276703\\
0.155088816834995 24.1844681786508\\
0.155735020238475 27.2870361200565\\
0.156381223641954 34.911802091144\\
0.157027427045433 47.0021458963241\\
0.157673630448912 30.2225133369156\\
0.158319833852391 22.5703177035531\\
0.15896603725587 33.176199294044\\
0.159612240659349 23.4863370046448\\
0.160258444062829 5.83126047626622\\
0.160904647466308 72.3982297389553\\
0.161550850869787 55.4065597416643\\
0.162197054273266 19.2084921818135\\
0.162843257676745 27.8010738153362\\
0.163489461080224 29.2362647383157\\
0.164135664483703 31.6480415362032\\
0.164781867887183 18.822446305504\\
0.165428071290662 17.3681320511482\\
0.166074274694141 10.1538961785276\\
0.16672047809762 15.0680142579928\\
0.167366681501099 10.7938637889589\\
0.168012884904578 72.6375093471021\\
0.168659088308058 31.1418516455107\\
0.169305291711537 15.6498939460077\\
0.169951495115016 54.8724682533304\\
0.170597698518495 15.2187626514485\\
0.171243901921974 1.66806901160491\\
0.171890105325453 37.2077745813871\\
0.172536308728932 39.9668468860506\\
0.173182512132412 37.6024641261634\\
0.173828715535891 50.8246181933628\\
0.17447491893937 35.7690826438841\\
0.175121122342849 45.5759797491662\\
0.175767325746328 7.49229264396284\\
0.176413529149807 24.587970463866\\
0.177059732553286 28.0941950558182\\
0.177705935956766 9.98035330685289\\
0.178352139360245 34.2887324481324\\
0.178998342763724 15.9792435207609\\
0.179644546167203 19.724599731318\\
0.180290749570682 31.6703682568636\\
0.180936952974161 5.66524054672839\\
0.18158315637764 56.520647388587\\
0.18222935978112 43.113401670477\\
0.182875563184599 20.6633183076065\\
0.183521766588078 17.3905871172083\\
0.184167969991557 26.9728268410922\\
0.184814173395036 14.7056851796368\\
0.185460376798515 38.2021318738324\\
0.186106580201995 12.4356976443407\\
0.186752783605474 18.2249267955744\\
0.187398987008953 12.7044773496929\\
0.188045190412432 33.8106819211121\\
0.188691393815911 21.2543231331659\\
0.18933759721939 24.9611255979253\\
0.189983800622869 25.3672353426042\\
0.190630004026349 2.37118407776397\\
0.191276207429828 30.5643276984053\\
0.191922410833307 33.7572248165865\\
0.192568614236786 17.8294948060138\\
0.193214817640265 13.0684836142863\\
0.193861021043744 36.905203826918\\
0.194507224447223 28.750324224663\\
0.195153427850703 23.284761176774\\
0.195799631254182 43.2528910406355\\
0.196445834657661 38.8133291155015\\
0.19709203806114 36.3238062449078\\
0.197738241464619 12.8980706164218\\
0.198384444868098 33.9398656767336\\
0.199030648271577 55.7777919861804\\
0.199676851675057 42.4663417368652\\
0.200323055078536 23.7381903387185\\
0.200969258482015 25.2688033676321\\
0.201615461885494 12.7303366523244\\
0.202261665288973 16.781815875548\\
0.202907868692452 56.3943466913141\\
0.203554072095931 39.0339201204293\\
0.204200275499411 19.0393753544074\\
0.20484647890289 21.0369594830897\\
0.205492682306369 31.2788366109296\\
0.206138885709848 19.3178765110486\\
0.206785089113327 22.861409654329\\
0.207431292516806 25.1211803674779\\
0.208077495920285 18.5333672865525\\
0.208723699323765 31.3326311820626\\
0.209369902727244 11.42218619731\\
0.210016106130723 39.0025639689776\\
0.210662309534202 10.4269224683754\\
0.211308512937681 42.8756388511854\\
0.21195471634116 20.7418243372214\\
0.21260091974464 24.0446539048779\\
0.213247123148119 29.1050317543313\\
0.213893326551598 3.93620284300953\\
0.214539529955077 15.033072714923\\
0.215185733358556 20.3328450944893\\
0.215831936762035 13.2879881586694\\
0.216478140165514 10.5600955010422\\
0.217124343568994 5.00951084686471\\
0.217770546972473 20.1682659115377\\
0.218416750375952 13.6442498963525\\
0.219062953779431 21.4466482489544\\
0.21970915718291 22.3391136807428\\
0.220355360586389 21.2354536394232\\
0.221001563989868 15.6505924228065\\
0.221647767393348 30.956004806859\\
0.222293970796827 20.2801982934451\\
0.222940174200306 19.331808358106\\
0.223586377603785 26.4547157469884\\
0.224232581007264 8.57320569451999\\
0.224878784410743 8.95098517911618\\
0.225524987814222 19.0054607198588\\
0.226171191217702 8.77838987862506\\
0.226817394621181 30.2787621639651\\
0.22746359802466 33.57764503448\\
0.228109801428139 31.2795854923886\\
0.228756004831618 2.61092761217469\\
0.229402208235097 15.9466127526344\\
0.230048411638577 13.9061538478725\\
0.230694615042056 13.5932976760035\\
0.231340818445535 4.22795355277937\\
0.231987021849014 33.2538188663569\\
0.232633225252493 33.1675800995312\\
0.233279428655972 27.1303189852985\\
0.233925632059451 13.3033780264429\\
0.234571835462931 10.8967108146918\\
0.23521803886641 28.1649037891307\\
0.235864242269889 7.54666355602159\\
0.236510445673368 27.5891248517962\\
0.237156649076847 14.1325729844618\\
0.237802852480326 28.7554695387058\\
0.238449055883805 31.4876291662157\\
0.239095259287285 19.6406489022269\\
0.239741462690764 10.3043715902295\\
0.240387666094243 15.9875663103187\\
0.241033869497722 5.22282042010907\\
0.241680072901201 14.9107725877228\\
0.24232627630468 22.2139507951686\\
0.242972479708159 25.8087358235951\\
0.243618683111639 13.460752597068\\
0.244264886515118 31.5216246929397\\
0.244911089918597 18.1133783662437\\
0.245557293322076 17.9560169752276\\
0.246203496725555 29.3006476596904\\
0.246849700129034 17.4316284911048\\
0.247495903532513 4.39881392393939\\
0.248142106935993 10.5883416194768\\
0.248788310339472 15.4541195200072\\
0.249434513742951 19.6139537103201\\
0.25008071714643 14.2295949723275\\
0.250726920549909 29.7131950140786\\
0.251373123953388 22.348063408883\\
0.252019327356868 30.4487265579989\\
0.252665530760347 17.9855895885888\\
0.253311734163826 12.4539567213172\\
0.253957937567305 3.24143756484591\\
0.254604140970784 14.1022275529205\\
0.255250344374263 15.4802389840492\\
0.255896547777742 3.71176108898459\\
0.256542751181222 13.3944805112993\\
0.257188954584701 24.119762840194\\
0.25783515798818 16.345980315162\\
0.258481361391659 24.2741965944947\\
0.259127564795138 25.5338869426033\\
0.259773768198617 29.9997765924616\\
0.260419971602096 20.99444329201\\
0.261066175005576 27.5488558058062\\
0.261712378409055 32.6157348029237\\
0.262358581812534 11.9424638332528\\
0.263004785216013 1.27847395435691\\
0.263650988619492 11.9798731857174\\
0.264297192022971 21.6142519769122\\
0.26494339542645 17.7401844214788\\
0.26558959882993 7.10164066585218\\
0.266235802233409 24.023005104695\\
0.266882005636888 7.76401200661601\\
0.267528209040367 31.5609046650801\\
0.268174412443846 15.9843444139518\\
0.268820615847325 22.0056042534704\\
0.269466819250805 8.07229540012102\\
0.270113022654284 19.6982192503965\\
0.270759226057763 9.14038707206829\\
0.271405429461242 19.350489631063\\
0.272051632864721 14.25658583519\\
0.2726978362682 21.0527244953986\\
0.273344039671679 13.7884645233019\\
0.273990243075159 21.7469981311669\\
0.274636446478638 33.5020175388496\\
0.275282649882117 64.1847983484969\\
0.275928853285596 38.8058385691865\\
0.276575056689075 102.152300836733\\
0.277221260092554 86.2188848635464\\
0.277867463496033 44.8495894066407\\
0.278513666899513 40.1381128984751\\
0.279159870302992 15.3865170279695\\
0.279806073706471 13.1773982150331\\
0.28045227710995 9.97823752956308\\
0.281098480513429 13.6534143422694\\
0.281744683916908 9.92233256238014\\
0.282390887320387 8.1001924138674\\
0.283037090723867 8.49946167755614\\
0.283683294127346 2.80532090230647\\
0.284329497530825 7.03920769544504\\
0.284975700934304 7.0356442056612\\
0.285621904337783 6.43229994123095\\
0.286268107741262 1.34499410240978\\
0.286914311144741 5.43652057568257\\
0.287560514548221 7.75398849515015\\
0.2882067179517 8.12590445821246\\
0.288852921355179 5.72204647935453\\
0.289499124758658 6.40023969431479\\
0.290145328162137 6.46885269526795\\
0.290791531565616 4.74014607191548\\
0.291437734969096 6.00949935818282\\
0.292083938372575 0.723028745398592\\
0.292730141776054 1.74217694401178\\
0.293376345179533 5.5911491727363\\
0.294022548583012 2.97734386655823\\
0.294668751986491 2.51955670677844\\
0.29531495538997 2.40574847836251\\
0.29596115879345 6.30790666654393\\
0.296607362196929 0.773252121907391\\
0.297253565600408 2.41798397056025\\
0.297899769003887 3.42359065524912\\
0.298545972407366 5.2264597059145\\
0.299192175810845 2.33873774035207\\
0.299838379214324 4.28080218271764\\
0.300484582617804 2.39147605058721\\
0.301130786021283 0.784161040857012\\
0.301776989424762 0.604941768294626\\
0.302423192828241 2.24031534681957\\
0.30306939623172 4.11898209324475\\
0.303715599635199 2.7101432291554\\
0.304361803038678 1.23001306039081\\
0.305008006442158 3.42815364456735\\
0.305654209845637 2.77875875101364\\
0.306300413249116 3.86858949073223\\
0.306946616652595 6.4377596157103\\
0.307592820056074 2.61582052905318\\
0.308239023459553 2.67685144048991\\
0.308885226863032 1.76777440895826\\
0.309531430266512 1.41134713369204\\
0.310177633669991 1.2288181466747\\
0.31082383707347 3.18208111211773\\
0.311470040476949 2.11903951599809\\
0.312116243880428 4.5884453368535\\
0.312762447283907 3.62036234181549\\
0.313408650687387 2.87053816842625\\
0.314054854090866 1.64245686710701\\
0.314701057494345 3.74230392640875\\
0.315347260897824 4.64070241288236\\
0.315993464301303 1.75518133674633\\
0.316639667704782 1.67709082725667\\
0.317285871108261 6.26830464287276\\
0.317932074511741 2.38109807620976\\
0.31857827791522 4.27699597073927\\
0.319224481318699 0.57983056009231\\
0.319870684722178 2.82193780694599\\
0.320516888125657 3.3359613600163\\
0.321163091529136 4.10676280359607\\
0.321809294932615 2.19670704634518\\
0.322455498336095 1.42555861680191\\
0.323101701739574 2.09542974264646\\
0.323747905143053 6.40746128315056\\
0.324394108546532 5.70012546378519\\
0.325040311950011 4.11065100076674\\
0.32568651535349 3.12238852469721\\
0.326332718756969 4.57994517032718\\
0.326978922160449 3.76713569495717\\
0.327625125563928 2.14011877517866\\
0.328271328967407 1.08480090797293\\
0.328917532370886 2.07973865218035\\
0.329563735774365 3.11682792133889\\
0.330209939177844 4.61350966837754\\
0.330856142581324 3.97801453748926\\
0.331502345984803 2.64539064814043\\
0.332148549388282 4.85832940560649\\
0.332794752791761 4.16092026748554\\
0.33344095619524 2.2183117328781\\
0.334087159598719 1.51285866562881\\
0.334733363002198 3.32100329786939\\
0.335379566405678 3.860298870455\\
0.336025769809157 2.47771823325238\\
0.336671973212636 1.89273358082858\\
0.337318176616115 2.94924361128424\\
0.337964380019594 0.91318251406972\\
0.338610583423073 0.776607302938529\\
0.339256786826552 0.953840986020467\\
0.339902990230032 2.05403745597918\\
0.340549193633511 1.86912433679852\\
0.34119539703699 1.20026303471934\\
0.341841600440469 1.63865145008854\\
0.342487803843948 4.09036021285127\\
0.343134007247427 3.17873480397756\\
0.343780210650906 4.20556990007005\\
0.344426414054386 1.10909726688309\\
0.345072617457865 2.54274050448519\\
0.345718820861344 2.86256343293379\\
0.346365024264823 3.97966822546811\\
0.347011227668302 0.381900908332173\\
0.347657431071781 1.84958148381077\\
0.34830363447526 4.17620785222891\\
0.34894983787874 1.745757022737\\
0.349596041282219 3.83081568041688\\
0.350242244685698 1.37110210139015\\
0.350888448089177 3.20982311203628\\
0.351534651492656 1.19736098676502\\
0.352180854896135 2.78471361041135\\
0.352827058299615 1.29245847068443\\
0.353473261703094 0.896133088487101\\
0.354119465106573 0.779968172371531\\
0.354765668510052 0.481714903464761\\
0.355411871913531 1.59394336231525\\
0.35605807531701 2.92961936746273\\
0.356704278720489 1.16207032342085\\
0.357350482123969 2.46164468683572\\
0.357996685527448 2.17564589545996\\
0.358642888930927 1.53503510898959\\
0.359289092334406 0.710537551488277\\
0.359935295737885 1.40472996200602\\
0.360581499141364 1.79280026527793\\
0.361227702544843 2.23054688424385\\
0.361873905948323 3.73234896050265\\
0.362520109351802 1.07871602865483\\
0.363166312755281 1.45020214323589\\
0.36381251615876 3.20242756290235\\
0.364458719562239 0.718110828383396\\
0.365104922965718 0.685520307508226\\
0.365751126369197 2.73923854966434\\
0.366397329772677 3.45044078792006\\
0.367043533176156 1.9019020483759\\
0.367689736579635 2.84137027280012\\
0.368335939983114 3.9780430502866\\
0.368982143386593 1.41423904830173\\
0.369628346790072 2.97867825917381\\
0.370274550193551 1.84487003898548\\
0.370920753597031 3.73505505852023\\
0.37156695700051 1.98624716686153\\
0.372213160403989 1.30644585334299\\
0.372859363807468 2.51183702310719\\
0.373505567210947 0.650120049040709\\
0.374151770614426 3.6148995828899\\
0.374797974017906 3.0320598828799\\
0.375444177421385 1.94958299643526\\
0.376090380824864 1.61977736865118\\
0.376736584228343 2.29045683678074\\
0.377382787631822 1.3209937053936\\
0.378028991035301 0.73151412163637\\
0.37867519443878 0.415235248848032\\
0.37932139784226 2.10837149947038\\
0.379967601245739 0.927611255399544\\
0.380613804649218 0.99224532376331\\
0.381260008052697 2.68741188174295\\
0.381906211456176 1.17316621608038\\
0.382552414859655 1.29063740762524\\
0.383198618263134 2.81939509862373\\
0.383844821666614 0.706028998575551\\
0.384491025070093 2.24045105987657\\
0.385137228473572 1.84328834525495\\
0.385783431877051 0.293625129940708\\
0.38642963528053 1.92599052295471\\
0.387075838684009 1.66757935306222\\
0.387722042087488 2.01817598341173\\
0.388368245490968 2.77247914045988\\
0.389014448894447 1.84534999571735\\
0.389660652297926 2.1415647611125\\
0.390306855701405 0.536150467666081\\
0.390953059104884 1.58132439104905\\
0.391599262508363 0.972859340646187\\
0.392245465911843 1.82950315109152\\
0.392891669315322 1.52022815437604\\
0.393537872718801 2.91227812046959\\
0.39418407612228 3.19951356307793\\
0.394830279525759 2.48711411287266\\
0.395476482929238 2.59842562970353\\
0.396122686332717 2.06783361674067\\
0.396768889736196 0.766525451516706\\
0.397415093139676 2.30077734878641\\
0.398061296543155 2.85989243646692\\
0.398707499946634 2.01715152143998\\
0.399353703350113 2.2747059676948\\
0.399999906753592 0.752499948200274\\
0.400646110157071 2.18209306478331\\
0.401292313560551 1.70115488444248\\
0.40193851696403 1.93574397058543\\
0.402584720367509 2.29574832240185\\
0.403230923770988 2.36133521914577\\
0.403877127174467 0.971733371845739\\
0.404523330577946 0.842535088141425\\
0.405169533981426 1.79515822750211\\
0.405815737384905 2.88654923170636\\
0.406461940788384 1.15816034576768\\
0.407108144191863 1.82181234170002\\
0.407754347595342 1.34598425657833\\
0.408400550998821 1.61346777151741\\
0.4090467544023 2.22798926442602\\
0.409692957805779 3.11614645136565\\
0.410339161209259 2.62363493224012\\
0.410985364612738 1.58217211946348\\
0.411631568016217 1.83324825371901\\
0.412277771419696 2.52971984245997\\
0.412923974823175 1.5486218644527\\
0.413570178226654 1.94936251767214\\
0.414216381630134 2.35961781737159\\
0.414862585033613 0.59538826232491\\
0.415508788437092 2.86446650842901\\
0.416154991840571 2.29741786657388\\
0.41680119524405 0.73813770760522\\
0.417447398647529 1.89498642280579\\
0.418093602051008 2.05851822272293\\
0.418739805454488 1.41311528944752\\
0.419386008857967 1.36360088956604\\
0.420032212261446 0.85540264341506\\
0.420678415664925 0.975813578858066\\
0.421324619068404 2.12960553844031\\
0.421970822471883 0.661453816993101\\
0.422617025875362 0.453657842759743\\
0.423263229278842 0.781538028273857\\
0.423909432682321 1.92280220042196\\
0.4245556360858 2.40882879856546\\
0.425201839489279 2.27683447432612\\
0.425848042892758 1.59361994801499\\
0.426494246296237 2.39602294512577\\
0.427140449699717 2.0236856272042\\
0.427786653103196 1.66048968780938\\
0.428432856506675 0.36773851257145\\
0.429079059910154 0.412542270141757\\
0.429725263313633 1.95608155230356\\
0.430371466717112 2.35041699039958\\
0.431017670120591 1.56591886357193\\
0.431663873524071 1.21237744790214\\
0.43231007692755 0.659528050050535\\
0.432956280331029 1.32933535344398\\
0.433602483734508 0.841741102845678\\
0.434248687137987 0.969200657969536\\
0.434894890541466 1.73447959823541\\
0.435541093944945 0.776928746249273\\
0.436187297348425 1.3051087212516\\
0.436833500751904 1.53477499512826\\
0.437479704155383 2.24981463928609\\
0.438125907558862 1.7956781188346\\
0.438772110962341 2.22203140633091\\
0.43941831436582 2.7723930697466\\
0.4400645177693 1.67632052180066\\
0.440710721172779 1.04807373385699\\
0.441356924576258 0.705044485264331\\
0.442003127979737 0.750974862575129\\
0.442649331383216 0.646866227234279\\
0.443295534786695 1.61866937399616\\
0.443941738190174 1.57097632256542\\
0.444587941593654 1.71201881936843\\
0.445234144997133 1.84355222775346\\
0.445880348400612 1.09928721061759\\
0.446526551804091 1.1547963749545\\
0.44717275520757 2.2619415920165\\
0.447818958611049 2.33502400252603\\
0.448465162014528 1.99565447185396\\
0.449111365418007 0.845006581965894\\
0.449757568821487 1.03684873980363\\
0.450403772224966 1.38595645207921\\
0.451049975628445 0.302087558182473\\
0.451696179031924 0.972218664978554\\
0.452342382435403 1.36213239780515\\
0.452988585838882 1.546053478586\\
0.453634789242362 1.02121808446356\\
0.454280992645841 1.75574089156512\\
0.45492719604932 0.566680307180413\\
0.455573399452799 0.976321553655513\\
0.456219602856278 1.26524529932198\\
0.456865806259757 1.57404455459192\\
0.457512009663236 0.69715310552746\\
0.458158213066716 0.597985402069265\\
0.458804416470195 0.894888726178129\\
0.459450619873674 0.610384196864278\\
0.460096823277153 0.499241537512633\\
0.460743026680632 1.25360662980373\\
0.461389230084111 0.518296499433098\\
0.46203543348759 1.44988708626737\\
0.46268163689107 1.11199134616313\\
0.463327840294549 2.25668148132397\\
0.463974043698028 0.532202415050683\\
0.464620247101507 1.31279710922593\\
0.465266450504986 1.64582598845387\\
0.465912653908465 1.35302905665549\\
0.466558857311945 1.26111510028193\\
0.467205060715424 0.928123187033963\\
0.467851264118903 0.571704893536023\\
0.468497467522382 0.725538458818733\\
0.469143670925861 1.89423624648825\\
0.46978987432934 1.85701710331416\\
0.470436077732819 2.08269013296239\\
0.471082281136298 2.31196403338988\\
0.471728484539778 1.12481404000712\\
0.472374687943257 1.36735285496924\\
0.473020891346736 0.690452476762982\\
0.473667094750215 1.21294426926427\\
0.474313298153694 1.37946762779772\\
0.474959501557173 0.751513954176695\\
0.475605704960653 1.15508950827132\\
0.476251908364132 0.76314325555597\\
0.476898111767611 0.158913618566118\\
0.47754431517109 0.60074320152601\\
0.478190518574569 1.00953089902641\\
0.478836721978048 2.04846073564674\\
0.479482925381527 1.96854570544475\\
0.480129128785007 1.8077225571795\\
0.480775332188486 1.06963897928172\\
0.481421535591965 0.821404853270594\\
0.482067738995444 0.819483832350558\\
0.482713942398923 0.138337859384226\\
0.483360145802402 0.652651453804438\\
0.484006349205881 0.413122909783976\\
0.484652552609361 1.2472398744584\\
0.48529875601284 1.1195588895167\\
0.485944959416319 1.08815575053613\\
0.486591162819798 0.663631367424769\\
0.487237366223277 1.54264956591874\\
0.487883569626756 1.49248506437091\\
0.488529773030236 1.65928013612616\\
0.489175976433715 1.76262834706856\\
0.489822179837194 0.418917332132204\\
0.490468383240673 0.977955378752194\\
0.491114586644152 1.32739700554426\\
0.491760790047631 0.506971053756864\\
0.49240699345111 0.714390060774613\\
0.49305319685459 0.671389331977566\\
0.493699400258069 0.921903356309492\\
0.494345603661548 1.26513593258554\\
0.494991807065027 1.29261112671189\\
0.495638010468506 0.736756875861268\\
0.496284213871985 1.15671826070134\\
0.496930417275464 0.812582838426367\\
0.497576620678943 1.95722060117285\\
0.498222824082423 0.311001364084641\\
0.498869027485902 0.527575095956734\\
0.499515230889381 1.22982336417501\\
0.50016143429286 1.12218615946259\\
0.500807637696339 1.03563243820151\\
0.501453841099818 0.254107155678854\\
0.502100044503298 0.843158362088244\\
0.502746247906777 0.661602544547667\\
0.503392451310256 1.06013392116883\\
0.504038654713735 1.38824768827587\\
0.504684858117214 0.537576450355432\\
0.505331061520693 0.605140035895602\\
0.505977264924172 0.847087791351354\\
0.506623468327652 0.658075051868123\\
0.507269671731131 1.35749678852935\\
0.50791587513461 1.21718120128038\\
0.508562078538089 0.583614567883485\\
0.509208281941568 1.07998563703401\\
0.509854485345047 0.84875036448577\\
0.510500688748527 0.747445739871242\\
0.511146892152006 0.973030596345655\\
0.511793095555485 1.28476590470711\\
0.512439298958964 1.57911840161405\\
0.513085502362443 0.775973959880471\\
0.513731705765922 1.25249501683307\\
0.514377909169401 1.12343498836841\\
0.515024112572881 0.222266581857365\\
0.51567031597636 0.621429802972263\\
0.516316519379839 0.998247916184011\\
0.516962722783318 1.74172118078166\\
0.517608926186797 1.27512466474228\\
0.518255129590276 1.43219859702314\\
0.518901332993755 1.14169433486242\\
0.519547536397235 0.407840284001937\\
0.520193739800714 1.61134515535338\\
0.520839943204193 1.1074309799798\\
0.521486146607672 1.18451589423023\\
0.522132350011151 0.825138079363482\\
0.52277855341463 1.28275925208529\\
0.523424756818109 1.1064750624318\\
0.524070960221589 0.737273435217458\\
0.524717163625068 0.913947199005302\\
0.525363367028547 0.849629130539014\\
0.526009570432026 1.43038533368272\\
0.526655773835505 0.788174740075481\\
0.527301977238984 0.528358768285364\\
0.527948180642464 0.449360831079533\\
0.528594384045943 0.749870099485466\\
0.529240587449422 1.29682153993769\\
0.529886790852901 0.9301420064413\\
0.53053299425638 0.480865390292153\\
0.531179197659859 1.24004167441147\\
0.531825401063338 1.2898996797754\\
0.532471604466818 0.589402834733128\\
0.533117807870297 0.724242744376988\\
0.533764011273776 0.729040181138189\\
0.534410214677255 0.78795399667026\\
0.535056418080734 1.38537440578016\\
0.535702621484213 0.703767299990641\\
0.536348824887692 0.28511840047476\\
0.536995028291172 0.325891071330874\\
0.537641231694651 1.59407648385203\\
0.53828743509813 1.58281796467354\\
0.538933638501609 0.970721685858028\\
0.539579841905088 0.907945160281146\\
0.540226045308567 1.03818433131126\\
0.540872248712047 1.46055981906705\\
0.541518452115526 0.769865840292542\\
0.542164655519005 0.998826338991049\\
0.542810858922484 0.85386504260856\\
0.543457062325963 0.856296828222158\\
0.544103265729442 1.25318890255271\\
0.544749469132921 0.597690989512483\\
0.5453956725364 0.626392346623642\\
0.54604187593988 0.351743680305397\\
0.546688079343359 0.681400488514345\\
0.547334282746838 0.363245731825461\\
0.547980486150317 0.838498986822734\\
0.548626689553796 1.17438689639528\\
0.549272892957275 1.3200135654037\\
0.549919096360755 0.892203954782212\\
0.550565299764234 0.657789183972127\\
0.551211503167713 1.17919730053998\\
0.551857706571192 0.188458606646279\\
0.552503909974671 0.92043644181222\\
0.55315011337815 1.34063530031744\\
0.553796316781629 0.949254834434578\\
0.554442520185109 0.479577982591556\\
0.555088723588588 0.48382095898566\\
0.555734926992067 0.748278112911567\\
0.556381130395546 0.803299481234299\\
0.557027333799025 0.614141192860145\\
0.557673537202504 0.766345539401273\\
0.558319740605983 0.623494888904786\\
0.558965944009463 0.528950429665798\\
0.559612147412942 1.00322856750449\\
0.560258350816421 0.67976898960748\\
0.5609045542199 0.806115689992094\\
0.561550757623379 0.529993555997453\\
0.562196961026858 0.693379427668648\\
0.562843164430337 0.917033629250204\\
0.563489367833817 0.938274841142488\\
0.564135571237296 0.777244538286135\\
0.564781774640775 0.648144789248364\\
0.565427978044254 0.737035398554472\\
0.566074181447733 0.80124474038549\\
0.566720384851212 0.702647788050921\\
};
\addlegendentry{Suspended};
\addplot [
color=green!50!black,
solid,
line width=0.6pt,
mark size=0.5pt,
mark=o,
mark options={solid}
]
table[row sep=crcr,x expr=\thisrow{x}*10]{
x y \\
0 0.80049501900913\\
0.000644670379887604 23.7436780588453\\
0.00128934075977521 6.27801145379132\\
0.00193401113966281 24.6121973610882\\
0.00257868151955042 20.2780013984331\\
0.00322335189943802 12.1931435858617\\
0.00386802227932563 21.2942935130416\\
0.00451269265921323 2.90002028355982\\
0.00515736303910083 12.5389077380775\\
0.00580203341898844 11.925877327049\\
0.00644670379887604 6.17999037969161\\
0.00709137417876365 3.08906904848315\\
0.00773604455865125 15.583872210052\\
0.00838071493853886 23.3738616078201\\
0.00902538531842646 107.889906262999\\
0.00967005569831406 28.7980550145895\\
0.0103147260782017 3.18482144912548\\
0.0109593964580893 13.3121518611209\\
0.0116040668379769 18.0652190935665\\
0.0122487372178645 12.8458857397283\\
0.0128934075977521 6.77617737678047\\
0.0135380779776397 16.3404181499597\\
0.0141827483575273 43.6517011573123\\
0.0148274187374149 8.22743312458415\\
0.0154720891173025 16.1827465113516\\
0.0161167594971901 9.50938542200359\\
0.0167614298770777 13.8964609591377\\
0.0174061002569653 10.4291602981113\\
0.0180507706368529 6.21802916450032\\
0.0186954410167405 1.39558958286694\\
0.0193401113966281 9.32179547149372\\
0.0199847817765157 8.43226928510773\\
0.0206294521564033 9.33520931265123\\
0.0212741225362909 6.79113666002204\\
0.0219187929161785 9.75237953584085\\
0.0225634632960661 12.7206378022742\\
0.0232081336759538 11.7620506839204\\
0.0238528040558414 12.8957449492325\\
0.024497474435729 0.65734405304918\\
0.0251421448156166 17.1601131480768\\
0.0257868151955042 10.6016189186961\\
0.0264314855753918 3.31120330970056\\
0.0270761559552794 4.78165925140407\\
0.027720826335167 4.60039262455616\\
0.0283654967150546 15.0980161555845\\
0.0290101670949422 14.2823534664556\\
0.0296548374748298 7.29444794732773\\
0.0302995078547174 10.2505692214109\\
0.030944178234605 3.21217024982688\\
0.0315888486144926 5.0637843324699\\
0.0322335189943802 2.54453640656838\\
0.0328781893742678 3.9033530665322\\
0.0335228597541554 16.4199575522844\\
0.034167530134043 5.58329611844672\\
0.0348122005139306 6.90840308127561\\
0.0354568708938182 1.38314678558135\\
0.0361015412737058 11.694358106231\\
0.0367462116535934 8.30363185704472\\
0.037390882033481 16.7458248695752\\
0.0380355524133686 4.67872239213263\\
0.0386802227932563 8.37510542939894\\
0.0393248931731439 12.1340936355018\\
0.0399695635530315 11.7096568539221\\
0.0406142339329191 14.7094593445341\\
0.0412589043128067 15.6797792460922\\
0.0419035746926943 14.2873855111827\\
0.0425482450725819 5.09863314760292\\
0.0431929154524695 1.29408349917552\\
0.0438375858323571 2.17692813279677\\
0.0444822562122447 8.47282242990871\\
0.0451269265921323 4.85653527511777\\
0.0457715969720199 3.40814818553406\\
0.0464162673519075 14.108065067783\\
0.0470609377317951 10.9370543284092\\
0.0477056081116827 6.76769594692056\\
0.0483502784915703 8.50719727068273\\
0.0489949488714579 12.8022795891106\\
0.0496396192513455 3.10866251074692\\
0.0502842896312331 9.46091619385574\\
0.0509289600111207 6.95067953075929\\
0.0515736303910083 6.62817468318981\\
0.0522183007708959 2.48042274319563\\
0.0528629711507835 11.4974618965585\\
0.0535076415306711 1.5653228090555\\
0.0541523119105587 0.471760550977192\\
0.0547969822904464 6.78083733953614\\
0.055441652670334 11.3133483010858\\
0.0560863230502216 2.20336878483607\\
0.0567309934301092 4.68347235585658\\
0.0573756638099968 12.3912188328867\\
0.0580203341898844 7.82318982877595\\
0.058665004569772 8.15074636959897\\
0.0593096749496596 5.14126272394371\\
0.0599543453295472 7.80208199410477\\
0.0605990157094348 7.5589954165185\\
0.0612436860893224 4.18292464245747\\
0.06188835646921 2.88990313206542\\
0.0625330268490976 3.81197351884164\\
0.0631776972289852 7.30608848675274\\
0.0638223676088728 4.05707217913864\\
0.0644670379887604 17.4774096084504\\
0.065111708368648 6.95356463017387\\
0.0657563787485356 9.60751819000019\\
0.0664010491284232 3.96176917219752\\
0.0670457195083108 6.79381270486478\\
0.0676903898881984 11.303655691409\\
0.068335060268086 11.0491662202188\\
0.0689797306479737 8.76607548675628\\
0.0696244010278612 4.17921204633183\\
0.0702690714077489 12.9165149884062\\
0.0709137417876365 5.06423292487987\\
0.0715584121675241 5.43737307360682\\
0.0722030825474117 9.01649204229286\\
0.0728477529272993 5.05508785384592\\
0.0734924233071869 12.5094153634526\\
0.0741370936870745 10.901999517398\\
0.0747817640669621 14.8997948791562\\
0.0754264344468497 9.38637714089897\\
0.0760711048267373 11.0295903339827\\
0.0767157752066249 4.33425070098124\\
0.0773604455865125 0.627526116877257\\
0.0780051159664001 18.9995327648491\\
0.0786497863462877 11.5125936174139\\
0.0792944567261753 8.29442233475901\\
0.0799391271060629 4.17764835424419\\
0.0805837974859505 2.79514195199007\\
0.0812284678658381 6.59988745036828\\
0.0818731382457257 6.68041323641456\\
0.0825178086256133 12.4694217839956\\
0.0831624790055009 4.0266244824948\\
0.0838071493853885 14.4613989351269\\
0.0844518197652761 20.6310406240888\\
0.0850964901451638 8.01086996844233\\
0.0857411605250513 3.88730462578998\\
0.086385830904939 10.6865939369468\\
0.0870305012848266 8.89926609696144\\
0.0876751716647142 6.46260253126807\\
0.0883198420446018 16.2604320669792\\
0.0889645124244894 6.72004992246644\\
0.089609182804377 6.51521476573058\\
0.0902538531842646 0.536476195906258\\
0.0908985235641522 9.13508876625975\\
0.0915431939440398 6.48163249405125\\
0.0921878643239274 5.96987887082272\\
0.092832534703815 6.46017699989922\\
0.0934772050837026 7.00692139047183\\
0.0941218754635902 7.24252216600348\\
0.0947665458434778 2.50255817193181\\
0.0954112162233654 6.73592988917748\\
0.096055886603253 13.3062954108859\\
0.0967005569831406 15.4407053905733\\
0.0973452273630282 8.89572836217222\\
0.0979898977429158 9.17903802475463\\
0.0986345681228035 6.66656131040683\\
0.099279238502691 6.16777591147155\\
0.0999239088825786 20.5008872947815\\
0.100568579262466 8.1978607836509\\
0.101213249642354 7.21377023797832\\
0.101857920022241 8.95488321354492\\
0.102502590402129 5.39668654926914\\
0.103147260782017 5.3387499895777\\
0.103791931161904 11.6683256383218\\
0.104436601541792 9.66794816405518\\
0.105081271921679 12.9084441180404\\
0.105725942301567 13.2631157003664\\
0.106370612681455 8.70105083667166\\
0.107015283061342 7.134989866031\\
0.10765995344123 9.21188853824458\\
0.108304623821118 3.15515955295109\\
0.108949294201005 9.92237615057354\\
0.109593964580893 11.884076559722\\
0.11023863496078 11.1537963418336\\
0.110883305340668 0.0633688821388846\\
0.111527975720556 2.09572120949572\\
0.112172646100443 13.6783199092661\\
0.112817316480331 13.9243391431845\\
0.113461986860218 3.74450871356587\\
0.114106657240106 10.2144628719702\\
0.114751327619994 5.94863593610563\\
0.115395997999881 7.78760810917485\\
0.116040668379769 10.1600342984115\\
0.116685338759656 7.1433533986035\\
0.117330009139544 11.7998243974731\\
0.117974679519432 3.16861917861329\\
0.118619349899319 7.56120042541353\\
0.119264020279207 5.42144874056954\\
0.119908690659094 11.8893580552381\\
0.120553361038982 7.73458138679085\\
0.12119803141887 2.31079569787104\\
0.121842701798757 0.635889230945006\\
0.122487372178645 4.66148125232187\\
0.123132042558532 4.93565716314339\\
0.12377671293842 8.18383730547116\\
0.124421383318308 7.67558245722631\\
0.125066053698195 4.41889287126404\\
0.125710724078083 10.1608988526423\\
0.12635539445797 4.43469761880013\\
0.127000064837858 13.1136596720078\\
0.127644735217746 5.30059477447585\\
0.128289405597633 6.84753385774453\\
0.128934075977521 10.9893041831921\\
0.129578746357408 6.53817771536159\\
0.130223416737296 7.95753189220099\\
0.130868087117184 0.708861810631658\\
0.131512757497071 2.12171732493322\\
0.132157427876959 8.55816291113064\\
0.132802098256846 4.11503703358091\\
0.133446768636734 8.50442811368868\\
0.134091439016622 4.22562346017925\\
0.134736109396509 9.55244220575661\\
0.135380779776397 3.24467426082691\\
0.136025450156284 2.30506677238583\\
0.136670120536172 10.7168033223017\\
0.13731479091606 12.0593482905896\\
0.137959461295947 7.93471597726382\\
0.138604131675835 7.15552832160477\\
0.139248802055722 7.22971822751015\\
0.13989347243561 3.73370930460459\\
0.140538142815498 5.47838538035819\\
0.141182813195385 6.43915319804855\\
0.141827483575273 8.3180810968316\\
0.142472153955161 4.8630397507521\\
0.143116824335048 2.57953453780212\\
0.143761494714936 5.27791239645357\\
0.144406165094823 1.75430898281819\\
0.145050835474711 1.77917644871607\\
0.145695505854599 12.7480100728091\\
0.146340176234486 4.97837543178456\\
0.146984846614374 5.39136764559738\\
0.147629516994261 11.3479338758678\\
0.148274187374149 9.81266241987485\\
0.148918857754037 7.79547852717349\\
0.149563528133924 7.34523752797419\\
0.150208198513812 8.57014998769841\\
0.150852868893699 2.28987915334588\\
0.151497539273587 4.7028585280468\\
0.152142209653475 6.1976152743542\\
0.152786880033362 3.04589702991115\\
0.15343155041325 0.610451448028173\\
0.154076220793137 6.72983487118173\\
0.154720891173025 9.95669935593555\\
0.155365561552913 15.0344277457924\\
0.1560102319328 9.36610098385743\\
0.156654902312688 4.06242732222173\\
0.157299572692575 2.02977604076901\\
0.157944243072463 10.6670508276547\\
0.158588913452351 7.45669941575807\\
0.159233583832238 3.54249116483051\\
0.159878254212126 7.20795171004468\\
0.160522924592013 4.60846142607973\\
0.161167594971901 5.64712204909584\\
0.161812265351789 7.76655501707502\\
0.162456935731676 12.5136796672888\\
0.163101606111564 5.62009899078621\\
0.163746276491451 7.9645531069955\\
0.164390946871339 5.47000627466999\\
0.165035617251227 1.13251719290041\\
0.165680287631114 10.1602179800731\\
0.166324958011002 3.43517999683876\\
0.16696962839089 6.66977354359397\\
0.167614298770777 7.51683341095842\\
0.168258969150665 10.6721519207065\\
0.168903639530552 11.9681887768744\\
0.16954830991044 2.06328237206813\\
0.170192980290328 6.24075655517712\\
0.170837650670215 7.74267743924415\\
0.171482321050103 7.05884533552058\\
0.17212699142999 3.53773412005827\\
0.172771661809878 10.0851494273271\\
0.173416332189766 9.18084132534284\\
0.174061002569653 2.25003709520827\\
0.174705672949541 2.1715098506823\\
0.175350343329428 13.6715244145977\\
0.175995013709316 10.921990524947\\
0.176639684089204 13.3758296506939\\
0.177284354469091 1.19283111445569\\
0.177929024848979 3.99716909079\\
0.178573695228866 13.5872368815098\\
0.179218365608754 1.47594941065567\\
0.179863035988642 8.04258744483703\\
0.180507706368529 9.94540594119959\\
0.181152376748417 5.18594857254091\\
0.181797047128304 2.62450965873275\\
0.182441717508192 3.53662498368821\\
0.18308638788808 5.94880432453865\\
0.183731058267967 12.8248721987972\\
0.184375728647855 10.162203859539\\
0.185020399027742 5.61270851786008\\
0.18566506940763 5.55604861044592\\
0.186309739787518 1.92918455292211\\
0.186954410167405 6.37560901230465\\
0.187599080547293 6.98738963432164\\
0.18824375092718 5.40637384605267\\
0.188888421307068 2.28169424661466\\
0.189533091686956 14.0812980786093\\
0.190177762066843 3.8276584168243\\
0.190822432446731 5.83196972439776\\
0.191467102826618 7.6731612216808\\
0.192111773206506 4.63045869747202\\
0.192756443586394 11.014054050295\\
0.193401113966281 2.56654119504493\\
0.194045784346169 9.15898737100869\\
0.194690454726056 1.71689386631132\\
0.195335125105944 11.9771381468807\\
0.195979795485832 5.94881275306222\\
0.196624465865719 9.31394491098237\\
0.197269136245607 4.56952904614015\\
0.197913806625494 8.715462410098\\
0.198558477005382 4.14947680445222\\
0.19920314738527 2.83287871625471\\
0.199847817765157 3.01844274689321\\
0.200492488145045 8.27336365707421\\
0.201137158524932 7.00873992720853\\
0.20178182890482 2.3443206581307\\
0.202426499284708 2.0935844809493\\
0.203071169664595 7.16500015998182\\
0.203715840044483 10.4405215847428\\
0.204360510424371 8.10518413867236\\
0.205005180804258 0.629508134866132\\
0.205649851184146 1.38271169752317\\
0.206294521564033 4.096648183146\\
0.206939191943921 11.3750864839138\\
0.207583862323809 2.75449882272253\\
0.208228532703696 7.76479787660261\\
0.208873203083584 8.09124893808591\\
0.209517873463471 4.76508119279426\\
0.210162543843359 3.71922987002435\\
0.210807214223247 7.22754584531863\\
0.211451884603134 2.05861834521282\\
0.212096554983022 2.48245028181132\\
0.212741225362909 6.62817797273017\\
0.213385895742797 5.76169191636522\\
0.214030566122685 7.05737697982741\\
0.214675236502572 3.62762834873607\\
0.21531990688246 7.11685217798311\\
0.215964577262347 3.63373513878754\\
0.216609247642235 6.18188710905355\\
0.217253918022123 3.70762245349794\\
0.21789858840201 9.1454425778596\\
0.218543258781898 3.70694570135466\\
0.219187929161785 4.65222776956634\\
0.219832599541673 5.45466172027854\\
0.220477269921561 4.34005332086976\\
0.221121940301448 5.55824148383354\\
0.221766610681336 9.34253073104169\\
0.222411281061223 8.79696152575099\\
0.223055951441111 6.09394815982482\\
0.223700621820999 4.41029587174214\\
0.224345292200886 9.23533743650115\\
0.224989962580774 7.31461712311531\\
0.225634632960661 6.81849578963141\\
0.226279303340549 9.68587054680793\\
0.226923973720437 3.25372593222813\\
0.227568644100324 6.77200962637797\\
0.228213314480212 6.15138994847177\\
0.228857984860099 0.823201502174101\\
0.229502655239987 3.03481522612997\\
0.230147325619875 8.30024581558839\\
0.230791995999762 3.85123910574203\\
0.23143666637965 9.93421084404825\\
0.232081336759538 3.89283108748871\\
0.232726007139425 6.02144920841027\\
0.233370677519313 8.18989607703603\\
0.2340153478992 2.63519284938621\\
0.234660018279088 5.40319778060529\\
0.235304688658976 8.66350999981472\\
0.235949359038863 2.22861739160277\\
0.236594029418751 8.58070575826595\\
0.237238699798638 8.58041562412159\\
0.237883370178526 10.9896759676439\\
0.238528040558414 6.46310791456334\\
0.239172710938301 3.15394884401769\\
0.239817381318189 9.50842605353114\\
0.240462051698076 8.01232543247287\\
0.241106722077964 7.45042321456648\\
0.241751392457852 6.80578404635548\\
0.242396062837739 13.3035394992917\\
0.243040733217627 5.40861002733829\\
0.243685403597514 6.0470105870582\\
0.244330073977402 5.10650063417769\\
0.24497474435729 5.76820561244042\\
0.245619414737177 2.73604513660263\\
0.246264085117065 6.5359578404003\\
0.246908755496952 14.7088045605093\\
0.24755342587684 2.77894304765259\\
0.248198096256728 5.50149503234249\\
0.248842766636615 3.09710696655233\\
0.249487437016503 7.45552353466997\\
0.25013210739639 4.69889184365449\\
0.250776777776278 12.0473372603615\\
0.251421448156166 9.10950888305282\\
0.252066118536053 4.1530666281361\\
0.252710788915941 6.49896542846075\\
0.253355459295828 8.1979322000024\\
0.254000129675716 3.51582472809504\\
0.254644800055604 6.05206595173544\\
0.255289470435491 9.69898747878693\\
0.255934140815379 7.5353916308165\\
0.256578811195266 8.8271297334282\\
0.257223481575154 4.60828018465914\\
0.257868151955042 4.92541591950439\\
0.258512822334929 7.25822005352505\\
0.259157492714817 6.35124991760432\\
0.259802163094705 4.90538055555266\\
0.260446833474592 9.10946810530522\\
0.26109150385448 0.745966913483809\\
0.261736174234367 9.24412271555748\\
0.262380844614255 4.48761440246527\\
0.263025514994143 7.32213583305411\\
0.26367018537403 9.46508575179408\\
0.264314855753918 7.94347321262591\\
0.264959526133805 7.64387292629628\\
0.265604196513693 7.65895392430392\\
0.266248866893581 6.31958282133975\\
0.266893537273468 4.98753561441151\\
0.267538207653356 6.64791141573189\\
0.268182878033243 4.6844382433945\\
0.268827548413131 5.29138121574427\\
0.269472218793019 4.8118516044263\\
0.270116889172906 13.0323815897757\\
0.270761559552794 18.4541761136808\\
0.271406229932681 20.6956072619394\\
0.272050900312569 20.8957646956228\\
0.272695570692457 38.1695468129519\\
0.273340241072344 20.8499736228562\\
0.273984911452232 23.2646778674372\\
0.274629581832119 10.4337936590863\\
0.275274252212007 7.90509869327542\\
0.275918922591895 6.92426565134766\\
0.276563592971782 5.42934619703491\\
0.27720826335167 3.7007535997841\\
0.277852933731557 2.98668588221665\\
0.278497604111445 2.12100676359357\\
0.279142274491333 2.82440987938231\\
0.27978694487122 2.90372405955639\\
0.280431615251108 2.81668976047752\\
0.281076285630995 1.41324598273987\\
0.281720956010883 2.7431771168658\\
0.282365626390771 1.04180706883456\\
0.283010296770658 1.04482426222762\\
0.283654967150546 1.54079304289686\\
0.284299637530433 2.07864701725319\\
0.284944307910321 0.464583812336767\\
0.285588978290209 0.406160311469414\\
0.286233648670096 0.461096109673804\\
0.286878319049984 0.699474424297162\\
0.287522989429871 1.53266313761356\\
0.288167659809759 0.853612426622153\\
0.288812330189647 0.693425666757923\\
0.289457000569534 1.22017948282543\\
0.290101670949422 0.884604498128507\\
0.290746341329309 1.03640875509901\\
0.291391011709197 1.00183819755305\\
0.292035682089085 1.76657441971913\\
0.292680352468972 0.998918577709992\\
0.29332502284886 1.25697045577276\\
0.293969693228748 1.17058396538156\\
0.294614363608635 0.738889169075242\\
0.295259033988523 1.82154144559774\\
0.29590370436841 0.546121602902056\\
0.296548374748298 0.893158988901443\\
0.297193045128186 0.592922659512263\\
0.297837715508073 0.928635933508363\\
0.298482385887961 1.01986226618377\\
0.299127056267848 1.40937236480226\\
0.299771726647736 0.198047215972309\\
0.300416397027624 0.298959998037797\\
0.301061067407511 0.243090722897482\\
0.301705737787399 0.860155761683604\\
0.302350408167286 1.5811589164101\\
0.302995078547174 1.39761509678547\\
0.303639748927062 1.42152657727328\\
0.304284419306949 0.542173187330023\\
0.304929089686837 0.867800696532369\\
0.305573760066724 0.931079385919458\\
0.306218430446612 0.955055392056504\\
0.3068631008265 1.2801854332621\\
0.307507771206387 0.952263247376982\\
0.308152441586275 1.46026386547671\\
0.308797111966162 0.707923688254081\\
0.30944178234605 1.1136504138521\\
0.310086452725938 0.758439481579991\\
0.310731123105825 0.119480242077851\\
0.311375793485713 1.07339403286304\\
0.3120204638656 0.602191963039999\\
0.312665134245488 0.283963284873001\\
0.313309804625376 1.17745295791165\\
0.313954475005263 0.75426217775074\\
0.314599145385151 1.41205222645771\\
0.315243815765038 0.348405617383719\\
0.315888486144926 0.676357035459768\\
0.316533156524814 1.98717199314953\\
0.317177826904701 0.784400842890882\\
0.317822497284589 1.22574727330574\\
0.318467167664476 0.82998228023392\\
0.319111838044364 0.568777460732759\\
0.319756508424252 0.744620498226406\\
0.320401178804139 0.404973535404773\\
0.321045849184027 0.74597457263373\\
0.321690519563914 0.753130719069079\\
0.322335189943802 0.582292985682328\\
0.32297986032369 2.02532289685962\\
0.323624530703577 1.67119013682844\\
0.324269201083465 1.25363580767851\\
0.324913871463352 0.304305546007632\\
0.32555854184324 2.32953022517373\\
0.326203212223128 2.85989428267969\\
0.326847882603015 1.76443811674173\\
0.327492552982903 2.26778287566544\\
0.328137223362791 1.04588554938745\\
0.328781893742678 4.30724754402023\\
0.329426564122566 4.00659142470723\\
0.330071234502453 1.68389293726702\\
0.330715904882341 0.409667804633431\\
0.331360575262229 0.987076265119708\\
0.332005245642116 0.307254730357235\\
0.332649916022004 0.258899130164941\\
0.333294586401891 0.570098121798004\\
0.333939256781779 0.423642951608449\\
0.334583927161667 0.403459642533674\\
0.335228597541554 0.378035380913632\\
0.335873267921442 0.304541347111464\\
0.336517938301329 0.462219165915638\\
0.337162608681217 0.40228422601857\\
0.337807279061105 0.780784448749557\\
0.338451949440992 0.349916863891111\\
0.33909661982088 1.4451797252232\\
0.339741290200767 0.175518297302254\\
0.340385960580655 0.56466485558449\\
0.341030630960543 0.973003316034806\\
0.34167530134043 0.607219098236972\\
0.342319971720318 1.34600587443797\\
0.342964642100205 1.04413572169752\\
0.343609312480093 1.02700222729915\\
0.344253982859981 1.07238498715292\\
0.344898653239868 1.03106813673753\\
0.345543323619756 0.949988879692206\\
0.346187993999643 1.36957883248656\\
0.346832664379531 0.0815862073740518\\
0.347477334759419 1.41759720057067\\
0.348122005139306 1.54849489343239\\
0.348766675519194 0.488271428879847\\
0.349411345899081 0.675286011517594\\
0.350056016278969 0.439757609600811\\
0.350700686658857 0.71866651195965\\
0.351345357038744 0.330915571738632\\
0.351990027418632 0.893377874323316\\
0.352634697798519 0.497915937972155\\
0.353279368178407 0.851746054270556\\
0.353924038558295 0.373458851600412\\
0.354568708938182 0.368492798263547\\
0.35521337931807 0.172352591438538\\
0.355858049697958 0.605715269807462\\
0.356502720077845 1.14069539394292\\
0.357147390457733 0.429158204571176\\
0.35779206083762 0.941845020098623\\
0.358436731217508 0.373447935365078\\
0.359081401597395 0.642022656924457\\
0.359726071977283 0.13138797959517\\
0.360370742357171 0.651764642786133\\
0.361015412737058 0.811262571723366\\
0.361660083116946 0.910553666416624\\
0.362304753496834 1.1677988600484\\
0.362949423876721 0.123885642497757\\
0.363594094256609 0.628468647746594\\
0.364238764636496 0.0179357820509832\\
0.364883435016384 0.93967259199272\\
0.365528105396272 0.348794000129821\\
0.366172775776159 0.387560079088756\\
0.366817446156047 0.876923229130027\\
0.367462116535934 0.365890802196315\\
0.368106786915822 0.396165177603618\\
0.36875145729571 0.480710232958751\\
0.369396127675597 1.14202112832987\\
0.370040798055485 0.913712100844539\\
0.370685468435372 0.351275804505446\\
0.37133013881526 0.50423157621166\\
0.371974809195148 0.414474121466853\\
0.372619479575035 0.457355007549653\\
0.373264149954923 0.72018777882058\\
0.37390882033481 0.237653058554195\\
0.374553490714698 0.325150366733996\\
0.375198161094586 0.321403165870613\\
0.375842831474473 0.161245057833877\\
0.376487501854361 0.50258230958835\\
0.377132172234248 0.680720144670306\\
0.377776842614136 0.249943257368383\\
0.378421512994024 0.932147198031224\\
0.379066183373911 0.727278711509292\\
0.379710853753799 0.378775825636743\\
0.380355524133686 0.895559159443504\\
0.381000194513574 0.9404590234534\\
0.381644864893462 0.379576683241323\\
0.382289535273349 1.08132686434935\\
0.382934205653237 1.53913630640118\\
0.383578876033124 0.183110695621112\\
0.384223546413012 0.944217127909611\\
0.3848682167929 0.285484739690819\\
0.385512887172787 0.586917036766296\\
0.386157557552675 0.823171198587144\\
0.386802227932563 0.144838765014088\\
0.38744689831245 0.770011376071273\\
0.388091568692338 0.307315751872114\\
0.388736239072225 0.698683167008646\\
0.389380909452113 1.59354759469669\\
0.390025579832001 0.678384029305205\\
0.390670250211888 0.6709837270385\\
0.391314920591776 0.676622718229717\\
0.391959590971663 0.123874709022687\\
0.392604261351551 0.496738409974821\\
0.393248931731439 0.524566734995718\\
0.393893602111326 0.148525296115843\\
0.394538272491214 0.554860189092968\\
0.395182942871101 0.62995678269079\\
0.395827613250989 0.674344309878319\\
0.396472283630877 0.671362234399977\\
0.397116954010764 0.623188312529085\\
0.397761624390652 1.01577520949601\\
0.398406294770539 1.07024013247175\\
0.399050965150427 1.54930380887963\\
0.399695635530315 0.192516239764158\\
0.400340305910202 0.429601328730772\\
0.40098497629009 0.0685211609656609\\
0.401629646669977 1.2273497834178\\
0.402274317049865 0.44318268676436\\
0.402918987429753 0.318819893006235\\
0.40356365780964 0.280479969533635\\
0.404208328189528 1.05301851340876\\
0.404852998569415 0.746167913992889\\
0.405497668949303 0.517627803076362\\
0.406142339329191 0.760480313603695\\
0.406787009709078 0.215911196658567\\
0.407431680088966 0.781665132670207\\
0.408076350468853 1.33178439123623\\
0.408721020848741 0.375715876704475\\
0.409365691228629 0.649014776334231\\
0.410010361608516 0.6678403847111\\
0.410655031988404 0.40899092137951\\
0.411299702368292 0.378587336000239\\
0.411944372748179 0.81565720019348\\
0.412589043128067 0.572336292779268\\
0.413233713507954 0.33497607590826\\
0.413878383887842 0.167139774408363\\
0.41452305426773 1.1320554955878\\
0.415167724647617 0.797446839789498\\
0.415812395027505 0.0699726229009259\\
0.416457065407392 0.427210120768071\\
0.41710173578728 0.444758801700545\\
0.417746406167168 0.153132392577136\\
0.418391076547055 0.710729614032087\\
0.419035746926943 0.796413608596821\\
0.41968041730683 1.78914708979198\\
0.420325087686718 1.07440463698603\\
0.420969758066606 0.556680074728552\\
0.421614428446493 0.843066845580875\\
0.422259098826381 0.778381375589838\\
0.422903769206268 1.22921698175061\\
0.423548439586156 0.889166742237648\\
0.424193109966044 0.451339223628232\\
0.424837780345931 1.07254214648733\\
0.425482450725819 0.353161455726013\\
0.426127121105706 0.510135831662938\\
0.426771791485594 0.652801662818585\\
0.427416461865482 0.358660144640214\\
0.428061132245369 0.534337449803021\\
0.428705802625257 1.73235730017814\\
0.429350473005144 0.674147233454724\\
0.429995143385032 0.459129719557215\\
0.43063981376492 1.31176028029328\\
0.431284484144807 0.618286327703995\\
0.431929154524695 0.538876863373054\\
0.432573824904582 0.678531217106803\\
0.43321849528447 0.903968009364529\\
0.433863165664358 0.920248414386715\\
0.434507836044245 0.370169431981573\\
0.435152506424133 0.91962715391236\\
0.43579717680402 0.45406431079828\\
0.436441847183908 0.40042571875501\\
0.437086517563796 1.00282091724888\\
0.437731187943683 0.652127140825808\\
0.438375858323571 0.366095686771744\\
0.439020528703458 0.599625251164657\\
0.439665199083346 0.406608681191374\\
0.440309869463234 0.642417462367762\\
0.440954539843121 0.758716829729723\\
0.441599210223009 0.430624013444005\\
0.442243880602896 0.391206724956904\\
0.442888550982784 0.868095915948399\\
0.443533221362672 0.250651740508221\\
0.444177891742559 1.01168645636494\\
0.444822562122447 0.487051640324719\\
0.445467232502335 0.804666267475564\\
0.446111902882222 1.05525183567181\\
0.44675657326211 0.398576927633433\\
0.447401243641997 0.348434197610115\\
0.448045914021885 0.664315081236\\
0.448690584401773 0.564044823180914\\
0.44933525478166 0.573372043345211\\
0.449979925161548 0.343016452937687\\
0.450624595541435 0.281725191636668\\
0.451269265921323 0.449871925270198\\
0.451913936301211 0.318315246590952\\
0.452558606681098 0.910980186355224\\
0.453203277060986 0.973101322132636\\
0.453847947440873 0.65861810822261\\
0.454492617820761 0.113081834896646\\
0.455137288200648 0.579993174091679\\
0.455781958580536 0.508099199010237\\
0.456426628960424 0.456487863082758\\
0.457071299340311 1.52066916488217\\
0.457715969720199 0.665793376400352\\
0.458360640100087 0.678725272495215\\
0.459005310479974 0.950171171017316\\
0.459649980859862 0.384263484981726\\
0.460294651239749 0.982104001092733\\
0.460939321619637 0.412856313161241\\
0.461583991999525 1.17277452187592\\
0.462228662379412 0.710814001389388\\
0.4628733327593 0.542980335978955\\
0.463518003139187 1.20797318811902\\
0.464162673519075 0.78547007756743\\
0.464807343898963 1.37831187001776\\
0.46545201427885 0.626970638139466\\
0.466096684658738 0.274852228348871\\
0.466741355038625 0.550916565660391\\
0.467386025418513 0.855463562594564\\
0.468030695798401 1.0159270232556\\
0.468675366178288 0.902566151772035\\
0.469320036558176 0.314441759605153\\
0.469964706938063 0.473593064669641\\
0.470609377317951 1.70663799048972\\
0.471254047697839 0.194170951371443\\
0.471898718077726 0.792578176426598\\
0.472543388457614 1.16287609546449\\
0.473188058837501 0.48287123715063\\
0.473832729217389 1.41110537425679\\
0.474477399597277 0.823275710643293\\
0.475122069977164 1.38306401729796\\
0.475766740357052 0.45884508230764\\
0.476411410736939 0.854826891939422\\
0.477056081116827 1.33694358701888\\
0.477700751496715 0.82251647173493\\
0.478345421876602 0.510581418131107\\
0.47899009225649 0.447151037325952\\
0.479634762636377 0.545429788227175\\
0.480279433016265 1.07508824702658\\
0.480924103396153 0.673346739218248\\
0.48156877377604 1.12649127905756\\
0.482213444155928 0.897083356660919\\
0.482858114535816 0.973547819286771\\
0.483502784915703 1.47557444755576\\
0.484147455295591 0.589115704211493\\
0.484792125675478 0.888740138032544\\
0.485436796055366 1.48110627637048\\
0.486081466435254 0.888057243017652\\
0.486726136815141 0.689967085782206\\
0.487370807195029 1.51878918940404\\
0.488015477574916 1.19253257012798\\
0.488660147954804 0.990067754533684\\
0.489304818334692 0.275377598590075\\
0.489949488714579 1.32302772672525\\
0.490594159094467 0.703301079392721\\
0.491238829474354 0.697621449651935\\
0.491883499854242 0.973291059355675\\
0.49252817023413 0.620557887949572\\
0.493172840614017 0.382176386152952\\
0.493817510993905 0.401089468238306\\
0.494462181373792 0.723173143094323\\
0.49510685175368 1.19507939117944\\
0.495751522133568 0.76813280815329\\
0.496396192513455 0.809318049421991\\
0.497040862893343 0.062548731595178\\
0.49768553327323 0.592840367239477\\
0.498330203653118 1.25069272113226\\
0.498974874033006 0.599353118490977\\
0.499619544412893 0.423461500672383\\
0.500264214792781 0.304996713626087\\
0.500908885172668 0.162846994727639\\
0.501553555552556 1.24406296616474\\
0.502198225932444 1.39108893105315\\
0.502842896312331 0.462934593360861\\
0.503487566692219 0.700100857078215\\
0.504132237072107 1.14741558453332\\
0.504776907451994 0.622988371384884\\
0.505421577831882 0.739375944794553\\
0.506066248211769 0.673027429924801\\
0.506710918591657 0.906209549576871\\
0.507355588971545 0.372137822979937\\
0.508000259351432 0.573448538524088\\
0.50864492973132 0.748254973238117\\
0.509289600111207 0.810970233568124\\
0.509934270491095 0.556931388865359\\
0.510578940870983 0.259802827839023\\
0.51122361125087 0.345582833416603\\
0.511868281630758 1.30556511839938\\
0.512512952010645 0.401622928321183\\
0.513157622390533 0.315538790414948\\
0.513802292770421 0.360935439918009\\
0.514446963150308 2.14960472706972\\
0.515091633530196 0.450979410198643\\
0.515736303910083 0.834105139846055\\
0.516380974289971 0.400661906594581\\
0.517025644669859 1.08821131444534\\
0.517670315049746 0.733622708007862\\
0.518314985429634 0.623443542683872\\
0.518959655809521 0.787080536146185\\
0.519604326189409 1.33951305194057\\
0.520248996569297 0.670634507954769\\
0.520893666949184 0.586126037999083\\
0.521538337329072 1.30571699942627\\
0.522183007708959 0.153894660894552\\
0.522827678088847 0.582358601375472\\
0.523472348468735 0.376406207110086\\
0.524117018848622 1.62982346439243\\
0.52476168922851 0.719071409610005\\
0.525406359608397 0.637391719643742\\
0.526051029988285 0.88185379673798\\
0.526695700368173 0.890869600148598\\
0.52734037074806 1.14599760534083\\
0.527985041127948 0.84225598055457\\
0.528629711507835 1.01910066330038\\
0.529274381887723 2.559298664482\\
0.529919052267611 1.60977533640563\\
0.530563722647498 0.276780397742104\\
0.531208393027386 0.374353872829372\\
0.531853063407273 0.498287583647445\\
0.532497733787161 1.00019690008488\\
0.533142404167049 1.31530468698885\\
0.533787074546936 1.16356694711317\\
0.534431744926824 1.74081673801238\\
0.535076415306711 1.70265923194714\\
0.535721085686599 0.62961317938733\\
0.536365756066487 0.799469382410005\\
0.537010426446374 0.298628343138299\\
0.537655096826262 0.643774520136915\\
0.538299767206149 1.07744932303716\\
0.538944437586037 1.01394403572202\\
0.539589107965925 1.06815147619992\\
0.540233778345812 0.924556183175069\\
0.5408784487257 1.61258171971306\\
0.541523119105588 1.00468798222385\\
0.542167789485475 0.961459194875559\\
0.542812459865363 0.780103907336954\\
0.54345713024525 0.732329684074356\\
0.544101800625138 0.461342221153143\\
0.544746471005026 1.6905174366817\\
0.545391141384913 0.650159800878327\\
0.546035811764801 0.728235645784788\\
0.546680482144688 1.62905324442157\\
0.547325152524576 1.83890598436375\\
0.547969822904464 1.14833042930035\\
0.548614493284351 0.150105365589467\\
0.549259163664239 1.14417186570589\\
0.549903834044126 0.743643805047877\\
0.550548504424014 0.215592021648915\\
0.551193174803902 0.184111129593355\\
0.551837845183789 0.450423129880062\\
0.552482515563677 1.69839212090223\\
0.553127185943564 2.04707667348212\\
0.553771856323452 1.3729550085532\\
0.55441652670334 0.576186274159884\\
0.555061197083227 1.54931822870691\\
0.555705867463115 1.50839471013326\\
0.556350537843002 0.550234278801904\\
0.55699520822289 1.47552756599213\\
0.557639878602778 2.40230591278178\\
0.558284548982665 1.15829713574153\\
0.558929219362553 1.61264286740853\\
0.55957388974244 1.52617912480769\\
0.560218560122328 1.48265364533559\\
0.560863230502216 1.1378672374547\\
0.561507900882103 1.49857633020345\\
0.562152571261991 1.2310233505757\\
0.562797241641878 0.955721405529259\\
0.563441912021766 1.35449486664605\\
0.564086582401654 2.47399410056024\\
0.564731252781541 1.48336996179433\\
0.565375923161429 1.15760578665136\\
0.566020593541317 2.76337359155614\\
0.566665263921204 1.40203955471507\\
};
\addlegendentry{Not suspended};
\end{axis}
\node at (6.2,-0.9) {\large{\textbf{b}}};
\begin{axis}[scale = 0.18*\textwidth/(4.5in),
xshift = -0.38*\textwidth,
yshift = 0.18*\textwidth,
width=4.52083333333333in,
height=3.565625in,
scale only axis,
xmin=0,
xmax=250,
xlabel={Distance ($\mu$m)},
xlabel style = {yshift=1.1ex},
ymin=0,
ymax=120,
ylabel={},
ylabel style = {yshift=-1.5ex},
legend style={fill=none,draw=none,legend cell align=left,at={(0.50,1.0)},anchor=north},
at=(plot1.right of south east),
anchor=left of south west,
axis x line*=bottom,
axis y line*=left,
minor xtick={50,100,...,200},
minor ytick={25,50,...,100}
]
\addplot [
color=blue,
densely dotted,
line width=1.0pt,
mark size=0.5pt,
mark=o,
mark options={solid}
]
table[row sep=crcr,y expr=\thisrow{y}/4]{
x y \\
0 0.278598865904589\\
6.46203403479147 143.110610121414\\
12.9240680695829 6.49188612177484\\
19.3861021043744 152.573282893612\\
25.8481361391659 111.332921384056\\
32.3101701739574 81.5191124754069\\
38.7722042087488 60.4056016685656\\
45.2342382435403 32.7774907005184\\
51.6962722783318 104.23843524198\\
58.1583063131233 12.7833493421319\\
64.6203403479147 36.7633575215851\\
71.0823743827062 52.9787491874336\\
77.5444084174977 50.5578375639892\\
84.0064424522892 67.2962500063884\\
90.4684764870806 416.177434529021\\
96.9305105218721 51.6040513193207\\
103.392544556664 80.0141804885212\\
109.854578591455 33.24721676615\\
116.316612626247 22.1082073640792\\
122.778646661038 37.8090981403369\\
129.24068069583 23.7104025595442\\
135.702714730621 44.9516152700414\\
142.164748765412 195.415815200243\\
148.626782800204 55.6858975290417\\
155.088816834995 90.5863998778204\\
161.550850869787 45.7212829153688\\
168.012884904578 18.1448759570638\\
174.47491893937 83.2490716897523\\
180.936952974161 26.5929751105273\\
187.398987008953 39.9787503189324\\
193.861021043744 22.629613475177\\
200.323055078536 28.6163400298128\\
206.785089113327 43.7726754835412\\
213.247123148119 32.9196280503952\\
219.70915718291 24.3154671913803\\
226.171191217702 51.464811029108\\
232.633225252493 62.3588629793612\\
239.095259287285 21.5446455917886\\
245.557293322076 18.8064389386513\\
};
\addplot [
color=green!50!black,
solid,
line width=1.0pt,
mark size=0.5pt,
mark=o,
mark options={solid}
]
table[row sep=crcr]{
0 0.80049501900913\\
6.44670379887604 23.7436780588453\\
12.8934075977521 6.27801145379132\\
19.3401113966281 24.6121973610882\\
25.7868151955042 20.2780013984331\\
32.2335189943802 12.1931435858617\\
38.6802227932562 21.2942935130416\\
45.1269265921323 2.90002028355982\\
51.5736303910083 12.5389077380775\\
58.0203341898844 11.925877327049\\
64.4670379887604 6.17999037969161\\
70.9137417876365 3.08906904848315\\
77.3604455865125 15.583872210052\\
83.8071493853885 23.3738616078201\\
90.2538531842646 107.889906262999\\
96.7005569831406 28.7980550145895\\
103.147260782017 3.18482144912548\\
109.593964580893 13.3121518611209\\
116.040668379769 18.0652190935665\\
122.487372178645 12.8458857397283\\
128.934075977521 6.77617737678047\\
135.380779776397 16.3404181499597\\
141.827483575273 43.6517011573123\\
148.274187374149 8.22743312458415\\
154.720891173025 16.1827465113516\\
161.167594971901 9.50938542200359\\
167.614298770777 13.8964609591377\\
174.061002569653 10.4291602981113\\
180.507706368529 6.21802916450032\\
186.954410167405 1.39558958286694\\
193.401113966281 9.32179547149372\\
199.847817765157 8.43226928510773\\
206.294521564033 9.33520931265123\\
212.741225362909 6.79113666002204\\
219.187929161785 9.75237953584085\\
225.634632960661 12.7206378022742\\
232.081336759538 11.7620506839204\\
238.528040558414 12.8957449492325\\
244.97474435729 0.65734405304918\\
};
\end{axis}
\end{tikzpicture}%
}
\caption{\textbf{The interface reflections are small.} \textbf{a}, The fiber-to-fiber transmission spectrum shows the grating bandwidth. It exhibits a slight ($1 \, \text{dB}$) variation at a free spectral range of about $2.9 \, \text{nm}$, both in standard and suspended silicon-on-insulator waveguides. \textbf{b}, The Fourier transform of this spectrum reveals potential cavities present in the wire. We compare the regular wire (green) to a wire with $85$ suspensions ($L_{\text{s}} = 25.4 \, \mu\text{m}$) and anchors ($L_{\text{a}} = 4.6 \, \mu\text{m}$) (blue). The distance-axis was calibrated with a group index of $n_{\text{g}} = 4.7$ and shows the waveguide length of $2.7 \, \text{mm}$ because of grating coupler reflections. These are weaker in the suspended case owing to the higher propagation loss. The $2.9 \, \text{nm}$ free spectral range (\textbf{a}) shows up at $90 \, \mu\text{m}$ (\textbf{b}, inset). Generally, the two Fourier spectra are nearly identical and do not exhibit notable peaks related to the suspension.}
\label{fig:passive}
\end{figure*}
\section*{Acknowledgement.} R.V.L. acknowledges the Agency for Innovation by Science and Technology in Flanders (IWT) for a PhD grant and A.B. the ITN-network cQOM for a postdoc grant. This work was partially funded under the FP7-ERC-InSpectra programme.
\section*{References}
{\scriptsize
|
\section{Introduction}
\label{sec:intro}
Gaarder \& Wolf~\cite{GaarderWolf} showed that perfect instantaneous output-feedback\footnote{By \emph{perfect instantaneous output-feedback} we refer to a model where each transmitter observes all previous channel outputs before it has to produce the next input. We henceforth refer to it also as \emph{perfect feedback.}} can increase the capacity of the two-user memoryless multiple-access channel (MAC) by enabling cooperation between the transmitters. The capacity region for general MACs with feedback is still unknown even for only two users. (A notable exception being Ozarow's capacity result for the two-user Gaussian MAC with perfect feedback \cite{Ozarow84}.)
The Gaarder-Wolf scheme has been extended by Cover \& Leung \cite{CoverLeung} who introduced the ideas of block-Markov coding and superposition coding. Specifically, in the Cover-Leung scheme, in each block~$b$, the transmitters send independent fresh data superposed on common update information belonging to the previous block~$(b-1)$. After observing the outputs in block~$b$, the receiver creates a list of all possible pairs of block-$b$ fresh data that is compatible (jointly typical) with these outputs. It also decodes the common update information. This common update information describes resolution information that allows the receiver to resolve its block-$(b-1)$ list, and thus to identify the fresh data that was sent in block~$(b-1)$. In order to be able to compute and send the block-$b$ common update information, the transmitters have to decode each other's fresh data sent in block~$(b-1)$ and calculate the receiver's block-$(b-1)$ list. They perform these tasks using their block-$(b-1)$ input
signals and the block-$(b-1)$ feedback signals. (In case of perfect feedback, the latters correspond to the receiver's channel outputs.) For some MACs with perfect feedback the presented Cover-Leung scheme is optimal and achieves capacity~\cite{Willems82}. For others, for example for the Gaussian MAC~\cite{Ozarow84}, it is strictly suboptimal~\cite{Ozarow84,BrossLapidoth, VenkatPradhan2011}.
The Cover-Leung scheme has been improved by relaxing the requirement that after the transmission of each block, the transmitters have to decode each other's fresh data sent in this block~\cite{BrossLapidoth,VenkatPradhan2011}. Instead, the decoding at the transmitters (and also at the receivers) is delayed, allowing the transmitters to gain more information about each other's message before decoding. This results in less stringent rate-constraints as compared to the original Cover-Leung scheme. To implement this idea, Bross \& Lapidoth \cite{BrossLapidoth} proposed to append to each block a two-way transmitters-exchange phase and to delay the transmitters' decoding thereafter.
Venkataramanan \& Pradhan \cite{VenkatPradhan2011} suggested to delay the transmitters decoding of the fresh data by an entire block.
In their scheme, in each block~$b$ the transmitters send two sorts of resolution information, common \emph{receiver-side resolution information} to resolve the receiver's uncertainty about the block-$(b-1)$ fresh data, and correlated \emph{transmitters-side resolution information} to resolve each transmitter's uncertainty about the other transmitter's block-$(b-1)$ fresh data.
Coding schemes were also presented for the MAC with generalized, noisy, or rate-limited\footnote{While for generalized \cite{Carleial} or noisy feedback the feedback signals are ``passively" produced in a memoryless way from the channel inputs and outputs, in the model for rate-limited feedback the receiver can \emph{actively} code over the feedback links.} feedback.
Carleial~\cite{Carleial} proposed a coding scheme for general discrete memoryless MACs with generalized feedback, which combines the Cover-Leung scheme with an optimal nofeedback scheme through rate-splitting.
Lapidoth \& Wigger~\cite{LapidothWigger} proposed a scheme for the two-user Gaussian MAC with noisy feedback. Their scheme can be viewed as a robustification of Ozarow's capacity-achieving perfect-feedback scheme~\cite{Ozarow84} to noisy feedback.
The main focus of this paper is on rate-limited feedback. For this model,
Shaviv \& Steinberg~\cite{SteinbergShaviv} proposed a coding scheme based on Carleial's extension of the Cover-Leung scheme and on Heegard-Berger source coding~\cite{HeegardBerger} to communicate over the feedback links. For sufficiently large feedback rates their scheme achieves Cover \& Leung's achievable region for perfect feedback~\cite{CoverLeung} (which in this case coincides with Carleial's achievable region).
In this paper, we propose a coding scheme for the two-user discrete memoryless MAC with common rate-limited feedback. Our coding scheme is based on the Venkataramanan-Pradhan scheme and on Heegard-Berger source coding~ \cite{HeegardBerger} over the feedback links. Our new region includes the Shaviv-Steinberg achievable region and this inclusion can be strict.
For sufficiently large feedback rates, our achievable region coincides with the Venkataramanan-Pradhan achievable region.
\section{Channel Model}
\label{sec:channel}
We consider the two-user discrete memoryless MAC with rate-limited feedback.
The setup is characterized by the triple of finite alphabets $(\set{X}_1,\set{X}_2,\set{Y})$, the conditional probability distribution $P_{Y|X_1X_2}$, and a nonnegative feedback rate $R_{\textnormal{fb}}$. At each time $t\in \mathbb{N}$, if $x_{1,t}\in \set{X}_1$ and $x_{2,t}\in \set{X}_2$ denote the signals sent by Transmitters~1 and~2, the receiver observes the channel output $y_t\in\set{Y}$ with probability
$P_{Y|X_1X_2}(y_t|x_{1,t},x_{2,t})$.
The goal of communication is that Transmitters~1 and~2 convey the independent messages $M_1$ and $M_2$ to the common receiver. The messages $M_1$ and $M_2$ are uniformly distributed over
$\mathcal{M}_1=\{1,\dots, \lfloor 2^{n R_1}\rfloor\}$ and $\mathcal{M}_2=\{1,\dots, \lfloor 2^{n R_2}\rfloor\}$, where $R_1$ and $R_2$ are the rates of transmission and $n$ is the blocklength.
We assume common rate-limited feedback from the receiver to both transmitters. Specifically, upon observing $Y_t$, the receiver can send a feedback signal $F_{t}\in\set{F}_{t}$ to both transmitters where $\set{F}_{t}$ denotes the finite alphabet of $F_{t}$.
The feedback signals $F_1,\ldots, F_n$ are of the form
\begin{IEEEeqnarray}{rCl}
F_{t}&=&\psi_{t}^{(n)}\left(Y_1,\dots,Y_t\right) ,\quad i\in \{1,2\}, \quad t \in \{1,\dots,n\},\label{eq:feedback_encoding}
\end{IEEEeqnarray}
for some feedback-encoding functions
$\psi_{t}^{(n)}\colon \set{Y}^t \to \set{F}_{t}$.
It is assumed that both transmitters receive the feedback signals perfectly whenever the former satisfy the rate constraint on the feedback links:
\begin{IEEEeqnarray}{rCl}
|\set{F}_{1}|\times \ldots \times |\set{F}_{n}|&\leq& 2^{n R_{\textnormal{fb}}}. \label{eq:feedback_constraint}
\end{IEEEeqnarray}
(The present feedback rate constraint is rather weak. One could imagine a stronger constraint where each sample $F_t$ has to satisfy $H(F_t)\leq R_{\textnormal{fb}}$. It can be easily shown that the two definitions are equivalent in terms of achievable rates.)
Notice that here the alphabets $\set{F}_1,\ldots, \set{F}_n$ are design parameters of the coding scheme.
Transmitter~$i$'s channel input at time~$t$, $X_{i,t},$ for $i\in\{1,2\}$, can depend on Message $M_i$ and the prior feedback signals $F_{1},\dots,F_{t-1}$:
\begin{IEEEeqnarray}{rCL}\label{eq:MACinputs}
X_{i,t}=\varphi_{i,t}^{(n)} (M_i,F_{1},\dots,F_{t-1} ),\quad t\in\{1,\dots,n\},\label{eq:encoding}
\end{IEEEeqnarray}
for some encoding functions of the form $\varphi_{i,t}^{(n)}:\mathcal{M}_i\times\set{F}_{1}\times \dots\times\set{F}_{t-1}\to \set{X}_i.$
The receiver bases its guess of its desired messages on the output sequence $Y_1,\ldots,Y_n$. That is, it produces
\begin{equation*}
(\hat{M}_1,\hat{M}_2)=\phi^{(n)}(Y_1,\ldots, Y_n),
\end{equation*}
for a decoding function ${\phi^{(n)}\colon \set{Y}^n \to\mathcal{M}_1\times \mathcal{M}_2}$. There is an error in the communication whenever $(\hat{M}_1,\hat{M}_2)\neq (M_1,M_2)$. The average probability of error is thus
\begin{IEEEeqnarray}{rCL}
\label{eq:error-mac}
P_{\textnormal{e}}^{(n)}\triangleq\textnormal{Pr}\big[ (\hat{M}_1,\hat{M}_2) \neq (M_1,M_2) \big].
\end{IEEEeqnarray}
We say that a rate pair $(R_1,R_2)$ is achievable over the MAC with common rate-limited feedback if there exists a sequence of encoding and decoding functions $\{\{\varphi_{1,t}^{(n)}\}_{t=1}^n,\{\varphi_{2,t}^{(n)}\}_{t=1}^n,\phi^{(n)}\}_{n=1}^\infty$ as described above, a sequence of feedback alphabets $\{\set{F}_{t}\}_{t=1}^n$ satisfying \eqref{eq:feedback_constraint}, and feedback-encoding functions of the form~\eqref{eq:feedback_encoding} such that $P_{\textnormal{e}}^{(n)}$ tends to zero as the blocklength $n$ tends to infinity.
When $R_\textnormal{fb}=0$, the feedback signals have to be deterministic and the setup is equivalent to a setup without feedback.
When $R_\textnormal{fb}\geq\log_2 |\set{Y}|$, the setup is equivalent to perfect-feedback.
\section{Achievable Region}
\label{sec:ach_reg}
\begin{theorem}[Achievable Region]
\label{thm}
Let $\set{W}$, $\set{U}_1$, $\set{U}_2$, $\set{V}_1$, $\set{V}_2$, $\set{Y}_{12}$, $\set{Y}_{1}$, and $\set{Y}_{2}$ be arbitrary finite sets. Also, let $(W,V_1,V_2,U_1,U_2,X_1, X_2, Y,Y_{12},Y_1,Y_2)$ and $(\tilde{W},\tilde{V}_1,\tilde{V}_2,\tilde{U}_1,\tilde{U}_2,\tilde X_1, \tilde X_2, \tilde Y,\tilde{Y}_{12},\tilde Y_1,\tilde Y_2)$ be two correlated random tuples over the product alphabets $\set{W}\times\set{U}_1\times\set{U}_2\times\set{V}_1\times\set{V}_2 \times \set{X}_1\times \set{X}_2 \times \set{Y}\times \set{Y}_{12} \times \set{Y}_1 \times \set{Y}_2$ satisfying the following two conditions:
\begin{enumerate}
\item The joint distributions of the two tuples coincide:
\begin{equation}\label{eq:cond31}
P_{WU_1U_2V_1V_2X_1X_2YY_{12}Y_1 Y_2} =P_{\tilde W \tilde U_1\tilde U_2\tilde V_1\tilde V_2\tilde X_1\tilde X_2\tilde Y\tilde Y_{12}\tilde Y_1 \tilde Y_2},
\end{equation}
and each of them factors as
\begin{IEEEeqnarray}{rCl}\label{eq:cond32}
\lefteqn{P_{WU_1U_2V_1V_2X_1X_2YY_{12}Y_1 Y_2}}\qquad \nonumber \\ & =& P_W P_{V_1V_2} P_{U_1|WV_1} P_{U_2| WV_2} \nonumber\\&&\hspace{0.5cm}\cdot P_{X_1| W U_1 V_1} P_{X_2| W U_2 V_2} P_{Y|X_1 X_2}\nonumber\\ & &\hspace{1cm}\cdot P_{Y_{12}|Y W} P_{Y_2|W Y Y_{12}} P_{Y_1|W Y Y_{12}}
\end{IEEEeqnarray}
where $P_{Y|X_{1}X_2}$ describes the channel law of our MAC.
\item Defining $\tilde{S}\eqdef (\tilde W, \tilde{V}_1, \tilde{V}_2, \tilde{Y}_{12})$,
the joint distribution over both tuples factors as
\begin{IEEEeqnarray}{rCl}\label{eq:cond33}
\lefteqn{P_{ W U_1U_2V_1V_2X_1X_2Y Y_{12} Y_1 Y_2\tilde S \tilde U_1\tilde U_2\tilde X_1\tilde X_2\tilde Y \tilde Y_1 \tilde Y_2}} \qquad\qquad \nonumber \\ & =& P_{\tilde S \tilde U_1 \tilde U_2 \tilde X_1 \tilde X_2 \tilde Y \tilde Y_1 \tilde Y_2} P_{V_1|\tilde S \tilde U_1}P_{V_2|\tilde S \tilde U_2}\nonumber \\ & &\hspace{1cm} \cdot P_{W U_1 U_2 X_1 X_2 Y_{12} Y_1 Y_2 Y|V_1 V_2}.\IEEEeqnarraynumspace
\end{IEEEeqnarray}
\end{enumerate}
All nonnegative rate pairs $(R_1,R_2)$ satisfying Constraints~\eqref{eq:ach_region} on top of next page are achievable.
\begin{figure*}
\begin{subequations}
\label{eq:ach_region}
\begin{IEEEeqnarray}{lCl}
R_{\textnormal{fb}}&\geq&\max\bigr\{I(Y_{12};Y |W X_1),I(Y_{12};Y |W X_2)\bigr\}+ I(Y;Y_1|Y_{12} X_1 W)+ I(Y;Y_2|Y_{12} X_2 W)\label{eq:fb_constraint}\\
R_1&\leq &I(X_1;Y|SU_1U_2 X_2)\nonumber\\&&+\min \{I(U_1 ;Y_2 Y_{12}|\tilde S\tilde Y_2 \tilde U_2\tilde X_2 W V_2 U_2 X_2), I(W;Y|\tilde W \tilde Y)+I(U_1;Y| W V_1 V_2 U_2)+ I(V_1;Y|\tilde Y\tilde S \tilde Y_1 \tilde Y_2\tilde U_2 W V_2)\} \\
R_2&\leq &I(X_2;Y|SU_1U_2 X_1)\nonumber\\&&+\min\{I(U_2;Y_1Y_{12}|\tilde S \tilde Y_1\tilde U_1\tilde X_1 W V_1 U_1 X_1),I(W;Y|\tilde W \tilde Y)+ I(U_2;Y| W V_1 V_2 U_1)+I(V_2;Y|\tilde Y \tilde S \tilde Y_1 \tilde Y_2 \tilde U_1 W V_1)\}\\
R_1+R_2&\leq &I(X_1X_2;Y|SU_1U_2)+I(W;Y|\tilde W \tilde Y)+ I(U_1 U_2;Y| W V_1 V_2)+I(V_1 V_2 ;Y|\tilde Y\tilde S\tilde Y_1 \tilde Y_2 W)\\
R_1+R_2&\leq &I(X_1X_2;Y|SU_1U_2)\nonumber\\&&+\min\{I(U_2;Y_1Y_{12}|\tilde S \tilde Y_1\tilde U_1\tilde X_1 W V_1 U_1 X_1),I(W;Y|\tilde W \tilde Y)+ I(U_2;Y| W V_1 V_2 U_1)+I(V_2;Y|\tilde Y \tilde S \tilde Y_1 \tilde Y_2\tilde U_1 W V_1)\}\nonumber\\&&+\min\{I(U_1 ;Y_2Y_{12}|\tilde S \tilde Y_2\tilde U_2\tilde X_2 W V_2 U_2 X_2),I(W;Y|\tilde W \tilde Y)+I(U_1;Y| W V_1 V_2 U_2)+ I(V_1;Y|\tilde Y\tilde S \tilde Y_1 \tilde Y_2\tilde U_2 W V_2)\}\IEEEeqnarraynumspace
\vspace*{-1mm}
\end{IEEEeqnarray}
\end{subequations}
\hrule
\vspace*{-5mm}
\end{figure*}
\end{theorem}
\begin{remark}Using time-sharing, it can be shown that also the convex hull of the region described in Theorem~\ref{thm} is achievable.
\end{remark}
\begin{remark}When choosing $V_1=V_2=Y_{12}=Y_1=Y_2=W=\emptyset$, $U_1=X_1$, and $U_2=X_2$, the achievable region in Theorem~\ref{thm} reduces to the nofeedback capacity region.
When $R_{fb}\geq \log_2|\set{Y}|$, Constraint~\eqref{eq:fb_constraint} is always satisfied. In this case, we can choose $Y=Y_{12}$ and $Y_1=Y_2=\emptyset$ to mimic the setup with perfect feedback.
If now we specialize the achievable region in Theorem~\ref{thm} to the choices $U_1=X_1$ and $U_2=X_2$, we recover the achievable region of Venkataramanan \& Pradhan for perfect feedback~\cite{VenkatPradhan2011}.
Finally, when choosing $V_1=V_2=\emptyset$, the tilded and the non-tilded tuples become independent and the achievable region in Theorem~\ref{thm} specializes to the set of all rate pairs $(R_1, R_2)$ satisfying
\begin{subequations}
\label{eq:v1v2empty}
\begin{IEEEeqnarray}{cCl}
R_{\textnormal{fb}}&\geq&\max_{i\in\{1,2\}}\bigr\{I(Y_{12};Y |W X_i)\bigr\}\nonumber\\
&&+ I(Y;Y_1|Y_{12} X_1 W)+ I(Y;Y_2|Y_{12} X_2 W)\\
R_1&\leq &I(X_1;Y|W U_1 X_2)+\min \{I(U_1;Y| W U_2)\nonumber\\&&+I(W;Y),I(U_1 ;Y_2 Y_{12}|W X_2)\} \\
R_2&\leq &I(X_2;Y|W U_2 X_1)+\min\{I(U_2;Y| W U_1)\nonumber\\&&+I(W;Y),I(U_2;Y_1Y_{12}| W X_1)\}\\
R_1+R_2&\leq &I(X_1X_2;Y|W U_1U_2)+\min\{I(U_2;Y| W U_1)\nonumber\\&&+ I(W;Y),I(U_2;Y_1Y_{12}| W X_1)\}\nonumber\\&&
+\min\{I(U_1;Y| W U_2)+I(W;Y),\nonumber\\&&I(U_1 ;Y_2Y_{12}| W X_2)\}\IEEEeqnarraynumspace\\
R_1+R_2&\leq &I(X_1X_2;Y)
\end{IEEEeqnarray}
\end{subequations}
where now $P_{WU_1U_2X_1X_2Y Y_{12}Y_1 Y_2}$ factors as
\begin{IEEEeqnarray}{rCl}
\lefteqn{P_{WU_1U_2X_1X_2Y Y_{12}Y_1 Y_2}} \nonumber\qquad \\ & =& P_W P_{U_1|W} P_{U_2| W} P_{X_1|U_1 W} P_{X_2|U_2 W} P_{Y|X_1 X_2}\nonumber\\&&\hspace{1cm}\cdot P_{Y_{12}|Y W} P_{Y_1|Y W Y_{12}} P_{Y_2|Y W Y_{12}} . \IEEEeqnarraynumspace
\end{IEEEeqnarray}
This region contains the achievable region by Shaviv \& Steinberg~\cite{SteinbergShaviv}.
\end{remark}
\section{Outline of Coding Scheme}
\label{sec:coding}
Let $W, U_1, U_2, V_1, V_2, X_1, X_2, Y, Y_{12}, Y_1, Y_2, \tilde W, \tilde U_1, \tilde U_2, \tilde V_1,\break \tilde V_2, \tilde X_1, \tilde X_2,\tilde Y, \tilde Y_{12}, \tilde Y_1,\tilde Y_2$ be as defined in Theorem~\ref{thm} so that they satisfy~\eqref{eq:cond31}--\eqref{eq:cond33}.
Fix a nonnegative rate pair $(R_1, R_2)$ that satisfies rate constraints~\eqref{eq:ach_region} with strict inequalities. Using for example Fourier-Motzkin Elimination, it can be shown that there must exist rates $R_1'>0, R_2'>0$ so that they satisfy $R_1> R_1'$ and $R_2> R_2'$ and the following nine conditions
\begin{subequations}
\begin{IEEEeqnarray}{cCl}
R_{\textnormal{fb}}&>&\max\bigr\{I(Y_{12};Y |W X_1),I(Y_{12};Y |W X_2)\bigr\}\nonumber\\&&+\sum_{i\in\{1,2\}} I(Y;Y_i|Y_{12} X_i W),\label{eq:rfb} \IEEEeqnarraynumspace\\
R_1-R'_1 &< &I(X_1;Y|SU_1U_2 X_2)\label{eq:r1-r'1} \\
R_2-R'_2 &< &I(X_2;Y|SU_1U_2 X_1)\label{eq:r2-r'2}\\
R_1\hspace*{-0.5mm}-\hspace*{-0.5mm}R'_1\hspace*{-0.5mm}+\hspace*{-0.5mm}R_2\hspace*{-0.5mm}-\hspace*{-0.5mm}R'_2\hspace*{-0.5mm} &< &I(X_1X_2;Y|SU_1U_2)\label{eq:r-r'}\\
R'_1&< & I(U_1 ;Y_2Y_{12}|\tilde S \tilde Y_2\tilde U_2\tilde X_2 W V_2 U_2 X_2)\label{eq:r'11}\IEEEeqnarraynumspace \\
R'_2&< & I(U_2;Y_1Y_{12}|\tilde S \tilde Y_1\tilde U_1\tilde X_1 W V_1 U_1 X_1)\label{eq:r'21}\IEEEeqnarraynumspace\\
R'_1&< & I(W;Y|\tilde W \tilde Y)+I(U_1;Y| W V_1 V_2 U_2)\nonumber\\&&
+ I(V_1;Y|\tilde Y\tilde S \tilde Y_1 \tilde Y_2\tilde U_2 W V_2)\label{eq:r'12}\\
R'_2&< & I(W;Y|\tilde W \tilde Y)+ I(U_2;Y| W V_1 V_2 U_1)\nonumber\\&&+I(V_2;Y|\tilde Y \tilde S \tilde Y_1 \tilde Y_2 \tilde U_1 W V_1)\label{eq:r'22}\\
R'_1 +R'_2&< & I(W;Y|\tilde W \tilde Y)+ I(U_1 U_2;Y| W V_1 V_2)\nonumber\\&&+I(V_1 V_2 ;Y|\tilde Y\tilde S \tilde Y_1 \tilde Y_2 W).\label{eq:r'1+r'2}
\end{IEEEeqnarray}
\end{subequations}
Fix also a constant $R_0$ satisfying:
\begin{IEEEeqnarray}{rCl}
R_0 &<& I(W;Y|\tilde{W},\tilde{Y}).\label{eq:r0}
\end{IEEEeqnarray}
We briefly describe a random code construction for which the average probability of error (averaged over codebooks, messages, and channel realizations) can be shown to tend to 0. A deterministic coding scheme achieving the same rates can then be obtained via standard arguments.
Our coding scheme is based on block-Markov and superposition coding, rate-splitting, sliding-window decoding, and Heegard-Berger coding on the feedback links. It extends over $B+2$ blocks.
For each $i\in\{1,2\}$, split Message $M_i$ into $B$ submessages $(m_{i}^{(1)}, \ldots, m_{i}^{(B)})$ each of rate $R_i/B$. For $b\in \{1,\dots,B\}$, split each of these submessages into a pair $(j_{i}^{(b)}, k_{i}^{(b)})$ of rates $R_i'/B$ and $(R_i-R_i')/B$, respectively. For blocks $b\in\{B+1,B+2\}$, set $j_1^{(b)}=k_1^{(b)}=j_2^{(b)}=k_2^{(b)}=1$.
For each block $b\in\{1,\dots,B+2\}$, at Transmitter $i$, for $i\in \{1,2\}$, the $j$-message $(j_i^{(b)})$ is transmitted using a feedback scheme and is going to be decoded at the other transmitter and the $k$-message $(k_i^{(b)})$ is transmitted without using the feedback and is decoded only at the receiver.
For each block $b\in \{1,\dots, B+2\}$, Transmitter~$i$, for $i\in\{1,2\},$ sends $\vect{x}_i^{(b)}(k_i^{(b)}|\vect{u}_i^{(b)}(j_i^{(b)}|\vect{v}_i^{(b)},\vect{w}^{(b)}))$.
The choice of the sequences $\vect{u}_i^{(b)}$, $\vect{v}_i^{(b)}$, and $\vect{w}^{(b)}$ is explained next.
After each block $b$, the receiver compresses the channel outputs $\vect{y}^{(b)}$ it observed for this block into $(\vect{y}_{12}^{(b)},\vect{y}_{1}^{(b)},\vect{y}_{2}^{(b)})$ using Heegard-Berger coding~\cite{HeegardBerger}. The receiver
uses the feedback links only once to send the same message $M_\textnormal{fb}^{(b)}$ which---as shortly explained ahead at the end of this section---is the triple $(r_{12}^{(b)},r_{1}^{(b)},r_{2}^{(b)})$, indices of bins containing the quantized output sequences $\vect{y}_{12}^{(b)}$, $\vect{y}_{1}^{(b)}$, and $\vect{y}_{2}^{(b)}$. Upon receiving $M_\textnormal{fb}^{(b)}$, Transmitter~$i$, for $i\in \{1,2\}$, reconstructs the sequence $\vect{y}_{12}^{(b)}$ by looking for a codeword $\vect{y}_{12}^{(b)}$ in bin $r_{12}^{(b)}$ jointly typical with $(\vect{w}^{(b)},\vect{x}_i^{(b)})$. Then, it looks for a codeword $\vect{y}_{i}^{(b)}$ in bin $r_{i}^{(b)}$ jointly typical with $(\vect{w}^{(b)},\vect{x}_i^{(b)},\vect{y}_{12}^{(b)})$.
For a given block~$b$, the messages $j_1^{(b)}$ and $j_{2}^{(b)}$ are at first transmitted using the $\vect{u}_1^{(b)}$- and $\vect{u}_2^{(b)}$-codewords of this block $b$. In contrast to the schemes by Cover \& Leung \cite{CoverLeung} or by Carleial \cite{Carleial}, the transmitters do not immediately decode each-other submessages $j_{1}^{(b)}$ or $j_{2}^{(b)}$ after learning the $(\vect{y}_{12}^{(b)},\vect{y}_{1}^{(b)},\vect{y}_{2}^{(b)})$ signals of block $b$. Instead, they wait for another block, where they exchange information helping them in the decoding. Specifically, at the end of each block $b\in \{1,\dots, B+1\}$, each Transmitter~$i$ computes \emph{transmitter-side} resolution information $\vect{v}_i^{(b+1)}$ as a (randomized) function of its block-$b$ codeword $\vect{u}_i^{(b)}$ and some common information $\vect{s}^{(b)}$ which is known to both transmitters. (The common information $\vect{s}^{(b)}$ consists of the $\vect{w}^{(b)}, \vect{v}_1^{(b)}, \vect{v}_2^{(b)}, \vect{y}_{12}^{(b)}
$ sequences; the former three are explained in the sequel.)
The resolution information $\vect{v}_i^{(b+1)}$ is then sent in block $b+1$. The initial sequences $(\vect{v}_1^{(1)},\vect{v}_2^{(1)})$ are drawn i.i.d. according to $P_{V_1V_2}$ and are known to everyone. In each block $b\in \{1,\dots,B+2\}$, the sequences $(\vect{v}_1^{(b)},\vect{v}_2^{(b)})$
are correlated, which makes that sending them can be more efficient than sending independent data. (In particular, Condition \eqref{eq:cond33} ensures that they have i.i.d. joint distribution $P_{V_1 V_2}$.)
After reception of $\vect{y}^{(b+1)}$, the receiver creates a list of the most likely $(j_1^{(b)}, j_2^{(b)})$ message pairs based on $\vect{y}^{(b)}$ and $\vect{y}^{(b+1)}$ (and all the information that it has for blocks $b$ and $b+1$).
Upon observing the feedback outputs in block~${b+1}$, Transmitter~$i$, for $\in\{1,2\}$, uses the sequences $\vect{y}_{12}^{(b)}$, $\vect{y}_i^{(b)}$, $\vect{y}_{12}^{(b+1)}$, and $\vect{y}_i^{(b+1)})$ (and all the information that it has for blocks $b$ and $b+1$) to estimate the decoder's list of highly-likely message-pairs $(j_1^{(b)}, j_2^{(b)})$. At the same time, it also estimates the other transmitter's $\vect{v}^{(b+1)}$-sequence and decodes the other transmitter's message $j^{(b)}$. (Notice that at this point, the receiver is hindered compared to the transmitters as it does know any of the $\vect{v}^{(b+1)}$-sequences.)
Conditions~\eqref{eq:r'11} and \eqref{eq:r'21} ensure that each transmitter decodes the other transmitter's $\vect{v}^{(b+1)}$-sequences correctly with high probability.
Therefore, besides having an estimate of the decoder's list, Transmitter $i$ also has an estimate of the position of the correct message pair therein. Let the index $\lambda_i^{(b+2)}$ describe this position. If there is no such index or if it exceeds $\lfloor 2^{nR_0}\rfloor$, then the index $\lambda_i^{(b+2)}$ is chosen uniformly at random from the set $\{1,\dots,\lfloor 2^{nR_0}\rfloor\}$.
If the feedback information and the $j$-messages were decoded correctly, with high probability we have $\lambda_1^{(b+2)}=\lambda_2^{(b+2)}$. We abuse notation and call this index $\lambda^{(b+2)}$. The two transmitters send this index jointly in block $b+2$ using a cooperation sequence $\vect{w}^{(b+2)}$ that plays the role of \emph{receiver-side} resolution information. For blocks $b\in\{1,2\}$, it is fixed and known to everyone.
Upon receiving $\vect{y}^{(b+2)}$, the receiver decodes the $\vect{w}^{(b+2)}$ codeword and the index $\lambda^{(b+2)}$ based on $\vect{w}^{(b+1)}$, $\vect{y}^{(b+1)}$, and $\vect{y}^{(b+2)}$, and uses $\lambda^{(b+2)}$ to identify the correct message-pair $(j_1^{(b)}, j_2^{(b)})$ within its list.
Condition~\eqref{eq:r0} ensures that $\vect{w}^{(b+2)}$ can be correctly decoded at the receiver. The receiver's list for block $b$ can be resolved by $\vect{w}^{(b+2)}$ with high probability if Conditions~\eqref{eq:r'12}-\eqref{eq:r'1+r'2} are satisfied.
Thereafter, the receiver also decodes with high probability the messages $k_1^{(b)}$ and $k_2^{(b)}$, encoded in the $\vect{x}_1^{(b)}$- and $\vect{x}_2^{(b)}$-codewords, based on $\vect{y}^{(b)}$ if Conditions~\eqref{eq:r1-r'1}-\eqref{eq:r-r'} are true.
To compress $\vect{y}^{(b)}$ at the end of block $b\in\{1,\ldots, B+1\}$, the receiver looks for a sequence $\vect{y}_{12}^{(b)}$ jointly typical with $\vect{y}^{(b)}$ and the decoded sequence $\vect{w}^{(b)}$. Then, for $i\in \{1,2\}$, it looks for a sequence $\vect{y}_{i}^{(b)}$ jointly typical with $(\vect{ y}_{12}^{(b)},\vect{y}^{(b)},\vect{w}^{(b)})$. Let $r_{12}^{(b)}$, $r_{1}^{(b)}$, $r_{2}^{(b)}$ denote the indices of the bins containing $\vect{y}_{12}^{(b)}$, $\vect{y}_{1}^{(b)}$, and $\vect{y}_{2}^{(b)}$, respectively. The Heegard-Berger coding~\cite{HeegardBerger} and Constraint~\eqref{eq:rfb} ensure that with high probability, Transmitter~$i$ can reconstruct $(\vect{y}_{12}^{(b)},\vect{y}_i^{(b)})$, for $i\in \{1,2\}$.
\section{Example: the Gaussian MAC}
\label{sec:gaussian}
\begin{figure*}
\addtocounter{equation}{-1}
\setcounter{equation}{24}
\begin{small}
\begin{subequations}
\label{eq:gaussian}
\begin{IEEEeqnarray}{rCl}
R_f&\geq&C\left(\frac{\sigma^2}{\sigma_{12}^2}+\frac{P}{\sigma_{12}^2}- \frac{P \theta}{\sigma_{12}^2} \left(1 - \alpha - \beta +\frac{ \beta^2 \theta
\lambda^2}{(1 - (1 - \alpha - \beta) \theta)}\right)\right)\nonumber\\
&&
+\sum_{i=1}^2 C\left(\frac{\sigma_{12}^2}{\sigma_{i}^2} \frac{\sigma^2 \left(1- (1 - \alpha - \beta) \theta\right) +
P \left((1 - (1 - \alpha - \beta) \theta)^2 - \beta^2 \theta^2 \lambda^2\right)}{(\sigma^2 + \sigma^2_{12})\left(1 - (1 - \alpha - \beta) \theta\right) +
P \left((1 - (1 - \alpha - \beta) \theta)^2 - \beta^2 \theta^2 \lambda^2\right)}\right)\\
R_1 &\leq& C\left(\frac{P (1-\theta)}{\sigma^2}\right)+\min\biggr\{C\left(\frac{P \theta\alpha}{\sigma^2+ \frac{\sigma_{12}^2\sigma_2^2 }{\sigma_{12}^2+\sigma_{2}^2}+ P (1-\theta)}
+\frac{ P \theta\beta (1+\lambda)} {\frac{\sigma^2_{12} \sigma^2_2}{\sigma^2_{12} + \sigma^2_2} + \sigma^2 + P (1-\theta)+P \alpha \theta }\right) ,C\left(\frac{\alpha P \theta}{2 P(1-\theta)+\sigma^2}\right)\nonumber\\&&+C\left(\frac{4 P\theta (1-\alpha-\beta)}{\sigma^2+2 P(1-\theta) +2 P \theta (\alpha+\beta (1+\lambda))}\right)+C\left(\frac{\beta P \theta (2 P (1-\theta)+\sigma^2)(1+\lambda) }{(2\alpha P \theta +2 P (1-\theta) +\sigma^2)(\sigma^2+2 P (1-\theta)+\alpha P \theta ) }\right)\biggr\},\\
R_2 &\leq& C\left(\frac{P (1-\theta)}{\sigma^2}\right)+\min\biggr\{C \left(\frac{P \theta\alpha}{\sigma^2+ \frac{\sigma_{12}^2\sigma_1^2 }{\sigma_{12}^2+\sigma_1^2}+ P (1-\theta)}
+\frac{ P \theta\beta (1+\lambda)} {\frac{\sigma^2_{12} \sigma^2_1}{\sigma^2_{12} + \sigma^2_1} + \sigma^2 + P (1-\theta)+P \alpha \theta }\right),C\left(\frac{\alpha P \theta}{2 P(1-\theta)+\sigma^2}\right)\nonumber\\
&&+C\left(\frac{4 P\theta (1-\alpha-\beta)}{\sigma^2+2 P(1-\theta) +2 P \theta (\alpha+\beta (1+\lambda))}\right)+C\left(\frac{\beta P \theta (2 P (1-\theta)+\sigma^2)(1+\lambda) }{(2\alpha P \theta +2 P (1-\theta) +\sigma^2)(\sigma^2+2 P (1-\theta)+\alpha P \theta ) }\right)\biggr\},\\
R_1+R_2&\leq & C\left(\frac{ 2 P (1-\theta)}{\sigma^2} \right)+C\left(\frac{2 P \theta (2-\alpha-\beta (1-\lambda))}{2 P (1-\theta )+\sigma^2}\right),\\
R_1+ R_2&\leq & C\left(\frac{ 2 P (1-\theta)}{\sigma^2} \right)+
\sum_{i=1}^2
\min\biggr\{C\left(\frac{P \theta\alpha}{\sigma^2+ \frac{\sigma_{12}^2\sigma_i^2 }{\sigma_{12}^2+\sigma_i^2}+ P (1-\theta)}+\frac{ P \theta\beta (1+\lambda)} {\frac{\sigma^2_{12} \sigma^2_i}{\sigma^2_{12} + \sigma^2_i} + \sigma^2 + P (1-\theta)+P \alpha \theta }\right),C\left(\frac{\alpha\theta P}{2 P(1-\theta)+\sigma^2}\right)\nonumber\\
&&+C\left(\frac{4 P\theta (1-\alpha-\beta)}{\sigma^2+2 P(1-\theta) +2 P \theta (\alpha+\beta (1+\lambda))}\right)+C\left(\frac{\beta P \theta (2 P (1-\theta)+\sigma^2)(1+\lambda) }{(2\alpha P \theta +2 P (1-\theta) +\sigma^2)(\sigma^2+2 P (1-\theta)+\alpha P \theta ) }\right)\biggr\},
\end{IEEEeqnarray}
\end{subequations}
\end{small}
\vspace*{-2mm}
\setcounter{equation}{24}
\hrule
\vspace*{-3mm}
\end{figure*}
\setcounter{equation}{12}
Consider a memoryless Gaussian MAC with symmetric input-power constraint $P$. The channel output is $Y=X_1+X_2+Z$, where $Z$ is zero-mean Gaussian with variance $\sigma^2>0$.
It can be shown that our coding scheme in Section~\ref{sec:coding} and Theorem~\ref{thm} in Section~\ref{sec:ach_reg} hold also for this Gaussian MAC.
Inspired by \cite{VenkatPradhan2011}, we propose the following choices.
Let $\alpha, \beta>0$ such that $\alpha+\beta <1$, $\theta \in [0,1]$, and $\sigma_{12},\sigma_1,\sigma_2>0$.
Also, let $\lambda \in[-1,1]$ so that
\vspace*{-1mm}
\begin{equation}
\label{eq:cond_lambda}
\lambda \leq \frac{P \theta \alpha}{\sigma^2+\sigma_{12}^2+ P \alpha \theta + 2 P (1-\theta)}.
\vspace*{-1mm}
\end{equation}
Now, let $W$, $A_1$, $A_2$, $I_{X_1}$, $I_{X_2}$, $\tilde W$, $\tilde A_1$, $\tilde A_2$, $\tilde I_{X_1}$, and $\tilde I_{X_2}$ be independent zero-mean standard Gaussians, and independent thereof and independent of each other, let $Z_{12},\tilde{Z}_{12}$ be zero-mean Gaussians of same variance $\sigma_{12}^2$, $Z_{1},\tilde{Z}_{1}$ be zero-mean Gaussians of same variance $\sigma_{1}^2$, and $Z_2,\tilde{Z}_2$ be zero-mean Gaussians of same variance $\sigma_2^2$,
and let $(\tilde V_1, \tilde V_2)$ be a centered bivariate Gaussian of covariance matrix
$\begin{pmatrix}1&\lambda\\\lambda&1\end{pmatrix}$.\\
Define for $i \in \{1,2\}$,
\vspace*{-1.5mm}
\begin{small}
\begin{IEEEeqnarray}{rCl}
\tilde U_i&\eqdef& \sqrt{\alpha} \tilde A_i+\sqrt{\beta} \tilde V_i+\sqrt{1-\alpha-\beta} \tilde W,\\
\tilde X_i&\eqdef & \sqrt{P (1-\theta)} \tilde I_{X_i}+ \sqrt{P \theta}\tilde U_i,\\
\tilde Y_i&\eqdef & \tilde Y+ \tilde Z_i,\\
\tilde Y_{12}&\eqdef & \tilde Y+ \tilde Z_{12}.
\end{IEEEeqnarray}
\end{small}
\vspace*{-1mm}
\\
Furthermore, define
\begin{small}
\vspace*{-1mm}
\begin{subequations}
\begin{IEEEeqnarray}{rCl}
V_1&\eqdef&\xi_1 \left(\tilde U_1-\sqrt{\beta } \tilde V_1-\sqrt{(1-\alpha -\beta)}\tilde W\right)/\sqrt{\alpha }+\xi_2 f(\tilde S)\IEEEeqnarraynumspace\\%\nonumber\\&=&\xi_1 \tilde A_1+\xi_2 f(\tilde S)\\
V_2&\eqdef&-\xi_1 \left(\tilde U_2-\sqrt{\beta } \tilde V_2-\sqrt{(1-\alpha -\beta)}\tilde W\right)/\sqrt{\alpha }-\xi_2 f(\tilde S) \IEEEeqnarraynumspac
\vspace*{-1mm}
\end{IEEEeqnarray}
\end{subequations}
\end{small}
\vspace*{-2mm}
\\
where the function $f$ is chosen as
\begin{small}
\begin{IEEEeqnarray}{rCl}
f(\tilde S)&=& f(\tilde W, \tilde V_1, \tilde V_2, \tilde Y_{12})\nonumber\\
&\eqdef&\frac{\tilde Y_{12}-\sqrt{\beta \theta P} (\tilde V_1+\tilde V_2)- 2 \sqrt{(1-\alpha-\beta)\theta P} \tilde W}{\sqrt{\sigma^2+\sigma_{12}^2+2 P \alpha \theta + 2 P (1-\theta)}}\IEEEeqnarraynumspac
\end{IEEEeqnarray}
\end{small}
and where $\xi_1,\xi_2 \in \Reals$ are chosen to satisfy
\begin{subequations}
\begin{small}
\begin{IEEEeqnarray}{rCl}
1&=& \xi_1^2+\xi_2^2+2\xi_1\xi_2 \sqrt{\frac{P \theta \alpha}{\sigma^2+\sigma_{12}^2+2 P \alpha \theta + 2 P (1-\theta)}},\IEEEeqnarraynumspace\label{eq:cond1}\\
\lambda&=&-2 \xi_1 \xi_2 \sqrt{\frac{P \theta \alpha}{\sigma^2+\sigma_{12}^2+2 P \alpha \theta + 2 P (1-\theta)}} -\xi_2^2.
\end{IEEEeqnarray}
\end{small}
\end{subequations}
\vspace*{-2mm}
\\(Condition~\eqref{eq:cond_lambda} ensures that such real $\xi_1$ and $\xi_2$ exist. In general, there are four possible choices for $\xi_1,\xi_2$. The specific choice of $\xi_1, \xi_2$ does not show up in the rate-constraints~\eqref{eq:gaussian} and does not change the set of achievable rates.)
For these choices define for $i \in \{1,2\}$,\vspace*{-1mm}
\begin{small}
\begin{IEEEeqnarray}{rCl}
U_i&\eqdef& \sqrt{\alpha} A_i+\sqrt{\beta} V_i+\sqrt{1-\alpha-\beta} W,\\
X_i&\eqdef & \sqrt{P (1-\theta)} I_{X_i}+ \sqrt{P \theta} U_i,\\
Y_i&\eqdef & Y+ Z_i,\\
Y_{12}&\eqdef & Y+ Z_{12}.
\vspace*{-1mm}
\end{IEEEeqnarray}
\end{small}
\indent Substituting the above choice into the rate-constraints of Theorem~\ref{thm}, we obtain that all nonnegative rate pairs $(R_1,R_2)$ satisfying Constraints~\eqref{eq:gaussian} on top of this page are achievable.
In~\eqref{eq:gaussian} we use the notation $C(x)\eqdef \frac12 \log(1+x)$.
\begin{figure}
\begin{center}
\includegraphics[height=5cm,width=8.5cm]{P100N20RF2.eps}
\end{center}
\vspace*{-5mm}
\caption{Achievable regions for the Gaussian MAC for $P/\sigma^2=5$ and $R_\textnormal{fb}=2$\label{figures2}}
\vspace*{-5mm}
\end{figure}
Figure~\ref{figures2} compares the achievable region in~\eqref{eq:gaussian}, to the nofeedback capacity region, the perfect feedback capacity region~\cite{Ozarow84}, the Cover-Leung~\cite{CoverLeung} and Venkataramanan-Pradhan~\cite{VenkatPradhan2011} regions for perfect feedback, and to the Shaviv-Steinberg region with rate-limited feedback~\cite{SteinbergShaviv}.
For the sake of simplicity we restrict to the case where only common feedback is present ($Y_1=Y_2=\emptyset, \sigma_1^2,\sigma_2^2\to \infty$) which reduces to Wyner-Ziv coding~\cite{WynerZiv} over the feedback links. In this case, we need to have $\sigma_{12}^2 \geq \frac{\sigma^2+P}{2^{2R_\textnormal{fb}}-1}$ and we see that our scheme is strictly better in terms of sum-rate than the Shaviv-Steinberg scheme.
In fact, based on extensive simulations, we conjecture that this is the case whenever
$\frac{P}{\sigma^2} < 2^{2 R_\textnormal{fb}}-2$,
which is equivalent to $\sigma_{12}^2 < \sigma^2$.
\section*{Acknowledgements}
The author would like to thank M.~Wigger for helpful discussions and the city of Paris for supporting this work under the ``Emergences'' program.
\addcontentsline{toc}{chapter}{\numberline{}Bibliography}
|
\section{Introduction}\label{sect:intro}
In the classical representation theory of finite semigroups, a key role is played by the so-called \emph{Munn rings}. These are rings of $m\times n$ matrices (where $m$ and $n$ need not be equal) with the familiar addition operation but with a \emph{sandwich} multiplication defined by $X\star Y=XAY$, where $A$ is a fixed $n\times m$ matrix. These rings are so named, because of Douglas Munn's 1955 paper \cite{Munn1955}, in which it was shown that: (1) the representation theory of a finite semigroup is determined by the representations of certain \emph{completely $0$-simple semigroups} arising from its ideal structure, and (2) the semigroup algebra of such a finite completely $0$-simple semigroup is isomorphic to an appropriate Munn ring over the group algebra of a naturally associated maximal subgroup; conditions were also given for such a Munn ring to be semisimple. (Here, the \emph{sandwich matrix} $A$ arises from the celebrated Rees structure theorem \cite{Rees1940} for completely $0$-simple semigroups.) Since their introduction in \cite{Munn1955}, Munn rings have been studied by numerous authors, and continue to heavily inflence the theory of semigroup representations: for classical studies, see \cite{Munn1957,Munn1955,Munn1960, Ponizovskii1956, LP1969,Hall1970, McAlister1971, McAlister1971a,McAlister1971b, Clifford1942,Clifford1960,CPbook}; for modern accounts, see for example \cite{IRS2011,GMS2009,Putcha1998,Putcha1996,OP1991,Steinberg2006,Steinberg2008,AMSV2009}, and especially the monographs \cite{Okninski1991,Okninski1998,Renner2005,Putcha1988,SteinbergBook}.
In the same year as Munn's article \cite{Munn1955} was published, William Brown introduced the so-called \emph{generalised matrix algebras}~\cite{Brown1955}, motivated by a connection with classical groups \cite{Brown1956,Brauer1937,Weyl1939}. These generalised matrix algebras are again rings of $m\times n$ matrices over a field, with multiplication determined by a fixed $n\times m$ \emph{sandwich matrix}. Whereas the sandwich matrix in a Munn ring is taken to be the structure matrix of a completely $0$-simple semigroup (and so has a certain prescribed form), Brown considered arbitrary sandwich matrices. As with Munn rings, these generalised matrix algebras have influenced representation theory to this day, and have been studied by numerous authors; see for example \cite{DL2004,Thrall1955,XW2010,LW2012,XW2014,GW2015,KX2001,Gavarini2008,KX1998}.
Shortly after the Munn and Brown articles \cite{Munn1955,Brown1955} appeared, Evgeny Lyapin's early monograph on semigroups~\cite{Lyapin} was published. In \cite[Chapter VII]{Lyapin}, we find a number of interesting semigroup constructions, including the following. Let $V$ and $W$ be arbitrary non-empty sets, and let $\theta:W\to V$ be an arbitrary (but fixed) function. Then the set $\T(V,W)$ of all functions $V\to W$ forms a semigroup, denoted $\T^\theta(V,W)$, under the operation $\star_\theta$ defined by $f\star_\theta g=f\circ\theta\circ g$.
If it is assumed that $V$ and $W$ are vector spaces (over the same field) and~$\theta$ a linear transformation, then the subset $\mathcal L(V,W)\sub\T(V,W)$ of all linear transformations $V\to W$ is a subsemigroup of $\T^\theta(V,W)$. This subsemigroup, denoted $\mathcal L^\theta(V,W)$ and referred to as a \emph{linear sandwich semigroup}, is clearly isomorphic to the underlying multiplicative semigroup of an associated generalised matrix algebra \cite{Brown1955}. As noted above, the addition on a generalised matrix algebra is just the usual operation, so these linear sandwich semigroups capture and isolate (in a sense) the more complex of the operations on the algebras.
The sandwich semigroups $\T^\theta(V,W)$ were first investigated in a series of articles by Magill and Subbiah \cite{Magill1967,MS1975,MS1978}, and more recent studies may be found in \cite{Sullivan_preprint,MGS2013,CC2008,WK2002}; most of these address structural concerns such as (von Neumann) regularity, Green's relations, ideals, classification up to isomorphism, and so on.
The linear sandwich semigroups $\mathcal L^\theta(V,W)$ have received less attention, though they have also been studied by a number of authors \cite{MGS2014,Kemprasit2002,Chinram2009,JCK2010}, with studies again focusing on basic structural properties. This is regrettable, because these semigroups display a great deal of algebraic and combinatorial charm, as we hope to show in the current article.
It is therefore our purpose to carry out a systematic investigation of the linear sandwich semigroups, bringing their study up to date, and focusing on modern themes, especially combinatorial invariant theory. As does Brown \cite{Brown1955}, we focus on the case that $V$ and $W$ are finite dimensional; in fact, we study the equivalent sandwich semigroups $\MmnA=\MmnA(\F)$ consisting of all $m\times n$ matrices over the field $\F$ under the operation~$\star_A$ defined by $X\star_AY=XAY$, where $A$ is a fixed $n\times m$ matrix.
We speculate that the difficulty (until now) of systematically investigating the linear sandwich semigroups may be due to the lack of a consistent theoretical framework for studying sandwich semigroups in more generality. In the case that $V=W$, the sets $\T(V,W)$ and $\mathcal L(V,W)$ are themselves semigroups (under composition); these are the \emph{full transformation semigroup} $\T_V$ \cite{GMbook,MGS2013,Gomes1987,GR2012,EMP2015, Howie1990,Howie1966,Howie1978} and the \emph{general linear monoid} $\mathcal L_V$ \cite{DG2014,Djokovic1968,Erdos1967,Dawlings81/82,AM2005, Dawlings1982,Gray2008,Waterhouse,Putcha2006, Laffey1983,Okninski1998}, respectively. In turn, the semigroups $\mathcal T^\theta(V,V)$ and $\mathcal L^\theta(V,V)$ are special cases of the \emph{semigroup variant} construction. The \emph{variant} of a semigroup $S$ with respect to an element $a\in S$ is the semigroup $S^a=(S,\star_a)$, with operation defined by $x\star_ay=xay$. Variants were first explicitly studied by Hickey in the 1980s \cite{Hickey1983,Hickey1986}, though (as noted above) the idea goes back to Lyapin's monograph \cite{Lyapin}; a more recent study may be found in \cite{KL2001}. The current authors developed the general theory of variants further in \cite{DE2}, and then used this as a starting point to explore the variants of the finite full transformation semigroups, obtaining a great deal of algebraic and combinatorial information about these semigroups. Unfortunately, the theory of semigroup variants does not help with studying the more general sandwich semigroups $\mathcal T^\theta(V,W)$ and $\mathcal L^\theta(V,W)$, since the underlying sets $\T(V,W)$ and $\mathcal L(V,W)$ are not even semigroups if $V\not=W$. One of the main goals of the current article, therefore, is to develop an appropriate general framework for working with arbitrary sandwich semigroups. Namely, if $V$ and $W$ are objects in a (locally) small category $\C$, and if $\theta\in\Hom(W,V)$ is some fixed morphism, then the set $\Hom(V,W)$ becomes a semigroup under the sandwich operation defined by $f\star_\theta g=f\circ\theta\circ g$, for $f,g\in\Hom(V,W)$. (In the case that $V=W$ and $\theta$ is the identity morphism, this construction reduces to the usual endomorphism monoid $\End(V)$.) The semigroups $\mathcal T^\theta(V,W)$ and $\mathcal L^\theta(V,W)$ arise when $\C$ is the category of sets (and mappings) or vector spaces (and linear transformations), respectively. In order to develop a general theory of sandwich semigroups in such categories, we first explain how many important semigroup theoretical techniques
extend to the more general categorical setting; we note that there is only a little overlap with the theory of Green's relations in categories developed in \cite{LS2012}, which focuses on issues more relevant to representation theory. In order to avoid any confusion arising from terminology conflicts between semigroup and category theory, rather than speak of (locally small) categories, we focus on the equivalently defined class of \emph{partial semigroups}, which are related to Ehresmann-style ``arrows only'' categories \cite{Ehresmann1965}. We hope that the general theory we develop will prove to be a useful starting point for future studies of sandwich semigroups in other categories.
The article is organised as follows.
In Section \ref{sect:partial}, we develop a general theory of sandwich semigroups in partial semigroups (i.e., locally finite categories), extending certain important semigroup theoretic notions (such as Green's relations, regularity and stability, the definitions of which are given in Section~\ref{sect:partial}) to the more general context.
In Section \ref{sect:preliminaries}, we gather results on the partial semigroup~$\M=\M(\F)$ of all (finite dimensional) matrices over the field $\F$, mainly focusing on regularity, stability and Green's relations, and we state some well-known results on (idempotent) generation and ideals of the general linear monoids~$\M_n$.
We begin our investigation of the linear sandwich semigroups $\MmnA$ in Section \ref{sect:MmnJ}, the main results of this section being: a characterisation of the regular elements (Proposition~\ref{prop:P1P2}); a description of Green's relations (Theorem~\ref{green_thm}) and the ordering on $\D$-classes (Propositions \ref{prop:DorderMmnJ}, \ref{prop:DorderP} and \ref{prop_maximalD}); a classification of the isomorphism classes of sandwich semigroups over $\Mmn$ (Corollary \ref{cor:MmnAcongMmnB}); and the calculation of $\rank(\MmnA)$ (Theorems~\ref{thm:rankMmnJ} and~\ref{thm:rankMmnJ_r=m}).
(Recall that the \emph{rank} of a semigroup $S$, denoted $\rank(S)$, is the minimum size of a generating set for $S$.)
Section \ref{sect:non-sandwich} explores the relationship between a sandwich semigroup $\MmnA$ and various (non-sandwich) matrix semigroups, the main structural results being Theorem \ref{thm:diamondsMmnJ} and Propositions~\ref{mono_prop} and~\ref{prop:MAN}.
We then focus on the regular subsemigroup $P=\Reg(\MmnA)$ in Section \ref{sect:RegMmnJ}, where we:
calculate the size of $P$ and various Green's classes (Proposition \ref{prop:DXJ_combinatorics} and Theorem \ref{inflation_thm}); classify the isomorphism classes of finite linear sandwich semigroups (Theorem \ref{thm:classification}); and calculate $\rank(P)$ (Theorem \ref{thm:rankP}).
In Section \ref{sect:EMmnJ}, we investigate the idempotent generated subsemigroup $\E_{mn}^A$ of $\MmnA$, where we: enumerate the idempotents of $\MmnA$ (Proposition \ref{prop:enumeration_E}); show that $\E_{mn}^A$ consists of $P\sm D$ and the idempotents from $D$, where $D$ is the maximal $\D$-class (Theorem \ref{thm:EmnJ}); and calculate $\rank(\E_{mn}^A)$ and $\idrank(\E_{mn}^A)$, showing in particular that these are equal (Theorem \ref{thm:rankEmnJ}).
(The \emph{idempotent rank} of an idempotent generated semigroup $S$, denoted $\idrank(S)$, is defined similarly to the rank, but with respect to idempotent generating sets for $S$.)
Finally, in Section \ref{sect:ideals}, we classify the proper ideals of $P$, showing that these are idempotent generated, and calculating their ranks and idempotent ranks, which are again equal (Theorem~\ref{thm:ideals}).
We note that all our results have applications to the \emph{variants} $\M_n^A$ of the full linear monoid $\M_n$ (in the case $m=n$), and to certain semigroups of linear transformations of restricted range or kernel (in the case that $\rank(A)$ is equal to one of $m,n$; see Remarks~\ref{rem:r=m} and \ref{rem:r=m2}).%
\begin{comment}
\newpage
\section{Old into}
; see for example
\cite{McAlister1971,CPbook,Renner2005,Putcha1988} and the many references to be found in \cite{AMSV2009}.
In addition, there is a vast literature on the full linear monoids themselves; see for example \cite{Sullivan2008,SS2014,DG2014,Djokovic1968,Erdos1967,Dawlings81/82,AM2005, Dawlings1982,Gray2008,Waterhouse,Putcha2006, Laffey1983} and especially the monograph \cite{Okninski1998}.
Many of these studies have been inspired by topics in transformation semigroup theory
\cite{MGS2013,Gomes1987,GR2012,Howie1990,Howie1966,Howie1978},
since the {full transformation semigroup} $\T_n$ (consisting of all self maps of an $n$-set) embeds naturally in $\M_n$ (for any $\F$), analogously to the way the {symmetric group} $\S_n$ embeds in~$\G_n$. Thus, research on the full linear monoids has concentrated on themes including (but in no way limited to): generators; singular matrices and products of idempotents; free idempotent generated semigroups and biordered sets; Green's relations, maximal subgroups and ideals; and the calculation of various combinatorial invariants such as ranks and idempotent ranks. (The \emph{rank} of a semigroup $S$, denoted $\rank(S)$, is the smallest cardinality of a generating set for $S$. The \emph{idempotent rank} of an idempotent generated semigroup $S$, denoted $\idrank(S)$, is the smallest cardinality of a generating set for $S$ consisting entirely of idempotents.)
In a recent article \cite{DE2}, the current authors considered the so-called \emph{variants} of the finite full transformation semigroups; see also \cite{Tsyaputa2004,Tsyaputa2005,GMbook} for prior studies. Recall that the variant of a semigroup $S$ with respect to an element $a\in S$ is the semigroup $S^a=(S,\star_a)$, where the operation $\star_a$ is defined by $x\star_ay=xay$ for all $x,y\in S$. Variants were first explicitly studied by Hickey \cite{Hickey1983,Hickey1986} (see also \cite{KL2001}), but the notion goes back to the more general concept of a \emph{sandwich semigroup}. It seems that concrete sandwich semigroups of functions and linear transformations were first mentioned in the 1960 monograph of Lyapin \cite{Lyapin}, though the latter are closely related to the so-called \emph{generalized matrix algebras} introduced by Brown \cite{Brown1955} in 1955; for recent studies of these algebras, see \cite{XW2010,LW2012,XW2014} and references therein. (We also note the similarities to the famous \emph{Rees matrix semigroups} \cite{Rees1940}.) If $V$ and $W$ are arbitrary non-empty sets, and if $\theta:W\to V$ is an arbitrary (but fixed) function, then the set $\T(V,W)$ of all functions $V\to W$ forms a semigroup under the operation $\star_\theta$ defined by $f\star_\theta g=f\theta g$ for all $f,g\in\T(V,W)$.\footnote{Throughout the article, we write $f(x)$ for the image of $x$ under the function $f$, and we compose functions right-to-left.} (When $V=W$, we obtain the variants studied in \cite{DE2}.) If it is assumed that $V$ and $W$ are vector spaces (over the same field) and~$\theta$ a linear transformation, then the set $\mathcal L(V,W)$ of all linear transformations $V\to W$ is a subsemigroup of $\T(V,W)$.
The sandwich semigroups $\T(V,W)$ were first systematically investigated in a series of articles by Magill and Subbiah \cite{Magill1967,MS1975,MS1978}; for more recent studies, see for example \cite{Sullivan_preprint,MGS2013,CC2008,WK2002}.
To the authors' knowledge, there exist only a small number of papers on the linear sandwich semigroups $\mathcal L(V,W)$, mostly concerning Green's relations and regularity \cite{MGS2014,Kemprasit2002,Chinram2009}; we are also aware of one brief article on variants of the full linear monoids~$\M_n$ \cite{JCK2010}.
It is the purpose of the current article to carry out a systematic investigation of the linear sandwich semigroups in the case that $V$ and $W$ are finite dimensional. (In fact, we study the equivalent sandwich semigroups $\MmnA$ consisting of all $m\times n$ matrices under the operation~$\star_A$ defined by $X\star_AY=XAY$, where $A$ is a fixed $n\times m$ matrix.)
We also identify an appropriate setting in which to carry out a more general study of sandwich semigroups: namely, in a suitably defined class of \emph{partial semigroups} (see Definition \ref{defn:S}) that are related to Ehresmann-style ``arrows only'' categories \cite{Ehresmann1965}.
In particular, our approach via partial semigroups leads to a general method for investigating regularity and Green's relations in arbitrary sandwich semigroups.
The article is organised as follows.
In Section \ref{sect:partial}, we develop a general theory of {sandwich semigroups} in the context of a suitably defined class of {partial semigroups}, extending certain important semigroup theoretic notions (such as Green's relations, regularity and stability, the definitions of which are given in Section~\ref{sect:partial}) to the context of partial semigroups and sandwich semigroups.
In Section \ref{sect:preliminaries}, we gather results on the partial semigroup~$\M$ of all (finite dimensional) matrices, mainly focusing on regularity, stability and Green's relations, and we state some well-known results on (idempotent) generation and ideals of $\M_n$.
We begin our investigation of the linear sandwich semigroups $\MmnA$ in Section \ref{sect:MmnJ}, the main results of this section being: a characterisation of the regular elements (Proposition~\ref{prop:P1P2}); a description of Green's relations (Theorem~\ref{green_thm}) and the order on $\D$-classes (Propositions \ref{prop:DorderMmnJ}, \ref{prop:DorderP} and \ref{prop_maximalD}); a classification of the isomorphism classes of sandwich semigroups over $\Mmn$ (Corollary \ref{cor:MmnAcongMmnB}); and the calculation of $\rank(\MmnA)$ (Theorems~\ref{thm:rankMmnJ} and~\ref{thm:rankMmnJ_r=m}).
Section \ref{sect:non-sandwich} explores the relationship between a sandwich semigroup $\MmnA$ and various (non-sandwich) matrix semigroups, the main structural results being Theorem \ref{thm:diamondsMmnJ} and Propositions~\ref{mono_prop} and~\ref{prop:MAN}.
We then focus on the regular subsemigroup $P=\Reg(\MmnA)$ in Section \ref{sect:RegMmnJ}, where we:
calculate the size of $P$ and various Green's classes (Proposition \ref{prop:DXJ_combinatorics} and Theorem \ref{inflation_thm}); classify the isomorphism classes of finite linear sandwich semigroups (Theorem \ref{thm:classification}); and calculate $\rank(P)$ (Theorem \ref{thm:rankP}).
In Section \ref{sect:EMmnJ}, we investigate the idempotent generated subsemigroup $\E_{mn}^A$ of $\MmnA$, where we: enumerate the idempotents of $\MmnA$ (Proposition \ref{prop:enumeration_E}); show that $\E_{mn}^A$ consists of $P\sm D$ and the idempotents from $D$, where $D$ is the maximal $\D$-class (Theorem \ref{thm:EmnJ}); and calculate $\rank(\E_{mn}^A)$ and $\idrank(\E_{mn}^A)$, showing in particular that these are equal (Theorem \ref{thm:rankEmnJ}).
Finally, in Section \ref{sect:ideals}, we classify the proper ideals of $P$, showing that these are idempotent generated, and calculating their ranks and idempotent ranks, which are again equal (Theorem~\ref{thm:ideals}).
We note that all our results have applications to the \emph{variants} $\M_n^A$ of the full linear monoid $\M_n=\M_{nn}$ (in the case $m=n$), and to certain semigroups of linear transformations of restricted range or kernel (in the case that $\rank(A)$ is equal to one of $m,n$; see Remarks \ref{rem:r=m} and \ref{rem:r=m2}).%
\footnote{The authors wish to thank Dr Attila Egri-Nagy for constructing the GAP \cite{GAP} code that enabled us to produce the eggbox diagrams from several figures in this article.}
\end{comment}
\section{Sandwich semigroups from partial semigroups}\label{sect:partial}
Recall that our main interest is in the linear sandwich semigroups $\MmnA=\MmnA(\F)$. The underlying set of $\MmnA$ is $\Mmn$, the set of all $m\times n$ matrices over the field $\F$, which is not itself a semigroup (unless $m=n$). However, $\Mmn$ is contained in $\M$, the set of \emph{all} (finite dimensional) matrices over $\F$. While $\M$ is still not a semigroup, it does have the structure of a
(small) category. As we will see, in order to understand the linear sandwich semigroups $\MmnA$, we need to move beyond just $m\times n$ (and $n\times m$) matrices, and gain a fuller understanding of the whole category $\M$. Some (but not all) of what we need
to know about $\M$ is true in a larger class of categories, and more general structures we call \emph{partial semigroups}, so we devote this section to the development of the general theory of these structures. We begin with the definitions.
\ms
\begin{defn}\label{defn:S}
A \emph{partial semigroup} is a $5$-tuple $(S,\cdot,I,\lam,\rho)$ consisting of a set $S$, a partial binary operation $(x,y)\mt x\cdot y$ (defined on some subset of $S\times S$), a set $I$, and functions $\lam,\rho:S\to I$, such that, for all $x,y,z\in S$,
\bit
\item[(i)] $x\cdot y$ is defined if and only if $\rho(x)=\lam(y)$,
\item[(ii)] if $x\cdot y$ is defined, then $\lam(x\cdot y)=\lam(x)$ and $\rho(x\cdot y)=\rho(y)$,
\item[(iii)] if $x\cdot y$ and $y\cdot z$ are defined, then $(x\cdot y)\cdot z=x\cdot (y\cdot z)$.
\eit
We say that a partial semigroup $(S,\cdot,I,\lam,\rho)$ is \emph{monoidal} if in addition to (i--iii),
\bit
\item[(iv)] there exists a function $I\to S:i\mt e_i$ such that, for all $x\in S$, $x\cdot e_{\rho(x)}=x=e_{\lam(x)}\cdot x$.
\eit
We say that a partial semigroup $(S,\cdot,I,\lam,\rho)$ is \emph{regular} if in addition to (i--iii),
\bit
\item[(v)] for all $x\in S$, there exists $y\in S$ such that $x=x\cdot y\cdot x$ and $y=y\cdot x\cdot y$.
\eit
\end{defn}
\ms
{\begin{rem}
We note that conditions (i--iv) amount to one of several equivalent ways to define (small) categories in an ``arrows only'' fashion. See for example Ehresmann's monograph \cite{Ehresmann1965}, and also \cite{Hollings2012} for a historical discussion of the connections between category theory and (inverse) semigroup theory.
\end{rem}
For a partial semigroup $(S,\cdot,I,\lam,\rho)$, and for $i,j\in I$, we write
\[
S_{ij} = \set{x\in S}{\lam(x)=i,\ \rho(x)=j} \AND S_i = S_{ii}.
\]
So $S=\bigcup_{i,j\in I}S_{ij}$. Note that if $x\in S$, then $x\cdot x$ is defined if and only if $\lam(x)=\rho(x)$. It follows that $S_i$ is a semigroup with respect to the induced binary operation (the restriction of $\cdot$ to $S_i\times S_i$) for each $i\in I$, but that $S_{ij}$ is not if $i\not=j$. We will often slightly abuse notation and refer to ``the partial semigroup $S$'' if the rest of the data $(S,\cdot,I,\lam,\rho)$ is clear from context. We also note that in what follows, we could allow~$S$ and~$I$ to be classes (rather than insist on them being sets); but we would still require $S_{ij}$ to be a set for each $i,j\in I$.
Note that, as is the case with semigroups, condition (v) is equivalent to the (ostensibly) weaker condition:
\bit
\item[(v)$'$] for all $x\in S$, there exists $z\in S$ such that $x=x\cdot z\cdot x$.
\eit
Indeed, with $z$ as in (v)$'$, one easily checks that $y=z\cdot x\cdot z$ satisfies the condition of (v).
If $S$ is monoidal, then $S_i$ is a monoid with identity $e_i\in S_i$ for each $i$.
If $S$ is not monoidal, then $S$ may be embedded in a monoidal partial semigroup $\Sone$ as follows: for each $i\in I$ we adjoin an element $e_i$ to $S_i$ and declare that $x\cdot e_i=x$ and $e_i\cdot y=y$ for all $x,y\in S$ with $\rho(x)=i$ and $\lam(y)=i$, if such an element $e_i\in S_i$ does not already exist. In particular, if $S$ is monoidal, then $S=\Sone$.
Obviously any semigroup is a partial semigroup (with $|I|=1$); in particular, all results we prove in this section concerning partial semigroups hold for semigroups.
A great number of non-semigroup examples exist, but we will limit ourselves to describing just a few.
\ms
\begin{eg}
As a trivial example, let $\set{S_i}{i\in I}$ be any set of pairwise disjoint semigroups. Then $S=\bigcup_{i\in I}S_i$ is a partial semigroup where we define $\lam,\rho:S\to I$ by $\lam(x)=\rho(x)=i$ for each $i\in I$ and $x\in S_i$, and $x\cdot y$ is defined if and only if $x,y\in S_i$ for some $i$, in which case $x\cdot y$ is just the product of $x,y$ in $S_i$. Note that this $S$ is regular (resp., monoidal) if and only if each $S_i$ is regular (resp., a monoid).
\end{eg}
\newcommand{\X}{\mathscr X}
\ms
\begin{eg}
Let $\X$ be some set, and $\P(\X)=\set{A}{A\sub\X}$ the power set of $\X$. The set $T_\X=\set{(B,f,A)}{A,B\sub\X,\ \text{$f$ is a function $A\to B$}}$ is a regular monoidal partial semigroup. We define $I=\P(\X)$, and $\lam(B,f,A)=B$ and $\rho(B,f,A)=A$, with $(D,g,C)\cdot(B,f,A)$ defined if and only if $B=C$, in which case $(D,g,C)\cdot(B,f,A)=(D,g\circ f,A)$.
\end{eg}
The previous example may be extended in a number of ways, by replacing functions $f:A\to B$ by other objects such as binary relations \cite{Thornton1982, Chase1979}, partial functions \cite{Sullivan1975,Chinram2008}, partial bijections \cite{Chinram2008b}, block bijections \cite{FL1998}, partial braids \cite{EL2004}, partitions \cite{Martin1994}, Brauer diagrams \cite{Brauer1937}, etc., or by assuming the functions $f:A\to B$ preserve some kind of algebraic or geometric structure on the sets $A,B$. The main example we will concentrate on in this article is as follows.
\ms
\begin{eg}\label{eg:M}
Let $\F$ be a field, and write $\M=\M(\F)$ for the set of all (finite dimensional, non-empty) matrices over~$\F$. Then $\M$ has the structure of a regular monoidal partial semigroup. We take $I=\mathbb N=\{1,2,3,\ldots\}$ to be the set of all natural numbers and, for $X\in\M$, we define $\lam(X)$ (resp., $\rho(X)$) to be the number of rows (resp., columns) of $X$. For $m,n\in\mathbb N$, $\Mmn=\Mmn(\F)$ denotes the set of all $m\times n$ matrices over $\F$, and forms a semigroup if and only if $m=n$. (Of course, $\M$ is isomorphic to a certain partial semigroup of linear transformations; we will have more to say about this later.)
\end{eg}
For the remainder of this section, we fix a partial semigroup $(S,\cdot,I,\lam,\rho)$, and we write $xy$ for the product $x\cdot y$ (whenever it is defined).
Note that we may define a second partial binary operation $\bullet$ on $S$ by
\[
x\bullet y=y\cdot x \qquad\text{for each $x,y\in S$ with $\rho(y)=\lam(x)$.}
\]
We see then that $(S,\bullet,I,\rho,\lam)$ is a partial semigroup (note the swapping of $\lam$ and $\rho$), and we call this the \emph{dual partial semigroup} to $(S,\cdot,I,\lam,\rho)$. As is frequently the case in semigroup theory, this duality will allow us to shorten several proofs.
Green's relations and preorders are crucial tools in semigroup theory (for general background on semigroups, see \cite{Hig,Howie}), and we will need to extend these to the partial semigroup setting.
If $x,y\in S$, then we say
\bit
\item $x\leq_\R y$ if $x=ya$ for some $a\in\Sone$,
\item $x\leq_\L y$ if $x=ay$ for some $a\in \Sone$,
\item $x\leq_\J y$ if $x=ayb$ for some $a,b\in\Sone$.
\eit
Note that if $x\leq_\R y$ (resp., $x\leq_\L y$), then $\lam(x)=\lam(y)$ (resp., $\rho(x)=\rho(y)$).
Note also that if $x\leq_\R y$, then $ux\leq_\R uy$ for any $u\in S$ with $\rho(u)=\lam(x)$; a dual statement holds for the $\leq_\L$ relation.
Finally, note that the use of $\Sone$ is merely for convenience since, for example, $x\leq_\R y$ means that $x=y$ or $x=ya$ for some $a\in S$.
All three of the above relations are preorders (i.e., they are reflexive and transitive).
If $\gK$ is one of $\R$, $\L$, $\J$, we write $\gK={\leq_\gK}\cap{\geq_\gK}$ for the equivalence relation on $S$ induced by $\gK$. So, for example, $x\R y$ if and only if $x=ya$ and $y=xb$ for some $a,b\in \Sone$.
We also define equivalence relations
\[
\H=\R\cap\L \AND \D=\R\vee\L.
\]
(The join $\ve\vee\eta$ of two equivalences $\ve$ and $\eta$ is the transitive closure of $\ve\cup\eta$, and is itself an equivalence.)
It is easy to see that $\D\sub\J$.
The duality mentioned above means that $x\leq_\R y$ in $(S,\cdot,I,\lam,\rho)$ if and only if $x\leq_\L y$ in $(S,\bullet,I,\rho,\lam)$, and so on.
Analogously to the definition for semigroups \cite[Definition A.2.1]{RSbook}, we say that the partial semigroup $S$ is \emph{stable} if for all $x,y\in S$,
\[
x\J xy \ \iff\ x\R xy \AND x\J yx\ \iff\ x\L yx.
\]
The following simple but crucial observation is proved in analogous fashion to the corresponding results for semigroups; see for example \cite[Proposition 2.1.3]{Howie} and \cite[Corollary A.2.5]{RSbook}.
\ms
\begin{lemma}\label{lem:Rol=LoR_S}
We have $\D=\R\circ\L=\L\circ\R$. If $S$ is stable, then $\D=\J$. \epfres
\end{lemma}
If $x\in S_{ij}$ and $\gK$ is one of $\R$, $\L$, $\J$, $\D$, $\H$, we write
\[
\xK = \set{y\in S}{x\gK y} \AND K_x=\xK\cap S_{ij} = \set{y\in S_{ij}}{x\gK y}.
\]
We call $\xK$ (resp., $K_x$) the \emph{$\gK$-class of $x$ in $S$} (resp., \emph{in $S_{ij}$}).
The next result is reminiscent of Green's Lemma, and may be proved in virtually identical fashion to \cite[Lemma 2.2.1]{Howie}.
\ms
\begin{lemma}\label{lem:GreensLemma_S}
Let $x,y\in S$.
\bit
\itemit{i} Suppose $x\R y$, and that $x=ya$ and $y=xb$ where $a,b\in\Sone$. Then the maps $\xL\to \yL:w\mt wb$ and $ \yL\to \xL:w\mt wa$ are mutually inverse bijections. These maps restrict to mutually inverse bijections $\xH\to \yH$ and $\yH\to \xH$.
\itemit{ii} Suppose $x\L y$, and that $x=ay$ and $y=bx$ where $a,b\in\Sone$. Then the maps $\xR\to \yR:w\mt bw$ and $\yR\to \xR:w\mt aw$ are mutually inverse bijections. These maps restrict to mutually inverse bijections $\xH\to \yH$ and $\yH\to \xH$.
\itemit{iii} If $x\D y$, then $\big|\xR\big|=\big|\yR\big|$, $\big|\xL\big|=\big|\yL\big|$ and $\big|\xH\big|=\big|\yH\big|$. \epfres
\eitres
\end{lemma}
Note that if $x,y\in S$ are such that $x\H y$, then $\lam(x)=\lam(y)$ and $\rho(x)=\rho(y)$. It follows that $\xH=H_x$ for all $x\in S$.
\ms
\begin{lemma}\label{lem:GreensLemma_Sij}
Let $x,y\in S_{ij}$.
\bit
\itemit{i} Suppose $x\R y$, and that $x=ya$ and $y=xb$ where $a,b\in\Sone$. Then the maps $L_x\to L_y:w\mt wb$ and $L_y\to L_x:w\mt wa$ are mutually inverse bijections. These maps restrict to mutually inverse bijections $H_x\to H_y$ and $H_y\to H_x$.
\itemit{ii} Suppose $x\L y$, and that $x=ay$ and $y=bx$ where $a,b\in\Sone$. Then the maps $R_x\to R_y:w\mt bw$~and $R_y\to R_x:w\mt aw$ are mutually inverse bijections. These maps restrict to mutually inverse bijections $H_x\to H_y$ and $H_y\to H_x$.
\itemit{iii} If $x\D y$, then $|R_x|=|R_y|$, $|L_x|=|L_y|$ and $|H_x|=|H_y|$.
\eitres
\end{lemma}
\pf Suppose $x\R y$, and that $x=ya$ and $y=xb$ where $a,b\in\Sone$. We first show that the map $f:L_x\to S:w\mt wb$ does indeed map $L_x$ into $L_y$. With this in mind, let $w\in L_x$. We already know that $wb\in \yL$, by Lemma \ref{lem:GreensLemma_S}(i). Also, $w=ux$ for some $u\in\Sone$, since $w\L x$. Now, $\lam(wb)=\lam(w)=i$, and also $\rho(wb)=\rho(uxb)=\rho(uy)=\rho(y)=j$, showing that $wb\in \yL\cap S_{ij}=L_y$, as required. By symmetry, it follows that $g:L_y\to S:w\mt wa$ maps $L_y$ into $L_x$. By Lemma \ref{lem:GreensLemma_S}(i), we see that $f\circ g$ and $g\circ f$ are the identity maps on their respective domains. This completes the proof of (i).
Next, note that (ii) follows from (i) by duality. Now suppose $x\D y$. So $x\R z\L y$ for some $z\in S$. Since $x\R z$, it follows that $\lam(z)=\lam(x)=i$; similarly, $\rho(z)=j$, so in fact, $z\in S_{ij}$. In particular, $R_x=R_z$ and $L_y=L_z$. The statement about cardinalities then follows from parts (i) and (ii). \epf
As is the case for semigroups \cite{Hig,Howie}, Lemma \ref{lem:Rol=LoR_S} means that the elements of a $\D$-class $D$ of $S$ or $S_{ij}$ may be grouped together in a rectangular array of cells, which (for continuity with semigroup theory) we call an \emph{eggbox}. We place all elements from $D$ in a box in such a way that $\R$-related (resp., $\L$-related) elements are in the same row (resp., column), and $\H$-related elements in the same cell. An example is given in Figure~\ref{fig:eggbox} below for a $\D$-class of the linear partial semigroup $\M(\mathbb Z_3)$.
We now come to the definition of the main objects of our study, the \emph{sandwich semigroups}.
\ms
\begin{defn}
Let $(S,\cdot,I,\lam,\rho)$ be a partial semigroup. Fix some $a\in S_{ji}$, where $i,j\in I$. Define a binary operation $\star_a$ on $S_{ij}$ by $x\star_a y=xay$ for each $x,y\in S_{ij}$. It is easily checked that $\star_a$ is associative. We denote by $\Sija=(S_{ij},\star_a)$ the semigroup obtained in this way, and call $\Sija$ the \emph{sandwich semigroup} of $S_{ij}$ with respect to $a$. (Note that when $i=j$, $\Sija=S_i^a$ is the well-known \emph{variant} \cite{Hickey1983,Hickey1986,KL2001} of $S_i$ with respect to $a\in S_i$.)
\end{defn}
Recall that an element $x$ of a semigroup $T$ is \emph{regular} if $x=xyx$ and $y=yxy$ for some $y\in T$ (or, equivalently, if $x=xzx$ for some $z\in T$). The set of all regular elements of $T$ is denoted by $\Reg(T)$, and we say $T$ is \emph{regular} if $T=\Reg(T)$. (In general, $\Reg(T)$ need not even be a subsemigroup of $T$.) Of crucial importance is that if any element of a $\D$-class $D$ of a semigroup $T$ is regular, then \emph{every} element of $D$ is regular, in which case every element of $D$ is $\L$-related to at least one idempotent (and also $\R$-related to a possibly different idempotent); the $\H$-class $H_e$ of an idempotent $e\in E(T)=\set{x\in T}{x=x^2}$ is a group, and $H_e\cong H_f$ for any two $\D$-related idempotents~$e,f\in E(T)$. When drawing eggbox diagrams, group $\gH$-classes are usually shaded grey (see for example Figure \ref{fig:M0...M3}). See \cite{Hig,Howie} for more details.
If $S$ is a regular partial semigroup, then the sandwich semigroups $\Sija$ need not be regular themselves (although all of the semigroups $S_i$ are), but the set $\RegSija$ of all regular elements of $\Sija$ forms a subsemigroup, as we now show.
\begin{prop}\label{prop:regularSija}
Let $(S,\cdot,I,\lam,\rho)$ be a regular partial semigroup. Then $\RegSija$ is a subsemigroup of $\Sija$ for all $i,j\in I$ and $a\in S_{ji}$.
\end{prop}
\pf Let $x,y\in\RegSija$, so $x=xauax$ and $y=yavay$ for some $u,v\in S_{ij}$. Since $S$ is regular, there exists $w\in S$ such that $(auaxayava)w(auaxayava)=(auaxayava)$. Then
\begin{align*}
(xay)a(vawau)a(xay) &= (xauaxay)a(vawau)a(xayavay) \\
&= x(auaxayava)w(auaxayava)y = x(auaxayava)y = xay,
\end{align*}
showing that $(x\star_ay)\star(v\star_aw\star_au)\star_a(x\star_ay)=x\star_ay$, and $x\star_ay\in\RegSija$. \epf
In order to say more about the regular elements and Green's relations of the sandwich semigroup $\Sija$, we define the sets
\[
P_1^a = \set{x\in S_{ij}}{xa\R x} \COMMA
P_2^a = \set{x\in S_{ij}}{ax\L x} \COMMA
P_3^a = \set{x\in S_{ij}}{axa\J x} \COMMA
P^a=P_1^a\cap P_2^a.
\]
The next result explains the relationships that hold between these sets; the various inclusions are pictured in Figure \ref{fig:P}.
\ms
\begin{figure}[ht]
\begin{center}
\scalebox{.8}{
\begin{tikzpicture}
\tikzset{venn circle/.style={draw,circle,minimum width=6cm,fill=#1,opacity=0.4}}
\tikzset{small venn circle/.style={draw,circle,minimum width=1cm,fill=#1,opacity=0.4}}
\node [venn circle = black,opacity=0.2] (C) at (4,0) {$$};
\node [venn circle = black,opacity=0.2] (A) at (0,0) {$$};
\node [venn circle = black,opacity=0.2] (B) at (2,-0.76393) {$$};
\node [small venn circle = black,opacity=0.2] (B) at (2,0) {$$};
\node at (2,0) {$R$};
\node at (-0.76393*2,0.76393*2) {$P_1^a$};
\node at (0.76393*2+4,0.76393*2) {$P_2^a$};
\node at (2,-0.76393*4) {$P_3^a$};
\node at (2,-1.2) {$P^a$};
\end{tikzpicture}
\qquad\qquad
\begin{tikzpicture}
\tikzset{venn circle/.style={draw,circle,minimum width=6cm,fill=#1,opacity=0.4}}
\tikzset{small venn circle/.style={draw,circle,minimum width=1cm,fill=#1,opacity=0.4}}
\node [venn circle = black,opacity=0.2] (C) at (4,0) {$$};
\node [venn circle = black,opacity=0.2] (A) at (0,0) {$$};
\node [venn circle = black,opacity=0.0] (B) at (2,-0.76393) {$$};
\node [small venn circle = black,opacity=0.2] (B) at (2,0) {$$};
\node at (2,0) {$R$};
\node at (-0.76393*2,0.76393*2) {$P_1^a$}; \node at (0.76393*2+4,0.76393*2) {$P_2^a$};
\node at (2,-1.2) {$P^a=P_3^a$};
\end{tikzpicture}
}
\caption{Venn diagrams illustrating the various relationships between the sets $P_1^a$, $P_2^a$, $P_3^a$, $P^a=P_1^a\cap P_2^a$ and $\Reg(\Sija)$ in the general case (left) and the stable case (right); for clarity, we have written $R=\Reg(\Sija)$.}
\label{fig:P}
\end{center}
\end{figure}
\ms
\begin{prop}\label{prop:Reg(Sija)}
Let $(S,\cdot,I,\lam,\rho)$ be a partial semigroup, and fix $i,j\in I$ and $a\in S_{ji}$. Then
\begin{itemize}\begin{multicols}2
\itemit{i} $\RegSija \sub P^a \sub P_3^a$,
\itemit{ii} $P^a=P_3^a$ if $S$ is stable.
\emc
\end{prop}
\pf If $x\in\RegSija$, then $x=xayax$ for some $y\in S_{ij}$, giving $x\R xa$ and $x\L ax$, so that $x\in P_1^a\cap P_2^a=P^a$. Next, suppose $x\in P^a=P_1^a\cap P_2^a$, so $x=xav=uax$ for some $u,v\in\Sone$. It follows that $x=uaxav$, so $x\J axa$ and $x\in P_3^a$. This completes the proof of (i).
Now suppose $S$ is stable, and let $x\in P_3^a$. So $x=uaxav$ for some $u,v\in\Sone$. It then follows that $x\J xa$ and $x\J ax$. By stability, it follows that $x\R xa$ and $x\L ax$, so that $x\in P_1^a\cap P_2^a=P^a$, completing the proof of (ii). \epf
\begin{rem}
The assumption of regularity (resp., stability) could be greatly weakened in Proposition~\ref{prop:regularSija} (resp., Proposition \ref{prop:Reg(Sija)}(ii)).
However, because the linear partial semigroup $\M$ is regular and stable
(see Lemmas \ref{lem:green<M} and \ref{lem:regularity}), we will not pursue this thought any further.
\end{rem}
We now show how the sets $P_1^a$, $P_2^a$, $P_3^a$ and $P^a=P_1^a\cap P_2^a$ may be used to relate Green's relations on the sandwich semigroups $\Sija$ to the corresponding relations on $S$.
To avoid confusion, if $\gK$ is one of $\R$, $\L$, $\J$, $\D$, $\H$, we write $\gKa$ for the Green's $\gK$-relation on $\Sija$. So, for example, if $x,y\in S_{ij}$, then
\bit
\item $x\gRa y$ if and only if [$x=y$] or [$x=y\star_a u=yau$ and $y=x\star_a v=xav$ for some $u,v\in S_{ij}$].
\eit
It is then clear that $\gRa\sub\R$, and the analogous statement is true for all of the other Green's relations.
If $x\in S_{ij}$, we write $K_x^a=\set{y\in S_{ij}}{x\gKa y}$ for the $\gKa$-class of $x$ in $\Sija$. Since $\gKa\sub\gK$, it follows that $K_x^a\sub K_x$ for all $x\in S_{ij}$.
\ms
\begin{thm}\label{thm:green_Sij}
Let $(S,\cdot,I,\lam,\rho)$ be a partial semigroup, and let $a\in S_{ji}$ where $i,j\in I$. If $x\in S_{ij}$, then
\ms
\begin{itemize}\begin{multicols}{2}
\itemit{i} $R_x^a = \begin{cases}
R_x\cap P_1^a &\text{if $x\in P_1^a$}\\
\{x\} &\text{if $x\in S_{ij}\sm P_1^a$,}
\end{cases}$
\itemit{ii} $L_x^a = \begin{cases}
L_x\cap P_2^a &\hspace{0.7mm}\text{if $x\in P_2^a$}\\
\{x\} &\hspace{0.7mm}\text{if $x\in S_{ij}\sm P_2^a$,}
\end{cases}
$
\itemit{iii} $H_x^a = \begin{cases}
H_x &\hspace{7.4mm}\text{if $x\in P^a$}\\
\{x\} &\hspace{7.4mm}\text{if $x\in S_{ij}\sm P^a$,}
\end{cases}$
\itemit{iv} $D_x^a = \begin{cases}
D_x\cap P^a &\text{if $x\in P^a$}\\
L_x^a &\text{if $x\in P_2^a\sm P_1^a$}\\
R_x^a &\text{if $x\in P_1^a\sm P_2^a$}\\
\{x\} &\text{if $x\in S_{ij}\sm (P_1^a\cup P_2^a)$,}
\end{cases}$
\itemit{v} $J_x^a = \begin{cases}
J_x\cap P_3^a &\hspace{2.2mm}\text{if $x\in P_3^a$}\\
D_x^a &\hspace{2.2mm}\text{if $x\in S_{ij}\sm P_3^a$.}
\end{cases}$
\end{multicols}\end{itemize}
\begin{comment}
hhh
\ms
\bmc2
\itemit{i} $R_x^a = \begin{cases}
R_x\cap P_1^a &\text{if $x\in P_1^a$}\\
\{x\} &\text{if $x\in S_{ij}\sm P_1^a$,}
\end{cases}$
\itemit{ii} $L_x^a = \begin{cases}
L_x\cap P_2^a &\hspace{0.7mm}\text{if $x\in P_2^a$}\\
\{x\} &\hspace{0.7mm}\text{if $x\in S_{ij}\sm P_2^a$,}
\end{cases}
\phantom{
\begin{cases}
a\\b\\c\\d
\end{cases}
}
$
\itemit{iii} $H_x^a = \begin{cases}
H_x &\hspace{8.2mm}\text{if $x\in P^a$}\\
\{x\} &\hspace{8.2mm}\text{if $x\in S_{ij}\sm P^a$,}
\end{cases}$
\itemit{iv} $D_x^a = \begin{cases}
D_x\cap P^a &\text{if $x\in P^a$}\\
L_x^a &\text{if $x\in P_2^a\sm P_1^a$}\\
R_x^a &\text{if $x\in P_1^a\sm P_2^a$}\\
\{x\} &\text{if $x\in S_{ij}\sm (P_1^a\cup P_2^a)$.}
\end{cases}$
\end{multicols}\end{itemize}
\end{comment}
Further, if $x\in S_{ij}\sm P^a$, then $H_x^a=\{x\}$ is a non-group $\gHa$-class of $\Sija$.
\end{thm}
\pf The proof of \cite[Proposition 3.2]{DE2} may easily be adapted to prove (i--iv) and the final statement about $\gHa$-classes. We now prove (v). Let $x\in S_{ij}$.
Suppose $y\in J_x^a\sm\{x\}$. So one of (a--c) and one of (d--f) holds:
\bitbmc2
\item[(a)] $y=sax$ for some $s\in S_{ij}$,
\item[(b)] $y=xat$ for some $t\in S_{ij}$,
\item[(c)] $y=saxat$ for some $s,t\in S_{ij}$,
\item[(d)] $x=uay$ for some $u\in S_{ij}$,
\item[(e)] $x=yav$ for some $v\in S_{ij}$,
\item[(f)] $x=uayav$ for some $u,v\in S_{ij}$.
\emc
Suppose first that (a) and (d) hold. Then $x\gLa y$. Since $x\not=y$, we deduce that $x\in P_2^a$ by (ii). Since $L_x^a=L_y^a$, we also have $y\in P_2^a$. Similarly, if (b) and (e) hold, then $x\gRa y$ and $x,y\in P_1^a$. One may check that any other combination of (a--c) and (d--f) implies $x,y\in P_3^a$. For example, if (a) and (e) hold, then
\[
y=sax=s(aya)v \AND x=yav=s(axa)v.
\]
In particular, we have shown that $|J_x^a|\geq2$ implies $x\in P_1^a\cup P_2^a\cup P_3^a$. By the contrapositive of this last statement, if $z\in S_{ij}\sm(P_1^a\cup P_2^a\cup P_3^a)$, then $J_z^a=\{z\}=D_z^a$, with the last equality following from (iv).
Next, suppose $x\in P_1^a\sm P_3^a$. In particular, $x\not\in P_2^a$ since $P_1^a\cap P_2^a\sub P_3^a$ by Proposition \ref{prop:Reg(Sija)}(i). Since $\gDa\sub\gJa$, we have $D_x^a\sub J_x^a$. Conversely, suppose $y\in J_x^a$. We must show that $y\in D_x^a$. If $y=x$, then we are done, so suppose $y\not=x$. As above, one of (a--c) and one of (d--f) holds. If (b) and (e) hold, then $y\in R_x^a=D_x^a$, the second equality holding by (iv). If any other combination of (a--c) and (d--f) holds then, as explained in the previous paragraph, $x$ (and $y$) would belong to $P_2^a$ or $P_3^a$, a contradiction. This completes the proof that $J_x^a\sub D_x^a$. A dual argument shows that $J_x^a=D_x^a$ if $x\in P_2^a\sm P_3^a$.
Finally, suppose $x\in P_3^a$. Let $z\in J_x\cap P_3^a$. So we have
\[
x=s'axat' \COMMA z=s''azat'' \COMMA z=u'xv' \COMMA x=u''zv'' \qquad\text{for some $s',s'',t',t'',u',u'',v',v''\in \Sone$.}
\]
We then calculate
$
z = u'xv' = u's'axat'v' = u's'a(s'axat')at'v' = (u's'as') \star_a x \star_a (t'at'v'),
$
and similarly $x = (u''s''as'') \star_a z \star_a (t''at''v'')$, showing that $z\gJa x$, and $J_x\cap P_3^a\sub J_x^a$. To prove the reverse inclusion, since we have already observed that $J_x^a\sub J_x$, it suffices to show that $J_x^a\sub P_3^a$. So suppose $y\in J_x^a$. If $y=x$, then $y\in P_3^a$, so suppose $y\not=x$. Then one of (a--c) and one of (d--f) above holds. If (a) and (d) hold, then
\[
y = sax = sa s'axat' = sa s'auayat',
\]
showing that $y\in P_3^a$. A similar argument covers the case in which (b) and (e) hold. As we observed above, any other combination of (a--c) and (d--f) implies that $y\in P_3^a$. This completes the proof. \epf
For a pictorial understanding of Theorem \ref{thm:green_Sij},
Figures \ref{fig:V3212_V3322} and \ref{fig:V2322_V2422} below give eggbox diagrams of various linear sandwich semigroups.%
Next, we show that stability of $S$ entails stability of all sandwich semigroups~$\Sija$.
\ms
\begin{prop}\label{prop:stabilityMija}
Let $(S,\cdot,I,\lam,\rho)$ be a stable partial semigroup. Then $\Sija$ is stable for all $i,j\in I$ and $a\in S_{ji}$.
\end{prop}
\pf Let $x,y\in S_{ij}$. We must show that
\[
x\gJa x\star_ay \ \iff\ x\gRa x\star_ay \AND x\gJa y\star_ax\ \iff\ x\gLa y\star_ax.
\]
By duality, it suffices to prove the first of these. Clearly, $x\gRa x\star_ay \ \implies\ x\gJa x\star_ay$. Conversely, suppose $x\gJa x\star_ay$. Then one of the following holds:
\vspace{-0.2cm}
\begin{multicols}2
\begin{enumerate}
\item[(i)] $x=xay$,
\item[(ii)] $x=xayav$ for some $v\in S_{ij}$,
\item[(iii)] $x=uaxay$ for some $u\in S_{ij}$,
\item[(iv)] $x=uaxayav$ for some $u,v\in S_{ij}$.
\end{enumerate}
\end{multicols}
Clearly, (i) or (ii) implies $x\gRa xay$. Next, suppose (iv) holds. Then $x\J xaya$, so that $x\R xaya$ by stability. In particular, (a) $x=xaya$ or (b) $x=xayaw$ for some $w\in S_{ij}$. If (a) holds, then $x=(xaya)aya$, so (b) holds with $w=aya$. In particular, $x=(x\star_ay)\star_aw$, completing the proof that $x\gRa x\star_ay$. Finally, if (iii) holds, then $x=ua(uaxay)ay$, so that case (iii) reduces to case (iv). The proof is therefore complete.~\epf
We conclude this section with a result that shows how regularity of the sandwich element implies close relationships between certain sandwich semigroups $\Sija$ and $S_{ji}^b$ and certain (non-sandwich) subsemigroups of $S_i$ and $S_j$.
\ms
\begin{thm}\label{thm:diamonds}
Let $(S,\cdot,I,\lam,\rho)$ be a partial semigroup and let $i,j\in I$. Let $a\in S_{ji}$ and $b\in S_{ij}$ be such that $a=aba$ and $b=bab$. Then
\bit
\itemit{i} $S_{ij}a$ and $aS_{ij}$ are subsemigroups of $S_i$ and $S_j$ (respectively),
\itemit{ii} $(aS_{ij}a,\star_b)$ and $(bS_{ji}b,\star_a)$ are monoids with identities $b$ and $a$ (respectively), and are subsemigroups of $S_{ji}^b$ and $\Sija$ (respectively),
\itemit{iii} the maps $aS_{ij}a\to bS_{ji}b:x\mt bxb$ and $bS_{ji}b\to aS_{ij}a:x\mt axa$ define mutually inverse isomorphisms between $(aS_{ij}a,\star_b)$ and $(bS_{ji}b,\star_a)$,
\itemit{iv} $a\RegSija a$ is contained in $\Reg(S_{ji}^b)$,
\itemit{v} the following diagrams commute, with all maps being homomorphisms:
\[
\includegraphics{CD1.pdf}
\]
\eit
\end{thm}
\pf Part (i) is clear, and parts (ii) and (iii) are easily checked. Next, suppose $x\in\RegSija$, so $x=xauax$ for some $u\in S_{ij}$. Then
$
axa = axauaxa = axabauabaxa = (axa)\star_b(aua)\star_b(axa),
$
giving (iv). Part (v) is all mostly easy to check. That $\Phi_1$ is a homomorphism follows from
$\Phi_1((xa)(ya))=axaya = axabaya = \Phi_1(xa)\star_b\Phi_1(ya)$.
It is clear that $\psi_1$ maps $\RegSija$ into $\Reg(S_{ij}a)$.
It follows from (iv) that $\phi_1$ maps $\Reg(S_{ij}a)$ into $\Reg(S_{ji}^b)$. \epf
\begin{rem}
Other relationships exist, such as $(baS_{ij}a,\cdot) = (bS_{ji}ba,\cdot)$, but these will not be explored any further.
\end{rem}
\section{The linear partial semigroup}\label{sect:preliminaries}
As noted earlier, to understand the linear sandwich semigroups $\MmnA$, it is crucial to first understand the partial semigroup $\M$. So in this section, we gather the required material on $\M$, showing how the general framework of Section \ref{sect:partial} applies in this case.
We fix a field $\F$ for the remainder of the article. For positive integers $m,n\in\mathbb N=\{1,2,3,\ldots\}$, we write $\M_{mn}=\Mmn(\F)$ for the set of all $m\times n$ matrices (i.e., all matrices with $m$ rows and $n$ columns) over $\F$. We write $\M=\M(\F)=\bigcup_{m,n\in\mathbb N}\Mmn$ for the set of all (finite dimensional, non-empty) matrices over $\F$. So $\M$ is a \emph{partial semigroup}, as noted in Example \ref{eg:M}.
By convention, we consider there to be a unique $m\times0$ and $0\times n$ matrix for any $m,n\geq0$, namely the empty matrix, which we denote by $\emptyset$. So $\M_{mn}=\{\emptyset\}$ if $m=0$ or $n=0$. But this is a matter of convenience, and we do not consider the empty matrix $\emptyset$ to be an element of $\M$.
We also write $\M_n=\M_n(\F)=\M_{nn}$ for any $n$%
, and denote by $\G_n=\G_n(\F)$ the group of $n\times n$ invertible matrices over $\F$. So $\M_n$ and $\G_n$ are the \emph{full linear monoid} and \emph{general linear group} of degree $n$. For background on the full linear monoids, the monograph \cite{Okninski1998} is highly recommended.
If $V$ and $W$ are vector spaces, we write $\Hom(V,W)$ for the set of all linear transformations from $V$ to~$W$. As usual, if $\al\in\Hom(V,W)$, we write $\im(\al)=\set{\al(v)}{v\in V}$ and $\ker(\al)=\set{v\in V}{\al(v)=0}$ for the \emph{image} and \emph{kernel} of $\al$. We write $\End(V)=\Hom(V,V)$ for the monoid of all endomorphisms of $V$ (i.e., all linear transformations $V\to V$), and $\Aut(V)$ for the group of all automorphisms of $V$ (i.e., all invertible endomorphisms of $V$). For $n\geq0$, we write $V_n=\F^n$ for the vector space of all $n\times1$ column vectors over~$\F$. We will identify $\M_{mn}$ with $\Homnm$ in the usual way. Namely, if $X\in\Mmn$, we write $\lam_X\in\Homnm$ for the linear transformation $\lam_X:V_n\to V_m$ defined by $\lam_X(v)=Xv$ for all $v\in V_n$. We will often prove statements about $\M_{mn}$ by proving the equivalent statement about $\Homnm$. When $m=n$, the map $X\to\lam_X$ determines an isomorphism of monoids $\M_n\to\Endn$, and its restriction to $\G_n\sub\M_n$ determines an isomorphism of groups $\G_n\to\Autn$. We write $\{e_{n1},\ldots,e_{nn}\}$ for the standard basis of $V_n$ ($e_{ni}$ has a $1$ in position $i$ and $0$'s elsewhere). We also write $W_{ns}=\Span\{e_{n1},\ldots,e_{ns}\}$ for each $0\leq s\leq n$. (We interpret $\Span\emptyset=\{0\}$, though the dimension of the ambient space must be understood from context.)
Our first aim is to characterise Green's relations ($\R$, $\L$, $\J$, $\D$, $\H$) and preorders ($\leq_\R$, $\leq_\L$, $\leq_\J$) on $\M$. Because $\M$ is monoidal (see Definition~\ref{defn:S}), $\M=\M^{(1)}$. So, for example, if $X,Y\in\M$ are two matrices (not necessarily of the same size), then $X\leq_\R Y$ if and only if $X=YA$ for some $A\in\M$.
Note that if $X\leq_\R Y$ (resp., $X\leq_\L Y$), then $X$ and $Y$ must have the same number of rows (resp., columns).
Let $X\in\M_{mn}$. For $1\leq i\leq m$ and $1\leq j\leq n$, we write $\row_i(X)$ and $\col_j(X)$ for the $i$th row and $j$th column of~$X$, respectively. We write $\Row(X)=\Span\{\row_1(X),\ldots,\row_m(X)\}$ and $\Col(X)=\Span\{\col_1(X),\ldots,\col_n(X)\}$ for the \emph{row space} and \emph{column space} of $X$, respectively, and we write $\rank(X)=\dim(\Row(X))=\dim(\Col(X))$ for the \emph{rank} of $X$.
Because of the transpose map $\M\to\M:A\mt A^\tr$, which is a bijection and satisfies $(AB)^\tr=B^\tr A^\tr$, the linear partial semigroup $\M$ is \emph{self-dual} (in the sense that it is anti-isomorphic to its own dual).
Since $\Row(X^\tr)=\Col(X)$, any statement about row spaces implies a corresponding dual statement about column spaces (and vice versa). (Without causing confusion, we will often blur the distinction between row vectors and column vectors, and think of $\Row(X)$ and $\Col(X)$ as subspaces of $\F^n$ and $\F^m$, respectively.)
The next result characterises Green's relations and preorders on $\M$ in terms of the parameters introduced above. An equivalent formulation in the special case of square matrices may be found in \cite[Lemma 2.1]{Okninski1998}.
\ms
\begin{lemma}\label{lem:green<M}
Let $X,Y\in\M$. Then
\begin{itemize}\begin{multicols}2
\itemit{i} $X\leq_\R Y \iff \Col(X)\sub\Col(Y)$,
\itemit{ii} $X\leq_\L Y \iff \Row(X)\sub\Row(Y)$,
\itemit{iii} $X\leq_\J Y \iff \rank(X)\sub\rank(Y)$,
\itemit{iv} $X\R Y \tiff \Col(X)=\Col(Y)$,
\itemit{v} $X\L Y \tiff \Row(X)=\Row(Y)$,
\itemit{vi} $X\J Y \tiff \rank(X)=\rank(Y)$.
\end{multicols} \end{itemize} \nss\ss
Further, $\M$ is stable, so $\J=\D$.
\end{lemma}
\pf Clearly, (iv--vi) follow from (i--iii). Note that (ii) is the dual of (i), which is true because
\begin{align*}
X\leq_\R Y &\ \iff\ \text{$X=YA$ for some $A\in\M$}\\
&\ \iff\ \text{every column of $X$ is a linear combination of the columns of $Y$}\\
&\ \iff\ \Col(X)\sub\Col(Y).
\end{align*}
For (iii),
if $X\leq_\J Y$, then $X=AYB$ for some $A,B\in\M$, giving $\rank(X)=\rank(AYB)\leq\rank(Y)$. Conversely, suppose $\rank(X)\leq\rank(Y)$, and say $X\in\Mmn$ and $Y\in\Mkl$. It is sufficient to show that $\lam_X=\al\circ\lam_Y\circ\be$ for some $\al\in\Homkm$ and $\be\in\Homnl$.
Put $r=\rank(X)$ and $s=\rank(Y)$. Choose bases $\B_1=\{u_1,\ldots,u_n\}$ and $\B_2=\{v_1,\ldots,v_l\}$ for $V_n$ and $V_l$ so that $\{u_{r+1},\ldots,u_n\}$ and $\{v_{s+1},\ldots,v_l\}$ are bases for $\ker(\lam_X)$ and $\ker(\lam_Y)$, respectively. Extend (if necessary) the linearly independent sets $\{\lam_Y(v_1),\ldots,\lam_Y(v_r)\}$ and $\{\lam_X(u_1,\ldots,\lam_X(u_r)\}$ arbitrarily to bases
\[
\B_3=\{\lam_Y(v_1),\ldots,\lam_Y(v_r),w_{r+1},\ldots,w_k\} \AND \B_4=\{\lam_X(u_1),\ldots,\lam_X(u_r),x_{r+1},\ldots,x_m\}
\]
for $V_k$ and $V_m$. Now let $\al\in\Homkm$ and $\be\in\Homnl$ be chosen arbitrarily so that
\[
\begin{array}{rclcrcll}
\al(\lam_Y(v_i)) \hspace{-.25cm}&=&\hspace{-.25cm} \lam_X(u_i), & &
\al(w_j) \hspace{-.25cm} &\in&\hspace{-.25cm} \Span\{x_{r+1},\ldots,x_m\}
&\qquad\text{for all $1\leq i\leq r$ and $r+1\leq j\leq k$,}
\\
\be(u_i) \hspace{-.25cm}&=&\hspace{-.25cm} v_i, & &
\be(u_j) \hspace{-.25cm} &\in&\hspace{-.25cm} \Span\{v_{s+1},\ldots,v_l\}
&\qquad\text{for all $1\leq i\leq r$ and $r+1\leq j\leq n$.}
\end{array}
\]
One easily checks that $\al\circ\lam_Y\circ\be=\lam_X$, by checking the respective actions on the basis $\B_1$ of $V_n$.
To prove stability, we must show that for all $X,Y\in\M$,
\[
X\J XY \tiff X\R XY \AND X\J YX \tiff X\L YX.
\]
By duality, it suffices to prove the first equivalence. Since $\R\sub\J$, it is enough to prove that $X\J XY \implies X\R XY$. Now, $\Col(XY)\sub\Col(X)$. But also $X\J XY$ gives $\dim(\Col(X))=\rank(X)=\rank(XY)=\dim(\Col(XY))$, so that $\Col(X)=\Col(XY)$, whence $X\R XY$. \epf
As we saw in Section \ref{sect:partial}, stability and regularity are very useful properties for a partial semigroup to have. Now that we know $\M$ is stable, let us show that $\M$ is also regular.
\ms
\begin{lemma}\label{lem:regularity}
The linear partial semigroup $\M$ is regular.
\end{lemma}
\pf Let $X\in\M_{mn}$. It suffices to show that there exists $\al\in\Hommn$ such that ${\lam_X=\lam_X\circ\al\circ\lam_X}$. Let $\B=\{v_1,\ldots,v_n\}$ be a basis of $V_n$ such that $\{v_{r+1},\ldots,v_n\}$ is a basis of $\ker(\lam_X)$. Extend (if necessary) the linearly independent set $\{\lam_X(v_1),\ldots,\lam_X(v_r)\}$ to a basis $\{\lam_X(v_1),\ldots,\lam_X(v_r),w_{r+1},\ldots,w_m\}$ of $V_m$. Let $\al\in\Hommn$ be any linear transformation for which $\al(\lam_X(v_i))=v_i$ for each $1\leq i\leq r$. Then one easily checks that $\lam_X=\lam_X\circ\al\circ\lam_X$ by calculating the action on the basis $\B$. \epf
As in Section \ref{sect:partial}, if $X\in\Mmn$ and $\gK$ is one of $\R$, $\L$, $\J$, $\D$, $\H$, we write $K_X=\set{Y\in\Mmn}{X\gK Y}$, and call $K_X$ the \emph{$\gK$-class} of $X$ in $\Mmn$. Note that all matrices from $K_X$ have the same dimensions. (We will have no need to consider the sets $\XK$ of \emph{all} matrices $\gK$-related to $X$.) Recall that $\G_k$ denotes the group of all invertible $k\times k$ matrices over $\F$. The next result gives an alternative description of various Green's classes in $\M$.
\ms
\begin{lemma}\label{lem:greenMmn}
Let $X\in\M_{mn}$. Then
\bit
\itemit{i} $R_X = \set{Y\in\M_{mn}}{\Col(X)=\Col(Y)} = X\G_n$,
\itemit{ii} $L_X = \set{Y\in\M_{mn}}{\Row(X)=\Row(Y)} = \G_mX$,
\itemit{iii} $D_X = J_X = \set{Y\in\M_{mn}}{\rank(X)=\rank(Y)} = \G_mX\G_n$.
\eitres
\end{lemma}
\pf For (i),
note that clearly $X\G_n\sub R_X$. By Lemma \ref{lem:green<M}, it remains to show the reverse inclusion, so suppose $Y\in R_X$. In particular, $X\J Y$, so $\rank(X)=\rank(Y)$. Put $r=\rank(X)$. We show that $\lam_Y=\lam_X\circ\al$ for some $\al\in\Autn$. Since $X\R Y$, we already know that $\lam_Y=\lam_X\circ\be$ for some $\be\in\Endn$. Let $\B_1=\{u_1,\ldots,u_n\}$ be a basis of $V_n$ such that $\{u_{r+1},\ldots,u_n\}$ is a basis of $\ker(\lam_Y)$. So $\{\lam_X(\be(u_1)),\ldots,\lam_X(\be(u_r))\}=\{\lam_Y(u_1),\ldots,\lam_Y(u_r)\}$ is a basis of $\im(\lam_Y)$. It follows that $\{\be(u_1),\ldots,\be(u_r)\}$ is linearly independent. We may therefore extend this set to a basis $\B_2=\{\be(u_1),\ldots,\be(u_r),v_{r+1},\ldots,v_n\}$ of $V_n$, where $\{v_{r+1},\ldots,v_n\}$ is a basis of $\ker(\lam_X)$. Now define $\al\in\Autn$ by
\[
\al(u_i) = \begin{cases}
\be(u_i) &\text{if $1\leq i\leq r$}\\
v_i &\text{if $r<i\leq n$.}
\end{cases}
\]
One easily checks that $\lam_Y=\lam_X\circ\al$. This completes the proof of (i).
Part (ii) is dual to (i).
For (iii),
clearly $\G_mX\G_n\sub J_X$, and the converse follows quickly from (i) and~(ii) and the fact that $\J=\D=\L\circ\R$. By Lemma \ref{lem:green<M}, this completes the proof. \epf
If $\gK$ is one of $\R$, $\L$, $\D=\J$, then the set $\Mmn/\gK$ of all $\gK$-classes of $\Mmn$ inherits a partial order:
\[
K_X \leq_\gK K_Y \ \iff \ X\leq_\gK Y.
\]
We typically write $\leq$ for the order $\leq_\J$ on the $\D=\J$-classes. Of importance is the fact that these classes form a chain:
\[
\DMmn0<\DMmn1<\cdots<\DMmn l,
\]
where $\DMmn s=\set{X\in\Mmn}{\rank(X)=s}$ for all $0\leq s\leq l=\min(m,n)$.
Figure \ref{fig:eggbox} pictures an eggbox diagram (as explained in Section \ref{sect:partial}) of the $\D$-class $D_1(\M_{23}(\mathbb Z_3))$ of all $2\times3$ matrices of rank $1$ over the field $\F=\mathbb Z_3=\{0,1,2\}$
(see Lemma \ref{lem:combinatorics_Mmn} for an explanation of the number and sizes of the $\R$-, $\L$- and $\H$-classes). The reader need not yet worry about the subdivisions within the eggbox; for now, it is enough to note that the matrices to the left (resp., top) of the vertical (resp., horizontal) divider satisfy the property that the first column (resp., row) spans the column space (resp., row space) of the matrix.
\begin{figure}[ht]
\begin{center}
\scalebox{.8}{
\begin{tikzpicture}[scale=0.9]
\dmat1{7.9}100000
\dmat1{7.1}200000
\dmat1{5.9}100100
\dmat1{5.1}200200
\dmat1{3.9}100200
\dmat1{3.1}200100
\dmat1{1.9}000100
\dmat1{1.1}000200
\dmat2{7.9}101000
\dmat2{7.1}202000
\dmat2{5.9}101101
\dmat2{5.1}202202
\dmat2{3.9}101202
\dmat2{3.1}202101
\dmat2{1.9}000101
\dmat2{1.1}000202
\dmat3{7.9}102000
\dmat3{7.1}201000
\dmat3{5.9}102102
\dmat3{5.1}201201
\dmat3{3.9}102201
\dmat3{3.1}201102
\dmat3{1.9}000102
\dmat3{1.1}000201
\dmat4{7.9}110000
\dmat4{7.1}220000
\dmat4{5.9}110110
\dmat4{5.1}220220
\dmat4{3.9}110220
\dmat4{3.1}220110
\dmat4{1.9}000110
\dmat4{1.1}000220
\dmat5{7.9}111000
\dmat5{7.1}222000
\dmat5{5.9}111111
\dmat5{5.1}222222
\dmat5{3.9}111222
\dmat5{3.1}222111
\dmat5{1.9}000111
\dmat5{1.1}000222
\dmat6{7.9}112000
\dmat6{7.1}221000
\dmat6{5.9}112112
\dmat6{5.1}221221
\dmat6{3.9}112221
\dmat6{3.1}221112
\dmat6{1.9}000112
\dmat6{1.1}000221
\dmat7{7.9}120000
\dmat7{7.1}210000
\dmat7{5.9}120120
\dmat7{5.1}210210
\dmat7{3.9}120210
\dmat7{3.1}210120
\dmat7{1.9}000120
\dmat7{1.1}000210
\dmat8{7.9}121000
\dmat8{7.1}212000
\dmat8{5.9}121121
\dmat8{5.1}212212
\dmat8{3.9}121212
\dmat8{3.1}212121
\dmat8{1.9}000121
\dmat8{1.1}000212
\dmat9{7.9}122000
\dmat9{7.1}211000
\dmat9{5.9}122122
\dmat9{5.1}211211
\dmat9{3.9}122211
\dmat9{3.1}211122
\dmat9{1.9}000122
\dmat9{1.1}000211
\dmat{10}{7.9}010000
\dmat{10}{7.1}020000
\dmat{10}{5.9}010010
\dmat{10}{5.1}020020
\dmat{10}{3.9}010020
\dmat{10}{3.1}020010
\dmat{10}{1.9}000010
\dmat{10}{1.1}000020
\dmat{11}{7.9}011000
\dmat{11}{7.1}022000
\dmat{11}{5.9}011011
\dmat{11}{5.1}022022
\dmat{11}{3.9}011022
\dmat{11}{3.1}022011
\dmat{11}{1.9}000011
\dmat{11}{1.1}000022
\dmat{12}{7.9}012000
\dmat{12}{7.1}021000
\dmat{12}{5.9}012012
\dmat{12}{5.1}021021
\dmat{12}{3.9}012021
\dmat{12}{3.1}021012
\dmat{12}{1.9}000012
\dmat{12}{1.1}000021
\dmat{13}{7.9}001000
\dmat{13}{7.1}002000
\dmat{13}{5.9}001001
\dmat{13}{5.1}002002
\dmat{13}{3.9}001002
\dmat{13}{3.1}002001
\dmat{13}{1.9}000001
\dmat{13}{1.1}000002
\foreach \x in {0,2,4,6,8}
\draw (0.75,\x+.5)--(20.25,\x+.5);
\foreach \x in {0,...,13}
\draw (\x*1.5+0.75,.5)--(\x*1.5+0.75,8.5);
\draw[line width=.5mm] (0.75,0.5)--(20.25,0.5)--(20.25,8.5)--(0.75,8.5)--(0.75,0.5)--(20.25,0.5);
\draw[line width=.5mm] (0.75,2.5)--(20.25,2.5);
\draw[line width=.5mm] (14.25,8.5)--(14.25,0.5);
\draw[|-|] (0.75,9.0)--(14.25,9.0);
\draw[|-|] (20.25,9.0)--(14.25,9.0);
\draw[|-|] (.25,.5)--(.25,2.5);
\draw[|-|] (.25,8.5)--(.25,2.5);
\draw(7.5,9.3)node{$\sub P_1\phantom{\sub}$};
\draw(17.25,9.3)node{$\not\sub P_1\phantom{\sub}$};
\draw(.25,5.5)node[left]{$\sub P_2\phantom{}$};
\draw(.25,1.5)node[left]{$\not\sub P_2\phantom{}$};
\end{tikzpicture}
}
\caption{An eggbox diagram of the $\D$-class $D_1(\M_{23}(\mathbb Z_3))$.
\label{fig:eggbox}
\end{center}
\end{figure}
So $\Mmn$ has $\min(m,n)+1$ $\D$-classes.
It will also be convenient to have some more combinatorial information about the number and size of certain $\gK$-classes. Recall that the $q$-factorials and $q$-binomial coefficients are defined by
\[
\qfact s = 1\cdot(1+q)\cdots(1+q+\cdots+q^{s-1}) = \frac{(q-1)(q^2-1)\cdots(q^s-1)}{(q-1)^s}
\]
and
\[
\qbin ms = \frac{\qfact m}{\qfact s \qfact{m-s}} = \frac{(q^m-1)(q^m-q)\cdots(q^m-q^{s-1})}{(q^s-1)(q^s-q)\cdots(q^s-q^{s-1})} = \frac{(q^m-1)(q^{m-1}-1)\cdots(q^{m-s+1}-1)}{(q^s-1)(q^{s-1}-1)\cdots(q-1)}.
\]
It is easy to check (and well-known) that when $|\F|=q<\infty$,
\[
|\G_s| = (q^s-1)(q^s-q)\cdots(q^s-q^{s-1}) = q^{{s\choose2}}(q-1)^s \qfact s.
\]
In what follows, a crucial role will be played by the matrices $J_{mns}\in\Mmn$ defined for $s\leq\min(m,n)$ by
\[
J_{mns} = \mat{I_s}{O_{s,n-s}}{O_{m-s,s}}{O_{m-s,n-s}} .
\]
Here and elsewhere, we write $I_s$ and $O_{kl}$ for the $s\times s$ identity matrix and $k\times l$ zero matrix (respectively). If the dimensions are understood from context, we just write $O=O_{kl}$.
So $J_{mns}$ is the $m\times n$ matrix with $1$'s in the first $s$ positions on the leading diagonal and $0$'s elsewhere. Note that if $s=m\leq n$ (resp., $s=n\leq m$), then the matrices $O_{m-s,s}$ and $O_{m-s,n-s}$ (resp., $O_{s,n-s}$ and $O_{m-s,n-s}$) are empty, and $J_{mns}=[I_s\ O_{s,n-s}]$ (resp., $\thmat{I_s}{O_{m-s,s}}$).
\ms
\begin{lemma}\label{lem:combinatorics_Mmn}
Suppose $|\F|=q<\infty$, and let $0\leq s\leq \min(m,n)$. Then
\bit
\itemit{i} $\DMmn s$ contains $\tqbin ms$ $\R$-classes,
\itemit{ii} $\DMmn s$ contains $\tqbin ns$ $\L$-classes,
\itemit{iii} $\DMmn s$ contains $\tqbin ms\tqbin ns$ $\H$-classes, each of which has size $|\G_s|=q^{{s\choose2}}(q-1)^s\qfact s$,
\itemit{iv} $|D_s(\Mmn)|=\tqbin ms\tqbin nsq^{{s\choose2}}(q-1)^s\qfact s$.
\eitres
\end{lemma}
\pf Parts (i) and (ii) follow immediately from parts (i) and (ii) of Lemma \ref{lem:greenMmn} and the well-known fact that $\tqbin ms$ is the number of $s$ dimensional subspaces of an $m$ dimensional vector space over $\F$. The number of $\H$-classes follows immediately from (i) and (ii). By Lemma \ref{lem:GreensLemma_Sij}, all the $\H$-classes in $\DMmn s$ have the same size, so it suffices to calculate the size of $H=H_{J_{mns}}$.
Let $X=\tmat ABCD\in H$, where $A\in\M_s$, $B\in\M_{s,n-s}$, and so on. Since $\Row(X)=\Row(J_{mns})$, we see that $B$ and $D$ are zero matrices. Considering column spaces, we see that~$C$ is also a zero matrix. It follows that $X=\tmat AOOO$, and also $\rank(A)=\rank(X)=\rank(J_{mns})=s$. Clearly every such matrix $X=\tmat AOOO$ with $\rank(A)=s$ belongs to $H$. The condition that $\rank(A)=s$ is equivalent to $A\in\G_s$, so it follows that~$|H|=|\G_s|$. Finally, (iv) follows from (iii).~\epf
Of course, by considering the size of $\Mmn$ when $|\F|=q<\infty$, we obtain the identity
\[
q^{mn} = \sum_{s=0}^l \qbin ms\qbin ns \qfact s(q-1)^sq^{{s\choose2}}.
\]
We conclude this section by stating some well-known results on the full linear monoids $\M_n$ and their ideals that we will require in what follows.
The set $E(\M_n)=\set{X\in\M_n}{X=X^2}$ of idempotents of~$\M_n$ is not a subsemigroup (unless $n\leq1$), but the
subsemigroup $\E_n=\la E(\M_n)\ra$ of $\M_n$ generated by these idempotents has a neat description. Namely, it was shown by Erdos \cite{Erdos1967} that any singular (i.e., non-invertible) matrix over~$\F$ is a product of idempotent matrices. This result has been reproved by a number of authors \cite{FL1992,Djokovic1968,Dawlings81/82,AM2005,Laffey1983}. The minimal number of (idempotent) matrices required to generate~$\E_n$ was determined by Dawlings \cite{Dawlings1982}. Recall that the \emph{rank} (resp., \emph{idempotent rank}) of a semigroup (resp., idempotent generated semigroup)~$S$, denoted $\rank(S)$ (resp., $\idrank(S)$), is the minimal size of a generating set (resp., idempotent generating set) for $S$. (The rank of a semigroup should not be confused with the rank of a matix.) If $U$ is a subset of a semigroup $S$, we write $E(U)=E(S)\cap U$ for the set of all idempotents from $U$.
\ms
\begin{thm}[Erdos \cite{Erdos1967}, Dawlings \cite{Dawlings81/82,Dawlings1982}]\label{thm_MnGn}
We have
\[
\E_n=\la E(\M_n)\ra = (\MnGn)\cup\{I_n\} \AND \MnGn= \la E(D_{n-1}(\M_n)) \ra .
\]
Further, if $|\F|=q<\infty$, then
\[
\epfreseq
\rank(\MnGn)=\idrank(\MnGn)=(q^n-1)/(q-1).
\]
\end{thm}
The previous result has been extended by Gray \cite{Gray2008} to arbitrary ideals of $\M_n$.
\ms
\begin{thm}[Gray \cite{Gray2008}]\label{thm_ideals_Mn}
The ideals of $\M_n$ are precisely the sets
\[
I_s(\M_n)=D_0(\M_n)\cup\cdots\cup D_s(\M_n)=\set{X\in\M_n}{\rank(X)\leq s} \qquad\text{for $0\leq s\leq n$,}
\]
and they form a chain: $I_0(\M_n)\sub\cdots\sub I_n(\M_n)$. If $0\leq s<n$, then $I_s(\M_n) = \la E(D_s(\M_n)) \ra$ is generated by the idempotents in its top $\D$-class. Further, if $|\F|=q<\infty$, then
\[\epfreseq
\rank(I_s(\M_n))=\idrank(I_s(\M_n))=\qbin ns.
\]
\end{thm}
Note that $I_n(\M_n)=\M_n$, $D_n(\M_n)=\G_n$ and $I_{n-1}(\M_n)=\MnGn$, so Theorem \ref{thm_MnGn} is a special case of Theorem \ref{thm_ideals_Mn} since $\tqbin n{n-1}=(q^n-1)/(q-1)$.
On several occasions, we will need to make use of the fact that the general linear group $\G_n$ may be generated by two matrices, as was originally proved by Waterhouse \cite{Waterhouse}; see also \cite{Gill2015}, where minimal generating sets for $\G_n$ are explored in more detail.
Probabilistic generation of matrix groups is considered in \cite{GK2000,BGK2008}, for example, though the context is usually for classical groups.
\ms
\begin{thm}[Waterhouse \cite{Waterhouse}]\label{thm:waterhouse}
If $|\F|<\infty$, then
\bit
\itemit{i} $\rank(\G_1)=1$, and $\rank(\G_n)=2$ if $n\geq2$,
\itemit{ii} $\M_n=\la\G_n\cup\{X\}\ra$ for any $X\in D_{n-1}(\M_n)$,
\itemit{iii} $\rank(\M_1)=2$, and $\rank(\M_n)=3$ if $n\geq2$. \epfres
\eitres
\end{thm}
For convenience, eggbox diagrams are given for the full linear monoids $\M_n(\mathbb Z_2)$ for $0\leq n\leq3$ in Figure \ref{fig:M0...M3} below. In the diagrams, group $\gH$-classes are shaded grey, and a label of {\tt k} indicates that the group $\gH$-class is isomorphic to $\G_k(\mathbb Z_2)$.
\begin{figure}[ht]
\begin{center}
\includegraphics[width=7.3mm]{M02.pdf}
\qquad
\includegraphics[width=7.5mm]{M12.pdf}
\qquad
\includegraphics[width=13.5mm]{M22.pdf}
\qquad
\includegraphics[width=21mm]{M32.pdf}
\caption{Egg box diagrams of the full linear semigroups $\M_0$, $\M_1$, $\M_2$, $\M_3$, all over $\mathbb Z_2$ (left to right).}
\label{fig:M0...M3}
\end{center}
\end{figure}
\section{Linear sandwich semigroups}\label{sect:MmnJ}
Now that we have gathered the required material on $\M$, we may begin our study of
the linear sandwich semigroups. From now on, we fix integers $m,n\geq1$ and an $n\times m$ matrix $A\in\M_{nm}$. As in Section \ref{sect:partial}, we denote by $$\MmnA=\MmnA(\F)=(\Mmn,\star_A)$$ the sandwich semigroup of $\Mmn$ under the operation $\star_A$ defined by
\[
X\star_A Y = XAY \qquad\text{for $X,Y\in\M_{mn}$.}
\]
We note that if $m=n$, then $\MmnA=\M_n^A$ is a \emph{variant} \cite{Hickey1983} of the full linear monoid~$\M_n$, so everything we prove about linear sandwich semigroups holds for such linear variants also.
We begin with a simple observation.
\ms\ms
\begin{lemma}\label{lem:MmnAMmnB}
\bit
\itemit{i} If $A\in\M_{nm}$, then $\M_{mn}^A\cong\M_{nm}^{A^\tr}$.
\itemit{ii} If $A,B\in\M_{nm}$ are such that $\rank(A)=\rank(B)$, then $\MmnA\cong\MmnB$.
\eitres
\end{lemma}
\pf It clear that $X\mt X^\tr$ defines an isomorphism $\M_{mn}^A\to\M_{nm}^{A^T}$, giving (i). Next, if $\rank(A)=\rank(B)$, Lemma \ref{lem:greenMmn} gives $A=UBV$ for some $U\in\G_m$ and $V\in\G_n$. But then one may check that $X\mt VXU$ defines an isomorphism $\MmnA\to\MmnB$, giving (ii). \epf
In particular, when studying the semigroup $\MmnA$ where $\rank(A)=r$,
we may choose any $A\in\Mnm$ of rank~$r$. For the rest of the article, we will therefore study the semigroup $\MmnJ$, where
\[
J=J_{nmr} = \mat{I_r}{O_{r,m-r}}{O_{n-r,r}}{O_{n-r,m-r}}\in\M_{nm}.
\]
From now on, unless otherwise specified, whenever a $k\times l$ matrix $X$ (with $k,l\in\{m,n\}$) is written in $2\times2$ block form, $X=\tmat ABCD$, we will be tacitly assuming that $A\in\M_r$ (from which the dimensions of $B,C,D$ may be deduced). So for example, we will usually just write $J=\tmat IOOO$.
For simplicity, we will write $\star$ for the operation $\star_J$ on $\MmnJ$, throughout. One easily verifies the rule
\[
\mat ABCD \star \mat EFGH = \mat{AE}{AF}{CE}{CF}.
\]
Also note that if $X=\tmat ABCD$, then
\[
XJ=\mat AOCO \in\M_m \COMMA JX=\mat ABOO\in\M_n \COMMA JXJ = \mat AOOO\in\M_{nm}.
\]
\ms
\begin{rem}\label{rem:r=m}
In the special case that $r=m\leq n$, we have $J=[I\ O]$, and the product in $\MmnJ$ satisfies $[A\ B] \star[E\ F] = [AE\ AF]$. But we just view this as a special case of the above rule, with the bottom rows --- i.e., $[C\ D]$, $[G\ H]$, $[CE\ CF]$ --- containing empty blocks. A dual statement holds in the case $r=n\leq m$.
In only one place will we need to consider the case in which $r=\min(m,n)$ separately (see Theorems~\ref{thm:rankMmnJ} and~\ref{thm:rankMmnJ_r=m}). If $r=m=n$, then $\MmnJ$ is precisely the full linear monoid $\M_n$; since all the problems we investigate have already been solved for $\M_n$, we will typically assume that $r=m=n$ does not hold, though our results are true for the case $r<m=n$ (corresponding to \emph{variants} of the full linear monoids~$\M_n$).
See Remark~\ref{rem:r=m2}, where the above observations are used to show that the sandwich semigroups $\MmnJ$ are isomorphic to certain well-known (non-sandwich) matrix semigroups in the case that $r=\min(m,n)$.
\end{rem}
Green's relations and the regular elements of the sandwich semigroup $\MmnJ$ were calculated in \cite{Chinram2009,Kemprasit2002}. We now show how these results may be recovered (and given a cleaner presentation) using the general theory developed in Section \ref{sect:partial}.
In particular, a crucial role is played by the sets
\[
P_1^J = \set{X\in\Mmn}{XJ\R X} ,\
P_2^J = \set{X\in\Mmn}{JX\L X} ,\
P_3^J = \set{X\in\Mmn}{JXJ\J X} ,\
P^J=P_1^J\cap P_2^J.
\]
For simplicity, we denote these sets simply by $P_1$, $P_2$, $P_3$, and $P=P_1\cap P_2$.
Certain special matrices from~$\Mmn$ will be very important in what follows. With this in mind, if $A\in\M_r$, $M\in\M_{m-r,r}$ and $N\in\M_{r,n-r}$, we write
\[
[M,A,N] = \mat{A}{AN}{MA}{MAN} \in\Mmn.
\]
One may check that when matrices of this form are multiplied in $\MmnJ$, they obey the rule
\[
[M,A,N]\star[K,B,L] = [M,AB,L].
\]
\ms
\begin{prop}\label{prop:P1P2}
~
\bit
\itemit{i} $P_1=\set{X\in\Mmn}{\rank(XJ)=\rank(X)}= \bigset{X\in\Mmn}{\Col(XJ)=\Col(X)}$,
\itemit{ii} $P_2=\set{X\in\Mmn}{\rank(JX)=\rank(X)}= \bigset{X\in\Mmn}{\Row(JX)=\Row(X)}$,
\itemit{iii} $P_3=P=\set{X\in\Mmn}{\rank(JXJ)=\rank(X)}$
\item[] $\phantom{P_3=P}=\bigset{[M,A,N]}{A\in\M_r,\ M\in\M_{m-r,r},\ N\in\M_{r,n-r}}$,
\itemit{iv} $P=\RegMmnJ$ is the set of all regular elements of $\MmnJ$, and is a subsemigroup of $\MmnJ$.
\eitres
\end{prop}
\pf Parts (i) and (ii) follow quickly from Lemma \ref{lem:green<M} (making crucial use of stability).
We now prove (iii). Since $\M$ is stable, Proposition~\ref{prop:Reg(Sija)} and Lemma \ref{lem:greenMmn} give
$
P_3 = P = \set{X\in\Mmn}{\rank(JXJ)=\rank(X)}.
$
Now let $X=\tmat ABCD\in\Mmn$. First, note that
\begin{align*}
X\in P_2 &\ \iff\ \Row(X)=\Row(JX)=\Row\tmat ABOO \\
&\ \iff\ \text{each row of $[C\ D]$ is a linear combination of the rows of $[A\ B]$} \\
&\ \iff\ \text{$[C\ D]=M[A\ B]=[MA\ MB]$ for some $M\in\M_{m-r,r}$.}
\intertext{Similarly,}
X\in P_1 &\ \iff\ \text{$\hmat BD= \hmat ACN=\hmat{AN}{CN}$ for some $N\in\M_{r,n-r}$.}
\end{align*}
Putting these together, we see that $X\in P=P_1\cap P_2$ if and only if $P=\tmat A{AN}{MA}{MAN}=[M,A,N]$, completing the proof of (iii).
For (iv),
Proposition \ref{prop:Reg(Sija)} gives $\RegMmnJ\sub P$.
Conversely, suppose $X=[M,A,N]\in P$. If $B\in\M_r$ is such that $A=ABA$ (see Lemma \ref{lem:regularity}), then it is easy to check that $X=X\star Y\star X$ where $Y=\tmat BCDE$ for any (appropriately sized) $C,D,E$, completing the proof that $P=\RegMmnJ$. The fact that $P$ is a subsemigroup follows immediately from Proposition \ref{prop:regularSija} and Lemma \ref{lem:regularity} (or directly from the rule $[M,A,N]\star[K,B,L]=[M,AB,L]$).~\epf
\begin{rem}
Part (iv) of the previous proposition also follows from \cite[Theorem 2.1]{Chinram2009}, but the rest of Proposition~\ref{prop:P1P2} appears to be new.
\end{rem}
Now that we have described the sets $P_1$, $P_2$, $P_3=P=P_1\cap P_2$, we may characterise Green's relations on~$\MmnJ$. As in Section \ref{sect:partial}, if $\gK$ is one of $\R$, $\L$, $\H$, $\D$, $\J$, we will write $\gKJ$ for the Green's $\gK$-relation on $\MmnJ$. Since $\MmnJ$ is not a monoid in general, these relations are defined, for $X,Y\in\Mmn$, by
\bit
\item $X\gRJ Y \ \iff \ [X=Y]$ or [$X=Y\star U$ and $Y=X\star V$ for some $U,V\in\Mmn$],
\eit
and so on.
Since $\M$ is stable, so too is $\MmnJ$, so we have $\gJJ=\gDJ$ (see Proposition \ref{prop:stabilityMija} and Lemmas \ref{lem:Rol=LoR_S} and \ref{lem:green<M}).
We will continue to write $\R$, $\L$, $\H$, $\D$, $\J$ for the relations on $\M$ defined in Section \ref{sect:preliminaries}. As in Section \ref{sect:partial}, if $\gK$ is one of $\R$, $\L$, $\H$, $\D=\J$, and if $X\in\Mmn$, we will write
\[
K_X=\set{Y\in\Mmn}{X\gK Y} \AND K_X^J=\set{Y\in\Mmn}{X\gKJ Y}
\]
for the $\gK$-class and $\gKJ$-class of~$X$ in $\Mmn$, respectively. As noted in Section \ref{sect:partial}, $\gKJ\sub\gK$ for each $\gK$, and so $K_X^J\sub K_X$ for each $X$.
The next result follows immediately from Theorem \ref{thm:green_Sij}. It also follows from Theorem 2.3, Lemma 2.4, and Corollaries 2.5--2.8 of \cite{Chinram2009}, but we prefer the current succinct description.
\ms
\begin{thm}\label{green_thm}
If $X\in\Mmn$, then
\ms
\bmc2
\itemit{i} $R_X^J = \begin{cases}
R_X\cap P_1 &\text{if $X\in P_1$}\\
\{X\} &\text{if $X\in \Mmn\sm P_1$,}
\end{cases}$
\itemit{ii} $L_X^J = \begin{cases}
L_X\cap P_2 &\hspace{0.7mm}\text{if $X\in P_2$}\\
\{X\} &\hspace{0.7mm}\text{if $X\in \Mmn\sm P_2$,}
\end{cases}
\phantom{
\begin{cases}
a\\b\\c\\d
\end{cases}
}
$
\itemit{iii} $H_X^J = \begin{cases}
H_X &\hspace{6.8mm}\text{if $X\in P$}\\
\{X\} &\hspace{6.8mm}\text{if $X\in \Mmn\sm P$,}
\end{cases}$
\itemit{iv} $D_X^J = \begin{cases}
D_X\cap P &\text{if $X\in P$}\\
L_X^J &\text{if $X\in P_2\sm P_1$}\\
R_X^J &\text{if $X\in P_1\sm P_2$}\\
\{X\} &\text{if $X\in \Mmn\sm (P_1\cup P_2)$.}
\end{cases}$
\emc
The sets $P_1,P_2$ are described in Proposition \ref{prop:P1P2}, and the sets $R_X,L_X,H_X,D_X$ in Proposition \ref{lem:greenMmn}. In particular, $R_X^J=L_X^J=H_X^J=D_X^J=\{X\}$ if $\rank(X)>r$. If $X\in\Mmn\sm P$, then $H_X^J=\{X\}$ is a non-group $\gHJ$-class of $\MmnJ$. \epfres
\end{thm}
Eggbox diagrams of some linear sandwich semigroups are given in Figures \ref{fig:V3212_V3322} and \ref{fig:V2322_V2422}. As usual, grey boxes indicate group $\gHJ$-classes; a label of {\tt k} on such a group $\gHJ$-class indicates isomorphism to $\G_k$. Note that the bottom diagram from Figure \ref{fig:V3212_V3322} is of a \emph{variant} of $\M_3(\Z_2)=\M_{33}(\Z_2)$. The diagrams in the pdf version of this article may be zoomed in a long way. The authors may be contacted for more such pictures.
\begin{figure}[ht]
\begin{center}
\includegraphics[width=\textwidth]{V3212.pdf}
\\~\\
\includegraphics[width=\textwidth]{V3322.pdf}
\caption{Egg box diagrams of the linear sandwich semigroups $\M_{32}^{J_{231}}(\Z_2)$ and $\M_{33}^{J_{332}}(\Z_2)$ (top and bottom, respectively).}
\label{fig:V3212_V3322}
\end{center}
\end{figure}
\begin{figure}[ht]
\begin{center}
\includegraphics[height=5cm]{V3222.pdf}
\qquad\qquad\qquad
\includegraphics[height=5cm]{V2422.pdf} \\~\\
\caption{Egg box diagrams of the linear sandwich semigroups $\M_{32}^{J_{232}}(\Z_2)$ and $\M_{24}^{J_{422}}(\Z_2)$ (left and right, respectively).}
\label{fig:V2322_V2422}
\end{center}
\end{figure}
Theorem \ref{green_thm} yields an intuitive picture of the internal structure of $\MmnJ$. Recall that the $\gD$-classes of $\Mmn$ are the sets $\DMmn s=\set{X\in\Mmn}{\rank(X)=s}$ for $0\leq s\leq l=\min(m,n)$. If $r<l$, then each of the $\gD$-classes $\DMmn{r+1},\ldots,\DMmn l$ separates completely into singleton $\gDJ$-classes in $\MmnJ$. (We will study these classes in more detail shortly.) Next, note that $\DMmn0=\{O\}\sub P$ (as the zero matrix clearly belongs to both $P_1$ and $P_2$), so $\DMmn0$ remains a (regular) $\gDJ$-class of $\MmnJ$.
Now fix some $1\leq s\leq r$. The $\gD$-class $\DMmn s$ is split into a single regular $\gDJ$-class, namely $\DMmn s\cap P$, and a number of non-regular $\gDJ$-classes. Some of these non-regular $\gDJ$-classes are singletons, namely those of the form $D_X^J=\{X\}$ where $X\in \DMmn s$ belongs to neither $P_1$ nor $P_2$. Some of the non-regular $\gDJ$-classes consist of one non-singleton $\gLJ$-class, namely those of the form $D_X^J=L_X^J=L_X\cap P_2$, where $X\in \DMmn s$ belongs to $P_2\sm P_1$; the $\gHJ$-classes contained in such a $\gDJ$-class are all singletons. The remaining non-regular $\gDJ$-classes contained in $\DMmn s$ consist of one non-singleton $\gRJ$-class, namely those of the form $D_X^J=R_X^J=R_X\cap P_1$, where $X\in \DMmn s$ belongs to $P_1\sm P_2$; the $\gHJ$-classes contained in such a $\gDJ$-class are all singletons. This is all pictured in Figure \ref{fig:green} for the $\D$-class $D_1(\M_{23})$ where $\F=\mathbb Z_3=\{0,1,2\}$ and $J=J_{321}=\left[\begin{smallmatrix}1&0\\0&0\\0&0\end{smallmatrix}\right]$; cf.~Figure \ref{fig:eggbox}.
\newcommand{\fillboxbig}[2]{\draw[fill=gray!30](#1,#2)--(#1+1,#2)--(#1+1,#2+1)--(#1,#2+1)--(#1,#2);}
\begin{figure}[ht]
\begin{center}
\scalebox{.8}{
\begin{tikzpicture}[scale=0.9]
\draw[fill=lightgray!50,line width=.5mm] (0.75,2.5)--(14.25,2.5)--(14.25,8.5)--(0.75,8.5)--(0.75,2.5)-- (14.25,2.5); \bdmat{1.0}{7.9}100000
\dmat{1.0}{7.1}200000
\dmat{1.0}{5.9}100100
\dmat{1.0}{5.1}200200
\dmat{1.0}{3.9}100200
\dmat{1.0}{3.1}200100
\rldmat{1}{1.5-1}000100
\rrdmat{1}{1.5-1}000200
\dmat2{7.9}101000
\dmat2{7.1}202000
\dmat2{5.9}101101
\dmat2{5.1}202202
\dmat2{3.9}101202
\dmat2{3.1}202101
\rldmat{2}{1.5-1}000101
\rrdmat{2}{1.5-1}000202
\dmat3{7.9}102000
\dmat3{7.1}201000
\dmat3{5.9}102102
\dmat3{5.1}201201
\dmat3{3.9}102201
\dmat3{3.1}201102
\rldmat{3}{1.5-1}000102
\rrdmat{3}{1.5-1}000201
\dmat4{7.9}110000
\dmat4{7.1}220000
\dmat4{5.9}110110
\dmat4{5.1}220220
\dmat4{3.9}110220
\dmat4{3.1}220110
\rldmat{4}{1.5-1}000110
\rrdmat{4}{1.5-1}000220
\dmat5{7.9}111000
\dmat5{7.1}222000
\dmat5{5.9}111111
\dmat5{5.1}222222
\dmat5{3.9}111222
\dmat5{3.1}222111
\rldmat{5}{1.5-1}000111
\rrdmat{5}{1.5-1}000222
\dmat6{7.9}112000
\dmat6{7.1}221000
\dmat6{5.9}112112
\dmat6{5.1}221221
\dmat6{3.9}112221
\dmat6{3.1}221112
\rldmat{6}{1.5-1}000112
\rrdmat{6}{1.5-1}000221
\dmat7{7.9}120000
\dmat7{7.1}210000
\dmat7{5.9}120120
\dmat7{5.1}210210
\dmat7{3.9}120210
\dmat7{3.1}210120
\rldmat{7}{1.5-1}000120
\rrdmat{7}{1.5-1}000210
\dmat8{7.9}121000
\dmat8{7.1}212000
\dmat8{5.9}121121
\dmat8{5.1}212212
\dmat8{3.9}121212
\dmat8{3.1}212121
\rldmat{8}{1.5-1}000121
\rrdmat{8}{1.5-1}000212
\dmat{9}{7.9}122000
\dmat{9}{7.1}211000
\dmat{9}{5.9}122122
\dmat{9}{5.1}211211
\dmat{9}{3.9}122211
\dmat{9}{3.1}211122
\rldmat{9}{1.5-1}000122
\rrdmat{9}{1.5-1}000211
\dmat{10.666}{8.0}010000
\dmat{10.666}{7.0}020000
\dmat{10.666}{6.0}010010
\dmat{10.666}{5.0}020020
\dmat{10.666}{4.0}010020
\dmat{10.666}{3.0}020010
\dmat{10.666}{2.0-1}000010
\dmat{10.666}{1.0-1}000020
\dmat{11.666}{8.0}011000
\dmat{11.666}{7.0}022000
\dmat{11.666}{6.0}011011
\dmat{11.666}{5.0}022022
\dmat{11.666}{4.0}011022
\dmat{11.666}{3.0}022011
\dmat{11.666}{2.0-1}000011
\dmat{11.666}{1.0-1}000022
\dmat{12.666}{8.0}012000
\dmat{12.666}{7.0}021000
\dmat{12.666}{6.0}012012
\dmat{12.666}{5.0}021021
\dmat{12.666}{4.0}012021
\dmat{12.666}{3.0}021012
\dmat{12.666}{2.0-1}000012
\dmat{12.666}{1.0-1}000021
\dmat{13.666}{8.0}001000
\dmat{13.666}{7.0}002000
\dmat{13.666}{6.0}001001
\dmat{13.666}{5.0}002002
\dmat{13.666}{4.0}001002
\dmat{13.666}{3.0}002001
\dmat{13.666}{2.0-1}000001
\dmat{13.666}{1.0-1}000002
\foreach \x in {4,6} \draw (0.75,\x+.5)--(14.25,\x+.5);
\foreach \x in {3,...,7} \foreach \y in {14.5,16,17.5,19} \draw (0.75+.15+\y,\x+.5)--(2.25-.15+\y,\x+.5);
\foreach \x in {1,...,8} \draw (\x*1.5+0.75,2.5)--(\x*1.5+0.75,8.5);
\foreach \x in {0.5,1,1.5,2,...,8,8.5} \draw (\x*1.5+0.75,2+.35-1)--(\x*1.5+0.75,1-.35-1);
\foreach \x in {1,2.5,4,5.5} \foreach \y in {0,-1} {\draw[line width=.5mm] (15-.6+\x,1-.35+\y)--(15+.6+\x,1-.35+\y)-- (15+.6+\x,1+.35+\y)--(15-.6+\x,1+.35+\y)--(15-.6+\x,1-.35+\y)--(15+.6+\x,1-.35+\y);}
\foreach \x in {1,2.5,4,5.5} {\draw[line width=.5mm] (15-.6+\x,1-.5+2)--(15+.6+\x,1-.5+2)-- (15+.6+\x,8.5)--(15-.6+\x,8.5)--(15-.6+\x,1-.5+2)--(15+.6+\x,1-.5+2);}
\draw[line width=.5mm] (0.75,2.5-.15-1)--(14.25,2.5-.15-1)--(14.25,0.5+.15-1)--(0.75,0.5+.15-1) --(0.75,2.5-.15-1)--(14.25,2.5-.15-1);
\draw[|-|] (0.75,9.0)--(14.25,9.0);
\draw[|-|] (20.25+1,9.0)--(14.25+1,9.0);
\draw[|-|] (.25,.5-1)--(.25,2.5-1);
\draw[|-|] (.25,8.5)--(.25,2.5);
\draw(7.5,9.3)node{$\sub P_1\phantom{\sub}$};
\draw(16.75+1.5,9.3)node{$\not\sub P_1\phantom{\sub}$};
\draw(.25,5.5)node[left]{$\sub P_2\phantom{}$};
\draw(.25,1.5-1.0)node[left]{$\not\sub P_2\phantom{}$};
\end{tikzpicture}
}
\caption{A $\gD$-class $D_1(\M_{23}(\mathbb Z_3))$ breaks up into $\gDJ$-classes in $\M_{23}^{J}(\mathbb Z_3)$, where $J=J_{321}$. Group $\gH^J$-classes are shaded grey; the idempotent of such a group is the upper of the two matrices. (cf.~Figure \ref{fig:eggbox}.)}
\label{fig:green}
\end{center}
\end{figure}
It will be important to have a description of the partial order $\leq$ on the $\gDJ$-classes of $\MmnJ$.
\ms
\begin{prop}\label{prop:DorderMmnJ}
Let $X,Y\in\Mmn$. Then $D_X^J\leq D_Y^J$ in $\MmnJ$ if and only if one of the following holds:
\ms
\bit
\begin{multicols}{2}
\itemit{i} $X=Y$,
\itemit{ii} $\rank(X)\leq\rank(JYJ)$,
\itemit{iii} $\Row(X)\sub\Row(JY)$,
\itemit{iv} $\Col(X)\sub\Col(YJ)$.
\end{multicols}
\end{itemize}
\end{prop}
\pf Note that $D_X^J\leq D_Y^J$ if and only if one of the following holds:
\ms
\bmc2
\item[(a)] $X=Y$,
\item[(b)] $X=UJYJV$ for some $U,V\in\Mmn$,
\item[(c)] $X=UJY$ for some $U\in \Mmn$,
\item[(d)] $X=YJV$ for some $V\in\Mmn$.
\emc
The equivalences (b) $\Leftrightarrow$ (ii), (c) $\Leftrightarrow$ (iii), and (d) $\Leftrightarrow$ (iv) all follow from Lemma \ref{lem:green<M}. \epf
The description of the order on $\gDJ$-classes of $\MmnJ$ from Proposition \ref{prop:DorderMmnJ} may be simplified in the case that one of $X,Y$ is regular.
\ms
\begin{prop}\label{prop:DorderP}
Let $X,Y\in\Mmn$.
\bit
\itemit{i} If $X\in P$, then $D_X^J\leq D_Y^J\iff\rank(X)\leq\rank(JYJ)$.
\itemit{ii} If $Y\in P$, then $D_X^J\leq D_Y^J\iff\rank(X)\leq\rank(Y)$.
\eit
The regular $\gDJ$-classes of $\MmnJ$ form a chain: $D_0^J<\cdots<D_r^J$, where $$D_s^J=\DMmn s\cap P=\set{X\in P}{\rank(X)=s} \qquad\text{for each $0\leq s\leq r$.}$$
\end{prop}
\pf As in the proof of Proposition \ref{prop:DorderMmnJ}, $D_X^J\leq D_Y^J$ if and only if one of (a--d) holds.
Suppose first that $X\in P$, so $X=XJZJX$ for some $Z\in\Mmn$. Then (a) implies $X=XJZ(JYJ)ZJX$, (c) implies $X=U(JYJ)ZJX$, and (d) implies $X=XJZ(JYJ)V$. So, in each of cases (a--d), we deduce that $\rank(X)\leq\rank(JYJ)$. So $D_X^J\leq D_Y^J$ implies $\rank(X)\leq\rank(YJY)$. Proposition \ref{prop:DorderMmnJ} gives the reverse implication.
Next, suppose $Y\in P$. Now, each of (a--d) implies $\rank(X)\leq\rank(Y)$. Conversely, if $\rank(X)\leq\rank(Y)$, then Proposition~\ref{prop:DorderMmnJ} gives $D_X^J\leq D_Y^J$, since $\rank(Y)=\rank(JYJ)$.
The statement about regular $\gDJ$-classes follows quickly from (ii).~\epf
The linear ordering on the regular $\gDJ$-classes may be seen by inspecting Figures \ref{fig:V3212_V3322} and \ref{fig:V2322_V2422}; see also Figure~\ref{fig:R}.
As an immediate consequence of Proposition \ref{prop:DorderP}, we may classify the isomorphism classes of sandwich semigroups on the set~$\Mmn$; the $m=n$ case of the next result was proved in \cite{JCK2010}.
\ms
\begin{cor}\label{cor:MmnAcongMmnB}
Let $A,B\in\Mnm$. Then $\MmnA\cong\MmnB$ if and only if $\rank(A)=\rank(B)$.
\end{cor}
\pf Put $r=\rank(A)$ and $s=\rank(B)$. By Proposition \ref{prop:DorderP} and Lemma \ref{lem:MmnAMmnB}(ii), $\MmnA\cong\Mmn^{J_{nmr}}$ and $\MmnB\cong\Mmn^{J_{nms}}$ have $r+1$ and $s+1$ regular $\D^A$- and $\D^B$-classes, respectively. So $\MmnA\cong\MmnB$ implies $r=s$. The converse was proved in Lemma~\ref{lem:MmnAMmnB}(ii).~\epf
\begin{rem}
It is possible to have $\M_{mn}^A\cong\M_{kl}^B$ even if $(m,n)\not=(k,l)$, although we would of course still need $\rank(A)=\rank(B)$ by Proposition \ref{prop:DorderP}. For example, if $O=O_{nm}$ is the $n\times m$ zero matrix, then $\M_{mn}^O$ is a \emph{zero semigroup} ($X\star Y=O_{mn}$ for all $X,Y\in\M_{mn}$). Two such zero semigroups $\Mmn^O$ and $\M_{kl}^O$ are isomorphic if and only if they have the same cardinality; that is, if and only if $\F$ is infinite or $\F$ is finite and $mn=kl$. We will return to the problem of distinguishing non-isomorphic $\M_{mn}^A$ and $\M_{kl}^B$ in Theorem \ref{thm:classification}. See Figure \ref{fig:V2202}.
\end{rem}
\begin{figure}[ht]
\begin{center}
\includegraphics[width=12cm]{V2202.pdf}
\caption{Egg box diagram of the linear sandwich semigroup $\M_{22}^{O_{22}}(\Z_2)$ or, equivalently, $\M_{21}^{O_{12}}(\F_4)$. Both are zero semigroups of size $16$.}
\label{fig:V2202}
\end{center}
\end{figure}
The next result describes the maximal $\gDJ$-classes of $\MmnJ$. See also Figures \ref{fig:V3212_V3322} and \ref{fig:V2322_V2422}.
\ms\ms\ms
\begin{prop}\label{prop_maximalD}
\bit
\itemit{i} If $r=\min(m,n)$, then $D_r^J=D_r\cap P=\set{X\in P}{\rank(X)=r}$ is the unique maximal $\gDJ$-class of~$\MmnJ$, and is a subsemigroup of $\MmnJ$.
\itemit{ii} If $r<\min(m,n)$, then the maximal $\gDJ$-classes of $\MmnJ$ are those of the form $D_X^J=\{X\}$ with $\rank(X)>r$.
\eitres
\end{prop}
\pf Part (i) follows immediately from Proposition \ref{prop:DorderP}(ii), the rule $[M,A,N]\star[K,B,L]=[M,AB,L]$, and the fact that $\G_r=D_r(\M_r)$ is a subgroup of $\M_r$.
For (ii), let $X\in\Mmn$. Suppose first that $\rank(X)>r$ and that $D_X^J\leq D_Y^J$. Then condition (ii) from Proposition \ref{prop:DorderMmnJ} does not hold, since $\rank(JYJ)\leq\rank(J)=r<\rank(X)$. Similarly, $\rank(JY)<\rank(X)$ and $\rank(YJ)<\rank(X)$, so neither (iii) nor (iv) holds. Having eliminated (ii--iv), we deduce that (i) must hold; that is, $X=Y$, so $D_X^J=\{X\}$ is indeed maximal.
Conversely, suppose
$\rank(X)\leq r$, and let $Y=\tmat {I_r}OOD$, where $D\not=O$. Then $\rank(Y)>r$, so $D_Y^J=\{Y\}$ is maximal by the previous paragraph. But also $JYJ=J$, and it follows that $\rank(X)\leq r=\rank(J)=\rank(JYJ)$, so that $D_X^J< D_Y^J=\{Y\}$, whence $D_X^J$ is not maximal. \epf
The description of the maximal $\gDJ$-classes from Proposition \ref{prop_maximalD} allows us to obtain information about generating sets for $\MmnJ$ and, in the case of finite $\F$, about $\rank(\MmnJ)$.
In order to avoid confusion when discussing generation, if $\Om\sub\Mmn$, we will write $\la \Om\raJ$ for the subsemigroup of $\MmnJ$ generated by $\Om$, which consists of all products $X_1\star\cdots\star X_k$, with $k\geq1$ and $X_1,\ldots,X_k\in \Om$. If $\Si\sub\M_k$ for some $k$, we will continue to write $\la\Si\ra$ for the subsemigroup of $\M_k$ generated by $\Si$. For convenience, we will state two separate results, according to whether $r=\min(m,n)$ or $r<\min(m,n)$. The next lemma will be useful as the inductive step in the proofs of both Theorems \ref{thm:rankMmnJ} and \ref{thm:rankMmnJ_r=m}. Recall that $\{e_{m1},\ldots,e_{mm}\}$ is the standard basis of $V_m=\F^m$.
\ms
\begin{lemma}\label{lem:ind_step}
Suppose $X\in \DMmn s$, where $0\leq s\leq l-1$ and $l=\min(m,n)$. Then $X=Y\star Z$ for some $Y\in \DMmn l$ and $Z\in\DMmn{s+1}$.
\end{lemma}
\pf
Let $\B=\{v_1,\ldots,v_n\}$ be a basis of $V_n$ such that $\{v_{s+1},\ldots,v_n\}$ is a basis of $\ker(\lam_X)$.
Consider the linear transformation $\be\in\Homnm$ defined by
\[
\be(v_i) = \begin{cases}
e_{mi} &\text{if $1\leq i\leq s$}\\
0 &\text{if $s<i<n$}\\
e_{mm} &\text{if $i=n$,}
\end{cases}
\]
noting that $\rank(\be)=s+1$. The proof now breaks into two cases, depending on whether $r<m$ or $r=m$.
{\bf Case 1.} Suppose first that $r<m$. Let $\al\in\Homnm$ be any linear transformation of rank $l$ that extends the map $e_{ni}\mt \lam_X(v_i)$ ($1\leq i\leq s$). One easily checks that $\al\circ\lam_J\circ\be=\lam_X$.
{\bf Case 2.} Now suppose $r=m$. Recall that we are assuming that $r=m=n$ does not hold,
so $r=m<n$. This time, define we let $\al$ be any linear transformation of rank $m=l$ that extends the map $e_{ni}\mt\lam_X(v_i)$ ($1\leq i\leq s$), $e_{nr}=e_{nm}\mt0$. Then, again, one easily checks that $\al\circ\lam_J\circ\be=\lam_X$. \epf
\ms
\begin{thm}\label{thm:rankMmnJ}
Suppose $r<l=\min(m,n)$. Then $\MmnJ=\la \Om\raa$, where $\Om=\set{X\in\Mmn}{\rank(X)>r}$. Further, any generating set for $\MmnJ$ contains $\Om$.
If $|\F|=q<\infty$, then
\[
\rank(\MmnJ)=|\Om|=\sum_{s=r+1}^{l} \qbin ms \qbin ns q^{{s\choose2}}(q-1)^s\qfact s.
\]
\end{thm}
\pf For convenience, we will assume that $l=m\leq n$. The other case will follow by duality. We will also denote $D_s(\Mmn)$ simply by $D_s$ for each $0\leq s\leq m$. Consider the statement:
\nss
\begin{quote}
$H(s)$: ~ $\la \Om\raa$ contains $D_s\cup\cdots\cup D_m=\set{X\in\Mmn}{\rank(X)\geq s}$.
\end{quote}
\nss
Note that $\Om=D_{r+1}\cup\cdots\cup D_m$, so $H(s)$ is clearly true for $r+1\leq s\leq m$. Lemma \ref{lem:ind_step} shows that $H(s+1)$ implies $H(s)$ for all $0\leq s\leq m-1$. So we conclude that $H(s)$ is true for all $0\leq s\leq m$.
In particular, $H(0)$ says that $\MmnJ=\la \Om\raa$.
Since $\{X\}$ is a maximal $\gDJ$-class for any $X\in \Om$, it follows that any generating set of $\MmnJ$ must contain~$\Om$. Thus, $\Om$ is the minimal generating set with respect to both size and containment, so $\rank(\MmnJ)=|\Om|$. The formula for $|\Om|$ with $|\F|$ finite follows from Lemma \ref{lem:combinatorics_Mmn}. \epf
In order to consider the case in which $r=\min(m,n)$, we first prove an intermediate result. There is a dual version of the following lemma (dealing with the case in which $r=n<m$), but we will not state it.
\ms
\begin{lemma}\label{lem:r=m}
If $r=m<n$, then
~\ \ \emph{(i)} $P_2=\MmnJ$, \qquad
\emph{(ii)} $P=P_1$ is a left ideal of $\MmnJ$, \qquad
\emph{(iii)} $\gLJ=\L$ in $\MmnJ$.
\end{lemma}
\pf Let $X\in\MmnJ$. As noted earlier, in the $2\times2$ block description, $X=\tmat ABCD$ (where $A\in\M_r$, and so on), the matrices $C$ and $D$ are empty (since $r=m$). So we write $X=[A\ B]$. Note that $J=\thmat IO$, so $JX=\thmat IO[A\ B]=\tmat ABOO$. It follows that $\Row(JX)=\Row(X)$ and, since $X\in\MmnJ$ was arbitrary, this completes the proof of (i).
We immediately deduce $P=P_1$ from (i). As in Proposition~\ref{prop:P1P2}, the regular elements of $\MmnJ$ are of the form $[A\ AN]$ where $A\in\M_r$ and $N\in\M_{r,n-r}$. We denote such a regular element by $[A,N]$. The proof of~(ii) concludes with the easily checked observation that $[A\ B]\star[C,N]=[AC,N]$.
Part (iii) follows quickly from (i) and Theorem \ref{green_thm}(ii). \epf
\ms
\begin{thm}\label{thm:rankMmnJ_r=m}
Suppose $r=\min(m,n)$ where $m\not=n$. If $|\F|=q<\infty$, then
\[
\rank(\MmnJ)
= \qbin Ll,
\]
where $L=\max(m,n)$ and $l=\min(m,n)$.
\end{thm}
\pf Again, it suffices to assume that $r=m<n$, so $l=m$ and $L=n$. We keep the notation of the previous proof.
Let $\Om$ be an arbitrary generating set for $\MmnJ$. Let $X\in D_m(\Mmn)$ be arbitrary. We claim that $\Om$ must contain some element of $L_X^J=L_X$. Indeed, consider an expression $X=Y_1\star\cdots\star Y_k$, where $Y_1,\ldots,Y_k\in\Om$. If $k=1$, then $Y_1=X\in L_X$ and the claim is established, so suppose $k\geq2$.
Since $D_m(\Mmn)$ is a maximal $\gDJ$-class, we must have $Y_k\in D_m(\Mmn)$. So $Y_k \gDJ X = (Y_1\star\cdots\star Y_{k-1})\star Y_k$, whence $Y_k\gLJ (Y_1\star\cdots\star Y_{k-1})\star Y_k = X$, by stability. By Lemma \ref{lem:r=m}(iii), this completes the proof of the claim.
In particular, $|\Om|$ is bounded below by the number of $\gL$-classes contained in $D_m(\Mmn)$, which is equal to $\tqbin nm$, by Lemma~\ref{lem:combinatorics_Mmn}. Since $\Om$ was an arbitrary generating set, it follows that $\rank(\MmnJ)\geq\tqbin nm=\tqbin Ll$.
To complete the proof, it remains to check that there exists a generating set of the desired cardinality.
For each $N\in\M_{m,n-m}$, choose some $A_N\in\G_r$ such that $\set{A_N}{N\in\M_{m,n-m}}$ generates $\G_m$, and put $X_N=[A_N,N]\in D_m^J$. (This is possible since $|\M_{m,n-m}|=q^{m(n-m)}\geq2$, and $\rank(\G_m)\leq2$ by Theorem~\ref{thm:waterhouse}.) It is easy to see that $\Om_1=\set{X_N}{N\in\M_{m,n-m}}$ is a cross-section of the $\L$-classes in $D_m^J$. Also, choose some cross-section $\Om_2=\set{Y_i}{i\in I}$ of the $\L$-classes contained in $D_m(\Mmn)\sm D_m^J$. Then
$
\Om = \Om_1\cup\Om_2
$
is a cross-section of the $\L$-classes contained in $D_m(\Mmn)$. Since, therefore, $|\Om|=\tqbin nm$, the proof will be complete if we can show that $\MmnJ=\la\Om\raJ$. By Lemma \ref{lem:ind_step}, it suffices to show that $\la\Om\raJ$ contains $D_m(\Mmn)$. So suppose $Z\in D_m(\Mmn)$. Assume first that $Z\in D_m^J$, and write $Z=[B,L]$, noting that $B\in\G_r$. Choose $N_1,\ldots,N_k\in\M_{m,n-m}$ such that $BA_L^{-1}=A_{N_1}\cdots A_{N_k}$. Then one easily checks that $Z=X_{N_1}\star\cdots\star X_{N_k}\star X_L$. Now, suppose $Z$ is not regular. Choose $i\in I$ such that $Z\L Y_i$. By Lemma \ref{lem:greenMmn}, $Z=UY_i$ for some $U\in\G_m$. But then $Z=[U\ V]\star Y_i$ for any $V\in\M_{m,n-m}$. Since $\rank(U)=m$, we have $[U\ V]\in D_m^J\sub\la\Om\raJ$, whence $Z\in\la\Om\raJ$, completing the proof. \epf
\begin{rem}
By inspecting Figures \ref{fig:V3212_V3322} and \ref{fig:V2322_V2422}, the reader may use Theorems \ref{thm:rankMmnJ} and \ref{thm:rankMmnJ_r=m} to locate the elements from a minimal generating set for $\MmnJ$.
\end{rem}
\section{Connection to (non-sandwich) matrix semigroups}\label{sect:non-sandwich}
Recall that $J=J_{nmr}=\tmat {I_r}{O_{r,n-r}}{O_{m-r,r}}{O_{m-r,n-r}}\in\M_{nm}$. Now let $K=J^T=J_{mnr}=\tmat {I_r}{O_{r,m-r}}{O_{n-r,r}}{O_{n-r,m-r}}\in\M_{mn}$. So Lemma \ref{lem:MmnAMmnB} says that $\MnmK$ and $\MmnJ$ are anti-isomorphic. Also, since $J=JKJ$ and $K=KJK$, Theorem \ref{thm:diamonds} says that we have the following commutative diagrams of semigroup homomorphisms where, for clarity, we write $\cdot$ for (non-sandwich) matrix multiplication:
\[
\includegraphics{CD2.pdf}
\]
In this section, we show that the various semigroups appearing in the above diagrams are all (equal to or isomorphic to) certain well-known (non-sandwich) matrix semigroups, and explore the consequences for the structure of the sandwich semigroups $\MmnJ$. First, we have a simple observation.
\ms
\begin{lemma}\label{lem:JMJ}
We have $(J\Mmn J,\star_K)=\Reg(J\Mmn J,\star_K)\cong(\M_r,\cdot)$.
\end{lemma}
\pf Let $X=\tmat ABCD\in\Mmn$. We have already observed that, whether $X$ is regular or not, $JXJ=\tmat AOOO\in\Mnm$. The result follows quickly from the fact that $\tmat AOOO\star_K \tmat EOOO=\tmat{AE}OOO$. \epf
For integers $k\geq1$ and $0\leq l\leq k$, we write
\begin{align*}
\cC_k(l) &= \set{X\in\M_k}{\col_{l+1}(X)=\cdots=\col_k(X)=O}, \\
\cR_k(l) &= \set{X\in\M_k}{\row_{l+1}(X)=\cdots=\row_k(X)=O}.
\intertext{(As before, without causing confusion, we write $O$ for any zero matrix when the dimensions are clear from context.) These matrix semigroups have been studied in a number of contexts (see for example \cite{NK2007,Sullivan2008}), along with their associated isomorphic semigroups of linear transformations}
\mathcal K_k(l) &= \bigset{\al\in\End(V_k)}{\ker(\al)\supseteq W_{kl}^\perp},\\
\mathcal I_k(l) &= \bigset{\al\in\End(V_k)}{\im(\al)\sub W_{kl}}.
\end{align*}
Here we have written $W_{kl}^\perp=\Span\{e_{k,l+1},\ldots,e_{kk}\}$. Clearly, $\Ckl$ and $\Rkl$ are anti-isomorphic.
\ms
\begin{lemma}
We have $\MMNJ=\Cmr$ and $\JMMN=\Rnr$.
\end{lemma}
\pf Let $X=\tmat ABCD\in\Mmn$. We have already observed that $XJ=\tmat AOCO\in\M_m$ and $JX=\tmat ABOO$, and the result quickly follows. \epf
\begin{rem}\label{rem:r=m2}
A typical element $X\in\Rkl$ may be written as $X=\tmat ABOO$, where $A\in\M_l$, $B\in\M_{l,k-l}$ and so on. One easily checks that multiplication of matrices in this form obeys the rule $\tmat ABOO \tmat EFOO = \tmat {AE}{AF}OO$. Comparing this to the discussion in Remark \ref{rem:r=m}, we see that $\Rkl$ is isomorphic to the sandwich semigroup $\M_{lk}^J$ where $J=J_{kll}\in\M_{kl}$. (A dual statement holds for the matrix semigroups $\Ckl$.) Thus, every result we obtain for linear sandwich semigroups leads to analogous results for the semigroups $\Rkl$ and $\Ckl$. For example, we deduce from Theorem \ref{thm:rankMmnJ_r=m} that
$
\rank(\Ckl) = \rank(\Rkl) = \tqbin kl
$
if $|\F|=q<\infty$.
Note that the sandwich semigroups $\MmnJ$ pictured in Figure \ref{fig:V2322_V2422} satisfy $r=\min(m,n)$, so Figure \ref{fig:V2322_V2422} essentially pictures eggbox diagrams of $\mathcal C_3(2)$ and $\mathcal R_4(2)$.
\end{rem}
\ms
\begin{rem}
Similarly, one may think of an arbitrary linear sandwich semigroup $\MmnJ$ itself as a (non-sandwich) matrix semigroup, as noted by Thrall in \cite{Thrall1955} and slightly adapted as follows. Consider the set $\mathscr M$ of all $(m+n-r)\times (m+n-r)$ matrices
that may be written in $3\times3$ block form
$\tmatt OOOBAODCO$,
where $A\in\M_r$, $D\in\M_{m-r,n-r}$ (and from which the dimensions of the other sub-matrices may be derived). One easily checks that the matrices from $\mathscr M$ multiply according to the rule
$\tmatt OOOBAODCO \tmatt OOOFEOHGO = \tmatt OOO{AF}{AE}O{CF}{CE}O$,
so that
$\tmat ABCD \mt \tmatt OOOBAODCO$
determines an isomorphism $(\Mmn,\star_J)\to(\mathscr M,\cdot)$. Note also that
\begin{align*}
\mathscr M &= \cR_{m+n-r}^*(m) \cap \cC_{m+n-r}(n)
\end{align*}
where here we write
$\cR_k^*(l) = \set{X\in\M_k}{\row_1(X)=\cdots=\row_{k-l}(X)=O}$.
(It is easily seen that the map $\tmat ABOO\to\tmat OOBA$ determines an isomorphism $\Rkl\to \cR_k^*(l)$.)
Since using this isomorphic copy $\mathscr M$ of $\MmnJ$ does not appear to confer any obvious advantage, we will make no further reference to it.
\end{rem}
The regular elements of $\Ckl$ and $\Rkl$
(and also of $\mathcal I_k(l)$ and $\mathcal K_k(l)$)
were classified in \cite{NK2007}. The next result, which gives a much simpler description of these regular elements, may be deduced from \cite[Theorems~3.4 and 3.8]{NK2007} (and vice versa), but we include a simple proof for convenience.
\ms
\begin{prop
\label{cor:RegMNJ}
The regular elements of the semigroups $\Cmr=\MMNJ$ and $\Rnr=\JMMN$ are given by
\begin{align*}
\Reg(\Cmr) = \Reg(\MMNJ) = PJ &= \set{X\in\MMNJ}{\rank(JX)=\rank(X)} \\
\Reg(\Rnr) = \Reg(\JMMN) = JP &= \set{X\in\JMMN}{\rank(XJ)=\rank(X)}.
\end{align*}
\end{prop}
\pf We just prove the second statement as the other is dual. Let $X=\tmat ABCD\in\Mmn$, and put $X'=\tmat ABOO\in\Mmn$. Then $JX=JX'=\tmat ABOO\in\M_n$ (where the zero matrices in the last expression have $n-r$ rows). Since $X'$ clearly belongs to $P_2$ (by Proposition \ref{prop:P1P2}), we have $\JMMN\sub JP_2$. Next, note that $KJ=J_{mmr}$, so that $KJY=Y$ for all $Y\in\M_{mn}$ of the form $Y=\tmat ABOO$. Now suppose $X\in\Mmn$ is such that $JX\in\RegJMMN$. As above, we may assume that $X=\tmat ABOO$. So $(JX)=(JX)(JY)(JX)$ for some $Y\in\Mmn$. But then
$
X = K(JX) = K(JXJYJX)=XJYJX = X\star Y\star X,
$
so that, in fact, $X\in \RegMmnJ=P$. This completes the proof that $\RegJMMN\sub JP$. The reverse inclusion is easily checked.
Now suppose $X=JY$ where $Y=\tmat A{AN}OO\in P$. Then $X=JY=\tmat A{AN}OO$ (with appropriately sized zero matrices), so $\rank(X)=\rank(JY)=\rank(Y)=\rank(JYJ)=\rank(XJ)$, where we have used Proposition~\ref{prop:P1P2}. Conversely, suppose $X\in\JMMN$ is such that $\rank(XJ)=\rank(X)$. As before, we may assume that $X=JY$ where $Y\in P_2$. Then $\rank(Y)=\rank(JY)=\rank(X)=\rank(XJ)=\rank(JYJ)$, so that $Y\in P$. This completes the proof. \epf
\begin{rem}
As always, the condition $\rank(JX)=\rank(X)$, for $X\in\Mmn$, is equivalent to saying that rows $\row_{r+1}(X),\ldots,\row_m(X)$ belong to $\Span\{\row_1(X),\ldots,\row_r(X)\}$, with a dual statement holding for the condition $\rank(XJ)=\rank(X)$. The regular elements of the corresponding semigroups of linear transformations are given by
\begin{align*}
\Reg(\K_m(r)) &= \set{\al\in\K_m(r)}{\im(\al)\cap W_{mr}^\perp=\{0\}}, \\
\Reg(\I_n(r)) &= \set{\al\in\I_n(r)}{\im(\al)=\al(W_{nr})}.
\end{align*}
\end{rem}
Putting together all the above, we have proved the following. (In the following statement, we slightly abuse notation by still denoting the map $\mathcal C_m(r)=\Mmn J\to\M_r$ by $\Phi_1$ and so on.)
\ms
\begin{thm}\label{thm:diamondsMmnJ}
We have the following commutative diagrams of semigroup epimorphisms:
\[
\includegraphics{CD3.pdf}
\]
~ \epfres
\end{thm}
The remaining results of this section concern the regular subsemigroup $P=\Reg(\MmnJ)$. From now on, we denote by $\phi=\phi_1\circ\psi_1=\phi_2\circ\psi_2$ the induced epimorphism $\phi:P\to\M_r$. Also, for $X=\tmat ABCD\in P$, we write $\Xb=\phi(X)=A$. The next result shows how the second commutative diagram from Theorem \ref{thm:diamondsMmnJ} may be used to identify $\RegMmnJ$ as a special kind of subdirect product of $\Reg(\Cmr)$ and $\Reg(\Rnr)$.
\ms
\begin{prop}\label{mono_prop}
There is an embedding
\[
\psi : \RegMmnJ\to \Reg(\Cmr)\times\Reg(\Rnr): X\mt (\psi_1(X),\psi_2(X))=(XJ,JX).
\]
As such, $P=\RegMmnJ$ is (isomorphic to) a pullback product of $PJ=\Reg(\Cmr)$ and $JP=\Reg(\Rnr)$. Namely,
\[
P\cong\im(\psi)=\bigset{(\psi_1(X),\psi_2(X))}{X\in P} = \bigset{(Y,Z)\in PJ\times JP}{\phi_1(Y)=\phi_2(Z)}.
\]
\end{prop}
\pf Clearly, $\psi$ is a homomorphism. Now let $X=[M,A,N]$ and $Y=[K,B,L]$ be elements of $P$
with $\psi(X)=\psi(Y)$. Then
\[
\left(\tmat{A}{O}{MA}{O},\tmat A{AN}OO \right) = \psi(X) = \psi(Y) = \left(\tmat{B}{O}{KB}{O},\tmat B{BL}OO \right).
\]
Comparing various coordinates, we deduce $A=B$, $MA=KB$ and $AN=BL$, giving
$X=\tmat A{AN}{MA}{MAN}=\tmat B{BL}{KB}{KBL}=Y$, completing the proof that $\psi$ is injective.
To prove the statement about $\im(\psi)$, let $X\in P$ and put $Y=\psi_1(X)=XJ$ and $Z=\psi_2(X)=JX$. Then $\phi_1(Y)=JY=JXJ=ZJ=\phi_2(Z)$. Conversely, suppose $(Y,Z)\in PJ\times JP$ satisfies $JY=ZJ$. Say $Y=UJ$ and $Z=JV$, where $U=[M,A,N]$ and $V=[K,B,L]$ belong to $P$. Then $JY=JUJ=\tmat AOOO$ and $ZJ=JVJ=\tmat BOOO$, giving $A=B$. But then $(Y,Z)=\psi(X)$, where $X=[M,A,L]\in P$. \epf
\begin{rem}
We note that the previous result does not lift to a similar identification of $\MmnJ$ as a pullback product of $\Cmr$ and $\Rnr$ because the induced map
\[
\Psi:\MmnJ\to\Cmr\times\Rnr:X\mt(\Psi_1(X),\Psi_2(X))=(XJ,JX)
\]
is not injective. Indeed, if $X=\tmat ABCD\in\Mmn$, then $\Psi(X)=\left(\tmat{A}{O}{C}{O},\tmat A{B}OO \right)$, with $\tmat ABCE$ mapping to the same pair for any other $E\in\M_{m-r,n-r}$.
\end{rem}
\ms
\begin{rem}
More generally, given a partial semigroup $(S,\cdot,I,\lam,\rho)$, the epimorphisms $\Psi_1$ and $\Psi_2$ from Theorem \ref{thm:diamonds}(v) allow for the definition of a map
\[
\Psi:(S_{ij},\star_a) \to (S_{ij}a,\cdot)\times (aS_{ij},\cdot):x\mt(xa,ax).
\]
To say that $\Psi$ is injective is to say that, for all $x,y\in S_{ij}$, $xa=ya$ and $ax=ay$ together imply $x=y$. Compare this to the notion of a \emph{weakly reductive} semigroup $S$, in which, for every $x,y\in S$, the assumption that $xa=ya$ and $ax=ay$ for all $a\in S$ implies $x=y$. See for example \cite[Definition 1.42]{Nagy2001}.
\end{rem}
We conclude this section with a simple but important observation that shows that $P=\RegMmnJ$ is a homomorphic image of the direct product of a rectangular band by the (non-sandwich) matrix semigroup~$\M_r$. (Recall that a \emph{rectangular band} is a semigroup of the form $S\times T$ with product $(s_1,t_1)(s_2,t_2)=(s_1,t_2)$.) Its proof is routine, relying on Proposition~\ref{prop:P1P2} and the rule $[M,A,N]\star[K,B,L] = [M,AB,L]$. For the statement, recall that the \emph{kernel} of a semigroup homomorphism $\phi:S\to T$ (not to be confused with the kernel of a linear transformation) is the congruence $\ker(\phi)=\set{(x,y)\in S\times S}{\phi(x)=\phi(y)}$. (A congruence on a semigroup $S$ is an equivalence relation $\sim$ for which $x_1\sim y_1$ and $x_2\sim y_2$ together imply $x_1x_2\sim y_1y_2$ for all $x_1,x_2,y_1,y_2\in S$; the quotient $S/{\sim}$ of all $\sim$-classes is a semigroup under the induced operation. The first homomorphism theorem for semigroups states that any semigroup homomorphism $\phi:S\to T$ induces an isomorphism $S/\ker(\phi)\cong \im(\phi)$.)
\ms
\begin{prop}\label{prop:MAN}
Consider the semigroup $U=\M_{m-r,r}\times\M_r\times\M_{r,n-r}$ under the operation $\diamond$ defined by
\[
(M,A,N)\diamond(K,B,L) = (M,AB,L).
\]
Define an equivalence $\sim$ on $U$ by
\[
(M,A,N) \sim (K,B,L) \ \ \iff \ \ \text{$A=B$, $MA=KB$ and $AN=BL$.}
\]
Then $\sim$ is a congruence on $U$, and the map
\[
\xi:U\to P=\RegMmnJ:(M,A,N)\mt[M,A,N]=\mat A{AN}{MA}{MAN}
\]
is an epimorphism with $\ker(\xi)={\sim}$. In particular, $P\cong U/{\sim}$. \epfres
\end{prop}
\section{The regular subsemigroup}\label{sect:RegMmnJ}
In this section, we continue to study the subsemigroup $P=\RegMmnJ$
consisting of all regular elements of~$\MmnJ$. Eggbox diagrams of $P=\Reg(\M_{43}^J(\Z_2))$ are given in Figure \ref{fig:R} for values of $0\leq\rank(J)\leq3$; more examples can be seen by inspecting the regular $\gDJ$-classes in Figures~\ref{fig:V3212_V3322} and \ref{fig:V2322_V2422}.
\begin{figure}[ht]
\begin{center}
\rotatebox[origin=c]{270}{
\includegraphics[width=0.55cm]{R4302.pdf}
\qquad
\includegraphics[width=1.3cm]{R4312.pdf}
\qquad
\includegraphics[width=1.85cm]{R4322.pdf}
\qquad
\includegraphics[width=2cm]{R4332.pdf}
}
\vspace{-4.5cm}
\caption{Egg box diagrams (drawn sideways) of the regular linear sandwich semigroups $P=\Reg(\M_{43}^{J}(\Z_2))$, where $\rank(J)=0,1,2,3$ (top to bottom).}
\label{fig:R}
\end{center}
\end{figure}
Comparing Figure \ref{fig:R} with Figure \ref{fig:M0...M3}, which pictures the full linear monoids $\M_r(\Z_2)$ for $0\leq r\leq 3$, an interesting pattern seems to emerge: namely, that $P=\Reg(\M_{43}^J(\Z_2))$ appears to be a kind of ``inflation'' of $\M_r$, where $r=\rank(J)$. One of the goals of this section is to explain this phenomenon, and we do so by further exploring the map $$\phi:P\to\M_r:X= [M,A,N]\mt \Xb=A$$ defined after Theorem \ref{thm:diamondsMmnJ}.
We also calculate $|P|$, $\rank(P)$, and the number and sizes of various Green's classes. As before, we assume that $J=J_{nmr}=\tmat {I_r}OOO\in\M_{nm}$. Since $\MmnJ$ is just a zero semigroup if $r=0$, we generally assume that $r\geq1$.
Now, Theorem \ref{green_thm} enables us to immediately describe Green's relations on $P=\RegMmnJ$. Since~$P$ is a regular subsemigroup of $\MmnJ$, the $\gR$, $\gL$, $\gH$ relations on $P$ are just the restrictions of the corresponding relations on $\MmnJ$ (see for example \cite{Hig,Howie}), and it is easy to check that this is also true for the $\gD=\gJ$ relation in this case. So if $X\in P$ and $\gK$ is one of $\gR$, $\gL$, $\gH$, $\gD$, we will continue to write $\gKJ$ for the $\gK$ relation on~$P$, and $K_X^J$ for the $\gKJ$-class of $X$ in $P$.
Parts (i--iv) of the next result also appear in \cite[Theorem~2.3]{Chinram2009}.
\ms
\begin{cor}\label{cor_green}
If $X\in P$, then
\bit
\itemit{i} $R_X^J =R_X\cap P = \set{Y\in P}{\Col(X)=\Col(Y)}$,
\itemit{ii} $L_X^J =L_X\cap P = \set{Y\in P}{\Row(X)=\Row(Y)}$,
\itemit{iii} $H_X^J = H_X\cap P=H_X = \set{Y\in P}{\Col(X)=\Col(Y)\text{\emph{ and }}\Row(X)=\Row(Y)}$,
\itemit{iv} $D_X^J =D_X\cap P = \set{Y\in P}{\rank(X)=\rank(Y)}$.
\eit
The $\gDJ$-classes of $P$ form a chain: $D_0^J<\cdots<D_r^J$, where $D_s^J=\set{X\in P}{\rank(X)=s}$ for each $0\leq s\leq r$.~\epfres
\end{cor}
Also, the regularity of $P$ means that $P$ inherits the stability property from $\MmnJ$.
The next result gives some combinatorial information about the size of $P$, and of various Green's classes in~$P$, in the case that $\F$ is finite. Recall that $\{e_{k1},\ldots,e_{kk}\}$ is the standard basis of $V_k=\F^k$ and that $W_{ks}=\Span\{e_{k1},\ldots,e_{ks}\}$ for each $0\leq s\leq k$.
\ms
\begin{prop}\label{prop:DXJ_combinatorics}
Suppose $|\F|=q<\infty$. Let $X\in P$ with $\rank(X)=s$. Then
\bit
\itemit{i} $|R_X^J|=q^{s(n-r)}q^{{s\choose2}}(q-1)^s\qfact s\tqbin rs$,
\itemit{ii} $|L_X^J|=q^{s(m-r)}q^{{s\choose2}}(q-1)^s\qfact s\tqbin rs$,
\itemit{iii} $|H_X^J|=|\G_s|= q^{{s\choose2}}(q-1)^s\qfact s$,
\itemit{iv}$D_X^J=D_s^J$ is the union of:
\begin{itemize}\begin{multicols}{3}
\itemit{a} $q^{s(m-r)}\tqbin rs$ $\gRJ$-classes,
\itemit{b} $q^{s(n-r)}\tqbin rs$ $\gLJ$-classes,
\itemit{c} $q^{s(m+n-2r)}\tqbin rs^2$ $\gHJ$-classes,
\end{multicols}\end{itemize}
\itemit{v} $|D_X^J|=|D_s^J|=q^{s(m+n-2r)}q^{{s\choose2}}(q-1)^s\qfact s\tqbin rs^2$.
\eit
Consequently,
$
\ds{|P|=|\RegMmnJ|=\sum_{s=0}^r q^{s(m+n-2r)}q^{{s\choose2}}(q-1)^s\qfact s\tqbin rs^2}.
$
\end{prop}
\pf We start with (i). Since $|R_X^J|=|R_Y^J|$ for all $Y\in D_X^J=D_s^J$, we may assume $X=J_{mns}$. Now, $\Col(X)=W_{ms}$. By Proposition~\ref{prop:P1P2} and Corollary \ref{cor_green}, we have
\begin{align*}
R_X^J &= \set{Y\in P}{Y\gRJ X} \\
&= \set{Y\in\Mmn}{\Col(X)=\Col(Y),\ \Col(Y)=\Col(YJ),\ \Row(Y)=\Row(JY)} \\
&= \set{Y\in\Mmn}{\Col(Y)=\Col(YJ)=W_{ms}},
\end{align*}
since if $\Col(Y)=W_{ms}$, then $Y$ is of the form $Y=\tmat AB{O_{m-s,r}}{O_{m-s,n-r}}$ for some $A\in\M_{sr}$ and $B\in\M_{s,n-r}$, in which case $JY=\tmat AB{O_{n-s,r}}{O_{n-s,n-r}}$ automatically has the same row space as $Y$.
Now consider some $Y\in R_X^J$. As noted above, we must have $Y=\tmat ABOO$ for some $A\in\M_{sr}$ and $B\in\M_{s,n-r}$. Since $YJ=\tmat AOOO$, the condition $\Col(YJ)=W_{ms}$ is equivalent to $\Col(A)=V_s$. In particular, there is no restriction on the entries of $B$, so $B$ may be chosen (arbitrarily, and independently of $A$) in $q^{s(n-r)}$ ways. Also, $\dim(\Row(A))=\dim(\Col(A))=s$. So $A$ may be specified by listing its rows (in order), which are $s$ linearly independent row vectors from $\F^r$. The number of possible choices for $A$ is therefore $(q^r-1)(q^r-q)\cdots(q^r-q^{s-1})=q^{{s\choose2}}(q-1)^s\qfact s\tqbin rs$. Multiplying these two values gives (i).
Part (ii) is dual to (i).
Part (iii) follows directly from Corollary \ref{cor_green}(iii) and Lemma \ref{lem:combinatorics_Mmn}(iii). Parts (a) and (b) of (iv) follow by dividing $|L_X^J|$ and $|R_X^J|$ by $|H_X^J|$, respectively. Part (c) follows from (a) and (b). Part (v) follows from (iii) and part (c) of (iv).
The formula for $|P|$ is obtained by adding the sizes of the $\gDJ$-classes.~\epf
\begin{comment}
The next result is very important in what follows. Its proof is routine, relying on Proposition \ref{prop:P1P2} and the rule $[M,A,N]\star[K,B,L] = [M,AB,L]$.
\ms
\begin{prop}\label{prop:MAN}
Consider the semigroup $U=\M_{m-r,r}\times\M_r\times\M_{r,n-r}$ under the operation $\cdot$ defined by
\[
(M,A,N)\cdot(K,B,L) = (M,AB,L)
\]
Define an equivalence $\sim$ on $U$ by
\[
(M,A,N) \sim (K,B,L) \ \ \iff \ \ \text{$A=B$, $MA=KB$ and $AN=BL$.}
\]
Then $\sim$ is a congruence on $U$, and the map
\[
\xi:U\to P=\RegMmnJ:(M,A,N)\mt[M,A,N]=\mat A{AN}{MA}{MAN}
\]
is an epimorphism with $\ker(\xi)={\sim}$. In particular, $P\cong U/{\sim}$. \epfres
\end{prop}
We continue to explore the structure of $P=\RegMmnJ$ by relating $P$ to various other matrix semigroups. If $X\in\Mkl$, we write $\row_i(X)$ and $\col_j(X)$ for the $i$th row and $j$th column of $X$ for each $1\leq i\leq k$ and $1\leq j\leq l$.
Now consider an element $X=\tmat ABCD\in\Mmn$. We have already observed that
\[
XJ=\mat AOCO \in\M_m \COMMA JX=\mat ABOO\in\M_n \COMMA JXJ = \mat AOOO\in\M_{nm}.
\]
It follows that
\begin{align*}
\MMNJ &= \set{XJ}{X\in\Mmn} = \set{Y\in\M_m}{\row_i(Y)=0\ (\forall r<i\leq m)} \\
\JMMN &= \set{JX}{X\in\Mmn} = \set{Y\in\M_n}{\col_i(Y)=0\ (\forall r<i\leq n)}.
\end{align*}
In particular, $\MMNJ$ and $\JMMN$ are subsemigroups (under the usual matrix multiplication) of $\M_m$ and $\M_n$ (respectively). These matrix semigroups (and also their isomorphic images under the maps $\M_m\to\Endm$ and $\M_n\to\Endn$, which consist of various linear transformations with restricted image or kernel) have been studied extensively {\red --- refs?}.
The regular elements of $\MMNJ$ and $\JMMN$
were classified in \cite{NK2007}. The next result, which gives a much simpler description of these regular elements, may be deduced from \cite[Theorem~3.4]{NK2007} (and vice versa), but we include a simple proof for convenience.
\ms
\begin{prop
\label{cor:RegMNJ}
The regular elements of the semigroups $\MMNJ$ and $\JMMN$ are given by
\begin{align*}
\Reg(\MMNJ) = PJ &= \set{X\in\MMNJ}{\rank(JX)=\rank(X)} \\
\Reg(\JMMN) = JP &= \set{X\in\JMMN}{\rank(XJ)=\rank(X)}.
\end{align*}
\end{prop}
\pf We just prove the second statement as the other is dual. Let $X=\tmat ABCD\in\Mmn$, and put $X'=\tmat ABOO\in\Mmn$. Then $JX=JX'=\tmat ABOO$ (where the zero matrices in the last expression have $n-r$ rows). Since $X'$ clearly belongs to $P_2$ (by Proposition \ref{prop:P1P2}), we have $\JMMN\sub JP_2$. Next, note that $J_{mnr}J=J_{mmr}$, and that $J_{mmr}Y=Y$ for all $Y\in\M_{mn}$ of the form $Y=\tmat ABOO$. Now suppose $X\in\Mmn$ is such that $JX\in\RegJMMN$. As above, we may assume that $X=\tmat ABOO$. So $(JX)=(JX)(JY)(JX)$ for some $Y\in\Mmn$. But then
\[
X = J_{mmr}X = J_{mnr}(JX) = J_{mnr}(JXJYJX)=XJYJX = X\star Y\star X,
\]
so that, in fact, $X\in \RegMmnJ=P$. This completes the proof that $\RegJMMN\sub JP$. The reverse inclusion is easily checked.
Now suppose $X=JY$ where $Y=\tmat A{AN}OO\in P$. Then $X=JY=\tmat A{AN}OO$ (with appropriately sized zero matrices), so $\rank(X)=\rank(JY)=\rank(Y)=\rank(JYJ)=\rank(XJ)$, where we have used Proposition~\ref{prop:P1P2}. Conversely, suppose $X\in\JMMN$ is such that $\rank(XJ)=\rank(X)$. As before, we may assume that $X=JY$ where $Y\in P_2$. Then $\rank(Y)=\rank(JY)=\rank(X)=\rank(XJ)=\rank(JYJ)$, so that $Y\in P$. This completes the proof. \epf
The next two propositions are the main structural results of this section, and show exactly how $P=\RegMmnJ$ is related to the (non-sandwich) matrix semigroups $PJ=\RegMMNJ$, $JP=\RegJMMN$, and $\M_r$.
{\red Introduce notation for $\set{Y\in\M_m}{\row_i(Y)=0\ (\forall r<i\leq m)}$? $\M_{m,(r)}$? Then the next result says $P$ embeds in $\M_{m,(r)}\times\M_{m,(r)}^*$.}
\ms
\begin{prop}\label{mono_prop}
There is an embedding
\[
\psi : P\to PJ\times JP : X\mt (XJ,JX).
\]
The image of $\psi$ is the set
\[
\im(\psi) = \bigset{(Y,Z)\in PJ\times JP}{JY=ZJ}.
\]
In particular, $P=\RegMmnJ$ is (isomorphic to) a subdirect product of $PJ=\RegMMNJ$ and $JP=\RegJMMN$.
\end{prop}
\pf Let $X=[M,A,N]$ and $Y=[K,B,L]$ be elements of $P$. Then
\[
\psi(X\star Y)=\psi(XJY)=(XJYJ,JXJY)=(XJ,JX)(YJ,JY)=\psi(X)\psi(Y),
\]
showing that $\psi$ is a homomorphism. Suppose now that $\psi(X)=\psi(Y)$. Then
\[
\left(\tmat{A}{O}{MA}{O},\tmat A{AN}OO \right) = \psi(X) = \psi(Y) = \left(\tmat{B}{O}{KB}{O},\tmat B{BL}OO \right).
\]
Comparing various coordinates, we deduce $A=B$, $MA=KB$ and $AN=BL$, giving $X=[M,A,N]=[K,B,L]=Y$, completing the proof that $\psi$ is injective.
To prove the statement about $\im(\psi)$, let $X\in P$ and put $Y=XJ$ and $Z=JX$. Then $JY=JXJ=ZJ$. Conversely, suppose $(Y,Z)\in PJ\times JP$ satisfies $JY=ZJ$. Say $Y=UJ$ and $Z=JV$, where $U=[M,A,N]$ and $V=[K,B,L]$ belong to $P$. Then $JY=JUJ=\tmat AOOO$ and $ZJ=JVJ=\tmat BOOO$, giving $A=B$. But then $(Y,Z)=\psi(X)$, where $X=[M,A,L]\in P$. This completes the proof that $\im(\psi)$ has the desired description, and also shows that $\im(\psi)$ is a subdirect product of $PJ$ and $JP$. \epf
The homomorphism $\psi$ from the previous result is built up out of the two epimorphisms
\[
\psi_1:P\to PJ:X\mt XJ \AND \psi_2:P\to JP:X\mt JX.
\]
The proof of the next result is routine, making use of the fact that $JXJ=\tmat AOOO$ where $X=[M,A,N]\in P$.
\ms
\begin{prop}\label{epi_prop}
The maps
\[
\phi_1:PJ\to\M_r:\tmat AO{MA}O \mt A \AND \phi_2:JP\to\M_r:\tmat A{AN}OO\mt A
\]
are epimorphisms, and the following diagram commutes:
\[
\begin{tikzcd}
P \arrow{r}{\psi_1} \arrow[swap]{d}{\psi_2} & PJ \arrow{d}{\phi_1} & \\
JP \arrow[swap]{r}{\phi_2} & \M_r
\end{tikzcd}
\]
Further, the induced map $\phi:P\to\M_r:[M,A,N]\to A$ is an epimorphism. In particular, $P=\RegMmnJ$ is (isomorphic to) \emph{pullback product} of $PJ=\RegMNJ$ and $JP=\RegJMN$ with respect to $\M_r$. Namely,
\[ \epfreseq
P\cong\im(\psi)=\set{(Y,Z)\in PJ\times JP}{\phi_1(Y)=\phi_2(Z)}.
\]
\end{prop}
\end{comment}
Recall that, for $X=[M,A,N]\in P$, we write $\Xb=\phi(X)=A\in\M_r$. We extend this notation to subsets of~$P$, so if $\Om\sub P$, we write $\Omb=\set{\Xb}{X\in\Om}$.
We now show how the epimorphism $\phi:P\to\M_r$ may be used to relate Green's relations on the semigroups $P$ and $\M_r$.
If $X,Y\in P$ and $\gK$ is one of $\gR$, $\gL$, $\gH$, $\gD$, we say $X\gKh Y$ if $\Xb\gK\Yb$ (in~$\M_r$). Denote by $\Kh_X=\phi^{-1}(K_{\Xb})=\set{Y\in P}{X\gKh Y}$ the $\gKh$-class of $X$ in~$P$. We first need a technical result.
\ms
\begin{lemma}\label{lem:|RhX|}
Let $X,Y\in P$. If $X\gDh Y$, then $|\Rh_X|=|\Rh_Y|$ and $|\Lh_X|=|\Lh_Y|$.
\end{lemma}
\pf By duality, it suffices to prove the statement about $\gRh$-classes. Now, $X\gDh Y$ means that $X\gLh W\gRh Y$ for some $W\in P$. Since $\Rh_Y=\Rh_W$, we may assume without loss of generality that $X\gLh Y$.
Write $X=[M,A,N]$ and $Y=[K,B,L]$. By definition $X\gLh Y$, means that $A\L B$ in $\M_r$, so $A=UB$ for some $U\in\G_r$ by Lemma \ref{lem:greenMmn}. Now let $Z=[M',A',N']\in\Rh_X$, and define $\al(Z)=[M'U,U^{-1}A',N']$. It is easy to check that for any other representation $Z=[M'',A'',N'']$, we have $[M''U,U^{-1}A'',N'']=[M'U,U^{-1}A',N']$, so that $\al(Z)$ is well-defined. Also,
\begin{align*}
Z\gRh X\ \implies\ A'\R A
\ \implies\ U^{-1}A'\R U^{-1}A =B
\ \implies\ \al(Z) \gRh Y.
\end{align*}
Thus $\al$ is a map $\al:\Rh_X\to\Rh_Y$. It is easy to check that $\Rh_Y\to\Rh_X:[M',A',N']\mt[M'U^{-1},UA',N']$ is the inverse mapping of $\al$. We conclude that $|\Rh_X|=|\Rh_Y|$. \epf
For the proof of the next result, we note that stability of $\M$ implies that $A^2\D A \iff A^2\H A$ for all $A\in\M_r$. We also use the fact that an $\H$-class $H$ of a semigroup $S$ is a group if and only if $x^2\in H$ for some (and hence for all) $x\in H$ \cite[Theorem 2.2.5]{Howie}.
Recall that a $k\times l$ \emph{rectangular band} is a semigroup of the form $S\times T$ with product $(s_1,t_1)(s_2,t_2)=(s_1,t_2)$, where $|S|=k$ and $|T|=l$. A $k\times l$ \emph{rectangular group} with respect to a group $G$ is a direct product of a $k\times l$ rectangular band with $G$.
For the proof of the next result (and elsewhere), it will be convenient to define a number of equivalence relations. For $A\in\M_r$, we define equivalences $\sim_A$ and $\approx_A$ on $\M_{m-r,r}$ and $\M_{r,n-r}$ (respectively) by
\[
M_1\sim_AM_2 \ \iff \ M_1A=M_2A \AND N_1\approx_AN_2 \ \iff \ AN_1=AN_2.
\]
\ms
\begin{thm}\label{inflation_thm}
Suppose $|\F|=q<\infty$. Let $X\in P=\RegMmnJ$ and put $s=\rank(X)$.
\bit
\itemit{i} $\Rh_X$ is the union of $q^{s(m-r)}$ $\gRJ$-classes of $P$.
\itemit{ii} $\Lh_X$ is the union of $q^{s(n-r)}$ $\gLJ$-classes of $P$.
\itemit{iii} $\Hh_X$ is the union of $q^{s(m+n-2r)}$ $\gHJ$-classes of $P$, each of which has size $|\G_s|=q^{{s\choose2}} (q-1)^s\qfact s$. The map $\phi:P\to\M_r$ is injective when restricted to any $\gHJ$-class of $P$.
\itemit{iv} If $H_{\Xb}$ is a non-group $\gH$-class of $\M_r$, then each $\gHJ$-class of $P$ contained in $\Hb_X$ is a non-group.
\itemit{v} If $H_{\Xb}$ is a group $\gH$-class of $\M_r$, then each $\gHJ$-class of $P$ contained in $\Hb_X$ is a group isomorphic to $\G_s$; further, $\Hb_X$ is a $q^{s(m-r)}\times q^{s(n-r)}$ rectangular group with respect to $\G_s$.
\itemit{vi} $\gDh=\gDJ$ and $\Dh_X=D_X^J=D_s^J=\set{Y\in P}{\rank(Y)=s}$ is the union of:
\begin{itemize}
\itemit{a} $\tqbin rs$\ $\gRh$-classes (and the same number of $\gLh$-classes) of $P$,
\itemit{b} $q^{s(m-r)}\tqbin rs$\ $\gRJ$-classes of $P$,
\itemit{b} $q^{s(n-r)}\tqbin rs$\ $\gLJ$-classes of $P$,
\itemit{d} $\tqbin rs^2$\ $\gHh$-classes of $P$,
\itemit{e} $q^{s(m+n-2r)}\tqbin rs^2$\ $\gHJ$-classes of $P$.
\eitres
\eitres
\end{thm}
\pf First observe that if $\rho:S\to T$ is an epimorphism of semigroups, and if $K$ is a $\gK$-class of $T$ where $\gK$ is one of $\gR$, $\gL$, $\gH$, then $\rho^{-1}(K)$ is a union of $\gK$-classes of $S$. Throughout the proof, we write
\[
X=[M,A,N]=\mat A{AN}{MA}{MAN},
\]
so $A\in\M_r$ satisfies $\rank(A)=\rank(JXJ)=\rank(X)=s$. We note that $\gDh=\gDJ$ immediately follows.
\bit
\item[(i)] By the first observation, it suffices to count the number of $\gRJ$-classes contained in $\Rh_X$.
Since $|\Rh_X|=|\Rh_Y|$ for all $Y\in D_s^J$ by Lemma \ref{lem:|RhX|}, it follows that each $\gRh$-class of $D_s^J$ contains the same number of $\gRJ$-classes. By Lemma \ref{lem:combinatorics_Mmn}, $D_s(\M_r)$ is the union of $\tqbin rs$ $\R$-classes (and the same number of $\L$-classes), so it follows that $D_s^J$ is the union of $\tqbin rs$ $\gRh$-classes (and the same number of $\gLh$-classes). By Proposition \ref{prop:DXJ_combinatorics}, $D_s^J$ is the union of $q^{s(m-r)}\tqbin rs$ $\gRJ$-classes. Dividing these, it follows that each $\gRh$-class of $D_s^J$ is the union of $q^{s(m-r)}$ $\gRJ$-classes.
\item[(ii)] This is dual to (i).
\item[(iii)] The statement concerning the number of $\gHJ$-classes contained in $\Hb_X$ follows immediately from (i) and~(ii), and the size of these $\gHJ$-classes was given in Proposition \ref{prop:DXJ_combinatorics}. Next, for any $B\in\M_r$ with $B\H A$, it is easy to check that $[M,B,N] \gHJ X$. So the set $\Om=\set{[M,B,N]}{B\in H_A}$ is contained in~$\gHJ_X$. Since $|\Om|=|H_A|=|\G_s|=|H_X^J|$, we see that $H_X^J=\Om$. For any $Z=[M,B,N]\in\Om=H_X^J$, we have $\phi(Z)=B$, so it follows that $\phi|_{H_X^J}$ is injective.
\item[(iv)] Suppose $H_A=H_{\Xb}$ is a non-group $\gH$-class of $\M_r$, and let $Y=[K,B,L]\in\Hh_X$ be arbitrary. Since $Y\gHh X$, it follows that $B=\Yb\gH \Xb=A$. Since $H_B=H_A$ is not a group, we have $B^2\not\in H_B$, whence $B^2\not\in D_B$ and $\rank(B^2)<\rank(B)=\rank(A)=s$. But then $Y^2=[K,B^2,L]\not\in D_s^J=D_Y^J$, so that $Y^2\not\in H_Y^J$, and we conclude that $H_Y^J$ is not a group.
\item[(v)] Suppose $H_{\Xb}$ is a group. Then $\Yb^2\in H_{\Xb}$ for any $Y\in \Hh_X$, so $\rank(Y\star Y)=\rank(\Yb^2)=\rank(\Yb)=\rank(Y)$, giving $Y\star Y\gDJ Y$, so that $Y\star Y\in H_Y^J$ and $H_Y^J$ is a group. By (iii), the restriction of $\phi$ to $H_Y^J$ yields an isomorphism onto $H_{\Yb}\cong\G_s$.
Let $E\in\M_r$ be the identity element of the group $H_A$.
Let $\sM_E\sub\M_{m-r,r}$ (resp., $\sN_E\sub\M_{r,n-r}$) be a cross-section of the $\sim_E$-classes (resp., $\approx_E$-classes) in $\M_{m-r,r}$ (resp., $\M_{r,n-r}$).
It is easy to check that every $Y\in\Hh_X$ may be uniquely represented as $Y=[K,B,L]$ for some $K\in\sM_E$, $B\in H_A$ and $L\in\sN_E$.
It follows that the map
\[
\sM_E\times H_A\times\sN_E\to\Hh_A:(K,B,L)\mt[K,B,L]
\]
is a well-defined isomorphism, where the (rectangular group) product on $\sM_E\times H_A\times\sN_E$ is defined by $(K_1,B_1,L_1)\cdot(K_2,B_2,L_2)=(K_1,B_1B_2,L_2)$. We have already observed that $H_A\cong\G_s$, and the dimensions of the rectangular band $\sM_E\times\sN_E$ follow from parts (i--iii) together with the observation that $\set{[K,B,L]}{B\in H_A}$ is an $\gHJ$-class contained in $\Hh_X$ for each $K\in\sM_E$ and $L\in\sN_E$.
\item[(vi)] We have already noted that $\gDh=\gDJ$. We proved (a) while proving (i), above. Parts (b), (c) and (e) were proved in Proposition \ref{prop:DXJ_combinatorics}. Part (d) follows from (a).
\epf
\eit
The previous result explains the ``inflation'' phenomenon discussed at the beginning of this section; see also Figure \ref{fig:R}. As an immediate corollary of Theorem \ref{inflation_thm}, we may now completely classify the isomorphism classes of finite linear sandwich semigroups.
\ms
\begin{thm}\label{thm:classification}
Let $\F_1$ and $\F_2$ be two finite fields with $|\F_1|=q_1$ and $|\F_2|=q_2$, let $m,n,k,l\geq1$, and let $A\in D_r(\M_{nm})$ and $B\in D_s(\M_{lk})$.
The following are equivalent:
\bit
\itemit{i} $\MMN\cong\MKL$,
\itemit{ii} one of the following holds:
\begin{itemize}
\itemit{a} $r=s=0$ and $q_1^{mn}=q_2^{kl}$, or
\itemit{b} $r=s\geq1$, $(m,n)=(k,l)$, and $q_1=q_2$.
\end{itemize}
\eit
Further, if $r\geq1$, then $\MMN\cong\MKL$ if and only if $\RegMMN\cong\RegMKL$.
\end{thm}
\pf Again, if $r\not=s$, then counting the regular $\D^A$- and $\D^B$-classes shows that $\MMN\not\cong\MKL$. For the remainder of the proof, we assume $r=s$.
Suppose first that $r=s=0$. Then $\MMN$ and $\MKL$ are both zero semigroups and so are isomorphic if and only if their sizes, $q_1^{mn}$ and $q_2^{kl}$, are equal.
For the remainder of the proof, we assume $r=s\geq1$, and write $D_t^A$ and $D_t^B$ for the relevant regular $\D^A$- and $\D^B$-classes in $\MMN$ and $\MKL$ for each $0\leq t\leq r=s$.
By Theorem \ref{inflation_thm}(v), any group $\H^A$-class contained in $D_1^A$ is isomorphic to $\G_1(\F_1)\cong\F_1^\times$, the multiplicative group of $\F_1$. Since $|\F_1^\times|=q_1-1$ and $|\F_2^\times|=q_2-1$, it follows that if $q_1\not=q_2$, then $\RegMMN\not\cong\RegMKL$ and, hence, $\MMN\not\cong\MKL$.
Now suppose $q_1=q_2$ (so $\F_1\cong\F_2$), and write $q=q_1$. By Theorem~\ref{inflation_thm}(vi), $D_1^A$ (resp., $D_1^B$) contains $q^{m-r}\tqbin r1$ $\R^A$-classes (resp., $q^{k-r}\tqbin r1$ $\R^B$-classes).
It follows that if $m\not=k$ (or, dually, if $n\not=l$), then $\RegMMN\not\cong\RegMKL$ and, hence, $\MMN\not\cong\MKL$.
Conversely, if (b) holds, then $\MMN\cong\MKL$ by Lemma \ref{lem:MmnAMmnB}(ii).
For the final statement, first note that $\MMN\cong\MKL$ clearly implies $\RegMMN\cong\RegMKL$. In the previous paragraph, we showed that the negation of (b) implies $\RegMMN\not\cong\RegMKL$. This completes the proof. \epf
\begin{rem}
Of course, if $\rank(A)=\rank(B)=0$, then $\RegMMN=\{O_{mn}\}\cong\RegMKL=\{O_{kl}\}$, regardless of $m,n,k,l,q_1,q_2$. So the final clause of Theorem \ref{thm:classification} does not hold for $r=0$.
\end{rem}
\ms
\begin{rem}\label{rem:infinite_classification}
The infinite case is not as straight-forward, since $|\M_{mn}(\F)|=|\F|$ for all $m,n\geq1$, and since it is possible for two non-isomorphic fields $\F_1,\F_2$ to have isomorphic multiplicative groups $\F_1^\times,\F_2^\times$ (for example, $\mathbb Q$ and $\mathbb Z_3(x)$ both have multiplicative groups isomorphic to $\mathbb Z_2\oplus F$, where $F$ is a free abelian group of countably infinite rank). So we have the following isomorphisms:
\bit
\item[(i)] $\MMN\cong\MKL$ if $m,n,k,l\geq1$, $|\F_1|=|\F_2|$, and $\rank(A)=\rank(B)=0$ --- indeed, both sandwich semigroups are zero semigroups of size $|\F_1|=|\F_2|$;
\item[(ii)] $\MMN\cong\M_{mn}^B(\F_2)$ if $\F_1^\times\cong\F_2^\times$ and $\rank(A)=\rank(B)=1$ --- indeed,
when $J=J_{nm1}=\tmat{I_1}OOO$, sandwich products $X\star_J Y$ involve only field multiplication and no addition:
\[
\left[
\begin{matrix}
a_{11} & a_{12} & \cdots & a_{1n} \\
a_{21} & a_{22} & \cdots & a_{2n} \\
\vdots & \vdots &\ddots &\vdots \\
a_{m1} & a_{m2} & \cdots & a_{mn} \\
\end{matrix}
\right]
\star_J
\left[
\begin{matrix}
b_{11} & b_{12} & \cdots & b_{1n} \\
b_{21} & b_{22} & \cdots & b_{2n} \\
\vdots & \vdots &\ddots &\vdots \\
b_{m1} & b_{m2} & \cdots & b_{mn} \\
\end{matrix}
\right]
=
\left[
\begin{matrix}
a_{11}b_{11} & a_{11}b_{12} & \cdots & a_{11}b_{1n} \\
a_{21}b_{11} & a_{21}b_{12} & \cdots & a_{21}b_{1n} \\
\vdots & \vdots &\ddots &\vdots \\
a_{m1}b_{11} & a_{m1}b_{12} & \cdots & a_{m1}b_{1n} \\
\end{matrix}
\right].
\]
\eitres
We leave it as an open problem to completely classify the isomorphism classes of linear sandwich semigroups over infinite fields. But we make two simple observations:
\bit
\item[(iii)] as in the proof of Theorem \ref{thm:classification}, if $\MMN\cong\M_{kl}^B(\F_2)$, then we must have $\rank(A)=\rank(B)$;
\item[(iv)] if $\MMN\cong\M_{kl}^B(\F_2)$ with $\rank(A)=\rank(B)=r\geq2$, we must have $\F_1\cong\F_2$ (since the maximal subgroups of $\MMN$ are isomorphic to $\G_s(\F_1)$ for $0\leq s\leq r$, and since $\G_s(\F_1)\cong\G_s(\F_2)$ implies $\F_1\cong\F_2$ for $s\geq2$ \cite{Dieudonne1971}).
\eit
\end{rem}
In what follows, the top $\gDJ$-class of $P=\RegMmnJ$ plays a special role. We write $D$ for this $\gDJ$-class, so
\[
D=D_r^J = \phi^{-1}(\G_r) = \set{X\in P}{\rank(X)=r}.
\]
As a special case of Theorem \ref{inflation_thm}(v), $D$ is a $q^{r(m-r)}\times q^{r(n-r)}$ rectangular group with respect to $\G_r$. Since~$D$ is the pre-image of $\G_r$ under the map $\phi:P\to\M_r$, we may think of $D$ as a kind of ``inflation'' of $\G_r$, the group of units of $\M_r$. In fact, more can be said along these lines. Recall again that the \emph{variant} of a semigroup~$S$ with respect to an element $a\in S$ is the semigroup $S^a$ with underlying set $S$ and operation $\star_a$ defined by $x\star_ay=xay$ for all $x,y\in S$.
Recall also that an element $a\in S$ of a (necessarily regular) semigroup $S$ is \emph{regularity preserving} if the variant $S^a$ is regular. The set $\RP(S)$ of all regularity preserving elements of $S$ was studied in \cite{KL2001,Hickey1983}; we will not go into the details here, but it was explained in \cite{KL2001} that $\RP(S)$ is a useful alternative to the group of units in the case that $S$ is not a monoid (as with $P$ when $r=m=n$ does not hold). Because of this, it is significant that $D$ is equal to $\RP(P)$, the set of all regularity preserving elements of $P=\RegMmnJ$, as we will soon see. We now state a result from \cite{KL2001} concerning regularity preserving elements. Recall that an element $u$ of a semigroup $S$ is a \emph{mididentity} if $xuy=xy$ for all $x,y\in S$ \cite{Yamada1955}; of course for such an element, $\star_u$ is just the original semigroup operation.
\ms
\begin{prop}[Khan and Lawson \cite{KL2001}]\label{RPS_prop}
Let $S$ be a regular semigroup.
\bit
\itemit{i} An element $a\in S$ is regularity preserving if and only if $a\gH e$ for some regularity preserving idempotent $e\in E(S)$. (In particular, $\RP(S)$ is a union of groups.)
\itemit{ii} An idempotent $e\in E(S)$ is regularity preserving if and only if $fe \gR f \gL ef$ for all idempotents $f\in E(S)$.
\itemit{iii} Any mididentity is regularity preserving. \epfres
\eitres
\end{prop}
In order to avoid confusion when discussing idempotents, if $\Om\sub\Mmn$, we will write
\[
\EJ(\Om)=\set{X\in\Om}{X=X\star X}
\]
for the set of idempotents from $\Om$ with respect to the $\star$ operation on $\MmnJ$. If $\Si\sub \M_k$ for some $k$, we will continue to write $E(\Si)=\set{A\in\Si}{A=A^2}$ for the set of idempotents from $\Si$ with respect to the usual matrix multiplication.
\ms\ms
\begin{lemma}\label{lemma_aea}
\bit
\itemit{i} $\EJ(\MmnJ)=\EJ(P)=\set{[M,A,N]}{A\in E(\M_r),\ M\in\M_{m-r,r},\ N\in\M_{r,n-r}}$.
\itemit{ii} $\EJ(D)=\set{[M,I_r,N]}{M\in\M_{m-r,r},\ N\in\M_{r,n-r}}$ is a $q^{r(m-r)}\times q^{r(n-r)}$ rectangular band.
\itemit{iii} Each element from $\EJ(D)$ is a mididentity for both $\MmnJ$ and $P$.
\itemit{iv} $D=\RP(P)$ is the set of all regularity-preserving elements of $P$.
\eitres
\end{lemma}
\pf Note that all idempotents are regular. If $X=[M,A,N]\in P$, then $X\star X=[M,A^2,N]$, so $X=X\star X$ if and only if $A=A^2$, giving~(i). Part (ii) follows from (i),
since $I_r$ is the only idempotent from the group~$\G_r=D_r(\M_r)$. Using (ii), it is easy to check by direct computation that $X\star Y\star Z=X\star Z$ for all $X,Z\in \Mmn$ and $Y\in\EJ(D)$, giving (iii). Finally, to prove (iv), note that by Proposition \ref{RPS_prop}(i), it suffices to show that $\EJ(\RP(P))=\EJ(D)$. By (iii) and Proposition \ref{RPS_prop}(iii), we have $\EJ(D)\sub\EJ(\RP(P))$. Conversely, suppose $X\in\EJ(\RP(P))$. Let $Y\in\EJ(D)$. By Proposition \ref{RPS_prop}(ii), and the fact that $\gLJ\sub\gDJ$, $X\star Y\gDJ Y$. It follows that $r=\rank(Y)=\rank(XJY)\leq\rank(X)\leq r$, giving $\rank(X)=r$, and $X\in\EJ(D)$.
This shows that $\EJ(\RP(P))\sub \EJ(D)$, and completes the proof.~\epf
We may now calculate the rank of $P=\RegMmnJ$ in the case of finite $\F$. For the following proof, recall from \cite{HHR} that the \emph{relative rank} $\rank(S:U)$ of a semigroup $S$ with respect to a subset $U\sub S$ is defined to be the minimum cardinality of a subset $V\sub S$ such that $S=\la U\cup V\ra$.
\ms
\begin{thm}\label{thm:rankP}
Suppose $|\F|=q<\infty$. If $1\leq r\leq\min(m,n)$ and we do not have $r=m=n$, then $$\rank(P)=\rank(\RegMmnJ)=q^{r(L-r)}+1,$$ where $L=\max(m,n)$.
\end{thm}
\pf Since $D$ is a subsemigroup of $P$ and $P\sm D$ is an ideal, it quickly follows that $\rank(P)=\rank(D)+\rank(P:D)$. It is well-known \cite{Ruskuc1994} that a rectangular group $R=(S\times T)\times G$ satisfies $\rank(R)=\max\big\{|S|,|T|,\rank(G)\big\}$. Since $D$ is a $q^{r(m-r)}\times q^{r(n-r)}$ rectangular group with respect to $\G_r$, and since $\rank(\G_r)\leq2$ by Theorem~\ref{thm:waterhouse}, it immediately follows that $\rank(D)=q^{r(L-r)}$. Since $\la D\raJ=D\not=P$ (as $r\geq1$), we have $\rank(P:D)\geq1$, so the proof will be complete if we can show that $P=\la D\cup\{X\}\raJ$ for some $X\in P$.
With this in mind, let $X\in D_{r-1}^J$ be arbitrary. Note that $\Db=\set{\Yb}{Y\in D}=\G_r$, and $\Xb\in D_{r-1}(\M_r)$. It follows from Theorem~\ref{thm:waterhouse} that $\M_r=\la\Db\cup\{\Xb\}\ra$. Now let $Y=[M,A,N]\in P$ be arbitrary. Choose $Z_1,\ldots,Z_k\in D\cup\{X\}$ such that $A=\Zb_1\cdots\Zb_k$. Then $Y=[M,I_r,N]\star Z_1\star\cdots\star Z_k\star[M,I_r,N]$, with $[M,I_r,N]\in D$. \epf
\begin{rem}\label{rem:r=m3}
If $r=0$, then $P=\{O\}$, while if $r=m=n$, then $P=\M_n$. So $\rank(P)$ is trivial in the former case, and well-known in the latter (see Theorem~\ref{thm:waterhouse}).
As in Remark \ref{rem:r=m2}, we deduce that
$
\rank(\Reg(\Ckl)) = \rank(\Reg(\Rkl)) = q^{l(k-l)} + 1
$
for $|\F|=q<\infty$.
\end{rem}
\section{The idempotent generated subsemigroup}\label{sect:EMmnJ}
In this section, we investigate the idempotent generated subsemigroup $\la\Ea(\MmnJ)\raa$ of $\MmnJ$; we write $\EmnJ$ for this idempotent generated subsemigroup. Our main results include a proof that $\EmnJ=(P\sm D)\cup \Ea(D)$ and a calculation of $\rank(\EmnJ)$ and $\idrank(\EmnJ)$; in particular, we show that these two values are equal.
Since the solution to every problem we consider is trivial when $r=0$, and well-known when $r=m=n$, we will continue to assume that $r\geq1$ and that $r=m=n$ does not hold. To simplify notation, we will write $E=\Ea(\MmnJ)=\Ea(P)$, so $\EmnJ=\la E\raa$.
We begin by calculating $|E|$ in the case of finite $\F$, for which we need the following formulae for $|E(D_s(\M_r))|$. Although the next result might already be known, we are unaware of a reference and include a simple proof for convenience.
\ms
\begin{lemma}\label{lem:EDmMr}
Suppose $|\F|=q<\infty$. If $0\leq s\leq r$, then $|E(D_s(\M_r))|=q^{s(r-s)}\tqbin rs $. Consequently,
\[
|E(\M_r)| = \sum_{s=0}^r q^{s(r-s)}\qbin rs.
\]
\end{lemma}
\pf To specify an idempotent endomorphism $\al\in\Endr$ of rank $s$, we first choose $W=\im(\al)$, which is a subspace of dimension $s$ and may be chosen in $\tqbin rs$ ways, and we note that $\al$ must map $W$ identically. If $\{v_1,\ldots,v_r\}$ is an arbitrary basis for $V_r$, such that $\{v_1,\ldots,v_s\}$ is a basis of $W$, then $\al$ may map each of $v_{s+1},\ldots,v_r$ arbitrarily into $W$, and there are $(q^s)^{r-s}$ ways to choose these images. \epf
\begin{prop}\label{prop:enumeration_E}
Suppose $|\F|=q<\infty$. If $0\leq s\leq r$, then $|\EJ(D_s^J)|=q^{s(m+n-r-s)}\tqbin rs$. Consequently,
\[
|\EJ(\MmnJ)| = \sum_{s=0}^r q^{s(m+n-r-s)}\qbin rs.
\]
\end{prop}
\pf
Parts (iv) and (v) of Theorem \ref{inflation_thm} say that an $\gHJ$-class $H_X^J\sub D_s^J$ is a group (so contains an idempotent) if and only if $H_{\Xb}$ is a group $\H$-class of $\M_r$, and that there are $q^{s(m-r)}\times q^{s(n-r)}$ idempotents of $D_s^J$ corresponding to each rank $s$ idempotent of $\M_r$, of which there are $q^{s(r-s)}\tqbin rs$ by Lemma \ref{lem:EDmMr}. The result quickly follows. \epf
We now describe the idempotent generated subsemigroup of $\MmnJ$.
\ms
\begin{thm}\label{thm:EmnJ}
We have $\EmnJ=\la\EJ(\MmnJ)\raJ=(P\sm D)\cup\EJ(D)$.
\end{thm}
\pf Suppose $X_1,\ldots,X_k\in E=\EJ(\MmnJ)$, and write $X_i=[M_i,A_i,N_i]$ for each $i$. So $A_i\in E(\M_r)$ for each $i$. Then $X_1\star\cdots\star X_k = [M_1,A_1\cdots A_k,N_k]$. If any of $A_1,\ldots,A_k$ belongs to $\MrGr$, then so too does $A_1\cdots A_k$, so that $X_1\star\cdots\star X_k\in P\sm D$. If all of $A_1,\ldots,A_k$ belong to $\G_r$, then $A_1=\cdots=A_k=I_r$, so $X_1\star\cdots\star X_k=[M_1,I_r,N_k]\in\EJ(D)$. This shows that $\EmnJ\sub(P\sm D)\cup\EJ(D)$. Conversely, it suffices to show that $P\sm D\sub\EmnJ$, so suppose $X\in P\sm D$, and write $X=[M,A,N]$. Since $X\not\in D$, we must have $\rank(A)=\rank(X)<r$. But then $A\in\MrGr$, so that $A=B_1\cdots B_l$ for some $B_1,\ldots,B_l\in E(\M_r)$ by Theorem \ref{thm_MnGn}. It follows that $X=[M,B_1,N]\star\cdots\star[M,B_l,N]$, with all $[M,B_i,N]\in E$. \epf
\begin{rem}
Recall (see Theorem \ref{thm_MnGn}) that ${\E_n=\la E(\M_n)\ra = (\MnGn)\cup\{I_n\}}$. Theorem \ref{thm:EmnJ} is a pleasing analogue of that result, since $\{I_n\}=E(\G_n)$, where $\G_n$ is the top $\gD$-class of $\M_n$. Also, $\G_n=G(\M_n)=\RP(\M_n)$ and, while $P$ has no group of units as it is not a monoid, it is still the case that $D=\RP(P)$.
\end{rem}
Now that we have described the elements of the semigroup $\EmnJ$, the next natural task is to calculate its rank and idempotent rank.
\ms
\begin{thm}\label{thm:rankEmnJ}
Suppose $|\F|=q<\infty$. Then $$\rank(\EmnJ)=\idrank(\EmnJ)=q^{r(L-r)}+(q^r-1)/(q-1),$$ where $L=\max(m,n)$.
\end{thm}
\pf As in the proof of Theorem \ref{thm:rankP}, we have $\rank(\EmnJ)=\rank(\Ea(D))+\rank(\EmnJ:\Ea(D))$. Since $\EJ(D)$ is a $q^{r(m-r)}\times q^{r(n-r)}$ rectangular band (see Lemma \ref{lemma_aea}(ii)), we again deduce from \cite{Ruskuc1994} that $\rank(\Ea(D))=\idrank(\Ea(D)) = q^{r(L-r)}$. So it remains to show that:
\bit
\item[(i)] there exists a set $\Om\sub E$ of size $(q^r-1)/(q-1)$ such that $\EmnJ=\la\Ea(D)\cup \Om\raa$, and
\item[(ii)] if $\Si\sub P$ satisfies $\EmnJ=\la\Ea(D)\cup \Si\raa$, then $|\Si|\geq(q^r-1)/(q-1)$.
\eit
By Theorem \ref{thm_MnGn}, we may choose some set $\Ga\sub E(\M_r)$ with $\la\Ga\ra=\MrGr$ and $|\Ga|=(q^r-1)/(q-1)$. For each $A\in\Ga$, choose any $M_A\in\M_{m-r,r}$ and $N_A\in\M_{r,n-r}$, and put $\Om=\set{[M_A,A,N_A]}{A\in\Ga}$. Since $\EmnJ=(P\sm D)\cup\EJ(D)$, the proof of (i) will be complete if we can show that $P\sm D\sub\la\Ea(D)\cup \Om\raa$. So let $X=[K,B,L]\in P\sm D$, and write $B=A_1\cdots A_k$ where $A_1,\ldots,A_k\in\Ga$. Then
\[
X=[K,I_r,L]\star[M_{A_1},A_1,N_{A_1}]\star\cdots\star[M_{A_k},A_k,N_{A_k}]\star[K,I_r,L]\in\la\Ea(D)\cup \Om\raa,
\]
as required. Next, suppose $\EmnJ=\la\Ea(D)\cup \Si\raa$, where $\Si\sub\EmnJ\sm \Ea(D)=P\sm D$. We will show that $\Sib$ generates $\MrGr$. Indeed, let $A\in\MrGr$ be arbitrary, and choose any $X\in P$ such that $\Xb=A$. Since $\rank(X)=\rank(A)<r$, it follows that $X\in P\sm D\sub\EmnJ$. Consider an expression $X=Y_1\star\cdots\star Y_k$, where $Y_1,\ldots,Y_k\in \Ea(D)\cup \Si$. Now, $A=\Xb=\Yb_1\cdots\Yb_k$. If any of the $Y_i$ belongs to $\Ea(D)$, then $\Yb_i=I_r$, so the factor $\Yb_i$ is not needed in the product $A=\Yb_1\cdots\Yb_k$. After cancelling all such factors, we see that $A$ is a product of elements from $\Sib$. Since $A\in\MrGr$ was arbitrary, we conclude that $\MrGr=\la\Sib\ra$. In particular, $|\Si|\geq|\Sib|\geq\rank(\MrGr)=(q^r-1)/(q-1)$, giving (ii). \epf
\begin{rem}\label{rem:r=m4}
As in Remarks \ref{rem:r=m2} and \ref{rem:r=m3}, we deduce from the results of this section that for $|\F|=q<\infty$,
\bit
\item $\Ckl$ (and $\Rkl$) has $\sum_{s=0}^l q^{s(k-s)}\tqbin ls$ idempotents,
\item the semigroup generated by $E(\Ckl)$ (and the semigroup generated by $E(\Rkl)$) has rank and idempotent rank equal to $q^{l(k-l)} + (q^l-1)/(q-1)$.
\eit
\end{rem}
\section{Ideals}\label{sect:ideals}
In this final section, we consider the ideals of $P=\RegMmnJ$. In particular, we show that each of the proper ideals is idempotent generated, and we calculate the rank and idempotent rank, showing that these are equal. Although the next result is trivial if $r=0$ and well-known if $r=m=n$ (see Theorem \ref{thm_ideals_Mn}), the statement is valid for those parameters.
\ms
\begin{thm}\label{thm:ideals}
The ideals of $P=\RegMmnJ$ are precisely the sets
\[
I_s^J=D_0^J\cup\cdots\cup D_s^J=\set{X\in P}{\rank(X)\leq s} \qquad\text{for $0\leq s\leq r$,}
\]
and they form a chain: $I_0^J\sub\cdots\sub I_r^J$. If $0\leq s<r$, then $I_s^J = \la \Ea(D_s^J) \raa$ is generated by the idempotents in its top $\gDJ$-class, and if $|\F|=q<\infty$, then
\[
\rank(I_s^J)=\idrank(I_s^J)=q^{s(L-r)}\qbin rs, \qquad\text{where $L=\max(m,n)$.}
\]
\end{thm}
\pf
For convenience, we will assume that $m\leq n$ throughout the proof, so that $L=n$. (The other case will follow by duality.)
More generally, it may easily be checked that if the $\gJ$-classes of a semigroup $S$ form a chain, $J_0<\cdots<J_k$, then the ideals of $S$ are precisely the sets $I_h=J_0\cup\cdots\cup J_h$ for $0\leq h\leq k$ (and these obviously form a chain). Now suppose $0\leq s<r$, let $\Ga\sub E(D_s(\M_r))$ be any idempotent generating set of $I_s(\M_r)$ (see Theorem~\ref{thm_ideals_Mn}), and put $\Om_\Ga=\set{[M,A,N]}{M\in\M_{m-r,r},\ A\in\Ga,\ N\in\M_{r,n-r}}$. If $X=[M,A,N]\in I_s^J$ is arbitrary, then $A=B_1\cdots B_k$ for some $B_1,\ldots,B_k\in\Ga$, and it follows that $X=[M,B_1,N]\star\cdots\star[M,B_k,N]\in\la\Om_\Ga\raJ$. Since $\Om_\Ga\sub\EJ(D_s^J)$, it follows that $I_s^J = \la \Ea(D_s^J) \raa$.
We now prove the statement about rank and idempotent rank. Suppose $\Om$ is an arbitrary generating set for~$I_s^J$ where $0\leq s<r$. Let $X\in D_s^J$ and consider an expression $X=Y_1\star\cdots\star Y_k$ with $Y_1,\ldots,Y_k\in\Om$. Since $X=X\star Z\star X$ for some $Z\in D_s^J$, we may assume that $k\geq2$. Since $I_{s-1}^J$ is an ideal of $I_s^J$ (we interpret $I_{s-1}^J=\emptyset$ if $s=0$), each of $Y_1,\ldots,Y_k$ must belong to $D_s^J=D_X^J$. In particular, $Y_k\gDJ X=(Y_1\star\cdots\star Y_{k-1})\star Y_k$. By stability, it then follows that $Y_k\gLJ (Y_1\star\cdots\star Y_{k-1})\star Y_k=X$. Since $X\in D_s^J$ was arbitrary, it follows that $\Om$ contains at least one element from each $\gLJ$-class contained in $D_s^J$, and there are $q^{s(n-r)}\tqbin rs$ such $\gLJ$-classes, by Theorem~\ref{inflation_thm}(vi). It follows that $\rank(I_s^J)\geq q^{s(n-r)}\tqbin rs
=q^{s(L-r)}\tqbin rs$.
Since $\idrank(S)\geq\rank(S)$ for any idempotent generated semigroup $S$, the proof will be complete if we can find an idempotent generating set of $I_s^J$ of the specified size. First, let $\Ga\sub E(D_s(\M_r))$ be such that ${\la\Ga\ra=I_s(\M_r)}$ and $|\Ga|=\tqbin rs$. Fix some $A\in\Ga$, and let $\sim_A$ and $\approx_A$ be the equivalence relations on $\M_{m-r,r}$ and $\M_{r,n-r}$ defined before Theorem \ref{inflation_thm}, and let $\sM_A$ and $\sN_A$ be cross-sections of the equivalence classes of $\sim_A$ and $\approx_A$. Let $\sM_A=\{M_1,\ldots,M_{q^{s(m-r)}}\}$ and $\sN_A=\{N_1,\ldots,N_{q^{s(n-r)}}\}$. (We know $\sM_A$ and $\sN_A$ have the specified sizes by Theorem \ref{inflation_thm}.) Put $Q=q^{s(n-r)}=q^{s(L-r)}$. (Recall that we are assuming $m\leq n$.) Extend $\sM_A$ arbitrarily to $\sM_A'=\{M_1,\ldots,M_Q\}$. Now put $\Om_A=\set{[M_i,A,N_i]}{1\leq i\leq Q}$. If $M\in\M_{m-r,r}$ and $N\in\M_{r,n-r}$ are arbitrary, then $M\sim M_i$ and $N\sim N_j$ for some $i,j$, and we have $[M,A,N]=[M_i,A,N_j]=[M_i,A,N_i]\star[M_j,A,N_j]\in\la\Om_A\raJ$. Now put $\Om=\bigcup_{A\in\Ga}\Om_A$. By the previous discussion, we see that $\la\Om\raJ$ contains $\Om_\Ga$, which is a generating set for $I_s^J$ (by the first paragraph of this proof), so $I_s^J=\la\Om\raJ$. Since $|\Om|=Q|\Ga|=q^{s(L-r)}\tqbin rs$, the proof is complete. \epf
\begin{rem}
Again, we may deduce a corresponding statement for the ideals of the matrix semigroups $\Reg(\Ckl)$ and $\Reg(\Rkl)$; the reader may supply the details if they wish.
\end{rem}
\subsection*{Acknowledgements}
The first named author gratefully acknowledges the support of Grant No.~174019 of the Ministry of Education, Science, and Technological Development of the Republic of Serbia, and Grant No.~1136/2014 of the Secretariat of Science and Technological Development of the Autonomous Province of Vojvodina.
The authors wish to thank Dr Attila Egri-Nagy for constructing the GAP \cite{GAP} code that enabled us to produce the eggbox diagrams from Figures \ref{fig:V3212_V3322}, \ref{fig:V2322_V2422}, \ref{fig:V2202} and \ref{fig:R}.
\footnotesize
\def-1.1pt{-1.1pt}
|
\section{Introduction}
Relativistic viscous hydrodynamics has been applied quite
successfully to study and understand various collective phenomena
observed in the evolution of the strongly interacting QCD matter,
with very high temperature and density, created in relativistic
heavy-ion collisions; see Ref.~\cite{Heinz:2013th} for a recent
review. The derivation of hydrodynamic equations is essentially a
coarse graining procedure whereby one obtains an effective theory
describing the long-wavelength low-frequency limit of the
microscopic dynamics of a system \cite{Heinz:2013th, Landau}.
Relativistic hydrodynamics is formulated as an order-by-order
expansion in powers of space-time gradients where ideal
hydrodynamics is of zeroth order \cite{Heinz:2013th}. The viscous
effects arising in the first-order theory, also known as the
relativistic Navier-Stokes theory \cite{Eckart:1940zz, Landau},
results in acausal signal propagation and numerical instability.
While causality is restored in the second-order Israel-Stewart (IS)
theory \cite{Israel:1979wp}, stability may not be guaranteed \cite
{Romatschke:2009im}. Consistent formulation of a causal theory of
relativistic viscous hydrodynamics and accurate determination of the
associated transport coefficients is currently an active research
topic \cite {Muronga:2003ta, El:2009vj, Denicol:2012cn,
Jaiswal:2014isa, Chattopadhyay:2014lya, Denicol:2010xn,
Jaiswal:2012qm, Jaiswal:2013fc, Bhalerao:2013aha, Denicol:2014vaa,
Denicol:2014mca, Jaiswal:2013npa, Jaiswal:2013vta, Bhalerao:2013pza,
Romatschke:2003ms, Florkowski:2010cf, Martinez:2010sc,
Florkowski:2013lza, Bazow:2013ifa, Nopoush:2014pfa,
Florkowski:2014bba, Florkowski:2014sfa, Florkowski:2014sda,
Prakash:1993bt, Wiranata:2012br, Wiranata:2012vv, Davesne:1995ms,
Wiranata:2014jda}.
The second-order IS equations can be derived in several ways \cite
{Romatschke:2009im}. For example, in the derivations based on the
second law of thermodynamics, the hydrodynamic transport
coefficients related to the relaxation times for bulk and shear
viscous evolution remain undetermined. While these transport
coefficients can be obtained in the derivations based on kinetic
theory \cite{Israel:1979wp, Muronga:2003ta}, the form of
non-equilibrium phase-space distribution function, $f(x,p)$, has to
be specified. Two most extensively used methods to determine $f(x,p)$
for a system which is close to local thermodynamic equilibrium are
(1) the Grad's 14-moment approximation \cite{Grad} and (2) the
Chapman-Enskog method \cite{Chapman}. Note that while Grad's
14-moment approximation has been widely employed in the formulation
of a causal theory of relativistic dissipative hydrodynamics \cite
{Israel:1979wp, Muronga:2003ta, El:2009vj, Denicol:2010xn,
Denicol:2012cn, Jaiswal:2012qm, Jaiswal:2013fc, Bhalerao:2013aha,
Denicol:2014vaa, Denicol:2014mca, Romatschke:2009im}, the
Chapman-Enskog method remains less explored \cite{Jaiswal:2013npa,
Jaiswal:2013vta, Bhalerao:2013pza, Jaiswal:2014isa,
Chattopadhyay:2014lya}. On the other hand, the Chapman-Enskog
formalism has been often used to extract various transport
coefficients of hot hadronic matter \cite{Prakash:1993bt,
Wiranata:2012br, Wiranata:2012vv, Davesne:1995ms, Wiranata:2014jda}
Although in both methods the distribution function is expanded
around its equilibrium value $f_0(x,p)$, it has been demonstrated
that the Chapman-Enskog method in the relaxation-time approximation
(RTA) leads to better agreement with both microscopic Boltzmann
simulations as well as exact solutions of the relativistic RTA
Boltzmann equation \cite {Jaiswal:2013npa, Jaiswal:2013vta,
Jaiswal:2014isa, Chattopadhyay:2014lya}. This may be attributed to
the fact that a fixed-order moment expansion, as required in Grad's
approximation, is not necessary in the Chapman-Enskog method.
Much of the research on the application of viscous hydrodynamics in
relativistic heavy-ion collisions is devoted to the extraction of
the shear viscosity to entropy density ratio, $\eta/s$, from the
analysis of the anisotropic flow data \cite{Romatschke:2007mq,
Song:2010mg, Schenke:2011bn}. Indeed the estimated $\eta/s$ has been
found to be close to the conjectured lower bound $\eta/s|_{\rm KSS}
= 1/4\pi$ \cite{Policastro:2001yc, Kovtun:2004de}. On the other
hand, a self-consistent and systematic study of the effect of the
bulk viscosity in numerical simulations of high-energy heavy-ion
collisions is relatively lacking. This may be attributed to the fact
that the hot QCD matter is assumed to be nearly conformal and
therefore the bulk viscosity is estimated to be much smaller
compared to the shear viscosity. However, in reality, QCD is a
non-conformal field theory and therefore bulk viscous corrections to
the energy-momentum tensor should not be ignored in order to
correctly understand the dynamics of the QCD system. Moreover, for
the range of temperature explored in heavy-ion collision experiments
at Relativistic Heavy-Ion Collider (RHIC) and Large Hadron Collider
(LHC), the magnitude and temperature dependence of the bulk
viscosity could be large enough to influence the space-time
evolution of the hot QCD matter \cite {Moore:2008ws,
Noronha-Hostler:2014dqa, Rose:2014fba, Ryu:2015vwa, Wiranata:2009cz,
Gavin:1985ph, Chakraborty:2010fr}.
It is important to note that the second-order transport
coefficients, appearing in the evolution equation for the bulk
viscous pressure, are less understood compared to those of the shear
stress tensor. While the relaxation time for the bulk viscous
evolution can be obtained by using the second law of thermodynamics
in a kinetic theory set up \cite{Jaiswal:2013fc, Bhalerao:2013aha},
this method fails to account for the important coupling of the bulk
viscous pressure with the shear stress tensor \cite
{Denicol:2014vaa, Denicol:2014mca, Jaiswal:2014isa}. On the other
hand, for finite masses and classical Boltzmann distribution, the
second-order transport coefficients corresponding to bulk viscous
pressure and shear stress tensor have been obtained by employing the
Grad's 14-moment approximation \cite{Denicol:2014vaa,
Denicol:2014mca} as well as the Chapman-Enskog method \cite
{Jaiswal:2014isa}, within a purely kinetic theory framework.
However, these transport coefficients still remain to be determined
for quantum statistics, i.e., for Bose-Einstein and Fermi-Dirac
distribution.
In this paper, we express the transport coefficients appearing in
the second-order viscous evolution equations with non-vanishing
masses for Bose-Einstein, Boltzmann and Fermi-Dirac distribution. We
obtain these transport coefficients using the Grad's 14-moment
approximation as well as the method of Chapman-Enskog expansion. In
addition, in the case of one-dimensional scaling expansion of the
viscous medium, we compare the results obtained using the above
methods with those obtained from the exact solution of massive 0+1d
relativistic Boltzmann equation in the relaxation-time approximation
\cite{Florkowski:2014sda}. We demonstrate that the results obtained
using the Chapman-Enskog method are in better agreement with the
exact solution of the RTA Boltzmann equation than those obtained
using the Grad's 14-moment approximation.
\section{Relativistic hydrodynamics}
In the absence of any conserved charges, i.e., for vanishing
chemical potential, the hydrodynamic evolution of a system is
governed by the local conservation of energy and momentum:
$\partial_\mu T^{\mu\nu}=0$. The energy-momentum tensor, $T^{\mu\nu}$,
which characterizes the macroscopic state of a system, can be
expressed in terms of the single-particle phase-space distribution
function, $f(x,p)$, and tensor decomposed into hydrodynamic degrees
of freedom \cite{deGroot},
\begin{equation}\label{NTD}
T^{\mu\nu} = \!\int\! dP \, p^\mu p^\nu f(x,p) = \epsilon u^\mu u^\nu
- (P+\Pi)\Delta^{\mu\nu} + \pi^{\mu\nu}.
\end{equation}
Here $dP\equiv gd^3p/[(2\pi)^3p^0]$ is the invariant momentum-space
integration measure, $g$ being the degeneracy factor, and $p^\mu$ is
the particle four-momentum. In the tensor decomposition, $\epsilon$,
$P$, $\Pi$, and $\pi^{\mu\nu}$ are the energy density, the
thermodynamic pressure, the bulk viscous pressure, and the shear
stress tensor, respectively. The projection operator
$\Delta^{\mu\nu}\equiv g^{\mu\nu}-u^\mu u^\nu$ is constructed such
that it is orthogonal to the hydrodynamic four-velocity $u^\mu$. The
metric tensor is assumed to be Minkowskian,
$g^{\mu\nu}\equiv\mathrm{diag} ({+}1,{-}1,{-}1,{-}1)$, and $u^\mu$
is defined in the Landau frame: $T^{\mu\nu} u_\nu=\epsilon u^\mu$.
The energy-momentum conservation equation, $\partial_\mu
T^{\mu\nu}=0$, when projected along and orthogonal to $u^\mu$ gives
the evolution equations for $\epsilon$ and $u^\mu$,
\begin{align}\label{evol}
\dot\epsilon + (\epsilon+P+\Pi)\theta - \pi^{\mu\nu}\sigma_{\mu\nu} &= 0, \nonumber\\
(\epsilon+P+\Pi)\dot u^\alpha - \nabla^\alpha (P+\Pi) + \Delta^\alpha_\nu \partial_\mu \pi^{\mu\nu} &= 0.
\end{align}
Here we have used the usual notation $\theta\equiv\partial_\mu u^\mu$
for the expansion scalar, $\dot A\equiv u^\mu\partial_\mu A$ for
the co-moving derivative, $\nabla^\alpha\equiv\Delta^{\mu\alpha}
\partial_\mu$ for the space-like derivative, and
$\sigma^{\mu\nu}\equiv(\nabla^\mu u^\nu + \nabla^\nu u^\mu)/2
-(\theta/3)\Delta^{\mu\nu}$ for the velocity stress tensor. The
energy density and the thermodynamic pressure can be written in
terms of the equilibrium phase-space distribution function $f_0$ as
\begin{align}
\epsilon_0 &= u_\mu u_\nu \!\int\! dP \, p^\mu p^\nu f_0, \label{ED}\\
P_0 &= -\frac{1}{3}\Delta_{\mu\nu} \!\int\! dP \, p^\mu p^\nu f_0, \label{TP}
\end{align}
where the equilibrium distribution function for vanishing chemical
potential is given by
\begin{equation}\label{EDF}
f_0 = \frac{1}{\exp(\beta u\cdot p) + a}.
\end{equation}
Here $u\cdot p\equiv u_\mu p^\mu$ and $a=-1,0,1$ for Bose-Einstein,
Boltzmann, and Fermi-Dirac gas, respectively. The inverse
temperature, $\beta\equiv1/T$, is determined by the matching
condition $\epsilon=\epsilon_0$.
For a system close to local thermodynamic equilibrium, the
non-equilibrium phase-space distribution function can be written as
$f=f_0+\delta f$, where $\delta f\ll f$. Using Eq.~(\ref {NTD}), the
bulk viscous pressure, $\Pi$, and the shear stress tensor,
$\pi^{\mu\nu}$, can be expressed in terms of $\delta f$ as \cite
{deGroot}
\begin{align}
\Pi &= -\frac{1}{3}\Delta_{\alpha\beta} \!\int\! dP \, p^\alpha p^\beta\, \delta f, \label{BVP}\\
\pi^{\mu\nu} &= \Delta^{\mu\nu}_{\alpha\beta} \!\int\! dP \, p^\alpha p^\beta\, \delta f, \label{SST}
\end{align}
where $\Delta^{\mu\nu}_{\alpha\beta}\equiv
(\Delta^{\mu}_{\alpha}\Delta^{\nu}_{\beta} +
\Delta^{\mu}_{\beta}\Delta^{\nu}_{\alpha})/2 -
(1/3)\Delta^{\mu\nu}\Delta_{\alpha\beta}$ is a traceless symmetric
projection operator orthogonal to $u^\mu$ and $\Delta^{\mu\nu}$. In
the following, we employ the expressions for $\delta f$ obtained
using the Grad's 14-moment approximation and the Chapman-Enskog like
iterative solution of the relativistic Boltzmann equation to obtain
expressions for the quantum transport coefficients associated with
viscous evolution.
\section{Viscous corrections to the distribution function}
Precise determination of the form of the non-equilibrium single
particle phase-space distribution function is one of the central
problems in statistical physics. For a system close to local
thermodynamic equilibrium, the problem reduces to determining the
form of the correction to the equilibrium distribution function.
Within the framework of relativistic hydrodynamics, the viscous
corrections to the equilibrium distribution function can be obtained
from two different methods: (1) the moment method and (2) the
Chapman-Enskog method. The moment method, more popularly known as
the Grad's 14-moment ansatz, is based on a Taylor-like expansion of
the non-equilibrium distribution in powers of momenta. On the other
hand, the Chapman-Enskog method relies on the solution of the
Boltzmann equation.
Ignoring dissipation due to particle diffusion, the Grad's 14-moment
approximation leads to
\begin{align}\label{G14M}
\delta f_{\rm G} =\, & \Big[\left\{E_0 + B_0m^2 + D_0(u\cdot p) - 4B_0(u\cdot p)^2\right\}\Pi \nonumber\\
&+ B_2p^\alpha p^\beta \pi_{\alpha\beta}\Big]f_0 \tilde f_0,
\end{align}
where $\tilde f_0 = 1-af_0$. The coefficients $E_0$, $B_0$, $D_0$
and $B_2$ are known in terms of $m$, $T$ and $u\cdot p$ and can be
expressed as
\begin{align}
B_2 =\,& \frac{1}{2J^{(0)}_{42}}, \label{B2}\\
\frac{D_0}{3B_0} =\,& 4\frac{J^{(0)}_{31}J^{(0)}_{20} - J^{(0)}_{41}J^{(0)}_{10}}
{J^{(0)}_{30}J^{(0)}_{10} - J^{(0)}_{20}J^{(0)}_{20}} \equiv C_2,\label{D0}\\
\frac{E_0}{3B_0} =\;& m^2 - 4\frac{J^{(0)}_{31}J^{(0)}_{30} - J^{(0)}_{41}J^{(0)}_{20}}
{J^{(0)}_{30}J^{(0)}_{10} - J^{(0)}_{20}J^{(0)}_{20}} \equiv C_1,\label{E0}\\
B_0 =\,& -\frac{1}{3C_1J^{(0)}_{21} + 3C_2J^{(0)}_{31} - 3J^{(0)}_{41} + 5J^{(0)}_{42}},\label{B0}
\end{align}
where the thermodynamic functions $J^{(r)}_{nq}$ are defined as
\begin{equation}\label{ICJ}
J_{nq}^{(r)} \equiv \frac{1}{(2q+1)!!}\!\int\! dP\,
(u\cdot p)^{n-2q-r}\,(\Delta_{\mu\nu}p^\mu p^\nu)^q f_0\tilde f_0.
\end{equation}
In the above equation, the indices $n-r$ and $q$ represents the
number of times $p^\mu$ and $\Delta_{\mu\nu}$ appear in the
integration, respectively.
An analogous expression for $\delta f$ can also be obtained using an
iterative Chapman-Enskog like solution of the relativistic Boltzmann
equation in the relaxation-time approximation \cite
{Romatschke:2011qp, Jaiswal:2014isa}. In absence of dissipation due
to particle diffusion, the Chapman-Enskog method leads to \cite
{Jaiswal:2014isa}
\begin{align}\label{DFCE}
\delta f_{\rm CE} = \frac{\beta }{u\cdot p}\bigg[&- \frac{1}{3\beta_\Pi}\left\{m^2-(1-3c_s^2)(u\cdot p)^2\right\}\Pi \nonumber\\
& + \frac{1}{2\beta_\pi}\;p^\mu p^\nu\pi_{\mu\nu}\bigg]f_0\tilde f_0,
\end{align}
where
\begin{align}
\beta_\Pi &= \frac{5}{3}\beta_\pi - (\epsilon+P)c_s^2, \label{betaPi}\\
\beta_\pi &= \beta\, J_{42}^{(1)}, \label{betapi}
\end{align}
and $c_s^2\equiv dP/d\epsilon$ is the speed of sound squared which
can be expressed as
\begin{equation}\label{cs2}
c_s^2 = \frac{\epsilon+P}{\beta J_{30}^{(0)}}.
\end{equation}
\section{Viscous evolution equations}
The second-order evolution equations for $\Pi$ and $\pi^{\mu\nu}$
can be derived by considering the co-moving derivative of
Eqs.~({\ref{BVP}) and ({\ref{SST}),
\begin{align}
\dot\Pi &= -\frac{1}{3}\Delta_{\alpha\beta} \!\int\! dP \, p^\alpha p^\beta\, \delta\dot f, \label{BVPE}\\
\dot\pi^{\langle\mu\nu\rangle} &= \Delta^{\mu\nu}_{\alpha\beta} \!\int\! dP \, p^\alpha p^\beta\, \delta\dot f. \label{SSTE}
\end{align}
Further, $\delta\dot f$ appearing in the above equations can be
simplified by rewriting the relativistic Boltzmann equation,
\begin{equation}\label{RBE}
p^\mu\partial_\mu f = C[f],
\end{equation}
in the form \cite{Denicol:2010xn}
\begin{equation}\label{RBER}
\delta\dot f = -\dot f_0 - \frac{1}{u\cdot p}\, p^\mu\nabla_\mu f + \frac{1}{u\cdot p}\, C[f].
\end{equation}
In the following, we consider the relaxation-time approximation for
the collision term in the Boltzmann equation,
\begin{equation}\label{RTAC}
C[f] = -\left(u\cdot p\right) \frac{\delta f}{\tau_{\rm eq}},
\end{equation}
where $\tau_{\rm eq}$ is the relaxation time. In general, in order
for the collision term in the above equation to respect the
conservation of particle four-current and energy-momentum tensor,
$\tau_{\rm eq}$ must be independent of momenta and $u^\mu$ has to be
defined in the Landau frame \cite{Anderson_Witting}.
Substituting $\delta\dot f$ from Eq.~(\ref{RBER}) into Eqs.~({\ref
{BVPE}) and ({\ref{SSTE}) along with the form of $\delta f$ given in
Eqs.~({\ref{G14M}) and ({\ref{DFCE}), and after performing the
integrations, we obtain
\begin{align}
\dot{\Pi} =& -\frac{\Pi}{\tau_{\rm eq}}
-\beta_{\Pi}\theta
-\delta_{\Pi\Pi}\Pi\theta
+\lambda_{\Pi\pi}\pi^{\mu\nu}\sigma_{\mu \nu }, \label{BULK}\\
\dot{\pi}^{\langle\mu\nu\rangle} =& -\frac{\pi^{\mu\nu}}{\tau_{\rm eq}}
+2\beta_{\pi}\sigma^{\mu\nu}
+2\pi_{\gamma}^{\langle\mu}\omega^{\nu\rangle\gamma}
-\delta_{\pi\pi}\pi^{\mu\nu}\theta \nonumber \\
&-\tau_{\pi\pi}\pi_{\gamma}^{\langle\mu}\sigma^{\nu\rangle\gamma}
+\lambda_{\pi\Pi}\Pi\sigma^{\mu\nu}, \label{SHEAR}
\end{align}
where $\omega ^{\mu \nu }\equiv \frac{1}{2}(\nabla^{\mu}u^{\nu}
-\nabla^{\nu }u^{\mu })$ is the vorticity tensor. Note that the
above form of the evolution equations for the bulk viscous pressure
and the shear stress tensor is exactly same for both Grad's
14-moment approximation ($\delta f_G$) and the Chapman-Enskog
expansion ($\delta f_{CE}$). Moreover, the expressions for the first
order transport coefficients, $\beta_{\Pi}$ and $\beta_{\pi}$, are
also identical for these two cases and are given in Eqs.~(\ref
{betaPi}) and (\ref{betapi}). However, the second order transport
coefficients appearing in the above equations are different for the
Grad's 14-moment method and the Chapman-Enskog method.
The transport coefficients in the case of Grad's 14-moment
approximation are calculated to be
\begin{align}
\delta^{\rm (G)}_{\Pi\Pi} &= 1 - c_s^2 - \frac{m^4}{9}\gamma^{(0)}_2, \label{Gcoeff1}\\
\lambda^{\rm (G)}_{\Pi\pi} &= \frac{1}{3} - c_s^2 + \frac{m^2}{3}\gamma^{(2)}_2, \label{Gcoeff2}\\
\delta^{\rm (G)}_{\pi\pi} &= \frac{4}{3} + \frac{m^2}{3}\gamma^{(2)}_2, \label{Gcoeff4}\\
\tau^{\rm (G)}_{\pi\pi} &= \frac{10}{7} + \frac{4m^2}{7}\gamma^{(2)}_2, \label{Gcoeff3}\\
\lambda^{\rm (G)}_{\pi\Pi} &= \frac{6}{5} - \frac{2m^4}{15}\gamma^{(0)}_2, \label{Gcoeff5}
\end{align}
where
\begin{align}
\gamma^{(0)}_2 &= (E_0+B_0m^2)J^{(0)}_{-20} + D_0J^{(0)}_{-10} - 4B_0J^{(0)}_{00}, \label{gamma0}\\
\gamma^{(2)}_2 &= \frac{J^{(0)}_{22}}{J^{(0)}_{42}}. \label{gamma2}
\end{align}
On the other hand, these transport coefficients in the case of
Chapman-Enskog method are obtained as
\begin{align}
\delta^{\rm (CE)}_{\Pi\Pi} &= -\frac{5}{9}\,\chi - c_s^2, \label{CEcoeff1}\\
\lambda^{\rm (CE)}_{\Pi\pi} &= \frac{\beta}{3\beta_\pi}\!\left(7J_{63}^{(3)}+2J_{42}^{(1)}\right) - c_s^2, \label{CEcoeff2}\\
\delta^{\rm (CE)}_{\pi\pi} &= \frac{5}{3} + \frac{7\beta}{3\beta_\pi}\, J_{63}^{(3)},\label{CEcoeff4}\\
\tau^{\rm (CE)}_{\pi\pi} &= 2 + \frac{4\beta}{\beta_\pi}\,J_{63}^{(3)}, \label{CEcoeff3}\\
\lambda^{\rm (CE)}_{\pi\Pi} &= -\frac{2}{3}\chi, \label{CEcoeff5}
\end{align}
where
\begin{equation}
\chi = \frac{\beta}{\beta_\Pi}\!\left[(1-3c_s^2)\!\left(J_{42}^{(1)}+J_{31}^{(0)}\right)
- m^2\!\left(J_{42}^{(3)}+J_{31}^{(2)}\right)\right].
\end{equation}
The integral functions $J_{nq}^{(r)}$ appearing in the
expressions for the transport coefficients satisfy the relations
\begin{align}
J_{nq}^{(r)} & = \frac{1}{(2q+1)}\left[m^2 J_{n-2,q-1}^{(r)} - J_{n,q-1}^{(r)}\right], \label{prop1} \\
J_{nq}^{(0)}&= \frac{1}{\beta}\left[-I_{n-1,q-1}^{(0)} + (n-2q)I_{n-1,q}^{(0)} \right],\label{prop2}
\end{align}
where,
\begin{equation}\label{ICI}
I_{nq}^{(r)} \equiv \frac{1}{(2q+1)!!}\!\int\! dP\,
(u\cdot p)^{n-2q-r}\,(\Delta_{\mu\nu}p^\mu p^\nu)^q f_0.
\end{equation}
Here we readily identify $I_{20}^{(0)}=\epsilon$ and
$I_{21}^{(0)}=-P$. Using Eqs.~(\ref{prop1}) and (\ref{prop2}), we
obtain the identities
\begin{align}
J_{42}^{(0)} &= \frac{m^2}{5}J_{21}^{(0)} - \frac{1}{5}J_{41}^{(0)}, \label{J042}\\
J_{31}^{(0)} &= -\frac{1}{\beta}(\epsilon+P). \label{J031}
\end{align}
To compute all the transport coefficients, we also need to determine
the integrals $J_{20}^{(0)}$, $J_{10}^{(0)}$, $J_{41}^{(0)}$,
$J_{21}^{(0)}$, $J_{30}^{(0)}$, $J_{-20}^{(0)}$, $J_{-10}^{(0)}$,
$J_{00}^{(0)}$, $J_{22}^{(0)}$, $J_{63}^{(3)}$, $J_{42}^{(1)}$,
$J_{42}^{(3)}$, and $J_{31}^{(2)}$. In the following, we obtain
expressions for these quantities in terms of modified Bessel
functions of the second kind.
\section{Transport coefficients}
Let us first simplify our equilibrium distribution function. Using
the result of summation of a infinite geometric progression,
\begin{equation}\label{GPS}
\frac{1}{1+x} = 1-x+x^2-x^3\cdots = \sum_{l=0}^{\infty}(-1)^lx^l, \quad \rm{for}~|x|<1,
\end{equation}
we obtain,
\begin{equation}\label{EDFGP}
f_0 = \frac{e^{-\beta (u\cdot p)}}{1+ae^{-\beta (u\cdot p)}} = \sum_{l=1}^{\infty}(-a)^{l-1}e^{-l\beta (u\cdot p)}.
\end{equation}
Similarly, using the result obtained after differentiating Eq. (\ref
{GPS}), we obtain
\begin{equation}\label{FTFGP}
f_0\tilde f_0 = \frac{e^{-\beta (u\cdot p)}}{(1+ae^{-\beta (u\cdot p)})^2} = \sum_{l=1}^{\infty}l(-a)^{l-1}e^{-l\beta (u\cdot p)}.
\end{equation}
Using the above results for $f_0$ and $f_0\tilde f_0$, the
thermodynamic integrals $I_{nq}^{(r)}$ and $J_{nq}^{(r)}$ can be
written as,
\begin{align}
I_{nq}^{(r)} =&\ \frac{gT^{n+2-r}z^{n+2-r}}{2\pi^2(2q+1)!!}\,(-1)^q\sum_{l=1}^{\infty}(-a)^{l-1}
\!\!\int_0^\infty\!\!\! d\theta\, \label{Inqr} \\
&\times(\cosh\theta)^{n-2q-r}(\sinh\theta)^{2q+2}\,\exp(-lz\cosh\theta), \nonumber\\
J_{nq}^{(r)} =&\ \frac{gT^{n+2-r}z^{n+2-r}}{2\pi^2(2q+1)!!}\,(-1)^q\sum_{l=1}^{\infty}l(-a)^{l-1}
\!\!\int_0^\infty\!\!\!\! d\theta\, \label{Jnqr} \\
&\times(\cosh\theta)^{n-2q-r}(\sinh\theta)^{2q+2}\,\exp(-lz\cosh\theta), \nonumber
\end{align}
where $z\equiv\beta m=m/T$.
The integral coefficients $I_{nq}^{(r)}$ and $J_{nq}^{(r)}$ can be
expressed in terms of modified Bessel functions of the second kind.
The integral representation of the relevant Bessel function is given
by
\begin{equation}\label{Bessel}
K_n(z) = \!\int_0^\infty\! d\theta \cosh(n\theta)\, \exp(-z\cosh\theta).
\end{equation}
The pressure, energy density and $J_{30}^{(0)}$, required to
calculate the velocity of sound, can be expressed in terms of Bessel
functions using the above equation
\begin{align}
P =&\, \frac{gT^4z^2}{2\pi^2}\!\sum_{l=1}^{\infty}\frac{1}{l^2}(-a)^{l-1}K_2(lz), \label{Prs_K}\\
\epsilon =&\, \frac{gT^4z^3}{2\pi^2}\!\sum_{l=1}^{\infty}\frac{1}{l}(-a)^{l-1}K_3(lz) - P, \label{eng_K}\\
J_{30}^{(0)} =&\, \frac{gT^5z^5}{32\pi^2}\!\sum_{l=1}^{\infty}l(-a)^{l-1}\Big[K_5(lz)+K_3(lz)-2K_1(lz)\Big]. \label{J300}
\end{align}
The integral functions $J_{nq}^{(r)}$ appearing in the transport
coefficients obtained using the Grad's 14-moment method can be
expressed as
\begin{align}\label{relint_G}
J_{21}^{(0)} =& -\frac{gT^4z^2}{2\pi^2}\!\sum_{l=1}^{\infty}\frac{1}{l}(-a)^{l-1}K_2, \nonumber\\
J_{20}^{(0)} =&\, \frac{gT^4z^3}{2\pi^2}\!\sum_{l=1}^{\infty}(-a)^{l-1}K_3 + J_{21}^{(0)}, \nonumber\\
J_{10}^{(0)} =&\, \frac{gT^3z^3}{8\pi^2}\!\sum_{l=1}^{\infty}l(-a)^{l-1}\Big[K_3-K_1\Big], \nonumber\\
J_{41}^{(0)} =& -\frac{gT^6z^6}{192\pi^2}\!\sum_{l=1}^{\infty}l(-a)^{l-1}\Big[K_6-2K_4-K_2+2K_0\Big], \nonumber\\
J_{-20}^{(0)} =&\, \frac{g}{2\pi^2}\!\sum_{l=1}^{\infty}l(-a)^{l-1}\Big[K_0-K_{i,2}\Big], \nonumber\\
J_{-10}^{(0)} =&\, \frac{gTz}{2\pi^2}\!\sum_{l=1}^{\infty}l(-a)^{l-1}\Big[K_1-K_{i,1}\Big], \nonumber\\
J_{22}^{(0)} =&\, \frac{gT^4z^4}{240\pi^2}\!\sum_{l=1}^{\infty}l(-a)^{l-1}\Big[K_4-8K_2+15K_0-8K_{i,2}\Big], \nonumber\\
J_{00}^{(0)} =&\, \frac{gT^2z^2}{4\pi^2}\!\sum_{l=1}^{\infty}l(-a)^{l-1}\Big[K_2-K_0\Big],
\end{align}
where the $(lz)$-dependence of $K_n$ and $K_{i,n}$ is implicitly
understood. The integral functions appearing in the transport
coefficients obtained using the Chapman-Enskog method are
\begin{align}\label{relint_CE}
J_{63}^{(3)} =& -\frac{gT^5z^5}{3360\pi^2}\!\sum_{l=1}^{\infty}l(-a)^{l-1}\Big[K_5-11K_3+58K_1 \nonumber\\
&-64K_{i,1}+16K_{i,3}\Big], \nonumber\\
J_{42}^{(1)} =&\, \frac{gT^5z^5}{480\pi^2}\!\sum_{l=1}^{\infty}l(-a)^{l-1}\Big[K_5\!-7K_3\!+22K_1\!-16K_{i,1}\Big], \nonumber\\
J_{42}^{(3)} =&\, \frac{gT^3z^3}{120\pi^2}\!\sum_{l=1}^{\infty}l(-a)^{l-1}\Big[K_3\!-9K_1\!+12K_{i,1}\!-4K_{i,3}\Big], \nonumber\\
J_{31}^{(2)} =& -\frac{gT^3z^3}{24\pi^2}\!\sum_{l=1}^{\infty}l(-a)^{l-1}\Big[K_3-5K_1+4K_{i,1}\Big].
\end{align}
Here the function $K_{i,n}(lz)$ is defined by the integral
\begin{equation}\label{kin}
K_{i,n}(lz) = \!\int_0^\infty\! \frac{d\theta}{(\cosh\theta)^n}\,\exp(-lz\cosh\theta).
\end{equation}
Note that the subscript $i$ in the above function is not an index
and just serves to distinguish it from the Bessel functions.
$K_{i,n}$ satisfy the following recurrence relation:
\begin{equation}\label{kinkn}
\frac{d}{dz}K_{i,n}(lz) = -lK_{i,n-1}(lz),
\end{equation}
which can be expressed in the integral form:
\begin{equation}\label{kinkn}
K_{i,n}(lz) = K_{i,n}(0) - l\!\int_0^z\! K_{i,n-1}(lz') dz'.
\end{equation}
We observe that by matching $K_{i,0}(lz)=K_0(lz)$, the above
recursion relation can be used to evaluate $K_{i,n}(lz)$ up to any
$n$.
Armed with the above expressions for $J_{nq}^{(r)}$ in terms of
series summation of the Bessel function, it is instructive to
calculate the ratio between the coefficient of bulk viscosity and
shear viscosity. In the relaxation-time approximation, this ratio is
given by $\zeta/\eta = \beta_\Pi/\beta_\pi$ \cite{Denicol:2014vaa,
Jaiswal:2014isa}. In the small-$z$ approximation, using the series
expansion of the Bessel functions in powers of $z$, we obtain
\begin{equation}\label{zetabyeta}
\frac{\zeta}{\eta} = \Gamma\left(\frac{1}{3}-c_s^2\right)^2 + {\cal O}(z^5),
\end{equation}
where $\Gamma=75$ for Boltzmann distribution and $\Gamma=48$ for
Fermi-Dirac distribution. However, in the case of Bose-Einstein
distribution (remember $a=-1$), we get $\Gamma=-15+36/(1+a)$, and
therefore it diverges as $\sim1/(1+a)$ up to the leading-order. It
is interesting to note that the above expression is similar to the
well known relation, $\zeta/\eta=15(1/3-c_s^2)^2$, derived by
Weinberg \cite {Weinberg}. The difference in the proportionality
constant may be attributed to the fact that here we have considered
a single component system whereas, motivated by applications to
cosmology, Weinberg considered a system composed of mixture of
radiation and matter.
\begin{figure}[t]
\begin{center}
\includegraphics[width=\linewidth]{zetabyeta.pdf}
\end{center}
\vspace{-0.8cm}
\caption{(Color online) $m/T$ dependence of the ratio $\zeta/\eta$
scaled by $(1/3-c_s^2)^2$ for Boltzmann (red solid line),
Bose-Einstein (blue dashed line) and Fermi-Dirac (green dotted
line) statistics. We also show results for two-flavor (brown
dashed-dotted line) and three-flavor QGP (purple
dashed-dotted-dotted line). Thin horizontal black dashed lines
corresponds to constant values $75$ and $48$.}
\label{ZetabyEta}
\end{figure}
In order to calculate $\zeta/\eta$ for the quark-gluon plasma (QGP),
we need to provide appropriate degeneracy factors for the integral
coefficients $I_{nq}^{(r)}$ and $J_{nq}^{(r)}$. For QGP, these
integral coefficients will transform as
\begin{align}\label{IJnqr}
I_{nq}^{(r)} &\to I_{nq}^{(r)}|_{g=g_q,a=1} + I_{nq}^{(r)}|_{g=g_g,a=-1}, \nonumber\\
J_{nq}^{(r)} &\to J_{nq}^{(r)}|_{g=g_q,a=1} + J_{nq}^{(r)}|_{g=g_g,a=-1},
\end{align}
where $g_q$ and $g_g$ are the degeneracy factors corresponding to
quarks and gluons respectively. These factors are given as
\begin{align}\label{degen}
g_q &= N_s \times N_{q\bar q} \times N_C \times N_f = 12N_f, \nonumber\\
g_g &= N_s \times \big(N_C^2 - 1\big) = 16,
\end{align}
where $N_f$ is the number of flavours, $N_C=3$ is the number of
colours, $N_s=2$ corresponds to spin degrees and $N_{q\bar q}=2$
denotes quark and anti-quark.
\begin{figure}[t]
\begin{center}
\includegraphics[width=\linewidth]{fig_0_B.pdf}
\end{center}
\vspace{-0.8cm}
\caption{(Color online) Time evolution of the the bulk viscous
pressure times $\tau$ (top) and the pressure anisotropy ${\cal
P}_L/{\cal P}_T$ scaled by that obtained using exact solution of
the RTA Boltzmann equation (bottom) for Boltzmann statistics.
The three curves in both panels correspond to three different
calculations: the exact solution of the RTA Boltzmann equation
\cite{Florkowski:2014sda} (red solid line), second-order viscous
hydrodynamics obtained using the Chapman-Enskog method (blue
dashed line) and using the Grad's 14-moment approximation (green
dotted line). For both panels we use $T_0=300$ MeV at
$\tau_0=0.5$ fm/c, $m=300$ MeV, and $\tau_{\rm
eq}=\tau_\pi=\tau_\Pi=0.25$ fm/c. The initial spheroidal
anisotropy in the distribution function, $\xi_0=0$, corresponds
to isotropic initial pressures with $\Pi_0=\pi_0=0$.}
\label{fig_0_B}
\end{figure}
In Fig.~\ref{ZetabyEta}, we show $m/T$ dependence of the ratio
$\zeta/\eta$ scaled by $(1/3-c_s^2)^2$ for various cases. We observe
that in accordance with the predictions from small-$z$ expansion,
Eq.~(\ref{zetabyeta}), the curve for Boltzmann statistics (red solid
line) and Fermi-Dirac statistics (green dotted line) saturates at
$75$ and $48$ (thin horizontal black dashed lines), respectively, at
small values of $z$. We also see that at very small-$z$,
Bose-Einstein statistics (blue dashed line) results in very large
values of the scaled viscosity ratio $(\zeta/\eta)/(1/3-c_s^2)^2$
indicating divergence. We see that this ratio for the QGP is
dominated by the Bose-Einstein statistics for two-flavor (brown
dashed-dotted line) as well as three-flavor QGP (purple
dashed-dotted-dotted line). This however does not imply that the
bulk viscosity of QGP is very large because a realistic calculation
with lattice QCD equation of state suggest that $z>0.6$ even for
temperatures as high as central RHIC and LHC energies \cite
{Romatschke:2011qp, Andersen:2011sf, Strickland:2014zka}. Moreover,
for large-$z$ we see that all three statistics lead to similar
results for the scaled viscosity ratio which suggests that there is
a relatively small effect coming from composition of the fluid.
\section{Boost-invariant 0+1d case}
In this section, we consider evolution in the case of transversely
homogeneous and purely-longitudinal boost-invariant expansion \cite
{Bjorken:1982qr}. For such an expansion all scalar functions of
space and time depend only on the longitudinal proper time
$\tau=\sqrt{t^2-z^2}$. Working in the Milne coordinate system,
$(\tau,x,y,\eta)$, the hydrodynamic four-velocity is given by
$u^\mu=(1,0,0,0)$. The energy-momentum conservation equation
together with Eqs.~(\ref {BULK}) and (\ref{SHEAR}) reduce to
\begin{align}
\dot\epsilon &= -\frac{1}{\tau}\left(\epsilon + P + \Pi -\pi\right) \, , \label{epsBj}\\
\dot\Pi + \frac{\Pi}{\tau_\Pi} &= -\frac{\beta_\Pi}{\tau} - \delta_{\Pi\Pi}\frac{\Pi}{\tau}
+\lambda_{\Pi\pi}\frac{\pi}{\tau} \, , \label{bulkBj}\\
\dot\pi + \frac{\pi}{\tau_\pi} &= \frac{4}{3}\frac{\beta_\pi}{\tau} - \left( \frac{1}{3}\tau_{\pi\pi}
+\delta_{\pi\pi}\right)\frac{\pi}{\tau} + \frac{2}{3}\lambda_{\pi\Pi}\frac{\Pi}{\tau} \, , \label{shearBj}
\end{align}
where $\pi\equiv-\tau^2\pi^{\eta\eta}$. Note that in this case the
term involving the vorticity tensor in Eq.~(\ref{SHEAR}) vanishes
and hence has no effect on the dynamics of the fluid. Also note that
the first terms on the right-hand side of Eqs.~(\ref{bulkBj}) and
(\ref {shearBj}) corresponds to the first-order terms
$\beta_\Pi\theta$ and $2\beta_\pi \sigma^{\mu\nu}$, respectively,
whereas the rest are of second-order.
\begin{figure}[t]
\begin{center}
\includegraphics[width=\linewidth]{fig_0_BE.pdf}
\end{center}
\vspace{-0.8cm}
\caption{(Color online) Same as Fig. \ref{fig_0_B} except here
we consider Bose-Einstein statistics.}
\label{fig_0_BE}
\end{figure}
\begin{figure}[t]
\begin{center}
\includegraphics[width=\linewidth]{fig_0_FD.pdf}
\end{center}
\vspace{-0.8cm}
\caption{(Color online) Same as Fig. \ref{fig_0_B} except here
we consider Fermi-Dirac statistics.}
\label{fig_0_FD}
\end{figure}
\begin{figure}[t]
\begin{center}
\includegraphics[width=\linewidth]{fig_100_B.pdf}
\end{center}
\vspace{-0.8cm}
\caption{(Color online) Same as Fig. \ref{fig_0_B} except here
we take $\xi_0=100$ corresponding to $\pi_0=2.86$ GeV/fm$^3$ and
$\Pi_0=0.138$ GeV/fm$^3$.}
\label{fig_100_B}
\end{figure}
\begin{figure}[t]
\begin{center}
\includegraphics[width=\linewidth]{fig_100_BE.pdf}
\end{center}
\vspace{-0.8cm}
\caption{(Color online) Same as Fig. \ref{fig_100_B} except here
we consider Bose-Einstein statistics.}
\label{fig_100_BE}
\end{figure}
\begin{figure}[t]
\begin{center}
\includegraphics[width=\linewidth]{fig_100_FD.pdf}
\end{center}
\vspace{-0.8cm}
\caption{(Color online) Same as Fig. \ref{fig_100_B} except here
we consider Fermi-Dirac statistics.}
\label{fig_100_FD}
\end{figure}
We simultaneously solve Eqs.~(\ref{epsBj})-(\ref{shearBj}) assuming
an initial temperature of $T_0=300$ MeV at the initial proper time
$\tau_0=0.5$ fm/c, with relaxation times $\tau_{\rm eq}=\tau_\Pi=
\tau_\pi=0.25$ fm/c corresponding to initial $\eta/s=1/4\pi$. We
solve these equations for two different initial pressure
configurations: $\xi_0=0$ corresponding to an isotropic pressure
configuration $\pi_0=\Pi_0=0$ and $\xi_0=100$ corresponding to a
highly oblate anisotropic pressure configuration. The anisotropy
parameter $\xi$ is related to the average longitudinal and
transverse momentum in the local rest frame as:
$\xi=\frac{1}{2}\langle p_T^2\rangle/\langle p_L^2\rangle-1$, where
$\langle\cdots\rangle \equiv \int d^3p f_0(\sqrt{p_T^2 + (1+\xi)
p_L^2},\Lambda)$ and $\Lambda$ is a temperature-like scale which can
be identified with the temperature of the system in the isotropic
equilibrium limit. For particle mass, we consider $m=300$ MeV which
roughly corresponds to the constituent quark mass. We solve
Eqs.~(\ref{epsBj})-(\ref {shearBj}) with transport coefficients
obtained using the Grad's 14-moment method,
Eqs.~(\ref {Gcoeff1})-(\ref{Gcoeff5}), as well as the Chapman-Enskog
method, Eqs.~(\ref{CEcoeff1})-(\ref {CEcoeff5}).
In Figs.~\ref{fig_0_B} -- \ref{fig_100_FD} we show the time
evolution of the bulk viscous pressure times proper-time (top) and
the pressure anisotropy ${\cal P}_L/{\cal P}_T\equiv
(P+\Pi-\pi)/(P+\Pi+\pi/2)$ scaled by that obtained using exact
solution of the RTA Boltzmann equation (bottom) for three different
calculations: the exact solution of the RTA Boltzmann equation \cite
{Florkowski:2014sda} (red solid line), second-order viscous
hydrodynamics obtained using the Chapman-Enskog method, Eqs.~(\ref
{CEcoeff1})-(\ref {CEcoeff5}), (blue dashed line) and using the
Grad's 14-moment approximation, Eqs.~(\ref{Gcoeff1})-(\ref
{Gcoeff5}), (green dotted line). Figures~\ref {fig_0_B} and \ref
{fig_100_B} are for Boltzmann statistics, Figs.~\ref{fig_0_BE} and
\ref{fig_100_BE} are for Bose-Einstein statistics, and Figs.~\ref
{fig_0_FD} and \ref{fig_100_FD} are for Fermi-Dirac statistics.
Figs.~\ref {fig_0_B} -- \ref {fig_0_FD} correspond to an isotropic
initial condition ($\xi_0=0$), while Figs.~\ref{fig_100_B} -- \ref
{fig_100_FD} correspond to a highly oblate anisotropic initial
condition ($\xi_0=100$).
From Figs.~\ref{fig_0_B} -- \ref{fig_100_FD}, we observe that,
compared to Grad's 14-moment approximation, the transport
coefficients obtained using the Chapman-Enskog method does a
marginally better job in reproducing the ${\cal P}_L/{\cal P}_T$
obtained using the exact solution of the RTA Boltzmann equation. On
the other hand, the result for $\tau\Pi$ obtained using the
Chapman-Enskog method shows better agreement with the exact solution
of the RTA Boltzmann equation than the Grad's 14-moment method.
\section{Conclusions and outlook}
In this paper we expressed the transport coefficients appearing in
the second-order viscous hydrodynamical evolution of a massive gas
using Bose-Einstein, Boltzmann and Fermi-Dirac statistics for the
equilibrium distribution function and Grad's 14-moment approximation
as well as the method of Chapman-Enskog expansion for the
non-equilibrium part. The second-order viscous evolution equations
are obtained by coarse graining the relativistic Boltzmann equation
in the relaxation-time approximation. We also obtained the ratio of
the coefficient of bulk viscosity to that of shear viscosity, in
terms of the speed of sound, for classical and quantum statistics as
well as for the QGP. We then considered the specific case of a
transversally homogeneous and longitudinally boost-invariant system
for which it is possible to exactly solve the RTA Boltzmann equation
\cite{Florkowski:2014sda}. Using this solution as a benchmark, we
compared the pressure anisotropy and bulk viscous pressure evolution
obtained by employing both the Chapman-Enskog method as well as the
Grad's 14-moment method. We demonstrated that the Chapman-Enskog
method is in better agreement with the exact solution of the RTA
Boltzmann equation compared to the Grad's 14-moment method. We found
that, while both methods give similar results for the pressure
anisotropy, the Chapman-Enskog method better reproduces the exact
solution for the bulk viscous pressure evolution.
At this juncture, we would like to clarify that we have used the
exact solution of the Boltzmann equation, in the relaxation-time
approximation, as a benchmark in order to compare different
hydrodynamic formulations. The relaxation-time approximation is
based on the assumption that the collisions tend to restore the
phase-space distribution function to its equilibrium value
exponentially. Although the microscopic interactions of the
constituent particles are not captured in this approximation, it is
reasonably accurate to describe a system which is close to local
thermodynamic equilibrium \cite{Dusling:2009df}. Looking forward, it
will be interesting to determine the impact of the quantum transport
coefficients, obtained herein, in higher dimensional simulations.
Moreover, it would also be instructive to see if the second-order
results derived herein could be extended to obtain third order
transport coefficients for quantum statistics \cite
{Jaiswal:2013vta}. We leave these questions for a future work.
\begin{acknowledgments}
A.J. acknowledges useful discussions with Gabriel Denicol, Bengt
Friman and Krzysztof Redlich. W.F. and E.M. were supported by Polish
National Science Center Grant No.~DEC-2012/06/A/ST2/00390. A.J. was
supported by the Frankfurt Institute for Advanced Studies (FIAS).
R.R. was supported by Polish National Science Center Grant
No.~DEC-2012/07/D/ST2/02125. M.S. was supported in part by U.S.~DOE
Grant No.~DE-SC0004104.
\end{acknowledgments}
|
\section{Introduction}\label{sec1} We consider the 1D scalar
conservation law associated to the conserved variable
$u(x,t)$,
\begin{eqnarray} \displaystyle \frac{\partial}{\partial t}
u(x,t) + \frac{\partial}{\partial x}f(u(x,t)) = 0,\;\; (x,\,t)\in
\mathbb{R}\times\,
\mathbb{R}^{+} \label{nonlin}\\ u(x,0)=u_{0}(x)\nonumber
\end{eqnarray}
where $f(u)$ is a non-linear flux function. The numerical
approximation for the solution of (\ref{nonlin}) is done by the
discretization of the spatial and temporal space into $N$ equispaced
cells $I_{i} = [x_{i-\frac{1}{2}}, x_{i+\frac{1}{2}}],\; i=0,1,\dots
N$ of length $\Delta x$ and $M$ equispaced intervals $[t^n,
t^{n+1}],\; n=0,1,\dots, M$ of length $\Delta t$ respectively.
Let $x_{i}=i\Delta x$ and $t^{n}=n\Delta t$ denote the cell center
of cell $I_{i}$ and the $n^{th}$ time level respectively then a
conservative numerical approximation for (\ref{nonlin}) can be defined
by \begin{equation} {u}^{n+1}_{i} = {u}_{i}^{n} - \lambda
\left(\mathcal{F}^n_{i+\frac{1}{2}}-
\mathcal{F}^n_{i-\frac{1}{2}}\right),\;\;\lambda = \frac{\Delta
t}{\Delta x}. \label{s2eq2} \end{equation} where
${u^n_{i}=u(x_{i},t^{n})}$ and $\mathcal{F}^n_{i\pm
\frac{1}{2}}$ is the numerical flux function defined at the cell
interface $x_{i \pm \frac{1}{2}}$ at time level $n$. The characteristics speed $a(u)= \frac{\partial
f(u)}{\partial u}$ associated with (\ref{nonlin}) can be approximated as,
\begin{equation}
\displaystyle a^n_{i+\frac{1}{2}}= \left\{\begin{array}{clc}
\displaystyle \frac{F^n_{i+1}-F^n_{i}}{{u}^n_{i+1} - {u}^n_{i}}& \mbox{if}
&{u}^n_{i+1} \neq {u}^n_{i},\\ & &\\ \displaystyle
\left. \frac{\partial f}{\partial u}\right|_{{u}^n_{i}} & \mbox{if}
&{u}^n_{i+1}=
{u}^n_{i} \end{array}\right.,\label{speed}
\end{equation}
where $F_{i}= f(u^n_{i})$. In general due to non-linearity of
(\ref{nonlin}), beyond a small finite time, even for a smooth initial
data the evolution of discontinuities in the solution $u(x,t)$ is
inevitable. Therefore, it is required to have a conservative
approximation of the solution with high accuracy and crisp resolution
of such discontinuities with out numerical oscillations. Contrary to
this need, most classical high order schemes despite of being linearly
Von-Neumann stable give oscillatory approximation for discontinuities
even for the trivial case of transport equation i.e., $f(u)=au,\,
0\neq a\in R $. Such oscillatory approximation can not be considered
as admissible solution since it violets the following global maximum
principle satisfied by the physically correct solution $u(x,t)$ of
(\ref{nonlin}) i.e.,
\begin{equation} \min(u_{0}(x)) \leq u(x,t) \leq \max(u_{0}(x)),
\forall (x,\;t)\in \mathbb{R}\times\, \mathbb{R}^{+}.
\label{mp}
\end{equation} In
order to overcome these undesired numerical instabilities, various
notion of non-linear stability are developed in the light of maximum
principle (\ref{mp}). Examples of Maximum principle satisfying schemes
are monotone schemes \cite{vanLeer1974, Crandall}, total variation
diminishing (TVD) schemes
\cite{harten1984,sweby1984,Yee1987,sanders1988,davis1987,
goodman1988,rkd2007,zhang2010}. Some uniformly high order
maximum-principle satisfying and positivity preserving schemes are
\cite{LaxLiu1998, zhangjcp2010,zhang2011}. There are other
non-oscillatory schemes which do not strictly follow maximum principle
but practically give excellent numerical results e.g., Essentially
non-oscillatory (ENO) and weighted ENO schemes see \cite{cwshu1999}
and references therein. It is known that among global maximum
principle satisfying schemes, the monotone and total variation
diminishing (TVD) schemes experience difficulties at data extrema. On
the one hand, such high order schemes locally degenerate to first
order accuracy at non-sonic data extrema and on the other hand, even
such a uniformly first order accurate schemes may exhibit induced
local oscillations at data extrema. In this work the focus is on the
construction of improved TVD schemes at smooth data extrema.
\section{Global maximum principle and data extrema}\label{sec2}
The above global maximum principle (\ref{mp}) satisfying monotone
and TVD schemes have been of great interest mainly due to excellent
convergence proofs for entropy solution
\cite{Sanders1983,Lefloach99} and \cite{chakra,yang1996}
respectively. The key idea is, any maximum principle satisfying
scheme produce a bounded solution sequence and convergence follows
due to compactness of solution sequence space \cite{LeVeque1992}. It
can be shown that monotone stable scheme $\Rightarrow$ TVD scheme
$\Rightarrow$ monotonicity preserving scheme (or Local extremum
diminishing (LED)) scheme \cite{harten1983,jameson1995}. Unfortunately,
monotone as well TVD schemes experience difficulty at data
extrema. The monotone stability relies on monotone data and
therefore a monotone scheme preserves the monotonicity of a data set
by mapping it to a new monotone data set but fails to preserve the
non-monotone solution region i.e., at data extrema. These monotone
schemes are criticized mainly due to barrier theorem which state
that a 'linear' three point monotone scheme can be at most first
order accurate \cite{godunov}. Later, second order 'non-linear'
conservative monotone schemes are constructed using limiters but
again by compromising on second order accuracy at extrema,
e.g. \cite{vanLeer1974}. The TVD stability mimics the maximum
principle as it relies on the condition that global extremum values
of solution must remain be bounded by global extremum values of
initial solution. In \cite{harten1983}, Harten gave the concept of
total variation diminishing scheme by measuring the variation of the
grid values as follows
\begin{definition}\label{def1} Conservative scheme
(\ref{s2eq2}) is called total variation diminishing if
\begin{equation}\label{tvddef} \displaystyle
\sum_{i=-\infty}^{\infty}\left|\Delta_{-}u_{i}^{n+1}\right|\leq
\sum_{i=-\infty}^{\infty} \left|\Delta_{-}u_{i}^{n}\right|
\end{equation} where
$\Delta_{+}{u}_{i} = \Delta_{-}{u}_{i+1} = {u}_{i+1} -{u}_{i}$.
\end{definition}
Note that that the definition \ref{def1} is global as it is defined on the whole
computational domain and ensures that global maxima or minima of
initial solution $u_{0}(x)$ will not increase or decrease
respectively. Such conservative TVD schemes are heavily
criticized because, even if they are higher order accurate in most
solution region, they give up second order of accuracy at non-sonic
critical values of the solution \cite{tadmor1988,chakra}\footnote{Sanders also defined the total
variation by measuring the variation of the reconstructed
polynomials and such TVD schemes can be uniformly high order
accurate \cite{sanders1988,zhang2010}.}. {\it We emphasize that these depressing results on
degeneracy of accuracy of TVD method are given for {\bf
conservative} schemes and in the above {\bf global sense}}. More
precisely the global nature of TVD definition (\ref{tvddef}) allows
shift in indices technique in $\sum$ sign and is extensively used
in different terms of the infinite sums in the TVD proofs of various
schemes and results in the literature including the following one
due to Harten \cite{harten1983}.
\begin{lemma}\label{lem1} A conservative scheme in
Incremental form (I-form)
\begin{equation}
{u}_{i}^{n+1}=u_{i}^{n}+\alpha_{i+\frac{1}{2}}({u}^{n}_{i+1}-{u}^{n}_{i})-
\beta_{i-\frac{1}{2}}({u}^{n}_{i}-{u}^{n}_{i-1})\label{iform}
\end{equation}
is TVD iff $\alpha_{i+\frac{1}{2}}\geq0, \beta_{i+\frac{1}{2}}\geq
0\;\mbox{and}\; \alpha_{i+\frac{1}{2}}+ \beta_{i+\frac{1}{2}}\leq
1,\; \forall i$. \cite{Thomas}
\end{lemma}
\subsection{Degenerate accuracy at extrema:}
In \cite{chakra}, proof for degeneracy to first order accuracy at
non-sonic critical points of solution i.e., points $u^{*}(x,t)$
s.t. $f^{'}(u^{*}(x,t)) \neq 0= u^{*}_{x}(x,t)$ is mainly based on
modified equation analysis and a {\it conservative} semi-discrete
version of Lemma \ref{lem1}. In \cite{tadmor1988}, using a trade off
between second order accuracy and TVD requirement along with shift in
indices technique, it is shown that second order accuracy must be
given up by a {\it conservative} TVD scheme at non sonic critical
values $u_{i}=u^{*}(x_{i},t)$ which corresponds to extreme values i.e.,
$[u(x_{i}+\Delta\,x,t)-
u(x_{i},t)].[u(x_{i},t)-u(x_{i}-\Delta\,x,t)]<0 \neq f^{'}(u_{i})$.
It is also worthy to note that problem of degenerate accuracy by
modern high resolution TVD schemes is also due to their construction
procedure. For example, the numerical flux function of flux limiters or
slope limiters based TVD schemes is essentially design in such a way
that it reduces to first order accuracy at extrema and high
gradient region by forcing limiter function to be zero see
\cite{Piperno,Dubey2013} and references therein. This makes it
impossible for a limiter based TVD schemes to achieve higher than
first order accuracy at solution extrema as well at steep gradient
region \cite{LeVeque1992}. Thus every high order TVD (in global sense
(\ref{tvddef}) ) scheme suffers from clipping error and cause flatten
approximation for smooth extrema though they sharply capture
discontinuities\cite{Laney}.
\subsection{Induced local oscillations:} Apart from compromise in uniform high accuracy, it is notable that
global maximum principle satisfying monotone and TVD schemes do not
necessarily ensure preservation of non-monotone data set i.e., for a
data set with extrema as demonstrated in Figure \ref{F1a}(b). In
particular first order monotone and TVD schemes with {\bf large
coefficient of numerical viscosity} can allow the occurrence of
induced oscillations at data extrema and formation of new local
extremum values as shown in Figure \ref{F1b}(a). This phenomena of
generation of local oscillations at extrema is reported and analyzed
for well known monotone and TVD three point {\it Lax-Friedrichs}
scheme in \cite{Breuss2010,jiequan2009,jiequan2011} similar to
Figure \ref{F1b}(a).
\begin{figure}[h]
\begin{tabular}{cc}
\hspace{-0cm} \includegraphics[%
scale=0.3]{monotonedata.pdf} & \includegraphics[%
scale=0.3]{nonmonotonedata.pdf}\\
(a) & (b)
\end{tabular}
\caption{\label{F1a} Induced oscillations may occur for
non-monotone data extrema. (a) Monotone stability rely on
monotone data sequence. (b) For non-monotone data set TVD
definition (\ref{tvddef}) is satisfied though updated approximation
is oscillatory.}
\end{figure}
\begin{figure}[h]
\begin{tabular}{cc}
\includegraphics[scale=0.45]{LxFimpulse.pdf} &
\includegraphics[scale=0.45]{Upimpulse.pdf}\\ (a) &
(b) \end{tabular} \caption{\label{F1b} Numerical
approximation of Linear transport equation (\ref{transport})
with impulsive initial data. (a) By monotone and TVD
Lax-Friedrichs scheme: Induced oscillations (b) First order two
point upwind scheme: Absence of induced oscillations.}
\end{figure}
It is interesting to note that two point monotone and TVD upwind
scheme does not exhibit induced oscillations in Figure
\ref{F1b}(b).
The main aim of this work is to construct uniformly non-oscillatory
shock capturing monotone and TVD methods with high accuracy for
non-sonic smooth extrema\footnote{It shows improved TVD
approximation in the region of degenerate accuracy reported in
\cite{chakra,tadmor1988}.}. In order to achieve it, a {\bf
non-conservative} formulation is done using framework of local
maximum principle (LMP) with the help of gradient ratio parameter.
The rest of paper is organized as: For completeness, in section
\ref{sec3}, local maximum principle (LMP) stability is defined for
two points schemes. It is shown in Lemma \ref{lem2}, that away from
sonic point LMP stability implies global monotone and TVD
stability. In section \ref{sec4}, we analyze representative
uniformly second order schemes in non-sonic region for
their TVD (or LMP) stability by converting them into two point
schemes. This yields computable bounds for the stability of these
scheme and are presented as main results in Theorems
\ref{thm1}-\ref{thm3}. These obtained TVD bounds ensure for {\it
second rrder TV stable approximation for smooth
solution with non-sonic critical point of extreme nature}. In
section \ref{sec5}, hybrid schemes are designed using the TVD
bounds and a shock detector technique. Numerical results are
given to support the theoretical results and claim. Conclusion and
future work is discussed in section \ref{sec6}.
\section{Local maximum principle (LMP) stability}\label{sec3}
It is clear from the above discussion that the global maximum principle
(\ref{mp}) satisfying monotone or TVD stability experience
difficulty in the presence of non-monotone data with extrema. These
two local phenomena i.e, induced oscillations and degeneracy in accuracy
at non-sonic extrema by monotone and TVD schemes motivate us to look
for their non-linear stability locally at non-sonic extrema. Consider the
following local maximum principle (LMP)\footnote{Also known as
upwind range condition \cite{Laney}} for scalar conservation law
(\ref{nonlin}), \begin{subequations} \label{urc} \begin{align}
\min_{x_{i-1}\leq x \leq x_{i}}u(x,t^{n})\leq u(x,t^{n+1})\leq
\max_{x_{i-1}\leq x \leq x_{i}}u(x,t^{n})&\;\; \mbox{if}\;
f^{'}(u) >0, \\ \min_{x_{i}\leq x \leq x_{i+1}}u(x,t^{n})\leq
u(x,t^{n+1})\leq \max_{x_{i}\leq x \leq x_{i+1}}u(x,t^{n})&\;\;
\mbox{if}\;f^{'}(u)<0. \end{align} \end{subequations} In case
of two point schemes, initial solution data will always be monotone
as either $u(x_{a},t)\leq u(x_{b},t),\; x_{a}\neq x_{b}$ or vice
verse thus the LMP condition (\ref{urc}) reduces to
\begin{subequations} \label{urc1} \begin{align} \min(u_{i-1}^{n},
u_{i}^{n}) \leq u_{i}^{n+1} \leq \max(u_{i-1}^{n}, u_{i}^{n})
&\;\; \mbox{if} f^{'}(u)>0,\\ \min(u_{i}^{n}, u_{i+1}^{n}) \leq
u_{i}^{n+1} \leq \max(u_{i}^{n}, u_{i+1}^{n}) &\;\; \mbox{if}\;
f^{'}(u)< 0. \end{align} \end{subequations} Thus away from
sonic point $u_{i}^{*}$ i.e., $a_{i+\frac{1}{2}}=f^{'}(u_{i}^{*})\neq0$ define,
\begin{definition}\label{def2} A numerical scheme is LMP stable (\ref{urc1}) if it can be written
as \begin{equation} \label{lmpscheme} u_{i}^{n+1} =
\left\{ \begin{array}{ll} \mathcal{C} u_{i}^{n} + \mathcal{D}
u_{i-1}^{n} &\;\; \mbox{if}\;\;\;\;\;\;\;\;\; 0< \lambda
a^n_{i+\frac{1}{2}}\leq 1,\\ \mathcal{C} u_{i}^{n} +
\mathcal{D} u_{i+1}^{n} &\;\; \mbox{if} \;\;\;\;\; -1\leq
\lambda a^n_{i+\frac{1}{2}}< 0, \end{array}\right.
\end{equation} where
coefficients $\mathcal{C}$and $\mathcal{D}$ are real functions such
that $\mathcal{C}\geq 0, \mathcal{D} \geq 0$, $\mathcal{C} +
\mathcal{D} = 1$.
\end{definition}
Scheme (\ref{lmpscheme}) is essentially a convex combination of two
point values of $u(:,t^{n})$ thus ensures that the updated solution
value of $u(x_{i}, t^{n+1})$ will remain be bounded by both point values
without introducing of new local maxima-minima. Note that two point
first order upwind scheme is a natural example of LMP stable scheme
with coefficients
\begin{subequations} \begin{align} \mathcal{C} = 1-\lambda
a_{i-\frac{1}{2}}, \mathcal{D} = \lambda a_{i-\frac{1}{2}} &\;\;
\mbox{if}\;\;\;\;\; 0< \lambda f^{'}(u_{i}^{n})\leq 1\\
\mathcal{C} = 1-\lambda a_{i+\frac{1}{2}}, \mathcal{D} = \lambda
a_{i+\frac{1}{2}} &\;\; \mbox{if}\;\;\;\;\; -1\leq \lambda
f^{'}(u_{i}^{n})< 0. \end{align}
\end{subequations}
This justifies the non-occurrence of local oscillation by first order
upwind scheme in Figure \ref{F1b}(b). From definition \ref{def2}
it follows,
\begin{lemma}\label{lem2} Local maximum principle stable
scheme (\ref{lmpscheme}) is global total variation diminishing stable.
\end{lemma}
\begin{proof} Using relation $\mathcal{C} =1-\mathcal{D}$, rewrite
(\ref{lmpscheme}) in the {\bf non-conservative}\footnote{Note that for problems with constant $f'(u)$ e.g. linear transport
equation (\ref{transport}), the form (\ref{lmptvd}) is
conservative. Also one can obtained a conservative approximation
from (\ref{lmptvd}) by defining $\mathcal{D}$ suitably such as
$\displaystyle \mathcal{D}=
\frac{\Delta_{-}f(u^{n}_{i})}{\Delta_{-}u_{i}^{n}},\; 0<\lambda
f'(u_{i}^{n})\leq 1$ as in \cite{rkdejde1}. In the persent work the
coefficient $\mathcal{D}$ comes out in such a form which results
in to a non-conservative approximation.} half incremental form as,
\begin{equation} \label{lmptvd} u_{i}^{n+1} =
\left\{ \begin{array}{ll} u_{i}^{n} - \mathcal{D}
(u_{i}^{n}-u_{i-1}^{n}) &\;\; \mbox{if}\;\;\;\;\;\;\;\;\; 0<
\lambda a^n_{i+\frac{1}{2}}\leq 1,\\ u_{i}^{n}+\mathcal{D}
(u_{i+1}^{n} - u_{i}^{n}) &\;\; \mbox{if} \;\;\;\;\; -1\leq
\lambda a^n_{i+\frac{1}{2}}< 0, \end{array}\right.
\end{equation}
where from Definition \ref{def2}, $0\leq \mathcal{C}\leq 1$ and
$0\leq \mathcal{D}\leq 1$. On appropriately choosing one of the
coefficients $\alpha$ or $\beta$ zero in I-form (\ref{iform}), the
Lemma \ref{lem1} shows approximation by half I-form (\ref{lmptvd})
is global TVD.
\end{proof}
\begin{remark} The LMP stability is defined only in non-sonic region
therefore by Lemma \ref{lem2}, LMP stability implies TVD stability
in non-sonic region i.e., away from sonic point $f^{'}(u^{*}) \neq 0$.
In this setting, from next section onward until stated the term LMP/TVD
stability implies global TVD stability (\ref{tvddef}) {\bf away
from sonic point}.
\end{remark}
\section{Bounds on high order TVD accuracy}\label{sec4}
In this section using definition \ref{def2}, it is shown that second
order total variation diminishing approximation is possible for the
solution with non-sonic extreme critical points. It follows from the
LMP stability bounds given for the representative second order
accurate Lax-Wendroff (LxW), Beam-Warming (BW) and Fromm schemes
respectively for scalar problem (\ref{nonlin}). Let the stencil
$[x_{i-s},x_{i+r}]$ of $r+s+1$ point scheme locally does not contain
sonic point $u^{*}(x,t)$ i.e., $f^{'}(u^*)\neq0$ and characteristics
speed at local cell interfaces is non-zero. Note that the case of
degenerate characteristic speed i.e., $a_{i+\frac{1}{2}}=0$ or/and
$a_{i-\frac{1}{2}}=0$ are not interesting as these schemes do not
necessarily preserve their uniform order of accuracy. We also
consider the wave speed slpit $a^{+}_{i+\frac{1}{2}} +
a^{-}_{i+\frac{1}{2}}= a_{i+\frac{1}{2}}$ such that
$a^{+}_{i+\frac{1}{2}}\geq0$ and $a^{-}_{i+\frac{1}{2}}\leq 0$. Note
that for $a_{i+\frac{1}{2}}>0 \Rightarrow
a_{i+\frac{1}{2}}=a^{+}_{i+\frac{1}{2}}>0, a^{-}_{i+\frac{1}{2}}=0$
whereas $a_{i+\frac{1}{2}}<0 \Rightarrow
a_{i+\frac{1}{2}}=a^{-}_{i+\frac{1}{2}}<0, a^{+}_{i+\frac{1}{2}}=0,$.
After dropping the superscript for time level $n$, following function
definitions and notations are used in the rest of the presentation.
Define the smoothness parameter as
\begin{equation}
\displaystyle r^{\pm}_{i} = r^{\pm}(F_{i})= \frac{\left(1\mp \lambda a^{\pm}_{i\mp\frac{1}{2}}\right)\Delta_{\mp}F^{\pm}_{i}}{\left(1 \mp
\lambda a^{\pm}_{i\pm\frac{1}{2}}\right)\Delta_{\pm}F^{\pm}_{i}},\label{gradratio}
\end{equation}
where the flux split $F^{+}_{i}+F^{-}_{i}=F_{i}$ is consistent with wave split and given by
\begin{equation}
F^{\pm}_{i+1} -F^{\pm}_{i}= a^{\pm}_{i+\frac{1}{2}}(u_{i+1}-u_{i}). \label{fluxsplit}
\end{equation}
Here the superscript $\pm$ sign of $r_{i}$ denotes the
positive/negative sign of wave speed. Also define the signum function
\begin{equation}
\sigma(x) = sgn(x) =\left\{\begin{array}{ll} +1 & \;if\; x\geq0,\\-1 & \;if\; x< 0.\end{array}\right.
\end{equation}
In order to analyze the local non-linear
stability of considered schemes we choose practically viable CFL like
condition
\begin{equation}
0<\lambda \max_{u}|f^{'}(u)|<1 \label{cflNo}
\end{equation}
Note that the choice $\displaystyle \lambda \max_{u}|f^{'}(u)|=0\Rightarrow f^{'}(u)$
corresponds to the case of degenerate characteristic speed or steady
state case.
\subsection{Centered Lax-Wendroff scheme}
\begin{theorem}\label{thm1}
Away from sonic point and under CFL condition (\ref{cflNo}), the
second order accurate Lax-Wendroff scheme for scalar conservation
law (\ref{nonlin}) is TVD in the solution region where
\[\displaystyle
r^{\pm}_{i}\in \left(-\infty,\, \kappa_{1}\left(\lambda
a^{\pm}_{i\mp\frac{1}{2}}\right)\right)\cup
\left(\gamma_{1}\left(\lambda
a^{\pm}_{i\mp\frac{1}{2}}\right),\,\infty\right)\]
where numbers $\kappa_{1} < 0$, $\gamma_{1}> 0$ depends on CFL number
for linear stability given by
\begin{equation}
\kappa_{1}(x) = -\frac{1-
x\sigma(x)}{1+x\sigma(x))}, \label{kappa1}
\end{equation}
\begin{equation}
\gamma_{1}(x) = \frac{x \sigma(x)}{2+ x\sigma(x)}.
\label{gamma1}
\end{equation}
\end{theorem}
\begin{proof}
Consider the numerical flux function of Lax-Wendroff (LxW) scheme
\begin{equation}
F^{n,LxW}_{i+ \frac{1}{2}}= \frac{1}{2}\left(F_{i+1} + F_{i}\right) -
\frac{\lambda\,a^2_{i+ \frac{1}{2}}}{2} \Delta_{+} {u}_{i}. \label{flxw}
\end{equation}
In order to ensure non sonic region, let the characteristics speed is
locally non-zero. Since LxW uses three point centred stencil $[x_{i-s},x_{i+r}],
r=s=1$, it suffice to assume $a_{i+\frac{1}{2}}\times a_{i-\frac{1}{2}}>0.$
\begin{itemize}
\item[] {\bf Case $f^{'}({u})> 0$:} Let $ a_{i \pm \frac{1}{2}}> 0$,
then the conservative approximation using (\ref{flxw}) can be
written as
\begin{equation} {u}^{n+1}_{i}= {u}_{i} - \left[\frac{\lambda
a_{i+\frac{1}{2}}}{2}\left(1-\lambda\,
a_{i+\frac{1}{2}}\right) \Delta_{+}{u}_{i}+ \frac{\lambda\,
a_{i-\frac{1}{2}}}{2}\left(1+\lambda\,
a_{i-\frac{1}{2}}\right) \Delta_{-}{u}_{i} \right],
\end{equation}
which can be written in the following non-conservative half Incremental form (\ref{lmptvd}),
\begin{equation} {u}^{n+1}_{i}= {u}_{i} - \left[\frac{\lambda
a_{i+\frac{1}{2}}}{2}\left(1-\lambda\,
a_{i+\frac{1}{2}}\right) \frac{\Delta_{+}{u}_{i}}
{\Delta_{-}{u}_{i}} +\frac{\lambda\,
a_{i-\frac{1}{2}}}{2}\left(1+\lambda\,
a_{i-\frac{1}{2}}\right) \right]
\Delta_{-}{u}_{i}. \label{Iflxw1}
\end{equation}
From Lemma \ref{lem2}, half I-from (\ref{Iflxw1}) will be TVD if,
\begin{equation} 0\leq \left[\lambda
a_{i+\frac{1}{2}}\left(1-\lambda\, a_{i+\frac{1}{2}}\right)
\frac{\Delta_{+}{u}_{i}} {\Delta_{-}{u}_{i}} +\lambda\,
a_{i-\frac{1}{2}}\left(1+\lambda\, a_{i-\frac{1}{2}}\right)
\right] \leq 2,
\end{equation}
which reduces to,
\begin{equation}
- \lambda\, a_{i-\frac{1}{2}}\left(1+\lambda\,
a_{i-\frac{1}{2}}\right) \leq \lambda\, a_{i+\frac{1}{2}}\left(1-\lambda\,
a_{i+\frac{1}{2}}\right) \frac{\Delta_{+}{u}_{i}}
{\Delta_{-}{u}_{i}} \leq 2- \lambda\, a_{i-\frac{1}{2}}\left(1+\lambda\,
a_{i-\frac{1}{2}}\right)\label{inq1}
\end{equation}
Note that $\lambda a_{i-\frac{1}{2}}(1-\lambda\,
a_{i-\frac{1}{2}})>0$ under discrete CFL condition,
\begin{equation}
0<\lambda\, \max_{i}{a_{i+\frac{1}{2}}} < 1, \label{cfl1}
\end{equation}
Hence inequality (\ref{inq1}) can be written as
\begin{equation}
-\frac{(1+\lambda\, a_{i-\frac{1}{2}})}{(1-\lambda\, a_{i-\frac{1}{2}})}
\leq
\frac{a_{i+\frac{1}{2}}(1-\lambda\,
a_{i+\frac{1}{2}})}{a_{i-\frac{1}{2}}(1-\lambda\,
a_{i-\frac{1}{2}})} \frac{\Delta_{+}{u}_{i}}
{\Delta_{-}{u}_{i}} \leq
\frac{2-\lambda a_{i-\frac{1}{2}}(1+\lambda\, a_{i-\frac{1}{2}})}
{\lambda\, a_{i-\frac{1}{2}}(1-\lambda\, a_{i-\frac{1}{2}})}
\end{equation}
or
\begin{equation}
-\frac{(1+\lambda\, a_{i-\frac{1}{2}})}{(1-\lambda\, a_{i-\frac{1}{2}})}
\leq
\frac{a_{i+\frac{1}{2}}(1-\lambda\,
a_{i+\frac{1}{2}})}{a_{i-\frac{1}{2}}(1-\lambda\,
a_{i-\frac{1}{2}})} \frac{\Delta_{+}{u}_{i}}
{\Delta_{-}{u}_{i}} \leq
\frac{2+\lambda\, a_{i-\frac{1}{2}}}{\lambda a_{i-\frac{1}{2}}},
\label{inq2b}
\end{equation}
Using definition (\ref{speed}) and flux wave split (\ref{fluxsplit}), Inequality (\ref{inq2b}) becomes
\begin{equation}
-\frac{(1+\lambda\, a^{+}_{i-\frac{1}{2}})}{(1-\lambda\, a^{+}_{i-\frac{1}{2}})}
\leq
\frac{(1-\lambda\,
a^{+}_{i+\frac{1}{2}})}{(1-\lambda\,
a^{+}_{i-\frac{1}{2}})} \frac{\Delta^{+}_{+}{F^{+}}_{i}}
{\Delta^{+}_{-}{F^{+}}_{i}} \leq \frac{2+\lambda\, a^{+}_{i-\frac{1}{2}}}{\lambda a^{+}_{i-\frac{1}{2}}}. \label{inq2c}
\end{equation}
Inequality (\ref{inq2c}) on inversion yields,
\begin{equation}
r^{+}_{i}< \kappa^{+}_{1}(\lambda\, a^{+}_{i-\frac{1}{2}})
\; \mbox{OR}\; r^{+}_{i} >
\gamma^{+}_{1}(\lambda a^{+}_{i-\frac{1}{2}}), \label{inq2d}
\end{equation}
where $$\displaystyle r^{+}_{i}
=\frac{(1-\lambda\, a^{+}_{i-\frac{1}{2}})
\Delta^{+}_{-}{F}_{i}}{(1-\lambda\,
a^{+}_{i+\frac{1}{2}})\Delta^{+}_{+}{F}_{i}},\; \kappa^{+}_{1}(\lambda
a^{+}_{i-\frac{1}{2}})= -\frac{(1-\lambda\,
a^{+}_{i-\frac{1}{2}})}{(1+\lambda\, a^{+}_{i-\frac{1}{2}})}\;
\mbox{and}\;\displaystyle \gamma^{+}_{1}(\lambda
a^{+}_{i-\frac{1}{2}})=\frac{\lambda\, a^{+}_{i-\frac{1}{2}}}{2+\lambda a^{+}_{i-\frac{1}{2}}}.$$\\
\item[] {\bf Case ${f^{'}(u)<0}$:} Let $a_{i\pm\frac{1}{2}}<0$, then the non-conservative I-form can be written
as
\begin{equation} {u}^{n}_{i} = {u}_{i} - \left[\frac{\lambda\,
a_{i+\frac{1}{2}}}{2}\left(1 -\lambda\,
a_{i+\frac{1}{2}}\right) + \frac{\lambda\,
a_{i-\frac{1}{2}}}{2}\left(1+\lambda\,
a_{i-\frac{1}{2}}\right)\frac{\Delta_{-}{u}_{i}}
{\Delta_{+}{u}_{i}}\right] \Delta_{+}{u}_{i}.\label{Iflxw2}
\end{equation}
Using Lemma \ref{lem2}, half I-from (\ref{Iflxw2}) will be TVD if,
\begin{equation}
0 \leq -\frac{\lambda\,
a_{i+\frac{1}{2}}}{2}\left(1 -\lambda\, a_{i+\frac{1}{2}}\right)
- \frac{\lambda\,
a_{i-\frac{1}{2}}}{2}\left(1+\lambda\,
a_{i-\frac{1}{2}}\right)\frac{\Delta_{-}{u}_{i}}
{\Delta_{+}{u}_{i}}\leq 1
\end{equation}
which can be written as
\begin{equation}
\lambda\,
a_{i+\frac{1}{2}}\left(1 -\lambda\, a_{i+\frac{1}{2}}\right)
\leq - \lambda\,
a_{i-\frac{1}{2}}\left(1+\lambda\,
a_{i-\frac{1}{2}}\right)\frac{\Delta_{-}{u}_{i}}
{\Delta_{+}{u}_{i}}\leq 2 + \lambda\,
a_{i+\frac{1}{2}}\left(1 -\lambda\, a_{i+\frac{1}{2}}\right). \label{inq3}
\end{equation}
The discrete CFL condition for $a(u)<0$ is,
\begin{equation}
-1 < \lambda a_{i+\frac{1}{2}}<0,\, \forall i.\label{cfl2}
\end{equation}
Therefore the quantity $-\lambda a_{i+\frac{1}{2}}(1+ \lambda
a_{i+\frac{1}{2}})>0.$ Divide Inequality (\ref{inq3}) by it and
using (\ref{speed}) yields,
\begin{equation}
-\frac{\left(1 -\lambda\, a_{i+\frac{1}{2}}\right)}{\left(1+ \lambda a_{i+\frac{1}{2}}\right)}
\leq \frac{\left(1+\lambda\,
a_{i-\frac{1}{2}}\right)}{\left(1+ \lambda a_{i+\frac{1}{2}}\right)}\frac{\Delta_{-}F_{i}}
{\Delta_{+}F_{i}}\leq -\frac{2 + \lambda\,
a_{i+\frac{1}{2}}\left(1 -\lambda\,
a_{i+\frac{1}{2}}\right)}{\lambda a_{i+\frac{1}{2}}\left(1+ \lambda
a_{i+\frac{1}{2}}\right)}
\end{equation}
or using flux wave split (\ref{fluxsplit})
\begin{equation}
-\frac{\left(1 -\lambda\, a^{-}_{i+\frac{1}{2}}\right)}{\left(1+ \lambda a^{-}_{i+\frac{1}{2}}\right)}
\leq \frac{\left(1+\lambda\,
a^{-}_{i-\frac{1}{2}}\right)}{\left(1+ \lambda a^{-}_{i+\frac{1}{2}}\right)}\frac{\Delta_{-}F^{-}_{i}}
{\Delta_{+}F^{-}_{i}}\leq \frac{\lambda\,a^{-}_{i+\frac{1}{2}}-2}{\lambda\;a^{-}_{i+\frac{1}{2}}}
\label{inq3a}
\end{equation}
Inequality (\ref{inq3a}) on inversion yields,
\begin{equation}
r^{-}_{i}<\kappa_{1}\left(\lambda a^{-}_{i+\frac{1}{2}}\right)\;
\mbox{OR}\; r^{-}_{i} > \gamma_{1}\left(\lambda a^{-}_{i+\frac{1}{2}}\right). \label{inq3b}
\end{equation}
where $\displaystyle r_{i}^{-}=\frac{\left(1+ \lambda
a^{-}_{i+\frac{1}{2}}\right)\Delta_{+}F^{-}_{i}}{\left(1+\lambda\,
a^{-}_{i-\frac{1}{2}}\right)\Delta_{-}F^{-}_{i}},\;
\kappa_{1}\left(\lambda a^{-}_{i+\frac{1}{2}}\right)=
\frac{-\left(1+\lambda a^{-}_{i+\frac{1}{2}}\right)}{\left(1-\lambda
a^{-}_{i+\frac{1}{2}}\right)}$ and $\displaystyle
\gamma_{1}\left(\lambda a^{-}_{i+\frac{1}{2}}\right)= \frac{-\lambda
a^{-}_{i+\frac{1}{2}}}{2 - \lambda\, a^{-}_{i+\frac{1}{2}}}$.\\
Condition (\ref{inq2d}) and (\ref{inq3b}) completes the proof.
\end{itemize}
\end{proof}
\subsubsection{LxW on Linear problem: Every extrema is non-sonic.}
In order to see the improvement in the TVD approximation at non-sonic
extrema, consider the linear transport equation
\begin{equation}
\frac{\partial}{\partial t}u(x,t) + a\frac{\partial}{\partial u}(\,u(x,t)) =0,\; a\neq0 \label{transport}
\end{equation}
In this case the smoothness parameter (\ref{gradratio}) reduces to $
r^{\pm}_{i} = \displaystyle \frac{\Delta_{\mp} u_{i}}{\Delta_{\pm}
u_{i}}$. Note that at point of extrema the measure of smoothness is
negative i.e, for transport equation (\ref{transport}), every
$r^{\pm}_{i} <0$ implies a {\it non-sonic extreme critical
point}. Following result follows from Theorem \ref{thm1}
\begin{cor}\label{cor1}
Under the linear stability condition $0<\lambda |a|\leq 1$, the
second order accurate Lax-Wendroff scheme for (\ref{transport}) is
total variation diminishing where $r_{i} \in
(\infty,\,\kappa_{1}(a\lambda))\cup(\gamma_{1}(a\lambda),\,\infty)$.
\end{cor}
\begin{figure}[!htb]
\begin{center}
\begin{tabular}{cc}
\includegraphics[%
scale=0.55]{LxWTVDregion.pdf}&\includegraphics[%
scale=0.55]{LxWTVDregionzoom.pdf}\\
(a) & (b)
\end{tabular}
\caption{\label{kappagammaLxW}(a) : LMP/TVD stability region (shaded) for
Lax-Wendroff scheme (b): Zoomed view.}
\end{center}
\end{figure}
In Figure \ref{kappagammaLxW}, the behavior of CFL number $a\lambda$
on dependent parameters $\kappa_{1}$ and $\gamma_{1}$ is shown. Note
that when $\lambda a \rightarrow 0^{+}$ the parameter $\gamma_{1}
\rightarrow 0^{+}$ whilst when $\lambda a \rightarrow 1^{-}$ the
parameter $\kappa_{1} \rightarrow 0^{-}$. In particular, for
$a|\lambda|=1$, definitions (\ref{kappa1}), (\ref{gamma1}) yield
non-TVD interval $[\kappa_{1}=0, \gamma_{1}=\frac{1}{3}]$. Note that
under linear CFL condition $0<\lambda a\leq1$, $\kappa_{1}(\lambda\,a
) \in (-1,0]$ and $ \gamma_{1}(\lambda\,a) \in (0,1/3)$. The following
result give CFL independent TVD bounds for LxW scheme
\cite{riteshEJDE}.
\begin{cor}\label{cor2}
The Lax-Wendroff scheme for (\ref{transport}) is total variation
diminishing under the linear stability condition $0<\lambda |a|\leq1,$
if $r_{i} =\in (\infty,\,-1)\cup(\frac{1}{3},\,\infty)$
\end{cor}
Thus it, can also be concluded from corollary \ref{cor1} and shaded
TVD region for LxW scheme in Figure \ref{kappagammaLxW} that {\it
except for $r_{i} \in [\kappa_{1},\gamma_{1}]$ the second order
accurate LxW scheme yields TVD approximation for all solution region
including extreme points with $r_{i}<0$.} More precisely,
\subsection{Upwind Beam-Warming scheme}
\begin{theorem}\label{thm2}
Away from sonic point and under CFL condition (\ref{cflNo}), second
order accurate Beam-Warming scheme is TVD for scalar conservation
law (\ref{nonlin}) in the solution region where \[ r^{\pm}_{i \mp 1}
\in \left[ \kappa_{2}\left(\lambda a^{\pm}_{i\mp\frac{1}{2}}\right),
\gamma_{2}\left(\lambda a^{\pm}_{i\mp\frac{1}{2}}\right)\right],\]
where parameter $\gamma_{2}$ and $\kappa_{2}$ defined as,
\begin{equation}
\kappa_{2}(x) = \displaystyle \frac{-(2 -x \sigma(x))}{x\sigma(x)} \label{kappa2}
\end{equation}
\begin{equation}
\gamma_{2}(x) = \displaystyle \frac{3-x\sigma(x)}{1-x\sigma(x)}\label{gamma2}
\end{equation}
\end{theorem}
\begin{proof}
\begin{itemize}
\item[]{\bf Case $f^{'}(u)>0$:} In this case, BW stencil use grid points
$[x_{i-s}, x_{i+r}], s=2, r=0$ thus to ensure locally non-sonic
region it suffice to let $a_{i-\frac{k}{2}} >0, k=-1,1,3$. The
numerical flux of Beam-Warming scheme is,
\begin{equation}
F^{n,BW}_{i+\frac{1}{2}}= F_{i} +
\frac{a_{i-\frac{1}{2}}}{2}\left(1- \lambda\,a_{i-\frac{1}{2}}\right)
\Delta_{-}{u}_{i},\;\; a_{i+\frac{1}{2}}>0. \label{bwflx1}
\end{equation}
Resulting non-conservative half I-form can be written as
\begin{equation} {u}^{n+1}_{i} = {u}_{i} -\left[\frac{\lambda\,
a_{i-\frac{1}{2}}}{2}\left(3-\lambda\,
a_{i-\frac{1}{2}}\right) - \frac{\lambda\,
a_{i-\frac{3}{2}}}{2}\left(1- \lambda\,
a_{i-\frac{3}{2}}\right)\frac{\Delta_{-}{u}_{i-1}}{\Delta_{-}{u}_{i}}
\right]\Delta_{-}{u}_{i} \label{Ifbw1}
\end{equation}
For half I-from (\ref{Ifbw1}) to be TVD, from Lemma \ref{lem2}
\begin{equation}
0 \leq \frac{\lambda\,a_{i-\frac{1}{2}}}{2}\left(3 -
\lambda\,a_{i-\frac{1}{2}}\right) - \frac{\lambda\,a_{i-\frac{3}{2}}}{2}\left(1 -
\lambda\,a_{i-\frac{3}{2}}\right)\frac{\Delta_{-}{u}_{i-1}}
{\Delta_{-}{u}_{i}}
\leq 1. \label{inq5}
\end{equation}
Compound Inequality (\ref{inq5}) can be written as,
\begin{equation}
\lambda a_{i-\frac{1}{2}}(3-\lambda a_{i-\frac{1}{2}})-2 \leq
\lambda a_{i-\frac{3}{2}}(1-\lambda
a_{i-\frac{3}{2}})\frac{\Delta_{-}{u}_{i-1}}{\Delta_{-}{u}_{i}}
\leq
\lambda
a_{i-\frac{1}{2}}(3-\lambda a_{i-\frac{1}{2}})
\end{equation}
Under CFL condition (\ref{cfl1}), i.e., $0<\lambda
a_{i+\frac{1}{2}}< 1,\; \forall i$ quantity $\lambda
a_{i-\frac{1}{2}}(1-\lambda a_{i-\frac{1}{2}})$ is positive, hence
(\ref {inq5}) can be written as,
\[
\frac{\lambda a_{i-\frac{1}{2}}(3-\lambda
a_{i-\frac{1}{2}})-2}{\lambda a_{i-\frac{1}{2}}(1-\lambda
a_{i-\frac{1}{2}})} \leq \frac{a_{i-\frac{3}{2}}(1-\lambda
a_{i-\frac{3}{2}})}{ a_{i-\frac{1}{2}}(1-\lambda
a_{i-\frac{1}{2}})}\frac{\Delta_{-}{u}_{i-1}}{\Delta_{-}{u}_{i}}
\leq \frac{ (3-\lambda a_{i-\frac{1}{2}})}{(1-\lambda
a_{i-\frac{1}{2}})}
\]
or
\[
\frac{\lambda a_{i-\frac{1}{2}}-2}{\lambda a_{i-\frac{1}{2}}} \leq
\frac{a_{i-\frac{3}{2}}(1-\lambda a_{i-\frac{3}{2}})}{
a_{i-\frac{1}{2}}(1-\lambda
a_{i-\frac{1}{2}})}\frac{\Delta_{-}{u}_{i-1}}{\Delta_{-}{u}_{i}}
\leq \frac{ (3-\lambda a_{i-\frac{1}{2}})}{(1-\lambda
a_{i-\frac{1}{2}})}
\]
On using flux wave split (\ref{fluxsplit}) we get
\[
\frac{\lambda a^{+}_{i-\frac{1}{2}}-2}{\lambda
a^{+}_{i-\frac{1}{2}}} \leq \frac{(1-\lambda
a_{i-\frac{3}{2}})}{(1-\lambda
a_{i-\frac{1}{2}})}\frac{\Delta_{-}F^{+}_{i-1}}{\Delta_{-}F^{+}_{i}}
\leq \frac{ (3-\lambda a^{+}_{i-\frac{1}{2}})}{(1-\lambda
a^{+}_{i-\frac{1}{2}})}
\]
Which, using (\ref{gradratio}) can be written as,
\begin{equation}
\kappa_{2}\left(\lambda\,a^{+}_{i-\frac{1}{2}}\right) \leq r^{+}_{i-1} \leq
\gamma_{2}\left(\lambda\,a^{+}_{i-\frac{1}{2}}\right) \label{inq6}
\end{equation}
where $\kappa_{2}$ and $\gamma_{2}$ are defined in (\ref{kappa2}) and (\ref{gamma2}).\\
\item[] {\bf Case $f^{'}(u)<0$:} In case of negative characteristics
speed, BW use stencil $[x_{i-s}, x_{i+r}], s=0, r=2]$ thus to
ensure locally non-sonic region it suffice to let
$a_{i+\frac{k}{2}} >0, k=-1,1,3$. In case of negative wave speed
the Beam-Warming flux is given by,
\begin{equation}
F^{n,BW}_{i+\frac{1}{2}} = F_{i+1} -
\frac{a_{i+\frac{3}{2}}}{2}\left(1+
\lambda\,a_{i+\frac{3}{2}}\right)\Delta_{+}{u}_{i+1},\;\; a_{i+\frac{1}{2}}>0.\label{bwflx2}
\end{equation}
Resulting non-conservative half I-form is,
\begin{equation} {u}^{n+1}_{i}= {u}_{i} +\left[
\frac{\lambda\,a_{i+\frac{3}{2}}}{2}\left(1 +
\lambda\,a_{i+\frac{3}{2}}\right)\frac{\Delta_{+}{u}_{i+1}}{\Delta_{+}{u}_{i}}
-
\frac{\lambda\,a_{i+\frac{1}{2}}}{2}\left(3 +
\lambda\,a_{i+\frac{1}{2}}\right) \right]\Delta_{+}{u}_{i}.\label{Ifbw2}
\end{equation}
Condition for (\ref{Ifbw2}) to be TVD is
\begin{equation}
0\leq \lambda\,a_{i+\frac{3}{2}}\left(1 +
\lambda\,a_{i+\frac{3}{2}}\right)\frac{\Delta_{+}{u}_{i+1}}{\Delta_{+}{u}_{i}} -
\lambda\,a_{i+\frac{1}{2}}\left(3 +\lambda\,a_{i+\frac{1}{2}}\right) \leq 2 \label{inq7}
\end{equation}
\begin{equation}
\lambda\,a_{i+\frac{1}{2}}\left(3 +
\lambda\,a_{i+\frac{1}{2}}\right) \leq \lambda\,a_{i+\frac{3}{2}}\left(1 +
\lambda\,a_{i+\frac{3}{2}}\right)\frac{\Delta_{+}{u}_{i+1}}{\Delta_{+}{u}_{i}}
\leq 2 +\lambda\,a_{i+\frac{1}{2}}\left(3 +
\lambda\,a_{i+\frac{1}{2}}\right)
\label{inq8}
\end{equation}
Note under CFL condition (\ref{cfl2}) $-1\leq
\lambda\,a_{i+\frac{1}{2}} <0,\, \forall i$, $\lambda\,
a_{i+\frac{1}{2}}\left(1+\lambda\,a_{i+\frac{1}{2}}\right)$ is
negative. Compound Inequality (\ref{inq8}) reduced to
\begin{equation}
\frac{2 +\lambda\,a_{i+\frac{1}{2}}\left(3
+\lambda\,a_{i+\frac{1}{2}}\right)}{\lambda\,
a_{i+\frac{1}{2}}\left(1+\lambda\,a_{i+\frac{1}{2}}\right)}
\leq
\frac{a_{i+\frac{3}{2}}\left(1 +
\lambda\,a_{i+\frac{3}{2}}\right)}{ a_{i+\frac{1}{2}}\left(1+\lambda\,a_{i+\frac{1}{2}}\right)}
\frac{\Delta_{+}{u}_{i+1}}{\Delta_{+}{u}_{i}}
\leq \frac{\left(3 +
\lambda\,a_{i+\frac{1}{2}}\right)}{\left(1+\lambda\,a_{i+\frac{1}{2}}\right)}
\end{equation}
or
\begin{equation}
\frac{\lambda\,a_{i+\frac{1}{2}}+2}{\lambda\,a_{i+\frac{1}{2}}} \leq
\frac{a_{i+\frac{3}{2}}\left(1 +
\lambda\,a_{i+\frac{3}{2}}\right)}{ a_{i+\frac{1}{2}}\left(1+\lambda\,a_{i+\frac{1}{2}}\right)}
\frac{\Delta_{+}{u}_{i+1}}{\Delta_{+}{u}_{i}}
\leq \frac{\left(3 +
\lambda\,a_{i+\frac{1}{2}}\right)}{\left(1+\lambda\,a_{i+\frac{1}{2}}\right)}
\label{inq9a}
\end{equation}
on using (\ref{fluxsplit})
\begin{equation}
\frac{\lambda\,a^{-}_{i+\frac{1}{2}}+2}{\lambda\,a^{-}_{i+\frac{1}{2}}} \leq
\frac{\left(1 +
\lambda\,a^{-}_{i+\frac{3}{2}}\right)}{ \left(1+\lambda\,a_{i+\frac{1}{2}}\right)}
\frac{\Delta_{+}F^{-}_{i+1}}{\Delta_{+}F^{-}_{i}}
\leq \frac{\left(3 +
\lambda\,a^{-}_{i+\frac{1}{2}}\right)}{\left(1+\lambda\,a^{-}_{i+\frac{1}{2}}\right)}
\label{inq9}
\end{equation}
Using (\ref{gradratio}), (\ref{kappa2}) and (\ref{gamma2}) Inequality
(\ref{inq9}) can be written as,
\begin{equation}
\kappa_{2}\left(\lambda\, a^{-}_{i+\frac{1}{2}}\right)
\leq r^{-}_{i+1} \leq
\gamma_{2}\left(\lambda\, a^{-}_{i+\frac{1}{2}}\right)
\end{equation}
\end{itemize}
\end{proof}
\subsubsection{BW on Linear problem}
\begin{cor}\label{cor3}
Beam-Warming scheme for (\ref{transport}) is total variation
diminishing under the linear stability condition $0<\lambda |a|
\leq1$, if the smoothness parameter $r^{\pm}_{i\mp 1} \in
[\kappa_{2}(\lambda a),\gamma_{2}(\lambda a)] $
\end{cor}
\begin{figure}[!\htb]
\begin{tabular}{cc}
\includegraphics[%
scale=0.55]{BWTVDregion.pdf}& \includegraphics[%
scale=0.55]{BWTVDregionzoom.pdf}\\
(a) & (b)
\end{tabular}
\caption{\label{kappagammaBW} (a) : LMP/TVD stability region (shaded) for
Beam-Warming scheme (b): Zoomed view.}
\end{figure}
In Figure \ref{kappagammaBW}, TVD region for Beam-Warming scheme is
shown as shaded region. It can be deduced that
\begin{enumerate}
\item $\lambda a\rightarrow 0^{+}$ Beam-Warming scheme yield TVD
approximation for $r_{i-\sigma(a)} \in (-\infty, 3]$.
\item $\lambda a\rightarrow 1^{-}$ Beam-Warming scheme yield TVD
approximation for $r_{i-\sigma(a)} \in [-1,\infty)$.
\end{enumerate}
Note that for $\lambda |a| \in (0,1]$ parameters $\kappa_{2}\in
(-\infty, -1]$ and $\gamma_{2} \in [3,\infty)$. Following CFL number
independent weaker TVD bounds can be concluded
\begin{cor}\label{cor4}
The Beam-Warming scheme for (\ref{transport}) is total variation
diminishing under the linear stability condition $0<\lambda |a|< 1,$
if $r^{\pm}_{i \mp 1} =[-1,3]$.
\end{cor}
\subsection{Fromm's scheme}
A less ocsillatory and second order accurate scheme is obtained by using a simple
average of LxW and BW flux i.e.,
\begin{equation}
F^{n,FROMM}_{i+\frac{1}{2}}= \frac{1}{2}\left(F^{n,LxW}_{i+\frac{1}{2}} + F^{n,BW}_{i+\frac{1}{2}}\right)\label{3rd}
\end{equation}
From Theorem \ref{thm1} and Theorem \ref{thm2} following result can be
proved,
\begin{theorem}\label{thm3}
Away from sonic point and CFL condition (\ref{cflNo}), the Fromm's scheme corresponding
to flux (\ref{3rd}) for scalar conservation law (\ref{nonlin}) is
TVD in the solution region where
\[\displaystyle
r^{\pm}_{i}\in \left(-\infty,\, \kappa_{1}\left(\lambda
a^{\pm}_{i\mp\frac{1}{2}}\right)\right)\cup
\left(\gamma_{1}\left(\lambda
a^{\pm}_{i\mp\frac{1}{2}}\right),\,\infty\right)\]
and
\[ r^{\pm}_{i \mp1} \in \left[ \kappa_{2}\left(\lambda
a^{\pm}_{i\mp\frac{1}{2}}\right),
\gamma_{2}\left(\lambda
a^{\pm}_{i\mp\frac{1}{2}}\right)\right].\] where
parameter $\kappa_{1}, \gamma_{1},\kappa_{2}$ and $ \gamma_{2}$ are
defined in (\ref{kappa1}), (\ref{gamma1}), (\ref{kappa2}) and
(\ref{gamma2}) respectively.
\end{theorem}
\section{Hybrid high order LMP/TVD stable schemes} \label{sec5}
It follows from the Theorem \ref{thm1}, \ref{thm2} and Theorem
\ref{thm3} that it is possible to achieve second or higher order TVD
approximation for most solution region including non-sonic exterma
where $r_{i}<0$. In order to demonstrate it numerically, we construct
hybrid schemes using a monotone/TVD scheme as {\bf complementary conservative
scheme (CCS)}. The following hybrid schemes are the natural choice
which satisfies the LMP/TVD bounds obtained in previous section and
thus ensures a LMP/TVD approximation. The second order accurate
LMP/TVD schemes use second order LxW and BW schemes in the region of
their LMP/TVD stability using bounds on smoothness parameter in
Theorem \ref{thm1} and \ref{thm2} respectively, otherwise use a
conservative conservative scheme (CCS).
\subsection{Centered scheme: LW-CCS}\label{algo1}
\begin{algorithmic}[1]
\IF{$r^{\pm}_{i}\leq \kappa_{1}$ OR $r^{\pm}_{i}\geq \gamma_{1}$} \STATE Update $u_{i}^{n+1} \gets$
LxW scheme
\ELSE \STATE Update $u_{i}^{n+1} \gets$ CCS.
\ENDIF
\end{algorithmic}
\subsection{Upwind Scheme: BW-CCS}\label{algo2}
\begin{algorithmic}[1]
\IF{($r^{\pm}_{i\mp 1}\geq \kappa_{2}$ AND
$r^{\pm}_{i\mp 1}\leq \gamma_{2}$)} \STATE Update $u_{i}^{n+1}
\gets$ BW scheme \ELSE \STATE Update $u_{i}^{n+1} \gets$ CCS.
\ENDIF
\end{algorithmic}
\subsection{Centred-Upwind scheme: FLWBW-CCS approximation}\label{algo3}
This LMP/TVD stable scheme can be obtained using Fromm scheme in
its region of LMP/TVD stability using bounds in \ref{thm3} along
with schemes \ref{algo1} and \ref{algo2} as follows
\begin{algorithmic}[1]
\IF{($r^{\pm}_{i}\leq \kappa_{1}$ OR $r^{\pm}_{i}\geq \gamma_{1}$)
\text{AND} ($r^{\pm}_{i\mp1}\geq \kappa_{2}$ AND
$r^{\pm}_{i\mp1}\leq \gamma_{2}$)} \STATE Update $u_{i}^{n+1}
\gets$ Fromm's scheme \ELSIF{$r^{\pm}_{i}\leq \kappa_{1}$ OR
$r^{\pm}_{i}\geq \gamma_{1}$} \STATE Update $u_{i}^{n+1} \gets$
LxW scheme \ELSIF{($r^{\pm}_{i\mp 1}\geq \kappa_{2}$ AND
$r^{\pm}_{i\mp 1}\leq \gamma_{2}$)} \STATE Update $u_{i}^{n+1}
\gets$ BW scheme \ELSE \STATE Update $u_{i}^{n+1} \gets$ CCS.
\ENDIF
\end{algorithmic}
Note that the scheme \ref{algo1}-\ref{algo3} are non-conservative as
they are based on TVD conditions on the smoothness parameter dedced
from the non-conservative form of studied schemes\footnote{except for
equations having constant characteristic speed e.g. linear transport
problem.}. Therefore they capture the steady shock accurately but may
produce moving shock at wrong location see results in Figure
\ref{Fig6a}(a). Note that incorrect shock location by scheme
\ref{algo1} is legging whereas by scheme \ref{algo2} it is leading to
the exact shock location in Figure \ref{Fig6a}(a). It is interesting
to see the scheme \ref{algo3} cancels the leading and legging errors
and gives exact shock location in Figure \ref{Fig6b}(a). This
phenomena of yielding wrong moving shock location by non-conservative
schemes along with the shock correction criteria is well explained in
\cite{Hou1994}. The idea for shock correction is to apply locally a
shock capturing conservative scheme in the vicinity of discontinuity
using a shock detector. It is therefore, to capture the moving shock
correctly, the following hybrid approach can be used,
\subsection{Shock Correction: SC-LW-CCS, SC-BW-CCS, SC-FLWBW-CCS hybrid schemes}\label{algo4}
\begin{algorithmic}[1]
\IF{Shock region} \STATE Update $u_{i}^{n+1} \gets$ use CCS,
\ELSE \STATE Update $u_{i}^{n+1} \gets$
with either of algorithm \ref{algo1}-\ref{algo3}. \ENDIF
\end{algorithmic}
\subsection{Extension to system of hyperbolic conservation
laws} \label{algo4sys} Consider the hyperbolic systems of
conservation law in one dimensions,
\begin{equation}
\frac{\partial}{\partial
t}\textbf{u}+\frac{\partial}{\partial x}\textbf{F(u)}=0, \label{1Dsys}
\end{equation}
where $\textbf{u}$ is vector of conserved quantities $u^{j},\,
i=1,2,\dots l$ and $\textbf{F}$ is the vector flux function. The above
proposed schemes for scalar case are extended to non-linear systems
(\ref{1Dsys}) in the natural manner using flux vector splitting and
average flux Jacobian matrix $A=\textbf{F}^{'}\textbf{(u)}$ of the
flux function. In particular for the numerical results presented in
next section, the Steger-Warming flux vector splitting is used for 1D
and 2D systems. The average Jacobian matrix is computed as follows,
$$A^{n}_{i+\frac{1}{2}}=A\left(\frac{\textbf{u}^n_{i+1}+\textbf{u}^n_{i}}{2}\right).$$
In order to compute the TVD bounds, the $n-$characteristic speeds associated
with system \ref{1Dsys})
\begin{equation}
a^{j}_{i+\frac{1}{2}}=\left\{\begin{array}{cc}
\frac{F^{j}_{i+1}-F^{j}_{i}}{u^{j}_{i+1}-u^{j}_{i}} &
{u^{j}_{i+1}-u^{j}_{i}}\neq 0
\\ \sigma(A_{i+\frac{1}{2}}) & else,\end{array}\right.j=1,2\dots l, \label{sysspeed}
\end{equation}
where $\sigma_{i+\frac{1}{2}}$ is the spectrum of eigen values of
$A_{i+\frac{1}{2}}$. In above computation (\ref{sysspeed}) the
nonphysical discrete wave speed caused by numerical overflow in case
of $u^{j}_{i+1}\approx u^{j}_{i}$ are corrected using following way
which is similar to the wave speed correction technique proposed in
\cite{jairaghu}. It is done as,
\begin{eqnarray}
a^{j}_{i+\frac{1}{2}}= \sigma_{max} \;\mbox{if}\; |a^{j}_{i+\frac{1}{2}}|\geq \sigma_{max},\\
a^{j}_{i+\frac{1}{2}}= \sigma_{min} \;\mbox{if}\; |a^{j}_{i+\frac{1}{2}}|\leq \sigma_{min},
\end{eqnarray}
where $\sigma_{max}$ and $\sigma_{min}$ refer to the local maxima and
minima of the magnitudes of characteristic speeds associated with
system (\ref{1Dsys}). For example, the one dimensional Euler equations
has the eigenvalues $u,\;u\pm c$ where $u$ and $c$ denotes fluid
velocity and the speed of sound respectively. In this case, we define
\begin{eqnarray}
\,\sigma_{\max}&=& \max{(\max{(|u|, |u-c|, |u+c|)_{i}} , \max{(|u|, |u-c|, |u+c|)_{i+1}) }},\\
\sigma_{\min}&=& \max{(\min{(|u|,|u-c|, |u+c|)_{i}} , \min{(|u|, |u-c|, |u+c|)_{i+1} )}}.
\end{eqnarray}
\subsection{Shock Sensor}
In order to locate the presence of discontinuities, a shock detector
proposed in \cite{Liushock} with some modification is used. A brief
detail on the shock switch is given below for the sake of completeness
of the discussion on numerical implementation.
\begin{itemize}
\item[Step 1:] Check multigrid ratio check
$$MR(i,h)=\frac{T_{C}(i,h)}{T_{F}(i,h) +\epsilon}$$ where
$T_{C}(i,2h)$ and $T_{F}(i,h)$ are the $(4^{th}, 5^{th}\;
\mbox{and}\; 6^{th})$ order truncation error sum on a coarse (with
N/2 grid points) and fine grid (with N points) respectively. The
derivatives in this step are calculated by by sixth order compact
scheme proposed in \cite{Lele1992}. The small parameter
$0<\epsilon<<1$ is used to avoid division by zero.
\item[Step 2:] Calculate the local ratio check at
the grid point $x_i$ which has multigrid ratio $MR(i,h)\leq 4$,
$$LR(i) = \left|\frac{\left(u^{'}_{R}\right)^2-\left(u^{'}_{L}\right)^2}{\left(u^{'}_{R}\right)^2 +
\left(u^{'}_{R}\right)^2 +\epsilon}\right| $$ where
$u^{'}_{L}=3u_{i}-4u_{i-1}+u_{i-2}$ and
$u^{'}_{R}=3u_{i}-4u_{i+1}+u_{i+2}$ left and right slope
respectively at grid point $x_i$.
\item[Step 3:] Use a cutoff value $\delta \in(0,1]$ to create a shock switch (SS) on the result of step 2. i.e.,
$$ SS(i)=\left\{\begin{array}{ll} 0, & \mbox{if}\;LR(i)<\delta\; \mbox{i.e., data is locally smooth around grid point}\; x_{i},\\
1,& \mbox{if} \;LR(i)\geq \delta\; \mbox{i.e., data is discontinuous around grid point}\; x_{i}.
\end{array}\right.$$
\end{itemize}
Note that the above shock detector has parameters $\epsilon$ and
$\delta$ which governs the sensitivity of shock switch. It is observed
in numerical computations that for larger value of parameters e.g.,
$\epsilon =1\times 10^{-2},\delta=0.8$ above shock switch is less
sensitive for mild shock or sharp turns and detects only strong shocks
whereas $\epsilon =1\times 10^{-8},\delta=0.2$ detect corners and mild
shock along with the strong shock. Also in case of non-linear systems,
it is observed that small oscillations may arise in the vicinity of
shock depending on the choice of shock parameters
$\epsilon,\delta$. It is therefore, to make it robust and less prone
to parameters $\epsilon,\delta$ a slight modification is done as
follows,
\begin{itemize}
\item[Step 4:] Treat neighbouring grid point $x_{i\pm 1}$ in
discontinuity region if $SS(i)=1$ i.e., use $SS(i\pm1)=1$ if
$SS(i)=1$.
\end{itemize}
\section{Numerical Results} \label{sec6}
In this section numerical results are presented for various benchmark
scalar and system test problems in both one and two
dimensions. Different smooth as well discontinuous initial conditions
are taken to show the performance of schemes in section \ref{sec5} in terms of
accuracy and discontinuity capturing respectively. Numerical results
show that the proposed hybrid scheme, due to improved accuracy at
extrema and steep gradient region, nicely approximates the smooth
region of solution with crisp resolution for rarefaction, contact and
shock discontinuities. Moreover it produces the total variation
diminishing numerical approximation.
\subsection{\bf Linear transport equation: every extrema is non-sonic}
Consider the linear transport equation
\begin{equation}
u_{t} + u_{x}=0,\; u(x,0)= u_{0}(x) \label{transport1}
\end{equation}
with periodic boundary condition. The exact solution of
(\ref{transport1}) equation convects with out changing the initial
shape of $u_{0}(x)$ and is given by $u(x,t) = u_{0}(x-t)$. Note that
in general number of extrema are finite and clipping error due to
degenerate accuracy at smooth extrema by existing high order monotone
and TVD method is visible only after a long period of time in form of
approximation of smooth extrema with corners or flatten profile see
Figure \ref{Fig1a}(a). Also due to degenerate accuracy at extrema and
steep gradient region their erratic convergence rate can be seen only
after a long time see Table \ref{LWflmT3linprb3}. It is therefore,
probably the transport equation (\ref{transport1}) the only test which
can be used to check the large time performance of any method. Since
the problem (\ref{transport1}) is linear thus discontinuities present
in the solution does not represents shock or rarefaction therefore
scheme \ref{algo3} can be directly applied for this test problem with
out shock switch. The numerical computation for problem
(\ref{transport1}) is done by using the first order upwind scheme as
complementary conservative scheme (CCS) in hybrid scheme \ref{algo3}
and results are shown by legend {\it method}.
\subsubsection{\bf Accuracy check: smooth initial condition}
Consider \ref{transport1}) along with the the following three different
initial conditions which comprises of smooth extrema, monotone region
with mild as well sharp turn.
\begin{enumerate}
\item[i] Smooth extrema
\begin{equation}
u(x,0) = \sin(\pi x), \; x\in
[-1,\;1].\label{Lin-IC1}
\end{equation}
The initial profile consists smooth extreme points at $x=\pm
\frac{1}{2}$ which preservingly convects to the right direction.
\item[ii] Smooth extrema with monotone data region
\begin{equation}
u(x,0) = \sin^4(\pi x), \; x\in [0,\;1].\label{Lin-IC2}
\end{equation}
This initial condition is taken from \cite{zhang2010} and has a
smooth extrema at $x=0.5$ with monotone solution regions with mild
turn towards the bottom.
\item[iii] Smooth extrema with steep gradient region
\begin{equation}\label{Lin-IC3}
u(x,0) = \left\{\begin{array}{ll}
e^{\frac{-1}{1-x^2}}\; & x\in [-1:1]\\
0\; & else
\end{array}\right.
\end{equation}
This initial condition has a smooth extrema at $x=0$. Compared to
initial condition (\ref{Lin-IC2}) it has high gradient monotone
region with sharp turns towards the bottom where $r\rightarrow 0+$
or $r>>1$ respectively. This is a good test to see the degenerate
convergence rate for any limiter based TVD scheme.
\end{enumerate}
For smooth initial data numerical solution plots are given at large
time level{ $\bf T_{s}$}\footnote{We consider large time $T_{s}$ as
the time level when corners are visible in the approximation of
smooth extrema by high order TVD methods such as in
\cite{harten1983,Dubey2013,sweby1984,Yee1987,davis1987}.}. The
convergence rate of scheme \ref{algo4} is given at various time
instance with varying $CFL$ number to show the robust and higher than
second order convergence rate of the method \ref{algo3}. In Figure
\ref{Fig1a}(a), numerical solution obtained by method corresponding
to IC (\ref{Lin-IC1}) is compared with high order TVD Lax-Wendroff
flux limited method (LxWFLM) \cite{Rider} with compressive Superbee
limiter \cite{roe1985some}. In Figure \ref{Fig1b} and Figure
\ref{Fig1c} approximate solution is given for transport problem
corresponding to initial conditions (\ref{Lin-IC2}) and
(\ref{Lin-IC3}) respectively. The total variation of the computed
solution obtained by method is compared with uniformly second order LxW, BW and Fromm schemes respectively for all three IC's in Figure
\ref{Fig1a}(b), \ref{Fig1b}(b) and \ref{Fig1c}(c).
From the numerical results in it is evident that problem of flattening
of smooth round shaped solution profile is removed due to improved
approximation of extreme points. Moreover it can be observed and
Figure \ref{Fig1b}(a) that this improvement more visible as $\lambda a
\rightarrow 1$ which support the improved TVD region for extrema of
LxW as discussed in the Corollary \ref{cor1}. Total variation plots show
that total variation for designed scheme \ref{algo3} is decreasing
whilst uniformly second order LxW and BW schemes do not produce TVD
solution though it remain total variation bounded (TVB).
In Table \ref{Tab1linprb1}, discrete maximum $L^{\infty}$ error
convergence rate is given for scheme \ref{algo3} as $L^{\infty}$ error
is the best indicator for checking the performance of any scheme in
terms of clipping error due to drop in accuracy at smooth extrema. In
Table \ref{Tab2linprb2} and Table \ref{Tab3linprb3}, error convergence
rates are given in terms of $L^{1}$ and $L^{\infty}$ error for method
with different choice of $CFL$ and time for linear test corresponding
to initial conditions (\ref{Lin-IC2}) and (\ref{Lin-IC3})
respectively. The numerical results show that the designed scheme
\ref{algo3} shows higher than second order convergence rate
independently of the choice of $CFL$ number or final time $T$. Also
due to improved approximation of extrema and steep gradient region,
the used method yields smooth approximation with out clipping error
which support the Corollary \ref{cor1} and \ref{cor2}.
\begin{figure}[!htb]
\begin{center}
\begin{tabular}{cc}
\includegraphics[%
scale=0.5]{methodlinprob1.pdf} &\includegraphics[%
scale=0.5]{tvplotlineartest1.pdf}\\
(a) at $T=30$. & (b) at $T=2$ with $CFL=0.95$.\\
\end{tabular}
\end{center}
\caption{\label{Fig1a} Solution of (\ref{transport1}) with IC
(\ref{Lin-IC1}) using $N=80$: (a) Flatten approximation for smooth
extrema by LxWFLM whereas proposed scheme \ref{algo4} preserve
solution with smooth extrema with out introducing corners. (b)
Comparison of total variation.}
\end{figure}
\begin{table}[!htb]
\begin{tabular}{|c|c|
\hline
T=2 & T=30\\
\hline
\begin{tabular}{c|c
CFL=0.5 & CFL=0.95\\
\hline
\begin{tabular}{ccc
N & $L^{\infty}$ error & Rate\\
\hline
20 & 9.1927e-03 & \dots \\
40 & 1.9456e-03 & 2.240 \\
80 & 3.7656e-04 & 2.369 \\
160 & 7.0744e-05 & 2.412 \\
320 & 1.2599e-05 & 2.489 \\
640 & 3.2485e-06 & 1.955\\
\end{tabular
&
\begin{tabular}{cc
$L^{\infty}$ error & Rate\\
\hline
1.4026e-03 & \dots \\
2.6640e-04 & 2.396 \\
5.6365e-05 & 2.241 \\
1.1332e-05 & 2.314 \\
2.6879e-06 & 2.076 \\
3.1057e-07 & 3.113 \\
\end{tabular
\end{tabular
&
\begin{tabular}{c|c
CFL=0.5 & CFL=0.95\\
\hline
\begin{tabular}{cc
$L^{\infty}$ error & Rate\\
\hline
6.3628e-02 & \dots \\
1.0892e-02 & 2.546 \\
1.8543e-03 & 2.554 \\
3.3400e-04 & 2.473 \\
4.1812e-05 & 2.998 \\
3.5142e-06 & 3.573\\
\end{tabular
&
\begin{tabular}{cc
$L^{\infty}$ error & Rate\\
\hline
1.1716e-02 & \dots\\
2.2095e-03 & 2.407 \\
4.4726e-04 & 2.305 \\
8.6187e-05 & 2.376 \\
1.4613e-05 & 2.560 \\
4.5598e-07 & 5.002 \\
\end{tabular
\end{tabular}\\
\hline
\end{tabular
\caption{\label{Tab1linprb1} {\it Consistent higher than second order $L^{\infty}$ convergence rate with the
mesh refinement corresponding to initial condition (\ref{Lin-IC1})}.}
\end{table}
\begin{figure}[!htb]
\begin{center}
\begin{tabular}{cc}
\includegraphics[%
scale=0.5]{methodlinprob2atT20.pdf} &\includegraphics[%
scale=0.5]{tvplotlineartest2.pdf}\\
(a) & (b)\\
\end{tabular}
\end{center}
\caption{\label{Fig1b} Solution for linear equation
(\ref{transport}) with IC (\ref{Lin-IC2}) (a) at time $T=20,\;
N=100$, Smooth approximation of extrema with reduced clipping
error: (b) TV plot up to $T=2$ using $N=80, CFL=0.95$.}
\end{figure}
\begin{table}[!htb]
\begin{tabular}{|c|c|}
\hline
CFL=0.5 & CFL=0.95\\
\hline
\begin{tabular}{ccc|cc}
\hline
N& $L^{1}$ Error & Rate & $L^{\infty}$ error & Rate\\
\hline
40 & 7.6962e-02 & \dots & 6.1348e-03 & \dots \\
80 & 1.7350e-02 & 2.149 & 1.1039e-03 & 2.474 \\
160 & 3.8066e-03 & 2.188 & 1.9596e-04 & 2.494 \\
320 & 7.3723e-04 & 2.368 & 2.4368e-05 & 3.008 \\
640 & 1.1693e-04 & 2.656 & 4.3613e-07 & 5.804 \\
1280 & 2.8601e-05 & 2.031 & 5.4189e-08 & 3.009 \\
2560 & 7.0685e-06 & 2.017 & 6.6891e-09 & 3.018 \\
\end{tabular}
&
\begin{tabular}{cc|cc}
\hline
$L^{1}$ Error & Rate & $L^{\infty}$ error & Rate\\
\hline
1.3179e-02 & \dots & 1.4315e-03 & \dots \\
3.4723e-03 & 1.924 & 2.7228e-04 & 2.394 \\
8.7436e-04 & 1.990 & 5.2159e-05 & 2.384 \\
2.0210e-04 & 2.113 & 8.6669e-06 & 2.589 \\
3.2508e-05 & 2.636 & 1.1141e-07 & 6.282 \\
8.0887e-06 & 2.007 & 1.3877e-08 & 3.005 \\
2.0182e-06 & 2.003 & 1.7315e-09 & 3.003 \\
\end{tabular}\\
\hline
\end{tabular}
\caption{\label{Tab2linprb2} {\it Order of convergence with the
mesh refinement at $T=20$ corresponding to initial condition (\ref{Lin-IC2})}.}
\end{table}
\begin{figure}[!htb]
\begin{center}
\begin{tabular}{cc}
\includegraphics[%
scale=0.5]{methodlinprob3atT6.pdf} &\includegraphics[%
scale=0.5]{tvplotlineartest3.pdf}\\
(a) & (b)\\
\end{tabular}
\end{center}
\caption{\label{Fig1c} Solution for linear equation
(\ref{transport}) with IC (\ref{Lin-IC3}) at time $T=6$ and
$N=100,\;CFL=0.95$. (a) Smooth approximation of extrema with no
clipping error by Scheme \ref{algo1}. (b) Comparison of total
variation.}
\end{figure}
\begin{table}[!htb]
\begin{tabular}{|c|c|}
\hline
I-Order Upwind & Second order TVD scheme\\
\hline
\begin{tabular}{c|cc|cc}
\hline
N& $L^{1}$ Error & Rate & $L^{\infty}$ error & Rate\\
\hline
80 & 2.5248e-01 & 0.583 & 1.6091e-02 & 1.482 \\
160 & 1.6210e-01 & 0.639 & 5.8392e-03 & 1.462 \\
320 & 9.7673e-02 & 0.731 & 2.2646e-03 & 1.367 \\
640 & 5.6346e-02 & 0.794 & 8.2206e-04 & 1.462 \\
1280 & 3.1355e-02 & 0.846 & 2.8328e-04 & 1.537 \\
2560 & 1.6889e-02 & 0.893 & 9.2241e-05 & 1.619 \\
\end{tabular}
&
\begin{tabular}{cc|cc}
\hline
$L^{1}$ Error & Rate & $L^{\infty}$ error & Rate\\
\hline
2.2319e-02 & 1.314 & 2.4725e-03 & 2.294 \\
1.6468e-02 & 0.439 & 1.3464e-03 & 0.877 \\
5.2704e-03 & 1.644 & 2.0531e-04 & 2.713 \\
2.3125e-03 & 1.188 & 8.7812e-05 & 1.225 \\
1.0757e-03 & 1.104 & 4.1735e-05 & 1.073 \\
4.1654e-04 & 1.369 & 1.0950e-05 & 1.930\\
\end{tabular}\\
\hline
\end{tabular}
\caption{\label{LWflmT3linprb3} {\it Order of convergence using I
order upwind and LxW flux limited TVD method with Superbee
limiter at $T=6, CFL=0.5$ corresponding to initial condition
(\ref{Lin-IC3})}.}
\end{table}
\begin{table}[!htb]
\begin{tabular}{|c|c|}
\hline
CFL=0.5 & CFL=0.95\\
\hline
\begin{tabular}{c|cc|cc}
\hline
N& $L^{1}$ Error & Rate & $L^{\infty}$ error & Rate\\
\hline
80 & 4.3378e-02 & 1.654 & 4.9870e-03 & 2.139\\
160 & 1.2992e-02 & 1.739 & 1.3187e-03 & 1.919 \\
320 & 3.2156e-03 & 2.014 & 2.6333e-04 & 2.324 \\
640 & 6.1867e-04 & 2.378 & 3.8218e-05 & 2.785 \\
1280 & 1.1353e-04 & 2.446 & 4.1315e-06 & 3.210 \\
2560 & 1.7115e-05 & 2.730 & 3.6114e-07 & 3.516 \\
\end{tabular}
&
\begin{tabular}{cc|cc}
\hline
$L^{1}$ Error & Rate & $L^{\infty}$ error & Rate\\
\hline
1.1681e-02 & 1.646 & 2.2055e-03 & 2.239 \\
3.2801e-03 & 1.832 & 5.0305e-04 & 2.132 \\
8.1011e-04 & 2.018 & 8.1325e-05 & 2.629 \\
1.7801e-04 & 2.186 & 1.3173e-05 & 2.626 \\
4.5829e-05 & 1.958 & 1.8927e-06 & 2.799 \\
1.2376e-05 & 1.889 & 2.5213e-07 & 2.908 \\
\end{tabular}\\
\hline
\end{tabular}
\caption{\label{Tab3linprb3} {\it Consistent higher than second order of convergence with
the mesh refinement at $T=6$ corresponding to initial condition (\ref{Lin-IC3})}}
\end{table}
\subsubsection{\bf Discontinuous initial condition}
In this test is taken from \cite{despres} originally used by Harten in
\cite{Harten1989}. Initial solution is complex in nature which
contains parts of smooth solution, mix discontinuities,
discontinuities of derivative in the interval $[-1,1]$. In Figure
\ref{Fig4}(a), numerical results is given by proposed method and in
Figure \ref{Fig4}(b) the total variation plot of the computed solution
by {\it method} is given.
\begin{equation}
\label{icprb3}
u_{0}(x) = \left\{\begin{array}{lcl}
2x + 2 -sin(3\pi(x-0.5))/6 & \mbox{if} &-1\leq x\leq 0.5, \\
(0.5-x)sin(1.5\pi(x-0.5)^2) & \mbox{if} &-0.5<x<1/6,\\
|sin(2\pi(x-0.5))| & \mbox{if} & 1/6< x<= 5/6,\\
2x - 2 -sin(3\pi(x-0.5))/6 & \mbox{if} & 5/6< x <=1.\\
\end{array}\right. \\
\end{equation}
It can be seen in Figure \ref{Fig4} that the proposed scheme \ref{algo3} yields TVD
solution with crisp capturing of the discontinuities with out clipping
error at smooth extrema. The error convergence rate is not shown for
this discontinuous test problem as discontinuities can only be
approximated with at most first order of accuracy \cite{cwshu2012}
\begin{figure}[httb!]
\begin{center}
\begin{tabular}{cc}
\includegraphics[scale=0.5]{methodlinearHartentest.pdf} & \includegraphics[%
scale=0.5]{tvplotlinprob3.pdf}\\
(a) & (b)
\end{tabular}
\caption{\label{Fig4} (a) High resolution oscillation free solution
for test problem \ref{icprb3} by scheme \ref{algo4} for data
$CFL=0.95,\, N=160,\, T=2.0 $. (b) Total variation decreasing plot of method}
\end{center}
\end{figure}
\subsection{\bf Non-linear case: Burgers equation}
Consider the Burgers equation
\begin{equation}
\displaystyle u_t + \left(\frac{u^2}{2}\right)_x=0, -a\leq x
\leq b \label{burgers}
\end{equation}
with initial condition $u_{0}(x)$ and periodic boundary conditions. It
is the non-linear nature of the equation (\ref{burgers}) that even for
smooth initial condition, the solution of (\ref{burgers}) eventually develops
discontinuities like rarefaction and shocks after breaking time $T_{b}$
given by
\begin{equation}
T_{b}=\frac{-1}{\min_{x}(u^{'}_0(x))} \label{Tbtime}
\end{equation}
Also the unique sonic point for Burgers equation (\ref{burgers}) is
$u^{*}=0$. It is therefore, Burgers equation is a good test to check
the performance of any scheme for smooth as well discontinuous
solutions profile at pre and post-shock time $T_{b}$ respectively. In
the numerical computation FORCE scheme is used as CCS in schemes
\ref{algo1} to \ref{algo3} and their respective hybrid shock corrected
analog in scheme \ref{algo4}.
\subsubsection{\bf Shock correction moving shock}
We first consider the following discontinuous initial condition to show
the non-conservative nature of schemes \ref{algo1}, \ref{algo2} and
\ref{algo3} as they yield solution with wrong location of moving shock.
\begin{equation}
u(x,0) = \left\{\begin{array}{cc} 1, & |x|\leq \frac{1}{3},\\ 0, &
else. \end{array}\right.\label{nonlin-IC2}
\end{equation}
The solution corresponding to IC (\ref{nonlin-IC2}), develops a
rarefaction wave and a {\it moving} shock which corresponds to initial
discontinuities at $x=-1/3$ and $x=1/3$ respectively. In Figure
\ref{Fig6a}(a) results obtained by second order Lax-Wendroff
(LW-FORCE) and Beam-Warming (BW-FORCE) schemes \ref{algo1} and
\ref{algo2} respectively are given. It is clear that shock location
given by LW-FORCE is legging behind whereas by BW-FORCE it is leading
ahead of exact shock location. Results by shock corrected schemes
SC-LW-FORCE and SC-BW-FORCE as described in \ref{algo4} are given in
\ref{Fig6a}(b) which show correct location of shock is recovered with
out loosing crisp resolution of left rarefaction. Numerical results in
Figure \ref{Fig6b}(a), obtained by scheme \ref{algo3}
(FLWBW-FORCE) shows that shock is crisply captured at right location
with high resolution for bottom and top of left rarefaction. In Figure
\ref{Fig6b}(b), results by SC-FLWBW-FORCE are given which show little
dissipative resolution for shock which is due to approximation by
dissipative FORCE scheme in the vicinity of shock.
\begin{figure}[httb!]
\begin{tabular}{c|c}
\includegraphics[scale=0.4]{LWFORCE-laneyCFL0p8.pdf} & \includegraphics[scale=0.4]{DetectLWFORCE-laneyCFL0p8.pdf}\\
\includegraphics[scale=0.4]{BWFORCE-laneyCFL0p8.pdf} & \includegraphics[scale=0.4]{DetectBWFORCE-laneyCFL0p8.pdf} \\
(a) & (b)\\
\end{tabular}
\caption{\label{Fig6a} Solution at $T =0.8$ using $CFL=0.8,\;
N=80$, (a) Wrong location of moving shock using Scheme
\ref{algo1} and \ref{algo2} (b) Shock correction using shock
switch.}
\end{figure}
\begin{figure}[httb!]
\begin{tabular}{c|c}
\includegraphics[scale=0.4]{FROMMLWBWFORCE-laneyCFL0p8.pdf}
&\includegraphics[scale=0.4]{DetectFROMMLWBWFORCE-laneyCFL0p8.pdf}\\
(a) & (b)\\
\end{tabular}
\caption{\label{Fig6b} Solution at $T =0.8$ using $CFL=0.8,\; N=80$,
(a) Scheme \ref{algo3} give correct shock location high
resolution for left rarefaction and crisp capturing for the moving
shock, (b) Solution by Scheme \ref{algo4}.}
\end{figure}
\subsubsection{\bf Accuracy Check: Smooth Initial conditions}
Consider three different smooth initial
conditions (IC) along with corresponding breaking time $T_{b}$
\begin{equation}
u(x,0) = 0.1+ sin^{4}(\pi x),\, x\in[0,\,1],\; T_{b} =
0.27803225 \label{nonlin-IC2a}
\end{equation}
\begin{equation}
u(x,0) = 0.1+\exp^{-x^{4}},\, x\in [-2:3], \; T_{b} =
0.65669683. \label{nonlin-IC2b}
\end{equation}
\begin{equation}
u(x,0) = \frac{1}{4}(1+ sin(\pi x)),\, x\in [-1:1], \;
T_{b}=\frac{4}{\pi}. \label{nonlin-IC2c}
\end{equation}
IC (\ref{nonlin-IC2a}) and (\ref{nonlin-IC2b}) does not contain any
sonic point whereas IC (\ref{nonlin-IC2c}) has a sonic point at
$x=-0.5$ since $u(-0.5, 0)=0$. The solution corresponding to IC
(\ref{nonlin-IC2a}) and (\ref{nonlin-IC2b}) develop a {\it moving}
shock followed by a rarefaction fan whereas the moving shock
corresponding to IC (\ref{nonlin-IC2c}) is separated by two
rarefaction fans. In Figure \ref{Fig1burgerIC2a}, \ref{Fig2burgerIC2b}
and \ref{burgerShuosherFig} the pre and post-shock solution of Burgers
equation obtained by the shock corrected hybrid Scheme \ref{algo4}
({\it SC-FLWBW-FORCE}) corresponding to IC (\ref{nonlin-IC2a}),
(\ref{nonlin-IC2b}) and (\ref{nonlin-IC2c}) respectively are
given. The total variation plots are also given for different choices
of CFL number $\lambda \max_{u}|f^{'}(u)|$. In Table
\ref{Tab1burgerIC2a} to Table \ref{burgerShuosherTab}, $L^{1}$ and $L^{\infty}$
errors are shown at pre-shock time $\displaystyle T_{b}$ using
different CFL number. Results show that the hybrid scheme nicely
approximated pre-shock solution with out clipping error and does not
introduce induced oscillations near shock in the post-shock
solution. Moreover purposed method yields a total variation
diminishing solution and shows a consistent convergence rate between
second and third order in both the norms.
\begin{figure}[!htb]
\begin{tabular}{cc}
\includegraphics[%
scale=0.5]{burgerIC1CFL0p95.pdf}&\includegraphics[%
scale=0.5]{burgerIC1tvplot.pdf}\\
(a) & (b)\\
\end{tabular}
\caption{\label{Fig1burgerIC2a} Burgers equation solution using {\it SC-FLWBW-FORCE}
corresponding to IC (\ref{nonlin-IC2a}): (a) No clipping error for
smooth extrema as well near shock zone $CFL=0.95, N=50$, (b) Effect of CFL on total variation of computed solution}
\end{figure}
\begin{table}[!htb]
\begin{tabular}{|c|c|c|}
\hline N & CFL=0.6 & CFL=0.9\\ \hline
\begin{tabular}{c}
\\
10\\
20\\
40\\
80\\
160\\
320\\
640\\
\end{tabular}
&
\begin{tabular}{cccc}
$L^{1}$ error & Rate &$L^{\infty}$ error & Rate\\
\hline
3.1074e-02 & \dots & 1.2720e-02 & \dots\\
1.3874e-02 & 1.1634 & 4.6798e-03 & 1.4426\\
2.0530e-03 & 2.7566 & 3.9157e-04 & 3.5791e\\
4.0690e-04 & 2.3350 & 6.9600e-05 & 2.4921e\\
8.6630e-05 & 2.2317 & 1.2510e-05 & 2.4760e\\
1.4700e-05 & 2.5591 & 1.2000e-06 & 3.3820e\\
2.4800e-06 & 2.5674 & 5.0000e-08 & 4.5850e\\
\end{tabular}
&
\begin{tabular}{cccc}
$L^{1}$ Error & Rate &$L^{\infty}$ error & Rate\\
\hline
2.7767e-02 & \dots & 1.0627e-02 & \dots\\
5.8287e-03 & 2.2521 & 2.5638e-03 & 2.0514\\
1.0398e-03 & 2.4868 & 2.2157e-04 & 3.5324\\
2.3176e-04 & 2.1657 & 3.1460e-05 & 2.8162\\
5.2120e-05 & 2.1527 & 3.9000e-06 & 3.0120\\
1.0380e-05 & 2.3280 & 4.1000e-07 & 3.2498\\
2.4200e-06 & 2.1007 & 5.0000e-08 & 3.0356\\
\end{tabular}\\
\hline
\end{tabular}
\caption{\label{Tab1burgerIC2a} Convergence rate corresponding to IC (\ref{nonlin-IC2a}) at time $T=T_{b}/2$}
\end{table}
\begin{table}[!htb]
\begin{tabular}{|c|c|}
\hline
CFL=0.45 & CFL=0.95\\
\hline
\begin{tabular}{ccccc}
N& $L^{1} error$ & Rate& $L^{\infty} error$ & Rate\\
\hline
10 & 3.1120e-01 & \dots & 1.0335e-01 & \dots\\
20 & 6.4051e-02 & 2.2805 & 2.7574e-02 & 1.9061\\
40 & 8.7373e-03 & 2.8740 & 2.9299e-03 & 3.2344\\
80 & 1.9428e-03 & 2.1690 & 4.6987e-04 & 2.6405\\
160 & 4.1491e-04 & 2.2273 & 6.1350e-05 & 2.9371\\
320 & 9.0750e-05 & 2.1928 & 6.7500e-06 & 3.1841\\
640 & 2.0870e-05 & 2.1205 & 7.3000e-07 & 3.2089\\
\end{tabular}
&
\begin{tabular}{cccc}
$L^{1} error$ & Rate & $L^{\infty} error$ & Rate\\
\hline
2.5190e-01 & \dots & 9.7632e-02 &\dots\\
4.7036e-02 & 2.4210 & 2.2831e-02 & 2.0964\\
1.2098e-02 & 1.9590 & 5.2950e-03 & 2.1083\\
2.3084e-03 & 2.3899 & 5.5543e-04 & 3.2530\\
4.2048e-04 & 2.4568 & 5.5800e-05 & 3.3153\\
8.6830e-05 & 2.2758 & 5.3900e-06 & 3.3719\\
1.9470e-05 & 2.1569 & 5.6000e-07 & 3.2668\\
\end{tabular}\\
\hline
\end{tabular}
\caption{\label{Tab1burgerIC2b} Third order $L^{\infty}$ convergence rate corresponding to IC (\ref{nonlin-IC2b}) at pre-Shock time Tb/2, $Tb=0.65669683$}
\end{table}
\begin{figure}[!htb]
\begin{center}
\begin{tabular}{cc}
\includegraphics[%
scale=0.45]{burgerIC2CFL0p8.pdf}&\includegraphics[%
scale=0.45]{burgerIC2tvplot.pdf}\\
(a) & (b)\\
\end{tabular}
\end{center}
\caption{\label{Fig2burgerIC2b} Solution corresponding to IC
(\ref{nonlin-IC2b}): (a) No clipping error for smooth extrema as
well near shock zone $CFL=0.8, N=80$, (b) Effect of CFL on total variation diminishing plot of computed solution.}
\end{figure}
\begin{figure}[!htb]
\begin{tabular}{cc}
\includegraphics[scale=0.45]{burgerShutestCFL0p6.pdf}& \includegraphics[scale=0.45]{burgerShutvplot.pdf}\\
(a) & (b)\\
\end{tabular}
\caption{\label{burgerShuosherFig} (a) Comparison of computed solution
of corresponding to IC (\ref{nonlin-IC2c}) at $Tb=1/4\pi$ $CFL=0.6, N=80$ (b) Total variation diminishing plot up to
$t=2Tb$}
\end{figure}
\begin{table}[!htb]
\begin{tabular}{|ccccc|cccc|}
\hline
& & CFL=0.45 & & & & CFL=0.9 & & \\
\hline
N& $L^{1}$ error& Rate & $L^{\infty}$ error& Rate & $L^{1}$ error& Rate & $L^{\infty}$ error& Rate\\
\hline
10 & 3.9495e-02 & 0.0000 & 1.0467e-02 & 0.0000 & 1.9593e-02 & 0.0000 & 4.7957e-03 & 0.0000\\
20 & 7.2108e-03 & 2.4534 & 1.8537e-03 & 2.4974 & 4.6493e-03 & 2.0753 & 1.1326e-03 & 2.0821\\
40 & 2.1212e-03 & 1.7653 & 3.7326e-04 & 2.3122 & 1.0845e-03 & 2.1000 & 2.4774e-04 & 2.1927\\
80 & 4.9995e-04 & 2.0850 & 7.0710e-05 & 2.4002 & 2.7441e-04 & 1.9826 & 4.2200e-05 & 2.5535\\
160 & 9.2620e-05 & 2.4324 & 1.7250e-05 & 2.0353 & 6.6870e-05 & 2.0369 & 8.3800e-06 & 2.3322\\
320 & 2.2230e-05 & 2.0588 & 4.1900e-06 & 2.0416 & 1.6580e-05 & 2.0119 & 1.7400e-06 & 2.2679 \\
640 & 5.5900e-06 & 1.9916 & 1.0400e-06 & 2.0104 & 4.1900e-06 & 1.9844 & 3.6000e-07 & 2.2730 \\
\hline
\end{tabular}
\caption{\label{burgerShuosherTab} Convergence rate for test case corresponding to IC (\ref{nonlin-IC2c}) at pre-shock time $t=2/\pi$ $CFL=0.8$}
\end{table}
\subsection{Buckley Leverett Equation}
Consider Buckley-Leverett equation which has convex-concave flux. This equation physically represents
the flow of a mixture of oil and water through a porous medium.
\begin{equation}
\frac{\partial u}{\partial t} +\frac{\partial f(u)}{\partial x} =0\label{buckley}
\end{equation}
The flux function is given by,
\begin{equation}
f(u) = \frac{u^2}{u^2 + \alpha(1-u)^2}.
\end{equation}
Here $\alpha$ is viscosity ratio and $u$ represents the saturation of water and
lies between $0$ and $1.$
\subsubsection{One moving shock }
Consider equation (\ref{buckley}) with $\alpha =\frac{1}{2}$ and initial condition
\begin{equation}
u(x,\,0)= \left\{
\begin{array}{ll}
1, & x<0,\\
0, & x>0.
\end{array}
\right. \label{ICbuckley1}
\end{equation}
The solution involves one single moving shock followed by an rarefaction wave.
\subsubsection{Two moving shock}
Consider equation (\ref{buckley}) with $\alpha =\frac{1}{4}$ and
subject to initial condition
\begin{equation}
u(x,\,0)= \left\{
\begin{array}{ll}
1, & -0.5\leq x\leq 0,\\
0, & elsewhere.
\end{array}
\right. \label{ICbuckley2}
\end{equation}
The solution involves two moving shocks, each followed by an
rarefaction wave. In numerical simulation flux limited high resolution
LxW TVD scheme \cite{Rider} is used in hybrid scheme \ref{algo4} as
CCS. The results corresponding IC (\ref{ICbuckley1}) and
(\ref{ICbuckley2}) are given in Figure \ref{Figbuckley}(a) and
\ref{Figbuckley}(b) respectively. Results show that the proposed
scheme sharply captures both the fast and slow shocks. The rarefaction
waves are also approximated with high resolution.
\begin{figure}[!htb]
\begin{tabular}{cc}
\hspace{-1cm}\includegraphics[%
scale =0.55,angle =0]{test1buckley.pdf}&\hspace{-1cm}\includegraphics[%
scale =0.55,angle =0]{test2buckley.pdf}\\
(a) & (b)\\
\end{tabular}
\caption{\label{Figbuckley}{\it Numerical solution using $CFL=0.8$ (a) $N=80$ at time $t=0.75$ (b) $N=100$ at time $t=0.4$: Sharp
resolution for rarefaction fans and slow and fast moving shocks.}}
\end{figure}
\subsection{\bf 1D Euler Equation}
The 1D Euler equations of the Gas dynamics is given by
\begin{equation}
\frac{\partial}{\partial
t}\textbf{u}+\frac{\partial}{\partial x}\textbf{F(u)}=0, \label{euler1d}
\end{equation}
where $\textbf{u}=\left(\begin{array}{c}\rho\\ \rho
u\\ E\end{array}\right)$ and
$\textbf{F(u)}=\left(\begin{array}{c}\rho u\\ \rho
u^2+p\\ (E+p)u \end{array}\right)$ denotes vector of conservative
variables and conservative fluxes respectively. Variables $\rho,u $
and $p$ represents density, velocity and pressure respectively . The
total energy $e$ is defined by,
\begin{equation}
e=\frac{p}{\gamma-1}+\frac{\rho u^2}{2}
\end{equation}
where $\gamma$ is the ratio of specific heat coefficients. We consider
the four shock tube problems modeled by (\ref{euler1d}) to check the
robustness of proposed scheme in section \ref{algo4sys}. These shock
tube tests check any method in capturing the contact and shock
discontinuity along with non-oscillatory high resolution approximation
for smooth extrema. In all the numerical test a simple high resolution
TVD flux limited centered (FLIC) \cite{toro2000,toro2009} scheme with
MINBEE limiter is used as CCS in \ref{algo4}. We denote results by
this scheme by FLWBW-FLIC instead SC-FLWBW-FLIC. Numerical results are
compared with FLIC to see the improvement in capturing the solution
profile by FLWBW-FLIC. Note that, the MINBEE limiter satisfies the
universal TVD stability region given in \cite{Dubey2013} and therefore
robustly works for both positive and negative characteristics speed
associated with system (\ref{euler1d})
\subsubsection{Shu-Osher shock tube test \cite{shuosher}}
\begin{equation}
(\rho,u,p)=\begin{array}{ll} (3.857143,2.629369,10.3333) & x<-4.0,\\
(1+0.2\sin(5x),0,1) & x\geq -4.0. \label{1dEulerIC1}
\end{array}
\end{equation}
This test depicts shock interaction with a sine wave in density. The
main challenge in this case is to capture both the complex small-scale
smooth flow and shocks. In Figure \ref{FigShuosher1DEuler} results are
presented an compared with FLIC scheme. It is evident from zoomed
figure \ref{FigShuosher1DEuler}(b) that the FLWBW-FLIC yields
oscillation free approximation for shock with higher resolution
compared to FLIC for complex oscillatory solution region about
$[0.5,2.5]$. It also capture the smooth region in around $[-3,0.5]$
with out clipping or flattening error which is due to improved
approximation of smooth extrema and steep gradient region.
\begin{figure}[!htb]
\begin{tabular}{cc}
\hspace{-1cm}\includegraphics[scale=0.55]{test4shockparameter2rho.pdf}&\hspace{-1cm}\includegraphics[scale=0.55]{zoomtest4rho.pdf}\\
(a) & (b)\\
\end{tabular}
\caption{\label{FigShuosher1DEuler} Numerical solution of shock
entropy wave interaction $CFL=0.8, N=800$ using shock switch
parameters $\epsilon=1\times10^{-8}, \delta=0.8$: high resolution of
smooth extrema and steep gradient region.}
\end{figure}
\subsubsection{Sod test tube}
\begin{equation}
(\rho,u,p)=\begin{array}{ll} (1\; kg/m^3 , 0\; m/s, 100,000\; N/m^2) & x<0\\
(0.125\; kg/m^3, 0\;m/s , 10,000\; N/m^2) & x\geq 0; \label{SodIC}
\end{array}, x\in[-10,10].
\end{equation}
This test problem has no sonic point but the contact and shock are
very close which cause a smeared approximation to the middle contact
discontinuity. In Figure \ref{Fig1Sod}, numerical results are given and
for different choice of shock switch parameters and compared with
FLIC. Results show that proposed FLWBW-FLIC crisply captures the
smooth rarefaction and contact discontinuity and shock more accurately
than high order TVD scheme FLIC with Minbee limiter.
\begin{figure}[!htb]
\begin{tabular}{cc}
\includegraphics[scale=0.4]{test1shockparameter3rho.pdf} &
\includegraphics[scale=0.4]{test1shockparameter2rho.pdf}\\
\includegraphics[scale=0.4]{test1shockparameter3P.pdf}&
\includegraphics[scale=0.4]{test1shockparameter2P.pdf}\\
\includegraphics[scale=0.4]{test1shockparameter3U.pdf}&
\includegraphics[scale=0.4]{test1shockparameter2U.pdf}\\
\includegraphics[scale=0.4]{test1shockparameter3E.pdf}&
\includegraphics[scale=0.4]{test1shockparameter2E.pdf}\\
(a) & (b) \\
\end{tabular}
\caption{\label{Fig1Sod} Solution Sod shock tube, $N=100, CFL=0.5$
after $67$ time steps at $T= 0.01$ second with shock switch
parameters $\epsilon=1\times10^{-8}$, (a) $\delta=0.4$, (b) $\delta=0.9$ }
\end{figure}
\subsubsection{Lax Tube}
\begin{equation}
(\rho,u,p)=\begin{array}{ll} (0.445\; kg/m^3 , 0.698\; m/s, 3.528\; N/m^2) & x<1,\\
(0.5\; kg/m^3, 0\;m/s , 0.571\; N/m^2) & x\geq 1; \label{LaxIC}
\end{array} x\in[0,2].
\end{equation}
Compared to Sod tube the shock in this case is very strong and mostly
use to check the robustness of any schemes. In Figure \ref{FigLax}
numerical results obtained by FLWBW-FLIC are given. It can be seen
that that the method capture the contact and the rarefaction wave with higher resolution
shock compared to FLIC for various choices of shock parameters.
\begin{figure}[!htb]
\begin{tabular}{ccc}
\raisebox{3cm}{\bf (a)} & \includegraphics[scale=0.36]{test3shockparameter1rho.pdf}& \includegraphics[scale=0.36]{test3shockparameter1E.pdf}\\
\raisebox{3cm}{\bf (b)} & \includegraphics[scale=0.36]{test3shockparameter4rho.pdf} & \includegraphics[scale=0.36]{test3shockparameter4E.pdf}\\
\raisebox{3cm}{\bf (c)}& \includegraphics[scale=0.36]{test3shockparameter3rho.pdf}& \includegraphics[scale=0.36]{test3shockparameter3E.pdf}\\
\raisebox{3cm}{\bf (d)}& \includegraphics[scale=0.36]{test3shockparameter2rho.pdf}& \includegraphics[scale=0.36]{test3shockparameter2E.pdf}\\
\end{tabular}
\caption{\label{FigLax} Density and energy plots Lax Shock tube case 2, $N=200,
CFL=0.8$ after $187$ time steps at $T= 0.32$ second with shock
switch parameters in row (a) $\epsilon=1\times10^{-2}, \delta=0.2$, (b)
$\epsilon=1\times 10^{-2}, \delta=0.8$, (c) $\epsilon=1\times
10^{-4}, \delta=0.4$ and (d) $\epsilon=1\times 10^{-8}, \delta=0.8$.}
\end{figure}
\subsubsection{Laney Test \cite{Laney}}
\begin{equation}
(\rho,u,p)=\begin{array}{ll} (1\; kg/m^3 , 0\; m/s, 100,000\; N/m^2) & x<0,\\
(0.01\; kg/m^3, 0\;m/s , 1,000\; N/m^2) & x\geq 0; \label{LaneyIC}
\end{array} x\in[-10,15].
\end{equation}
In this test the density and pressure state on the right side of
initial discontinuity is much smaller compared to the left
state. Therefore computationally, even small oscillations can lead to negative
density or pressure which results in to non-physical imaginary speed
of sound $c=\sqrt{\frac{\gamma p}{\rho}}$. This makes it an important
test to check the non-oscillatory nature of any numerical scheme. In
Figure \ref{FigLaney}, the density and pressure plots obtained by
FLWBW-FLIC are given and compared for shock switch parameters
$\epsilon,\delta$.
\begin{figure}[!htb]
\begin{tabular}{cc}
\includegraphics[scale=0.55]{test2shockparameter4rho.pdf}&\includegraphics[scale=0.55]{test2shockparameter5rho.pdf}\\
\includegraphics[scale=0.55]{test2shockparameter4P.pdf}&\includegraphics[scale=0.55]{test2shockparameter5P.pdf}\\
\includegraphics[scale=0.55]{test2shockparameter4E.pdf}&\includegraphics[scale=0.55]{test2shockparameter5E.pdf}\\
(a) & (b)\\
\end{tabular}
\caption{\label{FigLaney} Solution Laney test, $N=200, CFL=0.8$
after $111$ time steps at $T= 0.01$ second with shock switch
parameters $\epsilon=1e-8$ (a) $\delta=0.6$ (b) $\delta=0.9$ }
\end{figure}
Numerical results for 1D shock tube test problems show that the
corners of rarefaction and contact discontinuities are better resolved
by FLWBW-FLIC compared to centred flux limiter shock capturing TVD
scheme FLIC. The shock is also crisply captured compared to to FLIC
with Minbee limiter. The hybrid scheme presented in section
\ref{algo4sys} is robust and works for different choices of shock
parameters $\epsilon,\; \delta$.
\subsection{2D Euler Equation}
Consider the two-dimensional Euler equations for compressible gas
dynamics defined by the system
\begin{equation}
U_t+F(U)_x+G(U)_y=0,\label{euler2d}
\end{equation}
where \begin{equation}
U=\left( \begin{array}{cccc}
\rho \\
\rho u \\
\rho v \\
e \end{array} \right),
~ F=\left( \begin{array}{cccc}
\rho u\\
\rho u^2+p\\
\rho uv\\
u(e+p)\end{array} \right),
~ G=\left( \begin{array}{cccc}
\rho v\\
\rho uv\\
\rho v^2+p\\
v(e+p) \end{array} \right)
\end{equation}
Here $\rho$ is the density, $u$ and $v$ are velocity components in $x$ and $y$
direction respectively, $p$ is the pressure and $e$ is
the energy defined by,
\begin{equation}
e=\frac{p}{\gamma-1}+\frac{\rho(u^2+v^2)}{2}.
\end{equation}
The Riemann problem for Euler equation (\ref{euler2d}) can be defined by
considering the constant initial data in each quadrant of unit square
$[0,1]\times[0,1]$ with center at $(0.5,0.5)$. More precisely consider
(\ref{euler2d}) with initial data\\
\begin{equation}
(p,\rho,u,v)(x,y,t=0) = \left\{\begin{array}{cc}
(p_{1}, \rho_{1}, u_{1}, v_{1}) & \mbox{if}\; x>0.5, y>0.5\\
(p_{2}, \rho_{2}, u_{2}, v_{2}) & \mbox{if}\; x<0.5, y>0.5\\
(p_{3}, \rho_{3}, u_{3}, v_{3}) & \mbox{if}\; x<0.5, y<0.5\\
(p_{4}, \rho_{4}, u_{4}, v_{4}) & \mbox{if}\; x>0.5, y<0.5\\
\end{array}\right.
\end{equation}
In the 2D Riemann problems due to complex geometric wave pattern most
high resolution schemes experience problems in yielding oscillations
free crisp resolution to solution profile. Such 2D Riemann problem are
numerically solved by using positive scheme and Riemann solvers free
central schemes in \cite{LaxLiu1998} and \cite{Kurganov2002}
respectively. Recently some of the these Riemann problems are
considered to see the performance of a new finite volume adaptive
artificial viscosity method in \cite{Kurganov2012} and (with slight
changed geometry) a HLL Riemann solver in \cite{Vides2015643}
respectively. In this section numerical results are given for twelve
configurations. The one dimensional scheme presented in section
\ref{algo4sys} is extended to two dimensional Euler equation using the
Strang dimension by dimension splitting technique. In Figure
\ref{2DRP1} to Figure \ref{2DRP14} the contour plot of density are
given and compared with FORCE scheme for different test cases. In all
the figures contour plot by FORCE is given in column (a), and by
FLWBW-FORCE in column (b) with shock parameters $\epsilon =1\times
10^{-8}, \delta=0.6$. It is observed that small oscillations can occur
with if a flux limited TVD scheme is used in hybrid scheme in section
\ref{algo4sys} (Results are not shown here). Note that it is in
agreement with the comments in \cite{Kurganov2002}, that a over
compressive Minmod type limiter can lead to spurious oscillations for
the schemes proposed therein.
\begin{enumerate}
\item[] Configuration 1
\begin{equation}
\begin{array}{llll}
p_1=1 &p_2=0.4 &p_3=0.0439 &p4=0.15\\
\rho_1=1 &\rho_2=0.5197 &\rho_3=0.1072 &\rho_4=0.2579\\
u_1=0 &u_2=-0.7259 &u_3=-0.7259 &u_4=0.0\\
v_1=0 &v_2=0 &v_3=-1.4045 &v_4=-1.4045\\
\end{array}
\end{equation}
\begin{figure}[!htb]
\begin{tabular}{cc}
\includegraphics[scale=0.5]{FORCE1CL30.pdf} & \hspace{1cm} \includegraphics[scale=0.5]{methodwithFORCEconfig1ss9CL30.pdf}\\
(a)& (b)\\
\end{tabular}
\caption{\label{2DRP1} Configuration 1: Density contour plot (30 lines)
at $T=0.2$, sharp resolution for rarefaction and ripple in lower
rarefaction are sharply resolved in (b) compared to FORCE (a).}
\end{figure}
\item[] Configuration 2
\begin{equation}
\begin{array}{llll}
p_1=1 &p_2=0.4 &p_3=1 &p_4=0.4 \\
\rho_1=1 &\rho_2=0.5197 &\rho_3=1 &\rho_4=0.5197 \\
u_1=0 &u_2=-0.7259 &u_3=-0.7259 &u_4=0 \\
v_1=0 &v_2=0 &v_3=-0.7259 &v_4=-0.7259
\end{array}
\end{equation}
\begin{figure}[!htb]
\begin{tabular}{cc}
\includegraphics[scale=0.5]{FORCE2CL30.pdf} &\hspace{1cm} \includegraphics[scale=0.5]{methodwithFORCEconfig2ss9CL30.pdf}\\
(a)& (b)\\
\end{tabular}
\caption{\label{2DRP2} Configuration 2: Density contour plot (30
lines) at $T=0.25$ non-oscillatory sharp resolution for rarefaction
waves and corners in (b) compared to FORCE in (a)}
\end{figure}
\item[] Configuration 3
\begin{equation}
\begin{array}{llll}
p_1=1.5 &p_2=0.3 &p_3=0.029 &p_4=0.3\\
\rho_1=1.5 &\rho2=0.5323 &\rho_3=0.138 &\rho4=0.5323\\
u_1=0 &u_2=1.206 &u_3=1.206 &u_4=0\\
v_1=0 &v_2=0 &v_3=1.206 &v_4=1.206
\end{array}
\end{equation}
\begin{figure}[!htb]
\begin{tabular}{cc}
\includegraphics[scale=0.5]{FORCE3CL32.pdf} & \hspace{1cm}\includegraphics[scale=0.5]{methodwithFORCEconfig3ss9CL32.pdf}\\
(a)& (b)\\
\end{tabular}
\caption{\label{2DRP3} Configuration 3: Density contour plot (32 lines), All four shocks are sharply resolved with out induced oscillations in FLWBW-FORCE in (b) compared to FORCE in (a).}
\end{figure}
\item[] Configuration 4
\begin{equation}
\begin{array}{llll}
p_1=1.1 &p_2=0.35 &p_3=1.1 &p_4=0.35\\
\rho_1=1.1 &\rho_2=0.5065 &\rho_3=1.1 & \rho_4=0.5065\\
u_1=0 &u_2=0.8939 &u_3=0.8939 &u_4=0\\
v_1=0 &v_2=0 &v_3=0.8939 &v_4=0.8939
\end{array}
\end{equation}
\begin{figure}[!htb]
\begin{tabular}{cc}
\includegraphics[scale=0.5]{FORCE4CL30.pdf} & \hspace{1cm} \includegraphics[scale=0.5]{methodwithFORCEconfig4ss9CL30.pdf}\\
(a)& (b)\\
\end{tabular}
\caption{\label{2DRP4} Configuration 4: Density contour plot (30
lines) non-oscillatory, FLWBW-FORCE yield sharp resolution for shocks in (b) compared to FORCE in (a)}
\end{figure}
\item[] Configuration 5
\begin{equation}
\begin{array}{llll}
p_1=1 &p_2=1 &p_3=1 &p_4=1\\
\rho_1=1 &\rho_2=2 &\rho_3=1 & \rho_4=3\\
u_1=-0.75 &u_2=-0.75 &u_3=0.75 &u_4=0.75\\
v_1=-0.5 &v_2=0.5 &v_3=0.5 &v_4=-0.5
\end{array}
\end{equation}
\begin{figure}[!htb]
\begin{tabular}{cc}
\includegraphics[scale=0.5]{FORCE5CL30.pdf} & \hspace{1cm} \includegraphics[scale=0.5]{methodwithFORCEconfig5ss9CL30.pdf}\\
(a)& (b)\\
\end{tabular}
\caption{\label{2DRP5} Configuration 5: Density plot (30 Lines) all four contacts are poorly resolved
by FORCE (a), though FLWBW-FORCE yield sharp resolution to contacts with out oscillations in (b).}
\end{figure}
\item[] Configuration 6
\begin{equation}
\begin{array}{llll}
p_1=1 &p_2=1 &p_3=1 &p_4=1\\
\rho_1=1 &\rho_2=2 &\rho_3=1 &\rho_4=3\\
u_1=0.75 &u_2=0.75 &u_3=-0.75 &u_4=-0.75\\
v_1=-0.5 &v_2=0.5 &v_3=0.5 &v_4=-0.5
\end{array}
\end{equation}
\begin{figure}[!htb]
\begin{tabular}{cc}
\includegraphics[scale=0.5]{FORCE6CL32.pdf} &\hspace{1cm} \includegraphics[scale=0.5]{methodwithFORCEconfig6ss9CL32.pdf} \\
(a)& (b)\\
\end{tabular}
\caption{\label{2DRP6} Configuration 6: The ripples in NE and SW
quadrants are captured with comparable resolution with the one in
\cite{Kurganov2002} using FLWBW-FORCE (b) though the resolution for
contacts is little diffusive but much sharper compared to FORCE (a).}
\end{figure}
\item[] Configuration 7
\begin{equation}
\begin{array}{llll}
p_1=1 &p_2=0.4 &p_3=0.4 &p_4=0.4\\
\rho_1=1 &\rho_2=0.5197 &\rho_3=0.8 &\rho_4=0.5197\\
u_1=0.1 &u_2=-0.6259 & u_3=0.1 & u_4=0.1\\
v_1=0.1 &v_2=0.1 &v_3=0.1 &v_4=-0.6259\\
\end{array}
\end{equation}
\begin{figure}[!htb]
\begin{tabular}{cc}
\includegraphics[scale=0.5]{FORCE7CL30.pdf} &\hspace{1cm} \includegraphics[scale=0.5]{methodwithFORCEconfig7ss9CL30.pdf}\\
(a)& (b)\\
\end{tabular}
\caption{\label{2DRP7} Configuration 7: The contacts in South and West
are crisply resolved by FLWBW-FORCE (b). Moreover the rarefaction
corners in NE quadrants are significantly sharper than FORCE (a).}
\end{figure}
\item[] Configuration 8
\begin{equation}
\begin{array}{llll}
p_1=0.4 &p_2=1 &p_3=1 &p_4=1\\
\rho_1=0.5197 &\rho_2=1 &\rho_3=0.8 &\rho_4=1\\
u_1=0.1 &u_2=-0.6259 &u_3=0.1 &u_4=0.1\\
v_1=0.1 &v_2=0.1 &v_3=0.1 &v_4=-0.6259
\end{array}
\end{equation}
\begin{figure}[!htb]
\begin{tabular}{cc}
\includegraphics[scale=0.5]{FORCE8CL30.pdf} & \hspace{1cm} \includegraphics[scale=0.5]{methodwithFORCEconfig8ss9CL30.pdf} \\
(a)& (b)\\
\end{tabular}
\caption{\label{2DRP8} Configuration 8: The semi-circular wave front in NE is sharply resolved and resolution for contacts are comparable with the one in \cite{Kurganov2002}.}
\end{figure}
\item[] Configuration 9
\begin{equation}
\begin{array}{llll}
p_1=1& p_2=1& p_3=0.4& p_4=0.4\\
\rho1=1& \rho2=2& \rho3=1.039& \rho4=0.5197\\
u_1=0& u_2=0.0& u_3=0& u_4=0.0\\
v_1=0.3& v_2=-0.3& v_3=-0.8133& v_4=-0.4259\\
\end{array}
\end{equation}
\begin{figure}[!htb]
\begin{tabular}{cc}
\includegraphics[scale=0.5]{FORCE9CL30.pdf} &\hspace{1cm} \includegraphics[scale=0.5]{methodwithFORCEconfig9ss9CL30.pdf} \\
(a)& (b)\\
\end{tabular}
\caption{\label{2DRP9} Configuration 9: Again rarefaction and corners are resolved sharper than FORCE (a). The vertical contact is crisply captured by FLWBW-FORCE.}
\end{figure}
\item[] Configuration 10
\begin{equation}
\begin{array}{llll}
p_1=1 &p_2=1 &p_3=0.3333 &p_4=0.3333\\
\rho_1=1 &\rho_2=0.5 &\rho_3=0.2281 &\rho_4=0.4562\\
u_1=0 &u_2=0 &u_3=0 &u_4=0\\
v_1=0.4297 &v_2=0.6076 &v_3=-0.6076 &v_4=-0.4297
\end{array}
\end{equation}
\begin{figure}[!htb]
\begin{tabular}{cc}
\includegraphics[scale=0.5]{FORCE10CL32.pdf} & \hspace{1cm}\includegraphics[scale=0.5]{methodwithFORCEconfig10ss9CL32.pdf} \\
(a)& (b)\\
\end{tabular}
\caption{\label{2DRP10} Configuration 10: Resolution of contacts by FLWBW-FORCE is comparable with the one in \cite{Kurganov2002}.}
\end{figure}
\item[] Configuration 11
\begin{equation}
\begin{array}{llll}
p_1=1 &p_2=0.4 &p_3=0.4 &p_4=0.4\\
\rho_1=1 &\rho_2=0.5313 &\rho_3=0.8 &\rho_4=0.5313\\
u_1=0.1 &u_2=0.8276 &u_3=0.1 &u_4=0.1\\
v_1=0.0 &v_2=0.0 &v_3=0.0 &v_4=0.7276
\end{array}
\end{equation}
\begin{figure}[!htb]
\begin{tabular}{cc}
\includegraphics[scale=0.5]{FORCE11CL32.pdf} & \hspace{1cm} \includegraphics[scale=0.5]{methodwithFORCEconfig11ss9CL32.pdf} \\
(a)& (b)\\
\end{tabular}
\caption{\label{2DRP11} Configuration 11: Resolution of contact is SW quadrant, Shocks in SE and NW quadrants by FLWBW-FORCE (b) is better than that of the one in \cite{Kurganov2002, LaxLiu1998}.}
\end{figure}
\item[] Configuration 12
\begin{equation}
\begin{array}{llll}
p_1=0.4 & p_2=1& p_3=1.0& p_4=1.0\\
\rho1=0.5313& \rho2=1.0222& \rho3=0.8& \rho4=1.0\\
u_1=0.1& u_2=-0.6179& u_3=0.1& u_4=0.1\\
v_1=0.1 & v_2=0.1& v_3=0.1& v_4=0.8276
\end{array}
\end{equation}
\begin{figure}[!htb]
\begin{tabular}{cc}
\includegraphics[scale=0.5]{FORCE14CL32.pdf} &\hspace{1cm} \includegraphics[scale=0.5]{methodwithFORCEconfig14ss9CL32.pdf} \\
(a)& (b)\\
\end{tabular}
\caption{\label{2DRP14} Configuration 12: FLWBW-FORCE recovers the
ripples between NE shock and contact waves. The resolution for shock
and contacts are comparable with the second order scheme results in
\cite{Kurganov2002}.}
\end{figure}
\end{enumerate}
\textbf{Comments}
\section{Conclusion and Future work}\label{sec7}
In this work LMP/TVD bounds are obtained for uniformly second order
accurate schemes in non-conservative form. These bound show that
higher than second order TVD accuracy can be achieved at extrema and
steep gradient region in limiting sense i.e., when $r\rightarrow
0^{-}$. Based on the LMP/TVD bounds hybrid local maximum principle
satisfying schemes are constructed and applied on various benchmark
test problems. Numerical results show improvement in TVD approximation
of solution region with extreme points, smooth rarefaction as well
contact discontinuity compared to existing higher order TVD
method. For a separate work, the focus is on TVD bounds for multi-step
methods and efficient use of a shock detector. As, the algorithm
\ref{algo4} recovered the shock at right location for scalar case, it
would be interesting to devise a hybrid scheme for system by modifying
the wave speed choice in section \ref{algo4sys}, with out a shock
detector.
|
\section{Introduction}
\label{sec:Intro}
Previous works \cite{Kevin_2011_CDC, Galloway_PRS_13, Marshall_TAC_04, Kim20071426, Smith20051045, Marshall20063, 5160735} have demonstrated pertinence of dyadic pursuit interactions as a building block for collective control. \editKG{In \cite{Kevin_2011_CDC, Galloway_PRS_13}, the particular focus} is on a cyclic pursuit scheme in which each agent employs a constant bearing (CB) pursuit law \cite{Wei_Justh_PSK_09} with regards to exactly one other agent in the collective. In that context, it was demonstrated (for a range of control parameters) that the closed-loop dynamics admit circling relative equilibria (among other special solutions), with the corresponding formation shape determined by the choice of control parameters. However, both the location of the circumcenter (with respect to an inertial frame) and the radius of the circular orbit are determined by initial conditions rather than control parameters, which limits the effectiveness of the control methodology from a design perspective.
In the current work, we introduce a modified version of the CB control law, in which the pursuer is attentive to both a neighboring agent as well as to a beacon (which is assumed to be fixed in the current setting). Though the control law is not designed to stabilize a particular station-keeping range from the beacon, we demonstrate that in an $n$-agent collective where each agent $i$ employs this ``beacon-referenced'' CB control law with respect to agent $i+1$ and a common beacon, circling equilibria exist which are centered on the beacon position and have a radius determined by the control parameters. For the case where $n=2$, we analyze the stability of the associated circling equilibria by linearization of the dynamics, deriving stability conditions in terms of the control parameters.
While our approach is motivated by the numerous robotic station-keeping applications which require autonomous agents to orbit a specified location while maintaining a fixed formation shape and scale (e.g. search and rescue, environmental sensing, etc.), we also note that this work may provide insights into the mechanisms underlying collective behavior observed in nature. For example, the beacon-referenced cyclic pursuit analyzed in this paper may provide tools for modeling the ``explore-exploit'' behavior evidenced by some animal collectives (e.g. honeybees \cite{Seeley_Behav_Ecol_91}) searching for food sources.
\section{Modeling the interaction}
\label{sec:Model}
\subsection{Modeling the system}
As presented in \cite{Justh03steeringlaws, Justh_PSK_SCL04}, we model an agent as a unit-mass self-steering particle with twice-differentiable motion path in $\mathds{R}^2$. This allows us to use natural Frenet frame \cite{Nat_Frenet_Bishop} equations to describe the motion for a group of $n$ agents. By letting $\mathbf{r}_i$ denote the position of the $i$-th agent, underlying system dynamics can be expressed as
\begin{equation}
\begin{array}{rcl}
\dot{\mathbf{r}}_i &=& \nu_i \mathbf{x}_i \\
\dot{\mathbf{x}}_i &=& \nu_i u_i \mathbf{y}_i \\
\dot{\mathbf{y}}_i &=& - \nu_i u_i \mathbf{x}_i, \quad i = 1,2,\ldots,n.
\end{array} \label{Explicit_MODEL}
\end{equation}
Here $\mathbf{x}_i$ is the normalized velocity and $\mathbf{y}_i$ represents orthogonally rotated $\mathbf{x}_i$ in the counter-clockwise direction. Moreover, $\nu_i$ is the speed, and $u_i$ is the natural curvature viewed as a steering control. Alternatively, by packing $\mathbf{r}_i, \mathbf{x}_i, \mathbf{y}_i$ inside a matrix
\begin{equation}
g_i
\triangleq
\left[ \begin{array}{ccc}
\mathbf{x}_i & \mathbf{y}_i & \mathbf{r}_i \\
0 & 0 & 1
\end{array} \right]
\in SE(2),
\label{GrP_Defn_SE2}
\end{equation}
the natural Frenet frame equations \eqref{Explicit_MODEL} can be expressed as a left invariant dynamics on $SE(2)$. From this perspective, the dynamics of an agent can be expressed as
\begin{equation}
\dot{g}_i = g_i\xi_i = \nu_i g_i (X_0 + u_iX_2),
\label{Implicit_MODEL}
\end{equation}
where $X_0$ and $X_2$ represent standard basis elements of $\mathfrak{se}(2)$.
As discussed earlier, practical applications often require the collective of agents to maneuver with respect to some particular desired location. Therefore, in this work we introduce a beacon at location $\mathbf{r}_b \in \mathds{R}^2$, along with a fixed frame $[\mathbf{x}_b \; \mathbf{y}_b]$ attached to it (which is assumed to be the inertial reference frame, without loss of generality), and define $g_b \in SE(2)$ as in \eqref{GrP_Defn_SE2} to pack $\mathbf{r}_b$, $\mathbf{x}_b$ and $\mathbf{y}_b$ inside a single matrix.
\subsection{Directed graph and commutativity constraints}
Now we introduce the notion of an \textit{attention graph} \cite{Galloway_PRS_13} to describe which agent(s) a particular agent is paying attention to. By letting $\mathcal{N} = \{1,2,\ldots,n,b\}$ denote the node-set for the problem of our concern, the corresponding arc/edge-set ($\mathcal{A}$) is defined as
\begin{equation}
\mathcal{A}
=
\big\{
(i,i+1), (i,b)
\big|
i = 1,2,\ldots ,n
\big\},
\end{equation}
where addition in the index variables should be interpreted modulo $n$ throughout this paper. Clearly, this attention graph $\mathcal{G} = (\mathcal{N}, \mathcal{A})$ is weakly connected and devoid of any self loop.
Now, we formulate a reduction to the shape space, and introduce the following set of variables (possibly redundant) along arcs of the attention graph $\mathcal{G}$:
\begin{align}
\tilde{g}_{i,i+1} &= g_{i+1}^{-1} g_i
\label{GrP_Defn_Shape_1}
\\
\textrm{and,} \qquad
\tilde{g}_{ib} &= g_b^{-1} g_i.
\label{GrP_Defn_Shape_2}
\end{align}
It follows from the definition that $\tilde{g}_{i,i+1}$, $i = 1,2,\ldots,n$ are subject to the cycle closure constraint
\begin{equation}
\prod\limits_{i=1}^n \tilde{g}_{i,i+1}
=
\tilde{g}_{n1}\tilde{g}_{n-1,n} \cdots \tilde{g}_{23} \tilde{g}_{12}
=
\mathds{I}_3,
\label{constraint_CYCLE}
\end{equation}
where $\mathds{I}_3$ is the $3\times 3$ identity matrix. (See also \cite{Galloway_PRS_13}.) Moreover, by exploiting the commutative property, we have
\begin{equation}
\tilde{g}_{i,i+1} = \tilde{g}_{i+1,b}^{-1} \tilde{g}_{ib}, \qquad i = 1,2,\ldots,n-1.
\label{constraint_GROUP}
\end{equation}
This set of equations \eqref{constraint_GROUP} poses consistency conditions on the space of shape variables. We should note here that together \eqref{constraint_CYCLE} and \eqref{constraint_GROUP} ensure $\tilde{g}_{n1} = \tilde{g}_{1b}^{-1} \tilde{g}_{nb}$.
Also, from \eqref{Implicit_MODEL}, \eqref{GrP_Defn_Shape_1} and \eqref{GrP_Defn_Shape_2}, it follows that the shape dynamics can be expressed as
\begin{equation}
\begin{aligned}
\dot{\tilde{g}}_{ib} &= \tilde{g}_{ib} \xi_i,
\\
\textrm{and} \quad \dot{\tilde{g}}_{i,i+1} &= \tilde{g}_{i,i+1} \tilde{\xi}_{i,i+1},
\end{aligned}
\label{Shape_dynamics_Implicit}
\end{equation}
where $\tilde{\xi}_{i,i+1} = \xi_i - \tilde{g}^{-1}_{i,i+1} \xi_{i+1} \tilde{g}_{i,i+1}$. It is a straightforward exercise to show that the constraints \eqref{constraint_CYCLE}-\eqref{constraint_GROUP} will be satisfied for all future time if they are satisfied initially.
\subsection{Scalar shape variables}
Now, following the approach of earlier works \cite{Galloway_PRS_13}, we introduce scalar shape variables (a polar parametrization) to describe the state of an agent relative to the beacon and other agents. By letting \begin{equation}
R(\vartheta)
=
\left[
\begin{array}{cc}
\cos\vartheta & -\sin\vartheta \\ \sin\vartheta & \cos\vartheta
\end{array}\right]
\in SO(2)
\end{equation}
denote a counter-clockwise planar rotation through an angle $\vartheta$, we define a set of scalar shape variables $\rho_i$, $\rho_{ib}$, $\kappa_i$, $\theta_i$, $\kappa_{ib}$ and $\psi_i$ as
\begin{align}
& \rho_i = |\mathbf{r}_{i+1,i}|,
\qquad
&& \rho_{ib} = |\mathbf{r}_{b,i}|,
\nonumber \\
& R(\kappa_i)\mathbf{x}_i = \frac{\mathbf{r}_{i+1,i}}{|\mathbf{r}_{i+1,i}|},
\qquad
&& R(\theta_i)\mathbf{x}_i = -\frac{\mathbf{r}_{i,i-1}}{|\mathbf{r}_{i,i-1}|}
\label{Scalar_Shape_DEFN} \\
& R(\kappa_{ib})\mathbf{x}_i = \frac{\mathbf{r}_{b,i}}{|\mathbf{r}_{b,i}|},
&& R(\pi - \psi_i)\mathbf{x}_b = \frac{\mathbf{r}_{b,i}}{|\mathbf{r}_{b,i}|},
\nonumber
\end{align}
where $i=1,2,\ldots,n$ and $\mathbf{r}_{i,j} = \mathbf{r}_{i} - \mathbf{r}_{j}$ represents the position of $i$-th agent relative to the $j$-th agent (see Fig~\ref{Scalar_Shapes}).
\begin{figure}[h]
\begin{center}
\includegraphics[width=0.48\textwidth]{scalar_Shape_CPCB}
\caption{Illustration of the scalar shape variables ($\rho_i$, $\rho_{ib}$, $\theta_i$, $\kappa_i$ and $\kappa_{ib}$) used in our analysis of an $n$-agent system with a fixed beacon at location $\mathbf{r}_b$.}
\label{Scalar_Shapes}
\end{center}
\end{figure}
Using these recently introduced scalar shape variables, the consistency conditions \eqref{constraint_GROUP} can be expressed as
\begin{align}
&
R(\psi_i - \psi_{i+1})
=
R(\pi + \kappa_i - \theta_{i+1} + \kappa_{i+1,b} - \kappa_{ib})
\label{constraint_1_SHAPE}
\\
&
\rho_i \mathds{I}_2 = \rho_{ib}R(\kappa_{ib} - \kappa_i) + \rho_{i+1,b}R(\kappa_{i+1,b} - \theta_{i+1})
\label{constraint_2_SHAPE_final}
\end{align}
for $i = 1,2,\ldots,n-1$. In a similar way, the cycle closure constraint \eqref{constraint_CYCLE} can be expressed as
\begin{align}
&
R\big( \sum\limits_{i = 1}^{n} (\pi + \kappa_i - \theta_{i+1}) \big)
=
\mathds{I}_2
\label{Cycle_Closure_1}
\\
&
\sum\limits_{i=1}^n \rho_i R\big( \sum\limits_{j = 1}^{i} (\pi + \kappa_j - \theta_{j+1}) \big)
=
0.
\label{Cycle_Closure_2}
\end{align}
In actuality, certain simplifications allow us to consider only a subset of these variables and constraints. First, it is clear that for a particular choice of $\psi_1$, \eqref{constraint_1_SHAPE} provides an explicit representation for the scalar shape variables $\psi_i$, $i = 2,3,\ldots,n$ in terms of other shape variables, and therefore we can disregard those variables as well as constraint \eqref{constraint_1_SHAPE} from the following analysis. Furthermore, the choice of beacon frame $[\mathbf{x}_b \; \mathbf{y}_b]$ is arbitrary and the proposed feedback law (see Section~III) is invariant to any rotation of the beacon frame, and therefore there exists an $\mathcal{S}^1$ symmetry enabling us to exclude $\psi_1$ from further analysis. Finally, one can show that \eqref{constraint_2_SHAPE_final} together with \eqref{Cycle_Closure_1} and \eqref{Cycle_Closure_2} can be expressed more concisely as simply \eqref{constraint_2_SHAPE_final} (holding for $i=1,2,\ldots,n$) with \eqref{Cycle_Closure_1}. Therefore, the shape space can be parametrized by the scalar variables $\kappa_i, \kappa_{ib}, \theta_i, \rho_i, \rho_{ib}$, for $i=1,2,\ldots,n$, subject only to the positivity constraints on $\rho_{i}$ and $\rho_{ib}$ (necessary for the well-posedness of a feedback law introduced in the next section) and the constraints \eqref{constraint_2_SHAPE_final} - \eqref{Cycle_Closure_1} for $i=1,2,\ldots,n$, i.e.
\begin{align}
&
R\big( \sum\limits_{i = 1}^{n} (\pi + \kappa_i - \theta_{i+1}) \big)
=
\mathds{I}_2
\label{FINAL_Constraint_1}
\\
&
\rho_i \mathds{I}_2 = \rho_{ib}R(\kappa_{ib} - \kappa_i) + \rho_{i+1,b}R(\kappa_{i+1,b} - \theta_{i+1}).
\label{FINAL_Constraint_2}
\end{align}
Next we focus on representing the shape dynamics \eqref{Shape_dynamics_Implicit} in terms of scalar shape variables. By straightforward calculations, one can show that the shape dynamics are given in terms of the scalar variables by
\begin{align}
\dot{\rho}_i
&=
- \nu_i\cos\kappa_i - \nu_{i+1}\cos\theta_{i+1}
\nonumber \\
\dot{\kappa}_i
&=
- \nu_i u_i + \frac{1}{\rho_i}\big[ \nu_i\sin\kappa_i + \nu_{i+1}\sin\theta_{i+1}\big]
\nonumber \\
\dot{\theta}_i
&=
- \nu_i u_i + \frac{1}{\rho_{i-1}}\big[ \nu_{i-1} \sin\kappa_{i-1} + \nu_i \sin\theta_{i}\big]
\label{shape_dynamics_EXPLICIT} \\
\dot{\rho}_{ib}
&=
- \nu_{i}\cos\kappa_{ib}
\nonumber \\
\dot{\kappa}_{ib}
&=
- \nu_i u_i + \frac{\nu_i}{\rho_{ib}} \sin\kappa_{ib}, \; i=1,2,\ldots,n,
\nonumber
\end{align}
subject to the cycle closure constraint \eqref{FINAL_Constraint_1} and consistency conditions \eqref{FINAL_Constraint_2}.
\section{\editKG{A beacon-referenced CB pursuit law}}
\label{sec:CL_Dyna}
In this section we introduce a modified version of the CB pursuit law from \cite{Wei_Justh_PSK_09} that introduces an additional term referenced to the beacon bearing. We construct this feedback law as a convex combination of two fundamental building blocks, expressed as
\begin{equation}
u_i = (1 - \lambda)u_{CB}^i + \lambda u_B^i, \qquad \lambda \in [0,1]
\label{u_i_top_level}
\end{equation}
where $u_{CB}^i$ is given by the original CB pursuit law \cite{Wei_Justh_PSK_09} referenced to agent $i+1$, and $u_B^i$ represents the deviation from a desired bearing angle to the beacon. More specifically, we choose
\begin{align}
u_{CB}^i
&=
- \mu_i \left( R(\alpha_i)\mathbf{y}_i \cdot \frac{\mathbf{r}_{i,i+1}}{|\mathbf{r}_{i,i+1}|} \right)
- \frac{1}{\nu_i |\mathbf{r}_{i,i+1}|} \left( \frac{\mathbf{r}_{i,i+1}}{|\mathbf{r}_{i,i+1}|} \cdot R(\pi/2) \dot{\mathbf{r}}_{i,i+1} \right),
\label{u_CB_i}
\end{align}
with $\mu_i > 0$ being a control gain and the angle $\alpha_i \in S^1$ representing the desired offset between the $i$-th agent's heading and the current location of the $(i+1)$-th agent. For the beacon tracking component, we let
\begin{equation}
u_B^i
=
- \mu_i^b \left( R(\alpha_{ib})\mathbf{y}_i \cdot \frac{\mathbf{r}_{i,b}}{|\mathbf{r}_{i,b}|} \right),
\label{u_Beacon_i}
\end{equation}
where $\mu_i^b > 0$ is the corresponding control gain and the angle $\alpha_{ib} \in S^1$ is the desired offset between the current heading of the $i$-th agent and the bearing to the beacon location. Note that the neighbor tracking goal may conflict with the beacon referencing goal, and the parameter $\lambda$ maintains a balance between the beacon's influence and the influence of the neighboring agent $i+1$. In particular, for $\lambda = 0$, $u_i$ is simply the original CB pursuit law with no reference to the beacon.
In terms of scalar shape variables, the feedback law $u_i$ can be expressed as
\begin{align}
u_i
&=
\lambda \mu_i^b\sin(\kappa_{ib} - \alpha_{ib})
+
(1 - \lambda)\mu_i\sin(\kappa_i - \alpha_i)
+ \frac{1 - \lambda}{\rho_i}\left(\sin\kappa_i + \frac{\nu_{i+1}}{\nu_i}\sin\theta_{i+1}\right).
\label{u_i_shape}
\end{align}
\begin{remark}
As noted in \cite{Wei_Justh_PSK_09}, the last component of this feedback law \eqref{u_i_shape} can be interpreted as the angular speed at which the baseline between agent-$i$ and agent-$i+1$ is rotating around the $i$-th agent. Therefore it is plausible to evaluate the steering command $u_i$ without explicit measurement of distance between the agents, although it will require an appropriate sensing mechanism (mimicking the principle of compound eyes in visual insects).
\end{remark}
Before going into detailed analysis of relative equilibria, we introduce the following simplifying assumptions:
\begin{itemize}
\item[(A1)] The speed of the agents are equal and constant. Hence, without loss of generality, we can assume $\nu_i = 1$ for every $i \in \{1,2,\ldots,n\}$.
\item[(A2)] The controller gains ($\mu_i$ and $\mu_i^b$) are equal and common for all agents, i.e. $\mu_i = \mu_i^b = \mu $ for every $i \in \{1,2,\ldots,n\}$.
\item[(A3)] The bearing angles ($\alpha_{ib}$) with respect to the beacon are common for all agents, i.e. $\alpha_{ib} = \alpha_0 $ for every $i \in \{1,2,\ldots,n\}$.
\end{itemize}
Under assumptions (A1)-(A3), the closed loop shape dynamics \editKG{(\eqref{shape_dynamics_EXPLICIT} with \eqref{u_i_shape})} can be expressed as
\begin{align}
\dot{\rho}_i
&=
- \big( \cos\kappa_i + \cos\theta_{i+1} \big)
\nonumber \\
\dot{\kappa}_i
&=
- \mu \big[ (1 - \lambda) \sin(\kappa_i - \alpha_i) + \lambda \sin(\kappa_{ib} - \alpha_0) \big]
+ \frac{\lambda}{\rho_i} \big[ \sin\kappa_i + \sin\theta_{i+1}\big]
\nonumber \\
\dot{\theta}_i
&=
\dot{\kappa}_i
- \frac{1}{\rho_i}\big[ \sin\kappa_i + \sin\theta_{i+1}\big]
+ \frac{1}{\rho_{i-1}}\big[ \sin\kappa_{i-1} + \sin\theta_{i}\big]
\label{CL_dynamics_n_simplified} \\
\dot{\rho}_{ib}
&=
- \cos\kappa_{ib}
\nonumber \\
\dot{\kappa}_{ib}
&=
\dot{\kappa}_i
- \frac{1}{\rho_i}\big[ \sin\kappa_i + \sin\theta_{i+1}\big]
+ \frac{1}{\rho_{ib}} \sin\kappa_{ib}
\nonumber
\end{align}
for $i = 1,2,\ldots,n$, subject to the constraints \eqref{FINAL_Constraint_1}-\eqref{FINAL_Constraint_2}.
\section{Relative equilibria}
\label{sec:Rel_EQ_exist}
In this section we analyze the closed loop shape dynamics \eqref{CL_dynamics_n_simplified} to determine existence conditions and characterization of their equilibria (i.e. relative equilibria for the full dynamics \eqref{Explicit_MODEL} with \eqref{u_i_top_level}). At the extreme value of $\lambda = 0$, the shape dynamics simplify to the cyclic CB pursuit dynamics previously analyzed in \cite{Galloway_PRS_13}, while at the other extreme, inter-agent interaction is completely lost whenever $\lambda = 1$. Therefore we restrict $\lambda$ to lie in the open interval $(0,1)$ for the rest of our analysis.
From the form of $\dot{\rho}_{ib}$ and $\dot{\rho}_i$ in \eqref{CL_dynamics_n_simplified}, we can obtain necessary conditions at equilibrium given by
\begin{equation}
\kappa_{ib} = \pm \frac{\pi}{2},
\quad \textrm{and,} \quad
\theta_{i+1} = \pi \pm \kappa_i,
\label{eq_kappa_i_ib}
\end{equation}
for $i = 1,2,\ldots,n$. Similarly, by setting the dynamics of $\theta_i$ and $\kappa_{ib}$ equal to zero, we obtain (for $i = 1,2,\ldots,n$)
\begin{align}
\frac{1}{\rho_i}(\sin\kappa_i + \sin\theta_{i+1})
&=
\frac{1}{\rho_{i-1}}(\sin\kappa_{i-1} + \sin\theta_{i}),
\label{2nd_One}
\\
\frac{1}{\rho_i}(\sin\kappa_i + \sin\theta_{i+1})
&=
\frac{1}{\rho_{ib}} \sin\kappa_{ib}.
\label{3rd_One}
\end{align}
A straightforward calculation reveals that if $\kappa_{ib} = \pm \pi/2$ with $\kappa_i = \pi + \theta_{i+1}$, then \eqref{3rd_One} results in a contradiction since its left hand side vanishes to zero, contrary to a non-zero ($\pm 1/\rho_{ib}$) right hand side. Thus, at a relative equilibrium, we have
\begin{equation}
\theta_{i+1} = \pi - \kappa_i,
\quad
i = 1,2,\ldots,n.
\label{theta_soln}
\end{equation}
By introducing a new variable $\gamma_i$ defined as
\begin{equation}
\gamma_i \triangleq \frac{1}{\rho_i}\big(\sin\kappa_i + \sin\theta_{i+1}\big) = \frac{2}{\rho_i} \sin\kappa_i,
\end{equation}
we obtain the following condition
\begin{equation}
\gamma_i = \gamma_{i-1}
\qquad
i = 1,2,\ldots,n
\label{eq_gamma_constant}
\end{equation}
from \eqref{2nd_One}. This condition, along with \eqref{3rd_One}, gives rise to
\begin{equation}
\frac{\sin\kappa_{ib}}{\rho_{ib}} = \frac{\sin\kappa_{i-1,b}}{\rho_{i-1,b}}
\qquad
i = 1,2,\ldots,n,
\end{equation}
which in turn yields the equilibrium values of $\kappa_{ib}$ as
\begin{equation}
\kappa_{ib}
=
\left\{
\begin{array}{ll}
\pi/2 & \quad \forall i \in \{1,2,\ldots,n\}, \qquad \textrm{or} \\
-\pi/2 & \quad \forall i \in \{1,2,\ldots,n\}.
\end{array}
\right.
\label{kappa_ib_soln}
\end{equation}
Now setting the dynamics of $\kappa_i$ to zero, we obtain
\begin{align}
&\mu \Big[(1 - \lambda)\sin(\kappa_i - \alpha_i) + \lambda \sin(\kappa_{ib} - \alpha_0) \Big]
\nonumber \\
& \qquad
= \frac{\lambda}{\rho_i}(\sin\kappa_i + \sin\theta_{i+1})
= \gamma_i
\label{1st_One}
\end{align}
for $i = 1,2,\ldots,n$. As $\kappa_{ib} = \pm\pi/2$ and $\theta_{i+1} = \pi - \kappa_i$ at a relative equilibrium, \eqref{1st_One} yields an equilibrium value for $\rho_i$ given by
\begin{equation}
\rho_i
=
\frac
{2 \sin\kappa_{ib} \sin\kappa_i}
{\mu \left( \frac{1}{\lambda} - 1 \right) \sin\kappa_{ib} \sin(\kappa_i - \alpha_i) + \mu \cos \alpha_0 }.
\label{eq_rho_i}
\end{equation}
Similarly, by \eqref{3rd_One} an equilibrium value of $\rho_{ib}$ can be expressed as
\begin{equation}
\rho_{ib}
=
\frac{1}{\mu \left( \frac{1}{\lambda} - 1 \right) \sin\kappa_{ib} \sin(\kappa_i - \alpha_i) + \mu \cos \alpha_0}
=
\frac{\sin\kappa_{ib}}{\gamma_i}.
\label{eq_rho_ib_equidistant}
\end{equation}
As we have shown earlier in \eqref{eq_gamma_constant} that $\gamma_i = \gamma_{i-1}$ for every $i \in \{1,2,\ldots,n\}$, it follows from \eqref{eq_rho_ib_equidistant} that the agents will be equi-distant from the beacon at any relative equilibrium. Hence, \textbf{all relative equilibria for the system must be circling equilibria}. Moreover, as both $\rho_{i}$ and $\rho_{ib}$ are required to be positive, we have \textit{necessary} conditions for existence of circling equilibrium, given by
\begin{align}
&
\lambda \cos \alpha_0 + (1 - \lambda) \sin \kappa_{ib} \sin(\kappa_i - \alpha_i) > 0
\label{positive_1}
\\
\textrm{and,} \qquad &
\sin \kappa_{ib} \sin\kappa_i > 0.
\label{positive_2}
\end{align}
\subsection{Evaluating solutions for $\kappa_i$}
It is easy to check that at any (circling) equilibrium of the closed loop shape dynamics, \eqref{eq_gamma_constant} with \eqref{1st_One} implies that
\begin{equation}
\sin(\kappa_{i+1} - \alpha_{i+1})
=
\sin(\kappa_i - \alpha_i),
\label{angle_condition_n}
\end{equation}
for which we have two possible solutions given by
\begin{subnumcases}{\kappa_{i+1} - \alpha_{i+1} = }
\kappa_i - \alpha_i
\label{Solution_normal}
\\
\pi - (\kappa_i - \alpha_i)
\label{Solution_abnormal}
\end{subnumcases}
where $i \in \{1,2,\ldots,n\}$. Equilibrium values for $\kappa_i$ can therefore be derived from \eqref{Solution_normal}-\eqref{Solution_abnormal}, along with shape variable constraints \eqref{FINAL_Constraint_1}-\eqref{FINAL_Constraint_2} and positivity conditions \eqref{positive_1}-\eqref{positive_2}.
If we let $\alpha^*$ be the angle satisfying $\kappa_1 - \alpha_1 = \alpha^*$ \editKG{at equilibrium}, then by \eqref{Solution_normal}-\eqref{Solution_abnormal} $\kappa_2 - \alpha_2$ must assume either $\alpha^*$ or $\pi - \alpha^*$. This aspect of binary possibilities holds true for every agent and therefore results in the branching depicted in Fig~\ref{fig:tree_1}. This figure provides a graphical illustration of all possible solutions for \eqref{angle_condition_n}, with each branch representing a candidate solution for $\kappa_i$. By considering a particular branch of the tree with $M$ (where $M \in \{1,2,\cdots,n\}$) copies of $\alpha^*$ and $(n-M)$ copies of $\pi - \alpha^*$, we have
\begin{equation}
\sum\limits_{i=1}^n \kappa_i
=
(n-M)\pi + (2M-n)\alpha^* + \sum\limits_{i=1}^n \alpha_i.
\label{sum_kappa_alpha_diff}
\end{equation}
\input{4_Binary_Tree_KAPPA}
Now we focus on the shape variable constraints \eqref{FINAL_Constraint_1}-\eqref{FINAL_Constraint_2} to obtain solutions for $\alpha^*$. It is easy to check that the consistency conditions \eqref{FINAL_Constraint_2} hold true at any relative equilibrium of the closed loop dynamics. Additionally, by exploiting the relationship between equilibrium values of $\kappa_i$ and $\theta_i$, the cycle closure constraint \eqref{FINAL_Constraint_1} can be expressed as
\begin{equation}
\sum_{i=1}^{n} \kappa_i
=
m\pi, \qquad m \in \mathds{Z},
\label{Closure_Bye_Product_1}
\end{equation}
where $\mathds{Z}$ is the set of integers. By substituting \eqref{Closure_Bye_Product_1} into \eqref{sum_kappa_alpha_diff}, we have
\begin{equation}
(2M-n)\alpha^*
=
(m+M-n)\pi - \sum_{i=1}^n\alpha_i .
\label{cond_on_alpha_star}
\end{equation}
We summarize the preceding discussion in the following proposition.
\begin{proposition}
\label{prop:existenceProp}
Consider an $n$-agent cyclic CB pursuit system with beacon, whose shape dynamics is governed by \eqref{CL_dynamics_n_simplified} \editKG{parametrized by $\mu, \lambda$, and the CB parameters $\left\{\alpha_0, \alpha_1, \alpha_2, \ldots, \alpha_n \right\}$}. The following statements are true.
\begin{itemize}
\item[(a)] The only possible relative equilibria are circling equilibria.
\item[(b)] Whenever $\sin(\sum \alpha_i) \neq 0$, a circling equilibrium exists \textit{if and only if} there exists $m \in \mathds{Z}$ and $\sigma = (\sigma_1,\sigma_2,\ldots,\sigma_n) \in \{-1,1\}^n$ such that
\begin{itemize}
\item[(i)] \editKG{the cardinality $M$ of the subset $\{\sigma_i | \sigma_i = 1, i = 1,2,\ldots,n\}$ satisfies}
\begin{equation}
2M-n \neq 0,
\label{Prop_4_cond_on_M}
\end{equation}
and
\item[(ii)]
\begin{equation}
\begin{aligned}
& \lambda \cos \alpha_0 + (1 - \lambda) \sin \alpha^* > 0,
\\
& \sin \big( \alpha^* + \sigma_i\alpha_i \big) > 0, \qquad i = 1,2,\ldots,n
\end{aligned}
\label{Proposition_4}
\end{equation}
where $\alpha^*$ is given by
\begin{equation}
\alpha^* = \left( \frac{m+M-n}{2M-n} \right)\pi - \sum_{i=1}^{n} \left( \frac{\alpha_i}{2M-n} \right).
\label{alpha_star_soln}
\end{equation}
\end{itemize}
At equilibrium, we have \editKG{either $\kappa_{ib} = \pi/2, \; i=1,2,\ldots, n$ or $\kappa_{ib} = -\pi/2, \; i=1,2,\ldots, n$} and the equilibrium values of $\kappa_i$, $\rho_{ib}$, and $\rho_i$ given by
\begin{align*}
\kappa_i
&=
\editKG{\frac{\pi (1-\sigma_i)}{2} + (\sigma_i \alpha^* + \alpha_i) = \;}
\biggl\{
\begin{array}{ll}
\alpha^* + \alpha_i, &\qquad \text{if } \sigma_i = +1 \\
\pi - \alpha^* + \alpha_i, &\qquad \text{if } \sigma_i = -1
\end{array}\\
\rho_{ib}
&=
\frac{1}{\mu \lambda \cos \alpha_0 + \mu \left( 1 - \lambda \right) \sin\kappa_{ib} \sin\alpha^*}\\
\rho_{i}
&=
2\rho_{ib} \sin \kappa_{ib} \sin \kappa_i .
\end{align*}
\end{itemize}
\end{proposition}
\vspace{.3cm}
\begin{proof}
The first statement of the proposition directly follows from \eqref{eq_rho_ib_equidistant}.
\editKG{The preceding discussion has demonstrated that if a circling equilibrium exists, then $\kappa_{ib}$ and $\kappa_i$ must satisfy \eqref{kappa_ib_soln},\eqref{positive_1}-\eqref{positive_2}, \eqref{Solution_normal}-\eqref{Solution_abnormal}, and \eqref{Closure_Bye_Product_1}, and the equilibrium values for $\theta_i$, $\rho_i$\, and $\rho_{ib}$ can be expressed in terms of $\kappa_{ib}$ and $\kappa_i$ by \eqref{theta_soln}, \eqref{eq_rho_i}, and \eqref{eq_rho_ib_equidistant}. Further analysis of \eqref{positive_1}-\eqref{positive_2} demonstrated that if we let $\alpha^*$ denotes the angle difference $(\kappa_1 - \alpha_1)$, then for each $i$, we have one of two possibilities - either $\kappa_i = \alpha^* + \alpha_i$ or $\kappa_i = \pi - \alpha^* + \alpha_i$, which we represent by the binary tree in Fig~\ref{fig:tree_1}.}
We now consider a particular branch of the binary tree (Fig~\ref{fig:tree_1}), for which $\kappa_i = \alpha^* + \alpha_i$ for exactly $M$ agents ($1 \leq M \leq n$) and $\kappa_i = \pi - \alpha^* + \alpha_i$ for the remaining $n-M$ agents. Clearly, for the first set of $M$ agents we have $\sin \kappa_i = \sin (\alpha^* + \alpha_i)$, while the remaining agents will have $\sin \kappa_i = \sin (\alpha^* - \alpha_i)$, and therefore there exists some $\sigma = (\sigma_1,\sigma_2,\ldots,\sigma_n) \in \{-1,1\}^n$ such that $\sin\kappa_i = \sin (\alpha^* + \sigma_i \alpha_i)$ for every $i=1,2,\ldots,n$, and the cardinality of the set $\{\sigma_i | \sigma_i = 1, i = 1,2,\ldots,n\}$ is $M$. \editKG{This implies that \eqref{cond_on_alpha_star} will hold (as demonstrated in the previous discussion), and thus} for $\kappa_{ib} = \pi/2$, it is clear that the positivity conditions \eqref{positive_1}-\eqref{positive_2} can be expressed as \eqref{Proposition_4}, with \eqref{alpha_star_soln} following from \eqref{cond_on_alpha_star} as long as $2M - n \neq 0$.
It remains to be shown that \eqref{Proposition_4} also encompasses the case $\kappa_{ib} = -\pi/2$, so that \eqref{Proposition_4} is equivalent to \eqref{positive_1}-\eqref{positive_2}. For $\kappa_{ib} = -\pi/2$, the positivity conditions \eqref{positive_1}-\eqref{positive_2} simplify to
\begin{equation}
\begin{aligned}
&
\lambda \cos \alpha_0 - (1 - \lambda) \sin \alpha^* > 0
\\
&
\sin(\alpha^* + \sigma_i \alpha_i) < 0,
\end{aligned}
\label{CoNd__2}
\end{equation}
and we must show that there exists $\hat{m} \in \mathds{Z}$ such that \eqref{Proposition_4} with $\hat{m}$ substituted into \eqref{alpha_star_soln} is equivalent to \eqref{CoNd__2}. Choosing $\hat{m} = m+2M-n$ yields the desired result, and therefore statement (b) of the proposition is established. The characterization of the associated equilibrium values follows from the preceding discussion\editKG{, and also establishes the sufficiency of our existence conditions, completing the proof}.
\end{proof}
\begin{remark}
The possibility of having $2M-n= 0$ cannot be ruled out for an even number of agents, in which case \eqref{cond_on_alpha_star} can only be satisfied if $\sum\alpha_i$ is an integer multiple of $\pi$. This case corresponds to existence of a continuum of circling equilibria.
\end{remark}
\begin{remark}
Letting $\psi_{i,i+1}$ denote the angular separation between agent $i$ and $i+1$ at a circling equilibrium, we have
\begin{equation}
\cos\psi_{i,i+1}
=
\frac{\rho_{ib}^2 + \rho_{i+1,b}^2 - \rho_i^2}{2 \rho_{ib} \rho_{i+1,b}}
=\cos(2\kappa_i).
\end{equation}
Therefore the equilibrium value of angular separation between agent $i$ and agent $i+1$ is $2\kappa_i$.
\end{remark}
\begin{remark}
\label{remark:leftbranch}
If we consider the special case for which \eqref{Solution_normal} holds true for each pair of agents (i.e. the leftmost branch in Fig~\ref{fig:tree_1}), then we have $\sigma = (1,1,\dots,1)$, i.e. $M=n$. In this case \eqref{cond_on_alpha_star} simplifies to the form
\begin{equation}
\alpha^* = m \left( \frac{\pi}{n} \right) - \sum_{i=1}^{n} \left( \frac{\alpha_i}{n} \right).
\label{alpha_soln_special_case}
\end{equation}
\end{remark}
\section{Stability analysis for the two-agent system}
\label{sec:2_Agent}
Here we consider the case where $n=2$ and analyze the stability of the associated circling equilibria. For the two-agent system, our constraints \eqref{FINAL_Constraint_1}-\eqref{FINAL_Constraint_2} imply that $\rho_1 = \rho_2 \triangleq \rho$ and $\theta_i = \kappa_i$ for $i=1,2$. Thus the dynamics \eqref{CL_dynamics_n_simplified} simplify for the two-agent case to
\begin{align}
\label{CL_dynamics_2_simplified}
\dot{\rho}
&=
- (\cos\kappa_1 + \cos\kappa_2)
\nonumber \\
\dot{\kappa}_i
&=
- \mu \Big[(1 - \lambda)\sin(\kappa_i - \alpha_i) + \lambda \sin(\kappa_{ib} - \alpha_0) \Big]
+ \frac{\lambda}{\rho}(\sin\kappa_1 + \sin\kappa_2) \\
\dot{\rho}_{ib}
&=
- \cos\kappa_{ib}
\nonumber \\
\dot{\kappa}_{ib}
&=
\dot{\kappa}_i
- \frac{1}{\rho}(\sin\kappa_1 + \sin\kappa_2)
+ \frac{1}{\rho_{ib}} \sin\kappa_{ib}, \; i=1,2,
\nonumber
\end{align}
and are subject to the \editKG{constraint \eqref{FINAL_Constraint_1} which simplifies to}
\begin{align}
\rho \mathds{I}_2 - \rho_{1b} R(\kappa_{1b} - \kappa_1) - \rho_{2b} R(\kappa_{2b} - \kappa_{2}) &= 0.
\label{2_agent_CoNsTrAiNt_simplified}
\end{align}
\subsection{Existence of circling equilibria for the two-agent case}
For the two-agent system, we have only two possible branches in Fig~\ref{fig:tree_1}. The right-hand branch corresponds to $M=1$, for which we have $2M-n = 0$, and \textit{Proposition \ref{prop:existenceProp}} does not apply. However, if $\alpha_1 + \alpha_2 = \editKG{k}\pi$ for some $\editKG{k} \in \mathds{Z}$, then a continuum of equilibria exist with $\kappa_1$ and $\kappa_2$ satisfying $\kappa_1 + \kappa_2 = \pi + \alpha_1 + \alpha_2$.
For the left-hand branch in Fig~\ref{fig:tree_1}, we have $M=2$ (i.e. $2M-n \neq 0$) and therefore we can apply \textit{Proposition \ref{prop:existenceProp}} (as long as $\sin(\sum \alpha_i) \neq 0$) for which \eqref{alpha_star_soln} simplifies to
\begin{equation}
\alpha^* = m \left( \frac{\pi}{2} \right) - \sum_{i=1}^{2} \left( \frac{\alpha_i}{2} \right).
\label{alpha_soln_special_case_n2}
\end{equation}
\editKG{Thus we have (for $i=1,2$)
\begin{equation}
\sin(\alpha^* + \sigma_i \alpha_i)
=
\sin\left(m\frac{\pi}{2} + \alpha_i - \frac{\alpha_1+\alpha_2}{2}\right),
\end{equation}
and the second constraint} in \eqref{Proposition_4} requires $m= \pm 1$. We label these options as Type 1 and Type 2 two-agent circling equilibrium and summarize the resulting characterization in Table~\ref{Table_1_FULL}, where we have made use of the notation
\begin{align}
\alpha^+ = (\alpha_1+\alpha_2)/2, \quad \alpha^- = (\alpha_1-\alpha_2)/2.
\end{align}
\subsection{Stability analysis}
We analyze the stability of two-agent circling equilibria by linearizing the dynamics \eqref{CL_dynamics_2_simplified} about the equilibria described in the previous section. Following the line of thought from the stability analysis in \cite{Marshall_TAC_04}, we first demonstrate that the linearized dynamics will always have exactly one pair of pure imaginary eigenvalues resulting from the constraint equation \eqref{2_agent_CoNsTrAiNt_simplified}.
First, denoting $\xi = \left\{\kappa_1, \kappa_2,\rho, \kappa_{1b}, \kappa_{2b},\rho_{1b}, \rho_{2b}\right\}$ and the corresponding dynamics \eqref{CL_dynamics_2_simplified} by $f(\xi)$, we express \eqref{2_agent_CoNsTrAiNt_simplified} in terms of scalar constraints by
\begin{align}
g_1(\xi) &\triangleq \rho - \rho_{1b}\cos(\kappa_{1b}-\kappa_1) - \rho_{2b}\cos(\kappa_{2b}-\kappa_2) = 0, \nonumber \\
g_2(\xi) &\triangleq \rho_{1b}\sin(\kappa_{1b}-\kappa_1) + \rho_{2b}\sin(\kappa_{2b}-\kappa_2) = 0. \nonumber
\end{align}
We define the manifold on which these constraints are satisfied by
\begin{align}
M = \left\{\xi \in \mathbb{R}^{7n}: g_1(\xi) = g_2(\xi) = 0\right\},
\end{align}
which can be shown to be invariant under the dynamics \eqref{CL_dynamics_2_simplified}. If we let $\bar{\xi}$ denote a representative circling equilibrium for the two-agent case and let $\dot{\tilde{\xi}} = A \tilde{\xi}$ denote the linearization of the dynamics \eqref{2_agent_CoNsTrAiNt_simplified} about $\bar{\xi}$, then it was demonstrated in \cite{Marshall_TAC_04} that invariance of $M$ implies existence of a change of basis which will transform $A$ into upper-triangular form with a $2\times 2$ lower-right hand block. A suitable explicit form for the change of basis is given by $\phi = \Phi(\xi)$, where
\begin{align}
\phi_1 = \kappa_1, \phi_2=\kappa_2, \phi_3=\rho, \phi_4=\kappa_{1b}, \nonumber \\
\phi_5=\rho_{1b},\phi_6=g_1(\xi),\phi_7=g_2(\xi).
\end{align}
We note that the corresponding equilibrium $\bar{\phi} = \Phi(\bar{\xi})$ will have $0$'s for the last two components.
By a straightforward calculation, we have
\begin{align}
\label{eqn:dDotEqns}
\dot{g}_{1}(\xi) &= \frac{\partial g_1(\xi)}{\partial \xi}f(\xi) = -\frac{\sin \kappa_1 + \sin \kappa_2}{\rho}g_2(\xi), \nonumber \\
\dot{g}_{2}(\xi) &= \frac{\partial g_2(\xi)}{\partial \xi}f(\xi) = \frac{\sin \kappa_1 + \sin \kappa_2}{\rho}g_1(\xi),
\end{align}
from whence it follows that, under the change of basis, the dynamics linearized about $\bar{\phi}$ take the upper-triangular form
\begin{align}
\dot{\phi} = \left[\begin{array}{cc}
A_{11} & * \\
0_{2\times 5} & A_{22}
\end{array}\right]\phi,
\end{align}
with
\begin{align}
\label{eqn:A22matrix}
A_{22} &= \left[\begin{array}{cc}
0 & -\frac{\sin \kappa_1 + \sin \kappa_2}{\rho} \\
\frac{\sin \kappa_1 + \sin \kappa_2}{\rho} & 0 \end{array}\right]_{\xi = \Phi^{-1}(\phi)} \nonumber \\
&= \left[\begin{array}{cc}
0 & \mp \delta \\
\pm \delta & 0 \end{array}\right],
\end{align}
where $\delta \triangleq \mu\left(\cos\alpha_0 + \left(\frac{1-\lambda}{\lambda}\right)\cos\alpha^{+}\right)$. It is clear from \eqref{eqn:A22matrix} that $A_{22}$ has a pair of pure imaginary eigenvalues at $\lambda = \pm j\mu\delta$ (resulting from the constraint equation \eqref{2_agent_CoNsTrAiNt_simplified}). By analogy with the argument presented in \cite{Marshall_TAC_04}, we focus our stability characterization on the remaining five eigenvalues.
Returning to the original coordinates $\xi$, we proceed with our stability analysis by linearizing the dynamics about the Type 1 CCW circling equilibrium from Table \ref{Table_1_FULL}. One can show that the corresponding characteristic polynomial $P(x)$ is given by
\begin{align}
P(x) &= (x^2 + \delta^2)(x^2 +\lambda\editKG{\Psi} x + \lambda \delta^2)
\bigl(x^3 + \lambda\Psi x^2 + \delta^2 x + (1-\lambda)\mu \sin\alpha^{+} \delta^2\bigr),
\end{align}
where
\begin{align}
\Psi \triangleq \mu\left[\sin(\alpha_0) + \left(\frac{1-\lambda}{\lambda}\right)\sin(\alpha^{+})\right].
\end{align}
As expected based on the previous discussion, $P(x)$ has two pure imaginary roots at $x = \pm j\delta$. It is also clear that the roots of the quadratic term will have strictly negative real part if and only if $\Psi>0$. By the Routh-Hurwitz criterion, the roots of the cubic factor will be in the open left-half plane if and only if $\Psi>0$, $\sin\alpha^{+}>0$, and $\lambda\Psi\delta^2 - (1-\lambda)\mu \sin\alpha^{+} \delta^2 > 0$. Since this last condition simplifies to $\lambda\Psi - (1-\lambda)\mu \sin\alpha^{+} > 0$, which always holds if $\sin(\alpha_0) > 0$, our requirement for stability is given by $\sin\alpha^{+}>0$ and $\sin(\alpha_0) > 0$.
A similar analysis of the other possible circling equilibria (i.e. Type 1 CW and Type 2 CCW and CW) results in analogous stability conditions, which can be summarized in the following proposition.
\begin{proposition}
\label{prop:TwoAgentStability}
The Jacobian associated with the two-agent circling equilibria has two pure imaginary eigenvalues resulting from the constraint equation \eqref{2_agent_CoNsTrAiNt_simplified}. The remaining eigenvalues all have real parts less than zero if and only if
\begin{itemize}
\item $\sin(\alpha_0) > 0$ and $\sin \alpha^{+}>0$ in the Type 1 CCW case;
\item $\sin(\alpha_0) < 0$ and $\sin \alpha^{+}<0$ in the Type 1 CW case;
\item $\sin(\alpha_0) > 0$ and $\sin \alpha^{+}<0$ in the Type 2 CCW case;
\item $\sin(\alpha_0) < 0$ and $\sin \alpha^{+}>0$ in the Type 2 CW case.
\end{itemize}
\end{proposition}
\vspace {.25cm}
\begin{proof}
Follows from the discussion above.
\end{proof}
Note that these results are also summarized in Table~\ref{Table_1_FULL}.
\begin{table*}[t!]
\centering
\begin{adjustbox}{max width=\columnwidth}
\vspace{-0.1in}
\begin{tabular}{lllll}
\hline
&
\multicolumn{2}{c}{Type 1}
&
\multicolumn{2}{c}{Type 2}
\\
\hline
$\begin{array}{l}
\textrm{\textcolor{my_GREEN}{Existence Condition}}
\end{array}$
&
\multicolumn{2}{c}{
$\begin{array}{l}
\cos\alpha^- > 0
\\
\lambda \cos \alpha_0 + (1 - \lambda) \cos\alpha^+ > 0
\end{array}$
}
&
\multicolumn{2}{c}{
$\begin{array}{l}
\cos\alpha^- < 0
\\
\lambda \cos \alpha_0 - (1 - \lambda) \cos\alpha^+ > 0
\end{array}$
}
\\
\hline
&
$CCW$
&
$CW$
&
$CW$
&
$CCW$
\\
\cmidrule(l){2-5}
$\begin{array}{l}
\textrm{\textcolor{blue}{Chracterization}}
\\
\textrm{\textcolor{blue}{of}}
\\
\textrm{\textcolor{blue}{Equilibria}}
\end{array}$
&
$\begin{array}{l}
\kappa_{ib} = \frac{\pi}{2}
\\
\kappa_1 = \frac{\pi}{2} + \alpha^-
\\
\kappa_2 = \frac{\pi}{2} - \alpha^-
\\
\psi_{12} = \pi + 2\alpha^-
\end{array}$
&
$\begin{array}{l}
\kappa_{ib} = -\frac{\pi}{2}
\\
\kappa_1 = -\frac{\pi}{2} + \alpha^-
\\
\kappa_2 = -\frac{\pi}{2} - \alpha^-
\\
\psi_{12} = - \pi + 2\alpha^-
\end{array}$
&
$\begin{array}{l}
\kappa_{ib} = -\frac{\pi}{2}
\\
\kappa_1 = \frac{\pi}{2} + \alpha^-
\\
\kappa_2 = \frac{\pi}{2} - \alpha^-
\\
\psi_{12} = \pi + 2\alpha^-
\end{array}$
&
$\begin{array}{l}
\kappa_{ib} = \frac{\pi}{2}
\\
\kappa_1 = -\frac{\pi}{2} + \alpha^-
\\
\kappa_2 = -\frac{\pi}{2} - \alpha^-
\\
\psi_{12} = -\pi + 2\alpha^-
\end{array}$
\\
\cmidrule(l){2-5}
&
\multicolumn{2}{l}{
$\begin{array}{l}
\displaystyle \rho_i = \frac{2\lambda \cos\alpha^-}{\mu\lambda \cos \alpha_0 + \mu(1 - \lambda) \cos\alpha^+}
\\
\displaystyle \rho_{ib} = \frac{\lambda}{\mu \lambda \cos \alpha_0 + \mu (1 - \lambda) \cos\alpha^+}
\end{array}$
}
&
\multicolumn{2}{l}{
$\begin{array}{l}
\displaystyle \rho_i = \frac{2\lambda \cos\alpha^-}{\mu(1 - \lambda) \cos\alpha^+ - \mu \lambda \cos \alpha_0}
\\
\displaystyle \rho_{ib} = \frac{\lambda}{\mu (1 - \lambda) \cos\alpha^+ - \mu \lambda \cos \alpha_0 }
\end{array}$
}
\\
\hline
$\begin{array}{l}
\textrm{\textcolor{red}{Stability Condition}}
\end{array}$
&
$\begin{array}{l}
\sin \alpha_0 > 0
\\
\sin \alpha^+ > 0
\end{array}$
&
$\begin{array}{l}
\sin \alpha_0 < 0
\\
\sin \alpha^+ < 0
\end{array}$
&
$\begin{array}{l}
\sin \alpha_0 > 0
\\
\sin \alpha^+ < 0
\end{array}$
&
$\begin{array}{l}
\sin \alpha_0 < 0
\\
\sin \alpha^+ > 0
\end{array}$
\\
\hline
\end{tabular}
\end{adjustbox}
\caption{Characterization of circling equilibria for a two-agent system (whenever $\sin(2\alpha^+) \neq 0$), with $\alpha^+ \triangleq (\alpha_1 + \alpha_2)/2$ and $\alpha^- \triangleq (\alpha_1 - \alpha_2)/2$.}
\label{Table_1_FULL}
\end{table*}
\section{Implementation Results}
\label{sec:Implementation}
\subsection{Experimental setup}
We use Pioneer 3 DX (from Adept MobileRobots), a compact differential-drive mobile robot with reversible DC motors, high-resolution motion encoders, as the experimental platform. Onboard computation is done via 32-bit Renesas SH2-7144 RISC microprocessor, including the P3-SH microcontroller with ARCOS. \editKG{ARIA, a software library from the developer, provides an interface for controlling and receiving data from the robot, and communication with the robot for sending control commands (\textit{forward velocity} and \textit{turning rate}) is carried out via 802.11-b/g/n networking.}
\begin{figure}[h]
\begin{center}
\includegraphics[width=0.35\textwidth]{ISL_Robot}
\caption{Mobile robot based experimental platform (Pioneer 3 DX) with two-wheel differential and caster.}
\label{Robot_P3_DX}
\end{center}
\end{figure}
Algorithm implementation (i.e, feedback law computation) has been \editKG{carried out} in C++ using ROS (Robot Operating System), along with ROS-ARIA, as the interfacing robotics middleware. The experiments have been carried out in a laboratory environment equipped with a sub-millimeter accurate Vicon motion capture system (\textit{www.vicon.com}). The Dell workstation, which we use to evaluate control commands at $25Hz$, is connected to the Vicon server via a dedicated Ethernet connection.
\subsection{Two robots with asymmetric distribution on the circle}
The first experiment presented here involves two robots circling around the beacon in a counter-clockwise direction. The parameters $\alpha_1$ and $\alpha_2$ were selected as $5\pi/12$ and $-\pi/12$, respectively. These choices, by yielding the equilibrium value of angular separation as $\pi/2$, demonstrate that the proposed approach can give rise to asymmetric distribution of agents on the circle (at equilibrium). Moreover, the parameter values $\alpha_0 = \pi/3$ and $\lambda = 1/2$ yield the circling radius as
\begin{displaymath}
\rho_{ib} = \frac{1}{\mu\big( \cos(\pi/3) + \cos(\pi/6) \big)}, \quad i=1,2,
\end{displaymath}
and by choosing $\mu = 0.75 m^{-1}$ we get $\rho_{ib} = 0.9761 m$. The corresponding robot trajectories are shown in Fig~\ref{fig:2_1}, and the evolution of distance and angular separation for the agents are shown in Fig~\ref{fig:2_2} and Fig~\ref{fig:2_3}, respectively (refer \cite{ISL_Videos} for implementation videos). These figures show a quick convergence to the circling equilibrium (within $150 sec.$).
\begin{figure}[!ht]
\centering
\begin{subfigure}[b]{0.5\textwidth}
\centering
\includegraphics[width=0.9\textwidth]{TWO_Robot_Ft_Prnt}
\caption{Robot trajectories}
\label{fig:2_1}
\end{subfigure}
\begin{subfigure}[b]{0.5\textwidth}
\centering
\includegraphics[width=0.9\textwidth]{TWO_Robot_BEACON_Distance}
\caption{Distances from the beacon}
\label{fig:2_2}
\end{subfigure}
\begin{subfigure}[b]{0.5\textwidth}
\centering
\includegraphics[width=0.9\textwidth]{TWO_Robot_Angular_Sep}
\caption{Inter-agent angular separations}
\label{fig:2_3}
\end{subfigure}
\caption{Robot trajectories during implementation of \eqref{CL_dynamics_n_simplified}, along with evolution of relevant quantities (Note that, $n=2$, $\alpha_0 = \pi/3$, $\alpha_1 = 5\pi/12$, $\alpha_2 = -\pi/12$, $\lambda = 1/2$, $\mu = 0.75 m^{-1}$).}\label{fig:2_Agent}
\end{figure}
\begin{figure}[!hb]
\centering
\begin{subfigure}[b]{0.5\textwidth}
\centering
\includegraphics[width=0.9\textwidth]{FIVE_Robot_Ft_Prnt}
\caption{Robot trajectories}
\label{fig:5_1}
\end{subfigure}
\begin{subfigure}[b]{0.5\textwidth}
\centering
\includegraphics[width=0.9\textwidth]{FIVE_Robot_BEACON_Distance}
\caption{Distances from the beacon}
\label{fig:5_2}
\end{subfigure}
\begin{subfigure}[b]{0.5\textwidth}
\centering
\includegraphics[width=0.9\textwidth]{FIVE_Robot_Angular_Sep}
\caption{Inter-agent angular separations}
\label{fig:5_3}
\end{subfigure}
\caption{Robot trajectories during implementation of \eqref{CL_dynamics_n_simplified}, along with evolution of relevant quantities (Note that, $n=5$, $\alpha_0 = -\pi/6$, $\alpha_i = -\pi/4$, $\lambda = 1/2$, $\mu = 1.50 m^{-1}$). \editKG{A perturbation was applied to the system at $723 sec.$ into the experiment, and at $868 sec.$ the beacon was relocated to a new position.}}\label{fig:5_Agent}
\end{figure}
\subsection{Five robots with symmetric distribution on the circle}
Next we \editKG{choose control parameters which result in} five robots circling around the beacon in a clockwise direction\editKG{, with the robots distributed symmetrically around the circle. More specifically, we choose the same value for every $\alpha_i$ ($=-\pi/4$), and let $\alpha_0 = -\pi/6$ and $\lambda = 1/2$. Then} we have
\begin{displaymath}
\rho_{ib} = \frac{1}{\mu\big( \cos(-\pi/6) - \sin(-\pi/5 + \pi/4) \big)}, \quad i=1,\ldots,5,
\end{displaymath}
and a choice of $\mu = 1.50 m^{-1}$ yields \editKG{an equilibrium circle radius of} $\rho_{ib} = 0.9395 m$. The corresponding robot trajectories, along with the evolution of distance and angular separation of the agents, are shown in Fig~\ref{fig:5_Agent} (refer \cite{ISL_Videos} for implementation videos). In this experiment we introduced a perturbation to the system at $723 sec.$ into the experiment, and later (at $868 sec.$) the beacon has been moved to a new position. In both cases, the formation quickly converges back to the desired circling equilibrium.
\subsection{Discussion}
The results show some level of imperfection during implementation of the proposed feedback law. This deviation from theoretical predictions can be attributed to multiple factors. To start with, our theoretical analysis assumes the agents to be point particles where in reality they occupy significant space (width - $380 mm$, swing radius - $260 mm$). Also the placement of markers (necessary for measurement using Vicon motion capture system) introduces some error due to misalignment between center of the robot axle and origin of the body fixed frame. Furthermore, as only planar components of positions and heading were considered in measuring relevant quantities, a small slope (which has later been verified) in the lab floor acts as another source of error. However, in spite of these multiple sources of error, the proposed feedback mechanism \eqref{u_i_shape} is able to restrict the error margins well below the physical dimension of the agents.
\section{Conclusion and Future Work}
\label{sec:Concl}
We have introduced a modified version of the CB pursuit law which references a fixed beacon as well as a neighboring agent, and demonstrated that implementation in a cycle graph (with ``spokes'') yields an interesting set of closed-loop dynamics. Analysis of those dynamics reveals the existence of circling equilibria centered on the beacon, and it is of particular interest that a specific equilibrium radius emerges as a function of the control parameters. Future work will focus on extending our analysis to the 3-d setting, as well as consideration of scenarios with multiple beacons or slowly moving beacons.
\section*{Acknowledgments}
The authors would like to take this opportunity to thank P. S. Krishnaprasad and E. W. Justh for their valuable feedback and comments. They also appreciate the assistance from U. Halder in implementing the control law on mobile robots.
\bibliographystyle{IEEEtran}
|
\section{Introduction}
This article considers the question of estimating the angular velocity of a rigid body from signals from embedded sensors. This general question is of particular importance in various fields of engineering, and in particular for the problem of orientation control, as shown in numerous applications~\cite{salcudean1991,boskovic2000,silani2003,lovera2005} for spacecraft, unmanned aerial vehicles, guided ammunitions, to name a few.
In the literature, two types of methods have been proposed to address this question. First, one can directly measure the angular velocity by using a specific sensor. This straightforward solution requires a strap-down rate gyro~\cite{Titterton2004}. However, rate gyros being relatively fragile and expensive components, prone to drift, this solution is often discarded. The alternative is a \emph{two-step} approach. In the first step, attitude is determined from measurements of known reference vectors. Then, in the second step, attitude variations are used to estimate the angular velocity.
The first step is detailed in~\cite{crassidis2007}. In a nutshell, when two independent vectors are measured with vector sensors attached to a rigid body, the attitude of the rigid body can be found under the form the solution of the Wahba problem~\cite{wahba1965} which is a minimization problem having as unknown the rotation matrix from a fixed frame to the body frame. Thus, at any instant, full attitude information can be obtained~\cite{shuster1978,shuster1990,baritzhack1996,choukroun2003}. In principles, this is sufficient to perform the second step: once the attitude is known, angular velocity can be estimated from a time-differentiation. However, noises disturb this process. To address this issue, introducing \emph{a priori} information in the estimation process allows one to filter-out noises from the estimates. Following this approach, numerous observers based on the Euler equations have been proposed to estimate angular velocity from full attitude information \cite{salcudean1991,thienel2007,sunde2005,jorgensen2011}.
Besides this two-step approach, which requires measurements of two independent reference vectors, a more direct and less requiring solution can be proposed. In this paper, we expose an algorithm that directly uses the measurements of a \emph{single} vector and reconstructs the angular velocity in a simple manner, by means of a nonlinear observer. This is the contribution of this article. In a related philosophy, we have recently proposed an observer using the measurements from two linearly independent vectors as input~\cite{magnis2014a}. The present paper studies a similarly structured observer. However, due to the fact that here only a single vector measurement is employed, the arguments of proof are completely different, and result in a new and independent contribution.
The paper is organized as follows. In Section~\ref{sec : notations}, we introduce the notations and the problem statement. We analyze the attitude dynamics (rotation and Euler equations) and relate it to the measurements. In Section~\ref{sec : observer}, we define the proposed nonlinear observer. The observer has an extended state and uses output injection. To prove its convergence, the error equation is identified as a linear time-varying (LTV) system perturbed by a linear-quadratic term. Under a persistent excitation (PE) assumption, the LTV dynamics is shown to generate an exponentially convergent dynamics. This property, together with assumptions on the inertia parameters of the rigid body, reveal instrumental to conclude on the exponential uniform convergence of the error dynamics. Importantly, the PE assumption is proven to be automatically satisfied in the particular case of free-rotation. In details, in Section~\ref{sec:PE}, we establish that for almost all initial conditions, the PE assumption holds. This result stems from a detailed analysis of the various types of solutions to the free-rotation dynamics. Illustrative simulation results are given in Section~\ref{sec : simu}. Conclusions and perspectives are given in Section~\ref{sec : conclusion}.
\section{Notations and problem statement}
\label{sec : notations}
\subsection{Notations}
\textbf{Vectors in $\R{3}$} are written with small letters $x$. $|x|$ is the Euclidean norm of $x$. $\cro{x}$ is the skew-symmetric cross-product matrix associated with $x$, i.e. ${\forall y \in \R{3}, \ \cro{x}y = x \times y}$. Namely,
\begin{displaymath}
\cro{x} \triangleq \left(
\begin{array}{ccc}
0 & -x_3 & x_2 \\
x_3 & 0 & -x_1 \\
-x_2 & x_1 & 0
\end{array}
\right)
\end{displaymath}
where $x_1,x_2,x_3$ are the coordinates of $x$ in the standard basis of $\R{3}$. If $x$ is a unit vector, we have
\begin{displaymath}
\cro{x}^2 = x x^T - I
\end{displaymath}
\textbf{Vectors in $\R{6}$} are written with capital letters $X$. $|X|$ is the Euclidean norm of $X$. The induced norm on $6\times6$ matrices is noted $||\cdot||$. Namely,
\begin{displaymath}
||M|| = \max_{|X|=1} |MX|
\end{displaymath}
For convenience, we may write $X$ under the form
\begin{displaymath}
X = \left(X_1^T,X_2^T\right)^T
\end{displaymath}
with $X_1,X_2 \in \R{3}$. Note that $${|X|^2 = |X_1|^2+|X_2|^2}$$
\textbf{Frames} considered in the following are orthonormal bases of $\R{3}$.
\textbf{Rotation matrix.} For any unit vector ${u \in \R{3}}$ and any ${\zeta \in \mathbf{R}}$, $r_u(\zeta)$ designates the rotation matrix of axis $u$ and angle $\zeta$. Namely
\begin{displaymath}
r_u(\zeta) \triangleq \cos \zeta I + \sin \zeta \cro{u} + (1-\cos \zeta)u u^T
\end{displaymath}
\subsection{Problem statement}
Consider a rigid body rotating with respect to an inertial frame $\mathcal R_i$. Note $R$ the rotation matrix from $\mathcal R_i$ to a body frame $\mathcal R_b$ attached to the rigid body and $\o$ the corresponding angular velocity vector, expressed in $\mathcal R_b$.
Assuming that the body rotates under the influence of an external torque $\tau$ (which, is null in the case of free-rotation), the variables $R$ and $\o$ are governed by the following differential equations
\begin{align}
\dot R & = R \cro{\o} \label{eq R}\\
\dot \o & = J^{-1}\left(J\o \times \o + \tau \right) \triangleq E(\o) + J^{-1}\tau \label{eq euler}
\end{align}
where $J = \textrm{diag}(J_1,J_2,J_3)$ is the inertia matrix\footnote{Without restriction, we consider that the axes of $\mathcal R_b$ are aligned with the principal axes of inertia of the rigid body.}. Equation~\eqref{eq euler} is known as the set of Euler equations for a rotating rigid body \cite{landau1982}. The torque $\tau$ may result from control inputs or disturbances\footnote{In the case of a satellite e.g., the torque could be generated by inertia wheels, magnetorquers, gravity gradient, among other possibilities.}. We assume that $J$ and $\tau$ are known.
We assume that a constant reference unit vector $\mathring{\a}$ expressed in $\mathcal R_i$ is known, and that sensors arranged on the rigid body allow to measure the corresponding unit vector expressed in $\mathcal R_b$. Namely, the measurements are
\begin{equation}
\label{def : a,b}
\a(t) \triangleq R(t)^T \mathring{\a}
\end{equation}
For implementation, the sensors could be e.g. accelerometers, magnetometers, or Sun sensors to name a few \cite{magnis2014}. We now formulate some assumptions.
\begin{assump}
\label{hypothese O borné}
$\o$ is bounded : $|\o(t)| \leq \o_{\max}$ at all times
\end{assump}
\begin{assump}[persistent excitation]
\label{hyp:PE}
There exist constant parameters $T >0$ and $0 < \mu < 1$ such that $a(\cdot)$ satisfies
\begin{equation}
\label{eq:pe}
\frac{1}{T}\int_t^{t+T} \cro{a(\tau)}^T \cro{a(\tau)} d\tau \geq \mu I, \quad \forall t
\end{equation}
\end{assump}
The problem we address in this paper is the following.
\begin{Problem}
\label{Problem}
Under Assumptions~\ref{hypothese O borné}-\ref{hyp:PE}, find an estimate $\hat \o$ of $\o$ from the measurements $\a$ defined in \eqref{def : a,b}.
\end{Problem}
\begin{remark}[on the persistent excitation]
\label{rk:PE}
\eqref{eq:pe} is equivalent to
\begin{equation}
\label{eq:pe equivalent}
\frac{1}{T} \int_t^{t+T} \left(x^Ta(\tau)\right)^2 d\tau \leq 1-\mu, \quad \forall t, \quad \forall |x| = 1
\end{equation}
which is only possible if $a(\cdot)$ varies uniformly on every interval $[t,t+T]$. Without the PE assumption, Problem~\ref{Problem} may not have a solution. For example, the initial conditions
\begin{displaymath}
a(t_0) = \left(\begin{array}{c}
1 \\
0 \\
0
\end{array}\right), \quad
\o(t_0) = \left(\begin{array}{c}
w \\
0 \\
0
\end{array}\right)
\end{displaymath}
yield $a(t) = a(t_0)$ for all $t$, regardless of the value of $w$. Hence, the system is clearly not observable. Such a case is discarded by the PE assumption. Note that this assumption bears on the trajectory, hence on the initial condition $X(t_0)$ and on the torque $\tau$ only.
\end{remark}
\section{Observer definition and analysis of convergence}
\label{sec : observer}
\subsection{Observer definition}
The time derivative of the measurement $\a$ is
\begin{displaymath}
\label{dot ab}
\dot \a = \dot R^T \mathring{\a} = - \cro{\o} R^T \mathring{\a} = \a \times \o
\end{displaymath}
To solve Problem~\ref{Problem}, the main idea of the paper is to consider the reconstruction of the extended 6-dimensional state $X$ by its estimate $\hat X$
\begin{displaymath}
X = \left(
\begin{array}{c}
\a \\
\o
\end{array}\right), \quad
\hat X = \left(
\begin{array}{c}
\hat a \\
\hat \o
\end{array}\right)
\end{displaymath}
The state is governed by
\begin{displaymath}
\dot X = \left(
\begin{array}{c}
\a \times \o\\
E(\o) + J^{-1}\tau
\end{array}
\right)
\end{displaymath}
and the following observer is proposed
\begin{equation}
\label{def : observer}
\dot {\hat X} = \left(
\begin{array}{c}
\a \times \hat \o - k(\hat \a - \a)\\
E(\hat \o) + J^{-1}\tau + k^2 \a \times (\hat \a - \a)
\end{array}
\right)
\end{equation}
where $k>0$ is a constant (tuning) parameter. Note
\begin{equation}
\label{def erreur}
\tilde X \triangleq X - \hat X \triangleq \left(
\begin{array}{c}
\tilde \a \\
\to
\end{array}\right)
\end{equation}
the error state. We have
\begin{equation}
\label{eq : dot X tilde}
\dot{\tilde X} = \left(
\begin{array}{cc}
-k I & \cro{\a} \\
k^2 \cro{\a} & 0
\end{array}
\right)
\tilde X
+
\left(
\begin{array}{c}
0 \\
E(\o) - E(\hat \o)
\end{array}
\right)
\end{equation}
\subsection{Preliminary change of variables and properties}
The study of the dynamics \eqref{eq : dot X tilde} employs a preliminary change of coordinates. Note
\begin{equation}
\label{def : Z}
Z \triangleq \left(
\begin{array}{c}
\tilde \a \\
\frac{\to}{k}
\end{array}\right)
\end{equation}
yielding
\begin{displaymath}
\label{systeme Z perturbe}
\dot{Z} = kA(t)Z
+
\ll
\begin{array}{c}
0 \\
\frac{E(\o) - E(\hat \o)}{k}
\end{array}
\!\!\right)
\end{displaymath}
with
\begin{equation}
\label{def : A}
A(t) \triangleq \left(
\begin{array}{ccc}
- I & \cro{\a(t)} \\
\cro{\a(t)} & 0
\end{array}
\right)
\end{equation}
which we will analyze as an ideal linear time-varying (LTV) system
\begin{equation}
\label{systeme Z non perturbe}
\dot Z = k A(t) Z
\end{equation}
perturbed by the input term
\begin{equation}
\label{def : xi}
\xi \triangleq \ll
\begin{array}{c}
0 \\
\frac{E(\o) - E(\hat \o)}{k}
\end{array}
\!\!\right)
\end{equation}
We start by upper-bounding the disturbance \eqref{def : xi}.
\begin{proposition}[Bound on the disturbance]
For any $Z$, $\xi$ is bounded by
\begin{equation}
\label{prop : borne xi}
|\xi| \leq d(\sqrt{2}\o_{\max} |Z| + k |Z|^2)
\end{equation}
where $d$ is defined as
\begin{equation}
\label{def : d}
d \triangleq \max \left \{ \left|\frac{J_3-J_2}{J_1}\right|, \quad \left|\frac{J_1-J_3}{J_2}\right|, \quad \left|\frac{J_2-J_1}{J_3}\right| \right \}
\end{equation}
\end{proposition}
\begin{proof}
We have $$|\xi| = \frac{1}{k} |E(\o) - E(\hat \o)|$$
with, due to the quadratic nature of $E(\cdot)$,
\begin{align*}
& E(\o) - E(\hat \o) = J^{-1} \left( J\to \times \o + J\o \times \to - J\to \times \to \right) \\
& =
\left(
\begin{array}{c}
\frac{J_2-J_3}{J_1} (\o_2 \to_3 + \to_2 \o_3) \\
\frac{J_3-J_1}{J_2} (\o_3 \to_1 + \to_3 \o_1) \\
\frac{J_1-J_2}{J_3} (\o_1 \to_2 + \to_1 \o_2) \\
\end{array}
\right)
-
\left(
\begin{array}{c}
\frac{J_2-J_3}{J_1} \to_2 \to_3 \\
\frac{J_3-J_1}{J_2} \to_3 \to_1 \\
\frac{J_1-J_2}{J_3} \to_1 \to_2 \\
\end{array}
\right) \\
& \triangleq \delta_1 - \delta_2
\end{align*}
As a straightforward consequence
\begin{displaymath}
|\delta_2| \leq d|\to|^2
\end{displaymath}
Moreover, by Cauchy-Schwarz inequality
\begin{displaymath}
(\o_2 \to_3 + \to_2 \o_3)^2 \leq (\o_2^2+\o_3^2) (\to_2^3 + \to_3^2) \leq (\o_2^2+\o_3^2) |\to|^2
\end{displaymath}
Using similar inequalities for all the coordinates of $\delta_1$ yields
\begin{displaymath}
|\delta_1|^2 \leq 2 d^2 |\o|^2 |\to|^2 \leq 2 d^2\o_{\max}^2 |\to|^2
\end{displaymath}
Hence,
\begin{align*}
|\xi| \leq \frac{|\delta_1| + |\delta_2|}{k}& \leq d \sqrt{2} \o_{\max} \left|\frac{\to}{k}\right| + k d \left|\frac{\to}{k}\right|^2 \\
& \leq d (\sqrt{2} \o_{\max} |Z| + k |Z|^2)
\end{align*}
\end{proof}
\begin{remark}[on the quantity $d$]
As $J_1,J_2,J_3$ are the main moments of inertia of the rigid body, we have \cite{landau1982} (\S 32,9) $$J_i \leq J_j + J_k$$ for all permutations $i,j,k$ and hence ${0 \leq d \leq 1}$. Moreover, $d=0$ if and only if ${J_1=J_2=J_3}$. $d$ appears as a measurement of how far the rigid body is from an ideal symmetric body. For this reason, we call it \emph{distordance} of the rigid body. Examples:
\begin{itemize}
\item For a homogeneous parallelepiped of size ${l \times l \times L}$, with ${L \geq l}$, we have $$d = \frac{L^2-l^2}{L^2 + l^2}$$
\item For a homogeneous straight cylinder of radius $r$ and height $h$ we have $$d = \frac{\left|h^2-3r^2\right|}{h^2+3r^2}$$
\end{itemize}
\end{remark}
\subsection{Analysis of the LTV dynamics $\dot Z = k A(t) Z$}
\label{sec : Z'=kAZ}
The shape of $A(t)$ will appear familiar to the reader acquainted with adaptive control problems. Along the trajectories of~\eqref{systeme Z non perturbe} we have
\begin{displaymath}
\frac{d}{dt} |Z|^2 = -2k|Z_1|^2 = -Z^T C^T C Z
\end{displaymath}
with $${C \triangleq (\sqrt {2k} \quad 0)}$$ As will be seen in the proof of the following Theorem, the PE assumption will imply, in turn, that the pair $(kA(\cdot),C)$ is uniformly completely observable (UCO), which guarantees uniform exponential stability of the LTV system.
\begin{theorem}[LTV system exponential stability]
\label{thm Z' = kAZ}
There exists $0<c<1$ depending only on $T, \mu,k$ and $\o_{\max}$ such that the solution of \eqref{systeme Z non perturbe} satisfies for all integer $N\geq 0$ $$|Z(t)|^2 \leq c^{N} |Z(t_0)|^2, \quad \forall t\in [t_0+NT,t_0+(N+1)T]$$
for any initial condition $t_0, Z(t_0)$.
\end{theorem}
\begin{proof}
Along the trajectories of~\eqref{systeme Z non perturbe} we have
\begin{displaymath}
\frac{d}{dt} |Z|^2 = -2k|Z_1|^2 \leq 0
\end{displaymath}
which proves the result for $N = 0$. For all $t$
\begin{displaymath}
|Z(t+T)|^2 = |Z(t)|^2 - Z(t)^2 W(t,t+T) Z(t)
\end{displaymath}
where
\begin{displaymath}
W(t,t+T) \triangleq \int_t^{t+T} \phi(\tau,t)^T C^T C \phi(\tau,t) d\tau
\end{displaymath}
is the observability Gramian of the pair $(kA(\cdot),C)$ and $\phi$ is the transition matrix associated with~\eqref{systeme Z non perturbe}. Computing $W$ is no easy task. However, the output injection UCO equivalence result presented in \cite{ioannou1995} allows us to consider a much simpler system. Note
\begin{displaymath}
K(t) \triangleq \frac{\sqrt{k}}{\sqrt{2}} \left(\begin{array}{c}
I \\
-\cro{a(t)}
\end{array}\right)
\end{displaymath}
and
\begin{align*}
M(t) & \triangleq kA(t)+K(t) C \\
& = \left(\begin{array}{cc}
0 & k\cro{a(t)} \\
0 & 0
\end{array}\right)
\end{align*}
The observability Gramian $\widetilde W$ of the pair ${(M(\cdot),C)}$ is easily computed as
\begin{displaymath}
\widetilde W(t,t+T) = 2k \int_t^{t+T} \left(\begin{array}{cc}
I & \mathcal A(\tau,t) \\
\mathcal A(\tau,t)^T & \mathcal A(\tau,t)^T \mathcal A(\tau,t)
\end{array}\right)
d\tau
\end{displaymath}
where
\begin{displaymath}
\mathcal A(\tau,t) \triangleq k \int_t^{\tau} \cro{a(u)} du
\end{displaymath}
Such a Gramian is well known in optimal control and has been extensively studied e.g. in \cite{khalil1996}, Lemma~13.4. We have
\begin{itemize}
\item $\int_t^{t+T} k\cro{a(\tau)}^T k\cro{a(\tau)} d\tau \geq T k^2 \mu I, \quad \forall t$
\item $k\cro{a(\cdot)}$ is bounded by $k$
\item $\frac{d}{dt}k\cro{a(\cdot)}$ is bounded by $k\o_{\max}$
\end{itemize}
from which we deduce that there exists ${0 < \beta_1 < 1}$ depending on $T,\mu,k,\o_{\max}$ such that
\begin{displaymath}
\widetilde W(t,t+T) \geq \beta_1 I, \quad \forall t
\end{displaymath}
There also exists $\beta_2 > 0$ depending on $k,T$ such that ${\widetilde W(t,t+T) \leq \beta_2 I}$. From \cite{ioannou1995}, Lemma~4.8.1 (output injection UCO equivalence), $W(t,t+T)$ is also lower-bounded. More precisely, we have
\begin{displaymath}
W(t,t+T) \geq \frac{\beta_1}{2(1+\beta_2 T k)}I \triangleq (1-c) I
\end{displaymath}
with $0 < c < 1$. Assume the result is true for an integer ${N \geq 0}$. For any ${t \in [t_0+NT, t_0+(N+1)T]}$ we have
\begin{align*}
|Z(t+T)|^2 & = |Z(t)|^2 - Z(t)^TW(t,t+T) Z(t) \\
& \leq c |Z(t)|^2 \leq c^{N+1} |Z(t_0)|^2
\end{align*}
which concludes the proof by induction.
\end{proof}
\subsection{Convergence of the observer}
\label{sec : convergence}
Consider the quantity
\begin{equation}
\label{def : d*}
d^* \triangleq \frac{1-c}{2\sqrt 2 T \o_{\max}}
\end{equation}
where $c$ is defined in Theorem~\ref{thm Z' = kAZ}. The following Theorem, which is the main result of the paper, shows that if ${d < d^*}$, the observer~\eqref{def : observer} gives a solution to Problem~\ref{Problem}.
\begin{theorem}[main result]
\label{thm convergence obs}
We suppose that Assumptions~\ref{hypothese O borné}-\ref{hyp:PE} are satisfied and that
\begin{displaymath}
d < d^*
\end{displaymath}
where $d^*$ is defined in~\eqref{def : d*}. The observer~\eqref{def : observer} defines an error dynamics~\eqref{eq : dot X tilde} for which the equilibrium 0 is locally uniformly exponentially stable. The basin of attraction of this equilibrium contains the ellipsoid
\begin{equation}
\label{bassin d'attraction}
\left \{ \tilde X(0), \quad |\tilde \a (0)|^2 + \frac{|\to(0)|^2}{k^2} < r^2 \right \}
\end{equation}
with
\begin{equation}
\label{eq : def r}
r^2 \triangleq \frac{(1-c)^3}{8\sqrt{3}d^2T^3k^3}\left(1-\frac{2\sqrt 2 dT \o_{\max}}{1-c}\right)^2
\end{equation}
\end{theorem}
\begin{proof}
Consider the candidate Lyapunov function
\begin{displaymath}
V(t,Z) \triangleq Z^T \left( \int_t^{+\infty} \phi(\tau,t)^T \phi(\tau,t) d\tau \right)Z
\end{displaymath}
where $\phi$ is the transition matrix of system~\eqref{systeme Z non perturbe}. Let $(t,Z)$ be fixed. One easily shows that $kA(\cdot)$ is bounded by $k\sqrt 3$. Thus (see for example \cite{khalil2000} Theorem 4.12)
\begin{displaymath}
V(t,Z) \geq \frac{1}{2k\sqrt 3} |Z|^2 \triangleq c_1 |Z|^2 \triangleq W_1(Z)
\end{displaymath}
Moreover, Theorem~\ref{thm Z' = kAZ} implies that
\begin{align*}
V(t,Z) & = \sum_{N=0}^{+\infty} \int_{t+NT}^{t+(N+1)T} Z^T \phi(\tau,t)^T \phi(\tau,t) Z \\
& \leq T \sum_{N = 0}^{+\infty} c^N |Z|^2 = \frac{T}{1-c} |Z|^2 \\
& \triangleq c_2 |Z|^2 \triangleq W_2(Z)
\end{align*}
By construction, $V$ satisfies
\begin{displaymath}
\frac{\partial V}{\partial t}(t,Z) + \frac{\partial }{\partial Z}V(t,Z) kA(t) Z = -|Z|^2
\end{displaymath}
Hence, the derivative of $V$ along the trajectories of \eqref{systeme Z perturbe} is
\begin{align*}
\frac{d}{dt} V(t,Z) & = -|Z|^2 + \frac{\partial V}{\partial Z}(t,Z) \ \xi
\end{align*}
Using
\begin{align*}
\left|\frac{\partial}{\partial Z} V(t,Z)\right| & = 2 \left| \int_t^{+\infty} \phi(\tau,t)^T \phi(\tau,t) d\tau Z \right| \leq \frac{2T}{1-c} |Z|
\end{align*}
together with inequality \eqref{prop : borne xi} yields
\begin{displaymath}
\left|\frac{\partial V}{\partial Z}(t,Z) \ \xi \right| \leq \frac{2dT}{1-c} \left(\sqrt 2 \o_{\max} |Z|^2 + k |Z|^3\right)
\end{displaymath}
Hence
\begin{align*}
\frac{d}{dt} V(t,Z) & \leq - |Z|^2\left(1 - \frac{2\sqrt 2 dT \o_{\max}}{1-c} - \frac{2dTk}{1-c}|Z|\right) \\
& \triangleq -W_3(Z)
\end{align*}
By assumption $d < d^*$, which implies
\begin{align*}
1-\frac{2\sqrt 2dT \o_{\max}}{1-c} > 0
\end{align*}
We proceed as in \cite{khalil2000} Theorem~4.9. If the initial condition of \eqref{systeme Z perturbe} satisfies
\begin{align*}
|Z(t_0)| & < r \\
\Leftrightarrow |Z(t_0)| & < \frac{1-c}{2dkT} \left(1-\frac{2 \sqrt 2 dT\o_{\max}}{1-c}\right) \times \sqrt{\frac{c_1}{c_2}}
\end{align*}
then $W_3(Z(t_0)) >0$ and, while $W_3(Z(t)) >0$, $Z(\cdot)$ remains bounded by
\begin{align*}
|Z(t)|^2 & \leq \frac{V(t)}{c_1} \leq \frac{V(t_0)}{c_1} \leq \frac{c_2}{c_1} |Z(t_0)|^2
\end{align*}
which shows that
\begin{align*}
W_3(Z) \geq \left(1 - \frac{2 \sqrt 2 dT \o_{\max}}{1-c} - \frac{2dkT}{1-c} \sqrt{\frac{c_2}{c_1}}|Z(t_0)|\right) |Z|^2
\end{align*}
From \cite{khalil2000}, Theorem~4.10, \eqref{systeme Z perturbe} is locally uniformly exponentially stable. From \eqref{def : Z}, one directly deduces that the basin of attraction contains the ellipsoid \eqref{bassin d'attraction}.
\end{proof}
\begin{remark}
The limitations imposed on $\tilde \a(0)$ in \eqref{bassin d'attraction} are not truly restrictive because, as the actual value $\a(0)$ is assumed known, the observer may be initialized with ${\tilde \a(0) = 0}$. What matters is that the error on the unknown quantity $\o(0)$ can be large in practice.
\end{remark}
\section{PE assumption in free-rotation}
\label{sec:PE}
The PE Assumption~\ref{hyp:PE} is the cornerstone of the proof of the main result. It is interesting to investigate whether it is often satisfied in practice (we have seen in Remark~\ref{rk:PE} that it might fail). In this section we consider a free-rotation dynamics, namely $\tau = 0$. We will prove that Assumption~\ref{hyp:PE}, or equivalently condition~\eqref{eq:pe equivalent}, is satisfied for almost all initial conditions.
The following important properties hold.
\begin{itemize}
\item ${\omega^T J \omega}$ is constant over time (which implies that Assumption~\ref{hypothese O borné} is satisfied)
\item The \emph{moment of inertia} of the rigid body expressed in the inertial frame
\begin{equation}
\label{def : M}
\mathbf M \triangleq R(t)J\omega(t)
\end{equation}
is constant over time.
\item Thus, any trajectory ${t \mapsto \omega(t)}$ lies on the intersection of two ellipsoids
\begin{displaymath}
\omega^T J \omega = \omega(t_0)^T J \omega(t_0), \quad \omega^T J^2 \omega = \omega(t_0)^TJ^2 \omega(t_0)
\end{displaymath}
\end{itemize}
The analysis of the intersection of those ellipsoids is quite involved and has been extensively studied in e.g. \cite{landau1982}. It follows that there are four kinds of trajectories for the solutions $\omega$ of \eqref{eq euler}. We list them below, where $(\omega_1,\omega_2,\omega_3)$ are the coordinates of $\omega$ in the body frame.
\begin{description}
\item [Type 1]~ $\omega$ is constant, which is observed if and only if $\omega(t_0)$ is an eigenvector of $J$.
\item [Type 2]~ ${J_1>J_2>J_3}$ \emph{singular case}: $\omega_1(t)$ and $\omega_3(t)$ vanish, $\omega_2(t)$ tends to a constant when $t$ goes to infinity. This situation is observed only for a zero-measure set of initial condition $\omega(t_0)$.
\item [Type 3]~ ${J_1>J_2>J_3}$ \emph{regular case}: the trajectory is periodic and not contained in a plane. This situation is observed for almost all initial condition $\omega(t_0)$.
\item [Type 4]~ the trajectory is periodic and draws a non-zero diameter circle. This situation is observed if and only if two moments of inertia are equal and $\omega(t_0)$ is not an eigenvector of $J$.
\end{description}
Examples of such trajectories are given in Figures~\ref{fig:types 123}-\ref{fig:type 4} for various initial conditions.
\begin{figure}[!ht]
\includegraphics[width=\columnwidth]{types123}
\caption{Types 1, 2 and 3 trajectories in the case $J_1 > J_2 > J_3$ on an ellipsoid corresponding to a constant $|\mathbf M|$}
\label{fig:types 123}
\end{figure}
\subsection{Study of Type 1 and Type 2 solutions}
\label{sec : planar}
The simplest case one can imagine is when $\omega(t_0)$ (or simply $\omega$) is an eigenvector of $J$, namely for $i=1,2$ or $3$
\begin{displaymath}
J\omega = J_i\omega
\end{displaymath}
Note
\begin{displaymath}
R_0 \triangleq R(t_0), \quad w \triangleq |\omega|, \quad u \triangleq \frac{1}{w} R_0\omega
\end{displaymath}
\begin{proposition}
For all $t$, $R(t)$ writes
\begin{align*}
R(t) & = r_{u}(wt) R_0 \nonumber \\
& = \left(\c\, wt I + \ss wt \cro{u} + (1-\c\, wt ) u u^T\right)R_0
\end{align*}
where $\c, \, \textnormal{s}$ stand for $\cos, \, \sin$ respectively.
\end{proposition}
\begin{proof}
$R(t)$ and $r_u(wt)R_0$ have the same value $R(t_0)$ for $t = t_0$. Moreover,
\begin{align*}
\frac{d}{dt} r_{u}(wt)R_0 & = w \left(-\ss wt I + \c\, wt \cro{u} + \ss wt u u^T\right)R_0 \\
& = \left(\c\, wt \cro{u} + \ss wt \cro{u}^2 \right) w R_0 \\
& = \left(\c\, wt I + \ss wt \cro{u}\right) w \cro{u} R_0 \\
& = \left(\c\, wt I + \ss wt \cro{u} + (1-\c\, wt)u u^T\right)w \cro{u} R_0 \\
& = r_{u}(wt) w \cro{u} R_0 \\
& = r_u(wt)\cro{R_0\omega} R_0 = r_u(wt) R_0 \cro{\omega}
\end{align*}
Thus both functions satisfy~\eqref{eq R}, which concludes the proof by Cauchy-Lipschitz uniqueness theorem.
\end{proof}
It follows that for all $t$, $a(t)$ writes,
\begin{align}
\label{eq : a(t) planar}
a(t) & = R(t)^T \mathring{\a} \\
& = \c\, wt R_0^T\mathring{\a} - \ss wt R_0^T (u \times \mathring{\a}) + (1-\c\, wt) u^T\mathring{\a} R_0^Tu \nonumber
\end{align}
For this reason, we call \emph{planar rotation} the $R(\cdot)$ matrix generated by a Type 1 trajectory.
\begin{remark}
\label{rk : m et u}
The direction $u$ of the rotation can be simply computed from $\mathbf M$. We have
\begin{displaymath}
\mathbf M = RJ\omega = J_i R \omega = w J_i R R_0^T u = w J_iu
\end{displaymath}
which implies that
\begin{displaymath}
u = \frac{\mathbf M}{|\mathbf M|}
\end{displaymath}
\end{remark}
The impact of the planar nature of the rotation on the PE assumption is as explained in the next two subsections.
\subsubsection{Type 1 solution with $\mathbf M$ aligned with $\mathring{\a}$}
\label{sec:planar non pe}
Consider that $\mathring{\a}$ is aligned with $\mathbf M = R(t_0)J\omega(t_0)$. In this case $u = \pm \mathring{\a}$ (see Remark~\ref{rk : m et u}). Thus, \eqref{eq : a(t) planar} yields ${a(t) = R_0^T \mathring{\a}}$ constant over time. For any $T$ we have, for the unit vector ${x = R_0^T\mathring{\a}}$
\begin{displaymath}
\frac{1}{T} \int_0^T (a(s)^Tx)^2 ds = 1
\end{displaymath}
Thus, condition~\eqref{eq:pe equivalent} is not satisfied.
\subsubsection{Type 1 solution with $\mathbf M$ not aligned with $\mathring{\a}$}
\label{sec:planar pe}
Conversely, consider that $\mathring{\a}$ is not aligned with~$\mathbf M$. Define $v,z$ such that $(u,v,z)$ is a direct orthonormal basis of $\R{3}$. The decomposition of the unit vector $\mathring{\a}$ in this basis is given as
\begin{displaymath}
\mathring{\a} = a_1 u + a_2 v + a_3 z, \quad a_1^2 + a_2^2 + a_3^2=1, \quad \textrm{with} \quad a_1^2 < 1
\end{displaymath}
We have
\begin{align*}
a(t)= & R_0^T(a_1 u + (a_2 \c\, wt + a_3 \ss wt)v + (a_3 \c\, wt - a_2 \ss wt)z)
\end{align*}
For $T=\frac{2\pi}{w}$, any $t$ and any unit vector $${x = R_0^T(x_1 u + x_2 v + x_3 z)}$$ we have
\begin{align*}
& \frac{1}{T} \int_t^{t+T} \left(a(s)^T x\right)^2 ds = \\
& \frac{1}{T} \!\! \int_t^{t+T} \!\!\!\!\!\!\!\!\!\!\! \left(a_1 x_1 + (a_2 \c\, wt + a_3 \ss wt)x_2 + (a_3 \c\, wt - a_2 \ss wt)x_3\right)^2 \!ds \\
& = a_1^2 x_1^2 + \frac{a_2^2+a_3^2}{2}(x_2^2+x_3^2) \leq (1-\mu)
\end{align*}
with
\begin{displaymath}
\mu \triangleq \min \left(1-a_1^2,\frac{1+a_1^2}{2}\right) \in (0,1)
\end{displaymath}
Thus, condition~\eqref{eq:pe equivalent} is satisfied.
\subsubsection{Type 2 solutions}
As shown in \cite{landau1982}, the Type~2 solutions are characterized by $J_1>J_2>J_3$ and
\begin{displaymath}
\left|\omega_1(t_0)\right| = \sqrt{\frac{J_3(J_2-J_3)}{J_1(J_1-J_2)}} \left|\omega_3(t_0)\right| \neq 0
\end{displaymath}
which defines a zero-measure set. For this reason, they are called \emph{singular} solutions. In this case, $\omega(t)$ converges to a limit $\omega_{\infty} = (0,\pm w,0)$ when $t$ goes to infinity. The rotation $R(t)$ is thus asymptotically arbitrarily close to a planar rotation around $\mathbf M = R(t_0)J\omega(t_0)$. The same arguments as in Sections~\ref{sec:planar non pe},~\ref{sec:planar pe} show that condition~\eqref{eq:pe equivalent} is satisfied unless $R(t_0)J\omega(0)$ and $\mathring{\a}$ are aligned.
\subsection{Study of Type 3 and Type 4 solutions}
\label{sec:types 3 et 4}
In this section we will show that the Type 3 and Type 4 solutions satisfy the PE assumption. Both proofs relies on the following technical result.
\begin{proposition}[preliminary result]
\label{prop:preliminaire}
If condition~\eqref{eq:pe} is not satisfied, then for all $T>0$ and all $\varepsilon>0$ small enough, there exists $t$ such that for all ${y\in\R{3}}$, and all ${s \in [t,t+T]}$,
\begin{itemize}
\item $R(s)y$ remains between two planes orthogonal to $\mathring{\a}$ and distant by $\varepsilon|y|$
\item $R(s)^Ty$ remains between two parallel planes distant by $\varepsilon|y|$.
\end{itemize}
\end{proposition}
\begin{proof}
Consider $T > 0$ and $\mu$ such that
\begin{equation}
\label{eq : mu max}
0< \mu < \min\left(\frac{1}{4T\o_{\max}}, \frac{T\o_{\max}}{4} \right) < 1
\end{equation}
Assume that \eqref{eq:pe equivalent} is not satisfied. There exists $t,x$ such that $|x| = 1$ and \begin{equation}
\label{eq : non PE mu}
\frac{1}{T} \int_t^{t+T} \left(a(s)^T x\right)^2 ds \geq 1-\mu
\end{equation}
As will appear, one can use the bounded variations of $a(\cdot)$ due to its governing dynamics to establish a lower bound on the integrand. Note
\begin{displaymath}
h(s) \triangleq \left(a(s)^Tx\right)^2, \quad \forall s
\end{displaymath}
We will now show by contradiction that
\begin{displaymath}
h(s) \geq 1- 2 \sqrt{T\o_{\max} \mu}, \quad \forall s \in [t,t+T]
\end{displaymath}
Assume that there exists $s_0$ such that $${h(s_0) < 1-2 \sqrt{T\o_{\max} \mu}}$$ We have, for all $s$,
\begin{align*}
|\dot h(s)| & = \left|2 \dot a(s)^T x a(s)^T x\right| \\
& = \left|2(a(s)\times \omega)^T x a(s)^T x\right| \leq 2 \o_{\max}
\end{align*}
Assume $s_0 \leq t+\frac{T}{2}$ and note
\begin{displaymath}
s_1 \triangleq s_0+\sqrt{\frac{T \mu}{\o_{\max}}} \leq t+T
\end{displaymath}
We have, for any ${s \in [s_0,s_1]\subset [t,t+T]}$
\begin{align*}
h(s) & \leq h(s_0) + 2\o_{\max} (s-s_0) \\
& < 1-2\sqrt{T\o_{\max}\mu} + 2\o_{\max} (s-s_0)
\end{align*}
Hence
\begin{align*}
\int_t^{t+T} \left(a(s)^T x\right)^2 ds & < 1-\frac{1}{T}\sqrt{\frac{T\mu}{\o_{\max}}} \\
+ \frac{1}{T}\! \int_{s_0}^{s_1} \!\!\! & \left(1-2\sqrt{T\o_{\max} \mu} + 2\o_{\max} (s-s_0) \right)\! ds \\
& = 1-2\mu + \mu = 1-\mu
\end{align*}
which contradicts \eqref{eq : non PE mu}. The case ${s_0 > t+\frac{T}{2}}$ is analog with ${s \in [s_0-\sqrt{\frac{T\mu}{\o_{\max}}},s_0]\subset [t,t+T]}$. Finally, we have, for all $s$
\begin{displaymath}
0 < 1-2\sqrt{T\o_{\max} \mu} \leq \left(a(s)^Tx\right)^2 \leq 1
\end{displaymath}
which shows that the continuous function ${s\mapsto a(s)^Tx}$ is of constant sign, strictly positive without loss of generality. Thus, we have
\begin{displaymath}
0 < 1-2\sqrt{T\o_{\max} \mu} \leq a(s)^Tx \leq 1
\end{displaymath}
and in turn
\begin{equation}
\label{def : c}
|a(s)-x|^2 = 2-2a(s)^Tx \leq 4\sqrt{T\o_{\max} \mu} \triangleq \gamma \sqrt{\mu}
\end{equation}
Note $R_1$ a rotation matrix so that $${\mathring{\a} = R_1 x}$$ and, for all $s$, ${u(s), \xi(s)}$ such that
\begin{align*}
R(s) & \triangleq r_{u(s)}(\xi(s))R_1 \\% & = \left(\cos \xi(s) I + \sin \xi(s) \cro{u(s)} + \left(1-\cos \xi(s)\right) u(s)u(s)^T\right)R_1
\end{align*}
Note that $R(s)x = r_{u(s)}(\xi(s)) \mathring{\a}$. The next Lemma formulates that the rotation $R(s)$ is uniformly close to $r_{\mathring{\a}}(\xi(s))R_1$.
\begin{lemma}
\label{lemme : R presque planaire}
We have, for all $s\in[t,t+T]$ and all $y\in\R{3}$
\begin{equation}
\label{eq:R presque planaire}
\left|R(s) y - r_{\mathring{\a}}(\xi(s))R_1 y\right|^2 \leq 30 \gamma \sqrt{\mu} |y|^2
\end{equation}
where $\gamma$ is defined by~\eqref{def : c}.
\end{lemma}
\begin{proof}
Let ${s\in[t,t+T]}$. For clarity we may omit the $s$ dependency of $u$ and $\xi$. Note
\begin{align*}
\Delta & \triangleq R(s) - r_{\mathring{\a}}(\xi) R_1 \\
& = \left(\sin \xi \left(\cro{u} - \cro{\mathring{\a}}\right) + (1-\cos \xi) \left(u u^T - \mathring{\a} \mathring{\a}^T\right)\right) R_1
\end{align*}
If $\mathring{\a} = u(s)$, ${\|\Delta\| = 0 \leq 30\gamma\sqrt{\mu}}$. Otherwise, for $A=x$ we have, from~\eqref{def : c}
\begin{align*}
\left|\Delta A\right |^2 & = \left|R(s)x-r_{\mathring{\a}}(\xi)R_1x\right|^2 = \left|R(s)x-\mathring{\a}\right|^2 \\
& = \left|x-R^T(s)\mathring{\a}\right|^2 = |x-a(s)|^2 \leq \gamma\sqrt{\mu}
\end{align*}
Note $v,z$ so that ${(u,v,z)}$ is an orthonormal basis of $\R{3}$ write
\begin{displaymath}
\mathring{\a} = a_1 u + a_2 v + a_3 w, \quad a_1^2+a_2^2+a_3^2 = 1
\end{displaymath}
We have
\begin{align*}
\gamma\sqrt \mu & \geq \left|R(s)x - \mathring{\a}\right|^2 = \left|(r_u(\xi)-I) \mathring{\a}\right|\\
& = \left|(a_2(\c\, \xi-1) - a_3 \ss \xi) v + (a_2\ss \xi + a_3(\c\, \xi-1) )w\right|^2 \\
& = 4 (a_2^2+a_3^2) \sin^2 \frac{\xi}{2}
\end{align*}
Now, for $B = \frac{u \times x}{|u \times x|}$ we have
\begin{align*}
\left|\Delta B\right|^2 & = \frac{\sin ^2 \xi}{x_2^2+x_3^2} \left|u \times (u \times x)-x \times (u \times x)\right|^2 \\
& = \frac{\sin^2 \xi}{x_2^2+x_3^2}(1-x^Tu)^2 |u+x|^2 \\
& = \frac{\sin^2 \xi}{x_2^2+x_3^2}4(1-x_1^2)^2 \\
& \leq 16(x_2^2+x_3^2) \sin^2 \frac{\xi}{2} \leq 4\gamma\sqrt{\mu}
\end{align*}
For ${C = A \times B}$ we have
\begin{align*}
\left|\Delta C\right|^2 & = \left|QR(s)(A \times B) - P(s)(A \times B)\right|^2 \\
& = \left|QR(s)A \times QR(s) B - P(s) A \times P(s) B\right|^2 \\
& = \left|QR(s)A \times \Delta B + \Delta A \times P(s) B\right|^2 \\
& \leq 2(c\sqrt{\mu} + 4c \sqrt{\mu}) = 10 \gamma \sqrt{\mu}
\end{align*}
Finally, for any unit vector ${y = y_1A + y_2 B + y_3 C}$ we have
\begin{align*}
\left|\Delta y\right|^2 & = \left|y_1 \Delta A + y_2 \Delta B + y_3 \Delta C\right|^2 \\
& \leq 3 \left(y_1^2 \left|\Delta A\right|^2 + y_2^2 \left |\Delta B \right|^2 + y_3^2 \left|\Delta C\right|^2\right) \\
& \leq 3 (y_1^2+y_2^2+y_3^2) 10 \gamma\sqrt{\mu} = 30\gamma\sqrt{\mu} |y|^2
\end{align*}
which concludes the proof of Lemma~\ref{lemme : R presque planaire}.
\end{proof}
Note $\varepsilon = 2\sqrt{30\gamma\sqrt \mu}$ and consider any $y$ in $\R{3}$ and any $s$ in ${[t,t+T]}$. On the one hand, $r_{\mathring{\a}}(\xi(s))R_1y$ lies on a circle orthogonal to $\mathring{\a}$. On the other hand, $$\left|R(s)y - r_{\mathring{\a}}(\xi(s))R_1y\right| \leq \frac{\varepsilon}{2}|y|$$
This yields the first item of Proposition~\ref{prop:preliminaire} as ${\mu > 0}$ is arbitrary small. Rewriting the result of Lemma~\ref{lemme : R presque planaire} as
\begin{displaymath}
\left|R_1^Tr_{\mathring{\a}}(-\xi(s))y - R(s)^Ty\right|^2 \leq 30 \gamma\sqrt{\mu}|y|^2
\end{displaymath}
for any $s \in [t,t+T]$ and any $y$ yields the second item and concludes the proof.
\end{proof}
\subsubsection{Type 3 solutions}
\label{sec:type 3}
These solutions are characterized by $J_1 > J_2 > J_3$ and
$$|\omega_1(t_0)| \neq \sqrt{\frac{J_3(J_2-J_3)}{J_1(J_1-J_2)}} |\omega_3(t_0)|$$
In this case the trajectory of $\omega(\cdot)$ is closed and thus periodic of a certain period $\tau > 0$, and not contained in a plane. Assume that condition~\eqref{eq:pe equivalent} is not satisfied. We apply the second item of Proposition~\ref{prop:preliminaire} with $T = \tau$. For any $\varepsilon$ small enough, there exists $t$ such that for all $s \in [t,t+\tau]$
\begin{displaymath}
J \omega(s) = R^T(s) \mathbf M
\end{displaymath}
remains between two parallel planes and distant by $\varepsilon\left|\mathbf M\right|$. As $\omega(\cdot)$ is $\tau-$ periodic, this is true for all ${s \in \mathbf R}$. When $\varepsilon$ goes to 0, we conclude that the trajectory of $\omega(\cdot)$ remains in a plane, which is a contradiction. Thus, condition~\eqref{eq:pe equivalent} is satisfied, unconditionally on $R(t_0)$.
\subsubsection{Type 4 solutions}
\label{sec:type 4}
\begin{figure}[!ht]
\includegraphics[width=\columnwidth]{type4}
\caption{Type 4 trajectories in the case $J_1 = J_2 > J_3$ on an ellipsoid corresponding to a constant $|\mathbf M|$}
\label{fig:type 4}
\end{figure}
We now consider the case where $\omega(t_0)$ is not an eigenvector of $J$ and two moments of inertia are equal. In this case the trajectory $t \mapsto \omega(t)$ is a circle, as represented in Figure~\ref{fig:type 4}. Since it is contained in a plane, we can not apply directly the same technique as in Section~\ref{sec:type 3}. Without loss of generality, we study the case $J_1 = J_2 > J_3$ (the case ${J_1 > J_2=J_3}$ is analog). We thus consider a trajectory $\omega$ such that $\omega(0)$ satisfies
\begin{displaymath}
\left(\omega_1(t_0),\omega_2(t_0)\right) \neq (0,0), \quad \omega_3(t_0) \neq 0
\end{displaymath}
Following the extensive analysis exposed in \cite{landau1982}, we conveniently chose the inertial frame $(e_1,e_2,e_3)$ so that $e_3$ is aligned with $\mathbf M$, namely
\begin{displaymath}
e_3 = \frac{\mathbf M}{\left|\mathbf M\right|}
\end{displaymath}
For this choice of $e_3$ and in the case where ${J_1=J_2}$, equations \eqref{eq R}-\eqref{eq euler} simplify considerably and one can show that the rotation matrix satisfies for all $t$
\begin{equation}
\label{eq:R deux moments egaux}
R(t) = p \left(\begin{array}{ccc}
(\dots) & (\dots) & \c\, \xi_1 (t-t_1) \\
(\dots) & (\dots) & \ss \xi_1 (t-t_1)\\
\c\, \xi_2 (t-t_2) & \ss \xi_2 (t-t_2) & \frac{\sqrt{1-p^2} }{p}
\end{array}
\right)
\end{equation}
where $(\dots)$ designates terms that are irrelevant in the following analysis, $t_1,t_2$ are constant and
\begin{align*}
p & \triangleq \sqrt{\frac{J_1 ^2 \omega_1(t_0)^2 + J_1^2 \omega_2(t_0)^2}{J_1 ^2 \omega_1(t_0)^2 + J_2^2 \omega_2(t_0)^2 +J_3^2 \omega_3(t_0)^2} } \quad \in (0,1) \\
\xi_1 & \triangleq \sqrt{\omega_1(t_0)^2 + \omega_2(t_0)^2 + \frac{J_3^2}{J_1^2} \omega_3(t_0)^2} \quad > 0 \\
\xi_2 & \triangleq \left(\frac{J_3}{J_1}-1\right) \omega_3(t_0) \quad \neq 0
\end{align*}
We now show that condition~\eqref{eq:pe equivalent} is satisfied by contradiction. Assuming that it is not, one can apply the first item of Proposition~\ref{prop:preliminaire} with $$T = \max \left(\frac{2\pi}{\xi_1}, \frac{2\pi}{|\xi_2|}\right)$$ For $\varepsilon$ small enough, there exists $t$ such that for all ${s \in [t,t+T]}$ $R(s)e_3$ remains between two planes orthogonal to $\mathring{\a}$ and distant by $\varepsilon$. Moreover, expression~\eqref{eq:R deux moments egaux} yields for all $s$
\begin{displaymath}
R(s)e_3 = \left(\begin{array}{c}
p \cos \xi_1(s-t_1) \\
p \sin \xi_1(s-t_1) \\
\sqrt{1-p^2}
\end{array}\right)
\end{displaymath}
Simple geometric considerations show that
\begin{displaymath}
\sqrt{1-(\mathring{\a}^Te_3)^2} \leq \frac{\varepsilon}{2p}
\end{displaymath}
which yields $\mathring{\a} = \pm e_3$ when $\varepsilon$ goes to $0$. Hence for $\varepsilon$ small enough, and all $s \in[t,t+T]$
\begin{displaymath}
R(s)e_1 = \left(\begin{array}{c}
(\dots) \\
(\dots) \\
p \cos \xi_2 (s-t_2)
\end{array}
\right)
\end{displaymath}
remains between two planes orthogonal to $\mathring{\a} = \pm e_3$. Taking $\varepsilon < 2p$ yields a contradiction. The trajectories $R(t)e_1$ and $R(t)e_3$ are represented in Figure~\ref{fig:axes} for better visual understanding of the proof.
\begin{figure}[!ht]
\includegraphics[width = \columnwidth]{axes}
\caption{$R(t)e_3$ (dashed) and $R(t)e_1$ (solid) evolving on the unit sphere}
\label{fig:axes}
\end{figure}
\subsection{Conclusion}
In this section we have shown the following result.
\begin{theorem}
\label{thm:PE presque toujours vrai}
Consider the vector $$a(t) = R(t)^T \mathring{\a} $$ where $R(t)$ is a rotation matrix defined as the solution of the free-rotation dynamics \eqref{eq R}-\eqref{eq euler} with $\tau= 0$. Assumption~\ref{hyp:PE} is satisfied for almost all initial conditions ${(R(t_0),\omega(t_0))}$. It fails only in the cases listed below
\begin{enumerate}[(i)]
\item $\omega(t_0)$ is an eigenvector or $J$ and $R(t_0)J \omega(t_0)$ is aligned with $\mathring{\a}$, or
\item the eigenvalues of $J$ are of the form ${J_1 > J_2 > J_3}$, the coordinates of $\omega(t_0)$ in the trihedron of orthonormal eigendirections of $J$ satisfy
\begin{equation}
\label{eq : condition gamma}
\left|\omega_1(t_0)\right| = \sqrt{\frac{J_3(J_2-J_3)}{J_1(J_1-J_2)}} \left|\omega_3(t_0)\right|
\end{equation} and $R(t_0)J \omega(t_0)$ is aligned with $\mathring{\a}$.
\end{enumerate}
\end{theorem}
It follows that, except for the initial conditions listed in items $(i),(ii)$, the conclusion of Theorem~\ref{thm convergence obs} holds without requiring Assumption 2, which is automatically satisfied. Therefore, in almost all cases, observer~\eqref{def : observer} asymptotically reconstructs the desired angular velocity $\omega$.
\section{Simulation results}
\label{sec : simu}
In this section we illustrate the convergence of the observer and sketch the dependence with respect to the tuning gain $k$.
Simulations were run for a model of a CubeSat \cite{cubesat2014}. The rotating rigid body under consideration is a rectangular
parallelepiped of dimensions about ${20~\textrm{cm} \times 10~\textrm{cm} \times 10~\textrm{cm}}$ and mass $2$kg assumed to be slightly non-homogeneously distributed. The resulting moments of inertia are
\begin{displaymath}
J_1 = 87~\textrm{kg.cm}^2, \quad J_2 = 83~\textrm{kg.cm}^2, \quad J_3 = 37~\textrm{kg.cm}^2
\end{displaymath}
No torque is applied on this system, which is thus in free-rotation. Referring to Section~\ref{sec:PE}, we will consider Type 1 and Type 3 trajectories.
In this simulation the reference unit vector is the normalized magnetic field $\mathring{\a}$. The satellite is equipped with 3 magnetometers able to measure the normalized magnetic field $y_a$ in a magnetometer frame $\mathcal R_m$.
It shall be noted that, in practical applications, the sensor frame $\mathcal R_m$ can differ from the body frame $\mathcal R_b$ (defined along the principal axes of inertia) through a constant rotation $R_{m,b}$. With these notations, we have
\begin{displaymath}
a = R_{m,b}^T y_a
\end{displaymath}
which is a simple change of coordinates of the measurements.
For sake of accuracy in the implementation, reference dynamics and state observer~\eqref{def : observer}
were simulated using Runge-Kutta 4 method with sample period $0.01$s. The generated trajectories correspond to ${\o_{\max} \simeq 100}$~[rad/s].
\subsection{Noise-free simulations}
To emphasize the role of the tuning gain $k$, we first assume that the sensors are perfect i.e. without noise. Typical measurements for a general Type 3 trajectory are represented in Figure~\ref{fig:mesures}. As $J_1$ and $J_2$ are almost equal, the third coordinate is almost (but not exactly) periodic.
\begin{figure}[!t]
\includegraphics[width = \columnwidth]{mesures}
\caption{Typical measurements in the ideal noise-free case}
\label{fig:mesures}
\end{figure}
Figure~\ref{fig:convergence2} shows the convergence of the observer for various values of $k$.
\begin{figure}[!h]
\includegraphics[width=\columnwidth]{convergence}
\caption{Convergence of the observer for increasing values of $k$}
\label{fig:convergence2}
\end{figure}
Interestingly, large values of $k$ produce undesirable effects. This is a structural difference with the two reference vectors based observer previously introduced by the authors~\cite{magnis2014a}. The reason is that the convergence is guaranteed by a PE argument and not by a uniformly negative bound on eigenvalues.
In Figure~\ref{fig:nonPE} we represented the observer error for a case where the PE assumption is not satisfied, namely for a constant $\omega$ with $\mathbf M$ and $\mathring{\a} = (1,0,0)$ aligned. This is a singular case, as discussed earlier. Interestingly, the coordinates $\widetilde \omega_2$ and $\widetilde \omega_3$ converge to zero, while $\widetilde \omega_1$ converges to a constant value. This can easily be proved by using LaSalle invariance principle. Indeed, in this case, $\omega$ is constant and the measurements $a(\cdot)$ satisfy a LTI differential equation.
\begin{figure}[!ht]
\includegraphics[width=\columnwidth]{nonPE}
\caption{Without the PE assumption asymptotic convergence of the observer is lost, a bias remains.}
\label{fig:nonPE}
\end{figure}
\subsection{Measurement noise}
We now study the impact of measurement noise on the observer performance. The simulation parameters remain the same but we add Gaussian measurement noise with standard deviation $\sigma = 0.03~[\textrm{Hz}^{-\frac{1}{2}}]$. Typical measurements are represented in Figure~\ref{fig:mesures bruitees}.
\begin{figure}[!ht]
\includegraphics[width = \columnwidth]{mesures_bruit}
\caption{Vector measurement with additive noise}
\label{fig:mesures bruitees}
\end{figure}
The observer yields a residual error, about $5 \%$ in Figure~\ref{fig:convergence bruit} for $k=1$. Note that the measurement noise is filtered, thanks to a relatively low value of the gain $k$. For large values of $k$, the observer does not converge anymore (not represented).
\begin{figure}[!ht]
\includegraphics[width = \columnwidth]{convergence_bruit}
\caption{Observer performance under noisy measurement for $k=1$}
\label{fig:convergence bruit}
\end{figure}
\section{Conclusions and perspectives}
\label{sec : conclusion}
A new method to estimate the angular velocity of a rigid body has been proposed in this article. The method uses onboard measurements of a single constant vector. The estimation algorithm is a nonlinear observer which is very simple to implement and induces a very limited computational burden. At this stage, an interesting (but still preliminary) conclusion is that, in the cases considered here, rate gyros could be replaced with an estimation software employing cheap, rugged and resilient sensors. In fact, any type of sensors producing a 3-dimensional vector of measurements such as e.g., Sun sensors, magnetometers, could constitute one such alternative. Assessing the feasibility of this approach requires further investigations including experiments.
More generally, this observer should be considered as a first element of a class of estimation methods which can be developed to address several cases of practical interest. In particular, the introduction of noise in the measurement and uncertainty on the input torque (assumed here to be known) will require extensions such as optimal filtering to treat more general cases. White or colored noises will be good candidates to model these elements. Also, slow variations of the reference vector $\mathring{\a}$ should deserve particular care, because such drifts naturally appear in some cases.
On the other hand, one can also consider that this method can be useful for other estimation tasks. Among the possibilities are the estimation of the inertia $J$ matrix which we believe is possible from the measurements considered here. This could be of interest for the recently considered task of space debris removal~\cite{bonnal2013}.
Finally, recent attitude estimation techniques have favored the use of vector measurements \emph{together} with rate gyros measurements as inputs.
Among these approaches, one can find {\emph{i)} Extended Kalman Filters (EKF)-like algorithms e.g. \cite{choukroun2006,schmidt2008}, {\emph{ii)} nonlinear observers~\cite{mahony2008,martin2010,vasconcelos2008,tayebi2011,grip2011,trumpf2012}.
This contribution suggests that, here also, the rate gyros could be replaced with more in-depth analysis of the vector measurements.
\input{biblio.bbl}
\end{document}
|
\section{Introduction}
According to the color-screening model \cite{screening}, the dissociation
probability of the different quarkonium states ($c\bar{c}$ and $b\bar{b}$
mesons) due to the Quark-Gluon Plasma (QGP) is expected to provide essential
information about the properties of the system produced in heavy-ion collisions
(AA). Competing mechanisms called cold nuclear matter (CNM) effects (such as
gluon shadowing or coherent parton energy loss) can modify the quarkonium
production even in absence of the QGP \cite{CNM}, thus complicating the
interpretation of the results. Data from proton-nucleus collisions (pA) are
therefore necessary to disentangle these effects from the hot ones. Finally, new
measurements in pp collisions help to constrain the various models describing
the quarkonium production mechanisms.
The study of bottomonium production also complements the results obtained with
charmonia: for the latter system an important regeneration in AA collisions
might be expected at the LHC energies due to the large number of $c\bar{c}$ pair
produced, while this effect should be much smaller for the bottomonium
\cite{regeneration}. Moreover, the measurement of $\Upsilon$ allows a study in a
different Bjorken-$x$ range with respect to the J/$\psi$ and the theoretical
calculations for bottomonium are more robust due to the higher mass of $b$
quark.
ALICE \cite{ALICE} is the LHC experiment dedicated to the study of heavy-ion
collisions and has collected data in pp, p--Pb and Pb--Pb collisions.
At forward rapidity ($2.5<y<4$) quarkonia are reconstructed with the muon
spectrometer down to a transverse momentum
($p_\mathrm{T}$) equal to zero, exploiting their decay into $\mu^+\mu^-$.
\section{$\Upsilon$ production in pp collisions}
The $\Upsilon(1S)$ and $\Upsilon(2S)$ production cross sections have been
measured at forward rapidity in pp collisions at $\sqrt{s}=7$ TeV
\cite{ALICE_pp}. The rapidity dependence presented in Fig. \ref{fig: pp_sigma}
(left) shows a good agreement for both resonances with the measurements of
LHCb \cite{LHCb_pp} in the same range and complements the results obtained by
CMS at midrapidity \cite{CMS_pp_1,CMS_pp_2}.
\begin{figure}[htbp]
\centering
\includegraphics[width=0.45\textwidth]{pp_cross_section_y.pdf}
\includegraphics[width=0.42\textwidth]{pp_CSM_pt.pdf}
\caption{\small On the left, rapidity differential cross sections of
$\Upsilon(1S)$ and $\Upsilon(2S)$ measured by ALICE, LHCb \cite{LHCb_pp} and CMS
\cite{CMS_pp_1,CMS_pp_2}. On the right, $p_{\mathrm{T}}$ differential cross
section of $\Upsilon(1S)$ compared to three theoretical calculations
\cite{pp_th_1}.}
\label{fig: pp_sigma}
\end{figure}
In the right panel of Fig. \ref{fig: pp_sigma} the inclusive $\Upsilon(1S)$
production cross section as a function of $p_{\mathrm{T}}$ is compared to
Color Singlet Model (CSM) calculations which account for the feed-down from
higher mass states \cite{pp_th_1}. The leading order (LO) calculation
underestimates the data for $p_{\mathrm{T}} > 4$ GeV/c and falls too rapidly
with increasing $p_{\mathrm{T}}$. The $p_{\mathrm{T}}$ dependence of the
next-to-leading order (NLO) calculation is closer to the measurements, but the
prediction still underestimates the cross section over the full $p_{\mathrm{T}}$
range. A good agreement is achieved at a leading-$p_\mathrm{T}$
next-to-next-to-leading order (NNLO*), but over a limited $p_{\mathrm{T}}$ range
and with large theoretical uncertainties.
\section{$\Upsilon$ production in Pb--Pb collisions}
The effects of the hot and dense medium on the $\Upsilon(1S)$ production at
forward rapidity in Pb--Pb collisions at $\sqrt{s_\mathrm{NN}}=2.76$ TeV are
quantified by means of the nuclear modification factor ($R_{\mathrm{AA}}$),
defined as the meson yield in Pb--Pb divided by the production cross section in
pp collisions and the nuclear overlap function \cite{ALICE_PbPb}.
The $R_{\mathrm{AA}}$ in Fig. \ref{fig: Raa_Emerick} (left) shows a more
pronounced $\Upsilon(1S)$ suppression in central than in semiperipheral
collisions. Moreover, the rapidity dependence in the right panel suggests a
stronger suppression at forward than at midrapidity as it appears from the
comparison with the CMS point in $|y|< 2.4$ \cite{PbPb_CMS}.
\begin{figure}[htbp]
\centering
\includegraphics[width=0.45\textwidth]{Raa_th_centr_Emerick.pdf}
\includegraphics[width=0.45\textwidth]{Raa_th_rap_Emerick.pdf}
\caption{\small Inclusive $\Upsilon(1S)$ $R_{\mathrm{AA}}$ as a function
of the average number of participants (left) and rapidity (right) compared to
predictions from a transport model \cite{upsilon_Emerick}.}
\label{fig: Raa_Emerick}
\end{figure}
Predictions from a transport model are also shown in the same figures.
The calculation is based on a kinetic rate-equation approach in an evolving
QGP and include both suppression and regeneration effects
\cite{upsilon_Emerick}. CNM effects are calculated by varying an effective
absorption cross section between 0 and 2 mb, resulting in an uncertainty band.
This model underestimates the observed suppression, even if the centrality
dependence is fairly reproduced. The model predicts also an almost constant
$R_{\mathrm{AA}}$ as a function of the rapidity which is in disagreement with
the trend observed by ALICE and CMS.
Other predictions based on a dynamical model \cite{upsilon_Strickland} or
another transport model \cite{upsilon_Zhou}, not described here, show the same
difficulty to reproduce the ALICE data.
\section{$\Upsilon$ production in p--Pb collisions}
The nuclear modification factor ($R_{\mathrm{pPb}}$) measured in p--Pb
collisions at $\sqrt{s_{\mathrm{NN}}}=5.02$ TeV is used to determine the CNM
effects \cite{ALICE_pPb}. As shown in the two panels of Fig. \ref{fig: RpA}, the
inclusive $\Upsilon(1S)$ production is suppressed at forward rapidity (p-going
direction), while at backward rapidity (Pb-going direction) the measurement is
compatible with unity within uncertainties, disfavoring a strong gluon
antishadowing. At forward rapidity the $\Upsilon(1S)$ and J/$\psi$
$R_\mathrm{pPb}$ are rather similar. At backward rapidity, the J/$\psi$
measurement is systematically above that of $\Upsilon(1S)$, even if they are
still consistent within uncertainties \cite{jpsi_pPb}. Finally the
$R_\mathrm{pPb}$ measured by LHCb \cite{LHCb_pPb} is consistent within
uncertainties with the ALICE result.
\begin{figure}[htbp]
\centering
\includegraphics[width=0.45\textwidth]{upsilon_pPb_jpsi.pdf}
\includegraphics[width=0.45\textwidth]{upsilon_pPb_LHCb.pdf}
\caption{\small $R_{\mathrm{pPb}}$ of inclusive $\Upsilon(1S)$ as a function of
rapidity compared to the analogous measurements for the J/$\psi$ made by ALICE
\cite{jpsi_pPb} (left) and for $\Upsilon(1S)$ made by LHCb \cite{LHCb_pPb}
(right).}
\label{fig: RpA}
\end{figure}
In the left plot of Fig. \ref{fig: RpA_model}, the ALICE results have been
compared to a NLO Color Evaporation Model (CEM) calculation with shadowing
parametrized by EPS09 at NLO \cite{upsilon_EPS09} which tends to overestimate
the observed $\Upsilon(1S)$ $R_{\mathrm{pPb}}$. Coherent parton energy loss
calculations \cite{upsilon_eloss_1} with or without EPS09 are also shown: the
former reproduces the data at forward rapidity, while the latter is in better
agreement with the measurements at backward rapidity.
\begin{figure}[htbp]
\centering
\includegraphics[width=0.45\textwidth]{upsilon_pPb_model1.pdf}
\includegraphics[width=0.45\textwidth]{upsilon_pPb_model2.pdf}
\caption{\small Nuclear modification factor of inclusive $\Upsilon(1S)$ in
p--Pb collisions as a function of rapidity compared to several model
calculations
(\cite{upsilon_EPS09,upsilon_eloss_1,upsilon_eloss_2,upsilon_CGC}).}
\label{fig: RpA_model}
\end{figure}
In the right panel, the results are compared to a LO calculation of a
$gg\rightarrow\Upsilon g$ production with shadowing parametrization
\cite{upsilon_eloss_2}. The two bands show the uncertainties related to EPS09 LO
in the shadowing region and in the EMC region. A calculation at forward
rapidity based on the CGC framework coupled with a CEM production is also shown
\cite{upsilon_CGC}. Although this prediction only slightly
underestimates the $\Upsilon(1S)$ $R_{\mathrm{pPb}}$, it is not able to
reproduce the analogous J/$\psi$ measurement in the same rapidity range
\cite{jpsi_pPb}.
Finally, the forward-to-backward ratio ($R_\mathrm{FB}$) defined as the ratio
of the nuclear modification factors at forward and backward rapidities is shown
in Fig. \ref{fig: RFB}. It has the advantage to be independent from the pp cross
section which represents the main source of systematic uncertainties, but it can
only be measured in the restricted rapidity range $2.96<|y_\mathrm{cms}|<3.53$.
The $R_\mathrm{FB}$ is compatible with unity and is larger than that of the
J/$\psi$ \cite{jpsi_pPb}. All models describe the data within the present
uncertainties of the measurement.
\begin{figure}[htbp]
\centering
\includegraphics[width=0.45\textwidth]{upsilon_pPb_RFB1.pdf}
\includegraphics[width=0.45\textwidth]{upsilon_pPb_RFB2.pdf}
\caption{\small Forward-to-backward ratio of inclusive $\Upsilon(1S)$ yields
compared to the J/$\psi$ data \cite{jpsi_pPb} (left) and to the theoretical
models previously described (right).}
\label{fig: RFB}
\end{figure}
\section{Conclusions}
The measured inclusive production cross sections of $\Upsilon(1S)$ and
$\Upsilon(2S)$ at forward rapidity in pp collisions are in good agreement with
the results obtained by LHCb and complement those by CMS at midrapidity.
Both CSM LO and NLO calculations underestimate the data at large transverse
momentum, while the addition of the NNLO* contributions
helps to reduce this disagreement, but with large theoretical uncertainties.
The observed suppression of inclusive $\Upsilon(1S)$ in Pb--Pb collisions
increases with the centrality and rapidity as shown in the large domain covered
by ALICE and CMS. The suppression, larger than what predicted by the models
considered, might point to a significant dissociation of direct $\Upsilon(1S)$.
Finally, the $\Upsilon(1S)$ production in p--Pb collisions is suppressed at
forward rapidity, while at backward rapidity is consistent with unity. Models
including the nuclear modification of the gluon PDF or a contribution from
coherent parton energy loss tend to overestimate the measured $R_{\mathrm{pPb}}$
and cannot simultaneously describe the forward and backward rapidity
measurements.
|
\section{Introduction}
The classical term of \textit{unstable homeomorphism} ( now known as \textit{expansiveness}) first introduced by Utz in \cite{Ut50}, which is used to study the dynamical behavior saying roughly that every orbit can be accompanied by only one orbit with some certain constant.
It is clear that expansiveness implies the notion of sensitivity, which is the kernel in the definition of Devaney's chaos \cite{De89}. Hence expansiveness property involves a large class of dynamical systems exhibiting chaotic behavior, and nowadays an extensive literature has been developed on this property and its generalizations, see \cite{Ar14,AC14,APV13,Ei66,Ka93, KP99, KR69,Ma79,Mo12,MS10,Re65,Ut50,Wa00} and references therein for more knowledge.
\medskip
Among all the generalizations, the notion of \textit{$n$-expansiveness} originally introduced in \cite{Mo12} is an interesting one. Roughly speaking it loosens restriction to every orbit allows at most $n$ companion orbits with a certain constant. Note that the notion of positive $n$-expansiveness can be similarly defined when positive orbits are considered instead. Then the question that whether these generalized expansive systems can share the properties of the classical ones or not, was addressed naturally. It turns out that both positive and negative answers were provided in
\cite{Mo12}, and one particular result is that there are infinite compact metric spaces carrying positively $n$-expansive homeomorphisms for some $n\in\N$ (see \cite[Theorem 4.1]{Mo12}), which differ from the positively expansive ones.
Another natural question posed in \cite{Mo12} is that whether there are examples of compact metric spaces admitting fixed level of positively generalized expansive homeomorphisms, i.e. positively $n$- expansive homeomorphisms that are not positively $n-1$-expansive for some integer $n\ge 2$. Note that Morales partially solved this question by showing this is true for $n=2^k\ (k\in\N)$ (see \cite[Propostion 3.4]{Mo12}). By the same spirit we can ask this classification question for all the generalized expansiveness. In \cite{APV13} the authors gave an example of a $2$-expansive homeomorphism on surface which is not expansive, and the general examples are still open. It is worth mentioning that A.~Artigue \cite{Ar14} recently introduced another variant notion of expansiveness, say \textit{$(m,l)$-expansiveness} for given integer number $m> l\ge 1$, which presents a fine division among $n$-expansiveness (see \cite[Table 1]{Ar14} for basic hierarchy), but the examples to distinguish all different hierarchies are not available too.
According to the cardinality of companion orbits, Artigue and Carrasco-Olivera in \cite{AC14} further generalize expansiveness to \textit{$\aleph_0$-expansiveness}, where $\aleph_0$ is the first countable ordinal number, and they proved that $\aleph_0$-expansive homeomorphism is equivalent to another form of generalized expansive homeomorphism in the measurable sense (see \cite[Theorem 2.1]{AC14}).
\medskip
In this paper for simplicity we introduce the notion of \textit{essential $n$-expansiveness} (resp. \textit{essential $\aleph_0$-expansiveness}) to express $n$- but not $n-1$-expansiveness (resp. countable but not finite expansiveness), and the positively essential ones are similar to introduce. One can turn to Section \ref{sect:Def} for the precise definitions and their basic properties. In the sequential Section \ref{sect:LevPosExpan} and Section \ref{sect:LevExpan}, examples of all different levels of the generalized positive expansiveness and expansiveness are given, which completely solve the question left in \cite{Mo12} (see Theorem \ref{thm:EssPosNExpan} and Theorem \ref{thm:EssNExpan}). Among other things, when the space considered has countable cardinality, it turns out that there is no compact metric space carrying positively $n$-expansive homeomorphism for any $n\in\N$ (see Theorem \ref{thm:CountPosNExpan}), which extends the classical result in some sense (compare with \cite{KR69, MS10} and \cite[Theorem 4.1]{Mo12}); moreover, under the countable assumption we can also develop a characterization of $n$-expansiveness (see Theorem \ref{thm:NExpanChar}), and note that it is a natural generalization of Kato and Park \cite[Theorem 2.2]{KP99}.
Non-wandering points play an important role in the study of dynamical systems. Parallel to this classical theory, D.~Kwietniak \textit{et al} in \cite{KLOX15} introduced the notions of \textit{multi-non-wandering point} and the corresponding \textit{van der Waerden center} and \textit{depth}. In this paper as a corollary of Theorem \ref{thm:NExpanChar} we demonstrate that the van der Waerden depth is a countable ordinal and for every countable ordinal $\alpha$ there exists a compact metric system with van der Waerden depth equal to $\alpha$ (see Corollary \ref{cor:Van}). It answers positively a conjecture left open in \cite{KLOX15}.
\section{Definitions and basic properties}\label{sect:Def}
In this paper a \textit{topological dynamical system} (abbr. t.d.s.) is a pair $(X, T)$, where $X$ is a compact metric space and $T\colon X \to X$ is a homeomorphism from $X$ into itself. When discussing the positive notions, we may loosen $T$ to a continuous surjective map. Also, throughout this paper we denote $\N$, $\Z_+$, $\Z$ and $\R$ by the sets of positive integers,
nonnegative integers, integers and real numbers, respectively.
\medskip
Let $(X,T)$ be a compact metric t.d.s. with metric $d$. Fix $\delta>0$, we put
$$
\Gamma_\delta[x,T] =\{y\in X\colon d(T^n x, T^n y)\le\delta, \forall n\in \Z\}=\{x\}
$$
and
$$
\Phi_\delta[x,T] =\{y\in X\colon d(T^n x, T^n y)\le\delta, \forall n\in \Z_+\}=\{x\}
$$
when $T$ is not required to be a homeomorphism. We often write $\Gamma_\delta[x]$ or $\Phi_\delta[x]$ when the acting map $T$ is clear from the context.
\begin{de}[\cite{Ut50,Ei66}]\label{De:Expa}
A homeomorphism (resp. continuous surjective map) $T$ is said to be \textit{expansive} (resp. \textit{positively expansive}) if there is an \textit{expansive constant} $\delta>0$ for $T$ such that for every $x\in X$, $\Gamma_\delta[x]=\{x\}$ (resp. $\Phi_\delta[x]=\{x\}$).
\end{de}
In \cite{Mo12} Morales first introduced the notion of \textit{$n$-expansiveness}, which is a natural generalization of the usual expansiveness.
\begin{de}\label{de:nExpa}
Let $n\in\N$. A homeomorphism (resp. continuous surjective map) $T$ is said to be \textit{$n$-expansive} (resp. \textit{$n$-positively expansive}) if there is an \textit{$n$-expansive constant} $\delta>0$ for $T$ such that for every $x\in X$, $\Gamma_\delta[x]$ (resp. $\Phi_\delta[x]$) has at most $n$ elements.
\end{de}
Clearly $1$-expansiveness is just the classical expansiveness. Now by $\aleph_0$ denote the first countable cardinality. In the same spirit Artigue and Carrasco-Olivera \cite{AC14} extend the expansiveness to the following case:
\begin{de}\label{de:CountExap}
A homeomorphism (resp. continuous surjective map) $T$ is said to be \textit{$\aleph_0$-expansive} (resp. \textit{$\aleph_0$-positively expansive}) if there is an \textit{$\aleph_0$-expansive constant} $\delta>0$ for $T$ such that for every $x\in X$, $\Gamma_\delta[x]$ (resp. $\Phi_\delta[x]$) has at most countable elements.
\end{de}
\begin{de}\label{de:essential}
we call a homeomorphism $T$ is \textit{essentially $n$/$\aleph_0$-expansive} (resp. \textit{essentially positively $n$/$\aleph_0$-expansive}) if it is $n$/$\aleph_0$-expansive (resp. positively $n$/$\aleph_0$-expansive) and for any $\delta>0$ there is at least one point $x$ such that the cardinality of $\Gamma_\delta[x]$ (resp. $\Phi_\delta[x]$) is $n/\aleph_0$ .
\end{de}
It is easy to see that a homeomorphism $T$ is essentially $n$-expansive (resp. $\aleph_0$-expansive) if and only if it is $n$- but not $n-1$-expansive (resp. countable but not finite expansive). The equivalence for the corresponding positive cases are similar to achieve.
\begin{rem}\label{rem:DefExpa}
We have the following facts:
\begin{enumerate}
\item\label{rem:DefExpa:1} Another way to give the above concepts is to generalize the notion of generator introduced by Keynes and Robertson \cite{KR69}. That is, $T$ is $n/\aleph_0$-expansive if and only if there is a finite open cover $\alpha$ of $X$ for $T$ such that if for every bisequence $\{A_n\}_{n=-\infty}^{\infty}$ of members of $\alpha$, $\Card (\cap_{n=-\infty}^\infty T^{-n} \ol{A_n})$ is at most $n/\aleph_0$. Here $\Card(\cdot)$ means the cardinality of the set.
From this definition we can easily see that $n/\aleph_0$-expansiveness is a topological conjugacy invariant, and it is independent of the metric as long as the metric induces the topology of $X$ (although the $n/\aleph_0$-expansive constant does change).
\item\label{rem:DefExpa:2} It is clear that $n$-expansiveness implies $\aleph_0$-expansiveness for any $n\in\N$ and $n$-expansiveness implies $m$-expansiveness for any $m\ge n\in\N$.
\item\label{rem:DefExpa:3} A subsystem of an (essentially) (resp. positively) $n/\aleph_0$-expansive t.d.s. is (resp. positively) $n/\aleph_0$-expansive.
\end{enumerate}
\end{rem}
Let $(X,T)$ be a t.d.s. We say $x\in X$ is a \textit{periodic point} if $T^n x=x$ for some $n\in\N$, and a \textit{fixed point} if such $n=1$. Denote by $\textrm{Per}(X,T)$ (resp. $\textrm{Fix(X,T)}$) the set of all periodic (resp. fixed) points.
Now put
$$
\alpha(x)=\{y\in X\colon \lim_{i\to -\infty} T^{n_i} x\to y\}\ \mbox{and}\ \omega(x)=\{z\in X\colon \lim_{i\to +\infty} T^{n_i} x\to z\}
$$
Call some point $x$ has \textit{converging semi-orbits under $T$} if both $\alpha(x)$ and $\omega(x)$ consist of a single point. Put $\textrm{CS}(X,T)$ as the collection of all points having converging semi-orbits under $T$.
It is well known that under the classical expansiveness assumption $\textrm{Fix(X,T)}$ is finite (see \cite[Theorem 5.26]{Wa00}) and $\textrm{Per}(X,T)$ and $\textrm{CS}(X,T)$ are countable ( for instance \cite[Theorem 3.1]{Ut50} and \cite[Theorem 1]{Re65}, respectively). Now we improve these results to the generalized expansiveness.
\begin{thm}\label{thm:fact}
We have the following generalizations:
\begin{enumerate}
\item\label{fact:1} If $k\neq 0$ then $T$ is (resp. essentially) (resp. positively) $n/\aleph_0$-expansive if and only if so is $T^k$.
\item\label{fact:2} If $T$ is $n$-expansive, then $\emph{Fix}(X,T)$ is finite and $\emph{Per}(X,T)$ are countable,
\item\label{fact:3} If $T$ is $\aleph_0$-expansive, then $\emph{Fix}(X,T)$, $\emph{Per}(X,T)$ and $\emph{CS}(X,T)$ are countable.
\end{enumerate}
\end{thm}
\begin{proof}
\eqref{fact:1} We only prove that $T$ is $n/\aleph_0$-expansive if and only if so is $T^k$, and the other cases are similar. Since $T$ is continuous, there is $\ep>0$ such that whenever $d(x,y)<\ep$ then $d(T^i x, T^i y)<\delta$ for all $-k\le i\le k$ . Thus $\Gamma_\ep[x,T^k]\subset \Gamma_\delta[x,T]$ for all $x\in X$, which yields the necessity. On the other hand, we clearly have $\Gamma_\delta[x,T]\subset \Gamma_\delta[x,T^k]$, so the sufficiency holds.
\eqref{fact:2} Note that $\textrm{Per}(X,T)=\bigcup_{k\in\N} \textrm{Fix}(X,T^k)$, by \eqref{fact:1} it suffices to show $\textrm{Fix}(X,T)$ is finite whenever $T$ is $n$-expansive. Choose the $n$-expansive constant $\delta>0$ for $T$. Let $x_1,x_2,\dots,x_m\in X$ be such that $X=\bigcup_{i=1}^m B_{\delta/2}(x_i)$ due to compactness. If the contrary there are infinitely many fixed points $y_1,y_2,\dots$ in $B_{\delta/2}(x_{i_0})$ for some $1\le i_0\le m$. Put $A=\{y_1,y_2,\dots\}$. It is clear that for any $y_i\neq y_j\in A$ we have $d(T^l y_i, T^l y_j)\le \delta$ for each $l\in\Z$. However $\Card (A)$ is infinite which contradicts the definition of $\delta$. The proof ends.
\eqref{fact:3} Assume that $T$ is $\aleph_0$-expansive with $\aleph_0$-expansive constant $\delta>0$. First to claim that $\textrm{Fix}(X,T)$ and $\textrm{Per}(X,T)$ are countable. Similar as before it remains to prove $\textrm{Fix}(X,T)$ is countable whenever $T$ is $\aleph_0$-expansive. If not we apply the same manner to obtain a ball with radius $\delta/2$ containing uncountable fixed points, which yields a contradiction with the $\aleph_0$-expansive constant $\delta$. This proves the claim.
Now we show $\textrm{CS}(X,T)$ is also countable. Enumerate the countable set of fixed points as $z_1,z_2,\dots$. Consider the decomposition that
$$
\textrm{CS}(X,T)=\bigcup\nolimits_{i,j,k\in\N} \textrm{CS}(i,j,k),
$$
where
$$
\textrm{CS}(i,j,k)=\{x\in X\colon d(T^{-n} x, z_i)\le \delta/2 \mbox{ and } d(T^n x, z_j)\le \delta/2 \mbox{ for all } n\ge k\}.
$$
Clearly $\textrm{CS}(i,j,k)$ is compact. Since the countable union of countable sets is still countable, so if $\textrm{CS}(X,T)$ is uncountable, there are $i_0,j_0,k_0$ such that $\textrm{CS}(i_0,j_0,k_0)$ is uncountable.
On the other hand, by compactness we have
$$
\textrm{CS}(i_0,j_0,k_0)=\bigcup\nolimits_{s=1}^{\ t} B_{\ep_s}(x_s),
$$
where
$$
B_{\ep_s}(x_s)=\{y\in X\colon \mbox{ if } d(x_s, y)<\ep_s \mbox{ then } d(T^n x_s, T^n y)\le \delta/2 \mbox{ for all } |n|\le k_0\}.
$$
Thus there is $1\le s_0\le t$ such that $B_{\ep_{s_0}}(x_{s_0})$ contains uncountable elements. It implies that if $y_i\neq y_j\in B_{\ep_{s_0}}(x_{s_0})$ then $d(T^n y_i, T^n y_j)\le \delta$ holds for all $n\in\Z$, a contradiction with the choice of $\delta$. The proof is completed.
\end{proof}
We know that an interval or unit circle carries no expansive homeomorphisms (see for instance \cite{Re65} and \cite[Theorem 5.27]{Wa00}). Now we generalize these results to the case of $\aleph_0$-expansive homeomorphisms.
\begin{cor}\label{cor:nonexist-interval}
There is no $\aleph_0$-expansive homeomorphism of a compact interval.
\end{cor}
\begin{proof}
By Theorem \ref{thm:fact}\eqref{fact:1} we can assume that $T$ is orientation-preserving (if necessary replace $T$ by $T^2$). If $T$ is an $\aleph_0$-expansive homeomorphism acting on interval $[0,1]$, then it is easy to see that every point in $[0,1]$ has converging semi-orbit under $T$, which contradicts with Theorem \ref{thm:fact}\eqref{fact:3}.
\end{proof}
\begin{cor}\label{cor:nonexist-cicle}
There is no $\aleph_0$-expansive homeomorphism of an unit circle $S^1$.
\end{cor}
\begin{proof}
If $T$ is an $\aleph_0$-expansive homeomorphism on unit circle $S^1$, then by Corollary \ref{cor:nonexist-interval} and Remark \ref{rem:DefExpa}\eqref{rem:DefExpa:3} it has no fixed points.
By \cite[Theorem 6.18]{Wa00} there is a continuous surjection $\phi\colon S^1\lra S^1$ and a minimal rotation $S\colon S^1\lra S^1$ such that $\phi\, T=S\,\phi$, and for each $z\in S^1$ the set $\phi^{-1}(z)$ is either a point or closed interval. If each set $\phi^{-1}(z)$ is a point, then $\phi$ is a homeomorphism and $T$ is not $\aleph_0$-expansive because the minimal rotation $S$ is equicontinuous. Assume that for some $z_0$ the set $\phi^{-1}(z_0)$ is a closed interval of positive length. Since $\phi\, T=S\,\phi$, the sets $\{T^{-l}\phi^{-1}(z_0)\colon l\in\Z\}$ are mutually disjoint closed intervals. For any $\delta>0$ we can choose $N$ such that if $|l|\ge N$ the length of $T^{-l}\phi^{-1}(z_0)$ is less than $\delta$. Then by continuity of $T$ we can find a subinterval $A$ of $\phi^{-1}(z_0)$ with length less than some $\ep>0$ such that for any $a_1, a_2\in A$, $d(T^l a_1, T^l a_2)\le \delta$ for all $|l|\le N$. This implies that $\diam T^l A\le \delta$ for all $l\in \Z$. As $\Card (A)$ is uncountable, thus $\delta$ is not an $\aleph_0$-expansive constant for $T$. That is $T$ is not $\aleph_0$-expansive, a contradiction.
\end{proof}
Next we study the relationship between $\aleph_0$-expansive homeomorphisms and dimension. The definition and basic properties of dimension can be found in the book of Hurewicz and Wallman \cite{HW48}. Now we recall the notion of \textit{continuum-wise expansive homeomorphism}, which is another form of generalization first introduced by Kato \cite{Ka93}.
By a \textit{continuum} we mean a compact metric and connected non-degenerated space. A \textit{subcontinuum} is a continuum which is a subset of a space.
\begin{de}\label{de:cw-expan}
Let $(X,T)$ be a t.d.s. The homeomorphism $T$ is \textit{continuum-wise expansive} if there exists a constant $\delta>0$ such that for any nondegenerate subcontinuum $A$ of $X$, there is $n\in\Z$ such that $\diam (T^n A)>\delta$, where $\diam (A)=\sup \{d(a_1,a_2)\colon a_1,a_2\in A\}$.
\end{de}
The following lemma can be easily deduced from definitions, one can also refer to \cite{APV13,AC14}. Here we provide the details for the sake of completeness.
\begin{lem}\label{lem:CoutExp-cwExp}
If $(X,T)$ is $\aleph_0$-expansive then it is continuum-wise expansive.
\end{lem}
\begin{proof}
By definition if $X$ is $0$-dimensional then $T$ is always continuum-wise expansive. Now assume $\dim X>0$. Let $A$ be any non-degenerate subcontinuum of $X$ and $x\in A$. Then $\Card (A)$ contains uncountable elements by the non-degeneracy. Since $T$ is $\aleph_0$-expansive and assume the $\aleph_0$-expansive constant is $\delta>0$, then there exists a point $y\in A\setminus \Gamma_\delta[x]$. This implies that $d(T^n x, T^n y)>\delta$ for some $n\in\Z$, and then $\diam (T^n A)>\delta$. Thus $T$ is continuum-wise expansive with respect to $\delta$, completing the proof.
\end{proof}
An famous theorem by Ma\~n\`e \cite{Ma79} says that a compact metric space $X$ that admits an expansive homeomorphism $T$ is finite dimensional and every minimal set of $(X,T)$ is $0$-dimensional. Later Kato \cite[Theorem 5.2]{Ka93} proved that this theorem can be improved to the continuum-wise expansiveness case. By Lemma \ref{lem:CoutExp-cwExp} we immediately have the following theorem.
\begin{thm}\label{thm:dimension}
Let $(X,T)$ be an $\aleph_0$-expansive t.d.s. Then $\dim X<\infty$ and every minimal set of $T$ is $0$-dimensional.
\end{thm}
\begin{rem}\label{rem:minNExpan-minExpan}
Note that when $T$ is a $2$-expansive homeomorphism defined on a compact boundaryless surface with nonwandering set being the whole surface then $T$ is expansive (see \cite[Theorem A]{APV13}). So we may ask that if any minimal $n/\aleph_0$-expansive t.d.s. is expansive? If this were done, then coupled with Theorem of Ma\~n\`e we immediately have $X$ is zero-dimensional.
\end{rem}
\section{Levels of positive expansiveness}\label{sect:LevPosExpan}
It is a natural question as to whether a compact metric t.d.s. can admit an essentially positively $n$-expansive homeomorphism for any $n\in\N$. In \cite[Proposition 3.4]{Mo12} Morales gave a partial answer by showing that there is a t.d.s. $(X,T)$ which is positively $2^k$-expansive but not positively $(2^k-1)$-expansive for each $k\in\N$. Motivated by this example we here display a complete solution.
\begin{thm}\label{thm:EssPosNExpan}
There exists an essentially positively $n$-expansive t.d.s. $(X,T)$ for every $n\in\N$.
\end{thm}
\begin{proof}
Note that when $n=1$ it is just the classical positive expansiveness. Now Let $n\ge 2$. To begin with we recall a concrete construction of Denjoy homeomorphism of the circle $S^1$. Let $\alpha$ be an irrational number and $T_{\alpha}\colon S^1\to S^1,\ x\mapsto x+\alpha \ (\mbox{mod}\ 1)$. It is well known that $(S^1, T_\alpha)$ is minimal, i.e. the orbit closure of each point is the whole circle. Now fix $x_0\in S^1$, ``blow up'' each point of the orbit $\{T^k_\alpha x_0\colon k\in\Z\}$ by inserting arc $I_k= [a_k,b_k]$ (in the anticlockwise sense) such that for each $k\in \Z$,
\begin{itemize}
\item[(a)] $l(I_{k+1})=l(I_k)/2$ and $l(I_{-k})=l(I_k)$ for all $k\in\N\cup \{0\}$;
\item[(b)] $\sum_{k\in \Z} l(I_k)=1$, here $l(I)$ is the length of arc $I$.
\end{itemize}
We denote the expanded circle as $Y$ and the metric $d$ on it defined by $d(x,y)=\min \{l[x,y], l[y,x]\}$. Then there exists a monotone Denjoy homeomorphism $h\colon Y\to Y$ such that
$h(a_k)=a_{k+1},\ h(b_k)=b_{k+1},\ h(I_k)=I_{k+1}$ and $h(x)=T_\alpha(x)$ for all $x\in Y\setminus \cup_{k\in \Z} I_k$. A well known result says that every Denjoy map $h$ exhibits a unique minimal set $M_h$ which is isomorphic to a Cantor set, and in this case we have $M_h=Y\setminus \cup_{k\in\Z} (a_k, b_k)$.
To meet our needs, we modify the above construction by changing each $I_k$ to a set $A_k=\{a_k+l(I_k)\cdot i/(n-1)\colon i=0,1,\dots, n-1\}$ with cardinality $n$.
Denote the new space (which is a closed subset of $Y$) as $X$ and the homeomorphism is $T=h|_X$. Next we shall prove that $(X,T)$ is essentially positively $n$-expansive with respect to the metric $d|_X$. Let $0<\delta<l(I_0)/2$. We claim that $\Card (\Int(\Phi_\delta[x])\cap X)\le n-2$. Here $\Int(\cdot)$ denotes the interior operation.
To check this, we first show $\Int(\Phi_\delta[x])\cap M_h=\emptyset$ for all $x\in X$. Assume there is $y\in \Int(\Phi_\delta[x])\cap M_h$. Since $M_h$ is minimal there exists $k_s\to \infty$ with $h^{-k_s}(a_0)\to y$. By the construction we know $\{h^{-k}(I_0)\colon k\in\N\}$ is disjoint and $\sum_{k\in \Z} l(I_k)=1$, so $l(h^{-{k_s}}(I_0))\to 0$ as $s\to \infty$. It implies that $h^{-{k_s}}(I_0)\subset \Phi_\delta[x]$. Notice the fact that $h(\Phi_\delta[x])\subset \Phi_\delta[h(x)]$ we have $I_0\subset \Phi_\delta[h^{k_s}(x)]$, which is a contradiction with $0<\delta<l(I_0)/2$.
Now we check $\Card (\Int(\Phi_\delta[x])\cap (X\setminus M_h))\le n-2$. Otherwise there are at least $n-1$ distinct points $z_1,\dots,z_{n-1}\in \Int(\Phi_\delta[x])\cap (X\setminus M_h)$. Since $\Card (I_k\cap (X\setminus M_h))=n-2$ for each $k\in\Z$, without loss of generality we assume $z_1\in I_i, z_2\in I_j$ for some $i\neq j\in\Z$. As $\Phi_\delta[x]$ reduces to closed arc (possibly trivial) and $z_1,z_2\in \Int (\Phi_\delta[x])$, so the arc $[z_1, z_2]$ is contained in $\Phi_\delta[x]$. But there must have some point $z_0\in M_h\cap [z_1, z_2]$, which contradicts $\Int(\Phi_\delta[x])\cap M_h=\emptyset$ for all $x\in X$. Hence $\Card (\Int(\Phi_\delta[x])\cap (X\setminus M_h))\le n-2$ and the claims holds.
Since $\Phi_\delta[x]$ is a closed arc (possibly trivial), we have $\Card (\Phi_\delta[x]\cap M_h)\le 2$ and then $\Card (\Phi_\delta[x]\cap X)\le n$ for all $x\in X$. That is $(X,T)$ is positively $n$-expansive for the above $\delta$. Notice that by (a)(b) we also have $l(h^{k}(I_0))\to 0$ as $k\to \infty$, so for any $\delta>0$, there is $N\in\N$ such that $\diam (h^k(A_0))=\diam (A_k)=l(h^k(I_0))\le \delta$ for all $k\ge N$. This implies that $A_k\subset \Phi_\delta[a_k]\cap X$ and then $(X,T)$ is not positively $(n-1)$-expansive for $\delta$. Combining with $\Card (\Phi_\delta[x]\cap X)\le n$ we have $\Card (\Phi_\delta[a_k]\cap X)=n$, so $(X,T)$ is essentially positively $n$-expansive.
\end{proof}
A deep result in classical terms says that a compact metric space is finite once it carries a positively expansive homeomorphism, and several different proofs can be found in \cite{KR69, MS10} and the references therein. In Theorem \ref{thm:EssPosNExpan} we have shown that this is not true in the positive $n$-expansiveness case, that is there is an infinite t.d.s. $(X,T)$ carrying positively $n$-expansive homeomorphism. But if additionally $X$ is a countable space, we shall prove that the finiteness still holds.
Now we make some preparations. Call a point $x$ of $X$ an \textit{accumulation point} if $x\in \ol{X\setminus x}$. The collection of accumulation points of $X$ is said to be the \textit{derived set} of $X$, write as $X^d$. The derived set of $X$ of order $\alpha$ is recursively defined by the conditions: $X^{(1)}=X^d, \ X^{(\alpha+1)}=(X^{(\alpha)})^d$ and $X^{(\lambda)}=\bigcap_{\alpha<\lambda} X^{(\alpha)}$ if $\lambda$ is a limit ordinal. We denote the \textit{derived degree} of $X$ by $d(X)$ and $d(X)=\alpha$ if $X^{(\alpha)}\neq\emptyset$ and $X^{(\alpha+1)}=\emptyset$.
Note that a compact metric space $X$ is a countable set if and only if $d(X)$ exists and it is a countable ordinal number. Also, it is clear that if $d(X)=\alpha$ then $X^{(\alpha)}$ is a finite set. We leave readers to the book by Kuratowski \cite[p. 261]{Ku68} for more details.
\begin{thm}\label{thm:CountPosNExpan}
Let $(X,T)$ be a t.d.s. with countable infinite cardinality. Then $X$ carries no positively $n$-expansive homeomorphism $T$ for any $n\in\N$.
\end{thm}
\begin{proof}
Since $X$ is countable, we can choose arbitrarily small radii such that the balls below are open and closed. Let $d(X)=\alpha$ with $\alpha$ a countable ordinal and denote $X^{(\alpha)}=\{x_1,\dots,x_n\}$. By Theorem \ref{thm:fact}\eqref{fact:1} each $x_i$ can be assumed to be fixed point. Let $\ep>0$ be such that $B_\ep(x_1),\dots, B_\ep(x_n)$ are pairwise disjoint. Choose $0<\delta<\ep/2$ such that
\begin{equation*}
B_\delta(x_i)\subset B_\ep(x_i)\ \mbox{and} \ T B_\delta(x_i)\subset B_\ep(x_i). \eqno{(*)}
\end{equation*}
Note that each $B_\delta(x_i)$ contains countable infinite elements. Now we consider the following property
\begin{align*}
P(\alpha) \colon & \mbox{ if } d(X)=\alpha, \ \mbox{there is} \ \ 1\le i_\alpha\le n \ \mbox{such that} \\
& Y_{i_\alpha} =\{x\in X\colon \ol{\textrm{orb}^+(x, T)}\subset B_\delta(x_{i_\alpha})\} \ \mbox{is countable infinite.}
\end{align*}
Here $\ol{\textrm{orb}^+(x, T)}=\ol{\{x, T x, \dots\}}$ means the positive orbit closure of $x$ under $T$. Since for any $x,y\in Y_{i_\alpha}$ we have for each $l\in\Z_+$,
$$
d(T^l x,T^l y)<d(T^l x,x_{i_\alpha})+d(x_{i_\alpha},T^l y)<\ep,
$$
so if $P(\alpha)$ holds then $(X,T)$ is not positively $n$-expansive for any $n\in\N$. Next we would follow this idea to check the validity of $P(\alpha)$ by transfinite discussion.
\smallskip
Assume that $\alpha$ is not a limit ordinal, that is $\alpha=\beta+1$ for some ordinal number $\beta$. Then two cases are involved: (i) $X^{(\beta)}$ is pointwise periodic and (ii) there is a non-periodic point $y\in X^{(\beta)}$.
For the case (i), we claim that
$$
Z_\beta=\{x\in X\colon x\in X^{(\beta)} \ \mbox{and}\ \ol{\textrm{orb}^+(x, T)}\not\subset \cup_{i=1}^n B_\delta(x_i)\}
$$
is finite. In fact if the contrary then by $T(X^{(\beta)})\subset X^{(\beta)}$ we have a limit point $x_0\in X\setminus (\cup_{i=1}^n B_\delta(x_i))\cap X^{(\alpha)}=\emptyset$, a contradiction. Hence $B_\delta(x_i)\cap Z_\beta$ is finite for each $1\le i\le n$. On the other hand, by condition $(*)$ we get $Y_i\supset X^{(\beta)}\cap B_\delta(x_i)\setminus Z_\beta$, and note that $X^{(\beta)}\cap B_\delta(x_i)$ is countable infinite then so is $Y_i$ for every $1\le i\le n$. This case ends.
Now consider the case (ii). First we note that $X\setminus (\cup_{i=1}^n B_\delta(x_i))\cap X^{(\beta)}$ is finite. Then combine with the fact $T(X^{(\beta)})\subset X^{(\beta)}$ and the condition that $y\in X^{(\beta)}$ is not periodic, we can declare that there exist $m\in\N$ and $1\le i_\alpha\le n$ such that $T^m y\in Y_{i_\alpha}$, and then $\ol{\textrm{orb}^+(T^m y, T)}\subset Y_{i_\alpha}$. As the cardinality of $\ol{\textrm{orb}^+(T^m y, T)}$ is countable infinite, we have $P(\alpha)$ is true.
\smallskip
Assume that $\alpha$ is a limit ordinal number. Since $\bigcap_{\gamma<\alpha} X^{(\gamma)}=X^{(\alpha)}$, then there exists an ordinal $\gamma_{_0}<\alpha$ such that $X^{(\gamma)}\subset \cup_{i=1}^n B_\delta(x_i)$ for all $\gamma>\gamma_{_0}$. Note that $T(X^{(\gamma)})\subset X^{(\gamma)}$ for any $\gamma<\alpha$.
By the continuity of $T$, there is $\sigma>0$ such that if $z\in X^{(\gamma)}\cap B_\sigma(x_i)$ then $T z\in X^{(\gamma)}\cap B_{\delta}(x_i)$ for all $\gamma>\gamma_{_0}$ and all $i=1,\dots, n$. For this $\sigma>0$, we choose suitable $\gamma_{_0}<\gamma_{_1}<\alpha$ such that $X^{(\gamma_{_1})}\subset \cup_{i=1}^n B_\delta(x_i)$. Hence for each $1\le i\le n$ we have
$$T(X^{(\gamma_{_1})}\cap B_{\delta}(x_i))=T(X^{(\gamma_{_1})}\cap B_{\sigma}(x_i))\subset X^{(\gamma_{_1})}\cap B_{\delta}(x_i).$$
And inductively we see
$$T^k(X^{(\gamma_{_1})}\cap B_{\delta}(x_i))\subset X^{(\gamma_{_1})}\cap B_{\delta}(x_i) \ \ \mbox{for any} \ \ k\in\Z.$$
That is $X^{(\gamma_{_1})}\cap B_{\delta}(x_i)\subset Y_i$ and then $P(\alpha)$ is true. We are done.
\end{proof}
\begin{rem}
We point out that Theorem \ref{thm:CountPosNExpan} actually presents a speical class of essentially positively $\aleph_0$-expansive systems, and in this countable case, essentially positive $\aleph_0$-expansiveness is equivalent to positive non-$n$-expansiveness for any $n\in\N$. But in general they are not the same, for example see Corollary \ref{cor:nonexist-cicle}.
\end{rem}
\section{Levels of expansiveness}\label{sect:LevExpan}
Parallel to the previous section, we naturally ask if there exist examples to distinguish all the levels of expansiveness. It is not hard to check that the example of Theorem \ref{thm:EssPosNExpan} is essentially positively $n$-expansive but fails to be essentially $n$-expansive for any $n\ge 2$. In fact it is expansive (to check $\Card (\Gamma_\delta [x]\cap X)=1$). In \cite{Ar14} Artigue studied $(m,l)$-expansiveness ($m>l\ge 1$) and showed that there are $(4,2)$-expansive homeomorphisms that are not $(3,2)$-expansive (\cite[Proposition 4.3]{Ar14}), but we know that $(4,2)$-expansiveness implies $(2,1)$-expansiveness which is equivalent to expansiveness (\cite[Proposition 1.4]{Ar14}). Besides, in \cite{APV13} the authors worked out an example of a $2$-expansive but not expansive homeomorphism on surface, but the general case is still open. In this section we shall present a complete answer.
\subsection{Essential $n$-expansiveness}
Note that the next characterization is mainly inspired by Kato and Park \cite{KP99}, and for convenience we follow their notations and terms. Let $S=\{s_i\colon i\in\Z\}\cup \{s_\infty\}$ with $\lim_{i\to +\infty}s_i=\lim_{i\to -\infty}s_i=s_\infty$ and $s_i, s_\infty$ be points in the plane with $s_i\neq s_j\ (i\neq j),\ s_i\neq s_\infty$. We define a homeomorphism $g\colon S\to S$ by $g(s_i)=s_{i+1}$ and $g(s_\infty)=s_\infty$.
Let $r\in\N$ and $U(s_{-r}),\dots,U(s_r),U(s_\infty)$ be closed neighborhoods of $s_{-r},\dots,s_r,s_\infty$ in the plane respectively. Require $U(s_i)\cap U(s_i)=\emptyset,\ i\neq j$ and $S\subset \cup_{i=-r}^r U(s_i)\cup U(s_\infty)$. Set $V(=V_r)=\cup_{i=-r}^r U(s_i)\cup U(s_\infty)$ and call $V$ as a \textit{neighborhood system} of $S$. Assume a sequence $\{t_i\colon i\in\Z\}\subset V$ with $\lim_{i\to +\infty}t_i=\lim_{i\to -\infty}t_i=s_\infty$. Fix $d\in\Z_+$. We say $\{t_i\colon i\in\Z\}$ \textit{winds d-times around $S$} (with respect to $V$), if there is $k\in\Z_+$ satisfying that
\begin{enumerate}
\item[(a)] $k>2r$,
\item[(b)] $t_{jk+i}\in U(s_{-r+i})$ for all $0\le i\le 2r$, $1\le j\le d$, and
\item[(c)] $t_m\in U(s_\infty)$ for other $t_m$.
\end{enumerate}
For brevity we denote $w_S(\{t_i\colon i\in\Z\}; V)=d$. Here we remark that the conditions (a)-(c) have a small difference with the original ones in \cite{KP99}, but they are still available for some well chosen sequence $\{t_i\colon i\in\Z\}$. The intention of this change will be revealed when consider the van der Waerden depth (see \cite{KLOX15}) later.
Consider a countable t.d.s. $(X, T)$ with $S\subset X\subset V$ and $T\colon X\to X$ an extension of $g$. Let $d\in\Z_+$. We say a point $x\in X$ has \textit{winding number of d} if $\textrm{orb}(x,T)$ winds $d$-times around $S$ (with respect to $V$), and denote $w_S(x; T, V)=d$. It is clear that $w_S(s_i; T, V)=1$ for $s_i \in S\ (i\neq \infty)$ and $w_S(s_\infty; T, V)=0$. Let $Y$ be a subset of $X$, we write $w_S(Y;T,V)=\{w_S(x; T, V)\colon x\in Y\}$ and call $w_S(Y;T,V)$ is the \textit{winding set} of $Y$ around $S$.
\begin{thm}\label{thm:EssNExpan}
There exists an essentially $n$-expansive t.d.s. $(X,T)$ for every $n\in\N$.
\end{thm}
\begin{proof}
Note that the above t.d.s. $(S,g)$ is expansive. Now let $n\ge 2$. Assume $M$ is any infinite subset of prime numbers and $V$ is any neighborhood system of $S$. We shall construct a countable t.d.s. $(X,T)$ satisfying the following conditions:
\begin{enumerate}
\item[(a)] $S\subset X\subset V$, $d(X)=2$ and $X^{(2)}=\{s_\infty\}$;
\item[(b)] $T\colon X\to X$ is a homeomorphism and $T|_S=g\colon S\to S$;
\item[(c)] for any $x\in X\setminus \{s_\infty\}$, $w_S(x; T,V)\neq 0$ and $w_S(X\setminus S; T,V)\subset M$;
\item[(d)] for each pair $x_1,x_2\in X\setminus \{s_\infty\}$, we have either $\orb (x_1,T)=\orb (x_2,T)$ or $\orb (x_1,T)\cap \orb (x_2,T)=\emptyset$. Moreover,
\item[ ] \begin{itemize}
\item[(d1)] if $\orb (x_1,T)=\orb (x_2,T)$ then $w_S(x_1; T, V)= w_S(x_2;T, V)$, and
\item[(d2)] if $\orb (x_1,T)\cap \orb (x_2,T)=\emptyset$, then either $w_S(x_1; T, V)\neq w_S(x_2; T,V)$, or there are $n$, and only $n$, pairwise disjoint orbits, say $\orb (x_i,T)$, $i=1,\dots,n$, such that $w_S(x_i; T, V)= w_S(x_j; T,V)$ for every $1\le i,\ j\le n$;
\end{itemize}
\item[(e)] for any $\delta>0$, there exist $y_1,\dots, y_n\in X$ such that $\Gamma_\delta[y_i]=\{y_1,\dots,y_n\}$ for any $1\le j\le n$.
\end{enumerate}
It is easy to see that provided with the conditions above, $(X,T)$ is essentially $n$-expansive for any $n\ge 2$.
Now we give the construction. Denote $M=\{p_1,p_2,\dots\}$ with $p_i\neq p_j$ for any $i\neq j$. Choose a descending neighborhood systems family $V=V_1,\ V_2\dots$ of $S$ with $\bigcap_{i=1}^{+\infty} V_i=S$. For every $i\in \N$, add $n-1$ suitable decreasing neighborhood systems $W_1^i,\dots,W_{n-1}^i$ such that
$V_{i}\supset W_1^i\supset \dots\supset W_{n-1}^i\supset V_{i+1}$. Put $V_i=W_0^i,\ V_{i+1}=W_n^i$, then for each $i\in\N$ and $1\le m\le n$ we pick carefully a sequence $\{x_{m,j}^i\colon j\in\Z\}$ consisting of different points in $W_{m-1}^i\setminus W_m^i$, such that $\lim_{j\to +\infty}x_{m,j}^i=\lim_{j\to -\infty}x_{m,j}^i=s_\infty$ and $w_S(\{x_{m,j}^i\colon j\in\Z\}; V_i)=p_i$. Write $V_i=\bigcup_{j=-r_i}^{r_i} U_i(s_j)\bigcup U_i(s_\infty) $ and further we assume $\{x_{m,j}^i\colon m=1,\dots,n\}\subset U_i(s_{i_j})$ for each $i\in\N$ and $j\in\Z$.
Renumber all the sequences $\{x_{m,j}^i\colon j\in\Z\}$ well-ordered as $Y_1,Y_2,\dots$. We can see that each $Y_i$ is homeomorphic to $S$, $Y_i\cap S=Y_i\cap Y_j=\{s_\infty\}$ for $i\neq j$ and
$\lim_{i\to +\infty}H_d(Y_i,S)=0$. Take $X=\bigcup_{i\in\N} Y_i\cup S$ and define $T\colon X\to X$ by $T|_S=g\colon S\to S$, $T(x_{m,j}^i)=x_{m,j+1}^i$. Then $T$ is a homeomorphism and $d(X)=2$, $X^{(2)}=\{s_\infty\}$. It is not hard to see that $(X,T)$ satisfies all the conditions and so it is the desired system.
\end{proof}
In \cite{KP99} Kato and Park showed that $X$ admits an expansive homeomorphism if and only if its derived degree is not a limit ordinal number. We can improve this deep result to $n$-expansiveness case.
The following Lemma can be found in \cite[Proposition 2.1]{KP99}.
\begin{lem}\label{lem:Count}
Let $X,\ Y$ be two countable metric spaces with $d(X)=d(Y)=\alpha$. If $X^{(\alpha)}$ and $Y^{(\alpha)}$ are homeomorphic, then so are $X$ and $Y$.
\end{lem}
\begin{thm}\label{thm:NExpanChar}
Let $(X,T)$ be a t.d.s. with $d(X)=\alpha\ge 2$ and $n\in\N$. Then $X$ admits an essentially $n$-expansive homeomorphism if and only if $\alpha$ is not a limit ordinal number.
\end{thm}
\begin{proof}
Note that at the end part of the proof in Theorem \ref{thm:CountPosNExpan}, we in fact showed that if $\alpha$ is a limit ordinal number then $X$ admits no $n$-expansive homeomorphism for any $n\in\N$. So it remains to show the sufficiency.
Assume $\alpha\ge 2$ is not a limit ordinal. As $d(X)=\alpha$ we have $X^{(\alpha)}=\{a_1,\dots,a_n\}$. Choose suitable closed subsets $Z_1,\dots,Z_n$ of $X$ such that $X=\bigcup_{i=1}^n Z_i$, $Z_i\cap Z_j=\emptyset$ and $d(Z_i)=\{a_i\}$. Next we shall inductively construct an essentially $n$-expansive t.d.s. $(X_\alpha,T_\alpha)$ in the plane with $d(X_\alpha)=\alpha$ and $X^{(\alpha)}=\{s_\infty\}$. If this were done, then by Lemma \ref{lem:Count} and Remark \ref{rem:DefExpa}\eqref{rem:DefExpa:1} there exists an essentially $n$-expansive homeomorphism on every $Z_i$. It implies that we can find an essentially $n$-expansive homeomorphism on $X$ and the proof ends.
Now we construct the desired t.d.s. $(X_\alpha,T_\alpha)$, and similar as Theorem \ref{thm:EssNExpan} it suffices to fulfil the below property $P(\alpha)$:
\begin{enumerate}[ ]
\item Let $M$ be an infinite subset of prime numbers and $V$ be a neighborhood system of $S$ in the plane. Then we can construct a countable t.d.s. $(X_\alpha,T_\alpha)$ such that:
\item[ ]
\begin{enumerate}
\item[(a)] $S\subset X_\alpha\subset V$, $d(X)=\alpha$ and $X_\alpha^{(\alpha)}=\{s_\infty\}$;
\item[(b)] $T_\alpha\colon X_\alpha\to X_\alpha$ is a homeomorphism and $T_\alpha|_S=g\colon S\to S$;
\item[(c)] for any $x\in X_\alpha\setminus \{s_\infty\}$, $w_S(x; T_\alpha,V)\neq 0$ and $w_S(X_\alpha\setminus S;
T_\alpha,V)\subset M'$, where $M'=\{q_1\cdots q_t\colon q_i\in M, i=1,\dots,t, t\in\N\}$;
\item[(d)] for each pair $x_1,x_2\in X_\alpha\setminus \{s_\infty\}$, we have either $\orb (x_1,T_\alpha)=\orb (x_2,T_\alpha)$ or $\orb (x_1,T_\alpha)\cap \orb (x_2,T_\alpha)=\emptyset$. Moreover,
\item[ ] \begin{itemize}
\item[(d1)] if $\orb (x_1,T_\alpha)=\orb (x_2,T_\alpha)$ then $w_S(x_1; T_\alpha, V)= w_S(x_2;T_\alpha, V)$, and
\item[(d2)] if $\orb (x_1,T_\alpha)\cap \orb (x_2,T_\alpha)=\emptyset$, either $w_S(x_1; T_\alpha, V)\neq w_S(x_2; T_\alpha,V)$, or there are $n$, and only $n$, pairwise disjoint orbits, say $\orb (x_i,T_\alpha)$, $i=1,\dots,n$, such that $w_S(x_i; T_\alpha, V)= w_S(x_j; T_\alpha,V)$ for every $1\le i, \ j\le n$;
\end{itemize}
\item[(e)] for any $\delta>0$, there exist $y_1,\dots, y_n\in X$ such that $\Gamma_\delta[y_j]=\{y_1,\dots,y_n\}$ for any $1\le j\le n$.
\end{enumerate}
\end{enumerate}
It turns out that $P(2)$ is just the Theorem \ref{thm:EssNExpan}.
Now assume that $\alpha=\beta+1$. We consider two cases: (i) $\beta$ is a non-limit ordinal number; (ii) $\beta$ is a limit ordinal number.
For the case (i), by induction we assume $P(\beta)$ holds and the aim is to prove $P(\alpha)$ holds too. Let $M$ and $V$ be given as described in $P(\alpha)$. Divide $M=M_0\cup M_1\cup\cdots$ with $M_i\cap M_j=\emptyset$, $i\neq j$ and $\Card (M_i)=\infty$. Also, we list $M_0$ as $\{p_1,p_2,\dots\}$. Next select a descending family of neighborhood systems of $S$, say $V=V_1\supset V_2\supset\cdots$, such that $\bigcap_{i=1}^{+\infty}V_i=S$.
For each $i\in \N$, by induction we can construct a countable t.d.s. $(X_\beta^i, T_\beta^i)$ satisfying that:
\begin{enumerate}
\item[(a)] $S\subset X_\beta^i \subset V'_i\subset V_i$, $d(X_\beta^i)=\beta$ and $(X_\beta^i)^{(\beta)}=\{s_\infty\}$, where the well-chosen $V'_i$ is a sufficiently small neighborhood system and can be asked to meet the constraints related to $V_i$ and $p_i$ later;
\item[(b)] $T_\beta^i\colon X_\beta^i\to X_\beta^i$ is a homeomorphism and $T_\beta^i|_S=g\colon S\to S$;
\item[(c)] for any $x^i\in X_\beta^i\setminus \{s_\infty\}$, $w_S(x^i; T_\beta^i,V'_i)\neq 0$ and $w_S(X_\beta^i\setminus S;
T_\beta^i,V'_i)\subset M_i^*$;
\item[(d)] for each pair $x_1^i,x_2^i\in X_\beta^i\setminus \{s_\infty\}$, we have either $\orb (x_1^i,T_\beta^i)=\orb (x_2^i,T_\beta^i)$ or $\orb (x_1^i,T_\beta^i)\cap \orb (x_2^i,T_\beta^i)=\emptyset$. Moreover,
\item[ ] \begin{itemize}
\item[(d1)] if $\orb (x_1^i,T_\beta^i)=\orb (x_2^i,T_\beta^i)$ then $w_S(x_1^i; T_\beta^i, V'_i)= w_S(x_2^i;T_\beta^i, V'_i)$, and
\item[(d2)] $\orb (x_1^i,T_\beta^i)\cap \orb (x_2^i,T_\beta^i)=\emptyset$ then either $w_S(x_1^i; T_\beta^i, V'_i)\neq w_S(x_2^i; T_\beta^i,V'_i)$, or there are $n$, and only $n$, pairwise disjoint orbits, say $\orb (x_j^i,T_\beta^i)$, $j=1,\dots,n$, such that $w_S(x_j^i; T_\beta^i, V_i')= w_S(x_k^i; T_\beta^i, V_i')$ for every $1\le j,\ k\le n$;
\end{itemize}
\item[(e)] for any $\delta>0$, there exist $y_1^i,\dots, y_n^i\in X_\beta^i$ such that $\Gamma_\delta[y_j^i]=\{y_1^i,\dots,y_n^i\}$ for any $1\le j\le n$.
\end{enumerate}
Note that the sequence $\{s_j\colon j\in\Z\}\subset X_{\beta}^i$ and $w_S(s_j;T_{\beta}^i, V'_i)=1$ for each $i\in\N$ and $j\neq\infty$. To ensure the $n$-expansiveness, for each $i\in \N$ by well choosing the above $V'_i$ we can find a continuous embedding $\psi^i\colon X_\beta^i\to V_i$, such that $\psi^i(s_\infty)=s_\infty$, $w_S(\{\psi^i(s_j)\colon j\in
\Z\}; V_i)=p_i$ and $w_S(\{\psi^i(x^i_j)\colon j\in\Z\}; V_i)\in p_i\cdot M_i^*$ for any point $x^i_j\in X_\beta^i\setminus S$. Set $Z_\beta^i=\psi^i(X_\beta^i)$ and we may further assume $Z_\beta^i\cap Z_\beta^j=Z_\beta^i\cap S=\{s_\infty\}$ for any $i\neq j\in \Z$. Take $X_\alpha=\bigcup_{i=1}^{+\infty}Z_\beta^i\cup S$ and define $T_\alpha\colon X_\alpha\to X_\alpha$ by $T_\alpha|_S=g$ and $T_\alpha|_{Z_\beta^i}=\psi^i\circ T_\beta^i \circ (\psi^i)^{-1}$. It is easy to check that $(X_\alpha,T_\alpha)$ meets the conditions of $P(\alpha)$ and this case ends.
As for the case (ii), since $P(\beta)$ is not true as the necessity showed, then we assume $P(\gamma)$ is true for any $\gamma<\beta$. We attend to show $P(\alpha)$ is true too. The proof of this case is similar to the above one, and for the sake of completeness we present the details. Divide $M=M_0\cup M_1\cup\cdots$ with $M_i\cap M_j=\emptyset$, $i\neq j$ and $\Card (M_i)=\infty$, where $M_0=\{p_1,p_2,\dots\}$. Pick a well-ordered sequence of non-limit ordinals $\gamma_{_1}<\gamma_{_2}<\dots$ such that $\lim_{i\to +\infty}\gamma_{_i}=\beta$. Also, pick a descending family of neighborhood systems of $S$, say $V=V_1\supset V_2\supset\cdots $, such that $\bigcap_{i=1}^{+\infty}V_i=S$.
For each $i\in \N$, by induction we can construct a countable t.d.s. $(X_{\gamma_{_i}}, T_{\gamma_{_i}})$ satisfying that:
\begin{enumerate}
\item[(a)] $S\subset X_{\gamma_{_i}} \subset V'_i \subset V_i$, $d(X_{\gamma_{_i}})=\gamma_{_i}$ and $(X_{\gamma_{_i}})^{({\gamma_{_i}})}=\{s_\infty\}$ ($V'_i$ is similar as above);
\item[(b)] $T_{\gamma_{_i}}\colon X_{\gamma_{_i}}\to X_{\gamma_{_i}}$ is a homeomorphism and $T_{\gamma_{_i}}|_S=g\colon S\to S$;
\item[(c)] for any $x^i\in X_{\gamma_{_i}}\setminus \{s_\infty\}$, $w_S(x^i; T_{\gamma_{_i}},V'_i)\neq 0$ and $w_S(X_{\gamma_{_i}}\setminus S;
T_{\gamma_{_i}},V'_i)\subset M_i'$;
\item[(d)] for each pair $x_1^i,x_2^i\in X_{\gamma_{_i}}\setminus \{s_\infty\}$, we have either $\orb (x_1^i,T_{\gamma_{_i}})=\orb (x_2^i,T_{\gamma_{_i}})$ or $\orb (x_1^i,T_{\gamma_{_i}})\cap \orb (x_2^i,T_{\gamma_{_i}})=\emptyset$. Moreover,
\item[ ] \begin{itemize}
\item[(d1)] if $\orb (x_1^i,T_{\gamma_{_i}})=\orb (x_2^i,T_{\gamma_{_i}})$ then $w_S(x_1^i; T_{\gamma_{_i}}, V'_i)= w_S(x_2^i;T_{\gamma_{_i}}, V'_i)$, and
\item[(d2)] $\orb (x_1^i,T_{\gamma_{_i}})\cap \orb (x_2^i,T_{\gamma_{_i}})=\emptyset$ then either $w_S(x_1^i; T _{\gamma_{_i}}, V'_i)\neq w_S(x_2^i; T_{\gamma_{_i}},V'_i)$, or there are $n$, and only $n$, pairwise disjoint orbits, say $\orb (x_j^i,T_{\gamma_{_i}})$, $j=1,\dots,n$, such that $w_S(x_j^i; T_{\gamma_{_i}}, V_i')= w_S(x_k^i; T_{\gamma_{_i}}, V_i')$ for every $1\le j,\ k\le n$;
\end{itemize}
\item[(e)] for any $\delta>0$, there exist $y_1^i,\dots, y_n^i\in X_{\gamma_{_i}}$ such that $\Gamma_\delta[y_j^i]=\{y_1^i,\dots,y_n^i\}$ for any $1\le j\le n$.
\end{enumerate}
For each $i\in \N$, we choose a continuous embedding $\phi^i\colon X_{\gamma_{_i}}\to V_i$, such that $\phi^i(s_\infty)=s_\infty$, $w_S(\{\phi^i(s_j)\colon j\in
\Z\}; V_i)=p_i$ and $w_S(\{\phi^i(x^i_j)\colon j\in\Z\}; V_i)\in p_i\cdot M_i^*$ for any point $x^i_j\in X_{\gamma_{_i}}\setminus S$. Set $Z_{\gamma_{_i}}=\phi^i(X_{\gamma_{_i}})$ and further assume $Z_{\gamma_{_i}}\cap Z_{\gamma_{_i}}=Z_{\gamma_{_i}}\cap S=\{s_\infty\}$ for any $i\neq j\in \Z$. Take $X_\alpha=\bigcup_{i=1}^{+\infty}Z_{\gamma_{_i}}\cup S$ and define $T_\alpha\colon X_\alpha\to X_\alpha$ by $T_\alpha|_S=g$ and $T_\alpha|_{Z_{\gamma_{_i}}}=\phi^i\circ T_{\gamma_{_i}} \circ (\phi^i)^{-1}$. It is easy to check that $(X_\alpha,T_\alpha)$ fulfils the conditions of $P(\alpha)$ and we complete the whole proof.
\end{proof}
Denote $\mathcal{H}(X)$ all the homeomorphism on $X$. It is a metrizable space with metric defined by
$$
D(T_1,T_2)=\max \{d(T_1 (x), T_2 (x))\colon x\in X\}
$$
for any $T_1,T_2\in \mathcal{H}(X)$. For every $n\in
\N$ put
$$
\mathcal{E}_n(X)=\{T\in \mathcal{H}(X)\colon T \ \mbox{ is an } n \mbox{-expanive homeomorphism}\}.
$$
Then we have a direct corollary of Theorem \ref{thm:NExpanChar}.
\begin{cor}
Let $(X,T)$ be a countable t.d.s. with $d(X)=\alpha$ and $n\in\N$. If $\alpha$ is a non-limit ordinal number and $\alpha>\aleph_0$, then $\ol{\mathcal{E}_n(X)} \neq\mathcal{H}(X)$.
\end{cor}
\begin{proof}
Choose $x\in X^{(\aleph_0)}\setminus X^{(\aleph_0+1)}$ then there is an open and closed neighborhood $U$ of $x$ with derived degree $d(U)=\aleph_0$. Let $n\in\N$ and $0<\ep<d(X\setminus U,U)=\min \{d(y,z)\colon y\in X\setminus U, z\in U)\}$. If there is a homeomorphism $T$ such that $D(T,\id)<\ep$, then $T(U)=U$. That is $U$ is closed and $T$-invariant, then $(U, T|_U)$ forms a subsystem and $d(U)=\aleph_0$. By Theorem \ref{thm:NExpanChar} we know that $T|_U$ can not be $n$-expansive for any $n\in\N$, and then neither is $T$ by Remark \ref{rem:DefExpa}\eqref{rem:DefExpa:3}.
\end{proof}
\subsection{Van der Waerden depth}
Another consequence of Theorem \ref{thm:NExpanChar} is related to the notion of Van der Waerden depth, and to describe this we first recall some notions.
Let $(X,T)$ be a t.d.s. and $x\in X$. We say $x$ is a \textit{non-wandering point} if for each neighborhood $U$ of $x$ there is $n\in\N$ with $U\cap T^{-n}U\neq\emptyset$. Write $\Omega(X,T)$ the collection of all non-wandering points. It is well known that $(\Omega(X,T),T)$ forms a subsystem of $(X,T)$, and we can consider $\Omega(\Omega(X,T),T)$ in a natural way. Note that there exists a system $(X,T)$ such that $\Omega(\Omega(X,T),T)\neq \Omega(X,T)$ (see \cite{KP99} for example). More generally, by induction we set $\Omega_0(X,T)=X$, $ \Omega_1(X,T)=\Omega(\Omega_0(X,T),T)$, $\Omega_{\alpha+1}(X,T)=\Omega(\Omega_\alpha(X,T),T)$ and
$\Omega_{\lambda}(X,T)=\bigcap_{\alpha<\lambda} \Omega_\alpha(X,T)$ if $\lambda$ is a limit ordinal number.
A well known conclusion says that descending family of closed subsets in compact metric space is always at most countable, so there exists a countable ordinal $\alpha$
satisfying $\Omega_\alpha(X,T)=\Omega_{\alpha+1}(X,T)$. Denote the \textit{depth} of $(X,T)$ as
$$\textrm{depth}(X,T)=\min \{\alpha\colon \Omega_\alpha(X,T)=\Omega_{\alpha+1}(X,T)\}$$
and call $\Omega_\alpha(X,T)$ as the \textit{Birkhoff center} of $(X,T)$. It is well known that there exists a t.d.s. $(X,T)$ with $\textrm{depth}(X,T)=\alpha$ when $\alpha$ is a countable ordinal (see for instance \cite{Ne78} and \cite[Corollary 2.7]{KP99}).
Similar as above, the authors in \cite{KLOX15} introduced multi-non-wandering points and the van der Waerden center.
We say $x$ is \textit{multi-non-wandering} if for each neighborhood $U$ of $x$ and each $d\in\N$, there
is $k\in\N$ such that
$$U\cap T^{-k}U\cap T^{-2k}U\cap\dots \cap T^{-dk}U\neq\emptyset.$$
Denote by $\Omega^{(\infty)}(X,T)$ the set of all multi-non-wandering points. It is easy to see that $(\Omega^{(\infty)}(X,T),T)$ can form a subsystem of $(\Omega(X,T),T)$. Note that there also exists a system $(X,T)$ such that $\Omega^{(\infty)}(\Omega^{(\infty)}(X,T),T)\neq \Omega^{(\infty)}(X,T)$ (see \cite[Example 6.7]{KLOX15}). Inductively we put $\Omega_0^{(\infty)}(X,T)=X$, $ \Omega_1^{(\infty)}(X,T)=\Omega^{(\infty)}(\Omega_0^{(\infty)}(X,T),T)$, $\Omega_{\alpha+1}^{(\infty)}(X,T)=\Omega^{(\infty)}(\Omega_\alpha^{(\infty)}(X,T),T)$ and $\Omega_{\lambda}^{(\infty)}(X,T)=\bigcap_{\alpha<\lambda} \Omega_\alpha^{(\infty)}(X,T)$ when $\lambda$ is a limit ordinal number. We call $\Omega_{\lambda}^{(\infty)}(X,T)$ the \textit{van der Waerden center} of $(X,T)$ if $\Omega_{\alpha}^{(\infty)}(X,T)=\Omega_{\alpha+1}^{(\infty)}(X,T)$ and denote by $\textrm{depth}(X,T)$ the \textit{van der Waerden depth} of $(X,T)$ defined as
$$\textrm{depth}^{(\infty)}(X,T)=\min \{\alpha\colon \Omega_\alpha^{(\infty)}(X,T)=\Omega_{\alpha+1}^{(\infty)}(X,T)\}.$$
The same reason as above we know $\textrm{depth}^{(\infty)}(X,T)$ is a countable ordinal number.
Denote $\aleph_1$ the first uncountable ordinal number.
\begin{thm}\label{thm:Van}
Assume $\alpha<\aleph_1$ is a non-limit ordinal number and $n\ge 2$. Then there is a countable t.d.s. $(X_\alpha,T_\alpha)$ such that
\begin{enumerate}
\item $d(X_\alpha)=\emph{depth}^{(\infty)}(X_\alpha,T_\alpha)=\emph{depth}(X_\alpha,T_\alpha)=\alpha$, and
\item $(X_\alpha,T_\alpha)$ is essentially $n$-expansive.
\end{enumerate}
\end{thm}
\begin{proof}
We point out that the construction of Theorem \ref{thm:NExpanChar} is just what we desired. Here we only check the case that $\alpha=2$, and the general case is similar.
Note that $\alpha=2$ is the above Theorem \ref{thm:EssNExpan}. Let $X_2=\bigcup_{i\in\N} Y_i\cup S$ and $T_2\colon X_2\to X_2$, $T_2|_S=g\colon S\to S$, $T_2(x_{m,j}^i)=x_{m,j+1}^i$ be defined as in Theorem \ref{thm:EssNExpan}. It suffices to show
$$
\Omega_1(X_2,T_2)=\Omega_1^{(\infty)}(X_2,T_2)=S=\bigcup_{i\in\Z}\{s_i\}\cup\{s_\infty\}.
$$
To show this, we recall the decreasing family of neighborhood systems $V=V_0\supset\dots\supset V_{i}=W_0^i\supset W_1^i\supset \dots\supset W_{n-1}^i\supset V_{i+1}=W_n^i\supset\cdots$. So for any $x\in X_2\setminus S$, by the construction we have $x\in W_{m-1}^i\setminus W_m^i$ for some $m\in [1,n]$ and $i\in\N$. Let $0<\ep <d(x, S)/2$, where $d(x, S)=\min\{d(x,y)\colon y\in S\}$. Then there is $i_0\in\N$ such that $B_\ep(x)\cap V_i=\emptyset$ for all $i\ge i_0$. Hence
$
B_\ep(x)= B_\ep(x) \bigcap (\bigcup_{i=0}^{i_0} \bigcup_{m=1}^n \{x_{m,j}^i\colon j\in \Z\}).
$
On the other hand, $B_\ep(x)\bigcap \{x_{m,j}^i\colon j\in \Z\}$ is a finite set for each $m\in [1,n]$ and $i\in\N$, otherwise by compactness there is another limit point distinct with $s_\infty$, a contradiction with the assumption that $\lim_{j\to +\infty}x_{m,j}^i=\lim_{j\to -\infty}x_{m,j}^i=s_\infty$. It implies that we can choose small $0<\ep_0<\ep$ such that $B_{\ep_0}(x)=\{x\}$ is an open set, and then $x\notin \Omega_1(X_2,T_2)$. Finally the arbitrariness of $x$ yields $X_2\setminus S\cap \Omega_1(X_2,T_2)=\emptyset$.
Clearly $s_\infty\in \Omega_1^{(\infty)}(X_2,T_2)\subset \Omega_1(X_2,T_2)$. Now consider $s_j\ (j\neq\infty)$. As $\bigcap_{i=1}^{+\infty}V_i=S$ and $V_i=\bigcup_{l=-r_i}^{r_i} U_i(s_l)\bigcup U_i(s_\infty) $, there must exist some $i_1\in\N$ such that $-r_i\le j\le r_i$ and then $s_j\in U_i(s_j)$ for all $i\ge i_1$. By the construction, for each $d\in\N$ and $\delta>0$, there are $x\in X_2\setminus S$ and $i_2\in\N$ such that
the prime number $p_i>d$, $x\in U_i(s_j)\subset B_\delta(s_j)$ and $w_S(x; T_2, V_i)=p_i$ for all $i\ge i_2$. Choose $i\ge \max\{i_1,i_2\}$ and by the definition of winding number, we have some $k\in\Z_+$ and $x'\in X_2\setminus S$ such that $T_2^k x'\in U_i(s_j), \dots,T_2^{dk} x'\in U_i(s_j),T_2^{(d+1)k} x'\in U_i(s_j),\dots,T_2^{p_i\cdot k} x'\in U_i(s_j)$. This implies that
$$
B_\delta(s_j)\cap T_2^{-k} B_\delta(s_j)\cap T_2^{-2k} B_\delta(s_j)\cap\cdots\cap T_2^{-dk} B_\delta(s_j)\neq\emptyset,
$$
and then $s_j\in \Omega_1^{(\infty)}(X_2,T_2)$. As $s_j$ is arbitrary, we have $\Omega_1(X_2,T_2)=\Omega_1^{(\infty)}(X_2,T_2)=S$, which ends the proof.
\end{proof}
In \cite[Remark 6.8]{KLOX15} the authors left a conjecture that there exists a t.d.s. $(X,T)$ such that $\textrm{depth}^{(\infty)}(X,T)=\alpha$. Here we prove that the conjecture is true.
\begin{cor}\label{cor:Van}
Assume $\alpha<\aleph_1$ is an ordinal number. Then there is a countable t.d.s. $(X_\alpha,T_\alpha)$ such that
$$d(X_\alpha)=\emph{depth}^{(\infty)}(X_\alpha,T_\alpha)=\emph{depth}(X_\alpha,T_\alpha)=\alpha.$$
\end{cor}
\begin{proof}
By Theorem \ref{thm:Van} it remains to show the case that $\alpha$ is a limit ordinal number. Choose a sequence of non-limit ordinals $\alpha_1<\alpha_2<\dots<\alpha$ with $\lim_{i\to +\infty}\alpha_i=\alpha$. For each $i\in\N$, by Theorem \ref{thm:Van} there exists a countable t.d.s. $(X_{\alpha_i},T_{\alpha_i})$ in the plane such that $d(X_{\alpha_i})=\textrm{depth}^{(\infty)}(X_{\alpha_i},T_{\alpha_i})=\textrm{depth}(X_{\alpha_i},T_{\alpha_i})={\alpha_i}$ and
$\Omega_{\alpha_i}(X_{\alpha_i},T_{\alpha_i})=\Omega_{\alpha_i}^{(\infty)}(X_{\alpha_i},T_{\alpha_i})$. We can further require that the sequence $\{X_{\alpha_i} \}_{i=1}^{+\infty}$ are pairwise disjoint and $\lim_{i\to +\infty}H_d(X_{\alpha_i}, \{x_0\})=0$ for some point $x_0$ in the plane with $x_0\notin X_{\alpha_i}$. Set $X_\alpha=\bigcup_{i=1}^{+\infty}X_{\alpha_i} \cup \{x_0\}$ and $T_\alpha\colon X_\alpha\to X_\alpha$ as $T_\alpha|_{X_{\alpha_i}}=T_{\alpha_i},\ T_\alpha(x_0)=x_0$. It is easy to see that $(X_\alpha, T_\alpha)$ is what we want.
\end{proof}
\subsection{Essential $\aleph_0$-expansiveness} To end this paper, we give an example of essentially $\aleph_0$-expansive homeomorphism.
\begin{exam}\label{exam:CoutExpan}
There exists a t.d.s. $(X,T)$ such that $T$ is essentially $\aleph_0$-expansive.
\smallskip
Let $X=\{0\} \cup \left\{\frac{1}{n}: n \in \mathbb{N} \right\}$ with the subspace topology of the real line $\mathbb{R}$. Define $T: X \rightarrow X$ as
\begin{itemize}
\item $T(0)=0$ and $T(1)=1$;
\item $T\left(\frac{1}{2^n}\right)=\frac{1}{2^n+1}, \ldots, T\left(\frac{1}{2^{n+1}-1}\right)=\frac{1}{2^n}$ for each $n \in \mathbb{N}$.
\end{itemize}
Note that for any $\delta>0$ the set $\{y\in X\colon d(T^n 0, T^n y)\le\delta, \forall n\in \Z \}$ is countable infinite. That is what we need.
\end{exam}
\section*{Acknowledgments}
We thank Siming Tu for very useful suggestions.
J. Li was supported by NNSF of China (11371339) and R. Zhang
was supported by NNSF of China (11001071,11171320).
|
\section*{\refname}%
\begin{document}
\twocolumn[
\title{Goos-H\"anchen and Imbert-Fedorov shifts for Gaussian beams impinging on graphene-coated surfaces}
\author{Simon Grosche, Marco Ornigotti and Alexander Szameit}
\address{
Institute of Applied Physics, Friedrich-Schiller-Universit\"at Jena, \\ Max-Wien-Platz 1, D-07743 Jena, Germany \\
}
\begin{abstract}
We present a theoretical study of the Goos-H\"anchen and Imbert-Fedorov shifts for a fundamental Gaussian beam impinging on a surface coated with a single layer of graphene. We show that the graphene surface conductibility $\sigma(\omega)$ is responsible for the appearance of a giant and negative spatial Goos-H\"anchen shift.
\end{abstract}
\ocis{240.3695, 260.2110, }
]
When an optical beam impinges upon a surface, nonspecular reflection phenomena may occur, such as the Goos-H\"anchen (GH) \cite{refGH1,refGH2,refGH3} and Imbert-Fedorov (IF) \cite{refIF1,refIF2} shifts, resulting in an effective beam shift at the interface. A comprehensive review on beam shift phenomena can be found in Ref. \cite{res9}. Although Goos and H\"anchen published their work more than 60 years ago \cite{refGH1}, this field of research is still very active, and in the last decades a vast amount of literature has been produced on the subject, resulting not only in a better understanding of the underlying physical principles \cite{refIF3,refIF4,refIF5,refIF6,refIF7,refIF8,refIF10,res10}, but also in the careful investigation of the effects of various field configurations \cite{beam1,beam2,beam3,beam4,beam5} and reflecting surfaces \cite{surf1,surf2,surf3,surf4} on the GH and IF shifts.
In recent years, on a parallel trail, graphene attracted very rapidly a lot of interest, thanks to its intriguing properties \cite{ref1,ref2}. Its peculiar band structure and the existence of the so-called Dirac cones \cite{ref3}, for example, give the possibility to use graphene as a model to observe QED-like effects such as Klein tunneling \cite{ref5}, Zitterbewegung \cite{ref6}, the anomalous quantum Hall effect \cite{ref8} and the appearance of a minimal conductivity that approaches the quantum limit $e^2/\hbar$ for vanishing charge density \cite{ref9}. In addition, the reflectance and transmittance of graphene are determined by the fine structure constant \cite{ref10}, and a single layer of graphene shows universal absorbance in the spectral range from near-infrared to the visible part of the spectrum \cite{ref11}.
Among the vast plethora of applications, graphene also proved to be a very interesting system where to observe beam shifts. Very recently, in fact, the occurrence of GH shift in graphene-based structures has been reported, both for light beams (where giant GH shift has been observed \cite{ref13}) and for Dirac fermions \cite{ref14}. Despite all this, a full theoretical analysis of GH and IF shifts in a graphene-based structure has not been yet carried out.
In this Letter, we therefore present a theoretical analysis of the GH and IF shifts occurring for a monochromatic Gaussian beam impinging onto a glass surface coated with a single layer of graphene. The results of our investigations show on one hand, that the appearance of a giant GH shift is ultimately due to the graphene's surface conductivity $\sigma(\omega)$, and on the other hand, that the presence of the single layer of graphene introduces a dependence of the phases of the reflection coefficients on the incidence angle, thus resulting in a nonzero spatial GH shift also when total internal reflection does not occur.
We start our analysis by considering a monochromatic Gaussian beam with frequency $\omega=ck$ (with $k$ being the vacuum wave number), impinging on a dielectric surface characterized by the refractive index $n$ and coated with a single layer of graphene [Fig. \ref{fig1} (a)]. The graphene layer is characterized by the optical conductivity $\sigma(\omega)$, whose expression can be given in the following dimensionless form \cite{ref3}
\begin{equation}\label{sigma}
\sigma(\Omega)= \mathrm{i}\frac{4 \alpha}{\Omega}+ \pi \alpha \left[\Theta(\Omega-2)+\frac{\mathrm{i}}{2\pi} \ln \frac{(\Omega-2)^2}{(\Omega+2)^2}\right],
\end{equation}
where $\Omega = \hbar \omega/\mu$ is the dimensionless frequency, $\mu$ is the chemical potential, $\alpha\approx1/137$ is the fine structure constant \cite{ref6} and $\Theta(x)$ is the Heaviside step function \cite{nist}.
According to Fig.\,\ref{fig1}(b), we define three Cartesian reference frames: the laboratory frame $K=(O,x,y,z)$ attached to the reflecting surface, the (local) incident frame $K_i=(O,x_i,y,z_i)$ attached to the incident beam, and the (local) reflected frame $K_r=(O,x_r,y,z_r)$ attached to the reflected beam. These three reference frames are connected via a rotation of an angle $\theta$ around the $y$ direction \cite{PRA_Aiello}. The reflecting surface is located at $z=0$ , with the $z$-axis pointing towards the interface. With this choice of geometry, the incident beam comes from the region $z<0$ and propagates in the $x$-$z$ plane.
%
\begin{figure}[t!]
\centering
\includegraphics[width=0.5\textwidth]{Interface_AND_skizze.png}
\caption{(Color online) (a) Schematic representation of the considered surface. The graphene layer (red) is characterized by its surface conductivity $\sigma(\Omega)$, whose explicit expression is given by Eq. \eqref{sigma}. The dielectric substrate (green) is characterized by the refractive index $n$. (b) Geometry of beam reflection at the interface. The single graphene layer is located on the surface at $z=0$. The different Cartesian coordinate systems $K,K_i,K_r$ are shown.}
\label{fig1}
\end{figure}
The electric field in the incident frame can be then written, using its angular spectrum representation \cite{mandelWolf}, as follows:
\begin{equation}
\vett{E}_i(\vett{r})=\sum_{\lambda=1}^2\int d^2K\, \uvett{e}_{\lambda}(U,V,\theta) A_{\lambda}(U,V,\theta)e^{i\vett{k}_i\cdot\vett{r}_i},
\end{equation}
where $d^2K=dUdV$, $\uvett{e}_{\lambda}(U,V,\theta)$ is the local reference frame attached to the incident field \cite{brewster}, $A_{\lambda}(U,V,\theta)=\alpha_{\lambda}(U,V,\theta)A(U,V)$ and $\vett{k}_i\cdot\vett{r}_i=UX_i+VY_i+WZ_i$,
being $X_i=k_0x_i$ the normalized coordinate in the incident frame. $Y_i$ and $Z_i$ are defined in a similar manner. $\alpha_{\lambda}(U,V,\theta)=\uvett{e}_{\lambda}(U,V,\theta)\cdot\uvett{f}$ accounts for the projection of the beam's polarization $\uvett{f}=f_p\uvett{x}+f_s \uvett{y}$ (normalized according to $|f_p|^2+|f_s |^2=1$) onto the local basis, and $A(U,V)$ is the beam's spectral amplitude, which here is assumed to be Gaussian, i.e.,
\begin{equation}
A(U,V)=e^{-w_0^2(U^2+V^2)},
\end{equation}
being $w_0^2$ the spot size of the beam. In the remaining of the manuscript, we will consider only well collimated beams, namely the paraxial assumption $U,V\ll 1$ is implicitly understood.
Upon reflection, the electric field can be then written as follows:
\begin{equation}
\vett{E}_r(\vett{r}_r)=\sum_{\lambda =1}^2 \int d^2K\, \hat{\mathbf{e}}_\lambda (-U,V,\pi-\theta) \tilde{A}_\lambda(U,V,\theta) \mathrm{e}^{\mathrm{i}\vett{k}_r \cdot \vett{r}_r} ,
\label{eq:Er}
\end{equation}
where $\vett{k}_r \cdot \vett{r}_r=-UX_r+VY_r+WZ_r$ and
$\tilde{A}_\lambda(U,V,\theta)= r_\lambda(U,V,\theta) A_\lambda(U,V,\theta)$, with $r_{\lambda}(U,V,\theta)$ being the Fresnel reflection coefficients associated to the single plane wave component of the field \cite{BornWolf}. The minus sign in front of $U$ in
$\hat{\mathbf{e}}_{\lambda}$, as well as in $\vett{k}_r \cdot \vett{r}_r$ accounts for the specular reflection of the single plane wave component \cite{brewster}.
The presence of a single layer of graphene deposited on the dielectric surface modifies its reflection coefficients as follows \cite{transferMatrix}:
\begin{subequations}
\label{reflection}
\begin{align}
r_s(\theta)&=\frac{\cos\theta-\sqrt{n^2-\sin^2\theta}-\sigma(\Omega)}{\cos\theta+\sqrt{n^2-\sin^2\theta}+\sigma(\Omega)}
,\\
r_p(\theta)&=\frac{n^2\cos\theta-\sqrt{n^2-\sin^2\theta}\left[1-\sigma(\Omega)\cos\theta\right]}{n^2\cos\theta+\sqrt{n^2-\sin^2\theta}\left[1+\sigma(\Omega)\cos\theta\right] } ,
\end{align}
\end{subequations}
where $\theta$ is the incident angle, $n$ is the refractive index of the dielectric medium and $\sigma(\Omega)$ is the graphene's surface conductivity, as defined by Eq. \eqref{sigma}. The modulus $R_{\lambda}$ and phase
$\phi_{\lambda}$ of the reflection coefficients $r_{\lambda}=R_{\lambda}e^{i\phi_{\lambda}}$ (with $\lambda\in\{p,s\}$) are shown in Fig. \ref{fig2alt}, together with the correspondent quantities for the case of a simple dielectric surface without the graphene coating. While the presence of the graphene layer does not modify significatively $R_{\lambda}$ for neither $p$- or $s$-polarization (as it appears clear from Figs. \ref{fig2alt}(a) and (c), respectively), the change induced in the phases $\phi_{\lambda}$ of the reflection coefficients is considerable. For a normal air-glass interface, in fact, we have $\partial\phi_{\lambda}/\partial\theta=0$ being $\theta$ the angle of incidence. Here, instead, we have $\partial\phi_{\lambda}/\partial\theta\neq 0$.
A closer inspection of Eqs. \eqref{reflection}, moreover, reveals that such a novel $\theta$-dependence of the phases $\phi_{\lambda}$ is entirely due to the graphene conductivity $\sigma(\Omega)$.
\begin{figure*}[!t]
\begin{center}
\includegraphics[width=\textwidth]{r_absarg_single.pdf}
\caption{Modulus (left column) and phase (right column) of the reflection coefficients $r_{\lambda}=R_{\lambda}\exp{(i\phi_{\lambda})}$ for $p$-polarization (top row) and $s$-polarization (bottom row). In all graphs, the solid black line corresponds to the case of the graphene-coated surface, while the red dashed curve corresponds to the case without graphene coating. The refractive index of the bulk medium is chosen to be $n=1.5$.}
\label{fig2alt}
\end{center}
\end{figure*}
To compute the GH and IF shifts, we calculate the center of mass of the intensity distribution in the reflected frame, namely \cite{brewster}
\begin{equation}
\langle \vett{R}\rangle=\frac{\iint\limits_{-\infty}^{+\infty} \vett{R} |\mathbf{E}_r|^2 \mathrm{d}X_r \mathrm{d}Y_r}{\iint\limits_{-\infty}^{+\infty} |\mathbf{E}_r|^2 \,\mathrm{d}X_r \mathrm{d}Y_r}
= \langle X_r\rangle\uvett{X}_r+\langle Y_r\rangle\uvett{y}_r,
\label{eq:GHIFshiftDef}
\end{equation}
where $\vett{R}=(X_r,Y_r)^T$. Spatial ($\Delta$) and angular ($\Theta$) GH and IF shifts are then defined as follows:
\begin{subequations}
\begin{align}
\Delta_{GH}&=\langle X_r\rangle \Big|_{z=0},& \quad
\Theta_{GH}&=\frac{\partial\langle X_r\rangle}{\partial z},\\
\Delta_{IF}&=\langle Y_r\rangle \Big|_{z=0},& \quad
\Theta_{IF}&=\frac{\partial\langle Y_r\rangle}{\partial z}.
\end{align}
\end{subequations}
The explicit expressions of the GH and IF shifts for a fundamental Gaussian beam read, according to \cite{poly1}, as follows:
\begin{subequations}\label{shifts}
\begin{align}
\Delta_{GH} &=w_p\frac{\partial\phi_p}{\partial\theta}+w_s\frac{\partial\phi_s}{\partial\theta},\\
\Delta_{IF} &=-\cot\theta\Big[\frac{w_pa_s^2+w_sa_p^2}{a_pa_s}\sin\eta\nonumber\\
&+2\sqrt{w_pw_s}\sin(\eta-\phi_p+\phi_s)\Big],\\
\Theta_{GH}&=-\left(w_p\frac{\partial\ln R_p}{\partial\theta}+w_s\frac{\partial\ln R_s}{\partial\theta}\right),\\
\Theta_{IF}&=\frac{w_pa_s^2-w_sa_p^2}{a_pa_s}\cos\eta\cot\theta,
\end{align}
\end{subequations}
where $f_p=a_p$, $f_s=a_s\exp{(i\eta)}$ and $w_{\lambda}=a_{\lambda}^2R_{\lambda}^2/(a_p^2R_p^2+a_s^2R_s^2)$ (where $\lambda\in\{p,s\}$) is the fractional energy contained in each polarization state.
As suggested by Figs. \ref{fig2alt}(a) and (c), the changes in $R_{\lambda}$ introduced by the graphene layer are negligible. We therefore expect to observe no changes in the angular shifts $\Theta_{GH}$ and $\Theta_{IF}$, as they are functions of $R_{\lambda}$ solely. The spatial shifts $\Delta_{GH}$ and $\Delta_{IF}$, on the other hand, contain a dependence on the phases $\phi_{\lambda}$, and they are therefore affected by the presence of the graphene coating. Let us first discuss the IF shift. In this case $\phi_p-\phi_s$ is very close to $\pi$ [Figs. \ref{fig2alt}(b) and (d)], and the resulting spatial shift $\Delta_{IF}$ will be \emph{nonzero} (but very small) even for linear polarization, in contrast to the case without graphene.
More interesting is the case of the spatial GH shift. For a normal air-dielectric interface, one has $\partial\phi_{\lambda}/\partial\theta=0$ and therefore, according to Eq. (8a), $\Delta_{GH}=0$. It is in fact well known since the pioneering work of Goos and H\"anchen \cite{refGH1}, that $\Delta_{GH}\neq 0$ occurs only in total internal reflection, where $R_{\lambda}=1$ and $\partial\phi_{\lambda}/\partial\theta\neq0$. For the case of a graphene-coated surface, on the other hand, the phase $\phi_{\lambda}$ varies with $\theta$ for both $s$- and $p$-polarizations, as Figs. \ref{fig2alt} (b) and (d), respectively, show. In this case, then, we observe a \emph{nonzero} spatial GH shift even without total internal reflection.
The spatial GH shift $\Delta_{GH}$ occurring at a graphene-coated dielectric surface is depicted in Fig. \ref{fig3}(a) and (b) for $p$- and $s$-polarization, respectively. As can be seen, for both polarizations we have $\Delta_{GH}\neq 0$ although no total internal reflection takes place. In particular, $\phi_p$ varies very rapidly from $0$ to $-\pi$ in the vicinity of the Brewster angle $\theta_B$. This corresponds to a giant and negative spatial GH shift. On the other hand, $\phi_s$ varies very smoothly with $\theta$, thus resulting in a \emph{nonzero} (but very small) spatial GH shift for $s$-polarization.
In conclusion, we have presented a detailed theoretical analysis of GH and IF shifts of a Gaussian beam impinging onto a graphene-coated dielectric surface. Our analysis revealed that the main effect of the graphene layer is to introduce, through its surface conducibility $\sigma(\omega)$, a dependence of the phases $\phi_{\lambda}$ of the reflection coefficients on the incident angle $\theta$. This, ultimately, reflects in the appearance of a \emph{nonzero} spatial GH and IF shifts. In particular a giant and negative spatial GH shift in the vicinity of the Brewster's angle for $p$-polarization has been predicted, in ageeement with the recently published experimental results \cite{ref13}.
\begin{figure*}[t!]
\begin{center}
\includegraphics[width=1.\textwidth]{spGH_s-and-p.pdf}
\caption{Spatial GH shift $\Delta_{GH}$ for (a) $p$- polarization and (b) $s$-polarization for a graphene-coated surface. Since $\partial\phi_{\lambda}/\partial\theta\neq 0$, in both cases $\Delta_{GH}\neq 0$. In particular, since $\phi_p$ varies very rapidly with $\theta$ in the vicinity of the Brewster angle, the corresponding spatial shift for $p$-polarization [Panel (a)] is giant in modulus, and negative due to the fact that $\phi_p$ varies from $0$ to $-\pi$ [See Fig. \ref{fig2alt}(b)].}
\label{fig3}
\end{center}
\end{figure*}
The authors thank the German Ministry of Education and Science (ZIK 03Z1HN31) for financial support.
|
\section{Introduction}
True ternary fission of heavy nuclei, which has been discovered in the experiments
recently, occurs with much smaller probability ($\sim 10^{-3}$) compared to the binary fission
~\cite{PKVA10,PKKA10,PKAA11,PKVA12}.
These experimental studies of the decays in
$^{252}$Cf(sf,fff) \cite{PKVA10} and $^{235}$U(n$_{\rm th}$,fff) \cite{PKVA12} reactions
with two fission fragment coincidences
with two FOBOS-detectors \cite{PKVA10} placed at $180^{\circ}$,
using the missing mass approach, have established
the phenomenon of collinear cluster tripartition (CCT). The third product is not observed.
Only recently its dynamical properties could be investigated and it has been
concluded that it proceeds collinearly.
More details can be found, for example, in Refs.~\cite{DG74,Royer95,MB11}, and it is expected
that the investigation of true ternary fission will allow us to extend our knowledge about
fusion-fission processes.
In the present work, we consider one of the dominant modes of the CCT.
Because of small probability, there are only a few experimental measurements of CCT as
given in Refs.~\cite{PKVA10,PKKA10,PKAA11,PKVA12}.
Moreover, theoretical studies about ternary fission are very limited and
some early works on this topic can be found, for example, in
Refs.~\cite{MG76,DG74,Royer95,PGG04}.
In these early works it was found that the ternary fission process in heavy nuclei occurs
preferably in collinear geometry, which was confirmed by the recent theoretical studies in
Refs.~\cite{MB11,VVB12,TNS11}.
In Ref.~\cite{MB11}, the authors studied the difference between equatorial configurations and collinear
configurations in ternary fission by calculating the potential energies for geometries of three
fragments touching each other, i.e., a tri-nuclear system (TNS), where only the mass number of the
central nucleus changes.
On the other hand, the kinetic energies of the CCT products were evaluated in Ref.~\cite{VVB12}
to find that in most cases the velocity of the central fragment can be very small.
This may be responsible for missing the third product in the experiments of
Refs.~\cite{PKVA10,PKKA10,PKAA11,PKVA12}.
In Ref.~\cite{TNS11} the CCT process has been considered as two sequential binary fissions.
Namely, in the first stage the excited compound nucleus decays into two fragments in an asymmetric
channel, then the heavier fragment decays further into two fragments.
As a result, three fragments are obtained with comparable masses.
In Ref.~\cite{TNS11} only the yield of ternary fission fragments with comparable masses has been
considered as it is similar to the case observed in the experiments performed by the FOBOS
group~\cite{PKKA10}.
The theoretical results of the yield of \nuclide[80]{Ge} and \nuclide[84]{Se} isotopes as the first step
for the CCT products and the products \nuclide[70]{Ni}, \nuclide[74,76]{Zn}, and \nuclide[82]{Ge}
in the second step in a sequential fission process are in good agreement with some of the
corresponding experimental data on the mass distributions of the \nuclide[252]{Cf} decay.
This observation leads to the conclusion that these events can be associated with the sequential two
step mechanism of CCT.
However, the yields of \nuclide[70]{Ni}, \nuclide[82]{Ge}, and \nuclide[84]{Se}
in coincidence with ternary fission
masses $A = 130$--$150$ observed in Ref.~\cite{PKVA12} with a large probability were not fully
explained in Ref.~\cite{TNS11}.
Therefore, in the present paper we consider the mechanism of a sequential ternary fission with a
very short time between the ruptures of the two necks connecting the middle cluster of the collinear
TNS with its outer fragments.
This mechanism is the almost-simultaneous ternary fission as illustrated in Fig.~\ref{graph1}.
The main goal of the present work is to pursue theoretical analysis of the ternary fission channels
leading to the formation of the products of mass number $A = 132$--$140$.
\begin{figure}
\begin{center}
\includegraphics[width=0.3\textwidth]{Fig1.eps}
\end{center}
\vspace*{-0.5 cm}
\caption{(Color online) The CCT fission mechanism of a heavy nucleus in
a sequential decay~\cite{TNS11}.
$A_1$, $A_2$, and $A_3$ are mass numbers of the fragments formed in the tri-nuclear system.}
\label{graph1}
\end{figure}
The collinear configuration of the tri-nuclear system undergoing fission is defined as follows.
First, the three fragments are situated on \textit{one line} and the border nuclei are numbered as
``1'' and ``2'', while the middle nucleus is labeled as ``3'' as shown in Fig.~\ref{ThreeSph}.
Consequently, there is no nuclear interaction between the outer fragments ``1'' and ``2''.
However, the Coulomb interaction between them is taken into account because of its long-range
property.
In fact, it was found to have a nontrivial role in the decay of TNS.
The pre-scission barrier between fragments ``1'' and ``3'' decreases due to the Coulomb field
of the fragment ``2''.
For example, in the case of the sequential ternary fission of \nuclide[236]{U}, when \nuclide[132]{Sn}
forms as the fragment ``2'', the pre-scission barrier between the fragments ``2''
and ``3'' is smaller than
the one between the fragments ``1'' and ``3''~\cite{TNS11}.
Certainly the massive fragment ``2'' is separated at the first step, then occurs the rupture of the
second neck between fragments ``1'' and ``3''.
We will discuss the probability of the rupture at the second neck between ``1'' and ``3'', which
decreases with increasing the distance $R_{32}$ that induces the decrease of the Coulomb field
of the massive fragment ``2''.
The definitions of the variables of TNS used in this analysis can be found in Fig.~\ref{ThreeSph}.
\begin{figure}
\begin{center}
\includegraphics[width=0.65\textwidth]{Fig2.eps}
\vspace*{-4.0cm}
\end{center}
\caption{The variables of the tri-nuclear system used in the analysis of the interaction
energy between its fragments.
Here, $Z_i$ is the charge number of fragment $i$ ($i=1,2,3$) and $R_{ij}$ is the distance between
the mass centers of fragments $i$ and $j$.
}
\label{ThreeSph}
\end{figure}
\section{Theoretical approach: from a dinuclear system to a tri-nuclear system}
In order to explore the mechanisms of the CCT process, we apply the theoretical framework
of the dinuclear system (DNS) model~\cite{ACNPV93,ACNPV95,Volkov99,NFTA05}.
In the present work, we estimate the total energy of the interacting system by calculate the sum of
the binding energies of its constituents and the interaction potential energy between them.
The minima of the potential energy surfaces (PES) are found by the variation of the charge and mass
numbers of two fragments out of the three fragments and the distances between them.
The PES is the two dimensional driving potential which depends on the charge numbers of
two fragments of the collinear TNS. The distances $R_{13}$ and $R_{32}$ between centers of mass
of fragments are found from the minimum value of the nucleus-nucleus interaction.
The fission process is considered as a formation of the elongated mononucleus (for example,
superdeformed shape) which breaks down into two fragments as in the case of binary fission.
The formation of the third cluster in the neck region and the splitting of this system
into three fragments are related to the shape of the system such as hyperdeformation.
Furthermore, the assumption of the formation of a heavier nucleus as the third fragment
between the two main fission products is also introduced.
\subsection{Total potential energy of a tri-nuclear system}
The study on the landscape of the potential energy surface (PES) is carried out to find minima and
valleys, since, at local maxima, one can expect increased yields of the mass and charge distributions
in the TNS undergoing fission process.
It should be noted that the stage of transition from compound nucleus to the TNS configuration is not
analyzed.
Instead, we assume that TNS is formed during fission of the compound nucleus into a binary system.
This process can occur in the sense of energy conservation.
We refer to Ref.~\cite{CKT05} for the hyperdeformed \nuclide[236]{U} nucleus.
The PES is calculated as
\begin{eqnarray}
&& U(Z_1,A_1,\beta^{(1)},Z_2,A_2,\beta^{(2)},Z_3,A_3,\beta^{(3)},R_{13},R_{32})
\nonumber\\
&=& V_{\rm int}(Z_1,A_1,\beta^{(1)},Z_2,A_2,\beta^{(2)},Z_3,A_3,\beta^{(3)},R_{13},R_{32})
\nonumber \\ && \mbox{}
+Q_{\rm ggg}(Z_1,A_1,Z_3,A_3),
\label{Vtot}
\end{eqnarray}
where $Z_i$ and $A_i$ are the charge number and mass number of the $i^{\rm th}$
fragment of the TNS ($i=1,2,3$), respectively, and $R_{ij}$ is the distance between the
mass-centers of the $i^{\rm th}$ and $j^{\rm th}$ fragments.
Here, $\beta^{(i)}=\{ \beta_2^{(i)}, \beta_3^{(i)} \}$ is a set of deformation parameters
of the fragment $i$, where $\beta_2^{(i)}$ and $\beta_3^{(i)}$ represent the quadrupole
and octupole parts, respectively.
The interaction potential $V_{\rm int}$ between the fragments of TNS
can be written as
\begin{eqnarray}
&& V_{\rm int}(Z_1,A_1,\beta^{(1)},\beta^{(2)}, Z_3,A_3,\beta^{(3)},R_{13},R_{32})
\nonumber \\ &=&
\sum^{3}_{i<j}V_{ij}(Z_i,A_i,\beta^{(i)},Z_j,A_j,\beta^{(j)};R_{ij}),
\label{Vint}
\end{eqnarray}
where $V_{ij}$ is the two-body interaction potential between the nuclei ``$i$'' and ``$j$''.
It contains two parts, namely, the nuclear part $V^{ij}_{\rm nuc}$ and the Coulomb part $V^{ij}_C$,
so that
\begin{eqnarray}
&& V_{ij}(Z_i,A_i,\beta^{(i)},Z_j,A_j,\beta^{(j)};R_{ij})
\nonumber \\ &=&
V^{ij}_{\rm nuc}(Z_i,A_i,\beta^{(i)},Z_j,A_j,\beta^{(j)};R_{ij})
\nonumber\\ && \mbox{}
+V^{ij}_C(Z_i,A_i,\beta^{(i)},Z_j,A_j,\beta^{(j)};R_{ij}).
\label{Vij}
\end{eqnarray}
It is clear that $V^{12}_{\rm nuc} = 0$ since the fragments ``1'' and ``2'' are separated by the
fragment ``3'' and, therefore, there is no overlap of their nucleon densities.
The nuclear part of the nucleus-nucleus interaction $V^{ij}_{\rm nuc}$ is calculated by using
the double folding procedure~\cite{TNS11}, and the Coulomb part $V^{ij}_C$ is estimated by
the Wong expression~\cite{Wong73}.
In Eq.~(\ref{Vtot}), $Q_{\rm ggg}$ is the reaction balance energy in ternary fission, which is
written as
\begin{eqnarray}
&& Q_{\rm ggg}(Z_1,Z_3,A_1,A_3)
\nonumber \\
&=& B_1(Z_1,A_1) + B_2(Z_2,A_2) + B_3(Z_3,A_3)
\nonumber\\ && \mbox{}
-B_{CN}(Z_{CN},A_{CN}).
\label{Qgg}
\end{eqnarray}
The values of binding energies $B_i$ for ground states are taken from Refs.~\cite{AWT03,MNMS93}.
In order to calculate the mass and charge distributions of the TNS in the pre-scission state,
the minima and valleys of the PES are determined by computing the interaction potential $V_{\rm int}$
as a function of $(Z_1, A_1, Z_3, A_3, R_{13}, R_{32})$ since $(Z_2,A_2)$ can be defined through
$(Z_1, A_1, Z_3, A_3)$ and $R_{12}=R_{13}+R_{32}$.
This is done by taking $V_{\rm int}$ as a function of $R_{13}$ and $R_{32}$ for each configuration
of $\{Z_1,A_1; Z_3,A_3; Z_2,A_2\}$. (See Fig.~\ref{ThreeSph} for the geometry.)
In order to find the dominant cluster states of PES, the charge (and mass) numbers of the two
fragments are varied in the range of $2 < Z_1 < Z_{\rm CN}/2$ and
$2 < Z_3 < Z_{\rm CN}/2$ [$A_{1,\rm min} < A_1(Z_1) < A_{1,\rm max}$ and
$A_{3,\rm min} < A_3(Z_3) < A_{3,\rm max}$].
The charge and mass numbers of the third fragment can be found from the corresponding conservation
laws for them.
The distances $R_{13}$ and $R_{32}$ between interacting nuclei are then varied to find
$R^{(\rm min)}_{13}$ and $R^{(\rm min)}_{32}$ that correspond to the minimum values of the potential
wells $V_{13}$ and $V_{32}$, respectively.
It should be noted again that the potentials are affected by the Coulomb interaction
$V^{C}_{12}$ of the border fragments.
This process allows us to find the mass number $A_i$ that corresponds to the minimum value of
the PES for a given value of $Z_i$.
For example, the value of $A_1$ can be found by minimizing the PES for each value of $A_3$
at fixed values of $Z_1$ and $Z_3$.
From the set of the results calculated for PES as a function of
$(Z_1, A_1, Z_3, A_3, R_{13}, R_{32})$ we can establish the driving potential demonstrating the
configurations of the TNS with the well-pronounced cluster states having closed shells.
The three-dimensional driving potential $U_{\rm dr}(Z_1, A_1; Z_3, A_3)$ is determined by the
values of the PES in Eq.~(\ref{Vtot}) corresponding to the minimum values of the potential wells in the
nucleus-nucleus interaction $V_{\rm int}$ between neighbor fragments as a function of the distances
between their centers-of-mass:
\begin{eqnarray}
&& U_{\rm dr}(Z_1, A_1, \beta^{(1)},\beta^{(2)},Z_3, A_3, \beta^{(3)})
\nonumber \\ &=&
U(Z_1, A_1, \beta^{(1)},\beta^{(2)},Z_3, A_3, \beta^{(3)},R^{(\rm min)}_{13},R^{(\rm min)}_{32}).
\label{Udr}
\end{eqnarray}
A change of $A_i$ leads to the change of $Q_{\rm ggg}$ which depends on the binding
energies $B_i$.
As a result, $U_{\rm dr}$ is sensitive to the mass distribution between the TNS fragments.
\subsection{Probability of the yield of ternary fission fragments}
The mass and charge distributions of the TNS fragments are related to the driving potential
$U_{\rm dr}$.
Therefore, the knowledge of $U_{\rm dr}$ allows us to calculate the yield of the products of
ternary fission as in Ref.~\cite{TNS11}:
\begin{eqnarray}
Y(Z_1,A_1; Z_3, A_3) &=& P(Z_1,A_1; Z_3, A_3)
\nonumber \\ && \mbox{} \times
W_{13}(Z_1,A_1; Z_3, A_3)
\nonumber \\ && \mbox{} \times
W_{32}(Z_1,A_1; Z_3, A_3),
\label{Yield}
\end{eqnarray}
where $P(Z_1,A_1; Z_3, A_3)$ is the probability of the charge and mass distributions of the
TNS fragments.
The probability of the formation of a TNS, $P(Z_1,A_1; Z_3, A_3)$, can be found from the
condition of a statistical equilibrium as in Ref.~\cite{MS75},
i.e., the TNS has an equilibrium state before scission:
\begin{widetext}
\begin{equation}
P(Z_1,A_1; Z_3, A_3)=
P_{0}\exp\left[-U_{\rm dr}(Z_1,A_1; Z_3, A_3)/T_{\rm TNS}(Z_1,A_1; Z_3, A_3)\right],
\label{Pz}
\end{equation}
\end{widetext}
where $T_{\rm TNS}$ is the effective temperature of the TNS and
$U_{\rm dr}(Z_1,A_1; Z_3, A_3)$ is determined by the formula (\ref{Udr}).
The normalization coefficient for the yield probability is represented by $P_{0}$.
In Eq.~(\ref{Yield}), $W_{13}$ and $W_{32}$ are the decay probabilities of the TNS that are caused
by overcoming the pre-scission barriers $B_{13}$ and $B_{32}$ which correspond
to the separation of the first and second nuclei, respectively.
Their explicit expressions can be found as~\cite{TNS11}
\begin{eqnarray}
\label{W13}
W_{13}(Z_1,A_1; Z_3, A_3)&=&W^{0}_{13}\exp\left[-\frac{B_{13}}{T_{13}}\right],\\
W_{32}(Z_2,A_2; Z_3, A_3)&=&W^{0}_{32}\exp\left[-\frac{B_{32}}{T_{32}}\right],
\label{W32}
\end{eqnarray}
where $(B_{13},B_{32})$ and $(T_{13}, T_{32})$ are the pre-scission barriers and the effective temperatures on these barriers the corresponding parts of the TNS.
The barriers $B_{13}$ and $B_{32}$ prevent the separation of the outer fragments from the middle one.
These pre-scission barriers are defined by the depth of the nucleus-nucleus potential well
between neighbor fragments of the TNS.
Here, $W^{0}_{13}$ and $W^{0}_{32}$ are normalization coefficients for the corresponding
probability distributions.
The effective temperatures are determined by the excitation energy of the TNS which is generated by
the descent of the system from the saddle point in binary fission.
We assume that the third cluster is formed between the two parts of fissioning nuclei before their
splitting.
In this case, $E^*_{\rm TNS}(Z_1, A_1, Z_3, A_3)$, the excitation energy of the system,
is determined by the difference between the values at the saddle point and at the point of
the minimum driving potential with the considered charge and mass numbers of clusters:
\begin{equation}
E^*_{\rm TNS}(Z_1, A_1, Z_3, A_3)=E^*_{\rm CN}-U_{\rm dr}(Z_1,A_1;Z_3, A_3).
\end{equation}
The effective temperatures of the TNS, necks ``1-3'' and ``2-3'' are defined by the excitation
energies on the minimum of the driving potential and pre-scission barriers $B_{13}$ and $B_{32}$, respectively:
\begin{eqnarray}
T_{\rm TNS}^{} &=& \sqrt{\frac{12E^*_{\rm TNS}}{A_{CN}}},
\nonumber \\
T_{13} &=& \sqrt{\frac{12E^*_{13}}{(A_{1}+A_3)}} ,
\nonumber \\
T_{32} &=&\sqrt{\frac{12E^*_{32}}{(A_2+A_3)}},
\end{eqnarray}
where $E^*_{13}$ and $E^*_{32}$ are the excitation energies on the top of the pre-scission
barrier the DNS ``1-3'' and of the DNS ``2-3'', respectively. These excitation energies
are the result of sharing the TNS excitation energy between different degrees of freedom.
The parts of $E^*_{\rm TNS}$ corresponding to the decay axes $R_{13}$ and $R_{32}$
are estimated by assumption that their inertia masses are $A_{13}=A_1 (A_2+A_3)/A_{\rm CN}$ and
$A_{32}=A_2 (A_1+A_3)/A_{\rm CN}$, respectively. Then the values of $E^*_{13}$ and $E^*_{32}$
are found from the effective temperature of the TNS:
\begin{equation}
E^*_{\rm 13}(Z_1, A_1, Z_3, A_3)=\frac{T_{\rm TNS}^2 A_{13}}{12}-B_{13}, \\
\end{equation}
\begin{equation}
E^*_{\rm 32}(Z_1, A_1, Z_3, A_3)=\frac{T_{\rm TNS}^2 A_{32}}{12}-B_{32}.
\end{equation}
If the residual part of the TNS excitation energy
$E^*_{\rm res}=E^*_{\rm TNS}-E^*_{\rm 13}-E^*_{\rm 32}$ is larger than the energy $B_n$
for the emission of neutrons from the TNS fragments,
the ternary fission is accompanied by neutrons.
\section{\boldmath Investigation of tripartition in the \nuclide[252]{C\lowercase{f(sf)}} reaction}
\label{tripartition}
\subsection{Potential energy surface showing the cluster formation in TNS}
\label{Population}
In the experiment reported in Ref.~\cite{PKVA12}, the ternary products were formed in the
spontaneous fission of \nuclide[252]{Cf} and the yields of \nuclide[68]{Ni},
\nuclide[80-82]{Ge}, \nuclide[94]{Kr}, \nuclide[128,132]{Sn}, and \nuclide[144]{Ba} were obtained.
In the plot of the mass-mass distribution of two products (third one is missing) given in Fig.~10
of Ref.~\cite{PKVA12}, these events were found to form a rectangle, and the authors of
Ref.~\cite{PKVA12} assumed that the points in the right half of the rectangle likely
reflect the shell effects around $N=88$.
\begin{figure*}[t]
\begin{center}
\includegraphics[width=1.20\textwidth]{Fig3.eps}
\end{center}
\vspace*{-3.5 cm}
\caption{(Color online)
The potential energy surface of the \nuclide[252]{Cf(sf,fff)} reaction. The rectangle ``CCT''
shows the area of the mass numbers $Z_1 (A_1)$ and $Z_3 (A_3)$ which corresponds to the CCT products. The rectangle ``FFF'' shows the area of formation three fragments with the
close mass numbers. The solid and dashed lines show the TNS configuration having
$^{132}$Sn and $^{134}$Te, respectively, as the outer nucleus $Z_2$.}
\label{PES}
\end{figure*}
\begin{figure*}[t]
\begin{center}
\includegraphics[width=0.6\textwidth]{Fig4.eps}
\end{center}
\vspace*{-3.0 cm}
\caption{(Color online)
The pre-scission barriers $B_{\nuclide[]{Ni} \nuclide[]{Ca}}$ and $B_{\nuclide[]{Ca} \nuclide[]{Sn}}$
keeping TNS fragments together.}
\label{CaNiBarr}
\end{figure*}
The effect of the shell structure of the proton and neutron single-particle states
in the formation of a
tri-nuclear system and in its decay into the observed fission products is obviously seen in the
mass-mass distribution data of Ref.~\cite{PKVA12}.
This observation stimulates us to calculate PES, i.e., $V(Z_1,Z_3,A_1,A_3,R_{13},R_{32})$,
and the driving potential $U_{\rm dr}(Z_1,A_1,Z_3,A_3)$ for the intermediate system preceding to
their formation.
The products of a CCT-decay should be formed before being separated from the other part of the
system.
Our results for the PES are presented in Fig.~\ref{PES}.
Each point in the driving potential $U_{\rm dr}(Z_1,A_1,Z_3,A_3)$ for the TNS corresponds to a
specific configuration (channel), which consists of three interacting nuclei placed in one line as shown in Fig.~\ref{ThreeSph}.
In calculation of PES, the distances between the fragments are fixed at the values
corresponding to the minimum values of the corresponding wells in the
interaction potential between them (see Figs. \ref{CaNiBarr} and \ref{3DV1V2}).
The quadrupole deformation parameters of the first-excited $2^+$
state of nuclei \cite{Raman} are used in calculation of PES.
The rectangle ``CCT'' shows the area of the mass numbers $Z_1 (A_1)$ and $Z_3 (A_3)$ which corresponds to the CCT products. The rectangle ``FFF'' shows the area of formation three fragments with the close mass numbers. The solid and dashed lines show the TNS configuration having
$^{132}$Sn and $^{134}$Te, respectively, as the outer nucleus $Z_2$.
So, we can see the valley, which is the energetically minimum area ($Z_2=50$ and $Z_2=52$)
and corresponds to the $\nuclide[252]{Cf} \to \nuclide[]{f}_1 + \nuclide[]{f}_3 + \nuclide[132]{Sn}$ fission channel. The valley extends up to the area about $Z_3 = 28$.
As was mentioned earlier, $Z_3$ is the charge number of the middle cluster.
The phase space of the configurations corresponding to the blue color region is large.
Therefore, the probability of finding the TNS of configurations with a lower potential energy
is larger.
The configuration of $\nuclide[]{Ni} + \nuclide[]{Ca} + \nuclide[]{Sn}$ has large probability
compared with the $\nuclide[]{Ca} + \nuclide[]{Ni} + \nuclide[]{Sn}$ configuration since the PES value
of the latter configuration is about 12~MeV higher than that of the former configuration.
The calculations were performed to find the yield of the CCT products from the collinear geometry
based on the formula of Eq.~(\ref{Yield}).
In these calculations we found that the value of the pre-scission barrier plays the decisive role.
Therefore, in the next section we discuss the behavior of the barriers $B_{13}$ and $B_{32}$ for the
CCT channel of the $\nuclide[]{Ni} + \nuclide[]{Ca} + \nuclide[]{Sn}$ configuration.
\begin{figure*}[t]
\begin{center}
\includegraphics[width=0.8\textwidth]{Fig5.eps}
\end{center}
\vspace*{-5.0 cm}
\caption{(Color online)
The total nucleus-nucleus interaction potential
$V_{\rm int}$ as a function of inter-center distances $R_{13}$
and $R_{32}$ between fragments of the TNS with collinear geometry.}
\label{3DV1V2}
\end{figure*}
\subsection{The decrease of the pre-scission barrier due to the Coulomb field of outer fragments}
\label{barrier}
The mechanism of almost sequential ruptures of the two necks connecting the fragments
of a collinear TNS is suggested to explain the observed yields of heavy clusters such as
\nuclide[]{Ni}, \nuclide[]{Ge}, and \nuclide[]{Se} isotopes that appear with the product having
a mass number of $A = 138$--$148$ in the CCT of \nuclide[252]{Cf}~\cite{PKVA10,PKVA12}.
The PES shows the structure of valleys and local minima that correspond to the formation of heavy
clusters observed in experiments as shown in Fig.~\ref{PES}.
These fragments of a TNS should be emitted from the potential wells, and, therefore,
it is important to estimate the depths of the potential wells, since heavy clusters are allowed to exist during a definite long time.
In Fig.~\ref{CaNiBarr} the potential wells calculated for the TNS of
$\nuclide[]{Ni} + \nuclide[]{Ca} + \nuclide[]{Sn}$, which forms a linear chain, are presented
as functions of the distances between centers of the middle nucleus
(\nuclide[]{Ca}) and the outer nuclei (\nuclide[]{Ni} and \nuclide[]{Sn}).
The values of these nucleus-nucleus potentials are shifted by the values of $Q_{\rm ggg}$
as the contour plot of the PES [see Fig.~\ref{PES}
Eq. (\ref{Vtot})] to take into account the change of the intrinsic energy of the TNS.
\begin{figure}[t]
\begin{center}
\includegraphics[width=0.6\textwidth,clip=]{Fig6.eps}
\end{center}
\vspace*{-2 cm}
\caption{
The dependence of the pre-scission barrier $B_{13}$ for the decay of the binary system
$\nuclide[]{Ni} + \nuclide[]{Ca}$ on the distance $R_{32}$ due to the Coulomb interaction
of the massive third fragment \nuclide[]{Sn} in the collinear geometry.}
\label{Dwell}
\end{figure}
\begin{figure*}[t]
\begin{center}
\vspace*{3cm}
\includegraphics[width=1.0\textwidth,clip=]{Fig7.eps}
\end{center}
\vspace*{-3.3cm}
\caption{(Color online)
Theoretical results for the yield of the outer fragments \nuclide[A_1]{Z_1} and \nuclide[A_2]{Z_2} of
the TNS formed at the spontaneous fission of \nuclide[252]{Cf}.}
\label{YZ1Z2}
\end{figure*}
For the calculation of the interaction potential $V_{\nuclide[]{Ca} \nuclide[]{Sn}}$,
the distance $R_{\nuclide[]{Ni} \nuclide[]{Ca}}$ is fixed to the value corresponding
to the minimum of $V_{\nuclide[]{Ni} \nuclide[]{Ca}}$, while $V_{\nuclide[]{Ni} \nuclide[]{Ca}}$
potential is calculated at the fixed value of $R_{\nuclide[]{Ca} \nuclide[]{Sn}}$ that gives the
minimum of $V_{\nuclide[]{Ca} \nuclide[]{Sn}}$.
The results for the nucleus-nucleus interaction between the nuclei of the collinear TNS of
$\nuclide[]{Ni} + \nuclide[]{Ca} + \nuclide[]{Sn}$ as a function
of the independent variables $R_{13}$ ($R_{\nuclide[]{Ni} \nuclide[]{Ca}}$) and
$R_{32}$ ($R_{\nuclide[]{Ca} \nuclide[]{Sn}}$) are given by a three dimensional plot of the PES
in Fig.~\ref{3DV1V2}.
The contour lines of the PES presented in Fig.~\ref{PES} is calculated with the minimum value of the
nucleus-nucleus interaction at $R_{13}=11$~fm and $R_{32}=12$~fm in Fig.~\ref{3DV1V2}.
The decay of the TNS occurs due to the motion of the system along $R_{13}$ or $R_{32}$.
The height of the pre-scission barrier is smaller in the direction along $R_{32}$
($R_{\nuclide[]{Ca} \nuclide[]{Sn}}$ in Fig.~\ref{CaNiBarr}) and, therefore, the massive fragment
\nuclide[]{Sn} is separated first from the TNS. This result is obtained by the use of Eqs. (\ref{W13}) and (\ref{W32}).
If the residual $\nuclide[]{Ni} + \nuclide[]{Ca}$ part of the TNS does not decay, the binary decay would
be observed since the $\nuclide[]{Ni} + \nuclide[]{Ca}$ system is considered as a superdeformed
shape of \nuclide[48]{Cd}.
\begin{figure*}[t]
\begin{center}
\vspace*{3cm}
\includegraphics[width=1.0\textwidth,clip=]{Fig8.eps}
\end{center}
\vspace*{-2.8 cm}
\caption{(Color online)
Theoretical results for the yield of the outer \nuclide[A_1]{Z_1} and middle \nuclide[A_3]{Z_3} fragments
of the TNS formed at the spontaneous fission of \nuclide[252]{Cf}.}
\label{YZ1Z3}
\end{figure*}
The excitation energy of the residual $\nuclide[]{Ni} + \nuclide[]{Ca}$ system should be large enough
so that it decays into \nuclide[]{Ni} and \nuclide[]{Ca}, if these nuclei are observed as CCT products.
The probability of this event strongly depends on the position of the separated massive product, i.e.,
the \nuclide[]{Sn} nucleus.
The depth of the potential $V_{\nuclide[]{Ni} \nuclide[]{Sn}}$, which is the pre-scission barrier
$B_{\nuclide[]{Ni} \nuclide[]{Sn}}$, changes as a function of the distance $R_{32}$.
To show this phenomenon we estimate the change of the $B_{13}$ barrier, which is the difference
between the maximum (on the barrier) and the minimum values of $V_{\rm int}$ as a function of
$R_{13}$ in Eq.~(\ref{Vint}).
The dependence of the change of the barrier $B_{13}$ by the distance $R_{32}$ is reduced to a
simple form of
\begin{equation}
\Delta B_{13}(R_{32})=\frac{Z_1 Z_2 e^2}{(R^{(B)}_{13}+R_{32})}
-\frac{Z_1 Z_2 e^2}{(R^{\rm (min)}_{13}+R_{32})},
\end{equation}
where $e^2 = 1.44$~MeV$\cdot$fm.
The dependence of $B_{13}$ on $R_{32}$ is presented in Fig.~\ref{Dwell}.
The negative values mean the decrease of the depth of the potential well
($B_{13}(R_{32} \rightarrow \infty)+\Delta B_{13}(R_{32})$) in the interaction
of the $\nuclide[]{Ni} + \nuclide[]{Ca}$ system.
The main observation of the present work is that the presence of the third fragment is important to
cause the decay of the $\nuclide[]{Ni} + \nuclide[]{Ca}$ system in an easier way.
The presence of the third massive fragment \nuclide[]{Sn} makes the pre-scission barrier shallower
by $4$~MeV, and thus the decay probability of the $\nuclide[]{Ni} + \nuclide[]{Ca}$ system increases.
By taking into account the change of the pre-scission barrier one can obtain reasonable
results for the yields of the \nuclide[]{Ni} isotopes followed by the emission
of massive \nuclide[]{Sn} isotopes from the formula in
Eq.~(\ref{Yield}) that includes the effects of the pre-scission barriers $B_{13}$ and $B_{32}$.
The results are presented in Figs.~\ref{YZ1Z2} and \ref{YZ1Z3}.
In the former figure we use $Z_1$ ($A_1$) and $Z_2$ ($A_2$) axes for the plot, while in the latter
figure we use $Z_1$ ($A_1$) and $Z_3$ ($A_3$) axes.
Although the calculated yields of heavy clusters such as \nuclide[]{Ni}, \nuclide[]{Ge}, and
\nuclide[]{Se} isotopes are found to be in good agreement with the experimental data, there still
remains a difference between the mass numbers of the massive CCT products of \nuclide[252]{Cf}
observed in Refs.~\cite{PKVA10,PKVA12}, namely $A = 138$--$148$ which overlap with our results with
$A = 132$--$140$ presented in Figs.~\ref{YZ1Z2} and \ref{YZ1Z3}.
The strong yield of the products with mass numbers $A=132$--$140$ indicates the important role
of the magic number of neutrons at $N=82$.
This difference may be ascribed to that we use the tabulated masses of Refs.~\cite{AWT03,MNMS93}
to obtain the binding energies of nuclei.
This procedure, in fact, gives the binding energies of the ground states, but we may have deformed
nuclei at the scission point, which is highly probable for massive nuclei. But we should remind
the procedure of calculation of the PES by variation of the charge and mass numbers
of the two fragments of TNS ($2 < Z_1 < Z_{\rm CN}/2$ and
$2 < Z_3 < Z_{\rm CN}/2$ [$A_{1,\rm min} < A_1(Z_1) < A_{1,\rm max}$ and
$A_{3,\rm min} < A_3(Z_3) < A_{3,\rm max}$]). The depends of shell corrections on the
deformation should be studied for the most of the numerous (some thousands) combinations.
Since the primary goal of this work is to demonstrate the possibility of the formation of the \nuclide[]{Ni},
\nuclide[]{Ge}, and \nuclide[]{Se} isotopes and their yields in the CCT mechanism, we leave more
accurate and sophisticated description of the production of massive isotopes of $A = 138$--$148$ to
our future studies.
\section{Estimation of the kinetic energy of the middle fragment}
The range of the kinetic energy of the middle fragment ``3'' can be estimated by
applying the energy and momentum conservation laws.
For simplicity, we assume that the kinetic energy of the binary process is determined by the Coulomb
barrier of the nucleus-nucleus interaction.
The first step of the sequential collinear ternary fission is the separation of the right fragment ``2'' as
shown in Fig.~\ref{graph1}.
The sum of the kinetic energies of this fragment and the residual part of the TNS is the same as the
Coulomb repulsion between them, which leads to
\begin{eqnarray}
\frac{Z_1 Z_2 e^2}{R_{13} + R_{23} + d} + \frac{Z_3 Z_2 e^2}{R_{23} + d}
&=& \frac{m (A_1+A_3) v^2_{13}}{2} \nonumber\\
&+& \frac{m A_2 v^2_{2}}{2} \, ,
\\
m (A_1+A_3) v_{13}^{} + m A_2 v_{2}^{} &=& 0 \, ,
\end{eqnarray}
where $v_{13}^{}$ and $v_{2}^{}$ are the relative velocities of the DNS ``13'' and of the separated
fragment ``2'', respectively, in the laboratory frame.
The free parameter $d$ is introduced to decrease the sum of the total Coulomb barriers that can not
be larger than the reaction energy balance $Q_{\rm ggg}$ given in Eq.~(\ref{Qgg}).
The second step of the sequential collinear ternary fission is a decay of the DNS ``13'' into two
fragments ``1'' and ``3''.
The sum of their kinetic energies is then equal to the Coulomb repulsion between them so that
\begin{eqnarray}
&& \frac{Z_1 Z_3 e^2}{R_{13}} = \frac{m A_1 v'^{2}_{1}}{2} + \frac{m A_3 v'^{2}_{3}}{2} \, ,
\nonumber \, \\
&& m A_1 v'_{1} + m A_3 v'_{3} = 0 \, ,
\end{eqnarray}
where $v'_{1}$ and $v'_{3}$ are the velocities of the fragments ``1'' and ``3'', respectively, in the
moving frame with velocity $v_{13}^{}$ in the opposite direction to $v_2^{}$.
Therefore, we have
\begin{eqnarray}
v_1^{} &=& v'_{1} + v_{13}^{} \, ,
\nonumber \\
v_3^{} &=& -v'_{3} + v_{13}^{} \, .
\end{eqnarray}
\begin{figure*}[t]
\centering
\includegraphics[width=0.9\textwidth,clip=]{Fig9.eps}
\vspace*{-2.4 cm}
\caption{(Color online)
The contour map of the calculated velocity $v_3^{}$ (in cm/ns)
of the middle \nuclide[A_3]{Z_3} fragment of the TNS
formed at the spontaneous fission of \nuclide[252]{Cf} as a function of the mass numbers of the outer
fragments \nuclide[A_1]{Z_1} and \nuclide[A_2]{Z_2}. The negative values of $v_3^{}$
mean that its direction is opposite to the direction of $v_2$. }
\label{Fig9}
\end{figure*}
We then obtain $v_3^{}$ as a function of the mass numbers of the outer fragments \nuclide[A_1]{Z_1}
and \nuclide[A_2]{Z_2} and the results are presented in Fig.~\ref{Fig9}.
The negative values of $v_3^{}$ mean that its direction is opposite to the direction of $v_2$.
This figure also allows us to find the region of mass numbers $A_1$ and $A_2$ where the
velocity of the middle cluster is large enough to be registered in experiments.
One of the puzzles in the experimental data on the collinear tripartition presented in
Refs.~\cite{PKVA10,PKKA10,PKVA12} is the missing of the third fragment.
As can be understood from this analysis, the main physical reason of this phenomenon
is the smallness of the velocity of the ``missing" third product.
In Fig.~\ref{Fig9} one can see that the third product has a small velocity ($|v_3|<0.25$cm/ns)
for the case of
$A_1 = 60$--$80$ and $A_3 = 24$--$64$, which means that the range of mass numbers
for the massive fragment is $A_2 = 108$--$168$.
This region overlaps with the observed mass region, where a \nuclide[]{Ni}-like product with a mass
number of $A_1 = 68$--$72$ was observed with a massive product of
$A_2 = 136$--$144$~\cite{PKVA10,PKKA10,PKVA12}.
In the case of the symmetric masses, $A_1 \sim A_2 \sim A_3$, we have a small velocity of the middle
fragment $A_3$, namely, we get $v_3^{} = 0.3$--$0.4$~cm/ns.
The range of the mass numbers where the third middle fragment has an observable velocity
is found to be $A_1 = 100$--$120$ and $A_3 = 4$--$16$ (i.e., $A_2 = 116$--$148$) which corresponds
to the well-known ternary fission with emission of the light nuclei with a mass number of
$A_3 = 4$--$12$~\cite{Goenn04,GMK05}.
In the experiments reported in Refs.~\cite{Goenn04,GMK05}, all the three ternary fission products
were registered.
The other range of mass numbers of the outer fragments of TNS which allows to the observation of
the middle fragment is $A_1 = 104$--$112$ and $A_3 = 64$--$90$ ($A_2 = 50$--$84$).
The decay channel of $A_2 < 100$ has a very small probability to be realized since the pre-scission
barrier $B_{32}$ is sufficiently high.
Our analysis on the sequential true ternary fission shows that the possibility of observing the middle
fragment in this case is small.
\section{Conclusion}
In this work, we suggested a sequential ternary fission process with a very short time between
the ruptures of two necks connecting the middle cluster of the collinear tri-nuclear system.
The necessity of this mechanism is revealed in the decrease of the pre-scission barrier of the
residual part of the TNS due to the Coulomb field of the massive fragment being separated first.
This mechanism leads to the probability of about $10^{-3}$ for the yield of massive clusters such
as \nuclide[70]{Ni}, \nuclide[80-82]{Ge}, \nuclide[86]{Se}, and \nuclide[94]{Kr} produced with
a product of $A = 132$--$140$ in the CCT of \nuclide[252]{Cf}.
The yields of such products were observed in coincidence with a massive product of
$A = 138$--$148$ with a relatively large probability in the experiments of the FOBOS group at
the FLNR of JINR (Dubna).
To verify the realization of this mechanism, the total potential energy of the chain-like TNS was
calculated as a sum of $Q_{\rm ggg}$ and the nucleus-nucleus interaction potential energy between
its constituents.
The minima and valleys of the PES related to the shell effects in nuclei were determined by using the
binding energies obtained from the well-known mass tables of Refs.~\cite{AWT03,MNMS93} and the
calculation of the interaction potential for the charge and mass numbers of the three fragments as a
function of distances between their centers-of-mass.
The distances $R_{13}$ and $R_{32}$ between interacting nuclei were varied to find the minimum
values of the potential wells of $V_{13}$ and $V_{32}$, respectively, which are affected by the Coulomb
interaction $V^{C}_{12}$ of the border fragments.
The driving potential as a function of the charge and mass numbers of two fragments was obtained at
the values of the distances $R^{(\rm min)}_{13}$ and $R^{(\rm min)}_{32}$ that correspond to the
minimum values of the nucleus-nucleus interactions $V_{13}$ and $V_{32}$, respectively.
In order to find the dominant cluster states of the TNS, the driving potential
$U_{\rm dr}(Z_1,A_1,Z_3,A_3)$ was calculated for the values of the charge (mass) numbers of the
two fragments in the ranges of $2 < Z_1 < Z_{\rm CN}/2$ and $2 < Z_3 < Z_{\rm CN}/2$
[$A_{1,\rm min} < A_1(Z_1) < A_{1,\rm max}$ and the $A_{3,\rm min} < A_3(Z_3) < A_{3,\rm max}$].
The analysis of the results allows us to find the mass number $A_i$ corresponding to the minimum
value of the PES for a given value of $Z_i$.
The calculated total potential energy as a function of $(Z_1, A_1, Z_3, A_3)$ enabled us to
establish the three dimensional driving potential that demonstrates the configurations of TNS with
probable cluster states in the pre-fission states.
Finally, the contour lines of the three dimensional driving potential showed the structure of a valley
corresponding to the formation of the outer cluster with $Z_2 = 50$ or $N_2 = 82$ at the ternary fission,
which corresponds to the fission channel of
$\nuclide[252]{Cf} \to \nuclide[]{f}_1 + \nuclide[]{f}_3 + \nuclide[132]{Sn}$.
It was found that the valley extends up to the area of about $Z_3 = 28$ and the probability of a
configuration having lower potential energy for the TNS is large.
Therefore, the configuration of $\nuclide[]{Ni} + \nuclide[]{Ca} + \nuclide[]{Sn}$ has a
large probability in comparison with the configuration of $\nuclide[]{Ca} + \nuclide[]{Ni} + \nuclide[]{Sn}$
since the PES value of the latter configuration is about 12~MeV higher than that of the former
configuration.
The dependence of the velocity of the middle cluster on the mass numbers $A_1$ and $A_2$ was
also analyzed for the case of the collinear tripartition of \nuclide[252]{Cf}.
The main physical reason associated with the collinear tripartition is the smallness of the missed
third product.
We found that, in the range of the mass numbers $A_1 = 60$--$80$ and $A_2 = 132$--$140$, the
middle fragment has a very small velocity, which is in agreement with the observed range
of values presented in ~Refs. \cite{PKVA10,PKKA10,PKVA12}.
This means that it is indeed difficult to observe the middle product of a collinear tripartition of
\nuclide[252]{Cf} producing \nuclide[]{Ni} with the second product having a mass number of
$A_2 = 132$--$140$.
In the case of the symmetric masses, $A_1 \sim A_2 \sim A_3$, we have a small velocity of the middle
fragment $A_3$, namely, we get $v_3^{} = 0.3$--$0.4$~cm/ns. The smallness of the middle cluster
velocity may be a reason of its missing in the collinear tripartition in the $^{252}$Cf(sf,fff) \cite{PKVA10} and $^{235}$U(n$_{\rm th}$,fff) \cite{PKVA12} reactions.
The mass ranges of the two outer products, where the middle fragment can be observed, are
i) $A_1 = 100$--$120$ and $A_2 = 130$--$140$ which corresponds to the well-known ternary fission
with an emission of the light nucleus with $A_3 = 4$--$12$~\cite{Goenn04,GMK05}
and ii) $A_1 = 90$--$110$ and $A_2 = 100$--$132$.
\acknowledgments
We thank our colleagues, Prof. D. V. Kamanin and Prof. Yu. V. Pyatkov, for fruitful discussions.
A.K.N. is grateful to RFBR for partial support and W.v.O. thanks FLNR of JINR for their hospitality
during his stay in Dubna.
\newblock
A.K.N. was supported in part by the MSIP of the Korean Government under the Brain Pool
Program No. 142S-1-3-1034.
\newblock
The work of Y.O. was supported by the National Research Foundation of Korea under
Grant No.\ NRF-2013R1A1A2A10007294.
|
\section{Introduction}
\label{S:Introduction}
Consider the following special case of Dickson's lemma: for any two
functions $f,g$ on the natural numbers there are two numbers $i<j$
such that both $f$ and $g$ weakly increase on them, i.e., $f_i\le f_j$
and $g_i \le g_j$. By a combinatorial argument (due to the first
author) a simple bound for such $i,j$ is constructed. The
combinatorics is based on the finite pigeon hole principle and results
in a certain descent lemma. From the descent lemma one can prove
Dickson's lemma, then directly guess what the bound might be, and
finally verify it by an appropriate proof. We also extract (via
realizability) a bound from (a formalization of) our proof of the
descent lemma.
In its usual formulation, Dickson's lemma (for fixed functions) is a
$\Sigma^0_1$-formula. In contrast, we shall prove a quantifier-free
statement which implies Dickson's lemma in its usual form, but not
vice versa. Our proof can be carried out in the formal system of
Elementary Analysis \cite[p.144]{TroelstravanDalen88}, a conservative
extension of Heyting arithmetic with variables and quantifiers for
number-theoretic functions. In fact, we don't make use of the axiom
of choice at all. Furthermore, we can restrict induction to
quantifier-free formulas.
Dickson's lemma has many applications. For instance, it is used to
prove termination of Buchberger's algorithm for computing Gr\"obner
bases \cite{Buchberger70}, and to prove Hilbert's basis theorem
\cite{Simpson88}.
There are many other proofs of Dickson's lemma in the literature, both
with and without usage of non-constructive (or \inquotes{classical})
arguments. The original proof of Dickson \cite{Dickson13} and the
particularly nice one by Nash-Williams \cite{Nash-Williams63} (using
minimal bad sequences) are non-constructive, and hence do not
immediately provide a bound. But it is well known that by using some
logical machinery one can still read off bounds, using either
G\"odel's \cite{Goedel58} Dialectica translation as in Hertz
\cite{Hertz04} or Friedman's \cite{Friedman78} $A$-translation as in
\cite{BergerBuchholzSchwichtenberg02}. However, these bounds -- even
for the case of just two functions considered here -- heavily use
higher type (primitive recursive) functionals and are less perspicious
than the one obtained below.
The first constructive proof of Dickson's lemma has been given by
Sch\"utte and Simpson \cite{SchuetteSimpson85,Simpson88}, using
ordinal numbers and transfinite induction up to $\epsilon_0$. Similar
methods have been used by Sustik \cite{Sustik03} and Mart\'{i}n-Mateos
et al.\ \cite{MMRRAH11}. Since initial segments of transfinite
induction are used, these proofs when written in arithmetical systems
require ordinary induction on quantified formulas. A different
constructive proof has been given by Veldman \cite{Veldman04}. It
uses
dependent choice for $\Sigma_1$-formulas (with parameters), and
induction on $\Pi_2$-formulas. This proof also provides the basis of
Fridlender's \cite{Fridlender97} formalization in Agda. The
computational content of these proofs has not been studied; the bound
involved will be very different from the present one.
\section{A combinatorial proof of Dickson's lemma}
\label{S:Dickson}
We start with a finite pigeonhole principle, in two disjunctive forms.
The (rather trivial) proofs are carried out because they have
computational content which will influence the term extracted from a
formalization of our proofs in Section~\ref{S:Extract}.
\begin{lemma}[FPHDisj]
\label{L:FPHDisj}
$\forall_{m,f}(\ex_{i<j \le m} f_i=f_j \lor \ex_{j \le m} m \le f_j)$.
\end{lemma}
\begin{proof}
By induction on $m$. For $m=0$ the second alternative holds. In
case $m+1$ let $f_j$ be maximal among $f_0, \dots, f_{m+1}$. If
$m+1 \le f_j$ we are done. Else we have $f_j \le m$. Now we apply
the induction hypothesis to $f' \defeq f_0, \dots, f_{j-1}, f_{j+1},
\dots, f_{m+1}$. If two of them are equal we are done. Else $m \le
f_k$ for some $k \ne j$ and hence $f_j \le f_k$. If $f_j=f_k$ we
are done. Else we have $f_j<f_k$, contradicting the choice of $j$.
\end{proof}
Note that quantifier-free induction suffices here, since we only prove
a property of finite lists of natural numbers.
In the key lemma~\ref{L:Key} below we will need a somewhat stronger
disjunctive version of the pigeonhole principle. To this end we need
an injective coding $\pair {n} {m}$ of natural numbers which is
\inquotes{square-filling}, i.e.\ with the property
\begin{equation}
\label{E:CodeSqFill}
k^2 \le \pair {n} {m} \to k \le n \lor k \le m.
\end{equation}
This can be achieved by
\begin{alignat*}{5}
&\dots \\
&12 &\quad& 13 &\quad& 14 &\quad& 15 &\quad& \dots\\
&6 &\quad& 7 &\quad& 8 &\quad& 11 &\quad& \dots\\
&2 &\quad& 3 &\quad& 5 &\quad& 10 &\quad& \dots\\
&0 &\quad& 1 &\quad& 4 &\quad& 9 &\quad& \dots
\end{alignat*}
or explicitely
\begin{equation*}
\pair {n} {m} \defeq
\begin{cases}
n^2+m &\hbox{if $m<n$},
\\
m^2+m+n &\hbox{otherwise}.
\end{cases}
\end{equation*}
\begin{lemma}[FPHDisj2]
\label{L:FPHDisj2}
\begin{equation*}
\forall_{f,g,k}(
\ex_{i<j \le k^2}(f_i=f_j \land g_i=g_j) \lor
\ex_{j \le k^2} k \le f_j \lor
\ex_{j \le k^2} k \le g_j).
\end{equation*}
\end{lemma}
\begin{proof}
Fix $f,g,k$. Use Lemma~\ref{L:FPHDisj} with $s_i \defeq \pair {f_i}
{g_i}$ and $m \defeq k^2$. In the first case from $s_i=s_j$ we
obtain $f_i=f_j$ and $g_i=g_j$ by the injectivity of the coding. In
the second case we have some $j \le k^2$ with $k^2 \le s_j$. From
the square-filling property \eqref{E:CodeSqFill} of the coding we
obtain $k \le f_j$ or $k \le g_j$.
\end{proof}
As an immediate consequence we have
\begin{lemma}[Key]
\label{L:Key}
\begin{align*}
\forall_{f,g,n,k}(
&\ex_{n<i<j \le n+k^2+1}(f_i=f_j \land g_i=g_j) \lor {}
\\
&\ex_{n<j \le n+k^2+1} k \le f_j \lor
\ex_{n<j \le n+k^2+1} k \le g_j).
\end{align*}
\end{lemma}
\begin{proof}
Use Lemma~\ref{L:FPHDisj2} for $\lambda_i f_{n+1+i}$, $\lambda_i
g_{n+1+i}$ and $k$.
\end{proof}
Now we introduce some notation. $\mathrm{Mini}(f,n)$ is the first
argument where $f$ is minimal on $\{0, \dots, n\}$:
\begin{align*}
&\mathrm{Mini}(f,0) \defeq 0,
\\
&\mathrm{Mini}(f,n+1) \defeq
\begin{cases}
\mathrm{Mini}(f,n)
&\hbox{if $f_{\mathrm{Mini}(f,n)} \le f_{n+1}$},
\\
n+1
&\hbox{otherwise}.
\end{cases}
\end{align*}
We define functions $\Psi, \Phi, I$ and a formula $D$ with arguments
$f,g,n$. For readability $f,g$ are omitted.
\begin{equation}
\label{E:PsiPhiID}
\begin{split}
&\Psi(n) \defeq \max \{ f_{\mathrm{Mini}(g,n)}, g_{\mathrm{Mini}(f,n)} \},
\\
&\Phi(n) \defeq f_{\mathrm{Mini}(f,n)}+g_{\mathrm{Mini}(g,n)},
\\
&I(n) \defeq n+\Psi(n)^2 +1,
\\
&D(n) \defeq \ex_{i<j\le n}(f_i\le f_j \land g_i \le g_j).
\end{split}
\end{equation}
$D(n)$ expresses that $n$ is a bound for Dickson's lemma.
The next lemma states a crucial property of the function $I$: either
$I(n)$ already is a bound for Dickson's lemma, or else $\Phi$
decreases properly when going from $n$ to $I(n)$. Since this cannot
happen infinitely often, iteration of $I$ will finally give us the
desired bound.
\begin{lemma}[Descent]
\label{L:Descent}
$D(I(n)) \lor \Phi(I(n))<\Phi(n)$.
\end{lemma}
\begin{proof}
Use Lemma~\ref{L:Key} with $f$, $g$, $n$ and $\Psi(n)$. In the first
case we have $D(I(n))$. In the second case we have $n<j \le I(n)$
with $\Psi(n) \le f_j$; the third case is symmetric. Let $i \defeq
\mathrm{Mini}(g,n)$. Then $f_i \le \Psi(n)$. In case $g_i \le g_j$
we have $D(I(n))$ and are done. Therefore assume $g_j<g_i$. We
show (i) $\Phi(I(n)) \le \Phi(j)$ and (ii) $\Phi(j)<\Phi(n)$. From
$j \le I(n)$ we obtain (i). For (ii) we show
$f_{\mathrm{Mini}(f,j)}+g_{\mathrm{Mini}(g,j)} <
f_{\mathrm{Mini}(f,n)}+g_i$. Now $n<j$ implies
$f_{\mathrm{Mini}(f,j)} \le f_{\mathrm{Mini}(f,n)}$, and
$g_{\mathrm{Mini}(g,j)} \le g_j<g_i$.
\end{proof}
From Lemma~\ref{L:Descent} we construct a bound for Dickson's lemma.
Let
\begin{equation*}
I^0(n) \defeq n,\quad I^{m+1}(n) \defeq I(I^m(n)).
\end{equation*}
\begin{lemma}
\label{L:DescentCor}
$D(I^n(0)) \lor \Phi(I^n(0))+n \le \Phi(0)$.
\end{lemma}
\begin{proof}
Induction on $n$. Step $n \mapsto n+1$. Applying Lemma~\ref{L:Descent}
to $I^n(0)$ gives $D(I^{n+1}(0)) \lor
\Phi(I^{n+1}(0))<\Phi(I^n(0))$. In the second case we have
\begin{equation*}
\Phi(I^{n+1}(0))+n+1 < \Phi(I^n(0))+n+1
\le \Phi(0)+1
\end{equation*}
The latter inequality follows from the induction hypothesis, since
$D(I^n(0))$ implies $D(I^{n+1}(0))$.
\end{proof}
\begin{proposition}
\label{P:BoundDickson}
$D(I^{f_0+g_0+1}(0))$.
\end{proposition}
\begin{proof}
Apply Lemma~\ref{L:DescentCor} to $\Phi(0)+1$.
\end{proof}
This bound is far from optimal: already for
\begin{equation*}
f_n \defeq \begin{cases} 1 &\hbox{if $n=0$},\\ 0 &\hbox{else} \end{cases}
\qquad
g_n \defeq 0
\end{equation*}
with optimal bound $2$ we have
\begin{equation*}
I^{f_0+g_0+1}(0) = I^2(0) =I(I(0)) > I(0) = \Psi(0)^2 +1 = 2.
\end{equation*}
Can we extend this proof to show Dickson's lemma for finitely many
functions? For instance for three functions a corresponding version
of the key lemma holds:
\begin{align*}
\forall_{f,g,h,n,k}(
&\ex_{n<i<j \le n+k^4+1}(f_i=f_j \land g_i=g_j \land h_i=h_j) \lor {}
\\
&
\ex_{n<j \le n+k^4+1} k \le f_j \lor
\ex_{n<j \le n+k^4+1} k \le g_j \lor
\ex_{n<j \le n+k^4+1} k \le h_j)
\end{align*}
(Proof. Apply the original key lemma to $\pair f g, h, n$ and $k^2$).
We can also define a measure function $\Phi(n) \defeq
f_{\mathrm{Mini}(f,n)} + g_{\mathrm{Mini}(g,n)} + h_{\mathrm{Mini}(h,n)}$.
A natural candidate for $\Psi$ is
\begin{equation*}
\Psi(n) \defeq \max\{
f_{\mathrm{Mini}(g,n)},
f_{\mathrm{Mini}(h,n)},
g_{\mathrm{Mini}(f,n)},
g_{\mathrm{Mini}(h,n)},
h_{\mathrm{Mini}(f,n)},
h_{\mathrm{Mini}(g,n)} \}
\end{equation*}
and a natural candidate for $I$ is $I(n) \defeq n + \Psi(n)^4 +1$.
But the corresponding version of the descent lemma is false: let $n
\defeq 2$ and
\begin{align*}
f \defeq (0,1,1,1,0,f_5,\dots),
\\
g \defeq (1,0,1,0,1,g_5,\dots),
\\
h \defeq (1,1,0,0,0,h_5,\dots).
\end{align*}
Then $\Phi(n)=0$, $\Psi(n)=1$, $I(n)=4$, and we neither have $D(I(n))$
nor $\Phi(I(n))<\Phi(n)$. -- However, it may well be that a more
refined form of the present approach works. We leave this for future
research.
\section{Extracting computational content}
\label{S:Extract}
In the following, we demonstrate how a bound for Dickson's lemma can
be extracted from a proof of the existence of such a bound. The proof
we will use is essentially the one presented in
Section~\ref{S:Dickson}, i.e., it is based on the descent
lemma~\ref{L:Descent}. We will then apply the realizability
interpretation to obtain the bound. In fact, the bound will be
\emph{machine} extracted from a formalization of the existence proof.
In more detail, we shall use that $I$ is increasing (i.e., $n<I(n)$)
and that from $D(n)$ and $n<m$ we can infer $D(m)$. Then we prove the
existence of a bound by general induction with measure $\Phi$.
\subsection{General induction and recursion}
\label{SS:GenRec}
We first explain general induction w.r.t.\ a measure, and the
corresponding definition principle of general recursion.
General induction allows recurrence to all points \inquotes{strictly
below} the present one. In applications it is best to make the
necessary comparisons w.r.t.\ a measure function $\mu$; for simplicity
we restrict ourselves to the case where $\mu$ has values in the
natural numbers, and the ordering we refer to is the standard
$<$-relation. The principle of general induction then is
\begin{equation*}
\forall_{\mu, x} ( \Prog^{\mu}_x P x \to P x ),
\end{equation*}
where $\Prog^{\mu}_x P x$ expresses \inquotes{progressiveness
\index{progressive} w.r.t.\ $\mu$ and $<$, i.e.,
\begin{equation*}
\Prog^{\mu}_x P x \defequiv
\forall_x (
\forall_y(\mu y < \mu x \to P y) \to P x ).
\end{equation*}
It is easy to see that in our special case of the $<$-relation we can
prove general induction from structural induction. However, it will be
convenient to use general induction as a primitive axiom, for then the
more efficient general recursion constant $\grec$ will be extracted.
It is defined by
\begin{equation*}
\grec \mu x G = G x (
\lambda_y \ifthenelse{\mu y<\mu x}{\grec \mu y G}{\nullterm}),
\end{equation*}
where $\nullterm$ denotes a canonical inhabitant of the range. It is
easy to prove that $\grec$ is definable from an appropriate structural
recursion operator.
\subsection{Non-computational quantifiers}
\label{SS:NcQuant}
We now use general induction in our constructive proof of Dickson's
lemma. However, we have to be careful with the precise formulation of
what we want to prove. We are not interested in the pair $i,j$ of
numbers where both $f$ and $g$ increase, but only in a bound telling
us when at the latest this must have happened. Therefore the
existential quantifiers $\ex_{i,j}$ must be made \inquotes{uniform}
(i.e., non-computational); it will be disregarded in the realizability
interpretation. Such non-computational quantifiers have first been
introduced in \cite{Berger93a,Berger05}; in
\cite{SchwichtenbergWainer12} this concept is extended to all
connectives and discussed in detail. Let
\begin{equation*}
D'(n) \defeq \exnc_{i<j\le n}(f_i\le f_j \land g_i \le g_j).
\end{equation*}
Using this non-computational form of $D(n)$ we modify Lemma~\ref{L:Descent}
to
\begin{lemma}[Descent$^{\mathrm{nc}}$]
\label{L:DescentModif}
$D'(I(n)) \lor \Phi(I(n))<\Phi(n)$.
\end{lemma}
Note that the computational content of a proof of this lemma is that
of a functional mapping two unary functions and a number into a
boolean. From Lemma~\ref{L:DescentModif} we obtain as before a
modification of Proposition~\ref{P:BoundDickson} to
\begin{proposition}[Bound for Dickson's lemma]
\label{P:BoundDicksonModif}
\begin{equation*}
\forall_{f,g,n} \ex_k(
I(n) \le k \land
D'(k)).
\end{equation*}
\end{proposition}
\begin{proof}
By general induction with measure function $\Phi$. We fix $f,g$ and
prove progressiveness of the remaining $\forall_n \ex_k$-formula.
Therefore we can assume as induction hypothesis that for all $m$
with $\Phi(m)<\Phi(n)$ we have
\begin{equation*}
\ex_k(
I(m) \le k \land D'(k)).
\end{equation*}
We must show
\begin{equation*}
\ex_k(
I(n) \le k \land D'(k)).
\end{equation*}
By Lemma~\ref{L:DescentModif} we know $D'(I(n)) \lor
\Phi(I(n))<\Phi(n)$. In the first case we have $D'(I(n))$
and can take $k \defeq I(n)$. In the second case we apply the
induction hypothesis to $I(n)$. It provides a $k$ with $I(I(n)) \le
k$ and $D'(k)$. But $I(n) \le I(I(n))$ since
$n<I(n)$.
\end{proof}
\subsection{Formalization and extraction}
\label{SS:Extract}
The formalization\footnote{Available at
\texttt{git/minlog/examples/arith/dickson.scm}} (in
Minlog\footnote{See \url{http://www.minlog-system.de}}) of the proof
above is now routine. The term extracted from it is
\begin{verbatim}
[f,g,n](GRecGuard nat nat)(Phi f g)n
([n0,f1][if (cDesc f g n0) (I f g n0) (f1(I f g n0))])
True
\end{verbatim}
To explain this term we rewrite it in the notation above
\begin{equation*}
\lambda_{f,g,n}\grec \mu n G
\end{equation*}
with measure $\mu$ and step function $G$ defined by
\begin{align*}
\mu &\defeq \Phi,
\\
G(n,h) &\defeq
\begin{cases}
I(n) &\hbox{if $\mathtt{cDesc}(n)$, i.e., $D'(I(n))$},
\\
h(I(n)) &\hbox{otherwise, i.e., $\Phi(I(n)) < I(n)$},
\end{cases}
\end{align*}
where for readability we again omit the arguments $f,g$ from $\Phi, I,
\mathtt{cDesc}$. The functions $\Phi, I$ are defined as in
\eqref{E:PsiPhiID}, and $\mathtt{cDesc}$ is the computational content
of Lemma~\ref{L:DescentModif}:
\begin{verbatim}
[f,g,n][case (cKey f g n(f(Mini g n)max g(Mini f n)))
((DummyL nat ysum nat) -> True)
(Inr nn ->
[case nn
((InL nat nat)n0 ->
(cNatLeLtCases boole)(g(Mini g n))(g n0)True False)
((InR nat nat)n0 ->
(cNatLeLtCases boole)(f(Mini f n))(f n0)True False)])]
\end{verbatim}
Here \texttt{nn} is a variable of type $\typeN + \typeN$ with $\typeN$
the type of natural numbers, and \texttt{cNatLeLtCases}:
\begin{verbatim}
(Rec nat=>nat=>alpha=>alpha=>alpha)n
([n0,x,x0][case n0 (0 -> x0) (Succ n1 -> x)])
([n0,h,n1,x,x0][case n1 (0 -> x) (Succ n2 -> h n2 x x0)])
\end{verbatim}
is the computational content of the (simple) proof of
\begin{equation*}
\forall_{n,m}((n \le m \to P) \to (m<n \to P) \to P)
\end{equation*}
expressing case distinction w.r.t.\ $\le$ and $<$.
$\mathtt{cKey}$ is the computational content of Lemma~\ref{L:Key}:
\begin{verbatim}
[f,g,n,n0]
[case (cFPHDisjTwo([n1]f(Succ(n+n1)))([n1]g(Succ(n+n1)))n0)
((DummyL nat ysum nat) -> (DummyL nat ysum nat))
(Inr nn ->
Inr[case nn
((InL nat nat)n1 -> (InL nat nat)(Succ(n+n1)))
((InR nat nat)n1 -> (InR nat nat)(Succ(n+n1)))])]
\end{verbatim}
which uses \texttt{cFPHDisjTwo}:
\begin{verbatim}
[f,g,n][if (cFPHDisj(n*n)
([n0][if (g n0<f n0)
(f n0*f n0+g n0)
(g n0*g n0+g n0+f n0)]))
([ij](DummyL nat ysum nat))
([n0]
Inr[if (cCodeSqFill(f n0)(g n0)n)
((InL nat nat)n0)
((InR nat nat)n0)])]
\end{verbatim}
which in turn depends on \texttt{cCodeSqFill}:
\begin{verbatim}
[n,n0,n1](Rec nat=>nat=>boole)n([n2]False)
([n2,(nat=>boole),n3]
[case n3 (0 -> True) (Succ n -> (nat=>boole)n)])
n0
\end{verbatim}
and \texttt{cFPHDisj}:
\begin{verbatim}
[n](Rec nat=>(nat=>nat)=>nat@@nat ysum nat)n
([f](InR nat nat@@nat)0)
([n0,d,f]
[let n1
[if (f(Succ n0)<=f(Maxi f n0)) (Maxi f n0) (Succ n0)]
[if (Succ n0<=f n1)
((InR nat nat@@nat)n1)
[if (d([n2][if (n2<n1) (f n2) (f(Succ n2))]))
([ij]
(InL nat@@nat nat)
[if (right ij<n1)
ij
([if (left ij<n1)
(left ij)
(Succ left ij)]@Succ right ij)])
([n2]
[if (n2<n1)
((cNatLeCases nat@@nat ysum nat)(f n2)(f n1)
((InL nat@@nat nat)(0@0))
((InL nat@@nat nat)(n2@n1)))
((cNatLeCases nat@@nat ysum nat)(f(Succ n2))(f n1)
((InL nat@@nat nat)(0@0))
((InL nat@@nat nat)(n1@Succ n2)))])]]])
\end{verbatim}
To summarize, we have extracted a function which takes two functions
$f,g$ (suppressed for readability) and a number $n$ and yields a
bound. Notice that already with $n=0$ we obtain the desired bound for
Dickson's lemma. However, the inductive argument requires the general
formulation.
Our extracted bound $B(n) \defeq \grec \Phi n G$ satisfies
\begin{align*}
B(n) = \grec \Phi n G &= G n (
\lambda_m \ifthenelse{\Phi m<\Phi n}{\grec \Phi m G}{\nullterm})
\\
&=\begin{cases}
I(n) &\hbox{if $D'(I(n))$},
\\
B(I(n)) &\hbox{if $\Phi(I(n)) < I(n)$}.
\end{cases}
\end{align*}
by Lemma~\ref{L:DescentModif}, which also guarantees termination:
$B(n)$ will call itself at most $I(n)$ times. As long as the
iterations $I(n)$, $I^2(n)$, \dots, $I^m(n)$ decrease w.r.t.\ the
measure $\Phi$, the next iteration step is done. However, as soon as
Lemma~\ref{L:DescentModif} goes to its \inquotes{left} alternative
(i.e., $D'(I(n))$ holds), $I(n)$ is returned. Hence this extracted
bound differs from the \inquotes{guessed} one in
Proposition~\ref{P:BoundDickson} in that it does not iterate $I$ a
prescribed number of times (${f_0+g_0+1}$ many) at $0$, but stops when
allowed to do so by the outcome of Lemma~\ref{L:DescentModif}.
|
\section{Introduction} \renewcommand{\thetheorem}{\Alph{theorem}}
A C$^*$-algebra $A$ is said to be {\it nuclear} if for any C$^*$-algebra $B$ there is a unique C$^*$-cross norm on the algebraic tensor product $A \odot B$.
The notion of nuclearity was introduced by Takesaki \cite{Takesaki} in the 1960's.
On the other hand,
by remarkable works of Lance \cite{Lance_JFA}, Choi--Effros \cite{Choi-Effros}, and Kirchberg \cite{Kirchberg}, the nuclearity of a given C$^*$-algebra is known to be equivalent to the {\it completely positive approximation property} ({\it CPAP\/}), that is, the identity map can be approximated by a net of completely contractive positive maps that factor through matrix algebras.
This characterization is useful and plays an important role in various situations.
In this paper, we study `relative counterparts' of nuclearity and CPAP for inclusions of C$^*$-algebras.
It seems natural to formulate the `relative CPAP' of a given inclusion $B \subset A$ of C$^*$-algebras by the following asymptotically commuting diagram:
\[
\xymatrix @C=0.5cm @R=1.2cm
{
A
\ar[rr]^{\id} \ar[rd]_{\varphi_i} && A \\
& \lM_{n(i)} (B) \ar[ru]_{\psi_i}&
}
\]
Namely, $\varphi_i$ and $\psi_i$ are completely positive maps satisfying $a=\lim_i \psi_i \circ \varphi_i (a)$ for $a \in A$.
In this case, the CPAP of $B$ implies the one of $A$.
Moreover, there are many examples known to have this `relative CPAP'; for example,
crossed products by amenable discrete groups, group C$^*$-algebras of relative amenable discrete groups, tensor products with nuclear C$^*$-algebras, continuous fields of nuclear C$^*$-algebras, etc.
From the viewpoint of equivalence between nuclearity and CPAP, those inclusions should be `relatively nuclear', but the explicit formulation of `relative nuclearity' has never been established so far.
The aim of this paper is to formulate relative nuclearity for a given inclusion $B \subset A$ in such a way that it includes original nuclearity as a particular case when $B= \lC 1_A$, and is characterized by a kind of relative CPAP.
\medskip
It is widely known that nuclearity is to C$^*$-algebras what amenability is to von Neumann algebras.
Thus, motivated by Popa's formulation \cite{Popa} of relative amenability for von Neumann algebras, we will develop our theory of `relative nuclearity' by the use of C$^*$-correspondences, which are C$^*$-counterparts of bimodules over von Neumann algebras.
For C$^*$-algebras $A$ and $B$, an $A$-$B$ C$^*$-correspondence is given by a pair $(X, \pi_X)$, where $X$ is a Hilbert $B$-module and $\pi_X$ is a $*$-homomorphism from $A$ into the C$^*$-algebra of adjointable (right $B$-linear) operators on $X$.
$A$-$A$ C$^*$-correspondences are also called C$^*$-correspondences over $A$.
To define `relative nuclearity', let us introduce the notion of {\it universal factorization property} ({\it UFP\/}).
We say that a C$^*$-correspondence $(X, \pi_X)$ over $A$ has the UFP if for any C$^*$-algebra $B$ and any $*$-representation $\sigma :A \otimes_{\max} B \to \lB (H)$,
$\sigma$ factors through the image of the natural representation $\phi_X^H :A \otimes_{\max} B \to \lB (X \otimes_B H)$ (see \S\S \ref{ss-rel-UFP} for its precise definition) as follows.
\[
\xymatrix @C=2cm
{
A \us\otimes_{\rm max} B
\ar[d]_{\phi_X^H} \ar[r]^{ \sigma} & \lB( H) \\
\Imag \phi_X^H \ar@{-->}[ru]
}
\]
In our theory, the notion of UFP plays a role of the original definition of nuclearity.
Indeed, the nuclearity of $A$ is naturally equivalent to the UFP of $(H \otimes A, \pi_H \otimes 1)$ for some faithful $*$-representation $\pi_H :A \to \lB (H)$.
For a given unital inclusion $B \subset A$ with conditional expectation $E$,
we denote by $(L^2(A, E), \pi_E)$ the $A$-$B$ C$^*$-correspondence associated with $E$ given by the GNS-construction.
For simplicity, let us assume that $E$ is nondegenerate (i.e., $\pi_E$ is faithful) in the rest of this section.
We say that $(A, B, E)$ is {\it nuclear} if $(L^2(A, E) \otimes_B A, \pi_E \otimes 1)$ has the UFP.
In the case when $B = \lC 1_A$, $L^2(A, E) $ is a Hilbert space, and hence the nuclearity of $(A, \lC 1_A, E)$ is equivalent to the nuclearity of $A$.
Moreover, we prove the following theorem, which is a relative analogue of `nuclearity $\Leftrightarrow$ CPAP'.
\begin{theorem}\label{thm-A}
Let $B \subset A$ be a unital inclusion of $\rC^*$-algebras with conditional expectation $E$.
Then, $(A, B, E)$ is nuclear if and only if for any finite subset $\fF \subset A$ and $\varepsilon >0$, there exist $n,m \in \lN$ and completely positive maps $\varphi_k : A \to \lM_n (B)$ and $\psi_k : \lM_n (B) \to A$, $1 \leq k \leq m$ such that $\| a - \sum_{k=1}^m \psi_k \circ \varphi_k (a) \| < \varepsilon$ for $a \in \fF$, and each $\varphi_k$ and $\psi_k$ are of the form
\[
\varphi_k : a \mapsto \left[ \rule{0pt}{10pt} E (x_i^* a x_j ) \right]_{i,j=1}^n
\quad
\psi_k : \left[ \rule{0pt}{0pt}\, b_{ij}\, \right]_{i,j=1}^n \mapsto \sum_{i,j=1}^n y_i^* b_{ij} y_j
\]
for some $x_i, y_i \in A$, $1 \leq i \leq n$.
\end{theorem}
This theorem will be proved based on the following two observations concerning weak containment for C$^*$-correspondences:
The first one is that for given $A$-$B$ C$^*$-correspondences $(X, \pi_X)$ and $(Y, \pi_Y)$ with $A$ unital, the following are equivalent (see \S\S \ref{ss-weak-def} for the definition of weak containment):
\begin{itemize}
\item $(X, \pi_X)$ is weakly contained in $(Y, \pi_Y)$ with respect to the universal representation, written $(X, \pi_X) \prec_\univ (Y,\pi_Y)$.
\item For any $\xi \in X$, finite subset $\fF \subset A$ and $\varepsilon >0$, there exist $m \in \lN$ and $\eta_1, \dots, \eta_m \in Y$ such that $\| \i< \xi, \pi_X(a) \xi > - \sum_{k=1}^m \i< \eta_k, \pi_Y(a) \eta_k> \| < \varepsilon$ for $a\in \fF$.
\end{itemize}
The second one is that any C$^*$-correspondence $(X, \pi_X)$ over $A$ has the UFP if and only if $(A, \lambda_A) \prec_\univ (X, \pi_X)$ holds, where $(A, \lambda_A)$ is the identity C$^*$-correspondence over $A$ (see \S\S \ref{ss-rel-UFP}).
These observations also say that the nuclearity of $(A, B, E)$ is characterized by the condition that $(A, \lambda_A) \prec_\univ (L^2(A, E) \otimes_B A, \pi_E \otimes 1_A)$.
We point out that this is parallel to Popa's formulation \cite{Popa} of relative amenability for von Neumann algebras: an inclusion of von Neumann algebra $N \subset M$ is amenable if and only if ${}_M L^2(M)_M \prec {}_M L^2(M) \otimes_N L^2(M)_M$ holds.
\medskip
We also introduce the notion of {\it strong relative nuclearity}.
Roughly speaking, the strong nuclearity of a given triple $(A, B, E)$ is the property that each $\psi_k \circ \varphi_k$ in Theorem \ref{thm-A} can be chosen to be $B$-bimodule maps (see \S\S \ref{ss-rel-rel}).
This stronger notion seems technical, but almost all the examples of nuclear triples investigated in this paper are in fact strongly nuclear.
For example, a triple $(A, B, E)$ is strongly nuclear in any of the following cases:
\begin{itemize}
\item $A$ is nuclear, $B$ is finite dimensional, and the embedding $B \hookrightarrow A$ is full (i.e., $\ospan A b A =A$ for $b \in B \setminus \{ 0 \}$);
\item $A = B \otimes C$ with $C$ nuclear;
\item $B =C(X) \subset A' \cap A$ and $A$ is a continuous field of nuclear C$^*$-algebras over $X$;
\item $E$ is of Watatani index finite type;
\item $A=B \rtimes_\alpha \Gamma$, where $\Gamma$ is a discrete amenable group;
\item $B=C(X)$ and $A =C(X) \rtimes_\alpha \Gamma$, where $\alpha : \Gamma \act C(X)$ is amenable (\cite{Delaroche});
\item $A=\rC^*_\re (\Gamma)$ and $B=\rC^*_\re (\Lambda)$, where $\Lambda \triangleleft \Gamma$ is co-amenable (i.e., $\Gamma / \Lambda$ is amenable);
\item $A=\rC^*_\re (\cG)$ and $B = C (\cG^\unit)$, where $\cG$ is a locally compact Hausdorff amenable \'{e}tale groupoid whose unit space $\cG^\unit$ is compact.
\end{itemize}
We also show that a triple $(A, B, E)$ is nuclear in any of the following cases:
\begin{itemize}
\item $A$ is nuclear and the embedding $B \hookrightarrow A$ is full;
\item $A=\rC^*_\re (\Gamma)$ and $B=\rC^*_\re (\Lambda)$, where $\Lambda < \Gamma$ is co-amenable;
\item $E$ is of probabilistic index finite type;
\item $A=B \rtimes_\alpha \Gamma$, where $\alpha : \Gamma \act B$ is amenable.
\end{itemize}
Clearly, strong relative nuclearity implies relative nuclearity, but we do not know whether or not these two notions are actually different.
\medskip
As a (kind of) byproduct of our investigation of `relative nuclearity', we also obtain
Weyl--von Neumann--Voiculescu type results that partially generalize the ones due to Kasparov \cite{Kasparov} and Skandalis \cite{Skandalis}.
In particular, we prove the next characterization for weak containment (see Theorem \ref{thm-absorbing}).
\begin{theorem}\label{thm-B}
Let $A$ and $B$ be $\rC^*$-algebras with $A$ unital separable and $B$ $\sigma$-unital,
and $(X,\pi_X)$, $(Y,\pi_Y )$ be $A$-$B$ $\rC^*$-correspondences with $X$ countably generated and $\pi_X$ and $\pi_Y$ unital.
Then, $(X, \pi_X)\prec_\univ (Y,\pi_Y)$ if and only if $(X \oplus Y^\infty, \pi_X \oplus \pi_Y^\infty)$ and $(Y^\infty, \pi_Y^\infty)$ are approximately unitarily equivalent,
where $(Y^\infty, \pi_Y^\infty)$ is the countable infinite direct sum of $(Y, \pi_Y)$.
\end{theorem}
As applications, we obtain several results in $KK$-theory.
Firstly, we show that our strong relative nuclearity implies Germain's relative $K$-nuclearity \cite{Germain-fields}, which is a relative counterpart of Skandalis's $K$-nuclearity \cite{Skandalis}.
In \cite{Skandalis}, Skandalis proved that nuclearity implies $K$-nuclearity by using Kasparov's generalized Voiculescu theorem.
Similarly, we prove relative $K$-nuclearity by establishing a certain Weyl--von Neumann--Voiculescu type assertion under strong relative nuclearity (see \S\S \ref{ss-rel-K-K}). Then, we prove the following theorems:
\begin{theorem}\label{thm-C}
Let $\{(A_i, B, E_i) \}_{i \in \cI}$ be an at most countable family of unital inclusions of separable $\rC^*$-algebras $B \subset A_i$ with conditional expectations $E_i : A_i \to B$.
If each triple $(A_i, B, E_i)$ is strongly nuclear,
then the canonical surjection from
the full amalgamated free product $\bigstar_{B, i \in \cI} A_i$ onto the reduced one $\bigstar_{B, i \in \cI} (A_i, E_i)$ is a $KK$-equivalence.
\end{theorem}
\begin{theorem}\label{thm-D}
Let $\{(A_i, B, E_i) \}_{i \in \cI}$ be an at most countable family of unital inclusions of separable $\rC^*$-algebras $B \subset A_i$ with nondegenerate conditional expectations $E_i :A_i \to B$.
If each $A_i$ is nuclear and $B$ is finite dimensional, then the canonical surjection from $\bigstar_{B, i \in \cI} A_i$ onto $\bigstar_{B, i \in \cI} (A_i, E_i)$ is a $KK$-equivalence.
\end{theorem}
These two theorems follow from a somewhat more general and technical result (see \S\S \ref{ss-KK-CD}).
We note that Germain's result on free products of unital separable nuclear C$^*$-algebras \cite{Germain-duke} is a particular case of these theorems with $B=\lC$.
Finally, combing our result with Thomsen's result \cite{Thomsen} on $K$-theory of full amalgamated free products we obtain six term exact sequences in $K$-theory of reduced ones.
\medskip
This paper is organized as follows.
In \S \ref{sec-pre}, we recall basic facts on C$^*$-correspondences.
The definition of weak containment for C$^*$-correspondences is given in \S \ref{sec-weak}.
In that section, we also characterize weak containment by a certain approximation property for coefficients, mentioned above.
In \S \ref{sec-rel} we introduce the notion of UFP.
We then define relative nuclearity and strong one, and prove Theorem \ref{thm-A}.
We also see that our relative nuclearity is related to relative WEP recently introduced by Jian and Sepideh \cite{Jian-Sepideh}, as well as relative amenability for von Neumann algebras (\cite{Popa}\cite{Delaroche2}\cite{Ozawa-Popa}).
In \S \ref{sec-exam}, we see that the examples listed above are actually (strongly) nuclear.
\S \ref{sec-voic} is devoted to proving Weyl--von Neumann--Voiculescu type results.
Applications in $KK$-theory are given in the last three sections.
In \S \ref{sec-rel-K} we prove that strong relative nuclearity implies relative $K$-nuclearity.
The proof of Theorem \ref{thm-C} and Theorem \ref{thm-D} are given in \S \ref{sec-KK}.
In the final section, we establish six term exact sequences in $KK$-theory of reduced amalgamated free products and a $KK$-equivalence result for HNN-extensions.
\subsection*{Notation}
We basically follow the notation of Brown and Ozawa's book \cite{Brown-Ozawa}.
For a C$^*$-algebra $A$ we denote by $1_A$ the unit of the multiplier algebra $\cM (A)$ of $A$.
The symbols $\lB (H)$ and $\lK (H)$ stand for the set of bounded operators and the set of compact ones on a Hilbert space $H$, respectively.
We use the symbol $\odot$ to denote the algebraic tensor product over $\lC$.
For C$^*$-algebras $A$ and $B$ we denote by $A\otimes B$ and $A\otimes_{\rm max} B$ the minimal and the maximal tensor products, respectively.
For a von Neumann algebra $M$, we denote by $M_*$ the (unique) predual of $M$.
The von Neumann algebraic tensor product of $M$ and another von Neumann algebra $N$ is denoted by $M \botimes N$.
We denote by $\CP (A,B), \CCP(A, B)$ and $\UCP(A, B)$ the c.p.\ (completely positive) maps, the c.c.p.\ (completely contractive positive) maps, and the u.c.p.\ (unital completely positive) maps from $A$ into $B$, respectively.
For a linear map $\varphi : A \to B$ we denote by $\varphi^{(n)}$ the linear map $\varphi \otimes \id_{\lM_n } :\lM_n (A) \to \lM_n (B)$ with identification $\lM_n (A) = A \otimes \lM_n$, etc.
For these terminologies we refer the reader to \cite[Chapter 1-3]{Brown-Ozawa}.
For elements $x$ and $y$ in a normed space $X$ and $\varepsilon >0$ we write $x \approx_\varepsilon y$ when $\| x - y \| < \varepsilon$ holds.
The closed unit ball of $X$ is denoted by $\Ball (X)$.
\section*{Acknowledgment} The author wishes to express his gratitude to Professor Yoshimichi Ueda, who is his supervisor, for his continuous guidance and constant encouragement.
He is also grateful to Yuki Arano for many stimulating conversations.
\setcounter{theorem}{0}
\renewcommand{\thetheorem}{\arabic{section}.\arabic{theorem}}
\section{Preliminaries on C$^*$-correspondences}\label{sec-pre}
In this section, we fix notations and terminologies and recall basic facts on C$^*$-correspondences.
We refer the reader to Lance's book \cite{Lance} for Hilbert C$^*$-module theory.
\begin{definition}
An {\it inner product} $A$-module is a linear space $X$ with a right $A$-action which is compatible with scalar multiplication, i.e., $\lambda (\xi a)= (\lambda \xi ) a = \xi (\lambda a)$ for $\lambda \in \lC, \xi \in X, a \in A$ and an $A$-valued inner product $\i< \cdot, \cdot > : X \times X \to A$ satisfying the following conditions:
\begin{itemize}
\item[(1)] $\i< \xi, \lambda \eta + \mu \zeta >= \lambda \i< \xi , \eta > + \mu \i< \xi, \zeta >$ for $\xi, \eta, \zeta \in X$ and $\lambda, \mu \in \lC$,
\item[(2)] $\i< \xi, \eta a> = \i< \xi, \eta >a$ for $\xi, \eta \in X$ and $a\in A$,
\item[(3)] $\i< \xi, \eta >^*= \i< \eta, \xi >$ for $\xi, \eta \in X$,
\item[(4)] $\i< \xi, \xi >\geq 0$ for $\xi \in X$,
\item[(5)] $\xi =0$ if and only if $\i< \xi, \xi >=0$ for $\xi \in X$.
\end{itemize}
When $X$ is complete with respect to the norm $\| \xi \| = \| \i< \xi, \xi > \|^{1/2}$,
we call $X$ a {\it Hilbert $A$-module} or {\it Hilbert $\rC^*$-module over $A$}.
\end{definition}
Let $X$ be a Hilbert C$^*$-module over a C$^*$-algebra $A$.
When $A=\lC$, $X$ is a usual Hilbert space.
When we would like to emphasize the C$^*$-algebra $A$ of coefficients, we will write $\i< \cdot, \cdot >_A$.
A Hilbert $A$-module $X$ is said to be {\it countably generated} if there exists a countable subset $\{ \xi_n \}_{n=1}^\infty \subset X$ such that $\ospan \{ \xi_n a \mid a \in A, n \in \lN \} =X$.
Let $X$ and $Y$ be Hilbert $A$-modules.
A linear map $x : X \to Y$ is said to be {\it adjointable} if there exists a linear map $x^* :Y \to X$ which enjoys $\i< \eta, x \xi>= \i< x^* \eta, \xi>$ for all $\xi \in X, \eta \in Y$.
Note that adjointability implies $A$-linearity.
We denote by $\lL_A (X,Y)$ the set of adjointable maps from $X$ into $Y$ and set $\lL_A (X):=\lL_A (X,X)$.
Every adjointable map is automatically bounded and $\lL_A (X)$ forms a unital C$^*$-algebra with respect to the operator norm and the involution $x \mapsto x^*$.
For given vectors $\xi, \eta \in X$ we define the `rank one' operator $\theta_{\xi, \eta} \in \lL_A (X)$ by $\theta_{\xi, \eta} (\zeta ) =\xi \i< \eta, \zeta>$.
We denote by $\lK_A (X)$ the C$^*$-subalgebra of $\lL_A (X)$ generated by $\{ \theta_{\xi, \eta} \mid \xi, \eta \in X \}$.
Operators in $\lK_A (X)$ are called compact operators on $X$.
It is known that $\lK_A (X)$ is a C$^*$-ideal of $\lL_A (X)$ and $\lL_A(X)$ is isomorphic to the multiplier algebra of $\lK_A(X)$. We denote by $1_X$ the identity operator on $X$.
\begin{definition}
Let $A$ and $B$ be C$^*$-algebras.
An $A$-$B$ C$^*$-{\it correspondence} is a pair $(X, \pi_X)$ consisting of a Hilbert $B$-module $X$ and a $*$-homomorphism $\pi_X :A \to \lL_A (X)$, called the left action.
$A$-$A$ C$^*$-correspondences are also called {\it $\rC^*$-correspondences over} $A$.
We denote by $\Corr (A,B)$ the set of $A$-$B$ C$^*$-correspondences and set $\Corr (A):=\Corr(A,A)$.
\end{definition}
A $\rC^*$-correspondence $(X, \pi_X) \in \Corr (A, B)$ is said to be {\it unital} if $A$ is unital and $\pi_X$ is also a unital map, and {\it countably generated} if $X$ is countably generated as a Hilbert $B$-module.
We denote by $\Rep (A)$ the set of nondegenerate $*$-representations of $A$.
$A$-$B$ C$^*$-correspondences $(X,\pi_X)$ and $(Y,\pi_Y)$ are said to be {\it unitarily equivalent}, denoted by $(X, \pi_X) \cong (Y, \pi_Y)$, if there exists a unitary $U \in \lL_B (X, Y)$ such that $\pi_X(a)= U^* \pi_Y (a) U$ for $a\in A$.
When no confusion may arise, we will write $X$ for short instead of $(X, \pi_X)$.
\begin{definition}\label{def-coeff}
For $X \in \Corr (A, B)$ and $\xi, \eta \in X$ the mapping $A \ni a \mapsto \i< \xi, \pi_X(a) \eta>$ is called a {\it coefficient} of $X$.
We define the c.p.\ map $\Omega_\xi : A \to B$ by $\Omega_\xi (a)=\i< \xi, \pi_X(a ) \xi >$.
For a subset $S \subset X$ we denote by $\cF_S$ the convex hull of $\{ \Omega_\xi \mid \xi \in S \}$ in $\CP (A,B)$.
\end{definition}
\begin{definition}\label{def-int-tensor}
Let $X$ and $Y$ be Hilbert C$^*$-modules over $A$ and $B$, respectively, and $\varphi : A \to \lL_B (Y)$ be a c.p.\ map.
Then we can construct the Hilbert $B$-module $X\otimes_\varphi Y$ by separation and completion of $X\odot Y$ with respect to the $B$-valued semi-inner product (i.e., it satisfies the axiom of $B$-valued inner products except (5))
$
\i< \xi \otimes \eta, \xi' \otimes \eta' >:= \i< \eta, \varphi(\i< \xi, \xi'> ) \eta' >$ for $\xi,\xi'\in X$ and $\eta,\eta' \in Y.$
There are two $*$-homomorphisms:
\begin{align*}
\lL_A (X) \to \lL_B (X\otimes_\varphi Y); \quad x \mapsto x \otimes 1_Y\\
\varphi(A)'\cap \lL_B (Y) \to \lL_B (X\otimes_\varphi Y); \quad y \mapsto 1_X \otimes y
\end{align*}
satisfying that $(x\otimes 1_Y) (\xi \otimes \eta) =( x\xi )\otimes \eta$ and $(1_X\otimes y)( \xi \otimes \eta) =\xi \otimes (y \eta),$
for $\xi \in X$ and $\eta \in Y$.
Since these $*$-homomorphisms have mutually commuting ranges,
we will write $x\otimes y: =(x\otimes 1_Y) (1_X\otimes y)=(1_X \otimes y)(x\otimes 1_Y)$.
When $\varphi$ is a $*$-homomorphism,
the module $X\otimes_{\varphi} Y$ is called the {\it interior tensor product} of $X$ and $(Y, \varphi)$.
When no confusion may arise,
we may write $X\otimes_B Y=X\otimes_{\varphi} Y$.
Further assume that $Y=B$ and $\varphi :A \to B$ is surjective.
In this case, $X \otimes_\varphi B$ is called the {\it pushout} of $X$ by $\varphi$ and denoted by $X_\varphi$.
We also write $x_\varphi:=x \otimes 1_B $ for $x \in \lL_A (X)$.
\end{definition}
\begin{definition}
Let $X$ and $Y$ be Hilbert C$^*$-modules over $C$ and $D$, respectively.
The {\it exterior tensor product} of $X$ and $Y$ is the Hilbert $C\otimes D$-module given by separation and completion of $X \odot Y$ with respect to the $C \otimes D$-valued semi-inner product $\i< \xi \otimes \eta , \xi' \otimes \eta'> =\i< \xi, \xi'> \otimes \i< \eta, \eta '> \in C \otimes D$ for $\xi, \xi ' \in X$ and $\eta, \eta' \in Y$.
It is known that there exists a $*$-homomorphism $\iota :\lL_C (X) \otimes \lL_D ( Y) \to \lL_{C \otimes D} (X\otimes Y)$ such that $\iota (x \otimes y) (\xi \otimes \eta)= x\xi \otimes y \eta$ for $x \in \lL_C (X)$, $y \in \lL_D (Y)$, $\xi \in X$, $\eta \in Y$.
We will write $\iota (x\otimes y) = x\otimes y$ for short.
We note that when $(X, \pi_X) \in \Corr (A,C)$ and $(Y, \pi_Y) \in \Corr (B,D)$, we have $(X \otimes Y, \pi_X \otimes \pi_Y) \in \Corr (A \otimes B, C\otimes D)$.
\end{definition}
\begin{remark}
Let $X \in \Corr (A, B)$ and $Y \in \Corr (B, C)$ be given.
To simplify the notation, we use the same symbol $\pi_X \otimes 1_Y$ for the $*$-homomorphisms from $A$ into $\lL_C (X \otimes_B Y)$ and $\lL_{B \otimes C} (X \otimes Y)$.
\end{remark}
\begin{example}\label{ex-id}
Every C$^*$-algebra $A$ forms a Hilbert $A$-module with respect to the inner product $\i< a, b> =a^*b$.
It is not hard to see that $A\cong \lK_A (A)$.
Let $\lambda_A :A \to \lL_A (A)$ the canonical $*$-homomorphism given by the left multiplication.
The $(A, \lambda_A) \in \Corr (A)$ is called the {\it identity $\rC^*$-correspondence over} $A$.
\end{example}
\begin{example}
Let $\pi_H :A \to \lB (H)$ be a $*$-representation.
The C$^*$-correspondence $(H \otimes B, \pi_H \otimes 1_B)$ is called a {\it scalar representation}.
\end{example}
\begin{example}\label{ex-cond}
Let $B \subset A$ be an inclusion of C$^*$-algebras with a conditional expectation $E$ from $A$ onto $B$.
Then, $A$ naturally forms a right $B$-module by right multiplication.
We denote by $L^2 (A, E)$ the Hilbert $B$-module obtained from $A$ by separation and completion with respect to the $B$-valued semi-inner product $\i< a, b > := E(a^* b)$ for $a, b \in A$.
The left action $\pi_E : A \to \lL_B (L^2 (A,E))$ is given by the left multiplication.
When $A$ is unital, we denote by $\xi_E$ be the vector in $L^2(A, E)$ corresponding to $1_A$.
The triple $(L^2(A,E), \pi_E, \xi_E)$ is called {\it the GNS representation associated with} $E$.
The conditional expectation $E$ is said to be {\it nondegenerate} when $\pi_E$ is injective (or, equivalently, $a=0$ if and only if $E(xay)=0$ for all $x,y \in A$).
The conditional expectation $E$ is also said to be {\it faithful} when for any $a\in A$, $a=0$ if and only if $E(a^*a)=0$.
\end{example}
Let $\{\delta_i \}_{i=1}^n$ be the standard basis of $\lC^n$ and $\{ e_{ij} \}_{i,j=1}^n$ be the corresponding system of matrix units in $\lM_n$.
We denote by $\lC_n$ the Hilbert $\lM_n$-module $\lC^n$ equipped with the right action $\lC^n \times \lM_n \ni ( \xi, x) \mapsto {}^t x \xi \in \lC^n$, where ${}^t x$ is the transposed matrix of $x$, and the $\lM_n$-valued inner product defined by $\i<\delta_i, \delta_j>=e_{ij}$ for $1 \leq i,j \leq n$.
\begin{example}\label{ex-tensor}
We call $H_A:=\ell^2(\lN) \otimes A$ {\it the standard Hilbert module over} $A$.
Clearly, $H_A$ is isomorphic to the infinite direct sum $\bigoplus_{n=1}^\infty A$ of $A$ as Hilbert $A$-module.
For $(X, \pi_X) \in \Corr (A,B)$ and $n\in \lN$ we set
$(X^\infty, \pi_X^\infty) := (\ell^2(\lN) \otimes X, 1_{\ell^2(\lN)} \otimes \pi_X)$ and $(X^n, \pi_X^n):=(\lC^n \otimes X, 1_{\lC^n} \otimes \pi_X)$.
We also define $(X_n, \pi_{X_n}):= (X \otimes \lC_n, \pi_X \otimes 1_{\lC_n} ) \in \Corr (A, \lM_n (B) )$.
\end{example}
The next observation is standard, but important for us since it illustrates how c.p.\ maps that factors through matrix algebras (over C$^*$-algebras) appear.
\begin{remark}\label{rem-factor}
Let $(X, \pi_X) \in \Corr (A, B)$ and $(Y, \pi_Y) \in \Corr (B, C)$ be given.
For a given vector $\zeta \in X \otimes_B Y$ of the form $\sum_{i=1}^n \xi_i \otimes \eta_i$,
the coefficient $\Omega_\zeta : A \to C$ is equal to the composition of the c.p.\ maps $\varphi : A \to \lM_n (B)$ and $\psi : \lM_n (B) \to C$ defined by
\[
\varphi : a \mapsto \left[ \rule{0pt}{10pt} \i< \xi_i, \pi_X( a) \xi_j >_B \right]_{i,j=1}^n, \quad
\psi : \left[ \rule{0pt}{0pt}\, b_{ij}\, \right]_{i,j=1}^n \mapsto \sum_{i,j=1}^n \i< \eta_i, \pi_Y ( b_{ij} ) \eta_j >_C.
\]
\end{remark}
\renewcommand{\thetheorem}{\arabic{section}.\arabic{subsection}.\arabic{theorem}}
\section{Weak containment for C$^*$-correspondences}\label{sec-weak}
In this section we develop some general theory of weak containment for C$^*$-correspondences.
\subsection{Weak containment with respect to representations}\label{ss-weak-def}
\begin{definition}
Let $A$ be a C$^*$-algebra and $(H, \pi_H)$ and $(K, \pi_K)$ be $*$-representations of $A$.
We say that $(H,\pi_H)$ is {\it weakly contained in} $(K, \pi_K)$, written $(H, \pi_H) \prec (K,\pi_K)$ if $\ker \pi_K \subset \ker \pi_H$.
When no confusion may arise, we may write $H \prec K$ or $\pi_H \prec \pi_K$ for short.
\end{definition}
\begin{definition}\label{def-weak}
Let $A$ and $B$ be C$^*$-algebras.
For $(X,\pi_X) \in \Corr(A,B)$ and $(H, \pi_H) \in \Rep (B)$
we define the $*$-representation $\theta_X^H : A \otimes_{\rm max} \pi_H(B)' \to \lB (X\otimes_B H)$ by
\[
\theta_X^H (a\otimes x):= \pi_X(a)\otimes x, \quad a\in A, x\in \pi_H(B)'.
\]
We say that $(X, \pi_X )$ is {\it weakly contained in} $(Y,\pi_Y) \in \Corr(A,B)$ {\it with respect to} $(H,\pi_H)$, written $(X, \pi_X) \prec_{(H,\pi_H)} (Y,\pi_Y)$, if $(X\otimes_B H, \theta^H_X)$ is weakly contained in $(Y \otimes_B H, \theta^H_Y)$.
When no confusion may arise, we will write $X\prec_H Y$ for short.
In the case that $(H,\pi_H)$ is the universal representation of $B$, we write $X \prec_\univ Y$.
\end{definition}
\begin{remark}
The reader may think our definition of weak containment rather technical.
Hence, we will briefly explain why we formulated it as above.
Let $A$ and $B$ be C$^*$-algebras.
Fix $H \in \Rep (B)$ arbitrarily and set $M:=\pi_H(B)''$.
Then, weak containment for $A$-$B$ C$^*$-correspondences with respect to $H$ can be characterized by the one for corresponding $A$-$M$ bimodules in the following way.
Let $X, Y \in \Corr (A, B)$ be arbitrary.
Thanks to Lemma \ref{lem-normalrepn} below,
we can assume that $M \subset \lB (H)$ is of standard form (see, e.g., \cite{Haagerup}).
Then, the commutant $\pi_H(B)'$ is canonically isomorphic to the opposite algebra $M^\op$, and $\theta_X^H$ and $\theta_Y^H$ factor through the (right) normal tensor product $A \otimes_\nor M^\op$ (see \cite{Effros-Lance}).
Let $\rho_X$ and $\rho_Y$ be the representations of $A \otimes_\nor M^\op$ corresponding to $\theta_X^H$ and $\theta_Y^H$, respectively.
Then, it is clear that $\theta_X^H \prec \theta_Y^H$ if and only if $\rho_X \prec \rho_Y$.
We also note that this observation says that, in the case when $B=M$ and $X$ and $Y$ are selfdual,
our definition agrees with the one defined by Anantharaman-Delaroche and Havet \cite[Definition 1.7]{Delaroche-Havet}.
\end{remark}
\begin{remark}\label{rem-faithful}
Let $X, Y \in \Corr (A, B)$ and $(H, \pi) \in \Rep (B)$ be arbitrary.
Recall that the pushout of $X$ by $\pi$ is the Hilbert $\pi (B)$-module $X_\pi$.
Since $X \otimes_B H \cong X_\pi \otimes_{\pi(B)} H$, it follows that $X \prec_H Y$ if and only if $X_\pi \prec_H Y_\pi$ as $A$-$\pi(B)$ C$^*$-correspondences.
Thanks to this observation, we can reduce to the case when $(H, \pi_H)$ is faithful in some cases.
\end{remark}
\begin{lemma}[cf.\ {\cite[Lemma 3.8.4]{Brown-Ozawa}}]\label{lem-normalrepn}
Let $A$ and $B$ be $\rC^*$-algebras, and $X,Y \in \Corr(A,B)$ and $(H, \pi_H) \in \Rep(B)$ be given. Set $M=\pi_H(B)''$.
If $X \prec_H Y$ and $(K,\pi_K)$ is a normal representation of $M$,
then it follows that $X \prec_{(K,\pi_K\circ \pi_H)} Y$.
\end{lemma}
\begin{proof}
First we deal with the case that $K= H \otimes G$ and
$\pi_K : M \to \lB (H \otimes G); x\mapsto x\otimes 1_G$ for a Hilbert space $G$.
Fix $\sum_{k=1}^m a_k \otimes x_k \in A\odot \pi_K(M)'$ arbitrarily.
Let $P \in \lB (G)$ be an orthogonal projection of rank $n$.
Note that $\pi_K(M)' = M'\bar{\otimes}\lB (G)$ and $(1 \otimes P) \pi_K (M)' (1 \otimes P) \cong M' \otimes \lM_n$.
Let $\{ e_{ij} \}_{i,j=1}^n$ be a system of matrix units $\lM_n$ and $\sum_{i,j=1}^n x^{(k)}_{ij}\otimes e_{ij} \in M' \otimes \lM_n$ be the matrix representation of $(1\otimes P)x_k(1\otimes P)$ via the above isomorphism.
Since $\lB (X \otimes_B (H \otimes PG) ) \cong \lB (( X \otimes_B H) \otimes \lC^n ) \cong \lB (X \otimes_B H) \otimes\lM_n$, we have
\begin{align*}
& \left\| ( 1_X \otimes (1_H \otimes P) ) \sum_{k=1}^m \theta_X^{K}(a_k \otimes x_k) ( 1_X \otimes (1_H \otimes P) ) \right\|
_{\lB (X\otimes_B K)} \\
&\qquad =
\left\| \sum_{k=1}^m \theta_X^{K}(a_k \otimes ( (1_H \otimes P ) x_k ( 1_H \otimes P ) ) ) \right\|
_{\lB (X\otimes_B (H \otimes PG) )} \\
&\qquad =
\left\| \sum_{k=1}^m \sum_{i,j=1}^n \theta_X^H \left( a_k \otimes x_{ij}^{(k)} \right) \otimes e_{ij} \right\|
_{\lB (X\otimes_B H) \otimes \lM_n } \\
&\qquad =
\left\| (\theta_X^{H} \otimes \id_{\lM_n} ) \left( \sum_{k=1}^m \sum_{i,j=1}^n ( a_k \otimes x^{(k)}_{ij}) \otimes e_{ij} \right) \right\|
_{\lB (X\otimes_B H) \otimes \lM_n } \\
&\qquad \leq
\left\| (\theta_Y^{H } \otimes \id_{\lM_N} ) \left( \sum_{k=1}^m \sum_{i,j=1}^n ( a_k \otimes x^{(k)}_{ij}) \otimes e_{ij} \right) \right\|
_{\lB (Y \otimes_B H) \otimes \lM_n } \\
&\qquad \leq
\left\| \sum_{k=1}^m \theta_Y^{K}(a_k \otimes x_k) \right\|
_{\lB (Y \otimes_B K)}.
\end{align*}
Let $\{P_i \}_i \subset \lB (G)$ be a net of finite rank projections converges to $1_{G}$ strongly.
Then $\{1_X \otimes (1_H \otimes P_i) \}_i $ also converges to $1_{X\otimes K}$ strongly.
By the lower semi-continuity of operator norm we have
$\left\| \sum_{k=1}^m \theta_X^{K}(a_k \otimes x_k) \right\|
\leq
\left\| \sum_{k=1}^m \theta_Y^{K}(a_k \otimes x_k) \right\|$,
and hence we have $\theta_X^H \prec \theta_Y^H$.
Since every normal representation of $M$ is the cut-down of $\pi_K$ above by some projection in $\pi_K (M)'$, we are done.
\end{proof}
As we mentioned above, the following definition includes \cite[Definition 1.7]{Delaroche-Havet}.
\begin{definition}
Let $A$ be a C$^*$-algebra and $M$ be a von Neumann algebra.
For any two C$^*$-correspondences $X,Y \in \Corr (A,M)$ we say that $X$ is {\it weakly contained in} $Y$, written $X \prec Y$, if $X$ is weakly contained in $Y$ with respect to any (or some) faithful normal representations of $M$.
\end{definition}
\subsection{Characterization in terms of coefficients}\label{ss-weak-coe}
In this subsection, we prove Theorem \ref{thm-weak} below, which contains \cite[Proposition 2.3]{Delaroche-Havet} as a particular case that $B=M$ and $X$ is selfdual.
The proof below is based on the same idea as in Kirchberg's proof \cite{Kirchberg} for showing that C$^*$-nuclearity implies CPAP (see also \cite[Theorem 3.8.5]{Brown-Ozawa}).
\begin{theorem}\label{thm-weak}
Let $A$ and $B$ be $\rC^*$-algebras with $A$ unital.
Let $(X,\pi_X), (Y, \pi_Y) \in \Corr (A,B)$ and $(H, \pi_H ) \in \Rep (B)$ be given and
set $M:=\pi_H(B)''$.
Then, the following are equivalent:
\begin{itemize}
\item[$(1)$] $(X,\pi_X) \prec_{(H,\pi_H)} (Y,\pi_Y)$.
\item[$(2)$] For any $\xi \in X$ there exists a net of c.p.maps $\{\psi_i \}_i$ in $\cF_Y$ (see Definition \ref{def-coeff}) such that $\pi_H \circ \psi_i$ converges to $\pi_H \circ \Omega_\xi$ in the point $\sigma$-weak topology.
\item[$(3)$] For any $\xi \in X$ there exists a net of c.p.maps $\{\psi_i \}_i$ in $\cF_Y$ such that $\psi(1_A) \leq \Omega_\xi (1_A)$ and $\pi_H \circ \psi_i$ converges to $\pi_H \circ \Omega_\xi$ in the point $\sigma$-weak topology.
\item[$(4)$] $X\otimes_B M \prec Y \otimes_B M$.
\end{itemize}
\end{theorem}
The following technical lemmas originate in \cite[Lemma 2.2]{Delaroche-Havet} and are used to prove the implication (1) $\Rightarrow$ (2) in Theorem \ref{thm-weak}.
\begin{lemma}\label{prop-CS}
If $A$ and $B$ are unital $\rC^*$-algebras, $\varphi : A\to B$ is completely positive, and $f$ be a state on $B$,
then for any $a \in A$ and $b,c \in B$ it follows that
\
|f( b^*\varphi(a) c)| \leq \min \left\{ |f(b^* \varphi(1)b )|^{1/2} |f(c^*\varphi (a^*a)c )|^{1/2},
f(b^* \varphi(aa^*)b )|^{1/2} |f(c^*\varphi (1)c )|^{1/2} \right\}.
\]
\end{lemma}
\begin{proof}
Consider the $A$-$B$ C$^*$-correspondence $X = A\otimes_\varphi B$.
Then we have
$
f( b^*\varphi(a) c) = f (b^* \i< 1 \otimes 1, a\otimes 1 > c ) = f ( \i< 1 \otimes b, a\otimes c >).
$
Since $X\times X \ni (\xi, \eta) \mapsto f ( \i< \xi, \eta >) \in \lC$ defines a sesquilinear form, by the Cauchy--Schwarz inequality, we have
\[
|f ( \i< 1 \otimes b, a\otimes c >) |\leq |f( \i< 1\otimes b, 1 \otimes b >)|^{1/2} |f( \i< a \otimes c, a \otimes c >)|^{1/2} = | f(b^* \varphi(1)b ) |^{1/2} | f(c^* \varphi(a^*a)c ) |^{1/2}.
\]
Since $f ( \i< 1 \otimes b, a\otimes c >) = f ( \i< a^* \otimes b, 1\otimes c > )$ holds,
by the Cauchy--Schwartz inequality again, we get the desired inequality.
\end{proof}
\begin{lemma}
Let $A$ and $C$ be unital $\rC^*$-algebras and $\varphi : A \to C$ be a u.c.p.\ map.
Fix a faithful $*$-representation $C \subset \lB (H)$.
Let $\{ \phi_i \}_{i \in \cI}$ be a net in $\CP (A, C)$ which converges to $\varphi$ in the point $\sigma$-strong topology on $\CP (A, \lB (H))$ and set $c_i:= 2(1 + \phi_i (1_A))^{-1}$.
Then, $\phi'_i : A \to C; a \mapsto c_i \phi_i (a) c_i$ converges to $\varphi$ in the point $\sigma$-weak topology and satisfies that $\phi'_i (1_A) \leq 1_C$.
\end{lemma}
\begin{proof}
We first note that $\phi'_i (1_A) =4 \phi_i (1_A)(1 + \phi_i (1_A))^{-2} \leq 1$.
Let $\fF \subset A$ be a finite subset, $\cX \subset M_*$ be a finite subset of normal states,
and $\varepsilon>0$ be arbitrarily chosen.
By assumption, there exists $i_0 \in \cI$ such that
\[
| f ( \varphi (a) - \phi_i (a) ) | < \varepsilon /2, \quad 2 \| a \| |f ( (1- \phi_i (1_A))^2 ) |^{1/2} < \varepsilon/2, \quad | f (\phi_i (1_A)) - 1 | <1
\]
for all $a \in \fF$ and $f \in \cX$ as long as $i >i_0$.
By the previous lemma we have
\begin{align*}
| f (\phi'_i(a) - \phi_i(a) )|
&=
| f ( c_i^* \phi_i (a) c_i - \phi_i(a) ) | \\
&\leq
| f (c_i^* \phi_i(a) (c_i -1_C ) ) | + | f ((c_i^* -1_C ) \phi_i (a)) | \\
&\leq
|f(\phi_i (1_A) (1_C - c_i)^2)| ^{1/2}( | f (c_i^* \phi_i (aa^*) c_i)|^{1/2}+ |f (\phi_i (a^*a))|^{1/2}) \\
&\leq
|f(\phi_i (1_A) (1_C -c_i)^2)| ^{1/2} ( \| a \| + \|a \| f (\phi_i (1_A))^{1/2} )\\
&\leq 2 \| a \| |f(\phi_i (1_A) (1_C-c_i)^2)| ^{1/2} .
\end{align*}
Since $\phi_i (1_A) (1_C-c_i)^2=(1_C-\phi_i (1_A))^2 (1_C+\phi_i (1_A))^{-2} \phi_i (1_A) \leq (1_C-\phi_i (1_A))^2$,
we have $|f(\phi_i (1_A) (1_C-c_i)^2)| \leq | f ((1_C-\phi_i (1_A))^2) |.$
Hence, we get
$| f(\varphi (a) - \phi'_i(a)) | \leq | f (\varphi (a) - \phi_i (a) )| + | f (\phi_i(a) - \phi'_i (a)) | < | f (\varphi (a) - \phi_i (a) )| + 2 \| a \| |f ( (1_C- \phi_i (1_A))^2 ) |^{1/2} < \varepsilon$ as long as $i>i_0$.
\end{proof}
The next lemma specialized to the case that $B=B''$ is exactly \cite[Lemma 2.2]{Delaroche-Havet}.
\begin{lemma}\label{lem-cone}
Let $A$ and $B$ be $\rC^*$-algebras with $A$ unital.
Fix a nondegenerate faithful $*$-representation $B \subset \lB (H)$.
Let $\cF$ be a convex subset of $\CP(A,B)$ such that for any $b\in B$ and $\psi \in \cF$, the c.p.\ map $b^*\psi( \cdot )b$ also belongs to $\cF$.
If $\varphi \in \CP(A,B)$ belongs to the point $\sigma$-weak closure of $\cF$ in $\CP (A,\lB (H) )$,
then there exists a net $\{ \psi_i \}_i$ in $\cF$ such that $\psi_i(1_A) \leq \varphi(1_A)$ and it converges to $\varphi$ in the point $\sigma$-weak topology.
\end{lemma}
\begin{proof}
Set $b:=\varphi (1_A)$.
For each $n \in \lN$, we define the continuous function $f_n \in C_0 (0, \| b \| ]$ by
\[
f_n(x):=
\begin{cases}
n \sqrt{2x} & 0 \leq x \leq \frac{1}{2n},\\
(x + \frac{1}{2n} )^{-1/2} & \text{otherwise},
\end{cases}
\]
and set $b_n:= f_n (b) \in B$.
Let $e \in B''$ be the support projection of $b$ and set $C:=e B''e$.
Define $\theta_n \in \CP (A,B)$ by $\theta_n (a) := b_n \varphi (a) b_n$ for $a \in A$.
We first claim that $\theta_n$ converges to a u.c.p.\ map $\phi$ from $A$ into $C$ in the point $\sigma$-strong topology.
To see this, we fix a positive contraction $a \in A$ arbitrarily.
Since we have $0 \leq \varphi(a) \leq b$,
by the Douglas decomposition theorem (see e.g.\ \cite[Theorem 17.1]{Conway}),
there exists $c \in \lB (H)$ such that $\varphi(a)^{1/2} =c b^{1/2}$.
Since $\{ b_n b^{1/2} \}_n$ is an increasing sequence which converges $e$ strongly,
$\theta_n (a) = b_n b^{1/2} c^*c b^{1/2} b_n$ also converges to $e c^*c e$ in the same topology.
In the case when $a = 1_A$, the $c$ is equal to $e$, and hence $\phi (a):= e c^* c e$ defines the desired u.c.p.\ map.
We note that $\varphi = b^{1/2}\phi(\cdot) b^{1/2}$ holds.
Indeed, for any $a \in A$ we have
\[\varphi (a)=e \varphi (a) e =\lim_n b^{1/2}b_n \varphi (a) b_n b^{1/2} = \lim_n b^{1/2} \theta_n (a) b^{1/2} = b^{1/2}\phi (a) b^{1/2}.
\]
Thus, if we find a net $\phi_i'$ in $\cF$ in such a way that $\phi_i' (1_A) \leq \phi (1_A)=e$ and $\phi_i' (a)$ converges to $\phi (a)$ $\sigma$-weakly for $a \in A$, then $\psi_i:= b^{1/2} \phi_i'(\cdot) b^{1/2}$ gives the desired net.
Since $\cF$ is convex, the point $\sigma$-weak closure of $\cF$ in $\CP (A, \lB (H) )$ coincides with the point $\sigma$-strong closure of $\cF$.
This follows from the fact that for any finitely many elements $a_1, \dots, a_n \in A$ the set $\{ (\psi (a_1), \dots, \psi (a_n) )\in \bigoplus_{i=1}^n \lB (H) \mid \psi \in \cF \}$ is convex, and hence its $\sigma$-weak and $\sigma$-strong closures coincide.
Thus,
by the claim above, we can find nets $\varphi_i \in \cF$ and $n (i) \in \lN$ in such a way that $\phi_i:= b_{n(i)} \varphi_i (\cdot) b_{n(i)}$ converges to $\phi$ point $\sigma$-strongly.
The image of each $\phi_i$ is contained in $B \cap C$ since $b_n = e b_n$ holds for every $n \in \lN$.
By the preceding lemma, the net $\phi_i' =c_i \phi_i (\cdot) c_i $ with $c_i= 2 ( e+ \phi_i (1_A) )^{-1} \in C$ converges to $\phi$ in the point $\sigma$-weak topology and satisfies that $\phi_i' (1_A) \leq e$.
We show that $\phi_i'$ belongs to $\cF$.
Indeed, we have $ \phi'_i = c_i \phi_i (\cdot) c_i = c_i b_{n(i)} \varphi_i (\cdot) b_{n(i)}c_i $.
By the fact that $c_i$ is the norm limit of commutative polynomials of $e$ and $b_{n(i)} \varphi_i (1_A) b_{n(i)}$ and $b_{n(i)}e=b_{n(i)}$,
each $c_i b_{n(i)}$ is in $B$.
By the assumption of $\cF$, we get $\phi_i' \in \cF$.
\end{proof}
\if0
\begin{lemma}
Let $A$ and $B$ be unital $\rC^*$-algebras and $\varphi : A \to B$ be a u.c.p.\ map.
Fix a faithful $*$-representation $B \subset \lB (H)$.
Let $\{ \psi_i \}_{i \in \cI}$ be a net in $\CP (A, B)$ which converges to $\varphi$ in the point $\sigma$-strong topology on $\CP (A, \lB (H))$ and set $b_i:= 2(1 + \psi_i (1_A))^{-1}$.
Then, $\widehat{\psi}_i : A \to B; a \mapsto b_i \psi_i (a) b_i$ converges to $\varphi$ in the point $\sigma$-weak topology and satisfies that $\widehat{\psi}_i (1_A) \leq 1_B$.
\end{lemma}
\begin{proof}
We first note that $\widehat{\psi}_i (1_A) =4 \psi_i (1_A)(1 + \psi_i (1_A))^{-2} \leq 1$.
Let $\fF \subset A$ be a finite subset, $\cX \subset M_*$ be a finite subset of normal states,
and $\varepsilon>0$ be arbitrarily chosen.
By assumption, there exists $i_0 \in \cI$ such that
\[
| f ( \varphi (a) - \psi_i (a) ) | < \varepsilon /2, \quad 2 \| a \| |f ( (1- \psi_i (1))^2 ) |^{1/2} < \varepsilon/2, \quad | f (\psi_i (1)) - 1 | <1
\]
for all $a \in \fF$ and $f \in \cX$ as long as $i >i_0$.
By the previous lemma we have
\begin{align*}
| f (\widehat{\psi}_i(a) - \psi_i(a) )|
&=
| f ( b_i^* \psi_i (a) b_i - \psi_i(a) ) | \\
&\leq
| f (b_i^* \psi_i(a) (b_i -1 ) ) | + | f ((b_i^* -1 ) \psi_i (a)) | \\
&\leq
|f(\psi_i (1) (1-b_i)^2)| ^{1/2}( | f (b_i^* \psi_i (aa^*) b_i)|^{1/2}+ |f (\psi_i (a^*a))|^{1/2}) \\
&\leq
|f(\psi_i (1) (1-b_i)^2)| ^{1/2} ( \| a \| + \|a \| f (\psi_i (1))^{1/2} )\\
&\leq 2 \| a \| |f(\psi_i (1) (1-b_i)^2)| ^{1/2} .
\end{align*}
Since $\psi_i (1) (1-b_i)^2=(1-\psi_i (1))^2 (1+\psi_i (1))^{-2} \psi_i (1) \leq (1-\psi_i (1))^2$,
we have $|f(\psi_i (1) (1-b_i)^2)| \leq | f ((1-\psi_i (1))^2) |.$
Hence, we get
$| f(\varphi (a) - \widehat{\psi}_i(a)) | \leq | f (\varphi (a) - \psi_i (a) )| + | f (\psi_i(a) - \widehat{\psi}_i (a)) | < | f (\varphi (a) - \psi_i (a) )| + 2 \| a \| |f ( (1- \psi_i (1))^2 ) |^{1/2} < \varepsilon$ as long as $i>i_0$.
\end{proof}
The next lemma specialized to the case that $B=B''$ is exactly \cite[Lemma 2.2]{Delaroche-Havet}.
\begin{lemma}\label{lem-cone}
Let $A$ and $B$ be $\rC^*$-algebras with $A$ unital.
Fix a nondegenerate faithful $*$-representation $B \subset \lB (H)$.
Let $\cF$ be a convex subset of $\CP(A,B)$ such that for any $b\in B$ and $\psi \in \cF$, the c.p.\ map $b^*\psi( \cdot )b$ also belongs to $\cF$.
If $\varphi \in \CP(A,B)$ belongs to the point $\sigma$-weak closure of $\cF$ in $\CP (A,\lB (H) )$,
then there exists a net $\{ \varphi_i \}_i \subset \cF$ such that $\varphi_i(1) \leq \varphi(1)$ and it converges to $\varphi$ in the point $\sigma$-weak topology.
\end{lemma}
\begin{proof}
Since $\cF$ is convex, the point $\sigma$-weak closure $\overline{\cF}$ of $\cF$ in $\CP (A, \lB (H) )$ coincides with the point $\sigma$-strong closure of $\cF$.
This follows from the fact that for any finitely many elements $a_1, \dots, a_n \in A$ the set $\{ (\psi (a_1), \dots, \psi (a_n) )\in \bigoplus_{i=1}^n \lB (H) \mid \psi \in \cF \}$ is convex, and hence its $\sigma$-weak and $\sigma$-strong closures coincide.
Set $b:=\varphi(1)$ and let $e \in B''$ the support projection of $b$.
Define $\theta_n \in \CP (A,B'')$ by
$$
\theta_n (a) := (b+1/n)^{-1/2} \varphi (a) (b+1/n)^{-1/2}, \quad a \in A.
$$
For each $n \geq 1$, by the Kaplansky density theorem, we can find a bounded net $\{b_\lambda^{(n)} \}_\lambda$ of positive elements in $B$ which converges to $(b+1/n)^{-1/2}$ $\sigma$-strongly.
Since $b_\lambda^{(n)} \varphi (\cdot ) b_\lambda^{(n)}$ belongs to $\overline{\cF}$, so does $\theta_n$.
We claim that there exists $\theta \in \UCP (A, eB''e)$ such that $\theta_n$ converges to $\theta$ point $\sigma$-strongly in $\CP (A, \lB (H) )$.
To see this, we fix a positive contraction $a \in A$.
We then have $0 \leq \varphi (a) \leq b$.
By the Douglas decomposition theorem (see e.g. \cite[Theorem 17.1]{Conway}),
there exists $c \in \lB (H)$ such that $\varphi(a)^{1/2} =c b^{1/2}$.
We observe that $(b+1/n)^{-1/2}b^{1/2}$ is a bounded increasing net which converges to $e$ strongly.
This implies that $\theta_n(a) =(b+1/n)^{-1/2} \varphi (a)^{1/2}\varphi (a)^{1/2} (b+1/n)^{-1/2} =(b+1/n)^{-1/2}b^{1/2}c^*c b^{1/2}(b+1/n)^{-1/2}$ also converges to $ec^*ce$ strongly.
Because $A$ is spanned by its positive contractions, $\theta_n$ converges to a c.p.\ map $\theta$ in $\CP (A, eB''e)$.
Moreover, we have $\theta (1)= \lim_n b (b + 1/n)^{-1} =e$.
Thus, this $\theta$ is the desired one.
We next claim that there exists a net $\{ \phi_i \}_i $ in $\cF$ such that $b^{1/2}\phi_i ( \cdot ) b^{1/2}$ converges to $\theta$ point $\sigma$-strongly in $\CP (A, eB''e)$.
Take a finite subset $\fF \subset A$, $\varepsilon>0$, and a finite subset of normal states $\cX \subset \lB (H)_*$.
By the convexity of $\cF$ and the fact that $\theta = e \theta ( \cdot )e$ we can find $\psi \in \cF$
such that $\| \theta (a) - e \psi (a) e \|_f < \varepsilon$ for $a \in \fF$ and $f \in \cX$, where $\| x \|_f = f(x^*x)$.
Since $e=\lim_n b^{1/2} (b + 1/n)^{-1/2}$ and $(b + 1/n)^{-1/2}=\lim_\lambda b_\lambda^{(n)}$, there exists $n \in \lN$ and $\lambda$ such that $\| \theta (a) - b^{1/2}b_\lambda^{(n)} \psi (a)b_\lambda^{(n)}b^{1/2} \|_f < \varepsilon$ for $a \in \fF$ and $f \in \cX$.
The map $\phi:=b_\lambda^{(n)} \psi ( \cdot )b_\lambda^{(n)}$ forms the desired net.
We now apply the preceding lemma to the net $\phi_i':= b^{1/2}\phi_i ( \cdot ) b^{1/2} \in \CP (A, eB''e)$ and obtain a net $\{ \widehat{\phi_i'} \}_i$ which converges to $\theta$ point $\sigma$-weakly and satisfies that $\widehat{\phi_i'} (1) \leq e$.
Since $\varphi (a)=e \varphi (a) e \leftarrow b^{1/2}(b + 1/n)^{-1/2} \varphi (a) (b + 1/n)^{-1/2} b^{1/2} = b^{1/2} \theta_n (a) b^{1/2} \to b^{1/2}\theta (a) b^{1/2}$, the proof will be complete
if $b^{1/2}\widehat{\phi'_i }(\cdot) b^{1/2}$ belongs to $\cF$.
Indeed,
one has $b^{1/2}\widehat{\phi'_i }( \cdot )b^{1/2} = 4b^{1/2}c_i b^{1/2} \phi_i (\cdot)b^{1/2} c_i b^{1/2}$,
where $c_i$ is the inverse element of $e + b^{1/2}\phi_i (1)b^{1/2} $ in $eB''e$.
By the assumption of $\cF$ it suffices to show that $b^{1/2}c_i$ is in $B$.
This follows from the fact that $c_i$ is the norm limit of commutative polynomials of $e$ and $b^{1/2}\phi_i (1)b^{1/2}$ and that $b^{1/2}e=b^{1/2}$.
\end{proof}
\fi
\begin{proof}[Proof of Theorem \ref{thm-weak}]
As in Remark \ref{rem-faithful}, replacing $B$ and $X, Y$ by $\pi_H (B)$ and $X_{\pi_H}, Y_{\pi_Y}$ we may assume that $\pi_H$ is faithful and identify $B$ with $\pi_H (B)$.
We prove $(1) \Rightarrow (2):$
Fix $\xi \in X$, a finite subset $\fF \subset A$, a finite subset $\cX$ of normal states on $M$ and $\varepsilon >0$ arbitrarily.
Set $f:= | \cX |^{-1} \sum_{g \in \cX} g$ and let $(K, \pi_f, \xi_f)$ be the GNS-representation associated with $f$.
Since each $g \in \cX$ enjoys $g \leq |\cX| f$,
by the Radon--Nikodym theorem for states (see e.g. \cite[Proposition 3.8.3]{Brown-Ozawa}), there exists $x_g \in \pi_f(M)' = \pi_f (B)'$ such that $g (y) = \i< \xi_f, \pi_f (y) x_g \xi_f >$ for $y\in M$.
By Lemma \ref{lem-normalrepn} we have $X \prec_K Y$, i.e., $\theta_X^K \prec \theta_Y^K$.
Since the image of $Y \odot \xi_f$ is dense in $Y \otimes_B K$, applying Fell's characterization of weak containment \cite[Theorem 1.2]{Fell} to these $*$-representations of $A \otimes_{\max} \pi_f (B)'$ and the vector $\xi \otimes \xi_f \in X \otimes_B K$,
we find $\eta_1, \dots, \eta_n \in Y$ such that $| \i< \xi \otimes \xi_f, \theta_X^K (a \otimes x_g ) (\xi \otimes \xi_f) > - \sum_{i=1}^n \i< \eta_i \otimes \xi_f, \theta_Y^K (a \otimes x_g ) (\eta_i \otimes \xi_f)> | < \varepsilon$ for all $a\in \fF$ and $g \in \cX$.
We then have for $a\in \fF$ and $g \in \cX$
\begin{align*}
g \left( \sum_{i=1}^n \i< \eta_i, \pi_Y (a) \eta_i >_B \right)
&=\sum_{i=1}^n \i< \xi_f, \pi_f ( \i< \eta_i, \pi_Y (a) \eta_i >_B ) x_g \xi_f > \\
&=\sum_{i=1}^n \i< \eta_i \otimes \xi_f, \pi_Y (a) \eta_i \otimes x_g \xi_f > \\
&=\sum_{i=1}^n \i< \eta_i \otimes \xi_f, \theta_Y^K (a \otimes x_g ) ( \eta_i \otimes \xi_f) >.
\end{align*}
Since $g ( \Omega_\xi (a) ) = \i< \xi_f, \pi_f ( \i< \xi, \pi_X(a) \xi >_B ) x_g \xi_f > = \i< \xi \otimes \xi_f, \theta_X^K (a \otimes x_g )(\xi \otimes \xi_f) >$,
the c.p.\ map $\psi:=\sum_{k=1}^m \Omega_{\zeta_k} \in \cF_Y$ satisfies that $| g (\Omega_\xi (a) - \psi (a)) | < \varepsilon$ for $a \in \fF$ and $g \in \cX$.
The implication (2) $\Rightarrow$ (3) follows from Lemma \ref{lem-cone}.
We prove (3) $\Rightarrow$ (1):
Let $z \in \ker \theta^H_Y$ be given and show that $\theta^H_X (z)=0$.
It suffices to prove that $\theta^H_X (z) \xi \otimes \eta=0$ for every $\xi \in X$ and $\eta \in H$.
We fix $\varepsilon >0$ arbitrarily and take $w=\sum_{i=1}^n a_i \otimes x_i \in A \odot \pi_H(B)'$ in such a way that $\| z - w \|_{\rm max} < \varepsilon $.
By (3) we can choose $\psi = \sum_{k=1}^m \Omega_{\zeta_k} \in \cF_Y$ in such a way that $\psi (1) \leq \Omega_\xi (1)$ and $|\i< \eta, \sum_{i,j=1}^n \Omega_\xi (a_i^* a_j) x_i^* x_j \eta > -
\i< \eta, \sum_{i,j=1}^n \psi (a_i^* a_j) x_i^* x_j \eta >| < \varepsilon^2$.
Note that $ \sum_{k=1}^m \| \zeta_k \otimes \eta \|^2 = \i< \eta, \psi(1) \eta > \leq \i< \eta, \Omega_\xi (1) \eta > \leq \| \xi \otimes \eta \|^2$.
One has
\begin{align*}
&\| \sum_{i=1}^n \theta^H_X (a_i \otimes x_i) \xi \otimes \eta \|^2
=
\i< \eta, \sum_{i,j=1}^n\Omega_\xi (a_i^* a_j) x_i^* x_j \eta >
\approx_{\varepsilon^2}
\i< \eta, \sum_{i,j=1}^n \sum_{k=1}^m \Omega_{\zeta_k} (a_i^* a_j) x_i^* x_j \eta >\\
=
\sum_{k=1}^m \| \sum_{i=1}^n \theta^H_Y (a_i \otimes x_i) \zeta_k \otimes \eta \|^2
\leq
\varepsilon^2 \sum_{k=1}^m \| \zeta_k \otimes \eta \|^2
\leq
\varepsilon^2 \| \xi\otimes \eta \|^2.
\end{align*}
Thus, we have $\| \theta^H_X (z)(\xi \otimes \eta ) \| \leq \| z - w \|_{\rm max} + \| \theta^H_X (w) ( \xi \otimes \eta )\| \leq \varepsilon + \sqrt{2}\varepsilon$.
Since $\varepsilon$ is arbitrary, we get $\theta^H_X (z)=0$.
Finally, the canonical isomorphisms $(Y \otimes_B M) \otimes_M H \cong Y \otimes_B H$ and $(X \otimes_B M) \otimes_M H \cong X\otimes_B H$ imply that $\ker \theta_Y^H = \ker \theta_{Y\otimes_B M}^H $ and $\ker \theta_{X}^H= \ker \theta_{X \otimes_B M}^H$,
which proves $(1) \Leftrightarrow (4)$.
\end{proof}
\if0
\begin{remark}
The weak containments can be described in terms of a topology on (the unitary equivalence classes of) C$^*$-correspondences.
Let $X \in \Corr (A, B)$ and $H \in \Rep (B)$.
For finite subsets $\fF \subset A$ and $\cX = \{ \xi_1, \dots, \xi_n \} \subset X$ and a $\sigma$-weak open neighborhood $V$ of $0$ in $M:=\pi_H (B)''$ we define $\cV_H (X; \fF, \cX, V)$ by the set of $A$-$B$ C$^*$-correspondences $Y$ such that there exists $\eta_1, \dots, \eta_n \in Y$ such that $\pi_H (\i< \xi_i, \pi_X (a) \xi_j > - \i< \eta_i, \pi_Y (a) \eta_j > ) \in V$ for all $a\in \fF$ and $1\leq i,j \leq n$.
It is not hard to see that for each $X \in \Corr (A, B)$ the set $\cU_H (X):=\{ \cV (X; \fF, \cX, V) \mid \fF \subset A, \cX \subset X, 0 \in V \subset M: \text{open} \}$ forms a basis of neighborhoods of $X$.
We note that the topology $\cU_H$ on $\Corr (A, B)$ given by $\{ \cU_H (X) \}_X$ induces a topology on the set of unitarily equivalent classes of $A$-$B$ C$^*$-correspondences since $(X,\pi_X) \cong (Y,\pi_Y)$ implies $\cU_H (X)= \cU_H (Y)$.
\end{remark}
\fi
\begin{corollary}\label{cor-weak}
Let $A$ and $B$ be $\rC^*$-algebras with $A$ unital and $X,Y \in \Corr(A,B)$ be given.
The following are equivalent:
\begin{itemize}
\item[$(1)$] $X \prec_\univ Y$.
\item[$(2)$] For any $\xi \in X$ the c.p.\ map $\Omega_\xi$ belongs to the point norm closure of $\cF_Y$.
\item[$(3)$] For any $\xi \in X$ there exists a net $\{ \psi_i \}_i$ in $\cF_Y$ such that $\psi_i(1_A) \leq \Omega_\xi (1_A)$ and that $\lim_i \| \Omega_\xi (a) - \psi_i (a) \| =0$ for every $a\in A$.
\end{itemize}
\end{corollary}
\begin{proof}
Let $(H_u, \pi_u)$ be the universal representation of $B$.
Then, it is known that the enveloping von Neumann algebra $\pi_u (B)''$ is isomorphic to the second dual $B^{**}$.
Hence, the relative topology on $\pi_u(B)$ induced from the $\sigma$-weak topology on $\pi_u (B)''$ coincides with the weak topology.
Since $\cF_Y$ is convex, the point weak closure and the point norm closure of $\cF$ coincide (see e.g.\ \cite[Lemma 2.3.4]{Brown-Ozawa}).
Hence, the assertion follows from the previous theorem.
\end{proof}
\subsection{Elementary properties of weak containment}\label{ss-weak-prop}
In this subsection we establish some basic facts on weak containment.
\if0
Since the next proposition follows easily from Theorem \ref{thm-weak}, we omit its proof.
\begin{proposition}\label{prop-weak}
Let $A$ and $B$ be $\rC^*$-algebras with $A$ unital.
For $X,Y,Z \in \Corr (A , B)$ and $H \in \Rep (B)$ the following hold true{\rm :}
\begin{itemize}
\item[(1)] $X \prec_H Y$ and $Y \prec_H Z \Rightarrow X \prec_H Z$,
\item[(2)] $(X^\infty, \pi_X^\infty) \prec_\univ (X, \pi_X) \prec_\univ (X^\infty, \pi_X^\infty)$.
\item[(3)] Suppose that there exists a cyclic vector $\xi \in X$, i.e., $\lspan \pi_X (A) \xi B$ is norm dense in $X$.
Then $X \prec_H Y$ if and only if $\pi_H\circ \Omega_\xi$ belongs to the point $\sigma$-weak closure of $\{ \pi_H \circ \psi \mid \psi \in \cF_Y \}$.
\end{itemize}
\end{proposition}
\fi
For a Hilbert C$^*$-module over a unital C$^*$-algebra $B$, a vector $\xi \in X$ is said to be {\it normal} if $\i< \xi, \xi > =1_B$ holds.
We note that every C$^*$-correspondence $(A \otimes_\varphi B, \lambda_A \otimes 1_B)$ arising from a u.c.p.\ map $\varphi : A \to B$ admits the normal vector $1_A \otimes 1_B$.
\begin{proposition}\label{prop-unital}
Let $A$ and $B$ be unital $\rC^*$-algebras and $X,Y \in \Corr (A, B)$ and $H \in \Rep (B)$ be given.
Assume that $Y$ is unital and admits a normal vector $\eta \in Y$.
If $X \prec_K Y$ (resp. $X \prec_\univ Y$),
then for any $\xi \in X$ there exists $\psi_i \in \cF_Y$ with $\psi_i (1_A) =\Omega_\xi(1_A)$ such that $\pi_K \circ \psi_i $ converges to $\pi_K \circ \Omega_\xi$ point $\sigma$-weakly (resp. $\psi_i$ converges to $\Omega_\xi$ in the point norm topology).
\end{proposition}
\begin{proof}
By Theorem \ref{thm-weak}, there exists a net $\psi_i \in \cF_Y$ with $\psi_i (1_A) \leq \Omega_\xi (1_A)$ such that $\pi_K \circ \psi_i$ approaches $\pi_K \circ \Omega_\xi$.
Set $b_i:=\Omega_\xi (1_A) - \psi_i (1_A) \in B$ and $\zeta_i:=\eta b_i^{1/2}$.
For $a\in A$ and normal state $f \in \lB (K)_*$ we have
$f( \Omega_{\zeta_i} (a) )= f (b_i^{1/2}\i< \eta, \pi_Y(a) \eta >b_i^{1/2} ) \leq \| a \| f (b_i) \to 0$ and $\psi_i (1_A) + \Omega_{\zeta_i} (1_A)=\Omega_\xi(1_A)$.
Thus $\psi_i + \Omega_{\zeta_i} \in \cF_Y$ is the desired one.
The assertion in the case that $X \prec_\univ Y$ follows from Corollary \ref{cor-weak}.
\end{proof}
Let $A$ and $B$ be C$^*$-algebras with $A$ unital, and let $X, Y, Z \in \Corr (A, B)$ and $H \in \Rep (B)$.
We note that $X \prec_H Z$ holds whenever $X \prec_H Y$ and $Y \prec_H Z$ hold thanks to Theorem \ref{thm-weak}.
\begin{proposition}\label{prop-tensor}
Let $A, B, C$ and $D$ be $\rC^*$-algebras with $A$ and $C$ unital.
For $X, Y \in \Corr (A, B)$, $Z, W \in \Corr (C, D)$, $H \in \Rep (B)$ and $K \in \Rep (D)$ if $X \prec_H Y$ and $Z \prec_K W$ hold,
then we have $X \otimes Z \prec_{H \otimes K} Y \otimes W$ as $(A\otimes C)$-$(B\otimes D)$ $\rC^*$-correspondences.
\end{proposition}
\begin{proof}
By the remark above, it suffices to show the case when $(Z, \pi_Z)=(W, \pi_W)$.
We show that $\ker \theta_{Y \otimes Z}^{H \otimes H} \subset \ker \theta_{X \otimes Z}^{H \otimes K}$.
Fix $z \in \ker \theta_{Y \otimes Z}^{H \otimes H}$ with $\| z \| \leq 1$ arbitrarily.
By the polarization trick, we only have to show that $\i<\xi \otimes \zeta \otimes h \otimes k, \theta_{X \otimes Z}^{H \otimes H} (z) \xi \otimes \zeta \otimes h \otimes k > =0$ for all unit vectors $\xi \in X$, $\zeta \in Z$, $h \in H$, and $k \in K$.
For any $\varepsilon >0$, we can find a contraction $z_\varepsilon=\sum_l (a_l \otimes c_i ) \otimes x_l \in A \odot C \odot (\pi_H(B)' \botimes \pi_K (D)' )$ satisfying
$\| z - \sum_l (a_l \otimes c_i ) \otimes x_l \| < \varepsilon$ in $( A \otimes C) \otimes_{\max} ( \pi_H(B)' \botimes \pi_K (D)' )$.
Note that $\| \theta_{Y \otimes Z}^{H \otimes K} ( z_\varepsilon ) \| < \varepsilon$.
Since $A$ is unital and $X \prec_H Y$ holds, Theorem \ref{thm-weak} enables us to find $\eta_1, \dots, \eta_p \in Y$ in such a way that $\sum_{r=1}^p \i< \eta_r, \eta_r > \leq \i< \xi, \xi>$ and
\[
\left|
\sum_l \l< h \otimes k, x_l ( \pi_H ( \Omega_\xi (a_l) - \sum_{r=1}^p \Omega_{\eta_r} (a_l)) h ) \otimes (\pi_K ( \Omega_\zeta (c_l) ) k ) >
\right|
< \varepsilon.
\]
Now we have
\begin{align*}
& | \i<\xi \otimes \zeta \otimes h \otimes k, \theta_{X \otimes Z}^{H \otimes H} (z) \xi \otimes \zeta \otimes h \otimes k > |
\\ &\qquad \approx_{\varepsilon}
| \i<\xi \otimes \zeta \otimes h \otimes k, \theta_{X \otimes Z}^{H \otimes H} (z_\varepsilon) \xi \otimes \zeta \otimes h \otimes k >
\\ &\qquad =_{\phantom{\varepsilon}}
|\sum_l \i< h \otimes k, x_l ( \pi_H ( \Omega_\xi (a_l) ) h ) \otimes (\pi_K ( \Omega_\zeta (c_l) ) k ) > |
\\ &\qquad \approx_\varepsilon
| \sum_{r=1}^p \sum_l \i< h \otimes k, x_l ( \pi_H ( \Omega_{\eta_r} (a_l) ) h ) \otimes (\pi_K ( \Omega_\zeta (c_l) ) k ) > |
\\ &\qquad =_{\phantom{\varepsilon}}
|\sum_{r=1}^p \i< \eta_r \otimes \zeta \otimes h \otimes k, \theta_{Y \otimes Z}^{H \otimes K} ( z_\varepsilon ) \eta_r \otimes \zeta \otimes h \otimes k >
\\ &\qquad \leq_{\phantom{\varepsilon}}
\| \theta_{Y \otimes Z}^{H \otimes K} (z_\varepsilon ) \| \i< h \otimes k, \sum_{r=1}^p \i< \eta_r, \eta_r > h \otimes \i< \zeta, \zeta > k >
\\ &\qquad \leq _{\phantom{\varepsilon}}
\varepsilon \i< h \otimes k, \i< \xi, \xi> h \otimes \i< \zeta, \zeta > k >
\\ &\qquad \leq _{\phantom{\varepsilon}}
\varepsilon.
\end{align*}
Since $\varepsilon$ is arbitrary, we are done.
\end{proof}
The next useful proposition is a particular case of Proposition \ref{prop-tensor}.
\begin{proposition}\label{lem-tensor}
Let $A,B$ be $\rC^*$-algebras with $A$ unital and $X, Y \in \Cor (A, B)$ and $H \in \Rep (B)$.
If $X \prec_H Y$, then $(X_n, \pi_{X_n} ) \prec_{(H^n, \pi_H^{(n)})} (Y_n, \pi_{Y_n} )$ as $A$-$\lM_n(B)$ $\rC^*$-correspondences (see Example \ref{ex-tensor} for notations).
\end{proposition}
\begin{proposition}\label{prop-int-tensor}
Let $A, B$ and $C$ be $\rC^*$-algebras with $A$ and $B$ unital and $X, Y \in \Corr (A, B)$, $Z, W \in \Corr (B, C)$ and $K \in \Rep (C)$ be given.
Suppose that $Z\prec_K W$ and either $X \prec_{Z \otimes_C K} Y$ or $X \prec_{W \otimes_C K} Y$ holds.
Then $X \otimes_B Z \prec_{K} Y \otimes_B W$ as $A$-$C$ $\rC^*$-correspondences.
\end{proposition}
\begin{proof}
Suppose that $X \prec_{Z\otimes_C K} Y$.
We show that $X\otimes_B Z \prec_K Y \otimes_B Z \prec_K Y \otimes_B W$.
For any $\sum_i a_i \otimes x_i \in A \odot \pi_K(C)'$ we have
$\| \theta_{X\otimes_B Z}^K (\sum_i a _i \otimes x_i) \| = \| \theta_X^{Z\otimes_C K} (\sum_i a_i \otimes (1_Z \otimes x_i ) )\| \leq \|\theta_Y^{Z\otimes_C K} (\sum_i a_i \otimes (1_Z \otimes x_i ) )\| =\| \theta_{Y\otimes_B Z}^K (\sum_i a _i \otimes x_i) \|$.
Hence $X\otimes_B Z \prec_K Y \otimes_B Z$.
To see $Y \otimes_B Z \prec_K Y \otimes_B W$, let $\xi= \sum_{i=1}^n \eta_i \otimes \zeta_i \in Y \odot Z$ be arbitrary.
By Remark \ref{rem-factor},
the coefficient $\Omega_\xi$ is of the form $\Omega_\zeta \circ \Omega_\eta$ with $\eta=(\eta_1, \dots, \eta_n ) \in Y_n$ and $\zeta=(\zeta_1, \dots, \zeta_n ) \in Z^n$, where $\lM_n (B)$ acts on $Z^n$ by $\pi_Z \otimes \id_{\lM_n} : \lM_n (B) \to \lL_C (Z) \otimes \lM_n = \lL_ C( Z^n)$.
Since $(Z^n, \pi_Z \otimes \id_{\lM_n}) \prec_H (W^n, \pi_{W}\otimes \id_{\lM_n} )$ as $\lM_n (B)$-$C$ C$^*$-correspondences,
there exists a net $\psi_i \in \cF_{Z^n}$ such that $\pi_K \circ \psi_i \circ \Omega_\eta $ converges to $\pi_K \circ \Omega_\zeta \circ \Omega_\eta= \pi_K \circ \Omega_\xi$ in the point $\sigma$-weak topology.
Since $\psi_i \circ \Omega_\eta \in \cF_{X \otimes_B Z}$, we get $Y\otimes_B Z \prec_K Y \otimes_B W$.
When $X \prec_{W \otimes_C K} Y$, we can prove, in the same manner, that $X\otimes_B Z \prec_K X \otimes_B W \prec_K Y \otimes_B W$
\end{proof}
\begin{proposition}\label{prop-product-map}
Let $A$ and $B$ be $\rC^*$-algebras with $A$ unital.
Let $X, Y \in \Corr (A, B)$ and $H \in \Rep (B)$ be arbitrary and set $M=\pi_H (B)''$.
Then, $X \prec_H Y$ if and only if for any $\xi \in X$, the product map
$$
(\pi_H \circ \Omega_\xi ) \us \times_{\max} \iota_{M'} : A \us\otimes_{\max} M' \to \lB (H); \quad a \otimes x \mapsto \pi_H (\Omega_\xi (a) ) x
$$
factors through $\Imag \theta_Y^H$.
\end{proposition}
\begin{proof}
Thanks to Remark \ref{rem-faithful} we assume that $\pi_H$ is faithful and $B \subset \lB (H)$.
Suppose that $X \prec_H Y$.
For any $\sum_i a_i \otimes x_i \in A \odot M'$, $\xi \in X$, and $\eta, \zeta \in H$ one has
\begin{align*}
| \i< \eta, \sum_i \Omega_\xi (a_i) x_i \zeta >|
&= | \i< \xi \otimes \eta, \theta_X^H (\sum_i a_i \otimes x_i ) \xi \otimes \zeta >|
\leq \| \theta_Y^H (\sum_i a_i \otimes x_i ) \| | \xi \otimes \eta \| \| \xi \otimes \zeta \| \\
& \leq \| \theta_Y^H (\sum_i a_i \otimes x_i ) \| \| \xi \|^2 \| \eta \| \| \zeta \| ,
\end{align*}
which implies that $ \theta_Y ^H ( \sum_i a_i \otimes x_i ) \mapsto \sum_i \Omega_\xi (a_i) x_i$ is bounded and its norm is less than or equal to $\| \xi \|^2$.
Conversely, suppose that $\Phi: \Imag \theta_Y^H \to \lB (H); \theta_Y^H (a \otimes x) \mapsto \Omega_\xi (a) x$ is bounded.
Let $z \in \ker \theta_Y^H$ be arbitrarily fixed.
We show that $\i< \xi \otimes \eta, \theta_X^H (z) \xi' \otimes \eta' > =0$ for all $\xi, \xi ' \in X$ and $\eta, \eta' \in H$.
By the polarization trick we may assume that $\xi = \xi'$.
Let $\varepsilon >0$ be arbitrary and take $\sum_i a_i \otimes x_i \in A \odot M'$ in such a way that $\| z - \sum_i a_i \otimes x_i \|_{\max} < \varepsilon$.
We then have
\begin{align*}
| \i< \xi \otimes \eta, \theta_X^H (\sum_i a_i \otimes x_i ) \xi \otimes \eta' >|
&=| \i< \eta, \sum_i \i< \xi, \pi_X( a_i) \xi > x_i \eta' > |
\leq \i< \eta, \Phi (\sum_i a_i \otimes x_i ) \eta'> | \\
& \leq \|\Phi \| \| \theta_Y^H (\sum_i a_i \otimes x_i ) \| \| \eta \| \| \eta ' \|
< \varepsilon \|\Phi \| \| \eta \| \| \eta ' \|.
\end{align*}
Since $\varepsilon >0$ is arbitrary, we get $\theta_X^H (z)=0$, and hence $X \prec_H Y$.
\end{proof}
The next technical proposition will be used later.
The proof below is based on \cite[Proposition 3.6.5]{Brown-Ozawa}, called {\it The Trick}.
\begin{proposition}\label{prop-product-map2}
Let $A$ and $B$ be unital $\rC^*$-algebras, and $X,Y \in \Corr(A, B)$ and $H \in \Rep (B)$.
If $\pi_Y$ is unital and $X \prec_H Y$ holds, then for any normal vector $\xi \in X$ there exists a u.c.p.\ map $\Theta$ from $( \lL_B (Y) \otimes 1_H)'' \subset \lB (Y \otimes_B H)$ into $\pi_H (B)''$ such that $\Theta (\pi_Y (a) \otimes 1_H) = \pi_H \circ \Omega_\xi (a)$ for $a\in A$.
\end{proposition}
\begin{proof}
Thanks to the previous proposition there exists a u.c.p.\ map $\Phi : \Imag \theta_Y^H \to \lB (H)$ such that $\Phi ( \theta_Y^H ( a \otimes x) ) = \pi_H ( \i< \xi, \pi_X( a) \xi > )x$ for $ a\in A $ and $\xi \in X$.
Since $\Imag \theta_Y^H$ is a unital C$^*$-subalgebra of $\lB (Y \otimes_B H)$,
by Arveson's extension theorem we can extend $\Phi$ to a u.c.p.\ map $\Psi$ from $\lB (Y \otimes_B H)$ into $\lB (H)$.
Note that $\Imag \theta_Y^H$ is contained in the multiplicative domain of $\Psi$, i.e., $\Psi (abc) = \Phi (a) \Psi (b) \Phi(c)$ holds for $a, c \in \Imag \theta_Y^H$ and $b \in \lB (Y \otimes_B H)$.
Let $\Theta$ be the restriction of $\Psi$ to $(\lL_B (Y) \otimes 1_H )''$.
Then, for $x \in (\lL_B (Y) \otimes 1_H )''$ and $y \in M'$ we have $y\Theta (x) =\Phi ( \theta_Y^H (1_A \otimes y )) \Psi (x) = \Psi ( \theta_Y^H (1_A \otimes y )x) = \Psi (x \theta_Y^H (1_A \otimes y ))=\Psi (x) \Phi ( \theta_Y^H (1_A \otimes y )) =\Theta (x) y$, which implies $\Theta (x) \in \pi_H(B)''$.
\end{proof}
\section{Relative nuclearity}\label{sec-rel}
In this section we give the definition of relative nuclearity and prove Theorem \ref{thm-A}.
\subsection{Universal factorization property}\label{ss-rel-UFP}
To define relative nuclearity, we need the notion of universal factorization property, which plays a role of the original definition of nuclearity.
Recall the fact that for $\rC^*$-algebras $A$ and $B$ every nondegenerate representation $\sigma : A \odot B \to \lB (H)$ is of the form of $\sigma_A \times \sigma_B$, where $\sigma_A : A \to \lB (H)$ and $\sigma_B : B \to \lB (H)$ are unique $*$-homomorphisms having mutually commuting ranges and satisfying $\sigma (a \otimes b ) =\sigma_A (a) \sigma_B (b)$ (see e.g., \cite[Theorem 3.2.6]{Brown-Ozawa}).
\begin{definition}\label{def-UFP}
Let $A$ be a $\rC^*$-algebra.
We say that $(X,\pi_X)\in \Corr(A)$ has the {\it universal factorization property} ({\it UFP\/} for short) if it satisfies the following universal property:
For any $\rC^*$-algebra $B$, every $*$-representation $(H,\sigma) \in \Rep (A \otimes_{\max} B)$ is weakly contained in the $*$-representation $\phi_X^H: A \otimes_{\max} B \ni a \otimes b \mapsto \pi_X(a) \otimes \sigma_B (b) \in \lB (X\otimes _{\sigma_A} H)$.
In other words, there exists $*$-homomorphism $\Phi : \Imag \phi_X^H \to \lB( H)$ such that the following diagram commutes:
\[
\xymatrix @C=2cm
{
A \us\otimes_{\rm max} B
\ar[d]_{\phi_X^H} \ar[r]^{ \sigma} & \lB( H) \\
\Imag \phi_X^H \ar@{-->}[ru]_\Phi
}
\]
\end{definition}
Firstly, the UFP of a given C$^*$-correspondence is characterized in terms of weak containment.
\begin{proposition}\label{prop-UFP}
For any $\rC^*$-algebra $A$ and any $\rC^*$-correspondence $(X, \pi_X)$ over $A$,
$(X, \pi_X)$ has the UFP if and only if $(A,\lambda_A) \prec_\univ (X, \pi_X)$ holds.
\end{proposition}
\begin{proof}
Suppose that $(X, \pi_X)$ has the UFP.
Let $(H, \pi_H)$ be the universal representation of $A$.
Applying the UFP to $\pi_H \times \iota : A \odot \pi_H(A)' \to \lB (H)$,
where $\iota : \pi_H (A)' \hookrightarrow \lB (H)$ is the inclusion map, we get
$\ker \theta_X^H = \ker \phi_X^H \subset \ker (\pi_H \times_{\rm max} \iota) = \ker \theta_A^H$.
Conversely, suppose that $(A, \lambda_A) \prec_\univ (X, \pi_X)$ holds.
Let $B$ and $\sigma_A \times \sigma_B : A\odot B \to \lB (H)$ be given.
The condition that $(A, \lambda_A) \prec_\univ (X, \pi_X )$ implies
that $(A, \lambda_A) \prec_{(H, \sigma_A)} (X, \pi_X)$.
For any $\sum_i a_i \otimes b_i \in A \odot B $ we have
$ \| \sigma_A \times \sigma_B ( \sum_i a_i \otimes b_i ) \| =\| \theta_A^H (\sum_i a_i \otimes \sigma_B (b_i) ) \| \leq \| \theta_X^H ( \sum_i a_i \otimes \sigma_B (b_i) \| = \|\phi_X^H ( \sum_i a_i \otimes b_i ) \|$,
which implies $(H, \sigma) \prec (X \otimes_B H, \phi_X^H )$.
\end{proof}
We next translate the original definition of nuclearity into the language of C$^*$-correspondences by use of the notion of UFP.
We note that equivalence between nuclearity and CPAP follows from the next proposition together with Corollary \ref{cor-weak}.
\begin{proposition}\label{prop-nuclear}
Let $A$ be a $\rC^*$-algebra and $\pi_H : A \to \lB (H)$ be a faithful $*$-representation.
Then, the following are equivalent:
\begin{itemize}
\item[(1)] $A$ is nuclear.
\item[(2)] $(H \otimes A, \pi_H \otimes 1_A)$ has the UFP.
\item[(3)] $(A, \lambda_A) \prec_\univ (H \otimes A, \pi_H \otimes 1_A)$.
\end{itemize}
\end{proposition}
\begin{proof}
We prove (1) $\Rightarrow$ (2):
Suppose that $A$ is nuclear and fix a C$^*$-algebra $B$ and nondegenerate $*$-representation $\sigma: A \odot B \to \lB (K)$ arbitrarily.
We observe that $(H \otimes A )\otimes _{\sigma_A} K \cong H \otimes K$, and this isomorphism induces $\Imag \phi_{K \otimes A}^H \cong A \otimes \sigma_B (B) \subset \lB (K \otimes H)$.
The nuclearity of $A$ implies that $\Imag \phi_{K \otimes A}^H \cong A \otimes_{\max} \sigma_B (B)$.
By the universality of the maximal tensor product, the mapping $\Phi : \Imag \phi_{K \otimes A}^H \ni \phi_{K \otimes A}^H (a \otimes b ) \mapsto \sigma_A (a) \sigma_B (b) \in \lB (H)$ is bounded,
and hence $(H \otimes A, \pi_H \otimes 1_A)$ has the UFP.
Equivalence between (2) and (3) follows from the previous proposition.
We prove (2) $\Rightarrow$ (1):
Suppose that $(H \otimes A, \pi_H \otimes 1_A)$ has the UFP.
Let $\sigma : A\otimes_{\max} B \to \lB (H)$ be a faithful $*$-representation.
Since $\Imag \phi_{H \otimes A}^K \cong A \otimes \sigma_B(B)$ holds as above,
the UFP gives the inverse of the canonical surjection from $A\otimes_{\max} B$ onto $A \otimes B$, and hence $A$ is nuclear.
\end{proof}
\subsection{Relative nuclearity}\label{ss-rel-rel}
\begin{definition}\label{def-rel-nuc}
Let $B \subset A$ be an inclusion of $\rC^*$-algebras with conditional expectation $E :A \to B$.
We say that the triple $(A, B, E)$ is {\it nuclear via} $(Z,\pi_Z) \in \Corr (B)$ if the $\rC^*$-correspondence $(L^2(A,E) \otimes_B Z \otimes_B A, \pi_E \otimes 1_Z \otimes 1_A)$ has the UFP.
When $(Z, \pi_Z)=(B, \lambda_B)$, we say that $(A, B, E)$ is {\it nuclear}.
\end{definition}
Let $B \subset A$ be an inclusion of C$^*$-algebras and $E:A \to B$ be a nondegenerate conditional expectation.
As we will see in the next section, $(A, B, E)$ is nuclear whenever $A$ is nuclear and the embedding $B \hookrightarrow A$ is full (see Example \ref{ex-nuc}).
We do not know whether or not this still holds true when we remove the assumption of fullness,
but we can prove the nuclearity of $(A, B, E)$ via some C$^*$-correspondences over $B$.
This is the merit of considering nuclearity via C$^*$-correspondences over subalgebras.
We also mention that the `$(Z, \pi_Z)$' does not affect much in some cases.
For example, in the case when $B= \lC 1_A$, $L^2(A,E) \otimes_B Z$ is a usual Hilbert space.
Thus, the nuclearity of $(A, \lC 1_A, E)$ via some C$^*$-correspondence over $\lC$ is equivalent to the one of $A$ by Proposition \ref{prop-nuclear}.
Moreover, in \S\S \ref{ss-rel-WFP} and \S\S \ref{ss-rel-ame} we will see that the nuclearity of $(A, B, E)$ via some C$^*$-correspondence over $B$ implies the relative injectivity of $\pi_H (B)'' \subset \pi_H (A)''$ for any $(H, \pi_H) \in \Rep (A)$, and a relative weak expectation property of A.
The next theorem says that this `via version' of relative nuclearity is characterized by a kind of `relative CPAP'.
\begin{theorem}\label{thm-nuc-CPAP}
Let $B \subset A$ be a unital inclusion of $\rC^*$-algebras with conditional expectation $E$, and $(Z, \pi_Z)$ be a $\rC^*$-correspondence over $B$.
Then, $(A, B, E)$ is nuclear via $(Z, \pi_Z)$ if and only if for any finite subset $\fF \subset A$ and $\varepsilon >0$, there exist $n,m \in \lN$, $\varphi_k : A \to \lM_n (B)$ and $\psi_k : \lM_n (B) \to A$, $1 \leq k \leq m$ such that $\| a - \sum_{k=1}^m \psi_k \circ \varphi_k (a) \| < \varepsilon$ for $a \in \fF$, and each $\varphi_k$ and $\psi_k$ are of the form
\[
\varphi_k : a \mapsto \left[ \rule{0pt}{10pt} \i< \eta_i, (\pi_E(a) \otimes 1_Z) \eta_j > \right]_{i,j=1}^n
\quad
\psi_k : \left[ \rule{0pt}{0pt}\, b_{ij}\, \right]_{i,j=1}^n \mapsto \sum_{i,j=1}^n y_i^* b_{ij} y_j
\]
for some $\eta_1, \dots, \eta_n \in L^2(A, E) \otimes_B Z$ and $y_1, \dots, y_n \in A$.
\end{theorem}
\begin{proof
Suppose that $(A, B, E)$ is nuclear via $(Z, \pi_Z)$
and fix a finite subset $\fF \subset A$ and $\varepsilon >0$ arbitrarily.
By Proposition \ref{prop-UFP} and Proposition \ref{prop-nuclear},
we have $(A, \lambda_A) \prec_\univ (L^2(A,E) \otimes_B Z \otimes_B A, \pi_E \otimes 1_Z \otimes 1_A)$.
Since the identity map on $A$ is noting but $\Omega_{1_A} \in \cF_A$,
thanks to Theorem \ref{thm-weak}, we can find $m \in \lN$ and $\xi_1, \dots, \xi_m \in L^2(A, E) \otimes_B Z \otimes_B A$ in such a way that $\| a - \sum_{k=1}^m \Omega_{\xi_k} (a ) \| < \varepsilon$ for all $a\in \fF$.
Here we may assume that each $\xi_k$ is of the form $\sum_{i=1}^n \eta^{(k)}_i \otimes y^{(k)}_i \in (L^2(A, E) \otimes_B Z ) \odot A$.
By Remark \ref{rem-factor}, letting $\varphi_k (a) := \left[ \i< \eta^{(k)}_i, (\pi_E (a) \otimes 1) \eta^{(k)}_j > \right]_{i,j=1}^n$ for $a \in A$ and $\psi_k ([ b_{ij}]_{ij=1}^n ):=\sum_{i,j=1}^n y_i^{(k)*} b_{ij} y_j^{(k)}$ for $[b_{ij}]_{i,j=1}^n \in \lM_n (B)$,
we get $\| a - \sum_{k=1}^m \psi_k \circ \varphi_k (a) \| < \varepsilon$ for all $a \in \fF$.
The converse implication follows from Proposition \ref{prop-UFP}, Proposition \ref{prop-nuclear}, and Corollary \ref{cor-weak} again.
\end{proof}
\begin{proof}[Proof of Theorem \ref{thm-A}]
Suppose that $(A, B, E)$ is nuclear and fix a finite subset $\fF \subset A$ and $\varepsilon >0$ arbitrarily.
By the preceding theorem, we can find $n, m \in \lN$ and c.p.\ maps $\varphi_k : A \to \lM_n (B), \psi_k : \lM_n (B ) \to A$ satisfying that $\| a - \sum_{k=1}^m \psi_k \circ \varphi_k (a) \| < \varepsilon$ for $a \in A$.
We only have to modify $\varphi_k$'s.
For each $k$, $\varphi_k$ is of the form $a \mapsto [ \i< \eta_i, \pi_E(a) \eta_j > ]_{i,j=1}^n$ for some $\eta_1, \dots, \eta_n \in L^2(A, E)$.
Since $L^2(A,E)$ is the completion of $A$, we may assume that $\eta_i$ comes from an element $x_i \in A$.
Since $\i< x_i, \pi_E( a) x_j > = E (x_i^* a x_j)$ holds, we are done.
\end{proof}
The next proposition implies that if both $(A, B, E)$ and $B$ are nuclear, then so is $A$.
\begin{proposition}\label{prop-nuc-exact}
Let $A$ and $B$ be $\rC^*$-algebras with $A$ unital and $B$ nuclear.
If there exist $X \in \Corr (A, B)$ and $Y \in \Corr (B, A)$ such that $(X \otimes_B Y, \pi_X \otimes 1_Y) \in \Corr (A)$ has the UFP, then $A$ is nuclear.
\end{proposition}
\begin{proof}
Since $(X \otimes_B Y, \pi_X \otimes 1_Y)$ has the UFP,
by Proposition \ref{prop-UFP} we have $(A, \lambda_A) \prec_\univ (X \otimes_B Y, \pi_X \otimes 1_Y) $.
If $(H, \pi_H) \in \Rep (B)$ is a faithful representation,
then thanks to Proposition \ref{prop-nuclear} we have $(B, \lambda_B) \prec_\univ ( H \otimes B, \pi_H \otimes 1_A).$
Proposition \ref{prop-int-tensor} implies that $(X \otimes_B Y, \pi_X \otimes 1_Y) \prec_\univ ( X \otimes_B H \otimes Y, \pi_X \otimes 1_H \otimes 1_Y)$.
Since $( X \otimes_B H \otimes Y, \pi_X \otimes 1_H \otimes 1_Y)$ is a scalar representation of $A$, this implies the nuclearity of $A$.
\end{proof}
We next introduce the notion of strong relative nuclearity.
Let $B\subset A$ be an inclusion of $\rC^*$-algebras and $(X,\pi_X)$ be a $\rC^*$-correspondence over $A$.
A vector $\xi \in X$ is said to be $B${\it -central} if $\xi$ enjoys $\pi_X(b) \xi = \xi b$ for all $b \in B$.
We denote by $B' \cap X$ the set of $B$-central vectors in $X$.
For the identity C$^*$-correspondence $(A, \pi_A) \in \Corr (A)$ and $B\subset A$, the set of $B$-central vectors is nothing but the relative commutant $B' \cap A$.
We also note that every c.p.\ map in $\cF_{B'\cap X}$ forms a $B$-bimodule map.
\begin{definition}\label{def-CCPAP}
Let $B \subset A$ be an inclusions of $\rC^*$-algebras.
\begin{itemize}
\item We say that a $\rC^*$-correspondence $(X,\pi_X)$ over $A$ has the {\it $B$-central completely positive approximate property} ($B$-{\it CCPAP\/} for short) if there exists a net of c.c.p.\ maps $\psi_i \in \cF_{B'\cap X}$ such that $\lim_{i} \| a - \psi_i (a) \| =0$ for every $a\in A$.
\item Let $E :A \to B$ be a conditional expectation. We say that the triple $(A, B,E)$ is {\it strongly nuclear via} $(Z, \pi_Z) \in \Corr (B)$ if $(L^2(A,E) \otimes_BZ \otimes_B A, \pi_E \otimes 1_Z \otimes 1_A)$ has the $B$-CCPAP.
When $(Z, \pi_Z) =(B, \lambda_B)$, we say that $(A, B, E)$ is {\it strongly nuclear}.
\end{itemize}
\end{definition}
For a unital $A$ the $B$-CCPAP of a $\rC^*$-correspondence over $A$ implies the UFP.
Thus, every strongly nuclear triple $(A, B, E)$ is nuclear, but we do not know whether or not the converse is true.
\subsection{Relative WEP}\label{ss-rel-WFP}
We next discuss relative weak expectation property recently introduced by Jian and Sepideh \cite{Jian-Sepideh} in relation with our relative nuclearity.
\begin{definition}[{\cite[Proposition 3.3.6]{Brown-Ozawa}}]
An inclusion $B\subset A$ is said to be {\it relatively weakly injective} if the following equivalent conditions hold;
\begin{itemize}
\item[(1)] there exists a c.c.p.\ map $\varphi:A \to B^{**}$ such that $\varphi(b)=b$ for every $b\in B$;
\item[(2)] for every $*$-homomorphism $\pi:B\to \lB (H)$ there exists a c.c.p.\ map $\varphi:A \to \pi(B)''$ such that $\varphi(b)=\pi(b)$ for every $b\in B$;
\item[(3)] for every {\rm C}$^*$-algebra $C$ there is a natural inclusion
$
B\otimes_{\rm max} C \subset A\otimes_{\rm max}C.
$
\end{itemize}
\end{definition}
\begin{definition}[\cite{Jian-Sepideh}]
Let $A$ and $B$ be $\rC^*$-algebras.
Then, $A$ is said to have the $B$-WEP$_1$ (resp. $B$-WEP$_2$) if there exists $(X, \pi_X) \in \Corr (A, B)$ (resp. $(X, \pi_X) \in \Corr (A, B^{**} )$ with $X$ selfdual) such that $\pi_X $ is injective and the inclusion $\pi_X(A) \subset \lL_B (X)$ (resp. $\pi_X(A) \subset \lL_{B^{**}}(X)$ is relatively weakly injective.
\end{definition}
Note that Lance's WEP is exactly $\lC$-WEP$_1$.
In \cite{Jian-Sepideh} it was proved that $B$-WEP$_1$ implies $B$-WEP$_2$.
We note that the next proposition can be applied to every triple $(A, B, E)$ that is nuclear via some C$^*$-correspondence over $B$,
which is an analogue of the fact that `nuclearity $\Rightarrow$ WEP'.
\begin{proposition}\label{prop-BWEP}
Let $A$ and $B$ be $\rC^*$-algebras with $A$ unital.
If there exist $X \in \Corr (A, B)$ and $Y \in \Corr (B,A)$ such that $\pi_X$ is unital injective, and $(X\otimes_B Y, \pi_X \otimes 1_Y)$ has the UFP,
then the inclusion $\pi_X( A) \subset \lL_B (X)$ is relatively weakly injective.
In particular, $A$ has the $B$-WEP$_1$.
\end{proposition}
\begin{proof}
Let $(H, \pi_H) \in \Rep (A)$ be arbitrary.
We will show that there exists a u.c.p.\ map $\Psi : \lL_B (X) \to \pi_H (A)''$ such that $\Psi \circ \pi_X = \pi_H$.
The UFP of $X \otimes_B Y$ and Proposition \ref{prop-product-map2}
imply that there exists a u.c.p map $\Theta : \lL_A (X \otimes_B Y) \otimes 1_H \to \pi_H (A)''$ satisfying $\Theta ( \pi_X (a) \otimes 1_Y \otimes 1_H) = \pi_H (a)$.
Hence, the mapping $\Psi: \lL_B (X) \ni x \mapsto \Theta (x \otimes 1_Y \otimes 1_H) \in \pi_H (A)''$ is the desired one.
\end{proof}
\subsection{Relative amenability for von Neumann algebras}\label{ss-rel-ame}
In this subsection, we see that our relative nuclearity is related to relative amenability for von Neumann algebras \cite{Popa}\cite{Delaroche2}\cite{Ozawa-Popa}.
Let $N \subset M$ be an inclusion of finite von Neumann algebras and $\tau$ be a faithful normal tracial state.
Let $L^2(M)$ and $L^2(N)$ be the GNS Hilbert spaces for $\tau$ and $\tau |_N$, respectively and $\xi_\tau \in L^2(M)$ be the corresponding cyclic vector.
We may assume that $M \subset \lB (L^2(M))$.
Denote by $E_N$ the unique faithful normal $\tau$-preserving conditional expectation from $M$ onto $N$.
Let $e_N \in \lB (L^2(M))$ be the orthogonal projection onto $L^2(N) \subset L^2(M)$,
which is called the Jones projection and satisfies that $e_N x e_N = E_N (x) e_N$ for $x \in M$.
The basic extension $\i<M, e_N>$ is the von Neumann subalgebra of $\lB (L^2(M))$ generated by $M$ and $e_N$.
It is known that $\i< M, e_N>$ is the $\sigma$-weak closure of $\lspan \{ x e_N y \mid x, y \in M \}$ and becomes semifinite with the canonical faithful normal semifinite tracial weight $\Tr :x e_N y \mapsto \tau (xy)$.
\begin{theorem}[{\cite[Theorem 2.1]{Ozawa-Popa}}]
Let $M$ be a finite von Neumann algebra with a faithful normal tracial state $\tau$ and $Q, N \subset M$ be von Neumann subalgebras.
Then, the following are equivalent.
\begin{itemize}
\item[$(1)$] There exists a $N$-central state $\varphi$ on $\i< M, e_Q>$ such that $\varphi |_M = \tau$.
\item[$(2)$]There exists a $N$-central state $\varphi$ on $\i< M, e_Q>$ such that $\varphi$ is normal on $M$ and faithful on $\cZ (N' \cap M)$.
\item[$(3)$]There exists a conditional expectation $\Phi : \i< M, e_Q>$ onto $N$ such that $\Phi|_M = E_N$.
\item[$(4)$]There exists a net $\{ \xi_i \}_i$ in $L^2 \i< M, e_Q>$ such that $\lim_i \i< \xi_i, x \xi_i > =\tau (x)$ for $x \in M$ and $\lim_i \| [ u, \xi_i ] \|_2 =0$ for all $u \in \cU (N)$.
\end{itemize}
When any of these conditions holds, we say that $N$ is amenable relative to $Q$ inside $M$ and write $N \lessdot_M Q$.
\end{theorem}
When $M \lessdot_M N$ holds, we also say that $M$ is {\it amenable relative to} $N$.
The next lemma seems to be known among specialists, but we do give its proof for the reader's convenience.
\begin{lemma}
There exists a unitary $U$ from $L^2 \i< M, e_Q>$ onto $H:=L^2(M, E_Q) \otimes_Q L^2(M)$ that maps $x e_N y$ to $x \otimes y $ and satisfies that $ U \i< M, e_Q> U^*= (\lL_Q (L^2(M, E_Q) )\otimes 1_{L^2(M)})''$.
\end{lemma}
\begin{proof}
By \cite[Lemma 2.1]{Izumi-Longo-Popa} $ \lspan Me_N M$ is norm dense in $L^2\i< M, e_N>$.
For any finite sums $\sum_k x_k e_N y_k$, $\sum_l z_l e_N w_l \in \lspan M e_N M$ we have
$$
\Tr ( (\sum_k x_k e_N y_k)^*(\sum_l z_l e_N w_l) )
= \sum_{k,l} \tau ( y_k^* E_N ( x_k^*y_l) z_l) = \i<\sum_k x_k \otimes y_k \xi_\tau, \sum_l z_l \otimes w_l \xi_\tau >_H.
$$
Thus, the mapping $\lspan M e_N M \ni \sum_k x_k e_N y_k \mapsto \sum_k x_k \otimes y_k \xi_\tau \in H$ is bounded and extends to a unitary, which gives the isomorphism of Hilbert $M$-$M$ bimodules.
To see that $ U \i< M, e_Q> U^*= (\lL_Q (L^2(M, E_Q) )\otimes 1_{L^2(M)})''$, let $P_N \in \lL_N (L^2(M, E_N))$ the projection defined by $P_N x = E_N (x)$ for $x \in M$.
Then, it follows that $\lK (L^2(M, E_N)) = \ospan M P_N M$ and $U (x e_N y)U^*=x P_N y$ for $x, y \in M$.
Since $(\lK_N (L^2(M, E_N) )\otimes 1 )'' = (\lL_N (L^2(M, E_N) )\otimes 1 )''$,
we get $(\lL_Q (L^2(M, E_Q) )\otimes 1_{L^2(M)})''= ( \lspan M P_N M \otimes 1)'' =U( \lspan M e_N M \otimes 1 )''U^* = U \i<M, e_N> U^*$.
\end{proof}
Here we give a characterization of relative amenability in terms of weak containment.
\begin{proposition}\label{prop-rel-ame}
Let $Q, N \subset M$ be inclusions of finite von Neumann algebras and $\tau$ be a faithful normal tracial state on $M$.
Then, $N \lessdot_M Q$ if and only if $L^2(M, E_N) \prec L^2 (M, E_Q) \otimes_Q L^2 (M, E_N)$ as $M$-$N$ $\rC^*$-correspondences.
\end{proposition}
\begin{proof}
Put $X_N:=L^2(M,E_N)$ and $X_Q:=L^2(M, E_Q)$ and let $\xi_N \in L^2 (M,E_N)$ be the vector corresponding to $1$.
We note that $X_N \otimes_N L^2 (N) \cong L^2(M)$.
Suppose that $X_N$ is weakly contained in $X_Q \otimes_Q X_N$ with respect to the standard representation $N \subset \lB (L^2 (N, \tau|_N) )$.
By Proposition \ref{prop-product-map2} there exists a u.c.p map $\Phi$ from $(\lL_Q (X_Q) \otimes 1 \otimes 1)'' \subset \lB (X_Q \otimes_Q X_N \otimes_N L^2 (N) ) = \lB (X_Q \otimes_Q L^2 (M) )$ into $N$ satisfying that $\Phi (x \otimes 1 \otimes 1) = \Omega_{\xi_N} (x) =E_N (x)$ for $x\in M$.
Since $(\lL_Q (X_Q) \otimes 1 \otimes 1 )'' $ is isomorphic to $\i< M, e_Q>$, we are done.
Suppose $N \lessdot_M Q$.
Let $\xi_i \in L^2 \i< M, e_Q >$ be a net in (4) of the theorem above.
Let $z \in \ker \theta_{X_Q \otimes_Q X_N}^{L^2(N)}$ and $\varepsilon >0$ be arbitrary.
Take $w = \sum_k x_k \otimes y_k^\op \in M \odot N^\op$ such that $\| z - w \|_{\max} < \varepsilon$.
Note that $X_N \otimes_N L^2(N) \cong L^2(M)$.
For any $a, b \in M$ we have
\begin{align*}
| \i< a \xi_\tau, \theta_{X_N}^{L^2(N)} (w ) b \xi_\tau > |&= |\tau (\sum_k a^*x_k b y_k )|= \lim_i | \i< \xi_i,\sum_k a^*x_k b y_k \xi_i > | \\
&=\lim_i | \i< a \xi_i, \sum_k x_k b \xi_i y_k > | = \lim_i | \i< a \xi_i, \theta_{X_Q \otimes_Q X_N}^{L^2(N)} (w )b \xi_i >| \leq \varepsilon \|a \| \|b \|,
\end{align*}
which implies $\theta_{X_N}^{L^2(N)} (z) =0$.
\end{proof}
Let $N\subset M$ be an inclusion of von Neumann algebras.
Recall that $M$ is said to be {\it injective relative to} $N$ if there exists a norm one projection from $JN'J \subset L^2 (M)$ onto $M$ (\cite[Definition 3.1]{Delaroche2}),
where $J : L^2 (M) \to L^2 (M)$ is the modular conjugation.
When $M$ is finite, this is the case that $M \lessdot_M N$, that is, $M$ is amenable relative to $N$ (inside $M$).
\begin{proposition}\label{prop-relative-injective}
Let $B\subset A$ be a unital inclusion of $\rC^*$-algebras.
If there exists a unital $X\in \Corr(A,B)$ such that $(X \otimes_B A, \pi_X \otimes 1_A)$ has the UFP,
then for any $*$-representation $\pi_H :A \to \lB(H)$,
there exists a norm one projection from $\pi_H(B)'$ onto $\pi_H(A)'$.
Consequently, $\pi_H(A)''$ is injective relative to $\pi_H(B)''$.
\end{proposition}
\begin{proof}
We note that $X\otimes_B A \otimes_A H =X\otimes_B H$.
Applying the UFP of $X \otimes_B A$ to $\pi_H \times \iota : A\odot \pi_H(A)' \to \lB(H)$
we get the $*$-homomorphism from $\Imag \phi_X^H \subset \lB (X \otimes_B H)$ to $\lB (H)$. By Arveson's extension theorem, we obtain a u.c.p.\ map $\Phi: \lB (X\otimes_B H ) \to \lB (H)$ satisfying $\Phi ( \pi_X (a) \otimes x ) = \pi_H (a)x$ for $a \in A$ and $x\in \pi_H (A)'$.
Define $\Psi: \pi_H (B)' \to \lB (H)$ by $\Psi (x):= \Phi (1_X \otimes x)$.
Then we can prove that $\Psi$ is a norm one projection onto $\pi_H(A)'$ as in the proof of Proposition \ref{prop-product-map2}.
To see the second assertion, let $M:=\pi_H(A)''$ and $N:=\pi_H(B)''$.
We may assume that $M$ and $N$ are acting on the standard Hilbert space $L^2 (M)$.
Let $J$ be the modular conjugation on $L^2 (M)$.
We then have a u.c.p.\ map $\Phi : N' \to M'$.
Thus, $JN'J \ni J x J \mapsto J \Psi (x) J \in J M'J=M$ is the desired map.
\end{proof}
\subsection{Permanence properties}\label{ss-rel-per}
For Hilbert $\rC^*$-modules $X$ and $Y$ over $\rC^*$-algebras $A$ and $B$, respectively,
we denote by $X \toplus Y$ the Hilbert $A \oplus B$-module $X \oplus_{\rm alg} Y$ equipped with the inner product
$\i< \xi \oplus \eta, \xi' \oplus \eta >=\i< \xi, \xi' > \oplus \i< \eta, \eta' >$ for $\xi, \xi' \in X$ and $\eta, \eta' \in Y$.
The next proposition immediately follows from Theorem \ref{prop-UFP} and the definition of relative CCPAP.
\begin{proposition}
Let $B_i \subset A_i, i=1,\dots, n$ be inclusions of $\rC^*$-algebras with $A_i$ unital.
For $X_i \in \Corr (A_i)$ the following hold true.
\begin{itemize}
\item[$(1)$] The $X_1 \toplus X_2 \toplus \cdots \toplus X_n \in \Corr (\bigoplus_{i=1}^n A_i)$ has the UFP (resp. $\bigoplus_{i=1}^n B_i$-CCPAP) if and only if each $X_i$ has the UFP (resp. $B_i$-CCPAP).
\item[$(2)$] The $\bigotimes_{i=1}^n X_i \in \Corr (\bigotimes_{i=1}^n A_i)$ has the UFP (resp. $\bigotimes_{i=1}^n B_i $-CCPAP) if each $X_i$ has the UFP (resp. $B_i$-CCPAP).
\end{itemize}
\end{proposition}
\begin{proposition}[Direct sums and tensor products]\label{prop-sum-tensor}
Let $(A_i, B_i, E_i), 1\leq i \leq n$ be a finite family of unital inclusions of $\rC^*$-algebras with conditional expectations.
Then, the following hold true{\rm :}
\begin{itemize}
\item[$(1)$] $(\bigoplus_i A_i ,\bigoplus_i B_i, \bigoplus_i E_i )$ is nuclear (resp. strongly nuclear) if and only if so is each $( A_i, B_i, E_i)$.
\item[$(2)$] $(\bigotimes_i A_i ,\bigotimes_i B_i, \bigotimes_i E_i )$ is nuclear (resp. strongly nuclear) if so is each $( A_i, B_i, E_i)$.
\end{itemize}
\end{proposition}
\begin{proof}
Let $(A, B, E):=(\bigoplus_i A_i ,\bigoplus_i B_i, \bigoplus_i E_i )$ and $(\widetilde{A}, \widetilde{B}, \widetilde{E}):= (\bigoplus_i A_i ,\bigoplus_i B_i, \bigoplus_i E_i )$.
The assertions follow from the propositions above together with the next canonical isomorphisms
$L^2 (A, E ) \otimes_{B} A \cong \bigotimes_i (L^2 (A_i, E_i) \otimes_{B_i} A_i )$ and $L^2(\widetilde{A}, \widetilde{E}) \otimes_{\widetilde{B}} \widetilde{A} \cong \bigtoplus_i (L^2 (A_i, E_i) \otimes_{B_i} A_i )$.
\end{proof}
\begin{proposition}[Inductive limits]
Let $B \subset A$ be a unital inclusion of $\rC^*$-algebras with a conditional expectation $E :A \to B$.
If there exists an increasing net $1_A \in A_i, i \in \cI$, of unital $\rC^*$-subalgebras of $A$
such that $E(A_i) =A_i \cap B$ and $(A_i, A_i \cap B, E|_{A_i} )$ is nuclear and $\bigcup_i A_i$ is norm dense in $A$, then $(A, B, E)$ is nuclear.
\end{proposition}
\begin{proof}
Take finitely many elements $\{ x_1, \dots, x_n \} \subset A$ and $\varepsilon >0$ arbitrarily.
Since $\bigcup_i A_i$ is norm dense in $A$, there exists $i \in \cI$ and $y_1, \dots, y_n \in A_i$ such that $\| x_i - y_i \| <\varepsilon /3$ for $1\leq i \leq n$.
Set $B_i:=B \cap A_i$ and $E_i := E|_{A_i}$.
By the nuclearity of $(A_i, B_i, E_i )$ we can find a c.c.p.\ map $\psi \in \cF_{L^2(A_i, E_i) \otimes_{B_i} A_i}$ such that $\| \psi (y_i) - y_i \| <\varepsilon/3$ for $1 \leq i \leq n$.
Since $L^2(A_i, E_i) \otimes_{B_i} A_i$ can be embedded in $L^2(A, E) \otimes_B A$ canonically and $1_A =1_{A_i}$, the $\psi$ is a c.c.p.\ on $A$.
Now we get $\| x_i - \psi (x_i) \| \leq \| x_i - y_i \| + \| y_i - \psi (y_i) \| + \| \psi (y_i - x_i ) \| < \varepsilon$ for $1\leq i \leq n$.
\end{proof}
The following proposition can be shown in the same manner.
\begin{proposition}[Inductive limits with common subalgebras]
Let $B \subset A$ be a unital inclusion of $\rC^*$-algebras with a conditional expectation $E :A \to B$.
If there exists an increasing net $A_i, i \in \cI$, of unital $\rC^*$-subalgebras of $A$ containing $B$ such that $(A_i, B, E|_{A_i} )$ is strongly nuclear and $\bigcup_i A_i$ is norm dense in $A$, then $(A, B, E)$ is strongly nuclear.
\end{proposition}
\section{Examples}\label{sec-exam}
In this section, we give several examples of nuclear and strongly nuclear triples.
\subsection{Inclusions of nuclear C$^*$-algebras}\label{ss-exam-inc}
\begin{example}
For any $\rC^*$-algebra $A$ the triple $(A, A, \id )$ is strongly nuclear.
Further assume that $A$ is unital and $\varphi : A \to \lC$ is a nondegenerate state.
Then, the nuclearity of $A$ is equivalent to the strong nuclearity of $(A, \lC 1_A, \varphi )$ by Proposition \ref{prop-UFP} and Proposition \ref{prop-nuclear}.
\end{example}
The following lemma can be shown in the same manner as \cite[Lemma 2.21]{Dadarlat-Eilers}.
\begin{lemma}
Let $B \subset A$ be a unital inclusion of $\rC^*$-algebras and $\varphi$ be a state on $B$.
If the embedding $\iota : B \hookrightarrow A$ is full, that is, $\ospan AbA =A$ holds for all $b \in B \setminus \{0 \}$,
then for any finite subset $\fF \subset B$ and $\varepsilon >0$, there exist $n \in \lN$ and $a_1, \dots, a_n \in A$ such that $\| \varphi (b)1_A - \sum_{i=1}^n a_i^* b a_i \| < \varepsilon$ for $b \in \fF$.
\end{lemma}
\begin{example}[Inclusions of nuclear $\rC^*$-algebras]\label{ex-nuc}
Let $B \subset A$ be a unital inclusion of $\rC^*$-algebras with nondegenerate conditional expectation $E: A\to B$.
Take a faithful representation $\pi_H : B \to \lB (H)$.
Then, $A$ is nuclear if and only if $(A, B, E)$ is nuclear via $(H \otimes B, \pi_H \otimes 1)$.
Moreover, when the embedding $\iota : B \hookrightarrow A$ is full, the triple $(A, B, E)$ is nuclear.
Since $(L^2(A,E) \otimes_B H, \pi_E \otimes 1_H)$ is also a faithful representation of $A$,
$A$ is nuclear if and only if $(A, \id_A) \prec_\univ (( L^2(A,E) \otimes_B H) \otimes A , \pi_E \otimes 1_H \otimes 1_A)$ holds, thanks to Proposition \ref{prop-UFP} and Proposition \ref{prop-nuclear}.
The natural isomorphism $( L^2(A, E) \otimes_B H) \otimes A = L^2(A, E) \otimes_B (H \otimes B) \otimes_B A$ implies that this is the case when $(A, B,E)$ is nuclear via $(H\otimes B, \pi_H \otimes 1_A)$.
Now assume that $B$ is fully embedded.
Since every coefficient of $(H \otimes A, \pi_H \otimes 1_A)$ is approximated by c.p.\ maps of the form $b \mapsto \sum_{i,j} a_i^* \varphi (b_i^* b b_j) a_j$ for some state $\varphi$ and $a_i \in A, b_i \in B$,
by the preceding lemma and Corollary \ref{cor-weak}, we have $(H \otimes A, \pi_H \otimes 1_A) \prec_\univ (A, \iota)$.
Thus, thanks to Proposition \ref{prop-int-tensor} we have
\[
(A, \lambda_A) \prec_\univ (L^2(A, E) \otimes_{\pi_H \otimes 1_A} (H \otimes A), \pi_E \otimes 1_{H \otimes A}) \prec_\univ (L^2(A, E) \otimes_{ \iota} A, \pi_E \otimes 1_A).
\]
We note that if $A$ or $B$ is simple, then $\iota : B \to A$ is full.
\end{example}
\begin{example}[Continuous fields of nuclear C$^*$-algebras]
Let $A$ be a continuous filed of unital nuclear $\rC^*$-algebras on a compact Hausdorff space $X$,
that is, there is a unital embedding $C(X) \subset A' \cap A$, such that, the quotient maps $p^x : A \to A^x := A/ J_x$, where $J_x$ is the closed ideal generated by the kernel of the evaluation map $\ev_x : C (X) \to \lC$ at $x$,
satisfies that $X \ni x \mapsto \| p^x (a) \| \in \lC$ is continuous for each $a \in A$ and $A^x$ is nuclear for each $x \in X$.
Assume that there exists a conditional expectation $E : A \to C(X)$ such that for each $x \in X$ there exists a nondegenerate state $\varphi^x$ on $A^x$ satisfying $\ev_x \circ E = \varphi^x \circ p^x$.
Then, the triple $(A, C(X), E)$ is strongly nuclear (see \cite{Bauval}).
For the reader's convenience, we give a sketch of the proof of this observation.
Since every vector in $L^2(A, E) \otimes_{C(X) } A$ is $C(X)$-central, it suffices to show the nuclearity of $(A, B, E)$.
Fix $ a \in A$ and $\varepsilon >0$.
For each $x \in X$, the nuclearity of $A^x$ implies that there exist find finitely many vectors $\xi_{x,i} \in A \odot A \subset L^2(A, E) \otimes_{C(X) } A$ such that $\|p^x (a - \sum_{i} \i< \xi_{x,i}, a \xi_{x,i} >) \| < \varepsilon$.
Since $A$ is a continuous field, we can find an open neighborhood $U_x$ of $x$ such that $\|p^y (a - \sum_{i} \i< \xi_{x,i}, a \xi_{x,i} > ) \| < \varepsilon$ for $y \in U_x$.
By the compactness of $X$, there exists a finite subset $F \subset X$ such that $X= \bigcup_{x \in F} U_x$.
Let $\{ h_x \}_{x \in F}$ be a partition of unity for this covering.
Then, the vectors $\{ \xi_{x,i} h_x^{1/2} \}_{x \in F, i}$ in $L^2(A, E) \otimes_{C(X)} A$ do the job.
Indeed, we have
\begin{align*}
\| a - \sum_{x \in F} \sum_i \i< \xi_{x,i}h_x^{1/2}, a\xi_{x,i} h_x^{1/2}> \| &=\sup_{y \in X} \| \| p^y ( a - \sum_{x \in F} \sum_i \i< \xi_{x,i}, a \xi_{x,i} > h_x (y) ) \| \\ &\leq \sup_{y \in X} \sum_{x \in F} \| p^y ( a - \sum_i \i< \xi_{x,i}, a \xi_{x,i} > ) \| h_x (y) < \varepsilon.
\end{align*}
\end{example}
When $B$ is finite dimensional, relative nuclearity implies strong one thanks to the following proposition.
\begin{proposition}
Let $B \subset A$ be a unital inclusions of $\rC^*$-algebras with $B$ finite dimensional.
If $X \in \Corr (A)$ has the UFP, then $X$ has the $B$-CCPAP.
\end{proposition}
\begin{proof}
Write $B =\bigoplus_{r=1}^p \lM_{n(r)}$.
Let $\{ e^{(r)}_{ij} \}_{i,j=1}^{n(r)}$ be a matrix unit system of $\lM_{n(r)}$.
Take a finite subset $\fF \subset A$ and $\varepsilon >0$ arbitrarily.
By the the UFP, we can find $\xi_1, \dots, \xi_m \in X$ in such a way that
$$
\| e_{1i}^{(r)} a e_{j1 }^{(s)} - \sum_{k=1}^m \i< \xi_k, e_{1i}^{(r)} a e_{j1 }^{(s)} \xi_k > \| < \frac{\varepsilon}{p^2 \max_{1 \leq r \leq p} n(r)^2} \quad 1 \leq r,s \leq p,\; 1 \leq i,j \leq n(r)
$$
and $\sum_{k=1}^m \i< \xi_k, \xi_k > \leq 1$.
Set $ \eta_k := \sum_{r=1}^p \sum_{i=1}^{n(r)} e_{i1}^{(r)} \xi_k e_{1i}^{(r)}$ for $ 1 \leq k \leq m$.
Then, we have $e^{(r)}_{ij} \eta_k =e^{(r)}_{ij} e_{j1}^{(r)} \xi e_{1j}^{(r)} =e^{(r)}_{i1}\xi e_{1i}^{(r)}e_{ij}^{(r)} = \eta_k e_{ij}^{(r)}$, and hence $\eta_k$ is $B$-central.
For $a \in \fF$ we have
\begin{align*}
\sum_{k=1}^m \i< \eta_k, a \eta_k >
&= \sum_{k=1}^m\l< \sum_{r=1}^p \sum_{i=1}^{n(r)} e_{i1}^{(r)} \xi e_{1i}^{(r)}, a \sum_{s=1}^p \sum_{j=1}^{n(s)} e_{j1}^{(s)} \xi e_{1j}^{(s)} >
=\sum_{r, s=1}^p \sum_{i =1}^{n(r)} \sum_{j=1}^{n(s)} \sum_{k=1}^m e_{i1}^{(r)} \i< \xi, e_{1i}^{(r)} a e_{j1}^{(s)} \xi > e_{1j}^{(s)}\\
& \approx_\varepsilon \sum_{r, s=1}^p \sum_{i =1}^{n(r)} \sum_{j=1}^{n(s)} e_{i1}^{(r)} e_{1i}^{(r)} a e_{j1}^{(s)} e_{1j}^{(s)}
= \left(\sum_{r=1}^p \sum_{i =1}^{n(r)} e_{ii}^{(r)} \right) a \left( \sum_{s=1}^p \sum_{j=1}^{n(s)}e_{jj}^{(s)} \right)
=a
\end{align*}
We also have $\sum_{k=1}^m \i< \eta_k, \eta_k >= \sum_{r=1}^p \sum_{i=1}^{n(r)} e_{i1}^{(r)} \i< \xi, e_{11}^{(r)} \xi > e_{1i}^{(r)} \leq \sum_{i=1}^{n(r)} e_{ii}^{(r)} = 1_A$.
\end{proof}
The following proposition is important in terms of Theorem \ref{thm-D}.
\begin{proposition}\label{prop-fin-dim}
Let $B \subset A$ be a unital inclusion of $\rC^*$-algebras with nondegenerate conditional expectation $E$.
Suppose that $B$ is finite dimensional.
Then, $A$ is nuclear if and only if there exists a unital countably generated $(Z, \pi_Z) \in \Corr (B)$ such that $Z$ admits a $B$-central vector $\zeta \in Z$ with $\i< \zeta, \zeta > =1_B$ and $(A, B, E)$ is strongly nuclear via $(Z, \pi_Z)$.
\end{proposition}
\begin{proof}
The `only if' part follows from Proposition \ref{prop-nuc-exact}.
Suppose that $A$ is nuclear.
Take a faithful $*$-representation $\pi_H : B \to \lB (H)$.
By Example \ref{ex-nuc}, we have $(A, \lambda_A) \prec_\univ (L^2(A, E) \otimes_B (H \otimes B) \otimes_B A, \pi_E \otimes 1_{H \otimes B} \otimes 1_A )$.
Thanks to the previous proposition and Proposition \ref{prop-int-tensor}, it enoughs to find a unital countably generated $(Z, \pi_Z) \in \Corr (B)$ containing a $B$-central normal vector and satisfying that $(H \otimes B, \pi_H \otimes 1) \prec_\univ (Z, \pi_Z)$.
Write $B = \bigoplus_{r=1}^p \lM_{n(r)}$ and set $n:=\sum_{r=1}^p n(r)$.
Let $\iota : B \hookrightarrow \lM_n$ be the natural embedding.
Since $\lM_n$ is simple, this embedding is full, and hence we have $(H \otimes \lM_n, \pi_H \otimes 1) \prec_\univ (\lM_n, \iota)$ as $B$-$\lM_n$ C$^*$-correspondences by Example \ref{ex-nuc}.
Let $\Phi : \lM_n \to B$ be the canonical trace preserving conditional expectation.
We then have
\[
(H \otimes B, \pi_H \otimes 1_B) \prec_\univ (H \otimes L^2(\lM_n, \Phi ), \pi_H \otimes 1 ) \prec_\univ (L^2(\lM_n, \Phi), \pi_\Phi).
\]
Therefore, $(L^2(\lM_n, \Phi), \pi_\Phi \circ \iota ) \in \Corr (B)$ and $\xi_\Phi \in L^2(\lM_n, \Phi)$ are the desired ones.
\end{proof}
\subsection{Tensor products, subalgebras of finite index, and relative amenable groups}\label{ss-exam-other}
\begin{example}[Tensor products with nuclear C$^*$-algebras]
Let $A$ and $B$ be unital $\rC^*$-algebra and $f: A \to \lC$ be a nondegenerate state.
Then $A$ is nuclear if and only if the triple $(A \otimes B, \lC 1_A \otimes B, f \otimes \id_B)$ is strongly nuclear.
Indeed, if $A$ is nuclear, then there exists a net $\psi_i \in \cF_{H_f \otimes A}$ approximates $\id_{A \otimes B}$.
Since $X:=L^2 (A \otimes B, f \otimes \id_B) \otimes_B( A \otimes B )$ is isomorphic to $(H_f \otimes A) \otimes B$,
the $\psi_i \otimes \id_B$ belongs to $\cF_{B'\cap X}$.
Conversely, suppose that $(A \otimes B, B,f \otimes \id_B)$ is nuclear and take a net $\varphi_i \in \cF_X$ converges to $\id_{A\otimes B}$.
If $g$ be a state on $B$, then $A \ni a \mapsto (\id_A \otimes g ) \circ \psi_i (a \otimes 1_B)$ is nuclear and converges to $\id_A$.
\end{example}
\begin{example}[Finite Watatani index]
Let $1_A \in B \subset A$ be a unital inclusion of $\rC^*$-algebras
and assume that there exists a conditional expectation $E :A \to B$ of finite Watatani index,
that is, there exists a finite family of elements , called a quasi-basis, $\{ u_1, \dots, u_n \} \in A$ such that $a = \sum_{j=1}^n u_j E (u_j^* a) = \sum_{j=1}^n E(a u_j) u_j^* $ holds for $a \in A$.
Then $(A, B, E)$ is strongly nuclear.
Recall the fact that the element $e:=\sum_{i=1}^n u_i u_i^*$ (called the index of $E$) is an invertible element in $\cZ (A)=A'\cap A$ \cite[Lemma 2.3.1]{Watatani}.
Letting $\xi:=\sum_{i=1}^n u_i \otimes u_i^* e^{-1/2} \in L^2 (A, E) \otimes_B A$ we have
\begin{align*}
(\pi_E (a) \otimes_B 1_A ) \xi
&= \sum_{i=1}^n a u_i \otimes u_i^* e^{-1/2}
=\sum_{i=1}^n \sum_{j=1}^n u_j E(u_j^* a u_i ) \otimes u_i^* e^{-1/2} \\
&= \sum_{j=1}^n u_j \otimes \sum_{i=1}^n E(u_j^* a u_i ) u_i^* e^{-1/2}
=\sum_{j=1}^n u_j \otimes u_j^* a e^{-1/2} = \xi a,
\end{align*}
hence $\xi \in A' \cap ( L^2 (A, E) \otimes_B A)$.
Since $\i< \xi, \xi >= e^{-1} \sum_{i,j=1}^n u_i E(u_i^* u_j )u_j^*=1_A$, we get $\Omega_\xi (a) =a$ for all $a\in A$.
\end{example}
\begin{example}[Finite probabilistic index]
Let $B \subset A$ be a unital inclusion of C$^*$-algebras.
A conditional expectation $E : A \to B$ is said to be {\it finite probabilistic index finite} if there exists a constant $\lambda >0$ such that
$\lambda^{-1} E -\id _A$ is a positive map on $A$.
In this case, the triple $(A, B, E)$ is nuclear.
We show that $(X, \pi_X):= (L^2(A, E) \otimes_B A ,\pi_E \otimes 1_A)$ has the UFP.
By \cite[Theorem 1]{Frank-Kirchberg} there exists $\mu >0$ such that $\mu^{-1} E - \id _A$ is completely positive.
Let $C$ be any C$^*$-algebra and $\sigma: A \odot C \to \lB (H)$ be any nondegenerate representation.
Fix $\sum_{i=1}^n a_i \otimes c_i \in A \odot C$ and $\xi \in H$ arbitrarily.
We observe that $[\mu^{-1} E ( a_i^* a_j) - a_i^* a_j]_{i,j=1}^n \in \lM_n (A)$ is positive.
Hence, we get
\begin{align*}
\| \sum_{i=1}^n \sigma (a_i \otimes c_i) \xi \| ^2
&= \i< \xi, \sum_{i,j=1}^n \sigma_A (a_i^* a_j) \sigma_C (c_i^* c_j) \xi >
\leq \mu^{-1} \i< \xi, \sum_{i,j=1}^n \sigma_A ( E ( a_i^* a_j) ) \sigma_C (c_i^* c_j) \xi >\\
&=\mu^{-1} \| \sum_{i=1}^n \pi_E (a_i ) \xi_E \otimes 1_A \otimes \sigma_C (c_i) \xi \|^2
\leq \mu^{-1} \| \sum_{i=1}^n \phi_X^H (a_i \otimes c_i ) \|^2 \| \xi \|^2.
\end{align*}
Hence, the mapping $\sum_{i=1}^n \phi_X^H (a_i \otimes c_i ) \mapsto \sum_{i=1}^n \sigma (a_i \otimes c_i)$ is bounded.
\end{example}
\begin{remark}
We note that if $E :A \to B$ is Watatani index finite, then it is probabilistic index finite (\cite[Proposition 2.6.2]{Watatani}).
When $A$ is simple, the converse holds true by \cite[Theorem 3.2]{Izumi}.
Let $\Gamma$ be a finite group acting on a unital C$^*$-algebra $A$ by $\alpha$.
By averaging we obtain a conditional expectation $E_\alpha$ from $A$ into the fixed point algebra $A^\alpha$.
It is clear that $| \Gamma|^{-1} E_\alpha - \id_A$ is positive.
Hence, if $A$ is simple, then $(A, A^\alpha, E_\alpha)$ is strongly nuclear.
Thus, there are many example of strongly nuclear triples that are not ones arising from crossed products.
\end{remark}
\begin{example}[Co-amenable subgroups]
Let $\Lambda < \Gamma$ be an inclusion of discrete groups.
Recall that $\Lambda$ is said to be {\it co-amenable} in $\Gamma$ if there exists an left invariant mean on $\ell^\infty (\Gamma /\Lambda)$ for the action of $\Gamma$ on $\Gamma / \Lambda$ by the left multiplication.
When $\Lambda$ is a normal subgroup of $\Gamma$, this is equivalent to the amenability of the quotient group $\Gamma/ \Lambda$.
The compression map by the orthogonal projection onto $\ell^2(\Lambda) \subset \ell^2(\Gamma)$ gives a conditional expectation $E : \rC^*_\re (\Gamma) \to \rC^*_\re (\Lambda)$ such that $E(\lambda_g)=0$ whenever $g \notin \Lambda$.
Then, $( \rC^*_\re (\Gamma ), \rC^*_\re (\Lambda), E)$ is nuclear if and only if $\Lambda$ is co-amenable in $\Gamma$.
Moreover,
if $\Lambda$ is a co-amenable normal subgroup of $\Gamma$, then the triple is strongly nuclear.
Suppose that $\Lambda$ is co-amenable in $\Gamma$.
Fix a finite subset $\fF \subset \Gamma$ and $\varepsilon>0$ arbitrarily.
We denote by $\pi$ the quotient map from $\Gamma$ onto $\Gamma / \Lambda$.
By \cite[Theorem 4.1]{Greenleaf} there exist a finite subset $F \subset \Gamma / \Lambda$ such that
$|F|- |g^{-1}F \cap F | < \varepsilon | F| $ for $g \in \fF$.
Take a lift $G \subset \Gamma$ of $F$, i.e., $\pi (G) = F$ and $| F| =| G |$.
Letting $\xi=|F|^{-1/2}\sum_{f \in G} \lambda_{f} \otimes \lambda_{f}^* \in L^2 (\rC^*_\re (\Gamma) , E) \otimes_{\rC^*_\re (\Lambda)} \rC^*_\re (\Gamma)$ we have
$$
\i< \xi, \lambda_g \xi > = \frac{1}{|F|} \sum_{f, h\in G} \lambda_{f} E (\lambda_{f}^* \lambda_{g } \lambda_{h} ) \lambda_{h}^* = \frac{| F \cap g^{-1} F |}{| F|} \lambda_g
\approx_\varepsilon \lambda_g$$
for all $g \in \fF$.
In the case that $\Lambda$ is a normal subgroup, the vector $\xi$ is $\rC^*_\re (\Lambda)$-central.
Conversely, if $( \rC^*_\re (\Gamma ), \rC^*_\re (\Lambda), E)$ is nuclear, then the group von Neumann algebra $L (\Gamma)$ is injective relative to $L(\Lambda)$ by Proposition \ref{prop-relative-injective}.
By \cite[Corollary 7]{Monod-Popa}, this implies the co-amenability of $\Lambda$ in $\Gamma$.
\end{example}
\begin{remark}
By \cite{Monod-Popa}\cite{Pestov} there are triples of groups $K < H < G$ such that $K$ is co-amenable in $G$ but not in $H$.
Thus, the example above says that $A:=\rC^*_\re (G)$, $B:=\rC^*_\re (H)$, and $C:=\rC^*_\re (K)$ satisfy that $(A, C, E)$ is nuclear and $(B, C, E|_B)$ is not nuclear, where $E:A \to C$ is the canonical conditional expectation as above.
We also note that there is a conditional expectation from $A$ to $B$.
\end{remark}
\subsection{Crossed products}\label{ss-exam-cro}
Let $B$ be a unital $\rC^*$-algebra and $\alpha : \Gamma \act B$ be an action of a discrete group.
Let $B \rtimes_\alpha \Gamma$ and $B \rtimes_{\alpha, \re} \Gamma$ denote the full and reduced crossed products, respectively.
We denote by $\alpha^{**}$ the action of $\Gamma$ on $B^{**}$ induced from $\alpha$.
\begin{definition}[{\cite[D\'{e}finition 4.1]{Delaroche}}]
Let $B$ be a unital $\rC^*$-algebra and $\Gamma$ be a discrete group.
An action $\alpha : \Gamma \to \Aut (B)$ is said to be {\it amenable}
if there exists a net of functions $\varphi_i :\Gamma \to \cZ(B^{**})$ with a finite support such that
\begin{itemize}
\item $\sum_{g\in \Gamma} \varphi_i (g)^* \varphi_i (g)\leq 1$ for all $i$,
\item $\sum_{g \in \Gamma} \varphi_i (g)^* \alpha^{**}_f (\varphi_i (f^{-1}g))$ converges to $1$ $\sigma$-weakly for all $f \in \Gamma$.
\end{itemize}
\end{definition}
\begin{proposition}
Let $\alpha : \Gamma \act B$ be an action of discrete group and $A:=B \rtimes_{\alpha, \re} \Gamma$.
Then, the following are equivalent{\rm :}
\begin{itemize}
\item[$(1)$] The action $\alpha : \Gamma \act B$ is amenable.
\item[$(2)$] There exists a net of vectors $\xi_i$ in the linear span of $\{ \xi x \in L^2(A, E) \otimes_B A^{**} \mid \xi \in B' \cap ( L^2 (A,E) \otimes_B A), x \in \cZ (B^{**}) \}$ such that $\i< \xi_i, \xi_i > \leq 1$ and $\Omega_{\xi_i}(a)$ converges to $a$ $\sigma$-weakly for $a\in A$.
\item[$(3)$] The triple $(A, B, E)$ is nuclear.
\end{itemize}
When $B$ is commutative, these conditions are also equivalent to
\begin{itemize}
\item[$(4)$] The triple $(A, B, E)$ is strongly nuclear.
\end{itemize}
\end{proposition}
\begin{proof}
We first prove (1) $\Rightarrow$ (2): Take a net $\{ \varphi_i \}_i \in C_\cpt (\Gamma, \cZ (B^{**}))$ in the definition of amenable actions.
We set $\xi_i := \sum_{g \in \Gamma} u_g \otimes u_g^* \varphi_i (g) \in L^2 (A,E) \otimes_B A^{**}$.
Since $\varphi_i (g) \in \cZ (B^{**})$ we get $\xi_i \in B' \cap (L^2(A,E) \otimes_B A^{**})$.
For any $b \in B$ and $f \in \Gamma$ we have
\begin{align*}
\i< \xi_i, (\pi_X (b u_f) \otimes 1_{A^{**}}) \xi_i >
&= \sum_{g \in \Gamma} \sum_{h \in \Gamma} \i< u_g \otimes u_g^*\varphi_i (g) , b u_{fh} \otimes u_h^* \varphi_i (h)> \\
&= \sum_{g \in \Gamma} \varphi_i (g)^* u_g \left ( u_g^* b u_g \right ) u_{f^{-1}g}^*\varphi_i (f^{-1}g) \\
&= \left ( \sum_{g \in \Gamma} \varphi_i(g)^* \alpha_f (\varphi_i (f^{-1} g ) ) \right )b u_f,
\end{align*}
which converges to $b u_f$ by the choice of $\varphi_i$.
We also have $\i< \xi_i, \xi_i > =\sum_{g \in \Gamma} \varphi_i(g)^*\varphi_i (g) \leq 1$.
Since $A$ is the closed span of $\{ b u_g \mid b \in B, g \in \Gamma\}$, we get (2).
We next prove (2) $\Rightarrow$ (3): Approximating $\xi_i$ in (2) by vectors in $L^2(A,E) \otimes_B A$ we get $\eta_i \in L^2(A, E) \otimes_B A$ such that $\Omega_{\eta_i}$ converges to $\id_A$ point $\sigma$-weakly.
We show (3) $\Rightarrow$ (1): Assume that $(A, B , E)$ is nuclear.
Let $B \subset \lB (H)$ be the universal representation.
Then, $A$ is realized as a $\sigma$-weakly dense subalgebra of $B^{**}\bar{\rtimes}_{\alpha^{**}} \Gamma \subset \lB (H \otimes \ell^2(\Gamma))$.
By Proposition \ref{prop-relative-injective} $B^{**}\bar{\rtimes}_{\alpha^{**}} \Gamma$ is injective relative to $B^{**}$,
and hence $\alpha : \Gamma \act B$ is amenable thanks to \cite[Proposition 3.4]{Delaroche2}.
Now we suppose that $B$ is abelian. The implication (4) $\Rightarrow$ (3) is trivial.
We show (2) $\Rightarrow$ (4):
Since $\cZ (B^{**})=B^{**}=B''$ holds,
we can approximate the $\xi_i$ in (2) by $B$-central vectors in $L^2(A, E) \otimes_B A$.
\end{proof}
\begin{proposition}
Let $B$ and $\alpha : \Gamma \act B$ be as above.
If $\Gamma$ is amenable, then the $(A, B, E)$ is strongly nuclear.
\end{proposition}
\begin{proof}
Since $\Gamma$ is amenable, there exist F\o lner sets $F_i \subset \Gamma$.
If we put $\xi_i := | F_i |^{-1/2} \sum_{g \in F_i } u_g \otimes u_g^*$,
then $\xi_i$ is $B$-central and we have $\Omega_{\xi_i} (u_g) =|F|^{-1}|F \cap g^{-1} F|u_g $, which converges to $u_g$ in norm for all $g \in \Gamma$.
\end{proof}
\if0
\begin{proposition}[{cf.\ \cite[Proposition 2.4]{Ozawa-Popa}}]
Let $B$ be unital $\rC^*$-algebra and $\alpha : \Gamma \act B$ be an action of discrete group.
Let $E_B$ denote the canonical conditional expectation from $A:=B \rtimes_{\alpha, \re} \Gamma$ onto $B$.
If $A$ admits a $\Gamma$-invariant state $\varphi$ with the GNS-representation $(H_\varphi, \pi_\varphi, \xi_\varphi)$ faithful, then the following hold true.
\begin{itemize}
\item[$(1$)] There exists a nondegenerate conditional expectation $E_\Gamma$ from $A$ onto $\rC^*_\re (\Gamma)$.
\item[$(2)$] The group $\Gamma$ is amenable if and only if $(\rC^*_\re (\Gamma), E_\Gamma) \lessdot_A^\nuc (B, E_B)$.
\end{itemize}
\end{proposition}
\begin{proof}
Since $(H_\varphi, \pi_\varphi)$ is faithful, we assume that $A \subset \lB (H_\varphi)$.
To see the first assertion we realize $A$ as a subalgebra $\rC^* (\pi (A), u_g, g \in \Gamma)$ of $\lB (H_\varphi \otimes \ell^2 (\Gamma) )$,
where $\pi (a) (\xi \otimes \delta_g) = \alpha_g^{-1} (a) \xi \otimes \delta_g$ and $u_g = 1 \otimes \lambda_g$ for $\xi \in H_\varphi$ and $\delta \in \Gamma$.
We still denote by $\varphi$ the vector state on $\lB (H_\varphi)$ corresponding to $\xi_\varphi$.
Consider the u.c.p.\ map $\Phi:=\varphi \otimes \id : \lB (H_\varphi \otimes \ell^2 (\Gamma )) \to \lB (\ell^2(\Gamma))$.
Then, $\Phi (\pi (a)) =\varphi (a) 1$ and $\Phi (u_g) =\lambda_g$ for $a\in A$ and $g \in \Gamma$.
Hence $E_\Gamma :=\Phi |_A$ is a conditional expectation from $A$ onto $\rC^*_\re (\Gamma)$.
Suppose that $x \in A$ satisfies that $E_\Gamma (a x c )=0$ for all $a, c \in A$.
Take a sequence $x_n \in \lspan \{ \pi (a) u_g \mid a \in A, g \in \Gamma \}$.
For $g \in \Gamma$, let $b_g$ and $b_g^{(n)}$ denote the $g$-th Fourier coefficient of $x$ and $x_n$, respectively.
Let $\tau$ be the canonical tracial state on $\rC^*_\re (\Gamma)$.
Since $\lim_{n \to \infty} \| b_g - b_g^{(n)} \| =0$, we have
$\varphi ( a b_g c ) \gets \varphi (a b_g^{(n)} c ) = \tau ( \lambda_g^* E_\Gamma (a x_n \alpha_g^{-1} (c) ) ) \to \tau ( \lambda_g^* E_\Gamma (a x \alpha_g^{-1} (c) ) ) =0
$
for all $a,c \in A$ and $g\in \Gamma$, which implies $x=0$.
Thus, the $E_\Gamma$ is nondegenerate.
To see the second assertion, suppose that $\Gamma$ is amenable.
We set $X_B:=L^2(A, E_B)$ and $X_\Gamma:=L^2(A, E_\Gamma)$.
Then, $(A, B, E_B)$ is (strongly) nuclear, and hence we have $A \prec_\univ X_B \otimes_B A$.
By Proposition \ref{prop-int-tensor} we get $X_\Gamma \prec_\univ X_B \otimes_B X_\Gamma$.
Conversely, suppose that $(\rC^*_\re (\Gamma), E_\Gamma) \lessdot_A^\nuc (B, E_B)$.
By Proposition \ref{prop-product-map2} there exists a u.c.p.\ map $\Psi$ from $( \lL_B (X_B ) \otimes 1 \otimes 1 )'' \subset \lB (X_B\otimes_B X_\Gamma \otimes_{\rC^*_\re (\Gamma) } \ell^2 (\Gamma) )$ onto $L(\Gamma) \subset \lB (\ell^2 (\Gamma) )$.
Now consider an injective $*$-homomorphism $\theta : \ell^\infty (\Gamma) \to \lB (X_B \otimes_B X_\Gamma \otimes_{\rC^*_\re (\Gamma)} \ell^2 (\Gamma) )$ defined by $\theta (\delta_g) =u_g e_B u_g^* \otimes 1 \otimes 1$.
Now for any $\psi \in \ell^\infty (\Gamma)$ and $g \in \Gamma$ we have $\tau (\Psi( \theta ( g.\psi)) =\tau ( \Psi ( u_g \theta (f) u_g^*) ) =\tau (u_g \Psi (\theta (\psi ) )u_g^* ) = \tau ( \Psi ( \theta (\psi )))$, where $(g.\psi ) (h) = \psi (g^{-1} h)$.
Therefore, $\psi \mapsto \tau ( \Psi ( \theta (\psi )))$ defines a left invariant mean on $\Gamma$.
\end{proof}
\fi
\subsection{Groupoids}\label{ss-exam-grpd}
\begin{definition}
A {\it groupoid} $\cG$ is given by a unit space $\cG^{(0)}$ and range and source maps $r, s : \cG \to \cG^{(0)}$
and the multiplication $\cG^{(2)}:= \{ (g, h ) \in \cG \times \cG \mid s (g) = r (h) \} \to \cG; (g, h) \mapsto gh$
satisfying that
\begin{itemize}
\item $s (gh) =s(h)$ and $r (gh) =r (g)$ for $(g,h) \in \cG^\comp$,
\item $s(x) = r(x) = x$ for $x \in \cG^\unit$,
\item $g= r(g) g = g s (s)$ for $g \in \cG$,
\item every $g$ has the inverse $g^{-1} \in \cG$ such that $gg^{-1} =r (g)$ and $g^{-1}g=s(g)$.
\end{itemize}
\end{definition}
A {\it topological groupoid} is a groupoid with a topology such that range, source, multiplication and inverse maps are continuous.
Here the topology on $\cG^\comp$ is the relative topology in $\cG \times \cG$.
A topological groupoid is said to be {\it \'{e}tale} (or {\it $r$-discrete}) if $s$ and $r$ are local homermorphisms, i.e., for any $g \in \cG$ there exists a open neighborhood $U$ of $g$ such that $r|_U : U \to r (U)$ and $s|_U : U \to s(U)$ are homeomorphisms.
\medskip
In what follows, $\cG$ denotes a locally compact Hausdorff \'{e}tale groupoid.
For $x, y \in \cG^\unit$ we set $\cG_x := s^{-1}(x)$, $\cG^y:=r^{-1} (y)$, and $\cG_x^y:= \cG_x \cap \cG^y$.
We note that $\cG_x$ is discrete in $\cG$ and $\cG^\unit$ is clopen in $\cG$.
For $\varphi, \psi \in C_\cpt (\cG)$ we define the convolution product and the adjoint by
$$
\varphi * \psi (g) := \sum_{hk =g} \varphi (h) \psi (k), \quad \varphi^* (g) := \overline{\varphi (g^{-1})}
$$
Since $\cG^\unit$ is clopen in $\cG$, we have $C_\cpt ( \cG^\unit) \subset C_\cpt (\cG)$.
We note that for $\varphi \in C_\cpt (\cG)$ and $\psi \in C_\cpt (\cG^\unit)$ we have
$$
\varphi * \psi (g) = \varphi (g) \psi ( s(g)) \quad \psi* \varphi (g) = \psi (r(g)) \varphi (g).
$$
In particular, for $\psi, \psi' \in C_\cpt ( \cG^\unit)$ we get $\psi* \psi'(x) = \psi (x)\psi'(x)$.
Therefore, if $\cG^\unit$ is compact, the $\chi_{\cG^\unit}$ forms the unit of $(C_\cpt (\cG), *)$.
We also note that $C_\cpt (\cG)$ forms an inner product $C_\cpt (\cG^\unit)$-module in the following way:
for $\xi, \eta \in C_\cpt (\cG)$ and $\varphi \in C_\cpt (\cG^\unit)$, define
$$
(\xi \varphi ) (g) = \xi (g) \varphi (s (g)) \quad \i< \xi, \eta >(x) = \sum_{g \in \cG_x} \overline{\xi (g)} \eta (g).
$$
We denote by $L^2(\cG)$ the Hilbert $C_\cpt (\cG)$-module given by completion of $C_\cpt (\cG)$.
Define the {\it left regular representation} $\lambda : C_\cpt (\cG )\ni \varphi \mapsto \lambda (\varphi) \to \lL_{C_\cpt (\cG)} ( L^2 (\cG))$ by $\lambda ( \varphi) \psi:=\varphi * \psi$.
The {\it reduced groupoid $\rC^*$-algebra} $\rC^*_\re (\cG)$ of $\cG$ is the completion of the image of $\lambda$ in $\lL_{C_\cpt (\cG)} ( L^2(\cG))$.
As noted above $\rC^*_\re (\cG)$ is unital if $\cG^\unit$ is compact.
The restriction map from $C_\cpt (\cG)$ onto $C_\cpt (\cG^\unit)$ gives the conditional expectation $E$ from $\rC^*_\re (\cG)$ onto $C_\cpt ( \cG^\unit)$.
\begin{definition}
We say that $\cG$ is {\it amenable} if there exists a net of positive functions $\mu_i \in C_\cpt (\cG)$ such that
$$
\lim_i \sup_{g \in K} | \sum_{h \in \cG_{r (g)}} \mu_i (h) - 1 | =0, \quad \lim_i \sup_{g \in K} \sum_{h \in \cG_{r(g)}} | \mu_i (hg ) - \mu_i (h) | =0
$$
for all compact subset $K$ of $\cG$.
\end{definition}
\begin{lemma}[{\cite[Lemma 5.6.12]{Brown-Ozawa}}]
For $\varphi \in C_\cpt (\cG)$ we set $\| \varphi \|_{I,s}:= \sup_{x \in \cG^\unit} \sum_{ g \in \cG_x} | \varphi (g)|$.
Then, it follows that $\| \lambda (\varphi )\| \leq \max \{ \| \varphi \|_{I,s}, \| \varphi^* \|_{I,s} \}$.
\end{lemma}
To simplify our proofs, we use the following ad hoc notation and terminology.
For a subset $K \subset \cG$ and $x, y \in \cG^\unit$ we set $K_x:=K \cap \cG_x $, $K^y:=K \cap \cG^y$, and $K_x^y:=K_x \cap K^y$.
We also say that a subset $U \subset \cG$ is a {\it $\cG$-set} if $r|_U$ and $s|_U$ are homeomorphisms.
We note that if $U$ is a $\cG$-set, then $|U_x| \leq 1$ for all $x\in \cG^\unit$.
\begin{lemma}
For any compact subset $K \subset \cG$ we have $C_K:= \sup_{x \in \cG^\unit} \max \{ |K_x|, |K^x| \} < \infty$.
\end{lemma}
\begin{proof}
By the compactness of $K$, there exists a finite family of open $\cG$-sets $\{ U_i \}_{i \in \cI}$ which covers $K$.
Fix $x\in \cG^\unit$.
If $g \in K_x$ belongs to $g \in U_i$, then we have $(U_{i})_x =\{ g \}$.
This implies that $\sup_{x \in \cG^\unit} |K_x| \leq |\cI|$.
\end{proof}
\begin{lemma}
Let $V \subset \cG$ be an open $\cG$-set and $\nu \in C_\cpt (\cG)$ be arbitrary.
If $\supp (\nu ) \subset V$ holds,
then $\lambda_\nu^* \otimes \lambda_\nu$ is a $C_0 (\cG^\unit)$-central vector in $L^2( \cG) \otimes_{C_0 (\cG^\unit)} \otimes \rC^*_\re (\cG)$.
\end{lemma}
\begin{proof}
Set $r_V:=r|_V : V \to r (V)$ and $s_V:=s|_V: V \to s(V)$.
Fix $\varphi \in C_\cpt (\cG^\unit)$ arbitrarily.
Since $\cG$ is locally compact and Hausdorff, we can find a open subset $U \subset \cG$ such that $\supp (\nu) \subset U \subset \overline{U} \subset V$.
By Urysohn's lemma we can find a function $\psi \in C_\cpt ( \cG^\unit)$ such that $0 \leq \psi \leq 1$, $\psi |_{s(U)} =1$, and $\supp (\psi) \subset s (V)$.
Let $\theta:=s_V \circ (r_V)^{-1} : r (V) \to s (V)$.
Since $\supp ((\varphi \psi ) \circ \theta ) \subset r_V \circ (s_V)^{-1}( \supp (\psi)) \subset r (V)$,
we can extend $(\varphi \psi ) \circ \theta$ to $\widetilde{\varphi} \in C_\cpt (\cG^\unit)$ in such a way that $\widetilde{\varphi}(x)=0$ for $x \in \cG^\unit \setminus r (V)$.
We will show that $\varphi * \nu^* = \nu^* * \widetilde{\varphi}$.
We observe that both of $\varphi * \nu^*$ and $\nu^* * \widetilde{\varphi}$ vanish on $\cG \setminus U^{-1}$.
For $g \in U$ we get $\varphi* \nu^* (g^{-1}) = \varphi (r (g^{-1})) \nu^* (g^{-1}) =\varphi (s (g))\nu^*(g^{-1}) =(\varphi \psi) (s (g))\nu^*(g^{-1})= \widetilde{\varphi}( r(g)) \nu^*(g^{-1}) =\nu^*(g^{-1}) \widetilde{\varphi} ( s(g^{-1}))= \nu^* * \widetilde{\varphi} (g^{-1})$.
Thus, we have $\varphi * \nu^* = \nu^* * \widetilde{\varphi}$, and hence $\lambda( \varphi ) \lambda (\nu^*) \otimes \lambda (\nu) = \lambda ( \nu^* * \widetilde{\varphi}) \otimes \lambda (\nu ) = \lambda( {\nu}) ^* \otimes \lambda ( \widetilde{\varphi}*\nu) =\lambda (\nu)^*\otimes \lambda (\nu) \lambda (\varphi) $.
\end{proof}
\begin{theorem}
A locally compact Hausdorff \'{e}tale groupoid $\cG$ with $\cG^\unit$ compact is amenable if and only if $(\rC^*_\re (\cG), C (\cG^\unit), E)$ is strongly nuclear.
\end{theorem}
\begin{proof}
If $(\rC^*_\re (\cG), C (\cG^\unit), E)$ is strongly nuclear, then $\rC^*_\re (\cG)$ is nuclear by Proposition \ref{prop-nuc-exact}.
By \cite[Theorem 5.6.18]{Brown-Ozawa} this implies the amenability of $\cG$.
Conversely, suppose that $\cG$ is amenable.
Fix a finite subset $\fF \subset C_\cpt (\cG)$ and $0< \varepsilon <1/10$ arbitrarily.
We will show that there exists a $C (\cG^\unit)$-central vector $\xi \in L^2( \cG) \otimes_{C (\cG^\unit)} \rC^*_\re (\cG)$ such that
$ \| \lambda (\varphi ) - \Omega_\xi (\lambda (\varphi) ) \| < 9 \varepsilon $ for all $\varphi \in \fF$.
Since $\cG^\unit$ is compact, we may assume that $1 \in \fF$.
Set $K:=\bigcup_{\varphi \in \fF} \supp (\varphi) \cup \supp (\varphi)^{-1}$.
Then, $K$ is compact.
By the amenability of $\cG$, we can find $\mu \in C_\cpt (\cG)_+$ such that
\begin{equation*}
\sup_{ g \in K } | 1 - \sum_{h \in \cG_{r (g)} }\mu (h) | < \frac{ \varepsilon}{ C_K}, \quad
\sup_{ g \in K } \sum_{h \in \cG_{r(g)}} | \mu (h) - \mu (hg) | < \frac{\varepsilon^2}{ C_K^2}.
\end{equation*}
We note that for any $g \in K$ and $h \in \cG_{r (g)}$ it follows that $\mu (h), \mu (hg) \leq 3$.
Indeed, $\mu (h) \leq \sum_{h \in \cG_{r (g)}} \mu (h) \leq | \sum_{h \in \cG_{r (g)}} \mu (h) - 1 | + 1 \leq 2$ and
$\mu (hg) \leq | \mu (hg) - \mu (h) | + \mu (h) \leq 3$.
Since $\cG$ is locally compact and Hausdorff, we can find a compact subset $K_\mu$ and open subset $O$ in such a way that
$\supp (\mu ) \subset O \subset K_\mu$.
By continuity of $\mu^{1/2}$, there exists a finite open covering $\{ U_i \}_{i \in \cI}$ of $\supp (\mu)$ consisting of $\cG$-sets and $g_i \in U_i$ such that
$| \mu (g_i )^{1/2} - \mu (h)^{1/2} | < \varepsilon / (C_{K_\mu} C_K)$ for all $h \in U_i$ and $ i \in \cI$.
We may assume that $U_i \subset O \subset K_\mu$.
Let $\{ \nu_i \}_{i \in \cI}$ be a corresponding partition of unity and set $\beta := \sum_{ i\in \cI} \mu (g_i)^{1/2} \nu_i$.
We note that $\| \beta - \mu^{1/2} \|_\infty <\varepsilon/ (C_{K_\mu} C_K).$
Set $\theta_i:=\nu_i^{1/2}$ and $\xi:=\sum_{i \in \cI} \mu (g_i)^{1/2} \lambda ( \theta_i )^* \otimes \lambda( \theta_i) $.
By the preceding lemma this $\xi$ is $C ( \cG^\unit)$-central.
For $\varphi \in \fF$ we denote by $\widehat{\varphi}$ the element in $C_\cpt (\cG)$ corresponding to $\Omega_\xi (\lambda_\xi)$.
We fix $g \in K$ arbitrarily and set $y:=r (g)$.
For any $\psi \in C_\cpt (\cG)$ we define $\psi (\star):=0$.
When $(U_i)_y \neq \emptyset$ we denote by $h_i$ a unique element in $(U_i)_y$.
When $(U_i)_y =\emptyset$, we set $h_i:=\star$.
We also define $\star g:=\star$.
We show that
\begin{equation}\widehat{\varphi} (g) = \left( \sum_{i, j \in \cI} \mu (g_i )^{1/2}\mu (g_j)^{1/2} \nu_i (h_i) \nu_j ( h_ig) \right) \varphi (g). \label{eq-grpd}
\end{equation}
Firstly, letting $\varphi_{ij}:=\theta_i * \varphi * \theta_j^* |_{\cG^\unit}$ we have
$$\i< \xi, \lambda (\varphi) \xi >= \sum_{i,j \in \cI}\mu (g_i)^{1/2}\mu (g_j)^{1/2} \lambda (\theta_i) ^* E ( \lambda ( \theta_i * \varphi * \theta_j^* ) ) \lambda ( \theta_j) = \sum_{i,j \in \cI} \mu (g_i)^{1/2}\mu (g_j)^{1/2} \lambda ( \theta_i^* *\varphi_{ij} * \theta_j ).$$
Hence, for $i,j \in \cI$ we have
\begin{equation}
\theta_i^* *\varphi_{ij} * \theta_j (g)
=\sum_{h \in \cG^y} \sum_{k \in \cG^{s(h)}} \theta_i^* (h) \varphi_{ij}(k ) \theta_j (k^{-1}h^{-1}g)
=\theta_i (h_i) \varphi_{ij}(r(h_i)) \theta_j (h_ig). \label{eq-grpd-2}
\end{equation}
Set $z:=r (h_i)$.
The equation (\ref{eq-grpd-2}) implies that $h_i g$ belongs to $(U_j)^{z}$ whenever $\theta_i^* *\varphi_{ij} * \theta_j (g) \neq 0$.
In this case, we have
\begin{align*}
\varphi_{ij} (z)
&=\theta_i * \varphi * \theta_j^* (z)
=\sum_{h \in \cG^z} \sum_{k \in \cG^{s(h)}} \theta_i (h ) \varphi (k) \theta_j^* (k^{-1} h^{-1}z )\\
&=\theta_i (h_i) \sum_{k \in \cG^y} \varphi (k) \theta_j (h_i k)
=\theta_i (h_i) \varphi (g) \theta_j (h_ig).
\end{align*}
Thus, we get (\ref{eq-grpd}).
Next, we take a subset $\cJ \subset \cI$ in such a way that the sets $\cI_i:= \{j \in \cI \mid h_i = h_j \}, i \in \cJ$ satisfy that $\cI = \bigsqcup_{i \in \cJ} \cI_i$.
We then have
\begin{align*}
\sum_{i,j \in \cI}\mu (g_i)^{1/2}\mu (g_j)^{1/2}\nu_i (h_i) \nu_j (h_ig)
&= \sum_{i \in \cI}\mu (g_i)^{1/2}\nu_i (h_i) \sum_{j \in \cI} \mu (g_j)^{1/2} \nu_j (h_ig) \\
&= \sum_{i_0 \in \cJ} \sum_{i \in \cI_{i_0}} \mu (g_i)^{1/2}\nu_i (h_i) \sum_{j \in \cI} \mu (g_j)^{1/2} \nu_j (h_{i_0}g) \\
&= \sum_{i_0 \in \cJ} ( \sum_{i \in \cI_{i_0}} \mu (g_i)^{1/2}\nu_i (h_{(i_0)})) ( \sum_{j \in \cI} \mu (g_j)^{1/2}\nu_j (h_{i_0}g) )\\
&= \sum_{i \in \cJ} \beta (h_i) \beta (h_ig).
\end{align*}
By the choice of $\mu$ we have
$
| 1 - \sum_{i \in \cJ} \mu ( h_i) |= | 1 - \sum_{h \in \cG_y} \mu (h) | < \varepsilon/ C_K.
$
Since $\| \varphi \|_\infty \leq 1$, $|\cJ| \leq C_{K_\mu }$,
$| \beta (h_i) | \leq \| \beta - \mu^{1/2} \|_\infty + | \mu(u_{i, y} )^{1/2} | \leq 1 + \sqrt{3} \leq 3$, and $| \beta (h_ig) | \leq \| \beta - \mu^{1/2} \|_\infty + | \mu(u_{y}^{(i)} g )^{1/2} | \leq 1 + \sqrt{3} \leq 3$,
we have
\begin{align*}
&| \widehat{\varphi} (g) - \varphi (g) |\\
&\quad \leq | \sum_{i \in \cJ} \beta (h_i) \beta (h_ig) - \mu(h_i) | + \frac{ \varepsilon}{C_K} \\
&\quad\leq | \sum_{i \in \cJ} \beta (h_i) \beta (h_ig) - \mu(h_i)^{1/2}\mu (h_ig)^{1/2} |
+ | \sum_{i \in \cJ} \mu (h_i)^{1/2} ( \mu (h_ig )^{1/2} - \mu (h_i)^{1/2} ) |
+ \frac{ \varepsilon}{C_K}\\
&\quad\leq 6 C_{K_\mu} \| \beta - \mu^{1/2}\| _\infty
+ | \sum_{i \in \cJ} \mu (h_i)^{1/2} ( \mu (h_ig )^{1/2} - \mu (h_i)^{1/2} ) |
+ \frac{ \varepsilon}{C_K}\\
&\quad\leq \frac{ 6 \varepsilon}{C_K}
+ \sum_{i \in \cJ} \mu (h_i)^{1/2} | \mu (h_ig )^{1/2} - \mu (h_i)^{1/2} |
+ \frac{ \varepsilon}{C_K}.
\end{align*}
By the Cauchy--Schwartz inequality and the fact that $|a^{1/2} - b^{1/2} |^2 \leq | a^{1/2} - b^{1/2}| (a^{1/2} + b^{1/2}) =| a - b|$ for all positive real numbers $a,b >0$,
we have
\begin{align*}
\sum_{i \in \cJ} \mu (h_i )^{1/2} | \mu (h_ig )^{1/2} - \mu (h_i)^{1/2} ) |
& \leq \left( \sum_{i \in \cJ} \mu (h_i) \right)^{1/2}\left( \sum_{i \in \cJ} | \mu (h_ig )^{1/2} - \mu (h_i)^{1/2} |^2 \right)^{1/2} \\
& \leq \left( \sum_{h \in \cG_y} \mu (h) \right)^{1/2}\left( \sum_{k \in \cG_y} | \mu (kg ) - \mu (k) | \right)^{1/2}\\
& \leq \frac{2 \varepsilon}{C_K}.
\end{align*}
Therefore, we have
$| \widehat{\varphi} (g) - \varphi (g) | \leq 9\varepsilon/C_K$
for all $g\in K$.
Since $\supp (\varphi ) \subset K$ and $\supp (\widehat{\varphi} ) \subset K$,
we now have
\begin{align*}
\| \varphi - \widehat{\varphi} \|_{I,s}
&= \sup_{x \in \cG^\unit} \sum_{g \in \cG_x }| \varphi (g) - \widehat{\varphi} (g) |
= \sup_{x \in \cG^\unit} \sum_{g \in K_x }| \varphi (g) - \widehat{\varphi} (g) |\\
&= \sup_{g \in K} C_K | \varphi (g) - \widehat{\varphi} (g) |
\leq 9 \varepsilon.
\end{align*}
Since $K=K^{-1}$, we also have
$\| \varphi^* - \widehat{\varphi}^* \|_{I,s} \leq 9 \varepsilon$.
Therefore, we get $\| \lambda (\varphi) - \Omega_\xi (\lambda( \varphi) ) \| < 9 \varepsilon$ for all $\varphi \in \fF$.
Hence, we obtain a net of $C (\cG^\unit)$-central vectors $\{ \xi_i \}_i$ such that $\Omega_{\xi_i}$ converges to the identity map on the dense subspace $C_\cpt (\cG) \subset \rC^*_\re (\cG)$.
In particular, we have $\| \Omega_{\xi_i} \|^2 = \| \i< \xi_i, \xi_i > \| \leq 2$.
Thus, $\{ \Omega_{\xi_i} \}_i$ forms a bounded net, and hence converges on the whole $\rC^*_\re (\cG)$.
\end{proof}
\setcounter{theorem}{0}
\renewcommand{\thetheorem}{\arabic{section}.\arabic{theorem}}
\section{Weyl--von Neumann--Voiculescu type results}\label{sec-voic}
The main result in this section is Theorem \ref{thm-voic} below, which says that weak containment is characterized by a certain Weyl--von Neumann--Voiculescu type assertion.
Our proof is based on Arveson's argument \cite{Arveson}.
We use the condition (3) in Theorem \ref{thm-weak} and Corollary \ref{cor-weak} instead of Glimm's lemma.
We denote by $\{ \delta_n \}_{n=1}^\infty$ the canonical basis of $\ell^2( \lN)$ and by $p_n \in \lB (\ell^2( \lN))$ the orthogonal projection onto $\lC \delta_n$.
Recall that a $\rC^*$-algebra $B$ is said to be $\sigma$-{\it unital} if it admits a countable approximate unit.
Recall that every separable $\rC^*$-algebra is $\sigma$-unital.
The following lemma can be found in the proof of \cite[Lemma 10]{Kasparov}.
Thus, we give only a sketch of proof for the reader's convenience.
\begin{lemma}\label{lem-quasicentral}
Let $A$ and $B$ be $\rC^*$-algebras with $A$ unital separable and $B$ $\sigma$-unital.
Then, for any finite subset $\fF \subset A$ and $\varepsilon >0$ there exists a sequence of positive elements $\{ e_n \}_{n=1}^\infty \subset \lK_B (H_B)$ such that each $e_n$ has the form of $\sum_{i=1}^{d(n)} p_i \otimes b_i$ for some $b_1, \dots, b_{d(n)} \in B$, $\sum_{n=1}^\infty e_n^2 =1_{H_B}$ strictly, and
\begin{align*}
a - \sum_{n=1}^\infty e_n a e_n \in \lK_B (H_B) \;\; \text{for}\; a \in A,
\quad
\| b - \sum_{n=1}^\infty e_n b e_n \| < \varepsilon \;\;\text{for}\; b \in \fF,
\end{align*}
where the infinite sums converge strictly.
\end{lemma}
\begin{proof}[Sketch]
Set $\fF_1:=\fF$ and take an increasing sequence of finite subsets $\fF_1 \subset \fF_2 \subset \cdots \subset A$ such that $\bigcup_{n=1}^\infty \fF_n$ is norm dense in $A$.
Take a countable approximate unit $\{ v_n \}_{n \geq 1}$ of $B$.
Consider the following two separable $\rC^*$-subalgebras of $\lL_B (H_B)$,
$$
\cA:=\rC^* (1_{H_B}, A, p_n \otimes v_m, n\in \lN, m \in \lN), \quad \cJ:= \rC^* (p_n \otimes v_m, n\in \lN, m \in \lN).
$$
Since $\{ \sum_{i=1}^np_i \otimes v_n \}_{n=1}^\infty$ forms an approximate unit of $\cJ \subset \lK_B (H_B) \cong \lK (\ell^2 (\lN)) \otimes B$,
applying \cite[Theorem 1]{Arveson} to $\{ \sum_{i=1}^np_i \otimes v_n \}_n$ and $\cJ \triangleleft \cA$ we obtain a countable quasicentral approximate unit $\{ u_n \}_{n=1}^\infty$ contained in the convex hull of $\{ p_n \otimes v_m \}_{n,m} \subset \lK_B (H_B)$ satisfying that $\| [ a , (u_{n}- u_{n-1})^{1/2} ] \| < \varepsilon /2^n$ for $a \in \fF_n$.
Here we set $u_0:=0$.
Then, the $e_n:=(u_n-u_{n-1})^{1/2}$ is the desired one.
\end{proof}
\begin{lemma}\label{lem-glimm}
Let $A$ and $B$ be $\rC^*$-algebras with $A$ unital and $(Y, \pi_Y) \in \Corr(A,B)$ be unital.
Let $\varphi \in \CP ( A, \lL_B (H_B) )$, $(K,\pi_K) \in {\rm Rep}(B)$, and $b_1,\dots, b_n \in B$ be given.
For the compact operator $e = \sum_{i=1}^n p_i \otimes b_i \in \lK_B (H_B)$,
we define the c.p.\ map
$$
\psi : A\to \lK_B(H_B); \quad a \mapsto e^* \varphi (a) e.
$$
If $(A\otimes_\varphi H_B, \lambda_A \otimes 1_{H_B}) \prec_K Y$,
then for any finite subsets $\fF \subset A$, $\cX \subset \lB (H_B \otimes_B K)_*$, and $\varepsilon>0$
there exist $m \in \lN$ and $V \in \lL_B (H_B, Y \otimes \lC^m)$ such that $V^* V \leq e^*\varphi (1)e$ and
$$
\left| f ( ( \psi (a)- V^* \pi^m_Y (a) V)\otimes1_K ) \right| < \varepsilon, \quad a\in \fF, f\in \cX.
$$
When the given $(K, \pi_K)$ is the universal representation of $B$, the above $V$ can be chosen in such a way that $\| \psi(a) - V^*\pi_Y^m (a)V \| <\varepsilon $ for $a \in \fF$.
\end{lemma}
\begin{proof}
For each $a \in A$, the support and range of $\psi (a)$ are contained in $\lC^n \otimes B$, we can regard $\psi(a)$ as an element in $\lK_B (\lC^n \otimes B)$.
Put $\zeta:=( 1_A \otimes (\delta_1 \otimes b_1), \dots,1_A \otimes (\delta_n \otimes b_n) ) \in (A\otimes_\varphi H_B)_n$.
Then, via the isomorphisms $\lK_B (\lC^n \otimes B) \cong \lM_n (B)$ and $\lB ((\lC^n \otimes B) \otimes_B K) \cong \lB (K^n)$,
the operator $\psi (a)$ and $\psi (a) \otimes 1_K$ are identified with $\Omega_\zeta (a)$ and $\pi_{K}^{(n)} \circ \Omega_\zeta (a)$, respectively.
By Lemma \ref{lem-tensor}, we have $(A\otimes_\varphi H_B)_n \prec_{K^n} Y_n$.
Thus, there exist $\xi^{(1)}, \dots, \xi^{(m)} \in Y_n$ such that
\begin{equation}
\left|f \circ \pi_{K}^{(n)} \left( \Omega_\zeta (a) - \sum_{k=1}^m \i< \xi^{(k)}, \pi_{Y_n} (a) \xi^{(k)} >_{\lM_n (B)} \right) \right| < \varepsilon,\quad a\in \fF, f \in \cX \label{eq-Glimm}
\end{equation}
and $\sum_{k=1}^m \i< \xi^{(k)}, \xi^{(k)} > \leq \Omega_\zeta (1) =e^*\varphi(1) e$.
Write $\xi^{(k)}=(\xi_1^{(k)}, \dots, \xi_n^{(k)} )\in Y_n$ with $\xi_i^{(k)} \in Y$ and set
$\xi_i:=(\xi_i^{(1)}, \dots, \xi_i^{(m)} ) \in Y^m$.
Define $V \in \lL_B (H_B, Y^m)$ by
$V (\delta_i \otimes b) = \xi_i b$ for $1 \leq i \leq n$ and $V (\delta_i \otimes b)=0$ for $n<i$.
For $a \in A$, via the isomorphism $\lK_B (\lC^n \otimes B) \cong \lM_n (B)$, the operator $V^* \pi_Y^m (a) V$ is identified with $[ \i< \xi_i, \pi_Y^{m} (a) \xi_j >]_{i,j=1}^n$.
On the other hand, one has
\begin{align*}
\sum_{k=1}^m \i< \xi^{(k)} , \pi_{Y_n} (a) \xi^{(k)} >_{\lM_n (B)}
=\left[ \sum_{k=1}^m \i< \xi_i^{(k)}, \pi_Y (a) \xi_j^{(k)} >_B \right]_{i,j=1}^n
=\left[ \i< \xi_i, \pi_Y^{m} (a) \xi_j >_B \right]_{i,j=1}^n
\end{align*}
Hence, $V$ is the desired one.
In the case that $A \otimes_\varphi H_B \prec_\univ Y$,
by Corollary \ref{cor-weak},
we can choose $\xi^{(1)}, \dots, \xi^{(m)}$ in such a way that
$\| \Omega_\eta (a) - \sum_{k=1}^m \i< \xi^{(k)}, \pi_{Y_n} (a) \xi^{(k)} > \| < \varepsilon$ for $a \in \fF$ instead of (\ref{eq-Glimm})
\end{proof}
\begin{theorem}\label{thm-voic}
Let $A$ and $B$ be $\rC^*$-algebras with $A$ unital and $B$ $\sigma$-unital and $X$ be a countably generated Hilbert $B$-module.
For unital $(Y, \pi_Y) \in \Corr(A,B)$ and given $\varphi \in \CP ( A, \lL_B (X) )$ and $(K,\pi_K) \in {\rm Rep}(B)$, the following are equivalent:
\begin{itemize}
\item[$(1)$] $(A\otimes_\varphi X, \lambda_A \otimes 1_X) \prec_{(K,\pi_K)} (Y, \pi_Y)$.
\item[$(2)$] There exists a net $V_i \in \lL_B (X, Y^\infty)$ with $\|V_i^*V_i\| \leq \| \varphi (1_A) \| $ and $( V_i^*\pi^\infty_Y (a)V_i ) \otimes 1_K$ converges to $\varphi(a) \otimes 1_K$ in the $\sigma$-weak topology on $\lB (X \otimes_B K)$ for all $a \in A$.
\end{itemize}
Further suppose that $\varphi$ is unital, and $(K,\pi_K)$ is the universal representation of $B$.
Then, any of conditions above is also equivalent to
\begin{itemize}
\item[$(3)$] There exists a net of isometries $V_i \in \lL_B (X, Y^\infty)$such that $\lim_{i} \| V_i^* \pi^\infty_Y (a) V_i- \varphi (a) \| =0$ for all $a\in A$.
When $A$ is separable, we can choose the $V_i$ as a sequence satisfying that $V_i^* \pi^\infty_Y (a) V_i- \varphi (a) \in \lK_B (X)$ for all $a \in A$.
\end{itemize}
\end{theorem}
\begin{proof}
First, we show (1) $\Rightarrow$ (2):
It suffices to show the case that $X=H_B$.
Indeed, by Kasparov's stabilization theorem there exists a projection $P \in \lL_B (H_B)$ such that $PH_B\cong X$.
Hence, we can regard $\varphi$ as a map into $\lL_B (H_B)$.
Since $A \otimes_\varphi X =A\otimes_\varphi H_B$, we get $A\otimes_\varphi H_B \prec_K X$.
If we obtain a net $V_i \in \lL_B (H_B, Y^\infty)$ as in (2), then $V_iP \in \lL _B (X,Y^\infty)$ does the job.
Fix $\varepsilon>0$ and finite subsets $\fF \subset A$ and $\cX \subset \Ball ( \lB (H_B \otimes_B K)_*)$ arbitrarily.
It suffices to show that there exists $V \in \lL_B (H_B, Y^\infty )$ such that $\| V^*V \| \leq \| \varphi(1) \| $ and
$$
| f ((\varphi (a) - V^* \pi_Y^\infty (a) V) \otimes 1_K )| < \varepsilon, \quad a\in \fF, f\in \cX.
$$
When $A$ is separable, take an increasing sequence of finite subsets $\fF=\fF_1 \subset \fF_2 \subset \dots \subset A$ such that $A $ is the norm closure of $\bigcup_{n=1}^\infty \fF_n$. When $A$ is non-separable, set $\fF_n:=\fF$ for all $n\in \lN$.
Let $\cA$ be the separable $\rC^*$-subalgebra of $\lL_B (H_B)$ generated by $\bigcup_{n=1}^\infty \varphi (\fF_n)$.
By Lemma \ref{lem-quasicentral} we obtain a sequence of positive operators $\{ e_n \}_{n=1}^\infty$ such that each $e_n$ has the form of $\sum_{i=1}^{d(n)} p_i \otimes b_i$, $\sum_{n=1}^\infty e_n^2=1_{H_B}$ strictly, and
\begin{align*}
x - \sum_{n=1}^\infty e_n x e_n \in \lK_B (H_B) \;\; \text{for}\; x \in \cA,
\quad
\| \varphi (a) - \sum_{n=1}^\infty e_n \varphi (a) e_n \| < \varepsilon /2 \;\;\text{for}\; a \in \fF.
\end{align*}
Let $\psi_n : A \to \lL_B (H_B )$ be the c.p.\ map defined by $e_n$ as in Lemma \ref{lem-glimm}.
By Lemma \ref{lem-glimm} there exist $d(n) \in \lN$ and $W_n \in \lL_B ( H_B, Y^{d(n)})$ such that $W_n^*W_n \leq e_n \varphi (1) e_n$ and
\begin{equation}
\left| f \left( ( \psi_n(a) - (W^*_n \pi_Y ^{d(n)} (a) W_n ) \otimes 1_K \right) \right| < 2^{-n} \varepsilon, \quad a \in \fF_n, \; f \in \cX. \label{eq-voic}
\end{equation}
For any $\xi \in H_B$ and $N \in \lN$ one has $\sum_{n=1}^N \i< W_n \xi ,W_n\xi > \leq \| \varphi(1) \| \i< \xi, \sum_{n=1}^N e_n^2 \xi > \leq \| \varphi (1) \| \i< \xi, \xi>$, and hence $V:=\bigoplus_n W_n : H_B \to \bigoplus_n Y^{d(n)} \cong Y^\infty$ is well-defined and satisfies that $V^*V \leq \| \varphi (1) \| 1$.
We observe that that $V^* \pi^\infty_Y (a) V = \sum_{n=1}^\infty W_n^* \pi_Y^{d(n)} (a) W_n$ for $a\in A$.
Now for any $a\in \fF$ and $f\in \cX$ we have
\begin{align*}
&| f( ( \varphi(a) - V^*\pi_Y^\infty (a) V )\otimes 1_K) | \\
&\qquad \leq \| \varphi (a) - \sum_{n=1}^\infty e_n \varphi (a) e_n \|
+ |f( \sum_{n=1}^\infty ( \psi_n (a) - W_n^* \pi^{d(n)}_Y (a) W_n )\otimes 1_K ) |\\
&\qquad \leq \| \varphi (a) - \sum_{n=1}^\infty e_n \varphi (a) e_n \|
+ \sum_{n=1}^\infty \left| f \left( ( \psi_n(a) - W^*_n \pi_Y^{d(n)} (a) W_n) \otimes 1_K \right) \right|\\
&\qquad < \varepsilon.
\end{align*}
In the case that $A \otimes_\varphi H_B \prec_\univ Y$,
by Lemma \ref{lem-glimm} we can choose $W_n$ in such a way that
\begin{equation*}
\left \| \psi_n(a) - W^*_n \pi_Y^{d(n)} (a) W_n \right\| < 2^{-n} \varepsilon, \quad a\in \fF_n
\end{equation*}
instead of (\ref{eq-voic}).
This implies that $\| \varphi(a) - V^*\pi^\infty_Y (a) V \| <\varepsilon$ for all $a\in \fF$.
We also have $\sum_{n=1}^\infty \| \psi_n (a) - W_n^* \pi_Y^{d(n)} (a) W_n \| < \infty$ holds for all $a \in \bigcup_n \fF_n$.
Thus, if $A$ is separable, then $\varphi(a) - V^*\pi^\infty_Y (a) V \in \lK_B (H_B)$ for all $a\in A$. Moreover, if $\varphi$ is unital, then we can choose $V$ in such a way that $\| 1 - V^*V \| < \varepsilon<1$. Set $V_0:=V(V^*V)^{-1/2}$. Then $V_0$ is an isometry and enjoys that $\| \psi(a) - V_0^*\pi^\infty_Y(a) V_0 \| < \delta(\varepsilon)$ for $a\in \fF$, where $\delta(\varepsilon)$ is a positive number such that $\lim_{\varepsilon \searrow 0} \delta(\varepsilon)=0$, and hence we get (1) $\Rightarrow$ (3).
Finally, we prove (2) $\Rightarrow$ (1):
Suppose that we have a net $V_i \in \lL_B (X, Y^\infty)$ in (2).
Fix $\xi = \sum_{k=1}^m a_k \otimes \xi_k \in A\odot H_B$ arbitrarily.
Set $\zeta_i := \sum_{k=1}^m \pi_Y^{\infty} (a_k) V_i \xi_k \in Y^\infty$ and show that $\pi_K \circ\Omega_{\zeta_i}$ converges to $\Omega_\xi$ point $\sigma$-weakly.
Since $\{ \pi_K \circ\Omega_{\zeta_i} \}_i$ is norm bounded, it suffices to show convergence in the point weak operator topology.
For any $a\in A$ and $\eta \in K$ we have
\begin{align*}
\i< \eta, \pi_K\circ \Omega_{\zeta_i} (a) \eta >
&= \sum_{k,l=1}^m \i< \xi_k \otimes \eta, ( V_i^* \pi_Y^\infty (a_k a a_l ) V_i \xi_l ) \otimes \eta >\\
&\rightarrow \sum_{k,l=1}^m \i< \xi_k \otimes \eta, (\varphi (a_k a a_l ) \xi_l ) \otimes \eta >\\
&=\pi_K \circ \Omega_\xi (a).
\end{align*}
By Theorem \ref{thm-weak} we get $A\otimes_\varphi H_B \prec_K Y^\infty$.
Since $Y^\infty \prec_\univ Y$ holds, we are done.
\end{proof}
Here is a characterization of weak containment for unital countably generated C$^*$-correspondences.
\begin{corollary}
Let $A$ and $B$ be $\rC^*$-algebras with $A$ unital and $B$ $\sigma$-unital.
For unital $\rC^*$-correspondences $X,Y \in \Corr(A,B) $ with $X$ countably generated and a representation $K \in \Rep (B)$, $X$ is weakly contained in $Y$ with respect to $K$ if and only if there exists a net of contractions $V_i \in \lL_B (X, Y^\infty)$ and $V_i^* \pi_Y^\infty (a) V_i \otimes 1_K$ converges to $\pi_X(a) \otimes 1_K$ $\sigma$-weakly in $\lB (X \otimes_B K)$ for all $a\in A$.
\end{corollary}
\begin{corollary}\label{cor-voic}
Let $A$ and $B$ be $\rC^*$-algebras with $A$ unital and $B$ $\sigma$-unital.
For unital $\rC^*$-correspondences $(X,\pi_X), (Y,\pi_Y ) \in \Corr(A,B) $ with $X$ countably generated, $(X, \pi_X) \prec_\univ (Y,\pi_Y)$ if and only if there exists a net of isometries $V_i \in \lL_B (X, Y^\infty)$ such that $\lim_i \| V_i \pi_X(a) - \pi^\infty_Y (a) V_i \| =0$ for all $a\in A$.
When $A$ is separable, we may choose $V_i$'s as a sequence satisfies that $V_i \pi_X (a) - \pi^\infty_Y (a) V_i \in \lK_B (X, Y^\infty)$ for all $a\in A$.
\end{corollary}
\begin{proof}
The ``if part'' follows from the preceding theorem.
Let $V \in \lL_B (X, Y^\infty )$ be an isometry.
One has
\begin{align*}
(V \pi_X (a) - \pi^\infty_Y (a) V)^*(V \pi_X (a) - \pi^\infty_Y (a) V)
&=\pi_X (a^*) (\pi_X (a)- V^* \pi_Y^\infty (a) V)\\
&+ (\pi_X (a^*) - V^*\pi_Y^\infty (a^*)V )\pi_X (a) \\
&+ V^* \pi_Y^\infty (a^*a ) V - \pi_X (a^*a ).
\end{align*}
Thus, the assertion follows from Theorem \ref{thm-voic}.
\end{proof}
We note that $V_i$ in the preceding corollary satisfies $\pi^\infty_Y (a) V_iV_i^* - V_i V_i^* \pi^\infty_Y (a)$ converges to 0 in norm, and is compact if $A$ is separable.
Indeed,
$$
\pi_Y^\infty (a) V_iV_i^* - V_i V_i^* \pi_Y^\infty (a)
=(\pi_Y^\infty (a) V_i - V_i\pi_X (a) )V_i^* +V_i( \pi_X (a) V_i^* - V_i^* \pi_Y^\infty (a)).
$$
\begin{definition}
Let $A$ and $B$ are $\rC^*$-algebras. For $(X,\pi_X), (Y,\pi_Y) \in \Corr (A,B)$
we say that $(X,\pi_X)$ and $(Y,\pi_Y)$ are {\it approximately unitarily equivalent}, written $(X,\pi_X) \sim (Y,\pi_Y)$ if there exists a sequence of unitaries $U_n \in \lL_B (X,Y)$ such that $\pi_X (a) - U^*_n \pi_Y(a) U_n \in \lK_B (X)$ and $\lim_{n\to \infty} \|\pi_X (a) - U^*_n \pi_Y(a) U_n \|=0$ for all $a \in A$.
\end{definition}
We introduce a notation which is useful to prove the next theorem.
Let $X$ and $Y$ be Hilbert $B$-modules and $\varphi : A \to \lL_B (X)$ and $\psi :A\to \lL_B(Y)$ be maps.
For a subset $\fF \subset A$ and $\varepsilon>0$ we denote by $\varphi \sim_{(\fF, \varepsilon)} \psi$ if
there exists a unitary $U \in \lL_B (X,Y)$ such that $\varphi (a)- U^* \psi (a) U \in \lK_B (X)$ and $\| \varphi(a)- U^* \psi (a) U \| <\varepsilon$ for all $a\in \fF$.
We also write $\varphi \sim \psi$ if $\varphi \sim_{(\fF, \varepsilon)} \psi$ for any finite subset $\fF \subset A$ and $\varepsilon>0$.
\begin{theorem}\label{thm-absorbing}
Let $A$ and $B$ be $\rC^*$-algebras with $A$ unital and $B$ $\sigma$-unital.
For unital $\rC^*$-correspondences $(X, \pi_X)$, $(Y, \pi_Y) \in \Corr(A,B)$ with $X$ countably generated,
$(X, \pi_X) \prec_\univ (Y,\pi_Y)$ holds if and only if there exists a net of unitaries $U_i \in \lL_B (X \oplus Y^\infty,Y^\infty)$ such that $\lim_{i} \|\pi_X (a) \oplus \pi_Y^\infty (a) - U^*_i \pi_Y^\infty(a) U_i\|=0$ for all $a \in A$.
When $A$ is separable, this is the case that $(X \oplus Y^\infty, \pi_X \oplus \pi_Y^\infty) \sim (Y^\infty, \pi_Y^\infty)$.
\end{theorem}
\begin{proof}
Suppose that $(X, \pi_X) \prec_\univ (Y,\pi_Y)$.
We only prove the case that $A$ is separable since the proof for general $A$ proceeds in the same manner.
It suffices to show that $\pi_X \oplus \pi_Y^\infty \sim \pi_Y^\infty$.
Since $X^\infty \prec_\univ X \prec_\univ Y$ and $X^\infty$ is also countably generated,
we can apply the previous corollary and obtain a sequence of isometries $V_n \in \lL_B (X^\infty, Y^\infty)$ such that $V_n \pi_X^\infty (a) - \pi_Y^\infty (a) V_n$ is compact and converges to 0 in norm for all $a \in A$.
Put $P_n:=1-V_nV_n^* \in \lL_B (Y^\infty)$.
Then $U_n:=V_n\oplus P_n : X^\infty \oplus P_n Y^\infty\to Y^\infty$ is a unitary.
Fix $a\in A$. We have
\begin{align*}
\pi^\infty_Y (a) U_n - U_n ( \pi_X^\infty (a) \oplus P_n \pi_Y^\infty (a) P_n )
&=\pi^\infty_Y (a) P_n + \pi^\infty_Y (a) V_n - P_n \pi^\infty_Y (a) P_n - V_n \pi_X^\infty (a)\\
&=\pi^\infty_Y (a) V_nV_n^* - V_nV_n^* \pi^\infty_Y (a) + \pi^\infty_Y (a) V_n - V_n \pi_X^\infty (a),
\end{align*}
which is compact and converges to 0 by the remark above.
Hence, for any finite subset $\fF \subset A$ and $\varepsilon >0$ there exists $n \in \lN$ such that
$\pi_Y^\infty \sim_{(\fF, \varepsilon)} \pi_X^\infty \oplus P_n \pi_Y^\infty(\cdot) P_n$.
Now we have
$$
\pi_X \oplus \pi^\infty_Y \us\sim_{(\fF, \varepsilon)} \pi_X \oplus \pi_X^\infty \oplus P_n \pi_Y^\infty(\cdot) P_n \sim \pi_X^\infty \oplus P_n \pi_Y^\infty (\cdot) P_n \us\sim_{(\fF, \varepsilon)} \pi^\infty_X.
$$
This implies that $\pi_X\oplus \pi_Y^\infty \sim_{(\fF, 3\varepsilon)} \pi_Y^\infty$.
Conversely, suppose that $(X\oplus Y^\infty, \pi_X \oplus \pi_Y^\infty ) \sim (Y^\infty, \pi^\infty_Y)$.
Then, we have $(X\oplus Y^\infty , \pi_X \oplus \pi_Y^\infty ) \prec_\univ (Y^\infty, \pi^\infty_Y)$, and hence
$(X, \pi_X) \prec_\univ(X\oplus Y^\infty, \pi_X \oplus \pi_Y^\infty) \prec_\univ (Y^\infty, \pi_Y^\infty) \prec_\univ (Y, \pi_Y).
$
\end{proof}
\begin{definition}[{\cite[Definition 1.6]{Skandalis}}]
An $A$-$B$ C$^*$-correspondence $X$ is said to be {\it nuclear} if for any $n \in \lN$ and any $\xi = ( \xi_1, \dots, \xi_n) \in X_n=X\otimes \lC_n$ with $\| \xi \| \leq 1$,
the c.c.p.\ map $\Omega_\xi : A \to \lM_n (B); a \mapsto [ \i< \xi_i, \pi_X(a) \xi_j >_B ] _{i,j=1}^n$ is nuclear.
\end{definition}
The nuclearity of a given C$^*$-correspondence is characterized in terms of our weak containment.
Recall that a c.p.\ map $\theta : A \to B$ is said to be {\it factorable} if there exist $n \in \lN$ and c.p.\ maps $\varphi : A \to \lM_n$ and $\psi : \lM_n \to A$ such that $\theta = \psi \circ \varphi$.
The set of factorable maps from $A$ into $B$ is known to be convex.
By \cite[Proposition 3.8.2]{Brown-Ozawa} any c.c.p.\ map is nuclear if and only if it can be approximated by factorable c.p.\ maps in the point norm topology.
\begin{proposition}\label{prop-nuclear-bimodule}
Let $A$ and $B$ be $\rC^*$-algebras with $A$ unital.
Let $H \in \Rep (A)$ be faithful and $X \in \Corr (A, B)$ and $K \in \Rep (B)$ be given.
Then, $X \prec_\univ (H \otimes B, \pi_H \otimes 1_B)$ holds if and only if $X$ is nuclear.
\end{proposition}
\begin{proof}
Suppose that $X \prec_\univ H \otimes B$.
Since every c.c.p.\ map in $\cF_{(H\otimes B)_n}$ can be approximated by factorable maps for $n \in \lN$, $(H \otimes B, \pi_H \otimes 1_H)$ is nuclear,
and hence so is $X$ by Corollary \ref{cor-weak} and Lemma \ref{lem-tensor}.
Conversely, suppose that $X$ is nuclear and show the condition (2) in Corollary \ref{cor-weak}.
Fix $\xi \in X$ with $\| \xi \| \leq 1$, a finite subset $\fF \subset A$, and $\varepsilon >0$ arbitrarily.
Since $\Omega_\xi$ is a nuclear map, we may and do assume that $\Omega_\xi$ is of the form $\beta \circ \alpha$ for some c.c.p.\ maps $\alpha : A \to \lM_n$ and $\beta : \lM_n \to B$.
Let $(H_\alpha, \pi_\alpha, V_\alpha)$ be the Strinespring dilation of $\alpha$.
Since $(H_\alpha, \pi_\alpha)$ is weakly contained in $(H, \pi_H)$ and $\alpha=\Omega_\eta$ with $\eta = ( V_\alpha \delta_1, \dots, V_\alpha\delta_n ) \in (H_\alpha)_n$,
by Lemma \ref{lem-tensor} again, there exist $ \xi^{(1)}, \dots, \xi^{(m)} \in H \otimes \lC_n$ such that $\| \alpha (a) - \sum_{k=1}^m \Omega_{\xi^{(k)}} (a) \| < \varepsilon$ for $a\in \fF$.
Let $[ x_{ij} ]_{i,j=1}^n \in \lM_n (M)$ be the square root of the Choi matrix $[ \beta (e_{ij}) ]_{i,j=1}^n \in \lM_n (M)$.
Note that $\beta (e_{ij}) = \sum_{l=1}^n x_{li}^*x_{lj}$.
Write $\xi^{(k)}=(\xi^{(k)}_1, \dots, \xi^{(k)}_n )$ with $\xi_i^{(k)} \in H$.
Letting $\eta_l^{(k)}:= \sum_{i=1}^n \xi_k^{(i)} \otimes x_{li} \in H \otimes M$ we have
\begin{align*}
\beta \circ \alpha (a)
&\approx_\varepsilon \sum_{k=1}^m \beta ( \Omega_{\xi^{(k) }} (a) )
=\sum_{k=1}^m \beta \left( \left[ \l< \xi_i^{(k)}, \pi_H (a) \xi_j^{(k)} >\right]_{i,j=1}^n \right) \\
&=\sum_{k=1}^m \sum_{l=1}^n \sum_{i,j=1}^n \l< \xi_i^{(k)}, \pi_H (a) \xi_j^{(k)} > x_{li}^* x_{lj}
=\sum_{k=1}^m \sum_{l=1}^n \Omega_{\eta_l^{(k)}} (a)
\end{align*}
for every $a\in A$, which implies $\beta \circ \alpha$ belongs to the point norm closure of $\cF_{H \otimes A}$.
\end{proof}
We should remark that our results do not include original Voiculescu's theorem \cite{Voiculescu} as well as Kasparov's generalized one \cite[Theorem 5]{Kasparov} completely.
The following corollary follows from Theorem \ref{thm-voic} and is a particular case of \cite[Theorem 5]{Kasparov}.
\begin{corollary}
Let $A$ be unital separable $\rC^*$-algebra, $B$ be $\sigma$-unital, $\pi_H : A \to \lB (H)$ be a faithful representation with $H$ separable, and $\varphi : A \to \lL_B (H_B)$ be unital completely positive. Then $(A \otimes_\varphi H_B, \lambda_A \otimes 1_{H_B} ) $ is nuclear if and only if
there exists a sequence of isometries $V_n \in \lL_B (H_B)$ such that $\varphi (a) - V_n^* (\pi_H^\infty (a) \otimes 1_B ) V_n \in \lK_B (H_B)$ and $\lim_{n \to \infty } \| \varphi (a) - V^*_n (\pi_H^\infty (a) \otimes 1_B ) V_n \|=0$ and $a\in A$.
\end{corollary}
\renewcommand{\thetheorem}{\arabic{section}.\arabic{subsection}.\arabic{theorem}}
\section{Relative $K$-nuclearity}\label{sec-rel-K}
\subsection{Preliminaries on $KK$-theory}\label{ss-rel-K-pre}
In this section we prove that our strong relative nuclearity implies Germain's relative $K$-nuclearity.
Firstly, we recall some definitions and facts on $KK$-theory.
We refer to \cite{Blackadar} and \cite{Jensen-Thomsen} for $KK$-theory.
\begin{notation}
For a trivially graded $\rC^*$-algebra $B$, a {\it graded Hilbert $B$-module} is a Hilbert $B$-module $X$ such that there exist closed submodules $X_0$ and $X_1$ of $X$, called {\it even} and {\it odd} parts of $X$, such that $X=X_0 \oplus X_1$.
To make the grading clear, we will write $X=X_0 \hoplus X_1$.
An operator $x \in \lL_B (X_0 \hoplus X_1)$ is said to be of {\it degree} $i \in \{ 0, 1\}$ if $x X_j \subset X_{i+j \, (\text{mod } 2)}$.
A $*$-homomorphism $\phi: A \to \lL_B (X_0 \hoplus X_1)$ is said to be of degree 0 if $\phi (a)$ is of degree 0 for all $a\in A$.
In this case, there exist $*$-homomorphism $\phi_0: A \to \lL_B (X_0)$ and $\phi_1 : A \to \lL_B (X_1)$ such that $\phi (a) = \phi_0 (a) \oplus \phi_1 (a)$ for $a\in A$.
We will write $\phi =\phi_0 \hoplus \phi_1$.
\end{notation}
\begin{definition}
For (trivially graded) $\rC^*$-algebras $A$ and $B$, a {\it Kasparov $A$-$B$ bimodule} is a triple $(X, \phi, F )$ such that $X$ is a countably generated graded Hilbert $B$-module, $\phi :A \to \lL_B ( X)$ is a $*$-homomorphism of degree 0,
and $F \in \lL_B (X)$ is of degree 1 and satisfies the following condition:
\begin{itemize}
\item $[F, \phi (a)] \in \lK_B (X)$ for $a\in A$,
\item $(F-F^*) \phi (a) \in \lK_B (X)$ for $a\in A$,
\item $(1 - F^2) \phi (a) \in \lK_B (X)$ for $a \in A$.
\end{itemize}
If $[F, \phi (a)]=(F-F^*) \phi (a)=(1 - F^2) \phi (a)=0 $ for all $a\in A$,
we call $(X, \phi, F )$ {\it degenerate}.
We denote by $\lE (A,B)$ and $\lD (A, B)$ the set of Kasparov $A$-$B$ bimodules and degenerate ones, respectively.
\end{definition}
We say that two $A$-$B$ Kasparov bimodules $(X, \phi, F)$ and $(Y, \psi, G)$ are {\it unitarily equivalent}, denoted by $(X, \phi, F) \cong (Y, \psi, G)$, if there exists a unitary $U \in \lL (X , Y)$ of degree $0$ such that $\psi = U \phi (\cdot ) U^*$ and $F= UGU^*$.
For a C$^*$-algebra $B$ we set $IB:=B \otimes C[0,1]$.
We identify $IB$ with $C([0,1], B)$, the space of $B$-valued continuous functions on $[0,1]$.
For $ t \in [0,1]$ the {\it evaluation at} $t$ is the surjective $*$-homomorphism, still written $t$, from $IB$ onto $B$ defined by $t (f):= f (t)$ for $f \in IB$.
If $X$ is a Hilbert $IB$-module, then the pushout $X_t$ of $X$ by $t$ is a Hilbert $B$-module.
For $(X, \pi_X) \in \Corr (A, IB)$ we denote by $(\pi_X)_t$ the $*$-homomorphism $A \ni a \mapsto \pi_X(a)_t \in \lL_B (X_t)$.
For a Hilbert $A$-module $X$, we set $IX:=X\otimes C[0,1]$.
We also identify $IX$ with the Hilbert $IA$-module $C([0,1], X)$ of $X$-valued continuous functions on $[0,1]$ equipped with the inner product $\i< f, g >_{IA} (t):= \i< f (t), g(t) >_A$ for $f, g \in IX$.
The following proposition is probably well-known, but we give its proof for the reader's convenience.
\begin{proposition}\label{lem-path}
Let $X$ be a Hilbert $A$-module.
Then, for any strict continuous, norm bounded path $ \{ x_t \}_{0 \leq t \leq 1} \in \lL_A (X)$,
there exists a unique operator $x \in \lL_{IA} (IX)$ of which the evaluation at $t$ is $x_t$.
Conversely, for any element $y \in \lL_{IA} (IX)$, the evaluations $\{y_t \}_{0 \leq t \leq 1}$ of $y$ defines a strict continuous, norm bounded path in $\lL_B (X)$.
Moreover, an operator $z$ is in $\lK_{IA} (IX)$ if and only if the corresponding path is a norm continuous path in $\lK_A (X)$.
\end{proposition}
\begin{proof}
Let $ \{ x_t \}_{0 \leq t \leq 1} \in \lL_A (X)$ be strict continuous and norm bounded.
Fix $f \in IX$ arbitrarily.
Indeed, for any $t, s\in [0,1]$ we have
$
\| x_t f (t) - x_s f (s) \|
\leq \| (x_t - x_s ) f (t) \| + \| x_s \| \| f (t) - f (s) \| .
$
Since $ x_t $ is norm bounded and strict continuous, $t \mapsto x_t f (t) $ and $t \mapsto x^*_t f (t) $ define elements in $IX$.
Hence the mapping $x : f \mapsto [ t\mapsto x_t f (t) ]$ is the desired element in $\lL_{IA} (IX)$.
Conversely, let $y \in \lL_{IA} (IX)$ be given and $\{ y_t \}_{0\leq t\leq 1}$ be the corresponding evaluations. For $\xi \in X$, let $\xi \otimes 1 \in X\otimes C[0,1]= IX$ be the constant function. We then have $y_t \xi = [ y (\xi \otimes 1) ] (t)$ for all $t \in [0,1]$,
which implies $\{ y_t \}_{0 \leq t \leq 1}$ is strictly continuous.
The third assertion follows from the isomorphism $\lK (IX) \cong \lK (X) \otimes C[0,1] \cong C( [0,1], \lK (X) )$.
\end{proof}
Recall that for $\cX=(X, \phi, F) \in \lE (A, IB)$ the evaluation at $t \in [0, 1]$ is the Kasparov $A$-$B$ bimodule $\cX_t:=(X_t, \phi_t, F_t)$.
\begin{definition}
Two Kasparov $A$-$B$ bimodules $\cX$ and $\cY$ are said to be {\it homotopic} is there exists a Kasparov $A$-$IB$ bimodule $\cZ$ such that $\cX \cong \cZ_0$ and $\cY \cong \cZ_1$.
The $KK(A,B)$ is the set of homotopy equivalence classes of all $A$-$B$ Kasparov $A$-$B$ bimodules.
\end{definition}
Let $\cX=(X , \phi, F), \cY =(Y, \psi, G) \in \lE (A, B)$.
We denote by $[\cX]=[ X, \phi, F]$ and $[\cY]=[ Y, \psi, G]$ the elements in $KK (A, B)$ corresponding to $\cX$ and $\cY$, respectively.
The addition of $[\cX]$ and $[\cY]$ is defined by $[\cX]+[\cY]:=[\cX \oplus \cY]$, where $\cX\oplus \cY = (X \oplus Y, \phi \oplus \psi , F \oplus G)$.
All degenerate Kasparov bimodules are homotopic to the trivial bimodule $0=(0,0,0)$ and define the zero element in $KK(A, B)$.
Write $X=X_0 \hoplus X_1$ and set $-X:= X_1 \hoplus X_0$.
Let $U : X \to -X$ be the canonical isomorphism and set $\phi_-:=U \phi (\cdot) U^*$.
The inverse of $[\cX]$ is represented by $-\cX:=(-X, \phi_-, -F)$.
If $\phi : A \to B$ is a $*$-homomorphism, then $(B \hoplus 0, \phi \hoplus 0, 0 )$ defines a Kasparov $A$-$B$ bimodule.
We still denote by $\phi$ the element in $KK (A, B)$ corresponding to $(B \hoplus 0, \phi \hoplus 0, 0 )$.
Let $A$, $B$ and $C$ be $\rC^*$-algebras.
For ${\bf x} \in KK(A,B)$ and ${\bf y} \in KK (B, C)$, we denote by ${\bf x} \otimes_B {\bf y}$ the Kasparov product of ${\bf x}$ and ${\bf y}$.
Let $f : A \to B$ be a $*$-homomorphism.
Then, $f^* : KK (B, C) \to KK (A, C)$ is the group homomorphism given by $\lE(B,C) \ni (X, \phi, F) \mapsto f^*(X, \phi, F) :=(X, \phi \circ f, F) \in \lE (A, C)$.
Similarly, the $f_* : KK (C,A) \to KK (C, B)$ is defined by $\lE (C, A) \ni (Y, \psi, G) \mapsto f_* (Y, \psi, G):= (Y\otimes_f B, \psi \otimes 1, G \otimes 1) \in \lE (C, B)$.
Let ${\bf x} \in KK(B, C)$ be an element represented by $(X, \phi, F) \in \lE (B, C)$.
If $f : A \to B$ and $g : C \to D$ are $*$-homomorphisms, then we have $ f \otimes_B {\bf x} = f^* ({\bf x}) =[ X, \phi \circ f , F]$ and ${\bf x} \otimes_C g = g_* ({\bf x}) = [ X\otimes_g D,\phi \otimes 1, F \otimes 1]$.
\begin{definition}
An element ${\bf x} \in KK(A, B)$ is said to be a $KK$-{\it subequivalence} if there exists ${\bf y} \in KK( B, A)$ such that $\id_A ={\bf x} \otimes_B {\bf y}$.
When $\id_A ={\bf x}\otimes_B {\bf y}$ and $\id_B = {\bf y }\otimes_B {\bf x}$ hold, we say that ${\bf x}$ is a $KK$-{\it equivalence}.
In this case, the $A$ and $B$ are said to be $KK$-{\it equivalent}.
\end{definition}
If ${\bf x} \in KK(A, B)$ is a $KK$-equivalence, then for any $C$ the mappings ${\bf x} \otimes_B (\cdot ) :KK(B, C) \to KK(A, C)$ and $( \cdot ) \otimes_A {\bf x} : KK (C, A) \to KK (C, B)$ are isomorphisms.
In particular, we have $K_* (A) \cong K_* (B)$ if $A$ and $B$ are $\sigma$-unital.
\subsection{Strong relative nuclearity implies relative $K$-nuclearity}\label{ss-rel-K-K}
The main result of this subsection is Theorem \ref{thm-K-nuclear}.
This technical theorem plays an important in \S\S \ref{ss-KK-CD}.
For a Hilbert $A$-module $X$ we denote by $J_X$ the degree 1 unitary $\left[ \begin{smallmatrix} 0&1\\ 1& 0 \end{smallmatrix} \right] \in \lL_A (X\hoplus X)$.
Since $I(X^\infty)=(X^\infty) \otimes C[0,1]=\ell^2(\lN) \otimes X \otimes C[0,1]=(IX)^\infty$, we will write $IX^\infty$.
For $\rC^*$-correspondence $(X, \pi_X)$ we set $(IX, \pi_{IX}):= (IX, \pi_X \otimes 1_{C[0,1]})$.
\begin{definition}[{\cite[Definition 3.4]{Germain-fields}}]
Let $1_A \in B \subset A$ be a unital inclusion of $\rC^*$-algebras with a conditional expectation $E : A \to B$.
We say that $A$ is {\it $K$-nuclear relative to} $(B, E)$ if the $\rC^*$-correspondence $(X, \pi_X):= (L^2 (A, E) \otimes_B A, \pi_E \otimes 1_A)$ satisfies the following:
\begin{itemize}
\item[(i)] there exist unital $*$-homomorphisms $\pi^+$ and $\pi^-$ from $A$ into $\lL_{IA} (IX^\infty)$ such that $\cX=(IX^\infty \hoplus IX^\infty, \pi^+ \hoplus \pi^-, J_{IX^\infty} ) \in \lE (A, IA)$, i.e., $\pi^+ (a)- \pi^- (a) \in \lK_{IA} (IX^\infty)$ for all $a \in A$;
\item[(ii)] $\pi^+ (b) = \pi^- (b) =\pi_{IX}^\infty (b)$ for all $b \in B$;
\item[(iii)] the evaluation of $\cX$ at $t=1$ is degenerate, i.e., we have $\pi^+_1 = \pi^-_1$;
\item[(iv)] there exists a unitary $U \in \lL_A (A \oplus X^\infty, X^\infty)$ such that
$\pi^+_0 = U(\lambda_A \oplus \pi_X^\infty )U^*$ , $\pi^-_0= \pi_X^\infty$, and
and $U(1_A \oplus 0)= \xi \otimes 1_A$ for some $\xi \in L^2(A,E)^\infty$.
\end{itemize}
\end{definition}
\begin{proposition}
If $A$ is unital separable $\rC^*$-algebra, $B\subset A$ is a $\rC^*$-subalgebra,
and $X \in \Corr (A)$ has the $B$-CCPAP, then we can chose a sequence $\psi_n \in \cF_{B'\cap X}$ such that $\psi_n(1_A) =1_A$ for $n\in \lN$ and $\lim_{n \to \infty} \| a - \psi_n (a ) \| =0 $ for $a\in A$.
\end{proposition}
\begin{proof}
Take a countable dense set $\{ a_n \}_{n=1}^\infty$ in $A$ with $a_1 =1_A$.
Take $\psi_{n} \in \cF_{B' \cap X}$ in such a way that $\| a_k - \psi_n (a_k) \| < 2^{-n}$ for $ 1\leq k \leq n $ and $\psi_n (1_A) \leq 1_A$.
Then $\{ \psi_n \}_n $ enjoys the second assertion.
We next replace $\psi_n$ by a u.c.p.\ map.
Since $\psi_n (1_A) b = \psi_n (b) =b \psi_n (1_A)$ for $b \in B$, it follows that $\psi_n (1_A) \in B' \cap A$.
Define the u.c.p.\ map $\varphi_n \in \cF_{X}$ by $\varphi_n (a) := \psi_n (1_A)^{-1/2} \psi_n (a) \psi_n (1_A)^{-1/2}$.
Since $\psi_n (1_A)^{-1/2}$ is in $B' \cap A$ and converges to $1_A$ in norm,
$\{ \varphi_n \}_{n=1}^\infty$ is the desired one.
\end{proof}
For $X, Y \in \Corr (A, C)$ and $B \subset A$ we define the set ${}_B \lL_C (X, Y):= \{ x \in \lL_C(X, Y) \mid x\pi_X(b) = \pi_Y (b) x \text{ for } b \in B \}$ and ${}_B\lK_C (X, Y):=\lK_C (X, Y) \cap {}_B\lL_C (X, Y)$.
\begin{theorem}\label{thm-strongvoic}
Let $A$ be a unital separable $\rC^*$-algebra, $B$ be a $\rC^*$-subalgebra of $A$, and $(X,\pi_X) \in \Corr (A)$ be unital and have the $B$-CCPAP.
Then, there exists a sequence of isometries $V_n \in {}_B\lL_A (H_A, X^\infty)$ such that
\begin{itemize}
\item[$(1)$] $\lambda_A^\infty (a) - V_n^* \pi_X^\infty (a) V_n \in \lK_A (H_A)$ for $a\in A$,
\item[$(2)$] $\lim_{n \to \infty}\| \lambda_A^\infty (a) - V_n^* \pi_X^\infty (a) V_n \| =0$ for $a\in A$.
\end{itemize}
\end{theorem}
\begin{proof}
Since $A$ is separable, it suffices to show that for any finite subset $\fF \subset A$ and $\varepsilon>0$ there exists an isometry $V \in {}_B\lL_A (H_A, X^\infty)$ satisfying (1) and $\| \lambda_A^\infty (a) - V^* \pi_X^\infty (a) V \| <\varepsilon$ for $a\in \fF$.
Take an increasing sequence of finite subsets $\fF=\fF_1 \subset \fF_2 \subset \cdots$ of $A$ such that $A = \overline{\bigcup_{n=1}^\infty \fF_n}^{\| \cdot \|}$.
Write $H_A=\bigoplus_{n=1}^\infty A^{(n)}$ with $A^{(n)}:=A$.
Since $X$ has the $B$-CCPAP, for each $n \in \lN$ there exists $\psi_n \in \cF_{B'\cap X}$ such that $\psi_n(1_A) = 1_A$ and $\| a - \psi_n (a) \| < 2^{-n}\varepsilon $ for $a \in \fF_n$.
Write $\psi_n= \sum_{r=1}^{d(n)} \Omega_{\xi_n^{(r)}}$ with $\xi_n^{(r)} \in B'\cap X$.
Define $W_n \in \lL_A (A^{(n)}, X^{d(n)} )$ by $W_n 1_A =(\xi_n^{(1)}, \dots, \xi_n^{(d(n))})$.
We then have
$
\psi_n (a) = \sum_{r=1}^{d(n)} \i< \xi_n^{(r)}, \pi_X (a) \xi_n^{(r)} >_X = \i< W_n 1_A, \pi_X^{d(n)}(a) W_n1_A>_{A^{(n)}} = W_n^* \pi_X^{d(n)}(a) W_n
$
for $a\in A$.
Hence, $W_n$ is an isometry in ${}_B\lL_A (A^{(n)}, X^{d(n)})$.
Define $V \in \lL_A (H_A, X^\infty)$ by $V:=\bigoplus_{n=1}^\infty W_n : H_A \to \bigoplus_{n=1}^\infty X^{d(n)} \cong X^\infty$.
Then, it follows that $V$ is an isometry in ${}_B \lL_A (H_A, X^\infty)$.
Moreover, for any $a\in \bigcup_{n=1}^\infty \fF_n$, we have $\sum_{n=1}^\infty \| a - \psi_n(a) \| < \infty$, and hence $V$ satisfies (1).
By the construction of $V$ we also have $\| \lambda_A^\infty (a) - V^* \pi_X^\infty (a) V \| <\varepsilon$ for $a\in \fF$.
\end{proof}
\begin{corollary}\label{cor-cvoic}
Let $A$ be a unital separable $\rC^*$-algebra, $B$ be a $\rC^*$-subalgebra of $A$, and $(X,\pi_X) \in \Corr (A)$ be unital and have the $B$-CCPAP.
Then, there exists a sequence of unitaries $U_n \in {}_B\lL_A (H_A \oplus X^\infty, X^\infty)$ such that
\begin{itemize}
\item[$(1)$] $\pi_X^\infty (a) - U_n( \lambda_A^\infty(a) \oplus \pi_X^\infty (a) ) U^*_n \in \lK_A (X^\infty)$ for $a\in A$ and $n\in \lN$,
\item[$(2)$] $\lim_{n \to \infty}\| \pi_X^\infty (a) - U_n( \lambda_A^\infty(a) \oplus \pi_X^\infty (a)) U^*_n \| =0$ for $a\in A$.
\end{itemize}
\end{corollary}
\begin{proof}
Let $\{ V_n \}_n \subset \lL_A (H_A, X^\infty)$ be a sequence of isometries in the preceding theorem and put $P_n:=1 - V_nV_n^*$.
Note that $P_n \in \pi_X^\infty(B)'$.
Let $S_1,S_2 \in \lB (\ell^2(\lN))$ be the isometries defined by $S_1 \delta_n= \delta_{2n-1}$ and $S_2 \delta_n= \delta_{2x}$ for $n \geq 1$.
Then, the operator $T:=(S_1 \oplus S_2 ) \otimes 1_A : H_A \oplus H_A \to H_A$ is a unitary.
Define unitaries $W_n:=V_n \oplus P_n : H_A \oplus P_n X^\infty \to X^\infty$ and $U_n:=W_n (T\oplus 1_{X^\infty})(1_{H_A} \oplus W_n^*) : H_A\oplus X^\infty \to X^\infty$.
As in the proof of Theorem \ref{thm-absorbing} the sequence $\{ U_n\}_n$ satisfies (1) and (2).
Since each $U_n$ is the product of three unitaties intertwine the left actions of $B$, we have $U_n \in {}_B\lL_A (H_A \oplus X^\infty, X^\infty)$.
\end{proof}
For a Hilbert C$^*$-module $X$ and $\xi \in X$, we set $\xi^{(n)}:=\xi \otimes \delta_n \in X \otimes \ell^2 (\lN) =X^\infty$.
\begin{corollary}
Let $B \subset A$ and $X$ be as in the preceding corollary.
For any $\xi_0 \in B' \cap X$ with $\i< \xi_0, \xi_0> =1_A$ there exists a unitary $U \in {}_B\lL_A (H_A \oplus X^\infty, X^\infty)$ such that
\begin{itemize}
\item[$(1)$] $\pi_X^\infty (a) - U( \lambda_A^\infty(a) \oplus \pi_X^\infty (a) ) U^* \in \lK_A (X^\infty)$ for $a\in A$,
\item[$(2')$] $U(1_A^{(1)}\oplus 0) =\xi_0^{(1)}$.
\end{itemize}
\end{corollary}
\begin{proof}
We use the notation in the proof of Theorem \ref{thm-strongvoic}.
Replace $W_1 : A_1 \to X^{d(1)}$ in the proof of Theorem \ref{thm-strongvoic} by $\widetilde{W}_1 : A_1 \ni 1_A \mapsto \xi_0 \in X$.
Then $\widetilde{V}=\widetilde{W}_1 \oplus \bigoplus_{n=2}^\infty W_n \in {}_B\lL_A ( H_A, X^\infty)$ is also an isometry enjoys (1) in Theorem \ref{thm-strongvoic} and that $\widetilde{V} 1_A^{(1)} =\xi_0^{(1)}$.
Put $\widetilde{P}:=1 - \widetilde{V}\widetilde{V}^*$ and $\widetilde{W}:=\widetilde{V}\oplus \widetilde{P}$.
We note that the $T$ above satisfies that $T( 1_A^{(1)} \oplus 0)=1_A^{(1)}$.
Then $\widetilde{U}=\widetilde{W}(T \oplus 1_{X^\infty} )(1_{H_A}\oplus \widetilde{W}^*) \in {}_B\lL_A ( H_A \oplus X^\infty, X^\infty)$ is the desired unitary.
We only check $(2')$:
\begin{align*}
\widetilde{U}(1_A^{(1)}\oplus 0)
&=\widetilde{W}(T \oplus 1_{X^\infty} )(1_{H_A}\oplus \widetilde{W}^*)(1_A^{(1)}\oplus 0)
=\widetilde{W}(T \oplus 1_{X^\infty} )((1_A^{(1)}\oplus 0_{H_A}) \oplus 0_{X^\infty} ) \\
&=\widetilde{W}(1_A^{(1)}\oplus 0_{X^\infty})
=\widetilde{V}1_A^{(1)}
=\xi_0^{(1)}.
\end{align*}
\end{proof}
For a $\rC^*$-algebra $A$ we set $CA:=A\otimes C_0[0,1)$, where $C_0[0,1)=\{ f\in C[0,1] \mid f(1)=0 \}$. Note that $C_0[0,1)$ naturally forms a Hilbert $C[0,1]$-module.
\begin{theorem}\label{thm-K-nuclear}
Let $B \subset A$ be a unital inclusion of separable $\rC^*$-algebra,
$(X, \pi_X) \in \Corr(A)$ be countably generated, unital, and have the $B$-CCPAP,
$\xi \in B' \cap X$ be a fixed vector with $\i< \xi, \xi> =1_A$.
Set $(IX, \pi_{IX}):=(X\otimes C[0,1], \pi_X \otimes 1_{C[0,1]}) \in \Corr (A, IA)$.
Then, there exists a unitary $U \in \lL_{IA} (CA \oplus IX^\infty , IX^\infty )$ such that
\begin{itemize}
\item[$(1)$] the triple $\cX= (IX^\infty \hoplus IX^\infty, U ( \lambda_A \otimes 1_{C_0[0,1)} \oplus \pi_{IX}^\infty )U^* \hoplus \pi_{IX}^\infty, J_{IX^\infty })$ forms a Kasparov $A$-$IA$ bimodule.
\item[$(2)$] the unitary $U$ is in ${}_B\lL_{IA} (CA \oplus I X^\infty , IX^\infty )$, i.e., $U_t ( \lambda_A(b) \oplus \pi^\infty_X(b))U_t^*=\pi_X^\infty(b)$ for all $b\in B$ and $t\in [0,1)$.
\item[$(3)$] the evaluation $U_1$ of $U$ at $1$ equals $1_{X^\infty}$, and hence the evaluation of $\cX$ at 1 is degenerate.
\item[$(4)$] the evaluations $\{ U_t \}_{0 \leq t \leq 1}$ satisfies that
$$
\begin{cases}
U_t (\cos (\pi t) 1_A \oplus \sin (\pi t) \xi^{(1)}) &= \xi^{(1)} \quad\text{for} \quad 0 \leq t \leq 1/2 \\
U_t (0 \oplus \xi^{(1)})&= \xi^{(1)} \quad\text{for}\quad 1/2 \leq t < 1
\end{cases}
$$
\end{itemize}
\end{theorem}
\begin{proof}
By the $B$-CCPAP of $X$ there exists a unitary $V \in {}_B\lL_A (H_A \oplus X^\infty, X^\infty)$ such that $V (\lambda_A^\infty(a) \oplus \pi_X^\infty (a))V^*- \pi_X^\infty \in \lK_A (X^\infty)$ for $a\in A$ and $V (1_A^{(1)} \oplus 0)=\xi^{(1)}$.
We set $H^\circ :=\ell^2(\lN) \ominus \lC \delta_1$ and $H^\circ_{C[0,1]}:=H^\circ \otimes C[0,1]$.
By Kasparov's stabilization theorem, there exists a unitary $T \in \lL_{C[0,1]} (C_0[0,1) \oplus H^\circ_{C[0,1]}, H^\circ_{C[0,1]})$.
Set $W:=(V\otimes 1_{C[0,1]}) ( T \otimes 1_A \oplus 1_{IA} \oplus 1_{IX^\infty} ) (1_{CA} \oplus V^* \otimes 1_{C[0,1]}):$
\begin{align*}
CA \oplus IX^\infty
& \longrightarrow CA \oplus H_{IA} \oplus IX^\infty
= (C[0,1 ) \oplus H^\circ_{C[0,1]} ) \otimes A \oplus IA \oplus IX^\infty \\
&\longrightarrow (H_{C[0,1]}^\circ) \otimes A \oplus IA \oplus IX^\infty =H_{IA} \oplus IX^\infty \longrightarrow IX^\infty.
\end{align*}
By construction, it follows that $W \in {}_B\lL_{IA} (CA \oplus IX^\infty ,I X^\infty )$ and
\begin{equation}
W( \lambda_A(a) \otimes 1_{C_0[0,1)} \oplus \pi_{IX}^\infty(a) )W^* - \pi_{IX}^\infty (a) \in \lK_{IA}( I X^\infty ) \text{ for } a\in A, \label{eq-path}
\end{equation}
and $W(0 \oplus \xi^{(1)}\otimes 1_{C[0,1]}) =\xi^{(1)} \otimes 1_{C[0,1]}$.
Let $W_1 \in {}_B\lL_A (X^\infty)$ be the evaluation of $W$ at 1 and put $U:=(W_1^* \otimes 1_{C[0,1]}) W$.
Then $U$ also enjoys $U(0 \oplus \xi^{(1)}\otimes 1_{C[0,1]})=(W_1^* \xi^{(1)}) \otimes 1_{C[0,1]} = \xi^{(1)} \otimes 1_{C[0,1]}$ and (\ref{eq-path}) in which replaced $W$ by $U$.
Setting $\eta:=U_{1/2}(1_A \oplus 0)$ we have $\eta \in B' \cap X^\infty$, $\i< \eta, \eta > =1_A$, and $\i< \xi^{(1)} , \eta >=0$.
Hence we have $X^\infty = \xi^{(1)} A \oplus \eta A \oplus Y$ for a submodule $Y \subset X^\infty$.
Note that $\pi_X^\infty (B)$ commute with projections $\theta_{\xi^{(1)},\xi^{(1)}}$ and $\theta_{\eta, \eta}$.
Define $U_t' \in \lL_A (\eta A \oplus \xi^{(1)} A)$ for $0 \leq t \leq 1/2$ by
$U_t' \eta = \sin (\pi t) \eta \oplus \cos (\pi t) \xi^{(1)}$ and $U_t' (\xi^{(1)})= ( -\cos (\pi t) \eta ) \oplus \sin (\pi t ) \xi^{(1)}$.
($U'_t$ is the unitary given by the matrix $\left[ \begin{smallmatrix} \sin(\pi t) & - \cos (\pi t) \\ \cos (\pi t) & \sin (\pi t) \end{smallmatrix} \right] \in \lM_2 (A) \cong \lL_A (\eta A \oplus \xi^{(1)} A )$.)
Now $U'_t \oplus 1_Y$ is a norm continuous path of unitaties in $\lC 1 + {}_B\lK_A (X^\infty)$ and $U'_{1/2} \oplus 1_Y =1$.
Let $\widetilde{U} \in \lL_{IA}(CA \oplus IX^\infty, IX^\infty )$ be the unitary defined by the path $\{ (U'_t \oplus 1_Y) U_{1/2} \}_{ 0\leq t \leq 1/2} \cup \{ U_t \}_{1/2 \leq t \leq 1}$.
By Lemma \ref{lem-path}, $\widetilde{U}$ is well-defined.
Moreover, by construction we get $\widetilde{U} \in {}_B\lL_{IA}(CA \oplus IX^\infty, IX^\infty )$.
To see (1), it suffices to show that the path $\{ \widetilde{U}_t ( \lambda_A(a) \oplus \pi^\infty_X (a) )\widetilde{U}_t^* - \pi_X^\infty(a) \}_{ 0\leq t \leq 1}$ is a norm continuous path in $\lK_A (X^\infty)$.
This follows from Lemma \ref{lem-path} and definition of $\widetilde{U}$ again.
Thus, $\widetilde{U}$ is the desired unitary.
\end{proof}
As a corollary, we get a relative analogue of `nuclearity $\Rightarrow$ $K$-nuclearity' for {\it strongly} nuclear inclusions.
\begin{corollary}\label{cor-K-nuclear}
Let $B \subset A$ be unital inclusions of separable $\rC^*$-algebras with a conditional expectation $E :A \to B$.
If $(A, B, E)$ is strongly nuclear, then $A$ is $K$-nuclear relative to $(B,E)$.
\end{corollary}
\begin{proof}
Applying the previous theorem to $(X, \pi_X):=(L^2(A,E) \otimes_B A, \pi_E \otimes 1_A)$ and $\xi_E \otimes 1_A$ we get a unitary $U : CA \oplus IX^\infty \to IX^\infty$ satisfying the conditions in the theorem.
The $*$-homomorphisms $\pi^+:=U ( \lambda_A \otimes 1_{C_0[0,1)} \oplus \pi_{IX}^\infty )U^*$ and $\pi^-:= \pi_{IX}^\infty$ are the desired ones.
\end{proof}
\section{$KK$-equivalences of amalgamated free products}\label{sec-KK}
In this section, we prove Theorem \ref{thm-C} and Theorem \ref{thm-D}.
\subsection{Amalgamated free products of $\rC^*$-algebras}\label{ss-KK-ama}
Let $\cI$ be a set and $B$ be a unital $\rC^*$-algebra.
Let $\{(X_i, \pi_{X_i}) \}_{i \in \cI} \subset \Corr (B)$ be a family of unital $\rC^*$-correspondences over $B$ with $B$-central normal vectors $\xi_i$, i.e., one has $\pi_{X_i} (b) \xi_i = \xi_i b$ for $b \in V$ and $\i< \xi_i, \xi_i > =1_B$.
We define the index set $\cI_p$ for $p \in \lN$ by
$
\cI_p= \{ i: \{1, \dots, p \} \to \cI \mid i(k) \neq i(k+1) \text{ for } 1\leq k \leq p-1 \}.
$
The free product of $\{(X_i, \xi_i ) \}_{i \in \cI}$ is the Hilbert $B$-module $(X, \xi_0)$ given by
$$
X=B \oplus \bigoplus_{p \geq 1} \bigoplus_{i \in \cI_p} X_{i(1)}^\circ \otimes_B \dots \otimes_B X_{i(p)}^\circ,
$$
where $X_i^\circ = X_i \ominus \xi_i B$.
We will denote by $\xi_0$ the unit of $B$ in the first direct summand of $X$
and write $(X, \xi_0 )= \bigstar_{i \in \cI} (X_i, \xi_i)$.
We also define complemented submodules $X(\lambda, j)$ and $X(\rho, j)$ of $X$ for $j \in \cI$ by
\begin{align*}
X(\lambda, j) = \xi_0 B \oplus \bigoplus_{p \geq 1} \underset{i(1) \neq j}{\bigoplus_{i \in \cI_p}} X_{i(1)}^\circ \otimes_B \dots \otimes_B X_{i(p)}^\circ,\\
X(\rho, j) = \xi_0 B \oplus \bigoplus_{p \geq 1} \underset{i(p) \neq j}{\bigoplus_{i \in \cI_p}} X_{i(1)}^\circ \otimes_B \dots \otimes_B X_{i(p)}^\circ,
\end{align*}
and unitaries $v_i \in \lL_B (X, X_i \otimes_B X(\lambda, i))$ and $w_i \in \lL_B (X, X(\rho, i) \otimes_B X_i)$ by
\begin{align*}
v_i \xi_0 &= \xi_i \otimes \xi_0, \\
v_i (\eta_1 \otimes \cdots \otimes \eta_p)
&= \begin{cases}
\eta_1 \otimes \xi_0 & \text{ for } \eta_1 \in X_i^\circ, p=1\\
\eta_1 \otimes (\eta_2 \otimes \cdots \otimes \eta_p)& \text{ for } \eta_1 \in X_i^\circ, p \geq 2\\
\xi_i \otimes (\eta_1 \otimes \cdots \otimes \eta_p) & \text{ for } \eta_1 \notin X_i^\circ
\end{cases}\\
w_i \xi_0 &= \xi_0 \otimes \xi_i \\
w_i (\eta_1 \otimes \cdots \otimes \eta_p)
&= \begin{cases}
\xi_0 \otimes \eta_1& \text{ for } \eta_p \in X_i^\circ, p=1\\
( \eta_1 \otimes \cdots \otimes \eta_{p-1}) \otimes \eta_p & \text{ for } \eta_p \in X_i^\circ, p \geq 2\\
(\eta_1 \otimes \cdots \otimes \eta_p) \otimes \xi_i & \text{ for } \eta_1 \notin X_i^\circ .
\end{cases}
\end{align*}
We can now define $*$-homomorphisms $\lambda_i : \lL_B (X_i ) \to \lL_B(X)$ and $\rho_i : {}_B\lL_B (X_i) \to \lL_B (X)$ by
\begin{equation}
\lambda_i(x)=v_i^*( x\otimes 1_{X(\lambda,i)} )v_i,
\quad \rho_i(y) =w_i^* (1_{X(\rho,i)} \otimes y) w_i. \label{eq-left-right}
\end{equation}
Hence, $(X, \lambda_i \circ \pi_{X_i})$ forms a C$^*$-correspondence over $B$.
Further assume that we have a Hilbert $B$-module $Y$ and $Z \in \Corr (B, C)$ for a C$^*$-algebra $C$.
By using unitaries $1_Y \otimes v_i \in \lL_B (Y \otimes_B X, Y \otimes_B X_i \otimes_B X( \lambda, i ) )$ and $w_i \otimes 1_Z \in \lL_C (X \otimes_B Z, X (\rho, i) \otimes_B X_i \otimes_B Z) $ we can define $*$-homomorphisms $\lambda_i^Y : \lL_B (Y\otimes_B X_i) \to \lL_B (Y \otimes_B X); x \mapsto (1_Y \otimes v_i)^* (x \otimes 1_{X(\lambda, i)} ) (1_Y \otimes v_i)$ and $\rho_i^Z : {}_B \lL_C (X_i \otimes_B Z) \to \lL_C (X \otimes_B Z); y \mapsto (w_i \otimes 1_Z)^* ( 1_{X(\rho, i)} \otimes y ) (w_i \otimes 1_Z)$.
A direct computation will prove the following proposition.
\begin{proposition}\label{prop-left-right}
Under the notation above, it follows that
$$
\left[ \lambda_i (x) \otimes 1_Z, \rho_j^{Z} (y) \right] = \delta_{ij}\rho_i^Z ( [ x \otimes 1_Z, y] )(P_{X_i} \otimes 1_Z)
$$
for $x \in \lL_B (X_i)$, $y \in \lL_C (X_j \otimes_B Z)$, and $i,j \in \cI$, where $P_{X_i } \in \lL_B (X)$ is the orthogonal projection onto $X_i = \xi_0 B \oplus X_i^\circ \subset X$.
\end{proposition}
\begin{definition}[\cite{Voiculescu-free}]
Let $\{ (A_i, E_i ) \}_{i \in \cI}$ be a family of unital $\rC^*$-algebras with nondegenerate conditional expectations $E_i$ from $A_i$ onto a common $\rC^*$-subalgebra $B$ containing $1_{A_i}$.
Let $(L^2(A_i, E_i ), \pi_{i}, \xi_i ) \in \Corr (A_i, B)$ be the GNS representation for $(A_i, E_i)$.
Let $(X, \xi_0)$ be the free product Hilbert module $\bigstar_{i \in \cI} (L^2 (A_i, E_i ), \xi_i )$.
The reduced amalgamated free product of $\{ (A_i, E_i ) \}_{i \in \cI}$ over $B$ is given by the pair $( \bigstar_{B, i \in \cI} (A_i , E_i ), E)$, where $\bigstar_{B, i \in \cI} (A_i , E_i )$ is the $\rC^*$-subalgebra of $\lL_B (X)$ generated by $\lambda_i \circ \pi_{E_i} (A_i), i\in \cI$ and $E $ is the conditional expectation from $A$ onto $B$ given by $\Omega_{\xi_0} \in \cF_{X}$.
\end{definition}
We note that $L^2(A, E) =\xi_E B \oplus L^2(A,E)^\circ$, where $L^2 (A,E)^\circ$ the orthogonal complement of $\xi_E B$, and $\pi_E (B)$ reduces these subspaces.
\begin{definition}
Let $\{ A_i \}_{i \in \cI}$ be a family of $\rC^*$-algebras containing a common $\rC^*$-subalgebra $B$.
Then, the {\it full amalgamated free product} of $A_i , i \in \cI$ over $B$ is the $\rC^*$-algebra $\bigstar_{B, i \in \cI} A_i$, equipped with injective $*$-homomorphisms $\iota_i : A_i \to \bigstar_{B, i \in \cI} A_i$ such that $\iota_i (b) =\iota_j (b)$ for $b \in B$ and $i, j \in \cI$ and satisfying the following universal property{\rm :} for any $\rC^*$-algebra $C$ and $*$-homomorphisms $\pi_i : A_i \to C$ such that $\pi_i (b) = \pi_j(b)$ for $b \in B$ and $i,j \in \cI$, there exists a unique $*$-homomorphism $\bigstar_{ i \in \cI} \pi_i : \bigstar_{B, i \in \cI} A_i \to C$ such that $(\bigstar_{ i \in \cI} \pi_i )\circ \iota_i = \pi_i$ for $i \in \cI$.
\end{definition}
\subsection{Theorem C and Theorem D}\label{ss-KK-CD}
In this subsection we prove Theorem \ref{thm-KK} below, which contains Theorem \ref{thm-C} and Theorem \ref{thm-D} (see Proposition \ref{prop-fin-dim}).
(We refer to \S\S \ref{ss-rel-rel} for the definitions of strong relative nuclearity via C$^*$-correspondences and central vectors.)
\begin{theorem}\label{thm-KK}
Let $\{(A_i, B, E_i) \}_{i \in \cI}$ be an at most countable family of unital inclusions of separable $\rC^*$-algebras $B \subset A_i$ with conditional expectations $E_i : A_i \to B$.
If each triple $(A_i, B, E_i)$ is strongly nuclear via $\rC^*$-correspondence $(Z_i, \pi_{Z_i})$ over $B$ such that $Z_i$ is countably generated and admits a $B$-central vector $\zeta_i \in Z_i$ with $\i< \zeta_i, \zeta_i> =1_B$,
then the canonical surjection from
the full amalgamated free product $\bigstar_{B, i \in \cI} A_i$ onto the reduced one $\bigstar_{B, i \in \cI} (A_i, E_i)$ is a $KK$-equivalence.
\end{theorem}
We first establish several technical lemmas based on \cite{Germain-duke}\cite{Germain-fields}.
In what follows, $\cI$ denotes a countable set and $\{ (A_i, B, E_i) \}_{i \in \cI}$ is a family of unital inclusions of separable $\rC^*$-algebras with nondegenerate conditional expectations $E_i : A_i \to B$.
Let $A:= \bigstar_{i, \in \cI, B} A_i$ and $A_\re:= \bigstar_{i \in \cI, B} (A_i, E_i)$ be the full and the reduced amalgamated free products, respectively and let $\pi_\re : A \to A_\re$ be the canonical surjection.
The next lemma is probably well-known, but we give its proof for the reader's convenience.
\begin{lemma}\label{lem-free}
If $(Z,\pi_Z)$ is an $A$-$B$ $\rC^*$-correspondence such that there exists a subspace $\Gamma \subset Z$ such that $\pi_Z (A)\Gamma$ is norm dense in $Z$ and satisfying the freeness condition:
for any $p \in \lN$, $i \in \cI_p$, and $a_k \in \ker E_{i(k)} \subset A_{i(k)}, 1 \leq k \leq p$,
$$\i< \eta, \pi_Z ( a_1 \cdots a_p) \xi >=0 \quad \text{for } \xi, \eta \in \Gamma,$$
then the $\pi_Z$ factors through $A_\re$.
\end{lemma}
\begin{proof}
It suffices to show that $\ker \pi_\re \subset \ker \pi_Z$.
Fix $a \in \ker \pi_\re$ arbitrarily and take a sequence $\{ a_n \}_n \in *\text{-Alg}(A_i, i\in \cI ) \subset A$ such that $\lim_{n \to \infty} \| a - a_n \| =0$.
Since $\Gamma$ is a cyclic subspace for $\pi_Z (A)$ we only have to prove that $\i<\eta, \pi_Z (b^* a c) \xi >=0 $ for all $b,c \in *\text{-Alg}(\iota_i (A_i), i\in \cI)$ and $\xi,\eta \in \Gamma$.
It is known that $\pi_\re (b^*a_n c)$ is a sum of an element $b_n \in B$ and finitely many elements having the form of $x_{i(1)} \cdots x_{i(p)}$ for some $p \in \lN$, $i \in \cI_p$, and $x_{i(k)}\in A_{i(k)} \cap \ker E \subset A_\re$.
Since $\xi, \eta \in \Gamma$ we now have
$\| \i< \eta, \pi_Z (b^* a c ) \xi > \| \leftarrow \| \i< \eta, \pi_Z(b^*a_n c) \xi > \| = \|\i< \eta, \pi_Z (\iota (b_n)) \xi > \| \leq \| \xi \| \| \eta \| \| b_n \| = \|\xi \| \| \eta \| \| E (\pi_\re (b a_n c^*))\| \to 0$,
which implies that $\pi_Z (a) =0$.
\end{proof}
Let $Z_i \in \Corr (B)$ be unital faithful and $\zeta_i \in B' \cap Z_i$ be a normal vector, i.e., $\i< \zeta_i, \zeta_i>=1_B$.
Set $(X_i, \xi_i):= ( L^2(A_i, E_i) \otimes_B Z_i, \xi_{E_i} \otimes \eta_i )$ and $(X, \xi_0):=\bigstar_{i \in \cI} (X_i, \xi_i)$.
By the universality of $A$ we obtain the $*$-homomorphism $\pi_X:=\bigstar_{i \in \cI} \lambda_i \circ ( \pi_{E_i} \otimes 1_{Z_I}) : A \to \lL_B (X)$, where $\lambda_i : \lL_B (X_i) \to \lL_B (X)$ is as in (\ref{eq-left-right}).
We also set $(Y_i, \pi_{Y_i}):= (X_i \otimes_B A_i, \pi_{X_i} \otimes 1_{A_i} ) \in \Corr (A_i)$ and $(Y, \pi_Y):= (X \otimes_ B A, \pi_X \otimes 1_A) \in \Corr (A)$.
\begin{lemma}\label{lem-throughAr}
Under the notation above, there exists a $*$-homomorphism $\bar{\pi}_X : A_\re \to \lL_B (X)$ such that $\pi_X = \bar{\pi}_X \circ \pi_\re$.
\end{lemma}
\begin{proof}
Set $Z_i^\circ:= Z_i \ominus \zeta_i B$ and
define the closed submodule $\Lambda \subset X$ by
$$
\Lambda := \bigoplus_{ p \geq 2} \bigoplus_{j \in \cI_p} ( \xi_{E_{j(1)}} \otimes_B Z_{j(1)}^\circ) \otimes_B X_{j(2)}^\circ \otimes_B \cdots \otimes_B X_{j(p)}^\circ .
$$
We will show that the closed submodule $\Gamma := \xi_0 B \oplus \bigoplus_{i \in \cI}
( \xi_{E_i} \otimes_B Z_i^\circ)
\oplus \Lambda $ satisfies the condition in Lemma \ref{lem-free}.
Indeed, it is not hard to see that $\pi_X ( a_1 \cdots a_p ) \Gamma \perp \Gamma$ for all $p \geq 1 $, $i \in \cI_p$, and $a_k \in A^\circ_{i (k)}$.
Set $\fX_0:=\xi_0B$ and $\fX_p:=\bigoplus_{i \in \cI_p} X_{i(1)}^\circ \otimes_B \cdots \otimes_B X_{i(p)}^\circ$ for $p \geq 1$.
We also set $X_i^\bullet:=L^2(A_i, E_i)^\circ \otimes_B \zeta_i B =\overline{\pi_X (A_i^\circ) \xi_0}$ and $\fX_p^\bullet:=\bigoplus_{i \in \cI} X_{i(1)}^\bullet \otimes_B X_{i(2)}^\circ \otimes_B \cdots \otimes_B X_{i(p)}^\circ$ for $p \geq 1$.
Since $X=\bigoplus_{p \geq 0} \fX_p$, it suffices to show that $\ospan \pi_X(A)\Gamma$ contains $\fX_p$ for all $p \geq 0$.
The case that $p=0$ is trivial.
Suppose that $p \geq 1$ and $\fX_p \subset \ospan \pi_X(A) \Gamma$.
We observe that $\fX_{p+1}^\bullet \subset \overline{\pi_X(A) \fX_p}$ and $\fX_{p+1}\ominus \fX_{p+1}^\bullet = \bigoplus_{i \in \cI_{p+1}} ( L^2(A_{i(1)}, E_{i(1)})\otimes_B Z_{i(1)}^\circ) \otimes_B X_{i(2)}^\circ \otimes_B \cdots \otimes_B X_{i(p+1)}^\circ \subset \overline{ \pi_X(A)\Gamma}$,
which implies that $\fX_{p+1} \subset \ospan {\pi_(X) \Gamma}$.
By induction and Lemma \ref{lem-free}, we are done.
\end{proof}
\begin{lemma}[{cf.\ \cite[Proposition 4.2]{Germain-duke}}] \label{lem-throughAr2}
Under the notation above,
suppose that there exist $Z'_i \in \Corr (B)$ and unitary $W_i : X_i \otimes_B Z'_i \to X_i \otimes_B X_i$ satisfying that
\begin{itemize}
\item[(i)] $W_i ( \pi_{X_i} (b) \otimes 1_{Z'_i} ) W_i^* = \pi_{X_i} (b) \otimes 1_{X_i}$ for all $b \in B$;
\item[(ii)] $\xi_i \otimes \xi_i \in W_i (\xi_i B \otimes_B Z'_i)$.
\end{itemize}
Define $\phi_i : A_i \to \lL_B (X \otimes_B X)$ by $\phi_i (a) = \lambda_i^{X_i} (W_i (\pi_{X_i} (a) \otimes 1_{Z'_i})W_i^* ) \oplus \tau_i (a) \otimes 1_X$.
Then, the $*$-homomorphism $\phi:=\bigstar_{i \in \cI} \phi_i : A \to \lL_B(X \otimes_B X)$ factors through $A_\re$.
\end{lemma}
\begin{proof}
We set $V_i:=1_{X_i} \otimes v_i \in \lL_B (X_i \otimes_B X, X_i \otimes_B X_i \otimes_B X(\lambda, i ) )$.
Let $\fX_p, X_i^\bullet, \fX_p^\bullet$ and $\Lambda$ be as in the proof of Lemma \ref{lem-throughAr}.
Define the subspace $\Gamma_i \subset X_i\otimes_B X$ by
$$
\Gamma_i:=
V_i^* ( W_i \otimes 1_{X (\lambda, i)} ) \left( \xi_{E_i} \otimes_B Z_i \otimes_B Z'_i \otimes_B X (\lambda, i) \right).
$$
By (ii) the $\xi_0 \otimes \xi_0$ belongs to $\Gamma_i$.
Set $\Gamma_i^\circ := \Gamma_i \ominus \xi_0 \otimes_B \xi_0 B$.
We will show that the subspace
$
\Gamma:=
\xi_0 \otimes_B \xi_0 B \oplus \bigoplus_{i \in \cI} \Gamma_i^\circ
\oplus \Lambda \otimes_B X
$
satisfies the condition in Lemma \ref{lem-free}.
We note that
\begin{equation}
\Gamma \subset \left[ \fX_0 \oplus \fX_1 \oplus \bigoplus_{p=2}^\infty ( \fX_p \ominus \fX_p^\bullet ) \right] \otimes_B X.\label{eq-Gamma}
\end{equation}
Fix $i \in \cI$ and $a \in A_i^\circ$ arbitrarily and show that $\phi (a) \Gamma \perp \Gamma$.
By definition, we have
$$
\phi(a) = V_i^* (W_i \otimes 1_{X(\lambda, i)} ) ( \pi_{E_i} (a) \otimes 1_{Z_i} \otimes 1_{Z'_i} \otimes 1_{X(\lambda, i)} )(W_i \otimes 1_{X(\lambda, i)} )^* V_i \oplus \tau_i (a) \otimes 1_X,
$$
which implies that $\phi (a) \Gamma_i \perp \Gamma$.
For any $j \in \cI$ with $i \neq j$
the fact that $\phi (a) \Gamma_j^\circ \subset \fX_2^\bullet$ and the (\ref{eq-Gamma}) implies $\phi (a) \Gamma_j \perp \Gamma$.
We also have $\phi (\Lambda \otimes_B X) \perp \Gamma$ by (\ref{eq-Gamma}).
Next, let $q \geq 2$ and $a_{k} \in A_{i(k)}^\circ, 1 \leq k \leq q$ be arbitrary.
Because $\phi (a_1 \cdots a_q) \Gamma \subset \bigoplus_{p \geq 2} \fX_p^\bullet$ we get $\phi (a) \Gamma \perp \Gamma$ by (\ref{eq-Gamma}).
Finally, we prove that $ \Gamma_0:=\ospan \phi (A) \Gamma$ contains $\fX_p \otimes_B X$ for all $p \geq 0$.
The case of $p=0$ is trivial.
When $p=1$,
it follows that $X_i \otimes_B X \subset \Gamma_0$ since
\begin{align*}
\phi (A_i) \Gamma_i
=V_i \left [ W_i \left ( \pi_{X_i}(A_i) \xi_{E_i} \otimes Z_i \otimes_B Z'_i \right) \otimes_B X( \lambda, i) \right],
\end{align*}
which is dense in $V_i^* [ W_i (X_i \otimes_B Z'_i) \otimes_B X( \lambda, i) ] = X_i \otimes_B X.$
Since $\ospan ( \phi (A) \Lambda \otimes_B X )$ contains $\bigoplus_{p \geq 2} ( \fX_p \ominus \fX_p^\bullet) \otimes_B X$,
we only have to show that $\fX_p^\bullet \otimes_B X \subset \Gamma_0$ for $p \geq 2$.
This is done by induction.
Suppose that $\fX_p^\bullet \otimes_B X \subset \Gamma_0$ with $p \geq 2$.
We then have $\fX_p \otimes_B X \subset \Gamma_0$.
Since the restriction of $\phi (a)$ on $\fX_p \otimes_B X$ equals $\pi_X (a) \otimes 1_X$ for $a\in A$.
Thus, $\fX_{p+1}^\bullet \otimes_B X \subset ( \ospan \pi_X (A) \fX_p )\otimes_B X \subset \Gamma_0$.
Therefore, by induction we get $\Gamma_0= X\otimes_B X$.
\end{proof}
Suppose that for each $i \in \cI$ there exists a unitary $U_i : A_i \oplus Y_i \to Y_i$ satisfies that $U_i ( 1_{A_i} \oplus 0 ) = \xi_i \otimes 1_{A_i}$ and $U_i ( b \oplus \pi_{Y_i} (b) )U_i^* = \pi_{Y_i} (b)$ for all $b \in B$.
Note that $Y \cong Y_i \otimes_{A_i} A \oplus (X \ominus X_i) \otimes_B A$ for all $i \in \cI$.
We define $\psi_i : A_i \to \lL_{A_i} (Y_i)$ by $\psi_i (a) = U_i ( a \oplus \pi_{Y_i} (a) )U_i^*$ and set $\psi:=\bigstar_{i \in \cI} ( \psi_i \otimes 1_A \oplus \tau_i \otimes 1_A)$, where $\tau_i : A_i \to \lL_B (X \ominus X_i)$ is defined by $\tau (a):=\pi_X(a)|_{X\ominus X_i}$.
\begin{lemma}\label{lem-unitary}
Under the notation above, let $S_i \in \pi_{Y_i} (B)' \cap \lL_{A_i} (Y_i)$ be the isometry defined by $U_i |_{ 0 \oplus Y_i}$.
Then, the isometries $\rho_i^{A} (S_i \otimes 1_A) \in \lL_A (Y)$ satisfy that $P_A +\sum_{i \in \cI} \rho_i^A (S_i \otimes 1_A)\rho_i^A (S_i \otimes 1_A)^* =1_Y$, where $P_A \in \lL_A (Y)$ is the projection onto $\xi_0B \otimes_B A$.
Moreover, the unitary $U := P_A \oplus \bigoplus_{i \in \cI} \rho_i^A (S_i \otimes 1_A) : A \oplus Y\otimes \ell^2 (\cI) \to Y$ satisfies that $\psi (a) = U (a \oplus \pi_Y (a) \otimes 1_{\ell^2(\cI)} ) U^*$ for all $a \in A$.
\end{lemma}
\begin{proof}
We observe that the image of $\rho_i^A (S_i \otimes 1_A)$ is
$
\bigoplus_{p \geq 1} \bigoplus_{j \in \cI_p, j(p)=i} X_{j(1)}^\circ \otimes_B \cdots \otimes_B X_{j(p)}^\circ \otimes_B A.
$
Since these subspaces are mutually orthogonal and span $Y \ominus P_A Y$, we get the first and the second assertions.
To see the second assertion, it suffices to see that $\psi (a) = U (\lambda_A (a) \oplus \pi_Y (a) \otimes 1_{\ell^2 (\cI )}) U^*$ for all $i \in \cI$ and $a \in A_i$.
Thus, we fix $i \in \cI$ and $a\in A_i$.
Recall that $\psi (a)=[ ( U_i(\lambda_{A_i} (a) \oplus \pi_{Y_i} (a) ) U_i^* )\otimes 1_A ] \oplus \tau_i (a) \otimes 1_A = [ ( U_i \otimes 1_A)(\lambda_A(a) \oplus \pi_{Y_i} (a) \otimes 1_A )( U_i \otimes 1_A)^*] \oplus \tau_i (a) \otimes 1_A.$
Identifying $Y$ with $Y \otimes \lC \delta_i$ we have
\begin{align*}
\psi (a) U|_{A\oplus Y_i \otimes_{A_i} A \otimes \lC \delta_i }
&=\psi (a) (U_i \otimes 1_A)|_{A\oplus Y_i \otimes_{A_i} A } \\
&= ( U_i \otimes 1_A)(\lambda_A(a) \oplus \pi_{Y_i} (a) \otimes 1_A)|_{A\oplus Y_i \otimes_{A_i} A } \\
&=U (\lambda_A (a) \oplus \pi_Y (a) \otimes 1_{\ell^2(\cI)} )|_{A\oplus Y_i \otimes_{A_i} A \otimes \lC \delta_i }.
\end{align*}
By Proposition \ref{prop-left-right}, we also have
\begin{align*}
\psi (a) U |_{( X\ominus X_i ) \otimes_ B A \otimes \lC \delta_i }
&=( \tau_i (a) \otimes 1_A ) \rho_i^A (S_i \otimes 1_A) |_{( X\ominus X_i ) \otimes_ B A} \\
&=( \lambda_i (\pi_{X_i} (a)) \otimes 1_A ) \rho_i^A (S_i \otimes 1_A) |_{( X\ominus X_i ) \otimes_ B A} \\
&=\rho_i^A(S_i \otimes 1_A)( \lambda_i (\pi_{X_i} (a)) \otimes 1_A ) |_{( X\ominus X_i ) \otimes_ B A} \\
&=U ( \lambda_A (a) \oplus \pi_Y (a) \otimes 1_{\ell^2(\cI)} )|_{( X\ominus X_i ) \otimes_ B A \otimes \lC \delta_i }.
\end{align*}
Similarly, for any $j \in \cI$ with $j \neq i$, identifying $Y \otimes \lC \delta_j$ with $Y$ we have
\begin{align*}
\psi (a) U |_{Y \otimes \lC \delta_j }
&=( \tau_i (a) \otimes 1_A ) \rho_j^A (S_j \otimes 1_A) \\
&=( \lambda_i (\pi_{X_i} (a)) \otimes 1_A ) \rho_j^A (S_j \otimes 1_A) \\
&=\rho_j^A(S_j \otimes 1_A)( \lambda_i (\pi_{X_i} (a)) \otimes 1_A )\\
&=U ( \lambda_A (a) \oplus \pi_Y (a) \otimes 1_{\ell^2(\cI)} )|_{Y\otimes \lC \delta_j }.
\end{align*}
Hence, we get $\psi (a) U = U ( L(a) \oplus \pi_Y (a) \otimes 1_{\ell^2(\cI)} )$ for $a \in A_i$.
\end{proof}
We are now ready to prove Theorem \ref{thm-KK}.
\begin{proof}[Proof of Theorem \ref{thm-KK}]
Let $A:=\bigstar_{B, i \in \cI} A_i$ and $A_\re:=\bigstar_{B, i \in \cI} (A_i, E_i)$ and $\pi_\re : A \to A_\re$ be the canonical surjection.
We show that there exists $\cY \in \lE (A_\re, A)$ such that $\pi_\re \otimes_{A_\re} [ \cY] + \id_A = 0$ and $[ \cY ] \otimes_A \pi_\re + \id_{A_\re}=0$.
To simplify the notation we omit the canonical inclusion maps $\iota_i : A_i \to A$ and $\iota : B \to A$.
Put $(X_i, \pi_{X_i}, \xi_i):=(L^2(A_i, E_i) \otimes_B Z_i^\infty, \pi_{E_i} \otimes 1_{Z_i^\infty}, \xi_{E_i} \otimes \eta_i \otimes \delta_1)$ and
$(Y_i ,\pi_{Y_i} ):= (X_i \otimes_B A_i, \pi_{X_i} \otimes 1_{A_i}) \in \Corr (A_i)$.
Define $(X,\pi_X) \in \Corr (A, B)$ by $(X, \xi_0):=\bigstar_{i \in \cI} (X_i, \xi_i)$ and $\pi_X:=\bigstar_{ i\in \cI} (\lambda_i \circ\pi_{X_i})$, where $\lambda_i : \lL_B (X_i) \to \lL_B (X)$ is the canonical left action and set $(Y, \pi_Y):=(X \otimes_B A, \pi_X \otimes 1_A)$.
We identify $X_i$ with the canonical copy of $X_i$ in $X$ and $\xi_i$ with $\xi_0$.
We also use the notation that $A_i^\circ = \ker E_i$ and $X_i^\circ = X_i \ominus \xi_0 B$.
Since $Y_i =(L^2(A_i, E_i)\otimes_B \otimes Z_i \otimes_B A_i )^\infty$,
by Theorem \ref{thm-K-nuclear}, there exist unitaries $U^{(i)} \in \lL_{CA_i} (CA_i \oplus IY_i, IY_i)$ satisfying the following:
\begin{itemize}
\item[(1)] the triple $\cX_i=(IY_i \hoplus IY_i, U^{(i)} ( \lambda_{A_i} \otimes 1_{C_0[0,1)} \oplus \pi_{IY_i} )U^{(i)*} \hoplus \pi_{IY_i}, J_{IY_i})$ forms a Kasparov $A_i$-$IA_i$ bimodule;
\item[(2)] $\{ U^{(i)}_t \}_{0 \leq t <1}$ satisfies that $U_t ( \lambda_A(b) \oplus \pi^\infty_X(b))U_t^*=\pi_X^\infty(b)$ for all $b\in B$ and $t\in [0,1)$;
\item[(3)] the evaluation $U^{(i)}_1$ of $U$ at $1$ equals $1_{Y_i}$;
\item[(4)] the evaluations $\{U_t^{(i)} \}_{0\leq t \leq 1}$ enjoy that
$U^{(i)}_t ( \cos (\pi t) 1_{A_i} \oplus \sin (\pi t) \xi_i \otimes 1_{A_i} ) = \xi_i \otimes 1_{A_i}$ for $0 \leq t \leq 1/2$ and $U_t (0 \oplus \xi_i \otimes 1_{A_i})= \xi_i\otimes 1_{A_i}$ for $1/2 \leq t < 1$.
\end{itemize}
Let $\tau_i : A_i \to \lL_B (X\ominus X_i)$ be the compression of $\pi_{X} \circ \iota_i$ on $X\ominus X_i$.
We note that the isomorphism
$IY=IY_i \otimes_{IA_i} IA \oplus (X\ominus X_i) \otimes_B IA$
induces
$$
(IY, \pi_{IY})
\cong
\left( IY,
\us \bigstar_{i \in \cI} \left( \pi_{I Y_i} \otimes 1_{IA} \oplus \tau_i \otimes 1_{IA} \right) \right).
$$
Set $\psi^{(i)}:=U^{(i)} ( L_{A_i} \otimes 1_{C_0[0,1)} \oplus \pi_{IY_i})U^{(i)*} : A_i \to \lL_{IA_i} (IY_i)$.
By (2) it follows that $\psi^{(i)}|_B = \pi_{Y_i} |_B$ for all $i \in \cI$.
Hence, by the universality of $A$ we have the $*$-homomorphism
$$
\psi=\us\bigstar_{i \in \cI} \left( \psi^{(i)} \otimes 1_{IA} \oplus \tau_i \otimes 1_{IA} \right) : A \to \lL_{IA} (IY).
$$
We claim that the triple $\cX:=(IY \hoplus IY, \psi \hoplus \pi_{IY}, J_{IY} )$ defines a Kasparov $A$-$IA$ bimodule.
It suffices to show that for any $a \in A$, we have $\psi (a) - \pi_{IY} (a) \in \lK_{IA} (IY)$.
We may assume that $a$ has the form of $a_1 \cdots a_p$,
where $p \in \lN$ and $i \in \cI_p$ and $a_k \in A_{i(k)}$.
We then have
\begin{align*}
\psi (\prod_{k=1}^p a_k ) - \pi_{IY} (\prod_{k=l}^p a_l)
=\sum_{k=1}^p ( \prod_{n=1}^{k-1} \psi (a_n) ) ( \psi (a_k) - \pi_{IY} (a_k) ) ( \prod_{m=k+1}^p \pi_{IY} (a_m) ).
\end{align*}
This is compact since $\psi (a_k) - \pi_{IY} (a_k)$ equals 0 on $(X\ominus X_k) \otimes_B IA$ and equals $( \psi^{(i(k))} (a_k) - \pi_{IY_k} (a_k) ) \otimes 1_{IA} \in \lK_{IA_k} (IY_k) \otimes \lK_{IA} (IA)$ on $IY_k \otimes_{IA_k} IA$.
By (4) each unitary $U_0^{(i)}: A \oplus Y_i \to Y_i$ satisfies that $U_0^{(i)} (1_A \oplus 0 ) = \xi_i \ \otimes 1_{A_i}$.
Set $S_i := U^{(i)}_0|_{0\oplus Y_i}$, which is an isometry in ${}_B\lL_{A_i} (Y_i)$.
Let $\psi_t$ be the evaluation of $\psi$ at $t \in [0,1]$.
We observe that $\psi_0 (a) = \bigstar_{i \in \cI} ( U^{(i)}_0 ( L_{A_i} \oplus \pi_{Y_i} )U_0^{(i)*} \oplus \tau_i \otimes 1_A )$.
By Lemma \ref{lem-unitary} the operator $U := P_A \oplus \bigoplus_{i \in \cI} \rho_i^A (S_i \otimes 1_A) : A \oplus Y\otimes \ell^2 (\cI) \to Y$ is a unitary satisfying that $\psi_0 (a) = U (\lambda_A (a) \oplus \pi_Y (a) \otimes 1_{\ell^2(\cI)} ) U^*$ for all $a \in A$.
Thus, we have
\begin{align*}
\cX_0 &= ( Y \hoplus Y, \psi_0 \hoplus \pi_Y , J_{Y} ) = ( Y \hoplus Y,U (\lambda_A \oplus \pi_Y \otimes 1_{\ell^2 (\cI )}) U^* \hoplus \pi_Y , J_{Y} ) \\
& \cong \left(A \oplus Y \otimes \ell^2 (\cI) \hoplus Y, \lambda_A \oplus \pi_Y \otimes 1_{\ell^2 (\cI)} \hoplus \pi_Y, \begin{bmatrix} 0& U^* \\ U & 0 \end{bmatrix} \right).
\end{align*}
Since $U$ is a compact perturbation of $S:=\bigoplus_{i \in \cI} \rho_i^A (S_i \otimes 1_A) \lL_A (Y \otimes \ell^2 (\cI), Y)$,
this Kasparov bimodule is homotopic to
$$
(A \hoplus 0, \lambda_A \hoplus 0, 0 ) \oplus \left( Y \otimes \ell^2( \cI) \hoplus Y, \pi_Y \otimes 1_{\ell^2(\cI )} \hoplus \pi_Y, \begin{bmatrix} 0& S^* \\ S & 0 \end{bmatrix} \right).
$$
Since evaluations of $\cX_i$'s at 1 are degenerate, so is the one of $\cX$.
By Lemma \ref{lem-throughAr} there exists a $*$-homomorphism $\bar{\pi}_\re : A_\re \to \lL_A (Y)$ such that $\pi_X = \bar{\pi}_X \circ \pi_\re$.
Thus, letting
\[
\cY:= \left( Y \otimes \ell^2 (\cI) \hoplus Y, ( \bar{\pi}_X \otimes 1_A) \otimes 1_{\ell^2( \cI)} \hoplus \bar{\pi}_X \otimes 1_A , \begin{bmatrix} 0& S^* \\ S & 0 \end{bmatrix} \right)
\]
we have $0 = [ \cX_1] = [ \cX_0 ] = \id_A + \pi_\re \otimes_{A_\re} [\cY]$.
\medskip
We next prove that $[\cY ] \otimes_A \pi_\re + \id_{A_\re}=0$ in $KK (A_\re, A_\re)$.
Our proof depends on the argument in \cite[Section 4]{Germain-duke}.
We note that $[\cY ] \otimes_A \pi_\re$ is represented by
$$
\cY_{\pi_\re} = \left( X \otimes_B A_\re \otimes \ell^2 (\cI) \hoplus X \otimes_B A_\re, ( \bar{\pi}_X \otimes 1_{A_\re}) \otimes 1_{\ell^2( \cI)} \hoplus \bar{\pi}_X \otimes 1_{A_\re} , \begin{bmatrix} 0& S_{\pi_\re}^* \\ S_{\pi_\re} & 0 \end{bmatrix} \right).
$$
Let $\widetilde{\pi}_\re :=\pi_\re \otimes \id_{C[0,1]} :IA \to IA_\re$ and consider the pushout of $\cX$ by $\widetilde{\pi}_\re$,
which is the Kasparov $A$-$A_\re$ bimodule
$(I(X\otimes_B A_\re) \hoplus I(X\otimes_B A_\re), \psi_{\widetilde{\pi}_\re } \hoplus (\pi_X \otimes 1_{A_\re}) \otimes 1_{C[0,1]}, J_{I(X\otimes_B A_\re)}).$
We claim that the $*$-homomorphism $\psi_{\widetilde{\pi}_\re }$ factors through $A_\re$.
For any $ a \in A$ we have $ \|\psi_{\widetilde{\pi}_\re } (a) \| = \sup_{0 \leq t \leq 1} \|( \psi_{\widetilde{\pi}_\re })_t (a) \|$.
Since $ (\psi_{\widetilde{\pi}_\re })_t = (\psi_t)_{\pi_\re } : A \to \lL_{A_\re} (X\otimes_B A_\re )$,
it is sufficient to see that the $(\psi_t)_{\pi_\re }$ factors through $A_\re$ for every $t \in [0,1]$.
Furthermore, since $\bar{\pi}_X : A_\re \to \lL_B (X)$ is injective, it enoughs to show that $\psi_t \otimes 1_X: A \to \lL_B (Y \otimes_A X)= \lL_B (X\otimes_B X)$ factors through $A_\re$.
Recall that $\lambda_i^{X_i}: \lL_B (X_i \otimes_B X_i ) \to \lL_B (X_i \otimes_B X)$ is the $*$-homomorphism given by $x \mapsto (1_{X_i} \otimes v_i^*)(x \otimes 1_{X(\lambda, i)} ) (1_{X_i} \otimes v_i).$
Now we have
$$
\psi_t \otimes 1_X = \us\bigstar_{i \in \cI} \left[ \psi_t^{(i)} \otimes 1_X \oplus \tau_i \otimes 1_X \right]
=\us\bigstar_{i \in \cI} \left[ \lambda_{i}^{X_i}( \psi_t^{(i)} \otimes 1_{X_i}) \oplus \tau_i \otimes 1_X \right].
$$
Consider the natural isomorphism $T_i : (A_i \oplus Y_i ) \otimes_{A_i} X_i \cong X_i \oplus X_i \otimes_B X_i \cong X_i \otimes_B (B \oplus X_i)$ and set $W_t^{(t)}:=(U_t^{(i)} \otimes 1_{X_i} )\circ T_i^* : X_i \otimes_B (B \oplus X_i) \to X_i \otimes_B X_i$.
We then have
$$
\psi_t^{(i)} \otimes 1_{X_i} = W_{t}^{(i)} (\pi_{X_i} \otimes 1_{B \oplus X_i}) W_t^{(t)*}.
$$
We note that $\xi_i \otimes \xi_i=W_t^{(i)} (\xi_i \otimes (\cos (\pi t) 1_B \oplus \sin (\pi t ) \xi_i) ) $ for $0 \leq t \leq 1/2 $ and $\xi_i \otimes \xi_i =W_t^{(i)} ( \xi_i \otimes (0 \oplus \xi_i ) $ for $1/2 \leq t \leq 1$.
Hence, Lemme \ref{lem-throughAr2} says that $\psi_{\widetilde{\pi}_\re}$ factors through $A_\re$.
Thus, there exists a $*$-homomorphism $\phi : A_\re \to \lL_{IA_\re} (I(X\otimes_B A_\re))$ satisfies that $\psi_{\widetilde{\pi}_\re } = \phi \circ \pi_\re$.
Then $\cZ=(
I(X \otimes_B A_\re) \hoplus I(X\otimes_B A_\re),
\phi \hoplus (\bar{\pi}_X \otimes 1_{A_\re}) \otimes 1_{C[0,1]},
J_{I(X\otimes_B A_\re)}
)$ forms a Kasparov $A_\re$-$IA_\re$ bimodule and satisfies that $\pi_\re^* ( \cZ ) = \cX_{\widetilde{\pi}_\re}$.
We note that $\pi_\re^* ( \cZ_0)
=(\pi_\re^* (\cZ))_0
= ( \cX_{\widetilde{\pi}_\re} )_0
=(\cX_{\widetilde{\pi}_\re })_0$,
which consists of the Hilbert $\rC^*$-module $ A_\re \oplus (X \otimes_B A_\re) \otimes \ell^2 (\cI) \hoplus X \otimes_B A_\re$,
the left action $ \pi_\re \oplus ( \pi_X \otimes 1_{A_\re}) \otimes 1_{\ell^2 (\cI)} \hoplus ( \pi_X \otimes 1_{A_\re})$,
and the degree 1 operator $\left[ \begin{smallmatrix} 0& U_{\pi_\re}^* \\ U_{\pi_\re} & 0 \end{smallmatrix} \right]$.
Hence, the evaluation $\cZ_0$ is just
$$
\left( A_\re \oplus (X \otimes_B A_\re) \otimes \ell^2 (\cI) \hoplus X \otimes_B A_\re, \lambda_{A_\re }\oplus ( \bar{\pi}_X \otimes 1_{A_\re}) \otimes 1_{\ell^2 (\cI)} \hoplus ( \bar{\pi}_X \otimes 1_{A_\re}), \begin{bmatrix} 0& U_{\pi_\re}^* \\ U_{\pi_\re} & 0 \end{bmatrix} \right).
$$
Since $[ \cZ_1] =0$ and $U_{\pi_\re}$ is a compact perturbation of $S_{\pi_\re}$,
it follows that $0 = [\cZ_0]= \id_{A_\re} + [\cY_{\pi_\re}] = \id_{A_\re} + [\cY] \otimes_{A} \pi_\re$.
\end{proof}
\setcounter{theorem}{0}
\renewcommand{\thetheorem}{\arabic{section}.\arabic{theorem}}
\section{Applications}\label{sec-app}
In this section, we give some applications in $KK$-theory.
The next theorem follows from Thomsen's result on full amalgamated free products (\cite[Theorem 2.7]{Thomsen}) and Corollary D.
\begin{theorem
Let $A_1, A_2$, and $B$ be unital separable $\rC^*$-algebras, and $i_k :B \to A_k, k=1,2$ be unital embedding with nondegenerate conditional expectations $E_k : A_k \to i_k(B)$.
Set $A:=(A_1,E_1) \bigstar_B (A_2,E_2)$.
Let $j_k : A_k \to A, k=1,2$ be the canonical embeddings.
If each $A_i$ is nuclear and $B$ is finite dimensional, then for any separable $\rC^*$-algebra $D$ there are two cyclic six terms exact sequences:
\[
\begin{CD}
KK(D,B) @>(i_{1*},i_{2*}) >> KK(D,A_1) \oplus KK (D, A_2) @>j_{1*}- j_{2*}>>KK( D, A ) \\
@AAA @. @VVV \\
KK(SD,A ) @<j_{1*}- j_{2*}<< KK(SD,A_1 )\oplus KK (SD, A_2) @<(i_{1*},i_{2*}) <<KK( SD, B)\\
\end{CD}\]
\[ \begin{CD}
KK(B,D) @<i_1^*- i_2^*<< KK(A_1,D) \oplus KK (A_2,D) @<j_1^* + j_2^*<< KK( A, D) \\
@VVV @. @AAA \\
KK(A , SD ) @>j_1^* + j_2^*>> KK(A_1, SD )\oplus KK (A_2, SD) @>i_1^*- i_2^*>> KK( B,SD)
\end{CD}
\]
In particular, we have
$$
\begin{CD}
K_0(B) @>(i_{1*},i_{2*}) >> K_0(A_1) \oplus K_0 (A_2) @>j_{1*}- j_{2*}>>K_0( A ) \\
@AAA @. @VVV \\
K_1 (A ) @<j_{1*}- j_{2*}<< K_1 (A_1 )\oplus K_1 (A_2) @< << 0
\end{CD}
$$
\end{theorem}
We next discuss $KK$-equivalence of full and reduced HNN-extensions of $\rC^*$-algebras.
We refer to \cite{Ueda1}\cite{Ueda2} for the definition of HNN-extensions.
\begin{theorem}
Let $B \subset A$ be a unital inclusion of separable $\rC^*$-algebras and $\theta : B \to A$ be an injective $*$-homomorphism.
Assume that there exist conditional expectations $E :A \to B$ and $E_\theta : A \to \theta (B)$ such that the triple $(\lM_2(A), B \oplus \theta (B), E \oplus E_\theta)$ is strongly nuclear,
where
$$
B \oplus \theta (B) =\left \{ \begin{bmatrix} b_1 & 0 \\ 0& \theta (b_2) \end{bmatrix} \middle| b_1, b_2 \in B \right\}, \quad E \oplus E_\theta : \begin{bmatrix} a_1 & a_2 \\ a_3& a_4 \end{bmatrix} \mapsto \begin{bmatrix} E (a_1) & 0 \\ 0& E_\theta (a_4) \end{bmatrix}.
$$
Then the canonical surjection from the full HNN-extension $A \bigstar_B^\univ \theta$ onto the reduced one $(A,E) \bigstar_B ( \theta, E_\theta)$ is a $KK$-equivalence.
\end{theorem}
\begin{proof}
Let $\cA_{\rm f}$ and $\cA_\re$ be the full and reduced amalgamated free product $\lM_2 (A) \bigstar_{B \oplus B } \lM_2 (B)$ and $(\lM_2 (A), E \oplus E_\theta) \bigstar_{B \oplus B } (\lM_2 (B), E_1)$, where the inclusion $B \oplus B \to B \oplus \theta (B) \subset \lM_2(A)$ and the conditional expectation $E \oplus E_\theta$ are as above,
and the inclusion $\iota_1 : B \oplus B \to \lM_2 (B)$ and the conditional expectation $E_1 : \lM_2 (B) \to B$ are defined by
\begin{align*}
\iota_1 :(b_1 \oplus b_2 ) \mapsto \begin{bmatrix} b_1 & 0 \\ 0 & b_2 \end{bmatrix},
\quad E_1: \begin{bmatrix} b_1 & b_2 \\ b_3& b_4 \end{bmatrix} \mapsto \begin{bmatrix} b_1 & 0 \\ 0& b_4 \end{bmatrix}.
\end{align*}
We also set $\cB_{\rm f}:=A \bigstar_B^\univ \theta$ and $\cB_\re:=(A,E) \bigstar_B ( \theta, E_\theta)$.
By remarks following to \cite[Proposition 3.1]{Ueda2} and \cite[Proposition 3.3]{Ueda2},
there exist two $*$-isomorphism $\Phi_{\rm f} : \cA_{\rm f} \to \lM_2(\cB_{\rm f})$ and $\Phi_\re : \cA_\re \to \lM_2 (\cB_\re )$ such that $\sigma_\re ^{(2)} \circ \Phi_{\rm f} = \Phi_\re \circ \pi_\re$, where $\pi_\re : \cA_{\rm f} \to \cA_\re$ and $\sigma_\re: \cB_{\rm f} \to \cB_\re$ are canonical surjections.
Since the triple $(\lM_2 (B), B \oplus B , E_1)$ is also strongly nuclear by Proposition \ref{prop-sum-tensor},
$\pi_\re$ is a $KK$-equivalence.
Consider $*$-homomorphisms $\phi_\epsilon: \cB_\epsilon \to \lM_2 (\cB_\epsilon)$ and $\psi_ \epsilon : \lM_2 (\cB_\epsilon ) \to \cB_\epsilon$ for $\epsilon \in \{\rm f, r \}$ defined by
$$
\phi_\epsilon : b \mapsto \begin{bmatrix} b & 0 \\ 0 & 0 \end{bmatrix}, \quad \psi_\epsilon : \begin{bmatrix} b_1 & b_2 \\ b_3 & b_4 \end{bmatrix} \mapsto b_1,
$$
for $b, b_i \in \cB_\epsilon, i=1,2,3,4$.
Then we have the following commuting diagrams:
$$
\begin{CD}
\cA_{\rm f} @>\Phi_{\rm f} >> \lM_2 (\cB_{\rm f}) @<\phi_{\rm f}<< \cB_{\rm f} \\
@V\pi_\re VV @V\sigma_\re^{(2)} VV @VV\sigma_\re V \\
\cA_\re @>\Phi_\re >> \lM_2 (\cB_\re) @<\phi_\re<< \cB_\re \end{CD}
\qquad \quad
\begin{CD}
\cA_{\rm f} @>\Phi_{\rm f} >> \lM_2 (\cB_{\rm f}) @>\psi_{\rm f}>> \cB_{\rm f} \\
@V\pi_\re VV @V\sigma_\re^{(2)} VV @VV\sigma_\re V \\
\cA_\re @>\Phi_\re >> \lM_2 (\cB_\re) @>\psi_\re>> \cB_\re \end{CD}
$$
Let $\alpha \in KK(\cA_\re, \cA_{\rm f})$ be an element satisfying $\pi_\re \otimes_{\cA_\re} \alpha =\id_{\cA_{\rm f}}$ and $\alpha \otimes_{\cA_{\rm f}} \pi_\re = \id_{\cA_\re}$.
Set $\beta:=\Phi_\re^{-1} \otimes_{\cA_\re} \alpha \otimes_{\cA_{\rm f}} \Phi_{\rm f} \in KK (\lM_2 (\cB_\re), \lM_2 (\cB_{\rm f}) )$.
Then we have $\id_{\cB_{\rm f}} = \phi_{\rm f} \otimes_{\lM_2 (\cB_{\rm f} )} \sigma_\re^{(2)} \otimes_{\lM_2 (\cB_\re )} \beta \otimes_{\lM_2 (\cB_{\rm f})} \psi_{\rm f}
=\sigma_\re \otimes_{\cB_\re } \phi_\re \otimes_{\lM_2 ( \cB_\re )} \beta \otimes_{\lM_2(\cB_{\rm f})} \psi_{\rm f}$ and
$\id_{\cB_\re} =\phi_\re \otimes_{\lM_2 (\cB_\re)} \beta \otimes_{\lM_2 (\cB_{\rm f})} \sigma_\re^{(2)} \otimes_{\lM_2 (\cB_\re)} \psi_\re
=\psi_\re \otimes_{\lM_2 (\cB_\re)} \beta \otimes_{\lM_2 (\cB_{\rm f})} \psi_{\rm f} \otimes_{\lM_2 (\cB_{\rm f})} \sigma_\re$,
hence the $\phi_\re \otimes_{\lM_2 (\cB_\re)} \beta \otimes_{\lM_2 (\cB_{\rm f})} \psi_{\rm f} \in KK (\cB_\re, \cB_{\rm f})$ is the desired inverse element of $\sigma_\re$.
\end{proof}
We remark that when $B=\theta (B)$, the strong nuclearity of $(\lM_2 (A), B \oplus B, E \oplus E)$ follows from the one of $(A, B, E)$.
Combining the theorem above with \cite[Proposition 4.12]{Ueda2} we obtain the following six term exact sequence for reduced HNN-extensions.
\begin{corollary}
Let $B \subset A$ be a unital inclusion of separable $\rC^*$-algebras and $\theta : B \to A$ be an injective $*$-homomorphism.
Assume that there exist conditional expectations $E :A \to B$ and $E_\theta : A \to \theta (B)$ such that the triple $(\lM_2(A), B \oplus \theta (B), E \oplus E_\theta)$ is strongly nuclear, then we have the following six term exact sequence:
$$
\begin{CD}
K_0(B) @>(\theta_{1*}- \iota_{B*}) >> K_0(A) @>\iota_{A*}>>K_0( (A,E) \bigstar_B ( \theta, E_\theta)) \\
@AAA @. @VVV \\
K_1 ((A,E) \bigstar_B ( \theta, E_\theta) ) @<\iota_{A*}<< K_1 (A ) @<(\theta_{1*}- \iota_{B*}) <<K_1( B)
\end{CD}
$$
Here $\iota_B : B \to A$ and $\iota_A: A \to (A,E) \bigstar_B ( \theta, E_\theta) $ are inclusion maps.
\end{corollary}
|
\section{Introduction}
\hspace{5mm}
Minkowski spacetime has Poincar\'e group as an isometry. When the geometry fluctuates, then there is, in general, no exact isometry. It turns out, however, that when the metric tensor satisfies a certain appropriate asymptotic conditions, spacetime has an asymptotic symmetry. At spatial infinity, there is asymptotic Poincar\'e symmetry. At null infinity the asymptotic symmetry is enhanced, and ordinary global translations are extended to local translation (supertranslation) in the retarded time $u$. The resulting asymptotic symmetry group, which is composed of supertranslation and Lorentz group, is called BMS$_4$ group.\cite{BBM}\cite{Sachs}
In general relativity, conserved quantities such as Hamiltonian and angular momentum are represented as surface charges.\cite{Regge}\cite{BrownHenneaux}
Energy and angular momentum of asymptotically flat spacetime at spatial infinity were represented as surface integrals.\cite{Komar}
In 3 dimensional asymptotically anti-de Sitter spacetime, there is asymptotic infinite-dimensional conformal symmetry at the infinite boundary, and there exist associated surface charges at the boundary. The Dirac bracket algebra of these charges (direct sum of two Virasoro algebras) was derived in the Hamiltonian framework.\cite{BrownHenneaux}
On the null infinity of asymptotically flat spacetime, energy, momentum and angular momentum are not conserved due to non-vanishing fluxes via gravitational radiation. Difficulty was recognized in defining surface charges associated with these non-conserved quantities on the null infinity.\cite{Wald} It was argued that Hamiltonian framework is not applicable for defining surface charges corresponding to non-conserved quantities at null infinity, and that extra terms must be added to the variation equation for surface charges. By using covariant framework, such an extra term was proposed in \cite{Wald}. It was, however, not addressed in \cite{Wald} whether the surface charges correctly generate the Dirac bracket algebra of BMS$_4$ symmetry.
In \cite{BT2} Barnich and Troessaert carried out calculation of the Dirac bracket algebra of the extended BMS$_4$ by using covariant Lagrangian framework\cite{BB}. The extended BMS$_4$ group contains an infinite dimensional extension of the Lorentz group (superrotation), which may have singularities on the S$^2$ at null infinity.\cite{BT4}\cite{Barnich}\cite{Banks}\cite{Strominger3} In \cite{BT2} it was shown that the variation equation for the surface charges $Q$ is not integrable, and the Dirac bracket algebra of extended BMS$_4$\cite{BT2} at the null infinity was proposed by defining the algebra of the `integrable part' of the surface charges by modifying the non-integrable variation $ \Slash{\delta} Q$. An interesting proposal on algebra was made and it was shown that there appear central extensions which depend on the gravitational radiation field $C_{AB}$ on the null infinity. This proposal has not been justified.
In the analysis via the covariant framework, however, the variation equation for the surface charges is still not integrable with the above modification of $ \Slash{\delta} Q$, and the following questions arise.
Is it possible to obtain the surface charges and their algebra in the Hamiltonian framework?
Is the variation equation also non-integrable?
The purpose of this paper is to show that the algebra of BMS$_4$ group at null infinity can be also studied via Hamiltonian framework, and to reconsider the algebra of surface charges.
For this purpose we adopt a coordinate system, where boundaries of constant-time hypersurfaces cross the null infinity. The surface charges are obtained as the boundary terms which must be added to the Hamiltonian in order to make the calculation of the Poisson bracket of Hamiltonians well-defined, even if partial integrations are required\cite{Regge}\cite{BrownHenneaux}. The variation of the surface charge, $\delta Q$, is uniquely determined and this defines the Dirac bracket algebra of the surface charges as $\{Q[\xi], Q[\xi']\}_D=\delta_{\xi'} Q[\xi]+K[\xi,\xi']$, where $K[\xi,\xi']$ is a central extension.
In this paper, the constraint equation which determines the variation of the surface charges $\delta Q$ is obtained and found to be the same as the result \cite{BT2} in the covariant framework. This shows equivalence of the two frameworks for the present problem. It is confirmed that
if a gravitational radiation field $C_{AB}$ depends on the retarded time $u$, surface charges are not conserved and the equation for $\delta Q$ is not integrable.
While it is possible to add an appropriate additional term, {\em i.e.}, (\ref{extra}) to the variation equation to cancel the non-integrable piece, it is not possible to make the charges generate appropriate symmetry algebra. If $C_{AB}$ does not depend on $u$, however, the variation equation becomes integrable. In this case the charges are conserved and generate symmetry algebra.
Due to compatibility of the transformation rule of $C_{AB}$ and the condition $\partial_u C_{AB}=0$, symmetry algebra is restricted to global BMS$_4$ and there is no central extension. On the other hand, if $\partial_u C_{AB} \neq 0$ and hence charges are not conserved, appropriate charges which generate BMS$_4$ algebra are not found. Conclusion in this paper is different from that in \cite{BT2}.
We conclude this introduction with an outline of this paper.
In sec.2 a coordinate system defined by foliation of spacelike hypersurfaces which are asymptotically Euclidean Anti-de Sitter \cite{SolodukindeBoer} is introduced. Each hypersurface has an infinite boundary on the null infinity. In sec.3
Bondi frame metric without fixing the gauge of the metric component $g_{ut}$ is presented. The gauge of this component will not be fixed in this paper because except for gauge $g_{ut}=(1/16r^2) C_{AB}C^{AB} +O(r^{-2})$ its fixing generally breaks BMS$_4$ group. In sec.4 asymptotically flat metric in the coordinate system of sec.2 is obtained by a coordinate transformation. In sec.5 surface charge algebra and surface charges are obtained via Hamiltonian framework. It is shown that surface charges exist only for $C_{AB}$ independent of $u$. Sec.6 is left for summary.
\section{Asymptotically flat space via foliation of EAdS (and dS) hypersurfaces}
\hspace{5mm}
4d Minkowski space has two types of infinity. One is a spatial infinity (spi).
In standard coordinates $ds_{\text{standard}}^2=-dt^2+dr^2+r^2 \, (d\theta^2+\sin^2 \theta d\phi^2)$ this is defined by a limit $r \rightarrow \infty$ with fixed time $t$. The other is a null infinity ${\cal I}$ \cite{Penrose1}. In coordinates,
\begin{equation}
ds_0^2 =g_{\mu\nu} \, dx^{\mu} \, dx^{\nu}= -du^2-2du \, dr+r^2 \, (d\theta^2+\sin^2 \theta \, d\phi^2), \label{Minkowski1}
\end{equation}
where $u=t-r$ is a retarded time, it is defined by an unphysical metric obtained by a conformal rescaling with $\Omega=r^{-1}$.
\begin{equation}
d\tilde{s}_0^2=\tilde{g}_{\mu\nu} \, dx^{\mu} \, dx^{\nu}= \Omega^2 \, ds_0^2= -\Omega^2 \, du^2+2du \, d\Omega+ (d\theta^2+\sin^2 \theta \, d\phi^2) \label{ut}
\end{equation}
This metric $\tilde{g}_{\mu\nu}=\Omega^2 \, g_{\mu\nu}$ is regular and non-degenerate at infinity $\Omega=0$, and defines differential and topological structures of ${\cal I}$. \footnote{This is a future null infinity, ${\cal I}^+$.There also exists a past null infinity, ${\cal I}^-$, defined with an advanced time $v=t+r$. In this paper we will concentrate on ${\cal I}^+$ and denote it as ${\cal I}$, since an extension to ${\cal I}^-$ is straightforward.}
Asymptotically flat spacetimes are defined in a similar way, and their unphysical metrics $\tilde{g}_{\mu\nu}$ are well defined and conformally flat at ${\cal I}$. These spacetimes are also characterized by an infinite-dimensional symmetry group.\cite{BBM}\cite{Sachs} While Minkowski space has Poincar\'e symmetry as an isometry, curved spacetimes in general do not have isometries. However, asymptotically flat spacetimes have asymptotic symmetry group, BMS$_4$ group, at ${\cal I}$. Penrose interpreted this group as conformal isometries of ${\cal I}$.\cite{Penrose2} BMS$_4$ group is a direct product of a supertranslation group and a Lorentz group.
Locally, Minkowski space can be represented as foliations of Euclidean Anti-de Sitter spaces (EAdS) and de Sitter spaces (dS). \cite{SolodukindeBoer}
These foliations are defined (in the case of EAdS) by
\begin{eqnarray}
ds_1^2 &=& -dX_0^2+dX_1^2+dX_2^2+dX_3^2, \nonumber \\
&& (-\tau^2=-X_0^2+X_1^2+X_2^2+X_3^2).
\end{eqnarray}
With a parametrization
\begin{eqnarray}
X_0 &=& \tau \, \cosh \rho, \nonumber \\
X_1 &=& \tau \, \sinh \rho \, \sin \theta \cos \phi, \nonumber \\
X_2 &=& \tau \, \sinh \rho \, \sin \theta \sin \phi, \nonumber \\
X_3 &=& \tau \, \sinh \rho \, \cos \theta,
\end{eqnarray}
we obtain the metric inside the future ($\tau >0$) and the past ($\tau <0$) regions of the lightcone put at some vertex (origin).
\begin{equation}
ds_2^2 = -d\tau^2+\tau^2 \, (d\rho^2+\sinh^2 \rho \, (d\theta^2+\sin \theta \, d\phi^2)) \label{global}
\end{equation}
A hypersurface $\Sigma_{\tau}$ with constant $\tau$ is EAdS in global coordinates.
In the spacelike region outside the lightcone put at the origin we have a metric
\begin{equation}
ds_3^2 = d\sigma^2-\sigma^2 \, d\lambda^2+\sigma^2 \, \cosh^2 \lambda \, (d\theta^2+\sin^2 \theta \, d\phi^2).
\end{equation}
This metric corresponds to a foliation in terms of dS hypersurfaces
$-X_0^2+X_1^2+X_2^2+X_3^2=\sigma^2$ in global coordinates.
The whole Minkowski spacetime is obtained by patching up these local coordinates constructed as foliations of EAdS$_3$ and dS$_3$. It is expected that (at least) some class of asymptotically flat spaces can also be obtained by foliations. In this paper this kind of asymptotically flat space is studied.
In coordinates $\{u,r,\theta,\varphi\}$, null infinity ${\cal I}^+$ is reached by taking limit $r \rightarrow +\infty$ with $u=t-r$ fixed. The retarded time $u$ ranges over $-\infty < u < \infty$. In coordinates $\{\tau, \rho, \theta, \varphi\}$, however, ${\cal I}^+$ is defined by $\rho \rightarrow \infty$ with $u=\tau \, e^{-\rho}$ fixed. In the forward lightcone time $\tau$ is positive, and only half $u>0$ of ${\cal I}^+$ is covered. Hence $u=0$ is a horizon in the local coordinates. The advantage of the coordinates (\ref{global}) over (\ref{ut}) is that constant-time hypersurfaces $\Sigma_{\tau}$ cross at infinite boundary ($\rho \rightarrow +\infty$) with ${\cal I}$.
Hence the Hamiltonian framework can be used to study asymptotic symmetry algebra at ${\cal I}$.
\section{Symmetry of asymptotically-flat space in Bondi frame}
\hspace{5mm}
Asymptotically flat spacetime in Bondi frame\cite{BBM}\footnote{The coordinate (\ref{B}) with $\beta=-(1/16) \, C^{AB}C_{AB}$ is called Bondi frame in the literature. Instead, hereafter we will call (\ref{B}) with any $\beta$ Bondi frame.} has the following line element represented in the form of $1/r$ expansions.
\begin{eqnarray}
ds_{\text{Bondi}}^2 &=& -( 1-\frac{2 m}{r}-\frac{2 m_1}{r^2} +\cdots) \, du^2 -2 \, (1+\frac{\beta}{r^2}+\cdots ) \, dudr \nonumber \\
&&-2(U_A +\frac{1}{r} \, W_A+\cdots) \, du \, dx^A+ \gamma_{AB} \, dx^Adx^B, \label{B}\\
\gamma_{AB} &=& r^2 \, h_{AB}+ r \, C_{AB}+H_{AB}+\cdots \label{Bondi frame}
\end{eqnarray}
Here $A,B=z,\bar{z}$ and $x^z \equiv z$, $x^{\bar{z}} \equiv \bar{z}$ are complex coordinates for round $S^2$. It has standard metric, $h_{z\bar{z}}=2/(1+z\bar{z})^2$, $h_{zz} \equiv h_{\bar{z}\bar{z}}=0$. Fields
$m$, $\beta$, $U_A$, $W_A$, $C_{AB}$, $H_{AB}$ are functions of $u$, $z$ and $\bar{z}$.
$C_{AB}$ is traceless ($C_{z\bar{z}}=0$). These functions are determined by solving Einstein equation ($R_{\mu\nu}-\frac{1}{2} g_{\mu\nu}R=0$).
\begin{eqnarray}
U_A&=&-\frac{1}{2} \, D^B \, C_{BA}, \label{U}\\
\dot{ m} &=& -\frac{1}{8} \, \dot{C}^{AB} \, \dot{C}_{AB}-\frac{1}{4} \, \partial_u^2 \, \Big({H^A}_A-\frac{1}{2} \, C^{AB} C_{AB}\Big) -\frac{1}{2}\, D_A \, \dot{U}^A, \label{dm} \\
\beta &=&-\frac{1}{4} \, {H^A}_A+\frac{1}{16} \, C^{AB}C_{AB}, \label{beta} \\
\frac{3}{2} \, \dot{W}_A &=& -D_A \, m+\frac{1}{2} \, D_A \dot{\beta}-\frac{1}{2} \, D^B\dot{H}_{BA}+\frac{1}{2} D_A{\dot{H}^C}\, _C-\frac{1}{2} \, D_B D^B U_A \nonumber \\
&& +\frac{1}{2} \, D_BD_A U^B+\frac{1}{2}\partial_u(C_{AB}U^B)-\frac{1}{4} \, \dot{C}^{BC}D_AC_{BC} \nonumber \\ &&-\frac{1}{2} C^{BC}D_A \dot{C}_{BC}+\frac{1}{2} \, D^B \, \big(C_{BC} \, {{\dot{C}^C\,}}_A \big), \label{W}\\
H_{zz} &=& H_{zz}(z), \quad H_{\bar{z}\bar{z}}=H_{\bar{z}\bar{z}}(\bar{z}). \label{Hzz}
\end{eqnarray}
Here dot means a derivative with respect to $u$. Indices $A$, $B$... are raised and lowered by means of $h_{AB}$ and its inverse $h^{AB}$. $D_A$ is a covariant derivative with respect to $h_{AB}$, and non-zero Christoffel symbols are given by $\Gamma^z_{zz}=-2\, \bar{z}\, (1+z\bar{z})^{-1}$ and $\Gamma^{\bar{z}}_{\bar{z}\bar{z}}=-2z\, (1+z\bar{z})^{-1}$. There is also a $u$-flow equation for $m_1$, which is not presented here.
Function $\beta$ is gauge-fixed in several ways in the literature.\footnote{Our $\beta$ is different from that in \cite{BBM}.} For instance, a gauge $\beta=0$ is chosen in the null tetrad formalism\cite{Newman}. Another gauge $\beta= -(1/16) \, C_{AB}C^{AB} $ is chosen in \cite{BBM} so that $\sqrt{\text{det} \gamma_{AB}}=r^2 \, \sqrt{\text{det}h_{AB}}$. It can be shown that gauge condition such as
\begin{equation}
\beta=\lambda \, C^{AB} \, C_{AB}, \label{beta gauge}
\end{equation}
where $\lambda$ is a constant, cannot be preserved under BMS$_4$ group except for a value $\lambda=-1/16$. That is, gauge fixing (\ref{beta gauge}) generally breaks BMS$_4$ group. This will be shown at the end of this section. For this reason $\beta$ will {\em not} be fixed in this paper.
The eqs of motion (\ref{U})-(\ref{Hzz}) for gauge $\beta=-\frac{1}{16} \, C_{AB}C^{AB}$ are obtained in \cite{BBM}\cite{Sachs}\cite{Barnich}.
Asymptotically flat spacetime has BMS$_4$ group as asymptotic symmetry. This is composed of a group of (i) supertranslation and that of (ii) superrotation. Supertranslation (i) is a group of local translation in retarded time $u$. The corresponding asymptotic Killing vector $\xi^{\mu}$ is given by \cite{Sachs}\cite{BT2}
\begin{eqnarray}
\delta u &= & \xi^u \equiv f, \\
\delta r &=& \xi^r\equiv D^zD_z f +\frac{1}{r} \, (U^A \, D_A f-\frac{1}{4} \, C^{AB} \, D_A D_B f)+\cdots, \\
\delta z &=& \xi^z\equiv -\frac{1}{r} \, D^z f+\frac{1}{2 r^2} \, C^{zz} \, D_z \, f \label{z1}+\cdots, \\
\delta \bar{z} &=&\xi^{\bar{z}}\equiv -\frac{1}{r} \, D^{\bar {z}} f+\frac{1}{2 r^2} \, C^{\bar{z}\bar{z}} \, D_{\bar{z}} \, f+\cdots. \label{zb1}
\end{eqnarray}
Here $f=f(z,\bar{z})$ is a scalar function on a sphere. Subleading terms ($\cdots$) are suppressed here.
Superrotation (ii) corresponds to extension of Lorentz transformation in Minkowski space and has an asymptotic Killing vector, \cite{Sachs}
\begin{eqnarray}
\delta u &=& \xi^u \equiv \frac{1}{2} \, u \, D_z \zeta^z, \\
\delta r &=& \xi^r \equiv -\frac{1}{2} \, (r+u) \, D_z \zeta^z+\frac{u}{2r} \, \Big(U^z \, D_z^2 \zeta^z-U_z \, \zeta^z-\frac{1}{4} \, C^{zz} \, D_z^3\zeta^z\Big)+\cdots, \\
\delta z &=& \xi^z \equiv \zeta^z+\frac{u}{2r}\, \zeta^z -\frac{u}{4r^2} \, C^{zz} \, D^2_{z} \, \zeta^{z}+\cdots, \label{z2} \\
\delta \bar{z} &=& \xi^{\bar{z}} \equiv -\frac{u}{2r} \, h^{z\bar{z}} \, D^2_z \, \zeta^{z} -\frac{u}{4r^2} \, h_{z\bar{z}}\, C^{\bar{z}\bar{z}} \,\zeta^{z}+\cdots. \label{zb2}
\end{eqnarray}
Here $\zeta^z=\zeta^z(z)$ is a polynomial of $z$ up to quadratic order. There is also a similar transformation with a parameter function $\zeta^{\bar{z}}=\zeta^{\bar{z}}(\bar{z})$. For $\zeta^z=1,i,z,iz, z^2,iz^2$,
both transformations generate $SL(2,C)$ group.
It is argued that the transformation functions $\zeta^z(z)$ and $\zeta^{\bar{z}}(\bar{z})$ may be generalized to arbitrary (anti-) holomorphic functions, albeit singularities on $S^2$.\cite{Barnich}\cite{Strominger3} Then one obtains a direct product of two Virasoro groups. In this paper we will consider this extended symmetry group, and call it simply BMS$_4$ group.
Starting from (\ref{Bondi frame}), these transformations preserve the following asymptotic flatness condition.
\begin{eqnarray}
g_{uu} & \sim & -1+ {\cal O}(r^{-1}), \label{ASF1}\\
g_{ur} & \sim & -1+ {\cal O}(r^{-2}), \label{ASF2}\\
g_{AB} & \sim & r^2 \, h_{AB}+ {\cal O}(r^{1}), \label{ASF3}\\
g_{uA} & \sim & {\cal O}(1), \\
g_{rA} &= & 0, \\
g_{rr} &= & 0 \label{ASF6}
\end{eqnarray}
The function $\beta$ is the subleading term of (\ref{ASF2}): $g_{ur}=-1-\frac{\beta}{r^2}+\cdots$. Under BMS$_4$ group, $g_{ur}$ transforms as $\delta_{\xi} \, g_{ur}=\nabla_u \, \xi_r+\nabla_r \, \xi_u$. This gives a transformation rule of $\beta$. In the case of supertranslation,
\begin{equation}
\delta_f \, \beta= f \, \partial_u \, \beta+\frac{1}{4} \, C^{AB} \, D_AD_B \, f.
\label{delfBeta}
\end{equation}
For superrotation, we have
\begin{equation}
\delta_{\zeta} \, \beta= \zeta^z \, \partial_z \, \beta+D_z \zeta^z \, \beta+\frac{1}{2} \, D_z\zeta^z \, u\partial_u \, \beta+\frac{1}{8} \, u \, C^{zz} \, D_z^3\, \zeta^z. \label{delzetaBeta}
\end{equation}
Only for $\lambda=-\frac{1}{16}$, the gauge fixing condition (\ref{beta gauge}) is compatible with (\ref{delfBeta}) and (\ref{delzetaBeta}) and the transformation rules of $C_{AB}$, with are also derived via a definition $g_{zz} \equiv r \, C_{zz}+O(r^0)$ and $\delta g_{zz}=2\nabla_z \, \xi_z$,
\begin{eqnarray}
\delta_f \, C_{AB} &=& f \, \partial_u \, C_{AB}-2D_AD_B \, f +h_{AB} \, D_CD^C f, \label{delfC}\\
\delta_{\zeta} \, C_{zz} &=& \zeta^z \, D_z \, C_{zz}+2D_z \zeta^z \, C_{zz}+\frac{1}{2} \, D_z\zeta^z \, (u\partial_u-1) \, C_{zz}-u D_z^3 \zeta^z, \label{delzCzz} \\
\delta_{\zeta} \, C_{\bar{z}\bar{z}} &=& \zeta^z D_zC_{\bar{z}\bar{z}}+\frac{1}{2} \, D_z\zeta^z \, (u\partial_u-1) \, C_{\bar{z}\bar{z}}. \label{delzCzbzb}
\end{eqnarray}
Otherwise, gauge fixing (\ref{beta gauge}) with $\lambda \neq -1/16$ cannot be used so that $\beta$ must be left unfixed.
\section{Asymptotic symmetry in new coordinate system}
\hspace{5mm}
An asymptotically flat spacetime, which is described by coordinates $\{\tau,\rho,z,\bar{z}\}$ which are deformation of local coordinates (\ref{global}) of Minkowski space, also has asymptotic symmetry.
A coordinate transformation which connects the two metrics in both coordinates $\{r,u,z,\bar{z} \}$ and $\{\tau,\rho,z,\bar{z}\}$ is given by
\begin{eqnarray}
u &=& \tau \, e^{-\rho}, \label{tr1}\\
r &=& \tau \, \sinh \rho. \label{tr2}
\end{eqnarray}
The line element in coordinates $\{\tau, \rho, z,\bar{z}\}$ is obtained by substituting (\ref{tr1})-(\ref{tr2}) into (\ref{Bondi frame}). We will use a line element obtained from (\ref{Bondi frame}) by this transformation, instead of solving Einstein equation from scratch. Whether this new line element has BMS$_4$ symmetry or not must be checked once again independently, since the coordinate transformation changes the orders of terms in the variation of the line element, and does not enssure it. This will be discussed soon.
Half of future null infinity ${\cal I}^+$ ($u>0)$ is reached in the limit $r \rightarrow +\infty$ with $u$ fixed. This requires a fine-tuned limit in the $(\tau,\rho)$ coordinates.
\begin{equation}
\tau, \ \rho \rightarrow +\infty, \qquad
\tau \, e^{-\rho} =\, \text{fixed} \, (=u) \label{doublescaling}
\end{equation}
In this way the line element (\ref{Bondi frame}) will be rearranged into $1/r \sim e^{-2\rho}$ expansions, where coefficients are functions of $u$.
Line element in this frame obtained by the above coordinate transformation is given by
\begin{equation}
ds^2 = G_{\mu\nu} \, dx^{\mu}dx^{\nu},
\end{equation}
where $\mu, \nu=\tau, \rho, z,\bar{z}$ and
\begin{eqnarray}
G_{\tau\tau} &=& -1+\frac{4m}{\tau}\frac{e^{-3\rho}}{1-e^{-2\rho}}-\frac{4\beta}{\tau^2}\frac{e^{-2\rho}}{1-e^{-2\rho}}
+\cdots, \label{MetricG1}\\
G_{\tau \rho} &=& -\frac{4m}{1-e^{-2\rho}} e^{-3\rho}-\frac{4\beta e^{-4\rho}}{\tau(1-e^{-2\rho})^2}
+\cdots, \\
G_{\tau A} &=&
-e^{-\rho} \, \big[U_A+ \frac{2e^{-\rho}}{\tau \, (1-e^{-2\rho} )}\, W_A\big]+\cdots, \\
G_{\rho A} &=& \tau \,e^{-\rho} \, \big[U_A+\frac{2e^{-\rho}}{\tau \, (1-e^{-2\rho} )}\, W_A \big]+\cdots, \\
G_{\rho \rho} &=& \tau^2+\frac{4m\tau}{1-e^{-2\rho}} e^{-3\rho}+\frac{4\beta e^{-2\rho}}{(1-e^{-2\rho})^2}(1+e^{-2\rho})
+\cdots, \\
G_{AB} &=& \frac{1}{4} \, \tau^2 \, e^{2\rho}(1-e^{-2\rho})^2 \, h_{AB}+
\frac{1}{2} \tau \, e^{\rho} (1-e^{-2\rho}) \, C_{AB}+H_{AB}+\cdots \label{MetricG2}
\end{eqnarray}
Fields, $C_{AB}$, $U_A$, $W_A$, $\beta$ and $m$, have coordinate dependence like $C_{AB}(\tau \, e^{-\rho}, z,\bar{z})$.
Since the above metric (\ref{MetricG1})-(\ref{MetricG2}) is obtained from Bondi-frame metric by coordinate transformation, it also satisfies Einstein equation.
If $u=\tau \, e^{-\rho} (> 0)$ is kept fixed near large-$\rho$ region, the above metric takes the form of $1/e^{2\rho}$ expansion. Coefficient functions in the expansion of the metric tensor contain terms with negative powers of $u$.
Later, we will use this $1/e^{2\rho}$ expansion to derive a variation formula of surface charges $\delta \, Q$ at $\rho \rightarrow \infty$.
It is checked that when $\tau (\neq 0)$ is kept finite, terms which diverge as $\rho \rightarrow \infty$ cancel out. Since $u^{-1}= \tau^{-1} \, e^{\rho}$, this means that negative powers of $u$ do not appear in $\delta Q$. On the other hand, terms proportional to $u=\tau e^{-\rho}$ and $u^2=\tau^2 e^{-2\rho}$ may be created, by keeping terms which are higher orders in $e^{-\rho}$.
Under supertranslation, $\tau$ and $\rho$ change as
\begin{eqnarray}
\delta \, \tau &=& f \, \cosh \rho+e^{-\rho} \, D^zD_z f+ O(\tau^{-1} e^{-2\rho}), \label{taurho1}\\
\delta \, \rho &=& -f\tau^{-1} \sinh \rho+\tau^{-1}e^{-\rho} D^zD_z f + O(\tau^{-2} e^{-2\rho}). \label{taurho2}
\end{eqnarray}
Since $\delta \tau$ is large for finite $\tau$ and extremely large $\rho$, supertranslation is not a symmetry at ${\cal I}$ for finite $\tau$. However, when $\tau$ is taken to infinity at the same time, it becomes an asymptotic isometry.
The line element for Minkowski space in the new frame (\ref{global}) changes as follows.
\begin{equation}
\delta \, (ds_2^2) = -2\tau \, \sinh \rho \, D_z^2 f\, dz^2+2 \, e^{-\rho}(D_zf+D_z^2D^zf) \, dz \, (\tau \, d\rho-d\tau)+ \text{c.c} \label{delS0}
\end{equation}
In the case of Bondi frame there is only a single term, $-2r D_z^2f dz^2$, corresponding to the first term on the right above. Here an extra term (the second term) appears. These changes can be compensated by the changes of $C_{AB}$ and $U_{A}$. The above change (\ref{delS0}) vanishes for the four global translations $f=1, z/(1+z\bar{z}), \bar{z}/(1+z\bar{z}), (1-z\bar{z})/(1+z\bar{z})$, as it should.
Under superrotation, transformations of $\tau$ and $\rho$ are simply,
\begin{eqnarray}
\delta \, \tau &=& O(e^{-3\rho}), \label{taurho3}\\
\delta \, \rho &=& -\frac{1}{2} \, D_A \, \zeta^A+ O(\tau^{-1} e^{-3\rho}). \label{taurho4}
\end{eqnarray}
Hence superrotation is an asymptotic symmetry also in this frame.
In this case the change of the line element is
\begin{equation}
\delta \, (ds^2_2)= -\tau^2 \, e^{-\rho} \, \sinh \rho \, D_z^3\zeta^z \, dz^2+\text{c.c}.
\end{equation}
\section{Surface charge algebra of BMS$_4$}
\hspace{5mm}
The surface charges (generators) of BMS$_4$ group in four dimensions and their algebra have previously been studied via covariant approach in \cite{Barnich}\cite{BT2}. In this section, this task will be carried out by using the Hamiltonian framework\cite{Regge}\cite{BrownHenneaux}\cite{Kubota}. This is based on Hamiltonian (ADM) formulation of gravity\cite{ADM}. We choose $\tau$ as a time variable and a foliation of the spacetime with $\tau >0$ into constant-time space-like slices $\Sigma_{\tau}$. $\Sigma_{\tau}$ has a time-like normal vector $n_{\mu}=-\delta^{\tau}_{\mu}+{\cal O}(e^{-2\rho})$.
Each slice $\Sigma_{\tau}$ is a 3D Euclidean Anti-de Sitter space (EAdS$_3$) and has boundary at $\rho=+\infty$. $\rho=0$ is the center of this space. When taking the asymptotic limit $\rho \rightarrow +\infty$, we need to introduce a large cutoff $\rho=\rho_{\infty}$ in order to regularize divergences, and we obtain a cylinder with a radius $\rho_{\infty}$ and a height in the direction of increasing $\tau$. The time-like boundary at $\rho=\rho_{\infty}$ will be denoted as $\Sigma_{\rho_{\infty}}$.
In order to reach a point with positive value of $u$ on ${\cal I}^+$, $\tau$ must also be increased according to (\ref{doublescaling}), as the cutoff is removed, $\rho_{\infty} \rightarrow +\infty$. In this case one is forced to approach ${\cal I}^+$ along a null geodesic $u=$ const.\footnote{If we do not shift to hypersurfaces $\Sigma_{\tau}$ with increasing $\tau$ appropriately, then we will end up with the point $u=0$ of ${\cal I}^+$.} This, however, does not imply that canonical formalism on a null hypersurface is considered. Canonical commutation relations are imposed on a spacelike hypersurface $\Sigma_{\tau}$, and $\tau, \rho \rightarrow \infty$ limit will be taken afterwords.
\subsection{Hamiltonian}
\hspace{.5cm}
In ADM decomposition\cite{ADM}, the spacetime metric is arranged into the form
\begin{equation}
ds^2= G_{\mu\nu} \, dx^{\mu} \, dx^{\nu}=-(N^2-N_a N^a) \, d\tau^2
+\gamma_{ab} \, dx^a \, dx^b+2N_a \, dx^a \,d\tau, \label{ADM}
\end{equation}
where $a$ runs over $\rho, z,\bar{z}$, and $N^a=\gamma^{ab} \, N_b$ with $\gamma^{ab}$ being the inverse of $\gamma_{ab}=G_{ab}$.
In the case of the present solution, the lapse and shift functions $N$, $N_a$ depend on the fields.
\begin{eqnarray}
N_{\rho} &=& -4m e^{-3\rho}-(4m e^{-5\rho}+\frac{4\beta}{\tau} e^{-4\rho})+\cdots, \\
N_{A} &=& -e^{-\rho}U_A-\frac{2}{\tau} W_A (e^{-2\rho}+e^{-4\rho})+\cdots, \\
N &=& 1+(-\frac{2m}{\tau} e^{-3\rho}+\frac{2\beta}{\tau^2} e^{-2\rho})+\cdots \label{shiftA}
\end{eqnarray}
Action integral $S$ is given by
\begin{equation}
S= \frac{1}{16\pi G} \, \int d^4x \, \sqrt{-G} \, R.
\end{equation}
The momentum conjugate to $\gamma_{ab}$ is given by
\begin{equation}
\Pi^{ab}= \frac{1}{16\pi G} \, \sqrt{\gamma} \, (K^{ab}-\gamma^{ab} \, K),
\end{equation}
where $K^{ab}$ is extrinsic curvature of $\Sigma_{\tau}$: $K_{ab}= (1/2N)(\partial_{\tau} \gamma_{ab}-N_{a|b}-N_{b|a})$.
Symbol \lq $|a$' stands for a covariant derivative with respect to $\gamma_{ab}$.
On a spacelike surface $\Sigma_{\tau}$, canonical commutation relations are imposed on $\gamma_{ab}$ and $\Pi^{ab}$.
The Hamiltonian is given by
\begin{eqnarray}
H[N,N^a]&=& \int_{\Sigma_{\tau}} d\rho d^2z \, \Big[16\, \pi \, G\gamma^{-1/2} \, N \, \big(\Pi^{ab}\Pi_{ab}-\frac{1}{2} \, \Pi^2\big)- \frac{1}{16\pi G} \, \sqrt{\gamma} \, N ~^{(3)}R\Big] \nonumber \\
&& - \int_{\Sigma_{\tau}} d\rho d^2z \sqrt{\gamma} \, 2\, N_a \, \Big(\frac{1}{\sqrt{\gamma}}\, {\Pi^{ab}}\Big)_{|b}.
\end{eqnarray}
$^{(3)}R$ is the Ricci scalar constructed from $\gamma_{ab}$, and $\Pi= \gamma_{ab} \, \Pi^{ab}$.
The generator ${\cal H}[\xi]$ of the asymptotic symmetry is obtained by replacing $N$ and $N^a$ in $H[N,N^a]$, by $\bar{\xi}^{\tau}$ and
$\bar{\xi}^a$ as
\begin{eqnarray}
N & \rightarrow & \bar{\xi}^{\tau} \equiv N \, \xi^{\tau}, \nonumber \\
N^a & \rightarrow & \bar{\xi}^a \equiv N^a \, \xi^{\tau}+ \xi^a,
\end{eqnarray}
where $\xi^{\mu}=(\xi^{\tau}, \xi^a)$ are equal to $(\delta \tau, \delta \rho, \delta z, \delta \bar{z})$ in (\ref{z1})-(\ref{zb1}), (\ref{z2})-(\ref{zb2}) and (\ref{taurho1})-(\ref{taurho4}) which correspond to supertranslation $\xi^{\mu}=(\delta r, \delta u, \delta z, \delta \bar{z})$, and superrotation re-expressed in $\{\tau,\rho, z,\bar{z}\}$ coordinates.
\subsection{Variation equation for surface charges}
\hspace{.5cm}
Transformations of the canonical variables are expressed in terms of Poisson brackets: $\delta_{\xi} \gamma_{ab}= \{ \gamma_{ab}, {\cal H}[\xi]\}$, $\delta_{\xi} \, \Pi^{ab}= \{ \Pi^{ab}, {\cal H}[\xi]\}$. To compute the Poisson brackets, partial integration is in general necessary, and to make the Poisson bracket well-defined, an appropriate surface term $Q[\xi]$ must be added to ${\cal H}[\xi]$. Its variation must satisfy the following variation constraint.
\begin{multline}
\delta \, Q[\xi] =
\int_{S} d^2z \, \Big[ 2\, \bar{\xi}^a \, \Pi^{b\rho} \, \delta \gamma_{ab}+2\bar{\xi}^a \, \delta \, \Pi^{b\rho} \, \gamma_{ab} -\bar{\xi}^{\rho} \, \Pi^{ab} \, \delta \, \gamma_{ab}\Big] \\
+\frac{1}{16\pi G} \, \int_S d^2z \, \sqrt{\gamma} \, S^{abc\rho} \, \Big[\ \bar{\xi}^{\tau} \, (\delta \, \gamma_{ab})_{|c}-\partial_c \, \bar{\xi}^{\tau} \, \delta \, \gamma_{ab}\Big] \label{delQ}
\end{multline}
Here $S$ is 2-dim sphere at fixed $u$ and $r \rightarrow \infty$.
$S^{abcd}$ is defined by
\begin{equation}
S^{abcd}= \frac{1}{2} \, (\gamma^{ac} \, \gamma^{bd}+\gamma^{ad} \, \gamma^{bc}-2 \, \gamma^{ab} \, \gamma^{cd}).
\end{equation}
Hamilton's eqs are given as follows.
\begin{eqnarray}
\delta_{\xi} \, \gamma_{ab} &=& \bar{\xi}_{a|b}+\bar{\xi}_{b|a}+\frac{32\pi G}{\sqrt{\gamma}} \, \bar{\xi}^{\tau}\, (\Pi_{ab}-\frac{1}{2} \, \gamma_{ab} \, \Pi), \\
\delta_{\xi} \, \Pi^{ab} &=& -\frac{32\pi G}{\sqrt{\gamma}} \, \bar{\xi}^{\tau} \, \big(\Pi^{ac} \, {\Pi_c}^b-\frac{1}{2} \, \Pi \, \Pi^{ab}\big)
+\frac{8\pi G}{\sqrt{\gamma}} \, \bar{\xi}^{\tau}\big(\Pi^{cd}\Pi_{cd}-\frac{1}{2}\Pi^2\big) \, \gamma^{ab} \nonumber \\
&&-\frac{1}{16\pi G} \, \sqrt{\gamma} \, \bar{\xi}^{\tau} \,
\big(^{(3)} R^{ab}-\frac{1}{2}\gamma^{ab} \, ^{(3)} R\big)+\frac{1}{16\pi G}\sqrt{\gamma}\big(\bar{\xi}^{\tau |ab}-\gamma^{ab}{\bar{\xi}^{\tau |c}} \, _c\big)
\nonumber \\
&& +\sqrt{\gamma} \, \Big(\frac{1}{\sqrt{\gamma}} \, \Pi^{ab} \, \bar{\xi}^c \Big)_{|c}-\Pi^{ac} \, {N^b}_{|c}-\Pi^{bc} \, {N^a}_{|c}
\end{eqnarray}
In the case of AdS$_3$, the Hamiltonian generator ${\cal H}[\xi]$ generates a Poisson bracket algebra:
\begin{equation}
\ \{{\cal H}[\xi], {\cal H}[\xi']\} ={\cal H}[[\xi,\xi']]+K[\xi,\xi'],
\end{equation}
where $K[\xi,\xi']$ is a possible central extension. A strategy to evaluate this central extension is to replace Poisson bracket by Dirac bracket. Then the Hamiltonian and momentum constraints hold strongly, and the generator ${\cal H}[\xi]$ can be replaced by the surface charge $Q[\xi]$:
\begin{equation}
\ \{Q[\xi], Q[\xi']\}_D =Q[[\xi,\xi']]+K[\xi,\xi'].
\end{equation}
The left hand side of this equation may be evaluated as $\delta_{\xi'} \, Q[\xi]$, where $\delta_{\xi'}$ is a transformation associated with an asymptotic Killing vector $\xi'$. For AdS background, $Q[[\xi,\xi']]$ on the right hand side vanishes. Hence the central charge is evaluated as
\begin{equation}
K[\xi,\xi'] = \delta_{\xi'} \, Q[\xi]\big|_{\text{AdS}}. \label{defK}
\end{equation}
Here $\delta_{\xi'}$ stands for a variation associated with transformation $\xi'$.
\subsection{Charge algebra}
\hspace{.5cm}
Same prescription is used in the case of a four-dimensional asymptotically flat space on ${\cal I}$. The surface charges are denoted as $Q[f,\zeta^z, \bar{\zeta}^{\bar{z}}]$, where $f(z,\bar{z})$ is a parameter function for supertranslation and $\zeta^z(z)$, $\bar{\zeta}^{\bar{z}}(\bar{z})$ those for (extended) superrotations. The brackets of surface charges are given by
\begin{equation}
\ \Big\{Q[\xi], Q[\xi']\Big\}_D = \delta_{\xi'} \, Q[\xi] \label{QQdelQ}
\end{equation}
for $\xi=(f,\zeta, \bar{\zeta})$ and $\xi'=(f',\zeta', \bar{\zeta}')$.
Calculation is done by using Mathematica and only the results are now presented.\footnote{Here (non-)integrability of eq (\ref{delQ}) is not yet taken into account.}\footnote{In eq (\ref{tr}), the last equality is checked by actual calculation.}
\begin{multline}
\ \Big\{ Q[f_1,0,0], Q[f_2,0,0] \Big\}_D \equiv \delta_{f_2} \, Q[f_1,0,0] \\
= \frac{1}{16\pi G} \, \int_S d^2z \, h_{z\bar{z}} \,
\dot{C}^{AB} \, (f_2 \,D_AD_B \, f_1-f_1 \, D_AD_B \, f_2),
\label{tt}
\end{multline}
\begin{multline}
\ \Big\{ Q[0,\eta,0], Q[f,0, 0] \Big\}_D \equiv \delta_{f} \, Q[0,\eta,0] \\
= \frac{1}{16\pi G} \, \int_S d^2z \, h_{z\bar{z}} \, [\eta,f] \, \Big[4 \, m+\partial_u \, {H_A}^A -\frac{1}{2} \partial_u \, (C_{AB} \,
C^{AB}) \Big] \\
+\frac{1}{16\pi G} \, \int_S d^2z \, h_{z\bar{z}} \, \frac{1}{2} \Big[f \, D_z^3 \, \eta^z \, (u\partial_u-1)C^{zz}- u \, D_z\eta^z \, D_AD_B \, f \, \dot{C}^{AB}\Big] \\
+\frac{1}{16\pi G} \, \int_S d^2z \, h_{z\bar{z}} \, \Big[-\frac{3}{4}fD_z\eta^z \, \dot{ C}^{zz} \, C_{zz} -\frac{1}{2} \, f\eta^z \dot{C}^{zz} \, D_z C_{zz}\\+\frac{1}{4}fD_z\eta^z \, \dot{C}_{zz} \, C^{zz}-\frac{1}{2} \, f \eta^z \, \dot{C}_{zz} D_zC^{zz}\Big], \label{rt}
\end{multline}
\begin{equation}
\ \Big\{ Q[f,0,0], Q[0, \chi,0] \Big\}_D \equiv \delta_{\chi} \, Q[f,0,0] =-\delta_f \, Q[0,\chi,0], \label{tr}
\end{equation}
\begin{multline}
\ \Big\{Q[0,\eta,0], Q[0,\chi,0]\Big\}_D \equiv \delta_{\chi} \, Q[0,\eta,0] \\
= \frac{1}{16\pi G} \, \int_S d^2 z h_{z\bar{z}} \, [\eta,\chi]^z \, \Big[ -3 W_z+C_{zz} \, U^z+\frac{3}{4} \, D_z \, {H_A}^A-\frac{5}{16} \, D_z(C_{AB} C^{AB}) \\
-2u \, \partial_z \, m -\frac{1}{2} \, u \, D_z \, {\dot{H}_A}^A \\
+\frac{1}{4} \, u \, (2\, C^{zz} \, D_z \, \dot{C}_{zz} +\dot{C}_{zz} \, D_z \, C^{zz}+2\, C_{zz} \, D_z \, \dot{C}^{zz} +\dot{C}^{zz} \, D_z C_{zz}) \Big]\\
-\frac{1}{64\pi G} \, \int_S d^2 z h_{z\bar{z}} \, (D_z\eta^z \, D_z^3\chi^z-D_z\chi^z \, D_z^3 \, \eta^z) \, u \, (u\partial_u-1) \, C^{zz}
, \label{rr}
\end{multline}
\begin{multline}
\ \Big\{ Q[0,0,\bar{\chi}], Q[0,\eta,0]\Big\}_D \equiv \delta_{\eta} \, Q[0,0,\bar{\chi}] \\
=\frac{1}{16 \pi G} \, \int_S d^2z h_{z\bar{z}} \, \Big[\frac{1}{4} u\, D_z^3\eta^z \, D_{\bar{z}} \chi^{\bar{z}} \, (1-u\partial_u) \, C^{zz}-\frac{1}{4} u\,
D_{\bar{z}}^3\chi^{\bar{z}} \, D_{z} \eta^z\, (1-u\partial_u) \, C^{\bar{z}\bar{z}}\\
+\frac{1}{2}u \, D_z\eta^z D_{\bar{z}}\chi^{\bar{z}} \, (C_{zz}\, \dot{C}^{zz}-\dot{C}_{zz}\, C^{zz})
+\frac{1}{4} u\, \eta^z \, D_{\bar{z}} \chi^{\bar{z}}(\dot{C}_{zz}\, D_zC^{zz}+ \dot{C}^{zz}\, D_zC_{zz}) \\
- \frac{1}{4} u\, \chi^{\bar{z}} \, D_{z} \eta^{z}(\dot{C}_{zz}\, D_{\bar{z}}C^{zz}+ \dot{C}^{zz}\, D_{\bar{z}}C_{zz}) \Big].
\label{rbr}
\end{multline}
Here, $[\eta,f] \equiv -[f,\eta] \equiv \eta^z \, \partial_z \, f-\frac{1}{2} \, f \, D_z \eta^z$ and $[\eta,\chi]^z \equiv \eta^z \, \partial_z \, \chi^z-\chi^z \, \partial_z \, \eta^z$. There are also similar brackets involving $Q[0,0,\bar{\chi}]$ instead of $Q[0,\eta,0]$, which are not displayed here.
In the above calculation, there appeared terms proportional to $e^{\rho} \, \tau$, which are divergent as $\rho \rightarrow +\infty$. These can be dropped after partial integration over S$^2$, and the charges are finite. There also exist terms proportional to $\tau^{-2}$. However, those terms vanish in the $\tau \rightarrow +\infty$ limit.
In the calculation of $\{Q[0,\eta,0], Q[0,\chi,0]\}_D$ there is also a term
proportional to $h_{z\bar{z}} \, \chi^z(z) \, \eta^z(z) \, H_{zz}(z)$ in the integrand.
This is formally dropped by using a formula $D_{\bar{z}} \, D_z \, \eta^z(z)=-h_{z\bar{z}} \, \eta^z(z)$ and a partial integration with respect to the derivative $D_{\bar{z}}$. \footnote{Rigorously speaking, we must be careful in this procedure, because the surface terms for $z, \bar{z}$ integration cannot be dropped for polynomials $\eta^z$ and $\chi^z$. This issue needs to be further studied. This term might as well be dropped by assuming $H_{zz}(z)=0$. }
The above result shows that the variation of surface charge $\delta_{\xi} \, Q[\xi']$ satisfies a condition of anti-symmetry.
\begin{equation}
\delta_{\xi} \, Q[\xi'] =-\delta_{\xi'} \, Q[\xi] \label{antisym}
\end{equation}
This is a necessary condition for prescription (\ref{QQdelQ}) to work.
\subsection{Non-integrability of variation equation for surface charges}
\hspace{5mm}
Surface charges are defined by variation equation (\ref{delQ}). Integrability of this equation must be checked.
For arbitrary variations $\delta$ of fields, eq (\ref{delQ}) for supertranslation reads
\begin{equation}
\delta Q[f,0,0] = \frac{1}{16\pi G} \, \int_S d^2z \, h_{z\bar{z}} \, f \, \big[4 \delta M+\frac{1}{2} \, \dot{C}_{AB} \, \delta \, C^{AB}\big], \label{delQf}
\end{equation}
and that for superrotation reads\footnote{On the right hand side of (\ref{delQeta}) there also exists a divergent term in the integrand:\\ $-e^{2\rho} \,u \, h_{z\bar{z}} \, \eta^z(z) \, \delta \, U_z$. This term, however, drops out after partial integration. }
\begin{multline}
\delta Q[0,\eta,0] =\frac{1}{16\pi G} \, \int_S d^2z \, h_{z\bar{z}} \, \eta^z \, \Big[\delta (-2L_z-2uD_z M)-\frac{1}{4} \, u \, D_z(\dot{C}_{AB} \, \delta \, C^{AB})\Big]. \label{delQeta}
\end{multline}
These equations for variations of charges coincide with those obtained via covariant framework\cite{BT2}.
Here $M$ is a gauge invariant mass,\footnote{Definition $\delta_f \, g_{uu} \rightarrow 2\delta_fm/r$ yields $\delta_f m=f\dot{m}-\dot{U}^AD_Af+(1/4) \dot{C}^{AB}D_AD_B f$. This is true only when $\beta=-(1/16)C_{AB}C^{AB}$. For other choice of $\beta$, it is $M$, not $m$, that transforms as (\ref{delfM}). Similar statement is valid for $\delta_{\zeta} m$ and $\delta_{\zeta} \, M$.}
{\it i.e.}, mass which does not depend on the choice of $\beta$,
\begin{equation}
M \equiv m+\frac{1}{4} \partial_u {H_A}^A-\frac{1}{8} \, \partial_u (C_{AB}C^{AB}).
\end{equation}
This transforms under supertranslation as
\begin{eqnarray}
\delta_f \, M &=& -\frac{1}{8} \, f \, \partial_u C_{AB} \partial_u C^{AB}+\frac{1}{4} \, D^AD^B (f \partial_u C_{AB}) \nonumber \\
&=& f \, \partial_u M+\frac{1}{4} \, D^AD^B(f \, \partial_u \, C_{AB})-\frac{1}{4} \, f \, D^AD^B \, \partial_u \, C_{AB}, \label{delfM}
\end{eqnarray}
and under superrotation as
\begin{multline}
\delta_{\zeta} M= \zeta^z \, \partial_z \, M+\frac{3}{2} \, D_z\zeta^z \, M+\frac{1}{2} \, D_z\zeta^z \, u\partial_u M-\frac{1}{2} \, D_z^2\zeta^z \, u\partial_uU^z \\
+\frac{1}{2}\zeta^z u\partial_uU_z
+\frac{1}{8}\, D_z^3\zeta^z \, (u\partial_u+1) C^{zz}. \label{delzetaM}
\end{multline}
Similarly, angular momentum $L_A$ is defined by\footnote{By comparing transformation rules, it is found that $L_z$ is related to $N_z$ defined in \cite{BT2} by $L_z=-N_z-\frac{1}{32} D_z (C_{AB} C^{AB})$.}
\begin{equation}
L_A \equiv \frac{3}{2}W_A-\frac{1}{2} \, C_{AB}U^B-\frac{3}{8} \, D_A {H_A}^A+\frac{5}{32} \, D_A (C_{BC}C^{BC}) \label{LA}
\end{equation}
and transforms under supertranslation as
\begin{multline}
\delta_f \, L_z= -3M \, \partial_z f+f \, \dot{L}_z -\frac{1}{4} \, D_z^3f \, C^{zz}-\frac{3}{4} \, D_z \, (D_zf \, D_z \, C^{zz})\\-\frac{3}{4} \, D_z(D^z)^2f \, C_{zz}
+\frac{3}{4} \, D_zf \, (D^z)^2 \, C_{zz}-\frac{1}{4}(D^z)^2f D_zC_{zz}\\+D_zf \, (-\frac{3}{8} \, C_{zz}\dot{C}^{zz}+\frac{1}{8}\dot{C}_{zz}C^{zz}).
\end{multline}
Its transformation rule under superrotation is given by
\begin{eqnarray}
\delta_{\zeta} \, L_{z} &=& \frac{1}{2} \, u \, D_z\zeta^z \, \dot{L}_z+\zeta^z \, D_zL_z+2D_z\zeta^z \, L_z-\frac{3}{2} \, M \, D_z^2\zeta^z \nonumber \\&&-\frac{1}{2} \, \zeta^z \, H_{zz}-\frac{1}{8} \, u \, D_z^4\zeta^z \, C^{zz}+\frac{3}{8} \, u \, D_z^2\zeta^z \, (D^z)^2\, C_{zz}
-\frac{3}{8}u\, D_z^3 \,\zeta^z \, D_zC^{zz}\nonumber \\&&-\frac{3}{8} u D_z^2 \zeta^z D_z^2C^{zz}-\frac{3}{16}uD_z^2\zeta^zC_{zz}\dot{C}^{zz}+\frac{1}{16}uD_z^2\zeta^z\dot{C}_{zz}C^{zz}, \\
\delta_{\bar{\zeta}} \, L_z&=& \frac{1}{2} u\, D_{\bar{z}} \, \zeta^{\bar{z}} \, \dot{L}_z+\zeta^{\bar{z}} \, D_{\bar{z}}L_{z}-\frac{3}{2} \, D_{\bar{z}} \, \zeta^{\bar{z}} \, L_z-\frac{3}{2} \, h_{z\bar{z}} \, \zeta^{\bar{z}} \, M \nonumber \\ &&
+h_{z\bar{z}} \, \Big[\frac{1}{2} \, ( h^{z\bar{z}})^2 \, H_{zz}D_{\bar{z}}^2\zeta^{\bar{z}}-\frac{3}{8} \, u \, \zeta^{\bar{z}} \, (D^z)^2C_{zz} -\frac{1}{16}u\, \zeta^{\bar{z}} \, C^{zz}\dot{C}_{zz}\nonumber \\
&& -\frac{1}{8}u\, D_{\bar{z}}^3\zeta^{\bar{z}} \, D^{\bar{z}} C^{\bar{z}\bar{z}}+\frac{3}{8}u \, \zeta^{\bar{z}} \, D_z^2 C^{zz}+\frac{3}{16}u \, \zeta^{\bar{z}} \, C_{zz}\dot{C}^{zz} \Big]. \label{delzbLz}
\end{eqnarray}
Due to the last terms in (\ref{delQf})-(\ref{delQeta}),
the variation equations are in general not integrable.
One prescription to ensure integrability will be to require $\dot{C}_{AB}=0$. Then the surface charges become independent of $u$, and conserved. In the next subsection this prescription will be studied.
Another prescription will be to add to (\ref{delQf})-(\ref{delQeta})
an additional term such as
\begin{equation}
\delta \, Q^{\text{additional}}[\xi]=\frac{-1}{32\pi G} \, \int_S d^2 z \, h_{z\bar{z}} \, \dot{C}^{AB} \, \delta \, C_{AB} \, \xi^{u} \label{extra}
\end{equation}
to drop the second terms in (\ref{delQf}) and (\ref{delQeta}) as was proposed in \cite{Wald}. By integrating the new variation equation $\delta \hat{Q}[\xi]= \delta Q[\xi]+\delta Q^{\text{additional}}[\xi]$, {\em i.e.},
\begin{eqnarray}
\delta \hat{Q}[f,0,0]&=& \frac{1}{16\pi G} \, \int_S d^2z \, h_{z\bar{z}} \,4 \, f \, \delta M, \label{delhatQf}\\
\delta \hat{Q}[0,\eta,0]&=&\frac{1}{16\pi G} \, \int_S d^2z \, h_{z\bar{z}} \, \eta^z \, \delta (-2L_z-2uD_z M), \label{delhatQeta}
\end{eqnarray}
the surface charges $\hat{Q}[f,0,0]$ and $\delta Q[0,\eta,0]$ are obtained,
\begin{equation}
\hat{Q} [0,0,f] \equiv \frac{1}{16\pi G} \, \int d^2z \, h_{z\bar{z}}\, 4\, f \, M, \label{hQf}
\end{equation}
\begin{equation}
\hat{Q}[0,\eta,0] \equiv \frac{1}{16\pi G} \, \int d^2 z h_{z\bar{z}} \, \eta^z \, \Big[ -2 \, L_z-2u\partial_z M
\Big]. \label{hQeta}
\end{equation}
These charges coincide with the \lq integrable part of the surface charge' proposed in \cite{BT2}.\footnote{In eq (3.5) of \cite{BT2}, in order to define the algebra, the right hand side of (\ref{QQdelQ}) is modified by decomposing the variation of surface charge $\delta \, Q[\xi]$ into an integrable part $\delta \, Q^{\text{int}}[\xi]$ and a non-integrable part $\delta Q^{\text{nonint}}[\xi]$, and replacing the non-integrable part $\delta_{\xi'}\, Q^{\text{nonint}}[\xi]$ by $\delta_{\xi}\, Q^{\text{non}}[\xi']$. The left hand side is interpreted as $\{Q^{\text{int}}[\xi], Q^{\text{int}}[\xi']\}_D$.}
The algebra of these surface charges must be computed by using (\ref{QQdelQ}) with $Q[\xi]$ replaced by $\hat{Q}[\xi]$ and via the transformation rules (\ref{delfM})-(\ref{delzbLz}).
In this case, however, it can be shown that the anti-symmetry property (\ref{antisym}) is not respected.
For instance, by using (\ref{delfM}) for the supertranslation of $M$, the following Dirac bracket is obtained.
\begin{eqnarray}
&&\{\hat{Q}[f,0,0], \hat{Q}[f',0,0]\}_D \equiv
\delta_{f'} \, \hat{Q}[f,0,0] \nonumber \\
&&= \frac{1}{16\pi G} \, \int_S \, d^2z \, h_{z\bar{z}} \, \{ 4ff' \, \dot{M}-2 \, (D^A \, f)(D^B \, f') \, \dot{C}_{AB}-(f \, D^AD^B \, f') \, \dot{C}_{AB}\} \nonumber \\
&&\neq -\{\hat{Q}[f',0,0], \hat{Q}[f,0,0]\}_D
\end{eqnarray}
Hence, clearly, this prescription to add an additional term to the variation equation does not work for obtaining charges which generate algebra.
\subsection{Conserved surface charges and algebra}
\hspace{.5cm}
In this subsection, condition $\dot{C}_{AB}=0$ will be imposed on field $C_{AB}$.\footnote{ It is discussed in
\cite{BT2} that, with the modification of the algebra as described in footnote 11, the standard bms$_4$ charges are conserved in the absence of news and they generate a centerless algebra. In \cite{BTcurrent} a current algebra for spatial components in the absence of news is obtained. } Then the variation equations (\ref{delQf})-(\ref{delQeta}) become integrable.The surface charge of supertranslation, $Q[f,0,0]=Q^{\text{transl}} [f]$, and that of superrotation, $Q[0,\eta,0]=Q^{\text{rot}}[\eta]$ read
\begin{equation}
Q^{\text{transl}} [f] \equiv \frac{1}{16\pi G} \, \int d^2z \, h_{z\bar{z}}\, 4\, f \, M \Big|_{\dot{C}_{AB}=0}, \label{Qtra}
\end{equation}
\begin{equation}
Q^{\text{rot}}[\eta] \equiv \frac{1}{16\pi G} \, \int d^2 z h_{z\bar{z}} \, \eta^z \, \Big[ -2 \, L_z-2u\partial_z M
\Big]\Big|_{\dot{C}_{AB}=0}, \label{Qrot}
\end{equation}
where condition $\dot{C}_{AB}=0$ is imposed. Except for this constraint these charges are the same as (\ref{hQf}) and (\ref{hQeta}).
These charges are conserved, if $D_z^3\, \eta^z=0$. This is checked by using (\ref{dm}), (\ref{W}), (\ref{LA}).
By using definitions (\ref{Qtra})-(\ref{Qrot}) and results (\ref{tt})-(\ref{rbr}) and by setting $\dot{C}_{AB}=0$, the algebra of surface charges is expressed as follows. $\bar{Q}^{\text{rot}}[\bar{\eta}]$ is a complex conjugate of (\ref{Qrot}).
\begin{eqnarray}
\Big\{ Q^{\text{transl}}[f_1], Q^{\text{transl}}[f_2] \Big\}_D = 0,
\label{tt2}
\end{eqnarray}
\begin{equation}
\ \Big\{ Q^{\text{rot}}[\eta], Q^{\text{transl}}[f] \Big\}_D
= Q^{\text{transl}} [[\eta,f]]
-\frac{1}{16\pi G} \, \int_S d^2z \, h_{z\bar{z}} \, \frac{1}{2} \, f \, D_z^3 \, \eta^z \, C^{zz}
,\label{rt2}
\end{equation}
\begin{multline}
\Big\{Q^{\text{rot}}[\eta], Q^{\text{rot}}[\chi]\Big\}
= Q^{\text{rot}}[ [\eta,\chi] ] \\
+\frac{1}{64\pi G} \, \int_S d^2 z h_{z\bar{z}} \, (D_z\eta^z \, D_z^3\chi^z-D_z\chi^z \, D_z^3 \, \eta^z) \, u \, C^{zz}, \label{rr2}
\end{multline}
\begin{multline}
\ \Big\{ \bar{Q}^{\text{rot}}[\bar{\chi}], Q^{\text{rot}}[\eta]\Big\}_D
=\frac{1}{16 \pi G} \, \int_S d^2z h_{z\bar{z}} \, \Big[\frac{1}{4} u\, D_z^3\eta^z \, D_{\bar{z}} \chi^{\bar{z}} \, C^{zz}-\frac{1}{4} u\,
D_{\bar{z}}^3\chi^{\bar{z}} \, D_{z} \eta^z \, C^{\bar{z}\bar{z}}
\Big].
\label{rbr2}
\end{multline}
Let us note that there are central extensions which depend on field $C_{AB}$. However, the condition $\partial_u \, C_{AB}=0$ contradicts with the transformation rule for $C_{AB}$ (\ref{delzCzz}), if $D_z^3 \zeta \neq 0$. Hence BMS$_4$ algebra is not extended and the above central extensions all vanish.
\begin{equation}
\ \{Q[\xi], Q[\xi']\}_D =Q[[\xi,\xi']] \label{simple}
\end{equation}
The above algebra satisfies Jacobi identity:
$ \{Q[\xi_1], \{Q[\xi_2], Q[\xi_3]\}_D\}_D+\{Q[\xi_2], \{Q[\xi_3], Q[\xi_1]\}_D\}_D
+\{Q[\xi_3], \{Q[\xi_1], Q[\xi_2]\}_D\}_D=0$.
This is checked via $
\{Q[\xi_1], \{Q[\xi_2], Q[\xi_3]\}_D\}_D=- \delta_{\xi_1} \, \{Q[\xi_2], Q[\xi_3]\}_D$,
and the right hand side is evaluated by substituting variation of fields into the integrand of $\{Q[\xi_2], Q[\xi_3]\}_D$.
\section{Summary}
\hspace{5mm}
In this paper surface charges for supertranslation $Q^{\text{transl}}$ (\ref{Qtra}) and superrotations $Q^{\text{rot}}$ ($\bar{Q}^{\text{rot}}$) (\ref{Qrot}) on ${\cal I}$ and the algebra of these charges are studied via Hamiltonian framework.
If a field $C_{AB}$ is independent of $u$, these surface charges generate algebra (\ref{simple}) and the central extensions of the algebra vanish.
It would be desirable, if appropriate modification of surface charges $Q[\xi]$ could be found which generate BMS$_4$ algebra also for the case $\dot{C}_{AB} \neq 0$ so that the right hand side of the algebra would coincide with the changes of the surface charges $\delta_{\xi'} \, Q[\xi]$ under transformation.
We obtained surface charges on the $u \geq 0$ part of ${\cal I}^+$, and their Dirac bracket algebra (\ref{tt2})-(\ref{rbr2}). The charges are expressed in terms of radiation fields $C_{AB}$ in Bondi frame (\ref{Bondi frame}). In Bondi frame, however, there is no horizen at $u=0$ and this is not a special point on ${\cal I}^+$. Hence it is expected that the above bracket algebra also holds on the $u <0$ part of ${\cal I}^+$. On the other hand, since the space-like region ($r > t$) is foliated by time-like hypersurface, at present it is not possible to check this directly. This issue needs further investigation.
\newpage
\setcounter{section}{0}
\renewcommand{\thesection}{\Alph{section}}
|
\section{Introduction}
Before the discovery of a non-vanishing reactor angle, discrete symmetries were deeply implemented in the construction of flavour models to explain the flavour puzzle. In particular, it was a common feature of this class of models the prediction in first approximation of the PMNS matrix with a vanishing reactor angle and a maximal atmospheric one (e.g. Ref.~\cite{Altarelli:2010gt}).
With a sizable reactor angle~\cite{GonzalezGarcia:2012sz}, these models underwent to a severe loss of attractiveness. To achieve a model in agreement with the new data, a few strategies have been followed~\cite{Luhntalk}: introduction of additional parameters in preexisting minimal models; implementation of features that allow next order corrections only in specific directions in the flavour space; search for alternative mixing patterns or flavour symmetries that lead already in first approximation to $\theta_{13}\neq 0^\circ$ (see for example Refs.~\cite{Altarelli:2009gn} and references therein). In summary, the latest neutrino data can still be described in the context of discrete symmetries, but at the prize of fine-tunings and/or eccentric mechanisms.
Sum rules among neutrino masses and mixing angles are usually present in these models and are useful as tests at experiments~\cite{Ballett:2013wya}. Furthermore, studies on flavour violating observables~\cite{Feruglio:2008ht}, on the connection with astroparticle physics~\cite{Bertuzzo:2009im}, on the parameter running~\cite{Dighe:2006sr} and on the role of the CP symmetry~\cite{Feruglio:2012cw} have been performed to fully workout this framework. On the other side, the scalar and messenger sectors of these models are in general very complicated~\cite{Altarelli:2005yp}, it is not easy to provide a successful description of the quark sector~\cite{Feruglio:2007uu}, and the selection of a specific discrete symmetry usually does not follow from a more general criterium~\cite{Altarelli:2006kg}, but it is just a matter of taste.
Even if it is still worth to search for a realistic model based on discrete symmetries, the many drawbacks suggest to investigate alternative approaches: here the focus will be on continuous symmetries such as the simplest Abelian $U(1)$ or non-Abelian groups.
\boldmath
\section{Abelian models}
\unboldmath
\label{u1mod}
Models based on the Abelian $U(1)$ group are sometimes preferred with respect to those based on discrete symmetries for a series of reasons. First of all, the Abelian $U(1)$ group is an element already present in the Standard Model (SM) and in many beyond SM (BSM) theories. Furthermore, it has been shown much time ago that the quark sector~\cite{Froggatt:1978nt} is easily described in this context. In addition, the formulation of a model based on the $U(1)$ symmetry, in the supersymmetric context as the holomorphicity of the superpotential simplifies the construction of the Yukawa interactions, is simple and elegant:\\
- The flavour symmetry acts horizontally on leptons and the charges can be written as $e^c\sim(n_1^{\rm R},n_2^{\rm R},0)$, with $n_1^{\rm R}>n_2^{\rm R}>0$, for the $SU(2)_{\rm L}$ lepton singlets and as $\ell\sim(n_1^{\rm L},n_2^{\rm L},0)$ for the $SU(2)_{\rm L}$ lepton doublets. The Higgs fields are taken to be $U(1)$-singlets.\\
- Once leptons have $U(1)$ charges, the Yukawa terms are no longer invariant under the action of the flavour symmetry. To formally recover the invariance, a new scalar field, the flavon $\theta$, can be introduced that transforms non-trivially only under $U(1)$, with charge $n_\theta$. Then, the Yukawa Lagrangian can be written as
\begin{equation}
-\mathcal{L}_Y =\,(y_e)_{ij}\,\ell_i\,H_d\, e^c_j \left(\dfrac{\theta}{\Lambda}\right)^{p_e}+(y_\nu)_{ij}\,\dfrac{\ell_i\ell_j H_u H_u}{\Lambda_{\rm L}}\left(\dfrac{\theta}{\Lambda}\right)^{p_\nu} +\text{h.c.}
\label{Yukawas}
\end{equation}
where $\Lambda$ is the cut-off of the effective flavour theory and $\Lambda_{\rm L}$ the scale of the lepton number violation. $(y_e)_{ij}$ and $(y_\nu)_{ij}$ are free parameters, taken to be complex and with modulus of order 1. $p_e$ and $p_\nu$ are suitable powers of the dimensionless ratio $\theta/\Lambda$ necessary to enforce the invariance under the flavour symmetry.
Without loss of generality, we can fix $n_\theta=-1$; consequently, $n_1,n_2>0$ to assure that the Lagrangian expansion is well defined.
Here, neutrino masses are described by the effective Weinberg operator (see i.e. Refs.~\cite{Altarelli:2000fu,Altarelli:2012ia}).\\
- Once the flavon and the Higgs fields develop non-vanishing vacuum expectation values (VEVs), the flavour and electroweak symmetries are broken and mass matrices arise from the Yukawa Lagrangian. Defining, $\epsilon\equiv\mean{\theta}/\Lambda<1$, a useful parametrisation for the Yukawa matrices is given by
$Y_e=F_{e^c}\,y_e\,F_{\ell}$ and $Y_\nu=F_{\ell}\,y_\nu\,F_{\ell}$, where $F_f=\mathrm{diag}(\epsilon^{n_{f1}},\epsilon^{n_{f2}},\epsilon^{n_{f3}})$. Following Ref.~\cite{Altarelli:2012ia,Bergstrom:2014owa}, the charges will be taken to be integers.
\boldmath
\subsection{Specific $U(1)$ models}
\label{sec:SpecificU1}
\unboldmath
In the following, focussing on constructions where neutrino masses are described by the Weinberg operator, two specific models will be considered: the model $A$ representative of anarchical constructions and the model $H$ representative of hierarchical ones,
\begin{center}
\begin{tabular}{ccc}
{Anarchy ($A$)}:& $e_{\rm R}\sim(3,1,0)$\,, & $\ell_{\rm L}\sim(0,0,0)$\,, \\
{Hierarchy ($H$)}:& $e_{\rm R}\sim(8,3,0)$\,, & $\ell_{\rm L}\sim(2,1,0)$\,.
\end{tabular}
\end{center}
$A$ encodes the idea that an even structureless mass matrix can lead to a correct description of neutrino data: this mass matrix is characterised by entries that are random numbers which, under the additional requirement of basis invariance, leads to a unique measure of the mixing matrix -- the Haar measure~\cite{Hall:1999sn,deGouvea:2012ac,Lu:2014cla}. It has been claimed that such matrix generically prefers large mixings~\cite{Hall:1999sn} and that the observed sizable deviation from a zero reactor angle seems to favour anarchical models when compared to other more symmetric constructions \cite{deGouvea:2012ac}. However, as discussed in \refcite{Espinosa:2003qz}, how much a large value of a parameter is preferred can depend strongly on the definition of ``preferred'' and of ``large''.
It has been suggested in Ref.~\cite{Altarelli:2012ia,Bergstrom:2014owa} (see Ref.~\cite{Hirsch:2001he} for an earlier study) that the performances of anarchical models, formulated in a $U(1)$ context giving no charges to the left-handed fields, in reproducing the 2012 neutrino data are worse than those of models constructed upon the $U(1)$ flavour symmetry. In the latter ones, the small neutrino parameters are due to the built-in hierarchies and not due to chance. The construction $H$ considered in Ref.~\cite{Bergstrom:2014owa} has been shown by mean of the Bayesian inference to be the best one to describe the data, among all the possible $U(1)$ models.
The textures for the charged leptons $Y_e$ and neutrino $Y_\nu$ Yukawa matrices are as follows:
\begin{equation}
\begin{aligned}
A:\quad
Y_e&=\left( \begin{matrix} \epsilon^3 & \epsilon & 1 \\ \epsilon^3 & \epsilon & 1 \\ \epsilon^3 & \epsilon & 1 \end{matrix}\right)\,,\;
Y_\nu= \left( \begin{matrix} 1 & 1 & 1 \\ 1 & 1 & 1 \\ 1 & 1 & 1 \end{matrix}\right)\,,\\
H:\quad
Y_e&= \left( \begin{matrix} \epsilon^{10} & \epsilon^6 & \epsilon^2 \\ \epsilon^9 & \epsilon^5 & \epsilon \\ \epsilon^8 & \epsilon^4 & 1 \end{matrix} \right) ,\;
Y_\nu= \left( \begin{matrix} \epsilon^4 & \epsilon^3 & \epsilon^2 \\ \epsilon^3 & \epsilon^2 & \epsilon \\ \epsilon^2 & \epsilon & 1 \end{matrix} \right)\,.
\end{aligned}
\label{NewModels}
\end{equation}
In the spirit of $U(1)$ models, the coefficients in front of $\epsilon^n$ are expected to be complex numbers with absolute values of ${\cal O}(1)$ and arbitrary phases. As $Y_\nu$ is a symmetric matrix, the total number of parameters that should be considered in the analysis is $30$, from the Yukawa matrices, plus the unknown value of $\epsilon$.
While for the details of the analysis we refer to the original publication \refcite{Bergstrom:2014owa}, here we just comment on the results of the comparison between $A$ and $H$. Both the models have a $\chi^2$-minimum of zero and therefore a $\chi^2$-analysis can never exclude any of the models or be meaningful to compare them. On the other side, a Bayesian analysis allows instead a quantitative comparison of the models: the ratio between the logarithms of the evidences of $H$ normalised to the evidence of $A$, i.e., the Bayes factor between $H$ and $A$, is $\log B \simeq 3\div4.5$, depending on the prior on $\epsilon$. The uncertainty on the logarithms of the Bayes factors is about $0.2$. Accordingly to the Jeffreys scale, these values translate in a moderate evidence in favour of $H$ with respect to $A$.
Regarding the data adopted here, improved measurements of the oscillation parameters cannot further discriminate between the models. Instead, there are other observables which could be accurately measured in future experiments, and in principle could be used to distinguish between the models. These are primarily the CP-phase $\delta$ and observables related to the values of neutrino masses ($m_ {ee}$, $m_\beta$, $\Sigma$). See \refcite{Bergstrom:2014owa} for details.
\boldmath
\section{Non-Abelian models}
\unboldmath
\label{MLFVmod}
One of the main problematics of dealing with Abelian symmetries is the fact that the three fermion generations are independent from each other, translating in the large number of free parameters. On the other hand, in the context of non-Abelian symmetries, when fermions transform with multidimensional representations, the three families are connected one to the others, reducing the number of free parameters and therefore increasing predictivity.
Non-Abelian continuous symmetries have also been deeply investigated in the flavour sector, but mainly connected to the Minimal Flavour Violation (MFV)~\cite{Chivukula:1987py} ansatz: i.e. the requirement that
all sources of flavour violation in the SM and BSM are described at low-energies uniquely in terms
of the known fermion masses and mixings. Several distinct models formulated in this framework~\cite{D'Ambrosio:2002ex,Cirigliano:2005ck,Gavela:2009cd,Alonso:2011jd,Davidson:2006bd,Alonso:2011yg,Alonso:2012fy,Alonso:2013mca,Alonso:2013nca} turn out to be consistent with TeV new physics.
The power of MFV descends from the fact that it exploits the symmetries that the SM itself contains in a
certain limit: that of massless fermions. For example, in the case of the Type I Seesaw mechanism with three
RH neutrinos added to the SM spectrum, the flavour symmetry of the full Lagrangian, when Yukawa couplings and
the RH neutrino masses are set to zero, is $G_f=G^q_f\times G^\ell_f$ with $G^q_f=U(3)_{Q_L}\times U(3)_{U_R}\times U(3)_{D_R}$ and $G^\ell_f=U(3)_{\ell_L}\times U(3)_{E_R}\times U(3)_{N}$.
Under the flavour symmetry group $G_f$ fermion fields transform as
\begin{equation}
\begin{gathered}
Q_L\sim(3,1,1)_{G_f^q}\,,\qquad
U_R\sim(1,3,1)_{G_f^q}\,,\qquad
D_R\sim(1,1,3)_{G_f^q}\,,\\
\ell_L\sim(3,1,1)_{G_f^\ell}\,,\qquad
E_R\sim(1,3,1)_{G_f^\ell}\,,\qquad
N_R\sim(1,1,3)_{G_f^\ell}\,.
\end{gathered}
\label{FermionTransf}
\end{equation}
The Yukawa Lagrangian for the Type I Seesaw mechanism, then, reads:
\begin{equation}
-\mathscr{L}_Y=\ov{Q}_LY_DHD_R+\ov{Q}_LY_U\tilde{H}U_R+\ov{\ell}_LY_EHE_R+\ov{\ell}_LY_\nu\tilde{H}N_R+\ov{N}^c_R\dfrac{M_N}{2}N_R+{\mbox{\rm h.c.}}
\label{Lagrangian}
\end{equation}
To introduce $\mathscr{L}_Y$ without explicitly breaking $G_f$, the Yukawa matrices $Y_i$ and the mass matrix for the
RH neutrinos $M_N$ have to be promoted to be spurion fields transforming under the flavour symmetry as:
\begin{equation}
Y_U\sim(3,\bar3,1)_{G_f^q},\,
Y_D\sim(3,1,\bar3)_{G_f^q},\,
Y_E\sim(3,\bar3,1)_{G_f^\ell},\,
Y_\nu\sim(3,1,\bar3)_{G_f^\ell},\,
M_N\sim (1,1,\bar6)_{G_f^\ell}.
\label{YTransf}
\end{equation}
The quark masses and mixings are correctly reproduced once the quark spurion Yukawas get background values as
$Y_U =V^\dag\,{\bf y}_U$ and $Y_D ={\bf y}_D$ where ${\bf y}_{U,D}$ are diagonal matrices with Yukawa eigenvalues as diagonal entries, and $V$ a unitary matrix
that in good approximation coincides with the CKM matrix.
When discussing the MFV ansatz in the leptonic sector one has at disposal three different spurions, as can be
evinced from the list in Eq.~(\ref{YTransf}). The number of parameters that can be introduced in the model through
these spurions is much larger than the low energy observables. This in general prevents a direct link among neutrino
parameters and FV observables. The usual way adopted in the literature to lower the number of parameters consists
in reducing the number of spurions from three to two: for example Ref.~\cite{Cirigliano:2005ck} takes $M_N\propto\mathbbm{1}$; in Ref.~\cite{Gavela:2009cd}, a two-family RH neutrino model is considered with $M_N\propto\sigma_1$;
finally in Ref.~\cite{Alonso:2011jd} $Y^\dag_\nu Y_\nu\propto\mathbbm{1}$ is assumed.
An unifying description for all these models can be obtained by introducing the Casas-Ibarra parametrization: in the basis of diagonal mass matrices for RH neutrinos, LH neutrinos and charged
leptons, the neutrino Yukawa coupling can be written as
\begin{equation}
Y_\nu=\dfrac{1}{v}U\sqrt{\hat m_\nu}R\sqrt{\hat M_N},
\label{CIpar}
\end{equation}
where $v$ is the electroweak vev, the hatted matrices are light and heavy neutrino diagonal mass matrices, $U$ refers
to the PMNS mixing matrix and $R$ is a complex orthogonal matrix, $R^T R=\mathbbm{1}$. A correct description of lepton
masses and mixings is achieved assuming that $Y_E$ acquires a background value parametrised by a diagonal matrix,
$Y_E={\bf y}_E\equiv{\rm diag}(y_e,\,y_\mu,\,y_\tau)$, while the remaining spurion, $M_N$ or $Y_\nu$, accounts for the neutrino masses and the PMNS matrix (see
Refs.~\cite{Cirigliano:2005ck,Gavela:2009cd,Alonso:2011jd}).
\boldmath
\subsection{Dynamical Yukawas}
\label{Sect:DynamicalYukawas}
\unboldmath
Despite of the phenomenological success, it has to be noticed that, however, MFV does not provide by itself
any explanation of the origin of fermion masses and/or mixing, or equivalently does not provide any explanation
for the background values of the Yukawa spurions. This observation motivates the studies performed in
Refs.~\cite{Alonso:2011yg,Alonso:2012fy,Alonso:2013mca,Alonso:2013nca}, where the Yukawa spurions are promoted to dynamical scalar fields: the case in which a one-to-one correlation among Yukawa couplings and fields is assumed, $Y_i\equiv \mean{{\bf Y}_i}/\Lambda_f$, is discussed at length. The scalar potential constructed out of these fields was studied in Refs.~\cite{Alonso:2011yg,Alonso:2012fy,Alonso:2013mca,Alonso:2013nca}, considering renormalisable operators (and adding also lower-order non-renormalisable terms for the quark case): these effective Lagrangian expansions are possible under the assumption that the ratio of the flavon vevs and the cutoff scale of the theory is smaller than 1, condition that is always satisfied but for the top Yukawa coupling. In this case a non-linear description would be more suitable. We focus here only on the lepton sector, while for the quark sector we refer to the original article in \refcite{Alonso:2011yg}.
\\
\noindent{\bf - Two-family case -}
It is instructive and interesting to start with a toy model with only two generations~\cite{Alonso:2012fy}. Under the assumption of degenerate RH neutrino masses, $M_1=M_2\equiv M$, only the spurion fields $Y_E$ and $Y_\nu$ are promoted to dynamical fields, ${\bf Y}_E$ and ${\bf Y}_\nu$, and the flavour symmetry in this case is $G_f^\ell=U(2)_{\ell_L}\times U(2)_{E_R}\times O(2)_{N}$.
Only five independent invariants can be obtained at the renormalisable level:
\begin{equation}
\begin{gathered}
{\mbox{\rm tr}}\left[{\bf Y}_E{\bf Y}_E^\dagger\right],\,\,\,\,\,
{\mbox{\rm tr}}\left[{\bf Y}_\nu {\bf Y}_\nu^\dagger\right],\,\, \,\,\,
{\mbox{\rm tr}}\left[\left({\bf Y}_E{\bf Y}_E^\dagger\right)^2\right],\,\,\,\,\,
{\mbox{\rm tr}}\left[\left({\bf Y}_\nu {\bf Y}_\nu^\dagger\right)^2\right],\,\,\,\,\,
{\mbox{\rm tr}}\left[\left({\bf Y}_\nu\sigma_2{\bf Y}_\nu^\dagger\right)^2\right].
\end{gathered}
\label{LeptonInvariants}
\end{equation}
All the terms account for the lepton masses, but the last one that fixes the mixing angle. By adopting the Casas-Ibarra parametrisation in Eq.~(\ref{CIpar}) and minimising the scalar potential with respect to the angle $\theta$ and the Majorana phase $\alpha$, the following two conditions result:
\begin{equation}
(y^2-y'^2)\sqrt{m_{\nu_2}m_{\nu_1}}\sin2\theta\cos2\alpha=0\,,\qquad\qquad
\tan2\theta=\sin2\alpha\frac{y^2-y'^2}{y^2+y'^2}
\frac{2\sqrt{m_{\nu_2}m_{\nu_1}}}{m_{\nu_2}-m_{\nu_1}}\,,
\end{equation}
where $y$ and $y'$ are two parameters of $Y_\nu$ (see Ref.~\cite{Alonso:2012fy} for details). The first condition implies a maximal Majorana phase, $\alpha=\pi/4$ or $\alpha=3\pi/4$, for a non-trivial mixing angle. However, this does not imply observability of CP violation at experiments,
as the relative Majorana phase among the two neutrino eigenvalues is $\pi/2$. The second condition above represents a link among the size of mixing angle and the type of the neutrino spectrum: a large mixing angle is obtained from almost degenerate masses, while a small angle follows in the hierarchical case.
\\
\noindent{\bf - Generalisation to the three-family case -}
Moving to the realistic scenario of three families, but still considering $G_f^\ell$ containing a $O(2)_{N}$ factor, $N_R$ ($N'_R$) is a doublet (singlet) of $O(2)_{N}$. Correspondingly, the neutrino Yukawa field accounts for two components: a doublet and a singlet of $O(2)_{N}$, ${\bf Y}_\nu\sim(3,1,\bar2)$ and ${\bf Y}'_\nu\sim(3,1,1)$.
The leptonic flavour Lagrangian is given in this case by
\begin{equation}
-\mathscr{L}_Y=\ov{\ell}_LY_EHE_R+\ov{\ell}_LY'_\nu\tilde{H}N'_R+\ov{\ell}_LY_\nu\tilde{H}N_R+\dfrac{M'}{2}\ov{N}^{\prime c}_RN'_R+\dfrac{M}{2}\ov{N}^c_R\mathbbm{1} N_R+{\mbox{\rm h.c.}}\,.
\label{Lagrangian2}
\end{equation}
Once the Yukawa flavons develop vevs, the light neutrino mass matrix is generated:
\begin{equation}
m_\nu=\dfrac{v^2}{M'}Y'_\nu Y^{\prime T}_\nu+\dfrac{v^2}{M}Y_\nu Y^T_\nu\,.
\end{equation}
A total of nine independent invariants at the renormalisable level can be constructed in this case, namely
\begin{equation}
\begin{gathered}
{\mbox{\rm tr}}\left[{\bf Y}_E{\bf Y}_E^\dagger\right]\,,\quad
{\mbox{\rm tr}}\left[{\bf Y}_\nu{\bf Y}_\nu^\dagger\right]\,,\quad
{\bf Y}^{\prime \dag}_\nu {\bf Y}'_\nu\,,\quad
{\mbox{\rm tr}}\left[\left({\bf Y}_E{\bf Y}_E^\dagger\right)^2\right]\,, \quad
{\mbox{\rm tr}}\left[\left({\bf Y}_\nu{\bf Y}_\nu^\dagger\right)^2\right]\,,\\
{\mbox{\rm tr}}\left[{\bf Y}_E{\bf Y}_E^\dagger{\bf Y}_\nu{\bf Y}_\nu^\dagger\right]\,, \qquad
{\mbox{\rm tr}}\left[{\bf Y}_\nu{\bf Y}_\nu^T{\bf Y}_\nu^*{\bf Y}_\nu^\dag\right]\,,\qquad
{\bf Y}^{\prime \dag}_\nu {\bf Y}_E {\bf Y}^\dag_E {\bf Y}'_\nu \,,\qquad
{\bf Y}^{\prime \dag}_\nu {\bf Y}_\nu {\bf Y}^\dag_\nu {\bf Y}'_\nu \,.
\end{gathered}
\end{equation}
When considering the minimisation of the scalar potential, there are four analytical solutions:
\begin{equation}
\begin{aligned}
1)\begin{cases}
\tan2\theta_{12}=z/z'\\
m_{\nu_1}\neq m_{\nu_2}\\
m_{\nu_3}=0
\end{cases}
2)
\begin{cases}
\theta_{12}=\pi/4\\
m_{\nu_1}= m_{\nu_2}\neq m_{\nu_3}\\
\alpha=\pi/4
\end{cases}
3)\begin{cases}
\theta_{23}=\pi/4\\
m_{\nu_1}\neq m_{\nu_2} = m_{\nu_3}\\
\alpha=\pi/4
\end{cases}
4)\begin{cases}
\tan2\theta_{23}=z/z'\\
m_{\nu_2}\neq m_{\nu_3}\\
m_{\nu_1}=0
\end{cases}
\end{aligned}
\label{configurations}
\end{equation}
Case 1 (4) describes an inverse (direct) hierarchical spectrum and only one sizable mixing angle, the solar (atmospheric) one. In case 2, the light neutrinos ${\nu_1}$ and ${\nu_2}$ are degenerate and both mass orderings (hierarchical or degenerate) can be accommodated, while a maximal solar angle is predicted. Finally, case 3 corresponds to degenerate $\nu_2$ and $\nu_3$: a realistic scenario points to three almost degenerate neutrinos. Note that cases 2 and 3 encompass two degenerate neutrinos and the relative Majorana phase between the two degenerate states is $\pi/2$.
Cases 1-4 only account for one sizable angle. Realistic configurations with three non-trivial angles, however, follow in a straightforward way when interpolating between these four cases, at the prize of non-exact solutions that depend on the parameters of the of the scalar potential. The setup appears very promising, though, as all three angles can be naturally non-vanishing and moreover the number of free parameters is smaller than the number of observables, leading to predictive scenarios in which mixing angles and Majorana phases are linked to the spectrum.\\
It is finally interesting to consider the case with three degenerate RH neutrinos\cite{Alonso:2013mca,Alonso:2013nca}. The flavour symmetry is $G_f=U(3)_{\ell_L}\times U(3)_{E_R} \times O(3)_{N}$ and the basis of invariants is composed of the operators in Eq.~(\ref{LeptonInvariants}). The study of the extrema of these invariants has been presented in Ref.~\cite{Alonso:2013nca} and from the minimisation of the potential it follows that one of the possible configurations is 3) in Eq.~(\ref{configurations}). In the normal or inverse hierarchical case, two of the light neutrinos are degenerate in mass and a maximal angle and a maximal Majorana phase arise in their corresponding sector. On the other hand, if the third light neutrino is almost degenerate with the other two, then the perturbations split the spectrum and a second sizable angle arises~\cite{Alonso:2013nca}.
In summary, these results indicate that a realistic solution for the Flavour Puzzle in the lepton sector requires three RH neutrinos, two of which must be degenerate. All three light neutrinos would therefore acquire masses, and the precise values of the mixing angles and Majorana phases are related to the specific light mass spectrum, result that is almost exclusively a feature of continuous non-Abelian symmetries.
\bigskip
\section{Acknowledgments}
I acknowledge partial support by European Union FP7 ITN INVISIBLES (Marie Curie Actions, PITN-GA-2011-289442), by the Juan de la Cierva programme (JCI-2011-09244) and by the Spanish MINECOs ``Centro de Excelencia Severo Ochoa'' Programme under grant SEV-2012-0249. Finally, I thank the organisers and the conveners of the ``NuPhys2014: Prospects in Neutrino Physics'' conference for the kind invitation and for their efforts in organising this enjoyable meeting.
\providecommand{\href}[2]{#2}\begingroup\raggedright |
\section{Introduction}
The majority of stars form in clustered environments \citep[e.g.,][]{lada2003, portegies2010}, where close encounters between stars are frequent. These close encounters can perturb or destroy planetary systems \citep[e.g.,][]{spurzem2009, boley2012}. Even a mild perturbation can break up a marginally stable planetary system and can even affect the shortest-period planets in the system \citep[e.g.,][]{hao2013}, which can result in strong planet-planet scattering, physical collisions between planets and their host stars or other planets \citep[e.g.,][]{Rasio1996, Chatterjee2008, Juric2008, Nagasawa2011}, long-term secular evolution \citep[e.g.,][]{malmberg2007_kozai, parker2009_kozai, malmberg2011}, and the ejection of planets from the system. In addition, mass loss due to stellar evolution of single or binary host stars can result in loss of planetary companions through similar interactions \citep[e.g.,][]{veras2011, veras2012, adams2013, voyatzis2013, nowak2013}. Even the Galactic tidal field can indirectly play a role in the disruption of planetary systems when a wide stellar companion is present \citep[e.g.,][]{kaib2013}. \cite{verasscattering2012} demonstrate that the dynamical and stellar evolution of isolated planetary systems alone cannot account for the observed free-floating planet (FFP) population in the Galactic field, and that close encounters in star clusters are an important source of FFPs in the field.
Although only a small number of FFP candidates have been detected so far, they are potentially abundant in the Galactic disk \cite[e.g.,][]{strigari2012}. Exoplanets that are too distant from stars to directly affect their observable properties are most easily detected using microlensing \cite[e.g.,][]{mao1991, gould1992, abe2004, beaulieu2006, gaudi2012}. Microlensing surveys also have the potential to discover FFPs \citep[e.g.,][]{distefano2012},
and results imply that there are roughly twice as many FFPs in the Solar neighbourhood than there are main-sequence stars \citep{sumi2011}. In addition, deep imaging surveys can be used to detect young planetary-mass objects near the deuterium burning limit in young star clusters \citep[e.g.,][]{lucas2006, caballero2007, bihain2009, quanz2010, penaramirez2011, scholz2012, delorme2012}. These observational studies help us further constrain the origin and fate of these FFPs provided that a good understanding of their dynamics is known.
FFPs are thought to have been ejected from their system with velocities of typically $0.1-10$~km\,s$^{-1}$, as a result of planet-planet scattering ({\em delayed ejection}) or immediately after a close encounter with a passing star ({\em prompt ejection}), and the ejection velocities from the latter process tend to be higher \citep{malmberg2011}. When a FFP is ejected in a star cluster, it may escape immediately if its ejection velocity exceeds the local escape velocity of the star cluster, or it may remain bound to its host cluster for millions of years, until it escapes through ejection or evaporation. During its life in a star cluster, a FFP can experience multiple close encounters with other stars before escaping, and may even be re-captured by another star \citep{kouwenhoven2010, malmberg2011, moeckel2011, parker2012, peretskouwenhoven}.
Direct $N$-body simulations of single-planet systems in star clusters have shown that many are disrupted and that the resulting FFPs can remain in these star clusters for many millions of years \citep[e.g.,][]{hurley2002, parker2012, craig2013}. Fly-by simulations mimicking the evolution of multi-planet systems in star clusters confirm that the survival of these systems depends strongly on the properties of the stellar environment and the semi-major axes of the planets, but also demonstrated that planetary multiplicity itself plays an important roles as planets in perturbed systems also mutually interact \citep[e.g.,][]{Chatterjee2012, hao2013, liu2013}.
The aim of this study is to analyse the dynamical properties of FFPs in low-mass star clusters, with a particular focus on close encounters between the members of the star cluster (single stars, binary stars, and FFPs). This article is organised as follows. The methods and assumptions are described in \S~\ref{section:method}. The results are presented in \S~\ref{section:results}. Finally, we draw the conclusions and describe our future work in \S~\ref{section:conclusions}.
\section{Method and assumptions} \label{section:method}
\subsection{Initial conditions for the star clusters}
\begin{table}
\caption{Initial conditions for the modelled star clusters. \label{table:starcluster} }
\begin{tabular}{ll}
\hline\hline
Quantity & Value \\
\hline
Number of stars \& binaries & $N=N_{\rm s}+N_{\rm b} = 500$, 1000, 2000 \\
Half-mass radius & $r_{\rm hm} = 0.38$, 0.77, 1.54~pc\\
Dynamical model & \cite{plummer1911} \\
Virial ratio & $Q=1/2$ \\
Tidal field & Galactic Solar orbit \\
\hline
Initial mass function & \cite{kroupatoutgilmore}, $0.2-5M_\odot$ \\
Binary fraction & $\mathcal{B}=N_{\rm b}/N = 0\%$, $20\%$, $50\%$ \\
Semi-major axis distr. & $f_a(a) \propto a^{-1}$; ($10^{-6}-10^{-3})\timesr_{\rm vir}$\\
Mass ratio distr. & Random pairing \\
Eccentricity distr. & $f_e(e) = 2e$ ($0\leq e<1$) \\
Orbital orientation & Random \\
\hline
Planet-to-star ratio & $\mathcal{R}=N_{\rm p}/N_{\rm sb}=0.5, 1, 2$ \\
Planet mass & $M_p = 1 M_J$ \\
Density distribution & Equivalent to that of stars\\
Planet ejection velocity & $M_{\rm s}$, $f_{\rm ej}$ (see Figure~\ref{figure:ejectionvelocities} and \S~\ref{section:planetinitial}) \\
\hline\hline
\end{tabular}
\end{table}
The properties of our model star clusters are summarised in
Table~\ref{table:starcluster}. We create star clusters with
\cite{plummer1911} density and velocity distributions. We carry out simulations of open star clusters with $N=N_{\rm s}+N_{\rm b}=500$, 1000 and 2000 systems, where $N_{\rm s}$ and $N_{\rm b}$ represent the number of single stars and
binary systems in the star cluster, respectively, and $N_{\rm sb}=N_{\rm s}+2N_{\rm b}$ is the total
number of stars in the cluster. The initial values for the virial radii are $r_{\rm vir}=0.5$, 1.0, and 2.0~pc, which are typical for open clusters \citep[e.g.,][]{lada2003}, and correspond to intrinsic half-mass radii of $r_{\rm hm} = 0.38$, 0.77 and 1.54 pc, since for the Plummer model $r_{\rm vir}\approx 1.30\,r_{\rm hm}$ \citep[e.g.,][]{heggiehut}. Each star cluster is initially in
virial equilibrium, i.e., $Q=|K/P|=1/2$, where $K$ and $P$ are the total kinetic
and potential energies, respectively. As a consequence, the corresponding initial (one-dimensional)
velocity dispersion $\sigma(r)$ as a function of distance to the cluster centre $r$
is
\begin{equation} \label{eq:velocitydispersion}
\sigma^2(r) = \frac{GM}{6r_{\rm hm}} \left( 1+ \frac{r^2}{r_{\rm hm}^2} \right)^{-1/2} \ ,
\end{equation}
where $G$ the gravitational constant
and $M$ the total cluster mass \citep{heggiehut}.
Following \cite{malmberg2007_encounters}, the stellar masses are drawn from the \cite{kroupatoutgilmore} initial mass function (IMF) in the mass range $0.2-5~M_\odot$. There is evidence that the upper mass limit may depend on the mass of the star cluster \citep[e.g.,][]{weidner2004, weidner2013}, this dependence is still under debate \citep[e.g.,][]{cervino2013a, cervino2013b}. We therefore follow \cite{malmberg2007_encounters} in choosing a constant upper mass limit, while realising the possibility that realistic clusters may host a more massive star that can alter the dynamics of the FFP population.
We carry out the simulations with binary fractions $\mathcal{B}=N_{\rm b}/N=N_{\rm b}/(N_{\rm s}+N_{\rm b})=0\%$, $20\%$, and $50\%$. The individual components of the binary systems are
randomly paired from the IMF \citep[see][for details]{kouwenhoven2006, kouwenhovenpairing}. For this method of pairing stars into binary systems the total mass of the star cluster equals $N\langle M \rangle (1+\mathcal{B})$, where $\langle M \rangle$ is the average stellar mass. The
adopted semi-major axis distribution is $f_a(a)\propto a^{-1}$, which corresponds to a flat distribution in $\log a$, also known as
\"{O}pik's law \citep[e.g.,][]{vanalbada1968,vereshchagin1987,poveda2004, kouwenhoven2005, kouwenhoven2007}. Semi-major axes are drawn from this distribution in the range
$(10^{-6}-10^{-3})\timesr_{\rm vir}$, which corresponds to $1-1000$~AU for the star clusters with $r_{\rm hm}=0.38$~pc. We do
not include very tight binary systems as they are relatively inert and effectively act as
single stars. We also do not include very wide binary systems, as their individual components experience encounters very similar to those of single stars. Moreover, many of these wide binary systems are easily destroyed \citep[e.g.,][]{parker2009_binaries}. We choose a thermal eccentricity distribution $f_e(e)=2e$ for $0 \leq e < 1$ \citep{heggie1975}, and all orbits are assigned a random spatial orientation and a random orbital phase.
We include an external tidal field by modelling the star clusters on a Galactic circular orbit in the solar neighbourhood. The tidal radii of the star clusters are then given by
\begin{equation} \label{eq:tidalfield}
r_t(M) = \left[ \frac{GM}{4A(A-B)} \right]^{1/3} \approx 1.41 \left(\frac{M}{M_\odot}\right)^{1/3} \ {\rm pc} \ ,
\end{equation}
where $A=14.4$\,km\,s$^{-1}$ kpc$^{-1}$ and $B=-12.0$\,km\,s$^{-1}$ kpc$^{-1}$ are the Oort
constants \citep[e.g.,][]{binneytremaine}, and $M$ is the total mass of the cluster. This external tidal field enhances the dissolution of the cluster, in particular for low-mass and low-density star clusters. In our grid of models the tidal radii are $r_t=6.0-11.5$~pc, depending on total mass of the star cluster.
\subsection{Initial conditions for the free-floating planets} \label{section:planetinitial}
\begin{figure}
\centering
\includegraphics[width=0.45\textwidth,height=!]{ps/veject.pdf}
\caption{The cumulative distribution of velocities at which planets escape from their host
stars, as obtained by \protect\cite{malmberg2011}. The four curves indicate the
results for models with different host star masses, and represent the
distributions for prompt ejections (during the encounter) and delayed
ejections (after the encounter). For comparison, we also model the evolution of FFP populations with zero ejection velocities, in addition to the four velocity distributions shown here.
\label{figure:ejectionvelocities} }
\end{figure}
In addition to the stellar population we also add a population of FFPs. The
initial positions of these planets are drawn from the same distribution as that of
the stars. Since FFPs are thought to have been ejected from planetary systems,
they have an initial position distribution which is statistically identical to
that of their host stars. It can be argued that planets are preferably ejected
from their host systems near the cluster centre, as the close encounter frequency
with other stars is significantly larger in this region. On the other hand, from
delayed ejection in perturbed multi-planet systems, we can expect an initial
spatial distribution of ejected planets that is similar to that of the stars in
the star cluster when the delay time is larger than the crossing time, which is $1-3$~Myr for our cluster models. Fly-by experiments by \cite{malmberg2011}, \cite{hao2013} and \cite{liu2013} have demonstrated that, although some close encounters result in prompt ejections, perturbed planetary systems can eject planets up to tens of millions of years after a close encounter has taken place, at a location that is unrelated to the place where the close encounter took place.
Moreover, we make the assumption that all FFPs are free-floating at the start of our simulations, and that none of the stars have planetary companions. Hence, we ignore the possibility of new FFPs being ejected from perturbed planetary systems in the star clusters. Although this may seem a strong constraint to the applicability of our study, we show in Section~\ref{section:results} that many of our results are also applicable to populations of FFPs that are ejected at different times. Moreover, our results can also be use to calculate encounter probabilities for individual FFPs, and to estimate the time at which an individual FFPs may escape.
Two dynamical limits can be considered here: (i) the limit where all planets have zero ejection velocity with respect to their host stars, and (ii) the limit where all planets have ejection velocities that are substantially larger than the star cluster's escape velocity. The latter case is trivial and results in the immediate ejection of all FFPs. In the former case, both the stellar and planetary populations initially have identical position and velocity distributions. Subsequently, energy exchange between the bodies, result in mass segregation and also in the gradual ejection of all planets from the system.
The adopted ejection velocity distributions $f_{\rm ej}$ are based on the work of \cite{malmberg2011}, who studied the evolution of planetary systems consisting of four gas-giant planets orbiting a star in a star cluster. They studied the prompt ejection
velocities of planets escaping during close encounters with other stars ($f_{\rm ej}=P$), and the delayed ejections as a result of planet-planet scattering in a perturbed system ($f_{\rm ej}=D$). Their simulations were carried out using host stars with masses of $M_{\rm s}=0.6M_\odot$ and $M_{\rm s}=1.0M_\odot$. The
cumulative ejection velocities for the different models are shown in
Figure~\ref{figure:ejectionvelocities}. For comparison, we also carry out
simulations with zero ejection velocities, $f_{\rm ej}=Z$, where the initial velocity
distributions of planets and stars are identical. We study the dynamical evolution
of the FFP population for each of these five distributions $f_{\rm ej}$. As the two delayed ejection velocity distributions in Figure~\ref{figure:ejectionvelocities} are similar, the results for the models with $M_{\rm s}=1.0M_\odot$ and $f_{\rm ej}=D$ are omitted in most figures and discussion throughout the remaining part of this paper.
Each planet is assigned a mass of $M_P=1~M_J \approx 10^{-3}M_\odot$. Since these planets are of very low mass compared to the stars, the actual value of
the mass does not affect the evolution of the stellar population and star
cluster as a whole, the star-planet encounter rate, or the rate of planet
capture by stars. It may, however, slightly affect the planet-planet encounter
properties.
For the same reason, the total number of FFPs does not affect the evolution of the stellar population in the star cluster. It simply scales the total number of close encounters and capture events involving planets. Inspired by the findings of \cite{sumi2011} we adopt an initial planet-to-star ratio of $\mathcal{R}(0)\equivN_{\rm p}(0)/N_{\rm sb}(0)=2$ in our default model, and also carry out simulations with $\mathcal{R}(0)=0.5$, and $\mathcal{R}(0)=1$, for comparison.
As planets are presumed to have been ejected from the planetary system in which they formed, the velocity dispersion of the FFP population is larger than that of the stellar population. The initial velocity of each FFP therefore constructed by combining two velocity vectors. The main component, $\vec{v}_p(r)$, representing the velocity of the host star at the moment of ejection, is drawn from the the Plummer distribution (see Eq.~\ref{eq:velocitydispersion}). The second velocity component, $\vec{v}_e$, is drawn from the distribution of ejection velocities at which the FFP escaped from its host star:
\begin{equation} \label{eq:totalvffp}
\vec{v}_{\rm FFP}(r) = \vec{v}_p + \vec{v}_e \ .
\end{equation}
The resulting initial velocity dispersion for the FFP population is then
\begin{equation} \label{eq:supervirial}
\sigma_{\rm FFP}^2 = \sigma_p^2(r) + \sigma_e^2 \ ,
\end{equation}
because we draw $\vec{v}_p$ and $\vec{v}_e$
independently and with random spatial orientations.
The (one-dimensional) velocity dispersion of the stars at the half-mass radius (Eq.~\ref{eq:velocitydispersion}) ranges between $\sigma_p(r_{\rm hm})\approx 0.3$~km\,s$^{-1}${} (for $N_{\rm sb}=500$ and $r_{\rm hm}=1.54$~pc) and $\sigma_p(r_{\rm hm}) \approx 1.5$~km\,s$^{-1}${} (for $N_{\rm sb}=3000$ and $r_{\rm hm}=0.38$~pc). The planet ejection velocities are typically in the range $0.1-5$~km\,s$^{-1}$, and can therefore contribute substantially to the velocity dispersion of the FFPs in the star clusters. We therefore expect a relatively large fraction of the FFPs to escape at early times from low-density star clusters.
\subsection{Simulation parameters}
We use the NBODY6 package \citep{aarseth1999,aarseth2003} to carry out the simulations. Stellar evolution is modelled following the prescriptions of \cite{eggleton1989,eggleton1990} and \cite{hurley2000}.
The NBODY6 code was modified to suit
the requirements of (i) the inclusion of the free-floating planets, and (ii)
identification of the close encounters.
In a star cluster with a virial radius $r_{\rm vir}$ consisting of $N$ single stars, when a pair of stars approaches within a distance $r_{\rm close}\approx 4r_{\rm vir}/N$ their velocities are significantly perturbed \cite[e.g.,][]{aarseth2003}. In our grid of simulations, $r_{\rm close}$ is in most cases several hundred to somewhat more than a thousand astronomical units. Inspired by the previous work of \cite{malmberg2007_encounters}, we record encounters within a distance of $r_{\rm enc} \leq 1000$~AU. Encounters with hard (i.e., tight) binaries are registered as such, while for encounters with soft (i.e., wide) binary systems, the encounters with individual stars are registered as encounters with single stars. The individual components of binary systems with periastron separations smaller than 1000~AU experience encounters on a regular basis, according to this definition, but are excluded from the analysis of the close encounter properties. All simulations are carried out until the star cluster has completely dissolved,
such that we are able to register all close encounters that occur during the
lifetime of each star cluster. In order to account for statistical effects, we simulate ten realisations for each model.
\section{Results} \label{section:results}
\subsection{Star cluster evolution and membership}
\subsubsection{Star cluster dynamics}
\begin{figure}
\centering
\includegraphics[width=0.45\textwidth,height=!]{ps/ns_t.pdf}\\
\includegraphics[width=0.45\textwidth,height=!]{ps/nsnp_t.pdf}\\
\caption{{\em Top}: the evolution of the number of cluster member stars $N_{\rm sb}$ (blue
curve) and planets $N_{\rm p}$ (red curve) as a function of time. This particular example shows the results for the model with $N_{\rm sb}=2000$, $\mathcal{B}=0\%$, $\mathcal{R}(0)=1$, and $r_{\rm hm}=0.38$~pc. The distribution at which the FFPs are assumed to have been ejected from their host star is identical to the model with $f_{\rm ej}=P$ (for $1M_\odot$) in Figure~\ref{figure:ejectionvelocities}.
{\em Bottom}:
the planet-to-star ratio $\mathcal{R}(t)=N_{\rm p}(t)/N_{\rm sb}(t)$ as a function of time $t$. The red line indicates the best linear fit (Eq.~\ref{eq:ratiodefinition}).
\label{figure:members} }
\end{figure}
\begin{figure}
\centering
\includegraphics[width=0.53\textwidth,height=!]{ps/rbar_05_nsm1p2.pdf}\\[-47ex]
\quad\quad\quad\quad\quad(a) $r_{\rm hm}=0.38~{\rm pc}$\\
\quad\quad\quad\quad\quad\quad\quad$f_{\rm ej}=P$ ($1M_\odot$)\\[40ex]
\includegraphics[width=0.53\textwidth,height=!]{ps/mp21_nsnp1000.pdf}\\[-47ex]
\quad\quad\quad\quad\quad\quad\quad\quad\quad\quad\quad\quad\quad\quad\quad(b) $N_{\rm p}=1000$\\
\quad\quad\quad\quad\quad\quad\quad\quad\quad\quad\quad\quad\quad\quad\quad\quad\quad\quad\quad $f_{\rm ej}=P$ ($1M_\odot$)\\[40ex]
\includegraphics[width=0.53\textwidth,height=!]{ps/rbar_05_nsnp1000.pdf}\\[-47ex]
\quad\quad\quad\quad\quad\quad\quad\quad\quad\quad\quad\quad\quad\quad\quad\quad\quad(c) $r_{\rm hm}=0.38~{\rm pc}$\\
\quad\quad\quad\quad\quad\quad\quad\quad\quad\quad\quad\quad\quad\quad\quad\quad\quad$ N_{\rm p}=1000$\\[40ex]
\caption{The fitted linear parameters $\mathcal{R}_f(0)$ vs. $\alpha_f$ for planet-to-star ratio as a function of time (Eq.~\ref{eq:ratiodefinition}) for all models. {\em Top}: fixed $r_{\rm hm}$ and ejection velocity distribution of planets, for clusters with different $N_{\rm sb}=N_{\rm s}+N_{\rm b}$ and $N_{\rm p}$. {\em Middle}: fixed $N_{\rm p}$ and ejection velocity distribution of planets, for different $r_{\rm hm}$ and $N_{\rm sb}$. {\em Bottom}: fixed $r_{\rm hm}$ and $N_{\rm p}$, for different planet ejection velocity distributions and different $N_{\rm sb}$. The adopted initial planet ejection velocity distributions include prompt ejection ($P$), delayed ejection ($D$), and zero ejection velocity ($Z$); see Figure~\ref{figure:ejectionvelocities}. The symbols and their colours represent the different initial conditions for the simulations, as indicated in the legends. Each symbol represents the average result for an ensemble of star clusters.
\label{figure:memfit} }
\end{figure}
\begin{figure}
\centering
\includegraphics[width=0.53\textwidth]{ps/rbar_05_np1000.pdf}\\[-47ex]
\quad\quad(a) $r_{\rm hm}=0.38$~pc \\ [43ex]
\includegraphics[width=0.53\textwidth]{ps/rbar_10_np1000.pdf}\\[-47ex]
\quad\quad(b) $r_{\rm hm}=0.77$~pc \\ [43ex]
\includegraphics[width=0.53\textwidth]{ps/rbar_20_np1000.pdf}\\[-47ex]
\quad\quad(c) $r_{\rm hm}=1.54$~pc \\ [43ex]
\caption{The relation between the half-number time $t_{\rm hp}$ of the number of planets in the clusters and the half-number time $t_{\rm hs}$ of the number of stars in the clusters, for different sets of star clusters models and ejection velocity distributions. All models have $N_{\rm p}=1000$. The three panels show the results for half-mass radii of $r_{\rm hm}=0.38$, 0.77, and 1.54~pc. Each symbol represents the average result for an ensemble of ten star clusters with a certain $N_{\rm sb}=N_{\rm s}+N_{\rm b}$. The colours indicate the different initial planet ejection velocity distributions for prompt ejection ($P$), delayed ejection ($D$), and zero ejection velocity ($Z$); see also Figure~\ref{figure:ejectionvelocities}.\label{figure:thf} }
\end{figure}
The dynamical evolution of the stellar population in the star cluster is practically unaffected by the presence of the planets, as the planets are three
orders of magnitude lower in mass than the stars, and therefore effectively behave as test particles. The evolution of the stellar and planetary populations over time is illustrated in Figure~{\ref{figure:members}}, which shows that the number of FFP members in the cluster decreases faster than the number of stellar members. The reason for this twofold: (i) the FFP population is initially to some degree supervirial (see Eq.~\ref{eq:supervirial}), which results in the rapid escape of many of the FFPs with a high initial velocity, and (ii) dynamical interactions result in FFPs obtaining higher velocities than stellar components, resulting in the preferential loss of planets.
The initial escape velocity $v_{\rm esc}$ for the Plummer model as a function of distance to the cluster centre $r$ is
\begin{equation} \label{eq:escapevelocity}
v_{\rm esc}(r) = \left( \frac{2GM}{r} \right)^{1/2} \left( 1+\frac{r_{\rm hm}^2}{r^2} \right)^{-3/4}
\end{equation}
\citep{heggiehut}, where $M$ is the total cluster mass and $G$ the gravitational constant. At the half-mass radius ($r=r_{\rm hm}$) this equation reduces to $v_{\rm esc} \approx 0.84 \sqrt{GM/r_{\rm hm}}$. For our models its value typically ranges between $v_{\rm esc}(r_{\rm hm})\approx 0.7$~km\,s$^{-1}${} ($N_{\rm sb}=500$ and $r_{\rm hm}=1.54$~pc) and $v_{\rm esc}(r_{\rm hm}) \approx 3.6$~km\,s$^{-1}${} ($N_{\rm sb}=3000$ and $r_{\rm hm}=0.38$~pc). Although a number of FFPs are able to escape the cluster immediately, the FFPs with a relatively small ratio $\sigma_e/\sigma_p(r)$ are able to remain part
of the star cluster for some time, but are typically removed at earlier times than the stars. The stellar population is barely affected by these encounters, but the typical velocity of the FFPs in the star cluster rapidly increases. Two-body encounters typically occur on a relaxation time. For stars with a mass close to the average stellar mass in a Plummer sphere, the (initial) relaxation time $t_{\rm hm}$ at the half-mass radius is
\begin{equation} \label{eq:relaxationtime}
t_{\rm hm} = \frac{0.206\nsbr_{\rm hm}^{3/2}}{\sqrt{GM}\ln \Lambda} \ ,
\end{equation}
where $N_{\rm sb}$ is the number of stellar-mass objects, $M=N_{\rm sb}\langleM_{\rm s}\rangle(1+\mathcal{B})$ the total cluster mass, and $\ln\Lambda \approx \ln N_{\rm sb}$ is the Coulomb logarithm \citep{binneytremaine, heggiehut}. The half-mass relaxation time in our simulations ranges from $t_{\rm hm} \approx 7$~Myr (for $N_{\rm sb}=3000$ and $r_{\rm hm}=0.38$~pc) to $t_{\rm hm}\approx 30$~Myr (for $N_{\rm sb}=500$ and $r_{\rm hm}=1.54$~pc).
Encounters between the star cluster members also result in energy equipartition, also roughly on a timescale
\begin{equation}
t_{\rm ms} = \left( \frac{\langle M_{\rm sb} \rangle}{M_{\rm max}} \right)t_{\rm hm} \ ,
\end{equation}
\citep[e.g.,][]{spitzer1987} where $\langle M_{\rm sb} \rangle=\langleM_{\rm s}\rangle(1+\mathcal{B})$ is the average mass of a particle participating in the mass segregation process (a single star or a binary star) and $M_{\rm max}$ is the mass of the most massive particle in the cluster. In our simulations, $\langle M_{\rm s} \rangle / M_{\rm max} \approx 0.1$, and the mass segregation timescale therefore ranges from 0.3~Myr to 3~Myr for the set of models studied in this paper \citep[but may be shorter if the star clusters are initially substructured, see][]{allison2009a, allison2009b}. These timescales, as well as the dissolution time of the star clusters, also depend on the range of stellar masses present in the cluster \citep[e.g.,][]{kouwenhoven2014}. FFPs do not contribute to the mass segregation process in the sense that they do not affect the stellar population. However, they do obtain higher velocities as a result of close encounters with stars, also roughly on a timescale $t_{\rm hm}$.
\subsubsection{Analytic estimates for the survival times}
The FFP population can be separated into three categories: (i) FFPs that escape the cluster immediately, (ii) FFPs that are ejected from the cluster as a result of strong three-body encounters, and (iii) FFPs that evaporate from the cluster following a series of weak encounters and an interaction with the Galactic tidal field. Whether or not a FFP can escape from the star cluster immediately depends on a combination of factors, but most importantly the ratio between the FFP's initial ejection velocity and the escape velocity at the initial location of the FFP. Since we set up the FFP planets following a Plummer distribution, the initial cumulative number of planets $N(r)$ as a function of distance $r$ to the cluster centre is given by:
\begin{equation} \label{eq:planetdistribution}
N(r) = N \left( 1+\frac{r_{\rm hm}^2}{r^2} \right)^{-3/2} \ ,
\end{equation}
where $N$ is the total number of planets in the system. The number of FFPs that escape at early times can then be obtained by combining Eqs.~\ref{eq:escapevelocity} and~\ref{eq:planetdistribution}, and integrating over the entire cluster. Escapers are removed when they reach a distance of $2r_t$ (Eq.~\ref{eq:tidalfield}). For high-velocity FFPs this escape time is roughly $2r_t(M)/\sigma_{\rm FFP}\approx 1-10$~Myr.
It is possible to estimate the evolution of the FFPs beyond the initial phase of escape analytically. The linear relation between $t_{\rm hs}$ and $t_{\rm hp}$ in Figure~\ref{figure:thf} is remarkable, and can be explained as follows. The number of stellar members of the cluster decreases with time, and this trend can to first order be approximated as linear:
\begin{equation} \label{eq:starloss}
N_{\rm sb}(t)= N_{\rm sb}(0)\left( 1-\frac{t}{t_{\rm diss}}\right)
\end{equation}
\citep[e.g.,][]{baumgardt2003, heggiehut, lamers2005}.
With this simple relation, the time at which half of the stars have escaped from the star cluster is $t_{\rm hs}=t_{\rm diss}/2$. The planet-to star ratio $\mathcal{R}(t)\equiv N_{\rm p}(t)/N_{\rm sb}(t)$ decreases rapidly in a certain interval $0<t<t_1$ as high-velocity planets escape. Beyond this time, a certain fraction $x\equiv N_{\rm p}(t_1)/N_{\rm p}(0)$ of the FFPs remains. The expression for $\mathcal{R}(t)$ for $t>t_1$ is thus
\begin{equation} \label{eq:ratiostarloss}
\mathcal{R}(t) = \frac{N_{\rm p}(t_1)}{N_{\rm sb}(t_1)} \left(\frac{t_{\rm diss}-t}{t_{\rm diss}-t_1}\right) \ .
\end{equation}
The number of FFPs in the star cluster as a function of time can then be computed using Eqs.~\ref{eq:starloss} and~\ref{eq:ratiostarloss}:
\begin{equation} \label{eq:solution}
\begin{array}{ll}
N_{\rm p}(t) & = \mathcal{R}(t)N_{\rm sb}(t) \\
& =\frac{N_{\rm p}(t_1)N_{\rm sb}(0)}{N_{\rm sb}(t_1)}\left(\frac{t_{\rm diss}-t}{t_{\rm diss}-t_1}\right) \left(1-\frac{t}{t_{\rm diss}}\right) \\
& = N_{\rm p}(t_1)\left(\frac{t_{\rm diss}-t}{t_{\rm diss}-t_1}\right)^2 \ .
\end{array}
\end{equation}
The time $t_{\rm hp}$ at which half of the planets have escaped from the star cluster can then be obtained by solving Eq.~\ref{eq:solution} for $N_{\rm p}(t)/N_{\rm p}(0) = 1/2$:
\begin{equation} \label{eq:halfplanethalfstar}
t_{\rm hp} = t_{\rm hs}\left( 2 - \sqrt{\frac{2}{x}} \right) + t_1\sqrt{\frac{1}{2x}} \ ,
\end{equation}
where $x\equivN_{\rm p}(t_1)/N_{\rm p}(0)$ indicates the fraction of FFPs that remain bound to the star cluster during the time interval $[0,t_1]$. The above equation is valid for $N_{\rm p}(t_1)/N_{\rm p}(0)>1/2$, as larger initial escape fractions result in $t_{\rm hp}<t_1$. In the limit where all FFPs are initially virialized, so that $x=1$ and $t_1=0$, Eq.~\ref{eq:halfplanethalfstar} reduces to
\begin{equation}
t_{\rm hp} = (2-\sqrt{2})t_{\rm hs} \approx 0.59t_{\rm hs}
\end{equation}
In other words, if the FFP population in a star cluster initially has a velocity dispersion similar to the velocity dispersion of the star cluster, the planets typically escape at 60\% of the remaining lifetime of the star cluster. This result is independent of the star cluster parameters, and provides a good estimate of the fate of a planetary population, under the condition that Eq.~\ref{eq:starloss} is a reasonable approximation.
From Figure~\ref{figure:thf} we can estimate the fraction of planets $(1-x)$ that escapes during the initial time $0<t<t_1$ by solving Eq.~\ref{eq:halfplanethalfstar} for $x$:
\begin{equation} \label{eq:x}
x = \frac{1}{2}\left(\frac{2t_{\rm hs}-t_1}{2t_{\rm hs}-t_{\rm hp}}\right)^2 \approx \frac{1}{2}\left(\frac{2t_{\rm hs}}{2t_{\rm hs}-t_{\rm hp}}\right)^2 \ .
\end{equation}
This estimate is only valid when all FFPs are generated at $t=0$, which is generally not the case in realistic star clusters. However Eqs.~\ref{eq:halfplanethalfstar} and~\ref{eq:x} but it can also be used to statistically estimate long a FFP typically remains a member of a star cluster.
The results in Figure~\ref{figure:thf} are well described by Eq.~\ref{eq:halfplanethalfstar}. The relation between $t_{\rm hp}$ and $t_{\rm hs}$ is indeed linear, although slightly steeper in the regime where the star clusters can retain the majority of the FFPs for a substantial amount of time. The relation breaks down when many FFPs escape at early times. The time $t_{\rm hp}$ depends strongly on the ejection velocity distribution of the FFPs. A comparison between the panels in Figure~\ref{figure:thf} also shows that the FFPs escape earlier from clusters with a smaller velocity dispersion (i.e., a larger $r_{\rm hm}$). For $f_{\rm ej}=P$ ($1M_\odot$) the ejection velocity of the planets is largest, and therefore $t_{\rm hp}$ is small; many planets escape within $10-20$~Myr.
\subsubsection{Escape of the planet population}
Due to the preferred ejection of FFPs from the system, the planet-to-star ratio $\mathcal{R}(t)\equiv N_{\rm p}(t)/N_{\rm sb}(t)$ rapidly decreases over time. The evolution of $\mathcal{R}(t)$ is roughly linear for most of the time, and we therefore quantify this evolution with a linear fit:
\begin{equation} \label{eq:ratiodefinition}
\mathcal{R}(t) \equiv \frac{N_{\rm p}(t)}{N_{\rm sb}(t)} \approx \mathcal{R}_f(0) \left[ 1 + \alpha_f t \right] \ ,
\end{equation}
where $t$ is the time in units of Myr and $\alpha$ in units of Myr$^{-1}$. In the ideal case where this dependence is truly linear, the quantity $\mathcal{R}_f(0)$ is the initial planet-to-star-ratio, and the quantity $\alpha_f \approx d\mathcal{R}(t)/dt$ indicates the relative escape rate of planets in units of Myr$^{-1}$. Although this derivative varies slightly with time, it is to first order constant. Under this assumption, the timescale at which all planets are removed from the star cluster is $t_{\rm diss} \approx -\alpha_f^{-1}$. Apart from statistical fluctuations, $\alpha_f$ is independent of the initial number of FFPs, $N_{\rm p}$. Since the number of planets decreases more quickly than the number of stars, $\alpha_f$ is always negative. The quantity $\alpha_f$ can be used to estimate the number of planets in the star cluster at any time.
In the case of a FFP population that is initially close to virial equilibrium with the stellar population, which is the case in Figure~\ref{figure:members}, the fitted value $\mathcal{R}_f(0)$ is close to (but not equal to) the initial planet-to-star ratio. When planets are ejected from their host systems with high velocities, a large number escape immediately, and those in the low-velocity tail of the distribution follow Eq.~\ref{eq:ratiodefinition}. Due to early escape, the fitted value $\mathcal{R}_f(0)$ is smaller than the initial planet-to-star ratio $\mathcal{R}(0)$. This difference is thus an indicator of the initial fraction of high-velocity FFPs.
The fitted parameters $\mathcal{R}_f(0)$ and $\alpha_f$ for the different models are plotted in Figure~{\ref{figure:memfit}}. The top panel shows the values for the models with $r_{\rm hm}=0.38$~pc and $f_{\rm ej}=P$ ($1M_\odot$). All models have $-1.4\times 10^{3} < \alpha_f < -0.5\times 10^{-3}$, which indicates that all planets are ejected from the clusters on timescales of roughly $0.7-2$~Gyr. For models with identical values of $N$ the scatter in $\mathcal{R}_f(0)$ is small and concentrates around the initial planet-to-star ratio $\mathcal{R}(0)$. The time scale $t_{\rm diss}$ at which all planets escape from the cluster increases with increasing $N_{\rm sb}$, which corresponds to the dependence of the relaxation time $t_{\rm hm}$ on $N_{\rm sb}$ (Eq.~\ref{eq:relaxationtime}). In addition, the results are (apart from statistical differences) independent of the number of FFPs, $N_{\rm p}$.
Figure~{\ref{figure:memfit}b} shows a comparison for clusters with identical $N_{\rm p}$ and $f_{\rm ej}$, but with a varying number of stars $N_{\rm sb}$ (different symbols) and half-mass radii $r_{\rm hm}$ (different colors). Initially, $\mathcal{R}(0)=N_{\rm p}/N_{\rm sb}$ varies between 0.33 and 2. The ejection timescale for the FFPs ranges from $t_{\rm diss} \approx 700$~Myr for $N_{\rm sb}=500$ to $t_{\rm diss}=2$~Gyr for $N_{\rm sb}=3000$. Larger values of $r_{\rm vir}$ result in a smaller fitted value $\mathcal{R}_f(0)$ as compared to the initial conditions, as more FFP speeds are initially above the escape velocity (Eq.~\ref{eq:escapevelocity}) and therefore escape at early times.
For the clusters with the largest number of particles (where most FFPs are initially below the escape velocity), $\alpha_f$ increases with increasing $N_{\rm sb}$, indicating that clusters with a smaller radius eject their planets faster, due to their shorter relaxation time.
Figure~\ref{figure:memfit}c shows the dependence of $\mathcal{R}_f(0)$ for star clusters with identical $r_{\rm hm}$ and $N_{\rm p}$, but with different numbers of stars $N_{\rm sb}=N(1+\mathcal{B})$ and ejection velocity distributions $f_{\rm ej}$. After this initial phase of escape of high-velocity FFPs, the ratio $\mathcal{R}(t)$ decreases roughly linearly with time. As this initial phase is short, and as $\alpha$ is independent of $N_{\rm p}$ beyond this time, the corresponding timescale for the escape of the entire planetary population is still well-approximated with $t_{\rm diss}=-\alpha_f^{-1}$, irrespective of the choice for $f_{\rm ej}$.
The time at which the planet-to-star ratio $\mathcal{R}(t)$ drops to half the value it had {\rm after the initial rapid escape phase}, is $t_{\rm diss}/2$.
We quantify this half-life time for the stellar population with $t_{\rm hs}$, and for the FFP population with $t_{\rm hp}$.
The relation between $t_{\rm hs}$ and $t_{\rm hp}$ for the different star clusters is shown in Figure~\ref{figure:thf}. From top to bottom, the three panels show the results for models with $r_{\rm hm}=0.38$~pc, 0.77~pc, and 1.54~pc, respectively. The initial number of planets in each cluster is $N_{\rm p}=1000$. Since planets escape faster due to their initially supervirial state and/or two-body relaxation, all star clusters have $t_{\rm hs} > t_{\rm hp}$, and both quantities are independent of $N_{\rm p}$. Note that the choice for $N_{\rm p}$ does not affect the results in the figure, as the planets effectively behave as test particles.
\subsection{Close encounters}
During the simulations we record all close ($< 1000$~AU) encounters between the members of the star cluster. Two-body encounters can be of the type star-star (S-S), star-planet (S-P), and planet-planet (P-P). The effect of gravitational focusing for S-S encounters is largest, while it is almost negligible for P-P. Three-body encounters occur regularly and almost exclusively involve stellar binary systems (SS-S and SS-P). Three-body encounters that result in the formation of a dynamical binary system almost always result in the ejection of the body with the lowest mass. The formation of an SS binary is therefore much more common than a binary of type SP (a planet orbiting a star), while those of type PP (binary planets) only rarely occur. Moreover, FFPs tend to escape earlier than stars. As a result, close encounters of type SP-S are more common than those of type SP-P, even when initially $N_{\rm s}<N_{\rm p}$. Three-body encounters involving more than one planet in a bound system (SP-P and PP-S) are rare, since this first requires the formation of these systems by dynamical capture. As initially all planets in the star clusters are free-floating\footnote{Close encounters involving star-planet system (SP) may be common in realistic star clusters as a large fraction of stars is likely to host one or more planets. In our simulations, however, all stars are initialised without planetary companions.}, the encounter frequencies for the interactions SS-SP, SS-PP, SP-PP and PP-PP are negligible, primarily because capture of FFPs into SP-type systems is rare, while double planets (PP) almost never form. The only four-body encounters that occur frequently are those between two stellar binary systems (SS-SS), particularly when primordial binaries are present. Despite these arguments, it should be noted that planetary systems may be common in real star clusters, and our results are therefore lower limits for the encounter frequencies of planet-hosting stars in open clusters.
The encounter frequencies at a given time can be estimated using the abundances of the different types of systems. Under the assumption that all particles have a similar distribution in phase-space, and ignoring the effects of gravitational focusing, the number of encounters $\mathcal{N}$ of the different types can be estimated through the following proportionalities:
\begin{equation} \label{eq:encounterrates}
\begin{array}{lll}
\mathcal{N}_{S-S} & \propto N_{\rm s}^2 & = (1-\mathcal{B})^2N^2\\
\mathcal{N}_{S-P} & \propto N_{\rm s}N_{\rm p} & = (1-\mathcal{B})NN_{\rm p} \\
\mathcal{N}_{P-P} & \propto N_{\rm p}^2 \\
\mathcal{N}_{SP-S} & \propto N_{\rm sb}N_{\rm s}N_{\rm p} & = (1-\mathcal{B}^2)N^2 N_{\rm p} \\
\mathcal{N}_{SS-S} & \propto N_{\rm s}N_{\rm b} & = (1-\mathcal{B})\mathcal{B} N^2 \\
\mathcal{N}_{SS-P} & \propto N_{\rm b}N_{\rm p} & = \mathcal{B} N N_{\rm p} \\
\mathcal{N}_{SS-SS} & \propto N_{\rm b}^2 & = \mathcal{B}^2 N^2 \quad\quad .\\
\end{array}
\end{equation}
In addition, all encounter rates depend on the choice of the encounter criterion $\delta r < p_m$ and the size of the star cluster $r_{\rm vir}$ through the approximate proportionality $\mathcal{N} \propto (p_m/r_{\rm vir}(t))^3$ at time $t$.
These estimates are valid for short periods of time where the populations do not change substantially though the formation or destruction of gravitationally bound two-body systems or dynamical mass segregation.
\subsubsection{Encounter rates for dynamical populations} \label{section:encounterrates}
\begin{table*}
\caption{The frequency of the various types of encounters during the lifetime of a subset of the star clusters. All values and their errors represent the average and the standard deviation for an ensemble of ten realisations. Several types of encounters are very rare (such as SS-SP), and are therefore not listed here. \label{table:encounterrates} }
\begin{tabular}{p{0.0mm}|p{3mm}|p{0mm}|p{3mm}|p{2mm}|p{2mm}|p{2mm}|c|c|c|c|c|c|c|}
\hline\hline
ID & $N$&$\mathcal{B}$&$N_{\rm p}$&$r_{\rm vir}$&$M_{s}$&$f_{\rm ej}$&S-S&S-P&P-P&SS-S&SP-S&SS-P&SS-SS\\
& & \% & & pc &$M_\odot$& & (1) & (2) & (3) & (4) & (5) & (6) & (10) \\
\hline
1 & 500& 0&1000&0.5&1.0&P&$872\pm119$&$357\pm183$&$77\pm83$&$93\pm15$&$1\pm1$&$7\pm5$&$2\pm2$\\
2 &1000& 0&2000&0.5&1.0&P&$4374\pm461$&$2590\pm1139$&$656\pm648$&$254\pm31$&$2\pm2$&$17\pm8$&$7\pm3$\\
3 &2000& 0&4000&0.5&1.0&P&$22880\pm2105$&$20316\pm6745$&$6585\pm5374$&$707\pm87$&$7\pm5$&$57\pm17$&$13\pm4$\\
\hline
4 &2000&20&4000&0.5&1.0&P&$26265\pm2044$&$22959\pm7741$&$7700\pm5901$&$7599\pm796$&$29\pm13$&$2848\pm956$&$587\pm100$\\
5 &2000&50&4000&0.5&1.0&P&$33814\pm3462$&$28254\pm7824$&$9523\pm6539$&$18817\pm1555$&$90\pm23$&$7472\pm2144$&$2715\pm277$\\
\hline
6 &2000& 0&1000&0.5&1.0&P&$21931\pm1350$&$4723\pm1450$&$377\pm287$&$726\pm72$&$1\pm1$&$16\pm7$&$13\pm4$\\
7 &2000& 0&2000&0.5&1.0&P&$22458\pm2016$&$9796\pm3265$&$1602\pm1296$&$736\pm86$&$3\pm2$&$33\pm16$&$11\pm6$\\
\hline
8 &2000& 0&4000&1.0&1.0&P&$8190\pm745$&$4548\pm2073$&$1169\pm1232$&$457\pm56$&$1\pm1$&$28\pm9$&$14\pm6$\\
9 &2000& 0&4000&2.0&1.0&P&$3286\pm385$&$1051\pm593$&$190\pm241$&$255\pm34$&$1\pm1$&$9\pm9$&$7\pm3$\\
\hline
10&2000& 0&4000&0.5& -- &Z&$21916\pm1826$&$42707\pm2314$&$27029\pm1348$&$695\pm78$&$15\pm5$&$139\pm24$&$11\pm4$\\
11&2000& 0&4000&0.5&1.0&D&$21875\pm1670$&$36927\pm7276$&$20778\pm5758$&$694\pm77$&$10\pm7$&$119\pm29$&$13\pm4$\\
12&2000& 0&4000&0.5&0.6&D&$22070\pm2225$&$36351\pm3993$&$19765\pm3203$&$726\pm51$&$12\pm6$&$126\pm17$&$12\pm4$\\
13&2000& 0&4000&0.5&0.6&P&$19951\pm7343$&$25574\pm9194$&$10612\pm3806$&$627\pm227$&$8\pm5$&$98\pm36$&$9\pm5$\\
\hline\hline
\end{tabular}
\end{table*}
The frequency of occurrence of the various types of encounters for a subset of the modelled star clusters with different initial conditions are listed in Table~\ref{table:encounterrates}. The quantities represent averaged results for an ensemble of ten realisations of each model. Note these are cumulative values over the entire lifetime of the star clusters. The relations in Eq.~\ref{eq:encounterrates} therefore only provide first-order estimates. Model~3 is the reference model that is used for comparison with models that have different initial conditions.
Models $1-3$ represent the results for star clusters with different $N$ and identical initial planet-to-star ratios, and shows that the number of encounters $\mathcal{N}$ grows faster than the number of particles in the star cluster. This is because these three clusters have the same $r_{\rm hm}$ and therefore a stellar density and a total lifetime that increase with $N$. For the high-density star clusters the ratio $\mathcal{N}_{P-P}/\mathcal{N}_{S-S}$ is substantially larger than for low-density star clusters. The reason for this is that the ejection velocity for the planets is identical for all models, and therefore planets in the low-mass star clusters are supervirial and escape at earlier times. Although the initial binary fraction is zero for these three models, several binary systems (SS and SP) form and have close encounters with the other members of the star clusters.
Models $3-5$ show the results for models with an identical $N=N_{\rm s}+N_{\rm b}$ and $N_{\rm p}$, but with different binary fractions $\mathcal{B}$. The total number of individual stars in the clusters increases as $N_{\rm sb} = N_{\rm s}+2N_{\rm b} = N(1+\mathcal{B})$, and the total cluster mass follows the same proportionality. The number of S-S and S-P encounters initially decreases with increasing $\mathcal{B}$, but due to higher mass density, the destruction of binary systems and the longer lifetime of star clusters with a higher $\mathcal{B}$, the cumulative number of these types of encounters increases with increasing $\mathcal{B}$. For similar reasons, the P-P encounter rate, which is initially independent of $\mathcal{B}$, increases with $\mathcal{B}$ due to dependence of the dissolution timescale on $\mathcal{B}$. The total number of SS-S, SS-P, and SS-SS encounters increase strongly with $\mathcal{B}$, although the occasional formation and destruction of binary systems changes the proportionality of Eq.~\ref{eq:encounterrates} mildly.
The dependence of the the encounter frequencies on the number of planets $N_{\rm p}$ can by studied by comparing models~3, 6, and 7. As expected, the encounters involving only stars (S-S, SS-S, SS-SS) are independent of $N_{\rm p}$. The number of encounters between stars and planets (S-P, SS-P) are proportional to $N_{\rm p}$, while for the planet-planet encounters (P-P) the proportionality is $N_{\rm p}^2$. The number of SP-S encounters is well described by Eq.~\ref{eq:encounterrates} with proportionality constant $7.5\times 10^{-10}$.
Models 3, 8 and 9 represent identical star clusters apart from their initial size $r_{\rm vir}$. Since the initial virial radius of model~8 is twice as large as as that of model~3, we expect all encounters involving only stars (S-S, SS-S, SS-SS) to occur roughly eight times less frequently. The number of encounters between FFPs occur even less frequently because the ratio between the planet ejection velocities and the stellar velocity dispersion becomes larger as $r_{\rm vir}$ increases, resulting in the initial escape of a larger number of FFPs.
Models 3 and $10-13$ represent five star clusters with (statistically) identical stellar populations, but with different planet ejection velocity distributions $f_{\rm ej}$. Model~10 represents the special case where all planets are assumed to be ejected from their host star with zero velocity, i.e., the FFP population in this model is initially in virial equilibrium. As expected, the number of encounters of types S-S, SS-S, SS-SS are statistically identical. All encounters involving FFPs depend on the ability of the star cluster to retain the escaping planetary population. The number of encounters involving FFPs (S-P, P-P, SS-P) increases when the clusters can retain the FFPs for longer times, such as when the typical ejection velocity (Figure~\ref{figure:ejectionvelocities}) is lower. The number of FFPs that are captured into SP-systems is largest when the ejection velocities are smallest. Models in which the FFPs have lower initial velocities therefore result in a larger number of encounters of type SP-S.
\subsubsection{The encounter number distribution}
\begin{figure*}
\centering
\begin{tabular}{p{0.5\textwidth}p{0.5\textwidth}}
\includegraphics[width=0.5\textwidth]{ps/2000_05_0_2v4_m1p2_counts.pdf}&
\includegraphics[width=0.5\textwidth]{ps/2000_05_0_2v4_m1p2p_counts.pdf}\\
\includegraphics[width=0.5\textwidth]{ps/2000_05_05_2v4_m1p2_counts.pdf}&
\includegraphics[width=0.5\textwidth]{ps/2000_05_05_2v4_m1p2p_counts.pdf}\\
\end{tabular}
\caption{The distribution of the total number of close ($<1000$~AU) encounters per star for model~3 ({\em top}) and model~5 ({\em bottom}) during the entire lifetime of the star clusters. Both star cluster models initially contain $N=N_{\rm s}+N_{\rm b}=2000$ (single or binary) stellar members with a binary fraction $\mathcal{B}=0\%$ (model~3) or $\mathcal{B}=50\%$ (model~5). Both have an initial planet-to-star ratio $\mathcal{R}=2$ and an initial half-mass radius $r_{\rm hm}=0.38$~pc. Encounter distributions are shown for encounters experienced by single stars ({\em left}) and by FFPs ({\em right}). The different colours indicate different types of encounters. The horizontal axis indicates how many encounters are experienced by a cluster member, and the vertical axis represents the number of cluster members that experience that number of encounters.
Each of the histograms represents the average of an ensemble of ten realisations.
\label{figure:ecrate} }
\end{figure*}
\begin{figure}
\includegraphics[width=0.53\textwidth]{ps/np_1000_frac0c_P_Mall.pdf}
\caption{
The number of FFPs that never approach a star or a binary system within 1000~AU during the entire lifetime of the star cluster. The horizontal axis represents the different models, arranged by the initial number density in the star cluster. Each model initially contains $N_{\rm p}=1000$ free-floating planets. The adopted initial planet ejection velocity distributions include prompt ejection ($P$), delayed ejection ($D$), and zero ejection velocity ($Z$), and are indicated with the different symbols (see Figure~\ref{figure:ejectionvelocities}). FFPs our lowest-density models ($N_{\rm sb}=N_{\rm s}+N_{\rm b}=500$; $\mathcal{B}=0\%$; $r_{\rm vir}=2$~pc) almost never encounter a star or binary system, while many FFPs in the highest-density clusters ($N_{\rm sb}=2000$; $\mathcal{B}=50\%$; $r_{\rm vir}=0.5$~pc) do approach one or more massive cluster members within 1000~AU. In the latter clusters, the FFPs almost all FFPs experience a close encounter for the model in which the planet ejection velocity is zero ($f_{\rm ej}=Z$).
\label{figure:ffp_with_massive}}
\end{figure}
In our simulations we also record the total number of close encounters experienced by each individual object. This number depends strongly on the location of the object in the star cluster and on the time it spends in the cluster before escaping. The (combined) total number of encounters experienced by all the objects of type S, P, SS, and SP is:
\begin{equation} \label{eq:totalencounters}
\begin{array}{ll}
\mathcal{N}_{\rm S,tot} &= 2\mathcal{N}_{S-S} + \mathcal{N}_{S-P} + \mathcal{N}_{SP-S} + \mathcal{N}_{SS-S} \\
\mathcal{N}_{\rm P,tot} &= \mathcal{N}_{S-P} + 2\mathcal{N}_{P-P} + \mathcal{N}_{SS-P}\\
\mathcal{N}_{\rm SS,tot} &= \mathcal{N}_{SS-S} + \mathcal{N}_{SS-P} + 2\mathcal{N}_{SS-SS} \\
\mathcal{N}_{\rm SP,tot} &= \mathcal{N}_{SP-S} \quad\quad \\
\end{array}
\end{equation}
where we have ignored the encounter types that rarely occur, such as $\mathcal{N}_{SP-P}$, $\mathcal{N}_{SP-SP}$ and $\mathcal{N}_{SS-SP}$ (see also below). Note that during an encounter between two objects of the same type, both objects experience an encounter. Using Eqs.~\ref{eq:encounterrates}, and under the assumption that gravitational focusing is negligible (which may only be a reasonable assumption for encounters involving low-mass bodies), these can be expressed as
\begin{equation} \label{eq:totalencounterssimpler}
\begin{array}{ll}
\mathcal{N}_{\rm S,tot} &\propto N^2(1-\mathcal{B})[2-\mathcal{B}+N_{\rm p}(N^{-1}+1+\mathcal{B})] \\
\mathcal{N}_{\rm P,tot} &\propto N_{\rm p} (2N_{\rm p}+N) \\
\mathcal{N}_{\rm SS,tot} &\propto N^2\mathcal{B}(1+N_{\rm p}/N)\\
\mathcal{N}_{\rm SP,tot} &\propto (1-\mathcal{B}^2)N^2N_{\rm p}\\
\end{array}
\end{equation}
at an instantaneous time $t$, under the assumption that the different populations have the same distribution in phase-space. In the remaining part of this section we discuss the encounters experienced during the entire lifetime of the star clusters, and since all variables in Eq.~\ref{eq:totalencounterssimpler} change with time, these expressions only provide first-order approximations.
The distributions of the number of close encounters per star and per planet, split up into the different components in Eq.~\ref{eq:totalencounters}, are shown in Figure~\ref{figure:ecrate}, for models~3 and~5 (cf. Table~\ref{table:encounterrates}). The left-hand and right-hand panels show the encounters experienced by the stars and by the planets, respectively. Note that in this figure we show the number of encounters experienced by each body (instead of the number of two-body encounters occurring); each encounter corresponds to two individual bodies experiencing an encounter. The area under each histogram corresponds to the values in Table~\ref{table:encounterrates} when the encountering cluster members are of different type (e.g., S-P), and are double the values in Table~\ref{table:encounterrates} when the encountering members are of the same type (e.g., P-P).
The top-left panel of Figure~\ref{figure:ecrate} shows the distribution of the number of encounters experienced by single stars in model~3. Based on our initial conditions, we expect to first order that the total number of encounters of each type is $2\mathcal{N}_{S-S} \propto 2N_{\rm s}$, $\mathcal{N}_{S-P}\propto N_{\rm p}$, and $\mathcal{N}_{SS-S}\approx 0$, for encounters with other single stars, with FFPs, and with binary systems, respectively. Note that for our choice of the initial conditions, we expect the distributions for the star-star and star-planet encounters to be roughly the same. However, due to the large fraction of FFP escapers at the early stage, the evolution of the star cluster, the formation of new dynamical binaries, and gravitational focusing, the final results differ. Approximately 70~stars (3.5\% of the total number of single stars) experience only one encounter with another star, and a slightly higher number of single stars experiences only one encounter with a FFP. Several single stars, on the other hand, experience over a hundred encounters with other stars, while none of the single stars encounters more than a hundred FFPs. Close encounters of single stars with binary systems are substantially less frequent, as the initial binary fraction is zero and only few binary systems are formed at later times. Approximately 3.5\% of the single stars experience close encounters with one or more dynamically formed binary systems, and most of these occur after the clusters have over half of their stars.
For model~5 (bottom-left panel in Figure~\ref{figure:ecrate}), the number of single stars with a close encounter with one other single star is about 100 (10\%). Although one would expect this number to be lower due to the smaller number of single stars in this model, this value can be explained as a result of the higher mass density in the star cluster, the disruption of wide binaries, particularly in the cluster centre \citep[see, e.g.,][]{grijs2013}, and the longer cluster lifetime, as compared to model~3. Approximately 130 single stars (13\%) experience a close encounter with a FFP. The number of single stars experiencing a close encounter with a binary system is now substantial, due to the initial condition $\mathcal{B}=50\%$. Single stars having many encounters with binary systems occur less frequently compared to single-single encounters, which is partially due to the disruption of the widest binaries over time, and partially due to the fact that binary systems tend to increase the velocities of single stars after an encounter, such that they are more likely to escape the star cluster.
The distribution of close encounters experienced by the planets in models~3 and~5 are shown in the right-hand panels of Figure~\ref{figure:ecrate}. For model~3, approximately 400 (10\%) of the FFPs experience encounters with only one FFP, and a lower fraction (5\%) of the FFPs have only one encounter with a star. All FFPs in the cluster experience fewer than 60 encounters with another star, and fewer than 30 encounters with another FFP. FFPs experience on average more encounters with single stars and with FFPs in model~5 (w.r.t. model~3), since the number density for this model is larger. Again the encounter distributions for encounters with single stars and with FFPs are different. The difference in S-P and P-P encounters between the two models is partially a result of the higher mass concentration, and partially a result of the longer lifetime of model~5. Close encounters between stars and binaries do not occur frequently in model~3 because of the zero initial binary fraction, but as dynamical binaries form, several hundreds of FFPs have a close encounter with a binary system. Encounters of the type SS-P are more frequent in model~5, and the majority of the FFPs have one or more encounters with a binary system. Although the initial binary fraction in model~5 is 50\%, fewer FFPs experience multiple close encounters with binaries than with single stars.
The encounter distributions for the different types of cluster members in all models have in common that few bodies experience many encounters, while many bodies experience few (or no) encounters. The former are mostly bodies that remain part of the star clusters until it is nearly dissolved, while the latter escape or migrate the the cluster outskirts at early times. This means that statistical averages, such as the results in Eqs.~\ref{eq:totalencounters} and~\ref{eq:totalencounterssimpler} normalised by the number of bodies, should be avoided when estimating, for example, the typical effect of close encounters on existing planetary systems.
The fraction of FFPs that never experience a close encounter with a stellar-mass particle (single or binary) is shown in Figure~\ref{figure:ffp_with_massive} for the different initial conditions. Note that the initial number density along the horizontal axis is proportional to the initial mass density, as we adopt the same IMF in all models. Thus this result also represent the correlation between FFPs with no encounters and cluster initial mass density. In the lowest-density star clusters none of the FFPs experience a close encounter, while in the higher-density star clusters the fraction depends strongly on the initial velocity distribution of the FFPs. The FFP populations that are initially in virial equilibrium (the filled circles in Figure~\ref{figure:ffp_with_massive}) provide lower limits, and for these populations the fraction of non-encountering FFPs ranges from 0\% to 100\%, primarily depending on the star cluster density.
\subsubsection{Encounter times} \label{section:encountertime}
\begin{figure}
\centering
\begin{tabular}{c}
\includegraphics[width=0.53\textwidth]{ps/np_1000_tf_10all.pdf}\\
\includegraphics[width=0.53\textwidth]{ps/np_1000_tf_50all.pdf}\\
\includegraphics[width=0.53\textwidth]{ps/np_1000_tf_90all.pdf}\\
\end{tabular}
\caption{
The times at which 10\% ({\em top}), 50\% ({\em middle}), and 90\% ({\em bottom}) of the encounters have occurred, for all models. The horizontal axis represents the initial number density in the star cluster, from the lowest-density models ($N_{\rm sb}=N_{\rm s}+N_{\rm b}=500$; $\mathcal{B}=0\%$; $r_{\rm vir}=2$~pc) on the left to the highest-density clusters ($N_{\rm sb}=2000$; $\mathcal{B}=50\%$; $r_{\rm vir}=0.5$~pc) on the right. The vertical axis represents the time at which a certain fraction of encounters have occurred (cf. Eq.~\ref{eq:specifichalfencounters}).
Different colours represent different encounter types between stars (S), FFPs (P), binary systems (SS) and captured planetary systems (SP). The adopted initial planet ejection velocity distributions include prompt ejection ($P$), delayed ejection ($D$), and zero ejection velocity ($Z$), and are indicated with the different markers (see also Figure~\ref{figure:ejectionvelocities}). The data are derived from an ensemble of ten model realisations.
\label{figure:tfit}}
\end{figure}
The cumulative number of encounters $\mathcal{N}_{X-Y}(t)$ between members of two types of populations $X$ and $Y$ should follow
\begin{equation} \label{eq:generalhalfencounters}
\mathcal{N}_{X-Y} \propto \int_0^{t_{\rm diss}} N_X(t)N_Y(t)dt
\end{equation}
where $N_X$ and $N_Y$ are the number densities of these populations, respectively. When making the assumption that the proportionality remains constant (i.e., when we ignore mass loss, mass segregation, core collapse, and cluster expansion), the time $t_h$ at which half of the encounters have occurred can be obtained by solving $\mathcal{N}_{X-Y}(t_h)=\mathcal{N}_{X-Y}(t_{\rm diss})/2$ for $t_h$. Approximate values for $t_h$ for the various types of encounters can be derived by substituting Eqs.~\ref{eq:starloss} and/or~\ref{eq:solution} into Eq.~\ref{eq:generalhalfencounters}, giving
\begin{equation} \label{eq:specifichalfencounters}
\begin{array}{lll}
t_{h,S-S} & \approx (1-2^{-1/3})t_{\rm diss} & \approx 0.21t_{\rm diss} \\
t_{h,S-P} & \approx (1-2^{-1/4})t_{\rm diss} & \approx 0.16t_{\rm diss} \\
t_{h,P-P} & \approx (1-2^{-1/5})t_{\rm diss} & \approx 0.13t_{\rm diss}\\
\end{array}
\end{equation}
where we have for simplicity assumed that the FFP population is initially virialized, that $\mathcal{B}=0$, and that Eq.~\ref{eq:starloss} holds . The expressions can be generalized for the cases with $N_{\rm p}(t_1)<N_{\rm p}(0)$ and $t_1>0$ using Eq.~\ref{eq:ratiostarloss}, and will result in smaller values of $\mathcal{N}(t_h)$ for the encounters involving planets. Note that Eq.~\ref{eq:specifichalfencounters} is independent of the number of stars and the number of planets, if the assumptions stated above hold.
The timescales at which $10\%$, $50\%$ and $90\%$ of the encounters of the specific types occur, are shown in Figure~\ref{figure:tfit} for all modelled star clusters. Note that we model each star cluster until it has completely dissolved. The times at which these events occur vary with the initial properties of the star clusters and their FFP populations, and also with the types of encounters. In general, however, $10\%$ of the encounters occur in the first 3~Myr, $50\%$ occur within 30~Myr, and 90\% of the encounters occur within 100~Myr.
The middle panel shows the times at which half of the encounters have occurred, and its values can be approximated with Eq.~\ref{eq:specifichalfencounters}. Although $t_{\rm diss}$ depends on the star cluster properties (particularly on $r_{\rm vir}$, $N_{\rm sb}$ and $\mathcal{B}$), the ratios $t_{h,S-S}/t_{h,S-P}\approx 1.31$ and $t_{h,S-S}/t_{h,P-P}\approx 1.62$ are independent of these, which explains why the differences in $t_h$ in Figure~\ref{figure:tfit} are roughly constant in logarithmic units. The real differences in the figure are slightly larger than in Eq.~\ref{eq:specifichalfencounters} because several of the assumptions are violated, as a result of expansion and early loss of FFPs in the systems. This can be observed in the top panel, which shows that the 10\% encounter time of the P-P interactions is substantially smaller than the simple approximation suggests for the models with high planet ejection velocity. For these cases, a better expression may be derived by not adopting $N_{\rm p}(t_1)=N_{\rm p}(0)$ and $t_1=0$ in Eq.~\ref{eq:solution}, as done above. However, even though a large number of planets may escape at early times, the majority of the P-P encounters occur between the FFPs that remain in the star clusters for a long time, and therefore the results in the middle and bottom panels depend only mildly on $f_{\rm ej}$. The encounters involving massive components (such as SS-S and SS-SS) generally occur at later times than those with low-mass components, as the former remain part of the star cluster for longer times. Figure~\ref{figure:tfit} also demonstrates that encounters occur at later times for systems with increasing $N_{\rm sb}$, $\mathcal{B}$, or $r_{\rm vir}$, since these systems have larger relaxation times (Eq.~\ref{eq:relaxationtime}) and generally have longer dissolution times.
\subsection{Encounter velocities and periastron distances}
\subsubsection{Encounter velocity distributions}
\begin{figure*}
\begin{tabular}{cc}
\includegraphics[width=0.5\textwidth]{ps/2000_05_0_2v4_m1p2_vinf.pdf} &
\includegraphics[width=0.5\textwidth]{ps/2000_05_05_2v4_m1p2_vinf.pdf}\\
(a) $N=2000$, $\mathcal{B}=0\%$~$\mathcal{R}=2$, $r_{\rm hm}=0.38$~pc&
(b) $N=2000$, $\mathcal{B}=50\%$~$\mathcal{R}=2$, $r_{\rm hm}=0.38$~pc\\
\includegraphics[width=0.5\textwidth]{ps/2000_05_0_2v4_m1p2_vp.pdf} &
\includegraphics[width=0.5\textwidth]{ps/2000_05_05_2v4_m1p2_vp.pdf}\\
(c) $N=2000$, $\mathcal{B}=0\%$, $\mathcal{R}=2$, $r_{\rm hm}=0.38$~pc &
(d) $N=2000$, $\mathcal{B}=50\%$, $\mathcal{R}=2$, $r_{\rm hm}=0.38$~pc\\
\end{tabular}
\caption{The velocity-at-infinity ($v_\infty$) distributions of the close encounters, for model~3 ({\em top-left}) and model~5 ({\em top-right}), and the periastron velocity ($v_p$) distributions of the close encounters for model~3 ({\em bottom-left}) and model~5 ({\em bottom-right}).
The histograms represent the simulation results, and the curves are best fits to Eq.~\ref{eq:maxwellian} for $v_\infty$ and to Eq.~\ref{eq:maxwellianoffset} for $v_p$.
The data represent the combined results for ten realisations.
\label{figure:incomingvelocity} }
\end{figure*}
The trajectories of two approaching unbound bodies in isolation are hyperbolic, and have an associated velocity-at-infinity $v_\infty$. The velocity-at-infinity distribution, $f(v_\infty)$, of an ensemble of two-body encounters in a star cluster is approximately a Maxwell-Boltzmann distribution:
\begin{equation} \label{eq:maxwellian}
f(v_\infty)
= \left(\pi v_m^2\right)^{-3/2}
\exp \left(-\frac{v_\infty^2}{v_m^2} \right) 4\piv_\infty^2
\end{equation}
where $v_m$ is the most probable value of $v_\infty$. The corresponding mean value is $\langle v_\infty \rangle = 2\pi^{-1/2}v_m$. Eq.~\ref{eq:maxwellian} is generally a good approximation and is independent of mass of the bodies involved in the encounter, provided that the cluster is not mass segregated and that the star cluster's velocity distribution is isotropic. For an encounter with impact parameter $b$, the velocity at periastron can be expressed as
\begin{equation}
v_p = v_\infty \sqrt{ \frac{e+1}{e-1} }
\quad\quad {\rm with} \quad\quad
e = \sqrt{1+\left(\frac{bv_\infty^2}{GM}\right)^2} \ ,
\end{equation}
where $e$ is the hyperbolic eccentricity. The distribution of impact parameters follows $f(b)\propto b$. When the combined mass of the bodies is small, such as for P-P encounters, the equation reduces to $v_p=v_\infty$. For a close, low-$v_\infty$ encounter between massive bodies, the expression reduces to $v_p\approx 2GM/bv_\infty$.
We approximate the distribution of the velocities at periastron, $f(v_p)$ with a modified Maxwell-Boltzmann distribution that has a velocity offset $v_o$ and a dispersion parameter $v_m$:
\begin{equation} \label{eq:maxwellianoffset}
f(v_p)
\propto \left(\pi v_m^2\right)^{-3/2}
\exp \left(-\frac{(v_p-v_o)^2}{v_m^2} \right) 4\pi(v_p-v_o)^2 \ ,
\end{equation}
where $v_o=v_p-v_\infty > 0$ is the velocity increase during the encounter and $v_m$ is the most probable periastron velocity.
The histograms in Figure~\ref{figure:incomingvelocity} show the distributions of $v_\infty$ ({\em top}) and $v_p$ ({\em bottom}) for the various types of encounters that occur in star cluster models~3 ({\em left-hand panels}) and~5 ({\em right-hand panels}), and the corresponding best fits to Eqs.~\ref{eq:maxwellian} and~\ref{eq:maxwellianoffset}. The curves represent the distributions integrated over all locations in the star clusters, for their entire evolution. Since the encounter velocities depend both on the location within the star cluster and on time, Eqs.~\ref{eq:maxwellian} and~\ref{eq:maxwellianoffset} are approximations. Nevertheless, since most encounters occur within the first $\sim 100-200$~Myr, the approximations are reasonable.
The strongest deviations from the Maxwellian distribution are seen for the largest values of $v_\infty$ and $v_p$, indicating interactions in the centre of the star cluster, and for the smallest values of $v_\infty$ and $v_p$, indicating encounters in the cluster outskirts and during the latest phases of the evolution when the stellar density is low.
Since planetary systems have $v_p\approxv_\infty$, the corresponding curves for the P-P encounters are very similar in all four panels, and well-fitted by Eq.~\ref{eq:maxwellian}. All encounters involving planets (P-P, S-P, SS-P, and P-P) are qualitatively very similar, both for model~3 and model~5. As a result of gravitational focusing, the typical velocity at periastron increases as the total mass of the encounters increases.
With respect to the Maxwellian fit, excess S-S interactions with small $v_\infty$ occur. These interactions are most frequent in the low-density regions in the outskirts of the clusters and also at late stages of evolution, when the effect of gravitational focusing on low-velocity stars is more prominent. The deviation from a Maxwellian distribution is much stronger for the SS-S interactions. Since model~3 does not have primordial binaries, all binaries involved in the SS-S encounters are formed through capture. As these binaries are generally formed in the cluster centre, and as their components are generally of high mass, Eq.~\ref{eq:maxwellian} does not provide a good approximation for the distribution. Since $\mathcal{B}=50\%$ for model~5, the effect of gravitational focusing and longevity of the binary population is now much better visible. The SS-S and SS-SS show deviations from the Maxwell-Bolzmann distributions, because of gravitational focusing and because they tend to remain longer in the star cluster.
Note that the dynamically-formed binary systems in model~3 have on average substantially larger $v_p$ than the primordial binaries in model~5, although they are fewer in number. As these dynamical binaries in model~3 are generally more massive than the primordial binaries in model~5, their encounters with other stars and FFPs are stronger, and therefore their periastron velocities are larger.
There are about ten times more SP-S encounters in model~5 than in model~3, indicating that the presence of binary stars in star clusters enhances the star-planet captures. The distribution $f(v_\infty)$ for SP-S encounters is a scaled-down version of that of the S-S encounters, which may be expected since the planetary companions do not play a role in the encounter trajectory or encounter rates.
The best-fitting values of $v_m$ and $v_o$ in Eqs.~\ref{eq:maxwellian} and~\ref{eq:maxwellianoffset} are plotted in Figure~\ref{figure:vpfit} for the different types of encounters, for all simulated models. Note, however, that we did not include the bad fits, which result either from distributions that are substantially different from Eq.~\ref{eq:maxwellian}, or from low-number statistics.
Most encounters occur during the early phase of star cluster evolution when mass segregation is small, and during this period we expect $v_m$ to be directly proportional to the velocity dispersion of the star cluster (Eq.~\ref{eq:velocitydispersion}). For star cluster members near the half-mass radius $v_m \propto \sigma(r=r_{\rm hm}) \propto (N_{\rm sb}/r_{\rm hm})^{1/2}$. This trend is clearly seen for most fits to $f(v_\infty)$ plotted in Figure~\ref{figure:vpfit}, and is independent of $f_{\rm ej}$, $N_{\rm p}$, and the type of encounters. Since the P-P encounters have $v_p \approx v_\infty$, we obtain the same relation for $f(v_p)$. For the encounters between massive components the fitted values for $v_m$ are larger, and $v_m$ typically increases with increasing component masses.
Since FFPs barely change their velocities during P-P encounters, we expect $v_o\approx 0$. The scatter for the P-P interactions results from the fact that we have fitted the distributions $f(v_\infty)$ and $f(v_p)$ to all encounters integrated over the entire volume of the star cluster and its entire evolution. The value of $v_o$ depends only mildly on $\mathcal{B}$, $N_{\rm sb}$, $f_{\rm ej}$ and $r_{\rm hm}$. As the fitted value $v_m$ takes into account the dependency of $f(v_p)$ on $v_\infty$, the value of $v_o$ is primarily determined by the mass of the encountering components. Since gravitational focusing is stronger for encounters between massive bodies, the fitted value $v_o=v_p-v_\infty$ increases with the combined mass of the members involved in the encounter.
The fitting results for $f(v_p)$ are combined in Figure~\ref{figure:vpvcfit}. The value of $v_m$ is primarily determined by the velocity dispersion of the star cluster, while $v_o$ is mostly determined by combined mass of the encountering bodies. These trends are seen in Figure~\ref{figure:vpvcfit}: models with identical velocity dispersion lie roughly on vertical lines, while the same types of encounters (S-S, S-P, etc) lie roughly on a horizontal line. Apart from some scatter, the values of $v_o$ of most of the encounter types indeed lie roughly on a horizontal line in Figure~\ref{figure:vpfit}, although there is a trend of $v_o$ increasing with $v_m$ (and therefore with increasing $N_{\rm sb}$). This trend is strongest for the P-P encounters.
The results in Figure~\ref{figure:vpvcfit} can be used to estimate the periastron velocity distributions for other star cluster masses and for encounters between members with other masses, using Eq.~\ref{eq:maxwellianoffset}. It can also be used to statistically estimate the probability of obtaining a certain periastron velocity for a single free-floating planet in a star cluster.
\begin{figure}
\centering
\includegraphics[width=0.53\textwidth]{ps/np_1000_vinfmax.pdf}\\
\includegraphics[width=0.53\textwidth]{ps/np_1000_vpmax.pdf}\\
\includegraphics[width=0.53\textwidth]{ps/np_1000_vcut.pdf}\\[-2ex]
\caption{Properties of the velocity-at-infinity distributions $f(v_\infty)$ and periastron velocity distributions $f(v_p)$ for the different types of encounters for the modelled set of star clusters. The horizontal axis represents the initial stellar density, as in Figure~\ref{figure:tfit}. {\em Top:} the fitted parameter $v_m$ (Eq.~\ref{eq:maxwellian}) to the measured distribution $f(v_\infty)$. The two other panels show the best-fitting parameters $v_m$ ({\em middle}) and $v_o$ ({\em bottom}) in Eq.~\ref{eq:maxwellianoffset} for the measured distribution $f(v_p)$.
All models resemble model~3 (cf. Figure~\ref{figure:incomingvelocity}). Different colours represent different encounter types. Data with fitting errors larger than 0.5 are omitted. Different markers indicate the initial planet ejection velocity distributions, including prompt ejection ($P$), delayed ejection ($D$), and zero ejection velocity ($Z$); cf. Figure~\ref{figure:ejectionvelocities}). Data are derived from an ensemble of ten model realisations.\label{figure:vpfit}}
\end{figure}
\begin{figure}
\centering
\includegraphics[width=0.53\textwidth]{ps/rbar_05_vpvcnp1000m1p2.pdf}\\
\caption{Properties of the periastron velocity distributions $f(v_p)$ for the different types of encounters for the modelled set of star clusters. The fitted parameters $v_m$ and $v_o$ to Eq.~\ref{eq:maxwellianoffset} are indicated along the horizontal and vertical axes, respectively, for the models with $r_{\rm hm}=0.38$~pc, $N_{\rm p}=1000$ and with ejection velocity distributions corresponding to the prompt-$1M_\odot$ distribution in Figure~\ref{figure:ejectionvelocities}. The different markers indicate the different types of encounters between stars (S), FFPs (P), binary systems (SS), and captured planetary systems (SP). The values of $N_{\rm sb}=N_{\rm s}+N_{\rm b}$ are indicated with the different colors. \label{figure:vpvcfit}}
\end{figure}
\subsubsection{Encounter periastron distance distribution} \label{section:encounterperiastron}
\begin{figure}
\centering
\includegraphics[width=0.53\textwidth]{ps/2000_05_0_2v4_m1p2_p.pdf}\\[-47ex]
\quad\quad\quad\quad(a) $N=2000$, $\mathcal{B}=0\%$, $\mathcal{R}=2$, $r_{\rm hm}=0.38$~pc\\[45ex]
\includegraphics[width=0.53\textwidth]{ps/2000_05_05_2v4_m1p2_p.pdf}\\[-47ex]
\quad\quad\quad\quad(b) $N=2000$, $\mathcal{B}=50\%$, $\mathcal{R}=2$, $r_{\rm hm}=0.38$~pc \\[45ex]
\caption{The periastron distance distributions $f(p)$ of the close
encounters for model~3 ({\em top}) and model~5 ({\em bottom}). Different colours represent different types of encounters between stars (S), FFPs (P), binary systems (SS) and captured planetary systems (SP). The solid histograms indicate the number of close encounters in each bin, and the dashed lines are linear fits. Encounter types that occur less than 200 times (e.g., SS-SS) are omitted. The curves represent the combined results for ten realisations. \label{figure:periastron} }
\end{figure}
\begin{figure}
\centering
\includegraphics[width=0.53\textwidth]{ps/np_1000_pslope_norm_rhoall.pdf}\\
\caption{The fitted proportionality parameter $\alpha$ (Eq.~\ref{eq:periastrondistribution}) for the periastron distance distributions $f(p)$ for the different star clusters. The horizontal axis represents the initial stellar density, as in Figure~\ref{figure:tfit}. The adopted initial planet ejection velocity distributions are indicated with the different markers, and include prompt ejection ($P$), delayed ejection ($D$), and zero ejection velocity ($Z$); see Figure~\ref{figure:ejectionvelocities}. The different types of encounters between stars (S), FFPs (P), binary systems (SS) and captured planetary systems (SP) are indicated with the colors. Results are shown for a combined ensemble of ten realisations of each model. \label{figure:pfit}}
\end{figure}
For hyperbolic orbits between two objects with masses $M_1$ and $M_2$, the
relation between the impact parameter $b$, the periastron distance $p$, and
the velocity at infinity $v_\infty$ is given by:
\begin{equation}
b = p \sqrt{1 + \frac{2GM}{p v_\infty^2} } \ ,
\end{equation}
where $M=M_1+M_2$ is the combined mass of the two objects. For two approaching bodies in a virialized star cluster, the velocity at infinity $v_\infty$ is typically the
velocity dispersion in the star cluster multiplied by a factor $\sqrt{2}$. When both bodies have a negligible mass, such as in the case of a planet-planet encounter, the equation reduces to $p=b$, while for a close encounter between two large masses the effect of gravitational focusing is important and the equation reduces to $p = (bv_\infty)^2(2GM)^{-1}$. As the distribution over impact parameters scales as $f_b(b)\propto b$, the corresponding distribution for $p$ is $f(p)\propto p$ for distant encounters. The corresponding distribution for close encounters can be obtained with transformation of variables, which results in $f(p) \propto v_\infty^2(2GM)^{-1}$. In other words, $f(p)$ is independent of $p$ in this regime.
Figure~\ref{figure:periastron} shows the periastron distance distributions $f(p)dp$ for the different types of encounters for models~3 and~5. The distributions are linear over most of the range in $p$, and we therefore describe the results using the functional form
\begin{equation} \label{eq:periastrondistribution}
f(p) = \mathcal{N} \left( \alpha p + \beta \right)
= \mathcal{N} \left( \alpha (p-p_m/2) + p_m^{-1}\right)
\end{equation}
to the distributions for the different types of encounters, where $\mathcal{N}$ represents the total number of encounters of a specific type, $\alpha$ is in units of AU$^{-2}$ and $\beta$ in units of AU$^{-1}$. Normalization to the total number encounters gives $\beta=p_m^{-1}-\alpha p_m/2$, where $p_m=1000$~AU is the largest periastron distance that is taken into consideration. In the limit that the masses of the encountering bodies are negligible, $f(p)=2\mathcal{N} p/p_m^2$, while in the limit of very close encounters between massive bodies, $f(p)=\beta=\mathcal{N}/p_m$. The corresponding cumulative distribution is
\begin{equation}
F(p) = \mathcal{N} \left( \frac{\alpha p(p-p_m)}{2} + \frac{p}{p_m} \right) \ ,
\end{equation}
which indicates the cumulative number of encounters smaller than $p$.
Note that in this approximation the values of $\alpha$ (and $\beta$) are independent of the number of encounters, and also independent of $N_{\rm sb}$, $N_{\rm p}$, and $\mathcal{B}$. For encounters between more massive bodies, however, $\alpha$ (and therefore also $\beta$) depend on $v_\infty$, and the latter quantity may have a different distribution in space (due to mass segregation) and time (due to the preferred escape of low-mass members) than for P-P encounters.
In realistic star clusters $f(p)$ is a combination of weak, intermediate, and strong encounters, and for the combination of all encounters of a certain type, integrated over the entire star cluster and over time. Nevertheless, $f(p)$ is still reasonably well approximated by the expression in Eq.~\ref{eq:periastrondistribution}. Several of the distributions in Figure~\ref{figure:periastron} show an overabundance of periastron approaches below 10~AU with respect to Eq.~\ref{eq:periastrondistribution}. This overabundance is exclusively seen for encounters involving single stars (S-S, S-P, and S-SS) in model~5, and is related to the evolution of primordial binary systems, their interactions with other star cluster members, and subsequent decay.
Figure~\ref{figure:pfit} shows the fitted values of $\alpha$ (Eq.~\ref{eq:periastrondistribution}) for all models.
Since the vast majority of the P-P encounters are weak, the fitted value for $\alpha$ is close to the predicted value in the unfocused limit, $\alpha=2p_m^{-2}=2\times 10^{-6}$~AU$^{-2}$, although the scatter around this value is considerable, particularly for the low-density star clusters for which a large fraction of the FFPs escape immediately. Larger masses for the encountering bodies generally result in smaller values of $\alpha$, which is consistent with the increased gravitational focusing. In addition, when the star cluster mass increases and/or its radius decreases, the velocity dispersion (and hence $v_\infty$) becomes larger, such that gravitational focusing is less important and the measured values of $\alpha$ are larger.
The typical time that two encountering members spend near periastron is $\Delta t = 2p/v_p$. For weak encounters this is $\Delta t \approx 2b/v_\infty$ and for strong encounters this is $\Delta t \approx bv_\infty/GM$.
This timescale is important in studying the effects of encounters when one of the two members, for example a single star, has a planetary system, as it describes, given a planet orbiting a star with a certain semi-major axis, whether the encounter is impulsive or adiabatic, and whether it is close or tidal \citep[see the extensive study of][for further analysis]{spurzem2009}.
\section{Conclusions and Discussion} \label{section:conclusions}
Free-floating planets (FFPs) may be abundant in young star clusters, as close encounters between stars can destabilise planetary systems, which can result in direct or delayed ejection of planets from their host planetary systems. In this article we have presented a study of the dynamical evolution of FFP populations in various types of star clusters. We have carried out $N$-body simulations to characterise how the FFP population evolves and we have studied the properties of close ($<1000$~AU) encounters between single stars, binary stars, and FFPs. Our results can be summarised as follows:
\begin{enumerate}
\item A certain fraction of the FFPs escape from their star clusters shortly after they are ejected from their host stars, while the remaining FFPs escape at much later times. Beyond the time of early escape, the planet-to-star ratio decreases linearly with time, and the times $t_{\rm hs}$ and $t_{\rm hp}$ at which respectively half of the stars and planets escape from the star clusters are related through the simple relation $t_{\rm hp}\approx 0.6t_{\rm hs}$. For individual planets with low ejection velocities this means that they are likely to escape from the star cluster at a 40\% earlier time than their host star.
\item Many FFPs ejected from their host star system at early times experience tens of close ($<1000$~AU) encounters with other stars and FFPs in the star cluster before escaping from the cluster. The fraction of FFPs that leave the cluster without any close encounter increases with increasing initial velocity distribution and with decreasing stellar density.
\item Typically half of the encounters of all types occur within 30~Myr, while only 10\% of the encounters occur after the first 100~Myr. The ratios of the timescales at which half of the number of star-star, star-planet, and planet-planet encounters occur are $t_{h,S-S}:t_{h,S-P}:t_{h,P-P}\approx1:0.77:0.62$.
\item The velocity-at-infinity distributions, $f(v_\infty)$, of the encounters are well approximated with Maxwell-Boltzmann distributions. The periastron velocity distributions for all types of encounters are well fitted by distributions $f(v_p-v_o)$ similar to a Maxwell-Boltzmann distribution, with a velocity offset $v_o$. The most frequent velocity $v_m$ in this distribution is proportional to the velocity dispersion of the star cluster, and the offset $v_o$ is primarily determined by the combined mass of the encountering bodies of a certain type.
\item The distribution over periastron distances is linear, $f(p)=\alpha p + \beta$ for the encounters is linear over most of the values of $p$. In our case we have recorded all approaches within $p_m=1000$~AU. In that case, distant encounters have $\alpha=2\times 10^{-6}$~AU$^{-2}$ and $\beta=0$~AU$^{-1}$. In the case of close encounters, $f(p)=\beta=\mathcal{N}/p_m$, where $\mathcal{N}$ is the total number of encounters and $p_m$ is the close encounter limit. Also when we combine the encounters for the entire life of the star clusters, the linear fit is good, except for very small distances ($p<10$~AU), for which approaches are more common.
\end{enumerate}
Our study is intended to obtain a general overview of the dynamical evolution of a FFP population in star clusters. We have made several assumptions that should be kept in mind when making realistic predictions for existing star clusters, or when interpreting observational data. First, a critical assumption we made is that the FFP space density is identical to the stellar density. In reality, the ejection velocities of FFPs may depend in an intricate way on the position and velocity of the star at the moment when it is perturbed. This is especially the case when planets are ejected promptly after their host system experiences a close encounter. One would then expect planets to be preferably ejected when their host star is near the star cluster centre. In the case of delayed ejection, however, this may well be a good approximation, since the stars are well-mixed at the moment of planet ejection.
Second, in our initial model setup we have also made the crude assumption that all planets are initially free-floating. We also neglect any additional FFPs produced by the close encounters between stars in our model, as well as the possible capture of FFPs by planetary systems. Although these is an unrealistic assumptions, particularly in the case of perturbed planetary systems that may eject planets tens of millions of years later \citep[e.g.,][]{malmberg2011, hao2013}, it is a good approximation if most planet ejections occur at early times. Moreover, many of the analytical and computational results presented in Section~\ref{section:results} are easily scalable to any value of $N_{\rm p}$. These can therefore also be used to evaluate the probabilities for close encounters and escape for individual planets ($N_{\rm p}=1$). When convolved with an time-dependent FFP production rate, estimates for the dynamical behaviour of the entire population of FFP can subsequently be obtained.
In a follow-up study we will analyse more realistic initial conditions, ideally by modelling the full $N$-body evolution of decaying multi-planet systems in young star clusters, potentially using the AMUSE framework \citep[][]{portegies2013, pelupessy2013, cai2015},
Specifically, we will analyse the dynamical evolution of a FFP population where the FFPs have a stronger preference to be generated in the cluster centre, and where they are ejected at appropriate times resulting from close encounters with other cluster members and the decay of perturbed multi-planet systems.
\section*{Acknowledgments}
We wish to thank Sourav Chatterjee for carefully reading the manuscript and providing useful comments that helped to improve this paper.
We wish to thank Rainer Spurzem and Sverre Aarseth for their advice regarding the adjustment of the NBODY6 software package for our study. L.W. and X.C.Z. were supported by the Department of Astronomy at Peking University and the Kavli Institute for Astronomy and Astrophysics.
M.B.N.K. was supported by the Peter and Patricia Gruber Foundation through the PPGF fellowship, by the Peking University One Hundred Talent Fund (985), and by the National Natural Science Foundation of China (grants 11010237, 11050110414, 11173004). This publication was made possible through the support of a grant from the John Templeton Foundation and National Astronomical Observatories of Chinese Academy of Sciences. The opinions expressed in this publication are those of the author(s) do not necessarily reflect the views of the John Templeton Foundation or National Astronomical Observatories of Chinese Academy of Sciences. The funds from John Templeton Foundation were awarded in a grant to The University of Chicago which also managed the program in conjunction with National Astronomical Observatories, Chinese Academy of Sciences.
R.P.C. was supported by the Swedish Research Council (grants 2012-2254 and 2012-5807). M.B.D. was supported by the Swedish Research Council (grants 2008-4089 and 2011-3991).
|
\section{Introduction}
In the theory of dynamical systems on the circle, there is the following very natural question: let $a,b\in \widetilde{\mathrm{Homeo}}_+(S^1)$ be lifts on the real line of two orientation-preserving circle homeomorphisms, and we know their rotation (or, more precisely, translation) numbers $\mathop{\mathrm{rot}^{\sim}}(a),\mathop{\mathrm{rot}^{\sim}}(b)\in\mathbb{R}$. What can be said about the translation number of their composition~$ab$?
Another, more general, form of the same question was studied in a work~\cite{JN} of Jankins and Neumann. It have had topological origins: the question of classification of 3-manifolds, admitting at the same time a Seifert fibration and a codimension one foliation transverse to it. By the moment of Jankins--Neumann's work, the only non-studied case was the one of a manifold fibered over a 2-sphere. Via the study of the corresponding holonomy maps, this have led them to the following question:
\begin{question}\label{q:JN}
Given $a_1, a_2, \dots, a_n \in [0,1]$, $n\ge 3$, when do there exist lifts $f_1, f_2, \dots, f_n$ of orientation-preserving homeomorphisms of the circle with $\mathop{\mathrm{rot}^{\sim}}(f_i)=a_i$, such that $f_1\dots f_n=id$?
\end{question}
They have suggested a conjectural answer to this question, also proving that it suffices to establish their conjecture for $n=3$ and that for $n=3$ their conjecture holds at least $99.9\%$ of the volume of the set. The set they have proposed for $n=3$ was later called the Jankins-Neumann ziggurat due to its stepwise nature (see~Fig.~\ref{f1}).
Their conjecture was proven by Naimi in~\cite{Nai}:
\begin{Th}[Naimi; conjecture of Jankins-Neumann]\label{zigJN}
The set defined in Question~\ref{q:JN} for $n=3$ is the union of parallelepipeds
$$
[0;\frac{a}{m}]\times [0;\frac{m-a}{m}]\times [0;\frac{1}{m}]
$$
for all coprime $0<a<m$, and of their images under all the permutations of the coordinates.
\end{Th}
Later, Calegari and Walker attacked the question of the rotation number of the composition from the dynamical point of view. They have obtained an ``algorithmic'' description for analogous sets for any positive composition of two homeomorphisms with given rotation numbers:
\begin{figure}[h]
\noindent\centering{
\includegraphics[width=120mm]{fig_JN.pdf} }
\caption{Jankins-Neumann ziggurat (picture credit: Jankins-Neumann~\cite{JN})}
\label{f1} \end{figure}
\begin{Th}[Calegari--Walker, \mbox{\cite{Caleg}}]\label{t:w}
For any word $w$ in the alphabet $a,b$, one has
$$
\{\mathop{\mathrm{rot}^{\sim}}(w(a,b)) \mid \mathop{\mathrm{rot}^{\sim}}(a)=x, \mathop{\mathrm{rot}^{\sim}}(b)=y\} = [r_{w}(x,y), R_{w}(x,y)]
$$
for certain functions $r_w, R_w:\mathbb{R}\to \mathbb{R}$, and one has $r_w(x,y)=-R_w(-x,-y)$. If the word $w$ is positive (i.e. contains no $a^{-1}$ or $b^{-1}$), there is an explicit algorithm to compute the functions $r_w$, $R_w$ at any rational point $(x,y)$.
\end{Th}
This theorem implies, in particular, that the ``ziggurat'' of possible rotation numbers of the composition is described by its upper boundary, the graph of the function~$R_{ab}(x,y)$. An immediate remark is that the function $R_{ab}(x,y)-x-y$ is $\mathbb{Z}^2$-periodic, thus to understand $R_{ab}(x,y)$ it suffices to study it on the unit square $[0,1)^2$. Moreover, starting from their dynamical approach, Calegari and Walker have obtained an explicit formula for $R_{ab}$ in terms, different from those of Jankins and Neumann:
\begin{Th}[{\itshape ab}~Theorem, \mbox{\cite{Caleg}}]\label{t:ab}
\begin{equation}\label{formula}
\forall x,y \quad R_{ab}(x, y) = \sup_{\frac{p_1}{q}\leq x, \frac{p_2}{q}\leq y} \frac{p_1 + p_2 +1}{q}.
\end{equation}
\end{Th}
The purpose of the present text is twofold. First, the Jankins--Neumann ziggurat clearly exhibits some fractal nature (see Fig.~\ref{f1},~\ref{ab}). We study the geometry of this ziggurat, and in particular show that it is indeed the case: that the set of its vertices is self-similar under some simple projective transformations.
Secondly, as it was mentioned earlier, the theorem of Calegari and Walker and the Jankins--Neumann conjecture have the sets described in a different way. It is known that these two descriptions are equivalent (in particular, Calegari--Walker's Theorem~\ref{t:ab} gives an alternative proof of the Jankins--Neumann conjecture). However, we have found a very interesting proof of the passage between the two, and it seems that this way of proving the equivalence was not previously known, and is shorter than existing one. We present this passage in Sec.~\ref{s:equivalence}. Also, we discuss the corollaries of this comparison for the Calegari--Walker formula and the function~$R_{ab}$.
\section{Vertices and the self-similarity}\label{stat}
A first immediate remark is the passage between the definitions of the ziggurat by Jankins--Neumann and by Calegari--Walker:
\begin{pro}\label{zigCW}
The triple $(x, y, z) \in [0;1)^3$ of translation numbers can be represented by $f_1, f_2, f_3 \in \widetilde{\mathrm{Homeo}}_+(S^1)$ such that $f_1f_2f_3=\mathop{\mathrm{id}}$ if and only if $z \in [0; R_{ab}(1 - x , 1 - y) - 1]$.
\end{pro}
In this section, we will be working with the Jankins--Neumann ziggurat, denoting it by~$\tilde{\mathcal Z}$. Naimi's Theorem~\ref{zigJN} then implies that
\begin{multline}\label{e:zigJN}
\tilde \mathcal Z = \bigcup_{\mathop{\mathrm{gcd}}(a,m) = 1} \Pi\left( \frac{a}{m},\frac{m-a}{m}, \frac{1}{m}\right) \cup \bigcup_{\mathop{\mathrm{gcd}}(a,m) = 1} \Pi\left( \frac{a}{m},\frac{1}{m}, \frac{m-a}{m}\right) \\
\cup \bigcup_{\mathop{\mathrm{gcd}}(a,m) = 1} \Pi\left( \frac{1}{m},\frac{a}{m}, \frac{m-a}{m}\right)
\end{multline}
where $\Pi(a,b,c) = [0;a]\times [0;b]\times [0;c]$.
To state the self-similarity result, let us first introduce the notion: we call point $X$ a \emph{vertex} of a ziggurat $Z$, that is the union of parallelepipeds, if $X \in Z$ and there is no $\Pi(a,b,c) \subset Z$ such that $X \in \Pi(a,b,c)\setminus \{(a,b,c)\}$. In terms of $R_{ab}$, a \emph{vertex} is a point $(x,y,R_{ab}(x,y))\in [0,1)^3$, such that for any $x'\le x$ and $y'\le y$ inequality $x'+y'<x+y$ implies $R(x',y')<R(x,y)$.
Now, we can list the vertices of the Jankins--Neumann ziggurat: a direct corollary of the Naimi's theorem is
\begin{pro}\label{p:list}
The vertices of the Jankins--Neumann ziggurat are points of the three families, that differ by the permutation of the coordinates:
\begin{itemize}
\item $\{(\frac{a}{m},\frac{m-a}{m},\frac{1}{m})\}$, with coprime $m>a>0$;
\item $\{(\frac{m-a}{m},\frac{1}{m},\frac{a}{m})\}$, with coprime $m>a>0$;
\item $\{(\frac{1}{m},\frac{m-a}{m},\frac{a}{m})\}$, with coprime $m>a>0$.
\end{itemize}
These three families lie respectively on the planes $x+y=1$, $x+z=1$ and $y+z=1$. Moreover, they lie respectively inside the triangles $ABD$, $ACD$ and $BCD$, where $A$, $B$ and $C$ are respectively the points at unit distance on the axes $Ox$, $Oy$ and $Oz$, and $D=(1/2,1/2,1/2)$ is the only common point of all the three families.
\end{pro}
\begin{rem}\label{r:min}
Formally speaking, to ensure that all the points in the above list are the indeed the vertices, one should check that neither of the parallelepipeds listed in the Jankins--Neumann conjecture is contained in any other. Though, such a check is almost immediate (and we will do it in Sec.~\ref{s:equivalence}).
\end{rem}
\begin{cor}\label{c:list}
Translating Proposition~\ref{p:list} on the language of $R_{ab}$, we get for its (stepped) graph the families of vertices
\begin{itemize}
\item $\{(\frac{m-a}{m},\frac{a}{m},1+\frac{1}{m})\}$ with coprime $m>a>0$;
\item $\{(\frac{a}{m},1-\frac{1}{m},1+\frac{a}{m})\}$ with coprime $m>a>0$;
\item $\{(1-\frac{1}{m},\frac{a}{m},1+\frac{a}{m})\}$ with coprime $m>a>0$.
\end{itemize}
\end{cor}
Our {first} result, Theorem~\ref{ss} below, states that these points form a self-similar set. But before stating it, we would like to have additional geometric intuition on that set of vertices. Namely, at first glance it seems natural to decompose the ziggurat on Fig.~\ref{f1} into three parts ``near $A$'', ``near $B$'', ``near $C$''. However, as Proposition~\ref{p:list} shows, it is much more important to decompose the vertices into the three families listed in this proposition: the vertices that are on the triangles $ABD$, $ACD$ and $BCD$ respectively.
{Also, we would like to deduce conclusions for the function $R_{ab}$. Thus it is interesting to consider the projection of the ziggurat on the $xy$ plane (marking the level surfaces and discontinuity lines): see Fig.~\ref{proj}. Marking only the vertices on this projection, we get Fig.~\ref{vert}.} Such a projection sends the vertices that correspond to the first family on the line $x+y=1$, and the second and the third families become separated by the diagonal $x=y$: the second comes below the diagonal, while the third one comes above.
\begin{figure}[h] \begin{multicols}{2}
\hfill \includegraphics[width=70mm]{fig2.pdf} \hfill \caption{Graph of $R_{ab}$, seen from the top} \label{proj} \hfill \includegraphics[width=70mm]{vertices.pdf} \hfill \caption{Projection of its vertices on the $xy$-plane, with the axis of symmetry marked} \label{vert}
\end{multicols} \end{figure}
Now, let $\Delta$ be the set that comes from the projection of vertices from the third family. As we have claimed before, this set is then self-similar:
\begin{figure}
\noindent\centering{
\includegraphics[width=70mm]{vertices-invariance.pdf} \hfill \caption{The set $\Delta$ and its self-similarity} \label{map}}
\end{figure}
\begin{Th}[Self-similarity]\label{ss}
The set $\Delta$ is self-similar with respect to two projective transformations
$$
T_1(x,y)=\left( \frac{x}{1+x}, \frac{x+y}{1+x}\right), \quad T_2(x,y)=\left( \frac{1}{2-x}, \frac{1+y-x}{2-x}\right),
$$
namely, one has
$$
\Delta=T_1(\Delta) \sqcup T_2(\Delta) \sqcup \{\left(1/2,1/2\right)\}.
$$
\end{Th}
\begin{rem}\label{r:plane}
{As all the vertices of the third family lie on the plane $y+z=1$ of the triangle $ACD$, their projection to the $xy$ plane is an affine transformation. Hence, the set of the vertices of the third family is also self-similar under projective transformation that are lifts of $T_1$ and of $T_2$ on this plane. }
\end{rem}
\begin{rem}
{Note that the set $\Delta$ is symmetric with respect to the line $x=1/2$. It is an immediate observation if one considers the Jankins--Neumann ziggurat, coming from its full symmetry under the permutation of the coordinates. Though, it is much less clear if one studies the level curves of the function~$R_{ab}$: for any vertex and its symmetric image the projection of the ziggurat on the $xy$ plane collapses one of the sides of the corresponding parallelepiped, but the collapsed sides are different. This is why this symmetry does not appear on Fig.~\ref{proj}.}
\end{rem}
{Strangely enough, we did not find any direct way of establishing Theorem~\ref{ss} in the dynamical terms, that is, starting with the explicit formula~\eqref{formula} and without at first deducing the full list of the vertices (in a way it is done in Sec.~\ref{s:equivalence}). That is, one could imagine that having a vertex $(x,y)=(\frac{p_1}{q},\frac{p_2}{q})$ and considering its $T_1$ or $T_2$-image, formula~\eqref{formula} would allow to show that this image is also a vertex. Unfortunately, without stating the full list of vertices first such an argument does not seem to work.
Instead, we obtain Theorem~\ref{ss} as a corollary of the Jankins--Neumann original description of the ziggurat (and the associated list of vertices from Proposition~1).}
Another conclusion for the function $R_{ab}$, following from the Jankins--Neumann description is the following
\begin{pro}\label{c:rational}
The function $R_{ab}$ takes only rational values; in formula~\eqref{formula}, for any $x$ and $y$ the supremum is a maximum.
\end{pro}
This statement generalizes the Rationality Theorem of Calegari and Walker~\cite[Theorem 3.2]{Caleg}, giving the same rationality conclusion for the rational points~$(x,y)$.
It is also an interesting remark (though \emph{a posteriori} almost immediate to prove) that the projections of the vertices are indeed aligned along the lines that are ``visible'' on Fig.~\ref{vert}. The following theorem formalizes this statement; to state it, consider two families of lines. Namely, the ``green'' family of lines passing through the point $(0,1)$ and having slopes $(-1/m)$, $m=1,2,\dots$, and the ``red'' family of lines passing through the point $(1,1)$ and having slopes $1/k$, $k=1,2,\dots$.
It is easy to check that the lines from these families are given by equations $y=\frac{2m-1}{m}-\frac{x}{m}$ and $y=\frac{k-1}{k}+\frac{x}{k}$ respectively. Let $\Delta'$ be the set of intersection points of lines of green family with lines of red family. Then, we have the following theorem, illustrated by~Fig.~\ref{f:lines}.
\begin{Th}[Alternative vertex set description]\label{lines}
$\Delta$ is the part of $\Delta'$ formed by the points with the least possible ordinate for given abscise:
$$
\Delta=\{(x,y)\in \Delta' \mid \forall y'<y \quad (x,y')\notin \Delta'\}.
$$
\end{Th}
\begin{rem}
Again, as in Remark~\ref{r:plane}, the construction of this theorem can be lifted on the $ACD$ plane containing the third family of vertices. Making such a lift, one notices that the vertical line starting from an intersection point corresponds to a vertical edge of the parallelepiped, starting from the corresponding vertex. The ``least possible ordinate'' rule then corresponds to the fact that the intersection points with non-least ordinate lift to the points that belong to the corresponding edge, and thus that are not vertices.
\end{rem}
\begin{figure}
\noindent\centering{
\includegraphics[width=120mm]{vertices-lines.pdf} }
\caption{Lines listed in Theorem~\ref{lines}. Green and red families pass through the points (0,1) and (1,1) respectively. Bold black points correspond to the points of $\Delta$, the points of $X$ between diagonals $x=y$ and $x+y=1$ are marked by smaller grey points to illustrate additional aligning. Dashed vertical lines, starting in the points of~$\Delta$, illustrate the ``least possible ordinate'' condition.}
\label{f:lines} \end{figure}
\begin{figure}
\noindent\centering{
\includegraphics[width=120mm]{lines-in-theory.pdf} }
\caption{Fig. \ref{f:lines} after projective transformation $(x,y) \mapsto \left(\frac{x}{1 - y}, \frac{1-x}{1-y}\right)$. Some of dashed lines are not drawn, the extensions of some dashed lines are drawn dotted to illustrate passing through the origin.}
\label{f:strtn} \end{figure}
\begin{proof}[Proof of Theorems \ref{ss} and \ref{lines}]
First
apply a projective transformation $Q$ that sends $(0,1)=[0:1:1]$ and $(1,1)=[1:1:1]$ to the points at infinity $V=[0:1:0]$ and $H=[1:0:0]$, corresponding to the vertical and horizontal directions respectively, and that sends $V$ to the origin. It is easy to see that it is given by the formula
$$Q: (x,y) \mapsto \left(\frac{x}{1 - y}, \frac{1-x}{1-y}\right).$$
or, equivalently, by
$$
Q:[x:y:t]\mapsto [x:t-x:t-y].
$$
The result of its application is shown on Fig.~\ref{f:strtn} (the reader perhaps will find this figure even more convincing than the formal arguments below).
It is easy to see that it sends a vertex $\left(\frac{r}{q}, \frac{q-1}{q} \right)$ (due to Proposition~\ref{p:list} and its Cor.~\ref{c:list}) the set $\Delta$ is formed by such points with coprime $0<r<q$) to the point $(r, q-r)$, and hence $Q(\Delta)$ is exactly the subset of $\mathbb{N}^2$ formed by points with coprime coordinates.
Next, an immediate check gives that the image of the line at infinity is the line $x=y$, the point $[a:1:0]$ (corresponding to the slope~$1/a$) being sent to the point~$(a,-a)$. Thus, the green family (as these are lines passing through $(0,1)$ and having slope $(-1/m)$) becomes the family of the vertical lines $x=m$, while the red family (as these are lines passing through $(0,1)$ and having slope~$(1/k)$) becomes the family of horizontal lines $y=k$.
Hence the set $\Delta'$ is sent exactly to~$\mathbb{N}^2$. Finally, as the point $V$ is sent to the origin, the ``least possible ordinate'' condition after the transformation~$Q$ becomes exactly the coprimality condition; this concludes the proof of Theorem~\ref{lines}.
To prove Theorem~\ref{ss}, note that in new coordinates the transformations $T_1$ and $T_2$ take form
$$
\hat T_1(a,b) = (a, a+b); \, \text{ and } \,
\hat T_2(a,b) = (a+b, b)
$$
and the self-similarity of $\Delta$ becomes obvious: in the new coordinates, it is the Euclid's algorithm!
\end{proof}
\begin{figure}[h] \begin{multicols}{2}
\hfill \includegraphics[width=70mm]{Figure1.pdf} \hfill \caption{The ab-ziggurat (picture credit: Calegari-Walker~\cite[Fig.~1]{Caleg})} \label{ab} \hfill \includegraphics[width=70mm]{aabab.pdf} \hfill \caption{The abaab-ziggurat (picture credit:~\cite[Fig.~2]{Caleg})} \label{aabab}
\end{multicols} \end{figure}
To conclude this section, we would like to state some related interesting questions:
\begin{question}
Is there any direct dynamical proof of $T_{1,2}$-invariance of the set $\Delta$ or of its vertical symmetry? Do the transformations $T_{1,2}$ admit any dynamical interpretation?
\end{question}
\begin{question}\footnote{After this text was finished, we became aware that Subhadip Chowdhury has managed to show ``asymptotic'' projective self-similarity near the ``fringes'' of the unit square for the ziggurats associated to some other positive words . }
What happens for ziggurats associated to other positive products? For instance, the $abaab$-ziggurat also seems to have some fractal, and possibly self-similar nature, see Fig.~\ref{aabab}.
\end{question}
\begin{question}
Would the ziggurats in higher dimensions (for larger number of different homeomorphisms being multiplied) still look self-similar? (For the Jankins--Neuman higher-dimensional ziggurat, the proof of Theorem~\ref{ss} applies verbatim.)
\end{question}
\begin{question}
What happens with the ziggurat if the defining Jankins--Neumann-like relation includes \emph{all} the homeomorphisms more than once; for instance, for the relation $abcbac=\mathop{\mathrm{id}}$?
\end{question}
\section{Equivalence}\label{s:equivalence}
This section is devoted to a way of deducing the Jankins--Neumann conjecture from Calegari--Walker's formula~\eqref{formula}. To make such a deduction, let
\begin{equation}\label{e:zR}
\mathcal Z = \{(x,y,z) \mid 0\le z\le R_{ab}(1-x,1-y)-1\}.
\end{equation}
It is easy to see that the Jankins--Neumann conjecture is equivalent to that $\mathcal Z=\tilde \mathcal Z$. The Calegari--Walker formula allows to represent $\mathcal Z$ as a union of parallelepipeds. Indeed, the inequalities $\frac{p_1}{q}\le x$, $\frac{p_2}{q}\le y$ implying $R_{ab}(x,y)\ge \frac{p_1+p_2+1}{q}$ give for the set $\mathcal Z$ the representation
\begin{multline}\label{e:tz}
\mathcal Z = \bigcup_{p_1,p_2,q \in \mathbb{N}, \atop p_1,p_2\le q} \{(x,y,z) \mid 1-x\le \frac{p_1}{q},\, 1-y\le \frac{p_2}{q} , \, z \le \frac{p_1+p_2+1}{q}-1, \, x,y,z\ge 0\}
\\
= \bigcup_{p_1,p_2,q \in \mathbb{N},\atop p_1,p_2 \le q} \Pi\left( 1-\frac{p_1}{q}, 1-\frac{p_2}{q}, \frac{p_1+p_2 + 1}{q}-1\right).
\end{multline}
Take $p_1'=q-p_1, p_2'=q-p_2$. Then,
$$
\Pi\left( 1-\frac{p_1}{q}, 1-\frac{p_2}{q}, \frac{p_1+p_2 + 1}{q}-1\right) = \Pi\left( \frac{p_1'}{q}, \frac{p_2'}{q}, \frac{q+1-(p_1'+p_2')}{q}\right).
$$
Finally, denoting $p_3':=q-p_1'-p_2'+1$, we get
\begin{equation}\label{e:zigCW}
\mathcal Z= \bigcup_{p_1',p_2',p_3',q \in \mathbb{N}, \atop p_1'+p_2'+p_3'= q+1} \Pi\left( \frac{p_1'}{q}, \frac{p_2'}{q}, \frac{p_3'}{q}\right).
\end{equation}
To deduce Naimi's theorem (Jankins--Neumann conjecture) from the Calegari--Walker formula, we have to show that $\mathcal Z$ and $\tilde \mathcal Z$ coincide. The list of parallelepipeds in~\eqref{e:zigCW} contains all the parallelepipeds from the Jankins--Neumann conjecture, but also some others. Let us call a parallelepiped good if it is, up to the permutation of the coordinates, of the form $\Pi(\frac{1}{q}, \frac{p}{q}, \frac{q-p}{q})$ with coprime $p$ and $q$. Then, to prove the coincidence of $\mathcal Z$ and $\tilde \mathcal Z$ we have to show that any parallelepiped in~\eqref{e:zigCW} that is not a good one, is contained in one of the good ones. For the obvious monotonicity reasons, it suffices to check that its vertex $\left( \frac{p_1'}{q}, \frac{p_2'}{q}, \frac{p_3'}{q}\right)$ is contained there.
Before proceeding to prove this, we notice (as it was already promised in Remark~\ref{r:min} earlier) that the list of good parallelepipeds is minimal: no good parallelepiped is contained in another one. Indeed, for any of the parallelepipeds listed in~\eqref{e:zigCW}, the sum of any two of three coordinates does not exceed~$1$. As for a good parallelepiped $\Pi\left( \frac{p}{q}, \frac{q-p}{q}, \frac{1}{q}\right)$ the sum of its first two coordinate equals~$1$, it could be contained only in a parallelepiped of the form $\Pi\left( \frac{p'}{q'}, \frac{q'-p'}{q'}, \frac{1}{q'}\right)$ with $\frac{p'}{q'}=\frac{p}{q}$ and $q'<q$. But as $p$ and $q$ are coprime, it is impossible.
Now, let $p_1+p_2+p_3=q+1$, with $p_1,p_2,p_3>1$ (thus, $\Pi\left(\frac{p_1}{q},\frac{p_2}{q},\frac{p_3}{q} \right)$ is not good). We are going then to find a good parallelepiped that contains $\Pi( \frac{p_1}{q}, \frac{p_2}{q}, \frac{p_3}{q})$. Without loss of generality, we can assume $p_1\le p_2\le p_3$. Then, we will be looking for coprime $m<n$ such that
$$
\Pi( \frac{p_1}{q}, \frac{p_2}{q}, \frac{p_3}{q})\subset \Pi( \frac{1}{n}, \frac{m}{n}, \frac{n-m}{n}),
$$
or, what is the same,
$$
\frac{p_1}{q} \le \frac{1}{n}, \quad \frac{p_2}{q}\le \frac{m}{n}, \quad \frac{p_3}{q}\le \frac{n-m}{n}.
$$
Or, equivalently,
\begin{equation}\label{eq:mn}
n\le \left[\frac{p_1}{q}\right], \quad \frac{p_2}{q}\le \frac{m}{n} \le 1- \frac{p_3}{q}.
\end{equation}
Denote $N:=\left[\frac{p_1}{q}\right]$. Then, the desired~\eqref{eq:mn} could be reformulated in the following way: we want to prove that there is a fraction $\frac{m}{n}$ of denominator less than $N$ that belongs to the interval $[\frac{p_2}{q} , \frac{q-p_3}{q}]$. Hence it is natural to consider \emph{Farey sequence} of order~$N$.
Recall that the Farey sequence $F_N$ of order $N$ is the sequence of completely reduced fractions between~$0$ and~$1$ which have denominators less than or equal to~$N$, arranged by increasing. In the proof of the equivalence of two ziggurats we will use one of its properties, namely, for any two consecutive fractions $\frac{a}{b}<\frac{c}{d}$ in any of the Farey sequences, their denominators are coprime, and moreover, $\frac{c}{d}-\frac{a}{b}=\frac{1}{bd}.$ As for details, we refer the reader to ~\cite{F2} (as well as to the original historical papers~\cite{F0,F1}).
Suppose there is no $\frac{m}{n} \in F_N$ between $\frac{p_2}{q}$ and $\frac{q-p_3}{q}$. Then, take the two consecutive fractions $\frac{a}{b}, \frac{c}{d} \in F_N$ such that $\frac{a}{b}<\frac{p_2}{q}\le\frac{q- p_3}{q}<\frac{c}{d}$. In this case,
$$
\frac{p_2}{q}-\frac{a}{b}\ge \frac{1}{bd}, \quad \frac{c}{d}-\frac{q - p_3}{q}\ge \frac{1}{dq}.
$$
Thus,
\begin{multline} \label{eq:bdq}
\frac{q - p_3}{q} - \frac{p_2}{q}=\left(\frac{c}{d}-\frac{a}{b}\right) - \left(\frac{c}{d}-\frac{q - p_3}{q}\right) - \left(\frac{p_2}{q}-\frac{a}{b}\right) \le\\
\le \frac{1}{bd} - \frac{1}{bq} - \frac{1}{dq} = \frac{q-(b+d)}{bd\cdot q}.
\end{multline}
If both $b,d \ge 2$, then $bd \ge b+d\ge N+1 > \frac{q}{p_1}$, we see that the right hand side of~\eqref{eq:bdq} can be estimated as
$$
\frac{q-(b+d)}{bd\cdot q} < \frac{q-\frac{q}{p_1}}{\frac{q}{p_1}\cdot q} = \frac{p_1-1}{q}.
$$
Thus, $\frac{q - p_3}{q} - \frac{p_2}{q}< \frac{p_1-1}{q}$, contradicting the equation $p_1+p_2+p_3=q+1$.
Finally we have to consider the possibilities $b=1$ or $d=1$. Though, as $\frac{p_2}{q} + \frac{p_3}{q} < 1$, we have $\frac{p_2}{q}<\frac{1}{2}$ and hence $d \ge 2$. Finally, if $b=1$ and thus $\frac{a}{b} = 0, \frac{c}{d}=\frac{1}{N}$, we have
$$
1-\frac{p_3}{q}=\frac{p_2+p_1-1}{q}\ge \frac{2p_1-1}{q} \ge \frac{3}{2} \cdot \frac{p_1}{q}\ge \frac{3}{2} \cdot \frac{1}{N+1}.
$$
Then, the inequality $1-\frac{p_3}{q}<\frac{c}{d}=\frac{1}{N}$ is impossible, as it would imply $\frac{3}{2(N+1)} < \frac{1}{N}$ and hence $N< 2$. Though, $\frac{p_1}{q}\le \frac{1}{2}$ and hence $N\ge 2$.
\begin{figure}[h] \begin{multicols}{2}
\hfill \includegraphics[width=60mm]{z-proj-med-triangles-v2.pdf} \hfill \caption{The $\Delta_n$ triangles marked on the view on the ab-ziggurat from the top} \label{f:t-red} \hfill \includegraphics[width=60mm]{Delta-triangles-v3.pdf} \hfill \caption{Triangle $\Delta_{\frac{a}{b},\frac{c}{d}}$ and two its descendants $\Delta_{\frac{a}{b},\frac{a+c}{b+d}}$ and $\Delta_{\frac{a+c}{b+d},\frac{c}{d}}$.} \label{f:induction}
\end{multicols} \end{figure}
These contradictions conclude the proof of the equivalence.
Finally, we would like to note a beautiful connection of ziggurat and Farey series. Consider the set of triangles on the plane $Oxy$, enclosed by lines $x+y=1$, $x=1/n$ and $y=n/(n+1)$, where $n=2,3,\dots$. Let us denote $n$th such triangle by~$\Delta_n$:
\[
\Delta_n := \{(x,y)\in [0;1)^2 \mid x+y>1,\, x< 1/n,\, y< n/(n+1) \}
\]
These triangles naturally appear on the view from the top of the ab-ziggurat (Jankins-Neumann up to a linear transformation): see Fig.~\ref{proj} and~\ref{f:t-red}; on the latter they are marked by bold red lines. It is easy to see from the Jankins--Neumann's description, that all the vertices project into the diagonal $x+y = 1$ or outside these triangles.
Now, looking on a Fig.~\ref{f:t-red}, we see that each of these triangles is decomposed into a rectangle and two more triangles, each of these triangles decomposes into a rectangle and two more, etc. Each new rectangle corresponds to a vertex $(x,1-x)$ with rational $x$, and it is interesting to study the order, in which these vertices appear, and also prove it formally. To do so, note that any of these triangles is of the form
$$\Delta_{\textstyle{\frac{a}{b},\frac{c}{d}}}:=\{(x,y)\in [0;1)^2 \mid x+y>1,\, x< \frac{c}{d},\, y< 1-\frac{a}{b} \} $$
for some $\frac{a}{b}<\frac{c}{d}$.
Let us show that the fractions $\frac{a}{b}$ and $\frac{c}{d}$ are adjacent in one of the Farey series, and the rectangle that subdivides this triangle starts from a point with the abscise that is the \emph{mediant} (``freshman sum'') $\frac{a+c}{b+d}$ of $\frac{a}{b}$ and~$\frac{c}{d}$; see Fig.~\ref{f:induction}. Indeed, the subdivisions that we are observing are given by removing of the slices $\{(x,y) \mid R_{ab}(x,y)-1\ge r\}$ for some $r>0$. As $R_{ab}(\frac{p}{q},1-\frac{p}{q})-1=\frac{1}{q}$ for coprime $p,q$, the vertices appearing in any such slice are exactly those whose abscises belong to $[\frac{1}{r}]$-th Farey sequence. And when the triangle $\Delta_{\textstyle{\frac{a}{b},\frac{c}{d}}}$ is subdivided, it corresponds to the subdivision of an interval $[\frac{a}{b},\frac{c}{d}]$ of a Farey sequence; {as it is well-known (see~\cite{F2}), it is the mediant that subdivides it}.
Thus, the Farey sequences are not only a good tool for the study of the ziggurat, but they appear in this study in a natural way. We conclude this section by noticing an interesting consequence of the above arguments: we see that two fractions are neighbours in some Farey series if and only if the rectangles, starting at the corresponding points, have common segment.
\section*{Acknowledgements}
Author is exceedingly grateful to Danny Calegari, Walter Neumann, Alden Walker and Subhadip Chowdhury for their comments and cooperation and to Victor Kleptsyn for discussions and assistance.
|
\section{Introduction}
\subsection{Notations} Throughout this paper $R$ is a commutative Noetherian ring, often local, of prime characteristic $p$. Let $q=p^e$, where $e \in \mathbb{N} = \{ 0, 1, \ldots \}$. Consider the $e$th Frobenius homomorphism $F^e:R\to R$ defined $F(r)=r^q$, for all $r \in R$. For an $R$-module $M$, an $e$th Frobenius action (or Frobenius operator) on $M$ is an additive map $\phi: M \to M$ such that $\phi(rm)= r^{p^e}\phi(m)$, for all $r \in R, m \in M$. For any $e \geq 0$, we let $R^{(e)}$ be the $R$-algebra defined as
follows: as a ring $\R{e}$ equals $R$ while the $R$-algebra structure
is defined by $r \cdot s = r^{q} s$, for all $r \in R,\, s \in
R^{(e)}$. Also,
$\R{e}$ as an $\R{e}$-algebra is simply $R$ as an $R$-algebra. Similarly, for an $R$-module $M$, we can define
a new $R$-module structure on $M$ by letting $r * m = r^{p^e}m$, for all $r\in R, m\in M$. We denote this
$R$-module by $\M{e}$.
Consider now an $e$th Frobenius action, $\phi: M \to M$, on $M$, which
is no other than an $R$-module homomorphism $\phi : M \to \M{e}$. Such
an action naturally defines an $R$-module homomorphism $f_{\phi}:
\R{e} \otimes_R M \to M$, where $f_{\phi} (r \otimes m) = r \phi(m)$,
for all $r \in R, m \in M$. Here, $\R{e}$ has the usual structure
(i.e., without twisting) as
an $R$-module given by $\R{e}=R$ on the left, while on the right we
have the twisted module structure via the Frobenius action.
Let $\mathscr{F}^e(M)$ be the collection of all $e$th Frobenius operators on
$M$. The $R$-module structure on $\mathscr{F}^e{(M)}$ is given by viewing
$\M{e}$ as an $R$-module without twisting, that is,
$(r\phi)(x) = r\phi(x)$ for every $r \in R,\, \phi \in \mathscr{F}^e{(M)}$
and $x \in M$.
\begin{Def}\label{def:sfe}
We define {\it the algebra of Frobenius operators} on $M$ by
$$\mathscr{F}{(M)} = \oplus_{e\geq 0} \mathscr{F}^e{(M)},$$
with the multiplication on $\mathscr{F}{(M)}$ determined by composition of
functions; that is, if $\phi \in \mathscr{F}^e{(M)}, \psi \in \mathscr{F}^{e'}{(M)}$
then $\phi \psi:=\phi\circ \psi \in \mathscr{F}^{e+e'}(M)$. Hence, in
general, $\phi \psi \neq \psi \phi$.
\end{Def}
Note that $\mathscr{F}^0{(M)} = {\rm End}_R(M)$, which is a subring of
$\mathscr{F}(M)$. Naturally, each $\mathscr{F}^e{(M)}$ is a module over
$\mathscr{F}^0{(M)}$. Since $R$ maps canonically to $\mathscr{F}^0{(M)}$, this makes
$\mathscr{F}^e{(M)}$ an $R$-module by restriction of scalars. Note that $(\phi
\circ r)(m) = \phi(rm) = (r^q\phi)(m)$, for all $r \in R, m \in
M$. Therefore, $\phi r = r^q \phi$, for all $r\in R, \, \phi \in
\mathscr{F}^e{(M)}$, $q=p^e$.
\subsection{The Frobenius Complexity}
The main concept studied in this paper is the Frobenius complexity of
a local ring $R$, which was introduced in~\cite{EY}. In fact, the
results in this subsection, if not referenced otherwise, are taken
from \cite{EY}. We first need to review the
definition of the complexity of a graded ring.
\begin{Def}Let $A= \oplus_{e\geq 0} A_e$ be a $\mathbb{N}$-graded ring, not
necessarily commutative.
\begin{enumerate}
\item Let $G_e(A)= G_e$ be the subring of $A$ generated by the elements of
degree less or equal to $e$. (So $k_0 = 0$.)
We agree that $G_{-1} = A_0$.
\item We use $k_e=k_e(A)$ to denote the minimal number of homogeneous
generators of $G_e$ as a subring of $A$ over $A_0$. We say that $A$
is {\it degree-wise finitely generated} if $k_e < \infty$ for all
$e$. We agree that $k_{-1} = 0$.
\item For a degree-wise finitely generated ring $A$, we say that a set
$X$ of homogeneous elements of $A$ minimally generates $A$ if for all
$e$, $X_{\leq e} =\{ a \in X: deg(a) \leq e \}$ is a minimal set of
generators for $G_e$ with $k_e = |X_{\leq e} |$ for every $e \ge 0$. Also,
let $X_e= \{ a \in X: deg(a)=e \}$.
\end{enumerate}
\end{Def}
\begin{Prop} \label{prop:minimal-gen}
With the notations introduced above,
let $X$ be a set of homogeneous elements of $A$. Then
\begin{enumerate}
\item The set $X$ generates $A$ as a ring over $A_0$ if and only if
$X_{\le e}$ generates $G_e$ as a ring over $A_0$ for all $e \ge 0$
if and only if the image of $X_{e}$ generates
$\frac{A_{e}}{(G_{e-1})_{e}}$ as an $A_0$-bimodule for all $e \ge 0$.
\item Assume that $A$ is degree-wise finitely generated
$\mathbb{N}$-graded ring and $X$ generates $A$ as a ring over $A_0$.
The set $X$ minimally generates $A$ as a ring over $A_0$ if and
only if $|X_{e}|$ is the minimal number of generators (out of all
homogeneous generating sets) of $\frac{A_{e}}{(G_{e-1})_{e}}$ as an
$A_0$-bimodule for all $e \ge 0$.
\end{enumerate}
\end{Prop}
\begin{Cor} \label{cor:minimal-gen}
Let $A$ be a degree-wise finitely generated $\mathbb{N}$-graded ring
and $X$ a set of homogeneous elements of $A$. Then
\begin{enumerate}
\item The minimal number of generators of $\frac{A_{e}}{(G_{e-1})_{e}}$ as an
$A_0$-bimodule is $k_e - k_{e-1}$ for all $e \ge 0$.
\item If $X$ is generates $A$ as a ring over $A_0$ then
$|X_e| \ge k_e - k_{e-1}$ for all $e \ge 0$.
\end{enumerate}
\end{Cor}
\begin{Def}
\label{degreewise}
Let $A$ be a degree-wise finitely generated ring.
The sequence $\{k_e\}_e$ is called the {\it growth} sequence for $A$. The {\it complexity} sequence is given by $\{ c_{e}(A)= k_{e}-k_{e-1} \}_{e\geq 0}$.
The {\it complexity} of $A$ is
$$\inf \{ n \in \mathbb{R}_{> 0}: c_e(A)=k_{e} - k_{e-1} = O(n^e) \}$$
and it is denoted by $\operatorname{cx}(A)$. If there is no $n >0$ such that $c_e(A)= O(n^e)$, then we say that $\operatorname{cx}(A)=\infty$.
\end{Def}
\begin{Def}\label{def:nearly-onto}
Let $A$ and $B$ be $\mathbb N$-graded rings and $h \colon A \to B$ be
a graded ring homomorphism. We say that $h$ is \emph{nearly onto} if
$B = B_0[h(A)]$ (that is, $B$ as a ring is generated by $h(A)$ over $B_0$).
\end{Def}
\begin{Thm
\label{thm:nearly-onto}
Let $A$ and $B$ be $\mathbb N$-graded rings that are degree-wise
finitely generated. If there
exists a graded ring homomorphism $h \colon A \to B$ that
is nearly onto, then $c_e(A) \ge c_e(B)$ for all $e \ge 0$.
\end{Thm}
\begin{Def}
Let $A$ be a $\mathbb{N}$-graded ring such that there exists a ring homomorphism $R \to A_0$, where $R$ is a commutative ring. We say that $A$ is a (left) $R$-{\it skew algebra} if $aR \subseteq Ra$ for all homogeneous elements $a \in A$. A right $R$-skew algebra can be defined analogously. In this paper, our $R$-skew algebras will be left $R$-skew algebras and therefore we will drop the adjective `left' when referring it to them.
\end{Def}
\begin{Cor}
\label{interpretation}
Let $A$ be a degree-wise finitely generated $R$-skew algebra such that
$R=A_0$. Then $c_{e}(A)$ equals the minimal number of generators of
$\frac{A_{e}}{(G_{e-1})_{e}}$ as a left $R$-module for all $e$.
\end{Cor}
We are now in position to state the definition of the Frobenius
complexity of a local ring of prime characteristic.
\begin{Def} \label{def-FCX}
Let $(R,\fm,k)} %{\text{$(R,\fm,k)$}$ be a local ring of prime characteristic $p$.
We define the {\it Frobenius complexity} of the ring $R$ by
$$\operatorname{cx}_F(R) = \log _p (\operatorname{cx}(\mathscr{F}{(E)})).$$
Also, denote $k_e(R) : = k_e (\mathscr{F}{(E)})$, for all $e$, and call these
numbers the {\it Frobenius growth sequence} of $R$. Then $c_e= c_e(R):
= k_{e}(R)-k_{e-1}(R)$ defines the {\it Frobenius complexity sequence}
of $R$. If the Frobenius growth sequence of the ring $R$ is eventually
constant (i.e., $\operatorname{cx}(\mathscr{F}{(E)}) = 0$), then the Frobenius complexity of
$R$ is set to be $-\infty$.
If $\operatorname{cx}(\mathscr{F}(E)) = \infty$, the Frobenius complexity if $R$ is set to
be $\infty$.
\end{Def}
Katzman, Schwede, Singh and Zhang have introduced an important $\mathbb{N}$-graded ring in their paper~\cite{KSSZ}, which is an example of an $R$-skew algebra. We will study the complexity of this skew-algebra in this section, and apply these results to the complexity of the ring $R$ in subsequent sections.
\begin{Def}[\cite{KSSZ}]
Let $\mathscr{R}$ be an $\mathbb{N}$-graded commutative ring of prime
characteristic $p$ with $\mathscr{R}_0 =R$.
Define $T(\mathscr{R}):= \oplus_{e\geq 0} \mathscr{R} _{p^e-1}$, which is an
$\mathbb{N}$-graded ring by
$$ a *b = ab^{p^e}$$
for all $a \in\mathscr{R} _{p^e-1} ,\, b\in \mathscr{R} _{p^{e'}-1}$.
The degree $e$ piece of $T(\mathscr{R})$ is $T_e(\mathscr{R})=\mathscr{R} _{p^e-1}$.
\end{Def}
A number of results have been proved about the Frobenius complexity of
a local ring and they are summarized below.
\begin{Thm}[\cite{EY}, Corollary 2.12, Theorems 4.7, 4.9]
Let $(R,\fm,k)} %{\text{$(R,\fm,k)$}$ be a local ring.
\begin{enumerate}
\item
If $R$ is $0$-dimensional then $\operatorname{cx}_F(R) = -\infty$.
\item
If $R$ is normal, complete and has dimension at most two, then
$\operatorname{cx}_F(R) \leq 0$.
\item
If $R$ is normal, complete and has a finitely geneated anticanonical
cover, then $\operatorname{cx}_F(R) <\infty$.
\end{enumerate}
\end{Thm}
In addition the following holds.
\begin{Thm}[\cite{KSSZ} Proposition 4.1 and \cite{EY} Theorem 4.5]
If $(R,\fm,k)} %{\text{$(R,\fm,k)$}$ is normal and $\mathbb{Q}$-Gorenstein, then the order of its canonical module in the divisor class group is relatively prime to $p$ if and only if $\operatorname{cx}_F(R) =-\infty$.
\end{Thm}
As in \cite{EY}, we will also use the following notations and
terminologies in the sequel:
For an integer $a \in \mathbb N$, if $a= c_{n}p^{n} + \cdots + c_1p +
c_0$ with $0 \leq c_i \leq p-1$ for all $0 \le i \le n$,
then we use $a = \overline{c_{n} \cdots c_0}$ to denote the base $p$
expression of $a$.
Also, we write $a \vert_e$ to denote the remainder of $a$
when dividing to $p^e$. Thus, if $a = \ol{c_{n} \cdots c_0}$
then $a \vert_e = \overline{c_{e-1} \cdots c_0}$,
which we refer to as the $e$th truncation of $a$.
Put differently, $a \vert_e = a-\floor{\frac a{p^e}}p^e$, in which
$\floor{\frac a{p^e}}$ is the floor function of $\frac a{p^e}$.
When adding up integers $a_i \in \mathbb N$ with $1 \le i \le m$, all
written in base $p$ expressions, we can talk about the carry over to
digit corresponding to $p^e$, which is simply
$\floor{\frac {a_1 \vert_e + \dotsb + a_m\vert_e}{p^e}}$.
These notations depend on the choice of $p$, which should be clear
from the context.
For any positive integers $p$ and $m$ (with $p$ prime), denote by
$M_{p,m}(i)$ (or simply $M(i)$ if $p$ and $m$ are understood) the rank
of $(R[x_1,\,\dotsc,\, x_m]/(x_1^p,\,\dotsc,\, x_m^p))_i$ over $R$, for
all $i \in \mathbb Z$. This is clearly independent of $R$.
Observe that $M_{p,m} =0$ exactly when $i > d(p-1)$ or $i < 0$.
In fact, all $M_{p,m}(i)$ can be read off from the following Poincar\'e
series (actually a polynomial):
\[
\sum_{i = -\infty}^{\infty} M_{p,m}(i)t^i = \left(\frac{1-t^p}{1-t}\right)^m
=\left(1+\dotsb + t^{p-1}\right)^m.
\]
\subsection{Determinantal rings} \label{sec:det}
In this paper we consider the determinantal ring $K[X]/I$ where
$X$ is an $m \times n$ matrix of indeterminates and $I$ is the ideal
of all the $2 \times 2$ minors of $X$ and $K$ a field.
This ring is isomorphic to the Segre product of
$K[x_1,\dotsc, x_{m}]$ and $K[y_1,\dotsc, y_{n}]$.
Recall that, for $\mathbb N$-graded commutative rings
$A = \oplus_{i \in \mathbb N}A_i$ and $B = \oplus_{i \in \mathbb N}B_i$
such that $A_0 = R = B_0$, their Segre product is
\[
A \segre B = \oplus_{i \in \mathbb N}(A_i \otimes_R B_i),
\]
which is a ring under the natural operations.
\begin{Def}
\label{def:segre-complete}
Let $S_{m,n}$ denote the completion of
$K[x_1,\dotsc, x_{m}] \segre K[y_1, \dotsc, y_{n}]$
with respect to the ideal generated by all homogeneous elements of
positive degree, in which $K$ is a field and $m > n \ge 2$.
It is easy to see that
\begin{align*}
S_{m,n} &\cong \prod_{\alpha \in \mathbb N^{m},\,
\beta \in \mathbb N^{n},\, |\alpha| = |\beta|}
Kx^{\alpha}y^{\beta} \\
&=\left\
\sum_{|\alpha| = |\beta|}a_{\alpha,\, \beta}x^{\alpha}y^{\beta}
\; \Big | \;
a_{\alpha,\, \beta} \in K, \, \alpha \in \mathbb N^{m},\,
\beta \in \mathbb N^{n}\right\}
\subset K[[x_1, \dotsc,x_{m}, y_1, \dotsc, y_{n}]].
\end{align*}
Let $\mathscr R_{m,n}$ be the anticanonical cover of $S_{m,n}$.
\end{Def}
The anticanonical cover of such a ring was described by Kei-ichi Watanabe.
\begin{Thm} [{\cite[page 430]{Wa}}] \label{thm:Watanabe}
Let $K$ be a field and $m > n \ge 2$.
The anticanonical cover of the Segre product of $K[x_1, \dotsc,
x_{m}]$ and $K[y_1, \dotsc, y_{n}]$ is isomorphic to
\[
\bigoplus_{i \in \mathbb N}
\left(\bigoplus_{\alpha \in \mathbb N^{m},\,
\beta \in \mathbb N^{n},\, |\alpha| - |\beta|=i(m-n)}
Kx^{\alpha}y^{\beta}\right),
\]
in which the grading is governed by $i$.
Here, for $\alpha = (a_1, \dotsc, a_m)$ and $\beta = (b_1, \dotsc,
b_{n})$ we denote $x^{\alpha} = x_1^{a_1} \dotsm x_m^{a_m}$ and
$y^{\beta} = y_1^{b_1} \dotsm y_{n}^{b_{n}}$.
\end{Thm}
It follows from Theorem~\ref{thm:Watanabe} that
\[
\mathscr R_{m,n} \cong \bigoplus_{i \in \mathbb N}
\left(\prod_{\alpha \in \mathbb N^{m},\,
\beta \in \mathbb N^{n},\, |\alpha| - |\beta|=i(m-n)}
Kx^{\alpha}y^{\beta}\right),
\]
in which the grading is governed by $i$.
\begin{Lma}[\cite{EY}]
\label{lma:nearly-onto}
Let $A$ and $B$ be degree-wise finitely generated $\mathbb N$-graded
commutative rings and
$h \colon A \to B$ be a graded ring homomorphism.
\begin{enumerate}
\item The homomorphism $h$ is nearly onto if and only if $B_i$
is generated by $h(A_i)$ as a $B_0$-module for all $i \in \mathbb N$
(that is, $B$ is generated by $h(A)$ as a $B_0$-module).
\item If $A$ and $B$ have prime characteristic $p$ and $h$
is nearly onto, then the induced graded homomorphism
$T(h) \colon T(A) \to T(B)$ is nearly onto.
\end{enumerate}
\end{Lma}
\begin{Cor}\label{cor:nearly-onto}
Let $A$ and $B$ be $\mathbb N$-graded commutative rings of prime
characteristic $p$.
If there exists a graded ring homomorphism $h \colon A \to B$ that
is nearly onto, then $c_e(T(A)) \ge c_e(T(B))$ for all $e \ge 0$.
\end{Cor}
\begin{Prop}[{Compare with \cite[Proposition~5.5]{EY}}]
\label{prop:nearly-onto}
Let $K$, $S_{m,n}$ and $\mathscr{R}_{m,n}$ be as in Definition~\ref{def:segre-complete}
with $m > n \ge 2$.
Then there are nearly onto graded ring homomorphisms from $\mathscr{R}_{m,n}$ to
$V_{m-n}(K[x_1, \dotsc, x_{m}])$ and vice versa, in which
$V_{m-n}(K[x_1, \dotsc, x_{m}])$ denotes the $(m-n)$-Veronese subring
of $K[x_1, \dotsc, x_{m}]$.
\end{Prop}
\begin{proof}
In light of Definition~\ref{def:segre-complete} and
Theorem~\ref{thm:Watanabe}, we simply assume
\begin{align*}
\mathscr R_{m,n} = \bigoplus_{i \in \mathbb N}
\left(\prod_{\alpha \in \mathbb N^{m},\,
\beta \in \mathbb N^{n},\, |\alpha| - |\beta|=i(m-n)}
Kx^{\alpha}y^{\beta}\right).
\end{align*}
Define $\phi \colon \mathscr{R}_{m,n} \to V_{m-n}(K[x_1, \dotsc, x_{m}])$
and $\psi \colon V_{m-n}(K[x_1, \dotsc, x_{m}]) \to \mathscr{R}_{m,n}$ by
\begin{align*}
\phi(f(x_1,\dotsc, x_m,\, y_1, \dotsc, y_{n}))
&= f(x_1,\dotsc, x_m,\, 0, \dotsc, 0) \in K[x_1, \dotsc, x_{m}]\\
\text{and} \qquad
\psi(g(x_1,\dotsc, x_m)) &= g(x_1,\dotsc, x_m) \in \mathscr{R}_{m,n},
\end{align*}
for all $f(x_1,\dotsc, x_m,\, y_1, \dotsc, y_{n}) \in \mathscr{R}_{m,n}$
and all $g(x_1,\dotsc, x_m) \in V_{m-n}(K[x_1, \dotsc, x_{m}])$.
It is routine to verify that both $\phi$ and $\psi$ are graded ring
homomorphisms.
As $\phi \circ \psi$ is the identity map, we see that $\phi$ is onto
and hence nearly onto.
Finally, note that for every $i \in \mathbb N$, $(\mathscr{R}_{m,n})_i$ is
generated by $\psi(V_{m-n}(K[ x_1, \dotsc, x_{m}])_i)
= \psi(K[ x_1, \dotsc, x_{m}]_{i(m-n)})$ as a module over
$(\mathscr{R}_{m,n})_0 = S_{m,n}$. So $\psi$ is nearly onto, completing the proof.
\end{proof}
\begin{Thm}\label{thm:segre}
Let $K$, $S_{m,n}$ and $\mathscr{R}_{m,n}$ be as in
Definition~\ref{def:segre-complete}
with $m > n \ge 2$.
\begin{enumerate}
\item Then $\mathscr{R}_{m,n}$ and $V_{m-n}(K[x_1,\dotsc, x_{m}])$ have the same
complexity sequence.
\item If $K$ has prime characteristic $p$, then $T(\mathscr{R}_{m,n})$ and
$T(V_{m-n}(K[x_1,\dotsc, x_{m}]))$ have the same complexity sequence.
\item If $K$ has prime characteristic $p$, then
\[
\operatorname{cx}(\mathscr{F}(E_{m,n})) = \operatorname{cx}(T(\mathscr{R}_{m,n})) = \operatorname{cx}(T(V_{m-n}(K[x_1,\dotsc, x_{m}]))),
\]
in which $E_{m,n}$ stands for the injective hull of the residue field
of $S_{m,n}$. Consequently,
\[
\operatorname{cx}_F(S_{m,n}) = \log_p\operatorname{cx}(T(V_{m-n}(K[x_1,\dotsc, x_{m}]))).
\]
\end{enumerate}
\end{Thm}
\begin{proof}
This follows from Corollary~\ref{cor:nearly-onto},
Proposition~\ref{prop:nearly-onto} and \cite[Theorem~3.3]{KSSZ}.
\end{proof}
In summary, to compute the Frobenius complexity of $S_{m,n}$ with
$m > n \ge 2$, it suffices to study $T(V_{r}(K[x_1,\dotsc, x_{m}]))$
with $r = m-n$ (hence $0< r \le m-2$).
The next section is devoted to the study of $T(V_{r}(K[x_1,\dotsc,
x_{m}]))$, more generally with $1 \le r,\,m \in \mathbb N$.
\section{Investigating $T(V_r(R[x_1,\,\dotsc,\,x_m]))$}
Let $R$ be a commutative ring of prime characteristic $p$ and $r,\,m$
positive integers.
In this section, we study $T(V_r(R[x_1,\,\dotsc,\,x_m]))$. In
particular, we are interested in when it is finitely
generated over $R$, as well as how to compute its complexity.
To simplify notation, denote the following (with $R$, $p$, $m$ and $r$
understood):
\begin{itemize}
\item $\mathscr{R} := R[x_1,\,\dotsc,\,x_m]$.
\item $\mathscr{V} := V_r(\mathscr{R}) = V_r(R[x_1,\,\dotsc,\,x_m])$.
\item $T: = T(\mathscr{V}) =T(V_r(R[x_1,\,\dotsc,\,x_m]))$.
\item $G_e:= G_{e}(T)$.
\item $T_e:=T_e(\mathscr{V}) = T_e(V_r(R[x_1,\,\dotsc,\,x_m])) =
\mathscr{R}_{r(p^e-1)} = (R[x_1,\,\dotsc,\,x_m])_{r(p^e-1)}$.
As there are several gradings going on, when we say the degree of a
monomial, we agree that it refers to its (total) degree in $\mathscr{R} =
R[x_1,\,\dotsc,\,x_m]$. Thus a monomial in $T_e$ is a monomial of
(total) degree $r(p^e-1)$.
Note that $T_e=\mathscr{R}_{r(p^e-1})$ is an $R$-free (left) module with a basis
consisting of monomials of (total) degree $r(p^e-1)$.
In particular, $T_0 = R$.
\end{itemize}
Fix any $e \in \mathbb N$. We see that $G_{e-1} = G_{e-1}(T)$ is
an $R$-free (left) module with a basis consisting of monomials that
can be expressed as products (under $*$, the multiplication of $T$) of
monomials of degree $r(p^i-1)$ where
$i \leq e-1$. So all such monomials of total degree $r(p^e-1)$ form an
$R$-basis of $(G_{e-1})_e$.
In conclusion, $\frac{T_e}{(G_{e-1})_e}$ is free as a left $R$-module
with a basis given by monomials of degree $r(p^e-1)$ that cannot be
written as products (under $*$) of monomials of degree $r(p^i-1)$,
with $i \leq e-1$. We will refer to this basis as the \emph{monomial
basis} of $\frac{T_e}{(G_{e-1})_e}$. By Corollary~\ref{interpretation},
we see $c_e(T) = \operatorname{rank}_R(\frac{T_e}{(G_{e-1})_e})$.
As $c_0(T) = 0$ and $c_1(T) = \operatorname{rank}_R(T_1) = \operatorname{rank}_R(\mathscr{R}_{r(p-1)})$, we
may assume $e \ge 2$ in the following discussion.
Let $\alpha = (a_1, \dotsc, a_m) \in \mathbb N^m$ such that
$|\alpha| := a_1 + \dotsb + a_m = r(p^e-1)$, so that
$x^{\alpha} := x_1^{a_1} \cdots x_m^{a_m}$ is a monomial in $T_e$
(i.e., of degree $r(p^e-1)$). This monomial
$x^{\alpha}$ belongs to $(G_{e-1})_e$ if and only if it can be decomposed as
\[
x^{\alpha} = x^{\alpha'} * x^{\alpha''}= x^{\alpha' + p^{e'}\alpha''}
\]
for some $x^{\alpha'} \in T_{e'},\, x^{\alpha''}\in T_{e''}$
with $1 \leq e',\,e'' \le e-1$ and $e' +e'' =e$.
In other words, $x^{\alpha} \in (G_{e-1})_e$ if and only if there is
an equation
\[
{\alpha} = {\alpha' + p^{e'}\alpha''}
\]
for some $\alpha',\,\alpha'' \in \mathbb N^m,\,
1 \leq e' \le e-1,\, e'+e''=e$ with
$|\alpha'| =r(p^{e'}-1)$ and $|\alpha''| = r(p^{e''}-1)$,
which is equivalent to the existence of equations
\[
a_i = a_i' + p^{e'}a_i'' \quad \text{for all} \quad i \in \{1,\dotsc, m\}
\]
for some $(a_1',\dotsc,a_m'),\,(a_1'',\dotsc,a_m'') \in \mathbb N^m,\,
1 \leq e' \le e-1,\, e'+e''=e$ with
$\sum_{i=1}^m a'_i =r(p^{e'}-1)$ and $\sum_{i=1}^m a''_i = r(p^{e''}-1)$.
Now it is routine to see that the above holds if and only if there
exist $(a_1',\dotsc,a_m') \in \mathbb N^m$ and $1 \leq e' \le e-1$ with
$\sum_{i=1}^m a'_i =r(p^{e'}-1)$ such that
\[
a_i\vert_{e'} \le a_i' \le a_i \ \text{ and }\ a_i\vert_{e'} = a_i'\vert_{e'}
\quad \text{for all}
\quad i \in \{1,\dotsc, m\},
\]
which can be seen to be equivalent to the existence of an integer
$1 \leq e' \leq e-1$ such that
\[
a_1 \vert_{e'} + \cdots + a_m \vert_{e'} \le r(p^{e'}-1),
\]
which is equivalent to the existence of an integer
$1 \leq e' \leq e-1$ such that
\[
\floor{\frac {a_1 \vert_{e'} + \cdots + a_m \vert_{e'}}{p^{e'}}}
\le \floor{\frac {r(p^{e'}-1)}{p^{e'}}}.
\]
Note that the backward implications of
the last two equivalences rely on the fact that
$a_1\vert_{e'} + \cdots + a_m \vert_{e'}$ and $r(p^{e'}-1)$
are in the same congruence class modulo $p^{e'}$; the backward
implications of the next to last equivalence also relies on the fact
$a_i\vert_{e'} \equiv a_i \mod p^{e'}$ for all $i$, which allows us to
reverse-engineer $(a_1',\dotsc,a_m') \in \mathbb N^m$ as desired.
With the argument above, we establish the following result. (Again, the fact
$a_1\vert_{i} + \cdots + a_m \vert_{i} \equiv r(p^{i}-1) \mod p^{i}$
is needed in part~(2) of the following proposition.)
\begin{Prop} \label{prop:good/bad}
Consider $T = T(V_r(R[x_1,\,\dotsc,\,x_m]))$, in prime characteristic $p$.
\begin{enumerate}
\item For any monomial $x_1^{a_1} \cdots x_m^{a_m} \in T_e$ with $e \ge 1$,
the following are equivalent.
\begin{itemize}
\item $x_1^{a_1} \cdots x_m^{a_m} \in G_{e-1}(T)$.
\item There exists an integer $i$, $1 \le i
\le e-1$, such that the carry-over to the digit associated
with $p^{i}$ is less than or equal to $\floor{\frac {r(p^{i}-1)}{p^{i}}}$
when $a_1 + \dotsb + a_m$ is calculated in base $p$.
\end{itemize}
\item For any monomial $x_1^{a_1} \cdots x_m^{a_m} \in T_e$ with $e \ge 1$,
the following are equivalent.
\begin{itemize}
\item $x_1^{a_1} \cdots x_m^{a_m} \notin G_{e-1}(T)$.
\item $a_1 \vert_{i} + \cdots + a_m \vert_{i} = r(p^{i}-1) +
d_{i}p^{i}$ with $1 \le d_{i} \in \mathbb N$ for all $1 \le i \le e-1$.
\item The carry-over to the digit associated with $p^{i}$ is greater than
$\floor{\frac {r(p^{i}-1)}{p^{i}}}$ for all $1 \le i \le e-1$
when $a_1 + \dotsb + a_m$ is calculated in base $p$.
\end{itemize}
\end{enumerate}
\end{Prop}
\begin{Prop} \label{prop:count}
For $T = T(V_r(R[x_1,\,\dotsc,\,x_m]))$, $c_e(T)$ is the number of monomials
$x_1^{a_1} \cdots x_m^{a_m} \in T_e$ such that
the carry-over to the digit associated
with $p^{i}$ is bigger than $\floor{\frac {r(p^{i}-1)}{p^{i}}}$
for all $1 \le i\le e-1$
when $a_1 + \dotsb + a_m$ is calculated in base $p$.
\end{Prop}
Using the criteria given in Proposition~\ref{prop:good/bad}, we are
able to determine precisely when
$T(V_r(R[x_1,\,\dotsc,\,x_m]))$ is finitely generated over $T_0 = R$.
\begin{Thm}\label{thm:mr}
Let $T = T(V_r(R[x_1,\,\dotsc,\,x_m]))$, with $r,\,m,\,R$ as above.
\begin{enumerate}
\item If $r \ge m-1$, then $T$ is generated by $T_1$ over $T_0$ (that
is, $c_e(T) = 0$ for all $e \ge 2$).
\item If $r < m-1$, then $c_e(T) > 0$ (i.e.,
$T_e$ is not generated by lower degree) for all $e \ge 1$.
\item The ring $T(V_r(R[x_1,\,\dotsc,\,x_m]))$ is finitely generated over
$R$ if and only if $r \ge m-1$.
\end{enumerate}
\end{Thm}
\begin{proof}
Evidently, we only need to prove (1) and (2).
(1) Suppose, on the contrary, that for some $e \ge 2$ there exists a
monomial $x_1^{a_1} \cdots x_m^{a_m} \in T_e$ that does not belong to
$G_{e-1}(T)$. Then by Proposition~\ref{prop:good/bad}
\[
a_1 \vert_{i} + \cdots + a_m\vert_{i} \ge r(p^{i}-1) + p^{i}
\]
for all $1 \le i \le e-1$. However, the assumption $r \ge m-1$
implies
\[
a_1 \vert_{i} + \cdots + a_m \vert_{i} \le m(p^{i}-1) \le
(r+1)(p^{i}-1) < r(p^{i}-1) + p^{i}.
\]
We get a contradiction.
(2) As $c_1(T) > 0$ is clear, we assume $e \ge 2$. Consider
\[
x_1^{p^e-1}\dotsm x_{r-1}^{p^e-1}
x_r^{p^e-p^{e-1}-1}x_{r+1}^{p^{e-1}-1}x_{r+2}^{1}
\in \mathscr{R}_{r(p^e-1)} = T_e.
\]
Now it is routine to see that the carry-over to the digit associated
with $p^{i}$ is $\floor{\frac {r(p^{i}-1)}{p^{i}}} +1$
for all $1 \le i \le e-1$ when $a_1 = p^e-1,\, \dotsc,\,a_{r-1} =
p^e-1$, $a_r = p^e-p^{e-1}-1$, $a_{r+1}={p^{e-1}-1}$, $a_{r+2}={1}$
and $a_i = 0$ (for $r+2 < i \le m$) are added up in base $p$. This
verifies
$x_1^{p^e-1}\dotsm x_{r-1}^{p^e-1}x_r^{p^e-p^{e-1}-1}x_{r+1}^{p^{e-1}-1}x_{r+2}
\notin G_{e-1}(T)$ and hence $c_e(T) > 0$.
\end{proof}
\section{Computing $c_e(T(V_r(R[x_1,\,\dotsc,\,x_m])))$}
\label{sec:computing-ce}
Let $R$, $m$, $r$, $\mathscr{R}$, $\mathscr{V}$ and $T$ be as in last section and keep
the notations.
In particular, $T = T(V_r(R[x_1,\,\dotsc,\,x_m]))$ is an $\mathbb
N$-graded ring. For simplicity, denote $c_e(T)$ by $c_{m,r,e}$ or
simply by $c_e$ since $r$ and $m$ are understood. (It should be clear
that $c_e(T(V_r(R[x_1,\,\dotsc,\,x_m])))$ is independent of $R$. Also
note that $c_1 = \operatorname{rank}_R (\mathscr{R}_{r(p-1)}) = \binom{r(p-1)+m-1}{m-1}$.)
Fix an integer $e \ge 2$.
The goal is to count the number of monomials that produce
the monomial basis of $\frac{T_e}{(G_{e-1})_e}$.
First, we set up some notations.
Let $\alpha = (a_1, \dotsc, a_m) \in \mathbb N^m$ with
$|\alpha| := a_1 + \dotsb + a_m = r(p^e -1)$.
For each $n \in [1,\,m] := \{1,\dotsc,m\}$, write
$a_n = \ol{\cdots a_{n,i} \cdots a_{n,0}}$
in base $p$ expression. Then, for each $i \in [0,\,e-2] := \{0, \dotsc,
e-2\}$, denote
\[
\alpha_i := (a_{1,i}, \dotsc, a_{m, i})\in \mathbb N^m,
\]
which can be referred to as the vector of the digits corresponding to
$p^i$. Also denote
\[
\alpha_{e-1} :=
\left(\floor{\frac{a_{1}}{p^{e-1}}}, \dotsc,\floor{\frac{a_{m}}{p^{e-1}}}\right)
= \left(a_1 - a_1 \vert_{e-1}, \dotsc, a_m - a_m \vert_{e-1}\right)
\in \mathbb N^m.
\]
Moreover, for each $i \in \{0,\dotsc,e-1\}$, let $f_i(\alpha)$ denote the
carry-over to the digit corresponding to $p^{i}$ when computing
$\sum_{i=1}^ma_i$ in base $p$. In other words,
\[
f_i(\alpha) := \floor{\frac{a_1 \vert_{i} + \dotsb + a_m \vert_{i}}{p^i}}.
\]
Note that $f_{0}(\alpha) = 0$.
Then denote $f(\alpha) :=(f_{e-1}(\alpha), \dotsc, f_{0}(\alpha))
\in \mathbb N^e$.
Finally, denote
\[
d(\alpha) := (d_{e-1}(\alpha), \dotsc, d_0(\alpha)) : =
f(\alpha) - \left(\floor{\frac {r(p^{e-1}-1)}{p^{e-1}}},
\dotsc, \floor{\frac {r(p^{0}-1)}{p^{0}}}\right)\in \mathbb Z^e,
\]
so that $d_i(\alpha) = f_i(\alpha)-\floor{\frac {r(p^{i}-1)}{p^{i}}}$ for
all $i \in [0,\,e-1] := \{0,\dotsc,e-1\}$. Note that $d_0(\alpha) = 0$.
Moreover, for all $i \in [0,\,e-2]$, we have
\begin{align*}
d_{i+1}(\alpha)
& =
\floor{\frac{a_1 \vert_{i+1} + \dotsb + a_m \vert_{i+1}}{p^{i+1}}} -
\floor{\frac {r(p^{i+1}-1)}{p^{i+1}}} \\
& \overset{\dagger}{=}
\floor{\frac{|\alpha_i| + f_i(\alpha)}{p}} -
\floor{\frac{r(p-1) + \floor{\frac {r(p^{i}-1)}{p^{i}}}}{p}} \\
& \overset{\ddagger}{=}
\frac 1p \left[(|\alpha_i| + f_i(\alpha)) -
\left(r(p-1) + \floor{\frac {r(p^{i}-1)}{p^{i}}}\right)\right] \\
& =
\frac 1p \left[|\alpha_i|
+ \left(f_i(\alpha))- \floor{\frac {r(p^{i}-1)}{p^{i}}}\right) - r(p-1)\right]\\
& =
\frac 1p \big[|\alpha_i| + d_i(\alpha) - r(p-1)\big].
\end{align*}
Note that $\overset{\dagger}{=}$ follows from how we compute the
carry overs to digit corresponding $p^{i+1}$, while
$\overset{\ddagger}{=}$ follows from the fact that
$|\alpha_i| + f_i(\alpha) \equiv r(p-1) + \floor{\frac
{r(p^{i}-1)}{p^{i}}} \mod p$ since they are all congruent to the
(same) number representing the digit associated with $p^i$ in the base
$p$ expression of $r(p^e-1)$ and $r(p^i-1)$.
Let $\alpha = (a_1, \dotsc, a_m) \in \mathbb N^m$ with
$|\alpha| = r(p^e -1)$ as above
and let $\delta = (d_{e-1}, \dotsc,d_0)\in \mathbb Z^e$ with $d_{0} = 0$.
By what we have established above, we see
\begin{align*}
d(\alpha) = d
& \iff d_i(\alpha) = d_i,\,\forall i \in [1,\,e-1] \\
& \iff d_{i+1}(\alpha) = d_{i+1},\,\forall i \in [0,\,e-2] \\
& \iff \frac 1p \big[|\alpha_i| + d_i(\alpha) - r(p-1)\big]
= d_{i+1},\,\forall i \in [0,\,e-2] \\
& \iff
|\alpha_i| + d_i(\alpha) - r(p-1) = d_{i+1}p,\,\forall i \in [0,\,e-2] \\
& \iff
|\alpha_i| + d_i(\alpha) = r(p-1) + d_{i+1}p,\,\forall i \in [0,\,e-2] \\
& \overset{*}{\iff}
|\alpha_i| + d_i = r(p-1) + d_{i+1}p,\,\forall i \in [0,\,e-2] \\
& \iff
|\alpha_i| = r(p-1) + d_{i+1}p - d_i,\,\forall i \in [0,\,e-2].
\end{align*}
Note that $\overset{*}{\implies}$ holds because the assumption (i.e.,
antecedent) of this implication already implies $d(\alpha) =\delta$,
while $\overset{*}{\impliedby}$ follows from an easy induction on
$i$ (in light of the established equation
$d_{i+1}(\alpha)=\frac 1p \big[|\alpha_i| + d_i(\alpha) - r(p-1)\big]$).
Furthermore, the assumption $|\alpha| = r(p^e -1)$ (together with
$d(\alpha) = \delta$) translates to the following
\[
|\alpha_{e-1}| + f_{e-1}(\alpha)
= \floor{\frac{a_1 + \dotsb + a_m}{p^{e-1}}}
= \floor{\frac{r(p^e-1)}{p^{e-1}}}
= r(p-1) + \floor{\frac {r(p^{e-1}-1)}{p^{e-1}}},
\]
which is obtained by examining summations $a_1 + \dotsb + a_m$ and
$\overbrace{(p^e-1) + \dotsb + (p^e-1)}^{r \text{ terms}}$ in base $p$.
Therefore
\[
|\alpha_{e-1}|
= r(p-1) + \floor{\frac {r(p^{e-1}-1)}{p^{e-1}}} - f_{e-1}(\alpha)
= r(p-1)-d_{e-1}(\alpha) = r(p-1)-d_{e-1}.
\]
In summary, with $\alpha \in \mathbb N^m$ and $\delta \in \mathbb Z^e$ with
$d_0 = 0$ as above, we conclude that
$|\alpha| = r(p^e-1)$ and $d(\alpha) = \delta$ if and only if
\[
|\alpha_{e-1}| = r(p-1)-d_{e-1}
\quad \text{and} \quad
|\alpha_i| = r(p-1) + d_{i+1}p - d_{i} \quad
\text{for all} \quad i \in \{0, \dotsc, e-2\}.
\]
Now we are ready to formulate $c_e = c_e(T)$ for
$T = T(V_r(R[x_1,\,\dotsc,\,x_m]))$. This result generalizes
\cite[Proposition~3.7]{EY}. Since $c_e = 0$ for all $e \le 2$ when $m
\le r+1$, the formula in the following proposition is most meaningful
when $m-r-1 > 0$.
\begin{Prop}\label{prop:c_e}
For $T = T(V_r(R[x_1,\,\dotsc,\,x_m]))$, we have the following formula:
\begin{align*}
c_{e}&= \sum_{\substack{
(d_{e-1},\, \dotsc,\, d_1,\, d_{0}=0) \in \mathbb N^{e}
\\d_i \ge 1 \text{ for } 1 \le i \le e-1}}
\left(P_m\left(r(p-1)-d_{e-1}\right)
\prod_{i=0}^{e-2} M_{p,m}(r(p-1) + d_{i+1}p - d_{i})\right) \\
&= \sum_{\substack{
(d_{e-1},\, \dotsc,\, d_1,\, d_{0}=0) \in \mathbb N^{e}
\\1 \le d_i \le m-r-1 \text{ for } 1 \le i \le e-1}}
\left(\binom{r(p-1)-d_{e-1}+m-1}{m-1}
\prod_{i=0}^{e-2} M_{p,m}(r(p-1) + d_{i+1}p - d_{i})\right)
\end{align*}
for all $e \ge 2$,
where $P_m(i)$ denotes $\operatorname{rank}_R(R[x_1,\,\dotsc,\,x_m]_i)$, i.e.,
$P_m(i) = \binom{m+i-1}{i} = \binom{m+i-1}{m-1}$.
\end{Prop}
\begin{proof}
Fix any $e \ge 2$ and adopt the notations set up above.
Consider $x^{\alpha} = x_1^{a_1} \dotsm x_m^{a_m} \in T_e$. By
Proposition~\ref{prop:good/bad}, $x^{\alpha} \notin G_{e-1}(T)$ if and
only if
\[
d_i(\alpha) \ge 1 \quad \text{for all} \quad i \in \{1,\,\dotsc,\,e-1\}.
\]
To determine $c_e$, we need to find the number of monomials with the
above property, as stated in Proposition~\ref{prop:count}.
This is equivalent to counting the number of $\alpha \in \mathbb N^m$
such that $|\alpha| = r(p^e -1)$ and $d_i(\alpha) \ge 1$ for all
$i \in [1,\,e-1]$.
Fix any $\delta = (d_{e-1}, \dotsc,d_0)\in \mathbb N^e$ with $d_{0} = 0$
and $d_i \ge 1$ for all $i \in [1,\,e-1]$. We intend to find the
number of $\alpha \in \mathbb N^m$ such that $|\alpha| = r(p^e -1)$
and $d(\alpha) = \delta$, which can be written as
\[
\card{\{\alpha \in \mathbb N^m :
|\alpha| = r(p^e -1) \text{ and } d(\alpha) = \delta\}},
\]
in which $\card X$ stands for the cardinality of any set $X$.
For each $i \in [1,\,e-2]$, the number of ways to realize
$|\alpha_i| = r(p-1) + d_{i+1}p - d_{i}$ is given as follows:
\[
\card{\{\alpha_i \in [0,\,p-1]^m :
|\alpha_i| = r(p-1) + d_{i+1}p - d_{i}\}}
=M_{p,m}(r(p-1) + d_{i+1}p - d_{i}).
\]
The number of ways to realize
$|\alpha_{e-1}| = r(p-1) - d_{e-1}$ is given as follows:
\[
\card{\{\alpha_{e-1} \in N^m :
|\alpha_{e-1}| = r(p-1) - d_{e-1}\}}
=P_{m}(r(p-1) - d_{e-1}).
\]
Therefore, the number of $\alpha \in \mathbb N^m$ such that
$|\alpha| = r(p^e -1)$ and $d(\alpha) = \delta$ is governed by the
following formula:
\begin{multline*}
\card{\{\alpha \in \mathbb N^m :
|\alpha| = r(p^e -1) \text{ and } d(\alpha) = \delta\}} \\
=P_m\left(r(p-1)-d_{e-1}\right)\prod_{i=0}^{e-2} M_{p,m}(r(p-1) + d_{i+1}p - d_{i}).
\end{multline*}
Observe that if $m-r-1 \le 0$, then
\[
\card{\{\alpha \in \mathbb N^m :
|\alpha| = r(p^e -1) \text{ and } d(\alpha) = \delta\}} = 0,
\]
which follows from $M_{p,m}(r(p-1) + d_{1}p - d_{0}) = 0$ since
$r(p-1) + d_{1}p - d_{0} \ge r(p-1) + p = (r+1)(p-1)+1 \ge
m(p-1)+1$; also see Theorem~\ref{thm:mr}(1).
We further observe that, whenever there exists $d_i > m-r-1 >0$ for
some $i \in [1,\,e-1]$, then
\[
\card{\{\alpha \in \mathbb N^m :
|\alpha| = r(p^e -1) \text{ and } d(\alpha) = \delta\}} = 0.
\]
Indeed, pick the least $i \in [1,\,e-1]$ such that $d_i > m-r-1 >0$
and we get $r(p-1) + d_{i}p - d_{i-1} \ge m(p-1)+1$ and hence
$M_{p,m}(r(p-1) + d_{i}p - d_{i-1}) = 0$.
Put differently, when adding $m$ many non-negative integers to
$r(p^e-1)$, the carry overs to digits associated with $p^i$ can not exceed
$\floor{\frac {r(p^{i}-1)}{p^{i}}} + m - r-1$.
Finally, exhausting all $\delta = (d_{e-1}, \dotsc,d_0)\in \mathbb N^e$
with $d_{0} = 0$ and $d_i \ge 1$ for $i \in [1,\,e-1]$, we can
formulate $c_{e}=c_e(T(R[x_1, \dotsc, x_m]))$ as follows:
\begin{align*}
c_{d,e} &
= \sum_{\substack{
(d_{e-1},\, \dotsc,\, d_1,\, d_{0}=0) \in \mathbb N^{e}
\\d_i \ge 1 \text{ for } 1 \le i \le e-1}}
\card{\{\alpha \in \mathbb N^m : |\alpha| = r(p^e-1) \text{ and }
d(\alpha) = (d_{e-1},\, \dotsc,\, d_1,\, d_{0})\}} \\
& = \sum_{\substack{
(d_{e-1},\, \dotsc,\, d_1,\, d_{0}=0) \in \mathbb N^{e}
\\d_i \ge 1 \text{ for } 1 \le i \le e-1}}
\left(P_m\left(r(p-1)-d_{e-1}\right)
\prod_{i=0}^{e-2} M_{p,m}(r(p-1) + d_{i+1}p - d_{i})\right)\\
&= \sum_{\substack{
(d_{e-1},\, \dotsc,\, d_1,\, d_{0}=0) \in \mathbb N^{e}
\\1 \le d_i \le m-r-1 \text{ for } 1 \le i \le e-1}}
\left(\binom{r(p-1)-d_{e-1}+m-1}{m-1}
\prod_{i=0}^{e-2} M_{p,m}(r(p-1) + d_{i+1}p - d_{i})\right),
\end{align*}
which verifies the equations.
\end{proof}
Next, we outline a method that allows us compute $c_{e} =
c_e(T(V_r(R[x_1,\,\dotsc,\,x_m])))$ for any $m,\,r$ with $m \ge r+2$,
in which $R$ may have any prime characteristic $p$.
(Note that, if $m \le r+1$, then $c_{e}= 0$ for all $e \ge 2$, see
Theorem~\ref{thm:mr}.)
The following generalizes \cite[Discussion~3.8]{EY}.
\begin{Dis}\label{dis:m,r,p}
Fix any positive integers $r,\,m$ such that $r+1 < m$, any prime
number $p$, and any ring $R$ with characteristic $p$.
Let $\mathscr{R}=R[x_1,\,\dotsc,\,x_m]$.
We describe a way to determine $c_{e} = c_e(T(V_r(\mathscr{R})))$ explicitly
as follows:
For every $e \ge 0$, denote
\[
X_e :=
\begin{bmatrix}
X_{e,1} \\
\vdots \\
X_{e,m-r-1}
\end{bmatrix}_{(m-r-1) \times 1},
\]
in which
\begin{align*}
X_{e,n} &:= \sum_{\substack{
(d_{e+1}=n,\, d_{e-1},\,\dotsc,\, d_1,\,d_{0}=0) \in \mathbb N^{e+2}
\\1 \le d_i \le m-r-1 \text{ for } 0 \le i \le e}}
\prod_{i=0}^{e} M_{p,m}(r(p-1) + d_{i+1}p - d_{i})
\end{align*}
for all $n \in \{1,\,\dotsc,\,m-r-1\}$.
With these notations, it is straightforward to see that, for all
$i \in [1,\,m-r-1]$,
\begin{align*}
X_{e+1,i} &= \sum_{j=1}^{m-r-1}M_{p,m}(r(p-1) + ip - j)X_{e,j}
\end{align*}
In other words, $X_{e+1}$ can be computed recursively:
\[
X_{e+1} = U \cdot X_e,
\]
where
\[
U:= \begin{bmatrix}
u_{ij}
\end{bmatrix}_{(m-r-1) \times (m-r-1)}
\quad \text{with} \quad
u_{ij}:=M_{p,m}(r(p-1) + ip - j).
\]
Therefore,
\[
X_e = U^{e}\cdot X_0 \quad \text{for all} \quad e \ge 0.
\]
With $m,\,r$ and $p$ given, both $X_0$ and
$U= (u_{ij})_{(m-r-1) \times (m-r-1)}$ can be determined explicitly.
Accordingly, we can compute $X_e = U^{e}\cdot X_0$ explicitly for
all $e \ge 0$.
Finally, for all $e \ge 2$, we can determine $c_{e} =c_{e}(T(V_r(\mathscr{R})))$
explicitly, as follows:
\begin{align*}
c_{e}
&= \sum_{\substack{
(d_{e-1},\, \dotsc,\, d_1,\, d_{0}=0) \in \mathbb N^{e}
\\1 \le d_i \le m-r-1 \text{ for } 1 \le i \le e-1}}
\left(P_{m}(r(p-1)-d_{e-1})\prod_{i=0}^{e-2} M_{p,m}(r(p-1)
+ d_{i+1}p - d_{i})\right)\\
&= \sum_{n = 1}^{m-r-1}\left(P_{m}(r(p-1)-n)
\sum_{\substack{
(d_{e-1}=n,\, \dotsc,\, d_1,\, d_{0}=0) \in \mathbb N^{e}
\\1 \le d_i \le m-r-1 \text{ for } 1 \le i \le e-2}}
\prod_{i=0}^{e-2} M_{p,m}(r(p-1) + d_{i+1}p - d_{i})\right)\\
&=\sum_{n=1}^{m-r-1} P_{m}(r(p-1)-n) X_{e-2,n}
= \sum_{n=1}^{m-r-1} \binom{r(p-1)-n+m-1}{m-1} X_{e-2,n}\\
&=Y_0 \cdot U^{e-2} \cdot X_0,
\end{align*}
where $Y_0 := \begin{bmatrix}
\binom{r(p-1)-1+m-1}{m-1} & \cdots & \binom{r(p-1)-(m-r-1)+m-1}{m-1} \\
\end{bmatrix}_{1 \times (m-r-1)}$.
Consequently, $\operatorname{cx}(T(V_r(\mathscr{R})))$ can be computed.
\end{Dis}
\begin{Def}
In what follows, we call
\[
U(p,r,m)=U:= \begin{bmatrix}
u_{ij}
\end{bmatrix}_{(m-r-1) \times (m-r-1)}
\quad \text{with} \quad
u_{ij}:=M_{p,m}(r(p-1) + ip - j)
\]
as the {\it determining matrix} for $p, r, m$.
\end{Def}
\begin{Thm}\label{thm:m=r+2}
Consider $T = T(V_r(R[x_1,\dotsc,x_m]))$ as above with $m=r+2$. Then
$c_e(T) = \binom{rp}{m-1}\binom{p+m-2}{m-1}^{e-2}\binom{p+m-3}{m-1}$
for all $e \ge 2$ and $\operatorname{cx}(T) = \binom{p+m-2}{m-1}$.
\end{Thm}
\begin{proof}
Adopting all the notations introduced in Discussion~\ref{dis:m,r,p},
we see
\begin{align*}
&X_0 = M_{p,m}(r(p-1) + p) = M_{p,m}(p-2) = P_m(p-2) = \binom{p+m-3}{m-1} >0, \\
&U = M_{p,m}((r+1)(p-1)) = M_{p,m}(p-1) = P_m(p-1) =\binom{p+m-2}{m-1} >0,\\
&Y_0 = P_m(r(p-1)-1) = \binom{r(p-1)-1+m-1}{m-1} = \binom{rp}{m-1}> 0.
\end{align*}
Here we use the fact $M_{p,m}(i) = M_{p,m}(m(p-1)-i)$ for all $i$.
Therefore, for all $e \ge 2$, we obtain
\[
c_e =\binom{r(p-1)+m-2}{m-1}\binom{p+m-2}{m-1}^{e-2}
\binom{p+m-3}{m-1},
\]
which establishes
\[
\operatorname{cx}(T(V_r(R[x_1,\dotsc,x_m))) = \binom{p+m-2}{m-1}
\]
when $m=r+2$.
\end{proof}
\subsection{The Frobenius complexity as $p \to \infty$.} \label{p-to-infinity}
We will maintain the notations from this section, including the
condition $m \geq r+2$ and $r > 0$. The following results are
straightforward and left to the reader. We will comment on their
proofs only when necessary.
\begin{Lma}
\label{calc}
Fix $m>0$ an integer and $p$ a prime number.
\begin{enumerate}
\item
$M_{p,m}(i)= M_{p.m}(m(p-1)-i).$
\item
$M_{p,m}(i) \leq M_{p,m}(j)$ if $0 \leq i \leq j \leq \lceil m(p-1)/2
\rceil$ or $\lceil m(p-1)/2 \rceil \leq j \leq i \leq m(p-1)$.
\end{enumerate}
\end{Lma}
\begin{Lma}
\label{calc2}
For any integers $i, j$ such that $1 \leq i,j \leq m-r-1$, we have
$$ p -3 < p \leq r(p-1)+pi-j \leq m(p-1)-p+3,$$ for all $p \gg 0$.
\end{Lma}
\begin{Def}
For any $t \times s$ matrix $A=(a_{ij})$ with nonnegative entries,
where $t,s$ are positive integers, define $\abs A = \min \{ a_{ij}
\}$ and $\Vert A \Vert = \max \{ a_{ij} \}$.
\end{Def}
The following Lemma is a consequence of Lemmata~\ref{calc} and~\ref{calc2}.
\begin{Lma}
\label{abs}
Given $m$ and $r$, we have the following inequalities:
$$\binom{m-1+p-3}{m-1} \leq \abs U \leq \Vert U \Vert \leq
\binom{m-1+ \lceil \frac{m(p-1)}{2} \rceil}{m-1}$$
for the determining matrix $U = U(p,r,m)$ for all $p \gg 0$.
\end{Lma}
\begin{Lma}
Let $A, B$ be matrices with nonnegative entries of sizes $l \times t$,
respectively $t \times s$, with $l, t, s$ positive integers. Then
\[
t \abs A \cdot \abs B \le \abs {A \cdot B} \le
\Vert{A \cdot B} \Vert \leq t \Vert A \Vert \cdot \Vert B \Vert.
\]
\end{Lma}
Now, let us recall that (cf.~Discussion~\ref{dis:m,r,p})
$$c_e = Y_0 \cdot X_{e-2}= Y_0 \cdot U^{e-2} \cdot X_0,$$
where
\[
X_0 =
\begin{bmatrix}
X_{0,1} \\
\vdots \\
X_{0,m-r-1}
\end{bmatrix}_{(m-r-1) \times 1}
\quad \text{with} \quad
X_{0,i} = M_{p,m}( r(p-1)+ip)
\]
and
\[
Y_0 = \begin{bmatrix}
\binom{r(p-1)-1+m-1}{m-1} & \cdots & \binom{r(p-1)-(m-r-1)+m-1}{m-1} \\
\end{bmatrix}_{1 \times (m-r-1)}.
\]
\begin{Lma}
\label{pos}
For all $p,\,m,\,r$ as above, both $X_0$ and $Y_0$ are non-zero.
\end{Lma}
\begin{proof}
Indeed, $m \ge r+2$ implies $0 \le r(p-1)+p < r(p-1)+2(p-1) \le
m(p-1)$, which implies $X_{0,1} = M_{p,m}(r(p-1)+p) > 0$.
On the other hand, $r(p-1)-1 \ge 0$ implies $r(p-1)-1+m-1 \ge m-1$,
which implies $\binom{r(p-1)-1+m-1}{m-1} > 0$.
\end{proof}
Moreover, both $X_0$ and $Y_0$ have all positive entries
for $p \gg 0$. In fact we can be more precise.
\begin{Lma}
\label{pos}
If $p \geq m-r$, then both $X_0$ and $Y_0$ have all positive entries.
\end{Lma}
\begin{proof}
If $p \geq m-r$, then $0 \leq r(p-1)+ip \leq m(p-1)$ and hence
$M_{p,m}(r(p-1)+ip) > 0$, for all $i =1, \dotsc, m-r-1$.
On the other hand, note that $r(m-r)- m+1 =
-(r-1)(r-m+1) \ge 0$ for all $r = 1, \dotsc, m-2$.
Consequently, if $p \ge m-r$ then for all $i = 1, \dotsc, m-r-1$,
\begin{align*}
r(p-1)-i+m-1 &\ge r(p-1)-(m-r-1) +m-1 \\
&=(rp-m+1) + m-1 \ge (r(m-r) - m +1) + m-1 \ge m-1,
\end{align*}
which leads to $\abs{Y_0} > 0$.
\end{proof}
\begin{Prop} \label{prop:bounds}
We have
$$c_e \leq (m-r-1)^{e-1} \cdot \Vert Y_0 \Vert
\cdot \Vert U \Vert ^{e-2} \cdot \Vert X_0 \Vert$$
and
$$(m-r-1)^{e-1} \cdot \abs{Y_0} \cdot \abs{U}^{e-2} \cdot \abs{X_0}
\leq c_e.$$
(In fact $(m-r-1)^{e-3} \cdot \Abs{Y_0} \cdot \abs{U}^{e-2} \cdot \Abs{X_0}
\leq c_e$.)
Therefore we have that
$$ (m-r-1) \abs{U} \leq \operatorname{cx} (T(V_r(\mathscr{R})) \leq (m-r-1)\Vert U \Vert$$
for $p \gg 0$, where $\mathscr{R} = R[x_1,\,\dotsc,\,x_m]$.
\end{Prop}
\begin{Cor} \label{cor:limit}
Let $\mathscr{R} = R[x_1,\,\dotsc,\,x_m]$. If $p \gg 0$, then
$$ (m-r-1)\binom{m-1+p-3}{m-1} \leq \operatorname{cx} (T(V_r(\mathscr{R})) \leq
(m-r-1) \binom{m-1+ \lceil \frac{m(p-1)}{2} \rceil}{m-1}$$
and therefore
$\lim_{p \to \infty} \log_p\operatorname{cx} (T(V_r(\mathscr{R})) = m-1$.
\end{Cor}
This corollary motivates the definition of Frobenius complexity in characteristic zero, which is given in Section 4, see Definition~\ref{FC0}.
\subsection{Perron-Frobenius} \label{subsec:pf}
We would like to summarize a few things
about square matrices with positive real entries. Any such matrix
admits a real positive eigenvalue $\lambda$ such that all other
eigenvalues have absolute value less than $ \lambda$. We will refer to
this eigenvalue as the Perron root or Perron-Frobenius eigenvalue of
the matrix. This eigenvalue is a simple root of the characteristic
polynomial of the matrix. Moreover, an eigenvector for
$\lambda$ either has all entries positive or has all entries
negative. See \cite{Pe} and \cite{Fr}.
Let $p \gg 0$. Since $U$ has only positive entries by
Lemma~\ref{abs}, let $\lambda$ be the Perron-Frobenius eigenvalue for $U$.
There exists an invertible matrix $P$ such that
$$U = P D P^{-1}$$
where $D$ is the Jordan canonical form of $U$. (We may also take $D$
to be the rational canonical form of $U$ over $\mathbb R$ if we prefer
to stay within $\mathbb R$.)
Without loss of generality, the left upper
corner of $D$ is $\lambda$ (thus all the other entries of the first
row or first column are $0$); that is,
\[
D = \begin{bmatrix}
\lambda & 0 \\
0 & D_1
\end{bmatrix}_{(m-r-1) \times (m-r-1)}
\]
with $D_1$ being an $(m-r-2) \times (m-r-2)$ matrix whose eigenvalues
are all less than $\lambda$ in absolute value.
Hence the first column (row) of $P$
($P^{-1}$) is an eigenvector of $U$ ($U^T$) for $\lambda$.
Thus, without loss of generality, we may assume that the first column of
$P$ and (consequently) the first row of $P^{-1}$ have all positive entries.
Lastly, since both $Y_0$ and $X_0$ are non-zero, the first entries of
both $Y_0 P$ and $P^{-1} X_0$ are positive.
Write $Y_0 P = [a, \ A]$ and $P^{-1} X_0 = [b, \ B]^T$ in block form.
Now the fact that $\lambda$ is the largest eigenvalue in absolute
value implies that
\[
c_e = Y_0 U^{e-2} X_0 = (Y_0 P) D^{e-2} (P^{-1} X_0)
= ab\lambda^{e-2} + AD_1^{e-2}B
=ab\lambda^{e-2}+o(\lambda^{e}).
\]
Thus
\[
\operatorname{cx}(T(V_r(\mathscr{R})))= \lambda.
\]
(The above argument applies as long as $p,\,m,\,r$ are such that
$U$ is all positive, since $X_0$ and $Y_0$ are always non-zero.)
\section{Frobenius complexity of determinantal rings}
In this section, we combine what we have obtained to derive results on
the Frobenius complexity of determinantal rings.
In particular, we translate the results on $T(V_r(R[x_1,\dotsc,x_m]))$ to
$S_{m,n}$ with $m > n \ge 2$.
\begin{Thm}\label{thm:f-com-det}
Let $K$, $S_{m,n}$ and $\mathscr{R}_{m,n}$ be as in Section~\ref{sec:det}
(cf.~Definition~\ref{def:segre-complete}) with $m > n \ge 2$.
Further assume that $K$ is a field of prime characteristic $p$.
Let $E_{m,n}$ denote the injective hull of the residue field of $S_{m,n}$.
\begin{enumerate}
\item The ring of Frobenius operators of $S_{m,n}$ (i.e.,
$\mathscr{F}(E_{m,n})$) is never finitely generated over $\mathscr{F}_0(E_{m,n})$.
\item When $n = 2$, we have $\operatorname{cx}_F(S_{m,2}) = \log_p\binom{p+m-2}{m-1}$.
\item We have $\lim_{p \to \infty}\operatorname{cx}_F(S_{m,n}) = m-1$.
\item For $p \gg0$ or whenever the determining matrix $U=U(p, m, m-n)$
has all positive entries, we have $\operatorname{cx}_F(S_{m,n}) = \log_p (\lambda)$,
in which $\lambda$ is the Perron root for $U$.
\end{enumerate}
\end{Thm}
\begin{proof}
(1) Since $m-n \le m-2$, we see that $T(V_{m-n}(K[x_1,\dotsc,x_{m}]))$
is not finitely generated over $T_0(V_{m-n}(K[x_1,\dotsc,x_{m}]))$ by
Theorem~\ref{thm:mr}(2).
Thus $\mathscr{F}(E_{m,n})$) is not finitely generated over $\mathscr{F}_0(E_{m,n})$
by Theorem~\ref{thm:segre}(1).
(2) By Theorem~\ref{thm:segre}(3) and Theorem~\ref{thm:m=r+2},
\[
\operatorname{cx}_F(S_{m,2}) = \log_p\operatorname{cx}(T(V_{m-2}(K[x_1,\dotsc, x_{m}])))
= \log_p\binom{p+m-2}{m-1}.
\qedhere
\]
(3) This follows from Corollary~\ref{cor:limit}.
(4) This is a straightforward consequence of the discution in
Subsection~\ref{subsec:pf}.
\end{proof}
\begin{Rem} We like to point out the following:
\begin{enumerate}
\item Also note that, for every $m > 2$,
\[
\lim_{e \to \infty} c_e(\mathscr{F}(E_{m,2}))
= \lim_{e \to \infty} c_e(T(V_{m-2}(K[x_1, \dotsc,x_m]))) = \infty.
\]
\item Moreover, there exists an onto (hence nearly onto) graded ring
homomorphism from
$T(V_{m-n}(K[x_1, \dotsc,x_m]))$ to $T(V_{m-n}(K[x_1,\dotsc,x_{m-n+2}]))$.
Thus by Corollary~\ref{cor:nearly-onto},
\[
c_e(T(V_{m-n}(K[x_1, \dotsc,x_m])))
\ge c_e(T(V_{m-n}(K[x_1,\dotsc,x_{m-n+2}])))
\]
for all $ e \ge 0$.
Hence $c_e(\mathscr{F}(E_{m,n})) \ge c_e(\mathscr{F}(E_{m-n+2,2}))$ for all $ e \ge 0$
and consequently
\[
\lim_{e \to \infty} c_e(\mathscr{F}(E_{m,n})) = \infty
\]
for all $m > n \ge 2$.
\end{enumerate}
\end{Rem}
\subsection{Example}
We will illustrate our method with a concrete example. We are going to
use freely the notations established so far (especially the ones in
Section~\ref{sec:computing-ce}).
Let $r=2,\, m=5$ and $K$ be a field of characteristic $p=3$.
We are going to compute $c_e = c_e(T(V_2(K[x_1, \ldots, x_5])))$,
which in turn equals $c_e(\mathscr{F}(E_{5,3}))$ by Theorem~\ref{thm:segre}.
As in Discussion~\ref{dis:m,r,p}, we have
\[
X_e= U^e\cdot X_0 \quad \text{for all} \quad e \geq 0,
\]
in which
\begin{align*}
X_e &= \begin{bmatrix}
X_{e,1} \\
X_{e,2}
\end{bmatrix}, \\
X_0 &= \begin{bmatrix}
X_{0,1} \\
X_{0,2}
\end{bmatrix}
=\begin{bmatrix}
M_{3,5}(7) \\
M_{3,5}(10)
\end{bmatrix}
=\begin{bmatrix}
30 \\
1
\end{bmatrix}, \\
U &= \begin{bmatrix}
M_{3,5}(6) & M_{3,5}(5) \\
M_{3,5}(9) & M_{3,5}(8)
\end{bmatrix}
= \begin{bmatrix}
45 & 51 \\
5 & 15
\end{bmatrix}.
\end{align*}
Note that $U$ has all positive entries and the eigenvalues of $U$ are
$2(15+ 2\sqrt{30})$ and $2(15-2\sqrt{30})$.
At this point, we can apply the Theorem~\ref{thm:f-com-det}(4) above
directly and determine the Frobenius complexity of $S_{5,3}$ by
observing that the Perron root of $U$ is $2(15+2\sqrt{30})$.
However, for illustrative purposes let us compute $U^e$. This is
accomplished by diagonalizing $U$.
Skipping the details, we get
\[
U^e= \begin{bmatrix}
(15+4\sqrt{30})y_e
+(-15+4\sqrt{30})z_e
& 51(y_e-z_e)\\
5(y_e-z_e)
& (-15+4\sqrt{30})y_e +(15+4\sqrt{30})z_e
\end{bmatrix},
\]
in which
\[
y_e:=\frac{1}{\sqrt{15}} \cdot 2^{-\frac{7}{2}+e} \cdot (15+2\sqrt{30})^e
\quad \text{and} \quad
z_e:= \frac{1}{\sqrt{15}} \cdot 2^{-\frac{7}{2}+e}\cdot (15-2\sqrt{30})^e.
\]
Thus, for $e \geq 0$, we obtain
\begin{align*}
X_{e, 1} &= 30( (15+4\sqrt{30})y_{e}
+(-15+4\sqrt{30})z_{e})
+ 51(y_{e}+z_{e}),\\
X_{e,2} &= 150(y_{e}- z_{e}) +
(-15+4\sqrt{30})y_{e} +
(15+4\sqrt{30})z_{e}.
\end{align*}
Lastly, for $e \geq 2$, we have (cf.~Discussion~\ref{dis:m,r,p})
\begin{equation*}\label{eq:c}
c_e = c_e(T(V_2(K[x_1, \ldots, x_5])))
= \binom{7}{4}X_{e-2, 1} + \binom{6}{4}X_{e-2, 2},
\end{equation*}
which allows us to compute
$c_e(T(V_2(K[x_1, \ldots, x_5])))$ which equals $c_e(\mathscr{F}(E_{5,3}))$.
Therefore we are led to the following Proposition.
\begin{Prop}
When $p=3$, $\operatorname{cx}_F(S_{5,3}) = \log_3(2(15+2\sqrt{30}))$.
\end{Prop}
At conclusion of the paper, we would like to introduce the definition
of the Frobenius complexity for rings of characteristic zero, which is
motivated by Corollary~\ref{cor:limit} and Theorem~\ref{thm:f-com-det}(3).
As the definition involves rings that may not be local, we first
extend our Definition~\ref{def-FCX} by defining the Frobenius complexity
of a (not necessarily local) ring $R$ of prime characteristic $p$ as
$\operatorname{cx}_F(R) : = \log_p(\operatorname{cx}(\mathscr{C}(R)))$. (When $(R,\fm,k)} %{\text{$(R,\fm,k)$}$ is F-finite
complete local, $\mathscr{C}(R)$ and $\mathscr{F}(E(k))$ are opposite as graded rings;
so $\operatorname{cx}(\mathscr{C}(R)) = \operatorname{cx}(\mathscr{F}(E(k)))$ and we do have an extension of the definition.)
\begin{Def} \label{FC0}
Let $R$ be a ring (of characteristic zero) such that $R/pR \neq 0$ for
almost all prime number $p$.
When the limit $\lim_{p \to \infty} \operatorname{cx}_F(R/pR)$ exists, we call it
\emph{the Frobenius complexity} of $R$.
\end{Def}
It is natural to ask under what conditions, if any at all, the
Frobenius complexity exists. The case of
$R = \mathbb Z[X_1,\,\dotsc,\,X_n]/I$ and
$R = \mathbb Z[[X_1,\,\dotsc,\,X_n]]/I$ are particularly interesting.
If $R$ is a finitely generated algebra over a field $k$ of characteristic
zero, we could descend $R$ to a finitely generated $A$-algebra
$R_A$ (where $A$ is a subring of $k$ that is finitely generated over
$\mathbb Z$ containing the defining data of $R$) and study the the
Frobenius complexity of $R_A$.
|
\section{Introduction}
Random matrices were introduced to describe the spectrum of heavy nuclei \cite{wigner51}, but over the years it has been realized that they have the power to describe a large universality class of many-body strongly interacting systems, with applications ranging from nuclear physics to string theory, from 2-D quantum gravity to condensed matter physics, from statistical physics to econo-physics, from neuroscience to chaos theory, from number theory to integrable systems, and so \cite{mehtabook, natobook, akemannbook}.
Although in different applications the matrices can have different interpretations, we can use the partition function and the free energy as unifying concepts
\begin{equation}
{\cal Z} \equiv C_N \int {\cal D} {\bf M} {\rm e}^{-{1 \over g} W ( {\bf M})} \; , \qquad
{\cal F} \equiv \ln {\cal Z} \; ,
\label{partfunc}
\end{equation}
since, as usual, they can be used as generators for all physical observables ($C_N$ is an normalization constant introduced for later convenience). Here $g$ is a coupling constant, ${\bf M}$ is an $N \times N$ Hermitian matrix and $W ( {\bf M} )$ is the (matrix-valued) action. One can also consider different symmetry classes for the matrices, for instance orthogonal or symplectic, but we will concentrate just on the Hermitian case.
Typically, matrix models are considered in a basis invariant way, that is, the action in (\ref{partfunc}) is assumed invariant under a $U(N)$ rotation of the matrix: $ W( {\bf M}) = \mbox{tr}\, V ( {\bf M} )$. This choice is often motivated by physical reasons, but in any case invariant matrix models enjoy several powerful techniques for their study, which provides even non-perturbative results: from the use of orthogonal polynomials to the Riemann-Hilbert formulation, from the Coulomb gas approach to the mapping into Toda lattices, and so forth \cite{mehtabook, natobook, akemannbook}.
The versatility of random matrices is a consequence of a large universality, captured by a quadratic potential and valid for any polynomial $V(x)$. This universality is signaled by the correlated eigenvalue statistics known as Wigner-Dyson (WD) statistics. Its main feature is {\it level repulsion}: a characteristic suppression of the probability of finding two eigenvalues arbitrarily close. To capture this effect, one considers the level spacing distribution, that is, the distribution of the distance between eigenvalues. All standard, invariant matrix models present the same curve, hallmark of the universal WD level spacing, which is shared by a variety of strongly interacting systems.
The WD statistics for the eigenvalues has emerged as a defining signature of extended states for the eigenfunctions. Physically, one understands this connection because the eigenvalues interact only through their eigenvectors and thus, to repel one-another, their eigenstates need to have a finite overlap. Mathematically, this connection is well captured by the invariant matrix models, since the eigenstate and eigenvalue contributions decouple in the partition function:
\begin{equation}
{\cal Z} = C_N \int {\cal D} {\bf M} {\rm e}^{-{1 \over g} \mbox{tr}\, V ( {\bf M} )}
= C_N \int {\cal D} {\bf U} \int {\displaystyle\rm\mathstrut d} {\rm \Lambda} \Delta^2 ( {\bf \Lambda} )
{\rm e}^{- {1 \over g} \sum_j^N V (\lambda_j) } \; ,
\label{factZ}
\end{equation}
where the Hermitian matrix has been decomposed in its eigenvalues and eigenvectors as ${\bf M} = {\bf U}^\dagger {\bf \Lambda} {\bf U}$, with ${\bf \Lambda}$ a diagonal matrix of entries $\lambda_j$ and where $\Delta ( {\bf \Lambda} ) \equiv \prod_{j>l}^N (\lambda_j - \lambda_l)$ is the Van der Monde determinant which emerges as the Jacobian in changing integration variables from the matrix elements to the eigenvalue/eigenvector representation. We remark that this Jacobian introduces an effective interaction between the eigenvalues, which leads to a Coulomb Gas picture:
\begin{equation}
{\cal Z} = C_N \int {\cal D} {\bf U} \int {\displaystyle\rm\mathstrut d} {\rm \Lambda}
{\rm e}^{- {1 \over g} \sum_j^N V (\lambda_j)
+ 2 \sum_{j>l}^N \ln \left| \lambda_j - \lambda_l \right|} \; ,
\label{factZ1}
\end{equation}
in which the eigenvalues are interpreted as one-dimensional particles, confined by the potential $V(x)$ and interacting through a logarithmic repulsion. The balance of these two contrasting forces results in an equilibrium configuration for which the eigenvalues spread around the minima of $V(x)$.
The factorization in (\ref{factZ}) has lead to the conclusion that the eigenvector distribution is independent from $V({\bf M})$ and thus uniform over the $N$-dimensional sphere spanned by the Hilbert space. This behavior characterizes invariant matrix models as describing extended systems, i.e. systems where in any energy window at least one of the eigenvectors has all finite entries. This means that, when the matrix model is applied to mesoscopic systems, it only describes conducting states, i.e. situations where the disorder is not strong enough to localize the wavefunctions.
Recently, it was realized and shown that this expectation is violated when the eigenvalue distribution deviates from the WD \cite{franchini14}. In particular, if two eigenvalues are separated by a gap, their eigenvectors cannot mix. Without reference to a preferred basis, localization means that an eigenvector can move only within a portion of the whole Hilbert space and that an arbitrary, not-too-strong, perturbation cannot change this. If the eigenvalues of an invariant matrix model distribute over $n$ disconnected intervals, each of which containing $m_j$ ($j=1 \ldots n$) eigenvalues, each eigenvector is localized over the $m_j$-dimensional portion of the Hilbert space spanned by the eigenvectors of the other eigenvalues in that interval. This phenomenon corresponds to a spontaneous breaking of $U(N)$ rotational symmetry into $\prod_{j=1}^n U(m_j)$. As in any spontaneous symmetry breaking (SSB), the directions of these patches of Hilbert space are random and uniform (they can be induced by an external small symmetry breaking term), but once the system has chosen them, it is hard to change them and impossible in the thermodynamic limit.
In \cite{franchini14} the SSB of rotational symmetry was proven for matrix model with potentials able to separate the eigenvalues in disconnected patches, the so-called {\it multi-cut solutions}.
We are going to show that there exists another class of invariant matrix models, where the external potential is so weak that it combines with the interparticle interaction to generate an effective potential induced self-consistently by the eigenvalue distribution. The net effect of this process is that the partition function can be decomposed in the contributions of several saddle points, corresponding to different equilibrium configurations and thus of different effective potentials felt by the eigenvalues.
\section{The Weakly Confined Matrix Model}
We will consider a $U(N)$ invariant ensemble of random matrix characterized by a non-polynomial confining potential \cite{muttalib93,kravtsov95}, that is, with a potential of the form
\begin{equation}
V ( x ) \simeq {1 \over 2 \kappa} \ln^2 | x | \; , \qquad \qquad
| x | \to \infty \; ,
\label{logsquare}
\end{equation}
where $\kappa$ controls the strength of the interaction. This soft confinement justifies the name Weakly Confined Matrix Models (WCMM) for these systems.
The logarithmic behavior sets these potentials apart from the usual matrix models, where the potentials are polynomial in the matrix eigenvalues. Most importantly, this difference implies that the WCMM do not belong to the WD universality class. In fact, the level spacing is intermediate between the WD (approached for $\kappa \to 0$) and the Poissonian distribution (corresponding at the $\kappa \to \infty$ limit), typical of a completely localized system \cite{muttalib93}. Moreover, the density-density correlation function of the WCMM (in the unfolded coordinates) matches the expectations for the Anderson model at the metal/insulator transition \cite{kravtsov00}, at least in the perturbative regimes.
These observations have lead over the years to the conjecture that the WCMM can spontaneously break rotational invariance. In \cite{bogomolny97} it was shown that in the $\kappa \to \infty$ limit the eigenvalues crystallize on a regular lattice, so that they do not interact anymore. This observation was used in \cite{pato00} to argue that this eigenvalue behavior reflects the complete localization of eigenvectors, corresponding to a breaking of $U(N)$ symmetry into a $U(1)^N$, consistently with what expected from the level spacing behavior. Starting from these results, we will show what happens to the eigenvalue distribution at finite $\kappa$ and argue how this behavior reflects a critical breaking of $U(N)$, consistent with the conjectured relation between WCMM and the Anderson Metal/Instulator Transition (MIT).
The Riemann-Hilbert problem associated to potentials of the type (\ref{logsquare}) cannot be solved. Moreover, these potentials have an indeterminate moment problem, that is, the same set of orthogonal polynomials have the same set of moments with respect to several different measures.
Nonetheless, the WCMM can be exactly solved using the technique of orthogonal polynomials and in this way all the spectral characteristics have been calculated. The WCMM were actually introduced as the first example of a matrix model solvable using $q$-deformed (Hermite or Laguerre) polynomials, where the deformation parameter is determined by $\kappa$ \cite{muttalib93}. We stress once more that, because of the indetermination of the moment problem, different potentials sharing the same asymptotic behavior (\ref{logsquare}) are solved by the same set of polynomials.
For the moment, let us assume that the potential is {\it exactly} log-square and that the eigenvalues are confined to the positive semi-axis. We perform the change of variables
\begin{equation}
\lambda_j = {\rm e}^{\kappa x_j } \; ,
\label{expmap}
\end{equation}
in the partition function, getting
\begin{equation}
{\cal Z} = C_N {\rm Vol} \left[ U(N) \right] \kappa^N \int \prod_{j=1}^N {\displaystyle\rm\mathstrut d} x_j
\prod_{n<m} \left( {\rm e}^{\kappa x_n} - {\rm e}^{\kappa x_m} \right)^2 \:
{\rm e}^{-{\kappa \over 2} \sum_{l=1}^N \left[ x_l^2 - 2 x_l \right] } \; ,
\end{equation}
where ${\rm Vol} \left[ U(N) \right] = \prod_{n=1}^N { \left( 2 \pi \right)^{n} \over n!}$ is the volume of the unitary group.
Working in a simplex where the eigenvalues are sorted in increasing order ($x_j > x_l$ for $j >l$), we can write
\begin{equation}
{\cal Z} = C_N {\rm Vol} \left[ U(N) \right] \kappa^N N! \int \prod_{j=1}^N {\displaystyle\rm\mathstrut d} x_j
\prod_{n<m} \left[ 1 - {\rm e}^{\kappa (x_n - x_m)} \right]^2 \:
{\rm e}^{-{\kappa \over 2} \sum_{l=1}^N \left[ x_l^2 - 2 (2l - 1 )x_l \right] } \; .
\label{ZCN}
\end{equation}
In \cite{bogomolny97} it was noted that in the $\kappa \to \infty$ limit, all the exponentials in the Van der Monde determinant can be neglected and the eigenvalues crystallize:
\begin{eqnarray}
\lim_{\kappa \to \infty} {\cal Z} & = & C_N {\rm Vol} \left[ U(N) \right] \kappa^N N!
\exp \left[ {\kappa \over 2} \sum_{n=1}^N \left(2 n - 1 \right)^2 \right]
\int \prod_{j=1}^N {\displaystyle\rm\mathstrut d} x_j {\rm e}^{-{\kappa \over 2} \sum_{l=1}^N \left( x_l +1 -2 l \right)^2 }
\nonumber \\
& = & C_N {\rm Vol} \left[ U(N) \right] N!
{\rm e}^{ {\kappa \over 6} N (4 N^2 -1)} \left( 2 \pi \beta \right)^{N/2} \; .
\label{Z0}
\end{eqnarray}
In the large $\kappa$ limit, the eigenvalue distribution factorizes and the interaction between the eigenvalues can be exactly accounted by an new effective potential, which has the net effect of localizing each eigenvalue on a perfect lattice with a quadratic potential of width $1 \over \sqrt{\kappa}$. In \cite{bogomolny97} this phenomenology was discussed in details and it was shown that the eigenvalues become effectively uncorrelated in the $\kappa \to \infty$ limit.
In \cite{pato00} it was further shown that this crystallization of the eigenvalues tends to localize the eigenvectors as well. In fact the largest eigenvalue becomes exponentially larger than its closest neighbor and this means that rotating the diagonal matrix with a generic, uniformly distributed, unitary matrix produces a matrix with negligible off-diagonal elements. One can thus trace out the row and column corresponding to this eigenvalue and the same procedure can be applied again to the new highest eigenvalue. This structure was further confirmed by a numerical analysis in which a set of random matrices was generated with the crystallized eigenvalue distribution. These matrices were perturbed by multiplying the off-diagonal elements by a factor slightly less than unity. The resulting eigenvector distribution showed a significant deviation from the Porter-Thomas, indicating a localization of the eigenvectors. These arguments were the first direct proof that the crystallization of the eigenvalues affects the distribution of eigenvectors and tends to localize them. We can characterize this phenomenon as a breaking of the original $U(N)$ symmetry into a $U^N (1)$.
\section{Finite $\kappa$ analysis}
We saw that in the $\kappa \to \infty$ limit, the particles realize a sort of Mott-insulator that pins them to a lattice.
In (\ref{ZCN}) let us set
\begin{equation}
C_N \equiv {{\rm e}^{-{\kappa \over 6} N (4N^2-1)} \over \left( 2 \pi \beta \right)^{N/2} {\rm Vol} \left[ U(N) \right]}
\end{equation}
and define $x_n = 2 n -1 + t_n$:
\begin{equation}
{\cal Z} = \left( {\kappa \over 2 \pi} \right)^{N/2}
\int \prod_{j=1}^N {\displaystyle\rm\mathstrut d} t_j
\prod_{n<m} \left[ 1 - {\rm e}^{-2\kappa (m-n)} {\rm e}^{\kappa (t_n - t_m)} \right]^2 \:
{\rm e}^{- {\kappa \over 2} \sum_{l=1}^N t_l^2 } \; .
\label{Z}
\end{equation}
If the eigenvalues are ordered, for $\kappa \to \infty$, the term ${\rm e}^{-\kappa (m-n)}$ vanishes. It can be checked (and it will become more apparent as we proceed) that by removing the square from the Van der Monde determinant, we also remove the terms responsible for the other ordering of particles. Thus, in this way can work within one simplex without worrying about particle ordering anymore:
\begin{equation}
{\cal Z} = \left( {\kappa \over 2 \pi } \right)^{N/2} N!
\int \prod_{j=1}^N {\displaystyle\rm\mathstrut d} t_j
\prod_{m=2}^N \prod_{n=1}^{m-1} \left[ 1 - {\rm e}^{-2\kappa n } {\rm e}^{-\kappa (t_m - t_{m-n})} \right] \:
{\rm e}^{- {\kappa \over 2} \sum_{l=1}^N t_l^2 } \; .
\label{Zprime}
\end{equation}
For large $\kappa$ we can now use ${\rm e}^{-2 \kappa n}$ as an expansion parameter. In fact, we see that the effect of each term in the Van der Monde is to add a linear displacement to the quadratic potential, and thus it changes the equilibrium position of the particles. For $n=1$, a pair of nearest neighboring eigenvalues move from $2n -1$ and $2n +1$ to $2n$. For $n=2$ next-to-nearest neighbors move from $2n-1$ and $2n+3$ respectively to $2n$ and $2n +2$, and so on. After integration, each Gaussian integral contributes with a factor ${\rm e}^{-\kappa n}$. Thus, we introduce the expansion parameter $q \equiv {\rm e}^{-\kappa}$.
At zeroth order, from (\ref{Z0}), of course we have
\begin{equation}
{\cal Z}_0 = \left( {\kappa \over 2 \pi} \right)^{N/2}
\int \prod_{j=1}^N {\displaystyle\rm\mathstrut d} t_j {\rm e}^{- {\kappa \over 2} \sum_{l=1}^N t_l^2 }
= 1 \: .
\end{equation}
At first order in $q$ we have
\begin{eqnarray}
{\cal Z}_1 & = & {\cal Z}_0 - q^2 \left( {\kappa \over 2 \pi} \right)^{N/2} \sum_{m=2}^{N}
\int \prod_{j=1}^N {\displaystyle\rm\mathstrut d} t_j {\rm e}^{- {\kappa \over 2} \sum_{l=1}^N t_l^2 + \kappa ( t_{m-1} - t_{m} ) }
\nonumber \\
& = & {\cal Z}_0 - (N-1) q \; .
\end{eqnarray}
At the next order things start to complicate a bit, because we have three types of contributions: from two pairs of distinct nearest neighbors, from triplets of nearest neighbors and from pairs of next-to-nearest:
\begin{eqnarray}
{\cal Z}_2 & = & {\cal Z}_1
+ q^4 \left( {\kappa \over 2 \pi} \right)^{N/2} \int \prod_{j=1}^N {\displaystyle\rm\mathstrut d} t_j
{\rm e}^{- { \kappa \over 2} \sum_{l=1}^N t_l^2} \times
\nonumber \\
& & \times \left[
\sum_{m_1=4}^{N} {\rm e}^{\kappa ( t_{m_1 -1} - t_{m_1} )}
\sum_{m_2=2}^{m_1-2} {\rm e}^{\kappa ( t_{m_2 - 1} - t_{m_2} ) }
- (1-1) \sum_{m=3}^{N} {\rm e}^{\kappa ( t_{m-2} - t_{m} ) } \right]
\nonumber \\
& = & {\cal Z}_0 + {(N-2)(N-3) \over 2} q^2 .
\end{eqnarray}
We see that the last two contributions amount to the same process, but with equal and opposite weight.
Additional terms can be worked out similarly, but it is easy to see that the final result is
\begin{eqnarray}
{\cal Z} & = & \prod_{n=1}^{N-1} \left( 1 - q^n \right)^{N-n}
\label{Zqseries} \\
& = & 1 - (N-1) q + {(N-2)(N-3) \over 2} q^2 + {\rm O} (q^2) \; ,
\nonumber
\end{eqnarray}
which coincides with the exact result obtained in \cite{tierz04}, using the technique of orthogonal polynomials (in this case, the Stieltjes-Wigert polynomials), once the overall normalization $C_N$ is taken into account. We see that the first line of (\ref{Zqseries}) derives directly from the Van der Monde in (\ref{Zprime}) and allows for a straightforward interpretation of the partition function.
\section{The Energy Landscape}
Starting from the $\kappa \to \infty$ configuration, in which each eigenvalue $x_j$ has equilibrium position at $2j-1$ ($j=1 \ldots N$), each term in the Van der Monde ads a linear contribution to the quadratic potential and thus shifts the positions pairwise closer by one unity. All the configurations reached this way are stable by construction (small fluctuations are confined by the potential) and thus constitutes saddle points of the partition function. The terms of the Van der Monde connecting the different configurations have a natural interpretation as an instanton.
Every equilibrium configuration with center of mass at $m$ is realized by the action of the instantons: all these configurations have the same leading weight (given by $C_N$ and growing like ${\rm e}^{ {2 \over 3} \kappa N^3}$) and differ only by subleading contributions, with a power of $q$ equal to the number of steps each eigenvalue traveled to reach this configuration, starting from the $\kappa \to \infty$ one. In some of these configurations, the equilibrium point for more than one eigenvalue can coincide (in fact, the action of every allowed instanton brings all eigenvalues to oscillate around the same point $m$). Thus, there is an exponential number (in $N$) of possible equilibrium configurations. In each of them, eigenvalues are only subjected to the effective potential and do not interact. Hence, the eigenvalue correlation is a purely instantonic effect. In fact, these instantons have a fermionic nature, in that each of them contributes with a minus sign and only one instanton connecting two given eigenvalues is allowed.
The partition function (\ref{Zqseries}) has a natural interpretation in terms of the instantons, since only the largest eigenvalue can be connected to the smallest by a instanton with weight $-q^{N-1}$, while there are two ways to connect a pair of eigenvalues $N-2$ apart and so on, concluding with the $N-1$ instantons connecting nearest neighboring eigenvalues.
This newly discovered energy landscape is characteristic of a complex system and calls for the usual tools employed in the analysis of these systems. Intriguingly, the mathematical framework we just outlined shares many similarities with a recently discussed glassy system, of hard spheres in infinite dimensionality \cite{charbonneau14}.
In \cite{franchini14}, it was argued that instantons like the ones we discussed are responsible for restoring the broken symmetries. Since the $\kappa \to \infty$ configurations correspond to a breaking of $U(N)$ symmetry into $U(1)^N$, each instanton can be interpreted as restoring the $SU(2)$ component connecting the two eigenvalues.
\section{Conclusions}
We showed that the partition function of the weakly confined matrix models can be decomposed into a exponential number of independent equilibrium configurations, corresponding to different equilibrium points for each eigenvalue of the matrix. The different configurations are connected by instantons, which can be interpreted as generators of the the symmetries in which the $U(N)$ invariance is broken, according to \cite{franchini14}.
The WCMM is conjecture to represent a toy model for the Anderson Metal Insulator Transition. The structure unveiled supports this conjecture and draws a novel connection between localization problems and systems with complex energy landscapes, such as the Sherrington-Kirkpatrick model \cite{replicabook}.
Moreover, the analytically continued WCMM is connected to Chern-Simons theories \cite{tierz04} and SUSY string theories, such as ABJM \cite{marinobook}. In particular, the ABJM partition function in the localization limit corresponding to the matrix models can be interpreted in terms of a free Fermi gas at finite temperatures \cite{marino12}: it seems that this is a gas of instantons and thus the rich energy landscape should be analyses in this field theory formulation as well.
We considered models with an exact log-normal potentials: more general potentials can be considered, but we believe that the qualitative structure of the energy landscape will not change. Such generalizations include potentials which can be written as polynomials in logarithm of the matrix. A more interesting extension will be the treatment to the case of eigenvalues over the whole real axis, where non-local correlations between eigenvalues at opposite sides of the origin have been observed \cite{kravtsov95}, and interpreted in terms of a Luttinger liquid in a Hawking-Hartle bath \cite{franchini09}.
|
\section{Introduction}
One of the famous, yet intriguing feature of quantum mechanics is the wave-particle duality. This is often captioned by Bohr's complementarity principle. It states that the wave aspect and the particle aspect are
complementary in nature, in the sense that if an experiment clearly reveals the wave
nature, it will completely hide the particle aspect and vice-versa \cite{bohr}.
The complementarity principle has been a subject of debate since the time
of its inception when Einstein proposed his famous recoiling slit experiment
(see e.g. \cite{tqeinstein}).
Since then, attempts have been made to give a quantitative meaning to the
complementarity principle, in the context of interference experiments
\cite{wootters, greenberger, englert}. The idea is to investigate how much
of each aspect, wave or particle, can be seen at the same time.
In two-path interference experiment,
either in a two-slit experiment or in a two-path Mach-Zehnder interferometer,
the principle of complementarity is quantitatively represented by the
Englert-Greenberger-Yasin (EGY) relation
\begin{equation}
{\mathcal V}^2 + {\mathcal D}^2 \le 1,
\label{egy}
\end{equation}
where ${\mathcal V}$ is the visibility of the interference pattern and ${\mathcal D}$ is a measure of path distinguishability or which-path information.
For two-path interference, the quantum system (quanton) may arrive at the detector along two different paths. If the experimenter determines which path the quanton has traveled through without ambiguity (i.e., $\mathcal{D}=1$), then no interference fringes will be seen (i.e., $\mathcal{V}=0$). On the other hand, a non-zero ambiguity in the which-path information (i.e., $\mathcal{D}\neq 1$), of the experimenter, may retain a non-vanishing fringe visibility (i.e., $\mathcal{V} \neq 0$).
Thus, the knowledge of which-path information or the path distinguishability limits the interference visibility ${\mathcal V}$ in an interference experiment, according to the above complementarity relation. This relation has been demonstrated experimentally with atoms \cite{Durr98}, nuclear magnetic resonance \cite{Peng03, Peng05}, faint lasers \cite{Schwindt99}, and also with single photons \cite{Jacques08}. Further, the complementarity relation has been extended to the more general case of an asymmetric interferometer where only a single output port is considered and this duality holds \cite{Li12}. Recently, the duality relation has also been investigated in the presence of non-locality \cite{Peruzzo12} and quantum entanglement \cite{Jakob03}.
Intuitively, the complementarity relation, between the wave and the particle nature of the quantons, is expected to hold in multi-path or multi-slit experiments too. Several attempts have been made to quantitatively formulate
the complementarity principle in multi-path experiments \cite{jaeger, durr, bimonte, englertmb}. However, a derivation of a loophole free generalized complementarity relation, for multi-slit quantum interference experiment, is still demanding.
The underlying problem, for that, is the absence of
exact analytical forms of interference fringe visibility and path distinguishablity, which are strictly complementary to each other, for an $n$-path quantum interferometer. One may seek to resolve this difficulty by relaxing the conventional signatures of particle and wave natures of the quantons, such as which-path information and fringe visibility.
While the fringe visibility being the most used signature of wave nature of the quantum particles, it is certainly not the only one to capture the essence. For example, recently, the wave-particle duality in two-slit experiments has been shown to be equivalent to the entropic uncertainty relation \cite{coles}, and the wave and particle natures are represented with entropic quantifiers.
In this paper we derive a generalized complementarity relation from such an alternative perspective. In our approach, we quantify the wave nature in terms quantum coherence, which has been proposed recently in the context of quantum information theory \cite{coherence}. On the other hand, the particle nature connected to the which-path information or the path distinguishability, is quantified by the upper bound of the success probability in the unambiguous quantum state discrimination (UQSD) \cite{Helstrom76, Bae15, Pati05}. Remarkably, the quantum coherence and the path distinguishability are truly complementary in nature. That means, an increase in quantum coherence is always associated with a decrease in path distinguishability and vice versa.
With the proposed quantifiers, we derive a generalized complementarity relation for arbitrary $n$-slit scenario for both pure and mixed quanton states. We show that, the sum of (normalized) quantum coherence and path distinguishability in the complementarity relation, exactly, equals to one for every pure states and upper bounded by the same in the case of mixed states. Our duality relation then gives a
justification to the measure of quantum coherence as it truly brings out the wave nature of the quanton at its heart.
The paper is organized as follows. In section \ref{sec:coh}, we introduce the quantifiers of quantum coherence and unambiguous quantum state discrimination which quantitatively capture the wave nature and particle nature of the quantum system, respectively. The new duality relation, for pure quanton and detector state is derived in \ref{sec:pureCR}. Then we generalize the duality relation for mixed states of arbitrary dimension.
Finally, we conclude in section \ref{sec:concl}.
\section{Quantum Coherence and UQSD \label{sec:coh}}
\subsection{Quantum coherence}
Coherence is a fundamental feature of quantum physics which signifies the possible superposition between the orthogonal quantum states. Again, it is largely believed that, the quantum superposition is the manifestation of wave nature of quantum particles. Thus, the quantum coherence has a strong correspondence with the wave nature of a quantum particle. Though a rigorous study of coherence has been carried out in quantum optics in terms of quasi-probabilities, a generalized quantification of quantum coherence was absent until recently. In Ref. \cite{coherence}, Baumgratz {\em et al.}, proposed a reliable quantifier of quantum coherence from quantum information theoretic approach. The framework, to quantify coherence, is based on the characterization of the set of incoherent quantum states ($\mathcal{I}$).
For a given incoherent basis $\{ |i\rangle \}$, the incoherent states are defined as $\sigma^{\mathcal{I}}=\sum_i p_i |i\rangle \langle i| \in \mathcal{I}$, where $p_i$s are non-negative probabilities with $\sum_i p_i=1$. The
incoherent operations are the completely positive trace preserving (CPTP) maps $\Lambda^{\mathcal{I}}$, those transform $\sigma^{\mathcal{I}}\rightarrow \Lambda^{\mathcal{I}}(\sigma^{\mathcal{I}}) \in \mathcal{I}$ to an incoherent state which is, again, diagonal in the incoherent basis. The maximally coherent state, of dimension $n$, is defined as $|\Psi\rangle = {1\over\sqrt{n}}\sum_{i=1}^n |i\rangle$ and the coherence of such state is used as reference to compare the coherence in the other states. For a given incoherent basis, the reliable quantifier of quantum coherence $\mathcal{C}(\rho)$, is a function of the state $\rho$, and should satisfy \cite{coherence}: (1) $\mathcal{C}(\rho)=0$ if and only if $\rho \in \mathcal{I}$, (2) $\mathcal{C}(\rho)$ is non-increasing under incoherent operations, i.e., $\mathcal{C}(\rho) \geqslant \mathcal{C}(\Lambda^{\mathcal{I}} \rho)$, (3) $\mathcal{C}(\rho)$ is non-increasing on an average under selective incoherent measurement, i.e., $\mathcal{C}(\rho)\geqslant \sum_m q_m \mathcal{C}(\rho_m)$,
where $\rho_m=\frac{1}{q_m}K_m \rho K_m^\dag$, \ $q_m=\mbox{Tr}(K_m \rho K_m^\dag)$, $K_m$'s are the Kraus operators and $K_m \mathcal{I} K_m^\dag \subseteq \mathcal{I}$, (4) $\mathcal{C}(\rho)$ is non-increasing under convex mixing of density matrices, i.e., $\mathcal{C}(\sum_k q_k \rho_k)\leqslant \sum_k q_k \mathcal{C}(\rho_k)$. The functions of density matrix those satisfy these properties are relative entropy of coherence, $l_1$-norm of coherence \cite{coherence} and skew information of coherence \cite{Girolami}. In this work, we shall use $l_1$-norm of coherence to quantify the wave nature of the quantons, defined as
\begin{equation}
\mathcal{C}(\rho)=\mbox{min}_{\sigma^\mathcal{I} \in \mathcal{I}}\parallel \rho-\sigma^\mathcal{I} \parallel_{l_1},
\end{equation}
where $\parallel A \parallel_{l_1}=\sum_{ij}|A_{ij}|$ and the minimization is carried out over the set of all incoherent states $\sigma^\mathcal{I}$. It can easily be seen that the minimization is achieved for the $\sigma^\mathcal{I}=\sum_i \langle i | \rho | i \rangle |i \rangle \langle i|$ and then the coherence quantitatively becomes the sum of the absolute values of the off-diagonal elements of the
density matrix of a system, i.e.,
\begin{equation}
\mathcal{C}(\rho)=\sum_{i\neq j} |\rho_{ij}|
\end{equation}
with $\rho_{ij}=\langle i | \rho | j \rangle$.
Using this measure, we define normalized quantum coherence
as
\begin{equation}
{\mathcal C}(\rho) = {1\over n-1}\sum_{i\neq j} |\rho_{ij}| ,
\label{coherence}
\end{equation}
where $n$ is the dimensionality of the Hilbert space. Hereafter we call normalized coherence as ``coherence'', for simplicity. In what follows, we shall show that the coherence captures the wave nature of a quanton in a multi-path quantum interference scenario.
\subsection{Unambiguous Quantum State Discrimination \label{sec:uqsd}}
Quantum state discrimination has various important applications in quantum information theory
\cite{Helstrom76, Bae15, Pati05}. In the quantum state discrimination, the task is to find out which state an experimenter has in her possession to the best of her ability \cite{Helstrom76}. In quantum mechanics the existence of non-orthogonal states adds further difficulties to the problem, in addition to the statistical mixing of the quantum states. The commonly used strategies to discriminate non-orthogonal states can be divided in two broader classes, those are the ambiguous \cite{Helstrom76} and the unambiguous \cite{uqsd} quantum state discrimination. In ambiguous state discrimination, one always has an answer but with a probability of being wrong and the task is to minimize this probability. On the other hand, in unambiguous state discrimination, one guarantees to never be wrong but sometime may have nonanswer, that is to say one does not know. In this case the task is to minimize the probability of a nonanswer. The unambiguous state discrimination is particularly interesting for the cases where the states analyzed are mixed \cite{Rudolph03}. However, in what follows, we shall stick to UQSD for pure states as it suffices for our analysis.
UQSD was first formulated for unambiguously discriminating between two
non-orthogonal states \cite{uqsd}. Consider a state which could be either
$|d_1\rangle$ or $|d_2\rangle$, with equal probability. The probability with
which one can tell {\em for sure} which of the two states is the given one,
is bounded by \cite{uqsd}
\begin{equation}
P_2 \le 1 - |\langle d_1|d_2\rangle|.
\label{p2}
\end{equation}
One can also precisely specify the condition in which the success probability is
\begin{equation}
P_2 = 1 - |\langle d_1|d_2\rangle|,
\label{idp}
\end{equation}
which is the so-called IDP (Ivanovic-Dieks-Peres) limit. The two states,
$|d_1\rangle$ and $|d_2\rangle$, cannot be unambiguously discriminated
with a probability larger than the IDP limit, even in principle.
This shows that the IDP limit sets a fundamental limit to distinguish
two non-orthogonal states $|d_1\rangle$ and $|d_2\rangle$.
UQSD was later generalized to $n$ non-orthogonal states \cite{zhang}.
Consider a quantum state prepared in one of the $n$ states $|d_1 \rangle, ..., |d_n \rangle$ in an $n$-dimensional Hilbert space with corresponding probabilities $p_1,...,p_n$. The states are in general non-orthogonal. To find out which of the $n$ states, the given state is, one needs to perform one or more quantum measurements. The upper bound for the success probability, of unambiguous discrimination among the $n$ quantum states, is given by \cite{zhang}
\begin{equation}
P_n \le 1 - {1\over n-1}\sum_{i\neq j} \sqrt{p_ip_j} |\langle d_i|d_j\rangle|.
\label{pn}
\end{equation}
Clearly, for orthogonal quantum states, there exists a quantum measurement strategy for which the success probability of UQSD reaches maximum to 1.
For non-orthogonal states, there may exist a measurement strategy for which the
success probability of UQSD is
\begin{equation}
P_n = 1 - {1\over n-1}\sum_{i\neq j} \sqrt{p_ip_j} |\langle d_i|d_j\rangle|.
\label{pnmax}
\end{equation}
However, since the above does not represent the optimal success probability,
this upper bound may, in general, not be achievable.
In the next section, we shall use this upper bound of the success probability for UQSD as a measure of which-path information and that in turn can witness the particle nature of the quanton.
\section{Complementarity of Coherence and Path Distinguishability \label{sec:pureCR}}
\subsection{Duality relation for pure quanton and detector states}
Let us now consider the case of a $n$-slit quantum interference with pure
quantons. In $n$-slit interference, if $|\psi_i\rangle$ is the possible
state of the quanton if it takes the $i$'th slit or $i$'th path, then state of
the quanton, after crossing the slit, can be described solely in terms
of $|\psi_i\rangle$, and it may be treated as a basis state. Thus the
state of the quanton can be represented in terms of $n$ basis states
$\{|\psi_1\rangle,|\psi_2\rangle,\dots,|\psi_n\rangle \}$, where each of
the states represents each slit or path.
\begin{equation}
|\Psi \rangle = c_1|\psi_1\rangle + c_2|\psi_2\rangle +
\dots + c_n|\psi_n\rangle ,
\label{ent}
\end{equation}
and $c_i$ is the amplitude of taking the i'th path.
In an interference experiment,
if one wants to know which of the $n$ slits the quanton passes through,
or which of the $n$ paths does the quanton take, one needs to perform
a quantum measurement.
In quantum measurements, according to von Neumann, the first process is
to let a detector interacts with a quanton and gets entangled with it
\cite{neumann}. Then, the quantum measurements may be performed on the
detector state to infer about the properties of the quanton.
In general, the controlled unitary operations are used to correlate the
quanton and detector in an interference experiment. For a quanton
state $| \Psi\rangle$ and controlled unitary
interaction $U(|\psi_i\rangle |0_d\rangle) \mapsto |\psi_i\rangle |d_i\rangle$,
where $|0_d\rangle$ is the initial detector state, the combined
quanton-detector state becomes
\begin{equation}
|\Psi\rangle = c_1|\psi_1\rangle \otimes |d_1\rangle + c_2|\psi_2\rangle \otimes |d_2\rangle +
\dots , c_n|\psi_n\rangle \otimes |d_n\rangle,
\label{ent}
\end{equation}
where $|d_i\rangle$ is the state of the which-path detector if the quanton
went through the $i$'th path, and $\sum_{i=1}^n|c_i|^2=1$.
For simplicity, we consider the detector states
$\{|d_i\rangle\}$ to be normalized, but not necessarily orthogonal.
Now, if one tries to acquire knowledge about which path the quanton took, it shall reduce the coherence of the quanton.
The left out coherence present in the quanton will correspond to the coherence of the reduced density matrix of the quanton, where the latter is given by
\begin{equation}
\rho_s = \sum_{i=1}^n\sum_{j=1}^nc_ic_j^* \ \langle d_j|d_i\rangle \ |\psi_i\rangle\langle\psi_j|,
\end{equation}
after tracing over the detector states.
For a given interferometric set-up the set $\{ |\psi_i\rangle\}$ forms the incoherent bases. The coherence can now be calculated for the particle using the reduced
density matrix given above, as
\begin{eqnarray}
{\mathcal C} &=& {1\over n-1}\sum_{i\neq j}|\langle\psi_i|\rho_s|\psi_j\rangle|
\nonumber\\
&=& {1\over n-1}\sum_{i\neq j} |c_i||c_j||\langle d_j|d_i\rangle|.
\label{Cn}
\end{eqnarray}
It is interesting to note that if the detector states $\{|d_i\rangle \}$ form a mutually orthogonal basis, the reduced system states becomes diagonal in the incoherent basis and hence, has vanishing coherence. This implies that, in this situation, the
wave nature of the quanton cannot be seen. However, the situation will be different if the detector states are not mutually orthogonal to each other and in the reduced density matrix, the off-diagonal elements not necessarily be vanishing. Thus, the wave aspect of the quanton will acquire non-zero value as the quantum coherence is non-vanishing. But, the coherence of the quanton will be certainly reduced than that of the original state before the measurement interaction was turned on.
Now, let us focus on the problem of path distinguishability or which-path information which is attributed to the particle nature of the quantons.
Since, through the controlled unitary interaction, each of the path is marked with a detector state $| d_i \rangle$, the path distinguishability is equivalent to discriminating the detector states. In other words, if the quanton passes through the $i$'th path, the resulting the detector states becomes $|d_i\rangle$ with the probability $|c_i|^2$. Now distinguishing all these detector states $\{|d_i\rangle \}$ with the corresponding probabilities $\{|c_i|^2 \}$, is equivalent to distinguishing the paths the quanton chooses in the interferometric set-up. If the detector states $\{|d_i\rangle \}$ are mutually orthogonal, then the states can be distinguished with unit probability. In this case we will know which path the system has taken with certainty. However, the interesting case appears when $\{|d_i\rangle \}$ are not mutually orthogonal and in that case we have partial knowledge about which path information.
In general, the best strategy to distinguish between non-orthogonal states
is unambiguous quantum state discrimination (UQSD)
\cite{uqsd, jaeger2, bergou}.
In UQSD, the success probability with which non-orthogonal pure states can
be {\em unambiguously} distinguished depends on the measurement strategies
employed. One would like to know which strategy yields the maximum
success probability. The optimal success probability of unambigiously
distinguishing between $n$ non-orthogonal states is not known. However,
the success probability in the UQSD between the detector states $\{ | d_i \rangle \}$ with corresponding probabilities $\{|c_i|^2 \}$, is {\em bounded by} \cite{zhang}
\begin{equation}
P_n \le 1 - {1\over n-1}\sum_{i\neq j} |c_i||c_j| |\langle d_i|d_j\rangle|.
\label{pn}
\end{equation}
Note that the probabilities $|c_i|^2$ are decided by the initial superposition in the quanton state.
The upper bound
given by (\ref{pn}) may not be achievable in practice in many situations.
However, it can still serve as a measure of the degree to which the $n$ states
are distinguishable, namely the $n$ states cannot be distinguished with a
probability larger than that given by the upper bound of (\ref{pn}).
Here we {\em define} the path-distinguishability ${\mathcal D}_Q$ as the upper bound of the success
probability with which the $n$ paths of the particle can be distinguished
without any error. The subscript $Q$ is added to differentiate it from the
${\mathcal D}$ used in \cite{englert} and elsewhere in the literature.
This is just the upper bound of the success probability with which the states $\{|d_i\rangle\}$
can be unambiguously discriminated, which is the saturation limit of
(\ref{pn}). Thus, the path-distinguishability, for $n$-path interference, can
be defined as
\begin{equation}
{\mathcal D}_Q := 1 - {1\over n-1}\sum_{i\neq j} |c_i||c_j| |\langle d_i|d_j\rangle|.
\label{D}
\end{equation}
The path-distinguishability can take values between 0 and 1. For all mutually orthogonal detector states $\{|d_i\rangle\}$, one has ${\mathcal D}_Q=1$.
Now with the quantum coherence left in the reduced quanton state, in Eq. (\ref{Cn}) and the path-distinguishability, defined in Eq. (\ref{D}),
we get a general $n$-slit duality relation, for arbitrary pure quantons, as
\begin{equation}
{\mathcal C} + {\mathcal D}_Q = 1 .
\label{duality}
\end{equation}
Using the ${\mathcal C}$ and ${\mathcal D}_Q$ as the quantifiers of the wave and the particle nature of the quanton, respectively, the
Eqn. (\ref{duality}) puts a bound on how much of wave nature and particle nature
a system can display at the same time. This can be treated as a
quantitative statement of Bohr's principle, for $n$-path interference, using the measures of quantum coherence and path distinguishability. Note, the ${\mathcal C}$ and ${\mathcal D}_Q$ are truly complementary in nature where an increase in one results in a decrease in the other.
\subsubsection{Two-slit interference}
In a two-path interference with equally probable paths we have
$|c_1|=|c_2|=\frac{1}{\sqrt{2}}$, and hence the path-distinguishability becomes
${\mathcal D}_Q=1-|\langle d_1|d_2\rangle|$ which, interestingly, is just the
IDP limit. The coherence in (\ref{Cn}) reduces to
\begin{equation}
{\mathcal C} = |\langle d_1|d_2\rangle|.
\label{C2}
\end{equation}
However, for a double-slit experiment, the interference visibility,
defined as ${\mathcal V} \equiv {I_{max} - I_{min} \over I_{max} + I_{min} }$,
where $I_{max}$ and $I_{min}$ represent the maximum and minimum intensity
in neighboring fringes, respectively, is just
${\mathcal V} = |\langle d_1|d_2\rangle|$ \cite{tqeinstein}. Therefore, the fringe
visibility is just equal to the coherence ${\mathcal C}$. The duality
relation (\ref{duality}) now becomes
\begin{equation}
{\mathcal V} + {\mathcal D}_Q = 1 .
\end{equation}
In practical scenario there may be other factors which reduce the visibility of
fringes in interference experiments. So in the general case, the above relation will be an
{\em inequality}, saturating to equality.
This inequality has been derived before \cite{3slit}, and is completely
equivalent to the EGY duality relation (\ref{egy}) with the
recognition that ${\mathcal D_Q} = 1 - \sqrt{ 1 - {\mathcal D}^2}$
\cite{3slit}.
Therefore, for the two-slit case, the new duality relation (\ref{duality})
reduces to the EGY relation (\ref{egy}).
\subsubsection{Three-slit interference}
For the three-slit interference, the path-distinguishability becomes
${\mathcal D}_Q = 1 - (|c_1||c_2| |\langle d_1|d_2\rangle | +
|c_2||c_3| |\langle d_2|d_3\rangle | + |c_1||c_3| |\langle d_1|d_3\rangle |)$,
and the coherence reduces to
\begin{equation}
{\cal C} = |c_1||c_2||\langle d_1|d_2\rangle|
+|c_2||c_3| |\langle d_2|d_3\rangle |+|c_1||c_3| |\langle d_1|d_3\rangle|.
\label{C3}
\end{equation}
This coherence can be shown to be related to the ideal interference
visibility by the relation \cite{3slit}
\begin{equation}
{\mathcal C} = {2{\mathcal V}\over 3 - {\mathcal V}}.
\end{equation}
The duality relation reduces to
\begin{equation}
{\mathcal D}_Q + {2{\mathcal V}\over 3 - {\mathcal V}} = 1,
\end{equation}
which is exactly the same as the duality relation derived for the
three-slit interference \cite{3slit}. One should of course realize that
in a non-ideal situation, the fringe visibility will be reduced, and the the above relation will be an {\em inequality}.
Thus, for the 3-slit case, the new duality relation reduces to an earlier
relation derived independently.
\subsection{Generalization of duality relation for mixed states \label{sec:mixCR}}
We now extend the preceding analysis to the situations where a certain amount
of mixedness is introduced in density matrix of the quanton state, say $\rho=\sum_{ij} \rho_{ij}|\psi_i \rangle \langle \psi_j|$.
This may happen if the quantum system is exposed to environment. Interference experiments have been
carried out with large molecules, where the interaction with the environment,
although minimized, is not fully avoided \cite{c60, c70}. In a scenario where the initial detector state is chosen to be pure, the combined density matrix of the quanton and the path-detector after the controlled unitary (the measurement interaction), may be written as
\begin{equation}
\rho_{sd} = \sum_{i=1}^n\sum_{j=1}^n \rho_{ij}|\psi_i\rangle\langle\psi_j| \otimes |d_i\rangle\langle d_j|,
\label{rhom}
\end{equation}
where the quanton state becomes entangled with the detector. Note that the combined state is also mixed since the initial quanton state is mixed. As in the pure state scenario, the wave nature in the post-interaction quanton state can be quantified with the quantum coherence present in the reduced quanton state.
By tracing out over the path-detector states in (\ref{rhom}), one gets
the reduced density matrix for the quanton
\begin{equation}
\rho_{s} = \sum_{i=1}^n\sum_{j=1}^n \rho_{ij}|\psi_i\rangle\langle\psi_j|
\langle d_j|d_i\rangle .
\label{rhomr}
\end{equation}
The coherence can now be calculated as
\begin{eqnarray}
{\mathcal C} &=& {1\over n-1}\sum_{i\neq j}|\langle\psi_i|\rho_{s}|\psi_j\rangle|
\nonumber\\
&=& {1\over n-1}\sum_{i\neq j} |\rho_{ij}||\langle d_j|d_i\rangle|.
\label{Cm}
\end{eqnarray}
Note that, before the interaction between quanton and detector, the coherence is ${\mathcal C}={1\over n-1}\sum_{i\neq j} |\rho_{ij}|$, which is reduced after the interaction due to the factors $|\langle d_j|d_i\rangle| \leqslant 1$.
Again the particle nature of the quantons can be expressed, quantitatively, by the path-distinguishability or which-path information. That is nothing but how well an experimenter can distinguish between the detector states $\{|d_i\rangle \}$ with the corresponding probabilities $\{ \rho_{ii} \}$, where the probabilities are determined by the initial quanton state.
One may carry out UQSD on the state $|d_i\rangle$ as before. Since a state
$|d_i\rangle$ appears with a probability $\rho_{ii}$, the path distinguishability for
$n$-path interference which is the upper bound of success probability in UQSD, in the mixed case, can be written as
\begin{eqnarray}
{\mathcal D}_Q &=& 1 - {1\over n-1}\sum_{i\neq j} \sqrt{\rho_{ii}\rho_{jj}} |\langle d_i|d_j\rangle|.
\label{Dm}
\end{eqnarray}
For a given quanton state, the path distinguishability and the quantum coherence of the quanton depend on the choice of initial state of the which-path detector and the measurement interaction. Now from (\ref{Cm}) and (\ref{Dm}), we get
\begin{equation}
{\mathcal C} + {\mathcal D}_Q + {1\over n-1}\sum_{i\neq j}
(\sqrt{\rho_{ii}\rho_{jj}} - |\rho_{ij}|)|\langle d_j|d_i\rangle| = 1 .
\label{c+ds}
\end{equation}
Since every principal 2x2 sub matrix of (\ref{rhom}) is positive
semi-definite \cite{horn}, we have
\begin{equation}
\sqrt{\rho_{ii}\rho_{jj}} - |\rho_{ij}| \ge 0,
\label{eq:se}
\end{equation}
for arbitrary $i$ and $j$. This in turn implies that
${1\over n-1}\sum_{i\neq j}
(\sqrt{\rho_{ii}\rho_{jj}} - |\rho_{ij}|)|\langle d_j|d_i\rangle| \ge 0$. The equality holds only for the initial pure quantum systems.
Thus, (\ref{c+ds}) leads to the following duality relation between the
quantum coherence and the path distinguishability which are the quantifiers of wave nature and particle nature of a quanton respectively, given by
\begin{equation}
{\mathcal C} + {\mathcal D}_Q \le 1.
\label{c+d}
\end{equation}
The above is the generalized version of (\ref{duality}) and applicable for any mixed quanton state.
One can easily see that, in an experiment where the mutual overlap between the detector states are simultaneously increased or decreased, these two quantities become truly complementary in nature. There, an increase in the path distinguishability
inevitably reduces the quantum coherence in the quanton state, and vice versa.
Our analysis can further be extended in the case where one has the initial detector in a mixed state, as well. For the controlled-unitary operation as the measurement interaction $U=\sum_{i}|\psi_i\rangle \langle \psi_i| \otimes U_i$ and initial detector state $\rho_d$, the combined quanton-detector state after the interaction becomes
\begin{equation}
\rho_{sd}=\sum_{ij}\rho_{ij}|\psi_i\rangle \langle \psi_j|\otimes U_i \rho_d U_i^\dag.
\end{equation}
If the measurement interaction leads to mutually orthogonal detector states, then $\mbox{Tr}\left( U_i\rho_d U_i^\dag U_j\rho_d U_j^\dag\right)=0$ for $\forall \ i\neq j$. In such situation, one may easily see that $\mbox{Tr}\left( U_i \rho_d U_j^\dag \right)=0$, for $\forall \ i\neq j$, holds. The reduced quanton state after the interaction can be written as
\begin{equation}
\rho_s^\prime = \sum_{i,j}^n \rho_{ij} \ \mbox{Tr}\left( U_i \rho_d U_j^\dag \right) \ |\psi_i\rangle \langle \psi_j|.
\end{equation}
Clearly, for a ``good'' measurement interaction for which $\mbox{Tr}\left( U_i \rho_d U_j^\dag \right)=0$ for $\forall \ i\neq j$, the quanton state reduces to an incoherent state and thus the wave nature becomes absent.
On the other hand if the $\mbox{Tr}\left( U_i \rho_d U_j^\dag \right) \neq0$, there remains non-vanishing quantum coherence in the reduced quanton state, as
\begin{equation}
\mathcal{C}^\prime =\frac{1}{n-1} \sum_{i\neq j} \left| \ \rho_{ij} \ \mbox{Tr}\left( U_i \rho_d U_j^\dag \right) \right|.
\end{equation}
For an initial detector state, let $\rho_d=\sum_k r_k |d_k \rangle \langle d_k|$ be the spectral decomposition. Then, we have $\left| \ \rho_{ij} \ \mbox{Tr}\left( U_i \rho_d U_j^\dag \right)\right|=\left|\rho_{ij} \ \sum_k r_k \langle d_{ki} | d_{kj} \rangle \right| \leqslant \sum_k r_k |\langle d_{ki} | d_{kj} \rangle | |\rho_{ij} |$, where we denote $U_i |d_k \rangle=|d_{ki} \rangle$. Thus we are lead to
\begin{equation}
\mathcal{C}^\prime \leqslant \frac{1}{n-1} \sum_k r_k \sum_{i\neq j} |\rho_{ij}| |\langle d_{ki} | d_{kj} \rangle |.
\label{eq:mnc}
\end{equation}
Now for a given initial detector state $|d_k\rangle$, the success probability of UQSD for the ensemble $\{\rho_{ii}, \ |d_{ki}\rangle \}$ is bounded by $\mathcal{D}_Q^k=1-\frac{1}{n-1}\sum_{i\neq j} \sqrt{\rho_{ii}\rho_{jj}} |\langle d_{ki} | d_{kj} \rangle |$, which represents the path distinguishability. Then, for the initial detector state $\rho_d=\sum_k r_k |d_k \rangle \langle d_k|$, the path distinguishability can be computed by averaging over the individual $\mathcal{D}_Q^k$s, i.e.,
\begin{align}
\mathcal{D}_Q^\prime &= \sum_k r_k \mathcal{D}_Q^k \nonumber \\
&=1-\frac{1}{n-1} \sum_k r_k \sum_{i\neq j} \sqrt{\rho_{ii}\rho_{jj}} |\langle d_{ki} | d_{kj} \rangle |.
\label{eq:mpd}
\end{align}
Now, using Eqs. (\ref{eq:se}), (\ref{eq:mnc}) and (\ref{eq:mpd}),
we arrive at the most general duality relation
\begin{equation}
\mathcal{C}^\prime + \mathcal{D}_Q^\prime \leqslant 1.
\label{eq:genDuality}
\end{equation}
Thus, Eqs. (\ref{duality}), (\ref{c+d}) and (\ref{eq:genDuality}) constitute the central results in our paper.
It is important to note that, for initial mixed quanton and mixed detector states,
a generalized complementarity relation holds, where the particle and wave natures are quantified via path distinguishability based on UQSD and the quantum coherence, respectively.
\section{Conclusions \label{sec:concl}}
In this paper, we have introduced a generalized duality relation for arbitrary dimensional multi-slit quantum interference experiments. To delineate the wave nature of the quanton, which passes through the interferometer, we define normalized quantum coherence based on the recently introduced quantifier of quantum coherence in the framework of quantum information theory \cite{coherence}. Since, both the interference and quantum coherence rely on quantum superposition of the quantum states, we claim that the proposed measure of (normalized) quantum coherence can be a quantifier of the wave-nature of the quanton, instead of traditional quantifier based on the interference fringe visibility. The particle nature of a quanton is associated with the which-path information acquired through the detection process, i.e., which-path detection. In this work, we quantify the which-path information or path distinguishability by identifying it with the upper bound of success probability in unambiguous quantum state discrimination \cite{Helstrom76, Bae15, Pati05, zhang} of the detector states, after the path detectors are placed and the measurement interaction is turned on.
Based on the normalized quantum coherence and the path distinguishability as the
quantifiers the wave and particle natures of a quanton respectively,
we derive a new duality relation,
which is a quantitative statement of Bohr's
principle of complementarity. For two-path and three-path interference we
have related quantum coherence to the fringe visibility and recovered the
corresponding known duality relations. Furthermore, we show that, in cases where
decoherence may introduce some mixedness in the density matrix of
the quanton as well as in the detector, the duality relation continues to hold. We hope that our results will have fundamental implications in understanding the quantum complementarity, in particular, the wave nature of a quantum system in terms of quantum coherence.
\section*{Acknowledgement}
The authors thank A. K. Rajagopal for many fruitful discussions. MNB gratefully acknowledges the support provided by the Centre for Theoretical Physics, Jamia Millia Islamia, New Delhi, during his visit, where this work was initiated. MAS thanks the University Grants Commission, India for financial support.
|
\section{Introduction}
As a $\tau$-charm factory, the BESIII experiment has collected the
world's largest sample of $e^+ e^-$ collision data at
center-of-mass~(CM) energies between 3.810~GeV and~4.600 GeV. In this
energy region, the charmoniumlike states and higher excited charmonium
states are produced copiously, which makes comprehensive studies
possible.
The charmoniumlike states discovered in recent years have drawn great
attention of both theorists and experimentalists for their exotic
properties, as reviewed \emph{e.g.} in Ref.~\cite{Brambilla:2010cs}. Being well
above the open charm threshold, the strong coupling of these states to
hidden charm processes makes their interpretation as conventional
charmonium states very difficult. On the other hand, the theory of the
strong interaction, Quantum Chromodynamics~(QCD), does not prohibit
the existence of exotic states beyond the quark model, \emph{e.g.} molecular
states, tetraquark states, hybrid states, \emph{etc}. Either the
verification or the exclusion of the existence of such states will
help to evaluate the quark model and better understand QCD. Even
though some states have been identified as higher excited charmonium
states, such as the $\psi(4040)$, $\psi(4160)$, and $\psi(4415)$,
their large widths and the interference with each other make their
precise study complicated. In addition, the relationship between the
charmoniumlike states and higher excited charmonium states is still
not clear. The precise knowledge of the time-integrated luminosity is
essential for quantitative analysis of these states.
In this paper, we present a measurement of the integrated luminosity
based on the analysis of the Bhabha scattering process
\mbox{$e^+e^-\rightarrow (\gamma) e^+ e^-$}. A similar method has been
used in the luminosity measurement of $\psi(3770)$ data at
BESIII~\cite{3773}. The process has a simple and clean signature and
a large production cross section, which allows for a small systematic
and a negligible statistical uncertainty. A cross check of the result
is performed by analyzing the di-gamma process \mbox{$e^+ e^-
\rightarrow\gamma\gamma$}.
\section{The detector}
BESIII is a general purpose detector which covers 93\% of the solid
angle and operates at the $e^+ e^-$ collider BEPCII. A detailed
description of the facilities is given in Ref.~\cite{BESIII}. The
detector consists of four main components: (a) A small-cell,
helium-based main drift chamber (MDC) with 43 layers provides an
average single-hit resolution of 135~$\mu$m, and a momentum resolution
of 0.5\% for charged tracks at 1~GeV/$c$ in a 1~T magnetic field. (b)
An electro-magnetic calorimeter~(EMC), consisting of 6240 CsI(Tl)
crystals in a cylindrical structure (barrel and two endcaps). The
energy resolution for 1.0~GeV photons is 2.5\%~(5\%) in the barrel~(endcaps), while the position resolution is 6~mm~(9~mm) in the barrel~(endcaps). (c) A time-of-fight system~(TOF), constructed of 5~cm thick
plastic scintillators, arranged in 88 detectors of 2.4~m length in two
layers in the barrel and 96 fan-shaped detectors in the endcaps. The
barrel (endcap) time resolution of 80~ps~(110~ps) provides 2$\sigma
K/\pi$ separation for momenta up to about 1.0~GeV/$c$. (d) A muon
counter~(MUC), consisting of nine layers of resistive plate chambers
in the barrel and eight layers for each endcap. It is incorporated in
the iron return yoke of the superconducting magnet. Its position
resolution is about 2~cm. A {\sc geant4}~\cite{geant41,geant42} based
detector simulation package has been developed to model the detector
response. Due to the crossing angle of the beams at the
interaction point, the $e^+e^-$ CM system is slightly
boosted with respect to the laboratory frame.
\section{Data sample and Monte Carlo simulation}
Twenty-one data samples have been taken at CM energies between 3.810~GeV and 4.600~GeV. Six of the data sets exceed the others in
accumulated statistics by an order of magnitude. These samples were
taken on the peaks of charmoniumlike states, like the Y(4260),
Y(4360), and Y(4630), or higher excited charmonium states, like
$\psi(4040)$, and $\psi(4415)$, in order to study these resonances and
their decays in great detail. The data samples taken at the other CM
energies serve as scan points to study the behavior of the cross
section around these resonances. All individual data samples are
listed in Table~\ref{inf-Bhabha}.
\begin{table*}[!htbp]
\centering \caption{\label{inf-Bhabha} Center-of-mass energy,
luminosity obtained from the nominal measurement~($L$), cross check
results~($L_{\rm ck}$), and relative differences between the two
results. The uncertainties are statistical only. Superscripts
indicate separate samples acquired at the same CM energy.}
\begin{tabular}{lrrc}
\hline
CM energy~(GeV) &$L$ ($\rm pb^{-1}$) &$L_{\rm ck}$ ($\rm pb^{-1}$) &Relative difference~(\%)\\
\hline
3.810 &50.54$\pm$0.03 &50.11$\pm$0.08 & $-0.85$$\pm$0.17\\
3.900 &52.61$\pm$0.03 &52.57$\pm$0.08 & $-0.08$$\pm$0.17\\
4.009 &481.96$\pm$0.01 &480.54$\pm$0.23 & $-0.30$$\pm$0.05\\
4.090 &52.63$\pm$0.03 &52.37$\pm$0.08 & $-0.49$$\pm$0.17\\
4.190 &43.09$\pm$0.03 &43.08$\pm$0.08 & $-0.03$$\pm$0.20\\
4.210 &54.55$\pm$0.03 &54.27$\pm$0.09 & $-0.62$$\pm$0.18\\
4.220 &54.13$\pm$0.03 &54.22$\pm$0.09 & $+0.17$$\pm$0.18\\
4.230$^1$ &44.40$\pm$0.03 &44.64$\pm$0.08 & $+0.54$$\pm$0.20\\
4.230$^2$ &1047.34$\pm$0.14 &1041.56$\pm$0.37 & $-0.56$$\pm$0.04\\
4.245 &55.59$\pm$0.04 &55.52$\pm$0.09 & $-0.13$$\pm$0.18\\
4.260$^1$ &523.74$\pm$0.10 &524.57$\pm$0.26 & $+0.16$$\pm$0.06\\
4.260$^2$ &301.93$\pm$0.08 &301.11$\pm$0.20 & $-0.28$$\pm$0.08\\
4.310 &44.90$\pm$0.03 &45.29$\pm$0.08 & $+0.87$$\pm$0.19\\
4.360 &539.84$\pm$0.10 &541.38$\pm$0.28 & $+0.29$$\pm$0.06\\
4.390 &55.18$\pm$0.04 &55.27$\pm$0.09 & $+0.16$$\pm$0.18\\
4.420$^{1}$ &44.67$\pm$0.03 &44.77$\pm$0.08 & $+0.22$$\pm$0.20\\
4.420$^{2}$ &1028.89$\pm$0.13 &1029.63$\pm$0.37 & $+0.07$$\pm$0.04\\
4.470 &109.94$\pm$0.04 &109.51$\pm$0.13 & $-0.39$$\pm$0.13\\
4.530 &109.98$\pm$0.04 &109.47$\pm$0.13 & $-0.46$$\pm$0.13\\
4.575 &47.67$\pm$0.03 &47.57$\pm$0.08 & $-0.21$$\pm$0.18\\
4.600 &566.93$\pm$0.11 &563.45$\pm$0.28 & $-0.62$$\pm$0.06\\
\hline
\end{tabular}
\end{table*}
At each energy point, one million Bhabha events were generated using the {\sc
babayaga3.5}~\cite{babayaga} generator with the options presented in
Table~\ref{para}. For the {\sc babayaga3.5} generator, the uncertainty in
calculating the cross section is 0.5\% , which meets the demand of the total
uncertainty of luminosity measurement. The kinematic distributions of the final state
particles from the {\sc babayaga3.5} generator are consistent with those from
data. In the simulation, the scattering angles of the final state particles were
limited to a range from 20 degrees to 160 degrees, which slightly exceeds the
sensitive volume of the detector, in order to save on computing resources. An
energy threshold of 0.04~GeV was applied on the final state particles. The
acolinearity of the events has not been constrained. Finally, the generation
was taking into account the running of the electromagnetic coupling
constant and final state radiation~(FSR).
\begin{table}[!htbp]
\centering
\caption{\label{para} Options for the {\sc babayaga3.5} generator used to generate
the simulated MC data samples.}
\begin{tabular}{lcc}
\hline
Parameters & Value \\
\hline
Ebeam & 2.130 GeV or others \\
MinThetaAngle & 20$^{\circ}$ \\
MaxThetaAngle & 160$^{\circ}$ \\
MinimumEnergy & 0.04 GeV \\
MaximumAcollinearity & 180$^{\circ}$ \\
RunningAlpha & 1 \\
FSR switch & 1 \\
\hline
\end{tabular}
\end{table}
To study the background and optimize the event selection criteria, an
inclusive Monte Carlo~(MC) sample corresponding to a luminosity of
500~$\rm pb^{-1}$ at the CM energy of 4.260~GeV was generated, in which
the QED processes, the continuum production of hadrons, and the
initial state radiation~(ISR) to $J\slash\psi$ and $\psi(3686)$
resonance process were included.
The {\sc babayaga3.5} generator was used to simulate the QED processes,
signal and background. Other processes, such as the decays of the
$J\slash\psi$, were generated with specialized models that have been
packaged and customized for the BESIII Offline Software
System~(BOSS)~(see \cite{prg} for an overview).
\section{Event selection and result}
Signal candidates are required to have exactly two oppositely charged
tracks. The tracks must originate from a cylindrical volume, centered
around the interaction point, which is defined by a radius of 1~cm
perpendicular to the beam axis and a length of $\pm10$~cm along the beam
axis. In addition, the charged tracks are required to be within
$|\cos\theta|<0.8$, where $\theta$ is the polar angle, measured by the
MDC. Without applying further particle identification, the tracks are
assigned as electron and positron depending on their charge. The
deposited energies of electron and positron in EMC must be larger than
$\frac{\sqrt{s}}{4.26} \times 1.55$~(GeV) to remove the di-muon
background, where $\sqrt{s}$ is the CM energy in GeV; the momenta of
electron and positron are required to be larger than
$\frac{\sqrt{s}}{4.26} \times 2$~(GeV/c), to suppress background
events from lighter vector resonances produced in the ISR process,
such as $J\slash\psi$, $\psi(3686)$ and other resonances, decaying
into $e^{+}e^{-}$ pairs. For the data sample with a CM energy of 3.810
or 3.910 GeV, the effect of the remaining $\psi(3686)$ events is
studied by applying a 20\% larger momentum requirement, and is
found to be negligible. The requirements on the deposited energies and
momenta are not optimized in detail, as the number of the signal
events in such an analysis is large enough. All the variables
mentioned above are determined in the initial $e^{+}e^{-}$ CM
frame. The ratio of the number of remaining background events to
the number of signal events, estimated from the inclusive MC sample,
is found to be less than $2\times10^{-4}$, which is negligible. Thus
all the selected events are taken as Bhabha events.
Figure~\ref{4260}
shows the comparisons between data and MC simulation for the kinematic
variables of the leptons by taking data at the CM energy of 4.260~GeV
as an example. Reasonable agreement is observed in the angular and
momentum distributions. The striking difference between data and
simulation found in the distributions of energies deposited by the
leptons in the EMC emerges from imperfections in the simulation of
the energy response of individual detector channels. At the CM
energies analyzed in this work, a single shower in the calorimeter can
be so energetic that the deposited energy per crystal exceeds the
dynamic range of the analog-to-digital converter~(ADC), causing
individual ADC channels to saturate.
In the analysis presented in this report, conditions applied on the
energy deposits are not affected. Relevant deviations between data
and MC are considered as contributions to the systematic uncertainties.
\begin{figure*}[!htbp]
\centering
\subfigure{
\label{4260_1}
\includegraphics[width=0.31\textwidth]{Ep.pdf}}
\subfigure{
\label{4260_3}
\includegraphics[width=0.31\textwidth]{ctp.pdf}}
\subfigure{
\label{4260_5}
\includegraphics[width=0.31\textwidth]{Pp.pdf}}
\subfigure{
\label{4260_2}
\includegraphics[width=0.31\textwidth]{Em.pdf}}
\subfigure{
\label{4260_4}
\includegraphics[width=0.31\textwidth]{ctm.pdf}}
\subfigure{
\label{4260_6}
\includegraphics[width=0.31\textwidth]{Pm.pdf}}
\caption{Comparison between data and MC simulation at the CM
energy of 4.260~GeV. The top row is for positron and the bottom
row for electron. From left to right, the plots show the
distribution of deposited energy in EMC, the distribution of the
cosine of the polar angle measured by the MDC, and the
distribution of the track momentum from the MDC. Black dots with
error bars illustrate data and the red ones are MC simulation.
Note that the y-axis is in logarithmic scale and the MC is
normalized to data by the number of events for each sub-plot.
When drawing the distribution of one variable, the requirements on
the other variables are applied.}
\label{4260}
\end{figure*}
The integrated luminosity is calculated with
\begin{eqnarray}
\label{eq1}
L &=& \frac{N^{\rm obs}_{\rm Bhabha}}{\sigma_{\rm Bhabha}\times\epsilon},
\end{eqnarray}
where $N^{\rm obs}_{\rm Bhabha}$ is the number of observed Bhabha
events, $\sigma_{\rm Bhabha}$ is the cross section of the Bhabha
process, and $\epsilon$ is the efficiency determined by analyzing the
signal MC sample. The cross sections are calculated with the
{\sc babayaga3.5} generator using the parameters listed in
Table~\ref{para} and decrease with increasing energies. The
efficiencies are almost independent of the CM energy, as intended by
the choice of relative conditions on lepton momenta and deposited
energies. The luminosity results calculated with Equation~\ref{eq1}
are listed in Table~\ref{inf-Bhabha}. The statistical accuracy of the
resulting integrated luminosity is better than 0.1\% at all energy
points.
\section{Systematic uncertainty}
The following sources of systematic uncertainties are considered: the
uncertainty of the tracking efficiency, the uncertainty related to the
requirements on the kinematic variables, the statistical uncertainty
of the MC sample, the uncertainty of the beam energy measurement, the
uncertainty of the trigger efficiency, and the systematic uncertainty
of the event generator.
To estimate the systematic uncertainty related to the tracking
efficiency, the Bhabha event sample is selected using information from
the EMC only, without using the tracking information in the MDC. The selection
criteria are: at least two clusters in the EMC for each candidate, and
the two most energetic clusters are assumed to originate from the $\rm
e^{+}e^{-}$ pair; the deposited energies of the two clusters are
required to be larger than $\frac{\sqrt{s}}{4.26}\times 1.8$~(GeV). At
CM energies above 4.420~GeV, the requirement is changed to
$\frac{\sqrt{s}}{4.26}\times 1.55$~(GeV). This adjustment allows to
avoid additional systematic uncertainties which would be introduced by the deviation
of data and simulation in the deposited energy in the EMC, as
discussed in Sec.IV. The polar angle of each cluster is required to be
within $|\rm cos \theta^{EMC}|<0.8$, where $\theta^{\rm EMC}$ is the
polar angle measured by the EMC; to remove the background from the
di-photon process, $\Delta\phi$ is required to be in the range of
[$-40^\circ,-5^\circ$] or [$5^\circ,40^\circ$], where
$\Delta\phi=|\phi_{1}-\phi_{2}|-180^\circ$ and $\phi_{1,2}$ are
the azimuthal angles of the clusters in the EMC boosted to the CM
frame. The efficiency that the selected Bhabha events
pass through the track requirements applied in the nominal analysis is
calculated for both data and MC sample, and the difference between
them is taken as the systematic uncertainty connected to the tracking
efficiency.
The systematic uncertainty in the requirement on the polar angle is
estimated by changing the requirement from \mbox{$|\rm
cos\theta|<0.8$} to $|\rm cos\theta|<0.7$. The difference between
the resulting luminosity and nominal one is taken as the associated
systematic uncertainty. The systematic uncertainty caused by the
requirement on the energy deposited in the EMC is estimated by
changing the requirement from $\frac{\sqrt{s}}{4.26} \times
1.55$~(GeV) to $\frac{\sqrt{s}}{4.26} \times 1.71$~(GeV). The
systematic uncertainty caused by the requirement on the momentum is
estimated by changing the requirement from $\frac{\sqrt{s}}{4.26}
\times 2$~(GeV/c) to $\frac{\sqrt{s}}{4.26} \times 2.06$~(GeV/c). The
ranges are picked as these cause the largest deviations from the
nominal luminosity result near the requirements applied.
The statistical uncertainty of the efficiency determined from MC
simulations is 0.25\%. The CM energy is determined using
\mbox{$e^+e^-\rightarrow (\gamma) \mu^+ \mu^-$} events. The invariant mass of
the di-muon system is calculated taking into account ISR and FSR
effects~\cite{cms}. The difference between the CM energy listed in
Table~\ref{inf-Bhabha} and the one measured with di-muon process is
about 2~MeV, and the corresponding systematic uncertainty is estimated
by changing the CM energy by 2~MeV in the MC simulation. The trigger
efficiency for the Bhabha process is 100\% with an uncertainty of less
than 0.1\%~\cite{tri}. The theoretical uncertainty of the cross
section calculated by the {\sc babayaga3.5} generator is given as
0.5\%~\cite{babayaga}.
The same systematic uncertainty estimation method is applied to all
the sub-samples. The largest relative uncertainty among them is taken
as the associated uncertainty for all the sub-samples. The systematic
uncertainties considered in this work are summarized in
Table~\ref{sys}. By assuming the sources of the systematic
uncertainties to be uncorrelated, the total uncertainty is calculated
as 0.97\% by adding the contributions in quadrature.
\begin{table}[!htbp]
\centering
\caption{ \label{sys} Summary of the systematic uncertainties.}
\begin{tabular}{lc}
\hline
Source & Relative uncertainty (\%) \\
\hline
Tracking efficiency&0.39\\
Energy requirement &0.09\\
Momentum requirement &0.43\\
Polar angle requirement &0.38\\
MC statistics &0.25\\
Beam energy &0.42\\
Trigger efficiency &0.10\\
Generator &0.50\\
\hline
Total &0.97\\
\hline
\end{tabular}
\end{table}
\section{Cross check}
To verify the result, a cross check with di-gamma events is
performed. The event selection criteria are the same as those used in
estimating the systematic uncertainty caused by the tracking efficiency,
except for the requirement on $\Delta\phi$. In order to reduce the Bhabha
background, the $\Delta\phi$ is required to be in the range of
[$-0.8^\circ,0.8^\circ$], since photons are not deflected in the
magnetic field.
The luminosity results of this cross check~($L_{\rm ck}$) are shown in
Table~\ref{inf-Bhabha}, together with the relative differences to the
nominal ones. Both results are well consistent for all individual
measurements, indicating the robustness of the result.
\section{Summary}
The integrated luminosity of the data samples taken at BESIII for studying the
charmoniumlike states and higher excited charmonium states is
measured to a precision of 1\% with Bhabha events. The total
uncertainty is dominated by the systematic uncertainty. A cross check
with di-gamma events is performed and the results are consistent with
each other. The result presented here is essential for future
measurements of cross sections with these data, and it has already
been used in the discovery of charged charmoniumlike
states~\cite{3900,4020,3885,4025}.
\begin{acknowledgements}
The BESIII collaboration would like to thank the staff of BEPCII and
the IHEP computing center for their strong support. This work is
supported in part by National Key Basic Research Program of China
under Contract No.~2015CB856700; National Natural Science Foundation
of China (NSFC) under Contracts Nos.~11125525, 11235011, 11322544,
11335008, 11425524; the Chinese Academy of Sciences (CAS) Large-Scale
Scientific Facility Program; Joint Large-Scale Scientific Facility
Funds of the NSFC and CAS under Contracts Nos.~11179007, U1232201,
U1332201; CAS under Contracts Nos.~KJCX2-YW-N29, KJCX2-YW-N45; 100
Talents Program of CAS; INPAC and Shanghai Key Laboratory for Particle
Physics and Cosmology; German Research Foundation DFG under Contract
No.\ Collaborative Research Center CRC-1044; Istituto Nazionale di
Fisica Nucleare, Italy; Ministry of Development of Turkey under
Contract No.~DPT2006K-120470; Russian Foundation for Basic Research
under Contract No.~14-07-91152; U.S.\ Department of Energy under
Contracts Nos.\ DE-FG02-04ER41291, DE-FG02-05ER41374,
DE-FG02-94ER40823, DESC0010118; U.S.\ National Science Foundation;
University of Groningen (RuG) and the Helmholtzzentrum fuer
Schwerionenforschung GmbH (GSI), Darmstadt; WCU Program of National
Research Foundation of Korea under Contract No. R32-2008-000-10155-0.
\end{acknowledgements}
|
\section{Introduction}
The NLopt (Non-Linear Optimization) library (v2.4.2) \cite{johnson2010nlopt} is a rich collection of optimization routines and algorithms, which provides a platform-independent interface for their use for global and local optimization. The library has been widely used for practical implementations of optimization algorithms as well as for benchmarking new algorithms.
The work in this paper is based on the $IACO_\mathbb{R}$-LS algorithm proposed by Liao, Dorigo et al. \cite{iacor:algo} This algorithm introduced the local search procedure in the original $IACO_\mathbb{R}$ technique, specifically Mtsls1 by Tseng et al.\cite{mtsls} for local search. The $IACO_\mathbb{R}$ was an extension of the $ACO_\mathbb{R}$ algorithm for continuous optimization with the added advantage of a variable size solution archive. The premise of our work lies in improving the local search strategy adopted by $IACO_\mathbb{R}$-LS, by allowing algorithms other than Mtsls1 to be used for local search.
We present a comparison of using various algorithms from the NLopt library for local search procedure in the $IACO_\mathbb{R}$-LS algorithm. In order to introduce a hybrid approach for local search, we use a parameter that probabilistically determines whether to use the Mtsls1 algorithm or the NLopt library algorithm. In case of stagnation, we switch between Mtsls1 or the NLopt algorithm based on the algorithm being used in the previous iteration. The objective is to rigorously analyze the effect of using various optimization algorithms in the local search procedure for $IACO_\mathbb{R}$-LS, as well as provide results on benchmark functions to enable a naive researcher to choose an algorithm easily. To the best of our knowledge, available works in literature have not provided exhaustive comparisons using optimization algorithm libraries on ant colony based approaches, other than \cite{rios2013derivative}. However, surveys on state-of-art in multi-objective evolutionary algorithms \cite{zhou2011multiobjective}, differential evolution \cite{das2011differential} and real-parameter evolutionary multimodal optimization \cite{das2011real} have appeared in literature.
The rest of the paper is organized as follows. Section \ref{sec:hybrid} discusses our hybrid approach which allows using Mtsls1 alongwith an NLopt library algorithm for local search phase of $IACO_\mathbb{R}$-LS. This is followed by a discussion on the NLopt library in Section \ref{sec:nlopt}. We present our results and a discussion in Section \ref{sec:results}, followed by the conclusions in Section \ref{sec:conclusion}.
\section{Hybrid Local Search using Mtsls1 and NLopt algorithms }
\label{sec:hybrid}
We begin by introducing the Mtsls1 algorithm, and the motivation to develop a hybrid approach for local search. This is followed by a description of our algorithm which uses the hybrid local search using Mtsls1 and the algorithms from the NLopt library.
The Multi-Trajectory Local Search, or Mtsls1 algorithm \cite{mtsls} exploits the search space across multiple paths. The approach has evolved to many variants, notable among which are the self-adaptive evolution by Zhao et al. \cite{zhao2011self}, multi-objective optimization \cite{tseng2009multiple} and dynamic search trajectories by Snyman et al. \cite{snyman1987multi}.
Mtsls1 searches along one dimension, and optimum value of one dimension is used as starting point for the next dimension. At each dimension, Mtsls1 tries to move by a step size $s$ along one dimension, and evaluates the change in the function value. If the function value decreases, then new point is used for optimization along the next dimension, If the function value increases, then algorithm goes back to the starting point and moves by a factor of the step size, $0.5*s$ towards negative direction and evaluates the function. Again the function value is compared and based on minimum value of the function, the optimum point is provided.
We propose an hybrid local-search approach which incorporates the non-gradient based Mtsls1, alongwith an algorithm from the NLopt library as part of the $IACO_\mathbb{R}-LS$ technique. Our approach offers a choice between selecting either of the two, based on a probabilistically determined choice. Ths is indicated by the algorithm parameter $P(nlopt)$. In case this probabilistic choice fails to provide any improvement after a specific number of iterations, we switch the algorithm being used based on the algorithm used in the previous iteration. The parameters $ctr_{localsearch}$ and $thresh_{localsearch}$ have been used in our algorithm implement this, as we select a different local search algorithm when $ctr_{localsearch}$ crosses $thresh_{localsearch}$. This ensures that our local-search approach does not stagnate, and also gives our approach an ``adaptive'' flavor.
All algorithms from NLopt library are used as part of the hybrid local search approach. The Nlopt algorithms meant for global optimization are allowed as many function evaluations as set for global search, but for local search, maximum allowed function evaluations in a single local search call is set to $160$. It may be noted here that the method for approximating the gradient is directly linked to algorithm's ability to escape local minima. Solomon \cite{salomon1998evolutionary} opines that \textit{``if, however, the gradient is estimated by independent trials with a distance along each axis, the difference between both classes of algorithms almost vanishes.''} Hence, our computations of gradient are based on approximating the derivative using central differences. By using this method, the maximum number of function evaluations would effectively be ($2*n$ \textit{(for gradient)}$+1$ \textit{(function evaluation)})$*160$) for local search, for an $n$-dimensional problem. Our approach is illustrated in Algorithm \ref{ouralgo}; note that hybrid local-search approach is incorporated at Steps (\ref{ourlocalsearch})-(\ref{ourlocalsearch2}), for additional details reader may refer to \cite{mtsls}.
\begin{algorithm}[H]
\caption{Hybrid $IACO_\mathbb{R}$ Algorithm}\label{ouralgo}
\begin{algorithmic}[1]
\Procedure{Hybrid $IACO_\mathbb{R}$}{Probability ($p$), constant parameter ($\zeta$), initial archive size ($\alpha$), growth ($\gamma$), maximum archive size ($\alpha_{max}$), function tolerance ($\tau$),
maximum failure ($Fail_{max}$), maximum stagnation iterations ($Siter_{max}$), dimensions ($N$), termination criteria ($Tc$), probability of switching to NLopt algorithm ($P(nlopt)$), maximum local search iterations ($thresh_{localsearch}$)
\State \text{Initialize }$\alpha$ \text{ solutions}
\State \text{Evaluate initial solutions}
\While{($Tc$ not satisfied)}
\If {($Fail_{(i,best)} < Fail_{max}$)}
\State \text{Local search from }$Sol_{best}$
\ElsIf{($Fail_{(i,random)} < Fail_{max}$)}
\State \text{Local search from }$Sol_{random}$
\EndIf
\If {$P(nlopt) == 0$}\Comment{Local Search}\label{ourlocalsearch}
\State \text{Use Mtsls1}
\ElsIf{$P(nlopt) == 1$}
\State \text{Use NLopt algo}
\Else
\If{$ctr_{localsearch} < thresh_{localsearch}$}
\If{$rand()<P(nlopt)$}
\State \text{Use NLopt algo}
\Else
\State \text{Use Mtsls1}
\EndIf
\Else
\If{\text{Last iteration used Mtsls1}}
\State \text{Use NLopt algo}
\Else
\State \text{Use Mtsls1}
\EndIf
\EndIf
\EndIf \label{ourlocalsearch2}
\If{\text{No improvement in solution}}
\State \text{Increment }$Fail_i$
\EndIf
\Comment{Continued...}
\algstore{myalg}
\end{algorithmic}
\end{algorithm}
\begin{algorithm}[H]
\begin{algorithmic}[1]
\algrestore{myalg}
\If{$rand() < p$}\Comment{Generate new solution}
\State \text{Sample best Gaussian for each dimension}
\If{New soln is better}\Comment{Exploitation}
\State \text{Substitute new solution for }$Sol_{best}$
\EndIf
\Else
\ForAll{$j \in [1:\alpha]$}\Comment{Exploration}
\State \text{Sample Gaussian along each dimension }$j$
\If{new solution is better}
\State \text{Substitute new solution for }$Sol_{j}$
\EndIf
\EndFor
\EndIf
\If{$Iter_{curr}$ is a multiple of $\gamma$ and $\alpha < \alpha_{max}$}\Comment{Archive Growth}
\State \text{Initialize new solutions using }$S_{new} = S_{new} + rand() \cdot(S_{best}-S_{new})$
\State \text{Add new solution to archive}
\State \text{Increment }$\alpha$
\EndIf
\If{$ctr_{global search}$==$Siter_{max}$}\Comment{Restart}
\State \text{Re-initialize solution set without } $S_{best}$
\EndIf
\EndWhile
\EndProcedure
\end{algorithmic}
\end{algorithm}
\section{The NLopt Library }
\label{sec:nlopt}
The NLopt library optimization algorithms are partitioned into four categories as shown in Figure \ref{nlopt_libs}; algorithms in each category are listed in Table \ref{nlopt_algos}. For the sake of brevity, each algorithm has been assigned a numeric identifier in Table \ref{nlopt_algos} (Col. \textit{``ID''}) which is used to refer to them in the subsequent sections of this paper.
\begin{figure}[h]
\centering
\includegraphics[scale=0.8]{nlopt_library.pdf}
\caption{Categories of algorithms in the NLopt library}
\label{nlopt_libs}
\end{figure}
\begin{table*}[htbp]
\centering
\caption{NLopt algorithms}
\scalebox{0.7}{
\begin{tabular}{|c|c|l|l|}
\hline
\multicolumn{4}{|c|}{\textbf{Summary of Nlopt Algorithms}} \\
\hline
\multicolumn{1}{|c|}{S. No. } & \multicolumn{1}{|c|}{ID} & \multicolumn{1}{|c|}{Algorithm} & \multicolumn{1}{|c|}{Code} \\
\hline
\hline
\multicolumn{4}{|c|}{\textit{\textbf{Global Search Algorithms (Non Derivative Based)}}} \\
\hline
1 & A0 & DIRECT & NLOPT\_GN\_DIRECT \\
2 & A1 & DIRECT-L & NLOPT\_GN\_DIRECT\_L \\
3 & A2 & Randomized DIRECT-L & NLOPT\_GN\_DIRECT\_L\_RAND \\
4 & A3 & Unscaled DIRECT & NLOPT\_GN\_DIRECT\_NOSCAL \\
5 & A4 & Unscaled DIRECT-L & NLOPT\_GN\_DIRECT\_L\_NOSCAL \\
6 & A5 & Unscaled Randomized DIRECT-L & NLOPT\_GN\_DIRECT\_L\_RAND\_NOSCAL \\
7 & A6 & Original DIRECT version & NLOPT\_GN\_ORIG\_DIRECT \\
8 & A7 & Original DIRECT-L version & NLOPT\_GN\_ORIG\_DIRECT\_L \\
9 & A19 & Controlled random search (CRS2) with local mutation & NLOPT\_GN\_CRS2\_LM \\
10 & A20 & Multi-level single-linkage (MLSL), random & NLOPT\_GN\_MLSL \\
11 & A22 & Multi-level single-linkage (MLSL), quasi-random & NLOPT\_GN\_MLSL\_LDS \\
12 & A35 & ISRES evolutionary constrained optimization & NLOPT\_GN\_ISRES \\
13 & A36 & Augmented Lagrangian method & NLOPT\_AUGLAG \\
14 & A37 & Augmented Lagrangian method for equality constraints & NLOPT\_AUGLAG\_EQ \\
15 & A38 & Multi-level single-linkage (MLSL), random & NLOPT\_G\_MLSL \\
16 & A39 & Multi-level single-linkage (MLSL), quasi-random & NLOPT\_G\_MLSL\_LDS \\
17 & A42 & ESCH evolutionary strategy & NLOPT\_GN\_ESCH \\
\hline
\multicolumn{4}{|c|}{\textit{\textbf{Global Search Algorithms (Derivative Based)}}} \\
\hline
18 & A8 & Stochastic Global Optimization (StoGO) & NLOPT\_GD\_STOGO \\
19 & A9 & Stochastic Global Optimization (StoGO), random & NLOPT\_GD\_STOGO\_RAND \\
20 & A21 & Multi-level single-linkage (MLSL), random & NLOPT\_GD\_MLSL \\
21 & A23 & Multi-level single-linkage (MLSL) quasi-random & NLOPT\_GD\_MLSL\_LDS \\
\hline
\multicolumn{4}{|c|}{\textit{\textbf{Local Search Algorithms (Non Derivative Based)}}} \\
\hline
22 & A12 & Principal-axis, PRAXIS & NLOPT\_LN\_PRAXIS \\
23 & A25 & COBYLA & NLOPT\_LN\_COBYLA \\
24 & A26 & NEWUOA unconstrained optimization via quadratic models & NLOPT\_LN\_NEWUOA \\
25 & \textit{A27*} & \textit{Bound-constrained optimization via NEWUOA-based quadratic models} & \textit{NLOPT\_LN\_NEWUOA\_BOUND} \\
26 & A28 & Nelder-Mead simplex algorithm & NLOPT\_LN\_NELDERMEAD \\
27 & A29 & Sbplx variant of Nelder-Mead & NLOPT\_LN\_SBPLX \\
28 & A30 & Augmented Lagrangian method & NLOPT\_LN\_AUGLAG \\
29 & A32 & Augmented Lagrangian method for equality constraints & NLOPT\_LN\_AUGLAG\_EQ \\
30 & A34 & BOBYQA bound-constrained optimization via quadratic models & NLOPT\_LN\_BOBYQA \\
\hline
\multicolumn{4}{|c|}{\textit{\textbf{Local Search Algorithms (Derivative Based)}}} \\
\hline
31 & \textit{A10**} & \textit{Original L-BFGS by Nocedal et al.} & \textit{NLOPT\_LD\_LBFGS\_NOCEDAL}\\
32 & A11 & Limited-memory BFGS & NLOPT\_LD\_LBFGS \\
33 & A13 & Limited-memory variable-metric, rank 1 & NLOPT\_LD\_VAR1 \\
34 & A14 & Limited-memory variable-metric, rank 2 & NLOPT\_LD\_VAR2 \\
35 & A15 & Truncated Newton & NLOPT\_LD\_TNEWTON \\
36 & A16 & Truncated Newton with restarting & NLOPT\_LD\_TNEWTON\_RESTART \\
37 & A17 & Preconditioned truncated Newton & NLOPT\_LD\_TNEWTON\_PRECOND \\
38 & A18 & Preconditioned truncated Newton with restarting & NLOPT\_LD\_TNEWTON\_PRECOND\_RESTART \\
39 & A24 & Method of Moving Asymptotes (MMA) & NLOPT\_LD\_MMA \\
40 & A31 & Augmented Lagrangian method & NLOPT\_LD\_AUGLAG \\
41 & A33 & Augmented Lagrangian method for equality constraints & NLOPT\_LD\_AUGLAG\_EQ \\
42 & A40 & Sequential Quadratic Programming (SQP) & NLOPT\_LD\_SLSQP \\
43 & A41 & CCSA with simple quadratic approximations & NLOPT\_LD\_CCSAQ \\
\hline
\multicolumn{4}{|c|}{\textit{* - This algorithm has not been considered in this study as runtime errors were encountered during execution.}}\\
\multicolumn{4}{|c|}{\textit{** - The original algorithm is not part of NLopt library, after minor modification it has been made part of $A11$.}}\\
\hline
\end{tabular}%
}
\label{nlopt_algos}%
\end{table*}%
The global search algorithms can be categorized into derivative and non-derivative based algorithms. All global optimization algorithms require bound constraints to be specified on the optimization parameters \cite{mtsls}. The DIviding RECTangles (DIRECT) algorithm proposed by Jones et al. \cite{jones1993lipschitzian, finkel2003direct} is based on dividing the search space into hyperrectangles and searching simultaneously at the global and local level. We consider the DIRECT (A0), unscaled DIRECT (A3) and original DIRECT (A6) versions of the algorithm. A locally biased variant of the DIRECT approach proposed by Gablonsky et al. \cite{gablonsky2001locally} is the DIRECT-L, which is expected to perform well on functions with a single global minima and few local minima. We consider the DIRECT-L (A1), randomized DIRECT-L (A2), unscaled DIRECT-L (A4), unscaled randomized DIRECT-L (A5) and the original DIRECT-L (A7) versions of this algorithm in our study.
The Controlled Random Search (CRS) algorithm with local mutation by Kaelo et al. \cite{kaelo2006some} is similar to the idea behind genetic algorithms, where an initial population evolves across generations to converge to the minima. This algorithm uses an evolution strategy similar the Nelder Mead algorithm \cite{lagarias1998convergence}. The version of the CRS algorithm provided in the NLopt library supports bound constraints and starts with an initial population size of $10*(n+1)$ for an $n$ dimensional problem. We use the CRS2 with local mutation (A19) for our study.
The Multi-Level Single Linkage (MLSL) algorithm by Kan and Timmer \cite{kan1987stochastic} is based on selecting multiple start points initially at random, and then using clustering heuristics to traverse the search space effeciently without redundancy. The algorithm configuration in the NLopt library allows for sampling four random points by default, with default function and variable tolerances set to $10^{-15}$ and $10^{-7}$ respectively. We have used the non-derivative-based random MLSL (A20) and quasi-random MLSL (A22) for global search, their derivative-independent versions (A38 and A39), and derivative-based versions (A21 and A23) respectively.
The Improved Stochastic Ranking Evolution Strategy (ISRES) algorithm by Runarsson and Yao \cite{runarsson2005search} supports optimization with both linear and nonlinear constraints (A35). The algorithm uses a mutation rule with log-normal step size and an update rule similar to the Nelder Mead method. The default configuration for the initial population size in the NLopt library is $20*(n+1)$ for an $n$-dimensional function. Another evolutionary algorithm available in the library and used in our study is the ESCH algorithm by Santos et al. \cite{santos2010designing} which supports only linear bound constraints (A42). The STOchastic Global Optimization algorithm (StoGO) by Madsen et al. \cite{madsen1998global, zertchaninov1998c, gudmundsson1998parallel} is a derivative-based global search algorithm which supports only bound constraints. We consider the original StoGO (A8) and its randomized variant (A9).
In the category of non-gradient based local search methods, the Constrained Optimization BY Linear Approximation (COBYLA) by Powell \cite{powell1994direct} supports non-linear equality and inequality constraints (A25). It is based on linearly approximating the objective function using a simplex of $(n+1)$ points for an $n$-dimensional problem. Another version of this algorithm which supports bound constraints is Bound Optimization BY Quadratic Approximation, referred to as BOBYQA \cite{powell2009bobyqa} (A34). Its enhanced version is NEWUOA \cite{powell2006newuoa, powell2008developments, powell2007developments} (A26), with support for constrained and unconstrained problems. The PRincipal AXIS algorithm (PRAXIS) by Brent \cite{brent1973algorithms} primarily supports unconstrained optimization (A12). Other algorithms available in this category in the NLopt library that have been used in this study include the well known Nelder-Mead Simplex method \cite{nelder1965simplex} (A28), and its subplex variant by Rowan \cite{rowan1990functional} (A29).
We now present a brief summary of the derivative based algorithms for local search. The Broyden-Fletcher-Goldfarb-Shanno (BFGS) \cite{wright1999numerical} algorithm (A11) has been a classical optimization algorithm belonging to the class of approximate Newton methods. Along with several variants \cite{wei2004superlinear, wan2014new, liu1989limited, li2001modified}, BFGS has been widely used in the domain of unconstrained optimization. The Method of Moving Asymptotes (MMA) algorithm (A24) by Svanberg \cite{svanberg2002class} is based on locally approximating the gradient of the objective function and a quadratic penalty term. It is an enhanced version of the original Conservative Complex Separable Approximation (CCSA) algorithm \cite{svanberg2002class} (A41), which provides for pre-conditioning of the Hessian in the version available as part of the NLopt library. The Sequential Quadratic Programming (SQP) by Kraft \cite{kraft1988software, kraft1994algorithm} (A40) is available for both linear and non-linear equality and inequality constraints. The preconditioned truncated Newton method \cite{dembo1983truncated} allows for using gradient information from previous iterations, which provides for faster convergence with the trade-off for greater memory requirement. We have used the truncated Newton method (A15), with restart (A16), pre-conditioned Newton method (A17) and pre-conditioned with restart (A18) for our study.
A limited memory variable metric algorithm by Vl{\v{c}}ek and Luk{\v{s}}an \cite{vlvcek2006shifted} is available with rank-1 (A13) and rank-2 (A14) methods in the NLopt library. Also, an Augmented Lagrangian algorithm by Conn et al. \cite{conn1991globally, birgin2008improving} is available for all categories including gradient/non-gradient and global/local search. This algorithm combines the objective and associated constraints into a single function with a penalty term. This is solved separately as another problem without non-linear constraints, to finally converge to the desired solution. Variants of this algorithm to consider penalty function for only equality constraints is also available in the library. We have used the Augmented Lagrangian method (A30), a version with equality constraint support (A32), and the corresponding derivative based versions A31 and A33 respectively. We now present the results of our comparative analysis using these algorithms alongwith Mtsls1 for local search in the following section.
\section{Study and Discussion }
\label{sec:results}
We evaluate the performance of our approach by comparing its performance with the algorithms featured in SOCO and CEC 2014 benchmarks. The results are categorized into three subsections, presenting the results on the SOCO, CEC 2014 benchmarks, and a comparison of the standalone performance of the NLopt algorithms with our hybrid approach. For our study, we use $P(nlopt)$ as 0.6, and parameters for the NLopt algorithms as $x_{tol_{rel}} = x_{tol_{abs}} = 1e-7$, $f_{tol_{rel}} = f_{tol_{abs}} = 1e-15$ (for description of these parameters, the reader may refer to \cite{johnson2010nlopt}). Parameters specific to algorithms have been used with their default configuration, while ranking parameters for performance have been elucidated in \cite{chenproblem},
\subsection{Results on SOCO benchmarks}
We used the 50-dimensional versions of the 19 benchmark functions suite shown in Table \ref{soco_benchmark}. Functions $F1-F6$ were originally proposed for the special session on large scale global optimization organized for the IEEE 2008 Congress on Evolutionary Computation (CEC 2008) \cite{socof1-f6}, Functions $F7-F11$ were proposed at the ISDA 2009 Conference. Functions $F12-F19$ are hybrid functions that combine two functions belonging to $F1-F11$. Some properties of the benchmark functions are listed in Table \ref{soco_benchmark}. The detailed description of these functions is available in \cite{socof1-f19, socof1-press}.
\begin{table}
\caption{SOCO Benchmark functions}
\label{soco_benchmark}
\centering
\scalebox{0.7}{
\begin{tabular}{|c|c|c|c|c|c|c|c|}
\hline ID & Name & Analytical Form & \pbox{20cm}{Uni(U)/\\Multi(M)\\Modal} & Sep. & Rotated & \pbox{20cm}{Easily \\ optimized \\ dimension-\\ wise}\\
\hline $F1$ & Shift Sphere & $\sum_{i=1}^D z_i^2 + f_{bias}, z=x-o$ & U & Y & N & Y\\
\hline $F2$ & Shifted Schwefel 2.21 & $max_i \lbrace |z_i|, 1 \leq i \leq D \rbrace + f_{bias}, z=x-o$ & U & N & N & N \\
\hline $F3$ & Shifted Rosenbrock & $\sum_{i=1}^{D-1} (100 (z_i^2+z_{i+1})^2+(z_i-1)^2))+ f_{bias}, z=x-o$ & M & N & N & Y\\
\hline $F4$ & Shift. Rastrigin &$\sum_{i=1}^D (z_i^2 - 10 cos(2 \pi z_i)+10)+ f_{bias}, z=x-o$ & M & Y & N & Y\\
\hline $F5$ & Shift. Griewank & $\sum_{i=1}^D \frac{z_i^2}{4000} - \prod_{i=1}^D cos(\frac{z_i}{\sqrt{i}})+1+f_{bias}, z=x-o$ & M & N & N & N\\
\hline $F6$ & Shift. Ackley & \pbox{20cm}{$-20 e^{-0.2 \sqrt{\frac{1}{D}\sum_{i=1}^Dz_i^2}} -e^{\frac{1}{D}\sum_{i=1}^Dcos(2 \pi z_i)}+$ \\ $20+e+f_{bias}, z=x-o$} & M & Y & N & Y\\
\hline $F7$ & Shift. Schwefel 2.22 & $\sum_{i=1}^D |z_i| + \prod_{i=1}^D |z_i|, z=x-o$ & U & Y & N & Y\\
\hline $F8$ & Shift Schwefel 1.2 & $\sum_{i=1}^D (\sum_{j=1}^i z_j)^2, z=x-o$ & U & N & N & N\\
\hline $F9$ & Shift. Extended $F_{10}$ & \pbox{20cm}{$\sum_{i=1}^{D-1}f_{10}(z_i,z_{i+1})+f_{10}(z_D,z_1), z=x-o$\\where $f_{10}=(x^2+y^2)^{0.25} (sin^2(50(x^2+y^2)^{0.1})+1)$}& U & N & N & Y\\
\hline $F10$ & Shift. Bohachevsky & \pbox{20cm}{$\sum_{i=1}^D z_i^2 + 2z_{i+1}^2-0.3cos(3\pi z_i)-0.4cos(4\pi z_{i+1})+$ \\ $0.7, z=x-o$ }& U & N & N & N\\
\hline $F11$ & Shift. Schafler & $\sum_{i=1}^{D-1} (z_i^2+z_{i+1}^2)^{0.25} (sin^2(50(z_i^2+z_{i+1}^2)^{0.1})+1), z=x-o$ & U & N & N & Y\\
\hline $F12$ & Hybrid Function & $F9$ + 0.25 $F1$ & M & N & N & N\\
\hline $F13$ & Hybrid Function & $F9$ + 0.25 $F3$ & M & N & N & N\\
\hline $F14$ & Hybrid Function & $F9$ + 0.25 $F4$ & M & N & N & N\\
\hline $F15$ & Hybrid Function & $F10$ + 0.25 $F7$ & M & N & N & N\\
\hline $F16$ & Hybrid Function & $F9$ + 0.5 $F1$ & M & N & N & N\\
\hline $F17$ & Hybrid Function & $F9$ + 0.75 $F3$ & M & N & N & N\\
\hline $F18$ & Hybrid Function & $F9$ + 0.75 $F4$ & M & N & N & N\\
\hline $F19$ & Hybrid Function & $F10$ + 0.75 $F7$ & M & N & N & N\\
\hline \end{tabular}
}
\end{table}
We applied the termination conditions used for SOCO, that is, the maximum number of function evaluations was $5000\times D$, where $D$ denotes the number of dimensions in which the function is considered. All the investigated algorithms were run 25 times on each function. We report error values defined as $f(x)-f(x^*)$, where $x$ is a candidate solution and $x^*$ is the optimal solution. Error values lower than $10^{-14}$ (this value is referred to as 0-threshold) are approximated to 0. Our analysis is based on either the whole solution quality distribution, or on the median and average errors. For the evaluation of our $IACO_\mathbb{R}$-Hybrid approach, we use the algorithm parameters as indicated in Table 2 of \cite{iacor:algo}.
The average and median errors obtained on the benchmark functions have been shown in Tables \ref{nlopt_avg_tab} and \ref{nlopt_med_tab} respectively. The algorithms from the NLopt library (used for hybridization within $IACO_\mathbb{R}$-Mtsls1 framework) have been indicated as the rows of the table using numeric identifiers provided in Table \ref{nlopt_algos}, while the SOCO benchmark functions are provided as the columns. These provide a comprehensive analysis of the performance of these algorithms on the functions; cases where the error is zero have been indicated in boldface. It may be noted here that only the Differential Evolution algorithm (DE) \cite{storn1997differential}, the co-variance matrix adaptation evolution strategy with increasing population size (G-CMA-ES) \cite{auger2005restart} and the real-coded CHC algorithm (CHC) \cite{eshelman1993chapter} have been considered as baseline algorithms for performance evaluation on SOCO benchmarks in \cite{socof1-f19, socof1-press}. Further, the source code for implementation of IACO$_\mathbb{R}$ is available online at http://iridia.ulb.ac.be/supp/IridiaSupp2011-008/.
\begin{landscape}
\begin{table*}
\caption{Average error for NLopt algorithms for 50D}
\label{nlopt_avg_tab}
\scalebox{0.5}{
\begin{tabular} {|c|c|c|c|c|c|c|c|c|c|c|c|c|c|c|c|c|c|c|c|}
\hline ALGOs & F1 & F2 & F3 & F4 & F5 & F6 & F7 & F8 & F9 & F10 & F11 & F12 & F13 & F14 & F15 & F16 & F17 & F18 & F19 \\
\hline A0 & 2.04E-13 & 6.23E+01 & 2.15E+02 & 2.64E+01 & 2.33E-14 & 1.43E+00 & 3.00E-02 & 1.95E+01 & 6.69E-01 & 1.51E-01 & 5.69E-02 & 1.17E-03 & 3.69E+02 & 7.07E+01 & 9.34E-01 & 1.08E-02 & 5.58E+03 & 3.91E+01 & \textbf{0.00E+00} \\
\hline A1 & 5.94E-14 & 2.51E-12 & 1.61E+02 & 7.44E+01 & 5.62E-03 & 8.90E-02 & 2.09E-13 & 1.76E+03 & 7.76E+00 & 1.89E-01 & 4.47E+01 & 6.05E-14 & 1.47E+02 & 5.23E+01 & 1.49E+00 & 4.54E-14 & 3.96E+03 & 5.92E+00 & 2.96E+00 \\
\hline A2 & \textbf{0.00E+00} & 5.11E+01 & 1.38E+02 & 4.26E+00 & 5.42E-03 & 1.41E-01 & \textbf{0.00E+00} & 1.80E-01 & 1.13E-01 & \textbf{0.00E+00} & 5.64E-02 & \textbf{0.00E+00} & 7.32E+00 & 2.26E+00 & \textbf{0.00E+00} & \textbf{0.00E+00} & 1.02E+03 & 4.93E-01 & \textbf{0.00E+00} \\
\hline A3 & 5.94E-14 & 1.28E-12 & 1.61E+02 & 7.39E+01 & 5.62E-03 & 8.90E-02 & 2.20E-13 & 1.76E+03 & 7.76E+00 & 1.89E-01 & 4.47E+01 & 6.05E-14 & 1.47E+02 & 5.43E+01 & 1.49E+00 & 4.54E-14 & 3.96E+03 & 6.15E+00 & 2.96E+00 \\
\hline A4 & 5.94E-14 & 1.28E-12 & 1.61E+02 & 7.39E+01 & 5.62E-03 & 8.90E-02 & 2.20E-13 & 1.76E+03 & 7.76E+00 & 1.89E-01 & 4.47E+01 & 6.05E-14 & 1.47E+02 & 5.43E+01 & 1.49E+00 & 4.54E-14 & 3.96E+03 & 6.15E+00 & 2.96E+00 \\
\hline A5 & 5.94E-14 & 1.28E-12 & 1.61E+02 & 7.39E+01 & 5.62E-03 & 8.90E-02 & 2.20E-13 & 1.76E+03 & 7.76E+00 & 1.89E-01 & 4.47E+01 & 6.05E-14 & 1.47E+02 & 5.43E+01 & 1.49E+00 & 4.54E-14 & 3.96E+03 & 6.15E+00 & 2.96E+00 \\
\hline A6 & 4.96E+00 & 3.69E+01 & 1.10E+04 & 1.14E+02 & 6.81E-01 & 8.90E-02 & 2.42E+00 & 5.51E+03 & 1.26E+02 & 4.69E-01 & 1.32E+02 & 1.30E+02 & 3.68E+04 & 7.88E+01 & 7.07E+00 & 8.53E+01 & 4.65E+03 & 3.91E+01 & 5.61E+00 \\
\hline A7 & \textbf{0.00E+00} & 8.27E-13 & 2.11E+02 & 1.13E+02 & 5.62E-03 & 8.90E-02 & 2.23E-14 & 5.51E+03 & 1.26E+02 & 2.98E-01 & 1.32E+02 & \textbf{0.00E+00} & 1.94E+02 & 7.87E+01 & 3.17E+00 & \textbf{0.00E+00} & 4.58E+03 & 3.91E+01 & 5.53E+00 \\
\hline A8 & \textbf{0.00E+00} & 3.57E-12 & 6.34E+01 & 7.56E-01 & \textbf{0.00E+00} & 6.89E-14 & \textbf{0.00E+00} & 2.48E-05 & \textbf{0.00E+00} & \textbf{0.00E+00} & 1.05E-02 & 4.25E-02 & 3.05E+00 & 5.50E-01 & \textbf{0.00E+00} & 5.36E-03 & 1.30E+02 & 4.64E-02 & \textbf{0.00E+00} \\
\hline A9 & \textbf{0.00E+00} & 3.57E-12 & 6.34E+01 & 7.56E-01 & \textbf{0.00E+00} & 6.89E-14 & \textbf{0.00E+00} & 2.48E-05 & \textbf{0.00E+00} & \textbf{0.00E+00} & 1.05E-02 & 4.25E-02 & 3.05E+00 & 5.50E-01 & \textbf{0.00E+00} & 5.36E-03 & 1.30E+02 & 4.64E-02 & \textbf{0.00E+00} \\
\hline A11 & \textbf{0.00E+00} & \textbf{0.00E+00} & 1.13E-11 & \textbf{0.00E+00} & \textbf{0.00E+00} & \textbf{0.00E+00} & \textbf{0.00E+00} & \textbf{0.00E+00} & \textbf{0.00E+00} & \textbf{0.00E+00} & \textbf{0.00E+00} & 7.27E-03 & 1.86E+00 & 4.37E-01 & \textbf{0.00E+00} & \textbf{0.00E+00} & 6.89E+00 & 1.63E-01 & \textbf{0.00E+00} \\
\hline A12 & \textbf{0.00E+00} & 1.53E-14 & 8.20E+01 & 3.98E-02 & \textbf{0.00E+00} & 1.14E-14 & \textbf{0.00E+00} & 3.07E-09 & \textbf{0.00E+00} & \textbf{0.00E+00} & 3.75E-02 & \textbf{0.00E+00} & 3.38E+00 & 3.90E-01 & \textbf{0.00E+00} & 6.64E-03 & 1.70E+02 & 2.62E-03 & \textbf{0.00E+00} \\
\hline A13 & \textbf{0.00E+00} & \textbf{0.00E+00} & 8.11E-14 & 1.75E+00 & \textbf{0.00E+00} & \textbf{0.00E+00} & \textbf{0.00E+00} & \textbf{0.00E+00} & \textbf{0.00E+00} & \textbf{0.00E+00} & 7.82E-02 & 5.15E-03 & 1.49E+00 & 1.12E+00 & \textbf{0.00E+00} & 4.29E-04 & 8.01E+01 & 1.02E+00 & \textbf{0.00E+00} \\
\hline A14 & \textbf{0.00E+00} & \textbf{0.00E+00} & 8.11E-14 & 1.75E+00 & \textbf{0.00E+00} & \textbf{0.00E+00} & \textbf{0.00E+00} & \textbf{0.00E+00} & \textbf{0.00E+00} & \textbf{0.00E+00} & 7.82E-02 & 5.15E-03 & 1.49E+00 & 1.12E+00 & \textbf{0.00E+00} & 4.29E-04 & 8.01E+01 & 1.02E+00 & \textbf{0.00E+00} \\
\hline A15 & \textbf{0.00E+00} & \textbf{0.00E+00} & 3.19E-01 & 5.97E-01 & \textbf{0.00E+00} & \textbf{0.00E+00} & \textbf{0.00E+00} & \textbf{0.00E+00} & 2.33E-01 & \textbf{0.00E+00} & 7.84E-01 & 1.71E-02 & 6.37E+00 & 2.85E+00 & \textbf{0.00E+00} & 1.51E-01 & 8.49E+01 & 2.57E+00 & \textbf{0.00E+00} \\
\hline A16 & \textbf{0.00E+00} & \textbf{0.00E+00} & 9.57E-01 & 1.68E-13 & \textbf{0.00E+00} & \textbf{0.00E+00} & \textbf{0.00E+00} & \textbf{0.00E+00} & 4.88E-02 & \textbf{0.00E+00} & 6.32E-04 & 6.00E-02 & 6.81E+00 & 1.69E+00 & \textbf{0.00E+00} & 1.84E-01 & 4.10E+01 & 1.30E+00 & \textbf{0.00E+00} \\
\hline A17 & \textbf{0.00E+00} & \textbf{0.00E+00} & 4.78E-01 & 7.96E-02 & 3.95E-04 & \textbf{0.00E+00} & \textbf{0.00E+00} & \textbf{0.00E+00} & 3.97E+00 & \textbf{0.00E+00} & 1.96E+00 & 2.60E-02 & 8.10E+00 & 1.75E+00 & \textbf{0.00E+00} & 2.00E+00 & 3.78E+01 & 3.10E+00 & \textbf{0.00E+00} \\
\hline A18 & \textbf{0.00E+00} & \textbf{0.00E+00} & 7.97E-01 & 1.59E+00 & 6.57E-13 & \textbf{0.00E+00} & \textbf{0.00E+00} & \textbf{0.00E+00} & 4.35E-01 & \textbf{0.00E+00} & 6.95E-01 & 6.98E-02 & 2.47E+00 & 1.42E+00 & \textbf{0.00E+00} & 1.68E-02 & 6.57E+01 & 3.52E-01 & \textbf{0.00E+00} \\
\hline A19 & 2.34E-13 & 1.72E+01 & 1.53E+02 & 5.66E+01 & 8.77E-03 & 1.37E-01 & 1.83E-06 & 2.95E+01 & 1.03E+02 & 2.24E+00 & 1.06E+02 & 4.71E+01 & 1.06E+02 & 6.97E+01 & 3.37E-01 & 1.35E+02 & 2.25E+02 & 4.67E+01 & 2.21E+00 \\
\hline A20 & 9.60E-11 & 2.56E+01 & 1.73E+02 & 9.99E+01 & 6.87E-10 & 1.80E+00 & 9.59E+00 & 4.58E-06 & 2.42E+02 & 1.09E+01 & 2.55E+02 & 5.88E+01 & 8.80E+01 & 1.06E+02 & 1.39E+01 & 1.30E+02 & 9.07E+02 & 7.73E+01 & 1.39E+01 \\
\hline A21 & \textbf{0.00E+00} & 3.75E+01 & 2.07E+02 & 3.98E+02 & \textbf{0.00E+00} & 1.10E+01 & 2.25E+07 & 3.06E+01 & 2.76E+02 & 1.18E+01 & 2.80E+02 & 7.64E+01 & 1.03E+02 & 3.22E+02 & 9.38E+02 & 1.66E+02 & 4.23E+02 & 1.48E+02 & 1.31E+01 \\
\hline A22 & 1.07E-10 & 2.61E+01 & 1.52E+02 & 1.04E+02 & 6.82E-10 & 1.78E+00 & 5.94E+00 & 4.90E-06 & 2.47E+02 & 1.08E+01 & 2.50E+02 & 5.97E+01 & 8.86E+01 & 8.88E+01 & 1.47E+01 & 1.30E+02 & 1.30E+03 & 7.11E+01 & 1.39E+01 \\
\hline A23 & \textbf{0.00E+00} & 3.70E-03 & 1.07E+02 & 3.03E+00 & \textbf{0.00E+00} & \textbf{0.00E+00} & \textbf{0.00E+00} & 5.52E-01 & 2.37E+00 & \textbf{0.00E+00} & 1.93E+00 & 4.39E-01 & 1.90E+01 & 3.32E+00 & 2.62E-13 & 1.58E+00 & 2.06E+01 & 2.69E+00 & \textbf{0.00E+00} \\
\hline A24 & \textbf{0.00E+00} & 6.37E-03 & 1.81E+02 & \textbf{0.00E+00} & \textbf{0.00E+00} & \textbf{0.00E+00} & \textbf{0.00E+00} & 5.60E-01 & \textbf{0.00E+00} & \textbf{0.00E+00} & 5.92E-02 & 1.54E-01 & 9.38E+00 & 1.92E+00 & \textbf{0.00E+00} & 1.08E-01 & 2.03E+02 & 2.39E-01 & \textbf{0.00E+00} \\
\hline A25 & \textbf{0.00E+00} & 3.68E-12 & 7.69E+01 & 4.78E-01 & \textbf{0.00E+00} & 6.91E-14 & \textbf{0.00E+00} & 3.34E-05 & \textbf{0.00E+00} & \textbf{0.00E+00} & 5.87E-02 & 7.04E-02 & 2.10E+00 & 9.85E-03 & \textbf{0.00E+00} & \textbf{0.00E+00} & 1.50E+02 & \textbf{0.00E+00} & \textbf{0.00E+00} \\
\hline A26 & \textbf{0.00E+00} & 3.31E-12 & 7.07E+01 & 7.96E-02 & \textbf{0.00E+00} & 6.82E-14 & \textbf{0.00E+00} & 3.37E-05 & 3.47E-02 & \textbf{0.00E+00} & 1.05E-02 & 8.72E-03 & 2.74E+00 & 2.88E-01 & \textbf{0.00E+00} & \textbf{0.00E+00} & 2.31E+01 & \textbf{0.00E+00} & \textbf{0.00E+00} \\
\hline A28 & \textbf{0.00E+00} & 3.40E-12 & 1.35E+02 & 6.77E-01 & \textbf{0.00E+00} & 7.25E-14 & \textbf{0.00E+00} & 5.09E-05 & 1.18E-01 & \textbf{0.00E+00} & 3.47E-02 & 9.32E-02 & 2.71E+00 & 3.66E-01 & \textbf{0.00E+00} & 1.82E-01 & 1.80E+02 & 1.59E-01 & \textbf{0.00E+00} \\
\hline A29 & \textbf{0.00E+00} & 3.54E-12 & 4.35E+01 & 6.37E-01 & \textbf{0.00E+00} & 6.79E-14 & \textbf{0.00E+00} & 2.93E-05 & 2.08E-03 & \textbf{0.00E+00} & 4.64E-02 & 7.51E-02 & 3.25E+00 & 4.33E-01 & \textbf{0.00E+00} & \textbf{0.00E+00} & 8.55E+01 & 3.18E-01 & \textbf{0.00E+00} \\
\hline A30 & \textbf{0.00E+00} & 3.68E-12 & 7.69E+01 & 4.78E-01 & \textbf{0.00E+00} & 6.91E-14 & \textbf{0.00E+00} & 3.35E-05 & \textbf{0.00E+00} & \textbf{0.00E+00} & 5.87E-02 & 7.04E-02 & 2.10E+00 & 9.85E-03 & \textbf{0.00E+00} & \textbf{0.00E+00} & 1.50E+02 & \textbf{0.00E+00} & \textbf{0.00E+00} \\
\hline A31 & \textbf{0.00E+00} & 6.37E-03 & 1.81E+02 & \textbf{0.00E+00} & \textbf{0.00E+00} & \textbf{0.00E+00} & \textbf{0.00E+00} & 5.60E-01 & \textbf{0.00E+00} & \textbf{0.00E+00} & 5.92E-02 & 1.54E-01 & 9.38E+00 & 1.92E+00 & \textbf{0.00E+00} & 1.08E-01 & 2.03E+02 & 2.39E-01 & \textbf{0.00E+00} \\
\hline A32 & \textbf{0.00E+00} & 3.68E-12 & 7.69E+01 & 4.78E-01 & \textbf{0.00E+00} & 6.91E-14 & \textbf{0.00E+00} & 3.35E-05 & \textbf{0.00E+00} & \textbf{0.00E+00} & 5.87E-02 & 7.04E-02 & 2.10E+00 & 9.85E-03 & \textbf{0.00E+00} & \textbf{0.00E+00} & 1.50E+02 & \textbf{0.00E+00} & \textbf{0.00E+00} \\
\hline A33 & \textbf{0.00E+00} & 6.37E-03 & 1.81E+02 & \textbf{0.00E+00} & \textbf{0.00E+00} & \textbf{0.00E+00} & \textbf{0.00E+00} & 5.60E-01 & \textbf{0.00E+00} & \textbf{0.00E+00} & 5.92E-02 & 1.54E-01 & 9.38E+00 & 1.92E+00 & \textbf{0.00E+00} & 1.08E-01 & 2.03E+02 & 2.39E-01 & \textbf{0.00E+00} \\
\hline A34 & \textbf{0.00E+00} & 3.57E-12 & 1.01E+02 & 7.56E-01 & \textbf{0.00E+00} & 6.89E-14 & \textbf{0.00E+00} & 2.94E-05 & \textbf{0.00E+00} & \textbf{0.00E+00} & 1.05E-02 & 2.60E-03 & 3.06E+00 & 5.50E-01 & \textbf{0.00E+00} & 5.36E-03 & 1.81E+02 & 4.64E-02 & \textbf{0.00E+00} \\
\hline A35 & 6.74E+02 & 1.17E+01 & 8.15E+06 & 2.44E+02 & 7.24E+00 & 4.01E+00 & 1.61E+01 & 6.14E+03 & 1.59E+02 & 7.39E+01 & 1.60E+02 & 2.92E+02 & 1.54E+06 & 1.67E+02 & 2.77E+01 & 2.10E+02 & 5.89E+02 & 7.42E+01 & 3.97E+01 \\
\hline A36 & \textbf{0.00E+00} & 3.57E-12 & 6.34E+01 & 7.56E-01 & \textbf{0.00E+00} & 6.89E-14 & \textbf{0.00E+00} & 2.48E-05 & \textbf{0.00E+00} & \textbf{0.00E+00} & 1.05E-02 & 4.25E-02 & 3.05E+00 & 5.50E-01 & \textbf{0.00E+00} & 5.36E-03 & 1.30E+02 & 4.64E-02 & \textbf{0.00E+00} \\
\hline A37 & \textbf{0.00E+00} & 3.57E-12 & 6.34E+01 & 7.56E-01 & \textbf{0.00E+00} & 6.89E-14 & \textbf{0.00E+00} & 2.48E-05 & \textbf{0.00E+00} & \textbf{0.00E+00} & 1.05E-02 & 4.25E-02 & 3.05E+00 & 5.50E-01 & \textbf{0.00E+00} & 5.36E-03 & 1.30E+02 & 4.64E-02 & \textbf{0.00E+00} \\
\hline A38 & \textbf{0.00E+00} & 3.57E-12 & 6.34E+01 & 7.56E-01 & \textbf{0.00E+00} & 6.89E-14 & \textbf{0.00E+00} & 2.48E-05 & \textbf{0.00E+00} & \textbf{0.00E+00} & 1.05E-02 & 4.25E-02 & 3.05E+00 & 5.50E-01 & \textbf{0.00E+00} & 5.36E-03 & 1.30E+02 & 4.64E-02 & \textbf{0.00E+00} \\
\hline A39 & \textbf{0.00E+00} & 3.57E-12 & 6.34E+01 & 7.56E-01 & \textbf{0.00E+00} & 6.89E-14 & \textbf{0.00E+00} & 2.48E-05 & \textbf{0.00E+00} & \textbf{0.00E+00} & 1.05E-02 & 4.25E-02 & 3.05E+00 & 5.50E-01 & \textbf{0.00E+00} & 5.36E-03 & 1.30E+02 & 4.64E-02 & \textbf{0.00E+00} \\
\hline A40 & \textbf{0.00E+00} & \textbf{0.00E+00} & 4.33E+00 & 9.66E-01 & \textbf{0.00E+00} & \textbf{0.00E+00} & \textbf{0.00E+00} & \textbf{0.00E+00} & 1.18E-01 & \textbf{0.00E+00} & 9.78E-01 & \textbf{0.00E+00} & 2.35E+00 & 1.24E+00 & \textbf{0.00E+00} & \textbf{0.00E+00} & 1.90E+02 & \textbf{0.00E+00} & \textbf{0.00E+00} \\
\hline A41 & \textbf{0.00E+00} & 9.23E-03 & 4.88E+01 & 3.98E-02 & \textbf{0.00E+00} & \textbf{0.00E+00} & \textbf{0.00E+00} & 7.20E-01 & 5.63E+00 & \textbf{0.00E+00} & 2.10E-02 & \textbf{0.00E+00} & 1.71E+01 & 1.54E-01 & \textbf{0.00E+00} & 5.57E-01 & 2.60E+02 & 8.07E-02 & \textbf{0.00E+00} \\
\hline A42 & \textbf{0.00E+00} & 4.48E-12 & 6.94E+01 & 1.59E-01 & \textbf{0.00E+00} & 6.92E-14 & \textbf{0.00E+00} & 3.40E-05 & \textbf{0.00E+00} & \textbf{0.00E+00} & 5.69E-02 & 1.29E-02 & 2.49E+00 & 4.80E-02 & \textbf{0.00E+00} & 5.36E-03 & 1.19E+02 & 1.21E-01 & \textbf{0.00E+00} \\
\hline
\end{tabular}
}
\end{table*}
\end{landscape}
\begin{landscape}
\begin{table*}
\caption{Median error for NLopt algorithms for 50D}
\label{nlopt_med_tab}
\scalebox{0.5}{
\begin{tabular} {|c|c|c|c|c|c|c|c|c|c|c|c|c|c|c|c|c|c|c|c|}
\hline ALGOs & F1 & F2 & F3 & F4 & F5 & F6 & F7 & F8 & F9 & F10 & F11 & F12 & F13 & F14 & F15 & F16 & F17 & F18 & F19 \\
\hline A0 & \textbf{0.00E+00} & 9.66E+01 & 8.19E+01 & 1.49E+01 & \textbf{0.00E+00} & 2.56E+00 & 3.61E-11 & 9.73E+00 & 9.34E-05 & \textbf{0.00E+00} & 1.49E-08 & 9.22E-04 & 5.32E+02 & 8.87E+01 & 1.41E-12 & 1.07E-02 & 7.61E+03 & 4.38E+01 & \textbf{0.00E+00} \\
\hline A1 & \textbf{0.00E+00} & 2.40E-12 & 1.11E+02 & 8.06E+01 & 7.40E-03 & 1.59E-01 & 3.74E-13 & 5.90E+02 & 1.26E+00 & 1.90E-12 & 2.67E+01 & 1.08E-13 & 1.53E+02 & 5.55E+01 & 2.18E-09 & 8.11E-14 & 5.35E+03 & 3.98E+00 & 1.16E-07 \\
\hline A2 & \textbf{0.00E+00} & 7.66E+01 & 3.10E+01 & \textbf{0.00E+00} & \textbf{0.00E+00} & 7.51E-14 & \textbf{0.00E+00} & 8.91E-03 & \textbf{0.00E+00} & \textbf{0.00E+00} & \textbf{0.00E+00} & \textbf{0.00E+00} & 3.77E+00 & \textbf{0.00E+00} & \textbf{0.00E+00} & \textbf{0.00E+00} & 8.59E+01 & \textbf{0.00E+00} & \textbf{0.00E+00} \\
\hline A3 & \textbf{0.00E+00} & 1.22E-12 & 1.11E+02 & 8.06E+01 & 7.40E-03 & 1.59E-01 & 3.92E-13 & 5.90E+02 & 1.26E+00 & 1.90E-12 & 2.67E+01 & 1.08E-13 & 1.53E+02 & 5.55E+01 & 2.18E-09 & 8.11E-14 & 5.35E+03 & 3.98E+00 & 1.16E-07 \\
\hline A4 & \textbf{0.00E+00} & 1.22E-12 & 1.11E+02 & 8.06E+01 & 7.40E-03 & 1.59E-01 & 3.92E-13 & 5.90E+02 & 1.26E+00 & 1.90E-12 & 2.67E+01 & 1.08E-13 & 1.53E+02 & 5.55E+01 & 2.18E-09 & 8.11E-14 & 5.35E+03 & 3.98E+00 & 1.16E-07 \\
\hline A5 & \textbf{0.00E+00} & 1.22E-12 & 1.11E+02 & 8.06E+01 & 7.40E-03 & 1.59E-01 & 3.92E-13 & 5.90E+02 & 1.26E+00 & 1.90E-12 & 2.67E+01 & 1.08E-13 & 1.53E+02 & 5.55E+01 & 2.18E-09 & 8.11E-14 & 5.35E+03 & 3.98E+00 & 1.16E-07 \\
\hline A6 & 8.26E+00 & 5.68E+01 & 1.91E+04 & 1.26E+02 & 1.06E+00 & 1.59E-01 & 4.04E+00 & 5.99E+03 & 1.62E+02 & 7.54E-01 & 1.63E+02 & 2.02E+02 & 6.13E+04 & 1.06E+02 & 1.04E+01 & 1.20E+02 & 6.31E+03 & 4.38E+01 & 9.32E+00 \\
\hline A7 & \textbf{0.00E+00} & 8.10E-13 & 2.55E+02 & 1.25E+02 & 7.40E-03 & 1.59E-01 & 2.32E-14 & 5.99E+03 & 1.62E+02 & 4.70E-01 & 1.63E+02 & \textbf{0.00E+00} & 2.59E+02 & 1.06E+02 & 4.67E+00 & \textbf{0.00E+00} & 6.22E+03 & 4.38E+01 & 9.19E+00 \\
\hline A8 & \textbf{0.00E+00} & 1.75E-12 & 1.88E+01 & \textbf{0.00E+00} & \textbf{0.00E+00} & 7.51E-14 & \textbf{0.00E+00} & 2.16E-05 & \textbf{0.00E+00} & \textbf{0.00E+00} & \textbf{0.00E+00} & \textbf{0.00E+00} & 2.06E+00 & \textbf{0.00E+00} & \textbf{0.00E+00} & \textbf{0.00E+00} & 4.48E+01 & \textbf{0.00E+00} & \textbf{0.00E+00} \\
\hline A9 & \textbf{0.00E+00} & 1.75E-12 & 1.88E+01 & \textbf{0.00E+00} & \textbf{0.00E+00} & 7.51E-14 & \textbf{0.00E+00} & 2.16E-05 & \textbf{0.00E+00} & \textbf{0.00E+00} & \textbf{0.00E+00} & \textbf{0.00E+00} & 2.06E+00 & \textbf{0.00E+00} & \textbf{0.00E+00} & \textbf{0.00E+00} & 4.48E+01 & \textbf{0.00E+00} & \textbf{0.00E+00} \\
\hline A11 & \textbf{0.00E+00} & \textbf{0.00E+00} & \textbf{0.00E+00} & \textbf{0.00E+00} & \textbf{0.00E+00} & \textbf{0.00E+00} & \textbf{0.00E+00} & \textbf{0.00E+00} & \textbf{0.00E+00} & \textbf{0.00E+00} & \textbf{0.00E+00} & \textbf{0.00E+00} & 1.17E-01 & \textbf{0.00E+00} & \textbf{0.00E+00} & \textbf{0.00E+00} & 8.55E-02 & \textbf{0.00E+00} & \textbf{0.00E+00} \\
\hline A12 & \textbf{0.00E+00} & 1.42E-14 & 1.73E+01 & \textbf{0.00E+00} & \textbf{0.00E+00} & 1.11E-14 & \textbf{0.00E+00} & 1.98E-09 & \textbf{0.00E+00} & \textbf{0.00E+00} & \textbf{0.00E+00} & \textbf{0.00E+00} & 2.02E+00 & \textbf{0.00E+00} & \textbf{0.00E+00} & \textbf{0.00E+00} & 2.91E+01 & \textbf{0.00E+00} & \textbf{0.00E+00} \\
\hline A13 & \textbf{0.00E+00} & \textbf{0.00E+00} & 2.38E-14 & \textbf{0.00E+00} & \textbf{0.00E+00} & \textbf{0.00E+00} & \textbf{0.00E+00} & \textbf{0.00E+00} & \textbf{0.00E+00} & \textbf{0.00E+00} & \textbf{0.00E+00} & \textbf{0.00E+00} & 1.92E-01 & \textbf{0.00E+00} & \textbf{0.00E+00} & \textbf{0.00E+00} & 4.12E-01 & \textbf{0.00E+00} & \textbf{0.00E+00} \\
\hline A14 & \textbf{0.00E+00} & \textbf{0.00E+00} & 2.38E-14 & \textbf{0.00E+00} & \textbf{0.00E+00} & \textbf{0.00E+00} & \textbf{0.00E+00} & \textbf{0.00E+00} & \textbf{0.00E+00} & \textbf{0.00E+00} & \textbf{0.00E+00} & \textbf{0.00E+00} & 1.92E-01 & \textbf{0.00E+00} & \textbf{0.00E+00} & \textbf{0.00E+00} & 4.12E-01 & \textbf{0.00E+00} & \textbf{0.00E+00} \\
\hline A15 & \textbf{0.00E+00} & \textbf{0.00E+00} & \textbf{0.00E+00} & \textbf{0.00E+00} & \textbf{0.00E+00} & \textbf{0.00E+00} & \textbf{0.00E+00} & \textbf{0.00E+00} & \textbf{0.00E+00} & \textbf{0.00E+00} & \textbf{0.00E+00} & \textbf{0.00E+00} & 3.95E-01 & \textbf{0.00E+00} & \textbf{0.00E+00} & \textbf{0.00E+00} & 9.71E+00 & \textbf{0.00E+00} & \textbf{0.00E+00} \\
\hline A16 & \textbf{0.00E+00} & \textbf{0.00E+00} & \textbf{0.00E+00} & \textbf{0.00E+00} & \textbf{0.00E+00} & \textbf{0.00E+00} & \textbf{0.00E+00} & \textbf{0.00E+00} & \textbf{0.00E+00} & \textbf{0.00E+00} & \textbf{0.00E+00} & \textbf{0.00E+00} & 4.05E+00 & \textbf{0.00E+00} & \textbf{0.00E+00} & \textbf{0.00E+00} & 4.19E+00 & \textbf{0.00E+00} & \textbf{0.00E+00} \\
\hline A17 & \textbf{0.00E+00} & \textbf{0.00E+00} & \textbf{0.00E+00} & \textbf{0.00E+00} & \textbf{0.00E+00} & \textbf{0.00E+00} & \textbf{0.00E+00} & \textbf{0.00E+00} & \textbf{0.00E+00} & \textbf{0.00E+00} & \textbf{0.00E+00} & \textbf{0.00E+00} & 3.99E+00 & \textbf{0.00E+00} & \textbf{0.00E+00} & \textbf{0.00E+00} & 7.04E-01 & \textbf{0.00E+00} & \textbf{0.00E+00} \\
\hline A18 & \textbf{0.00E+00} & \textbf{0.00E+00} & \textbf{0.00E+00} & \textbf{0.00E+00} & \textbf{0.00E+00} & \textbf{0.00E+00} & \textbf{0.00E+00} & \textbf{0.00E+00} & \textbf{0.00E+00} & \textbf{0.00E+00} & \textbf{0.00E+00} & \textbf{0.00E+00} & 3.52E-01 & \textbf{0.00E+00} & \textbf{0.00E+00} & \textbf{0.00E+00} & 6.19E-01 & \textbf{0.00E+00} & \textbf{0.00E+00} \\
\hline A19 & \textbf{0.00E+00} & 2.35E+01 & 9.07E+01 & 1.89E+01 & 1.45E-12 & 1.44E-12 & 1.19E-11 & 2.42E+01 & 1.15E+02 & \textbf{0.00E+00} & 1.19E+02 & 3.73E+01 & 1.20E+02 & 5.47E+01 & 2.86E-08 & 2.02E+02 & 2.32E+02 & 4.98E+01 & 5.50E-11 \\
\hline A20 & 1.27E-10 & 3.64E+01 & 4.08E+00 & 1.08E+02 & 8.32E-10 & 2.94E+00 & 2.55E+00 & 4.14E-06 & 3.28E+02 & 1.62E+01 & 3.47E+02 & 7.34E+01 & 9.12E+01 & 1.34E+02 & 1.76E+01 & 1.81E+02 & 3.26E+02 & 9.68E+01 & 2.11E+01 \\
\hline A21 & \textbf{0.00E+00} & 4.08E+01 & 6.79E+01 & 5.45E+02 & \textbf{0.00E+00} & 1.95E+01 & 1.17E+05 & 2.49E+01 & 3.94E+02 & 1.83E+01 & 3.86E+02 & 1.04E+02 & 1.12E+02 & 4.51E+02 & 1.22E+03 & 2.37E+02 & 3.89E+02 & 1.74E+02 & 1.56E+01 \\
\hline A22 & 1.39E-10 & 3.60E+01 & 9.27E+00 & 1.13E+02 & 7.58E-10 & 2.48E+00 & 1.35E+00 & 3.58E-06 & 3.28E+02 & 1.62E+01 & 3.47E+02 & 7.61E+01 & 9.44E+01 & 9.46E+01 & 1.82E+01 & 1.81E+02 & 5.25E+02 & 8.91E+01 & 1.90E+01 \\
\hline A23 & \textbf{0.00E+00} & 3.81E-12 & 2.67E+01 & \textbf{0.00E+00} & \textbf{0.00E+00} & \textbf{0.00E+00} & \textbf{0.00E+00} & 1.72E-01 & 7.05E-02 & \textbf{0.00E+00} & 1.07E-02 & \textbf{0.00E+00} & 1.83E+01 & 7.37E-08 & \textbf{0.00E+00} & \textbf{0.00E+00} & 2.26E+00 & \textbf{0.00E+00} & \textbf{0.00E+00} \\
\hline A24 & \textbf{0.00E+00} & 1.42E-14 & 5.35E+01 & \textbf{0.00E+00} & \textbf{0.00E+00} & \textbf{0.00E+00} & \textbf{0.00E+00} & 1.06E-01 & \textbf{0.00E+00} & \textbf{0.00E+00} & \textbf{0.00E+00} & \textbf{0.00E+00} & 8.55E+00 & \textbf{0.00E+00} & \textbf{0.00E+00} & \textbf{0.00E+00} & 5.12E+00 & \textbf{0.00E+00} & \textbf{0.00E+00} \\
\hline A25 & \textbf{0.00E+00} & 1.78E-12 & 1.88E+01 & \textbf{0.00E+00} & \textbf{0.00E+00} & 6.79E-14 & \textbf{0.00E+00} & 2.73E-05 & \textbf{0.00E+00} & \textbf{0.00E+00} & \textbf{0.00E+00} & \textbf{0.00E+00} & 3.41E-01 & \textbf{0.00E+00} & \textbf{0.00E+00} & \textbf{0.00E+00} & 2.87E+01 & \textbf{0.00E+00} & \textbf{0.00E+00} \\
\hline A26 & \textbf{0.00E+00} & 2.52E-12 & 1.72E+01 & \textbf{0.00E+00} & \textbf{0.00E+00} & 6.79E-14 & \textbf{0.00E+00} & 2.60E-05 & \textbf{0.00E+00} & \textbf{0.00E+00} & \textbf{0.00E+00} & \textbf{0.00E+00} & 3.36E-01 & \textbf{0.00E+00} & \textbf{0.00E+00} & \textbf{0.00E+00} & 1.91E+00 & \textbf{0.00E+00} & \textbf{0.00E+00} \\
\hline A28 & \textbf{0.00E+00} & 1.78E-12 & 6.27E+01 & \textbf{0.00E+00} & \textbf{0.00E+00} & 7.51E-14 & \textbf{0.00E+00} & 3.75E-05 & \textbf{0.00E+00} & \textbf{0.00E+00} & \textbf{0.00E+00} & \textbf{0.00E+00} & 7.51E-01 & \textbf{0.00E+00} & \textbf{0.00E+00} & \textbf{0.00E+00} & 3.94E+01 & \textbf{0.00E+00} & \textbf{0.00E+00} \\
\hline A29 & \textbf{0.00E+00} & 1.82E-12 & 1.72E+01 & \textbf{0.00E+00} & \textbf{0.00E+00} & 6.79E-14 & \textbf{0.00E+00} & 2.14E-05 & \textbf{0.00E+00} & \textbf{0.00E+00} & \textbf{0.00E+00} & \textbf{0.00E+00} & 1.06E+00 & \textbf{0.00E+00} & \textbf{0.00E+00} & \textbf{0.00E+00} & 3.26E+01 & \textbf{0.00E+00} & \textbf{0.00E+00} \\
\hline A30 & \textbf{0.00E+00} & 1.78E-12 & 1.88E+01 & \textbf{0.00E+00} & \textbf{0.00E+00} & 6.79E-14 & \textbf{0.00E+00} & 2.74E-05 & \textbf{0.00E+00} & \textbf{0.00E+00} & \textbf{0.00E+00} & \textbf{0.00E+00} & 3.41E-01 & \textbf{0.00E+00} & \textbf{0.00E+00} & \textbf{0.00E+00} & 2.87E+01 & \textbf{0.00E+00} & \textbf{0.00E+00} \\
\hline A31 & \textbf{0.00E+00} & 1.42E-14 & 5.35E+01 & \textbf{0.00E+00} & \textbf{0.00E+00} & \textbf{0.00E+00} & \textbf{0.00E+00} & 1.06E-01 & \textbf{0.00E+00} & \textbf{0.00E+00} & \textbf{0.00E+00} & \textbf{0.00E+00} & 8.55E+00 & \textbf{0.00E+00} & \textbf{0.00E+00} & \textbf{0.00E+00} & 5.12E+00 & \textbf{0.00E+00} & \textbf{0.00E+00} \\
\hline A32 & \textbf{0.00E+00} & 1.78E-12 & 1.88E+01 & \textbf{0.00E+00} & \textbf{0.00E+00} & 6.79E-14 & \textbf{0.00E+00} & 2.74E-05 & \textbf{0.00E+00} & \textbf{0.00E+00} & \textbf{0.00E+00} & \textbf{0.00E+00} & 3.41E-01 & \textbf{0.00E+00} & \textbf{0.00E+00} & \textbf{0.00E+00} & 2.87E+01 & \textbf{0.00E+00} & \textbf{0.00E+00} \\
\hline A33 & \textbf{0.00E+00} & 1.42E-14 & 5.35E+01 & \textbf{0.00E+00} & \textbf{0.00E+00} & \textbf{0.00E+00} & \textbf{0.00E+00} & 1.06E-01 & \textbf{0.00E+00} & \textbf{0.00E+00} & \textbf{0.00E+00} & \textbf{0.00E+00} & 8.55E+00 & \textbf{0.00E+00} & \textbf{0.00E+00} & \textbf{0.00E+00} & 5.12E+00 & \textbf{0.00E+00} & \textbf{0.00E+00} \\
\hline A34 & \textbf{0.00E+00} & 1.75E-12 & 3.07E+01 & \textbf{0.00E+00} & \textbf{0.00E+00} & 7.51E-14 & \textbf{0.00E+00} & 1.99E-05 & \textbf{0.00E+00} & \textbf{0.00E+00} & \textbf{0.00E+00} & \textbf{0.00E+00} & 2.06E+00 & \textbf{0.00E+00} & \textbf{0.00E+00} & \textbf{0.00E+00} & 4.48E+01 & \textbf{0.00E+00} & \textbf{0.00E+00} \\
\hline A35 & 8.97E+02 & 1.47E+01 & 4.53E+06 & 3.12E+02 & 8.89E+00 & 6.35E+00 & 2.44E+01 & 6.37E+03 & 2.00E+02 & 9.82E+01 & 1.93E+02 & 3.88E+02 & 9.83E+05 & 2.45E+02 & 3.54E+01 & 2.89E+02 & 6.61E+02 & 9.09E+01 & 5.76E+01 \\
\hline A36 & \textbf{0.00E+00} & 1.75E-12 & 1.88E+01 & \textbf{0.00E+00} & \textbf{0.00E+00} & 7.51E-14 & \textbf{0.00E+00} & 2.16E-05 & \textbf{0.00E+00} & \textbf{0.00E+00} & \textbf{0.00E+00} & \textbf{0.00E+00} & 2.06E+00 & \textbf{0.00E+00} & \textbf{0.00E+00} & \textbf{0.00E+00} & 4.48E+01 & \textbf{0.00E+00} & \textbf{0.00E+00} \\
\hline A37 & \textbf{0.00E+00} & 1.75E-12 & 1.88E+01 & \textbf{0.00E+00} & \textbf{0.00E+00} & 7.51E-14 & \textbf{0.00E+00} & 2.16E-05 & \textbf{0.00E+00} & \textbf{0.00E+00} & \textbf{0.00E+00} & \textbf{0.00E+00} & 2.06E+00 & \textbf{0.00E+00} & \textbf{0.00E+00} & \textbf{0.00E+00} & 4.48E+01 & \textbf{0.00E+00} & \textbf{0.00E+00} \\
\hline A38 & \textbf{0.00E+00} & 1.75E-12 & 1.88E+01 & \textbf{0.00E+00} & \textbf{0.00E+00} & 7.51E-14 & \textbf{0.00E+00} & 2.16E-05 & \textbf{0.00E+00} & \textbf{0.00E+00} & \textbf{0.00E+00} & \textbf{0.00E+00} & 2.06E+00 & \textbf{0.00E+00} & \textbf{0.00E+00} & \textbf{0.00E+00} & 4.48E+01 & \textbf{0.00E+00} & \textbf{0.00E+00} \\
\hline A39 & \textbf{0.00E+00} & 1.75E-12 & 1.88E+01 & \textbf{0.00E+00} & \textbf{0.00E+00} & 7.51E-14 & \textbf{0.00E+00} & 2.16E-05 & \textbf{0.00E+00} & \textbf{0.00E+00} & \textbf{0.00E+00} & \textbf{0.00E+00} & 2.06E+00 & \textbf{0.00E+00} & \textbf{0.00E+00} & \textbf{0.00E+00} & 4.48E+01 & \textbf{0.00E+00} & \textbf{0.00E+00} \\
\hline A40 & \textbf{0.00E+00} & \textbf{0.00E+00} & 1.08E-09 & \textbf{0.00E+00} & \textbf{0.00E+00} & \textbf{0.00E+00} & \textbf{0.00E+00} & \textbf{0.00E+00} & \textbf{0.00E+00} & \textbf{0.00E+00} & \textbf{0.00E+00} & \textbf{0.00E+00} & 2.54E-01 & \textbf{0.00E+00} & \textbf{0.00E+00} & \textbf{0.00E+00} & 4.50E+01 & \textbf{0.00E+00} & \textbf{0.00E+00} \\
\hline A41 & \textbf{0.00E+00} & 1.42E-14 & 2.46E+01 & \textbf{0.00E+00} & \textbf{0.00E+00} & \textbf{0.00E+00} & \textbf{0.00E+00} & 2.09E-01 & \textbf{0.00E+00} & \textbf{0.00E+00} & \textbf{0.00E+00} & \textbf{0.00E+00} & 8.52E+00 & \textbf{0.00E+00} & \textbf{0.00E+00} & \textbf{0.00E+00} & 3.21E+01 & \textbf{0.00E+00} & \textbf{0.00E+00} \\
\hline A42 & \textbf{0.00E+00} & 1.78E-12 & 1.88E+01 & \textbf{0.00E+00} & \textbf{0.00E+00} & 7.51E-14 & \textbf{0.00E+00} & 3.15E-05 & \textbf{0.00E+00} & \textbf{0.00E+00} & \textbf{0.00E+00} & \textbf{0.00E+00} & 5.60E-01 & \textbf{0.00E+00} & \textbf{0.00E+00} & \textbf{0.00E+00} & 1.97E+01 & \textbf{0.00E+00} & \textbf{0.00E+00} \\
\hline
\end{tabular}
}
\end{table*}
\end{landscape}
Further, the box plots for the average and median errors are shown in Figures \ref{nlopt_avg} and \ref{nlopt_med} respectively. The horizontal axis represents the algorithms from the NLopt library referred by the numeric identifiers as given in Table \ref{nlopt_algos}, and the vertical axis represents the mean and median errors in the two cases. It may be noted that all plots indicating mean/median errors use a logarithmic scale. One can infer from the plots that in terms of average error, the choice of algorithms A6, A19, A20, A21, A22 and A35 provides better performance, whereas in case of median error, the algorithms A11, A13, A14, A15, A16, A17, A18 and A40 demonstrate superior performance.
\begin{figure}
\centering
\begin{subfigure}[b]{0.45\textwidth}
\includegraphics[scale=0.25]{avg.pdf}
\caption{Average error}
\label{nlopt_avg}
\end{subfigure}
\begin{subfigure}[b]{0.45\textwidth}
\includegraphics[scale=0.25]{median.pdf}
\caption{Median error}
\label{nlopt_med}
\end{subfigure}
\caption{Box plots for average and median errors obtained on the SOCO benchmark functions.}
\label{soco_error}
\end{figure}
The sum of average and median errors are shown in Figures (\ref{nlopt_avg_sum}) and (\ref{nlopt_med_sum}) respectively. The horizontal axis indicates the algorithm from the NLopt library, while the vertical axis shows the sum of the corresponding average and median errors. The number at the top of each bar in the graph indicates the number of times the global minima were obtained on the SOCO benchmark functions. In terms of average error, algorithms A11 and A40 are able to find the global minima for 13 and 12 functions out of the 19 benchmarks, respectively. In case of median error, algorithms A11, A15, A16, A17 and A18 are able to find global minima for 17 out of the 19 benchmark functions. None of the reference algorithms have achieved zero median error on as many functions.
\begin{figure}
\centering
\begin{subfigure}[b]{0.45\textwidth}
\includegraphics[scale=0.25]{avg_score.pdf}
\caption{Sum of average errors.}
\label{nlopt_avg_sum}
\end{subfigure}
\begin{subfigure}[b]{0.45\textwidth}
\includegraphics[scale=0.25]{median_score.pdf}
\caption{Sum of median errors.}
\label{nlopt_med_sum}
\end{subfigure}
\caption{Sum of average and median errors on SOCO benchmarks for the NLopt algorithms. The number above the bar represents the number of times optima were found.}
\label{fig:soco_error_ranks}
\end{figure}
Figure \ref{nlopt_rank} aims to provide a ranking of the algorithms from the NLopt library on our approach. The ranking is done on the basis of the sum of average and median errors obtained on the benchmark functions, which is indicated on the vertical axis. The bars represent the value of sum of average and median errors for the algorithms on the horizontal axis. The rank of the algorithm is indicated above the bars. From our analysis, the top three algorithms for our approach are A11 (limited memory BFGS), A16 (Truncated Newton method with restart) and A17 (Preconditioned Truncated Newton), all of which belong to the class of gradient-based methods for local search. The worst performing algorithm from our analysis is A21 (MLSL-random) which belongs to the category of gradient-based global search algorithms of the NLopt library.
\begin{figure}
\centering
\includegraphics[scale=0.4]{algo_ranking.pdf}
\caption{Ranking of all algorithms based upon the sum of average and median errors on SOCO benchmarks}
\label{nlopt_rank}
\end{figure}
Finally, we also present a comparison of the performance of the best 3 NLopt algorithms (A11, A16 and A17) on the SOCO benchmark functions vis-\`a-vis reference algorithms available in literature. These include the Differential Evolution algorithm (DE) \cite{storn1997differential} and its variants, the co-variance matrix adaptation evolution strategy with increasing population size (G-CMA-ES) \cite{auger2005restart}, the real-coded CHC algorithm (CHC) \cite{eshelman1993chapter}, Shuffle Or Update Parallel Differential Evolution (SOUPDE) \cite{weber2011shuffle}, $DE-D^40+M^m$ \cite{garcia2011role}, Generalized Opposition-based Differential Evolution (GODE) \cite{wang2011enhanced}, Generalized Adaptive Differential Evolution (GADE) \cite{yang2011scalability}, jDElscop \cite{brest2011self}, Self-adaptive Differential Evolution with Multi-Trajectory Search (SaDE-MMTS) \cite{zhao2011self}, MOS-DE \cite{latorre2011mos}, MA-SSW-Chains \cite{molina2011memetic}, Restart Particle Swarm Optimization with Velocity Modulation (RPSO-VM) \cite{garcia2011restart}, Tuned IPSOLS \cite{de2011incremental} EVO-PROpt \cite{duarte2011path}, EM323 \cite{gardeux2011em323} and VXQR \cite{neumaier2011vxqr} among others.
The box plots of average and median error when comparing with $IACO_\mathbb{R}$-Mtsls1 are shown in Figure \ref{soco_compare_avg_med}. It can be inferred that the error obtained using NLopt algorithms are much lower than the reference algorithms. A ranking of the best 3 NLopt algorithms with the reference algorithms is shown in Figure \ref{nlopt_rank_best}. The best performing algorithm was found to be the Tuned-IPSOLS, whereas CHC performs the worst.
\begin{figure*}[!htbp]
\centering
\begin{subfigure}[b]{0.45\textwidth}
\includegraphics[width=\textwidth]{soco_avg_50}
\caption{Average Error}
\label{fig:avg_50_soco_comp}
\end{subfigure}
\begin{subfigure}[b]{0.45\textwidth}
\includegraphics[width=\textwidth]{soco_median_50}
\caption{Median Error}
\label{fig:med_50_soco_comp}
\end{subfigure}
\caption{Plots comparing average and median error of SOCO functions on reference algorithms and best 3 NLopt algorithms.}
\label{soco_compare_avg_med}
\end{figure*}
\begin{figure}[h]
\centering
\includegraphics[scale=0.3]{soco_algo_ranking_50}
\caption{Ranking of best 3 NLopt and reference algorithms on SOCO benchmarks.}
\label{nlopt_rank_best}
\end{figure}
\subsection{Results on CEC 2014 benchmarks}
We now present the results on the CEC 2014 benchmarks \cite{liang2013problem}, where we have considered the 50-dimensional versions of the functions. The maximum number of function evaluations allowed was set to $10000 \times D$, where $D$ represents the number of dimensions in which the function is considered. The search range is $[-100,100]^D$, and uniform random initialization within the search space has been done. The algorithm was run $51$ times on each function; error values were defined as $f(x)-f(x^*)$, where $x$ is a candidate solution and $x^*$ is the optimal solution. Error values lower than $10^{-8}$ are approximated to 0. A summary of the benchmark functions is presented in Table \ref{tab:cec2014}.
\begin{table}[htbp]
\centering
\caption{CEC 2014 Benchmark Functions \cite{chenproblem}}
\scalebox{0.9}{
\begin{tabular}{|c|c|c|}
\hline
\textbf{Category} & \textbf{S. No. } & \multicolumn{1}{|c|}{\textbf{Function}} \\
\hline
\multirow{3}[0]{*}{Unimodal Functions} & 1 & Rotated High Conditioned Elliptic Function \\
& 2 & Rotated Bent Cigar Function \\
& 3 & Rotated Discus Function \\
\hline
\multirow{13}[0]{*}{Multimodal Functions} & 4 & Shifted and Rotated Rosenbrock’s Function \\
& 5 & Shifted and Rotated Ackley’s Function \\
& 6 & Shifted and Rotated Weierstrass Function \\
& 7 & Shifted and Rotated Griewank’s Function \\
& 8 & Shifted Rastrigin’s Function \\
& 9 & Shifted and Rotated Rastrigin’s Function \\
& 10 & Shifted Schwefel’s Function \\
& 11 & Shifted and Rotated Schwefel’s Function \\
& 12 & Shifted and Rotated Katsuura Function \\
& 13 & Shifted and Rotated HappyCat Function \\
& 14 & Shifted and Rotated HGBat Function \\
& 15 & Shifted and Rotated Expanded Griewank’s plus Rosenbrock’s Function \\
& 16 & Shifted and Rotated Expanded Scaffer’s F6 Function \\
\hline
\multirow{6}[0]{*}{Hybrid Function - 1} & 17 & Hybrid Function 1 (N=3) \\
& 18 & Hybrid Function 2 (N=3) \\
& 19 & Hybrid Function 3 (N=4) \\
& 20 & Hybrid Function 4 (N=4) \\
& 21 & Hybrid Function 5 (N=5) \\
& 22 & Hybrid Function 6 (N=5) \\
\hline
\multirow{8}[0]{*}{Composition Functions} & 23 & Composition Function 1 (N=5) \\
& 24 & Composition Function 2 (N=3) \\
& 25 & Composition Function 3 (N=3) \\
& 26 & Composition Function 4 (N=5) \\
& 27 & Composition Function 5 (N=5) \\
& 28 & Composition Function 6 (N=5) \\
& 29 & Composition Function 7 (N=3) \\
& 30 & Composition Function 8 (N=3) \\
\hline
\end{tabular}%
}
\label{tab:cec2014}%
\end{table}%
The average and median errors of the NLopt algorithms on the CEC benchmarks are shown in Figure \ref{fig:cec2014_errors}. The box plots showing average error are shown in Figure \ref{avg_cec2014}, while median error is shown in Figure \ref{median_cec2014}. It may be noted here that the plots are for the 50-dimensional versions of the functions. Further, the sum of average and median errors are shown in Figures \ref{avg_cec2014_rank} and \ref{median_cec2014_rank} respectively.
\begin{figure*}[!htbp]
\centering
\begin{subfigure}[b]{0.45\textwidth}
\includegraphics[width=\textwidth]{nlopt_avg_50}
\caption{Average Error}
\label{avg_cec2014}
\end{subfigure}
\begin{subfigure}[b]{0.45\textwidth}
\includegraphics[width=\textwidth]{nlopt_median_50}
\caption{Median Error}
\label{median_cec2014}
\end{subfigure}
\begin{subfigure}[b]{0.45\textwidth}
\includegraphics[width=\textwidth]{nlopt_avg_score_50}
\caption{Average Error Ranking}
\label{avg_cec2014_rank}
\end{subfigure}
\begin{subfigure}[b]{0.45\textwidth}
\includegraphics[width=\textwidth]{nlopt_median_score_50}
\caption{Median Error Ranking}
\label{median_cec2014_rank}
\end{subfigure}
\caption{Plots showing average and median error of NLopt algorithms with their ranking on CEC 2014 benchmarks.}
\label{fig:cec2014_errors}
\end{figure*}
The ranking of the algorithms on the CEC 2014 benchmarks is shown in Figure \ref{fig:algo_ranking_cec2014}. One can observe that A11 (limited memory BFGS) performs the best while A35 (ISRES evolutionary constrained optimization) performs the worst.
\begin{figure*}[!htbp]
\centering
\includegraphics[scale=0.25]{nlopt_algo_ranking_50}
\caption{Ranking of the NLopt algorithms on the CEC 2014 benchmarks.}
\label{fig:algo_ranking_cec2014}
\end{figure*}
We also provide a comparison of the top 3 NLopt algorithms (A11, A13 and A14) on the CEC 2014 benchmarks with other reference algorithms. These include the United Multi-Operator Evolutionary Algorithms (UMOEA) \cite{elsayed2014testing}, Success-History based Adaptive Differential Evolution using linear population size reduction (L-SHADE) \cite{tanabe2014improving}, Differential Evolution with Replacement Strategy (RSDE) \cite{xu2014differential}, Memetic Differential Evolution Based on Fitness Euclidean-Distance Ratio (FERDE) \cite{qu2014memetic}, Partial Opposition-Based Adaptive Differential Evolution (POBL-ADE) \cite{hu2014partial}, Mean-Variance Mapping Optimization (MVMO) \cite{erlich2014evaluating}, rmalschcma \cite{molina2014influence}, Opt Bees \cite{maia2014real}, Fireworks Algorithm with Differential Mutation (FWA-DE) \cite{yu2014fireworks}, Non-uniform Real-coded Genetic Algorithm (NRGA) \cite{yashesh2014non}, b3e3pbest \cite{bujok2014differential} and DE\_b6e6rl \cite{polakova2014controlled}.
The box plots for average and median errors are shown in Figure \ref{fig:cec2014_errors_comp}, specifically average error in Figure (\ref{avg_cec2014_comp}) and median error in Figure (\ref{median_cec2014_comp}) respectively. The range of average errors of the reference algorithms are relatively lower than the top 3 NLopt algorithms except on UMOEA, and also median error except MVMO and rmalschcma.
\begin{figure*}[!htbp]
\centering
\begin{subfigure}[b]{0.45\textwidth}
\includegraphics[width=\textwidth]{cec_avg_50}
\caption{Average Error}
\label{avg_cec2014_comp}
\end{subfigure}
\begin{subfigure}[b]{0.45\textwidth}
\includegraphics[width=\textwidth]{cec_median_50}
\caption{Median Error}
\label{median_cec2014_comp}
\end{subfigure}
\caption{Plots showing average and median error of best 3 NLopt and reference algorithms on CEC 2014 benchmarks.}
\label{fig:cec2014_errors_comp}
\end{figure*}
A relative ranking of these algorithms is shown in Figure \ref{fig:algo_ranking_cec2014_comp}. One can observe that the best performing algorithm is UMOEA, while the worst performing algorithm is FWA-DE.
\begin{figure*}[!htbp]
\centering
\includegraphics[scale=0.3]{cec_algo_ranking_50}
\caption{Ranking of the top 3 NLopt and reference algorithms on the CEC 2014 benchmarks.}
\label{fig:algo_ranking_cec2014_comp}
\end{figure*}
\subsection{Comparison of Hybrid approach with standalone algorithms used for Local Search}
We also provide a comparison of the standalone performance of the algorithms in Table \ref{nlopt_pval_tab}. The comparison is provided in terms of Wilcoxon Signed-Rank test \cite{wilcoxon1945individual}, which is a measure of the extent of statistical deviations in the results obtained using a particular approach. A p-value less than $0.05$ indicates that the results of our approach have a significant statistical difference with the results obtained using the algorithms being compared, whereas p-values greater than $0.05$ indicate non-significant statistical difference. The columns $(Wp)$ and $(Wn)$ denote the sum of signed ranks. The column $(N)$ indicates the number of instances for which there is a difference in the result between two algorithms.
It can be observed from Table \ref{nlopt_pval_tab} that on the SOCO benchmark functions, the NLopt algorithms give significant statistical difference on \textbf{all but one} (A34) in terms of average error, and on \textbf{all but two} (A6 and A7) in terms of median error. On the CEC 2014 benchmarks, there is significant statistical difference on all but \textbf{8} and \textbf{3} out of \textbf{42} algorithms in terms of average and median error respectively.
\begin{table*}
\centering
\caption{Wilcoxon signed rank test between hybrid and standalone approach}
\label{nlopt_pval_tab}
\scalebox{0.68}{
\begin{tabular} {|c|c|c|c|c|c|c|c|c|c|c|c|c|c|c|c|c|}
\hline & \multicolumn{8}{|c|} {\textbf{SOCO}} & \multicolumn{8}{|c|} {\textbf{CEC2014}} \\
\hline & \multicolumn{4}{|c|} {\textit{Average}} & \multicolumn{4}{|c|} {\textit{Median}} & \multicolumn{4}{|c|} {\textit{Average}} & \multicolumn{4}{|c|} {\textit{Median}} \\
\hline & \textit{Wp} & \textit{Wn} & \textit{n} & \textit{p} & \textit{Wp} & \textit{Wn} & \textit{n} & \textit{p} & \textit{Wp} & \textit{Wn} & \textit{n} & \textit{p} & \textit{Wp} & \textit{Wn} & \textit{n} & \textit{p} \\
\hline A0 & 7 & 183 & 19 & \textbf{3.98E-04} & 0 & 78 & 12 & \textbf{4.88E-04} & 26 & 250 & 23 & \textbf{6.58E-04} & 0 & 190 & 19 & \textbf{1.32E-04} \\
\hline A1 & 4 & 186 & 19 & \textbf{2.50E-04} & 0 & 91 & 13 & \textbf{2.44E-04} & 32 & 268 & 24 & \textbf{7.48E-04} & 0 & 153 & 17 & \textbf{2.93E-04} \\
\hline A2 & 0 & 190 & 19 & \textbf{1.32E-04} & 0 & 190 & 19 & \textbf{1.32E-04} & 26 & 274 & 24 & \textbf{3.96E-04} & 0 & 231 & 21 & \textbf{5.96E-05} \\
\hline A3 & 3 & 187 & 19 & \textbf{2.14E-04} & 0 & 91 & 13 & \textbf{2.44E-04} & 32 & 268 & 24 & \textbf{7.48E-04} & 0 & 153 & 17 & \textbf{2.93E-04} \\
\hline A4 & 3 & 187 & 19 & \textbf{2.14E-04} & 0 & 91 & 13 & \textbf{2.44E-04} & 32 & 268 & 24 & \textbf{7.48E-04} & 0 & 153 & 17 & \textbf{2.93E-04} \\
\hline A5 & 3 & 187 & 19 & \textbf{2.14E-04} & 0 & 91 & 13 & \textbf{2.44E-04} & 32 & 268 & 24 & \textbf{7.48E-04} & 0 & 153 & 17 & \textbf{2.93E-04} \\
\hline A6 & 0 & 190 & 19 & \textbf{1.32E-04} & 0 & 0 & 0 & 1.00E+00 & 33 & 243 & 23 & \textbf{1.41E-03} & 0 & 105 & 14 & \textbf{1.22E-04} \\
\hline A7 & 2 & 134 & 16 & \textbf{6.43E-04} & 0 & 0 & 0 & 1.00E+00 & 33 & 243 & 23 & \textbf{1.41E-03} & 0 & 91 & 13 & \textbf{2.44E-04} \\
\hline A8 & 0 & 190 & 19 & \textbf{1.32E-04} & 0 & 190 & 19 & \textbf{1.32E-04} & 61 & 404 & 30 & \textbf{4.20E-04} & 62 & 403 & 30 & \textbf{4.53E-04} \\
\hline A9 & 0 & 190 & 19 & \textbf{1.32E-04} & 0 & 190 & 19 & \textbf{1.32E-04} & 61 & 404 & 30 & \textbf{4.20E-04} & 62 & 403 & 30 & \textbf{4.53E-04} \\
\hline A11 & 1 & 135 & 16 & \textbf{5.31E-04} & 0 & 91 & 13 & \textbf{2.44E-04} & 109 & 216 & 25 & 1.50E-01 & 68 & 208 & 23 & \textbf{3.33E-02} \\
\hline A12 & 1 & 152 & 17 & \textbf{3.52E-04} & 1 & 135 & 16 & \textbf{5.31E-04} & 71 & 307 & 27 & \textbf{4.58E-03} & 72 & 306 & 27 & \textbf{4.94E-03} \\
\hline A13 & 0 & 153 & 17 & \textbf{2.93E-04} & 2 & 134 & 16 & \textbf{6.43E-04} & 109 & 242 & 26 & 9.12E-02 & 29 & 247 & 23 & \textbf{9.16E-04} \\
\hline A14 & 0 & 153 & 17 & \textbf{2.93E-04} & 2 & 134 & 16 & \textbf{6.43E-04} & 109 & 242 & 26 & 9.12E-02 & 29 & 247 & 23 & \textbf{9.16E-04} \\
\hline A15 & 0 & 136 & 16 & \textbf{4.38E-04} & 0 & 120 & 15 & \textbf{6.10E-05} & 93 & 313 & 28 & \textbf{1.23E-02} & 32 & 374 & 28 & \textbf{9.86E-05} \\
\hline A16 & 0 & 136 & 16 & \textbf{4.38E-04} & 0 & 120 & 15 & \textbf{6.10E-05} & 91 & 315 & 28 & \textbf{1.08E-02} & 34 & 344 & 27 & \textbf{1.96E-04} \\
\hline A17 & 3 & 150 & 17 & \textbf{5.03E-04} & 0 & 120 & 15 & \textbf{6.10E-05} & 93 & 285 & 27 & \textbf{2.11E-02} & 36 & 315 & 26 & \textbf{3.96E-04} \\
\hline A18 & 1 & 152 & 17 & \textbf{3.52E-04} & 0 & 105 & 14 & \textbf{1.22E-04} & 92 & 286 & 27 & \textbf{1.98E-02} & 29 & 271 & 24 & \textbf{5.46E-04} \\
\hline A19 & 15 & 175 & 19 & \textbf{1.28E-03} & 16 & 174 & 19 & \textbf{1.48E-03} & 147 & 288 & 29 & 1.27E-01 & 98 & 253 & 26 & 5.05E-02 \\
\hline A20 & 17 & 173 & 19 & \textbf{1.70E-03} & 7 & 183 & 19 & \textbf{3.98E-04} & 71 & 254 & 25 & \textbf{1.38E-02} & 62 & 238 & 24 & \textbf{1.19E-02} \\
\hline A21 & 0 & 153 & 17 & \textbf{2.93E-04} & 0 & 153 & 17 & \textbf{2.92E-04} & 53 & 325 & 27 & \textbf{1.08E-03} & 31 & 320 & 26 & \textbf{2.42E-04} \\
\hline A22 & 18 & 172 & 19 & \textbf{1.94E-03} & 13 & 177 & 19 & \textbf{9.67E-04} & 95 & 230 & 25 & 6.73E-02 & 78 & 247 & 25 & \textbf{2.30E-02} \\
\hline A23 & 0 & 153 & 17 & \textbf{2.93E-04} & 2 & 151 & 17 & \textbf{4.21E-04} & 66 & 340 & 28 & \textbf{1.81E-03} & 57 & 294 & 26 & \textbf{2.62E-03} \\
\hline A24 & 0 & 153 & 17 & \textbf{2.93E-04} & 0 & 153 & 17 & \textbf{2.93E-04} & 52 & 354 & 28 & \textbf{5.85E-04} & 44 & 307 & 26 & \textbf{8.38E-04} \\
\hline A25 & 0 & 190 & 19 & \textbf{1.32E-04} & 0 & 190 & 19 & \textbf{1.32E-04} & 99 & 336 & 29 & \textbf{1.04E-02} & 99 & 336 & 29 & \textbf{1.04E-02} \\
\hline A26 & 0 & 190 & 19 & \textbf{1.32E-04} & 0 & 190 & 19 & \textbf{1.32E-04} & 240 & 195 & 29 & 6.27E-01 & 229 & 236 & 30 & 9.34E-01 \\
\hline A28 & 0 & 190 & 19 & \textbf{1.32E-04} & 0 & 190 & 19 & \textbf{1.32E-04} & 83 & 382 & 30 & \textbf{2.11E-03} & 86 & 379 & 30 & \textbf{2.58E-03} \\
\hline A29 & 0 & 190 & 19 & \textbf{1.32E-04} & 0 & 190 & 19 & \textbf{1.32E-04} & 56 & 409 & 30 & \textbf{2.83E-04} & 88 & 377 & 30 & \textbf{2.96E-03} \\
\hline A30 & 0 & 190 & 19 & \textbf{1.32E-04} & 0 & 190 & 19 & \textbf{1.32E-04} & 99 & 336 & 29 & \textbf{1.04E-02} & 99 & 336 & 29 & \textbf{1.04E-02} \\
\hline A31 & 0 & 153 & 17 & \textbf{2.93E-04} & 0 & 153 & 17 & \textbf{2.93E-04} & 38 & 340 & 27 & \textbf{2.86E-04} & 33 & 318 & 26 & \textbf{2.96E-04} \\
\hline A32 & 0 & 190 & 19 & \textbf{1.32E-04} & 0 & 190 & 19 & \textbf{1.32E-04} & 99 & 336 & 29 & \textbf{1.04E-02} & 99 & 336 & 29 & \textbf{1.04E-02} \\
\hline A33 & 0 & 153 & 17 & \textbf{2.93E-04} & 0 & 153 & 17 & \textbf{2.93E-04} & 38 & 340 & 27 & \textbf{2.86E-04} & 33 & 318 & 26 & \textbf{2.96E-04} \\
\hline A34 & 38 & 115 & 17 & 6.84E-02 & 0 & 153 & 17 & \textbf{2.93E-04} & 253 & 182 & 29 & 4.43E-01 & 229 & 149 & 27 & 3.37E-01 \\
\hline A35 & 0 & 190 & 19 & \textbf{1.32E-04} & 9 & 181 & 19 & \textbf{5.39E-04} & 2 & 433 & 29 & \textbf{3.17E-06} & 0 & 435 & 29 & \textbf{2.56E-06} \\
\hline A36 & 0 & 190 & 19 & \textbf{1.32E-04} & 0 & 190 & 19 & \textbf{1.32E-04} & 61 & 404 & 30 & \textbf{4.20E-04} & 62 & 403 & 30 & \textbf{4.53E-04} \\
\hline A37 & 0 & 190 & 19 & \textbf{1.32E-04} & 0 & 190 & 19 & \textbf{1.32E-04} & 61 & 404 & 30 & \textbf{4.20E-04} & 62 & 403 & 30 & \textbf{4.53E-04} \\
\hline A38 & 0 & 190 & 19 & \textbf{1.32E-04} & 0 & 190 & 19 & \textbf{1.32E-04} & 61 & 404 & 30 & \textbf{4.20E-04} & 62 & 403 & 30 & \textbf{4.53E-04} \\
\hline A39 & 0 & 190 & 19 & \textbf{1.32E-04} & 0 & 190 & 19 & \textbf{1.32E-04} & 61 & 404 & 30 & \textbf{4.20E-04} & 62 & 403 & 30 & \textbf{4.53E-04} \\
\hline A40 & 0 & 136 & 16 & \textbf{4.38E-04} & 0 & 136 & 16 & \textbf{4.38E-04} & 99 & 226 & 25 & 8.75E-02 & 62 & 214 & 23 & \textbf{2.08E-02} \\
\hline A41 & 0 & 153 & 17 & \textbf{2.93E-04} & 0 & 153 & 17 & \textbf{2.93E-04} & 71 & 335 & 28 & \textbf{2.65E-03} & 38 & 313 & 26 & \textbf{4.79E-04} \\
\hline A42 & 0 & 190 & 19 & \textbf{1.32E-04} & 0 & 190 & 19 & \textbf{1.32E-04} & 37 & 428 & 30 & \textbf{5.79E-05} & 37 & 428 & 30 & \textbf{5.79E-05} \\
\hline
\end{tabular}
}
\end{table*}
\section{Conclusions }
\label{sec:conclusion}
This paper presented an exhaustive analysis of using optimization algorithms from the NLopt library in combination with the Mtsls1 algorithm within the $IACO_\mathbb{R}$-Mtsls1 framework for continuous global optimization. The results on SOCO and CEC 2014 benchmark functions present a ready reference on the performance of these approaches and would be of help to a researcher in deciding on a choice among these algorithms. The nature of functions on which these algorithms perform better can also be inferred from the results. A relative ranking of these approaches has also been provided based on the total error obtained in using them, which would provide a measure of the versatility of the algorithms.
The results of our analysis have been summarized in Table \ref{final_res}. On both the benchmark function sets, the hybrid $IACO_\mathbb{R}$-Mtsls1 with gradient-based local search performs better (A11, A16, A17 for SOCO and A11, A13, A14 for CEC). The best results are obtained using hybridization with BFGS on both benchmarks. On SOCO benchmarks, the hybrid approach outperforms the original $IACO_\mathbb{R}$-Mtsls1, as it has been able to achieve zero median error on \textbf{17} out of \textbf{19} functions, which has not been achieved by any other algorithm considered. We believe that the analysis presented in this paper would be of use to the research community at large.
\begin{table}[H]
\centering
\caption{Summary of algorithm rankings.}
\scalebox{0.7}{
\begin{tabular}{|c|c||c|c|}
\hline
\textbf{Function} & \textbf{Ranking} & \textbf{NLopt} & \textbf{State-of-Art} \\
\hline
\textbf{SOCO} & Best & A11 & Tuned IPSOLS \\
\textbf{} & Worst & A21 & CHC \\
\hline
\textbf{CEC 2014} & Best & A11 & UMOEA \\
\textbf{} & Worst & A35 & FWA-DE \\
\hline
\multicolumn{4}{|l|}{\textit{A11: Limited Memory BFGS }} \\
\multicolumn{4}{|l|}{\textit{A21: Multi-Level Single Linkage, random}} \\
\multicolumn{4}{|l|}{\textit{A35: ISRES Evolutionary Constrained Optimization}} \\
\multicolumn{4}{|l|}{\textit{Tuned IPSOLS: Incremental Particle Swarm for Large Scale Optimization}} \\
\multicolumn{4}{|l|}{\textit{UMOEA: united Multi Operator Evolutionary Algorithms}} \\
\multicolumn{4}{|l|}{\textit{FWA-DE: FireWorks Algorithm with Differential Evolution}} \\
\hline
\end{tabular}%
}
\label{final_res}%
\end{table}%
\bibliographystyle{plain}
|
\section{Introduction}
\label{intro}
The X-ray, GeV gamma-ray, and TeV gamma-ray skies have been deeply investigated by various telescopes. However, the MeV gamma-ray sky has not been extensively studied yet after the Imaging Compton Telescope (COMPTEL) onboard the {\it Compton Gamma-Ray Observatory} ({\it CGRO}) satellite scanned the sky in 1990s \citep{sch00_comptel}. This means that huge discovery space is buried in this energy range.
Measurement of MeV gamma rays is hampered by huge background events. Under this difficulty, the COMPTEL achieved the sensitivity of $\sim10^{-10}\ {\rm erg\ cm^{-2}\ s^{-1}}$ in the MeV band and discovered $\sim$63 gamma-ray sources (32 of them were gamma-ray bursts) at 0.75--30 MeV \citep{sch00_comptel}. The COMPTEL has found 10 extragalactic active galactic nuclei (AGNs) and 5 unidentified high latitude objects \citep{sch00_comptel}.
To open a new observational window in the MeV sky, various missions have been proposed. {\it ASTRO-H} \citep{tak14}, which will be launched in 2015, will have a sub-MeV instrument, the soft gamma-ray detector \citep[SGD,][]{taj10,wat12} covering at 40--600~keV. The SGD will provide precise spectra with higher sensitivity than previous instruments by introducing the concept of narrow-Field-of-View Compton telescope which can dramatically reduce background events \citep{tak01,tak04}. Other instruments are also proposed such as CAST \citep{nak12}, DUAL \citep{von12}, GRIPS \citep{gre12}, SMILE \citep{tak11}, COSI \citep{kie14}, and ASTROGAM\footnote{\url{http://astrogam.iaps.inaf.it/index.html}}. Moreover, the instrumental performance were tested with ballon experiments \citep{tak11,ban11}. These future missions will unveil the MeV sky in near future. In addition, a new image reconstruction algorithm that has improved sensitivity to multiple point-like sources has been proposed toward those future all-sky missions \citep{ike14}. The new era of the MeV gamma-ray astronomy is approaching.
A fundamental question to those missions is how many sources one can expect with their designed instruments. In the extragalactic sky, the dominant population in the MeV sky is still uncertain because the origin of the cosmic MeV gamma-ray background from 0.3~MeV to 100~MeV is not well established \citep[see][for a recent review]{ino14_cgb}. Seyferts which make up the cosmic X-ray background (CXB) cannot explain this MeV background because of the assumed cutoff at a few hundred keV \citep[e.g.][]{ued03,gil07}. Above 100~MeV, the total gamma-ray background is known to be predominantly composed of blazars \citep[e.g.][]{ino09,aje15}, star-forming galaxies \citep[e.g.][]{ack12_stb}, and radio galaxies \citep[e.g.][]{ino11}. Although the MeV background may comprises these GeV gamma-ray populations as well, the lower energy part of the MeV background spectrum is smoothly connected to the CXB spectrum and shows softer \citep[photon index of $\Gamma\sim$ 2.8][]{fuk75,wat97,wei00} than the GeV component \citep[photon index of $\Gamma\sim$ 2.4][]{ack15_cgb}, indicating their different origins.
Several candidates have been considered as the origin of the MeV background. One was the nuclear-decay gamma rays from Type Ia supernovae \citep{cla75,zdz96,wat99}. However, the measurements of the cosmic Type Ia supernovae rates show that the rates are not enough to explain the observed flux \citep[e.g.][]{ahn05_sn,str05,hor10,rui15}.
Seyferts may naturally explain the MeV background up to a few tens of MeV and the smooth connection to the CXB \citep[][hereinafter I08]{sch78,fie93,ste99,ino08}. Non-thermal electrons in coronae can generate the power-law spectrum in the MeV band after the thermal cut-off via the Compton scattering of disk photons (I08).
Flat-spectrum radio quasars (FSRQs) whose peak in the spectrum locates at MeV energies \citep{blo95,sam06} are also expected to significantly dominate the MeV background \citep[][hereinafter A09]{aje09}.
Radio galaxies have been also discussed as one of the origins of the MeV background \citep{str76}. Recent studies have revealed that both of lobe \citep{mas11} and core \citep{ino11} emissions from radio galaxies could contribute only $\sim10$\% of the MeV gamma-ray background flux.
Dark matters has also been discussed as the origin of the MeV background \citep[e.g.][]{ahn05_dm1,ahn05_dm2}. Those MeV mass dark matter particle candidates are less natural than GeV-TeV dark matter candidates. Therefore, Seyferts and FSRQs can be regarded as potential astrophysical origins of the MeV background.
The purpose of this paper is to investigate prospects of extragalactic observations by future MeV instruments, especially of statistical aspects of AGNs. We focus only on Seyferts and FSRQs in this paper. For this purpose, we use recent luminosity functions (LFs) and spectral models of Seyferts and FSRQs in literature. Since a quantitative estimate based only on the COMPTEL results is not easy due to the small number of detected sources, we adopt recent X-ray or gamma-ray luminosity functions and spectral models of Seyferts and FSRQs.
This paper is organized as follows. MeV gamma-ray emission models of Seyferts and FSRQs are described In Section \ref{sec:sed}. LFs of Seyferts and FSRQs are presented in Section \ref{sec:lf}. In Section \ref{sec:count}, results of the expected source counts and redshift distributions are presented. Discussions and conclusions are given in Section \ref{sec:dis_con}. Throughout this paper, we adopt the standard cosmological parameters of $(h, \Omega_M , \Omega_\Lambda) = (0.7, 0.3, 0.7)$.
\section{Spectra of Active Galactic Nuclei in the MeV gamma-ray band}
\label{sec:sed}
\subsection{Seyferts}
The X-ray spectra of Seyfert are phenomenologically explained by a combination of following components: a primary power-law continuum with a cutoff at $\sim$0.3~MeV in the form of $E^{-\Gamma}\exp(-E/E_c)$, absorption by surrounding gas, emission lines, a reflection component, and a soft excess of emission at $\lesssim2$~keV \citep[e.g.][]{dad08}. Relative fractions of these components vary with sources. Physically Comptonization of disk photons in a corona above the accretion disk generate the primary power-law continuum \citep[see e.g.][]{kat76,poz77,sun80}. The temperature of the corona roughly determines the position of the spectral cutoff and the photon index of the intrinsic continuum together with the optical depth \cite[see e.g.][]{zdz94}. The Compton reprocessed emission and bound-free absorption of the primary continuum by surrounding cold matter generate the reflection component \citep{lig88,mag95}.
\begin{figure}
\centering
\includegraphics[bb=50 30 410 282,width=8cm]{sed_template_v1.eps}
\caption{Spectral templates for Seyferts and FSRQs. Solid, dashed, dotted, and dot-dashed curve corresponds \citet[][U03]{ued03} Seyfert spectral model with a thermal cutoff at 0.3~MeV with $N_H=10^{20}\ {\rm cm^2}$, U03 Seyfert spectral model with a thermal cutoff at 0.3~MeV with $N_H=10^{25}\ {\rm cm^2}$, \citet[][I08]{ino08} Seyfert spectral model with a non-thermal tail in the MeV band with $N_H=10^{20}\ {\rm cm^2}$, and \citet[][A09]{aje09} FSRQ spectral model having a peak at 1~MeV. The shape is assumed to be independent of luminosities for these models. \label{fig:sed_1} }
\end{figure}
Although coronae with temperature of $\sim0.1$~MeV do not produce significant MeV gamma-ray emission, some spectral models predict a power-law tail after the thermal cutoffs (e.g. I08). If the corona is composed of thermal and non-thermal populations, a MeV tail will appear after the cut-off. Such non-thermal electrons may exist if magnetic reconnection heats the corona \citep{liu02} as in the Solar flares \citep[e.g.][]{shi95} and Earth's magnetotail \citep{lin05}. The MeV background can be explained by the Seyferts explaining the CXB which have non-thermal electrons in the coronae having $\sim4$\% of the total electron energy (I08).
Observationally, the cutoff energy of Seyferts are constrained at $\gtrsim0.2$~MeV \citep{ric11}. The Oriented Scintillation Spectroscopy Experiment ({\it OSSE}) clearly detected emission up to 0.5~MeV in the spectrum of the brightest Seyfert NGC~4151 \citep{joh97}. This measurement constrained non-thermal fraction to be $\lesssim$15\% \citep{joh97}. Interestingly, future radio observations are able to probe those non-thermal tail through synchrotron emission \citep{lao08,ino14}.
In this paper, as shown in Figure~\ref{fig:sed_1}, two primary spectral models are applied for Seyferts as in \citet{ino13}. One is thermal spectral model having a power-law continuum with a cutoff \citep[see e.g.][hereinafter U03]{ued03}. We adopt $\Gamma=1.9$ and $E_c=0.3$~MeV. The other is thermal plus non-thermal spectral model (see I08 for details) which could explain the MeV background spectrum. We adopt the same parameters as in I08 with the non-thermal photon index of 2.8, but setting the cutoff at 0.3~MeV. For the Compton reflection component, we use a Compton reflection model \citep{mag95}, assuming a solid angle of $2\pi$, an inclination angle of $\cos i = 0.5$, and solar abundance for all elements. To calculate absorbed spectra, we use an absorption model called "wabs" developed for the XSPEC package.
\subsection{FSRQs}
\begin{figure}
\centering
\includegraphics[bb=50 30 410 282,width=8cm]{sed_template_v2.eps}
\caption{Spectral templates for FSRQs, but for A12 models. Solid, dashed, dotted, and dot-dashed curve corresponds to the 0.1-100~GeV luminosity of $L_{\gamma}=10^{48.9}$, $10^{48.2}$, $10^{47.8}$, and $10^{46.8}\ {\rm erg\ s^{-1}}$, respectively, based on \citet{aje12}. \label{fig:sed_2} }
\end{figure}
Multi-wavelength observations allowed us to study overall spectra of blazars from radio to gamma-ray. The observed blazar spectra consists of two broadband emission components \citep{ulr97,fos98,ghi98,kub98,abd10_blz}. One is synchrotron radiation component and the other is inverse Compton (IC) component in which electrons scatter internal synchrotron radiation \cite[see e.g.][]{jon74} or external radiation \cite[see e.g.][]{der93,sik94}.
Blazars are divided into two categories by their optical spectra. Those are BL Lacertae objects (BL Lacs) and FSRQs. In the case of BL Lacs, the highest observable energy part of the synchrotron emission appears in the X-ray band. X-ray spectra of BL Lacs show soft spectra with a photon index of $\Gamma\sim2-3$. In the case of FSRQs, the lowest observable energy part of the IC emission appears in the X-ray band showing harder spectra. Although BL Lacs are not expected to the dominant contributors of the MeV background, FSRQs may make up the whole MeV background (A09). Hence, in this paper, we do not consider BL Lacs but only FSRQs for blazars.
We consider two spectral energy distribution (SED) models for FSRQs. One is based on the {\it Swift}-BAT FSRQs (A09), while the other is based on the {\it Fermi}-LAT FSRQs \citet[][hereinafter A12]{aje12}. Following A09, the SED is given by a double power-law model:
\begin{equation}
\frac{dN}{dE}\propto\left[\left(\frac{E}{E_b}\right)^{\Gamma_1}+\left(\frac{E}{E_b}\right)^{\Gamma_2}\right]^{-1},
\end{equation}
where we set $E_b=1$ MeV, $\Gamma_1=1.6$, and $\Gamma_2=2.9$ following A09 (see Figure~\ref{fig:sed_1}). The average photon index of the {\it Swift}-BAT FSRQs is $1.6\pm0.3$ at 15--55 keV (A09). Although theoretical models predict a spectral turnover at the MeV band for FSRQs \citep[see e.g.][]{ino96}, $E_b$ and $\Gamma_2$ in the MeV band are observationally unconstrained. Hence, these parameters are set FSRQs to make up the MeV background. Here, the latest catalog of the {\it Fermi}-LAT sources reported detection of 467 FSRQs. The mean of their photon index above 0.1 GeV of those gamma-ray bright FSRQs is typically $\sim2.4$ \citep{ack15_3LAC}, which is harder than that of the MeV background. Combining with the {\it Fermi}-LAT and {\it Swift}-BAT detected FSRQs, the average SED is modelled in Fig. 9 of A12 which are shown in Figure~\ref{fig:sed_2}. We also adopt these SEDs as the other FSRQ template. However, as discussed in the next section, the whole MeV background is not be able to be explained by the A12 FSRQ model.
\section{Luminosity Functions of Active Galactic Nuclei}
\label{sec:lf}
\begin{table*}
\caption{The parameters of the AGN Luminosity Functions \label{XLF-parameters}}
\begin{center}
\begin{tabular}{cccc}
{} & \cite{ued03} & \cite{aje09} & \cite{aje12}\\
{} & Seyfert & FSRQ & FSRQ\\
{} & 2-10~keV & 15-55~keV & 0.1-100~GeV\\
\hline
$A$$^a$ & $50.4\pm3.3$ & $0.533\pm0.104$ & $0.0306\pm0.0023$\\
$\log_{10}$$L^*$$^b$ & $43.94_{-0.26}^{+0.21}$ & $44.0$ & $\log_{10}(0.84\pm0.49)$+48.0\\
$\gamma_1$ & $0.86\pm 0.15$ & - & $0.21\pm0.12$\\
$\gamma_2$ & $2.23\pm 0.13$ & $3.45\pm0.20$ & $1.58\pm0.27$ \\
$z_c^*$ & $1.9$ & - & $1.47\pm0.16$\\
$\log_{10}$$L_a$$^b$ & $44.6$ & - & -\\
$\alpha$ & $0.335\pm 0.07$ & - & $0.21\pm0.03$\\
$p_1$ & $4.23\pm0.39$ & $3.72\pm0.50$ & $7.35\pm1.74$\\
$p_2$ & $-1.5$ & $-0.32\pm0.08$ & $-6.51\pm1.97$\\
$\mu$ & - & - & $2.44\pm0.01$\\
$\sigma$ & - & - & $0.18\pm0.01$\\
\hline
\multicolumn{4}{l}{$^{\rm a}$: In units of $10^{-7} {\rm Mpc}^{-3}$.} \\
\multicolumn{4}{l}{$^{\rm b}$: In units of ${\rm erg/s}$.} \\
\end{tabular}
\end{center}
\end{table*}
To obtain the expected source counts and redshift distributions of AGNs in the MeV gamma-ray band, a LF is required. We briefly review LFs of Seyferts and FSRQs in this section.
\subsection{Seyferts}
The cosmological evolution of Seyferts is investigated by various X-ray surveys \citep[see e.g. U03;][]{has05,gil07,ued14}. Based on these surveys, the observed X-ray luminosity functions (XLFs) are known to be well represented by luminosity-dependent density evolution (LDDE) models \citep[see e.g.][and references therein]{ued14}.
In this study, we follow the U03 LDDE XLF at 2-10~keV for the Seyfert models, since the I08 Seyfert model is based on U03. We note that U03 modelled the evolution of X-ray emitting AGNs. The comoving number density $\rho_X$ in the LDDE is
\begin{equation}
\rho_X(L_X, z,N_{\rm H}) = \rho_X(L_X,0)f(L_X,z)\eta(N_{\rm H};L_X, z),
\end{equation}
where $L_X$ is the X-ray luminosity at 2-10~keV before the absorption, $z$ is the redshift, and $N_{\rm H}$ is the neutral hydrogen column density. $\rho_X(L_X, 0)$ is the AGN XLF at present day characterized by the faint-end slope index $\gamma_1 $, the bright-end slope index $\gamma_2$, and the break luminosity $L^*$, as:
\begin{equation}
\rho_X(L_X,0)=A \left[ \left( \frac{L_X}{L^*}
\right)^{\gamma_1} + \left( \frac{L_X}{L^*} \right)^{\gamma_2} \right]^{-1} \ ,
\end{equation}
where $A$ is the normalization parameter having a dimension of volume$^{-1}$.
The density evolution function $f(L_X,z)$ is given as:
\begin{eqnarray}
f(L_X,z)=\left\{\begin{array}{ll}
(1+z)^{p_1} & z \le z_c(L_X), \\
(1+z_c(L_X))^{p_1}
\left( \frac{1+z}{1+z_c(L_X)} \right)^{p_2} & z > z_c(L_X), \\
\end{array}\right.
\end{eqnarray}
where $z_c$ is the redshift of evolutionary peak, given as
\begin{eqnarray}
z_c(L_X)=\left\{\begin{array}{ll}
z_c^* & L_X \ge L_a, \\
z_c^*(L_X/L_a)^\alpha & L_X < L_a. \\
\end{array}\right.
\end{eqnarray}
The function $\eta(N_{\rm H};L_{\rm X},z)$ is the distribution of absorption column density given in the following form in the XLF \citep{ued03}:
\begin{eqnarray}
\eta(N_{\rm H};L_{\rm X},z)
&=\left\{\begin{array}{ll}
2-\frac{5+2\epsilon}{1+\epsilon}\psi(L_{\rm X},z)& 20.0 \leq \log N_{\rm H} < 20.5, \\
\frac{1}{1+\epsilon}\psi(L_{\rm X},z) & 20.5 \leq \log N_{\rm H} < 23.0, \\
\frac{\epsilon}{1+\epsilon}\psi(L_{\rm X},z) & 23.0 \leq \log N_{\rm H} < 24.0,
\end{array}\right.
\end{eqnarray}
where $\epsilon=$1.7 and $\psi(L_{\rm X},z)={\rm min}\{\psi_{\rm max}, {\rm max}[0.47-0.1(\log L_{\rm X} - 44.0), 0])\},$
for which $\psi_{\rm max}=({1+\epsilon})/({3+\epsilon})$.
We note the integration of $\eta(N_{\rm H};L_{\rm X},z)$ between $20.0\le\log N_{\rm H}\le24.0$ is unity.
The parameters obtained by the fit to the observed data of X-ray AGNs in U03 are shown in Table \ref{XLF-parameters}. We set the minimum and maximum of the X-ray luminosity as $L_{X,\rm min}=10^{41.5}$ and $L_{X,\rm max}=10^{48.0}$ erg s$^{-1}$, respectively, the same as in U03. The redshift range is set to be $0.0 \le z \le 5.0$, otherwise noted. To explain the CXB, the fraction of the Compton thick AGNs between $24.0\le\log N_{\rm H}<25.0$ is set to be the same as that of the population at $23.0\le\log N_{\rm H}<24.0$ following U03.
Recently, \citet{air10} suggested another evolution model for AGNs, the luminosity and density evolution (LADE) model. Although the distribution of absorption column density is not available for the LADE, we can test that model assuming the same absorption column density distribution as in U03. The overall source counts in the MeV band significantly decrease by a factor of $\sim3$ at the flux range of $10^{-11}$--$10^{-12}\ {\rm erg\ cm^{-2}\ s^{-1}}$. However, the latest study of X-ray AGN evolution \citep{ued14} found that the AGN XLF is not well described with the LADE model but with LDDE model. When we adopt the latest LDDE model \citep{ued14}, the Seyfert source counts in the MeV band decrease by a factor of $\sim40$\% at the flux range of $10^{-11}$--$10^{-12}\ {\rm erg\ cm^{-2}\ s^{-1}}$. As the I08 Seyfert model is based on U03, we follow the U03 LDDE model in this paper.
\subsection{FSRQs}
In the GeV gamma-ray band, A12 and \citet{aje14} have recently shown that the gamma-ray luminosity functions (GLFs) of FSRQs and BL Lacs can be described by LDDE models based on the {\it Fermi}--LAT FSRQs and BL Lacs, which was indicated before the launch of {\it Fermi} \citep[e.g.][]{nar06,ino09}. With the A12 GLF model, FSRQs can explain up to $\sim30$\% of the MeV background. In the X-ray band, A09 have investigated the XLF of FSRQs using the 3-year {\it Swift}--BAT AGN survey data including 26 FSRQs. A pure luminosity evolution (PLE) model successfully reproduced the observed distribution of X-ray FSRQs. Assuming this PLE model and the spectral model in \S. 2.2, FSRQs explained the whole MeV background.
\begin{figure*}[t]
\centering
\includegraphics[bb=50 30 410 282,width=12cm]{count_panel.eps}
\caption{Cumulative source counts in the entire sky as a function of MeV gamma-ray flux (in $S$) of Seyferts and FSRQs. The five panels correspond to different photon energy bins as indicated in the panels. The solid, dashed, dotted, and dot-dashed curves are predictions based on \citet[][U03]{ued03} for Seyferts, \citet[][I08]{ino08} for Seyferts, \citet[][A09]{aje09} for FSRQs, and \citet[][A12]{aje12} for FSRQs. The horizontal thin double-dot-dashed line corresponds to one source detection level. Due to the thermal cutoff at 0.3~MeV in the spectral model of U03, detection of Seyferts above 3~MeV is not expected in the given flux ranges.
\label{fig:count} }
\end{figure*}
These two recent FSRQ XLF and GLF models of A09 and A12 are adopted in this paper. The FSRQ XLF at 15--55~keV by A09 is given in the form of the PLE. The comoving number densities $\rho_X$ in the PLE is given as:
\begin{equation}
\rho_X(L_X, z) = \rho_X(L_X/e(z),0),
\end{equation}
where $L_X$ is the X-ray luminosity at 15-55~keV here. The local XLF is described as:
\begin{equation}
\rho_X(L_X,0)=A \left( \frac{L_X}{L^*} \right)^{-\gamma_2}.
\end{equation}
The evolution factor $e(z)$ is independent of luminosity given by:
\begin{equation}
e(z)=(1+z)^{p_1 +p_2z}
\end{equation}
The values of the parameters for the A09 XLF are shown in Table \ref{XLF-parameters}. The minimum luminosity is also assumed to evolve with redshift as
\begin{equation}
L_{X, \mathrm{min}}(z)=L_{X, \mathrm{min,0}}\times e(z),
\end{equation}
where $L_{X, \mathrm{min,0}}$ is the minimum luminosity at $z=0$. We set $L_\mathrm{X,min,0}=3\times10^{44} \ {\rm erg \ s^{-1}}$, $L_{X,\rm max}$=10$^{50.0}$ erg s$^{-1}$, $z_{\rm min}=0.0$, and $z_{\rm max}=6.0$ (A09), otherwise noted.
On the contrary, the A12 FSRQ GLF is given in the form of LDDE model. The comoving number densities $\rho_\gamma$ in the LDDE is given as:
\begin{equation}
\rho_\gamma(L_\gamma, z, \Gamma) = \rho_\gamma(L_\gamma,0)f(L_\gamma,z)\theta(\Gamma),
\end{equation}
where $L_\gamma$ is the gamma-ray luminosity at 0.1--100~GeV and $\Gamma$ is the intrinsic photon index between 0.1--100~GeV. $\rho_\gamma(L_\gamma, 0)$ is the FSRQ GLF at present. This is characterized as:
\begin{equation}
\rho_\gamma(L_\gamma,0)=A \left[ \left( \frac{L_\gamma}{L^*}
\right)^{\gamma_1} + \left( \frac{L_\gamma}{L^*} \right)^{\gamma_2} \right]^{-1}.
\end{equation}
The density evolution function $f(L_\gamma,z)$ is given by:
\begin{equation}
f(L_\gamma,z)= \left[ \left( \frac{1+z}{1+z_c(L_\gamma)} \right)^{p_1} + \left( \frac{1+z}{1+z_c(L_\gamma)} \right)^{p_2} \right]^{-1}
\end{equation}
where $z_c$ is given as $z_c(L_\gamma)=z_c^* \cdot (L_\gamma/10^{48})^\alpha$
The function $\theta(\Gamma)$ describes the distribution of photon index, which is given by a Gaussian form in the GLF (A12):
\begin{equation}
\theta(\gamma)=\exp\left[-\frac{(\Gamma - \mu)^2}{2\sigma^2}\right],
\end{equation}
where $\mu$ and $\sigma$ are the mean and the dispersion of the Gaussian distribution. However, we adopt the average SED of FSRQs (see Fig. \ref{fig:sed_2}) not power-law SEDs. Therefore, we include the integrated $\theta(\Gamma)$ over $\Gamma$ between 1.8 and 3.0 which is independent of luminosity and redshift.
The parameters for the A12 GLF are shown in Table \ref{XLF-parameters}. The limits of gamma-ray luminosities, redshifts, and photon indices are set as $L_{\gamma,\rm min}$=10$^{44}$ erg s$^{-1}$, $L_{\gamma,\rm min}$=10$^{52}$ erg s$^{-1}$, $z_{\rm min}$=0.01, and $z_{\rm max}$=6, the same as in A12, otherwise noted.
\section{Source Counts and Redshift Distribution}
\label{sec:count}
\begin{figure*}[t]
\centering
\includegraphics[bb=50 30 410 282,width=12cm]{z_dist_panel_Flim_1e-11.eps}
\caption{Expected cumulative redshift distribution of AGNs in the entire sky assuming the limiting sensitivity of $10^{-11}\ {\rm [erg\ cm^{-2}\ s^{-1}]}$. The five panels correspond to different photon energy bins as indicated in the panels. The solid, dashed, dotted, and dot-dashed curves are predictions based on \citet[][U03]{ued03} for Seyferts, \citet[][I08]{ino08} for Seyferts, \citet[][A09]{aje09} for FSRQs, and \citet[][A12]{aje12} for FSRQs. The horizontal thin double-dot-dashed line corresponds to one source detection level.\label{fig:z_1e-11} }
\end{figure*}
The cumulative source counts in the entire sky at a sensitivity of $S$, $N(>S)$, are given by integrating the luminosity function where the luminosity range is set to be $L_{\rm min}(S_{E_1, E_2},z)\le L\le L_{\rm max}$. $S_{E_1,E_2}$ represents the sensitivity in a given energy range between $E_1$ and $E_2$ in the unit of ${\rm [erg\ cm^{-2}\ s^{-1}]}$ and $L_{\rm min}(S_{E_1, E_2},z)$ is the minimum detectable luminosity with a given sensitivity $S_{E_1,E_2}$ and a redshift $z$. Figure. \ref{fig:count} shows the cumulative source counts in the entire sky as a function of sensitivity limit at a given energy band in the entire sky predicted based on U03 for Seyferts with the thermal cutoff, I08 for Seyferts with the non-thermal tail, A09 for {\it Swift}--BAT FSRQs, and A12 for {\it Fermi}--LAT FSRQs. The surface number density is given by dividing the number by $4\pi$~steradian. The energy bands are divided into five as 0.3--1~MeV, 1--3~MeV, 3--10~MeV, 10--30~MeV, and 30--100~MeV as indicated in panels. Hereinafter, we investigate required sensitivities for detection of several hundred sources. Several hundred samples may allow us to study cosmological evolution of a population \citep[see U03 for X-ray AGNs, A12 for gamma-ray FSRQs,][for gamma-ray BL Lacs]{aje14}. We also consider the sensitivity limit of $10^{-11}\ {\rm erg\ cm^{-2}\ s^{-1}}$ at the MeV band which is an order of magnitude better than the sensitivity of the COMPTEL and almost the same as the expected sensitivity of the next generation MeV instruments such as ASTROGAM.
Based on the U03 Seyfert model, $\sim20$~Seyferts at 0.3--1~MeV with the sensitivity limit of $10^{-11}\ {\rm erg\ cm^{-2}\ s^{-1}}$ are expected, which would be useful to investigate the position of the thermal cutoff energies of Seyferts. Once we can achieve the sensitivity limit of $10^{-12}\ {\rm erg\ cm^{-2}\ s^{-1}}$, $\sim1400$~Seyferts at 0.3--1~MeV. However, only $\sim10$~Seyferts are expected at 1--3~MeV. As the U03 Seyfert model relies on the thermal cutoff spectral model at 0.3~MeV, we can not expect any detection of U03-type Seyferts above several MeV.
If the non-thermal tails of Seyferts make up the MeV background (I08), more Seyferts at the MeV band would be expected. Based on the I08 Seyfert model, we can expect $\sim60$, $\sim10$, $\sim5$, and a few Seyferts at 0.3--1, 1--3, 3--10, and 10--30~MeV, respectively, with the sensitivity limit of $10^{-11}\ {\rm erg\ cm^{-2}\ s^{-1}}$. By achieving the sensitivity limit of $10^{-12}\ {\rm erg\ cm^{-2}\ s^{-1}}$, we can expect a reasonable number of detections as $\sim1800$, $\sim400$, $\sim140$, $\sim40$, and $\sim10$ Seyferts at 0.3--1, 1--3, 3--10, 10--30, and 30--100~MeV, respectively. Therefore, the sensitivity limit of $\sim10^{-12}\ {\rm erg\ cm^{-2}\ s^{-1}}$ is required to detect several hundred Seyferts even if Seyferts dominate the MeV background radiation.
If the origin of the MeV background is FSRQs (A09), we can expect $\sim90$, $\sim50$, $\sim30$, $\sim10$, and $\sim5$ FSRQs at 0.3--1, 1--3, 3--10, 10--30, and 30--100~MeV, respectively, with the sensitivity limit of $10^{-11}\ {\rm erg\ cm^{-2}\ s^{-1}}$. Once the sensitivity limit of $4\times10^{-12}\ {\rm erg\ cm^{-2}\ s^{-1}}$ is achieved, $\sim720$, $\sim420$, $\sim260$, $\sim90$, and $\sim40$ FSRQs at 0.3--1, 1--3, 3--10, 10--30, and 30--100~MeV will be detected, respectively, because of strong positive evolution. However, based on the latest A12 FSRQ GLF, with which the whole MeV background can not be explained solely by FSRQs, the expected number will be $\sim220$, $\sim260$, $\sim390$, $\sim370$, and $\sim380$ FSRQs at 0.3--1, 1--3, 3--10, 10--30, and 30--100~MeV, respectively, with the sensitivity limit of $10^{-11}\ {\rm erg\ cm^{-2}\ s^{-1}}$. Moreover, at the sensitivity limit of the COMPTEL, $\sim10^{-10}\ {\rm erg\ cm^{-2}\ s^{-1}}$, about 10~FSRQs are expected based on the A12 GLF, which is consistent with the number of the COMPTEL blazars. The difference between FSRQ models mainly comes from the different evolutionary history.
When the sensitivity limit of $10^{-12}\ {\rm erg\ cm^{-2}\ s^{-1}}$ is achieved in the lowest energy band, expected detection rate for Seyferts and FSRQs will be one Seyfert per 22--28~deg$^2$ and one FSRQ per 4--40~deg$^2$, respectively, depending on assumed models. To avoid the source confusion, we may need angular resolutions of several degrees. ASTROGAM will achieve the FWHM of the angular resolution measure of $\sim4$~deg at 0.3~MeV \footnote{\url{http://astrogam.iaps.inaf.it/index.html}}. We note that COMPTEL had those of $\sim8$~deg at 0.5~MeV \citep{sch93}.
It is also important to know how distant AGNs can be observed by future missions. Since FSRQs have a spectral peak at the MeV band, future MeV missions will be able to detect distant FSRQs. In the GeV band, the most distant blazar confirmed by {\it Fermi} is PKS~0537-286 at $z=3.1$ at $\ge100$~MeV \citep{ack15_3LAC}. \citet{rom04} and \citet{rom06} have reported one EGRET blazar candidate at $z\sim5.48$, but it has not been confirmed so far by {\it Fermi}. It is also expected that {\it Fermi} may eventually detect blazars at $z>6$ \citep{ino11_highz}. \citet{tak13} have reported a candidate {\it Fermi} gamma-ray blazar at $z\sim3-4$.
\begin{figure*}[t]
\centering
\includegraphics[bb=50 30 410 282,width=12cm]{z_dist_panel_Flim_1e-12.eps}
\caption{The same as Figure. \ref{fig:z_1e-11}, but for the limiting sensitivity of $10^{-12}\ {\rm [erg\ cm^{-2}\ s^{-1}]}$. \label{fig:z_1e-12} }
\end{figure*}
Figure. \ref{fig:z_1e-11} shows the cumulative redshift distribution $N(>z)$ as a function of redshift at a given energy band in the entire sky assuming the sensitivity limit of $10^{-11}\ {\rm erg\ cm^{-2}\ s^{-1}}$ predicted based on U03 Seyfert model, I08 Seyfert model, A09 FSRQ model, and A12 FSRQ model. Here, we set $z_{\rm max}=8$ as the most distant known quasar is at $z=7.085$ \citep{mor11}. The energy bands were divided into five as 0.3--1~MeV, 1--3~MeV, 3--10~MeV, 10--30~MeV, and 30--100~MeV as indicated in panels. Figure. \ref{fig:z_1e-12} shows the same but assuming the sensitivity limit of $10^{-12}\ {\rm erg\ cm^{-2}\ s^{-1}}$.
For the Seyfert models, the detectable Seyferts will be only local Seyferts. Even if the sensitivity limit of $10^{-12}\ {\rm erg\ cm^{-2}\ s^{-1}}$ is achieved, the most distant Seyfert will be at $z\sim1$. On the other hand, for the FSRQ models, we can expect $z\sim3$--4 FSRQs with the sensitivity limit of $10^{-11}\ {\rm erg\ cm^{-2}\ s^{-1}}$. Once the sensitivity limit of $10^{-12}\ {\rm erg\ cm^{-2}\ s^{-1}}$ is achieved, $z\sim6$ FSRQs would be detectable even at the MeV band.
\section{Discussions and Conclusions}
\label{sec:dis_con}
In this paper, we studied the expected number counts and redshift distributions of AGNs (Seyferts and FSRQs) for future MeV missions based on recent AGN LFs and spectral models as Seyferts (I08) and FSRQs (A09) are discussed as the plausible origins of the cosmic MeV gamma-ray background. For Seyferts, we assume two primary spectral models. We considered a thermal spectral model with a cutoff at 0.3~MeV (U03) and a thermal plus non-thermal spectral model (I08). We adopt the (U03) LDDE XLF for both Seyfert models. For the thermal plus non-thermal spectral model (I08), a non-thermal component appears after the thermal cutoff at 0.3~MeV with the photon index of 2.8 to explain the MeV background up to a few tens of MeV. For FSRQs, we adopt two models. One is based on the {\it Swift}--BAT detected FSRQs (A09) and the other is based on the {\it Fermi}--LAT detected FSRQs (A12). The A09 FSRQ model can explain the entire MeV background solely by FSRQs, while the A12 FSRQ model makes up up to $\sim30$\% of the MeV background by FSRQs.
Since a thermal cutoff exists in spectra in the U03 Seyfert model, we can not expect any detections at $\gtrsim1$~MeV even with the sensitivity limit of $10^{-12}\ {\rm erg\ cm^{-2}\ s^{-1}}$. In contrast, if the origin of the MeV background is non-thermal tail from Seyferts (I08), we can expect several hundred Seyferts at the MeV band with the sensitivity limit of $10^{-12}\ {\rm erg\ cm^{-2}\ s^{-1}}$. Since each Seyfert is faint, we can detect only nearby ($z\lesssim1$) Seyferts even with the sensitivity limit of $10^{-12}\ {\rm erg\ cm^{-2}\ s^{-1}}$.
If FSRQs make up the whole MeV background (A09), the sensitivity of $\sim4\times10^{-12}\ {\rm erg\ cm^{-2}\ s^{-1}}$ is needed to detect several hundreds of FSRQs at the MeV gamma-ray band. However, based on the latest FSRQ GLF (A12), the sensitivity limit of $10^{-11}\ {\rm erg\ cm^{-2}\ s^{-1}}$, which is almost the same as the expected sensitivity of the next generation MeV telescopes, would be enough to detect several hundreds of FSRQs. The difference between the two FSRQ models comes from the different evolutionary history.
Future MeV observational windows will range about three orders of magnitude in energy. With that wide energy range, the dominant population of the cosmic MeV gamma-ray background radiation would change with the energy. Furthermore, as there are uncertainties of luminosity functions and spectral models, the sensitivity of several times of $10^{-12}\ {\rm erg\ cm^{-2}\ s^{-1}}$ would be desirable to detect several hundred AGNs.
It is important to know the origin of the MeV background because the expected source counts depends on it. To observationally unveil the origin of the MeV background, we need to resolve the sky into point sources as in soft X-ray. Here, the expected sensitivity of the next generation MeV instruments such as ASTROGAM is $\sim10^{-11}\ {\rm erg\ cm^{-2}\ s^{-1}}$. With this sensitivity, the expected resolved fraction will be a few percent of the total background in the I08 Seyfert scenario, while it will be $\sim$5\% in the A09 FSRQ scenario. Thus, the large fraction of the MeV sky will not be resolved even with the next generation instruments. However, as about one hundred to several hundred FSRQs are expected, we will be able to obtain typical MeV spectra and cosmological evolution of FSRQs and theoretically estimate their contribution to the MeV background more robustly. On the other hand, it would be hard for Seyferts since we can expect only about ten sources.
The angular power spectrum of the MeV sky will provide a unique opportunity to unveil the origin of the MeV sky \citep{ino13}. The SGD onboard {\it ASTRO-H} will be able to probe it up to 0.6~MeV. Although the angular power spectrum with the sensitivity of the SGD is dominated by the Poisson term $C_l^p$, a significant difference of the $C_l^p$ of Seyferts \citep{ino08} and FSRQs \citep{aje09} is expected. This is because FSRQs are brighter but fewer than Seyferts. Anisotropies of the MeV sky would allow us to understand the origin of the MeV background and be useful to design future missions.
\bigskip
The authors would like to thank Francesco Massaro for careful reading and constructive comments. The authors also thank Yasushi Fukazawa and Kazuhiro Nakazawa for useful comments. Y.I. acknowledges support by the JAXA international top young fellowship.
|
\section{Introduction\label{sec:intro}}
Many code packages have been created to perform rapid Monte Carlo integration of the phase space for collider physics processes, to simulate the effects of hadronization, and to emulate the response of detectors such as ATLAS and CMS. The most commonly used event generator is arguably MadGraph5+aMC@NLO \cite{Alwall:2014hca}, with Pythia \cite{Sjostrand:2007gs}, Herwig++ \cite{Bahr:2008pv}, Sherpa \cite{Gleisberg:2003xi} and WHiZaRD \cite{Reuter:2014ema} also commonly used for event generation. To simulate the detector response, including efficiency and isolation details, PGS \cite{Carena:2000yx} and Delphes \cite{Selvaggi:2014mya} are the two most commonly used packages for theorists/phenomenologists, with a modified version of Delphes being employed by the CheckMATE \cite{Drees:2013wra} code package, while full detector simulators are available only to experimental groups. The collection and analysis of the events passing through the generation-through-detector-simulation process is often left to the individual, with some packages, such as MadAnalysis \cite{Conte:2014xya}, existing to assist in the reproduction of experimental analyses, with varying learning curves.
These tools are important for phenomenologists for two purposes -- proposing new analyses that will be sensitive to a well motivated model or signature, and reproducing experimental analyses for a particularly motivated model to determine exclusions of parameter space. One of the challenges of performing these studies is in producing a sufficiently large number of events, especially of standard model (SM) backgrounds, to achieve an appropriate coverage of the full phase space where the analysis focuses. A good rule-of-thumb in generating backgrounds to reproduce an experimental analysis is to generate a sufficiently large number of events that the per-event weight is equal to the inverse of the integrated luminosity. Thus, any process with a nanobarn cross section at LHC8 with 20/fb integrated luminosity, would require 20 million events to be generated. Techniques can be employed at the event generation stage to focus on the region of phase space of the SM backgrounds which overlap with the signal of interest, however this limits the reusability of these backgrounds for scans over parameter space that result in varying kinematic signatures. As a result, many analyses would require a large number of events to be generated.
Common output file formats from the generation--hadronization--simulation process, including the Les Houches file format \cite{Alwall:2006yp} and the ROOT file format \cite{Antcheva:2009zz}, produce large file sizes that can be prohibitive to store for researchers without significant computing resources. Alternatively, the LHC Olympics (LHCO) file format is a minimalist output file format that can be produced by both PGS and Delphes detector simulation packages, and has the added bonus of shielding the user from all truth information regarding the events, thus eliminating the possibility of user bias in an analysis. This file format allows a user to store a large number of events with relatively little disk space use.
This article introduces a new analysis package, entitled Seer, for the rapid analysis of an arbitrary number of LHCO files (limited by computer memory only, tested with over 850 files of 50k events each, simultaneously).\footnote{A similar effort has been made with the CutLHCO program \cite{Walker:2012vf}. In addition, MadAnalysis can perform many of the same functions, though Seer offers some differences in both the input and output, and provides some different features.} The goal of this package is to provide a simple way to combine multiple event generation files for many different processes, apply trigger requirements and kinematic cuts, and analyze and histogram the results. Many common kinematic variables are already implemented, including the popular stransverse mass ($M_{T2}$) \cite{Cheng:2008hk,Lester:1999tx,Barr:2003rg} variable and the Razor variables\footnote{Razor calculations use the script provided by CMS at \url{https://twiki.cern.ch/twiki/bin/view/CMSPublic/RazorLikelihoodHowTo}} \cite{Rogan:2010kb,Chatrchyan:2012uea} used to analyze SUSY decays, allowing usefulness out-of-the-box. Since the package employs ROOT libraries, modification of the code to add user defined kinematic calculations and cuts is simple and easy for new users.
The installation of Seer is covered in Section \ref{sec:install}. Following this, an in-depth discussion regarding the use of Seer is covered in Section \ref{sec:use}, focusing on the ``out-of-the-box" version. Instructions for modifying Seer to creatie user-defined cuts and kinematic variables is discussed in Section \ref{sec:mod}. Lastly, a series of examples are provided showing very basic analyses of Seer, and examples of the final output histograms.
Of note, a variety of font face formats have been adopted to differentiate between references to files, references to text settings, and references to parts of the Seer code. As a brief legend, files use a {\fontfamily{phv}\selectfont helvetica font}, text settings use a \texttt{true-type font}, and references to code use a {\small\texttt{small true-type font}}.
\section{Installing Seer\label{sec:install}}
Installing Seer is straightforward. Seer has been tested on OS X 10.7+ and Linux distributions using the GNU C++ compilers. Future versions may address other compiler options.
\subsection{Pre-requisites}
Seer requires ROOT libraries, as much of the functionality uses classes defined within ROOT. As a result, ROOT needs to be installed. Instructions regarding the installation of ROOT can be found at \url{https://root.cern.ch/drupal/content/downloading-root}. Seer requires that the \texttt{\$ROOTSYS} system variable be correctly setup in order to function.
\subsection{Co-requisites}
Seer includes a copy of the MT2/stransverse mass calculator developed by Z. Han and H.C. Cheng \cite{Cheng:2008hk,Lester:1999tx,Barr:2003rg}. To implement alternative versions of this code, the {\fontfamily{phv}\selectfont mt2\_bisect.cpp} and {\fontfamily{phv}\selectfont mt2\_bisect.h} files should be replaced in the Seer main directory.
Any extra C++ code a user wishes to add can be placed in the Seer directory and added to the {\fontfamily{phv}\selectfont makefile}. The name of the header file needs to be added to the \texttt{\$HEADERS} variable in the {\fontfamily{phv}\selectfont makefile}, and the object file to create needs to be added to the \texttt{\$OBJECTS} variable, in addition to adding the appropriate include statement in the Seer code.
\subsection{Unpacking}
Seer must first be unpacked. This can be performed by navigating to the directory containing the tarball and running the command {\em tar -zxvf seer.tar.gz}. This will unpack all the files and directories needed by Seer. The following files are mandatory for Seer to compile and run:
\begin{itemize}
\item {\fontfamily{phv}\selectfont seer.h} \& {\fontfamily{phv}\selectfont seer.cpp} -- header and C++ file for the main script.
\item {\fontfamily{phv}\selectfont eventweights.h} \& {\fontfamily{phv}\selectfont eventweights.cpp} -- header and C++ file for the script that pre-scans the LHCO files for cross section and number of events.
\item {\fontfamily{phv}\selectfont calculations.h} \& {\fontfamily{phv}\selectfont calculations.cpp} -- header and C++ file for the cuts/trigger/analysis class {\small\texttt{event}}.
\item {\fontfamily{phv}\selectfont plots.h} \& {\fontfamily{phv}\selectfont plots.cpp} -- header and C++ file for the plotting class {\small\texttt{plot}}.
\item {\fontfamily{phv}\selectfont style.h} -- style file containing a number of settings for the histograms.
\item {\fontfamily{phv}\selectfont mt2\_bisect.h} \& {\fontfamily{phv}\selectfont mt2\_bisect.cpp} -- header and C++ file for the MT2 calculator.
\item {\fontfamily{phv}\selectfont seer\_cuts.txt} -- contains all the cut and trigger settings
\item {\fontfamily{phv}\selectfont seer\_fakes.txt} -- contains all the fake rate settings
\item {\fontfamily{phv}\selectfont seer\_files.txt} -- contains the list of LHCO files
\item {\fontfamily{phv}\selectfont seer\_plots.txt} -- contains the plot settings
\item {\fontfamily{phv}\selectfont seer\_settings.txt} -- contains all settings for the constants that can be used in the calculations
\end{itemize}
\subsection{Compiling}
Seer is compiled by navigating to the directory in a terminal window and entering the $make$ command.
\section{Using Seer\label{sec:use}}
Seer is run via a single command in a terminal, {\fontfamily{phv}\selectfont ./seer.exe}. All details of the run are determined by the settings in five text files: three primary files ({\fontfamily{phv}\selectfont seer\_files.txt}, {\fontfamily{phv}\selectfont seer\_cuts.txt} and {\fontfamily{phv}\selectfont seer\_plots.txt}) and two secondary files ({\fontfamily{phv}\selectfont seer\_fakes.txt}, and {\fontfamily{phv}\selectfont seer\_settings.txt}). The general process followed by Seer is given in Figure \ref{fig:flowchart}. A brief summary of the files is included below, and detailed information is provided in individual sections for each file.\\
\begin{figure*}[!t]
\centering
\mbox{\includegraphics[width=\textwidth]{Flowchart.pdf}}
\caption{Flow chart of Seer, describing where input settings are used, and the main flow of processes, and the files that contain each of the main functions.
\label{fig:flowchart}}
\end{figure*}
\noindent{\fontfamily{phv}\selectfont seer\_files.txt}
This is the first file that should be adjusted when starting an analysis. It contains the list of all LHCO files in the analysis, as well as details on their groupings. More information regarding this file is provided in Section \ref{sec:files}.\\
\noindent{\fontfamily{phv}\selectfont seer\_cuts.txt}
This is the second file that should be adjusted. It contains the details of the tagging, the experimental trigger thresholds, fine tuning details of the detector limits, and details of the cuts. More information regarding this file is provided in Section \ref{sec:cuts}.\\
\noindent{\fontfamily{phv}\selectfont seer\_plots.txt}
This is the last primary file to adjust. It contains the settings required to determine what type of histogram to make, as well as the details of the histogram and aspects of the output file names. More information regarding this file is provided in Section \ref{sec:plots}.\\
\noindent{\fontfamily{phv}\selectfont seer\_fakes.txt}
This is a secondary file, as the settings within are useful only for advanced users. Seer can be used to handle fakes instead of the detector simulator used. More information regarding this file is provided in Section \ref{sec:fakes}.\\
\noindent{\fontfamily{phv}\selectfont seer\_settings.txt}
Some additional settings are contained in this file, including numerical constants used by Seer to add additional smearing to jet and missing transverse energy kinematics. More information regarding this file is provided in Section \ref{sec:settings}.\\
\subsection{Seer Output}
The output of Seer is two files -- one plot file and one text file. The naming system for both of these files is identical except for the extension, and follows the scheme:
\begin{verbatim}
(plot_type)_(name_string)_n######.(file_type)
\end{verbatim}
where \texttt{(plot\_type)} is fixed in {\fontfamily{phv}\selectfont plots.cpp} and depends only on the \texttt{(plotnum)} setting in {\fontfamily{phv}\selectfont seer\_plots.txt}, \texttt{(name\_string)} is the setting in {\fontfamily{phv}\selectfont seer\_plots.txt} that allows users to add a unique or descriptive text identifier to a particular analysis, and the sequence of numbers is the total sum of events of all files included in {\fontfamily{phv}\selectfont seer\_files.txt}. The output file type for the histogram depends on the \texttt{(file\_type)} setting in {\fontfamily{phv}\selectfont seer\_files.txt}, while the text output file is in a .txt format. More information regarding these settings is given in Section \ref{sec:plots}. Note that running Seer multiple types with the same \texttt{(plotnum)} and same files in {\fontfamily{phv}\selectfont seer\_files.txt} will result in overwriting previous files. Users need to use the \texttt{(name\_string)} field to differentiate between applying different cuts or other analysis details that do not change the file name.
The plot output file contains a figure of the histograms of the various groupings of files included in {\fontfamily{phv}\selectfont seer\_files.txt}. The text output file contains a header with a copy of every line in the five input settings files. Following that, it provides the following information:
\begin{enumerate}
\item Individual File Summaries: for each file, a section starting with the text ``Now running event file..." provides the file path and tag in a descriptive sentence, then gives a summary of the results of the analysis.
\begin{itemize}
\item File Tag -- tag given in {\fontfamily{phv}\selectfont seer\_files.txt}
\item File Event Weight -- the event weight of the individual file (ignoring multiple files with same tag)
\item Combined Event Weight -- the event weight used when accounting for combined files ($w_f$)
\item Untagged Events -- the number of untagged events, and corresponding cross section ($\sigma_{\mathrm{untagged}} = n_{\mathrm{untagged}} \times w_f$)
\item Cut Events -- the number of events that were tagged but were cut in the analysis, and corresponding cross section ($\sigma_{\mathrm{cut}} = n_{\mathrm{cut}} \times w_f$)
\item Tagged \& Uncut Events -- the number of events that were tagged and passed the cuts of the analysis, and corresponding cross section ($\sigma_{\mathrm{included}} = n_{\mathrm{included}} \times w_f$)
\item Total Events -- the sum of the Untagged, Uncut and Tagged \& Uncut events, and corresponding cross section.
\item Cross Section -- the final, included cross section passed to the histogram, including associated theoretical error ($\delta \sigma_{\mathrm{included}} = \sigma_{\mathrm{included}}/\sqrt{n_{\mathrm{included}}}$; if no events pass cuts/tagging, $\delta\sigma_{\mathrm{included}} = w_f$ is used)
\end{itemize}
\item Group Summaries: includes all events from all files with the same tag in {\fontfamily{phv}\selectfont seer\_files.txt}.
\begin{itemize}
\item Group Tag -- tag of the group of files, as defined in {\fontfamily{phv}\selectfont seer\_files.txt}
\item Untagged Events -- the number of untagged events, and corresponding cross section ($\sigma_{\mathrm{untagged}} = n_{\mathrm{untagged}} \times w_f$)
\item Cut Events -- the number of events that were tagged but were cut in the analysis, and corresponding cross section ($\sigma_{\mathrm{cut}} = n_{\mathrm{cut}} \times w_f$)
\item Tagged \& Uncut Events -- the number of events that were tagged and passed the cuts of the analysis, and corresponding cross section ($\sigma_{\mathrm{included}} = n_{\mathrm{included}} \times w_f$)
\item Total Events -- the sum of the Untagged, Uncut and Tagged \& Uncut events, and corresponding cross section.
\item Cross Section -- the final, included cross section passed to the histogram, including associated theoretical error ($\delta \sigma_{\mathrm{group}} = \sqrt{\sum \delta\sigma_{\mathrm{individual}}^2}$)
\end{itemize}
\item Overall Summary: a list of the results from all files combined. This is the same information as in the group summaries, except now it includes all groups combined.
\item Breakdown of Cuts: a list of the cross section removed by each individual cut. Since multiple cuts can remove an event, these are not exclusive cross sections.
\item Number checks: a weight sum and total number of events check to make sure everything adds up correctly
\item Histogram results: a tab separated text list of the bin centre and cross section values for each histogram (in fb)
\end{enumerate}
Users who are interested in the total cross section before and after cuts can read this information directly from the relevant group summary in the text file. Users who are interested in distributions have a choice between the visual histogram and the text of the histogram values. Of note, the backgrounds in the text file are not summative, unlike the background histograms in the plot file, as discussed in Section \ref{sec:plots}.
The following sections describe in great detail the text files that Seer uses to read in all of the settings for the run. It is important to note that a line can be commented out by using double hashtags, \#\#.
\subsection{{\fontfamily{phv}\selectfont seer\_files.txt} \label{sec:files}}
Seer checks this file for the list of all LHCO files to be used in the analysis. For each file/line, there are four required entries and two optional entries. The format of these entries is as follows:
\begin{verbatim}
(tag) (d/s) (fake#) (file_location) (cross_section) (n_events)
\end{verbatim}
The \texttt{(cross\_section)} and \texttt{(n\_events)} entries are {\em optional}, meaning that they can be excluded from file without causing problems with the operation of Seer. These values are used to scale the per-event weighting. By default, Seer will read these values from the LHCO file directly if they are not set, but a user may wish to rescale the cross section and can do so to override the automatic scan performed by Seer. Further details of the entries in the {\fontfamily{phv}\selectfont seer\_files.txt} file are provided in Table \ref{tab:files}.
\begin{table}
\caption{Explanation of the settings for the {\fontfamily{phv}\selectfont seer\_files.txt} file.\label{tab:files}}
\noindent\begin{tabularx}{\textwidth}{|>{\bfseries}lX|}
\hline
Parameter & Explanation\\
\hline
\texttt{(tag)} & This should be an alphanumeric (including symbols ()-+=,.\#\%\/\_). Multiple sequential entries of identical tags will result in these files being grouped together, with cross sections being combined based on the status of the \texttt{(d/s)} flag. Note that the tag choice here will also be the text entry in the legend for that file.\\
\texttt{(d/s)} & This is a flag with two possible values: ``d" or ``s". A choice of ``d" indicates that this file has a different phase space or process from other files with the same tag, and the cross section should be summed. A choice of ``s" indicates that this file duplicates the same phase space as the previous file in the list, and so the total cross section should be averaged between all files with an ``s" and the same tag. \\
\texttt{(fake\#)} & This is a flag for adding in a pseudo-fake rate, and is a number between 0 and 9, inclusive. This is discussed further in Sec. \ref{sec:fakes}.\\
\texttt{(file\_location)} & This is the path (full or relative) to the LHCO file, including the file name.\\
\texttt{(cross\_section)} & This is a cross section value (in pb) that rescales the total cross section for that individual file only. This can be used to rescale total cross sections as desired, or to combine multiple different processes with the same tag, as discussed below.\\
\texttt{(n\_events)} & This is the number of events that corresponds to the cross section value. This entry must accompany the \texttt{(cross\_section)} entry when it is present. Values can be either pre- or post- matching (e.g. the \texttt{(cross\_section)} and \texttt{(n\_events)} settings 5.31 \hspace{1mm} 50000 would be the same as 3.5955 \hspace{1mm} 33856).\\
\hline
\end{tabularx}\\
\end{table}
To better explain the behaviour of Seer, it is useful to examine some example scenarios to clarify the interplay between the (tag) and (d/s) settings.\\
\vspace{2mm}
\begin{tabularx}{\textwidth}{>{\bfseries}lX}
Scenario 1:& Generated $p p \rightarrow W^+ W^-$ and $p p \rightarrow Z$, listed with a common tag (e.g. ``background") and with the flag ``d". This results in a single histogram being generated combining the events from both files, and with $\sigma_{tot} = \sigma_{WW}+\sigma_{Z}$.\\
Scenario 2:& Generated 10 files of $p p \rightarrow W^+ W^-$, listed with a common tag and the flag ``s". This results in a single histogram with total cross section averaged over all files ($\sigma_{tot} = \sum \sigma_i \times n_i / \sum n_i$), and the per-event cross section weighted by the inverse of the combined sum of events from all files. Note that these files have to be generated with the exact same phase space limits, otherwise combining them will not be valid.\\
Scenario 3:& Generated 10 files of $p p \rightarrow \ell^+ \ell^-$ each with different, non-overlapping ranges of $m_{\ell\ell}$ (such as 0-100, 100-200, 200-300, 300-400, etc...), listed with a common tag and with the flag ``d". This results in a single histogram with the combined total cross section ($\sigma_{tot} = \sum \sigma_i$). This method results in significantly better sensitivity over the full range of phase space as compared with 10 files with identical phase space.\\
\end{tabularx}
\vspace{2mm}
Files with unique tags will always be treated independently, and be plotted with a their unique tag. Thus, if there is only one file for a given tag, the choice of the \texttt{(d/s)} flag is unimportant.
There is currently no way to automatically combine multiple files of each of two separate processes with the same tag (i.e. Scenario 1 but with multiple files of each of the processes). This is why the cross section and event number overrides are implemented. In order to use this correctly, the user must manually reduce the cross section to the per-event value following the formula:
\begin{equation}
\sigma_{i}^\prime = \displaystyle\frac{\sigma_{i} \times n_{i}}{\displaystyle\sum_{j=1}^k n_{j}}
\end{equation}
where $\sigma_{i}$ is the cross section listed in the LHCO file. If all files have the same number of events, then this works out to simply $\sigma_{i}/k$.
Determining the per-event cross section is only slightly more complicated when users are employing matching and merging algorithms in the generation stage. Currently, the MG5+aMC@NLO plus Pythia package lists the total cross section before matching and the total events generated before matching/merging, as well as the events after matching/merging, in the LHCO file. In this case, the relationship $\sigma_{before}/n_{before} = \sigma_{after}/n_{after}$ holds (where {\em before}/{\em after} refers to before/after merging/matching), and so the per-event weight can be determined with $\sigma_{before}/n_{before}$, which is calculated at the MG5+aMC@NLO stage.
\clearpage
\subsection{{\fontfamily{phv}\selectfont seer\_cuts.txt} \label{sec:cuts}}
In the {\fontfamily{phv}\selectfont seer\_cuts.txt} file, four different types of entries are possible: tagging, trigger, ignores and cuts. These will each be addressed separately below. The general format for the {\fontfamily{phv}\selectfont seer\_cuts.txt} file is given by:
\begin{verbatim}
Signal ## Do not alter this line!
##TAGGING
(nlow) (nhigh) (type), pt (ptlow) (pthigh), eta (etalow) (etahigh)
Extra ## Do not alter this line!
##TRIGGER
trigger (type) pt (pt1) (pt2) (pt3) (pt4)
##IGNORES
ignore (type) with (pt/eta) between (kinlow) (kinhigh)
ignore (type) with (pt/eta) over (kinlow)
##CUTS
addcut (type) with (variable) between (kinlow) (kinhigh)
addcut (type) with (variable) over (kinlow)
\end{verbatim}
Changeable settings are enclosed in parentheses. The explanations for each section (tagging, trigger, ignore, and cut) are given below, along with the explanation of each relevant setting entry.
\subsubsection{Tagging}
Tagging determines the particle combinations that Seer uses in its calculations. Tagging criteria are added after the ``Signal" entry in the {\fontfamily{phv}\selectfont seer\_cuts.txt} file and before the ``Extra" entry. When passing events to the cut section for analysis, only events that meet precisely the criteria in this list will be included.
\begin{table}
\caption{Tagging settings definitions and explanations.\label{tab:tagging}}
\noindent\begin{tabularx}{\textwidth}{|>{\bfseries}lX|}
\hline
Parameter & Explanation\\
\hline
\texttt{(nlow)} & This is the minimum number of that type of particle.\\
\texttt{(nhigh)} & This is the maximum number of that type of particle.\\
\texttt{(type)} & This is the type of particle - options include:\\
& ele -- electron\\
& muo -- muon\\
& lep -- lepton, includes muon and electrons.\\
& tau -- hadronic taus\\
& ljt -- light jets (does not included $b$-tagged jets or hadronic tau)\\
& bjt -- $b$-tagged jets\\
& jet -- all hadronic jets (includes hadronic taus, light jets and bjets)\\
& gam -- photons\\
& all -- all particle types combined\\
\texttt{(ptlow)} & the minimum amount of transverse momentum (in GeV) for that specific type of particle\\
\texttt{(pthigh)} & the maximum amount of transverse momentum (in GeV) for that specific type of particle (99999 is often used to represent a value large enough that there is no upper cutoff in $p_T$)\\
\texttt{(etalow)} & the minimum pseudorapidity for that specific type of particle (typically negative value)\\
\texttt{(etahigh)} & the maximum pseudorapidity for that specific type of particle (typically positive value)\\
\hline
\end{tabularx}\\
\end{table}
The format for an entry is:
\begin{verbatim}
(nlow) (nhigh) (type), pt (ptlow) (pthigh), eta (etalow) (etahigh)
\end{verbatim}
These entries are not inclusive to particles that are not listed in the tag list. The seer process goes through the list of tagging entries and marks off every particle entry in an event as tagged. Once all tagging entries have been identified, the presence of any untagged particle in the event listing will result in the entire event being identified as ``untagged". There are two ways to make an analysis more inclusive (e.g. inclusive for jets) -- add an extra entry of 0-99 of that particle type with a wide range of kinematics, or add an entry to the ignore list (discussed in Section \ref{sec:ignore}). To make it clear how this works, several examples are included below.
\noindent{\bf Example 1:} This example will only perform an analysis on events that contain between 1 and 2 leptons (either muons or electrons) that have $10 < p_T < 99999$ GeV and within $|\eta| < 2.5$, as well as any number of light jets between 0 and 99 that have $10 < p_T < 40$ GeV, and within $|\eta| < 5$. Note, the presence of a $\tau_h$ or $b$-tagged jet of any $p_T$, or a light jet with $p_T>40$ GeV would cause the entire event to be rejected from the analysis.
\begin{verbatim}
##TAGGING
1 2 lep, pt 10 99999, eta -2.5 2.5
0 99 ljt, pt 10 40, eta -5 5
\end{verbatim}
\noindent{\bf Example 2:} This example will include only events with exactly two high $p_T$ light jets ($200 < p_T < 99999$) and then be inclusive to all light jets with $p_T$ below 100 GeV. Note the double entry of the same type of particle. Seer will tag the items on the list in the order they are listed, and in order from high to low $p_T$.
\begin{verbatim}
##TAGGING
2 2 ljt, pt 200 99999, eta -2.5 2.5
0 99 ljt, pt 10 100, eta -5 5
\end{verbatim}
\subsubsection{Trigger}
Triggers are implemented in Seer as minimum thresholds for $p_T$ required for an event to be considered for analysis. These are effectively treated as step-function cuts at the threshold values. All triggers listed in {\fontfamily{phv}\selectfont seer\_cuts.txt} are combined, so any given event will be included in the analysis if it passes the requirements for at least one trigger. To effectively include all events, or rather to turn off the trigger, all of the trigger lines should be commented out (double hash tag, \#\#, preceding the lines).
The format for trigger entries is:
\begin{verbatim}
trigger (type) pt (pt1) (pt2) (pt3) (pt4)
\end{verbatim}
The text for ``trigger" and ``pt" are both necessary and are not settings to be changed. The \texttt{(type)} entry is based on the list of possible triggers, not based on the type of particles. The currently implemented types of triggers are listed in Table \ref{tab:triggers}. There are up to four $p_T$ thresholds that can be set. For trigger types of a single particle type (like jets, which can have , this should be the $p_T$ ordered list of the thresholds. For triggers of a combination of types
\begin{table}
\caption{Implemented trigger types, with explanations. \label{tab:triggers}}
\noindent\begin{tabularx}{\textwidth}{|>{\bfseries}lX|}
\hline
Trigger & Explanation\\
\hline
ele & electron, four numerical options list the $p_T$ ordered trigger thresholds (a value of 0 turns that option off)\\
muo & muon, four numerical options list the $p_T$ ordered trigger thresholds (a value of 0 turns that option off)\\
gam & photon, four numerical options list the $p_T$ ordered trigger thresholds (a value of 0 turns that option off)\\
tau & hadronic tau, four numerical options list the $p_T$ ordered trigger thresholds (a value of 0 turns that option off)\\
met & missing transverse energy, only the first numerical option should list the missing $E_T$ threshold, the rest should be set to 0\\
ljt & light jet, four numerical options list the $p_T$ ordered trigger thresholds (a value of 0 turns that option off)\\
jet & jet, includes $b$ tagged jets, four numerical options list the $p_T$ ordered trigger thresholds (a value of 0 turns that option off)\\
bjt & $b$ tagged jets, four numerical options list the $p_T$ ordered trigger thresholds (a value of 0 turns that option off)\\
tae & single hadronic tau plus electron, numerical option \texttt{(pt1)} is the $p_T$ threshold of the $\tau_h$, numerical option \texttt{(pt2)} is the $p_T$ threshold of the electron (\texttt{(pt3)} and \texttt{(pt4)} should be set to 0)\\
tam & single hadronic tau plus muon, numerical option \texttt{(pt1)} is the $p_T$ threshold of the $\tau_h$, numerical option \texttt{(pt2)} is the $p_T$ threshold of the muon (\texttt{(pt3)} and \texttt{(pt4)} should be set to 0)\\
jat & single jet (includes $b$ tagged jets) plus missing $E_T$, numerical option \#1 is the $p_T$ threshold of the jet, numerical option \texttt{(pt2)} is the $\ensuremath{\not\!\!E_T}\xspace$ threshold (\texttt{(pt3)} and \texttt{(pt4)} should be set to 0)\\
tat & single $\tau_h$ plus missing $E_T$, numerical option \texttt{(pt1)} is the $p_T$ threshold of the $\tau_h$, numerical option \texttt{(pt2)} is the $\ensuremath{\not\!\!E_T}\xspace$ threshold (\texttt{(pt3)} and \texttt{(pt4)} should be set to 0)\\
eam & single electron and single muon, numerical option \texttt{(pt1)} is the $p_T$ threshold of the electron, and \texttt{(pt2)} is the $p_T$ threshold of the muon (\texttt{(pt3)} and \texttt{(pt4)} should be set to 0)\\
\hline
\end{tabularx}
\end{table}
\subsubsection{Ignore\label{sec:ignore}}
Seer has an extra feature called ``ignore", which allows fine tuning adjustment of how Seer treats particle entries in an event. This is particularly useful when using a single, inclusive particle type in which there are different kinematic limits for each. For example, the ATLAS detector can detect electrons with pseudorapidities within $|\eta| < 2.47$, and muons within $|\eta| < 2.5$. Asking Seer to tag general leptons within $|\eta| < 2.5$ would nominally include any electrons with $2.47 < |\eta| < 2.5$. Other uses include specifying whether particle information is discarded, as in \cite{atlas_conf_2013_047}, where it is stated that ``Following this step, all jet candidates with $|\eta|>2.8$ are discarded." This indicates that the event is retained even where jet candidates occur outside $|\eta| = 2.8$, but that the information on those jets are not included in calculations of the kinematics that characterize the event. This kind of treatment is not handled in the Delphes detector simulation.
An ignore setting tells Seer to {\em remove} a particle listing that meets that criteria from the event list, and from all sub-lists within Seer, but does not otherwise change the rest of the event. For example, an electron that is ignored is removed from the electron lists, the lepton lists, and the inclusive ``all" list. Since ignores are run before tagging and trigger, as shown in Figure \ref{fig:flowchart}, particles that are ignored will not participate in those processes.
There are two possible entry types for ignore statements, which are:
\begin{verbatim}
ignore (type) with (pt/eta) over (kinlow)
ignore (type) with (pt/eta) between (kinlow) (kinhigh)
\end{verbatim}
where text entries not enclosed in parentheses are text strings that are required and should not be changed. An ``over" statement indicates that any particles with kinematic ($p_T$ or $\eta$) values larger than the (kinlow) listed will be ignored. A ``between" statement indicates that any particles with kinematic ($p_T$ or $\eta$) values between \texttt{(kinlow)} and \texttt{(kinhigh)} will be ignored. In this case, the \texttt{(eta)} is absolute valued. Thus setting \texttt{eta over 2.0} would be $|\eta|>2.0$. Further descriptions of the limits for these settings is listed in Table \ref{tab:ignore}.
\begin{table}
\caption{Description and explanation of options for ignore feature.\label{tab:ignore}}
\noindent\begin{tabularx}{\textwidth}{|>{\bfseries}lX|}
\hline
Parameter & Explanation\\
\hline
\texttt{(type)} & This can be any particle type from the following list: ele, muo, lep, tau, jet, ljt, bjt, gam.\\
\texttt{(pt/eta)} & This is a text string that is either ``pt" or ``eta" and tells Seer what the kinlow and kinhigh values represent.\\
\texttt{(kinlow)} & This is the lower threshold for the ignore statement.\\
\texttt{(kinhigh)} & This is the upper threshold for the ignore statement.\\
\hline
\end{tabularx}
\end{table}
{\bf Special Note:} If a user is interested in being inclusive to a particular type of particle (assuming the event signature and trigger is based on other particle states), perhaps jets with $p_T< 30$~GeV, this can be done with a tagging statement of e.g. ``0 99 jets, pt 0 30, eta -2.5 2.5". If the user were to calculate the total $H_T$ of events using this method, these low $p_T$ jets would be included in the $H_T$ calculation. Alternatively, if a user were to use an ignore statement of ``ignore jet with pt between 0 30", then events would be similarly inclusive of these jets however all Seer calculations would be blind to these jets -- the low $p_T$ jets would not be included in total $H_T$ calculations or any other calculations that involve jets. Thus, a user needs to be aware of the effect of using ignore versus tagging statement on inclusiveness of events.
\subsubsection{Cuts}
One of the primary functions of Seer is in the application of cuts. Many types of cut options have been implemented by default. The addition of user-defined cuts is discussed in Section \ref{sec:mod}. This section discusses only the default cuts. To clarify possible confusion between cuts and ignores: cuts {\em remove the entire event} if the criteria is met, where as ignores {\em remove only the specific particle listing within the event} if the criteria is met while leaving the rest of the event untouched.
\begin{table}
\caption{Description and explanation of options for addcuts feature.\label{tab:cuts}}
\noindent\begin{tabularx}{\textwidth}{|>{\bfseries}lX|}
\hline
Parameter & Explanation\\
\hline
\texttt{(type)} & This can be any particle type from the following list: ele, muo, lep, tau, jet, ljt, bjt, gam, met, all.\\
\texttt{(variable)} & For all particle lists (ele through gam in the above list), there is a common set of cut variables. For all and met, there are special cut variables. See Tables \ref{tab:kin1} and \ref{tab:kin2}.\\
\texttt{(kinlow)} & This is the lower threshold for the cut statement.\\
\texttt{(kinhigh)} & This is the upper threshold for the cut statement.\\
\hline
\end{tabularx}
\end{table}
The format for cut entries is:
\begin{verbatim}
addcut (type) with (variable) over (kinlow)
addcut (type) with (variable) between (kinlow) (kinhigh)
\end{verbatim}
This format is very similar to the format for the Ignores listed previously. However, there are many more options available for the \texttt{(type)} and \texttt{(variable)} listings. Further explanations for the constraints on these settings are given in Table \ref{tab:cuts}. Furthermore, the types of \texttt{(variable)} options are dependent on the particle \texttt{(type)} chosen. Details of the kinematic cuts available for each particle type are listed in Tables \ref{tab:kin1} and \ref{tab:kin2}.
Cuts that do not necessarily have a single particle type, cuts that depend on multiple types of particles, and cuts that are created by the user, require a \texttt{(type)} setting of ``all". The eta, pt, num, ht, im2 and mas cuts, discussed in Table \ref{tab:kin1} are also implemented for ``all". All addition cuts that are also defined only for ``all" are listed in Table \ref{tab:kin2}. The ``met" setting for \texttt{(type)} only has one possible \texttt{(variable)} option, which is ``pt".
The following are a few examples for cuts that can be implemented:\\
\noindent{\bf Example 1:} Removing all events with light jet $p_T<50$~GeV is accomplished with the line \texttt{addcut ljt with pt between 0 50}.\\
\noindent{\bf Example 2:} Removing all events with any combination of dilepton invariant mass within 20 GeV of the $Z$ mass is accomplished with the line \texttt{addcut lep with im2 between 72 112}.\\
\noindent{\bf Example 3:} Removing all events $\ensuremath{\not\!\!E_T}\xspace > 50$~GeV is accomplished with the line \texttt{addcut met with pt over 50}.\\
\begin{table}
\caption{Cut variables common to particle (type) selections: ele, muo, lep, tau, jet, ljt, bjt and gam.\label{tab:kin1}}
\noindent\begin{tabularx}{\textwidth}{|>{\bfseries}lX|}
\hline
Cut Type & Explanation\\
\hline
eta & Pseudorapidity of any particle of that type. If any particle of that type has a pseudorapidity value within the specified range, the entire event is cut.\\
pt & Transverse momentum of any particle of that type. If any particle of that type has a transverse momentum value within the specified range, the entire event is cut. This is the only cut type currently implemented for the ``met" particle type.\\
pt1 & Transverse momentum of only the leading particle of that type. Cuts events based only on the $p_T$ of the highest $p_T$ particle of that type.\\
pt2 & Transverse momentum of only the subleading particle of that type. Cuts events based only on the $p_T$ of the second highest $p_T$ particle of that type.\\
num & Multiplicity of particles of that type. Cuts based on the total number of particles of that type in the event (this is highly influenced by the use of tagging versus ignores for inclusivity).\\
ht & Scalar sum of the $p_T$ of all particles of that type. If the total scalar sum of the $p_T$ of all particles of that type is within the specified range, the event is cut.\\
im2 & Invariant mass of any combination of two particles of that type (for leptons, the pairing must be opposite sign). If any pairing of particles of that type has an invariant mass within the range, the event is cut.\\
mas & Total invariant mass of all particles of that type. If the total invariant mass of all particles of that type lies within the range, the event is cut.\\
\hline
\end{tabularx}
\end{table}
\begin{table}
\caption{Cut variables options available only to the (type) ``all".\label{tab:kin2}}
\noindent\begin{tabularx}{\textwidth}{|>{\bfseries}lX|}
\hline
Cut Type & Explanation\\
\hline
mt2 & Stransverse mass variable. This only works if there are exactly 2 particles in the event (a user may use ignores to neglect low energy jets, for example), and will cut the event if there are more than 2 particles. It calculates the $M_{T2}$ variable using the code developed by \cite{Cheng:2008hk}.\\
maa & Leading diphoton mass. This first checks to see if there are at least two photons in the event, and then cuts based on the invariant mass of the leading two photons. There is some overlap in purposes between this and the im2 option in the previous table. For im2, it searches through all possible combinations of diphotons, but will not provide a cut if there is only one diphoton. This will cut the event if there is only one photon, and only looks at the leading two photons.\\
mll & Same as maa but for leptons.\\
mjj & Same as maa but for jets (all types).\\
mtl & Transverse mass of the only lepton (if only one), or the lowest $p_T$ lepton in the case of multiple leptons.\\
mcol & Collinear invariant mass of exactly two leptons (there can be any number of other particles), including taus. This calculates the invariant mass in the assumption that the missing transverse momentum is comprised from neutrinos that are collinear with the two leptons.\\
osl & Opposite sign (OS) leptons. Behaviour depends on the value of kinlow (kinhigh ignored). If kinlow is 0, all events without an OS lepton pairing are cut. If kinlow is 1, all events with an OS lepton pairing are cut.\\
sfl & Same flavour, opposite sign (SFOS) leptons. Behaviour depends on the value of kinlow (kinhigh ignored). If kinlow is 0, all events without an SFOS lepton pairing are cut. If kinlow is 1, all events with an SFOS lepton pairing are cut.\\
ofl & Opposite flavour, opposite sign (OFOS) leptons. Behaviour depends on the value of kinlow (kinhigh ignored). If kinlow is 0, all events without an OFOS lepton pairing are cut. If kinlow is 1, all events with an OFOS lepton pairing are cut.\\
ssl & Same sign (SS) leptons. Behaviour depends on the value of kinlow (kinhigh ignored). If kinlow is 0, all events without an SS lepton pairing are cut. If kinlow is 1, all events with an SS lepton pairing are cut.\\
\hline
\end{tabularx}
\end{table}
\clearpage
\subsection{{\fontfamily{phv}\selectfont seer\_plots.txt}\label{sec:plots}}
Seer can create either standard one-variable histograms, or can plot two-variable histograms. For one-variable histograms, the $y$-axis contains either cross section values (in fb) or a fraction of events, while the $x$-axis contains the binned kinematic variable value. Note that bins values are not divided by bin width, so Seer does not plot $d\sigma/dx$. Two-variable histograms have binned kinematic variables on both the $x$- and $y$-axes. The way that the bin values are plotted in two-variable histograms depends on whether there is one event type added or multiple event types. For a single event type (i.e. a single signal file), the histogram is plotted with a colour mapping for the cross section, with an associated scale. For multiple event types (i.e. signal files plus different backgrounds), Seer plots a pseudo-Dalitz style plot using the {\small\texttt{box}} option for the {\small\texttt{TH2}} class in ROOT.
For one-variable histograms, Seer distinguishes between two different kinds of events: signal and background. Backgrounds are additive, in that each histogram is the sum of the bin values of all smaller cross section histograms, plus the bin values for the cross section of the specified background. The order of histograms is thus from largest to smallest background, and the bin values for the largest cross section is in fact the sum of all backgrounds, so the user does not need to manually combine backgrounds to estimate the total background cross section. Backgrounds are plotted with a solid fill colour.
Files considered to be part of the signal are not additive, but are re-ordered to plot from largest total cross section to smallest. In addition, signal histograms do not use a fill colour and are only a contour, and are plotted on top of all the backgrounds. Signals are distinguished from backgrounds based on the text string in the ``tag" entry in the {\fontfamily{phv}\selectfont seer\_files.txt} file.
The relevant entries in the {\fontfamily{phv}\selectfont seer\_plots.txt} file are, where adjustable settings are enclosed in parentheses:
\begin{verbatim}
Disable = (disable)
PlotType = (plotnum)
PlotType2 = (plotnum2)
PlotMaxX = (xmax)
PlotMinX = (xmin)
PlotMaxY = (ymax)
PlotMinY = (ymin)
NumBinX = (nbins)
NumBinY = (nbins)
ChooseLn = (log)
Normalize = (norm)
PlotTxt = (name_string)
SigPref = (signal_string)
FileType = (file_type)
\end{verbatim}
The explanations for each of the entries in this file are given in Table \ref{tab:plots}.
\begin{table}
\caption{Plotting settings definitions and explanations.\label{tab:plots}}
\noindent\begin{tabularx}{\textwidth}{|>{\bfseries}lX|}
\hline
Parameter & Explanation\\
\hline
\texttt{(disable)} & This is a flag to disable the plotting features of Seer (1 = disable, 0 = enable). The text output is still produced. Seer has a maximum of 50 histograms (distinct \texttt{(tag)} entries in {\fontfamily{phv}\selectfont seer\_files.txt}), and this flag will necessarily need to be implemented if this number is exceeded.\\
\texttt{(plotnum)} & This is the kinematic variable that is plotted on the $x$-axis. The list of possible kinematic variables is included as comments in the {\fontfamily{phv}\selectfont seer\_plots.txt} file. The list is too long to include in this document.\\
\texttt{(plotnum2)} & This is the kinematic variable that is plotted on the $y$-axis, similar to \texttt{(plotnum)}. A value of 0 turns this feature off, and Seer plots a one-variable histogram.\\
\texttt{(xmax)} & This is the maximum value on the $x$-axis.\\
\texttt{(xmin)} & This is the minimum value on the $x$-axis.\\
\texttt{(ymax)} & This is the maximum value on the $y$-axis, which is cross section in units of fb. (Future functionality will include typing ``auto" for this setting.)\\
\texttt{(ymin)} & This is the minimum value on the $y$-axis, which is cross section in units of fb. (Future functionality will include typing ``auto" for this setting.)\\
\texttt{(nbinx)} & This is the number of bins to plot between \texttt{(xmin)} and \texttt{(xmax)}.\\
\texttt{(nbiny)} & This is the number of bins to plot between \texttt{(ymin)} and \texttt{(ymax)}.\\
\texttt{(log)} & Only considers values of 0 or 1. A value of 1 sets the $y$-axis or $z$-axis to use a logscale. A value of 0 sets the $y$-axis or $z$-axis to use a linear scale.\\
\texttt{(norm)} & Only considers values of 0 or 1. A value of 1 normalizes the binned events to the total binned cross section, and the $y$-axis ($z$-axis) contains event fraction values. A value of 0 sets the $y$-axis ($z$-axis) to plot cross section values.\\
\texttt{(name\_string)} & This adds a text string to the output file names. Useful for distinguishing between subsequent runs with identical files, but different settings.\\
\texttt{(signal\_string)} & This is the identifying text of the prefix of the signal file tags. For example, if the tag line in {\fontfamily{phv}\selectfont seer\_files.txt} for all signal files is ``sig", this setting should be set to ``sig". All such files will be plotted differently, as discussed above.\\
\texttt{(file\_type)} & This tells Seer the file extension for the produced histogram. Common options are ``pdf", ``eps" and ``png". Possible options are the same as the types of files known to ROOT.\\
\hline
\end{tabularx}
\end{table}
\clearpage
\subsection{{\fontfamily{phv}\selectfont seer\_fakes.txt}\label{sec:fakes}}
For event simulators, fake rates are typically handled internally. In Delphes, there is a probability that a jet will fake an electron, for example. Fake backgrounds can have event rates similar to that of rare signals if there is a combination of a very large background with a very small fake rate. Estimating the cross section that overlaps with the signal can be very challenging using events from the detector simulator; if the fake rate is 1/1000, then out of 50000 events generated, approximately 50 events will be useful. Generating enough statistics thus takes a long time.
To improve on this, Seer has a fake-rate multiplier implemented. In effect, it allows for the possibility of using a much larger event fraction while simulating the effect of the fake rate as a re-scaling of the overall cross section. For example, if a user is interested in the contribution of $W+$jets to $e^+e^-$ signals, and has generated events of $W+nj$, the cross section can be rescaled by the fake rate, and one of the jets can be randomly selected to be added to the electron lists.
{\bf Note: This can over-estimate the fake background if the fake rate is included in both the detector simulator and Seer. Thus, if Seer is going to handle fake rates, it is required that the fake rate be turned off in the detector simulator.}
Control over this process is handled by the {\fontfamily{phv}\selectfont seer\_fakes.txt} file. This file is structured as follows, where all setting entries are enclosed in parentheses:
\begin{verbatim}
Rates
(rstate) (fstate) (efficiency)
Schema
scheme 0
scheme 1
addreal (nlow) (nhigh) (rstate), pt (ptlow) (pthigh), eta (etalow) (etahigh)
addfake (rstate) (fstate), pt (ptlow) (pthigh), eta (etalow) (etahigh)
scheme 2
scheme 3
scheme 4
scheme 5
scheme 6
scheme 7
scheme 8
scheme 9
\end{verbatim}
There are 10 total fake schemes that can be implemented. It is recommended that scheme 0 be left empty, as a ``no fakes" option. Note that the scheme number is the same number that should accompany the file in {\fontfamily{phv}\selectfont seer\_files.txt}. These setting entries are defined in Table \ref{tab:fake}. There can only be one fake rate defined (under Rates) for each combination of \texttt{(rstate)} and \texttt{(fstate)}, but there can be multiple entries of the same combination of \texttt{(rstate)} and \texttt{(fstate)} as an addfake.
\begin{table}
\caption{Fake rate setting definitions and explanations.\label{tab:fake}}
\noindent\begin{tabularx}{\textwidth}{|>{\bfseries}lX|}
\hline
Parameter & Explanation\\
\hline
\texttt{(rstate)} & This is a real particle state from the list of lep, ele, muo, bjt, gam, tau, ljt.\\
\texttt{(fstate)} & This is a fake particle state, where the possibilities are limited depending on the corresponding rstate. Real light jets (ljt) can fake ele, tau, bjt and gam. Real photons (gam) can fake ele. Real electrons (ele) can fake photons.\\
\texttt{(efficiency)} & This is the probability of the rstate faking the fstate.\\
\texttt{(nlow)} & This is the minimum number of a real particle\\
\texttt{(nhigh)} & This is the maximum number of a real particle\\
\texttt{(ptlow)} & This is the minimum $p_T$ of a real particle\\
\texttt{(pthigh)} & This is the maximum $p_T$ of a real particle\\
\texttt{(etalow)} & This is the minimum $\eta$ of a real particle (etalow = -etamax for symmetric)\\
\texttt{(etahigh)} & This is the maximum $\eta$ of a real particle\\
\hline
\end{tabularx}
\end{table}
The ``addreal" entries list the requirements needed for Seer to consider the faking of an event. In the example of $W+nj$ faking $e^+e^-$, the assumption is that one electron comes from the decay of the $W$ and the other one is a faked light jet. Thus, the user would set \texttt{addreal 1 1 ele, pt 10 99999, eta -2.5 2.5}, which will look for any event that already has exactly one electron (if this is the only addreal, then the other states are inclusive) within the specified kinematic range. If one is found, then Seer will pass the event along to determine the fake particle.
The ``addfake" entries determines how Seer fakes the state. First, Seer will look among all of the \texttt{(rstate)} particles in the event and identify which ones meet the kinematics listed. It will then randomly select a particle from that list to turn into the \texttt{(fstate)} type, removing the particle from the appropriate lists and adding it to the fake state list(s). In the example of $W+nj$ faking $e^+e^-$, the user would set \texttt{addfake ljt ele, pt 10 99999, eta -2.5 2.5}, which will take one of the light jets from within that kinematic range and turn it into an electron. In the case of an electron, the charge of the electron is determined randomly, as well. Thus, it is possible that a fake electron from $W+nj$ will result in either same sign electrons or opposite sign electrons.
It is important to note the event weight for {\bf all} events within the file, whether a faked particle is present in the event or not, is scaled by the product of all the fake rate efficiencies for each ``addfake" entry in the scheme being used for that file. Continuing with the same example, if the fake rate of jets as electrons is 0.0001, then any file that uses that scheme will rescale the total cross section by 0.0001. Thus, any instances of $e^+e^-$ within those files will have their weight rescaled. This is why processes that fake the signal need to be separated from processes that produce the signal when using the Seer faking system, and why fake rates need to be turned off at the detector simulator level.
\subsection{{\fontfamily{phv}\selectfont seer\_settings.txt}\label{sec:settings}}
There are a number of constants that Seer uses for various calculations. The numerical value of these settings can be adjusted in the {\fontfamily{phv}\selectfont seer\_settings.txt} file. The format for this file is:
\begin{verbatim}
mlsp = (mass_lsp)
wmas = (W_mass)
wwid = (W_width)
zmas = (Z_mass)
zwid = (Z_width)
tmas = (t_mass)
twid = (t_width)
jeer = (jE_smear)
jmer = (jM_smear)
jher = (jh_smear)
metA = (MET_scale)
metP = (MET_pileup)
metM = (MET_minbias)
\end{verbatim}
Further information about these settings, as well as the default values for them, is listed in Table \ref{tab:settings}.
\begin{table}
\caption{Explanations and default values for the \label{tab:settings}}
\noindent\begin{tabularx}{\textwidth}{|>{\bfseries}llX|}
\hline
Parameter & Default & Explanation\\
\hline
\texttt{(mass\_lsp)} & 0 & sets the value of the mlsp variable (mass of the LSP setting used for MT2 calculations)\\
\texttt{(W\_mass)} & 80.385 & sets the value of the wmass variable (currently not used)\\
\texttt{(W\_width)} & 2.085 & sets the value of the wwidth variable (currently not used)\\
\texttt{(Z\_mass)} & 91.1876 & sets the value of the zmass variable (currently not used)\\
\texttt{(Z\_width)} & 2.4952 & sets the value of the zwidth variable (currently not used)\\
\texttt{(t\_mass)} & 173.21 & sets the value of the tmass variable (currently not used)\\
\texttt{(t\_width)} & 2.00 & sets the value of the twidth variable (currently not used)\\
\texttt{(jE\_smear)} & 0.00 & sets the value of the jet\_E\_error variable (used to add additional $p_T$ smearing)\\
\texttt{(jM\_smear)} & 0.00 & sets the value of the jet\_m\_error variable (used to add additional jet mass smearing)\\
\texttt{(jh\_smear)} & 0.00 & sets the value of the jet\_eta\_error variable (used to add additional $\eta/\phi$ smearing)\\
\texttt{(MET\_scale)} & 0.75 & sets the value of the met\_scale variable (used to scale MET smearing)\\
\texttt{(MET\_pileup)} & 30 & sets the value of the met\_pileup variable (average \# of pileup events for MET smearing)\\
\texttt{(MET\_minbias)} & 20 & sets the value of the met\_min\_bias variable in GeV (minimum bias energy for MET smearing)\\
\hline
\end{tabularx}
\end{table}
The $M_{T2}$ calculator included with Seer requires an initial guess of the mass of the LSP particle, which is set in this file. The $W$, $Z$, and $t$ mass settings are currently not used within any calculations. However, their inputs may be useful for any tagging or kinematics that depend on this values. An example where this may be useful would be the cuts employed in the ATLAS tri-lepton search for supersymmetry, where the phase space is separated into bins based on the dilepton invariant mass being smaller, larger or within a window about the $Z$ pole mass.
Additionally, Seer allows for the possibility of adding greater smearing to the jets and missing transverse energy values. All smearing values are calculated using a Gaussian distributed random value centred at zero with widths dependent on the type of jet being smeared. For hadronic jets (applies to light jets, $b$ jets and $\tau_h$ jets), additional jet $p_T$ will be smeared with a width of $\sigma_E=(jE\_smear)\times p_T^j$, jet mass will be smeared with a width $\sigma_m = (jM\_smear)\times m_j$ (useful particularly for large $R$ fat jets), and the $\eta$ and $\phi$ values will be independently smeared with a width $\sigma_R = (jh\_smear)\times \eta(\phi)$ (where \texttt{(jE\_smear)}, \texttt{(jM\_smear)} and \texttt{(jh\_smear)} are the relevant quantities from {\fontfamily{phv}\selectfont seer\_settings.txt}). For missing transverse energy, smearing is performed independently on the $x$ and $y$ components with a width given by the formula from \cite{ATL-PHYS-PUB-2013-004}:
\begin{equation}
\sigma_{\not\!E_T} = A(0.40+0.09\sqrt{P})\sqrt{\sum E_T +P\times M}
\end{equation}
where $A$ is an overall scaling factor given in Seer by the variable \texttt{metA} (used to account for existing $\ensuremath{\not\!\!E_T}\xspace$ smearing in the detector analysis), $P$ is the average number of pileup events given in Seer by the variable \texttt{metP}, and $M$ is a minimum bias energy given in Seer by the variable \texttt{metM}. The sum of visible transverse energies should be in GeV, as should the value of $M$. The numeric factors are fitted values from \cite{ATL-PHYS-PUB-2013-004}.
\subsection{Command Line Input/Output}
Seer is run using the command {\fontfamily{phv}\selectfont ./seer.exe} in the main folder. The seer terminal output provides most of the same information as is in the text file output with a few minor differences. The differences are:
\begin{itemize}
\item The descriptions of the settings files are more descriptive, rather than an identical copy of the lines in the file.
\item Individual file results are not listed, only group summaries and the overall summary is listed.
\item The histogram information is not listed.
\end{itemize}
The information in the text output file is meant to be useful for recreating the analysis, while the terminal output is meant to be a more descriptive explanation and summary of the analysis.
There are three possible arguments that can be input (individually or in any combination) which will force Seer to run with extra text output to the terminal (not to the output text file). These three arguments represent three modes:
\begin{itemize}
\item debugmode - used to print out information about nearly every step of the analysis process. This is useful to use when Seer crashed mid-run to determine where the problem lies.
\item rejectmode - used to print out a detailed output of all particles within an event {\em only} when that event is rejected. This is useful for analyzing each event that is cut to determine if the cuts are working as intended.
\item verbosemode - turns on output of the individual file summaries into the terminal output, and output of an initial verification of the cross section and event counts in the LHCO files.
\end{itemize}
Note that the debugmode and rejectmode modes produce an excessive amount of text, and it is recommended that they be used only in conjunction with a test LHCO file that contains around 10 events. Creating such an LHCO file is easy and does not require modifying the cross section or event count entries within the file. Such a test file can be created by making a copy of any of the signal or background files being examined and highlighting and deleting all but the first 10 events. (Note: deleting the extra carriage return at the end of the LHCO file can result in problems within Seer.)
\section{Modifying Seer\label{sec:mod}}
There are two aspects of Seer that a user may wish to modify. The first is user-defined cuts, and the second is user-defined plot variables. Both of these require an understanding of how the information is stored in Seer, and then an understanding of how that information will be interpreted and used within Seer. This will be discussed in this section.
\subsection{Accessing information in {\fontfamily{phv}\selectfont calculations.cpp}}
Seer stores the information for each particle within a series of vectors of the TLorentzVector class in ROOT. This information is stored only for the current event, and is cleared upon reading in a new event. The following vectors of particles are available:\\
\begin{lstlisting}
std::vector<TLorentzVector> v_gam; // photons
std::vector<TLorentzVector> v_jet; // all jets (includes b-tagged and hadronic tau jets)
std::vector<TLorentzVector> v_bjt; // only b-tagged jets
std::vector<TLorentzVector> v_ljt; // only non-b-tagged and non-hadronic tau jets
std::vector<TLorentzVector> v_tau; // only hadronic taus
std::vector<TLorentzVector> v_lep; // both electrons and muons
std::vector<TLorentzVector> v_lne; // only negatively charged leptons (both electrons and muons)
std::vector<TLorentzVector> v_lpo; // only positively charged leptons (both electrons and muons)
std::vector<TLorentzVector> v_ele; // all electrons
std::vector<TLorentzVector> v_ene; // only negatively charged electrons
std::vector<TLorentzVector> v_epo; // only positively charged electrons
std::vector<int> v_ele_charge; // the charge of electrons in v_ele, in the same order as v_ele
std::vector<TLorentzVector> v_muo; // all muons
std::vector<TLorentzVector> v_mne; // only negatively charged muons
std::vector<TLorentzVector> v_mpo; // only positively charged muons
std::vector<int> v_muo_charge; // the charge of muons in v_muo, in the same order as v_muo
std::vector<TLorentzVector> v_met; // missing transverse energy
std::vector<TLorentzVector> v_all; // all particles, except missing tranverse energy
std::vector<int> v_lep_charge; // the charge of leptons in v_lep, in the same order as v_lep
\end{lstlisting}
There are six types of particle entries in LHCO files, based on the number in the second column for the particle listing (referred to as tag). These options are 0 (photon), 1 (electron), 2 (muon), 3 (hadronic tau), 4 (jet), and 6 (missing transverse energy). Furthermore, the sign of the entry for number of tracks indicates the charge for electrons and muons, and the b-tag field indicates whether a jet is a b-tagged jet. This is the information that is used to determine which vectors to add the particle entries. These vectors are {\em not} exclusive, such that almost every particle entry exists in multiple lists. The reason for this methodology is to make modifying and adding new calculations more easy for users, at the expense of being more memory intensive.
Users have full access to the TLorentzVector functions for each of the vectors of these objects. The list of functions is available on \url{https://root.cern.ch/root/html/TLorentzVector.html}.
Note: If the user defines their own vector class object, this should be cleared within the {\small\texttt{event\_destroy()}} function in {\fontfamily{phv}\selectfont calculations.cpp}.
\subsection{User defined kinematic variables}
The following steps will walk users through the process of adding code to calculate a new kinematic variable to Seer.
\begin{enumerate}
\item[Step 1:] The function needs to be defined within {\fontfamily{phv}\selectfont calculations.h}. For an example, a calculation of the leading dijet plus leading lepton invariant mass is discussed through the rest of these steps. The function could be defined with the code:
\begin{lstlisting}
virtual double calc_Mjjl();
\end{lstlisting}
\item[Step 2:] Following this, the function needs to be defined within {\fontfamily{phv}\selectfont calculations.cpp} after the text ``This area is for user defined functions." Templates are provided in the code to understand the format. Continuing with the example, the code could be:
\begin{lstlisting}
double event::calc_Mjjl() {
if ( v_jet.size() >= 2 && v_lep.size() >=1 ) {
return (v_jet.at(0)+v_jet.at(1)+v_lep.at(0)).M();
}
return -99999;
}
\end{lstlisting}
In this example, a few important things should be noted:
\begin{enumerate}
\item The function must first check that the jet and lepton vectors have the appropriate size or the user will have errors during calculation that will stop Seer from functioning.
\item The particle vectors are ordered from largest to smallest $p_T$, and {\small\texttt{v\_jet.at(0)}} and {\small\texttt{v\_jet.at(1)}} are the leading and subleading jet. For other possible combinations of jets from a larger list of jets, the user will need to define boolean conditions to select the desired jets from the list.
\item This example does not distinguish between b-tagged jets, hadronic tau jets, and light jets. To use only light jets, the user could instead use {\small\texttt{v\_ljt}} instead of {\small\texttt{v\_jet}}.
\item Returning a value of 0 if there are insufficient jets and leptons to calculate a mass can result in the 0th bin being incorrectly filled with events that should not be plotted. Using a value that is unreasonable for the plotting scale will ensure that such events do not appear on the histogram.
\end{enumerate}
\item[Step 3:] To add plotting capabilities, allowing this new function as a plotting variable, two changes must be made. The first is to add the function to the list of variables output by the function {\small\texttt{calc\_ps(int hist\_type)}} within {\fontfamily{phv}\selectfont calculations.cpp}. The function needs to be added to the end of the list of else-if statements with an incremented reference number (see Section \ref{sec:plots} for more details on the reference number, \texttt{(plotnum)}). Continuing with the example, the code could look like:
\begin{lstlisting}
else if ( hist_type == 49 ) {
event_ps = calc_Mjjl();
}
\end{lstlisting}
\item[Step 4:] The second change to allow plotting requires modifying the {\fontfamily{phv}\selectfont plots.cpp} file. The area to modify, along with a template, is provided following the text string ``USER AREA." Using the same \texttt{(plotnum)} value from {\small\texttt{calc\_ps()}}, the {\small\texttt{name\_text}}, {\small\texttt{title\_text}}, {\small\texttt{yaxis\_text}} and {\small\texttt{xaxis\_text}} strings must be modified as needed. The {\small\texttt{name\_text}} variable is a string that will be added to the filename of the output files (see Section \ref{sec:plots} for more information). The {\small\texttt{title\_text}} variable is a string that will nominally be added as a title to the ROOT canvas, however this is currently disabled in the style file, {\fontfamily{phv}\selectfont style.h}. The {\small\texttt{yaxis\_text}} and {\small\texttt{xaxis\_text}} are ROOT-LaTeX strings that will become the labels on the y and x axes, respectively. Currently, the $y$-axis in Seer output plots is only cross section values in units of fb, so the {\small\texttt{yaxis\_text}} variable does not need to be changed. Continuing with the example, this code could look like:
\begin{lstlisting}
else if ( plot_type == 49 ) {
name_text = "mjjl";
title_text = "Dijet+Lepton Mass";
yaxis_text = "#sigma (fb)";
xaxis_text = "M_{jjl} (GeV)";
}
\end{lstlisting}
\end{enumerate}
At this point, all necessary changes have been made to create plots of with this variable. The number 49 is now an accessible \texttt{(plotnum)} in {\fontfamily{phv}\selectfont seer\_plots.txt}. This kinematic variable can also be implemented for use as a cut calculation, but this is a more nuanced issue that will be discussed in the next section.
\subsection{User defined cuts}
User defined cuts can be implemented using two methods. The first is using a boolean function that cuts based on some (complex) criteria. The second is by using a kinematic function. The key to implementing cuts is in setting {\bf both} the {\small\texttt{cuttest}} and {\small\texttt{intcuttest}} variables to be true if the condition is met. The logic used in Seer is that if {\small\texttt{cuttest}} is TRUE, the event will {\bf not} be included in the analysis, and it will not be passed to the plotting function.
The variable {\small\texttt{intcuttest}} is reset for each run through the different cuts implemented, distinguishing which cut settings eliminate the event (since an event may be cut by multiple cut conditions). Alternatively, {\small\texttt{cuttest}} is not reset through each cut, and so it simply represents an overall test whether the event is cut. This is why both variables must be set to be TRUE.
The following steps outline the procedure to define a boolean function to Seer.
\begin{enumerate}
\item[Step 1:] The function needs to be defined within {\fontfamily{phv}\selectfont calculations.h}. The naming scheme for boolean functions used for cuts is a prefix of {\small\texttt{cut\_}}. As an example, a function that cuts based on number of jets with a complex combination of $p_T$ thresholds is discussed through the rest of these steps. The function could be defined with the code:
\begin{lstlisting}
virtual bool cut_jets();
\end{lstlisting}
Note that this function does not make use of the \texttt{(kinlow)} and \texttt{(kinhigh)} variables from {\fontfamily{phv}\selectfont seer\_cuts.txt}, and is therefore a hard coded function. Users may wish to access information from \texttt{(kinlow)} and \texttt{(kinhigh)} for a more nuanced cut. To do this, the user could define the function as:
\begin{lstlisting}
virtual bool cut_jets(double kin_low, double kin_high);
\end{lstlisting}
If the user expects integer input for different cases, based only on the value of \texttt{(kinlow)}, the function could be defined as:
\begin{lstlisting}
virtual bool cut_jets(int kin_low);
\end{lstlisting}
Note that this would require recasting the double valued {\small\texttt{kin\_low}} variable as an integer at the point of use of the function. The following steps assume that the second of these three function definitions is implemented.
\item[Step 2:] Following this, the function needs to be defined within {\fontfamily{phv}\selectfont calculations.cpp} after the text ``This area is for user defined functions." Templates are provided in the code to understand the format. Continuing with the example, the code could be:
\begin{lstlisting}
bool event::cut_jets(double kin_low, double kin_high) {
if ( v_jets.size() >= 4 && v_tau.size() == 0 ) {
if ( v_jet.at(0).Pt() > kin_high && v_jet.at(1).Pt() > kin_low && v_jet.at(3).Pt() > 50 ) {
return true;
}
}
return false;
}
\end{lstlisting}
For this code, a few important things should be noted:
\begin{enumerate}
\item This function requires more than 4 jets with 0 tau jets, thus it is inclusive to b-tagged jets and light jets, but not hadronic taus.
\item This function requires that the leading jet has a $p_T$ larger than the value of kin\_high, and a subleading jet with a $p_T$ larger than the value of {\small\texttt{kin\_low}}.
\item This function requires that the fourth jet ({\small\texttt{v\_jet.at(3)}}) has a $p_T>50$~GeV. Since the jet list is $p_T$ ordered, this also means that the third jet ({\small\texttt{v\_jet.at(2)}}) will have a $p_T>50$~GeV.
\end{enumerate}
\item[Step 3:] Following the definition of the function, the cut must be implemented within the {\small\texttt{run\_cut()}} function in {\fontfamily{phv}\selectfont calculations.cpp}. Templates have been provided following the text string ``This area is for user defined cuts." Adding the cut first requires defining a unique string that will be used to identify the cuts in {\fontfamily{phv}\selectfont seer\_cuts.txt}, which will be tested against the variable {\small\texttt{kin\_type}}. The second issue requires understanding exactly the condition of the function which will result in cuts. This will be more clear in the context of continuing the example, which can be implemented with the code:
\begin{lstlisting}
if ( kin_type == "myjet" ) {
if ( !cut_jets(kin_low,kin_high) ) {
if ( debugmode ) {
std::cout << "Event rejected because of a myjet cut" << std::endl;
}
cuttest = true;
intcuttest = true;
}
}
\end{lstlisting}
A few important things should be noted:
\begin{enumerate}
\item This script sets {\small\texttt{cuttest}} to TRUE if the return of the {\small\texttt{cut\_jets()}} function is FALSE. This is an important distinction. Thus, events that don't have jets falling within the kinematic constraints defined in the function will be removed. Close attention to the logic is necessary to properly code these cuts. When {\small\texttt{cuttest}} is TRUE, the event is removed.
\item The {\small\texttt{debugmode}} feature has been implemented to address the complexity of these cuts. Running in {\small\texttt{debugmode}} on a reduced sample of events ($\sim10$ events is a good number) is useful for testing the behaviour.
\item This is not optimized coding, written in order to make understanding the behaviour easier.
\end{enumerate}
\end{enumerate}
The procedure for performing a cut on user defined variables is very similar. Steps 1 and 2 from the previous section need to be followed to create the kinematic calculation function. Following this, Step 3 from this section outlines the basic procedure to implement the cut. There is a template provided within the function {\small\texttt{run\_cut()}} in {\fontfamily{phv}\selectfont calculations.cpp}. For the example of the dijet plus lepton invariant mass function defined in the previous section, the following code would implement a cut based on the {\small\texttt{kin\_low}} and {\small\texttt{kin\_high}} values:
\begin{lstlisting}
if ( kin_type == "JJL" ) {
if ( v_jets.size() >= 2 && v_lep.size() >= 1 ) {
testvalue = calc_Mjjl();
if ( testvalue > kin_low && testvalue < kin_high ) {
if ( debugmode ) {
std::cout << "Event rejected because of a JJL cut of any particles with " << testvalue << std::endl;
}
cuttest = true;
intcuttest = true;
}
}
else { // This else statement will result in removal of all events that do not have at least 2 jets and at least 1 lepton.
cuttest = true;
intcuttest = true;
}
}
\end{lstlisting}
In this code, the calculated kinematic variable is passed to the {\small\texttt{testvalue}} variable. This is to save on computation time, which is especially important for more complex codes like the stransverse mass, {\small\texttt{calc\_MT2()}, and razor, {\small\texttt{calc\_MR()}, functions, which would add significant computing time to compute once for each of the comparisons to kin\_low and kin\_high. Also note that the dijet plus lepton mass can only be calculated on events with at least two jets and at least one lepton, so an extra logical test has been added to cut events that don't meet these minimum requirements.
In either of the two cases presented, adding these functions in the region outlined opens up the ability to add addcut lines to the {\fontfamily{phv}\selectfont seer\_cuts.txt} file such as:
\begin{verbatim}
addcut all with myjet between 300 400
addcut all with JJL between 0 500
\end{verbatim}
Based on the definition of the myjet cut, the first line would remove event if the there were fewer than 4 jets, the leading jet had a $p_T < 400$~GeV, the subleading jet had a $p_T<300$~GeV, or if the third and fourth leading jets had $p_T < 50$~GeV. Any or multiple of those conditions would result in removal of the event from the analysis. The JJL cut is more straightforward, as it will remove events in which the dijet plus lepton invariant mass is between 0 and 500~GeV.
\subsection{Other modifications to Seer}
Users may wish to make other modifications to Seer. A few of these changes will be addressed here.
\subsubsection{Changes to histogram colour schemes}
Currently, Seer checks the number of histograms to plot and chooses from a few spectral choices. For a different colour scheme, users can alter the settings within the constructor of the {\small\texttt{plot}} class within {\fontfamily{phv}\selectfont plots.cpp}. The colours available in ROOT are listed at \url{https://root.cern.ch/root/html/TColorWheel.html}. The three different colour scenarios are for 1-9 histograms, 9-15 histograms, and more than 15 histograms. It is not recommended that users plot more than 15 histograms, however, except in very specialized situations, as the lines plots will become increasingly busy and difficult to read. However, Seer can effectively handle up to 50 overlapping histograms.
\subsubsection{Changing canvas/axis settings}
The canvas and axis for the output histogram is set at the end of the constructor of the {\small\texttt{plot}} class within {\fontfamily{phv}\selectfont plots.cpp}. This area is labeled by a comment string stating ``Canvas/Axis settings". This area controls the offset and size of the axis labels, as well as the axis tick settings. Some default settings also exist within the {\fontfamily{phv}\selectfont style.h} file. Modifications to {\fontfamily{phv}\selectfont style.h} should be handled by users who are very experienced with ROOT. See the ROOT documentation on \url{root.cern.ch} for more information on each of these settings.
\subsubsection{Changing the legend}
The upper right corner of the legend defaults to the upper-right-most corner of the canvas, and has a width of 19\% of the canvas. The height of the legend is the larger value of 20\% or $3.5\% \times n_{hist}$, where $n_{hist}$ is the number of histograms included. Changes to the legend location and size can be made at the start of the {\small\texttt{generate\_plot()}} function within the {\small\texttt{plot}} class in {\fontfamily{phv}\selectfont plots.cpp}.
\section{Examples}
\subsection{Analysis Prototyping and Data Visualization}
One of Seer's primary functions is for analysis prototyping and data visualization, where kinematic distributions of signals and backgrounds can be explored with minimal effort. By adjusting a few settings files, many different tagging permutations and cuts can be implemented and visualized, producing publication quality figures.
In a recent study of light pseudoscalar production at LHC13 \cite{Kozaczuk:2015bea}, Seer was used to analyze SM and beyond the SM production of dilepton events in association with $b$ jets generated with MG5+aMC@NLO. A total of 40.7 million events were generated, and heavy states ($W$, $Z$, $a$, $t$) were decayed democratically over all lepton generations ($e$, $\mu$, $\tau$). Intermediate figures from this study are included in Figure \ref{fig:mumu}, which shows di-muon event production at the LHC with at least 1 $b$-tagged jet, and inclusive to light jets with $p_T<40$~GeV. Backgrounds of $\gamma^*/Z + 0-2 b + 0-2j$, $W^+W^- + 0-2b + 0-2j$, $W+0-2b+0-2j$ and $WZ+0-2b+0-2j$ are included, as well as for a BSM pseudoscalar with an enhanced coupling to down-type fermions over SM Higgs couplings, for masses of the pseudoscalar $m_a = 50,80$~GeV. Dimuon decays arise from direct dimuon production, but also as decay products from $\tau^+\tau^-$ production.
\begin{figure*}[tb]
\centering
\mbox{\includegraphics[width=0.45\textwidth]{ht_1u1u_0c_n27206616.pdf}
\includegraphics[width=0.45\textwidth]{lptmax_1u1u_0c_n27206616.pdf}}\\
\mbox{
\includegraphics[width=0.45\textwidth]{ptlmet_1u1u_0c_n27206616.pdf}
}
\caption{Kinematic distributions of the signal and background for pseudoscalar mediated dilepton production, including backgrounds, with a $\mu^+\mu^-$ signature, with no cuts made. Upper left figure is the total scalar sum of visible transverse momentum, upper right figure is the leading lepton $p_T$, and lower figure is the scalar sum of the lepton $p_T$ and the $\ensuremath{\not\!\!E_T}\xspace$. Distributions show distinctive differences in the distributions between signal and background.\label{fig:mumu}}
\end{figure*}
To illustrate how Seer was used for this project, the contents of the settings files are listed below. The contents of the {\fontfamily{phv}\selectfont seer\_files.txt} file was:
\begin{verbatim}
## Single W plus 0-2 bjets plus light jets
W+0-2b+jets d 1 ./lhcofiles/bkgs/wmbb/run_01.lhco 5.81426 41257
...
W+0-2b+jets d 1 ./lhcofiles/bkgs/wpbb/run_01.lhco 9.76014 40566
...
## W+Z plus 0-2 bjets
WZ+0-2b+jets d 0 ./lhcofiles/bkgs/wmzbb/run_01.lhco 0.000426996 50000
...
WZ+0-2b+jets d 0 ./lhcofiles/bkgs/wpzbb/run_01.lhco 0.000693746 50000
...
## WW plus 0-2 bjets, includes ttbar production
WW+0-2b+jets d 0 ./lhcofiles/bkgs/wwbb/run_01.lhco 3.8891 50000
...
## gamma*/Z plus 0-2 bjets plus jets
Z+0-2b+jets d 0 ./lhcofiles/bkgs/llbb/run_01.lhco 1.35258 32926
...
## Signal files.
MA=80,gd=25 d 0 ./lhcofiles/signal/set_80_5_1.lhco 0.923511 50000
...
\end{verbatim}
The ellipses refer to multiple entries of similar files. For the backgrounds, events were broken down into subsets of diagrams, which necessitates overriding the automatic event weight calculator in Seer and adding the cross section (in pb) and number of events at the end of each line (note the use of the \texttt{d} flag when overriding the automatic event weight calculator). Backgrounds involving a single $W$ decaying to leptons do not meet the dilepton criteria, and so a light jet must fake either an electron or a tau lepton to contribute to the signal. This was calculated by using Seer to estimate fakes instead of Delphes.
Dimuon events do not suffer from large fake lepton backgrounds, but $b$-jet faking was a significant issue. For the $e\mu$ analysis, however, jets were considered to fake electrons, and so the single $W$+jets channel could produce events that faked the signal. In this case, the flag of \texttt{1} for the single $W$ backgrounds shown above corresponded to scheme 1 of {\fontfamily{phv}\selectfont seer\_fakes.txt}, which contained:
\begin{verbatim}
Rates
jet ele 0.0001
jet tau 0.001
jet bjt 0.001
Schema
scheme 0
scheme 1
addreal 1 1 muo, pt 10 99999, eta -2.5 2.5
addreal 1 2 bjt, pt 10 99999, eta -2.5 2.5
addfake jet ele, pt 10 99999, eta -2.5 2.5
scheme 2
...
\end{verbatim}
This is the entry for the $e\mu$ analysis, where the muon is expected from the decay of the $W$, and the analysis requires that there is at least 1 $b$-jet. For events without any $b$-jets, a second entry of \texttt{addfake jet bjt, pt 10 99999, eta -2.5 2.5} can be added to the same scheme. For each event that meets that criteria, these settings instruct Seer to randomly select a jet with $p_T > 10$~GeV and $|\eta| < 2.5$ and switch its type to an electron. The fake rate for this is $\not\!\epsilon_j^e = 0.0001$, and this factor is then applied to all such events. Not included here for brevity is a similar inclusion of dilepton events with zero $b$-jets, in which a light jet fakes a $b$-jet.
The {\fontfamily{phv}\selectfont seer\_cuts.txt} file for the dimuon analysis was as follows:
\begin{verbatim}
## New Cut Card - double number signs at the start of the line indicate a comment
## Note: B-jets and hadronic taus are also considered jets unless they are explicitly tagged/removed!
Signal ## Do not alter this line!
2 2 muo, pt 10 99999, eta -2.5 2.5
1 2 bjt, pt 10 99999, eta -2.5 2.5
Extra ## Do not alter this line!
##THIS IS CMS TRIGGER
trigger ele pt 35 0 0 0
trigger ele pt 23 12 0 0
trigger muo pt 25 0 0 0
trigger muo pt 17 10 0 0
trigger gam pt 80 0 0 0
trigger gam pt 40 25 0 0
trigger tau pt 59 59 0 0
trigger jet pt 657 0 0 0
trigger jet pt 247 247 247 0
trigger jet pt 113 113 113 113
trigger tae pt 45 19 0 0
trigger tam pt 40 15 0 0
trigger jat pt 180 123 0 0
trigger tat pt 86 65 0 0
trigger bjt pt 237 0 0 0
trigger eam pt 23 10 0 0
trigger eam pt 12 23 0 0
##IGNORES
ignore ele with pt between 0 10
ignore ele with eta over 2.47
ignore muo with pt between 0 10
ignore muo with eta over 2.4
ignore gam with pt between 0 10
ignore gam with eta over 2.5
ignore jet with pt between 0 40
ignore jet with eta over 2.5
ignore bjt with pt between 0 10
ignore bjt with eta over 2.5
##CUTS
addcut all with osl over 0
##Below are the cuts for the 2u case
addcut lep with pt over 50
addcut all with ht over 120
addcut all with htl over 120
\end{verbatim}
The signal tagging required exactly two muons and either one or two $b$-jets. An estimate of the CMS trigger thresholds was included, and detector limitations for electrons, muons and jets were implemented in the \texttt{ignore} section.
The cuts included in this list are: an opposite sign lepton cut (eliminating same-sign lepton pairs); a $p_T<50$~GeV on the leading lepton cut (events with $p_T>50$~GeV were removed); a total scalar sum of $p_T$ of visible objects of $H_T<120$~GeV cut; and a scalar sum of the missing transverse momentum and $p_T$ of the leptons cut of $\not\!\!H_T^\ell < 120$.
The {\fontfamily{phv}\selectfont seer\_plots.txt} file for Figure \ref{fig:mumufin} included the settings:
\begin{verbatim}
PlotType = 35
PlotType2 = 0
PlotMaxX = 150
PlotMinX = 0
PlotMaxY = 10000
PlotMinY = 0.0001
NumBinX = 150
NumBinY = 30
ChooseLn = 1
Normalize = 0
PlotTxt = 1u1u1
SigPref = MA
FileType = pdf
\end{verbatim}
Setting \texttt{PlotType2} to 0 creates a single variable histogram, and the \texttt{NumBinX} value sets the bin width to be 1~GeV. Based on the \texttt{SigPref} entry, all entries in {\fontfamily{phv}\selectfont seer\_files.txt} that have a tag starting with ``MA" are treated as signal files and plotted as outlines, while all others are treated as backgrounds. Based on the \texttt{FileType} entry, the histogram was output as a portable document format (pdf) file.
\begin{figure*}[tb]
\centering
\mbox{\includegraphics[width=0.45\textwidth,clip]{llmass_1u1u_ac_n40707579.pdf}}
\caption{Invariant mass distribution of pseudoscalar mediated $\mu^+\mu^-$ production, as well as SM backgrounds, with cuts made to enhance signal over the background.\label{fig:mumufin}}
\end{figure*}
With the small bin sizes in Figure \ref{fig:mumufin}, the event rates per bin show a significant statistical fluctuation in bins where few backgrounds passed the tagging and cut requirements. This motivated the need for generating 40.7 million events, in order to get an accurate estimate of the backgrounds in each bin. For the ROOT file format, the file size for 50k events is larger than 500~MB. Performing this same analysis using information stored in ROOT files would require more than 400~GB of disk space. The time taken for the generation of these plots was approximately 20 minutes (MacBook Pro 2.66GHz i7, 8GB 1066 DDR3 RAM, 5400 RPM HD).
\begin{figure*}[tb]
\centering
\mbox{
\includegraphics[width=0.45\textwidth]{mll_v_lptmax_1e1u_n27088452.pdf}}\\
\mbox{
\includegraphics[width=0.45\textwidth]{mll_v_lptmax_1e1u_n24404624.pdf}
\includegraphics[width=0.45\textwidth]{mll_v_lptmax_1e1u_n2683828.pdf}}
\caption{Examples of the types of 2D histograms that can be generated by Seer. The box histogram (top) is produced when multiple tags are included in the {\fontfamily{phv}\selectfont seer\_files.txt} list. Box sizes represent the relative bin value. The coloured spectral histograms (lower) are produced when only a single tag is listed in the {\fontfamily{phv}\selectfont seer\_files.txt} list. Without setting Seer to produce normalized cross sections, the $z$-axis values are $\sigma$ values measured in fb.\label{fig:2d}}
\end{figure*}
Seer also allows for alternative ways of visualizing the events. The plots in Figure \ref{fig:2d} were produced by making a few changes to the settings files in Seer, but using the same input files. The signal is switched to the $e+\mu$ channel, which does not have the large invariant mass peak, by adjusting the appropriate signal lines in the {\fontfamily{phv}\selectfont seer\_cuts.txt} file, shown below:
\begin{verbatim}
Signal ## Do not alter this line!
1 1 ele, pt 10 99999, eta -2.5 2.5
1 1 muo, pt 10 99999, eta -2.5 2.5
1 2 bjt, pt 10 99999, eta -2.5 2.5
\end{verbatim}
Using this tagging, Figure \ref{fig:2d} was produced by combining all four background types into a single tag (bkg), and setting {\fontfamily{phv}\selectfont seer\_plots.txt} to:
\begin{verbatim}
PlotType = 35
PlotType2 = 5
PlotMaxX = 300
PlotMinX = 0
PlotMaxY = 300
PlotMinY = 0
NumBinX = 30
NumBinY = 30
ChooseLn = 1
Normalize = 0
PlotTxt = 1e1u
SigPref = MA
FileType = pdf
\end{verbatim}
Additionally, all of the backgrounds were given an identical tag in {\fontfamily{phv}\selectfont seer\_files.txt}, and only one signal set was included. The top of Figure \ref{fig:2d} was produced with both the signal and background entries in {\fontfamily{phv}\selectfont seer\_files.txt}, while the lower two were produced by commenting out the signal or background, respectively. This visual representation of the phase space is useful for identifying cut regions.
\subsection{Reproducing Experimental Analyses}
Seer can also be used to test MC signal data against existing experimental results. As with any such endeavour, the precise details of the detector simulation (such as in Delphes) and method for accounting for NLO effects can have a significant impact on the results. Included in this section are the results of using Seer to reproduce three experimental SUSY searches: 2-6 jet inclusive search \cite{TheATLAScollaboration:2013fha}, tri-lepton search \cite{ATLAS:2013rla}, di-lepton search \cite{TheATLAScollaboration:2013hha}. The routines to perform the cuts for each of these studies are implemented in Seer already in {\fontfamily{phv}\selectfont calculations.cpp}, allowing users to perform these same searches. In each case, the details of the Delphes settings used are not included, as this is outside the scope of Seer.
Another challenge in reproducing experimental searches is in accounting for combined exclusion limits. The experimental groups typically use the $CL_s$ method for determining exclusion constraints, which accounts for signal excesses from multiple signal regions, and accounts for correlations between systematic uncertainties present. The correlation matrix for the systematic uncertainties are not currently provided by any of the experimental groups, and so theorists are unable to perform this same analysis. In the studies reproduced here, exclusion contours were generated by assuming a boolean addition of the exclusions from each independent signal region. This naturally results in a more conservative contour, leading to a natural negative bias, especially near regions that are excluded by multiple signal regions simultaneously.
\subsubsection{ATLAS 2-6 jet inclusive search}
The 2-6 jet inclusive search \cite{TheATLAScollaboration:2013fha} that was reproduced involved the mSUGRA scan over $m_{1/2}$ vs $m_0$, with $\tan\beta = 30$, $A_0 = -2m_0$ and $\mu>0$. Contrary to the conference note, which used Herwig++ to generate events, this reproduction employed MG5+aMC@NLO to generate 50k events of the fully inclusive di-sparticle cross section (all possible pairings of sparticles) at leading order, without any additional hard element jets. These states were then decayed and hadronized using the standard Pythia 6.4 implementation with MG5+aMC@NLO, and Delphes 3.0 was employed to simulate detector effects. The parameters $m_0$ and $m_{1/2}$ were scanned over the values $m_0 \sim 700-5700$~GeV (step size 500 GeV) and $m_{1/2} \sim 300-900$~GeV (step size 100 GeV).
The search was reproduced using the following {\fontfamily{phv}\selectfont seer\_cuts.txt} settings:
\begin{verbatim}
Signal ## Do not alter this line!
2 99 jet, pt 10 99999, eta -5 5
Extra ## Do not alter this line!
trigger jat pt 80 100 0 0
##IGNORES
ignore ele with pt between 0 10
ignore ele with eta over 2.47
ignore muo with pt between 0 10
ignore muo with eta over 2.4
ignore gam with pt between 0 10
ignore gam with eta over 2.5
ignore bjt with pt between 0 20
ignore bjt with eta over 2.5
ignore ljt with pt between 0 20
ignore ljt with eta over 4.5
##CUTS
## A-loose
addcut all with jetA over 1
## A-medium
## addcut with jetA over 2
## B-medium
## addcut with jetB over 1
## B-tight
## addcut with jetB over 2
## C-medium
## addcut with jetC over 1
## C-tight
## addcut with jetC over 2
## D
## addcut with jetD over 1
## E-loose
## addcut with jetE over 1
## E-medium
## addcut with jetE over 2
## E-tight
## addcut with jetE over 3
\end{verbatim}
where the appropriate signal region (A through E) was uncommented separately for each run of Seer. All LHCO files for each data point were included in {\fontfamily{phv}\selectfont seer\_files.txt}, resulting in a total of 10 runs of Seer, once for each signal region. Due to the large number of LHCO used, the plotting features were disabled in {\fontfamily{phv}\selectfont seer\_plots.txt}, as distributions of events were irrelevant and only the overall summary provided in the text file output was needed. A script, {\fontfamily{phv}\selectfont analysis.cpp} (included in the {\fontfamily{phv}\selectfont Results} folder of Seer) was used to read through the output files and generate a Mathematica script to plot the exclusion contour.
The results of the calculations using Seer are shown in Figure \ref{fig:ex_047}. The leading order cross sections agree with the values in the conference note to within about 20\%, depending on the point in parameter space, but are systematically lower. Due to the large number of QCD events in the fully inclusive cross section, the effect of ISR jets and other NLO effects can be quite sizeable. Instead of incorporating them, these effects were accounted for using a K-factor of 35\%. \footnote{This value is chosen based on the K-factors listed on \url{https://twiki.cern.ch/twiki/bin/view/Sandbox/TestTopic232} for the mSUGRA benchmark points. Ideally, the K-factor will vary based on the particular parameter points. A flat K-factor was chosen as a sufficient approximation to vet the Seer code.} The excellent agreement between Seer and the experimental results suggests that Seer is sufficiently capable of analyzing and reproducing the experimental study.
\begin{figure*}[tb]
\centering
\mbox{\includegraphics[width=0.45\textwidth,clip]{Conf_047_compare.pdf}}
\caption{Exclusion plot for inclusive jet+$\not\!\!{E}_T$ production in mSUGRA from Seer (orange at LO, black with K-factor), as compared to the observed limit from the originating ATLAS study (red) \cite{TheATLAScollaboration:2013fha}.\label{fig:ex_047}}
\end{figure*}
\subsubsection{ATLAS tri-lepton search}
The tri-lepton search that was reproduced \cite{ATLAS:2013rla} focused on a simplified version of the pMSSM, where the squarks, gluons and Higgsinos are decoupled. This leaves three remaining relevant parameters, $M_1 \sim M_{\tilde{\chi}_1^0}$ and $M_2 \sim M_{\tilde{\chi}_2^0} \sim M_{\tilde{\chi}_1^\pm}$ which are varied, and a fixed $\tan\beta = 10$. The process $pp \rightarrow \tilde{\chi}_1^\pm \tilde{\chi}_2^0$ was generated with 50k events, and the same process of production, decay, hadronization and detector simulation was performed as discussed in the previous section.
The {\fontfamily{phv}\selectfont seer\_cuts.txt} settings used to reproduce this study were:
\begin{verbatim}
Signal ## Do not alter this line!
3 99 lep, pt 10 99999, eta -2.5 2.5
0 99 jet, pt 20 99999, eta -2.5 2.5
0 99 gam, pt 10 99999, eta -2.5 2.5
Extra ## Do not alter this line!
##THIS IS ATLAS TRIGGER
trigger ele pt 25 0 0 0
trigger ele pt 15 15 0 0
trigger muo pt 20 0 0 0
trigger muo pt 10 10 0 0
trigger gam pt 60 0 0 0
trigger gam pt 20 20 0 0
trigger jet pt 400 0 0 0
trigger jet pt 165 165 165 0
trigger jet pt 110 110 110 110
trigger jat pt 70 70 0 0
trigger tat pt 35 45 0 0
trigger eam pt 20 10 0 0
trigger eam pt 10 20 0 0
##IGNORES
ignore ele with pt between 0 10
ignore ele with eta over 2.47
ignore muo with pt between 0 10
ignore muo with eta over 2.4
ignore gam with pt between 0 10
ignore gam with eta over 2.47
ignore jet with pt between 0 20
ignore jet with eta over 2.5
addcut bjt with num over 0
addcut all with a3lp over 1
\end{verbatim}
where the numerical value of the \texttt{addcut all with a3lp over 1} line was varied from 1-6 for each signal region. All other aspects of the analysis were similar as to the 2-6 jets study discussed previously. The results of this are shown in Figure \ref{fig:ex_035}.
In this case, the exclusion region found using Seer is in excellent agreement for $m_{\tilde{\chi}_2^0} - m_{\tilde{\chi}_1^0} \gg m_Z$. In the region where the spectrum compresses, multiple signal regions provide exclusions. This results in an enhanced exclusion using the full power of the $CL_s$ method, as discussed, but an inability to precisely reproduce the exclusions using Seer. This is a limitation that cannot be resolved currently without more information regarding the correlations of systematic uncertainties in the experimental signal regions.
The original study employed the NLO cross sections from Prospino 2.0 \cite{Beenakker:1996ed}, however the reproduction in Seer uses LO cross sections. Appropriate K-factors in this region are approximately 10-20\%, based on a randomized parameter point check in Prospino over the parameter space explored. An overall 15\% K-factor has been included as a separate contour in Figure \ref{fig:ex_035}. Regions where the K-factor significantly increases the exclusion typically have high acceptance rates. Alternatively, where the K-factor exclusion line is close to the LO exclusion, the acceptance rates are low.
\begin{figure*}[tb]
\centering
\mbox{\includegraphics[width=0.45\textwidth,clip]{Conf_035_compare.pdf}}
\caption{Exclusion plot for chargino+neutralino production in the three lepton final state from Seer (orange at LO, black with K-factor), as compared to the observed limit from the originating ATLAS study (red) \cite{ATLAS:2013rla}. Lightest neutralino masses smaller than 10 GeV were not generated as part of this analysis.\label{fig:ex_035}}
\end{figure*}
\subsubsection{ATLAS di-lepton search}
The di-lepton search that was reproduced \cite{ATLAS:2013rla} focused on a simplified version of the pMSSM, where right handed sleptons (degenerate $\tilde{e}_R$ and $\tilde{\mu}_R$) decay directly to a lepton and a bino-type lightest neutralino. This results in a similar three parameter subset of the pMSSM, with a scan over the mass of the sleptons $m_{\tilde{\ell}}$ and the bino mass $M_1 \sim m_{\tilde{\chi}_1^0}$, and a fixed $\tan\beta=10$. The process $pp \rightarrow \tilde{\ell}_R^+ \tilde{\ell}_R^-$ was generated with 50k events using the same generation, decay, hadronization and detector simulation discussed previously.
The {\fontfamily{phv}\selectfont seer\_cuts.txt} settings used to reproduce this study were:
\begin{verbatim}
Signal ## Do not alter this line!
2 2 lep, pt 10 99999, eta -2.5 2.5
Extra ## Do not alter this line!
##THIS IS ATLAS TRIGGER
trigger ele pt 25 0 0 0
trigger ele pt 15 15 0 0
trigger muo pt 20 0 0 0
trigger muo pt 10 10 0 0
trigger gam pt 60 0 0 0
trigger gam pt 20 20 0 0
trigger jet pt 400 0 0 0
trigger jet pt 165 165 165 0
trigger jet pt 110 110 110 110
trigger jat pt 70 70 0 0
trigger tat pt 35 45 0 0
trigger eam pt 20 10 0 0
trigger eam pt 10 20 0 0
##IGNORES
ignore ele with eta over 2.47
ignore muo with eta over 2.4
ignore gam with pt between 0 20
ignore gam with eta over 2.4
ignore jet with pt between 0 20
ignore jet with eta over 4.5
addcut all with a2lp over 2
\end{verbatim}
where the numerical value of the \texttt{addcut all with a2lp over 1} line was varied from 1-5 for each signal region. All other aspects of the analysis were similar as to the 2-6 jets study discussed previously.
Very similar exclusion regions were found, however for the region of large $m_{\tilde{\ell}}$ the Seer exclusion does not lie within the $\pm 1 \sigma$ boundary for the exclusion. This is due to the very low uncertainties in the signal, which results from the exclusive nature of the search (all events are vetoed for the presence of any jet with $p_T>20$~GeV). One particular explanation for the difference in the exclusions is that Delphes employs a flat efficiency for lepton identification. This is one area where CheckMATE has improved upon the standard Delphes implementation available with the MG5+aMC@NLO package by incorporating many more details regarding lepton tagging and identification. None-the-less, Seer performs exceptionally well in reproducing this study, even with the standard Delphes treatment of leptons.
A similar K-factor to the tri-lepton study was included in this analysis, to show the dependence on the signal cross section. Appropriate K-factors found using Prospino are on the order of 10-20\%, however the jet veto employed in this analysis would suggest that an appropriate K-factor is more challenging to determine. Additionally, and since the K-factor is parameter point dependent, this value is included purely as a qualitative descriptor of the effect of increasing the signal cross section on the exclusion results.
\begin{figure*}[tb]
\centering
\mbox{\includegraphics[width=0.45\textwidth,clip]{Conf_049_compare.pdf}}
\caption{Exclusion plot for right handed slepton pair production from Seer (orange at LO, black with K-factor), as compared to the observed limit from the originating ATLAS study (red) \cite{TheATLAScollaboration:2013hha}. Slepton masses smaller than 100 GeV were not generated as part of this analysis.\label{fig:ex_049}}
\end{figure*}
\section{Summary}
Seer is a useful tool for analyzing LHCO output files for collider processes that have gone through a detector simulation like PGS or Delphes. With a simple structure to use, and an easy-to-modify script system for the implementation of new cuts and kinematic variables, Seer can be rapidly employed for exploring kinematic distributions as well as an analysis of trigger response, cut efficiencies, and even full analyses of experimental processes.
There are a number of primary uses for this tool. The most obvious would be in visualizing kinematic distributions and exploring the kinematics of a signal. The second would be in reproducing experimental analyses. The third is in testing new kinematic variables and distributions. For example, stransverse variables are of particular interest for SUSY searches, and new variants are still being developed. The Seer code structure provides a simple framework that allows for rapid implementation of these new test variables.
Alternative code packages such as MadAnalysis 5 and CheckMATE include different features and different interfaces, with different learning curves. Seer presents an easily usable, easily modifiable, powerful and effective way to analyze the results of simulation data.
\section*{Acknowledgements}
The author would like to thank Alejandro de la Puente and Jonathan Kozaczuk for encouragement to share this package, and Carlos Wagnar's tongue-in-cheek encouragement with his statement, ``There is no point in naming it if you aren't going to release it," or something to that effect. Additionally, the author would like to thank David Morrissey for his patience and encouragement while Seer was being written, and J.P. Archambault for contributions to the style code for ROOT. This work was supported in part by the National Science and Engineering Research Council of Canada~(NSERC).
|
\section{Introduction}
The quark model can reproduce the behavior of observables such as the spectrum and the magnetic moments, but it neglects pair-creation (or continuum-coupling) effects.
Above threshold, this coupling leads to strong decays; below threshold, it leads to virtual $q \bar q - q \bar q$ components in the meson wave function and shifts of the physical mass with respect to the bare mass.
The unquenching of the quark model for mesons is a way to take these components into account.
Pioneering work on the unquenching of meson quark models was done by Van Beveren, Dullemond and Rupp using a t-matrix approach \cite{vanBeveren:1979bd,vanBeveren:1986ea}, while T\"ornqvist and collaborators \cite{Ono:1983rd,Tornqvist} used their unitarized QM.
These methods were used (with a few variations) by several authors to study the influence of the meson-meson continuum on meson observables.
We mention, as an example, the study of the scalar meson nonet ($a_0$, $f_0$, etc.) of Ref. \cite{vanBeveren:1986ea,Tornqvist:1995kr} in which the loop contributions are given by the hadronic intermediate states that each meson can access. It is via these hadronic loops that the bare states become "dressed" and the hadronic loop contributions totally dominante the dynamics of the process. A very similar approach was developed by Pennington in Ref. \cite{Pennington:2002}, in which the dynamical generation of the scalar mesons by initially inserting only one "bare seed", was investigated. The study of Ref. \cite{Geiger:1989yc} demonstrates that the effects of the $q \bar q$ sea pairs in meson spectroscopy is simply a renormalization of the meson string tension.
Eichten {\it et al.} explored the influence of the open-charm channels on the charmonium properties, using the Cornell coupled-channel model \cite{Eichten:1974af} to assess departures from the single-channel potential-model expectations.
In this contribution, we discuss some of the latest applications of the UQM to the study of meson observables. Finally, we discuss the spectroscopy of hybrid mesons in Coulomb Gauge QCD.
\section{UQM }
\subsection{Formalism}
\label{Sec:formalism}
In the unquenched quark model for mesons \cite{bottomonium,charmonium,Ferretti:2013vua,Ferretti:2014xqa}, the meson wave function is made up the valence $q \bar q$ configuration plus a sum over the continuum components as
\begin{eqnarray}
\label{eqn:Psi-A}
\mid \psi_A \rangle &=& {\cal N} \left[ \mid A \rangle
+ \sum_{BC \ell J} \int d \vec{K} \, k^2 dk \, \mid BC \ell J;\vec{K} k \rangle \right.
\nonumber\\
&& \hspace{2cm} \left. \frac{ \langle BC \ell J;\vec{K} k \mid T^{\dagger} \mid A \rangle }
{E_a - E_b - E_c} \right] ~,
\end{eqnarray}
where $T^{\dagger}$ stands for the $^{3}P_0$ quark-antiquark pair-creation operator \cite{bottomonium,charmonium,Ferretti:2013vua,Ferretti:2014xqa}, $A$ is the meson, $B$ and $C$ represent the intermediate virtual mesons, $E_a$, $E_b$ and $E_c$ are the corresponding energies, $k$ and $\ell$ the relative radial momentum and orbital angular momentum between $B$ and $C$ and $\vec{J} = \vec{J}_b + \vec{J}_c + \vec{\ell}$ is the total angular momentum.
It is worthwhile noting that in Refs. \cite{bottomonium,charmonium,Ferretti:2013vua,Ferretti:2014xqa,Kalashnikova:2005ui}, the constant pair-creation strength in the operator (\ref{eqn:Psi-A}) was substituted with an effective one, to suppress unphysical heavy quark pair-creation.
In the UQM \cite{bottomonium,charmonium,Ferretti:2013vua,Ferretti:2014xqa}, the matrix elements of an observable $\hat O$ can be calculated as
\begin{equation}
O = \left\langle \psi_A \right| \hat O \left| \psi_A \right\rangle \mbox{ },
\end{equation}
where $\left| \psi_A \right\rangle$ is the state of Eq. (\ref{eqn:Psi-A}).
The result will receive a contribution from the valence part and one from the continuum component, which is absent in naive QM calculations.
\subsection{$c \bar c$ and $b \bar b$ spectra with self-energy corrections in the UQM}
In Refs. \cite{bottomonium,charmonium,Ferretti:2013vua,Ferretti:2014xqa}, the method was used to calculate the $c \bar c$ and $b \bar b$ spectra with self-energy corrections, due to continuum coupling effects. In the UQM, the physical mass of a meson,
\begin{equation}
\label{eqn:self-trascendental}
M_a = E_a + \Sigma(E_a) \mbox{ },
\end{equation}
is given by the sum of two terms: a bare energy, $E_a$, calculated within a potential model
\cite{Godfrey:1985xj}, and a self energy correction,
\begin{equation}
\label{eqn:self-a}
\Sigma(E_a) = \sum_{BC\ell J} \int_0^{\infty} k^2 dk \mbox{ } \frac{\left| M_{A \rightarrow BC}(k) \right|^2}{E_a - E_b - E_c} \mbox{ },
\end{equation}
computed within the UQM formalism.
Our results for the self energies of charmonia \cite{charmonium,Ferretti:2014xqa} and bottomonia \cite{bottomonium,Ferretti:2013vua,Ferretti:2014xqa} show that the pair-creation effects on the spectrum of heavy mesons are quite small. Specifically for charmonium and bottomonium states, they are of the order of $2 - 6\%$ and $1 \%$, respectively.
The relative mass shifts, i.e. the difference between the self energies of two meson states, are in the order of a few tens of MeV.
However, as QM's can predict the meson masses with relatively high precision in the heavy quark sector, even these corrections can become significant.
These results are particularly interesting in the case of states close to an open-flavor decay threshold, like the $X(3872)$ and $\chi_b(3P)$ mesons.
For example, in our picture the $X(3872)$ can be interpreted as a $c \bar c$ core, working as a "seed" [the $\chi_{c1}(2^3P_1)$], plus the virtual meson-meson continuum. In Ref. \cite{Ferretti:2014xqa}, we showed that the probability to find the $X(3872)$ in its core or continuum components is approximately $45\%$ and $55\%$, respectively.
\subsection{Loops of hybrid mesons}
The wave function of a meson can be written as
\begin{equation}
\label{eqn:higher-Fock}
\left| \Psi \right\rangle = \left| q \bar q \right\rangle + \left| q \bar q q \bar q \right\rangle
+ \left| q \bar q g \right\rangle + ... \mbox{ },
\end{equation}
where $\left| q \bar q \right\rangle$ is the quark-antiquark component, $\left| q \bar q q \bar q \right\rangle$ the tetraquark or molecular component and $\left| q \bar q g \right\rangle$ the quark-antiquark-gluon (hybrid) component.
In the QM, conventional mesons are in general described by the $\left| q \bar q \right\rangle$ valence component. Nevertheless, there are also attempts to accommodate exotic states as $\left| q \bar q q \bar q \right\rangle$ or $\left| q \bar q g \right\rangle$ states.
In the QM formalism, $\left| q \bar q \right\rangle$ and $\left| q \bar q g \right\rangle$ mesons are described by the non relativistic Hamiltonian \cite{Isgur:1984bm}
\begin{equation}
\label{eqn:Hstring}
H^\nu = - \frac{1}{2 \mu} \frac{\partial^2}{\partial r^2} + \frac{\ell (\ell+1)}{2 \mu r^2} + E^\nu(r) \mbox{ },
\end{equation}
where
\begin{equation}
\label{eqn:Enu}
E^\nu(r) = - \frac{\tau}{r} + \beta r + \frac{\nu \pi}{r} + C \mbox{ },
\end{equation}
is the potential describing $q \bar q$ mesons ($\nu = 0$), the first hybrid surface ($\nu = 1$), the second hybrid surface ($\nu = 2$), and so on.
The symbols $\ell$ and $\mu$ in Eq. (\ref{eqn:Enu}) stand for the orbital angular momentum of the state and the reduced mass of the $q \bar q$ system, respectively.
In the UQM formalism, the wave function of a meson can be written as the superposition of the components of Eq. (\ref{eqn:higher-Fock}).
Up to now, we have considered in our calculations \cite{bottomonium,charmonium,Ferretti:2013vua,Ferretti:2014xqa} only the first two terms, $\left| q \bar q \right\rangle$ and $\left| q \bar q q \bar q \right\rangle$.
They are the most important for ground state and lower-lying mesons, close to the first open-flavor decay thresholds.
Nevertheless, at higher energies, the effects of ground state open-flavor meson loops, like $D \bar D$ or $D \bar D^*$ in the $c \bar c$ sector (see Ref. \cite{charmonium}), become less important.
In the $c \bar c$ sector, the first hybrid $c \bar c g$ mesons lie at energies of 4.2 GeV, approximately.
Thus, above 4 GeV, the introduction of hybrid loops could be crucial to understanding the higher lying mesons' structure and spectrum.
In the QM formalism, the coupling between a hybrid meson $\mathcal H$ and a quarkonium state $Q$ is given by \cite{LeYaouanc:1984gh}
\begin{equation}
\label{eqn:Hibrid-coupling}
\left\langle \mathcal H \right| V \left| Q \right\rangle = \left\langle q \bar q g \right| V \left| q \bar q \right\rangle \mbox{ },
\end{equation}
where $V$ is an interaction that annihilates the constituent gluon $g$.
The coupling of Eq. (\ref{eqn:Hibrid-coupling}) can be used to calculate the contribution of hybrid loops to the self-energy of higher-lying mesons \cite{FS01}.
\begin{figure}[htbp]
\centerline
\includegraphics[width=7cm,angle=270]{ccg}}
\caption{Charmonium (solid boxes) and charmonium hybrid spectrum compared with data (where available) or lattice computations. Single dashed boxes are the $c{\bar c}g$ hybrids dominated by the $P$-wave quarks, all other have the $Q{\bar Q}$ pair in the relative $S$-wave orbital. Picture from Ref. \cite{Guo:2008yz}; APS copyright.}
\label{Fig:F1}
\end{figure}
\begin{figure}[htbp]
\centerline
\includegraphics[width=7cm,angle=270]{bbg}}
\caption{Same as Fig. \ref{Fig:F1} for bottomonium. Picture from Ref. \cite{Guo:2008yz}; APS copyright.}
\label{Fig:F2}
\end{figure}
\section{Hybrid mesons in Coulomb Gauge QCD}
Hybrid states have been studied in several models \cite{Isgur:1984bm,Horn:1977rq,Szczepaniak:1995cw,Simonov:2000ky,Wang:2006ida,Buisseret:2006wc}, including lattice QCD calculations (for example, see Ref. \cite{Morningstar:1999rf}).
The most commonly studied hybrids are composed of a quark, an antiquark and a gluon.
Particular interest has grown because of the recent discovery of several new states, in particular in the $c \bar c$ sector, probably including a hybrid resonance, the $Y(4260)$, with $1^{--}$ quantum numbers. This meson was discovered by BaBar in ISR production $J/\Psi \pi^+ \pi^-$ \cite{Aubert:2005rm} and then confirmed by CLEO \cite{CLEO:2006aa0} and BELLE \cite{Pakhlova:2010ek}.
The observation of $Y (4260)$ by ISR strongly suggests that it is a vector meson. In direct production, CLEO also observed the $Y (4260)$ decaying into $J/\Psi \pi^0\pi^0$ and $J/\Psi K^+ K^-$ \cite{CLEO:2006aa}. Although there may be not a large amount of data to allow an interpretation of the $Y (4260)$ based on its decay properties, the present state of hadronic theory is not up to the task.
Since no conventional charmonium vectors are expected at this mass, the $Y (4260)$ may be dominantly a hybrid meson.
Conventional heavy quarkonia are well described by non-relativistic QCD, thus one can expect that hybrids containing heavy quarks can be treated similarly, considering gluonic excitations in presence of slow quarks. Moreover, in physical gauge, the dynamical gluons can be separated from the instantaneous Coulomb-type forces that act between color charges, thus while the non-abelian Coulomb potential is expected to be responsible for binding and confinement, the remaining, transverse gluon excitations could bring contribution to the spectrum.
In Refs. \cite{Guo:2007sm,Guo:2008yz}, hybrid mesons were studied in a variational approach to QCD in the Coulomb gauge. In this approach, a confining linear potential emerged from the Dyson-Schwinger equations, at least at the hadronic scale. In a first stage, this potential was used to calculate the spectrum of the gluelump, which is an idealized system defined as gluonic excitations bounded to a static, localized color octet source, such as, for example, a very heavy quark and antiquark \cite{Guo:2007sm}. The next step was to introduce the quark-antiquark dynamics to calculate the spectrum of heavy hybrid mesons \cite{Guo:2008yz}. See Figs. \ref{Fig:F1} and \ref{Fig:F2}, where the results are compared with data (where available) or lattice calculations.
It is worthwhile noting that the lowest mass charmonium hybrid multiplet was predicted to lie at an energy of 4.476 GeV and to be composed by states with $J^{PC} = 1^{--}$; $(0; 1; 2)^{-+}$ \cite{Guo:2008yz}; the multiplet includes also an exotic state, with $J^{PC} = 1^{-+}$. This four state hybrid multiplet was also recently identified in lattice simulations \cite{Dudek:2011tt}.
In the bottomonium case, the lowest mass hybrid multiplet lies at an energy of 11.055 GeV and is composed, as in the case of $c \bar c g$, by states with $J^{PC} = 1^{--}$, $0^{-+}$, $1^{-+}$ and $2^{-+}$. The $1^{--}$ state could be identified with the resonance $Y_b(10888)$ observed by Belle in $e^+e^-$ annihilation \cite{Chen:2010}.
|
\section{Introduction}
Digital topology adapts tools from geometric and algebraic topology to
the study of digital images. In this paper, we consider questions of
pointed homotopy in digital topology.
We give an example showing that homotopy equivalence
between digital images $(X,c_u)$ and $(Y,c_v)$ does not imply
pointed homotopy equivalence between these images.
This example is then used to illustrate a new variant on the pointed homotopy of digital loops.
We present an alternate
version of the digital fundamental group that appears to have
advantages over the version introduced in~\cite{Boxer99}.
We correct the argument of~\cite{Boxer05a} for the assertion that
homotopy equivalent connected digital images $(X,\kappa)$ and $(Y,\lambda)$
have isomorphic fundamental groups $\Pi_1^{\kappa}(X,x_0)$ and
$\Pi_1^{\lambda}(Y,y_0)$.
Much of the material in section~\ref{prelim} is quoted or paraphrased from
~\cite{BoxKar2}.
\section{Preliminaries}
\label{prelim}
\subsection{General Properties}
\label{dig-con}
Let {\bf Z} be the set of integers.
A {\em (binary) digital image} is a pair $(X,\kappa)$, where
$X \subset {\bf Z}^n$ for some positive integer~$n$, and $\kappa$ is
some adjacency relation for the members of $X$.
Adjacency relations commonly used in the study of
digital images in ${\bf Z}^n$ include the following ~\cite{Han}.
For an integer $u$ such that $1 \leq u \leq n$, we define an
adjacency relation as follows. Points
\[ p \, = \, (p_1,p_2,\ldots,p_n), ~ q \, = \, (q_1,q_2,\ldots, q_n) \]
are $c_u$-adjacent~\cite{Boxer05b} if
\begin{itemize}
\item $p \neq q$, and
\item there are at most $u$ distinct indices $i$ for which
$|p_i - q_i| \, = \, 1$, and
\item for all indices $i$, if $|p_i - q_i| \neq 1$ then $p_i \, = \, q_i$.
\end{itemize}
We often denote a $c_u$-adjacency in ${\bf Z}^n$ by the number of points that
are $c_u$-adjacent to a given point in ${\bf Z}^n$. E.g.,
\begin{itemize}
\item in ${\bf Z}^1$, $c_1$-adjacency is 2-adjacency;
\item in ${\bf Z}^2$, $c_1$-adjacency is 4-adjacency and $c_2$-adjacency is 8-adjacency.
\item in ${\bf Z}^3$, $c_1$-adjacency is 6-adjacency, $c_2$-adjacency is 18-adjacency, and $c_3$-adjacency is 26-adjacency.
\end{itemize}
More general adjacency relations appear in~\cite{Herman}. The work in \cite{Staecker} treats digital images as abstract sets of points with arbitrary adjacencies without regard for their embeddings in $\mathbf Z^n$.
\begin{Definition} {\rm \cite{Boxer}}
\label{dig-int}
Let $a, b \in {\bf Z}$, $a < b$. A {\rm digital interval}
is a set of the form
\[ [a,b]_{{\bf Z}} ~=~ \{z \in {\bf Z} ~|~ a \leq z \leq b\} \]
in which $c_1$-adjacency is assumed. $\Box$
\end{Definition}
\begin{comment}
A digital image $(X,\kappa)$ is \emph{$\kappa$-connected}~\cite{Herman} (when $\kappa$ is understood, we simply say $X$ is \emph{connected})
if and only if for
every pair of points $\{x,y\} \subset X$, $x \neq y$, there exists a set
$P \, = \, \{x_0, x_1,\ldots, x_c\} \subset X$ of $c+1$ distinct points
such that $x = x_0$, $x_c = y$,
and $x_i$ and $x_{i+1}$ are adjacent, $i \in \{0,1,\ldots, c-1\}$.
The set $P$ is a {\em path}.
We say $c$ is the {\em length} of $P$.
We will abuse the terminology by also calling a
$(c_1,\kappa)$-continuous function (where continuity is as defined below at
Definition~\ref{cont-connect})
$f: [0,m]_{\bf Z} \rightarrow P$ such that $f(0)=x$, $f(m)=y$, a
{\em $\kappa$-path in $X$ from $x$ to $y$}.
\end{comment}
The following generalizes an earlier definition of~\cite{Rosenfeld}.
\begin{Definition}
\label{cont-connect}
{\rm \cite{Boxer99}}
Let $(X,\kappa)$ and $(Y,\lambda)$ be digital images. Then the function
$f: X \rightarrow Y$ is $(\kappa,\lambda)$-continuous if and only if
for every pair of $\kappa-$adjacent points
$x_0, x_1 \in X$, either $f(x_0)$ $= f(x_1)$, or $f(x_0)$ and $f(x_1)$
are $\lambda-$adjacent. $\Box$
\end{Definition}
See also~\cite{Chen94,Chen04}, where similar concepts are named
{\em immersion}, {\em gradually varied operator}, or
{\em gradually varied mapping}.
A {\em path} from $p$ to $q$ in $(X, \kappa)$ is a $(2,\kappa)$-continuous function
$F : [0,m]_{\bf Z} \rightarrow X$ such that $F(0)=p$ and $F(m)=q$. For a given path $F$, we define the reverse path, $F^{-1}: [0,m]_{\bf Z} \rightarrow X$ defined by
$F^{-1}(t)=F(m-t)$. A {\em loop} is a path $F: [0,m]_{\bf Z} \rightarrow X$ such that $F(0)=F(m)$.
\subsection{Digital homotopy}
\label{dig-htpy}
Intuitively, a homotopy between continuous functions $f,g: X \rightarrow Y$
is a continuous deformation of, say, $f$ over a time period
until the result of the deformation coincides with $g$.
\begin{Definition}{\rm (\cite{Boxer99};} see also~{\rm \cite{Khalimsky})}
\label{htpy-2nd-def}
Let $X$ and $Y$ be digital images.
Let $f,g: X \rightarrow Y$ be $(\kappa,\lambda)-$continuous functions
and suppose there is a positive integer $m$ and a function
\[ F: X \times [0,m]_{{\bf Z}} \rightarrow Y \]
such that
\begin{itemize}
\item for all $x \in X$, $F(x,0) = f(x)$ and $F(x,m)$ $= g(x)$;
\item for all $x \in X$, the induced function
$F_x: [0,m]_{{\bf Z}} \rightarrow Y$ defined by
\[ F_x(t) ~=~ F(x,t) \mbox{ for all } t \in [0,m]_{{\bf Z}}, \]
is $(c_1,\lambda)-$continuous;
\item for all $t \in [0,m]_{{\bf Z}}$, the induced function
$F_t: X \rightarrow Y$ defined by
\[ F_t(x) ~=~ F(x,t) \mbox{ for all } x \in X, \]
is $(\kappa,\lambda)-$continuous.
\end{itemize}
Then $F$ is a {\rm digital $(\kappa,\lambda)-$homotopy between} $f$ and
$g$, and $f$ and $g$ are $(\kappa,\lambda)${\rm -homotopic in} $Y$. If $m=1$, then $f$ and $g$ are homotopic \emph{in 1 step}.
If, further, there exists $x_0 \in X$ such that $F(x_0, t)= F(x_0,0)$ for
all $t \in [0,m]_{\bf Z}$, we say $F$ is a {\rm pointed homotopy}. If $g$ is a constant function, we say $F$ is a
{\rm nullhomotopy}, and $f$ is {\rm nullhomotopic}.
$\Box$
\end{Definition}
The notation $f~\simeq_{( \kappa, \lambda)} ~g$
indicates that functions $f$ and $g$ are digitally
$(\kappa,\lambda)-$homotopic in $Y$. If $\kappa=\lambda$, we abbreviate this as $f\simeq_\kappa g$. When the adjacencies are understood we simply write $f\simeq g$.
Digital homotopy is an equivalence relation among digitally continuous
functions~\cite{Khalimsky,Boxer99}.
Let $H: [0,m]_{\bf Z} \times [0,n]_{\bf Z} \rightarrow X$
be a homotopy between paths
$f, g: [0,m]_{\bf Z} \rightarrow X$. We say $H$ {\em holds the endpoints fixed} if
$f(0)=H(0,t)=g(0)$ and
$f(m)=H(m,t)=g(m)$ for all
$t \in [0,n]_{\bf Z}$. If $f$ and $g$ are loops,
we say $H$ is {\em loop preserving} if
$H(0,t)=H(m,t)$ for all $t \in [0,n]_{\bf Z}$.
Notice that if $f$ and $g$ are loops and
$H$ holds the endpoints
fixed, then $H$ is a loop preserving pointed homotopy
between $f$ and $g$.
As in classical topology, we say two digital images $(X,\kappa)$ and $(Y,\lambda)$ are \emph{homotopy equivalent} when there are continuous functions $f:X\to Y$ and $g:Y\to X$ such that $g\circ f \simeq_{(\kappa,\lambda)} 1_X$ and $f\circ g \simeq_{(\lambda,\kappa)} 1_Y$.
\subsection{Digital fundamental group}
If $f$ and $g$ are paths in $X$
such that $g$ starts where $f$ ends,
the {\em product} (see~\cite{Khalimsky})
of $f$ and $g$, written $f * g$, is, intuitively, the path obtained
by following $f$, then following $g$.
Formally, if $f: [0,m_1]_{{\bf Z}} \rightarrow X$,
$g: [0,m_2]_{{\bf Z}} \rightarrow X$, and $f(m_1)=g(0)$, then
$(f * g): [0,m_1+m_2]_{{\bf Z}} \rightarrow X$ is defined by
\[(f * g)(t) = \left\{
\begin{array}{ll}
f(t) & \mbox{if } t \in [0,m_1]_{{\bf Z}}; \\
g(t - m_1) & \mbox{if } t \in [m_1,m_1+m_2]_{{\bf Z}}.
\end{array}
\right . \]
Restriction of loop classes to loops defined on the
same digital interval would be undesirable. The following
notion of {\em trivial extension} to permit a loop
to ``stretch" within the same pointed homotopy class. In
section~\ref{ec-section}, we will introduce a different method of
``stretching'' a loop within its pointed homotopy class.
Intuitively, $f'$ is a trivial extension of $f$ if $f'$ follows the same
path as $f$, but more slowly, with pauses for rest
(subintervals of the domain on which $f'$ is constant).
\begin{Definition}
\label{triv-extension}
{\rm \cite{Boxer99}}
Let $f$ and $f'$ be loops in a pointed digital image $(X,x_0)$.
We say
$f'$ is a {\rm trivial extension of} $f$ if there are sets of paths
$\{f_1, f_2, \ldots, f_k\}$ and $\{F_1, F_2, \ldots, F_p\}$ in $X$ such that
\begin{enumerate}
\item $0<k \leq p$;
\item $f = f_1 * f_2 * \ldots * f_k$;
\item $f' = F_1 * F_2 * \ldots * F_p$;
\item there are indices
$1 \leq i_1 < i_2 < \ldots < i_k \leq p$ such that
\begin{itemize}
\item $F_{i_j} = f_j$, $1 \leq j \leq k$, and
\item $i \not \in \{i_1, i_2, \ldots, i_k\}$ implies $F_i$ is a
trivial loop. $\Box$
\end{itemize}
\end{enumerate}
\end{Definition}
This notion lets us compare the digital homotopy properties of loops
whose domains may have differing cardinality, since if $m_1 \leq m_2$,
we can obtain \cite{Boxer99} a trivial extension of a loop
$f:[0,m_1]_{{\bf Z}} \rightarrow X$ to
$f':[0,m_2]_{{\bf Z}} \rightarrow X$ via
\[ f'(t) = \left\{
\begin{array}{ll}
f(t) & \mbox{if } 0 \leq t \leq m_1; \\
f(m_1) & \mbox{if } m_1 \leq t \leq m_2.
\end{array}
\right. \]
Observe that every digital loop $f$ is a trivial extension of itself.
\begin{Definition}
\label{loop-class}
{\rm (\cite{Han}, correcting an earlier definition in \cite{Boxer05a}).}
Two loops $f_0,f_1$ with the same base point $p \in X$
belong to the same loop class~$[f]_X$ if they have trivial extensions
that can be joined by a homotopy $H$ that keeps the endpoints fixed.
\end{Definition}
It was incorrectly asserted as Proposition~3.1 of~\cite{Boxer05a}
that the assumption in Definition~\ref{loop-class}, that the homotopy keeps the endpoints fixed,
could be replaced by the weaker assumption
that the homotopy is loop-preserving; the error was pointed out in \cite{Boxer06}.
Membership in the same loop class in $(X,x_0)$ is
an equivalence relation among loops
~\cite{Boxer99}.
The digital fundamental group is derived from
a classical notion of algebraic topology (see \cite{Massey,Munkres,Spanier}).
The version discussed in this section is that developed in~\cite{Boxer99}.
The next result is used in \cite{Boxer99} to show the product operation of our
digital fundamental group is well defined.
\begin{Proposition}
\label{well-defined}
{\rm \cite{Boxer99,Khalimsky}}
Let $f_1,f_2,g_1,g_2$ be digital loops based at $x_0$ in a pointed digital
image $(X,x_0)$, with $f_2 \in [f_1]_X$ and
$g_2 \in [g_1]_X$. Then $f_2 * g_2 \in [f_1 * g_1]_X$. \qed
\end{Proposition}
Let $(X,x_0)$ be a pointed digital image;
{\em i.e.}, $X$ is a digital image, and $x_0 \in X$.
Define $\Pi_1(X,x_0)$ to be the set of loop classes
$[f]_X$ in $X$ with base point $x_0$. When we wish to emphasize an adjacency relation $\kappa$, we denote this set by $\Pi_1^\kappa(X,x_0)$.
By Proposition~\ref{well-defined}, the {\em product} operation
\[ [f]_X \cdot [g]_X~=~[f * g]_X \]
is well defined on $\Pi_1(X,x_0)$;
further, the operation $\cdot$ is associative on
$\Pi_1(X,x_0)$~\cite{Khalimsky}.
\begin{Lemma}
\label{ident-elt}
{\rm \cite{Boxer99}}
Let $(X,x_0)$ be a pointed digital image.
Let ${\overline{x_0}}: [0,m]_{{\bf Z}} \rightarrow X$ be a constant loop
with image $\{x_0\}$. Then $[{\overline{x_0}}]_X$ is an identity element for
$\Pi_1(X, x_0)$. \qed
\end{Lemma}
\begin{Lemma}
\label{inverse}
{\rm \cite{Boxer99}}
If $f: [0,m]_{{\bf Z}} \rightarrow X$ represents an element of $\Pi_1(X,x_0)$,
then the reverse loop $f^{-1}$
is an element of $[f]_X^{-1}$ in $\Pi_1(X,x_0)$. \qed
\end{Lemma}
\begin{Theorem}
{\rm \cite{Boxer99}}
$\Pi_1(X,x_0)$ is a group under the $\cdot$ product operation,
the {\em fundamental group of} $(X,x_0)$.
\qed
\end{Theorem}
\begin{comment}
\begin{itemize}
\item $H(x,0) ~=~x$ for all $x \in X$; and
\item $H(x,m) ~=~ x_0$ for all $x \in X$.
\end{itemize}
If we also have $H(x_0,t) ~=~ x_0$ for all $t \in [0,m]_{{\bf Z}}$, we say
$X$ is {\rm pointed $k$-contractible}.
\end{Definition}
\end{comment}
\begin{Theorem}
\rm{\cite{Boxer99}}
\label{induced}
Suppose $F: (X, \kappa, x_0) \rightarrow (Y,\lambda, y_0)$
is a pointed continuous function. Then
$F$ induces a homomorphism
$F_*: \Pi_1^{\kappa}(X,x_0) \rightarrow
\Pi_1^{\lambda}(Y,y_0)$
defined by $F_*([f])=[F \circ f]$. $\Box$
\end{Theorem}
\section{Homotopy equivalent images that aren't pointed homotopy equivalent}
In~\cite{Boxer05a}, it was asked if, given digital images $(X,\kappa)$ and
$(Y,\lambda)$ that are homotopy equivalent, must $(X,x_0,\kappa)$ and
$(Y,y_0,\lambda)$ be pointed homotopy equivalent for arbitrary
base points $x_0 \in X$, $y_0 \in Y$? The paper~\cite{Staecker} gives an
example, not using any of the $c_u$-adjacencies,
that answers this question in the negative.
It is desirable to have an example that uses $c_u$-adjacencies. In this
section, we give such an example by modifying that of ~\cite{Staecker}.
\begin{Example}
\label{the-ex}
Let $X = \{x_i\}_{i=0}^{10} \subset {\bf Z}^2$ where
$x_0=(2,0)$, $x_1=(1,1)$, $x_2=(0,2)$, $x_3=(-1,2)$, $x_4=(-2,1)$,
$x_5=(-2,0)$, $x_6=(-2,-1)$, $x_7=(-1,-2)$, $x_8=(0,-2)$, $x_9=(1,-1)$,
$x_{10}=(0,0)$. Let $Y=X \setminus \{x_0\} = \{x_i\}_{i=1}^{10}$. We consider both $X$ and $Y$ as digital images with $c_2$-adjacency.
See Figure~1. $\Box$
\end{Example}
\begin{figure}
\label{2b-htpc-single}
\begin{center}\begin{tikzpicture}[scale=.35]
\filldraw[fill=white, xshift=2cm,yshift=4cm]
(45:1.2) \foreach \x in {135,225,315,45} { -- (\x:1.2) };
\filldraw[fill=white, xshift=2cm,yshift=6cm]
(45:1.2) \foreach \x in {135,225,315,45} { -- (\x:1.2) };
\filldraw[fill=white, xshift=2cm,yshift=8cm]
(45:1.2) \foreach \x in {135,225,315,45} { -- (\x:1.2) };
\filldraw[fill=white, xshift=4cm,yshift=2cm]
(45:1.2) \foreach \x in {135,225,315,45} { -- (\x:1.2) };
\filldraw[fill=white, xshift=4cm,yshift=10cm]
(45:1.2) \foreach \x in {135,225,315,45} { -- (\x:1.2) };
\filldraw[fill=white, xshift=6cm,yshift=2cm]
(45:1.2) \foreach \x in {135,225,315,45} { -- (\x:1.2) };
\filldraw[fill=white, xshift=6cm,yshift=6cm]
(45:1.2) \foreach \x in {135,225,315,45} { -- (\x:1.2) };
\filldraw[fill=white, xshift=6cm,yshift=10cm]
(45:1.2) \foreach \x in {135,225,315,45} { -- (\x:1.2) };
\filldraw[fill=white, xshift=8cm,yshift=4cm]
(45:1.2) \foreach \x in {135,225,315,45} { -- (\x:1.2) };
\filldraw[fill=white, xshift=8cm,yshift=8cm]
(45:1.2) \foreach \x in {135,225,315,45} { -- (\x:1.2) };
\filldraw[fill=white, xshift=10cm,yshift=6cm]
(45:1.2) \foreach \x in {135,225,315,45} { -- (\x:1.2) };
\node () at (2cm,4cm) {$x_{6}$};
\node () at (2cm,6cm) {$x_{5}$};
\node () at (2cm,8cm) {$x_{4}$};
\node () at (4cm,2cm) {$x_{7}$};
\node () at (4cm,10cm) {$x_{3}$};
\node () at (6cm,2cm) {$x_{8}$};
\node () at (6cm,6cm) {$x_{10}$};
\node () at (6cm,10cm) {$x_{2}$};
\node () at (8cm,4cm) {$x_{9}$};
\node () at (8cm,8cm) {$x_{1}$};
\node () at (10cm,6cm) {$x_{0}$};
\end{tikzpicture}
\end{center}
\caption{A figure $X=\{x_i\}_{i=0}^{10}$ and its subset $Y=X\setminus \{x_0\}$
that are homotopic but not pointed homotopic as images in ${\bf Z}^2$ with $c_2$-adjacency}
\end{figure}
\begin{Proposition}
\label{htpy-equiv}
Let $X$ and $Y$ be the images of Example~\ref{the-ex}. Then
$X$ and $Y$ are $(c_2,c_2)$-homotopy equivalent.
\end{Proposition}
\begin{proof}
Let $f: X \rightarrow Y$ be defined by
\[ f(x_i) = \left \{ \begin{array}{ll}
x_{i+1} & \mbox{ if } 0 \leq i \leq 9; \\
x_1 & \mbox{ if } i=10.
\end{array}
\right . \]
Let $g: Y \rightarrow X$ be the inclusion map. Clearly, both $f$ and $g$ are
$(c_2, c_2)$-continuous.
The function $H: X \times [0,1]_Z \rightarrow X$ defined by
\[ H(x_i,t) = \left \{ \begin{array}{ll}
f(x_i) = g \circ f(x_i) & \mbox{ if } t=0; \\
x_i & \mbox{ if } t=1,
\end{array}
\right .
\]
is clearly a $(c_2,c_2)$-homotopy between $g \circ f$ and $1_X$.
The function $K: Y \times [0,1]_Z \rightarrow Y$ defined by
\[ K(x_i,t) = \left \{ \begin{array}{ll}
f(x_i) = f \circ g(x_i) & \mbox{ if } t=0
\mbox{ and } 1 \leq i \leq 10; \\
x_i & \mbox{ if } t=1
\mbox{ and } 1 \leq i \leq 10,
\end{array}
\right .
\]
is clearly a $(c_2,c_2)$-homotopy between $f \circ g$ and $1_Y$.
Thus, $(X,c_2)$ and $(Y,c_2)$ are homotopy equivalent.
\end{proof}
\begin{Proposition}
\label{Ymust-be-id}
Let $Y=\{x_i\}_{i=1}^{10}$ be as above.
Let $h: (Y, c_2) \rightarrow (Y,c_2)$ be a continuous map such
that $h(x) = x$ for some $x\in Y$ and $h$ is $(c_2,c_2)$-homotopic to $1_Y$ in 1 step.
Then $h=1_Y$.
\end{Proposition}
\begin{proof}
For convenience, we prove the statement in the case where $x=x_1$. Since $(Y,c_2)$ is a simple cycle of 10 points, the same argument will work for any other value of $x$.
Since $h$ is $(c_2,c_2)$-homotopic to $1_Y$ in 1 step,
$h(x_i)$ and $x_i$ are $c_2$-adjacent or equal for all $i$.
Suppose $h \neq 1_Y$. Since $h(x_1)=x_1$,
by $c_2$-continuity, $h(x_i) \in \{x_{i-1},x_i\}$ for $2 \leq i \leq 10$,
and since $h \neq 1_Y$, there is a $j_0$ such that $2 \leq j_0 \leq 10$
and $h(x_j)=x_{j-1}$ for
$j_0 \leq j \leq 10$. In particular, $h(x_{10})=x_9$, so we have a discontinuity
since the $c_2$-adjacent points $x_1$ and $x_{10}$ do not have $c_2$-adjacent
images under $h$. Since $h$ was assumed continuous, the contradiction leads
us to conclude that $h=1_Y$.
\end{proof}
\begin{comment}
\begin{Corollary}
\label{restriction-not-restricted}
Let $Y=\{x_i\}_{i=1}^{10}$ be as above.
Let $h: (Y, c_2) \rightarrow (Y,c_2)$ be a continuous map such
that $h(x_1) = x_1$ and $h$ is homotopic to $1_Y$. Then $h=1_Y$.
\end{Corollary}
{\em Proof}: Let $H: Y \times [0,m]_Z \rightarrow Y$ be a $(c_2,c_2)$-homotopy
from $h$ to $1_Y$. Without loss of generality,
$H(y,0)=y$ and $H(y,m)=h(y)$ for all $y \in Y$.
Suppose $h \neq 1_Y$. Then there is a first step $H_t$ of the homotopy
$H: Y \times [0,m]_Z \rightarrow Y$ between $1_Y$ such that
$H_t \neq 1_Y$. Since $H_0=1_Y$, our choice of $t$ implies that
$0 < t \leq m$ and $H(y,u)=y=1_Y(y)$ for all $y \in Y$ and $0 \leq u < t$.
Then the function $H': Y \times [0,1]_Z \rightarrow Y$ defined by
$H'(y,v) = H(y,t-1+v)$ satisfies $H'(y,0)= H(y,t-1)=y$ and
$H'(y,1)= H(y,t)=h(t)$. Thus, $H'$ is a homotopy between $1_Y$ and
$H_t \neq 1_Y$ in 1 step. This contradiction of
Proposition~\ref{must-be-id} establishes our assertion. $\Box$
\end{comment}
A similar argument shows the following.
\begin{Corollary}
\label{Xmust-be-id}
Let $X=\{x_i\}_{i=0}^{10}$ be as above.
Let $h: (X, c_2) \rightarrow (X,c_2)$ be a continuous map such
that $h(x_0) = x_0$ and $h$ is homotopic in 1 step to $1_X$. Then $h=1_X$. \qed
\end{Corollary}
\begin{Proposition}
\label{proof-of-example}
Let $X= \{x_i\}_{i=0}^{10}$ and $Y=X\setminus \{x_0\}$ be as above. Then for any
$x \in X$ and $y \in Y$,
$(X,x)$ and $(Y, y)$ are not pointed $(c_2,c_2)$-homotopy equivalent.
\end{Proposition}
\begin{proof}
Suppose otherwise. Then for some $x \in X$ and $y \in Y$, there are
$(c_2,c_2)$-continuous pointed
maps $f: (X,x) \rightarrow (Y, y)$ and $g: (Y,y) \rightarrow (X, x)$
such that $f \circ g$ is pointed homotopic to $1_X$ and
$g \circ f$ is pointed homotopic to $1_Y$.
First we argue that $g\circ f$ must in fact equal $1_X$. Since $f$ and $g$ are pointed maps we have $g\circ f(x) = x$, and our pointed homotopy from $g\circ f$ to $1_X$ will fix $x$ at all stages. If $g\circ f$ were not $1_X$, then there would be some final stage $h$ of the pointed homotopy from $g\circ f$ to $1_X$ for which $h \neq 1_X$ but $h$ is pointed homotopic to $1_X$ in one step. This is impossible by Proposition \ref{Ymust-be-id}, and so we conclude that $g\circ f = 1_X$. Similarly, using Corollary \ref{Xmust-be-id}, we have $f\circ g = 1_Y$.
Since $f \circ g = 1_Y$ and $g\circ f = 1_X$, it follows that $X$ and $Y$ are $(c_2,c_2)$-isomorphic images, which is impossible, as $X$ and $Y$ have different cardinalities. The assertion
follows.
\end{proof}
Example \ref{the-ex} is an image in $\mathbf Z^2$ with $c_2$-adjacency that exhibits interesting pointed homotopy properties. We remark that images exist in $\mathbf Z^2$ with $c_1$-adjacency with similar properties. The image in Figure \ref{4adj-ex} exhibits the same behavior as that of Example \ref{the-ex}.
\begin{figure}
\begin{center}
\begin{tikzpicture}[scale=.3]
\filldraw[fill=white, xshift=2cm,yshift=2cm]
(45:1.2) \foreach \x in {135,225,315,45} { -- (\x:1.2) };
\filldraw[fill=white, xshift=2cm,yshift=4cm]
(45:1.2) \foreach \x in {135,225,315,45} { -- (\x:1.2) };
\filldraw[fill=white, xshift=2cm,yshift=6cm]
(45:1.2) \foreach \x in {135,225,315,45} { -- (\x:1.2) };
\filldraw[fill=white, xshift=2cm,yshift=8cm]
(45:1.2) \foreach \x in {135,225,315,45} { -- (\x:1.2) };
\filldraw[fill=white, xshift=4cm,yshift=2cm]
(45:1.2) \foreach \x in {135,225,315,45} { -- (\x:1.2) };
\filldraw[fill=white, xshift=4cm,yshift=8cm]
(45:1.2) \foreach \x in {135,225,315,45} { -- (\x:1.2) };
\filldraw[fill=white, xshift=6cm,yshift=2cm]
(45:1.2) \foreach \x in {135,225,315,45} { -- (\x:1.2) };
\filldraw[fill=white, xshift=6cm,yshift=6cm]
(45:1.2) \foreach \x in {135,225,315,45} { -- (\x:1.2) };
\filldraw[fill=white, xshift=6cm,yshift=8cm]
(45:1.2) \foreach \x in {135,225,315,45} { -- (\x:1.2) };
\filldraw[fill=white, xshift=8cm,yshift=2cm]
(45:1.2) \foreach \x in {135,225,315,45} { -- (\x:1.2) };
\filldraw[fill=white, xshift=8cm,yshift=4cm]
(45:1.2) \foreach \x in {135,225,315,45} { -- (\x:1.2) };
\filldraw[fill=white, xshift=8cm,yshift=6cm]
(45:1.2) \foreach \x in {135,225,315,45} { -- (\x:1.2) };
\filldraw[fill=white, xshift=8cm,yshift=8cm]
(45:1.2) \foreach \x in {135,225,315,45} { -- (\x:1.2) };
\end{tikzpicture}
\end{center}
\caption{An image in $Z^2$ with $c_1$-adjacency having the same properties as in Example \ref{the-ex}.\label{4adj-ex}}
\end{figure}
Let $X$ be the digital image in Example \ref{the-ex}, and define two loops $f,g: [0,10]_Z \to X$ as follows:
\begin{align*}
f &= (x_1, x_2, \dots, x_9, x_{10}, x_1) \\
g &= (x_1, x_2, \dots, x_9, x_0, x_1)
\end{align*}
These loops are equivalent in $\Pi_1(X,x_1)$: consider the following trivial extensions
\begin{align*}
f' &= (x_1, x_2, x_3, \dots, x_9, x_{10}, x_1, x_1) \\
g' &= (x_1, x_1, x_2, \dots, x_8, x_9, x_0, x_1)
\end{align*}
These loops $f'$ and $g'$ are homotopic in one step, and so $f$ and $g$ are equivalent in $\Pi_1(X,x_1)$. Notice that the one-step equivalence above uses trivial extensions at the base point $x_1$. That is, there is some $t$ with $f'(t)=f'(t+1) = x_1$, and likewise for $g'$. In fact this is necessary for any equivalence between $f$ and $g$, as the following proposition shows:
\begin{Proposition}
Let $X$ be as in Example~\ref{the-ex}. Let $f$ and $g$ be the loops described above.
Let $f',g':[0,k]_Z \to X$ be trivial extensions of $f$ and $g$ that are homotopic by $H(t,s):[0,k]_Z \times [0,n]_Z \to X$. Then there is some time $p \in [0,n]_Z$
and intermediate stage of the homotopy $H$, i.e.,
$h: [0,k]_{Z} \rightarrow X$ defined by
$h(t) = H(t,p)$, such that $h(k-1)=h(k)= x_1$.
Similarly there is some $q \in [0,n]_Z$ and intermediate stage of the
homotopy $H$, i.e., $l: [0,k]_{Z} \rightarrow X$ defined by $l(t)=H(t,q)$,
such that $l(0)=l(1)=x_1$.
\end{Proposition}
\begin{proof}
We will prove the first statement; the second follows similarly. Suppose
that no intermediate loop $h$ obeys $h(k-1)=h(k)=x_1$.
Then we have $H(k-1,s) \neq x_1$ for all $s$. We must in particular have
$f'(k-1) \neq x_1$, and so $f'(k-1) = x_{10}$ since $f'$ is a trivial extension
of $f$.
Thus, considering $H(k-1,s)$ for various $s$ gives a path from
$H(k-1,0) = f'(k-1) = x_{10}$ to $g'(k-1) = x_0$ which never
passes through $x_1$. Because of the structure of our image $X$,
this path must at some point pass through $x_9$.
Thus there is some $r$ with $H(k-1,r) = x_9$. But $H(k,r) = x_1$
since all stages of $H$ are loops at $x_1$. This contradicts continuity
of $H$ from $H(k-1,r)$ to $H(k,r)$ since $x_9$ is not adjacent to $x_1$ in $X$.
\end{proof}
Thus we see that $f$ and $g$ are equivalent as loops in $\Pi_1(X,x_1)$, but this equivalence requires trivial extensions at the base point. This suggests a finer equivalence relation than the one used for the fundamental group, one in which loops are equivalent only by homotopies that do not extend the base point. Specifically, we call a loop $f$ \emph{tight at the basepoint (TAB)} $x_0$ when there is no $t$ with $f(t)=f(t+1)=x_0$. Two TAB loops are called TAB equivalent when there are TAB trivial extensions that are homotopic by a homotopy which is TAB in each stage.
Thus our example loops $f$ and $g$ above are equivalent in $\Pi_1(X,x_1)$, but not TAB equivalent, because any homotopy of trivial extensions must have a non-TAB intermediate stage. The equivalence classes using the TAB relation seem to have interesting and subtle structure, but they do not naturally form a group with respect to the product operation, as we show below.
Consider the product of $f$ and the reverse of $g$, which has the form:
\[ f*g^{-1} = (x_1,x_2, \dots, x_9,x_{10},x_1,x_0,x_9,\dots,x_2,x_1) \]
Note that $f*g^{-1}$ is nullhomotopic, using only TAB loops as intermediate steps. The first step of the nullhomotopy is as follows:
\begin{align*}
(x_1,x_2, \dots, x_9,x_{10},x_1,x_0,x_9,\dots,x_2,x_1) \mbox{ to}\\
(x_1,x_2, \dots, x_9,x_9,x_0,x_0,x_9,\dots,x_2,x_1),
\end{align*}
and then the loop deforms continuously to the constant map
$(x_1,x_1, \ldots, x_1)$ in an obvious way.
Since $f$ and $g$ are not TAB equivalent, but $f*g^{-1}$ is pointed nullhomotopic, the TAB relation, which is finer than the equivalence used in $\Pi_1(X,x_1)$, cannot be used to define a group. Nevertheless the TAB equivalence provides subtle and interesting information about loops in our space.
\begin{comment}
\section{Tight at basepoint homotopy}
In this section, we introduce a variant on the digital fundamental group,
motivated by Example~\ref{the-ex}.
\begin{Definition}
\label{TAB-def}
Let $(X,\kappa,p)$ be a pointed digital image. A $\kappa$-loop
$f: [0,m]_Z \rightarrow X$ with $m>1$ and $f(0)=f(m)=p$ is
{\rm tight at the basepoint (TAB)} if $0 \leq i < m$
implies $\{f(i),f(i+1)\} \neq \{p\}$. We also consider any constant loop at $p$ to be
TAB.
\end{Definition}
Next, we define a TAB homotopy of TAB loops.
\begin{Definition}
\label{TAB-htpy}
Let $(X,\kappa,p)$ be a pointed digital image.
Let $f,g: [0,m]_Z \rightarrow X$ be TAB $\kappa$-loops with basepoint $p$.
A homotopy $H$ from $f$ to $g$ that holds the endpoints fixed
is a {\rm TAB homotopy} if each
intermediate loop $H_t$, $0 \leq t \leq m$, is a TAB loop. Then we say
$f$ and $g$ are {\rm TAB homotopic}.
\end{Definition}
\begin{Proposition}
\label{equiv-rel}
Let $(X,\kappa,p)$ be a pointed digital image.
TAB homotopy among $\kappa$-loops into $(X,\kappa,p)$ based at $p$ is an
equivalence relation.
\end{Proposition}
{\em Proof}: Elementary, and left to the reader. $\Box$
Next, we define a somewhat more general kind of homotopy relationship. Some
care must be used for the case in which one of the loops is trivial.
\begin{Definition}
\label{TAB-equiv-rel}
Let $(X,\kappa,p)$ be a pointed digital image such that the $\kappa$-component
of $p$ in $X$ has more than one point.
Let $f$ and $g$ be TAB $\kappa$-loops with basepoint $p$. We say $f$ and $g$
are {\rm TAB equivalent} if there are trivial extensions
$\overline{f}$ of $f$ and $\overline{g}$ of $g$ that are TAB and
TAB homotopic.
\end{Definition}
\begin{Theorem}
\label{TAB-equiv-rel-thm}
TAB equivalence is an equivalence relationship.
\end{Theorem}
{\em Proof}: The details of the proof are similar to those for ordinary
homotopy.
\begin{itemize}
\item Reflexivity follows from the use of trivial homotopies. I.e.,
if $f: [0,m]_Z \rightarrow X$ is a loop, then the map
$H: [0,m]_Z \times [0,1]_Z \rightarrow X$ defined by
$H(x,t)=f(x)$ is a TAB equivalence from $f$ to $f$.
\item Symmetry follows by reversing homotopies.
If $h: [0,m]_Z \times [0,n]_Z \rightarrow X$ is a TAB homotopy
between trivial extensions $\overline{f}$ and $\overline{g}$ of
$p$-based loops $f$ and $g$, respectively, then the map
$k: [0,m]_Z \times [0,n]_Z \rightarrow X$ defined by
$k(x,t) = h(x,n-t)$ is easily seen to be a TAB homotopy between
$\overline{g}$ and $\overline{f}$.
\item Transitivity follows by combining homotopies, being careful where
necessary to preserve the TAB property. E.g., if further trivial
extensions are needed to ensure that all loop representatives are
defined on the same interval, these trivial extensions can be taken as
constant on intervals not including the endpoints of the domain.
$\Box$
\end{itemize}
Given a TAB loop $f$ based at $p \in X$, we denote the equivalence class
of $f$ discussed above by $[f]$. We denote by
$\Pi_1^{TAB,\kappa}(X,p)$, or
$\Pi_1^{TAB}(X,p)$ when the adjacency $\kappa$ is understood,
the set of all such TAB loop classes.
In parallel with the fundamental group, we show how to make a group out of
$\Pi_1^{TAB,\kappa}(X,p)$.
\begin{Definition}
\label{*-op}
For $[f],[g] \in \Pi_1^{TAB,\kappa}(X,p)$, let
$[f]*[g] \in \Pi_1^{TAB,\kappa}$ be as follows.
\begin{enumerate}
\item If $f: [0,m]_Z \rightarrow X$ and $g: [0,n]_Z \rightarrow X$ are
nonconstant loops, then
$[f]*[g] = [h]$, where $h: [0,m+n]_Z \rightarrow X$
is defined by
\[ h(t) = \left \{ \begin{array}{ll}
f(t) & \mbox{if } 0 \leq t \leq m; \\
g(t-m) & \mbox{if } m \leq t \leq m+n.
\end{array} \right . \]
\item If $f$ is a constant loop, then $[f]*[g]=[g]$.
\item If $g$ is a constant loop, then $[f]*[g]=[f]$.
\end{enumerate}
\end{Definition}
It is easy to see that in all cases, $[f]*[g]$ is well defined, and that the
$*$ operation preserves the TAB property.
\begin{Theorem}
\label{TAB-group}
$\Pi_1^{TAB,\kappa}(X,p)$ is a group, using the $*$ operation.
\end{Theorem}
{\em Proof}: Let $e: \{0\} \rightarrow X$ with $e(0)=p$ be the trivial loop at $p$. Then $[e]$ is an identity element for $\Pi_1^{TAB,\kappa}(X,p)$.
Associativity follows easily from Definition~\ref{*-op}.
For inverses: given some TAB $\kappa$-loop $f: [0,k]_Z \to X$, define $[f]^{-1} = [f']$, where $f'$ is the reverse loop $f'(t) = f(k-t)$. Then we have the following.
\begin{itemize}
\item Clearly, $[e]^{-1}=[e]$.
\item Let $f: [0,k]_Z \rightarrow X$ be a nontrivial TAB $\kappa$-loop
based at $p$. Then $[f]*[f]^{-1} = [h]$, where
\[ h(t) = \left \{ \begin{array}{ll}
f(t) & \mbox{for } 0 \leq t \leq k; \\
f'(t-k) = f(2k-t) & \mbox{for } k \leq t \leq 2k.
\end{array} \right . \]
It is easily seen that $h$ is a well defined TAB loop based at $p$.
We need to show that $h$ is homotopic to a constant loop at $p$ via
a TAB homotopy. Consider the
function $H: [0,2k]_Z \times [0,k-1]_Z \rightarrow X$ given by
\[ H(t,s) = \left \{ \begin{array}{ll}
p & \mbox{if } t=0; \\
f(1) & \mbox{if } 1 \leq t \leq s+1: \\
f(t-s) & \mbox{if } s+1 \leq t \leq k; \\
f(2k-t-s) & \mbox{if } k \leq t \leq 2k-s-2; \\
f(1) & \mbox{if } 2k - s - 2 < t \leq 2k-1; \\
p & \mbox{if } t=2k.
\end{array} \right . \]
It is easily verified that $H$ is well defined.
Further, it follows from the assumption that $f$ is a TAB loop that
$H$ is a TAB homotopy.
The final step of the homotopy $H$, corresponding to $s=k-1$, satisfies
\begin{equation}
\label{final-step}
H(t,k-1) = \left \{ \begin{array}{ll}
p & \mbox{if } t=0; \\
f(1) & \mbox{if } 1 \leq t \leq k; \\
f(1) & \mbox{if } k \leq t \leq k; \\
f(k-t+1) & \mbox{if } k \leq t \leq k-1; \\
f(1) & \mbox{if } k - 1 < t \leq 2k-1; \\
p & \mbox{if } t=2k,
\end{array} \right \}
=
\left \{ \begin{array}{ll}
p & \mbox{if } t=0; \\
f(1) & \mbox{if } 1 \leq t \leq 2k-1; \\
p & \mbox{if } t=2k.
\end{array} \right \}
\end{equation}
Note that we can extend the function $H$ by defining
$H': [0,2k]_Z \times [0,k]_Z \rightarrow X$ by
\[ H'(t,s) = \left \{ \begin{array}{ll}
H(t,s) & \mbox{if } 0 \leq s < 2k; \\
p & \mbox{if } s=2k.
\end{array} \right . \]
It follows from equation~(\ref{final-step}) that $H'$ is a TAB homotopy from
$h$ to the trivial loop based at $p$. Hence $[f']$ is a right inverse for $[f]$.
Similarly, $[f']$ is a left inverse for $[f]$, so $[f]$ and $[f']$ are inverses. $\Box$
\end{itemize}
We will see that the usual digital fundamental group and the TAB fundamental group can be significantly different, though they are of course related. When they are different, the TAB group has more structure than the usual group, as the following theorem shows.
\begin{Theorem}
\label{quot-thm}
$\Pi_1^{TAB}(X,p)$ is isomorphic to a quotient of $\Pi_1(X,p)$.
\end{Theorem}
\begin{proof}
Let $f: [0,m]_Z \rightarrow X$ be a $p$-based TAB loop. Then
$f$ is a $p$-based loop. Thus, we have the map
$j: \Pi_1^{TAB}(X,p) \rightarrow \Pi_1^{}(X,p)$ defined by
$j([f]) = [f]$.
It is easily seen that $j$ is a homomorphism. The function $j$ is in fact surjective, since any loop class includes a representative loop which is TAB.
Thus by the first isomorphism theorem, $\Pi_1^{}(X,p)$ is isomorphic to the quotient group $\Pi_1^{TAB}(X,p) / \ker(j)$.
\end{proof}
We use Example~\ref{the-ex} to show that the usual digital fundamental group and the TAB fundamental group can be significantly different. We saw in Proposition~\ref{htpy-equiv} that the images $(X,c_2)$ and $(Y,c_2)$ are homotopy equivalent, hence have isomorphic fundamental groups even though, by Proposition~\ref{proof-of-example}, these images are not pointed homotopy equivalent. Since $Y$ is a simple closed curve, the fundamental group of $Y$ (and thus of $X$) is isomorphic to $Z$. We will show that the TAB fundamental group of $X$ is not $Z$.
Let $(X,x_0)$ be the pointed space in Example \ref{the-ex}, and define two loops $f,g: [0,11]_Z \to X$ as follows:
\begin{align*}
f &= (x_1, x_2, \dots, x_9, x_{10}, x_1) \\
g &= (x_1, x_2, \dots, x_9, x_0, x_1)
\end{align*}
\begin{Proposition}\label{fg}
In $\Pi_1^{TAB}(X,x_1)$ we have $[f] \neq [g]$.
\end{Proposition}
\begin{proof}
Our proof is by contradiction, so we assume that $[f]=[g]$, that is, there are trivial extensions $f',g': [0,k]_Z \to X$ such that $f'$ and $g'$ are TAB homotopic, say by some homotopy $H:[0,k]_Z \times [0,s]_Z \to X$.
Since each stage of $H$ is a loop at $x_1$ we have $H(0,t) = H(k,t) = x_1$ for all $t$, and since $H$ is TAB we have $H(k-1,t) \neq x_1$ for all $t$. In particular $f'(k-1) \neq x_1$ and so $f'(k-1) = x_{10}$ since $f'$ is a trivial extension of $f$. Similarly $g'(k-1) = x_0$.
Now let $h(t) = H(k-1,t)$, so $h$ is a path from $h(0)=H(k-1,0)=f'(k-1)=x_{10}$ to $h(s) = H(k-1,s)=g'(k-1) = x_0$. Since $H$ is tight at the basepoint this path $h$ cannot pass through $x_1$. Thus there is some $t_0$ for which $h(t_0) = x_9$, since $x_9$ must be used by any path from $x_{10}$ to $x_0$ not passing through $x_1$.
But $h(t_0)=x_9$ implies that $H(k-1,t_0) = x_9$, which is not adjacent to $H(k,t_0)=x_1$, which contradicts continuity of $H$.
\end{proof}
The above argument easily generalizes to show that $\Pi_1^{TAB}(X,x_1)$ has considerably more structure than $Z$. In particular we have:
\begin{Proposition}
$\Pi_1^{TAB}(X,x_1)$ is not abelian.
\end{Proposition}
\begin{proof}
Consider the example loops $f$ and $g$ above, and the loop classes $[f]*[g]$ and $[g]*[f]$. By the definitions of $g$ and the operation $*$ we have $[f]*[g] = [h_1]$ where $h_1$ is a loop at $x_1$ having $x_0$ as its second-to-last point. Similarly we have $[g]*[f] = [h_2]$, where $h_2$ is a loop at $x_2$ having $x_{10}$ as its second-to-last point. Now exactly the same argument as in Proposition \ref{fg} shows that $[h_1] \neq [h_2]$, by following the second-to-last point through the homotopy.
Thus $[f]*[g] \neq [g]*[f]$, and so $\Pi_1^{TAB}(X,x_1)$ is not abelian.
\end{proof}
It seems likely that $\Pi_1^{TAB}(X,x_1)$ is a free group generated by $[f]$ and $[g]$, but we will not prove this.
\end{comment}
\section{A new formulation of the fundamental group}
\label{ec-section}
The equivalence relation of Definition~\ref{loop-class} used to define the fundamental group relies on trivial extensions, which are often cumbersome to handle. In this section we give an equivalent definition of the fundamental group which does not require trivial extensions. Our construction instead is based on \emph{eventually constant paths}. Let $\mathbb{N}= \{1,2,\dots\}$ denote the natural numbers,
and $\mathbb{N}^* = \{0\} \cup \mathbb{N}$. We consider $\mathbb{N}^*$ to be a digital image with 2-adjacency.
\begin{Definition}
Given a digital image $X$, a continuous function $f:\mathbb{N}^* \to X$ is called an \emph{eventually constant path} or \emph{EC path} if there is some point $c\in X$ and some $N \geq 0$ such that $f(x)=c$ whenever $x\geq N$. When convenient we abbreviate the latter by $f(\infty) = c$. The \emph{endpoints} of an EC path $f$ are the two points $f(0)$ and $f(\infty)$.
If $f$ is an EC path and $f(0)=f(\infty)$, we say $f$ is an \emph{EC loop}, and $f(0)$ is called the basepoint.
We say that a homotopy $H$ between EC paths is an \emph{EC homotopy} when the function $H_t: \mathbb{N}^* \rightarrow X$ defined by
$H_t(s) = H(s,t)$ is an EC path for all $t \in [0,k]_{Z}$. To indicate an EC homotopy, we write $f \simeq^{\mathrm{EC}} g$, or $f \simeq^{\mathrm{EC}}_{\kappa} g$ if it is desirable to
state the adjacency $\kappa$ of $X$.
We say an EC homotopy $H$ \emph{holds the endpoints fixed} when $H_t(0) = f(0) = g(0)$ and there is a $c \in \mathbb{N}^*$ such that
$n \geq c$ implies $H_t(n) = f(n) = g(n)$ for all $t$.
$\Box$
\end{Definition}
Not all homotopies of EC paths are EC homotopies, as the following example shows.
\begin{Example}
\label{wobbler}
Let $f,g: \mathbb{N}^* \rightarrow [0,1]_{Z}$ be defined by
$f(0)=g(0)=0$, $f(n)=g(n)=1$ for $n>0$.
Let $H: \mathbb{N}^* \times [0,2]_{Z} \rightarrow [0,1]_{Z}$ be defined by
$H_0=H_2=f=g$,
$ H_1(s)=0$ if $s$ is even, $H_1(s)= 1$ if $s$ is odd. Then $H$ is a homotopy
from $f$ to $g$ that is not an EC homotopy.
\end{Example}
\begin{proof} It is easy to see that $H$ is a homotopy. However, $H_1$
is not an EC path. The assertion follows.
\end{proof}
A familiar argument shows that EC homotopy is an equivalence relation.
\begin{Proposition}
\label{EC-equiv-relation}
EC homotopy and EC homotopy holding the endpoints fixed are equivalence relations among EC paths.
\end{Proposition}
\begin{proof} We give a proof without the assumption of endpoints being
held fixed. The same argument can be used with obvious modifications to
obtain the assertion for endpoints held fixed.
{\em Reflexive}: Given an EC path $f: \mathbb{N}^* \rightarrow X$, clearly the function
$H: \mathbb{N}^* \times \{ 0 \} \rightarrow X$ given by
$H(x,0) = f(x)$ shows $f \simeq^{\mathrm{EC}} f$.
{\em Symmetric}: If $H: \mathbb{N}^* \times [0,m]_{Z} \rightarrow X$ is an EC homotopy from
$f$ to $g$, then it is easy to see that the function
$H': \mathbb{N}^* \times [0,m]_{Z} \rightarrow X$ defined by
\[ H'(x,t) = \left \{ \begin{array}{ll}
H(x,m-t) & \mbox{if } 0 \leq t \leq m; \\
f(0) & \mbox{if } t \geq m,
\end{array} \right .
\]
shows $g \simeq^{\mathrm{EC}} f$.
{\em Transitive}: Suppose $H: \mathbb{N}^* \times [0,m_1]_{Z} \rightarrow X$ is
an EC homotopy from $f$ to $g$, and $K : \mathbb{N}^* \times [0,m_2]_{Z} \rightarrow X$
is an EC homotopy from $g$ to $h$. Then the function
$L: \mathbb{N}^* \times [0,m_1+m_2]_{Z} \rightarrow X$ defined by
\[ L(x,t) = \left \{ \begin{array}{ll}
H(x,t) & \mbox{if } 0 \leq t \leq m_1; \\
K(x,t-m_1) & \mbox{if } m_1 \leq t \leq m_2,
\end{array} \right .
\]
is an EC homotopy from $f$ to $h$.
\end{proof}
Homotopy of trivial extensions of loops can be easily stated in terms of EC homotopy of the corresponding EC loops. The latter formulation is preferable since it does not require trivial extensions, which obviates the need for several technical lemmas. For example the proof given below for Proposition \ref{EC-mult-well-defined} is much easier than the corresponding statement for trivial extensions (see \cite[Proposition 4.8]{Boxer}, which is only a sketch of a proof from \cite{Khalimsky}); and the proof given below for
Theorem~\ref{Pi1-iso-thm} is somewhat simpler, being
based on EC homotopy, than it would have been had we had to
construct trivial extensions.
Given a path $f: [0,m]_{Z} \rightarrow X$, we denote by
$f_{\infty}: \mathbb{N}^* \rightarrow X$ the function defined by
\[ f_{\infty}(n) = \left \{ \begin{array}{ll}
f(n) & \mbox{if } 0 \leq n \leq m; \\
f(m) & \mbox{if } n \geq m.
\end{array} \right .
\]
Given an EC path $g: \mathbb{N}^* \rightarrow X$, let
\[ N_g = \min\{ m \in \mathbb{N}^* \, | \, n \geq m \mbox{ implies } g(n)=g(m) \} \]
and let $g_-: [0,N_g]_{Z} = g|_{[0,N_g]_{Z}}$. We have the following.
\begin{Proposition}
\label{ec-extension-restriction}
Let $X$ be a digital image.
\rm{a)} Let $f: \mathbb{N}^* \rightarrow X$ be an EC path. Then $(f_-)_{\infty}=f$.
\rm{b)} Let $f: [0,m]_{Z} \rightarrow X$ be a path in $X$. Then $f$ is a trivial extension of $(f_{\infty})_-$. We have $f=(f_{\infty})_-$ if and only if either $m=0$ or $m>0$ and $f(m-1)\neq f(m)$.
\end{Proposition}
\begin{proof} These assertions are immediate consequences of the
definitions above.
\end{proof}
\begin{Lemma}\label{ec-homotopy}
Let $f,g:[0,m]_Z \to X$ be paths with $f\simeq g$. Then $f_\infty \simeq^{\mathrm{EC}} g_\infty$. If the homotopy from $f$ to $g$ holds the endpoints fixed, then so does the induced EC homotopy from $f_\infty$ to $g_\infty$.
\end{Lemma}
\begin{proof}
Let $H:[0,m]_Z \times [0,k]_Z \to X$ be a homotopy of $f$ to $g$. Consider $G: \mathbb{N}^* \times [0,k]_Z \to X$, defined as follows:
\[
G(s,t) = \begin{cases}
H(s,t) & \text{ if } s \le m \\
H(m,t) & \text{ if } s > m.
\end{cases}
\]
Clearly $G$ is an EC homotopy of $f_\infty$ to $g_\infty$. Further, $G$ holds the
endpoints fixed if $H$ does so.
\end{proof}
\begin{Lemma}
\label{cutoff}
Let $f$ and $g$ be EC homotopic EC paths in $X$.
Then $f_-$ and $g_-$ have homotopic trivial extensions.
If $f$ and $g$ are homotopic holding the endpoints fixed,
then $f_-$ and $g_-$ have trivial extensions that are homotopic holding
the endpoints fixed.
\end{Lemma}
\begin{proof} Let $N_f, N_g$ be as defined above. Without loss of generality,
$N_f \leq N_g$. Let $H: \mathbb{N} \times [0,m]_{Z} \rightarrow X$ be a homotopy from
$f$ to $g$. Let $H': [0,N_g] \times [0,m]_{Z} \rightarrow X$ be the restriction of $H$ to
$[0,N_g] \times [0,m]_{Z}$. It is easily seen that $H'$ is a homotopy between a
trivial restriction $f'$ of $f_-$ and the function $g_-$, where
$f': [0,N_g]_{Z} \rightarrow X$ is defined by
\[ f'(n) = \left \{ \begin{array}{ll}
f(n)=f_-(n) & \mbox{if } 0 \leq n \leq N_f; \\
f(N_f) & \mbox{if } N_f \leq n \leq N_g.
\end{array} \right .
\]
Further, if $H$ holds the
endpoints fixed, then so does $H'$.
\end{proof}
\begin{Lemma}\label{te-homotopy}
Let $f:[0,m]_Z\to X$ be a loop and $\bar f:[0,n]_Z \to X$ be a trivial extension of $f$. Then $f_\infty$ and $\bar f_\infty$ are EC homotopic with fixed endpoints.
\end{Lemma}
\begin{proof}
We will prove the Lemma in the case that $\bar f$ is obtained from $f$ by inserting a single trivial loop. The full result follows by induction. Specifically, let $f= f_1 * f_2$ and $\bar f = f_1 * c * f_2$, where $c$ is a trivial loop. Say that $f_1:[0,m]_Z \to X$ and $f_2: [0,n]_Z \to X$ and $c:[0,k]_Z \to X$. Then consider $H: \mathbb{N}^* \times [0,k]_Z \to X$ given by:
\[ H(s,t) = \begin{cases}
f_1(s) & \text{ if } 0 \leq s \leq m; \\
c(s-m) & \text{ if } m \le s \le m+t; \\
f_2(s-(m+t)) & \text { if } m+t \le s \le m+t+n; \\
x_0 & \text{ if } m+t+n \le s.
\end{cases}
\]
At time stage $t$ we have
$H_t = (f_1 * c_{|_{[0,t]_{Z}}} * f_2)_{\infty}$, so
$H$ is an EC homotopy of $f_\infty$ to $\bar f_\infty$ as desired. Further,
$H$ fixes the endpoints, since $H(0,t)=f_1(0)$ for all $t$ and $H(x,t)=f_2(n)$
for all $x \geq m+t+n$ and all $t$.
\end{proof}
\begin{Theorem}\label{ec-te-equiv}
Let $f$ and $g$ be loops in $X$ having some common basepoint $p$. Then there are trivial extensions $\bar f, \bar g$ of $f, g$ respectively with $\bar f \simeq \bar g$ with fixed endpoints if and only if $f_\infty$ and $g_\infty$ are EC homotopic with fixed endpoints.
\end{Theorem}
\begin{proof}
First we assume that there are trivial extensions $\bar f,\bar g$ with $\bar f\simeq \bar g$ fixing endpoints. Then by Lemmas \ref{te-homotopy} and \ref{ec-homotopy} we have $f_\infty \simeq^{\mathrm{EC}} \bar f_\infty \simeq^{\mathrm{EC}} \bar g_\infty \simeq^{\mathrm{EC}} g_\infty$ and all homotopies fix the endpoints as desired.
For the converse assume that $f_\infty \simeq^{\mathrm{EC}} g_\infty$ with fixed endpoints. Let $H:\mathbb{N}^* \times [0,k]_Z \to X$ be the EC homotopy. Since $H$ fixes the endpoints (at $p$) and has only finitely many stages, there must be some $M$ such that $H(s,t) = p$ for all $s\ge M$ and for all $t$.
Let $\bar f,\bar g:[0,M]_Z \to X$ be the restrictions of $f_\infty,g_\infty$ respectively to $[0,M]_Z$. Then $\bar f = f*c$ is a trivial extension of $f$, where $c$ is a trivial loop at $p$. Similarly $\bar g$ is a trivial extension of $g$.
Let $\bar H:[0,M]_Z \times [0,k]_Z \to X$ be the restriction of $H$ to $[0,M]_Z \times [0,k]_Z$. Then $H$ is a homotopy of $\bar f$ to $\bar g$ fixing the endpoints as desired.
\end{proof}
It is natural to overload the $*$ notation as follows.
\begin{Definition}
\label{EC-star}
For $x_0 \in X$, let $f_0, f_1: \mathbb{N}^* \rightarrow X$ be $x_0$-based EC loops in $X$.
Define $f_0 * f_1: \mathbb{N}^* \rightarrow X$ by
\[ f_0 * f_1(n) = \left \{ \begin{array}{ll}
f_0(n) & \mbox{if } 0 \leq n \leq N_{f_0}; \\
f_1(n-N_{f_0}) & \mbox{if } N_{f_0} \leq n.~~~~~ \Box
\end{array} \right .
\]
\end{Definition}
It is easily seen that $f_0 * f_1$ is well defined and is an EC loop in $X$.
The $*$ operator on EC loops has the following properties.
\begin{Proposition}
\label{star-props}
\begin{itemize}
\item Let $f,g: \mathbb{N}^* \rightarrow X$ be $x_0$-based EC loops, for some $x_0 \in X$. Then $f_-*g_-=(f*g)_-$.
\item Let $f: [0,m]_{Z} \rightarrow X$, $g: [0,n]_{Z} \rightarrow X$ be $x_0$-based EC loops, for some $x_0 \in X$. Then $f_{\infty} * g_{\infty} = (f*g)_{\infty}$.
\end{itemize}
\end{Proposition}
\begin{proof} These properties are simple consequences of
Definition~\ref{EC-star}.
\end{proof}
\begin{Lemma}\label{EC-mult-half}
Let $f, g, g'$ be EC loops in $X$ at a common basepoint, with $g\simeq^{\mathrm{EC}} g'$ holding the endpoints fixed. Then $f * g \simeq^{\mathrm{EC}} f * g'$ holding the endpoints fixed.
\end{Lemma}
\begin{proof}
Let $H:\mathbb{N}^* \times [0,m] \to X$ be the EC homotopy from $g$ to $g'$, and let $L:\mathbb{N}^* \times [0,m] \to X$ be given by
\[ L(s,t) = (f * H_t)(s). \]
Then $L$ is a EC homotopy from $f*g$ to $f*g'$ holding the endpoints fixed as desired.
\end{proof}
In order to prove Proposition~\ref{EC-mult-well-defined} below,
we must take care in how we mimic the proof of Lemma~\ref{EC-mult-half}
on the first factors of the * products, as shown by the following.
\begin{Example}
\label{htpy-discont-in-t}
Let $f,g: \mathbb{N}^* \rightarrow [0,1]_{Z}$ be defined by
\[ f(n)=g(n)= \left \{ \begin{array}{ll}
n & \mbox{if } n \in \{0,1,2\}; \\
1 & \mbox{if } n=3; \\
0 & \mbox{if } n > 3.
\end{array} \right .
\]
Then there is an EC homotopy
$H: \mathbb{N}^* \times [0,2]_{Z} \rightarrow [0,1]_{Z}$ from $f$ to $f$ such that the function
$K: \mathbb{N}^* \times [0,2]_{Z} \rightarrow [0,1]_{Z}$ defined by
$K(n,t) = H_t(n) * g(n)$ is not continuous in $t$, where
$H_t: \mathbb{N}^* \rightarrow [0,1]_{Z}$ is the induced function $H_t(n)=H(n,t)$.
\end{Example}
\begin{proof} Define $H(n,t)$ by $H(n,0)=f(n)=g(n)=H(n,2)$,
\[ H(n,1)= \left \{ \begin{array}{ll}
f(n) & \mbox{if } n \neq 5; \\
1 & \mbox{if } n=5.
\end{array} \right .
\]
It is easy to see that $H$ is a homotopy. However,
$K=H_0 * g = H_2 * g$ and $L=H_1*g$ are represented respectively by the sequences
\[ (K(0),K(1),K(2),\ldots)=(0,1,2,1,0,1,2,1,0,0,\ldots) \]
\[ (L(0),L(1),L(2),\ldots)=(0,1,2,1,0,1,0,1,2,1,0,0,\ldots) \]
In particular, $H_0*g(6)=2$ and $H_1 * g(6)=0$, so at $n=6$,
$H_t*g$ is not continuous in $t$.
\end{proof}
\begin{Proposition}
\label{EC-mult-well-defined}
Let $f,f',g,g'$ be EC loops in $X$ at a common basepoint such that
$f\simeq^{\mathrm{EC}} f'$ and $g \simeq^{\mathrm{EC}} g'$ with both homotopies
holding the endpoints fixed.
Then we have $f*g\simeq^{\mathrm{EC}} f' * g'$ holding the endpoints fixed.
\end{Proposition}
\begin{proof}
By Lemma \ref{EC-mult-half} we have $f*g \simeq^{\mathrm{EC}} f*g'$ holding the endpoints fixed.
By an argument similar to that of the proof of Lemma \ref{EC-mult-half} we will show that $f*g' \simeq^{\mathrm{EC}} f' * g'$.
Example~\ref{htpy-discont-in-t} shows that
$H_t*g'$ will not necessarily be continuous in $t$; however, this is easily fixed by inserting an extra constant segment in the first factor. In particular, let $H: \mathbb{N}^* \times [0,m]_{Z} \rightarrow X$ be an
EC homotopy from $f$ to $f'$ that holds the endpoints fixed. Let
$M = \max\{N_{H_t} \, | \, t \in [0,m]_{Z}\}$. For each $t \in [0,m]_{Z}$,
let $c_t : [0, M-N_{H_t}]_{Z} \rightarrow \{x_0\}$ be a constant function.
Then the function $K: \mathbb{N}^* \times [0,m]_{Z} \rightarrow X$
defined by
$K(n,t) = (H_t * c_t * g')(n)$
is an EC homotopy from $f*g'$ to $f' * g'$ that holds the endpoints
fixed.
Thus by transitivity of EC homotopy we have $f*g\simeq^{\mathrm{EC}} f'*g'$,
holding endpoints fixed.
\end{proof}
\begin{comment}
\begin{Proposition}
\label{EC-mult-well-defined}
Let $x_0 \in X$.
Let $f_0,f_1,g_0,g_1$ be $x_0$-based EC loops in $X$ such that
$f_0 \simeq^{\mathrm{EC}} f_1$ and $g_0 \simeq^{\mathrm{EC}} g_1$ with both homotopies
holding the endpoints fixed.
Then we have $f_0 * f_1 \simeq^{\mathrm{EC}} g_0 * g_1$ holding the endpoints fixed.
\end{Proposition}
\begin{proof}
By Lemma~\ref{cutoff}, $f_{0-}$ and $f_{1-}$ have trivial extensions
$\overline{f_{0-}}$ and $\overline{f_{1-}}$ that are homotopic holding the endpoints fixed, and
$g_{0-}$ and $g_{1-}$ have trivial extensions
$\overline{g_{0-}}$ and $\overline{g_{1-}}$ that are homotopic holding the endpoints fixed.
By Proposition~\ref{well-defined},
$\overline{f_{0-}} * \overline{g_{0-}}$ and
$\overline{f_{1-}} * \overline{g_{1-}}$ are homotopic holding the endpoints fixed. Since these
are respectively trivial extensions of $f_{0-}*g_{0-}$ and $f_{1-}*g_{1-}$, it
follows from Lemma~\ref{te-homotopy} that
$(f_{0-}*g_{0-})_{\infty}=f_0*g_0$ and
$(f_{1-}*g_{1-})_{\infty}=f_1*g_1$ are EC homotopic holding the endpoints fixed.
\end{proof}
\end{comment}
Let $G(X,x_0)$ be the set of all EC homotopy classes of EC loops in $X$ based at $x_0$.
\begin{Proposition}
\label{EC-grp}
$G(X,x_0)$ with the $\cdot$ operation defined by
$[f] \cdot [g] = [f*g]$ is a group.
\end{Proposition}
\begin{proof} By Proposition~\ref{EC-mult-well-defined}, the $\cdot$ operation
is closed and well defined on $G(X,x_0)$. Clearly, the EC pointed
homotopy class of the constant map $c(n)=x_0$ for all $n \in \mathbb{N}$ is the
identity element. Given an $x_0$-based EC loop $f: \mathbb{N}^* \rightarrow X$, the function
$g: \mathbb{N}^* \rightarrow X$ defined by
\[ g(n)= \left \{ \begin{array}{ll}
f(N_f -n) & \mbox{if } 0 \leq n \leq N_f; \\
x_0 & \mbox{if } n \geq N_f,
\end{array} \right .
\]
gives an inverse for $[f]$.
\end{proof}
We have the following analog of Theorem~\ref{induced}.
\begin{Theorem}
\label{ec-induced}
Suppose $F: (X, \kappa, x_0) \rightarrow (Y,\lambda, y_0)$
is a pointed continuous function. Then
$F$ induces a homomorphism
$F_*: G(X,x_0) \rightarrow
G(Y,y_0)$
defined by $F_*([f])=[F \circ f]$.
\end{Theorem}
\begin{proof} Given $x_0$-based EC loops $f,g: \mathbb{N} \rightarrow X$, we have, by
using Propositions~\ref{ec-extension-restriction} and ~\ref{star-props},
\begin{align*}
F([f*g])&=[F\circ (f*g)]= [F \circ ((f*g)_-)_{\infty} ]
= [((F \circ f_-) * (F \circ g_-))_{\infty}] \\
&= [(F \circ f_-)_{\infty} * (F \circ g_-)_{\infty}]
= [(F\circ f) * (F \circ g)].
\end{align*}
The assertion follows.
\end{proof}
The main result of this section is the following.
\begin{Theorem}
\label{iso-groups}
Given a digital image $X$ and a point $x_0 \in X$, the groups $G(X,x_0)$ and
$\Pi_1(X,x_0)$ are isomorphic.
\end{Theorem}
\begin{proof} Let $F: \Pi_1(X,x_0) \rightarrow G(X,x_0)$ be defined by
$F([f]_X)=[f_{\infty}]_X$, where $[f_{\infty}]_X$ is the set of EC loops that are $x_0$-based
in $X$ and are EC homotopic in $X$ to $f_{\infty}$ holding the endpoints fixed.
From Lemma~\ref{cutoff}, $F$ is one-to-one. Also, $F$ is onto, since
given an $x_0$-based EC loop $f$, we have $[f]=F([f_-])$.
From Proposition~\ref{EC-mult-well-defined}, $F$ is a homomorphism.
The assertion follows.
\end{proof}
\section{Homotopy equivalence and fundamental groups}
In the paper \cite{Boxer05a}, it is asserted that digital images that are
(unpointed) homotopy equivalent have isomorphic fundamental groups. However,
the proof of this assertion is
incorrect. Roughly, the flaw in the argument given in~\cite{Boxer05a} is
that insufficient care was given to making sure that a certain
homotopy between two loops holds the endpoints fixed.
In this section, we give a correction.
\begin{comment}
\begin{Proposition}
\label{change-base}
Suppose $F: (X,\kappa) \rightarrow (Y,\lambda)$ and $G: (Y,\lambda) \rightarrow (X,\kappa)$ are maps that
realize a homotopy equivalence, i.e., $F \circ G \simeq_{\lambda} 1_Y$ and
$G \circ F \simeq_{\kappa} 1_X$. Let $p \in X$, let $H: X \times [0,m]_{\mathbf Z} \rightarrow X$ be a homotopy between
$1_X$ and $G \circ F$, and let $f$ be a $p$-based loop in $X$.
Then there is a trivial extension $\overline{f}$ of $f$
and a homotopy from $\overline{f}$ to $(G \circ F) \circ \overline{f}$
that holds the endpoints fixed.
\end{Proposition}
{\em Proof}:
Let $q: [0,m]_Z \rightarrow X$ be the path from $p$ to $(G \circ F)(p)$ defined by
$q(t)=H(p,t)$. By taking a trivial extension of $f$ or of $q$ if necessary, we may
assume without loss of generality that
$f: [0,m]_{\bf Z} \rightarrow X$.
For $s \in [0,m]_{\bf Z}$, let
$q_s: [0,m]_{\bf Z} \rightarrow X$ be the path
from $q(0)=p$ to $q(s)=H(p,s)$ given by
$ q_s(t) = q(\min\{s,t\})$. Let
$\overline{f}: [0,3m]_{\bf Z} \rightarrow X$ be defined by
\[ \overline{f}(s) = \left \{ \begin{array}{ll}
p & \mbox{if } 0 \leq s \leq m; \\
f(s-m) & \mbox{if } m \leq s \leq 2m; \\
p & \mbox{if } 2m \leq s \leq 3m.
\end{array} \right .
\]
Clearly $\overline{f}$ is a trivial extension of $f$.
Let $K: [0,3m]_{\bf Z} \times [0,m]_{\bf Z}
\rightarrow X$ be the map
\[K(t,s) = (q_s * H_s * q_s^{-1})(t). \]
It is easily seen that $K$ is a homotopy from $\overline{f}$ to $(G \circ F) \circ \overline{f}$ that holds the endpoints fixed. $\Box$
\end{comment}
\begin{Theorem}
\label{change-basepoints}
{\rm \cite{Boxer99}}
Let $(X,\kappa)$ be a digital image and let $p,r$ be points of the same
$\kappa$-component
of $X$. Let $q$ be a $\kappa$-path in $X$ from $p$ to $r$. Then the induced function
$q_{\#}: \Pi_1^{\kappa}(X,p) \rightarrow \Pi_1^{\kappa}(X,r)$ defined by
$q_{\#}([f]) = [q^{-1}*f*q]$ is an isomorphism.
\qed
\end{Theorem}
Theorem~\ref{change-basepoints} was proven in ~\cite{Boxer99} for
the version of the fundamental
group based on finite loops. However, essentially the same argument
makes Theorem~\ref{change-basepoints} valid for the version of the
fundamental group based on EC loops, stated below.
\begin{Corollary}
\label{EC-change-basepoints}
Let $(X,\kappa)$ be a digital image and let $p,r$ be points of the same
$\kappa$-component
of $X$. Let $q$ be a $\kappa$-path in $X$ from $p$ to $r$. Then the induced function
$q_{\#}: \Pi_1^{\kappa}(X,p) \rightarrow \Pi_1^{\kappa}(X,r)$ defined for a $p$-based EC loop $f$
in $X$ by
$q_{\#}([f]) = [(q^{-1})_{\infty}*f*q_{\infty}]$, is an isomorphism.
\qed
\end{Corollary}
\begin{Theorem}
\label{Pi1-iso-thm}
Suppose $(X,\kappa)$ and $(Y,\lambda)$ are (not necessarily pointed) homotopy equivalent
digital images. Let $F: X \rightarrow Y$, $G: Y \rightarrow X$ be homotopy inverses.
Let $p \in X$. Then
$\Pi_1^{\kappa}(X,p)$ and $\Pi_1^{\lambda}(Y, F(p))$ are isomorphic groups.
\end{Theorem}
\begin{proof}
Let $F_{*}: \Pi_1^{\kappa}(X,p) \rightarrow \Pi_1^{\lambda}(Y,F(p))$ be
the homomorphism induced by $F$
according to Theorem~\ref{ec-induced}.
Let $r=(G \circ F)(p)$.
Let $G_{*}: \Pi_1^{\lambda}(Y,F(p)) \rightarrow \Pi_1^{\kappa}(X,r)$ be
the homomorphism induced by $G$
according to Theorem~\ref{ec-induced}.
Let $H: X \times [0,m]_Z \rightarrow X$ be a homotopy from $1_X$
to $G \circ F$.
Let $q$ be the path from $p$ to $r$ defined by
$q(t)=H(p,t)$.
For $s \in [0,m]_{\bf Z}$, let $q_s: [0,m]_{\bf Z} \rightarrow X$ be the path
from $q(0)=p$ to $q(s)=H(p,s)$ given by $ q_s(t) = q(\min\{s,t\})$.
For a $p$-based EC loop $f$ in $X$,
let $K: \mathbb{N}^* \times [0,m]_{Z} \rightarrow X$ be defined by
\[ K(n,t)=(q_t * (H_t \circ f_-) * (q_t)^{-1})_{\infty}(n).\]
Since $q_t$ is a path from $r$ to
$q(t) = H(p,t)=H_t(f(0)) = H_t(f_-(N_{f})) = (q_t)^{-1}(0)$,
$K$ is well defined and, for each $t$, the induced function
$K_t$ is a EC loop based at $p$. Also, if we let
$\overline{p}$ denote the constant EC loop at $p$, then
\[ K(n,0)=((q_0) * (H_0 \circ f_-) * (q_0)^{-1})_{\infty}(n) = \\
(\overline{p} * f_- * \overline{p})_{\infty}(n) =f(n) \]
and
\[ K(n,m) = (q_m * (H_m \circ f_-) * (q_m)^{-1}))_{\infty}(n) = \\
(q * (G \circ F \circ f_-) * q^{-1})_{\infty}(n).
\]
Therefore, $K$ is a EC homotopy from $f$ to
\[ (q * (G \circ F \circ f_-) * q^{-1})_{\infty} =
q_{\infty} * (G \circ F \circ f_-)_{\infty} * (q^{-1})_{\infty}
= \\
q_{\infty} * (G \circ F \circ f) * (q_{\infty})^{-1}\]
that keeps the endpoints fixed.
Let $q_{\#}: \Pi_1^{\kappa}(X,p) \rightarrow \Pi_1^{\kappa}(X,r)$ be defined by
$q_{\#}([f]) = [q_{\infty}*f*(q_{\infty})^{-1}]$.
By the conclusion of the previous paragraph, the function
$q_{\#} \circ G_* \circ F_*$ is the identity map on $\Pi_1^{\kappa}(X,p)$. We know
from Corollary~\ref{EC-change-basepoints} that $q_{\#}$ is an isomorphism.
It follows that $F_*$ is onto and $G_*$ is one-to-one. A similar argument
shows that $G_*$ is onto and $F_*$ is one-to-one. Therefore,
$F_*$ is an isomorphism.
\end{proof}
\section{Further remarks}
We have given the first example of two digital images with $c_u$-adjacencies
that are homotopy equivalent but not pointed homotopy equivalent. We have introduced a variant of the loop equivalence, based on the notion of
tight at the basepoint (TAB) pointed homotopy, and have explored properties of this notion.
We have given an alternate but equivalent approach to the digital
fundamental group based on EC loops that offers the advantage of
avoiding the often-clumsy use of trivial extensions.
We have
provided a correction to the faulty proof of ~\cite{Boxer05a} that
(unpointed) homotopy equivalent digital images have isomorphic
fundamental groups.
\begin{comment}
The condition that loops be tight at the basepoint is natural in the category of pointed digital spaces, but more general constructions could be considered in future work. For example, in a pointed space $(X,p)$ we could choose some point $q \in X - \{p\}$, and define a fundamental group in which all loops are required to be tight whenever they pass through $q$. (That is, $q$ must never appear as two consecutive points on the loop.) In our example space $(X,x_0)$ we had $\Pi^{TAB}_1(X,x_0) \neq \Pi_1(X,x_0)$, but this is due to the tightness restriction at $x_0$, not due to the fact that our loops are based at $x_0$.
It may be possible to define fundamental groups which have tightness restrictions on an arbitrary subset of the space. For example consider the following space $Y$:
\begin{center}
\begin{tikzpicture}[scale=.3]
\filldraw[fill=white, xshift=4cm,yshift=0cm]
(45:1.2) \foreach \x in {135,225,315,45} { -- (\x:1.2) };
\filldraw[fill=white, xshift=4cm,yshift=8cm]
(45:1.2) \foreach \x in {135,225,315,45} { -- (\x:1.2) };
\filldraw[fill=white, xshift=6cm,yshift=0cm]
(45:1.2) \foreach \x in {135,225,315,45} { -- (\x:1.2) };
\filldraw[fill=white, xshift=6cm,yshift=8cm]
(45:1.2) \foreach \x in {135,225,315,45} { -- (\x:1.2) };
\filldraw[fill=white, xshift=8cm,yshift=0cm]
(45:1.2) \foreach \x in {135,225,315,45} { -- (\x:1.2) };
\filldraw[fill=white, xshift=8cm,yshift=8cm]
(45:1.2) \foreach \x in {135,225,315,45} { -- (\x:1.2) };
\filldraw[fill=white, xshift=2cm,yshift=2cm]
(45:1.2) \foreach \x in {135,225,315,45} { -- (\x:1.2) };
\filldraw[fill=white, xshift=2cm,yshift=6cm]
(45:1.2) \foreach \x in {135,225,315,45} { -- (\x:1.2) };
\filldraw[fill=white, xshift=10cm,yshift=2cm]
(45:1.2) \foreach \x in {135,225,315,45} { -- (\x:1.2) };
\filldraw[fill=white, xshift=10cm,yshift=6cm]
(45:1.2) \foreach \x in {135,225,315,45} { -- (\x:1.2) };
\filldraw[fill=white, xshift=0cm,yshift=4cm]
(45:1.2) \foreach \x in {135,225,315,45} { -- (\x:1.2) };
\filldraw[fill=white, xshift=4cm,yshift=4cm]
(45:1.2) \foreach \x in {135,225,315,45} { -- (\x:1.2) };
\filldraw[fill=white, xshift=8cm,yshift=4cm]
(45:1.2) \foreach \x in {135,225,315,45} { -- (\x:1.2) };
\filldraw[fill=white, xshift=12cm,yshift=4cm]
(45:1.2) \foreach \x in {135,225,315,45} { -- (\x:1.2) };
\node () at (4cm,0cm) {$$};
\node () at (4cm,8cm) {$$};
\node () at (6cm,0cm) {$$};
\node () at (6cm,8cm) {$$};
\node () at (8cm,0cm) {$$};
\node () at (8cm,8cm) {$$};
\node () at (2cm,2cm) {$$};
\node () at (2cm,6cm) {$p$};
\node () at (10cm,2cm) {$$};
\node () at (10cm,6cm) {$q$};
\node () at (0cm,4cm) {$$};
\node () at (4cm,4cm) {$$};
\node () at (8cm,4cm) {$$};
\node () at (12cm,4cm) {$$};
\end{tikzpicture}
\end{center}
It seems clear that $\Pi_1(Y,p)$ and $\Pi_1(Y,q)$ are infinite cyclic, and $\Pi_1^{TAB}(Y,p)$ and $\Pi_1^{TAB}(Y,q)$ are free groups on two generators. But there is more structure to the space which is not captured by any one of these groups. In particular it also seems that enforcing tightness at both $p$ and $q$ simultaneously will give a larger group, possibly free on 4 generators. Some care would be needed in the definitions, since it does not seem clear how to mimic the argument given in Theorem \ref{TAB-group} to construct a homotopy for inverses which is tight on several points.
\end{comment}
|
\section{Introduction}
The Navier-Stokes equations for an incompressible viscous fluid of viscosity $\nu=1$ are given by
\begin{equation*}
\tag{NS}
\label{Navierstokes}
\left\{
\begin{array}{l}
u_t-\Delta u+u\cdot\nabla u+\nabla p=0 \quad \mbox{in} \quad X\times \left(0,\infty\right)\\
u\left(x,0\right)=\psi, \quad div \,u=0.
\end{array}
\right.
\end{equation*}
where, in this paper, $X=\mathbb{R}^3$ or $X=\mathbb{T}^3$.
It is known that given an initial condition of finite energy and which belongs to $\dot{H}^s\left(X\right)$
there is an interval of time
$\left(0,\eta\right)$, $\eta>0$, for which there is a unique smooth solution to (\ref{Navierstokes}) in
$C\left(\left[0,T\right];\dot{H}^s\left(X\right)\right)$. Let then $T>0$ be the largest
$\eta>0$ for which the unique solution with initial data $\psi\in \dot{H}^s\left(X\right)$ remains smooth. It is unknown whether $T<\infty$ or
$T=\infty$. In the case that $T<\infty$, there is the interesting question of estimating a rate at which
the $\dot{H}^s$-norm blows-up.
In \cite{CortissozMonteroPinilla}
the authors, based on ideas presented by Robinson, Sadowski and Silva in \cite{RobinsonSadowskiSilva}, showed an almost optimal
lower bound for the blow--up rate of solutions of the Navier--Stokes equations with periodic boundary conditions
on a bounded maximal interval of existence $\left(0,T\right)$, $T<\infty$,
when this solution belongs
to $\dot{H}^{\frac{3}{2}}(\mathbb{T}^3)\cap\dot{H}^{\frac{5}{2}}(\mathbb{T}^3)$.
To be more precise, it was shown that a regular solution of the Navier--Stokes equation
whose maximal interval of existence (or regularity) is $\left(0,T\right)$, must satisfy
\[
\left\|u\left(t\right)\right\|_{H^{\frac{3}{2}}\left(\mathbb{T}^3\right)}\geq \frac{c}{\sqrt{\left(T-t\right)\left|\log\left(T-t\right)\right|}},
\]
for a constant $c>0$. In this paper we go a little further and give a proof of the
expected optimal lower blow--up rate. Namely, we prove the following the following estimate
on the blow--up rate of putative singular solutions to the Navier--Stokes equations:
\begin{equation*}
\frac{C}{t^\frac{1}{2}}\leq \left\|u(T-t)\right\|_{\dot{H}^{\frac{3}{2}}(\mathbb{T}^3)}, \quad C>0.
\end{equation*}
The proof of this result requires a detailed inspection of the bounds on the nonlinear term
of the Navier--Stokes equations found in \cite{RobinsonSadowskiSilva},
and the application of an interpolation technique inspired by the method used by Hardy to prove Carlson's inequality (see \cite{LarsonPecaric}).
We must add that this problem using different techniques has been treated in the papers \cite{CheskidovZaya}
and \cite{Robinsonetal}.
The lower blow--up rates for putative singular solutions to the Navier--Stokes equations can be interpreted as a regularity criterion
for solutions of the equation (as they give a lower bound on the size of the maximal interval of existence).
These blow--up estimates were first stated for the $L^p$ spaces, $p > 3$,
without proof by Leray in his remarkable paper \cite{Leray}, and proved by Giga in \cite{Giga} via semigroup theory.
In this paper, we rather follow the elementary, and improve on, the proof on homogeneous Sobolev spaces given by Robinson, Sadowski and Silva for their
blow--up estimates.
On the other hand, there exists the related problem of investigating the possible blow-up behavior of solutions to the incompressible Euler equations:
\begin{equation*}
\tag{E}
\label{Euler}
\left\{
\begin{array}{l}
u_t+u\cdot\nabla u+\nabla p=0 \quad \mbox{in} \quad X\times \left(0,\infty\right)\\
u\left(x,0\right)=\psi, \quad \mbox{div} \!\left(u\right)=0.
\end{array}
\right.
\end{equation*}
In fact, recently in a very nice paper \cite{ChenPavlovic}, Chen and Pavlović showed the following (although they state their result
in $\mathbb{R}^3$ and we do in $\mathbb{T}^3$, our arguments apply in both cases, see Remark 1).
\begin{theorem}
\label{th:ChenPavlovic}
Let $u\left(x,t\right)$ be a solution of the periodic Euler equations in the class
\begin{equation}
\label{class}
C^1\left(\left[0,T\right],\dot{H}^{\frac{3}{2}+\delta}\left(\mathbb{T}^3\right)\right)\cap
C\left(\left[0,T\right],\dot{H}^{\frac{5}{2}+\delta}\left(\mathbb{T}^3\right)\right),\quad \delta>0,
\end{equation}
and let $T>0$ be the minimum time for which $u$ cannot be continued in the class (\ref{class}). Then, there exists a finite, positive constant
$C\left(\delta,\left\|u\left(0\right)\right\|_{L^2\left(\mathbb{T}^3\right)}\right)$ such that
\[
\left\|u\left(t\right)\right\|_{\dot{H}\left(\mathbb{T}^3\right)^{\frac{5}{2}+\delta}}\geq C\left(\delta,\left\|u\left(0\right)\right\|_{L^2\left(\mathbb{T}^3\right)}\right)
\left(\frac{1}{T-t}\right)^{2+\frac{2}{5}\delta}.
\]
\end{theorem}
The proof of this theorem given in \cite{ChenPavlovic} relies on obtaining a single exponential bound on the $H^s$ norms of a solution
of the Euler equations via a lenght parameter introduced by P. Constantin in \cite{Constantin}. In this paper we follow
the approach suggested in \cite{RobinsonSadowskiSilva} in conjunction with some ideas
presented in \cite{CortissozMonteroPinilla}, to give a less involved proof of Theorem \ref{th:ChenPavlovic}.
Part of this paper was written while the second author was visiting the Mathematics Department at Cornell University, and he is quite
grateful for their warm hospitality -and in particular to Prof. Tim Healy for
his encouragement. He also must acknowledge the support of Colciencias and his home institution,
the Universidad de los Andes for making this visit possible, and his advisor (the first named author of this paper)
for his encouragement, and his almost always insightful observations. The first author wants to thank the second author for
being a great student and colleague, and for all these wonderful years of shared mathematical enthusiasm.
He also wants to thank his home institution, the Universidad de los Andes, for providing an excellent research environment and
economic support (Proyecto Semilla P15.160322.009).
\section{The blow up rate for the Navier-Stokes equations}
The next statement is essentially the same given in \cite{RobinsonSadowskiSilva}.
The main difference is that we show a proof which includes the case when the solution belongs
to $\dot{H}^{\frac{3}{2}}(\mathbb{T}^3)\cap \dot{H}^{\frac{5}{2}}(\mathbb{T}^3)$. From now on, in this
paper we shall use the notation
\[
\left\|u\right\|_s:=\left\|u\right\|_{\dot{H}^s\left(\mathbb{T}^3\right)},
\]
and $\hat{u}_k$ refers to the Fourier wavenumber of wavevector $k$ of the function $u$.
\begin{theorem}
\label{th:Blow_up}
Let $u(x,t) = (u_1,u_2,u_3)$ be a solution Navier--Stokes equations whose maximum interval of existence is
$\left(0,T\right)$, $0<T<\infty$, and such that
$u\in C((0,T),\dot{H}^s(\mathbb{T}^3)\cap \dot{H}^{s+1}(\mathbb{T}^3))$, with $\displaystyle \frac{1}{2}<s<\frac{5}{2}$.
Then the following estimate holds
\begin{equation}
\label{eq:BlowUpRate}
\frac{C_s}{t^{\frac{1}{2}(s-\frac{1}{2})}} \leq \left\|u(T-s)\right\|_{\dot{H}^s(\mathbb{T}^3)}.
\end{equation}
\end{theorem}
\begin{proof}
First, we must recall the energy inequality found in \cite{RobinsonSadowskiSilva}:
\begin{equation}
\label{eq:EnergyInequality}
\frac{1}{2}\frac{d}{dt}(\left\|u(t)\right\|_s^2) + 4\pi^2\left\|u(t)\right\|_{s+1}\leq C_s
\left(
\sum_k \left|\widehat{u}_k\right|\left|k\right|^r
\right)\left\|u(t)\right\|_s\left\|u\right\|_{s+1-r},
\end{equation}
with $0\leq r\leq 1$. For the sake of completeness we will give a proof of this inequality below.
Now we pick $\displaystyle r=\frac{1}{2}\left(s-\frac{1}{2}\right)$,
and apply the interpolation technique employed by Hardy in his proof of Carlson's inequality (see \cite{LarsonPecaric}),
to the first factor on the right hand side of (\ref{eq:EnergyInequality}), to obtain:
\begin{align*}
\sum_k \left|\widehat{u}_k\right|\left|k\right|^{\frac{1}{2}\left(s-\frac{1}{2}\right)}=&
\sum_k \left|\widehat{u}_k\right|
\left|k\right|^{\frac{1}{2}\left(s-\frac{1}{2}\right)}\frac{\sqrt{a|k|^{s+\frac{1}{2}}+b|k|^{s+\frac{5}{2}}}}{\sqrt{a|k|^{s+\frac{1}{2}}+b|k|^{s+\frac{5}{2}}}}\\
\leq& \left(a\left\|u\right\|^2_s+b\left\|u\right\|^2_{s+1}\right)^{\frac{1}{2}}
\left(
\sum_k \frac{1}{a\left|k\right|^{s+\frac{1}{2}}+b\left|k\right|^{s+\frac{5}{2}}}
\right)^{\frac{1}{2}}\\
\leq& \left(a\left\|u\right\|^2_s+b\left\|u\right\|^2_{s+1}\right)^{\frac{1}{2}}
\left(
\frac{4\pi}{\sqrt{ab}}
\left(
\frac{\sqrt{a}}{\sqrt{b}}
\right)^{\frac{3}{2}-s}\int_0^\infty \frac{y^{\frac{3}{2}-s}}{1+y^2}dy
\right)^{\frac{1}{2}},
\end{align*}
if we choose $a=\left\|u(t)\right\|^2_{s+1}$ and $b=\left\|u(t)\right\|^2_{s}$ then the energy inequality (\ref{eq:EnergyInequality})
becomes
\begin{equation}
\label{eq:EnergyInequality2}
\frac{1}{2}\frac{d}{dt}(\left\|u(t)\right\|_s^2) + 4\pi^2 \left\|u(t)\right\|^2_{s+1}
\leq C_s \left\|u(t)\right\|_s^{\frac{s}{2}+\frac{3}{4}}\left\|u(t)\right\|_{s+1}^{\frac{5}{4}-\frac{s}{2}}\left\|u(t)\right\|_{\frac{s}{2}+\frac{5}{4}}.
\end{equation}
Now, observe that $\displaystyle \frac{s}{2}+\frac{5}{4}=
\left(
\frac{s}{2}-\frac{1}{4}
\right)s +
\left(
\frac{5}{4}-\frac{s}{2}
\right)(s+1)$, so by interpolation between homogeneous Sobolev spaces, we get
\begin{equation*}
\left\|u\right\|_{\frac{s}{2}+\frac{5}{4}}\leq \left\|u\right\|_s^{\frac{s}{2}-\frac{1}{4}}\left\|u\right\|_{s+1}^{\frac{5}{4}-\frac{s}{2}}.
\end{equation*}
Therefore, from inequality (\ref{eq:EnergyInequality2}) we obtain
\begin{equation*}
\frac{1}{2} \frac{d}{dt}(\left\|u(t)\right\|_s^2)+ 4\pi^2 \left\|u(t)\right\|^2_{s+1} \leq C_s \left\|u(t)\right\|_s^{s+\frac{1}{2}}
\left\|u(t)\right\|_{s+1}^{\frac{5}{2}-s}.
\end{equation*}
It is time to use Young's inequality $\displaystyle ab\leq \frac{a^p}{p}+\frac{b^q}{q}$, $\displaystyle \frac{1}{p}+\frac{1}{q}=1$,
with the choice $p=\frac{2\left(s+\frac{1}{2}\right)}{s-\frac{1}{2}}$ and $q=\frac{2}{\frac{5}{2}-s}$. We thus get
\[
\frac{1}{2} \frac{d}{dt}(\left\|u(t)\right\|_s^2)\leq c_s\left(\left\|u(t)\right\|_s^2\right)^{\left(1+\frac{1}{s-\frac{1}{2}}\right)}.
\]
Finally, by integrating between $T-t$ and $T$ the previous estimate, inequality (\ref{eq:BlowUpRate}) follows.
\end{proof}
\begin{remark}
Theorem \ref{th:Blow_up} is also valid when we consider the case of the whole space, i.e., for solutions
$u(t) \in \dot{H}^{\frac{3}{2}}(\mathbb{R}^3)\cap \dot{H}^{\frac{5}{2}}(\mathbb{R}^3)$, this
because all the calculations leading to its proof are valid on $\mathbb{R}^3$ if we change sums by integrals.
\end{remark}
As promised, we give a proof of inequality (\ref{eq:EnergyInequality}).
It is a consequence of the following lemma
(see the proof of Lemma 3.1 in \cite{RobinsonSadowskiSilva})
which gives an estimate
of the nonlinear term
\[
\left|\left(B\left(u,u\right),u\right)_{\dot{H}^{s}}\right|,
\]
where
\[
B\left(u,u\right)=P\left(u\cdot\nabla u\right),
\]
and $P$ is the Leray projector.
\begin{lemma}
\label{th:InterpLog}
For any \(s>1\) and \(0\leq r\leq 1\), we have
\begin{equation*}
\left|\sum_{k\in\dot{\mathbb{Z}^3}}\sum_{q\in\dot{\mathbb{Z}^3}}|k|^{2s}(k\cdot \widehat{u}_{k-q})(\widehat{u}_q\cdot\overline{\widehat{u}}_k)\right|
\leq c_s \left(\sum_{k\in\dot{\mathbb{Z}^3}} |k|^r|\widehat{u}_k|\right)\left\|u\right\|_{s}\left\|u\right\|_{s+1-r},
\end{equation*}
for all \( u\in \dot{H}^{s+1-r}(\mathbb{T}^3)\cap F^r\). Here
$\hat{u}_k$ denotes the Fourier wavenumber of $u$
with wavevector $k$, and $\overline{z}$
denotes the complex conjugate of $z$.
\end{lemma}
\begin{proof}
Since \(P\)
is self-adjoint and \(u(x,t)\)
is divergence free , we have that (see \cite{ConstantinFoias},
chapter 6 p. 53)
\begin{equation*}
\sum_{k\in \dot{\mathbb{Z}}^3}\sum_{q\in \dot{\mathbb{Z}}^3}|q|^s|k|^s(\widehat{u}_{k-q}\cdot q)(\widehat{u}_q\cdot \overline{\widehat{u}}_k) = 0.
\end{equation*}
Using the inequality ( see \cite{HardyLittlewoodPoyla}, p.39)
\begin{equation*}
\left|
|x|^s-|y|^s
\right|\leq s(2^s)|x-y|
\left(
|x-y|^{s-1}+|y|^{s-1}
\right),\quad s>1;
\end{equation*}
and the well know inequality
\begin{equation*}
(|x|+|y|)^r \leq |x|^r+|y|^r \textrm{ if } 0\leq r \leq 1,
\end{equation*}
we obtain:
\begin{align*}
&\left|\sum_{k\in\dot{\mathbb{Z}^3}}\sum_{q\in\dot{\mathbb{Z}^3}}|k|^{2s}(k\cdot \widehat{u}_{k-q})(\widehat{u}_q\cdot\overline{\widehat{u}}_k)\right|\\
&\quad\quad
=\left|\sum_{k\in\dot{\mathbb{Z}^3}}\sum_{q\in\dot{\mathbb{Z}^3}}|k|^{2s}(k\cdot
\widehat{u}_{k-q})(\widehat{u}_q\cdot\overline{\widehat{u}}_k)
-\sum_{k\in \dot{\mathbb{Z}}^3}\sum_{q\in
\dot{\mathbb{Z}}^3}|q|^s|k|^s(\widehat{u}_{k-q}\cdot
q)(\widehat{u}_q\cdot \overline{\widehat{u}}_k)
\right|\\
&\quad\quad \leq \sum_{k\in\dot{\mathbb{Z}^3}}\sum_{q\in\dot{\mathbb{Z}^3}}|k|^{s}|(q\cdot\widehat{u}_{k-q})|\,|\widehat{u}_q|\,|\widehat{u}_k|\,||k|^s- |q|^s|\\
&\quad\quad\leq s2^{s-1} \sum_{k\in\dot{\mathbb{Z}^3}}\sum_{q\in\dot{\mathbb{Z}^3}}|k|^{s}|(q\cdot\widehat{u}_{k-q})||\widehat{u}_q||\widehat{u}_k| |k-q|(|k-q|^{s-1}+|q|^{s-1})\\
&\quad\quad\leq s2^s \sum_{q\in\dot{\mathbb{Z}^3}}\sum_{k\in\dot{\mathbb{Z}^3}}|k|^{s}|k-q|^s|q||\widehat{u}_{k-q}||\widehat{u}_q||\widehat{u}_k|\\
&\quad\quad\leq s2^s \sum_{q\in\dot{\mathbb{Z}^3}}\sum_{k\in\dot{\mathbb{Z}^3}}|k|^{s}|k-q|^s|q|^r|q|^{1-r}|\widehat{u}_{k-q}||\widehat{u}_q||\widehat{u}_k|\\
&\quad\quad\leq s2^s \sum_{q\in\dot{\mathbb{Z}^3}}\sum_{k\in\dot{\mathbb{Z}^3}}|k|^{s}|k-q|^s|q|^r(|k-q|^{1-r}+|k|^{1-r})|\widehat{u}_{k-q}||\widehat{u}_q||\widehat{u}_k|\\
&\quad\quad\leq s2^{s+1} \sum_{q\in\dot{\mathbb{Z}^3}} |q|^r|\widehat{u}_q|\sum_{k\in\dot{\mathbb{Z}^3}}|k-q|^{s}|\widehat{u}_{k-q}||k|^{s+1-r}|\widehat{u}_k|\\
&\quad\quad\leq
s2^{s+1}\left(\sum_{k\in\dot{\mathbb{Z}^3}}|k|^r|\widehat{u}_k|\right)\left\|u\right\|_{s}\left\|u\right\|_{s+1-r},
\end{align*}
which is what we wanted to prove.
\end{proof}
The previous proof gives us also an lower bound on size on the maximal interval of existence. Indeed,
the following result holds.
\begin{corollary}
Let $u(x,t)$ be a solution of the Navier--Stokes equations with initial condition $u_0(x)\in \dot{H}^s(\mathbb{T}^3)$, $\frac{1}{2}<s<\frac{5}{2}$,
and let $T>0$ be the minimum time for blow--up. Then
\begin{equation}
\label{eq:Time}
\frac{K_s}{\left(\left\|u_0\right\|_s\right)^{\frac{4}{2s-1}}}\leq T.
\end{equation}
\end{corollary}
\section{The blow up rate for the Euler equations}
Estimate (\ref{eq:EnergyInequality}) can be used to provide the promised elementary proof of Theorem \ref{th:ChenPavlovic}. But before we present our
proof we will need the following estimate.
\begin{lemma}
Let
\[
\left\|u\right\|_{F^1}=\sum_{k\in \mathbb{Z}^3}\left|k\right|\left|\hat{u}_k\right|.
\]
There is a constant $c>0$ which only depends on $s$ such that
\[
\left\|u\right\|_{F^1}\leq c\left\|u\right\|_{L^2\left(\mathbb{T}^3\right)}^{\frac{s-\frac{5}{2}}{s}}\left\|u\right\|_{s}^{\frac{5}{2s}}.
\]
\end{lemma}
\begin{proof}
We have that
\begin{eqnarray*}
\sum_k\left|\hat{u}_k\right|\left|k\right|&=&\sum_k\left|\hat{u}_k\right|\left|k\right|\frac{\sqrt{a+b\left|k\right|^{2s-2}}}{\sqrt{a+b\left|k\right|^{2s-2}}}\\
&\leq& \left(\sum_k\left|\hat{u}_k\right|^2\left|k\right|^2\left(a+b\left|k\right|^{2s-2}\right)\right)^{\frac{1}{2}}
\left(\sum_k\frac{1}{a+b\left|k\right|^{2s-2}}\right)^{\frac{1}{2}}\\
&\leq&c_s\left(a\left\|u\right\|^2_1+b\left\|u\right\|_s^2\right)\left(\int_0^{\infty}\frac{x^2dx}{a+bx^{2s-2}}\right)^{\frac{1}{2}}\\
&\leq&c_s\left(a\left\|u\right\|^2_1+b\left\|u\right\|_s^2\right)\frac{1}{\sqrt{a}}\left(\frac{a}{b}\right)^{\frac{3}{2\left(2s-2\right)}}.
\end{eqnarray*}
We let $a=\left\|u\right\|^2_s$ and $b=\left\|u\right\|_{1}^2$ to obtain (for a new constant $c_s>0$)
\begin{eqnarray*}
\left\|u\right\|_{F^1}&\leq& c_s\left\|u\right\|_1\left(\frac{\left\|u\right\|_s}{\left\|u\right\|_1}\right)^{\frac{3}{2s-2}}\\
&=&c_s\left\|u\right\|_1^{\frac{2s-5}{2s-2}}\left\|u\right\|^{\frac{3}{2s-2}}_s.
\end{eqnarray*}
Now we use the Sobolev interpolation inequality
\[
\left\|u\right\|_1\leq \left\|u\right\|_{L^2\left(\mathbb{T}^3\right)}^{\frac{s-1}{s}}\left\|u\right\|^{\frac{1}{s}}_s,
\]
to obtain
\begin{eqnarray*}
\left\|u\right\|_{F^1}&\leq& c_s\left\|u\right\|_{L^2\left(\mathbb{T}^3\right)}^{\frac{2s-5}{2s}}\left\|u\right\|^{\frac{1}{s}\frac{2s-5}{2s-2}+\frac{3}{2s-2}}_s\\
&=&c_s\left\|u\right\|_{L^2\left(\mathbb{T}^3\right)}^{\frac{2s-5}{2s}}\left\|u\right\|^{\frac{5}{2s}}_s.
\end{eqnarray*}
\end{proof}
We are ready to prove the estimate of Chen and Pavlović. Indeed, proceeding as we did in the proof
of Theorem \ref{th:Blow_up}, for the Euler equations we obtain (i.e.,
by inequality (\ref{eq:EnergyInequality}) with $r=1$;
the extra positive term on the left-hand side does not appear, due to the lack of the diffusion term in the Euler equations) :
\[
\frac{d}{dt}\left\|u\right\|_s^2\leq c_s\left\|u\right\|_{F^1}\left\|u\right\|_s^2,
\]
(this is just equation (6.2) in \cite{RobinsonSadowskiSilva}),
and hence by the previous lemma we arrive at the differential inequality
\begin{eqnarray*}
\frac{d}{dt}\left\|u\right\|_s^2&\leq& c_s\left\|u\right\|_{L^2\left(\mathbb{T}^3\right)}^{\frac{2s-5}{2s}}\left\|u\right\|^{\frac{5}{2s}+2}_s.
\end{eqnarray*}
For a regular solution to Euler equation, it is well-known that for any $t\geq 0$,
\[
\left\|u\left(t\right)\right\|_{L^2\left(\mathbb{T}^3\right)}\leq \left\|u\left(0\right)\right\|_{L^2\left(\mathbb{T}^3\right)},
\]
so we obtain an inequality (here the constant involved depends on $\left\|u\left(0\right)\right\|_{L^2\left(\mathbb{T}^3\right)}$)
\[
\frac{d}{dt}\left\|u\right\|_s^2\leq c\left(\left\|u\left(0\right)\right\|_{L^2\left(\mathbb{T}^3\right)},s\right)\left\|u\right\|^{\frac{5}{2s}+2}_s.
\]
Let $s=\frac{5}{2}+\delta$. Then our inequality becomes
\[
\frac{d}{dt}\left\|u\right\|_{\frac{5}{2}+\delta}^2\leq c\left(\left\|u\left(0\right)\right\|_{L^2\left(\mathbb{T}^3\right)},\delta\right)
\left\|u\right\|^{2+\frac{1}{1+\frac{2}{5}\delta}}_{\frac{5}{2}+\delta}.
\]
Integrating the previous differential inequality from $t$ to $T$ (and assuming blow up at $T$) we get
\[
\frac{1}{\left\|u\left(t\right)\right\|^\frac{1}{1+\frac{2}{5}\delta}_{\frac{5}{2}+\delta}}\leq c\left(T-t\right),
\]
where $c$ is a constant that only depends on $\left\|u\left(0\right)\right\|_{L^2\left(\mathbb{T}^3\right)}$ and $\delta$. Solving for
$\left\|u\left(t\right)\right\|_{\frac{5}{2}+\delta}$ finishes the proof.
\vspace{.2in}
\noindent
{\bf Remark 2.} As commented before in the case of the Navier-Stokes equations, in the proofs in this section
it is possible to replace $\mathbb{T}^3$ by $\mathbb{R}^3$.
\section{Final comments: Some open questions}
Theorem \ref{th:Blow_up} includes the optimal lower bound for blow--up rates
when $u\in \dot{H}^{\frac{3}{2}}(\mathbb{T}^3)\cap\dot{H}^{\frac{5}{2}}(\mathbb{T}^3)$;
this particular case was missing in the proof given in \cite{RobinsonSadowskiSilva}, and in \cite{CortissozMonteroPinilla}
a non optimal bound was proved.
These bounds raise the following question: If there exists some $C>0$ such that $
\displaystyle \left\|u(T-t)\right\|_s\leq C t^{-\frac{1}{2}(s-\frac{1}{2})}$ ,
does $\left\|u(T-t)\right\|_s$ blow--up? Furthermore, a lower blow-up rate for $u\in \dot{H}^\frac{1}{2}(\mathbb{T}^3)$, for putative
blow--up solutions to the Navier--Stokes equations, is yet unknown.
|
\section{Introduction}
The classical nucleation theory (CNT) provides a theoretical link between the ice nucleation rates and the thermophysical properties of the ice-water system, i.e. the density of the solid phase, the diffusion coefficient of the liquid phase, the equilibrium pressure, and the interfacial energy \cite{ickes15}. Therefore, in cases where experimental measurements of the nucleation rate are available, CNT presents an indirect, theoretical approach for the estimation of the interfacial energy provided that the other above mentioned thermophysical properties are known.
For ice and water, the interfacial energy measurements reported in literature are restricted to the triple point temperature \cite{granasy02}. The estimation of the interfacial energy for lower temperatures in the supercooled region, was possible only due to ice nucleation rate measurements \cite{murray10} or molecular simulations of ice nuclei formation \cite{li11, espinoza16}.
In this work, a method for the estimation of ice-water interfacial energy from the experimental nucleation rates \cite{pant06, parsons06} will be investigated that is based on the evaluation of the slope of the logarithm of the measured nucleation rate $\ln J$ vs. the function $T^{-3}(\ln(S))^{-2}$ of the experimental temperature $T$, where $S$ is the ratio of water and ice saturation pressures. It will be referred to as the \textit{slope method} in the following text. Several authors estimated the ice-water interfacial energies based on the slope method \cite{murray10, murray12, manka12, bhabhe13} covering the temperature range 200 -- 240~K. I will study the validity of the ice-water interfacial energy estimates by the slope method in this paper. Particularly, I intend to show that the ice-water interfacial energy estimate of Murray \textit{et al.}~\cite{murray10} suffers from internal inconsistency and needs to be reconsidered. The nature of this inconsistency will be studied in detail to point out the weak spots of the slope method.
The paper is arranged in the following way. First, the equations of the slope method for interfacial energy estimation are derived in section \ref{se:method}. In section \ref{se:proof}, a proof is presented showing the inconsistency of the interfacial energy estimate by Murray \textit{et al.}~\cite{murray10}. In section \ref{se:discuss}, the source of the inconsistency is identified, a correct estimate of the ice-water interfacial energy is presented, and the limits of applicability of the slope method are discussed.
\section{Slope method} \label{se:method}
The slope method \cite{pant06, parsons06} for the estimation of the ice-water interfacial energy is based on the CNT nucleation rate equation \cite{kashchiev00}
\begin{equation} \label{eq:nr}
J = J_0 \exp \left( -\frac{W^{\star}}{kT} \right)
\end{equation}
relating the nucleation rate $J$ [m$^{-3}$s$^{-1}$] to the nucleation work of the critical cluster $W^{\star}$ [J]. The pre-factor $J_0$ [m$^{-3}$s$^{-1}$] reflects the kinetics of the cluster growth and $T$ [K] is absolute temperature. By using the CNT expression for the critical nucleation work $W^{\star} = \frac{16 \pi \gamma^3 v^2}{3 (kT \ln S)^2}$, Eq. (\ref{eq:nr}) can be rearranged to
\begin{equation} \label{eq:nrln}
\ln J = \ln J_0 - \frac{16 \pi \gamma^3 v^2}{3 k^3 T^3 (\ln S)^2}
\end{equation}
where $\gamma$ [J/m$^2$] is the ice-water interfacial energy, $v$ [m$^3$] is the molecular volume of the solid phase, and the ratio of saturation pressures $S = p^{eq}_{l}/p^{eq}_{s}$ represents the supersaturation of the liquid. Here, the vapor-liquid equilibrium pressure is denoted $p^{eq}_{l}$, and the vapor-solid equilibrium pressure is denoted $p^{eq}_{s}$. The vapor-solid equilibrium pressure of cubic ice \cite{shilling06} is used in this work to retain consistency with the work of Murray \textit{et al.}~\cite{murray10}. However, the structure of the clusters in ice nucleation is still under debate \cite{moore11}.
To transform Eq.~(\ref{eq:nrln}) into the relations of the slope method, the following assumptions are made. The interfacial energy is taken constant, i.e. independent of temperature. The density of the solid phase is taken constant. And the pre-factor is taken constant. Since the size of the temperature interval of the experimental nucleation data analyzed by the slope method is typically a few Kelvins, the assumptions of the constancy of the thermophysical properties over this narrow temperature interval are plausible, and Eq.~(\ref{eq:nrln}) can be written in the form
\begin{equation} \label{eq:nrln3}
\ln J = n + m t_s
\end{equation}
where the parameters
\begin{equation} \label{eq:n}
n = \ln J_0
\end{equation}
and
\begin{equation} \label{eq:m}
m = -\frac{16 \pi \gamma^3 v^2}{3k^3}
\end{equation}
are both due to the above assumptions constant, and
\begin{equation} \label{eq:ts}
t_s = \frac{1}{T^3 (\ln S)^2}
\end{equation}
is a scaled temperature. As a result, the slope method is simply a fit of the experimental ice nucleation rate data to the linear function~(\ref{eq:nrln3}) in $t_s$, giving the parameters $n$ and $m$. The ice-water interfacial energy $\gamma$ corresponding to the analyzed ice nucleation data is recovered from the slope $m$ according to Eq.~(\ref{eq:m}), and the pre-factor $J_0$ is given by the absolute value $n$ according to Eq.~(\ref{eq:n}).
\section{Inconsistency of interfacial energy estimates} \label{se:proof}
The validity of the slope method estimates of the ice-water interfacial energy based on the experimental nucleation data as published by Murray \textit{et al.}~\cite{murray10} will be investigated in this section.
I have identified inconsistent estimates of interfacial energy for two specific sets of ice nucleation data investigated by Murray \textit{et al.}~\cite{murray10}, i.e. their own experimental data and the data measured by Stan \textit{et al.}~\cite{stan09}, respectively. The two experimental data sets are plotted in Fig.~\ref{fig:mse}. By using the slope method, Murray \textit{et al.}~\cite{murray10} estimated ice-water interfacial energy $\gamma_M$ = 20.8 $\pm$ 1.2 mJ/m$^2$ from their own experimental data, and ice-water interfacial energy $\gamma_S$ = 23.7 $\pm$ 1.1 mJ/m$^2$ from the experimental data of Stan \textit{et al.}~\cite{stan09}, respectively. It is the gap between the two intervals, $\gamma_M$ and $\gamma_S$, that renders the two interfacial energy estimates inconsistent. In other words, as I will show below in detail, since the initial two experimental datasets used as input for the evaluation of the respective interfacial energies do overlap to a large extent, which is seen in Fig. \ref{fig:mse}, it is impossible to calculate two interfacial energy estimates by the slope method that form two disjunct intervals, as is the case with $\gamma_M$ and $\gamma_S$.
\begin{figure}[t]
\centering
\includegraphics[width=86mm]{expcmp2.pdf}
\caption{Experimental ice nucleation rates reported by Murray \textit{et al.}~\cite{murray10} and Stan \textit{et al.}~\cite{stan09}. The full lines show fits of the experimental data, and the dashed lines delimit the standard deviation bands. The fit of Murray \textit{et al.}~\cite{murray10} is used as presented in their Fig.~4. The data of Stan \textit{et al.}~\cite{stan09} as presented in their Fig.~8 are shown with the reported $\pm$0.4~$^{\circ}$C standard deviation.
}
\label{fig:mse}
\end{figure}
To prove the particular inconsistency in interfacial energy estimates $\gamma_M$ and $\gamma_S$ suggested above, I will frame general requirements for a physically-relevant deduction of the ice-water interfacial energy from the nucleation rate according to the CNT. Mathematically, the deductions will utilize elementary properties of continuous, strictly monotonic functions in a reductio ad absurdum type of logical proof. The slope method analysis will present a limiting case of this general scenario.
Let us denote $g$ a continuous, strictly monotonic function of temperature $g$: $T$ $\rightarrow$ $\gamma$, which stands for the functional dependence of the interfacial energy on temperature. Similarly, let $j$ be a continuous, strictly monotonic function of temperature $j$: $T$ $\rightarrow$ $J$, which stands for the functional dependence of the nucleation rate on temperature, e.g. in the form of Eq.~(\ref{eq:nr}). Let the temperature interval [$T_1$, $T_2$] be a subdomain of the two functions $g$ and $j$ that corresponds to the temperature range of the analyzed ice nucleation experiment. Both functions $g$ and $j$ are bijective by their above definitions. In other words, they present a one-to-one correspondence between the elements of subdomain [$T_1$, $T_2$] and the elements of functional images $g([T_1, T_2])$ and $j([T_1, T_2])$, respectively. Therefore, inverse function $j^{-1}$: $J$ $\rightarrow$ $T$ exists and it is bijective as well. Furthermore, composite function $g_j$ = $g(j^{-1})$: $J$ $\rightarrow$ $\gamma$ is bijective. Here, $g_j$ is the formal representation of a general theoretical method to deduce the interfacial energy from the nucleation rate in temperature range [$T_1$, $T_2$].
Now, let $\bm{J_1}$ = [$J_{11}$, $J_{12}$] and $\bm{J_2}$ = [$J_{21}$, $J_{22}$] be two intervals of nucleation rates that possess a non-empty intersection $\bm{J_0}$ = $\bm{J_1} \cap \bm{J_2} $ $\neq \emptyset$. Let $\bm{\gamma_1}$ = [$\gamma_{11}$, $\gamma_{12}$] and $\bm{\gamma_2}$ = [$\gamma_{21}$, $\gamma_{22}$] be $g_j$'s functional images of intervals $\bm{J_1}$ and $\bm{J_2}$, i.e. $\bm{\gamma_1}$ = $g_j(\bm{J_1})$ and $\bm{\gamma_2}$ = $g_j(\bm{J_2})$, respectively. Intervals $\bm{\gamma_1}$ and $\bm{\gamma_2}$ represent the ranges of interfacial energies deduced from the two nucleation rate data sets $\bm{J_1}$ and $\bm{J_2}$ according to a general method $g_j$. Then $\forall J \in \bm{J_0}$ it holds that $g_j(J) \in \bm{\gamma_1}$ and $g_j(J) \in \bm{\gamma_2}$. In other words, for any value of the nucleation rate $J$, which belongs to the intersection of intervals $\bm{J_1}$ and $\bm{J_2}$, the deduced value of interfacial energy according to method $g_j$ belongs to interval $\bm{\gamma_1}$ and also to interval $\bm{\gamma_2}$. Therefore, since element $\gamma = g_j(J)$ belongs to both intervals $\bm{\gamma_1}$ and $\bm{\gamma_2}$, intersection $\bm{\gamma_0}$ = $\bm{\gamma_1} \cap \bm{\gamma_2} $ has at least one element $\gamma$ and, therefore, $\bm{\gamma_0}$ must be non-empty, i.e. $\bm{\gamma_0} \neq \emptyset$.
For the particular case of the ice nucleation experimental data of Murray~\textit{et al.}~\cite{murray10} and Stan~\textit{et al.}~\cite{stan09} the intersection of the measured nucleation rate data sets is not empty, as can be seen in Fig.~\ref{fig:mse}; the measured values of the nucleation rate in the two datasets clearly overlap to a large extent. According to the above reasoning the range of interfacial energies deduced from the Murray~\textit{et al.}~\cite{murray10} nucleation data must have an non-empty intersection with the range of interfacial energies deduced from the Stan~\textit{et al.}~\cite{stan09} nucleation data provided that the method used to deduce the interfacial energy assures that the temperature dependency of the interfacial energy and the nucleation rate is strictly monotonic. The strict monotonicity of both the nucleation rate and the interfacial energy
is a natural requirement conforming to the physical reality; the nucleation rate decreases with increasing temperature and the interfacial energy increases with increasing temperature \cite{ickes15}.
However, the slope method assumes, rather unphysically, that the interfacial energy is constant, and not strictly monotonic as supposed in the general case above. Under such assumption, the considerations presented above result into an even stronger requirement on the deduced interfacial energies from two intersecting nucleation rate datasets. Any constant function can be viewed as a limit of a sequence of strictly monotonic functions, e.g. a sequence of linear functions $f_n = C + a_n (T - T_0)$ with slope $a_n$ decreasing as $1/n$ for $n \rightarrow \infty$. For every element of the sequence the above reasoning applies and the deduced interfacial energy ranges must possess a non-empty intersection. In the limit of the constant function the intersection shrinks to a single value of interfacial energy, common to both datasets. As a result, the slope method must produce interfacial energies equal to each other within their standard deviation ranges from two overlapping datasets of nucleation rates. However, we already stated that the estimates of the interfacial energy of the slope method, $\gamma_M$ and $\gamma_S$, as published by Murray \textit{et al.}~\cite{murray10} are not equal within their standard deviations, which is in contradiction with the outcome of the above reasoning. Therefore, the slope method procedure as preformed by Murray \textit{et al.}~\cite{murray10} must contain a hidden inconsistency, an erroneous step that introduces additional uncertainty. Its nature will be discussed in the following section.
\section{Discussion} \label{se:discuss}
The fit of the experimental ice nucleation-rate data to the linear function (\ref{eq:nrln3}) forms the basis of the slope method. The fit results in two parameters $n$ and $m$ that will be discussed in this section. First, the evaluation of the slope $m$ resulting in the estimation of ice-water interfacial energy (\ref{eq:m}) will be investigated with the goal of resolving the inconsistency found in Sec.~\ref{se:proof}. Second, the evaluation of the nucleation rate pre-factor from the term $n$ according to Eq.~(\ref{eq:n}) will be analyzed.
\subsection{Interfacial energy} \label{se:gamma}
Although Murray \textit{et al.}~\cite{murray10} do not specify the particular regression method that they used to perform the linear fit to their data yielding their reported slope $m_{OLS} = -$ (6.02 $\pm$ 0.36)$\times$10$^7$~K$^3$, their result can be reproduced with the widely used ordinary least squares (OLS) method. By using the OLS method implemented in Python (package \texttt{scipy.stats.linregress}) to fit the Murray \textit{et al.}~\cite{murray10} experimental data to the function (\ref{eq:nrln3}) we can check that the above mentioned value of $m_{OLS}$ is recovered. Therefore, let us assume that the OLS method was indeed used by Murray \textit{et al.}~\cite{murray10} in their analysis.
\begin{figure}[t]
\centering
\includegraphics[width=86mm]{mregr2.pdf}
\caption{Comparison of the ordinary least squares regression method (dashed line) and the orthogonal distance regression method (full line) for the estimation of the slope of the ice nucleation data of Murray \textit{et al.}~\cite{murray10}.
}
\label{fig:munc}
\end{figure}
The OLS method was derived with the assumption that the observation of the independent variable, i.e. the temperature in this case, is error-free. From this point of view the applicability of the OLS method is highly questionable for nucleation-rate experimental data, because the error in the temperature measurement of an ice nucleation experiment is typically larger than $\pm$0.4~$^{\circ}$C, and it is identified as the main source of uncertainty in the nucleation measurement \cite{stan09}.
The regression method derived with the assumption of non-zero observational errors in the independent variable is the orthogonal distance regression method (ODR), also known as errors-in-variables modeling, or total least squares \cite{huffel91}. The Python ODR implementation (\texttt{scipy.odr.odrpack}) applied to Murray \textit{et al.}~\cite{murray10} ice nucleation data results in the slope $m_{ODR}=-$ (8.04 $\pm$ 0.48)$\times$10$^7$~K$^3$.
The difference in the slope estimates between the OLS and ODR methods for the Murray \textit{et al.}~\cite{murray10} ice nucleation data is shown in Fig.~\ref{fig:munc}. The ice-water interfacial energy corresponding to the $m_{ODR}$ is $\gamma_{ODR}$ = 22.9 $\pm$ 0.5 mJ/m$^2$ according to the slope method, Eq. (\ref{eq:m}). And after including the uncertainty in the cubic ice sublimation pressure \cite{shilling06} ($\pm$ 0.8 mJ/m$^2$), the final estimate of ice-water interfacial energy using the ODR method is $\gamma_{ODR}$ = 22.9 $\pm$ 1.3 mJ/m$^2$. This new ODR estimate is roughly 10 \% higher that the OLS estimate $\gamma_{OLS}$ = 20.8 $\pm$ 1.2 mJ/m$^2$ calculated from $m_{OLS}$. The new value $\gamma_{ODR}$ is sufficiently close to the value $\gamma_S$ Murray \textit{et al.}~\cite{murray10} estimated from the nucleation data of Stan \textit{et al.}~\cite{stan09}, and therefore $\gamma_{ODR}$ does not suffer from the contradiction described in Sec.~\ref{se:proof}.
Note that the Stan data~\cite{stan09} are fitted correctly even with the OLS method that fails for fitting the Murray data \cite{murray10}.
We can check that the OLS method gives the slope $m = - 8.952\times$10$^7$~K$^3$ and the ODR method results in $m = - 8.967 \times$10$^7$~K$^3$, which both correspond to the interfacial energy $\gamma_S$ = 23.7 mJ/m$^2$ reported by Murray \textit{et al.}~\cite{murray10} for the nucleation data of Stan \textit{et al.}~\cite{stan09}.
The Stan data, as presented in Fig.~\ref{fig:mse} (and in their Fig.~8 \cite{stan09}), on the contrary to the raw data by Murray \textit{et al.}~\cite{murray10}, are already averaged from a set of more than 37 thousand freezing experiments, effectively removing the large uncertainty in experimental temperature measurement. The Stan data are therefore suitable for fitting even by the OLS method.
\subsection{Nucleation rate pre-factor} \label{se:j0}
\begin{figure}[t]
\centering
\includegraphics[width=86mm]{murray_j0.pdf}
\caption{Pre-factor $J_0$ estimated by Murray~\textit{et al.}~\cite{murray10} with the slope method (as given in their Table~2) from the ice nucleation data available in literature (red circles). Horizontal errorbars show the temperature range of the input experimental nucleation-rate datasets. The vertical errorbars denote the uncertainty of $J_0$ fit. The green full line shows the pre-factor according to the volume-diffusion controlled growth model \cite{kashchiev00}. The blue square shows the OLS-based fit of the pre-factor, and the the green diamond shows the ODR-based fit of the pre-factor from the experimental data of Murray~\textit{et al.}~\cite{murray10}.
}
\label{fig:mpf}
\end{figure}
The slope method concurrently fits two parameters, $n$ and $m$, from a given data set of experimental nucleation rates. The two parameters are therefore coupled and we can expect that the larger an error in the estimation of $m$ we make, the larger the corresponding error in $n$ we get. It was shown in the preceding section that the interfacial energy estimate $\gamma_M$ suffered from an error due to the inappropriate statistical processing of the experimental data. Therefore, an error was introduced in the corresponding estimate of nucleation rate pre-factor $J_0$ as well.
The difference between the pre-factor estimates using the OLS and the ODR regression methods is shown in Fig.~\ref{fig:mpf} for the experimental data of Murray \textit{et al.}~\cite{murray10}. The fitted parameters are $n_{OLS}$ = 75 $\pm$ 5, and $n_{ODR}$ = 94.5 $\pm$ 5, respectively. By evaluating the pre-factor according to Eq.~(\ref{eq:n}), we find that the OLS-based estimate of $J_0$ is roughly 9 orders of magnitude lower than the ODR-based result. It is the ODR-based estimate of $J_0$ that is within its uncertainty equal to an independent model of $J_0$, i.e. the volume-diffusion based description of the cluster growth \cite{kashchiev00} (p.~141), as shown in Fig.~\ref{fig:mpf}.
\subsection{Final remarks} \label{se:fr}
Not only the already discussed regression issue is the source of uncertainty in the slope method. The assumptions of the method itself, as summarized in Sec.~\ref{se:method}, play a role as well. The constancy of the physical properties assumed by the slope method is plausible for small temperature ranges of the analyzed experimental nucleation rates. As the temperature range of the experiment increases, the interfacial energy, ice density, and the pre-factor are getting less accurately approximated by a constant. It is therefore desirable for the slope method to analyze temperature ranges as small as possible. On the other hand, to evaluate the slope of the ice nucleation data from the cloud of scattered experimental nucleation rates one needs as large temperature ranges as possible, because the uncertainty of the fit of the slope gets larger as the temperature range of the analyzed data decreases. Obviously, both requirements, i.e. having the temperature interval small for the assumptions of the slope method to be valid and having the temperature interval large for a precise evaluation of the slope, cannot be simultaneously satisfied. And uncertainty due to this fact will be inevitably present in the results of the slope method.
Fig.~\ref{fig:mpf} shows the $J_0$ estimates from all the ice nucleation-rate datasets in the temperature range 235 -- 240~K available in literature as calculated by Murray \textit{et al.}~\cite{murray10}. A huge scatter in the pre-factor estimates spanning roughly 25 orders of magnitude can be observed. However, the pre-factor is only slightly temperature dependent in the case of nucleation in liquids \cite{kashchiev00} (p.~199). According to the above-mentioned diffusion-based growth model the pre-factor depends linearly on the self-diffusion coefficient of supercooled water, which was measured down to 238~K with accuracy less than 1\% \cite{price99}. Therefore, the volume-diffusion based growth model gives us an estimate of the pre-factor $J_0$ with negligible uncertainty compared to the uncertainty of 25 orders of magnitude reported in Murray \textit{et al.} ~\cite{murray10}. Obviously, the scatter of $J_0$ estimates as derived by Murray \textit{et al.}~\cite{murray10} and shown in Fig.~\ref{fig:mpf} is non-physical, and reflects the errors in implementing the slope method.
Finally, the slope method was also used by Manka \textit{et al.}~\cite{manka12} to estimate the ice-water energy of 15.6 mJ/m$^2$ from their ice nucleation data measured in the temperature range 202 -- 215~K. The authors do not report any uncertainty of their interfacial energy estimate; they just state "we analyzed all of our data using the same formalism" as Murray \textit{et al.}~\cite{murray10}. Therefore, a similar regression issue as in the above-discussed case of Murray \textit{et al.}~\cite{murray10} interfacial energy estimate is likely to arise. Also, the relatively large temperature interval of the Manka data collides with the assumptions of constancy of the thermophysical properties of the slope method. But most importantly, the CNT formulation used in the slope method as presented in Sec.~\ref{se:method} is not valid for ice nucleation data measured at such low temperatures due to the omitted pressure effect, as shown recently \cite{nemec13}. The pressure-related terms in the CNT formulation present a considerable contribution to the nucleation work and, therefore, account for a shift in the CNT-predicted ice nucleation rate at low temperatures. The errors in CNT formulation are then projected to errors in the ice-water interfacial energy estimates of the slope method.
\section{Conclusion}
A method for the estimation of the ice-water interfacial energy based on the classical nucleation theory and on the evaluation of the slope of the experimental ice nucleation rates vs. a scaled temperature was thoroughly analyzed in this work. A logical contradiction in the interfacial energy estimates was identified in the work of Murray \textit{et al.}~\cite{murray10} that is related to the linear regression algorithm utilized for the evaluation of the slope of experimental ice nucleation rate data. The contradiction was removed by using the orthogonal distance regression method for a proper evaluation of the slope instead of the ordinary least squares method. The corrected estimate of the interfacial energy is by 10 \% (2.1 mJ/m$^2$) higher than the original value reported by Murray \textit{et al.}~\cite{murray10}.
In the light of the findings of this work, estimation methods \cite{huang95, nemec13} utilizing the absolute values of the experimental nucleation rates instead of the temperature derivative, and using a theoretical model of the nucleation rate pre-factor instead of its fit from experimental nucleation data, present a safer way of deducing the interfacial energy from nucleation rate data avoiding uncertainties inherently contained in the slope method.
\section*{Acknowledgement}
The author would like to acknowledge the institutional support RVO:61388998 of the Institute of Thermomechanics, v.v.i., support by the Czech Science Foundation (project GAP101/10/1819), by the CENTEM project (reg. no. CZ.1.05/2.1.00/03.0088) co-funded by the ERDF as part of the Ministry of Education, Youth and Sports OP RDI programme and, in the follow-up sustainability stage, supported through CENTEM PLUS (LO1402) by financial means from the Ministry of Education, Youth and Sports under the National Sustainability Programme I, and by the POLYMEM project (reg. no. CZ.1.07/2.3.00/20.0107) co-funded by the ESF as part of the Ministry of Education, Youth and Sports ECOP programme.
|
\section*{Introduction}
\maketitle
\section{Introduction}
The most basic way to describe a graph is to consider its metric quantities
as for instance the diameter \cite{bollobas2004diameter},
the degrees, the distances \cite{bollobas1979graph}.
In case no other information is available, a good choice is to consider randomly drawn edges \cite{erdos1959random,erdos1960evolution} and determine the expected values and distribution of
those properties in such random graphs. More recently, the computer revolution and the pervasive presence of Internet and WWW, created a whole series of
complex networks in technological systems whose properties can be directly measured from data \cite{albert2002statistical}. All the real networks show particular structures of edges, making them definitely different from random graphs. Driven by such evidence, researchers recognized analogous structures in a variety of other cases, ranging from biology to economics and finance \cite{caldarelli2007scale}.
All these structures show lack of characteristic scale in the statistical distribution of the degree and small world effect, making therefore important to understand the basic principles at the basis of their formation \cite{barabasi1999emergence,caldarelli2002scale,krioukov2010hyperbolic}.
To bring order in this huge set of systems, it would be extremely useful if we could classify the various networks by means of some specific quantity differing from case to case.
In this quest of distinguishing universal from particular behavior we decide to consider the connection with the ``curvature'' of the graph.
Embedding spaces can have negative curvature (hyperbolic spaces), zero value of curvature (Euclidean spaces), or positive curvature (spherical spaces).
On the bases of the hyperbolicity measure \cite{gromov1987essays},
it is possible to extend such a measure of curvature for manifolds to discrete networks.
Hyperbolicity measure \cite{gromov1987essays} defines the curvature for an infinite metric space with bounded
local geometry, using a 4 points condition.
In detail, the hyperbolicity $\delta(x,y,v,w)$ of a $4$-tuple of vertices $\{x,y,v,w\}$ is defined as half the difference between the biggest two of the following sums:
\begin{equation} \label{eq:sums}
d(x,y)+d(v,w),\ \ \ d(x,v)+d(y,w), \ \ \ d(x,w)+d(y,v)
\end{equation}
where $d$ denotes the distance between two vertices. If the lengths of edges are integers (in our case, the length of each edge is assumed to be 1), the hyperbolicity of a $4$-tuple of vertices is always an integer or its half.
The hyperbolicity $\delta(G)$ of a graph $G$ is commonly defined as the maximum of the hyperbolicity of a $4$-tuple of vertices \cite{bermudo2011computing, bermudo2011hyperbolicity, Carballosa2013, cohen2013exact, kennedy2013hyperbolicity}. However, for the purposes of this work, also the average hyperbolicity of a $4$-tuple will have a significant role: to distinguish the two, we will use $\delta_{\text{worst}}$ and $\delta_{\text{avg}}$, respectively \cite{albert2014topological}.
This approach attracted the interest of the community, both in modeling this phenomenon \cite{shang2012lack}, and in classifying networks from the real world \cite{abu-ata2015metric}. For example, it has been argued that several properties of complex networks arise naturally, once a negative curvature of the space has been assumed \cite{krioukov2010hyperbolic}. Similarly, others investigated the role of hyperbolicity in a series of different networks \cite{albert2014topological} ranging from social networks in dolphins, to characters in books, with the aim of discovering essential edges in path of communication. In addition, by studying structural holes, it has been shown that most of these networks are essentially tree-like \cite{albert2014topological}.
Nevertheless, nobody provided an interpretation of what is measured by hyperbolicity, and in the literature there are few applications of this quantity.
This paper wants to fill this gap, by considering {\em hyperbolicity as a measure of how much a network is ``democratic''} (the larger the hyperbolicity value the larger the ``democracy'' in the network).
Indeed, we prove mathematical results that link a small hyperbolicity constant with the existence of a small set of vertices ``controlling'' many paths, and hence with a non-democratic network (implications are true in both directions).
As far as we know, this is the first measure of democracy in a complex network, apart from assortativity \cite{newman2002assortative,mixing2003newman}. In any case, our measure is quite different from the latter one, because it is based on shortest paths and not on neighbors: consequently, the new measure is global. Moreover, it is more robust: for instance, if we ``break'' all edges by ``adding a vertex in the middle'', the hyperbolicity of the graph does not change much, but the assortativity decreases drastically.
Starting from this interpretation, we derive consequences on the structure of biological, social, and technological networks, by computing $\delta_{\text{worst}}$ and $\delta_{\text{avg}}$ on a dataset made by 93\ complex networks.
This analysis confirms previous results showing that $\delta_{\text{worst}}$ is highly influenced by ``random events'', and it does not capture a specific characteristic of the network \cite{albert2014topological}. Differently from previous works, we will also be able to quantify this phenomenon, showing that the distribution of $\frac{2\delta_{\text{worst}}}{D}$ (where $D$ is the diameter of the graph) is approximately normal. The value of $\delta_{\text{avg}}$ is instead much more robust with respect to random events, and it allows us to effectively distinguish networks of different kind. Our classification will be different from the classification provided by assortativity \cite{newman2002assortative,mixing2003newman}: for instance, a network with few influential hubs not connected to each other is democratic if we consider assortativity, while it is aristocratic in our framework.
Finally, we introduce the hyperbolicity of vertex neighbors. This is done both starting from high-degree vertices and from random vertices. Our goal is to describe the ``influence area'' of such vertices. If a neighbor is not democratic, it means that this neighbor is in the influence area of the center (this implication is particularly evident when the center is a high-degree vertex).
If the starting vertex $v$ is random, the hyperbolicity of a neighbor grows quite irregularly, with local maxima when ``more influential'' vertices are reached, and local minima when the neighbor size corresponds to the influence area of such vertices. On the other hand, if the starting vertex is already influential, the plot grows almost linearly until the hyperbolicity of the whole graph is reached, and then it is constant. We have chosen to define the influence area of this vertex as the first neighbor where half this threshold is reached.
The size of the influence area of the vertex with highest degree provides a good way to classify networks. We are able to distinguish ``local'' networks, where each node creates connections to other nodes in order to reach their own goals, and ``global'' networks, where a common goal drives the creation of the network. For instance, peer-to-peer networks are local, because each node creates connections in order to download data, while metabolic networks are global, because the creation of links is driven by the global goal of making the cell alive. We show that influence areas in a local networks is quite small, containing about $\frac{n}{10}$ vertices, where $n$ is the total number of vertices. Conversely, influence areas in global networks have a much bigger size, close to $\frac{n}{3}$ vertices.
\section{Methods}
The main result of this paper is to show the link between the hyperbolicity of a graph and its ``democracy''. This interpretation is motivated by the following propositions: the first one shows that, if for some vertices $v,w$, $\max_{x,y} \delta(x,y,v,w)$ is not high, then there is a set of small diameter that ``controls'' all approximately shortest paths from $x$ to $y$. Consequently, a network with low hyperbolicity is not ``democratic'', because shortest paths are controlled by small sets.
\begin{proposition}
Let $v,w$ be two vertices in a network $G=(V,E)$, let $B_r(v)$ be the $r$-neighborhood of $v$ (that is, the set $\{u \in V:d(u,v) \leq r\}$), and let $B_s(w)$ be the $s$-neighborhood of $w$. Then, the diameter of the set $X=B_r(v) \cap B_s(w)$ is at most $2\max_{x,y} \delta(x,y,v,w)+r+s-d(v,w)$.
\end{proposition}
\begin{proof}
Let $x,y$ be two vertices in $X$. Then,
\begin{align*}
2\delta(x,y,v,w) \geq & d(x,y)+d(v,w)-\max(d(x,v)+d(y,w), \\
& d(x,w)+d(y,v)) \geq \\
\geq & d(x,y)+d(v,w)-(r+s).
\end{align*}
Taking the maximum over all possible $x,y \in X$, we obtain $D(X)=\max_{x,y \in X} d(x,y) \leq 2\max_{x,y}\delta(x,y,v,w)+r+s-d(v,w)$, where $D(X)$ is the diameter of the set $X$.
\end{proof}
The second proposition is a sort of converse: if there is a set of vertices controlling the shortest paths of a given $4$-tuple, the hyperbolicity of that $4$-tuple is low. Consequently, if the hyperbolicity is high, then there is not a small set of vertices controlling many shortest paths, and the network is democratic.
\begin{proposition}
Let $x,y,v,w$ be $4$-tuple of vertices, and let us assume that there exists a set $C\subseteq V$ of diameter $D$ such that all shortest paths between $x,y,v,w$ pass through $C$. Then, $\delta(x,y,v,w) \leq D$.
\end{proposition}
\begin{proof}
We may assume without loss of generality that $d(x,y)+d(v,w) \geq d(x,v)+d(y,w) \geq d(x,w)+d(y,v)$. Then, if we denote by $d(x,C):=\min_{c \in C} d(x,c)$,
\begin{align*}
2\delta(x,y,v,w) &= && d(x,y)+d(v,w) - d(x,v)-d(y,w) \\
& \leq && d(x,C)+D+d(C,y)+d(v,C)+D \\
&&& +d(C,w)-d(x,C)-d(C,v) \\
&&& -d(y,C)-d(C,w) \\
&=&& 2D.
\end{align*}
\end{proof}
These two results formalize our connection between the hyperbolicity constant and how democratic a complex network is. This work will confirm this interpretation by analyzing a dataset of 93\ graphs, made by 19\ biological networks, 32\ social networks, and 42\ technological networks, and it will draw conclusions on which networks are more democratic than others.
The first data check extends some of the activity already done \cite{krioukov2010hyperbolic,albert2014topological}: for each network in our dataset, we computed the distribution of $\frac{2\delta_{\text{worst}}}{D}$, where $D$ is the diameter of the graph (this value is always between $0$ and $1$ \cite{Fang2011}). With respect to the previous papers, we got more data referring to larger networks, and we therefore deal with lower statistical errors.
This is particularly important, since it is known that $\delta_{\text{worst}}$ does not capture structural properties of a network. Furthermore, its behavior is not robust,
in the sense that small modifications in the graph can significantly change the value of $\delta_{\text{worst}}$ \cite{albert2014topological}. Moreover, computing $\delta_{\text{worst}}$ is difficult task, since the best ``practical'' algorithm \cite{cohen2013exact} has running time $\O(n^4)$. We have been able to overcome these issues by focusing on $\delta_{\text{avg}}$, already considered in the literature \cite{albert2014topological}, but not deeply analyzed.
In particular, for each graph in the dataset, we have considered the ratio $\frac{2\delta_{\text{avg}}}{d_{\text{avg}}}$, where $d_{\text{avg}}$ is the average distance between two randomly chosen vertices \cite{Fang2011} (this parameter takes values in the interval $[0,1]$).
Although the exact computation of $\delta_{\text{avg}}$ is also hard, the value $\frac{2\delta_{\text{avg}}}{d_{\text{avg}}}$ can be easily approximated through sampling. More specifically, if we consider $N$ $4$-tuples of vertices with hyperbolicity $\delta_1,\dots,\delta_N$, and since $a_i=0 \leq \delta_i \leq 1=b_i$ for each $i$, by the Azuma-Hoeffding inequality \cite{hoeffding1963probability} we obtain:
\begin{align*}
\P\left(\left|\frac{\sum_{i=1}^N \delta_i}{N}-\delta_{\text{avg}}\right| \geq t \right) &\leq 2e^{\frac{-N^2t^2}{2\sum_{i=1}^N (b_i-a_i)^2}} \leq \\
& \leq 2e^{-\frac{Nt^2}{2}}.
\end{align*}
We choose to sample $N=10,000,000$ $4$-tuples of vertices, obtaining an estimate $\bar{\delta}_{\text{avg}}$. The previous inequality applied with $t=0.001$ yields:
\begin{align*}
\P\left(\left|\bar{\delta}_{\text{avg}}-\delta_{\text{avg}}\right| \geq 0.001 \right) &\leq 2e^{-5} \leq 0.7\%.
\end{align*}
This precision is more than sufficient for our purposes.
Finally, we have analyzed neighbors of a given vertex $v$, approximating their ratio $\frac{2\delta_{\text{avg}}}{d_{\text{avg}}}$ as before, in order to analyze the influence area of $v$. The size of the neighbor considered ranges from the degree of $v$ to $n$, with steps of $10$ vertices (where $n$ is the total number of vertices in the graph). We have chosen $v$ as the maximum degree vertex, which intuitively should have a big influence area, and we have used as a benchmark of comparison the same results from a random vertex.
We have defined the ``influence area'' of $v$ as the biggest neighbor where $\frac{2\delta_{\text{avg}}}{d_{\text{avg}}}$ is at most half than the same value in the whole graph. However, in order to avoid ``random deviations'' (especially, when the neighbor is small), in our experiments we have considered the fourth neighbor where this event has occurred. The purpose of this analysis is twofold: not only we define and compute the influence area of a vertex, but we also classify networks according to the size of this influence area.
\section{Results}
\subsection{Worst-Case Hyperbolicity.}
The most common approach is to consider the maximum hyperbolicity of a 4-tuple of vertices, that is, $\delta_{\text{worst}}(G)$. Despite some attempts in proving that real-world networks usually have low hyperbolicity \cite{kennedy2013hyperbolicity}, it soon became clear that the small values obtained are consequences of small world networks \cite{albert2014topological}, since $0 \leq 2\delta_{\text{worst}} \leq D$ (where $D$ is the diameter of the graph). In particular, in a dataset of small social and biological networks \cite{albert2014topological} there is no relation between $2\delta_{\text{worst}}$ and $D$. The ratio between these values varies between $25\%$ and $89\%$.
In this paper, we make a more detailed analysis, working with larger networks, between one hundred and several thousand of vertices.
The results obtained are shown in Figure \ref{fig:deltaworst}, which also contains specific results dealing only with social, biological and technological networks.
These results show that the distribution of the ratio $\frac{2\delta_{\text{worst}}}{D}$ is approximately Gaussian, both in the whole dataset and in each single kind of network. The average ratio is $0.521$, and the standard deviation is $0.085$. Moreover, a Chi-square goodness of fit test applied to the previous data does not reject the hypothesis that the distribution is Gaussian with mean $0.5$ and variance $0.085$, with a very high confidence level \cite{ross2010statistics}. This result confirms that the hyperbolicity of real-world networks is not much ``smaller than expected'', result already obtained in the past \cite{albert2014topological}. However, we are able to perform a further step: the Gaussian probability distribution makes us think that $\delta_{\text{worst}}$ is influenced by random events. Indeed it does not reflect particular characteristics of the network, since the same distribution arises from networks of different kinds.
\begin{figure*}
\includegraphics[width=\textwidth]{figure1.png}
\caption{The distribution of $\frac{2\delta_{\text{worst}}}{D}$ in the graphs in our dataset. The bar corresponding to the value $p$ contains all networks where $p-0.5 < \frac{2\delta_{\text{worst}}}{D} \leq p+0.5$.\label{fig:deltaworst}}
\end{figure*}
Social networks show a slightly different behavior, since many of them have a larger value of $\frac{2\delta_{\text{worst}}}{D}$, between $0.65$ and $0.75$. However, this is due to the presence of several financial (e-MID, a platform for interbank lending) networks, where the ratio is often $\frac{2}{3}$ or $\frac{3}{4}$ since the diameter is $3$ or $4$.
Despite this particular case, we may conclude that $\delta_{\text{worst}}$ is not a characteristic of the network, but it mainly depends on ``random events'' that have a deep impact on the value of $\delta_{\text{worst}}$.
This conclusion is further confirmed by the particular case of the e-MID networks: this parameter changed from $0.750$ in 2011 to $0.286$ in 2012, only because a simple path of length $3$ increased the diameter from $4$ to $7$.
\subsection{Average Hyperbolicity.}
In the past, the average hyperbolicity $\delta_{\text{avg}}$ of a 4-tuple of vertices was rarely analyzed: the only known result is that it is usually significantly smaller than $\delta_{\text{worst}}$ \cite{albert2014topological}.
However, we think that this parameter may provide very interesting results, because it is robust, in the sense that it does not change much if few edges of the graph are modified. Furthermore, it is easily approximable through sampling (while computing $\delta_{\text{worst}}$ takes time $\O(n^4)$).
Similarly to what we have done in the analysis of $\delta_{\text{worst}}$, we have considered the ratio $\frac{2\delta_{\text{avg}}}{d_{\text{avg}}}$, where $d_{\text{avg}}$ denotes the average distance in the network (also this parameter lies in the interval $[0,1]$).
The results obtained are plotted in Figure \ref{fig:deltaavg}.
\begin{figure*}
\includegraphics[width=\textwidth]{suppfig1.pdf}
\caption{The value $\frac{2\delta_{\text{avg}}}{d_{\text{avg}}}$ of all the networks in our dataset.\label{fig:deltaavg}}
\end{figure*}
The picture shows that the average hyperbolicity is usually an order of magnitude smaller than the average distance: in this sense, real-world networks are indeed hyperbolic. Moreover, it is possible to make a distinction between networks that are ``more democratic'', like the e-MID networks or the peer-to-peer networks (where the hyperbolicity is large), and networks that are ``more centralized'', like some social networks and most autonomous systems networks.
\subsection{Hyperbolicity of Neighbors.}
Since the hyperbolicity of a graph is closely related to the existence of a small part of the graph controlling most shortest paths, we have analyzed which subgraphs of a given graph have small hyperbolicity. Intuitively, these subgraphs should be ``less democratic'' than the whole graph, in the sense that they are contained in the ``influence area'' of a small group of vertices. In this analysis,
we have tried to spot the influence area of a single vertex, by measuring $\frac{2\delta_{\text{avg}}}{d_{\text{avg}}}$ on neighbors of $v$ in increasing order of size.
In order to prove the effectiveness of this approach, we have first tested a synthetic power-law graph \cite{lancichinetti2009benchmarks} made by three communities of $1000$ vertices each (see the lowest plot in Figure \ref{fig:deltaneighsinglegraphs}). We have computed the hyperbolicity of neighbors of the vertex $v$ with highest degree: we can see a local minimum close to the size of a community. In our opinion, this minimum appears because the neighbor is ``dominated by the community'', and consequently by the center $v$ of the community. This result confirms the link between the value of $\frac{2\delta_{\text{avg}}}{d_{\text{avg}}}$ and the influence area of a vertex.
Finally, we passed to the analysis of neighbors in real-world networks. The upper plots in Figure \ref{fig:deltaneighsinglegraphs} show the same results for one network of each kind:
\begin{itemize}
\item a social network, the General Relativity and Quantum Cosmology collaboration network;
\item a biological network, the yeast metabolic network;
\item a technological network, the peer-to-peer Gnutella network in 2004.
\end{itemize}
As a benchmark of comparison, we have also considered the hyperbolicity of neighbors of a random vertex.
\begin{figure*}
\includegraphics[width=\textwidth]{figure2.png}
\caption{The value of $\frac{2\delta_{\text{avg}}}{d_{\text{avg}}}$ for neighbours of a randomly chosen vertex (up left), or the maximum degree vertex (up right); $S_n$ is the number of vertices in the neighbours, while $S$ is the total number of vertices. Results are shown for a social network, a biological network, a technological network, and (below) a synthetic network.\label{fig:deltaneighsinglegraphs}}
\end{figure*}
The plots show that the hyperbolicity of a neighbor of the maximum degree vertex grows almost linearly with the neighbor size, until it converges to the hyperbolicity of the whole graph. Convergence time differs from graph to graph. In biological networks, convergence was reached at size close to $\frac{n}{2}$, while in the social and in the technological networks convergence is reached before. For neighbors of a random vertex, we outline a different behavior: at the beginning, the growth is not monotone, like in the previous case, and it is much more irregular. In our opinion, this is due to the fact that, when the neighbor grows, it reaches more and more ``influential'' vertices, and the first neighbor that touches such a vertex corresponds to a local maximum in the plot. After some steps, the hyperbolicity grows more and more regularly, because we have reached a very influential vertex $w$, and from that point on we are mainly considering the influence area of $w$, not of $v$. This issue is further confirmed by Figure \ref{fig:derivative}, where the derivative of the average hyperbolicity is shown. Hence, this experiment outlines two significant point: on the one hand, it allows us to define the influence area of $v$ as the neighbor where the first local maximum is reached; on the other hand, it gives motivations to the analysis of neighbors of influential vertices, because their influence area has a strong impact on the topology of the whole network.
\begin{figure}
\includegraphics[width=.48\textwidth]{figure3.png}
\caption{The derivative with respect to the neighbors proportion $S_n/S$ of the value of $\frac{2\delta_{\text{avg}}}{d_{\text{avg}}}$, in neighbors of the maximum degree vertex and of random vertices.
\label{fig:derivative}}
\end{figure}
For this reason, we have focused on the maximum degree vertex, and, in order to have more general results, we have analyzed all graphs in the dataset. Figure \ref{fig:firstneigh} shows the size of the maximum neighbor having ratio $\frac{2\delta_{\text{avg}}}{d_{\text{avg}}}$ at least half than the same ratio in the whole graph. Actually, in order to exclude random deviations from our analysis, we have plotted the fourth neighbor where the condition is satisfied.
The figure shows that the influence area of an individual is low in social and peer-to-peer networks, compared to biological or autonomous system network.
This standard behavior has few exceptions: first of all, protein-protein interaction networks (\texttt{string}, \texttt{ecoli.interaction}) are different from other biological networks, and the influence area is smaller. Furthermore, the social network \texttt{GoogleNW} contains a vertex with an enormous influence area: this network is the set of Google pages, and the central vertex $v$ considered is the page \texttt{www.google.com}, which clearly dominates all the others. Another particular case is the social network \texttt{facebook\_combined}: this network is a collection of ego-networks from Facebook, and links are made if common interests are retrieved. We think that this network is different from the others because it is a small subgraph of a bigger graph (where all Facebook users are considered), and the choice of the subgraph has a strong impact on the topology of the network, which does not reflect the standard behavior.
\begin{figure*}[h!]
\includegraphics[width=\textwidth]{suppfig2.pdf}
\caption{The size of the fourth neighbour of the maximum degree vertex having $\frac{2\delta_{\text{avg}}}{d_{\text{avg}}}$ at least half than the same value in the whole graph.\label{fig:firstneigh}}
\end{figure*}
\section{Discussion} \label{sec:IV}
In the literature, several works have analyzed the hyperbolicity of a complex network. They used this quantity in order to classify real-world networks and in order to draw conclusions about the impact of hyperbolicity on the network topology.
However, these works are mainly based on the analysis of $\delta_{\text{worst}}$, which has two drawbacks: it is not \emph{robust}, that is, small modifications on the network can have deep impacts on its value, and it is not \emph{scalable}, that is, it can be exactly computed only on small networks.
In this work, we confirmed these conclusions, and we have proposed a different approach: using $\delta_{\text{avg}}$ instead of $\delta_{\text{worst}}$, a parameter already considered in the literature. We interpreted this parameter as a measure of ``democracy'' in a network, and we classified different networks according to how democratic they are.
In particular, we have shown that technological autonomous system networks are less ``democratic'' than social or biological networks, in agreement with our intuition (since AS graphs have a ``built-in'' hierarchy, while in social networks everyone has the same role). Moreover, we have applied this concept to neighbors of influential nodes: we have seen that larger neighbors are more democratic, and the relationship between the size and the hyperbolicity is almost linear, until a threshold is reached. This analysis clearly outlines the influence area of a node, whose size strongly depends on the graph considered. We have shown that nodes have a rather small influence area in social and peer-to-peer networks, while in autonomous systems and biological networks the influence area can be close to half the graph. A possible explanation of this behavior is that the former networks are ``distributed'', in the sense that each node has a goal (downloading in peer-to-peer networks, and creating relationships in social networks), and edges are created locally by nodes that try to reach the goal. On the other hand, the latter networks have global goals (connecting everyone in the network, or making a cell live), and the creation of edges is ``centralized''. Our analysis is able to distinguish graphs of these two kinds.
These results prove that democracy in a complex network is well formalized by our definition of hyperbolicity, since the consequences of our interpretation are coherent with intuitive ideas.
Furthermore, we have provided an application of this interpretation, making it possible to define and analyze the influence area of a node.
As far as we know, this is the first work that provides this interpretation of the average hyperbolicity of a graph.
Possible applications include not only the classification of networks according to this parameter, but also the classifications of nodes in a network, or the classification of different communities. These communities might be democratic, if everyone has ``the same role'' and the hyperbolicity is high, or not democratic, if there is a group of few nodes that keeps the community together, making the average hyperbolicity small.
Future researches may address this characterization, and they may also classify nodes in a network according to their influence area, in order to distinguish important hubs from peripheral node. Finally, it would be interesting to compare this approach with other approaches to determine the centrality of a node (closeness centrality, betweenness centrality, eigenvector centrality, and so on), in order to better outline the features of this analysis.
\bibliographystyle{abbrv}
|
\section{Introduction}
\label{intro}
Active galactic nuclei (AGN) are very energetic extragalactic sources powered by accretion on supermassive black holes. Blazars are a class of AGN characterized by a relativistic jet pointing towards the observer. Because of the beaming, the emission originating in the jet dominates the observed spectrum.
The quasar 3C 273 is the brightest and hence one of the best monitored AGN. Located at a distance of $\mathrm{z} = 0.158$ \citep{Strauss1992}, it features interesting blazar-like (a jet with superluminal motion and high variability) and Seyfert-like (the strong blue bump, the soft X-ray excess, and variable emission lines) characteristics.
Several observational campaigns have been carried out on 3C 273, starting from the broad multiwavelength campaign by \citet{Lichti1995}, which made use of data spanning from radio to $\gamma$ rays, to more recent observations in X-rays. \citet{Courvoisier2003} described simultaneous observation of INTEGRAL, XMM-Newton, and RXTE performed in January 2003. \citet{Chernyakova2007} used simultaneous INTEGRAL and XMM-Newton data in the period 2003--2005 and compared their results to historical data to investigate the secular evolution of 3C 273. \citet{Liuliu2011} used XMM-Newton data to investigate the nature of the soft X-ray excess.
\begin{figure*}
\centering
\includegraphics[angle=-90, width=\hsize]{lightcurves3-eps-converted-to.pdf}
\caption{\textit{Top panel:} RXTE-PCA lightcurve in 0.2 - 75 keV. \textit{Middle panel:} Fermi-LAT lightcurve in 0.1 - 100 GeV. Area marked in light grey colour shows the selected flaring time epochs in both lightcurves. \textit{Bottom panel:} radio lightcurve at 37 GHz. Time units are days on bottom axis (Modified Julian Date) and decimal years on top axis.}
\label{fig_fltime}
\end{figure*}
Many works have been published to describe in detail the spectral and temporal characteristics of 3C 273, which has been widely used as a test case to validate theoretical models and predictions. \citet{Paltani1998} found two distinct variable components in the optical--ultraviolet band. \citet{Soldi2008} studied the temporal variability in a wide energy range. \citet{Grandi2004} were able to decouple the non-thermal (jet) and thermal (accretion flow) emission by studying the long-term spectral variability of BeppoSAX data. \citet{Marcowith1995} tested their model of relativistic electron--positron beam for blazar emission. \citet{Marscher1985} and \citet{Turler2000} modelled the radio emission in the context of synchrotron emission from relativistic electrons produced by shocks in the jet. \citet{Artyukh2012} investigated the physics and the condition of the core of 3C 273 on parsec scale. A detailed review has been written by \citet{Courvoisier1998} and a database collecting public data of the source over several decades is available online \citep{Turler1999b,Soldi2008}.
The aim of this work is to use contemporaneous X-ray to $\gamma$-ray observations obtained with INTEGRAL and Fermi for the first time since the CGRO campaign \citep{Lichti1995} to study the origin of the non-thermal spectral components. In Sect. \ref{data} we introduce the data used for this work and explain how the datasets have been buildt. In Sect. \ref{temp} we present the results of the cross-correlation analysis between lightcurves obtained with Fermi-LAT, RXTE-PCA, and at 37 GHz and compare our results with those of previous studies. We find that $\gamma$-ray and radio emission are correlated and that the X-ray component has a different origin.
In Sect. \ref{spec} we describe the high energy spectrum of 3C 273 and its variability. Data from RXTE-PCA (0.2 - 75 keV), JEM-X, ISGRI and SPI on board INTEGRAL (3 keV - 1 MeV), and Fermi-LAT (100 MeV - 100 GeV) are used. We selected two different spectral states based on the behaviour of the PCA and LAT lightcurves, and fit several models to their spectra to investigate the differences between the two states. We find that the best fit model is a two-component model, which can be interpreted as a Seyfert-like component dominating the X-ray emission, and blazar-like component dominating at $\gamma$ rays. In Sect. \ref{disc} we discuss the results and give a physical interpretation of the data, focusing on the blazar-like component, in the context of relativistic jet emission and synchrotron self Compton (SSC) model, and finally we summarize the results in sect. \ref{concl}.
\section{The data sample}
\label{data}
\revi{For this work we used data from RXTE-PCA, INTEGRAL and Fermi-LAT, which provide a broad spectral coverage from a few keV up to 100 GeV, and radio data collected at the Mets\"ahovi\ Radio Observatory\footnote{http://metsahovi.aalto.fi/en/}.}
\revi{Fermi-LAT \citep[0.1 - 100 GeV,][]{Atwood2009} data are used to study temporal and spectral variability at high energies. Thanks to its large field of view, Fermi-LAT has observed 3C 273 almost continuously. The SED is extracted with the Maximum Likelihood fitting technique commonly used to analyse Fermi data and the light curve is built using the LAT Aperture Photometric method assuming the average spectral index of 2.45 as listed in the Fermi-LAT 2nd Catalog \citep{Nolan2012} and a binning time of 1 day. The data spans from August 2008 to July 2012 and they are extracted using the Fermi Science Tools version v9r27p1 and the instrument response function P7\_V6.}
\revi{INTEGRAL \citep{Winkler2003} JEM-X (6 - 20 keV), ISGRI (17 - 210 keV) and SPI (50 - 770 keV) data are used for the spectral analysis. The JEM-X and ISGRI data are analysed with OSA 10 \footnote{http://www.isdc.unige.ch/integral/analysis}, while SPI spectra are built with the SPI\_TOOLSLIB tool \citep{Knodlseder2004} and the spectra extracted by a model fitting method. They span from January 2003 to January 2012.}
\revi{RXTE-PCA \citep[0.2 - 75 keV,][]{Jahoda1996} data are collected with the HEAVENS web interface\footnote{http://www.isdc.unige.ch/heavens/} \citep{Walter2010} and span from March 2005 to December 2011. During this period 3C 273 has been observed by RXTE-PCA every few days for $ \sim $ 2 ks with only seven long gaps (see Fig. \ref{fig_fltime} upper panel).}
\revi{The radio data at 37 GHz are obtained at the Mets\"ahovi\ Radio Observatory and span from September 2006 to June 2012. They are used for timing analysis to probe the correlation with the high energy emission. The flux density scale is set by observations of DR 21. Sources NGC 7027, 3C 274 and 3C 84 are used as secondary calibrators. A detailed description of the data reduction and analysis is given in \citet{Terasanta1998}. The error estimate in the flux density includes the contribution from the measurement rms and the uncertainty of the absolute calibration.}
\section{Temporal Analysis}
\label{temp}
\subsection{Cross-Correlation}
\label{temp_cc}
In the multiwavelength variability study of 3C 273 by \citet{Soldi2008}, correlations between the lightcurves obtained at various frequencies were studied. In particular hard X-rays ($\gtrsim 20~\mathrm{keV}$) and radio-mm were found to be correlated, while no correlation was found with lower energy X-rays ($\lesssim 20~\mathrm{keV}$) or radio-mm lightcurves. A difference between the hard ($\gtrsim 20~\mathrm{keV}$) and soft X-rays was also found in the fractional variability spectrum. This suggests a different origin for the X-rays above and below $\sim 20~\mathrm{keV}$.\\
In order to study the origin of the high energy emission we used lightcurves obtained with Fermi-LAT, RXTE-PCA and in the radio with the Mets\"ahovi\ Radio Telescope.
The lightcurves are shown in Fig. \ref{fig_fltime}.
The LAT and PCA lightcurves show a different behaviour. The LAT data show a stable quiescent flux at $\sim 0.3 \cdot 10^{-6}~\mathrm{photons~s^{-1}~cm^{-2}}$ with small flares randomly distributed and three very strong flares in September 2009 (MJD from 55071 to 55097). These strong flares are described in detail by \citet{Abdo2010} and \citet{Rani2013}. \citet{Rani2013} found a variability time scale of the order of $\sim 0.1$ day for the shortest flare and of $\sim 0.5$ -- 1 day for the others. The PCA lightcurve shows several consecutive flares with alternating high and low states, the rate varying from $\sim 5$ to $\sim 30~\mathrm{photons~s^{-1}~cm^{-2}}$.\\
\begin{figure}
\centering
\includegraphics[width=\hsize]{correlation2-eps-converted-to.pdf}
\caption{Cross-correlation between Fermi-LAT, RXTE-PCA and radio lightcurves. A positive lag for an "X-Y" correlation means a delay of Y with respect to X.}
\label{fig_corr}
\end{figure}
To study correlations, we used the discrete correlation function \citep[DCF, ][]{Edelson1988} with a time bin of 10 days. In a cross-correlation between the "X" and "Y" lightcurves, labelled "X-Y" in Fig. \ref{fig_corr}, a positive time lag means that the Y lightcurve is delayed with respect to X. Cross-correlation curves are shown in Fig. \ref{fig_corr}. The solid line represents the cross-correlation between the radio and the LAT data. The correlation is peaked at $\sim +120$ days with a coefficient of $\sim 0.63$ suggesting that $\gamma$ rays are leading the radio emission. Dashed and dotted lines represent the absence of significant correlation of the PCA lightcurve with the LAT or radio lightcurves.
This result is in agreement with the observation of \citet{Soldi2008}, suggesting that the hard X-ray photons come from the jet, whereas Softer X-rays instead are dominated by a different component.
\subsection{Using the IC response}
\label{temp_conv}
\begin{figure}
\centering
\includegraphics[angle=-90, width=\hsize]{response2m30-eps-converted-to.pdf}
\caption{Profile of the best radio response, the \textit{y} axis is in arbitrary scale. The maximum is located at 160 days after the beginning of the curve. The FWHM of the curve is 197 days. See text for details.}
\label{fig_resp}
\end{figure}
\begin{figure}
\centering
\includegraphics[angle=-90, width=\hsize]{radio_rlcsim18_bis-eps-converted-to.pdf}
\caption{Radio lightcurve (black) compared with the convolution obtained from the LAT lightcurve (red). Fermi-LAT lightcurve is shown in the bottom of the picture for comparison (not in scale). Vertical dotted line is MJD = 54700 days. See text for details.}
\label{fig_conv}
\end{figure}
The connection between the radio and the $\gamma$-ray emission in radio-loud AGN has been widely investigated proving that radio and $\gamma$-ray luminosities are correlated (see for example \citet{Padovani1993,Valtaoja1999,Jorstad2001,Lahteenmaki2003,Marscher2008,Hovatta2009,Ackermann2011}).
In order to better study the link between the GeV and the radio emission, we try to evaluate whether the observed radio variability could arise from a simple convolution of the Fermi lightcurve.
The response function would convolve each spike of the LAT lightcurve to a broader and delayed radio flare. The sum of all these broadened flares, which may be superimposed, should match the observed radio lightcurve. The response would then naturally introduce the delay seen with the cross-correlation of Fig. \ref{fig_corr}, caused by the cooling time of the electrons emitting synchrotron radiation and the expansion of the emitting region, becoming optically thin at lower energies.
As the convolution function is not known a priori, we rely on the shape of a typical outburst in 3C 273 at radio frequencies. \citet{Turler1999,Turler2000} developed a method -- in the frame of a shock-in-jet scenario -- to extract from the long-term multi-wavelength lightcurves the temporal evolution of a typical synchrotron outburst at any frequency. We use here the model lightcurve of an average outburst of 3C 273 at 37 GHz as derived from the observations obtained between 1980 and 2000 and with the approach and parameters described by \citet{Turler2007}. This model lightcurve rises relatively steeply to reach a maximum in less than a year before decaying gradually.\\
With this response function we could not reproduce the radio data because the peak of the response profile is too late and broad when compared to the observations. The observed radio lightcurve can only be reproduced by compressing the response by a factor of 2 in time scale and shifting it by $-30$ d. This compressed and shifted response, shown in Fig. \ref{fig_resp}, allow us to match the width and the position of the observed radio flares.
The necessity to shrink the response can be explained considering that the convolution smears the complex $\gamma$-ray variability, with the result that a sequence of short $\gamma$-ray flares separated by short time lag will be observed in the radio as a single broad flare. As explained above, the response curve was constructed on long-term multi-wavelength lightcurves, which do not include $\gamma$-ray data. It is hence possible that what was seen in the construction of the response as a single flare, i.e. a single shock wave in the modelling of \citet{Turler1999,Turler2000}, might actually be the superposition of several smaller shocks.\\
It has to be pointed out that the available LAT data start later than the radio data. So we cannot reproduce correctly the early radio signal. We added therefore some fake $\gamma$-ray data to reproduce the radio signal at MJD $\sim$ 54700.
We also needed to change the amplitude of the response from flare to flare to match correctly the two peaks in the radio band at MJD $\sim 55000$ and $\sim 55400$. To obtain a qualitatively good match between our synthetic light curve and the radio data we used four different response normalizations. There is no physical constraint for these responses to be the same for different radio outbursts: the physical details of the shocks are inside the response function, and different shocks may undergo different physical conditions. Table \ref{tab_pconv} shows the scaling factors used to modify the original response and the time intervals. The first response is used to build the synthetic lightcurve before the beginning of LAT data (relying on the fake $\gamma$-ray data introduced). The scaling factor is arbitrary chosen to be relative to the amplitude of the response in the first time interval.
With these prescription we are able to qualitatively reproduce the radio light curve after MJD $\sim 54700$. Figure \ref{fig_conv} shows the radio lightcurve (black line) and the curve derived from the LAT lightcurve with the described technique (red line). The LAT lightcurve is also drawn for comparison.
Since the maximum of the response is at $+160$ days (Fig. \ref{fig_resp}), the average lag between $\gamma$-ray and radio flares is therefore of the order of $140 \pm 20$ days.
\begin{table}
\caption{Time start and relative amplitude scaling of the several response profiles used to construct the synthetic lightcurve (red line in Fig. \ref{fig_conv}). See text for the details.}
\label{tab_pconv}
$$
\begin{array}{cc}
\hline\hline
\noalign{\smallskip}
\mathrm{Time\: interval\: in\: MJD} & \mathrm{Scaling\: factor} \\
\hline
\noalign{\smallskip}
53971 - 54710 & 1 \\
\hline
\noalign{\smallskip}
54710 - 54890 & 2.20 \\
\hline
\noalign{\smallskip}
54890 - 55340 & 1.04 \\
\hline
\noalign{\smallskip}
\mathrm{after}\:\: 55340 & 1.83 \\
\hline
\end{array}
$$
\end{table}
\section{Spectral Analysis}
\label{spec}
To differentiate and further investigate the nature of the X-rays and the hard X-ray to $\gamma$-ray components suggested by the cross-correlation analysis (Sect. \ref{temp}) we built the SED of 3C 273 in two different spectral states. We defined the \emph{PCA flaring state} as the time intervals when the source flux \revi{in the RXTE-PCA lightcurve} exceeds $18~\mathrm{photons~s^{-1}~cm^{-2}}$ and remain larger than this value for at least 9 days. Similarly we define the \emph{LAT flaring state} as the time intervals when the source flux \revi{of the Fermi-LAT lightcurve} goes beyond $1.2 \cdot 10^{-6}~\mathrm{photons~s^{-1}~cm^{-2}}$ and remain larger than this value for at least 3 days. In case of two consecutive flares we expanded the selected interval to the full width of the flaring period. The epochs defining the flaring states are shadowed in Fig. \ref{fig_fltime}.
For both spectral states, we built a multiwavelength spectrum from the X-rays to the $\gamma$ rays, using data from RXTE, INTEGRAL, and Fermi selected in the time intervals identified as the PCA and LAT flaring state s.
\revi{The signal-to-noise ratios are enough to build multiwavelength spectra for each of the datasets, but not to follow the temporal evolution of the spectral parameters during the individual flares. Moreover, INTEGRAL data are not available for the total duration of the flaring times identified.}
In the following discussion we will also compare our datasets with the SED obtained from the multiwavelength campaign on 3C 273 performed in 1991 \citep{Lichti1995} and built with data from GINGA (2 - 30 keV) and data from OSSE, COMPTEL and EGRET on board CGRO (20 keV - 10 GeV). X-ray and $\gamma$-ray data showed a break at $\sim 1~\mathrm{MeV}$. \citet{Lichti1995} tested several models on their whole SED which includes also radio, infrared, optical and ultraviolet data.
We first tested a simple cutoff powerlaw model $N(E) = N_{cp} E^{-\Gamma} \exp(-E/E_c)$, which is not able to fit the data: the $\chi_{red}^2$ is $\sim 48$ for the LAT flaring state, $\sim 37$ for the PCA flaring state\ and $\sim 4$ for the CGRO dataset. In the LAT and PCA flaring state s the flux and the curvature of the LAT data is clearly not consistent with an exponential cutoff, and also in the CGRO dataset\ EGRET data are clearly above the exponential cutoff.
We then used several models to fit the datasets: a \textit{broken powerlaw}, a \textit{cutoff powerlaw} and a \textit{log-parabola}. None of these models has been used in \citet{Lichti1995}, where the high energy data were fitted with an empirical function to model the break at $\sim 1~\mathrm{MeV}$.
The log-parabola model (see sect. \ref{spec_lpmodel}) is able to reproduce the typical inverse-Compton bumps measured by Fermi in blazar spectra, which show an intrinsic curvature \citep{Nolan2012} and can be interpreted with stochastic acceleration of the electrons \citep[and reference therein]{Massaro2004,Tramacere2011}.\\
Figures \ref{fig_bkp}, \ref{fig_lp} and \ref{fig_lpcop} show the multiwavelength spectra for the LAT and PCA flaring state s, plus the CGRO dataset\ from \citet{Lichti1995}. Tables \ref{tab_bkp}, \ref{tab_lp} and \ref{tab_lpcop} list the best fit parameters, reduced $\chi^2$ and degree of freedom for each dataset and each model tested. Errors are at $2.7 \: \sigma$ and the parameters without error are fixed.
In general the CGRO dataset\ is closer to the PCA flaring state\ whatever model is used to fit the data. The main difference is the overall luminosity, which is systematically lower than during the LAT and PCA flaring state s. \revi{We have also built the spectrum of the quiescent state, i.e. when the source is not flaring either in RXTE-PCA or in Fermi-LAT. It has a similar shape and is 2 times brighter than observed by CGRO.}
\begin{figure}
\centering
\includegraphics[angle=-90, width=\hsize]{fitted_res_bkp-eps-converted-to.pdf}
\caption{3C 273 multiwavelength spectra at different epochs fitted with the broken powerlaw model. The upper panel shows the SED and the best fit models (continuous lines), the lower panel shows the residuals from the best fit models.}
\label{fig_bkp}
\end{figure}
\begin{figure}
\centering
\includegraphics[angle=-90, width=\hsize]{fitted_res_lp-eps-converted-to.pdf}
\caption{Same as Fig. \ref{fig_bkp} for the log-parabola model.}
\label{fig_lp}
\end{figure}
\begin{figure}
\centering
\includegraphics[angle=-90, width=\hsize]{fitted_res_lpcop-eps-converted-to.pdf}
\caption{Same as Fig. \ref{fig_bkp} for the cutoff powerlaw + log-parabola model. The dashed lines shows the two components of the model for the LAT and PCA flaring state s.}
\label{fig_lpcop}
\end{figure}
\subsection{Broken powerlaw model}
\label{spec_bkpmodel}
Fig. \ref{fig_bkp} shows the three datasets with the best fit for a broken powerlaw model:
\begin{equation*}
\label{eq_bknpow}
N(E) = \begin{cases} N_{bp}~E^{-\Gamma_1} & \mbox{ at } E < E_b\\ N_{bp}~E_b^{\Gamma_2 - \Gamma_1} E^{-\Gamma_2} & \mbox{ at } E > E_b \end{cases}
\end{equation*}
The best fit parameters are reported in Table \ref{tab_bkp}. The fit by means of the broken powerlaw model does not provide a good description of the LAT and PCA flaring state s spectra, but the variation of the spectral indices between the two datasets gives some hints on the physics. This difference leads to a variation of the spectral break $\Delta \Gamma = \Gamma_2 - \Gamma_1$ from $\sim 0.7 \pm 0.1$ for the PCA flaring state\ to $\sim 1.0 \pm 0.05$ for the LAT flaring state, which could be related to a variation of the average slopes of the electron distribution (assuming the electron distribution modelled with a broken powerlaw too) from the LAT to the PCA flaring state. The break energy $E_b$ is larger for the LAT flaring state\ than for the PCA flaring state.
Both datasets show a bump at $\sim 20$ keV (in RXTE and JEMX data), clearly visible in the lower panel of Fig. \ref{fig_bkp}: to understand if this structure is real, we built the spectrum of the Crab nebula to compare the deviations from the best fit model for the two sources. As the deviations are of the same order, we conclude that this bump is likely due to a calibration issue.
\subsection{Log-parabola model}
\label{spec_lpmodel}
The log-parabola model is given by
\begin{equation}
\label{eq_logpar}
N(E) = N_{lp} \left( \frac{E}{E_{0}} \right) ^ {-\alpha - \beta \log \left( \frac{E}{E_{0}} \right)},
\end{equation}
\noindent where $N_{lp}$ is the photon flux at $E = E_{0}$, the $\alpha$ index denotes the slope at $E = E_{0}$ and the $\beta$ index models the curvature of the log-parabola.
Figure \ref{fig_lp} shows the three datasets with the best fit obtained with the log-parabola model. The best fit parameters are reported in Table \ref{tab_lp}. For the LAT flaring state\ the log-parabola alone is not able to fit the data adequately ($\chi_{red}^2 > 2$). The residuals indicate that the intrinsic curvature of the LAT data does not correspond to that of the complete SED.
The log-parabola model provides a good fit for the PCA flaring state, but it has to be pointed out that the LAT data in this case have worse statistics than for the LAT flaring state. Consisting of 3 points only, they do not show a clear intrinsic curvature. In this case the $\beta$ parameter can be determined only from the multiwavelength fit and it is impossible to state if an extra component is needed or not. A similar statement can be done for the CGRO dataset, where EGRET data do not have enough statistics to accurately constrain the shape of the spectrum at $\gamma$-ray energies.
\subsection{Cutoff powerlaw + log-parabola model}
\label{spec_lpcopmodel}
From the cross-correlation analysis (Sect. \ref{temp_cc}) we inferred that X-rays and $\gamma$ rays are dominated by different components. Interpreting the LAT data emitted by inverse-Compton processes in the jet, we tried to add an extra component in the form of a cutoff powerlaw at lower energies to model the X-ray emission in a similar way as assumed in Seyfert galaxies.
Figure \ref{fig_lpcop} shows the three datasets with the best fit for the cutoff powerlaw + log-parabola model. The best fit parameters are reported in Table \ref{tab_lpcop}. Adding the cutoff powerlaw component improves the quality of the fit for the LAT flaring state, supporting the hypothesis of an extra X-ray component. The parameters of the cutoff powerlaw component have been fixed to the typical values for Seyfert Galaxies \citep{Mushotzky1984,Turner1989} and we used the same values of spectral index and cutoff energy for the three datasets because fitting the datasets with six free parameters leads to a degeneracy, especially for the cutoff energy $E_c$. This model provides a good fit also for the PCA flaring state\ and for the CGRO dataset.\\
In order to constrain the best value for the cutoff powerlaw component we also tried to fit the X-ray data only (PCA, JEM-X, ISGRI and SPI) for the LAT flaring state\ with the cutoff powerlaw + log-parabola model. In this case, the parameters of the log-parabola component are fixed to the values that provide the best fit of LAT-Fermi data alone. The idea is to model the contribution of the blazar-like component fitting the $\gamma$-ray data alone and extrapolating its contribution at X-rays. Fitting the two components independently allows us to derive the best fit parameters for the Seyfert-like component, modelled as a cutoff powerlaw (Table \ref{tab_lpcop2}). It is not possible to follow the same procedure for the PCA flaring state\ due to lack of statistics of the LAT data.\\
We can also fit the whole multiwavelength spectrum using the spectral parameters of Table \ref{tab_lpcop2} for the LAT flaring state\ instead of the fixed parameters previously chosen (Table \ref{tab_lpcop}). In this way we obtain a new set of best-fit parameters and reduced $\chi^2$ for the cutoff powerlaw + log-parabola model (Table \ref{tab_lpcop3}). As expected the $\chi^2_{red}$ is better for the LAT flaring state, but it is worse for the PCA flaring state, suggesting that the cutoff powerlaw component may not have the same parameters in both datasets. The last line of Table \ref{tab_lpcop3} reports the best fit obtained leaving the cutoff powerlaw photon index $\Gamma$ free to vary.
In the PCA flaring state\ the variation of the spectral index $\Gamma$ plays an important role: Fig. \ref{fig_parcont} shows the dependency of the $\chi^2$ on the photon index $\Gamma$ and the cutoff energy $E_c$. A reasonable good fit in the PCA flaring state\ can be obtained only if $\Gamma \gtrsim 1.6$ . At $E_c = 360$ keV the minimum is located at $\Gamma = 1.66$ with $\chi^2_{red} = 1.02$, but the fit remains good for higher values of the photon index. The dependency on the cutoff energy is very weak: $\chi^2$ is almost constant for a large range of $E_c$ values. The data basically constrain only a lower limit to $E_c$.
\begin{figure}
\centering
\includegraphics[width=\hsize]{pca_lpcop_cont.jpeg}
\caption{Dependency of the $\chi^2$ on the photon index $\Gamma$ and cutoff energy $E_c$ of the cutoff powerlaw component for the PCA flaring state\ in the cutoff powerlaw + log-parabola model. The cross represents the best fit model. The lines from the internal to the external show confidence contour at $\sigma =$ 1, 2, 2.706 (90\%), 5.}
\label{fig_parcont}
\end{figure}
\begin{table*}
\caption{Best fit parameters, reduced $\chi^2$ and degree of freedom for the \textit{broken powerlaw} model. The first column identifies the dataset. See text for details.}
\label{tab_bkp}
\centering
\begin{tabular}{ccccccc}
\hline\hline
\noalign{\smallskip}
Spectrum & $\Gamma_1$ & $E_b (MeV)$ & $\Gamma_2$ & $N_{bp}$ & $\chi_{red}^2$ & $d.o.f.$ \\
\noalign{\smallskip}
\hline
\noalign{\smallskip}
LAT f.s. & $1.55 \pm 0.02$ & $(7.8_{-1.3}^{+1.6})$ & $2.61 \pm 0.04$ & $(2.64_{-0.02}^{+0.03}) 10^{-2}$ & 1.376 & 40 \\
\noalign{\smallskip}
\hline
\noalign{\smallskip}
PCA f.s. & $1.65 \pm 0.02$ & $(1.6_{-0.7}^{+1.0})$ & $2.41_{-0.08}^{+0.09}$ & $(3.80 \pm 0.02) 10^{-2}$ & 1.987 & 34 \\
\noalign{\smallskip}
\hline
\noalign{\smallskip}
CGRO & $1.62 \pm 0.01$ & $(2.2_{-0.9}^{+2.9})$ & $2.43_{-0.15}^{+0.17}$ & $(0.91 \pm 0.03) 10^{-2}$ & 0.718 & 16 \\
\noalign{\smallskip}
\hline
\end{tabular}
\end{table*}
\begin{table*}
\caption{Best fit parameters, reduced $\chi^2$ and degree of freedom for the \textit{log-parabola} model. The first column identifies the dataset. See text for details.}
\label{tab_lp}
\centering
\begin{tabular}{ccccccc}
\hline\hline
\noalign{\smallskip}
Spectrum & $E_{0}$ (keV) & $\alpha$ & $\beta$ & $N_{lp}$ & $\chi_{red}^2$ & $d.o.f.$ \\
\noalign{\smallskip}
\hline
\noalign{\smallskip}
LAT f.s. & 1 & $1.24 \pm 0.03$ & $0.107 \pm 0.004$ & $(1.58_{-0.15}^{+0.17}) 10^{-2}$ & 2.311 & 41 \\
\noalign{\smallskip}
\hline
\noalign{\smallskip}
PCA f.s. & 1 & $1.41 \pm 0.02$ & $(9.74_{-0.33}^{+0.35})10^{-2}$ & $(2.90_{-0.19}^{+0.20}) 10^{-2}$ & 1.060 & 35 \\
\noalign{\smallskip}
\hline
\noalign{\smallskip}
CGRO & 1 & $1.38 \pm 0.03$ & $(9.81_{-0.59}^{+0.66})10^{-2}$ & $(0.70 \pm 0.03) 10^{-2}$ & 1.290 & 17 \\
\noalign{\smallskip}
\hline
\end{tabular}
\end{table*}
\begin{table*}
\caption{Best fit parameters, reduced $\chi^2$ and degree of freedom for the \textit{cutoff powerlaw + log-parabola} model. The first column identifies the dataset. See text for details.}
\label{tab_lpcop}
\centering
\begin{tabular}{cccccccccc}
\hline\hline
\noalign{\smallskip}
Spectrum & $\Gamma$ & $E_c$ (keV) & $N_{cp}$ & $E_{0}$ (keV) & $\alpha$ & $\beta$ & $N_{lp}$ & $\chi_{red}^2$ & $d.o.f.$ \\
\noalign{\smallskip}
\hline
\noalign{\smallskip}
LAT f.s. & 1.7 & 150 & 0.028 & 1 & $0.75_{-0.08}^{+0.06}$ & $0.16 \pm 0.01$ & $(2.0 \pm 0.4) 10^{-3}$ & 0.971 & 41 \\
\noalign{\smallskip}
\hline
\noalign{\smallskip}
PCA f.s. & 1.7 & 150 & 0.028 & 1 & $1.10_{-0.07}^{+0.05}$ & $0.13 \pm 0.01$ & $(6.3 \pm 1.0) 10^{-3}$ & 1.266 & 35 \\
\noalign{\smallskip}
\hline
\noalign{\smallskip}
CGRO & 1.7 & 150 & 0.009 & 1 & $0.61_{-0.21}^{+0.20}$ & $0.19 \pm 0.03$ & $(0.34_{-0.18}^{+0.38}) 10^{-3}$ & 0.098 & 16 \\
\noalign{\smallskip}
\hline
\end{tabular}
\end{table*}
\begin{table*}
\caption{Best fit parameters, reduced $\chi^2$ and degree of freedom for the \textit{cutoff powerlaw + log-parabola} model at LAT flaring time. Here the fit is done using PCA, JEM-X, ISGRI and SPI data. Parameters of the log-parabola are obtained fitting LAT data alone and used here as frozen parameters. See text for details.}
\label{tab_lpcop2}
\centering
\begin{tabular}{cccccccccc}
\hline\hline
\noalign{\smallskip}
$\Gamma$ & $E_c (keV)$ & $N_{cp}$ & $E_{0}$ (MeV) & $\alpha$ & $\beta$ & $N_{lp}$ & $\chi_{red}^2$ & $d.o.f.$ \\
\noalign{\smallskip}
\hline
\noalign{\smallskip}
$1.55 \pm 0.04$ & $(3.7_{-1.1}^{+3.0}) 10^2$ & $(2.52_{-0.26}^{+0.29}) 10^{-2}$ & 100 & 2.33 & 0.19 & $2.63~10^{-11}$ & 0.934 & 35 \\
\noalign{\smallskip}
\hline
\end{tabular}
\end{table*}
\begin{table*}
\caption{Best fit parameters, reduced $\chi^2$ and degree of freedom for the \textit{cutoff powerlaw + log-parabola} model. using for the cutoff powerlaw the best fit parameters of Table \ref{tab_lpcop2}. Third line reports the best fit for the PCA flaring state\ if $\Gamma$ is not frozen. The first column identifies the dataset. See text for details.}
\label{tab_lpcop3}
\centering
\begin{tabular}{cccccccccc}
\hline\hline
\noalign{\smallskip}
Spectrum & $\Gamma$ & $E_c$ (keV) & $N_{cp}$ & $E_{0}$ (keV) & $\alpha$ & $\beta$ & $N_{lp}$ & $\chi_{red}^2$ & $d.o.f.$ \\
\noalign{\smallskip}
\hline
\noalign{\smallskip}
LAT f.s. & 1.55 & 370 & 0.0252 & 1 & $0.39_{-0.08}^{+0.11}$ & $0.19 \pm 0.01$ & $(1.6_{-0.4}^{+0.9}) 10^{-4}$ & 0.859 & 41 \\
\noalign{\smallskip}
\hline
\noalign{\smallskip}
PCA f.s. & 1.55 & 370 & 0.0252 & 1 & $1.34_{-0.08}^{+0.20}$ & $0.08_{-0.03}^{+0.01}$ & $(3.5_{-1.4}^{+2.9}) 10^{-3}$ & 2.057 & 35 \\
\noalign{\smallskip}
\hline
\noalign{\smallskip}
PCA f.s. & $1.66_{-0.05}^{+0.09}$ & 370 & 0.0252 & 1 & $1.27_{-0.06}^{+0.08}$ & $0.11 \pm 0.01$ & $(8.0_{-1.2}^{+1.5}) 10^{-3}$ & 1.021 & 34 \\
\noalign{\smallskip}
\hline
\end{tabular}
\end{table*}
\section{Discussion}
\label{disc}
The analysis performed on the lightcurves and energy spectra of 3C 273 suggests different origin for the X-ray emission with respect to the $\gamma$-ray emission. The $\gamma$-ray emission at $E \gtrsim 100$ MeV is compatible with the relativistically beamed SSC coming from the jet. This blazar-like origin is suggested by the cross-correlation between radio and $\gamma$ rays. The X-ray component is compatible with thermal photon emitted by the accretion disk and scattered in the X-ray domain by inverse-Compton processes, possibly on a hot corona close to the central black hole, as inferred for Seyfert galaxies.
At a first attempt, the Seyfert-like component has been modelled with the identical cutoff-powerlaw component in the LAT and PCA flaring state s, using the same parameters $\Gamma$, $E_c$ and $N_{cp}$ (Table \ref{tab_lpcop}). This means that we are able to explain the X-ray spectral variability with the variability of the blazar-like component only. However it is unlikely that the Seyfert-like component is not varying. In fact we can find slightly better fits using a different spectral index $\Gamma$ in the cutoff component (Table \ref{tab_lpcop3}). Furthermore, the CGRO dataset\ clearly shows that at the time of the multiwavelength campaign described in \citet{Lichti1995} the source was globally in a much lower state.
We note that \citet{Beaklini2014} reported a cross-correlation between 3C 273 Fermi-LAT data and radio data at 7 mm wavelength (43 GHz) claiming a delay between 120 and 170 days fully consistent with our estimate of $\sim$ 140 d (Sect. \ref{temp_conv}), and with similar conclusions on the link between radio and $\gamma$-rays. They explain the long-term variability of 3C 273 by variation of the Doppler factor in a 16-year jet-precessing model. The LAT flaring state\ lasts roughly $\sim 100$ days, and according to the interpretation of \citet{Beaklini2014} it is at the time when the Doppler factor is highest. The PCA flaring state\ spans a longer and not continuous time period: The simultaneous LAT data may thus be a composition of photons emitted with a noticeably different Doppler factor. Our data are thus broadly consistent with the precessing jet interpretation of these authors, although we cannot really test it with the chosen two different states.
\subsection{The convolution and the IC response}
\label{disc_conv}
Radio emission in radio loud AGN as 3C 273 is known to originate through synchrotron mechanism due to shocks in a relativistic jet \citep{Begelman1984,Marscher1985}. The response function used in the convolution (Sect. \ref{temp_conv}) hides the processes that happen in the propagation of the shock, like the cooling of the electrons and the IC mechanism. In the inner part of the jet the electrons are energetic enough to produce a flare in $\gamma$ rays through IC processes. Further out the flare is produced at longer wavelength because of the electron cooling and the expansion of the emitting region.
The relative success of our convolution attempt suggests that the radio outbursts are closely related to the $\gamma$-ray activity. The effect of the convolution is to smear out the complex variability structure of the Fermi lightcurve, resulting in only two apparently distinct radio outbursts. Our convolution thus implies that the radio outbursts would actually be a blend of emission from electron accelerated in several shorter shocks.
It has to be pointed out however that longer flares in radio band may be the result of the superimposition of shorter flares: \citet{Valtaoja1999} decomposed radio lightcurves (22 and 37 GHz band) of an AGN sample into exponential flares. \citet{Hovatta2009} used this technique to derive the variability of Doppler boosting and Lorentz factor of the jets from the time scales of the flares. These works show that it is possible to decompose a long radio flare into several sequential components. VLBI data would allow us to identify the different radio components and then it is possible to look for direct correlation with $\gamma$-ray data. \citet{Jorstad2001} used this method to look for correlation between radio and $\gamma$-ray properties in a sample of blazars.
When building the synthetic radio lightcurve we had to vary the amplitude of the IC response from flare to flare (Table \ref{tab_pconv}). There are several physical parameters which can affect the IC process and may vary from shock to shock producing the observed variation in the amplitude of the IC response, but probably the most important are the size of the shocked region and the injected electron distribution. The bigger is the size of the shock or the more energetic are the injected electrons, the larger is the number of scattered photons, and hence the more prominent is the flare in $\gamma$ rays. Different knots of the jet can be ejected at slightly different angles with respect to the observer's line of sight or with a slightly different Lorentz bulk factor, resulting in a different Doppler factor \citep{Jorstad2001}. The emission of the knots with higher Doppler factor is hence boosted and shifted to higher frequencies, increasing consequently the total $\gamma$-ray emission. Both these considerations can explain the observed variation of the IC response amplitude.
\subsection{Interpretation of the log-parabola model}
\label{disc_logpar}
The energy spectrum of the LAT and PCA flaring state s can be fitted with different parameter values. At first we note that the spectral index $\Gamma$ of the cutoff powerlaw and the spectral index $\alpha$ of the log-parabola are both harder in the LAT flaring state. The best fit of the LAT flaring state\ is obtained with $\Gamma = 1.55$ and $\alpha = 0.39$ while the best fit of the PCA flaring state\ is with $\Gamma = 1.66$ and $\alpha = 1.27$ (first and third lines of Table \ref{tab_lpcop3}, respectively).
In the PCA flaring state, the log-parabola model, because of the low statistics of the Fermi-LAT data, is sufficient to fit the dataset. Using the cutoff powerlaw + log-parabola model a degeneracy is added and the parameters can vary in a quite wide range.
The best fit yields a slightly steeper photon index $\Gamma$ of the cutoff component together with a lower photon index $\alpha$ of the log-parabola and a slightly increased curvature $\beta$. This means that with a softer cutoff powerlaw component the log-parabola index becomes harder in order to adjust the global slope of the spectrum (Table \ref{tab_lpcop3}). However exploring the range of the possible spectral parameter in the PCA flaring state\ the $\beta$ parameter is always smaller and the $\alpha$ parameter is always larger with respect to the LAT flaring state, meaning that the spectral variation between the two states is real.
It suggests a harder spectrum with stronger curvature in the LAT flaring state.
In order to find a possible physical interpretation of this difference we note that it is possible to show that a log-parabolic distribution of particles $n(\gamma)$ generates a log-parabolic photon spectrum using the $\delta$-assumption for the emission of a single particle. In this case the relation between the spectral indices is $\alpha = (s - 1)/2$ and $\beta = r/4$, where $s$ and $r$ are respectively the spectral index and the curvature in the log-parabolic particle distribution $n(\gamma)$.
Interpreting the log-parabola spectrum as the result of stochastic acceleration of the emitting electrons in a simple statistical approach, \citet{Tramacere2011} showed that $s \propto - \log ( \overline{\varepsilon} ) \left( \sigma_\varepsilon / \overline{\varepsilon} \right)^{-2}$ (it can be obtained from their Eq. 9) and $r \propto \left(n \left(\sigma_\varepsilon / \overline{\varepsilon} \right)^2 \right)^{-1}$, where $n$ is the mean number of acceleration steps, and $\overline{\varepsilon}$ and $\sigma_\varepsilon^2$ are respectively the mean and the variance of the energy gain in each acceleration step. The spectral parameters are therefore related to the acceleration mechanism through the coefficient of variation $\sigma_\varepsilon / \overline{\varepsilon}$. The larger is $\sigma_\varepsilon / \overline{\varepsilon}$, the softer and less curved is the spectrum.
Our data show that the log-parabola component of the LAT flaring state\ is harder and more curved than in the PCA flaring state, and also that the peak energy of the SED is higher in the LAT flaring state. In the stochastic acceleration interpretation, this means that the acceleration mechanism during the $\gamma$-ray flares is characterized by a lower coefficient of variation. This could result, for istance, because particles are uniformly boosted to higher energies through shocks.
\subsection{Synchrotron Self Compton variability}
\label{disc_ssc}
The SSC mechanism predicts constrains on the spectral variability which can be used to test the compatibility between the data and the model \citep[i.e. see][and reference therein for a detailed discussion on SSC]{Jones1974,Bloom1996}. \citet{Bloom1996} explored the consequences of the variations of the physical parameters of the emitting source (i.e. the magnetic field, the particle density distribution, the viewing angle) on the resulting spectrum in the frame of the SSC model.
We try to derive some informations on the physics of the X-rays and $\gamma$-rays emitting region in 3C 273 based on the differences between the LAT and PCA flaring state s assuming that the blazar-like emission modelled as a log-parabola is due entirely to the SSC mechanism, without any External Compton contribute.
In order to do this we introduce first Eqs. \ref{eq_lpepeak} and \ref{eq_lpmax} which give the value of the log-parabola peak energy $E_p$ and the SED flux at the peak $E_p$ \citep{Massaro2004}:
\begin{equation}
\label{eq_lpepeak}
E_{p} = E_{0} ~ 10^{(2 - \alpha)/2\beta}
\end{equation}
\begin{equation}
\label{eq_lpmax}
\log \left(E_p F_{E}(E_p) \right) = \log N_{lp} + \frac{(2 - \alpha)^2}{4 \beta}.
\end{equation}
The peak energy can be used to estimate the variation of the Lorentz factor $\gamma$ of the emitting electrons. In the SSC the radiation field is produced by synchrotron emission, so that the energy released by a single electron is peaking at a frequency of $\displaystyle \nu_c \propto \gamma^2 \frac{eB}{m_e c}$. This radiation field then interacts with the electrons themself, and photons are scattered to higher energies through IC scattering, increasing their frequency by another factor $\gamma^2$.
The ratio between the energy peaks of LAT and PCA flaring state s can provide information about the variation of $\gamma$ in the two spectral states. This ratio can be written as
\begin{equation}
\label{eq_epeakrat}
\frac{E_L^C}{E_P^C} \sim \frac{\gamma_L^2 E_L^S}{\gamma_P^2 E_P^S} \sim \left( \frac{\gamma_L}{\gamma_P} \right)^4 \frac{B_L}{B_P},
\end{equation}
\noindent where the subscripts L and P stand for LAT and PCA flaring state s, respectively, and the superscripts C and S stand for IC and synchrotron. Taking the energy peak of the log-parabola as $E^C$ and hence using Eq. \ref{eq_lpepeak} and the best fit parameters of the spectral states in Table \ref{tab_lpcop} we get $\gamma_L / \gamma_P \simeq 1.6$ if we assume $E_L^S= E_P^S$ or $\gamma_L / \gamma_P \simeq 1.2$ if we assume $B_L = B_P$. Particles would thus be accelerated to higher energies during the LAT flaring state.\\
To further investigate the physical parameters of the emitting region we can use the flux at the peak. We first rewrite the peak energy of the IC spectra in unit of $\displaystyle \epsilon = \frac{h\nu}{m_e c^2}$ as
\begin{equation}
\label{eq_epeak}
\epsilon_p = \frac{\delta_D}{1 + \mathrm{z}}\gamma^4 \varepsilon_B
\end{equation}
\noindent where $\displaystyle \varepsilon_B = \frac{\hbar eB}{m_e^2 c^3}$ and $\delta_D$ is the Doppler factor. The $\gamma$ factor here is that of the electrons emitting at the peak energy.
This photon energy is used in Eq. \ref{eq_flssc}, which gives the observed SSC $\nu F_\nu$ spectrum \citep[Ch. 7, eq. 7.97]{Dermerbook}
\begin{equation}
\label{eq_flssc}
\begin{array}{rcl}
\displaystyle E F_{E}(\epsilon) & = & \displaystyle \delta_D^4 \frac{c\sigma_T^2 R' U_B V'}{12 \pi d_L^2} \left( \frac{\epsilon}{\varepsilon_B} \right)^{3/2} \times\\
& & \displaystyle \int_0^{\min(\epsilon', 1/\epsilon')} d\epsilon_i' \epsilon_i'^{-1} n_e' \left( \sqrt{\frac{\epsilon'}{\epsilon_i'}} \right) n_e' \left( \sqrt{\frac{\epsilon_i'}{\varepsilon_B}} \right).
\end{array}
\end{equation}
Equation \ref{eq_flssc} is obtained in the $\delta$-function approximation and limiting the IC interation to the Thompson regime (upper limit of the integral). Prime quantities are measured in the rest frame of the source. $R'$ is the size of the source ($V' \sim R'^3$ being the volume), $\displaystyle U_B = \frac{B^2}{8\pi}$ is the energy density of the magnetic field. The electron distribution $n_e(\gamma)$ appears twice, the former to account for the IC scattering, the latter for the synchrotron emission. $\epsilon_i'$ is the energy of the synchrotron photons which are scattered.
Equations \ref{eq_epeak} and \ref{eq_flssc} relate the peak energy $E_p$ and the $\nu F_\nu$ spectrum to the physical quantities of the source. Varying the physical parameters we can explore how the peak energy and the $\nu F_\nu$ flux calculated at the peak energy vary thereby and compare these trends with the data.
According to Eq. \ref{eq_epeak} the observed peak energy depends on the Lorentz factor $\gamma$ of the electrons at the peak and the magnetic field $B$ (we do not consider here any variation of the Doppler factor $\delta_D$ of the emitting region). The value of the observed $E_p$ in LAT and PCA flaring state s is reproduced by fixing arbitrary one of the two parameters and setting properly the remaining one. Then we can compute the SSC $\nu F_\nu$ flux at the peak energy with Eq. \ref{eq_flssc}, provided that the remaining parameters are fixed.
To solve Eq. \ref{eq_flssc} we assume a log-parabolic electron distribution, for which we can derive the parameters $s$ and $r$ from the spectral fit as mentioned in Sect. \ref{disc_logpar}, and thus solve the integral numerically. We used the parameters of the best fit from Table \ref{tab_lpcop} for the LAT and PCA flaring state s to derive the parameters of the electron distribution.
Figure \ref{fig_varypar} shows these points together with the observations of the LAT and PCA flaring state s, Table \ref{tab_varypar} list the exact energy $E_p$ and associated SED flux, together with the corresponding values of the magnetic field $B$ and Lorentz factor $\gamma$.\\
\begin{figure}
\centering
\includegraphics[angle=-90, width=\hsize]{specerr3_nr-eps-converted-to.pdf}
\caption{Trend of variation of the SSC flux in the $E_p - E_p F_E (E_p)$ plane. The black points correspond to the LAT and PCA flaring state s. The red line connects the two points with different Lorentz factor $\gamma$, the blue line connects the two points with different magnetic field $B$. The parameters are listed in Table \ref{tab_varypar}. The dotted green line shows the trend associated to $\delta_D$ variation. See text for details.}
\label{fig_varypar}
\end{figure}
\begin{table}
\caption{List of the varying parameters $\gamma$ and $B$ used to draw the points in Fig. \ref{fig_varypar}. The remaining parameters are fixed to the following values: $\delta_D = 9$, $R' = 1.6 \; 10^{15}$ cm, $\displaystyle n_e' = 2.7 \; 10^{6} \; \mathrm{electrons}/\mathrm{cm}^3$, z = 0.158, $d_L = 2.30 \; 10^{27}$ cm, estimated for the redshift of 3C 273 in a flat universe ($\Omega_\Lambda = 1 - \Omega_M$) with $H_0$ = 71 km/s Mpc and $\Omega_M = 0.27$.}
\label{tab_varypar}
\centering
\begin{tabular}{cccc}
\hline\hline
\noalign{\smallskip}
$B$ (Gauss) & $\gamma$ & $E_p$ (keV) & $E_p F_E(E_p) \left( \frac{\mathrm{erg}}{\mathrm{cm}^2 s} \right)$ \\
\noalign{\smallskip}
\hline
\noalign{\smallskip}
1.4 & $2.8 \; 10^3$ & $8.0 \; 10^3$ & $2.4_{-1.7}^{+5.1} \; 10^{-9}$ \\
\noalign{\smallskip}
\hline
\noalign{\smallskip}
1.4 & $2.2 \; 10^3$ & $2.9 \; 10^3$ & $1.5_{-0.9}^{+2.3} \; 10^{-10}$ \\
\noalign{\smallskip}
\hline
\noalign{\smallskip}
2.3 & $2.5 \; 10^3$ & $8.0 \; 10^3$ & $7.9_{-6.8}^{+48.0} \; 10^{-9}$ \\
\noalign{\smallskip}
\hline
\noalign{\smallskip}
0.8 & $2.5 \; 10^3$ & $2.9 \; 10^3$ & $3.1_{-2.3}^{+20.0} \; 10^{-11}$ \\
\noalign{\smallskip}
\hline
\end{tabular}
\end{table}
We choose to investigate variation in $B$ and $\gamma$ because in our simple description, $E_p$ depends on these two physical parameters only. $\delta_D$ and $R'$ have been fixed to $\delta_D = 9$ \citep{Jorstad2005} and $R' = 1.6 \; 10^{15}$ cm \citep{Rani2013}, the remaining parameter $n_e'$ has been calculated to reproduce the observed flux. The total electron density $n_e'$ is the same in both states, i.e. the integral of the log-parabolic distribution $n_e'(\gamma)$ is forced to be constant adjusting the normalization constant. The electron density value needed to reproduce the observed flux is $n_e' = 2.7 \; 10^{6} \; \mathrm{electrons}/\mathrm{cm}^3$. This looks like a big value, but is low enough for an optically thin emitting region: the optical depth $\tau$ for inverse-Compton scattering in Thompson regime can be estimated to be $\tau \sim R' \sigma_T n_e' \sim 10^{-3}$.
Figure \ref{fig_varypar} shows that the observed variability between the LAT and PCA flaring state s is better described by a variable Lorentz factor $\gamma$ with a constant magnetic field B, rather than a constant $\gamma$ and a variable B. The confidence intervals are obtained accounting for the errors on the spectral parameters: the parameters are modified by their errors at $1 \sigma$ and then the flux is recalculated with Eq. \ref{eq_flssc}. This indicates that the variation of the Lorentz factor $\gamma$ of the electrons is likely responsible for the difference between the LAT and the PCA flaring state s. We note that variations of $R'$ or $n_e'$ do not produce a variation in $E_p$.
We investigated also if the variation could be explained by the variation of the Doppler factor $\delta_D$ of the emitting region. Different knots may have different Doppler factors because they may be emitted at a different angle or with different velocity, as assumed in the varying $\delta_D$-factor model of \citet{Beaklini2014}. To test this, we used the same electron distribution $n_e'(\gamma)$ for the LAT and PCA flaring state s assuming that the only varying parameter between the two states is $\delta_D$. In this case $E_p$ increases linearly with $\delta_D$, while $E_p F_E(E_p)$ increases as $\delta_D^4$ (see Eq. \ref{eq_flssc}, the trend shown by the dotted green line in Fig. \ref{fig_varypar}). This trend is not compatible with the data.
In conclusion, a variation of the electron distribution and consequently a variation of the electron Lorentz factor $\gamma$ corresponding to the peak of the electron energy distribution seems to be the most likely driver for the difference between the LAT and PCA flaring state s.
\subsection{Location of the $\gamma$-ray emitting region}
\label{disc_locgamma}
The spectrum observed by Fermi-LAT is not significantly absorbed up to $\sim 3$ GeV. Photons below 3 GeV could produce pairs by interacting with target photons of energies above ~ 0.3 keV. The latter are probably emitted from the accretion disk and corona within 100 gravitational radii from the central black hole. The $\gamma$-ray emitting region should be located outside of that region, which is not very constraining.
Since the sensitivity of Cherenkov telescopes is not sufficient to detect 3C 273 in the TeV domain, where the pair absorption signature should be present if the $\gamma$-ray emitting region is inside the broad line region, the position of the $\gamma$-ray source cannot be constrained with respect to the size of the broad line region \citep{Donea2003}.
\section{Summary and Conclusions}
\label{concl}
The high energy spectral variability of 3C 273 has been studied from $\sim 1$ keV up to $\sim 10$ GeV and compared with radio data. We presented Fermi-LAT, RXTE-PCA and Radio (37 GHz) lightcurves and multiwavelength spectra of RXTE-PCA, Fermi-LAT and JEMX, ISGRI, SPI on board INTEGRAL, built for two different spectral states, namely the LAT flaring state, i.e. when the source is flaring in Fermi-LAT, and the PCA flaring state, i.e. when the source is flaring in RXTE-PCA.
Timing analysis (Sect. \ref{temp}) indicates the presence of two different components in the considered energy band. Gamma ray data correlate with radio data, but not with X-ray data, confirming that the emissions in $\gamma$ rays and radio have a common origin, while the origin of the X-ray emission is different. The convolution between the $\gamma$-ray lightcurve and the SSC response function reproduces the radio signal (Sect. \ref{temp_conv}). This indicates that the data are compatible with the synchrotron mechanism in the frame of a shock-in-jet scenario \citep{Marscher1985,Turler1999,Turler2000}. The necessity to shrink the response function obtained by low energy data suggests that what is seen as a single flare in the radio lightcurve may be in this case the blended emission from several shorter shocks, which have been detected separately in the $\gamma$-ray lightcurve.
Spectral fitting of the high energy data (Sect. \ref{spec}) shows that a single model emission as a cutoff powerlaw or a log-parabola cannot fit adequately the whole SED from $\sim 1$ keV up to $\sim 10$ GeV. Knowing from timing analysis that X-rays and $\gamma$ rays have different origin, we used a two components spectral model: a cutoff powerlaw at low energies and a log-parabola at high energies, the former modelling the Seyfert-like emission, the latter modelling the jet emission. We found that for the LAT flaring state\ the powerlaw and log-parabola components are harder than for the PCA flaring state, and the LAT flaring state\ spectrum has a stronger curvature at the SED peak. The spectral variability between the two states can be understood if it is driven by a change in the electron distribution with a peak energy increasing by $\sim 30 \%$ during the LAT flaring state.\\
Unlike typical blazars, the high energy spectrum of 3C 273 is peaked in the MeV energy range, below the Fermi-LAT lower energy threshold. Even during the flare, the peak of the emission remains undetected. As there are no experiments sensitive enough in the MeV, it is difficult to study in detail the spectral properties of the $\gamma$-ray emission of 3C 273. Nevertheless, the shape of the $\gamma$-ray spectrum of 3C 273 is very similar to the shape of the GeV peaked blazar detected by Fermi, which are usually fitted with the log-parabola model, suggesting that the electron distribution looks universal in very different conditions.
\begin{acknowledgements}
This work is based on observation with INTEGRAL, an ESA project with instrument and science data center funded by ESA member states (especially the PI countries: Denmark, France, Germany, Italy, Switzerland, Spain), Czech Republic and Poland, and with the participation of Russia and the USA. The SPI project has been completed under the responsibility and leadership of CNES/France. The Mets\"ahovi\ team acknowledges the support from the Academy of Finland to our observing projects (numbers 212656, 210338, 121148, and others).
\end{acknowledgements}
|
\section{Distances and divergences for quantifying domain shift}
\label{app:distance}
Besides analyzing the performance drop when evaluating a model using source statistics on a target dataset, we consider the mismatch in model statistics directly.
We first take an ImageNet trained model and adapt it to each of the 95 conditions in IN-C.
To obtain a more exact estimate of the true statistics, we split the model into multiple stages with only few BN layers per stage and apply the following simple algorithm\footnote{%
Note that for simplicity, we do not reset the statistics of the remaining $(b_i - i)$ BN layers. This could potentially be adapted in future work.
}:
\begin{itemize}
\item Start with image inputs $\mathbf{z}^0_n \leftarrow \mathbf{x}_n$ from the validation set to adapt to, for each $n\in[50000]$.
\item Split the model into multiple stages,
$h(\mathbf{x}) = (f_{m} \circ \cdots \circ f_1)(\mathbf{x})$,
where each module $f_i$ can potentially contain one or multiple BN layers.
We denote the number of BN layers in the $i$-th module as $b_i$.
\item For each stage $i \in [m]$, repeat $b_i$ times:
$\mathbf{z}^i_n \leftarrow f_i(\mathbf{z}^{i-1}_n)$ for each $n$, and update the BN statistics in module $f_i(\mathbf{z}^{i-1}_n)$.
\item Return $h$ with adapted statistics.
\end{itemize}
Using this scheme, we get source statistics $\mu_s$ and $\Sigma_s$ for each layer and $\mu_t$ and $\Sigma_t$ for each layer and corruption.
In total, we get 96 different collections of statistics across network layers (for IN and the 95 conditions in IN-C).
For simplicity, we will not further index the statistics.
Note that all covariance matrices considered here are diagonal, which is a further simplification.
We expect that our domain shift estimates could be improved by considering the full covariance matrices.
In the following, we will introduce three possible distances and divergences which can be applied between source and target statistics to quantify the effect of common corruptions induced covariate shift.
We consider the Wasserstein distance, a normalized version of the Wasserstein distance, and the Jeffrey divergence.
\subsection{The Wasserstein distance}
\label{def:wasserstein}
Given a baseline ResNet-50 model with source statistics $\mu_s, \Sigma_s$ on IN, the Wasserstein distance (cf. \citealp{villani2008optimal}) between the train and test distribution with statistics $\mu_t, \Sigma_t$ is given as
\begin{equation}
W_2 (p_s, p_t)^2 = \| \bm{\mu}_s - \bm{\mu}_t \|_2^2 + \text{tr} \left( \bm{\Sigma}_s + \bm{\Sigma}_t - 2 \left(\bm{\Sigma}_t^{1/2} \bm{\Sigma}_s \bm{\Sigma}_t^{1/2} \right)^{1/2}\right).
\end{equation}
\subsection{The source-normalized Wasserstein distance}
\label{def:wasserstein-norm}
When estimated for multiple layers across the network, the Wasserstein distance between source and target depends on the overall magnitude of the statistics.
Practically, this means the metric is dominated by features with large magnitude (e.g. in the first layer of a neural network, which receives larger inputs).
To mitigate this issue, we normalize both statistics with the source statistics and define the normalized Wasserstein distance as
\begin{align}
\widetilde{W}_2^2 &= W_2^2\left(\bm{\Sigma}_s^{-1/2} \bm{\mu}_s, \mathbf{I}, \bm{\Sigma}_s^{-1/2} \bm{\mu}_t, \bm{\Sigma}_s^{-1} \bm{\Sigma}_t \right) \\
&= \Tr{\left(\mathbf{I} + \bm{\Sigma}_t \bm{\Sigma}_s^{-1} - 2 \bm{\Sigma}_t^{1/2} \bm{\Sigma}_s^{-1/2} \right)} + (\bm{\mu}_t - \bm{\mu}_s)^T \bm{\Sigma}_s^{-1} (\bm{\mu}_t - \bm{\mu}_s).
\end{align}
In the uni-variate case, the normalized Wasserstein distance $\widetilde{W}_2^2$ is equal to the Wasserstein distance $W_2^2$ between source and target statistics divided by $\sigma^2_s$:
\begin{align}
\widetilde{W}_2^2 &= W_2^2\left(
\frac{\mu_s}{\sigma_s},
1, \frac{\mu_t}{\sigma_s},
\frac{\sigma^2_t}{\sigma^2_s} \right) = 1 + \frac{\sigma^2_t}{\sigma^2_s} - 2 \frac{\sigma_t}{\sigma_s} + \frac{(\mu_t - \mu_s)^2}{\sigma^2_s}
= \frac{1}{\sigma^2_s} W_2^2(\mu_s, \sigma^2_s, \mu_t, \sigma^2_t).
\end{align}
\subsection{The Jeffrey divergence}
The Jeffrey divergence $J(p_s, p_t)$ between source distribution $p_s$ and target distribution $p_t$ is the symmetrized version of the Kullback-Leibler divergence $D_{KL}$:
\begin{align}
J(p_s, p_t) = \frac{1}{2} \left( D_{KL}(p_s \| p_t) + D_{KL}(p_t \| p_s) \right)
\end{align}
The Kullback-Leibler divergence between the $D$-dimensional multivariate normal source and target distributions is defined as
\begin{align}
D_{KL}(\mathcal{N}_t \| \mathcal{N}_s) &= \frac{1}{2} \left( \Tr{\left(\bm{\Sigma}_s^{-1} \bm{\Sigma}_t\right)} + (\bm{\mu}_s - \bm{\mu}_t)^\top \bm{\Sigma}_s^{-1} (\bm{\mu}_s - \bm{\mu}_t) - D + \ln{\left( \frac{\det \bm{\Sigma}_s}{\det \bm{\Sigma}_t}\right)} \right).
\end{align}
The Jeffrey divergence between the $D$-dimensional multivariate normal source and target distributions then follows as
\begin{align}
J(\mathcal{N}_t, \mathcal{N}_s) &= \frac{1}{4} \left( \Tr{\left(\bm{\Sigma}_s^{-1} \bm{\Sigma}_t\right)} + \Tr{\left(\bm{\Sigma}_t^{-1} \bm{\Sigma}_s\right)} + (\bm{\mu}_s - \bm{\mu}_t)^\top \left(\bm{\Sigma}_s^{-1} + \bm{\Sigma}_t^{-1}\right) (\bm{\mu}_s - \bm{\mu}_t) - 2D \right).
\end{align}
\subsection{Summary statistics and quantification of covariate shift between different IN-C conditions}
Given the 95 distances/divergences between the baseline (IN) statistics and 95 IN-C conditions, we first perform a layer-wise analysis of the statistics and depict the results in Figure~\ref{fig:wasserstein-layers}.
The unnormalized Wasserstein distance is sensitive to the magnitude of the source statistics and hence differs qualitatively from the results on the normalized Wasserstein distance and Jeffrey Divergence.
We appreciate that the most notable difference between source and target domains is visible in the ResNet-50 downsampling layers.
All three metrics suggest that the shift is mainly present in the first and final layers of the network, supporting the hypothesis that within the common corruption dataset, we have both superficial covariate shift which can be corrected by simple means (such as brightness or contrast variations) in the first layers, and also more ``high-level'' domain shifts which can only be corrected in the later layers of the network.
In Figure~\ref{fig:wasserstein-layers-fine}, we more closely analyze this relationship for different common corruptions.
We can generally appreciate the increased measures as the corruption severity increases.
\begin{figure}[hp!]
\begin{center}
\includegraphics[width=0.95\textwidth]{figs/ZZ-appendix/bars_Wasserstein_Distance.pdf}
\end{center}
\begin{center}
\includegraphics[width=0.95\textwidth]{figs/ZZ-appendix/bars_Wasserstein_Distance_normalized.pdf}
\end{center}
\begin{center}
\includegraphics[width=0.95\textwidth]{figs/ZZ-appendix/bars_Jeffrey_Divergence.pdf}
\end{center}
\caption{
Wasserstein distance,
normalized Wasserstein distance and Jeffrey divergence estimated among source and target statistics between different network layers.
We report the respective metric w.r.t. to the difference between baseline (IN) and target (IN-C) statistics and show the value averaged across all corruptions.
We note that for a ResNet-50 model, downsampling layers contribute most to the overall error.
}
\label{fig:wasserstein-layers}
\end{figure}
\begin{figure}[hp!]
\begin{center}
\includegraphics[width=0.95\textwidth]{figs/ZZ-appendix/grid_Wasserstein_Distance_normalized.pdf}
\end{center}
\begin{center}
\includegraphics[width=0.95\textwidth]{figs/ZZ-appendix/grid_Jeffrey_Divergence.pdf}
\end{center}
\caption{
Normalized Wasserstein distance and Jeffrey divergence across corruptions and layers in a ResNet-50.
\label{fig:wasserstein-layers-fine}
}
\end{figure}
\clearpage
\section{Notes on the experimental setup}
\label{app:experimental-setup}
\subsection{Practical considerations for implementing the method}
Our method is conceptually very easy to implement.
We generally recommend to first explore the easier variant of the algorithm where $N=0$, i.e., no source statistics are used.
As shown in our experiments, this setting works well if 100 or more target samples are available.
In this case, implementing the method boils down to enabling the training mode for all BN layers across the network.
We will discuss this option along with two variants important for application to practical problems:
Using exponential moving averaging (EMA) to collect target statistics across multiple batches, and using the source statistics as a prior.
\paragraph{Example implementation in PyTorch and caveats}
We encourage authors of robust models to always evaluate their models, and in particular baseline algorithms on both the train and test set statistics.
Implementation in both PyTorch, Tensorflow and other machine learning libraries is straightforward and adds only minimal overhead.
For PyTorch, adaptation is possible by simply adding
\begin{verbatim}
def use_test_statistics(module):
if isisinstance(module, nn._BatchNorm):
module.train()
model.eval()
model.apply(use_test_statistics)
\end{verbatim}
before starting a model evaluation.
For the adaptation to a full dataset, we provide a reference implementation with the source code release of this paper.
Also, in contrast to the convention of not shuffling examples during test time, \emph{make sure to enable dataset shuffling also during test time} in order to compute the correct statistics marginalized over class assignment.
\paragraph{Exponential moving averaging}
In practice, it might be beneficial to keep track of samples already encountered and use a running mean and variance on the test set to normalize new samples.
We can confirm that this technique closely matches the full-dataset adaptation case even when evaluating with batch size 1 and is well suited for settings with less powerful hardware, or in general settings where access to the full batch of samples is not possible.
Variants of this technique include the adaptation of the decay factor to discard statistics of samples encountered in the past (e.g. when the data domain slowly drifts over time).
\subsection{Notes on models}
Note that we only re-evaluate existing model checkpoints, and hence do not perform any hyperparameter tuning or adaptations to model training except for selecting the pseudo batchsize $N$ for the source domain.
Depending on the batch size and the architecture, model evaluations are done on one to eight Nvidia RTX 2080 GPUs (i.e., using 12 to 96 GB of memory) or up to four Nvidia V100 GPUs (128 GB of memory).
Since we merely re-evaluate trained models, it is also possible to work on less powerful hardware with less memory.
In these cases, the aggregation of batch normalization statistics has to be done across several batches using a variant of EMA.
\subsection{Hyperparameter tuning}
Our method is generally parameter-free if only target statistics should be considered for normalization.
This approach is generally preferred for larger batch sizes $n$ and should also be adapted in practice when a sufficient amount of samples is available.
For tuning $N$, we consider the pre-defined holdout corruptions in IN-C, including speckle noise, saturation, Gaussian blur and spatter using a grid search across different values for $N$.
\subsection{Notes on datasets}
In the main paper, we have used several datasets and provide more relevant information here:
\paragraph{ImageNet-C (IN-C)} For the evaluation on IN-C, we use the JPEG compressed images from \href{https://github.com/hendrycks/robustness}{github.com/hendrycks/robustness} as is advised by the authors to ensure reproducibility.
We note that \citet{ford2019adversarial} report a decrease in performance when the compressed JPEG files are used as opposed to applying the corruptions directly in memory without compression artefacts.
\paragraph{ObjectNet (ON)} We find that there are 9 classes with multiple possible mappings from ON to IN (see the list in Table~\ref{ON:ambiguous}); we discard these classes in our evaluation. Models trained on IN experience a large performance drop on the order of 40--45\% when tested on ON{}.
ON{} is an interesting test case for unsupervised domain adaptation since IN and ON{} are likely sampled from different distributions. ON intentionally shows objects from new viewpoints on new backgrounds.
\paragraph{ImageNet-V2 (IN-V2)} There are three test sets in IN-V2 that differ in \emph{selection frequencies} of the MTurk workers.
The selection frequency is given by the fraction of MTurk workers who selected an image for its target class.
For the ``MatchedFrequency'' dataset, images were sampled according to the estimated selection frequency of sampling of the original IN validation dataset.
For the ``Threshold0.7'' variant of IN-V2, images were sampled with a selection frequency of at least 0.7.
The ``TopImages'' was sampled from images with the highest selection frequency.
Although all three test sets were sampled from the same Flickr candidate pool and were labeled correctly and selected by more than 70\% of MTurk workers, the model accuracies on these datasets vary by 14\%.
The authors observe a systematic accuracy drop when comparing model performance on the original IN validation set and IN-V2 and attribute it to the distribution gap between their datasets and the original IN dataset.
They quantify the distribution gap by how much the change from the original distribution to the new distribution affects the considered model.
Engstrom et al. analyze the creation process of IN-V2 and identify statistical bias resulting from noisy readings of the selection frequency statistic as a main source of dropping performance \citep{engstrom2020identifying}. After correcting the bias, \citep{engstrom2020identifying} find that the accuracy drop between IN and IN-V2 measures only 3.6\% ± 1.5\% of the original 11.7\% ± 1.0\%.
\begin{table}[ht]
\centering
\small
\begin{tabular}{l l }
\toprule
ON class & IN classes \\
\midrule
wheel & wheel; paddlewheel, paddle wheel \\
helmet & football helmet; crash helmet \\
chair & barber chair; folding chair; rocking chair, rocker \\
still\_camera & Polaroid camera, Polaroid Land camera; reflex camera \\
alarm\_clock & analog clock; digital clock \\
tie & bow tie, bow-tie, bowtie; Windsor tie \\
pen & ballpoint, ballpoint pen, ballpen, Biro; quill, quill pen; fountain pen \\
bicycle & mountain bike, all-terrain bike, off-roader; bicycle-built-for-two, tandem bicycle, tandem \\
skirt & hoopskirt, crinoline; miniskirt, mini; overskirt \\
\bottomrule
\end{tabular}
\caption{Mapping between 9 ambiguous ON classes and the possible correspondences in IN. Different IN classes are separated with a semicolon.
}
\label{ON:ambiguous}
\end{table}
\subsection{Overview of models in torchvision}
In Table~\ref{tab:model_count}, we provide a list of the models we evaluate in the main paper, along with numbers of trainable parameters and BN parameters. Note that the fraction of BN parameters is at most at 1\% compared to all trainable parameters in all considered models.
\begin{table}[htp!]
\small
\begin{center}
\begin{tabular}{lrrr}%
\toprule
Model & Parameter Count & BN Parameters & Fraction (\%)\\
\midrule
\csvreader[head to column names]{data/parameters.csv}{
{\model
& \num[round-mode=places,
round-precision=2,
scientific-notation=true]{\params}
& \num[round-mode=places,
round-precision=2,
scientific-notation=true]{\bnparams}
& \num[round-mode=places,
round-precision=3]{\fraction} \\}\\
\bottomrule
\end{tabular}
\end{center}
\caption{%
Overview of different models with parameter counts.
We show the total number of BN parameters,
which is a sum of affine parameters.
}
\label{tab:model_count}
\end{table}
\subsection{Baseline corruption errors}\label{sec:appendix-mce}
In Table~\ref{tab:alexnet}, we report the scores used for converting top-1 error into the mean corruption error (mCE) metric proposed by \citet{hendrycks2018benchmarking}.
\input{figs/ZZ-appendix/alexnet}
\subsection{Software stack}
We use various open source software packages for our experiments, most notably Docker \citep{10.5555/2600239.2600241},
scipy and numpy \citep{2020SciPy-NMeth},
GNU parallel \citep{Tange2011a},
Tensorflow \citep{abadi2016tensorflow},
PyTorch \citep{paszke2017automatic} and
torchvision \citep{10.1145/1873951.1874254}.
\clearpage
\section{Additional results}
\label{app:results}
\subsection{Performance of SimCLRv2 models}
We evaluate the performance of 3 models from the SimCLRv2 framework with and without batchnorm adaptation. We test a ResNet50, a ResNet101 and a ResNet152, finetuned on 100\% of IN training data. Since our code-base is in PyTorch, we use the Pytorch-SimCLR-Converter \citep{convert-pytorch-simclr} to convert the provided checkpoints from Tensorflow to PyTorch. We notice a slight decline in performance when comparing the top-1 accuracy on the IN validation set, see Table~\ref{app:sim-clr-v2-IN}. For preprocessing, we disable the usual PyTorch normalization and use the PIL.Image.BICUBIC interpolation for resizing because this interpolation is used in the TensorFlow code (instead of the default PIL.Image.BILINEAR in PyTorch).
The BN adaptation results for the converted models are shown in Table~\ref{app:sim-clr-v2}. Adaptation improves the performance of the ResNet50 and the ResNet101 model, but hurts the performance of the ResNet152 model.
\input{figs/ZZ-appendix/sim_clr_v2_results_table.tex}
\subsection{Relationship between parameter count and IN-C improvements}
In addition to Fig.~3 in the main paper, we show the relationship between parameter count and IN-C mCE. In general, we see that the parameter counts correlates with corruption robustness since larger models have smaller mCE values.
\input{figs/YY-supplement/torchvision-param-vs-mce.tex}
\subsection{Per-corruption results on IN-C}
We provide more detailed results on the individual corruptions of IN-C for the most important models considered in our study in Fig.~\ref{fig:resnet_augmix_individual}.
The results are shown for models where the BN parameters are adapted on the full test sets.
The adaptation consistently improves the error rates on all corruptions for both vanilla and AugMix.
\input{figs/ZZ-appendix/individual-corruptions.tex}
\subsection{Qualitative analysis of similarities between common corruptions}
In this analysis, we compute a t-SNE embedding of the Wasserstein distances between the adapted models and the non-adapted model from Section 5, Fig. 4(i) of the main paper.
The results are displayed in Fig.~\ref{fig:wasserstein-tnse}.
We observe that the different corruption categories indicated by the different colors are grouped together except for the 'digital' category (pink).
This visualization shows that corruption categories mostly induce similar shifts in the BN parameters. This might be an explanation why training a model on Gaussian noise generalizes so well to other noise types as has been observed by \citet{Rusak2020IncreasingTR}: By training on Gaussian noise, the BN statistics are adapted to the Gaussian noise corruption and from Fig.~\ref{fig:wasserstein-tnse}, we observe that these statistics are similar to the BN statistics of other noises.
\input{figs/YY-supplement/wasserstein-tsne.tex}
\subsection{Error prediction based on the Wasserstein distance}
\label{app:error_prediction}
In Section 5, Fig.~4(i), we observe that the relationship between the Wasserstein distance and the top-1 error on IN-C is strikingly linear in the considered range of the Wasserstein distance.
Similar corruptions and corruption types (indicated by color) exhibit similar slope, allowing to approximate the expected top-1 error rate without any information about the test domain itself.
Using the split of the 19 corruptions into 15 test and 4 holdout corruptions \citep{hendrycks2018benchmarking}, we compute a linear regression model on the five data points we get for each of the holdout corruptions (corresponding to the five severity levels), and use this model to predict the expected top-1 error rates for the remaining corruptions within the corruption family. This scheme works particularly for the ``well defined'' corruption types such as noise and digital (\num{4.14}\% points absolute mean deviation from the real error. The full results are depicted in Table~\ref{tbl:predict-error}.
\input{figs/YY-supplement/table-predict-errors}
\subsection{Training details on the models trained with Fixup initialization and GroupNorm}
\label{app:fixup}
In Section 5 of the main paper, we consider IN models trained with GroupNorm and Fixup initialization.
For these models, we consider the original reference implementations provided by the authors.
We train ResNet-50, ResNet-101 and ResNet-152 models with stochastic gradient descent with momentum (learning rate 0.1, momentum 0.9), with batch size 256 and weight decay \num{1e-4} for 100 epochs.
\input{figs/04-results/ablation-priors-new.tex}
\input{figs/YY-supplement/all_results.tex}
\clearpage
\section{Analytical error model}\label{app:analytical}
We consider a univariate model in \textsection\ref{app:proof-sketch}--\ref{app:proof-proof} and discuss a simple extension to the multivariate diagonal case in \textsection\ref{app:proof-multi}.
As highlighted in the main text, the model qualitatively explains the overall characteristics of our experimental data.
Note that we assume a linear relationship between the Wasserstein distance and the error under domain shift, as suggested by our empirical findings.
\paragraph{Univariate model.}
We denote the source statistics as $\mu_s,\sigma_s^2$, the true target statistics as $\mu_t,\sigma^2_t$ and the
estimated target statistics as $\hat{\mu}_t,\hat{\sigma}_t^2$.
For normalization, we take a convex combination of the source statistics and estimated target statistics:
\begin{align}\label{definitionbarmu}
\Bar{\mu} = \frac{N}{N+n} \mu_s + \frac{n}{N+n} \hat{\mu}_t, \;
\Bar{\sigma}^2 = \frac{N}{N+n} \sigma^2_s + \frac{n}{N+n} \hat{\sigma}^2_t .
\end{align}
We now analyze the trade-off between using an estimate closer to the source or closer to the estimated target statistics.
In the former case, the model will suffer under the covariate shift present between target and source distribution.
In the latter case, small batch sizes $n$ will yield unreliable estimates for the true target statistics, which might hurt the performance even more than the source-target mismatch.
Hence, we aim to gain understanding in the trade-off between both options, and potential optimal choices of $N$ for a given sample size $n$.
As a metric of domain shift with good properties for our following derivation, we leverage the Wasserstein distance.
In \textsection\ref{sec:results} and \textsection\ref{app:error_prediction}, we already established an empirical link between domain shift measured in terms of the top-1 performance vs. the Wasserstein distance between model statistics and observed a linear relationship for case of common corruptions.
\renewcommand\theprop{1}
\begin{prop}[Bounds on the expected value of the Wasserstein distance between target and combined estimated target and source statistics]\label{prop:bounds}
We denote the source statistics as $\mu_s,\sigma_s^2$, the true target statistics as $\mu_t,\sigma^2_t$ and the biased estimates of the target statistics as $\hat{\mu}_t,\hat{\sigma}_t^2$.
For normalization, we take a convex combination of the source statistics and estimated target statistics as discussed in Eq.~\ref{definitionbarmu}.
At a confidence level $1-\alpha$, the expectation value of the squared Wasserstein distance
$W_2^2(\Bar{\mu}, \Bar{\sigma}, \mu_t, \sigma_t)$ between ideal and estimated target statistics w.r.t. to the distribution of sample mean $\hat{\mu}_t$ and sample variance $\hat{\sigma}^2_t$ is bounded from above and below
with $L \leq \mathbb{E}[W^2_2] \leq U$, where
\begin{equation}\label{univariatebound}
\begin{aligned}
L &= \left( \sigma_t - \sqrt{ \frac{N}{N+n} \sigma^2_s + \frac{n-1}{N+n} \sigma^2_t } \right)^2 + \frac{N^2}{(N+n)^2} \left(\mu_t - \mu_s \right)^2 + \frac{n}{(N+n)^2} \sigma_t^2 \\
U &= \; L + \sigma^5_t \frac{(n-1)}{2(N+n)^2} \left( \frac{N}{N+n} \sigma^2_s + \frac{1}{N+n} \chi^2_{1-\alpha/2, n-1} \sigma^2_t \right)^{-3/2}
\end{aligned}
\end{equation}
The quantity $\chi^2_{1-\alpha/2, n-1}$ denotes the left tail value of a chi square distribution with $n-1$ degrees of freedom, defined as $P\left(X \leq \chi^2_{1-\alpha/2, n-1}\right) = \alpha/2 \text{ for } X \sim \chi^2_{n-1}$.
\end{prop}
\subsection{Proof sketch}
\label{app:proof-sketch}
We are interested in the expected value of the Wasserstein distance defined in \eqref{def:wasserstein} between the target statistics $\mu_t, \sigma^2_t$ and the mixed statistics $\Bar{\mu}, \Bar{\sigma}^2$ introduced above in equation \eqref{definitionbarmu}, taken with respect to the distribution of the sample moments $\hat{\mu}_t$, $\hat{\sigma}_t^2$. The expectation value itself cannot be evaluated in closed form because the Wasserstein distance contains a term proportional to $\Bar{\sigma}$ being the square root of the convex combination of target and source variance.
In Lemma \ref{lemma1}, the square root term is bounded from above and below using Jensen's inequality and Holder's defect formula which is reviewed in Lemma \ref{holdersdefect}. After having bounded the problematic square root term, the proof of Proposition~\ref{prop:bounds} reduces to inserting the expectation values of sample mean and sample variance reviewed in Lemma \ref{lemma:moments}.
\subsection{Prerequisites}
\label{app:proof-reqisisites}
\begin{lemma}[Mean and variance of sample moments, following \citep{weisstein}]\label{lemma:moments}
The sample moments $\hat{\mu}_t,\hat{\sigma}_t^2$ are random variables depending on the sample size $n$.
\begin{align}
\hat{\mu}_t &= \; \frac{1}{n} \sum_{j=1}^n x_j, \quad
\hat{\sigma}_t^2 = \; \frac{1}{n} \sum_{j=1}^n (x_j - \hat{\mu}_t)^2 \; \text{ with } x_j \sim \mathcal{N} \left(\mu_t, \sigma^2_t \right).
\end{align}
For brevity, we use the shorthand $\mathbb{E}[\cdot]$ for all expectation values with respect to the distribution of $p(\hat{\mu}_t,\hat{\sigma}_t^2|n)$.
In particular, our computation uses mean and variance of $\hat{\mu}_t$ and $\hat{\sigma}_t^2$ which are well known for a normal target distribution:
\begin{align}
\hat{\mu}_t \sim \mathcal{N}\left(\mu_t, \frac{1}{n} \sigma^2_t \right), \; \mathbb{E}[\hat{\mu}_t] &= \; \mu_t, \;
\mathbb{V}[\hat{\mu}_t] = \frac{1}{n} \sigma^2_t \\
\frac{\hat{\sigma}^2_t}{\sigma_t^2/n} \sim \chi^2_{n-1}, \; \mathbb{E}[ \hat{\sigma}_t^2] &= \; \frac{n-1}{n} \sigma^2_t, \; \mathbb{V}[\hat{\sigma}^2_t] = \frac{\sigma_t^4}{n^2} \mathbb{V}\left[\frac{\hat{\sigma}^2_t}{\sigma_t^2/n}\right] =
\frac{\sigma_t^4}{n^2} \, 2(n-1).
\end{align}
The derivation of the variance $\mathbb{V}[\hat{\sigma}_t^2]$ in the last line uses the fact that the variance of a chi square distributed variable with $(n-1)$ degrees of freedom is equal to $2(n-1)$.
\end{lemma}
\begin{lemma}[Holder's defect formula for concave functions in probabilistic notation, following \citet{becker2012variance} ] \label{holdersdefect}
If the concave function $f: [a,b] \to \mathbb{R}$ is twice continuously differentiable and there are finite bounds $m$ and $M$ such that
\begin{align}
-M \leq f''(x) \leq -m \leq 0 \; \forall x \in [a,b],
\end{align}
then the defect between Jensen's inequality estimate $f\left(\mathbb{E}[X]\right)$ for a random variable $X$ taking values $x \in [a,b]$ and the true expectation value $\mathbb{E}[f(X)]$ is bounded from above by a term proportional to the variance of $X$:
\begin{align}
f \left( \mathbb{E}[X] \right) - \mathbb{E}[f(X)] \leq \frac{1}{2} M \mathbb{V}[X].
\end{align}
\end{lemma}
\begin{lemma}[Upper and lower bounds on the expectation value of $\Bar{\sigma}$] \label{lemma1}
The expectation value of the square root of the random variable $\Bar{\sigma}^2$ defined as
\begin{align}\label{definitionX}
\Bar{\sigma}^2 &= \frac{N}{N+n} \sigma^2_s + \frac{n}{N+n} \hat{\sigma}^2_t,
\end{align}
is bounded from above and below at a confidence level $1-\alpha$ by
\begin{align} \sqrt{\mathbb{E}\left[ \Bar{\sigma}^2 \right ]} - \frac{1}{2} M \mathbb{V}[\Bar{\sigma}^2] &\leq
\mathbb{E}\left[\sqrt{\Bar{\sigma}^2}\right] \leq
\sqrt{\mathbb{E}\left[\Bar{\sigma}^2\right]}\\
\sqrt{\mathbb{E}\left[\Bar{\sigma}^2\right]} &= \sqrt{\frac{N}{N+n} \sigma^2_s + \frac{n-1}{N+n} \sigma^2_t}, \\
\frac{1}{2} M \mathbb{V}[\Bar{\sigma}^2] &= \frac{(n-1)}{4(N+n)^2} \sigma^4_t \left( \frac{N}{N+n} \sigma^2_s + \frac{1}{N+n} \chi^2_{1-\alpha/2, n-1} \sigma^2_t \right).
\end{align}
The quantity $\chi^2_{1-\alpha/2, n-1}$ denotes the left tail value of a chi square distribution with $n-1$ degrees of freedom, defined as $P\left(X \leq \chi^2_{1-\alpha/2, n-1}\right) = \alpha/2 \; \text{ for } X \sim \chi^2_{n-1}$.
\begin{proof}
The square root function is concave, therefore Jensen's inequality implies the upper bound
\begin{align}
\mathbb{E}\left[\sqrt{\Bar{\sigma}^2}\right] \leq \sqrt{\mathbb{E}[\Bar{\sigma}^2]}.
\end{align}
The square root of the expectation value of $\Bar{\sigma}^2$ is computed using the expectation value of the sample variance as given in Lemma \ref{lemma:moments}.
\begin{align}
\sqrt{\mathbb{E}[\Bar{\sigma}^2]} &= \sqrt{\frac{N}{N+n} \sigma^2_s + \frac{n}{N+n} \frac{n-1}{n} \sigma^2_t}
= \sqrt{\frac{N}{N+n} \sigma^2_s + \frac{n-1}{N+n} \sigma^2_t}.
\end{align}
To state a lower bound, we use Holder's defect formula in probabilistic notation stated in Lemma \ref{holdersdefect}.
Holder's formula for concave functions requires that the random variable $\Bar{\sigma}^2$ can take values in the compact interval $[a,b]$ and that the second derivative of the square root function $f(\Bar{\sigma}^2) = \sqrt{\Bar{\sigma}^2}$, exists and is strictly smaller than zero in $[a,b]$.
Regarding the interval of $\Bar{\sigma}^2$, we provide probabilistic upper and lower bounds.
The ratio of sample variance and true variance divided by $n$ follows a chi square distribution with $n-1$ degrees of freedom.
At confidence level $1-\alpha$, this ratio lies between $\chi^2_{1-\alpha/2, n-1}$ and $\chi^2_{\alpha/2, n-1}$ which are defined as follows:
\begin{align}
\chi^2_{1-\alpha/2, n-1} &\leq \frac{\hat{\sigma}^2_t}{\sigma^2_t/n} \leq \chi^2_{\alpha/2, n-1}, \\
Pr(X \leq \chi^2_{1-\alpha/2, n-1}) &= \frac{\alpha}{2}, \;
Pr(X \geq \chi^2_{\alpha/2, n-1}) = \frac{\alpha}{2}.
\end{align}
Then at the same confidence level, the sample variance itself lies between the two quantiles multiplied by $\sigma_t^2/n$,
\begin{align}
\chi^2_{1-\alpha/2, n-1} \frac{\sigma_t^2}{n} &\leq \hat{\sigma}_t^2 \leq \chi^2_{\alpha/2, n-1} \frac{\sigma_t^2}{n},
\end{align}
and the random variable $\Bar{\sigma}^2$ lies in the interval
\begin{align} \label{definitionofa}
\Bar{\sigma}^2 \in [a,b] \text{ with } a &= \frac{N}{N+n} \sigma^2_s + \frac{1}{N+n} \chi^2_{1-\alpha/2, n-1} \sigma_t^2, \\
\text{ and } b &= \frac{N}{N+n} \sigma^2_s + \frac{1}{N+n} \chi^2_{\alpha/2, n-1} \sigma_t^2.
\end{align}
The variances and chi square values are all positive and therefore both $a$ and $b$ are positive as well, implying that the second derivative of the square root is strictly negative in the interval $[a,b]$.
\begin{align}
f(\Bar{\sigma}^2) &= \sqrt{\Bar{\sigma}^2}, \; f'(\Bar{\sigma}^2) = \frac{1}{2} (\Bar{\sigma}^2)^{-1/2}, \; f''(\Bar{\sigma}^2) = - \frac{1}{4} (\Bar{\sigma}^2)^{-3/2} < 0 \in [a,b].
\end{align}
Consequently the second derivative is in the interval $[M, m]$ at the given confidence level:
\begin{align}\label{computeM}
-M &\leq f''(\Bar{\sigma}^2) \leq -m \leq 0 \text{ for } \Bar{\sigma}^2 \in [a,b] \text{ with }
M = \frac{1}{4} a^{-3/2}, \; m = \frac{1}{4} b^{-3/2}.
\end{align}
The defect formula \ref{holdersdefect} states that the defect is bounded by
\begin{align}\label{defectformula}
\sqrt{\mathbb{E}[\Bar{\sigma}^2]} - \mathbb{E}[\sqrt{\Bar{\sigma}^2}] &\leq \frac{1}{2} M \mathbb{V}[\Bar{\sigma}^2].
\end{align}
The constant $M$ was computed above in \eqref{computeM}, and the variance of $\Bar{\sigma}^2$ is calculated in the next lines, using the first and second moment of the sample variance as stated in \ref{lemma:moments}.
\begin{equation}
\begin{aligned}\label{varianceX}
\mathbb{V}[\Bar{\sigma}^2] &= \mathbb{E}[(\Bar{\sigma}^2 - \mathbb{E}[\Bar{\sigma}^2])^2] = \mathbb{E}\left[\left(\frac{n}{N+n} \hat{\sigma}^2_t - \frac{n}{N+n} \frac{n-1}{n}\sigma^2_t \right)^2 \right] \\
&= \frac{n^2}{(N+n)^2} \mathbb{E}\left[ \left(\hat{\sigma}^2_t - \mathbb{E}[\hat{\sigma}^2_t\right)^2\right]
= \frac{n^2}{(N+n)^2} \mathbb{V}\left[\hat{\sigma}^2_t \right] \\
&= \frac{n^2}{(N+n)^2} \frac{2(n-1)}{n^2} \sigma^4_t = \frac{2(n-1)}{(N+n)^2} \sigma^4_t .
\end{aligned}
\end{equation}
Inserting $\mathbb{V}[\Bar{\sigma}^2]$ computed in \eqref{varianceX} and $M$ defined in \eqref{computeM} with $a$ as defined in \eqref{definitionofa} into the defect formula \eqref{defectformula} yields the lower bound:
\begin{equation}
\begin{aligned}
&\sqrt{\mathbb{E}[\Bar{\sigma}^2]} - \frac{1}{2} M \mathbb{V}[\Bar{\sigma}^2] \leq \mathbb{E}[\sqrt{\Bar{\sigma}^2}] \\
& \sqrt{\mathbb{E}[\Bar{\sigma}^2]} - \frac{1}{2} M \mathbb{V}[\Bar{\sigma}^2] \\
&= \sqrt{\mathbb{E}[\Bar{\sigma}^2]} - \frac{1}{2} \cdot \frac{1}{4} a^{-3/2} \frac{2(n-1)}{(N+n)^2} \sigma^4_t \\
&= \sqrt{\mathbb{E}[\Bar{\sigma}^2]} - \frac{(n-1)}{4(N+n)^2} \sigma^4_t \left(\frac{N}{N+n} \sigma^2_s + \frac{1}{N+n} \chi^2_{1-\alpha/2, n-1} \sigma_t^2\right)^{-3/2}.
\end{aligned}
\end{equation}
Assuming that source and target variance are of the same order of magnitude $\sigma$, the defect will be of order of magnitude $\sigma$: The factor $\mathbb{V}[X]$ scales with $\sigma^4$ and $M$ with $\sigma^{-3}$.
\end{proof}
\end{lemma}
\subsection{Proof of Proposition 1}
\label{app:proof-proof}
\begin{proof}
For two univariate normal distributions with moments $\mu_t, \sigma^2_t$ and $\Bar{\mu}, \Bar{\sigma}^2$, the Wasserstein distance as defined in \eqref{def:wasserstein} reduces to
\begin{align}
W_2^2 &= \sigma^2_t + \Bar{\sigma}^2 - 2 \Bar{\sigma} \sigma_t + (\Bar{\mu} - \mu)^2.
\end{align}
The expected value of the Wasserstein distance across many batches is given as
\begin{equation}
\begin{aligned}\label{expectedvaluewasserstein}
\mathbb{E}[W_2^2] =& \; \sigma_t^2 + \mathbb{E}[\Bar{\sigma}^2] - 2 \mathbb{E}[\Bar{\sigma}] \sigma_t + \mathbb{E}[(\mu_t - \Bar{\mu})^2] \\
=& \; \sigma^2_t + \frac{N}{N+n} \sigma^2_s + \frac{n}{N+n} \frac{n-1}{n} \sigma^2_t - 2 \sigma_t \mathbb{E}\left[\sqrt{\frac{N}{N+n} \sigma^2_s + \frac{n}{N+n} \hat{\sigma}^2_t}\right] \\
&+ \mathbb{E}\left[\left(\mu_t - \frac{N}{N+n} \mu_s - \frac{n}{N+n} \hat{\mu}_t \right)^2 \right]
\end{aligned}
\end{equation}
which can already serve as the basis for our numerical simulations.
To arrive at a closed form analytical solution, we invoke Lemma~\ref{lemma1} to bound the expectation value $\mathbb{E}\left[\Bar{\sigma}\right]$ in equation \eqref{expectedvaluewasserstein}.
\begin{align}
\label{boundsquareroot}
-2 \sigma_t \sqrt{\mathbb{E}\left[\Bar{\sigma}^2\right]} &\leq -2 \sigma_t
\mathbb{E}\left[\sqrt{\Bar{\sigma}^2}\right] \leq
-2 \sigma_t \sqrt{\mathbb{E}\left[\Bar{\sigma}^2\right]} -2 \sigma_t \left( -\frac{1}{2} M \mathbb{V}[\Bar{\sigma}^2] \right)
\end{align}
Apart from the square root term bounded in equation \eqref{boundsquareroot} above, the expectation value of the Wasserstein distance can be computed exactly. Hence the bounds on $\mathbb{E}\left[\Bar{\sigma}\right]$ multiplied by a factor of $(-2 \sigma^2_t)$ coming from equation \eqref{expectedvaluewasserstein} determine lower and upper bounds $L$ and $U$ on the expected value of $W_2^2$:
\begin{align}
L &\leq \mathbb{E}\left[W^2_2\right] \leq U = L + \sigma_t M \mathbb{V}[\Bar{\sigma}^2]
\end{align}
In the next lines, the lower bound is calculated:
\begin{equation}
\begin{aligned}
L &= \; \sigma^2_t + \frac{N}{N+n} \sigma^2_s + \frac{n-1}{N+n} \sigma^2_t - 2 \sigma_t \sqrt{\mathbb{E}\left[ \frac{N}{N+n} \sigma^2_s + \frac{n-1}{N+n} \sigma^2_t \right]} \\
&+ \left(\mu_t - \frac{N}{N+n} \mu_s\right)^2 - 2\left(\mu_t - \frac{N}{N+n} \mu_s \right) \frac{n}{N+n} \mathbb{E}[\hat{\mu}_t] + \frac{n^2}{(N+n)^2} \left( \mathbb{V}[\hat{\mu}_t] + \left(\mathbb{E}[\hat{\mu}_t] \right)^2 \right) \\
= & \; \sigma^2_t + \frac{N}{N+n} \sigma^2_s + \frac{n-1}{N+n} \sigma^2_t - 2 \sigma_t \sqrt{\frac{N}{N+n} \sigma^2_s + \frac{n-1}{N+n} \sigma^2_t} \\
&+ \left(\mu_t - \frac{N}{N+n} \mu_s\right)^2 - 2\left(\mu_t - \frac{N}{N+n} \mu_s \right) \frac{n}{N+n} \mu_t + \frac{n^2}{(N+n)^2} \left( \frac{1}{n} \sigma^2_t + \mu_t^2 \right) \\
= & \; \left( \sigma_t - \sqrt{ \frac{N}{N+n} \sigma^2_s + \frac{n-1}{N+n} \sigma^2_t } \right)^2 + \left( \mu_t - \frac{N}{N+n} \mu_s - \frac{n}{N+n} \mu_t \right)^2 + \frac{n}{(N+n)^2} \sigma_t^2 \\
= & \; \left( \sigma_t - \sqrt{ \frac{N}{N+n} \sigma^2_s + \frac{n-1}{N+n} \sigma^2_t } \right)^2 + \frac{N^2}{(N+n)^2} \left(\mu_t - \mu_s \right)^2 + \frac{n}{(N+n)^2} \sigma_t^2
\end{aligned}
\end{equation}
After having derived the lower bound, the upper bound is the sum of the lower bound and the defect term as computed in Lemma \ref{lemma1}.
\begin{equation}
\begin{aligned}
\mathbb{E}[W^2] &\geq U = L + \sigma_t M \mathbb{V}[\Bar{\sigma}^2] \\
&= L + \sigma_t \frac{1}{4} \left( \frac{N}{N+n} \sigma^2_s + \frac{n}{N+n} \chi^2_{1-\alpha/2, n-1} \frac{\sigma_t^2}{n} \right)^{-3/2} \frac{2(n-1)}{(N+n)^2} \sigma_t^4 \\
&= L + \left( \frac{N}{N+n} \sigma^2_s + \frac{1}{N+n} \chi^2_{1-\alpha/2, n-1} \sigma_t^2 \right)^{-3/2} \frac{(n-1)}{2(N+n)^2} \sigma_t^5.
\end{aligned}
\end{equation}
\end{proof}
Based on choices of the model parameters, the model qualitatively matches our experimental results. We plot different choices in Fig.~\ref{fig:simulation}.
\begin{figure}[hp]
\centering
\includegraphics[width=\textwidth]{figs/YY-supplement/theory-plots.pdf}
\caption{Overview of different parametrizations of the model.
We denote each plot with $(\mu_t - \mu_s, \sigma_t/\sigma_s)$ and report the lower bound $\sqrt{L}$ on the Wasserstein distance.
Parametrizations in columns four to seven produce qualitatively similar results we observed in our experiments, assuming a linear relationship between the Wasserstein distance and the error rate.
}
\label{fig:simulation}
\end{figure}
\subsection{Extension to multivariate distributions.}
\label{app:proof-multi}
We now derive a multivariate variant that can be fit to data from a DNN.
Due to the estimation of running statistics in the network, we have access to a diagonal approximation of the true covariance matrix.
We denote the diagonal covariance matrices with matrix elements $\sigma^2_i$ as
\begin{align}
(\Sigma_t)_{ii} &= (\sigma^2_t)_{i}, \; (\hat{\Sigma}_t)_{ii} = (\hat{\sigma}^2_t)_{i}, \; (\Sigma_s)_{ii} = (\sigma^2_s)_{i}
\end{align}
and extend our definition of the statistics used for normalization to $\Bar{\bm{\mu}}$ and $\Bar{\bm{\Sigma}}$:
\begin{align}
\Bar{\bm{\mu}} =& \; \frac{N}{N+n} \bm{\mu}_s + \frac{n}{N+n} \hat{\bm{\mu}}_t, \; \Bar{\bm{\Sigma}} = \frac{N}{N+n} \bm{\Sigma}_s + \frac{n}{N+n} \hat{\bm{\Sigma}}_t.
\end{align}
The Wasserstein distance between $\Bar{\bm{\mu}},\Bar{\bm{\Sigma}}$ and $\bm{\mu}_t, \bm{\Sigma}_t$ is then defined as
\begin{equation}
\begin{aligned}
W_2^2 =& \; \Tr{\bm{\Sigma}_t + \Bar{\bm{\Sigma}} - 2 \bm{\Sigma}_t^{1/2} \Bar{\bm{\Sigma}}^{1/2}} + (\bm{\mu}_t - \Bar{\bm{\mu}})^T (\bm{\mu}_t - \Bar{\bm{\mu}}) \\
&= \sum_{i=1}^D (\sigma^2_t)_i + (\Bar{\sigma}^2)_i - 2 (\Bar{\sigma})_i (\sigma_t)_i + \left((\mu_t)_i - (\Bar{\mu}_t)_i\right)^2 = \sum_{i=1}^D (W_2^2)_i
\end{aligned}
\end{equation}
Every component $(W_2^2)_i$ in the sum above is bounded by the univariate bound discussed above.
The multivariate Wasserstein distance which sums over the diagonal covariance matrix entries is then bounded by the sums over the individual bounds $L_i$ and $U_i$ given in \eqref{univariatebound}.
\begin{align}
L_i \leq (W_2^2)_i \leq U_i \Rightarrow \sum_{i=1}^D L_i \leq W_2^2 \leq \sum_{i=1}^D U_i.
\end{align}
\subsection{Limits of Proposition 1}
\paragraph{Limit $n \to \infty$}
In the limit of infinite batch size $n \to \infty$, upper and lower bounds on the expected Wasserstein distance between $\Bar{\mu}, \Bar{\sigma}^2$ and $\mu_t, \sigma^2_t$ both go to zero.
\begin{equation}
\begin{aligned}
\lim_{n \to \infty} L =& \lim_{n \to \infty} \left( \sigma_t - \sqrt{ \frac{N}{N+n} \sigma^2_s + \frac{n-1}{N+n} \sigma^2_t } \right)^2 + \frac{N^2}{(N+n)^2} \left(\mu_t - \mu_s \right)^2 + \frac{n}{(N+n)^2} \sigma_t^2 \\
=& (\sigma_t - \sigma_t)^2 = 0 \\
\lim_{n \to \infty} U =& \lim_{n \to \infty} L + \lim_{n \to \infty} \sigma^5_t \frac{(n-1)}{2(N+n)^2} \left( \frac{N}{N+n} \sigma^2_s + \frac{1}{N+n} \chi^2_{1-\alpha/2, n-1} \sigma^2_t \right)^{-3/2} = 0.
\end{aligned}
\end{equation}
The intuition behind this limit is that if a large number of samples from the target domain is given, $\hat{\mu}$ and $\hat{\sigma}^2$ approximate the true target statistics very well. As $\hat{\mu}$ and $\hat{\sigma}^2$ dominate $\Bar{\mu}$ and $\Bar{\sigma}^2$ for large $n$, the expected Wasserstein distance has to vanish.
\paragraph{Limit $N \to \infty$}
In the opposite limit $N \to \infty$, the expected value of the Wasserstein distance reduces to the Wasserstein distance between source and target statistics.
\begin{align}
\lim_{N \to \infty} \Bar{\mu} &= \mu_s, \; \lim_{N \to \infty} \Bar{\sigma}^2 = \sigma^2_s, \\
\Rightarrow \;
\lim_{N \to \infty} \mathbb{E}[W_2^2] &= \sigma^2_t + \sigma^2_s - 2 \sigma_t \sigma_s + (\mu_t - \mu_s)^2
= W_2^2 \left( \mu_s, \sigma^2_s, \mu_t, \sigma^2_t \right).
\end{align}
\paragraph{Limiting case $\mu_t = \mu_s$ and $\sigma^2_t = \sigma^2_s$}
When source and target domain coincide, and the statistics $\sigma^2_s = \sigma^2_t$ and $\mu_s = \mu_t$ are known, then the source target mismatch is not an error source.
However, one might assume that source and target domain are different even though they actually coincide. In this case, proceeding with our proposed strategy and using the statistics $\Bar{\mu}$ and $\Bar{\sigma}^2$, the bounds on the expected Wasserstein distance follow from setting $\sigma^2_t$ to $\sigma^2_s$ and $\mu_t$ to $\mu_s$ in Proposition~\ref{prop:bounds}.
\begin{equation}\label{source=target}
\begin{aligned}
\Bar{\mu} &= \frac{N}{N+n} \mu_t + \frac{n}{N+n} \hat{\mu}_t, \; \Bar{\sigma}^2 = \frac{N}{N+n} \sigma^2_t + \frac{n}{N+n} \hat{\sigma}^2_t, \; L \leq \mathbb{E}[W_2^2] \leq U \\
L &= \sigma^2_t \left(\frac{2N^2 + 4Nn - N + 2n^2}{(N+n)^2} -2 \sqrt{1 - \frac{1}{N+n}} \right), \\
U &= L + \sigma^2_t \frac{n-1}{2(N+n)^2} \left(\frac{N + \chi^2_{1-\alpha/2, n-1}}{N+n} \right)^{-3/2}.
\end{aligned}
\end{equation}
It could also be the case that the equality of source and target statistics is known but the concrete values of the statistics are unknown. In our model, this amounts to setting the number of pseudo samples $N$ to zero and assuming that source and target statistics are equal. Setting $N=0$ in equation \eqref{source=target} and keeping $n$ finite yields
\begin{align}
L &= 2 \sigma^2_t \left(1 - \sqrt{1 - \frac{1}{n}} \right), \; U = L + \sigma^2_t \frac{n-1}{2n^2} \left( \frac{\chi^2_{1-\alpha/2, n-1}}{n} \right)^{-3/2}.
\end{align}
\subsection{Bounds on the normalized Wasserstein distance}
The Wasserstein distance (cf. \textsection\ref{def:wasserstein}) between the interpolating statistics $\Bar{\mu}$, $\Bar{\sigma}^2$ and the target statistics can also be normalized by a factor of $\sigma_s^{-2}$.
Because $\sigma_s^{-2}$ is constant, the bounds on the expectation value of the unnormalized Wasserstein distance discussed in the previous subsections just have to be multiplied by $\sigma_s^{-2}$ to obtain bounds on the normalized Wasserstein distance (cf. \textsection\ref{def:wasserstein-norm}):
\begin{align}
\frac{L}{\sigma^2_s} \leq \widetilde{W}_2^2 = W_2^2 \left(\frac{\Bar{\mu}}{\sigma_s},, \frac{\Bar{\sigma}^2}{\sigma^2_s},
\frac{\mu_t}{\sigma_s},
\frac{\sigma_t^2}{\sigma_s^2} \right) = \frac{1}{\sigma^2_s} W_2^2(\Bar{\mu},
\Bar{\sigma}^2, \mu_t, \sigma^2_t) \leq \frac{U}{\sigma^2_s}.
\end{align}
\clearpage
\section{Full list of models evaluated on IN}
\label{apx:imagenet_model_descriptions}
The following lists contains all models we evaluated on various datasets with references and links to the corresponding source code.
\subsection{Torchvision models trained on IN}
Weights were taken from \url{https://github.com/pytorch/vision/tree/master/torchvision/models}
{\small
\begin{enumerate}
\item \texttt{alexnet} \citep{alexnet}
\item \texttt{densenet121} \citep{densenet}
\item \texttt{densenet161} \citep{densenet}
\item \texttt{densenet169} \citep{densenet}
\item \texttt{densenet201} \citep{densenet}
\item \texttt{densenet201} \citep{densenet}
\item \texttt{googlenet} \citep{googlenet}
\item \texttt{inception\_v3} \citep{inception}
\item \texttt{mnasnet0\_5} \citep{mnasnet}
\item \texttt{mnasnet1\_0} \citep{mnasnet}
\item \texttt{mobilenet\_v2} \citep{mobilenet}
\item \texttt{resnet18} \citep{he2016resnet}
\item \texttt{resnet34} \citep{he2016resnet}
\item \texttt{resnet50} \citep{he2016resnet}
\item \texttt{resnet101} \citep{he2016resnet}
\item \texttt{resnet152} \citep{he2016resnet}
\item \texttt{resnext50\_32x4d} \citep{xie2017aggregated}
\item \texttt{resnext101\_32x8d} \citep{xie2017aggregated}
\item \texttt{shufflenet\_v2\_x0\_5} \citep{shufflenet}
\item \texttt{shufflenet\_v2\_x1\_0} \citep{shufflenet}
\item \texttt{vgg11\_bn} \citep{vgg}
\item \texttt{vgg13\_bn} \citep{vgg}
\item \texttt{vgg16\_bn} \citep{vgg}
\item \texttt{vgg19\_bn} \citep{vgg}
\item \texttt{wide\_resnet101\_2} \citep{wideresnet}
\item \texttt{wide\_resnet50\_2} \citep{wideresnet}
\end{enumerate}
}
\iffalse
\subsection{EfficientNet Models \citep{tan2019efficientnet} }
Weights were taken from \url{https://github.com/tensorflow/tpu/tree/master/models/official/efficientnet}, the EfficientNet implementation in Pytorch was taken from \url{https://github.com/lukemelas/EfficientNet-PyTorch}.
{\small
\begin{enumerate}
\item \texttt{efficientnet-b0}
\item \texttt{efficientnet-b1}
\item \texttt{efficientnet-b2}
\item \texttt{efficientnet-b3}
\item \texttt{efficientnet-b4}
\end{enumerate}
}
\fi
\subsection{Robust ResNet50 models}
{\small
\begin{enumerate}
\item \texttt{resnet50 AugMix} \citep{hendrycks2019augmix} \url{https://github.com/google-research/augmix}
\item \texttt{resnet50 SIN+IN} \citep{geirhos2018imagenettrained} \url{https://github.com/rgeirhos/texture-vs-shape}
\item \texttt{resnet50 ANT} \citep{Rusak2020IncreasingTR}
\url{https://github.com/bethgelab/game-of-noise}
\item \texttt{resnet50 ANT+SIN} \citep{Rusak2020IncreasingTR}
\url{https://github.com/bethgelab/game-of-noise}
\item \texttt{resnet50 DeepAugment}
\citep{hendrycks2020many}
\url{https://github.com/hendrycks/imagenet-r}
\item \texttt{resnet50 DeepAugment+AugMix}
\citep{hendrycks2020many}
\url{https://github.com/hendrycks/imagenet-r}
\end{enumerate}
}
\subsection{SimCLRv2 models \citep{chen2020big}}
We used the checkpoints from \url{https://github.com/google-research/simclr} and converted them from TensorFlow to PyTorch with \url{https://github.com/tonylins/simclr-converter}, commit ID: 139d3cb0bd0c64b5ad32aab810e0bd0a0dddaae0.
{\small
\begin{enumerate}
\item \texttt{resnet50} FT100 SK=0 width=1
\item \texttt{resnet101} FT100 SK=0 width=1
\item \texttt{resnet152} FT100 SK=0 width=1
\end{enumerate}
}
\subsection{Robust ResNext models \citep{xie2017aggregated}}
Note that the baseline \texttt{resnext50\_32x4d} model trained on ImageNet is available as part of the \texttt{torchvision} library.
{\small
\begin{enumerate}
\item \texttt{resnext50\_32x4d WSL} \citep{mahajan2018exploring} \url{https://github.com/facebookresearch/WSL-Images/blob/master/hubconf.py}
\item \texttt{resnext101\_32x4d WSL} \citep{mahajan2018exploring} \url{https://github.com/facebookresearch/WSL-Images/blob/master/hubconf.py}
\item \texttt{resnext101\_32x8d Deepaugment+AugMix} \citep{hendrycks2020many} \url{https://github.com/hendrycks/imagenet-r}
\end{enumerate}
}
\subsection{ResNet50 with Group Normalization \citep{wu2018group}}
Model weights and training code was taken from \url{https://github.com/ppwwyyxx/GroupNorm-reproduce}
{\small
\begin{enumerate}
\item \texttt{resnet50 GroupNorm}
\item \texttt{resnet101 GroupNorm}
\item \texttt{resnet152 GroupNorm}
\end{enumerate}
}
\subsection{ResNet50 with Fixup initialization \citep{zhang2019fixup}}
Model weights and training code was taken from \url{https://github.com/hongyi-zhang/Fixup/tree/master/imagenet}.
For training, we keep all hyperparameters at their default values and note that in particular the batchsize of 256 is a sensitive parameter.
{\small
\begin{enumerate}
\item \texttt{resnet50 FixUp}
\item \texttt{resnet101 FixUp}
\item \texttt{resnet152 FixUp}
\end{enumerate}
}
\subsection{Effect of Pseudo Batchsize}
\begin{figure}[hp]
\begin{center}
\begin{tikzpicture}
\begin{axis}[%
tuftelike,
priorplot,
xmode=log,
xmin=1,
xmax=512,
ymin=50,
ymax=75,
legend pos=north east,
legend columns=3,
legend style={draw = none, at={(1.2,0.99), font=\tiny}, name=legend},
cycle list/Dark2-8,
title={Performance for optimal $N$}]
\foreach \modelname in {resnet, augmix, ant, antsin, sin}
{
\addplot+ [%
opacity=.95,
mark options = {solid},
discard if not={model}{\modelname}%
] table [
x=bsz,
y=mce,
]{data/batchsize/priors.tsv};
}
\legend{
ResNet,
AugMix,
ANT,
ANT+SIN,
SIN
}
\end{axis}
\begin{axis}[%
tuftelike,
at={(.5\textwidth,0)},
priorplot,
xmin=1,
xmax=512,
ymin=8,
ymax=1024,
ylabel={Pseudo Batch Size},
ymode=log,
xmode=log,
ytick= {1,2,4,8,16,32,64,128,256,512},
yticklabels={1, , ,8, , ,64, , ,512},
legend pos=north east,
legend columns=3,
legend style={draw = none, at={(1.1,0.99), font=\footnotesize}, name=legend},
cycle list/Dark2-8,
title={Best Pseudo Batchsize $N$}]
\foreach \modelname/\idx in {resnet/0, augmix/1, ant/2, antsin/3, sin/4}
{
\addplot+ [%
opacity=.95,
mark options = {solid},
] table [
x=bsz,
y=\modelname,
]{data/batchsize/priors-optimal-N.tsv};
}
\end{axis}
\end{tikzpicture}
\end{center}
\caption{%
Left: Performance for all the considered ResNet-50 variants based on the sample batch size.
The optimal $N$ is chosen according to the mCE on the holdout corruptions.
Right: Best choice for $N$ depending on the input batchsize $n$.
Note that in general for high values $n$, the model is generally more robust to the choice of $N$.
}
\begin{center}
\begin{tikzpicture}
\foreach \modelname/\ddx/\ddy in {resnet/0/0} {
\begin{axis}[tuftelike,priorplot,xmode=log,title=\modelname,at={(\ddx*0.5*\textwidth,-\ddy*0.31*\textwidth)}]
\addlegendimage{empty legend}
\foreach \ccol/\oppp in {
1/20,
2/30,
4/40,
8/50,
16/60,
32/70,
64/80,
128/90,
256/100
}
{
\edef\temp{\noexpand\addplot+ [
black!\oppp,
mark=none
] table [x=batchsize, y=\ccol]{data/priors/\modelname.tsv};
}
\temp
}
\legend{
\hspace{-.6cm}$N$,
1, 2, 4, 8, 16, 32, 64, 128, 256, 512
}
\end{axis}
}
\foreach \modelname/\ddx/\ddy in {augmix/1/1,
sin/0/1, ant/1/2, antsin/0/2} {
\begin{axis}[tuftelike,priorplot,xmode=log,title=\modelname,at={(\ddx*0.5*\textwidth,-\ddy*0.31*\textwidth)}]
\addlegendimage{empty legend}
\foreach \ccol/\oppp in {
1/20,
2/30,
4/40,
8/50,
16/60,
32/70,
64/80,
128/90,
256/100
}
{
\edef\temp{\noexpand\addplot+ [
black,
black!\oppp,
mark=none
] table [x=batchsize, y=\ccol]{data/priors/\modelname.tsv};
}
\temp
}
\end{axis}
}
\end{tikzpicture}
\end{center}
\caption{%
Effects of batch size $n$ and pseudo batch size $N$ for the various considered models.
We report mCE averaged across 15 test corruptions.
}
\label{fig:bsz_ant}
\end{figure}
We show the full results for considering different choices of $N$ for ResNet-50, Augmix, ANT, ANT+SIN and SIN models and display the result in Fig.~\ref{fig:bsz_ant}.
We observe a characteristic shape which we believe can be attributed to the way statistics are estimated.
We provide evidence for this view by proposing an analytical model which we discuss in \textsection\ref{app:analytical}.
\subsection{Mapping of ambiguous classes in ON to IN}
\label{app:objnet-mapping}
There are nine classes in ON that have multiple possible mappings to IN. We have discarded these classes in our evaluation. Out of the proposed 113 classes in ON that overlap with IN according to the ON publication, we therefore only test on 104 classes. The ambiguous classes and possible mappings are given in Table~\ref{ON:ambiguous}.
\begin{table}[ht]
\centering
\small
\begin{tabular}{l l }
\toprule
\\
ObjN class & IN classes \\
\midrule
wheel & wheel; paddlewheel, paddle wheel \\
helmet & football helmet; crash helmet \\
chair & barber chair; folding chair; rocking chair, rocker \\
still\_camera & Polaroid camera, Polaroid Land camera; reflex camera \\
alarm\_clock & analog clock; digital clock \\
tie & bow tie, bow-tie, bowtie; Windsor tie \\
pen & ballpoint, ballpoint pen, ballpen, Biro; quill, quill pen; fountain pen \\
bicycle & mountain bike, all-terrain bike, off-roader; bicycle-built-for-two, tandem bicycle, tandem \\
skirt & hoopskirt, crinoline; miniskirt, mini; overskirt \\
\end{tabular}
\caption{Mapping between 9 ambiguous ON classes and the possible correspondences in IN. Different IN classes are separated with a semicolon.
}
\label{ON:ambiguous}
\end{table}
\part{}
\maketitle
\begin{abstract}
Today's state-of-the-art machine vision models are vulnerable to image corruptions like blurring or compression artefacts, limiting their performance in many real-world applications.
We here argue that popular benchmarks to measure model robustness against common corruptions (like ImageNet-C) underestimate model robustness in many (but not all) application scenarios.
The key insight is that in many scenarios, multiple unlabeled examples of the corruptions are available and can be used for unsupervised online adaptation.
Replacing the activation statistics estimated by batch normalization on the training set with the statistics of the corrupted images consistently improves the robustness across 25 different popular computer vision models.
Using the corrected statistics, ResNet-50 reaches 62.2\% mCE on ImageNet-C compared to 76.7\% without adaptation.
With the more robust DeepAugment+AugMix model, we improve the state of the art achieved by a ResNet50 model up to date from 53.6\% mCE to 45.4\% mCE.
Even adapting to a single sample improves robustness for the ResNet-50 and AugMix models, and 32 samples are sufficient to improve the current state of the art for a ResNet-50 architecture.
We argue that results with adapted statistics should be included whenever reporting scores in corruption benchmarks and other out-of-distribution generalization settings.
\end{abstract}
\section{Introduction}\label{sec:introduction}
Deep neural networks (DNNs) are known to perform well in the independent and identically distributed (\emph{i.i.d.}{}) setting when the test and training data are sampled from the same distribution.
However, for many applications this assumption does not hold.
In medical imaging, X-ray images or histology slides will differ from the training data if different acquisition systems are being used.
In quality assessment, the images might differ from the training data if lighting conditions change or if dirt particles accumulate on the camera.
Autonomous cars may face rare weather conditions like sandstorms or big hailstones.
While human vision is quite robust to those deviations~\citep{geirhos2018noise}, modern machine vision models are often sensitive to such image corruptions.
We argue that current evaluations of model robustness underestimate performance in many (but not all) real-world scenarios.
So far, popular image corruption benchmarks like ImageNet-C [IN-C;~\citealp{hendrycks2018benchmarking}] focus only on ad hoc scenarios in which the tested model has zero prior knowledge about the corruptions it encounters during test time, even if it encounters the same corruption multiple times.
In the example of medical images or quality assurance, the image corruptions do not change from sample to sample but are continuously present over a potentially large number of samples.
Similarly, autonomous cars will face the same weather condition over a continuous stream of inputs during the same sand- or hailstorm.
These (unlabeled) observations can allow recognition models to adapt to the change in the input distribution.
Such unsupervised adaptation mechanisms are studied in the field of domain adaptation (DA), which is concerned with adapting models trained on one domain (the source, here clean images) to another for which only unlabeled samples exist (the target, here the corrupted images).
Tools and methods from domain adaptation are thus directly applicable to increase model robustness against common corruptions, but so far no results on popular benchmarks have been reported.
The overall goal of this work is to encourage stronger interactions
between the currently disjoint fields of domain adaptation and robustness towards common corruptions.
We here focus on one popular technique in DA, namely adapting batch normalization [BN;~\citealp{ioffe2015batch}] statistics~\citep{schneider2018multi, cariucci2017autodial, DBLP:journals/corr/LiWSLH16}.
In computer vision, BN is a popular technique for speeding up training and is present in almost all current state-of-the-art image recognition models.
BN estimates the statistics of activations for the training dataset and uses them to normalize intermediate activations in the network.
By design, activation statistics obtained during training time do not reflect the statistics of the test distribution when testing in out-of-distribution settings like corrupted images.
We investigate and corroborate the hypothesis that high-level distributional shifts from clean to corrupted images largely manifest themselves in a difference of first and second order moments in the internal representations of a deep network, which can be mitigated by adapting BN statistics, i.e. by estimating the BN statistics on the corrupted images.
We demonstrate that this simple adaptation alone can greatly increase recognition performance on corrupted images.
Our contributions can be summarized as follows:
\begin{itemize}
\item We suggest to augment current benchmarks for common corruptions with two additional performance metrics that measure robustness after partial and full unsupervised adaptation to the corrupted images.
\item We draw connections to domain adaptation and show that even adapting to a single corrupted sample improves the baseline performance of a ResNet-50 model trained on IN from \num{76.7}\% mCE to \num{71.4}\%.
Robustness increases with more samples for adaptation and converges to a mCE of \num{62.2}\%.
\item We show that the robustness of a variety of vanilla models trained on ImageNet [IN;~\citealp{russakovsky2015imagenet, deng2009imagenet}] substantially increases after adaptation, sometimes approaching the current state-of-the-art performance on IN-C without adaptation.
\item Similarly, we show that the robustness of state-of-the-art ResNet-50 models on IN-C consistently increases when adapted statistics are used.
We surpass the best non-adapted model (\num{52.3}\% mCE) by almost \num{7}\% points.
\item We show results on several popular image datasets and discuss both the generality and limitations of our approach.
\item We demonstrate that the performance degradation of a non-adapted model can be well predicted from the Wasserstein distance between the source and target statistics. We propose a simple theoretical model for bounding the Wasserstein distance based on the adaptation parameters.
\end{itemize}
\section{Measuring robustness against common corruptions}
The ImageNet-C benchmark~\citep{hendrycks2018benchmarking} consists of 15 test corruptions and four hold-out corruptions which are applied with five different severity levels to the \num{50000} test images of the \textsc{ilsvrc2012} subset of ImageNet~\citep{deng2009imagenet}.
During evaluation, model responses are assumed to be conditioned only on single samples, and are not allowed to adapt to e.g. a batch of samples from the same corruption.
We call this the ad hoc or non-adaptive scenario.
The main performance metric on IN-C is the mean corruption error (mCE), which is obtained by normalizing the model's top-1 errors with the top-1 errors of AlexNet~\citep{krizhevsky2012imagenet} across the $C = 15$ test corruptions and $S = 5$ severities (cf.~\citealp{hendrycks2018benchmarking}):
\begin{equation}
\text{mCE(model)} = \frac{1}{C} \sum_{c = 1}^{C}
\frac{\sum_{s=1}^{S} \text{err}^\text{model}_{c,s}}{\sum_{s=1}^{S} \text{err}_{c,s}^\text{AlexNet}}.
\end{equation}
Note that mCE reflects only one possible averaging scheme over the IN-C corruption types.
We additionally report the overall top-1 accuracies and report results for all individual corruptions in the supplementary material and the project repository.
In many application scenarios, this ad hoc evaluation is too restrictive.
Instead, often many unlabeled samples with similar corruptions are available, which can allow models to adapt to the shifted data distribution.
To reflect such scenarios, we propose to also benchmark the robustness of adapted models.
To this end, we split the \num{50000} validation samples with the same corruption and severity into batches with $n$ samples each and allow the model to condition its responses on the complete batch of images.
We then compute mCE and top-1 accuracy in the usual way.
We consider three scenarios:
In the \emph{ad hoc} scenario, we set $n = 1$ which is the typically considered setting.
In the \emph{full adaptation} scenario, we set $n = \num{50000}$, meaning the model may adapt to the full set of unlabeled samples with the same corruption type before evaluation.
In the \emph{partial adaptation} scenario, we set $n = 8$ to test how efficiently models can adapt to a relatively small number of unlabeled samples.
\section{Correcting Batch Normalization statistics as a strong baseline for reducing covariate shift induced by common corruptions}\label{sec:methods}
We propose to use a well-known tool from domain adaptation---adapting batch normalization statistics~\citep{cariucci2017autodial,DBLP:journals/corr/LiWSLH16}---as a simple baseline to increase robustness against image corruptions in the adaptive evaluation scenarios.
IN trained models typically make use of batch normalization [BN;~\citealp{ioffe2015batch}] for faster convergence and improved stability during training.
Within a BN layer, first and second order statistics $\mu_c, \sigma^2_c$ of the activation tensors $\mathbf{z}_c$ are estimated across the spatial dimensions and samples for each feature map $c$.
The activations are then normalized by subtracting the mean $\mu_c$ and dividing by $\sigma^2_c$.
During training, $\mu_c$ and $\sigma^2_c$ are estimated \emph{per batch}.
During evaluation, $\mu_c$ and $\sigma^2_c$ are estimated \emph{over the whole training dataset}, typically using exponential averaging~\citep{paszke2017automatic}.
Using the BN statistics obtained during training for testing makes the model decisions deterministic but is also problematic if the input distribution changes.
If the activation statistics $\mu_c, \sigma^2_c$ change for samples from the test domain, then the activations of feature map $c$ are no longer normalized to zero mean and unit variance, breaking a crucial assumption that all downstream layers depend on.
Mathematically, this \emph{covariate shift}\footnote{Note that our notion of internal covariate shift differs from previous work~\citep{ioffe2015batch,santurkar2018does}:
In \emph{i.i.d.}{} training settings,~\citet{ioffe2015batch} hypothesized that covariate shift introduced by changing lower layers in the network is reduced by BN, explaining the empirical success of the method.
We do not provide evidence for this line of research in this work: Instead, we focus on the covariate shift introduced (by design) in datasets such as IN-C, and provide evidence for the hypothesis that high-level domain shifts in the input partly manifests in shifts and scaling of \emph{internal} activations.}
can be formalized as follows:
\begin{defi}[Covariate Shift, cf.~\citealp{sugiyama2012machine,schoelkopf2012causal}]
There exists covariate shift between a source distribution with density $p_s: \mathcal X \times \mathcal Y \rightarrow \mathbb{R}^+$ and a target distribution with density $p_t: \mathcal X \times \mathcal Y \rightarrow \mathbb{R}^+$, written as $p_s(\mathbf{x},y) = p_s(\mathbf{x}) p_s(y|\mathbf{x})$ and $p_t(\mathbf{x},y) = p_t(\mathbf{x}) p_t(y|\mathbf{x})$, if $p_s(y|\mathbf{x}) = p_t(y|\mathbf{x})$ and $p_s(\mathbf{x}) \neq p_t(\mathbf{x})$
where $y \in \mathcal{Y}$ denotes the class label.
\end{defi}
\paragraph{Removal of covariate shift.}
If covariate shift (Def. 1) only
causes differences in the first and second order moments of the feature activations $\mathbf{z} = f(\mathbf{x})$, it can be removed by applying normalization:
\begin{equation}
p\left(\frac{f(\mathbf{x}) - \mathbb{E}_s[f(\mathbf{x})]}{\sqrt{\mathbb{V}_s[f(\mathbf{x})]}} \Big| \mathbf{x} \right) p_s(\mathbf{x}) \approx
p\left(\frac{f(\mathbf{x}) - \mathbb{E}_t[f(\mathbf{x})]}{\sqrt{\mathbb{V}_t[f(\mathbf{x})]}} \Big| \mathbf{x} \right) p_t(\mathbf{x}).
\end{equation}
Reducing the covariate shift in models with batch normalization is particularly straightforward: it suffices to estimate the BN statistics $\mu_t, \sigma^2_t$ on (unlabeled) samples from the test data available for adaptation.
If the number of available samples $n$ is too small, the estimated statistics would be too unreliable.
We therefore leverage the statistics $\mu_s, \sigma^2_s$ already computed on the training dataset as a prior and infer the test statistics for each test batch as follows,
\begin{equation}
\Bar{\mu} = \frac{N}{N + n} \mu_s + \frac{n}{N + n} \mu_t, \quad
\Bar{\sigma}^2 = \frac{N}{N + n} \sigma_s^2 + \frac{n}{N + n} \sigma_t^2.
\label{eq:Nn}
\end{equation}
The hyperparameter $N$ controls the trade-off between source and estimated target statistics and has the intuitive interpretation of a \emph{pseudo sample size} (p.~117,~\citealp{bishop}) for samples from the training set.
The case $N \to \infty$ ignores the test set statistics and is equivalent to the standard ad hoc scenario while $N = 0$ ignores the training statistics.
Supported by empirical and theoretical results (see results section and appendix), we suggest using $N\in[8,128]$ for practical applications with small $n < 32$.
\section{Experimental Setup}\label{sec:experiments}
\paragraph{Models.}%
We consider a large range of models (cf. Table~2, \textsection\ref{app:experimental-setup},\ref{apx:imagenet_model_descriptions})
and evaluate pre-trained variants of
DenseNet~\citep{densenet},
GoogLeNet~\citep{googlenet},
Inception and GoogLeNet~\citep{inception},
MNASnet~\citep{mnasnet},
MobileNet~\citep{mobilenet},
ResNet~\citep{he2016resnet},
ResNeXt~\citep{xie2017aggregated},
ShuffleNet~\citep{shufflenet},
VGG~\citep{vgg}
and Wide Residual Network [WRN,~\citealp{wideresnet}] from the \texttt{torchvision} library~\citep{10.1145/1873951.1874254}.
All models are trained on the \textsc{ilsvrc2012} subset of IN comprised of 1.2 million images in the training and a total of \num{1000} classes~\citep{russakovsky2015imagenet, deng2009imagenet}.
We also consider a ResNeXt-101 variant pre-trained on a 3.5 billion image dataset and then fine-tuned on the IN training set~\citep{mahajan2018exploring}.
We evaluate 3 models from the SimCLRv2 framework~\citep{chen2020big}.
We additionally evaluate the four leading methods from the ImageNet-C leaderboard, namely Stylized ImageNet training [SIN;~\citealp{geirhos2018imagenettrained}],
adversarial noise training [ANT;~\citealp{Rusak2020IncreasingTR}] as well as
a combination of ANT and SIN~\citep{Rusak2020IncreasingTR},
optimized data augmentation using AutoAugment [AugMix;~\citealp{hendrycks2019augmix,autoaugment2019}] and
Assemble Net~\citep{lee2020compounding}.
For partial adaptation, we choose $N \in \{2^{0},\cdots,2^{10}\}$ and select the optimal value on the holdout corruption mCE.
\paragraph{Datasets.}%
ImageNet-C [IN-C;~\citealp{hendrycks2018benchmarking}] is comprised of corrupted versions of the \num{50000} images in the IN validation set.
The dataset offers five severities per corruption type, for a total of 15 ``test'' and 4 ``holdout'' corruptions.
ImageNet-A [IN-A;~\citealp{DBLP:journals/corr/abs-1907-07174}] consists of unmodified real-world images which yield chance level classification performance in IN trained ResNet-50 models.
ImageNet-V2 [IN-V2;~\citealp{recht2019imagenet}] aims to mimic the test distribution of IN, with slight differences in image selection strategies.
ObjectNet [ON{};~\citealp{barbu2019objectnet}] is a test set containing \num{50000} images like IN organized in 313 object classes with 109 unambiguously overlapping IN classes.
ImageNet-R [IN-R;~\citealp{hendrycks2020many}] contains \num{30000} images with various artistic renditions of 200 classes of the original IN dataset.
Additional information on the used models and datasets can be found in \textsection\ref{app:experimental-setup}.
For IN, we resize all images to $256\times 256$px and take the center $224 \times 224$px crop.
For IN-C, images are already cropped.
We also center and re-scale the color values with $\mu_{RGB} = [0.485, 0.456, 0.406]$ and $\sigma=[0.229, 0.224, 0.225]$.
\section{Results}\label{sec:results}
\input{figs/04-results/ablation-bsz}
\paragraph{Adaptation boosts robustness of a vanilla trained ResNet-50 model.}\label{sec:result-vanilla}
We consider the pre-trained ResNet-50 architecture from the \texttt{torchvision} library and adapt the running mean and variance on all corruptions and severities of IN-C for different batch sizes.
The results are displayed in Fig.~\ref{fig:low-resource} where different line styles of the green lines show the number of pseudo-samples $N$ indicating the influence of the prior given by the training statistics.
With $N=16$, we see that even adapting to a single sample can suffice to increase robustness, suggesting that even the ad hoc evaluation scenario can benefit from adaptation.
If the training statistics are not used as a prior ($N=0$), then it takes around 8 samples to surpass the performance of the non-adapted baseline model (76.7\% mCE).
After around 16 to 32 samples, the performance quickly converges to 62.2\% mCE, considerably improving the baseline result.
These results highlight the practical applicability of batch norm adaptation in basically all application scenarios, independent of the number of available test samples.
\paragraph{Adaptation consistently improves corruption robustness across IN trained models.}
\label{sec:result-torchvision}
To evaluate the interaction between architecture and BN adaptation, we evaluate all 25 pre-trained models in the \texttt{torchvision} package and visualize the results in Fig.~\ref{fig:torchvision}.
All models are evaluated with $N=0$ and $n=\num{2000}$.
We group models into different families based on their architecture and observe consistent improvements in mCE for all of these families, typically on the order of 10\% points.
We observe that in both evaluation modes, DenseNets~\citep{densenet} exhibit higher corruption robustness despite having a comparable or even smaller number of trainable parameters than ResNets which are usually considered as the relevant baseline architecture.
A take-away from this study is thus that model architecture alone plays a significant role for corruption robustness and the ResNet architecture might not be the optimal choice for practical applications.
\paragraph{Adaptation yields new state of the art on IN-C for robust models.}\label{sec:result-sota}
\input{figs/04-results/sota_new}
We now investigate if BN adaptation also improves the most robust models on IN-C.
The results are displayed in Table~\ref{tbl:sota}.
All models are adapted using $n = \num{50000}$ (vanilla) or $n=\num{4096}$ (all other models) and $N=0$.
The performance of all models is considerably higher whenever the BN statistics are adapted.
The DeepAugment+AugMix reaches a new state of the art on IN-C for a ResNet-50 architecture of 45.4\% mCE.
Evaluating the performance of AugMix over the number of samples for adaptation (Fig.~\ref{fig:low-resource}, we find that as little as eight samples are sufficient to improve over AssembleNet~\citep{lee2020compounding}, the current state-of-the-art ResNet-50 model on IN-C without adaptation.
We have included additional results in \textsection\ref{app:results}.
\section{Analysis and Ablation Studies}
\label{ref:analysis}
\paragraph{Severity of covariate shift correlates with performance degradation.}
\input{figs/05-analysis/wasserstein-detailed}
The relationship between the performance degradation on IN-C and the covariate shift suggests an unsupervised way of estimating the classification performance of a model on a new corruption.
Taking the normalized Wasserstein distance (cf. \textsection\ref{app:distance}) between the statistics of the source and target domains\footnote{For computing the Wasserstein metric we make the simplifying assumption that the empirical mean and covariances fully parametrize the respective distributions.} computed on all samples with the same corruption and severity and averaged across all network layers, we find a correlation with the top-1 error (Fig.~\ref{fig:wasserstein-analysis-detailed} \emph{i--iii}) of both non-adapted (\emph{i}) and fully adapted model (\emph{ii}) on IN-C corruptions.
Within single corruption categories (noise, blur, weather, and digital), the relationship between top-1 error and Wasserstein distance is particularly striking: using linear regression, the top-1 accuracy of hold-out corruptions can be estimated with around 1--2\% absolute mean deviation (cf. \textsection\ref{app:error_prediction}) within a corruption, and with around 5--15\% absolute mean deviation when the estimate is computed on the holdout corruption of each category (see Fig.~\ref{fig:wasserstein-analysis-detailed}, typically, a systematic offset remains).
In Fig.~\ref{fig:wasserstein-analysis-detailed}\emph{(iv--v}), we display the Wasserstein distance across individual layers and observe that the covariate shift is particularly present in early and late downsampling layers of the ResNet-50.
\paragraph{Large scale pre-training alleviates the need for adaptation.}
\label{sec:ablation-wsl}
Computer vision models based on the ResNeXt architecture~\citep{xie2017aggregated} pretrained on a much larger dataset comprised of \num{3.5e9} Instagram images (IG-3.5B) achieve a 45.7\% mCE on IN-C~\citep{mahajan2018exploring,orhan2019robustness}.
We re-evaluate these models with our proposed paradigm and summarize the results in Table~\ref{tbl:wsl}.
While we see improvements for the small model pre-trained on IN, these improvements vanish once the model is trained on the full IG-3.5B dataset.
This observation also holds for the largest model, suggesting that training on very large datasets might alleviate the need for covariate shift adaptation.
\input{figs/05-analysis/wsl-effnet-table-and-plot-new}
\paragraph{Group Normalization and Fixup Initialization performs better than non-adapted batch norm models, but worse than batch norm with covariate shift adaptation.}
So far, we considered image classification models with BN layers and concluded that using training dataset statistics in BN generally degrades model performance in out-of-distribution evaluation settings.
We now consider models trained without BN and study the impact on corruption robustness, similar to~\citet{galloway2019batch}.
First, using Fixup initialization \citep{zhang2019fixup} alleviates the need for BN layers.
We train a ResNet-50 model on IN for 100 epochs to obtain a top-1 error of \num{24.23}\% and top-5 error of \num{ 7.56}\% (compared to 27.6\% reported by~\citet{zhang2019fixup} with shorter training, and the \num{23.87}\% obtained by our ResNet-50 baseline trained with BN).
The model obtains an IN-C mCE of \num{72.00}\% compared to \num{76.7}\% mCE of the vanilla ResNet-50 model and \num{62.2}\% mCE of our adapted ResNet-50 model (cf. Table~\ref{tbl:fixup}).
Additionally, we train a ResNet-101 and a ResNet-152 with Fixup initialization with similar results.
Second, GroupNorm [GN;~\citealp{wu2018group}] has been proposed as a batch-size independent normalization technique.
We train a ResNet-50, a ResNet-101 and a ResNet-152 architecture for 100 epochs and evaluate them on IN-C and find results very similar to Fixup.
\paragraph{Results on other datasets: IN-A, IN-V2, ObjectNet, IN-R}
\input{figs/05-analysis/in-in2}
\input{figs/05-analysis/object_net_mixed_inc_imagenet_r}
We use $N=0$ and vary $n$ in all ablation studies in this subsection.
The technique does not work for the case of ``natural adversarial examples'' of IN-A~\citep{DBLP:journals/corr/abs-1907-07174} and the error rate stays above 99\%, suggesting that the
covariate shift introduced in IN-A by design is more severe compared to the covariate shift of IN-C and can not be corrected by merely calculating the correct BN statistics.
We are not able to increase performance neither on IN nor on IN-V2, since in these datasets, no domain shift is present by design (see Fig.~\ref{fig:imagenet-v2}).
For ON, the performance increases slightly when computing statistics on more than 64 samples.
In Table~\ref{tbl:objectnet} (first and second column), we observe that the GroupNorm and Fixup models perform better than our BN adaptation scheme:
while there is a dataset shift in ON compared to IN, BN adaptation is only helpful for \textit{systematic} shifts across multiple inputs and this assumption is violated on ON.
As a control experiment, we sample a dataset ``Mixed IN-C'' where we shuffle the corruptions and severities.
In Table~\ref{tbl:objectnet} (third and fourth column), we now observe that BN adaptation expectedly no longer improves performance.
On IN-R, we achieve better results for the adapted model compared to the non-adapted model as well as the GroupNorm and Fixup models, see Table~\ref{tbl:objectnet} (last column). Additionally, on IN-R, we decrease the top-1 error for a wide range of models through adaptation (see Table~\ref{tbl:imagenet-r}).
For IN-R, we observe performance improvements for the vanilla trained ResNet50 when using a sample size of larger than 32 samples for calculating the statistics (Fig.~\ref{fig:imagenet-v2}, right-most plot)
\paragraph{A model for correcting covariate shift effects.}
\begin{wrapfigure}[14]{r}{0.3\textwidth}
\begin{center}
\includegraphics[width=0.3\textwidth]{figs/05-analysis/theory-figs.pdf}
\end{center}
\caption{%
The bound suggests small optimal $N$ for most parameters (i) and qualitatively explains our empirical observation (ii).
\label{eq:theory}
}
\end{wrapfigure}
We evaluate how the batch size for estimating the statistics at test time affects the performance on IN, IN-V2, ON{} and IN-R in Fig.~\ref{fig:imagenet-v2}.
As expected, for IN the adaptation to test time statistics converges to the performance of the train time statistics in the limit of large batch sizes, see Fig.~\ref{fig:imagenet-v2} middle.
For IN-V2, we find similar results, see Fig.~\ref{fig:imagenet-v2} left.
This observation shows that (\textit{i}) there is no systematic covariate shift between the IN train set and the IN-V2 validation set that could be corrected by using the correct statistics and (\textit{ii}) is further evidence for the \emph{i.i.d.}{} setting pursued by the authors of IN-V2.
In case of ON{} (Fig.~\ref{fig:imagenet-v2} right), we see slight improvements when using a batch size bigger than 128.
Choosing the number of pseudo-samples $N$ offers an intuitive trade-off between estimating accurate target statistics (low $N$) and relying on the source statistics (large $N$).
We propose a simple model to investigate optimal choices for $N$, disregarding all special structure of DNNs, and focusing on the statistical error introduced by estimating $\hat{\mu}_t$ and $\hat{\sigma}^2_t$ from a limited number of samples $n$.
To this end, we estimate upper ($U$) and lower ($L$) bounds of the expected squared Wasserstein distance $W_2^2$ as a function of $N$ and the covariate shift which provides good empirical fits between the estimated $W$ and empirical performance for ResNet-50 for different $N$ (Fig.~\ref{eq:theory}; bottom row).
Choosing $N$ such that $L$ or $U$ are minimized (Fig.~\ref{eq:theory}; example in top row) qualitatively matches the values we find, see \textsection\ref{app:analytical} for all details.
\begin{prop}[Bounds on the expected value of the Wasserstein distance between target and combined estimated target and source statistics]
We denote the source statistics as $\mu_s,\sigma_s^2$, the true target statistics as $\mu_t,\sigma^2_t$ and the biased estimates of the target statistics as $\hat{\mu}_t,\hat{\sigma}_t^2$.
For normalization, we take a convex combination of the source statistics and estimated target statistics as discussed in Eq.~\ref{eq:Nn}.
At a confidence level $1-\alpha$, the expectation value of the Wasserstein distance
$W_2^2(\Bar{\mu}, \Bar{\sigma}, \mu_t, \sigma_t)$ between ideal and estimated target statistics w.r.t. to the distribution of sample mean $\hat{\mu}_t$ and sample variance $\hat{\sigma}^2_t$ is bounded from above and below
with $L \leq \mathds{E}[W^2_2] \leq U$, where
\begin{equation*}
\begin{aligned}
L &= \; \left( \sigma_t - \sqrt{ \frac{N}{N+n} \sigma^2_s + \frac{n-1}{N+n} \sigma^2_t } \right)^2 + \frac{N^2}{(N+n)^2} \left(\mu_t - \mu_s \right)^2 + \frac{n}{(N+n)^2} \sigma_t^2 \\
U &= \; L + \sigma^5_t \frac{(n-1)}{2(N+n)^2} \left( \frac{N}{N+n} \sigma^2_s + \frac{1}{N+n} \chi^2_{1-\alpha/2, n-1} \sigma^2_t \right)^{-3/2}
\end{aligned}
\end{equation*}
The quantity $\chi^2_{1-\alpha/2, n-1}$ denotes the left tail value of a chi square distribution with $n-1$ degrees of freedom, defined as $P\left(X \leq \chi^2_{1-\alpha/2, n-1}\right) = \alpha/2 \text{ for } X \sim \chi^2_{n-1}$.
\emph{Proof:} See Appendix \textsection\ref{app:analytical}.
\end{prop}
\section{Related Work}\label{sec:related-work}
The IN-C benchmark~\citep{hendrycks2018benchmarking} has been extended to MNIST~\citep{mu2019mnist}, several object detection datasets~\citep{michaelis2019benchmarking} and image segmentation~\citep{Kamann2019BenchmarkingTR} reflecting the interest of the robustness community.
Most proposals for improving robustness involve special training protocols, requiring time and additional resources.
This includes data augmentation like Gaussian noise~\citep{ford2019adversarial}, optimized mixtures of data augmentations in conjunction with a consistency loss~\citep{hendrycks2019augmix}, training on stylized images~\citep{geirhos2018imagenettrained, michaelis2019benchmarking, Mikoajczyk2018DataAF} or against adversarial noise distributions~\citep{Rusak2020IncreasingTR}.
Other approaches tweak the architecture, e.g. by adding shift-equivariance with an anti-aliasing module,~\citep{zhang2019making} or assemble different training techniques~\citep{lee2020compounding}.
Unsupervised domain adaptation (DA) is a form of transductive inference where additional information about the test dataset is used to adapt a model to the test distribution.
Adapting feature statistics was proposed by~\citet{sun2017correlation} and follow up work evaluated the performance of adapting BN parameters in unsupervised \citep{DBLP:journals/corr/LiWSLH16, cariucci2017autodial} and supervised DA settings \citep{schneider2018multi}.
As an application example in medical imaging,~\citet{bug2017context} show that adaptive normalization is useful for removing domain shifts on histopathological data.
More involved methods for DA include self-supervised domain adaptation on single examples \citep{sun2019test} and pseudo-labeling \citet{french2017self}.
\citet{xie2020self} achieve the state of the art on IN-C with pseudo-labeling.
In work concurrent to ours, \citet{wang2020fully} also show BN adaptation results on IN-C. They also perform experiments on CIFAR10-C and CIFAR100-C and explore other domain adaptation techniques.
Robustness scores obtained by adversarial training can be improved when separate BN or GroupNorm layers are used for clean and adversarial images \citep{xie2020intriguing}.
The expressive power of adapting only affine BN parameters BN parameters was shown in multi-task \citep{rebuffi2017learning} and DA contexts \citep{schneider2018multi} and holds even for fine-tuning randomly initialized ResNets \citep{frankle2020training}.
Concurrent work shows additional evidence that BN adaptation yields increased performance on ImageNet-C~\citep{nado2020evaluating}.
\section{Discussion and Conclusion}\label{sec:discussion}
We showed that reducing covariate shift induced by common image corruptions improves the robustness of computer vision models trained with BN layers, typically by 10--15\% points (mCE) on IN-C.
Current state-of-the-art models on IN-C can benefit from adaptation, sometimes drastically like AugMix (\num{-14}\% points mCE).
This observation underlines that current benchmark results on IN-C underestimate the corruption robustness that can be reached in many application scenarios where additional (unlabeled) samples are available for adaptation.
Robustness against common corruptions improves even if models are adapted only to a single sample, suggesting that BN adaptation should always be used whenever we expect machine vision algorithms to encounter out-of-domain samples. Most further improvements can be reaped by adapting to 32 to 64 samples, after which additional improvements are minor.
Our empirical results suggest that the performance degradation on corrupted images can mostly be explained by the difference in feature-wise first and second order moments.
While this might sound trivial, the performance could also degrade because models mostly extract features susceptible to common corruptions~\citep{geirhos2020shortcut}, which could not be fixed without substantially adapting the model weights.
The fact that model robustness increases after correcting the BN statistics suggests that the features upon which the models rely on are still present in the corrupted images. The opposite is true in other out-of-domain datasets like IN-A or ObjectNet where our simple adaptation scheme does not substantially improve performance, suggesting that here the main problem is in the features that models have learned to use for prediction.
Batch Norm itself is not the reason why models are susceptible to common corruptions. While alternatives like Group Normalization and Fixup initialization slightly increase robustness, the adapted BN models are still substantially more robust.
This suggests that non-BN models still experience an internal covariate shift on corrupted images, but one that is now absorbed by the model parameters instead of being exposed in the BN layers, making it harder to fix.
Large-scale pre-training on orders of magnitude more data (like IG-3.5B) can remove the first- and second-order covariate shift between clean and corrupted image samples, at least partially explaining why models trained with weakly supervised training~\citep{mahajan2018exploring} generalize so well to IN-C.
Current corruption benchmarks emphasize ad hoc scenarios and thus focus and bias future research efforts on these constraints.
Unfortunately, the ad hoc scenario does not accurately reflect the information available in many machine vision applications like classifiers in medical computer vision or visual quality inspection algorithms, which typically encounter a similar corruption continuously and could benefit from adaptation.
This work is meant to spark more research in this direction by suggesting two suitable evaluation metrics---which we strongly suggest to include in all future evaluations on IN-C---as well as by highlighting the potential that even a fairly simple adaptation mechanism can have for increasing model robustness.
We envision future work to also adopt and evaluate more powerful domain adaptation methods on IN-C and to develop new adaptation methods specifically designed to increase robustness against common corruptions.
\section*{Broader Impact}
The primary goal of this paper is to increase the robustness of machine vision models against common corruptions and to spur further progress in this area. Increasing the robustness of machine vision systems can enhance their reliability and safety, which can potentially contribute to a large range of use cases including autonomous driving, manufacturing automation, surveillance systems, health care and others. Each of these uses may have a broad range of societal implications: autonomous driving can increase mobility of the elderly and enhance safety, but could also enable more autonomous weapon systems. Manufacturing automation can increase resource efficiency and reduce costs for goods, but may also increase societal tension through job losses or increase consumption and thus waste. Of particular concern (besides surveillance) is the use of generative vision models for spreading misinformation or for creating an information environment of uncertainty and mistrust.
We encourage further work to understand the limitations of machine vision models in out-of-distribution generalization settings. More robust models carry the potential risk of automation bias, i.e., an undue trust in vision models. However, even if models are robust to common corruptions, they might still quickly fail on slightly different perturbations like surface reflections. Understanding under what conditions model decisions can be deemed reliable or not is still an open research question that deserves further attention.
\begin{ack}
We thank
Julian Bitterwolf,
Roland S. Zimmermann,
Lukas Schott,
Mackenzie W. Mathis,
Alexander Mathis,
Asim Iqbal,
David Klindt,
Robert Geirhos,
other members of the Bethge and Mathis labs
and four anonymous reviewers
for helpful suggestions for improving our manuscript and providing ideas for additional ablation studies.
We thank the International Max Planck Research School for Intelligent Systems (IMPRS-IS) for supporting E.R. and St.S.;
St.S. acknowledges his membership in the European Laboratory for Learning and Intelligent Systems (ELLIS) PhD program.
This work was supported by the German Federal Ministry of Education and Research (BMBF) through the Tübingen AI Center (FKZ: 01IS18039A),
by the Deutsche Forschungsgemeinschaft (DFG) in the priority program 1835 under grant BR2321/5-2
and by SFB 1233, Robust Vision: Inference Principles and Neural Mechanisms (TP3), project number: 276693517.
The authors declare no conflicts of interests.
\end{ack}
\medskip
\small
\bibliographystyle{unsrtnat}
|
\section{Introduction and overview of the main results} \setcounter{equation}{0}
A function $f: \mathbb{R}^n\to \mathbb{R}$ is said to be log-concave if $-\log f: \mathbb{R}^n\to \mathbb{R}\cup\{+\infty\}$ is a convex function. It is well-known that the log-concave functions behave in many aspects akin to convex bodies (i.e., compact convex sets with nonempty interiors, and the set of all convex bodies is denoted by $\mathscr K^n$), which is considered as an analytic ``lifting" of geometry of convex bodies. In convex geometry, the Brunn-Minkowski theory (and its extensions) for convex bodies encompasses a large and growing range of fundamental results on the algebraic and geometric properties of convex bodies, and hence to study the parallel algebraic and geometric properties of log-concave functions is of great significance and in great demand. Recent years have witnessed that many results and notions in the Brunn-Minkowski theory (and its extensions) for convex bodies have found their functional analogues, including but not limited to the functional Blaschke-Santal\'{o} type inequality and its inverse \cite{AKM04,AS15,Ball88-2,BBF14,FM07,FM08, KM05}, (John, Lutwak-Yang-Zhang, and L\"owner) ellipsoids for log-concave functions \cite{AMJV18, FZ18,LSW19jga}, Rogers-Shephard type inequality and its reverse for log-concave functions \cite{Alo19, AAMJV19, AMJV16,Col06}, the affine surface areas
for log-concave functions \cite{AKSW12,CFGLSW16, CW14, CW15, CY16,LSW19}, the variation and Minkowski type problems related to log-concave functions \cite{CF13,CK15, Kla14, San16}, and (isoperimetric) inequalities related to log-concave functions \cite{ABM20, ABM, AFS20, Bob99, BL76, Lin17, Rot13}. Other contributions include e.g., \cite{ AM09,AS19, MR13} among others.
The celebrated Pr\'ekopa-Leindler inequality \eqref{PLineq} for log-concave functions has a formula similar to the famous dimensional-free Brunn-Minkowski inequality for convex bodies (see \eqref{p-BM-2-2} for $p=1$). Note that
the Pr\'ekopa-Leindler inequality indeed works for more general functions, see \cite{Gar02, Lei72,Pre71,Pre73,Pre75}.
The key ingredients in the Pr\'ekopa-Leindler inequality for log-concave functions are the total mass of a log-concave function $f$ given by \begin{equation}\label{totalmass}
J(f)=\int_{\mathbb{R}^n} f(x) \, dx,
\end{equation} where $\,dx$ denotes the Lebesgue measure in $\mathbb{R}^n$ and the Asplund sum $f\oplus g$ of two log-concave functions $f$ and $g$ defined by \begin{equation}\label{A-sum-6-23} f\oplus g(x)=\sup_{x=x_1+x_2} f(x_1)g(x_2), \end{equation} which is again a log-concave function. Likewise, the volume and the Minkowski addition for convex bodies (which is equivalent to \eqref{Lp-addition} for $p=1$ for convex bodies) are the key ingredients in the Brunn-Minkowski inequality, and their combination naturally results in an elegant variational formula (i.e., \eqref{variation-convex bodies} for $p=1$). Such a variational formula defines a crucial concept in convex geometry: the surface area measure $S_K$ for convex body $K\in \mathscr K ^n$. Notably, the Minkowski problem \cite{Minkowski1987, Minkowski1903},
aiming to characterize the surface area measure for convex bodies, can be solved by using the Euler-Lagrange equation based on the variational formula (i.e., \eqref{variation-convex bodies} for $p=1$).
Lifting to the functional setting, the first variation of the total mass of log-concave functions with respect to the Asplund sum has been established by Colesanti and Fragal\`a in their groundbreaking work \cite{CF13}, and from there one can see that things become way more complicated for log-concave functions. The challenge partially comes from the facts that the total mass is defined on $\mathbb{R}^n$ instead of compact sets ($S^{n-1}$ for convex bodies) and the family of log-concave functions contains too many ``unfavourable" log-concave functions (for example, those log-concave functions which are not smooth enough and/or whose domains are not nice enough).
The variational formula of the total mass for log-concave functions with respect to the Asplund sum produces a measure for log-concave functions which is similar to $S_K$ for $K\in \mathscr K^n$; such a measure was named as the surface area measure for a log-concave function in \cite{CF13} (which is also known as the moment measure of convex functions in \cite{CK15}). That is, for a log-concave function $f=e^{-\psi}$
with $0 <J(f) < \infty$, the moment measure of $\psi$ or the surface area measure of $f$, denoted by $\mu(f, \cdot)$, was defined as the push-forward measure of $e^{-\psi}\,dx$ under $\nabla\psi$ where $\nabla \psi$ denotes the gradient of $\psi$. That is, \begin{equation*}
\int_{\mathbb{R}^n} g(y) \, d\mu(f, y) = \int_{\mathbb{R}^n} g(\nabla \psi(x)) \, e^{-\psi(x)}\, dx
\end{equation*} for every Borel function $g$ such that
$g \in L^1(\mu(f, \cdot) )$ or $g$ is non-negative. In their groundbreaking work \cite{CK15}, Cordero-Erausquin and Klartag studied the Minkowski type problem for log-concave functions in order to characterize the moment measure.
Note that the Minkowski problem for log-concave functions in its full formulation was independently posed by Colesanti and Fragal\`a in \cite{CF13}, and finding solutions to this problem (in its full formulation) seems to be very important but highly intractable. Under certain conditions on log-concave functions (in particular, the essential continuity for the negative of their logarithms), Cordero-Erausquin and Klartag were able to solve the Minkowski problem for log-concave functions (see \cite[Proposition 1]{CK15} for necessity and \cite[Theorem 2]{CK15} for sufficiency and uniqueness). Santambrogio in \cite{San16} presented a solution to the Minkowski problem for log-concave functions by using the technique of the optimal mass transportation.
Although many concepts for log-concave functions have been studied, what brings into our attention is the missing of the ``$L_p$ addition" of log-concave functions and the corresponding variational formula. These shall be analogous to the $L_p$ addition \cite{Firey} and the related variational formula for convex bodies \cite{Lut93} which are the crucial elements in the rapidly developing $L_p$ Brunn-Minkowski theory for convex bodies. In the same manner as the Brunn-Minkowski theory for convex bodies, the $L_p$ theory for $p>1$ was developed based on the brilliant variational formula \eqref{variation-convex bodies} by Lutwak in his influential work \cite{Lut93}. Again, the variational formula \eqref{variation-convex bodies} relies on the combination of the $L_p$ addition for convex bodies
and the volume, which can be extended to all $0\neq p\in \mathbb{R}$ (see e.g., \cite{ZHY2016}). Naturally, the $L_p$ surface area measure $h_K^{1-p} S_{K}$ for a convex body $K\in \mathscr K^n_{(o)}$ (see Section \ref{Preliminaries} for notations) can be defined from \eqref{variation-convex bodies}, and consequently the $L_p$ Minkowski problem \cite{Lut93} characterizing the $L_p$ surface area measure for convex bodies
can be posed (see Problem \ref{lp-min-6-26} for its precise statement). The $L_p$ Minkowski problem is a milestone in convex geometry and receives enormous attention in many areas of mathematics. Contributions to the $L_p$ Minkowski problem include \cite{BLYZ13, chen,CW06,HLW15,HLYZ05,JLW15,JLZ16, LW13, Lut93, LO95, LYZ04, Uma03,Zhu15,Zhu15jfa,Zhu17} among others.
Our main goals in this paper are to establish a basic framework of the functional ``lifting" of the $L_p$ Brunn-Minkowski theory for convex bodies, and thence extend the scheme for log-concave functions to an $L_p$ setting. Based on the $L_p$ Asplund sum of log-concave functions which naturally generalizes the Asplund sum, for $p>1$, we establish a variational formula related to the total mass, define the $L_p$ surface area measure for log-concave functions, and study the $L_p$ Minkowski problem which characterizes the $L_p$ surface area measure for log-concave functions. It is our hope that these contributions provide some useful tools in the development of geometry of log-concave functions.
More specifically, Section \ref{p-sum of log-concave function} dedicates to study the properties for the $L_p$ Asplund sum of log-concave functions. Let $f=e^{-\varphi}\in \mathscr A_0$ and $g=e^{-\psi}\in \mathscr A_0$ (see Sections \ref{Preliminaries} and \ref{p-sum of log-concave function} for notations). For $p>1$ and $\alpha, \beta>0$, the $L_p$ Asplund sum $\alpha\cdot_p f\,\oplus_p\,\beta\cdot_p g$ may be formulated by
\begin{equation} \label{p-addition-6-23} \alpha\cdot_p f\,\oplus_p\,\beta\cdot_p g:=e^{-\phi^*} \ \ \ \mathrm{with} \ \ \ \phi:=\big[\alpha(\varphi^*)^p+\beta(\psi^*)^p\big]^{\frac{1}{p}},\end{equation} where $\varphi^*$ denotes the Fenchel conjugate of $\varphi$ given in \eqref{def-dual-1}. This definition is meaningful as it can be seen in Proposition \ref{close-Lp-A-sum} that $\alpha\cdot_p f\,\oplus_p\,\beta\cdot_p g\in \mathscr A_0$ for $p> 1$. We establish the following Pr\'ekopa-Leindler type inequality in Theorem \ref{prekopainequality}: {\em for $f, g\in\mathscr A_0$, $\lambda\in (0,1)$ and $p>1$, it holds that $$
J\big((1-\lambda)\cdot_p f\oplus_{p} \lambda\cdot_p g\big)\geq J(f)^{1-\lambda}J(g)^{\lambda} $$
with equality if and only if $f=g$ on $\mathbb{R}^n$.} We also prove the log-concavity of $J(f\oplus _p t\cdot_p g)$ on $t\in (0, \infty)$ and $J\big((1-t)\cdot_p f\oplus_p t\cdot_p g\big)$ on $t\in (0, 1)$ in Corollary \ref{concave-PL-1}.
In Section \ref{sectoion-4-6-23}, we define $\delta J_{p}(f, g)$ for $p>1$, the first variation of the total mass at $f\in \mathscr A_0$ along $g\in \mathscr A_0$ with respect to the $L_{p}$ Asplund sum, by \[\delta J_{p}(f,g):=\underset{t\rightarrow0^+}{\lim}\frac{J(f\oplus_{p} t\cdot_p g)-J(f)}{t},\] whenever the limit exists. To find an explicit formula to describe $\delta J_{p}(f,g)$ for all $f, g\in \mathscr A_0$ seems to be intractable and even impractical. However, when $f=g\in \mathscr A_0$ such that $J(f)>0$, we are able to prove in Lemma \ref{specialcaseoffg} that, for $p>1,$ $$
\delta J_{p}(f,f)= \frac{n}{p} \int_{\mathbb{R}^n} f \, dx + \frac{1}{p} \int_{\mathbb{R}^n} f \log f \, dx,
$$ which involves the entropy of $f$ (see \eqref{entropu-p>1} for more details). Our main result in this section is a Minkowski type inequality for $\delta J_{p}(f, g)$ established in Theorem \ref{minkowskiinequality}, namely, {\em
for $f \in \mathscr A_0$ and $g \in\mathscr A_0$ such that $J(f)>0$ and $J(g)>0$, one has, for $p> 1,$
$$
\delta J_{p}(f,g)\geq \delta J_{p}(f,f)+J(f)\log\Big(\frac{J(g)}{J(f)}\Big)
$$
with equality if and only if $f=g$.} This Minkowski type inequality is applied to obtain a unique determination of log-concave functions in Corollary \ref{cormink-uniq}.
Although, in general, an explicit formula for $\delta J_{p}(f,g)$ is not easy to get, under certain conditions on $f$ and $g$ (which differs from the one for $f=g$ in Section \ref{sectoion-4-6-23}), such a formula can be obtained. This result consists of our main contribution in Section \ref{section-varition-516} and is proved in Theorem \ref{variationformula}. Roughly speaking, it asserts that {\em for $p>1$, \begin{align}\label{p-derivative-6-23}
\delta J_{p}(f,g) =\frac{1}{p}\int_{\mathbb{R}^n}(\psi^*(y))^{p}(\varphi^*(y))^{1-p}\,d\mu(f, y) \end{align}
holds for $f=e^{-\varphi}$ and $g=e^{-\psi}$ smooth enough such that $(\varphi^*)^p-c (\psi^*)^p$ is a convex function for some constant $c>0$, and the function $\frac{1}{\alpha} (\varphi^*)^{\alpha}$ (understood as $\log \varphi^*$ when $\alpha=0$) is convex for some $\alpha\in [0, 1)$.} Please see more precise statements in Theorem \ref{variationformula} and Corollary \ref{verify-condition-1}. The proof of Theorem \ref{variationformula} is very technical and involves a lot of rather complicated analysis on the (first and second order) differentiability and the limit of the partial derivatives of the function $-\log (f\oplus_{p} t\cdot_p g)$. We would like to mention that, even in the case $p=1$ where $\phi$ defined in \eqref{p-addition-6-23} is linear, such analysis (for the differentiability and related limits) already exhibits its complexity as one can see in \cite{CF13}. The nonlinearity of $\phi$ defined in \eqref{p-addition-6-23} for $p>1$ does bring extra difficulty and makes our analysis even more complicated. In particular, additional conditions on $\varphi$ are needed, such as, \eqref{compatible-1} or those given in Corollary \ref{verify-condition-1}.
Our last contribution is the study of the $L_p$ Minkowski problem for log-concave functions, which will be presented in Section \ref{functional p Minkowski problem}. In fact, formula \eqref{p-derivative-6-23} suggests a natural way to define the $L_p$ surface area measure for the log-concave function $f=e^{-\varphi}$, denoted as $\mu_p(f, \cdot)$ and defined on $\Omega_{\varphi^*}=\{y\in \mathbb{R}^n: 0<\varphi^*(y)<+\infty\}$ (this set is always assumed to be nonempty). Under certain conditions on a given log-concave function $f=e^{-\varphi}$, the measure $\mu_p(f, \cdot)$ for $p\in \mathbb{R}$ is given by $$ \int_{\Omega_{\varphi^*}}g(y)\, d\mu_p (f, y)=\int_{\{x\in \mathbb{R}^n:\ \nabla \varphi(x)\in \Omega_{\varphi^*}\}}g(\nabla \varphi(x))(\varphi^*(\nabla \varphi(x)))^{1-p} e^{-\varphi(x)} \,dx $$ for every Borel function $g$ such that $g \in L ^ 1 (\mu_p (f,\cdot))$ or $g$ is non-negative. We pose the following $L_p$ Minkowski problem for log-concave functions: {\em for $p\in \mathbb{R}$, find the necessary and/or sufficient conditions on a finite nonzero Borel measure $\nu$ defined on $\mathbb{R}^n$ so that $\nu=\tau \mu_{p}(f,\cdot)$ holds for some log-concave function $f$ and $\tau\in\mathbb{R}$.} We provide a solution to this problem under the following mild conditions on $\nu$: $\nu$ is not supported in a lower-dimensional subspace of $\mathbb{R}^n$, $\nu(M_{\nu}\setminus L)>0$ holds for any bounded convex set $L\subset\mathbb{R}^n$ where $M_{\nu}$ is the interior of the convex hull of the support of $\nu$, and $ \int_{\mathbb{R}^n} |x|^p\, d\nu(x)<\infty.$ Note that these conditions for $\nu $ are all natural; see Section \ref{functional p Minkowski problem} for more details. Our solution to the $L_p$ Minkowski problem for log-concave functions is proved in Theorem \ref{solution-lp-mink-6-2}. That is, roughly speaking, {\em if $\nu$ is an even measure satistying the above conditions, for $p>1$, there exists an even log-concave function $f=e^{-\varphi}$, such that, \begin{eqnarray*}
\nu =\frac{\int_{\Omega_{\varphi^*}} \,d\nu(y)} {\int_{\Omega_{\varphi^*}} \,d \mu_{p}(f,y) } \mu_{p}(f, \cdot)\ \ \ \mathrm{on} \ \ \Omega_{\varphi^*}.
\end{eqnarray*} }
Finally, we would like to comment that if $\nu$ admits a density function with respect to the Lebesgue measure, say $\,d\nu(y)=h(y)\,dy$, and $f=e^{-\varphi}$ with the convex function $\varphi$ smooth enough, then finding a solution to the $L_p$ Minkowski problem for log-concave functions requires to search for a (smooth enough) convex function $\varphi$ satisfying the following Monge-Amp\`{e}re equation:
\begin{align*} h(y) = \tau \varphi^*(y)^{1-p} e^{\varphi^*(y)-\langle y, \nabla\varphi^*{y}\rangle} \det(\nabla^2\varphi^*(y)), \ \ \ \mathrm{for} \ y\in \Omega_{\varphi^*}, \end{align*} where $\tau\in\mathbb{R}$ is a constant and $\det(\nabla^2\varphi^*(y))$ denotes the determinant of the Hessian matrix of $\varphi^*$ at $y\in \mathbb{R}^n$.
\section{Preliminaries and Notations}\label{Preliminaries} \setcounter{equation}{0}
This section provides preliminaries and notations required for (log-concave) functions and convex bodies. More details can be found in \cite{BV10, Roc70,Sch14}. The paper \cite{Col17} by Colesanti is also an excellent source to learn various important topics related to log-concave functions.
Let $\mathbb{N}$ be the set of natural numbers and $n\in \mathbb{N}$ such that $n\geq 1$. In the $n$-dimensional Euclidean space $\mathbb{R}^n$, $\langle x , y \rangle$ stands for the standard inner product of $x,y\in\mathbb{R}^n$ and $|x|$ for the Euclidean norm of $x\in \mathbb{R}^n$. The origin of $\mathbb{R}^n$ is denoted by $o$. A set $E\subset\mathbb{R}^n$ is said to be origin-symmetric if $-x\in E$ for any $x\in E$. Let $B _r = \{ x \in \mathbb{R} ^n \, :\, |x| \leq r \}$ be the origin-symmetric ball with radius $r$. By $S^{n-1}$ we mean the unit sphere, i.e., the boundary of $B_1$. We also use $V(E), \mathrm{int}(E), \overline{E}$ and $\partial E$ to denote the volume, interior, closure and boundary of $E\subset \mathbb{R}^n$, respectively. By $\mathcal{H}^k$, we mean the $k$-dimensional Hausdorff measure.
A set $K\subset \mathbb{R}^n$ is said to be a convex body if $K$ is a compact convex set with nonempty interior. The family of convex bodies is denoted by $\mathscr K ^n$. With $\mathscr K_{(o)}^n$, we mean the collection of convex bodies containing $o$ in their interiors. We say $K\in \mathscr K_{(o)}^n$ is a dilation of $L\in \mathscr K_{(o)}^n$ if there exists a $\lambda>0$ such that $K=\lambda L$.
For any $K \in \mathscr K^n$, the {\it support function} of $K$, $h _K: \mathbb{R}^n\rightarrow \mathbb{R}$, which is a sublinear function on $\mathbb{R}^n$ and can be used to determine $K$ uniquely, takes the following form:
\begin{equation}\label{support-1}
h _K (x)= \sup_{ y \in K} \langle x, y \rangle \ \ \ \mathrm{for }\ x \in \mathbb{R} ^n.
\end{equation}
Let $\alpha, \beta>0$ be two constants. For $K, L\in \mathscr K_{(o)}^n$, define $\alpha \cdot_p K+_p\beta\cdot_p L$ for $p\geq 1$ (see e.g., \cite{Firey, Lut93, Sch14}) to be the convex body whose support function is given by \begin{equation}\label{Lp-addition} h_{\alpha \cdot_p K+_p\beta\cdot_p L}=\big(\alpha h_K^p+\beta h_L^p\big)^{\frac{1}{p}}. \end{equation} Of course, $\alpha \cdot_p K+_p\beta\cdot_p L$ in \eqref{Lp-addition} reduces to the Minkowski addition of convex bodies when $p=1$. In \cite{Lut93}, Lutwak proved the following $L_p$ Brunn-Minkowski inequality for $p>1$: \begin{eqnarray}\label{p-BM}
V(K+_pL)^{\frac{p}{n}}\geq V(K)^{\frac{p}{n}}+V( L)^{\frac{p}{n}}
\end{eqnarray}
with equality if and only if $K$ and $L$ are the dilation of each other. Moreover, Lutwak in \cite{Lut93} also established the following remark variational formula: for $K, L\in \mathscr K_{(o)}^n$ and $p> 1$, one has \begin{equation}\label{variation-convex bodies}
\frac{p}{n}\cdot \frac{d}{dt}V(K+_p t\cdot _p L)\Big|_{t=0^+}=\frac{1}{n}\cdot \int_{S^{n-1}}h_L^p(u)h_K^{1-p}(u)\, d S_{K}(u):=V_p(K,L).
\end{equation} Here $S_K$ is the surface area measure of $K$ on $S^{n-1}$, that is, $S_K(\eta)=\mathcal{H}^{n-1}(\nu_K^{-1}(\eta))$ for any Borel set $\eta\subset S^{n-1}$, where $\nu_K^{-1}$ is the reverse Gauss map of $K$. Note that both \eqref{p-BM} and \eqref{variation-convex bodies} works for $p=1$ as well, and inequality \eqref{p-BM} becomes the classical Brunn-Minkowski inequality (in this case, equality holds if and only if $K$ and $L$ agree up to a translation and a dilation). We refer \cite{Sch14} and references therein for more details.
This elegant formula \eqref{variation-convex bodies} introduces two important geometric objects: the $L_p$ mixed volume $V_p(K, L)$ for $K, L\in \mathscr K_{(o)}^n$ and the $L_p$ surface area measure of $K\in \mathscr K_{(o)}^n$ given by $\, d S_p(K, \cdot)=h_K^{1-p}\, d S_{K}.$ Associated to the $L_p$ mixed volume is the following $L_p$ Minkowski inequality for $p> 1$:
\begin{eqnarray*}\label{p-M}
V_{p}(K, L)\geq V(K)^{\frac{n-p}{n}}V(L)^{\frac{p}{n}},
\end{eqnarray*}
with equality if and only if $K$ and $L$ are the dilation of each other \cite{Lut93}. Related to the $L_p$ surface area measure is the following well-known $L_p$ Minkowski problem \cite{Lut93}.
\begin{problem}[The $L_p$ Minkowski problem for convex bodies]\label{lp-min-6-26} Let $p\in \mathbb{R}$. What are the necessary and sufficient conditions on a nonzero finite Borel measure $\mu$ on $S^{n-1}$ such that there exists a convex body $K
\subset \mathbb{R}^n$ satisfying $S_p(K, \cdot)=\mu$?\end{problem}
Throughout the paper, let $\varphi : \mathbb{R}^n\to\mathbb{R}\cup\{+\infty\}$ be a convex function, that is, $$\varphi(\lambda x+(1-\lambda)y)\leq \lambda \varphi(x)+(1-\lambda) \varphi(y)$$ holds for any $\lambda\in (0, 1)$ and $x, y\in \mathrm{dom} (\varphi)$, where $
\mathrm{dom} (\varphi)=\{x\in\mathbb{R}^n\,:\, \varphi(x)<+\infty\}$ denotes the domain of $\varphi.$ Denote by $\mathcal C$ the set of all convex functions from $\mathbb{R}^n$ to $\mathbb{R}\cup\{+\infty\}$. Clearly, $\mathrm{dom} (\varphi)$ is a convex set for any $\varphi \in \mathcal C$. A convex function $\varphi\in \mathcal C$ is said to be {\it proper} if $\mathrm{dom} (\varphi) \neq \emptyset$. By $\mathcal C^1(E)$ we mean the set of all functions $\varphi\in \mathcal C$ such that $\varphi$ is continuously differentiable on $E\subseteq \mathrm{int} ({\mathrm {dom}} (\varphi))$. Similarly, by $\varphi\in \mathcal C^ 2_+(E)$ we mean that $\varphi\in \mathcal C$ is twice continuously differentiable and its Hessian matrix is positive definite on $E\subseteq \mathrm{int} ({\mathrm {dom}} (\varphi))$.
Let $\varphi: \mathbb{R}^n\to\mathbb{R}\cup\{+\infty\}$ be a function (not necessary a convex function) such that $\varphi(x)<\infty$ for some $x\in \mathbb{R}^n$. The {\it Fenchel conjugate} of $\varphi$ is defined by:
\begin{equation}
\varphi^*(y)=\sup_{x\in\mathbb{R}^n}\big\{ \langle x , y \rangle -\varphi(x)\big\} \ \ \ \mathrm{for \ all} \ y \in \mathbb{R} ^n.\label{def-dual-1}
\end{equation}
For $\alpha>0$, \eqref{def-dual-1} yields that, for all $y\in \mathbb{R}^n$, \begin{equation} \label{const-mul-dual} (\alpha \varphi)^*(y)=\sup_{x\in\mathbb{R}^n}\big\{ \langle x , y \rangle -\alpha \varphi(x)\big\}=\alpha \sup_{x\in\mathbb{R}^n}\big\{ \langle x , y/\alpha \rangle -\varphi(x)\big\}=\alpha \varphi^*(y/\alpha). \end{equation} Clearly, $\varphi^{**}\leq \varphi$, $\varphi^*\in \mathcal C$, and $\varphi^*$ is always lower semi-continuous. It can also be checked that if $\varphi\leq \psi$, then $\varphi^*\geq \psi^*$. Moreover, if $\varphi\in \mathcal C$ is proper,
the Fenchel-Moreau theorem (or Fenchel biconjugation theorem) \cite{BV10, Roc70} yields that $(\varphi^*)^*=\varphi$ if and only if $\varphi$ is lower semi-continuous. It is obvious by \eqref{def-dual-1} that $\varphi ^*(0) = - \inf_{x\in \mathbb{R}^n} \varphi(x)$ for $\varphi\in \mathcal C$, and hence $\varphi^*$ is proper if $\inf _{x\in \mathbb{R}^n} \varphi(x) > - \infty$. Furthermore, if $\varphi\in \mathcal C$ is proper, then $\varphi^*(y)>-\infty$ for all $y\in\mathbb{R}^n$. We also let $e^{-\infty}=0$ by default (unless otherwise stated).
Denote by $\nabla \varphi(x)$ the gradient of $\varphi$ at $x\in {\mathrm {dom}}(\varphi)$ if $\varphi$ is differentiable at $x$. A convex function $\varphi\in \mathcal C$ may not be differentiable at each point in its domain, however it is always continuous on $\mathbb{R}^n\setminus \partial ({\mathrm {dom}} (\varphi))$ and is differentiable almost everywhere in $\mathrm{int} ({\mathrm {dom}} (\varphi))$. It can be checked that if $\varphi$ is differentiable at $x$, then the supremum in \eqref{def-dual-1} achieves its maximum if $y=\nabla \varphi(x)$. Hence, \begin{equation}
\varphi^*(\nabla \varphi(x))= \langle x , \nabla \varphi(x) \rangle -\varphi(x). \label{def-dual-2}
\end{equation} This relation will be used often in later context. Denote by $\nabla^2 \varphi$ the Hessian matrix of $\varphi$.
Throughout the paper, we shall consider two classes of functions $\mathscr L$ and $\mathscr A$ given by
\begin{eqnarray*}
\mathscr L &=& \Big\{ \varphi\in \mathcal C \ \big| \ \varphi \ \mathrm{is\ proper \ such\ that}\ \lim_{|x|\rightarrow+\infty}\varphi(x)=+\infty \Big\},
\\
\mathscr A& =&\Big\{f: \mathbb{R}^n\to\mathbb{R} \ \big| \ f=e^{-\varphi}, \ \ \varphi\in\mathscr L \Big\}.
\end{eqnarray*} It has been proved in \cite[Lemma 2.5]{CF13} that if $\varphi \in \mathscr L$, then for all $x\in\mathbb{R}^n$,
\begin{equation}\label{claimab}
\varphi(x)\geq a|x|+b,
\end{equation} where $a>0$ and $b$ are two constants, and
$\varphi ^*$ is proper and satisfies $\varphi ^*(y) > - \infty$ for all $y \in \mathbb{R}^n$.
A function $f: \mathbb{R}^n\rightarrow \mathbb{R}$ is said to be log-concave if $f=e^{-\varphi}$ for some convex function $\varphi\in \mathcal C$. We are interested in log-concave functions such that $0<J(f)<\infty$, where $J(f)$ is the total mass of $f$ defined in \eqref{totalmass}. It can be easily checked (see e.g.~\cite[p.~3835]{CK15}) that, for $f=e^{-\varphi}$ a log-concave function, $0<J(f)<\infty$ is equivalent to the facts that $\mathrm{dom}(\varphi)$ is not contained in a hyperplane and $\lim_{|x|\rightarrow+\infty}\varphi(x)=+\infty$ (in particular, $f\in \mathscr A$). It follows from \eqref{def-dual-1} and \eqref{claimab} that, if $f=e^{-\varphi}$ with $\varphi\in \mathcal C$ is non-degenerate (i.e., not vanishing on a set with positive Lebesgue measure), then $J(f)<\infty$ if and only if $o\in \mathrm{int}({\mathrm {dom}}(\varphi^*))$.
The surface area measure for a log-concave function $f$ (or the {\it moment measure} of $-\log f$ as in \cite{CK15}) plays important roles in applications. The following definition follows from \cite[Definition 1]{CK15} (see also \cite[Definition 4.1]{CF13}).
\begin{definition} \label{def-moment-measure} Let $f=e^{-\varphi}$
be a log-concave function. The surface area measure of $f$ (or the moment measure of $\varphi$), denoted by $\mu(f, \cdot)$, is the Borel measure on $\mathbb{R}^n$ such that $\mu (f,\cdot)$ is the push-forward measure of $e^{-\varphi(x)}\,dx$ under $\nabla \varphi$, that is,
\begin{equation}\label{moment-form-1} \int_{\mathbb{R}^n}g(y)\, d\mu(f, y)=\int_{\mathbb{R}^n}g(\nabla \varphi(x))e^{-\varphi(x)} \,dx, \end{equation} for every Borel function $g$ such that $g \in L ^ 1 (\mu (f,\cdot))$ or $g$ is non-negative. \end{definition}
It has been proved that $\varphi ^* \in L ^ 1 (\mu (f,\cdot))$, if $f = e ^ {-\varphi} \in \mathscr A$ with $\varphi^*\geq 0$ in \cite[Lemma 4.12]{CF13} or
if $0<J(e^{-\varphi})<\infty$ in \cite[Proposition 7]{CK15}, i.e., \begin{equation} -\infty<\int_{\mathbb{R}^n} \varphi^* (y) \, d\mu(f, y)=\int_{\mathbb{R}^n} \varphi^* (\nabla \varphi(x) ) \, e^{-\varphi(x)} \, dx < + \infty.\label{finitezza-1} \end{equation} Let $(\varphi \alpha)(x) =\alpha \varphi(\frac{x}{\alpha})$ for $\alpha>0$ and $\varphi 0 =\mathbf{I}_{\{0\}}$
where $\mathbf{I}_E$ is the indicatrix function of $E$ (i.e., $\mathbf{I}_E(x)=0$ if $x\in E$ and $\mathbf{I}_E=+\infty$ if $x\notin E$). By \eqref{const-mul-dual}, one sees $(\alpha \varphi)^*= \varphi^*\alpha $ for $\alpha>0$.
It can be easily checked from \eqref{A-sum-6-23} that, for real $\alpha, \beta>0$ and two log-concave functions $f=e^{-\varphi}$ and $g=e^{-\psi}$,
the Asplund sum of $\alpha \cdot f\oplus \beta\cdot g$ may also be formulated by \begin{equation}\label{A-sum-1}
\alpha\cdot f \oplus \beta\cdot g= e^{-\varphi \alpha \Box \psi \beta},
\end{equation}
where $\varphi \Box \psi$ is the {\it infimal convolution} of $\varphi$ and $\psi$ defined by $$
\varphi \Box \psi (x) = \inf _{y\in \mathbb{R}^n} \big \{ \varphi (x-y) + \psi (y) \big \}, \ \ \mathrm{for\ any}\ x\in \mathbb{R}^n.$$ It can be easily checked that
\begin{equation}\label{inf-conv-1}
(\varphi \alpha \Box \psi \beta)^* =\alpha \varphi^*+\beta \psi^*.
\end{equation}
The {\it Pr\'ekopa-Leindler inequality} implies that (see, e.g., \cite[Remark 3.3]{CF13}), for any integrable log-concave functions $f, g$ and $0<\lambda<1$, one has
\begin{eqnarray}\label{PLineq}
J\big((1-\lambda)\cdot f\oplus \lambda\cdot g\big) \geq J(f)^{1-\lambda} J(g)^{\lambda},
\end{eqnarray} with equality if and only if there exists $x_0\in\mathbb{R}^n$ such that $f(x)=g(x-x_0)$ (see, e.g., \cite{Dub77,Gar02}).
\section{The $L_{p}$ Asplund sum of log-concave functions
}\label{p-sum of log-concave function} \setcounter{equation}{0}
This section dedicates to the study of properties of the $L_{p}$ Asplund sum of log-concave functions. In particular, a Pr\'ekopa-Leindler type inequality related to the $L_{p}$ Asplund sum will be proved. We focus on $\mathscr L_0\subset \mathscr L$ and $\mathscr A_0\subset \mathscr A$, where $\mathscr A_0=\big\{ e^{-\varphi}: \varphi\in\mathscr L_0 \big\}$ with \begin{align*}
\mathscr L_0&=\Big\{ \varphi \in\mathscr L: \ \varphi \ \text{is\ non-negative and lower semi-continuous such that}\ \varphi(o)=0 \Big\}.
\end{align*} Clearly, if $\varphi\in \mathscr L_0$, then $\varphi$ has its minimum attained at the origin $o$.
From \eqref{def-dual-1}, one sees that if $\varphi(o)=0$, then \begin{equation}
\varphi^*(y)=\sup_{x\in\mathbb{R}^n}\big\{ \langle x , y \rangle -\varphi(x)\big\}\geq \langle o , y \rangle -\varphi(o)=0, \ \ \mathrm{for\ all\ } y \in \mathbb{R} ^n. \label{non-negative-6-26}\end{equation} Moreover, $\varphi^*(o)=0$ as \begin{equation}\label{dual-at-origin}
\varphi^*(o)=\sup_{x\in\mathbb{R}^n} -\varphi(x) =-\inf_{x\in \mathbb{R}^n} \varphi(x)=0. \end{equation}
The definition for the $L_{p}$ Asplund sum of log-concave functions is given below, with concentration on $p>1$.
\begin{definition}\label{def-lp-A-sum}
Let $f=e^{-\varphi}\in \mathscr A_0$ and $g=e^{-\psi}\in\mathscr A_0$. For $p>1$, define $ f\oplus_{p} g$, the $L_p$ Asplund sum of $f$ and $g$, by $ f\oplus_{p} g=e^{-\varphi \square_{p}\psi}$, where
\begin{equation}\label{prodottoL-no constants}
\varphi\square_{p} \psi=\left[
\big((\varphi^*)^p+(\psi^*)^p\big)^{\frac{1}{p}}\right]^*.
\end{equation}
\end{definition}
For convenience, if $\alpha>0$, let
\[(\varphi \cdot_p \alpha) (x)=( \varphi \alpha^{\frac{1}{p}})(x)=\alpha^{\frac{1}{p}} \varphi (\alpha^{-\frac{1}{p}} x) \ \ \mathrm{for\ any}\ x\in \mathbb{R}^n.\] For $\alpha,\beta> 0$, $p\geq1$, $f=e^{-\varphi}\in \mathscr A_0$ and $g=e^{-\psi}\in\mathscr A_0$, let $\alpha\cdot_p f\oplus_{p}\beta\cdot_p g=e^{-\varphi\cdot_p\alpha \square_{p}\psi\cdot_p\beta},$ where
\begin{equation}\label{prodottoL}
\varphi\cdot_p\alpha \square_{p}\psi\cdot_p\beta=\left[
\big(\alpha(\varphi^*)^p+\beta(\psi^*)^p\big)^{\frac{1}{p}}\right]^*.
\end{equation}
By \eqref{inf-conv-1} and \eqref{prodottoL}, $\alpha\cdot_p f\oplus_{p}\beta\cdot_p g$ for $p=1$ reduces to the Asplund sum $\alpha\cdot f\oplus\beta\cdot g$.
The following result asserts that the $L_p$ Asplund sum of log-concave functions is closed in $\mathscr A_0.$
\begin{proposition}\label{close-Lp-A-sum}
Let $f=e^{-\varphi}\in \mathscr A_0$ and $g=e^{-\psi}\in\mathscr A_0$. For $p>1$ and $\alpha, \beta>0$, let \begin{equation}\label{relation-p-sq-1} \phi=\big[\alpha(\varphi^*)^p+\beta(\psi^*)^p\big]^{\frac{1}{p}}.\end{equation} Then $\phi= \big(\varphi\cdot_p\alpha \square_{p}\psi\cdot_p\beta\big)^*$
and $\alpha\cdot_p f\,\oplus_p\,\beta\cdot_p g=e^{-\phi^*}\in\mathscr A_0.$
\end{proposition}
\begin{proof} Let $f=e^{-\varphi}\in \mathscr A_0$ and $g=e^{-\psi}\in\mathscr A_0$ which imply $\varphi, \psi \in\mathscr L_0$. By \eqref{non-negative-6-26} and \eqref{dual-at-origin}, $\varphi^*\geq 0$ and $\psi^*\geq 0$ with $\varphi^*(o)=\psi^*(o)=0$. Hence, $ \varphi\cdot_p\alpha \square_{p}\psi\cdot_p\beta$ for $p> 1$ given in \eqref{prodottoL} is well-defined and $\varphi\cdot_p\alpha \square_{p}\psi\cdot_p\beta\in \mathcal C$ is lower semi-continuous. Moreover, $\phi$ defined by \eqref{relation-p-sq-1} is non-negative and $\phi(o)=0$ is the minimum of $\phi$. This further yields that $ (\varphi\cdot_p\alpha \square_{p}\psi\cdot_p\beta)(o)= 0$ by \eqref{dual-at-origin}, and hence is proper. Note that $\varphi^*, \psi^*\in \mathcal C$ are lower semi-continuous. Consequently, for $p>1$, $\alpha, \beta>0$, $\phi$ is lower semi-continuous. Moreover, $\phi\in \mathcal C.$ To this end, for all $\lambda\in (0, 1)$ and $x, y\in {\mathrm {dom}}(\phi)$, the Minkowski inequality for norms yields that \begin{eqnarray*} \phi(\lambda x+(1-\lambda)y)\!\!
&=&\! \! \Big[\alpha\big(\varphi^*(\lambda x+(1-\lambda)y) \big)^p+\beta\big(\psi^*(\lambda x+(1-\lambda)y)\big)^p \Big]^{\frac{1}{p}}\\ \!\!&\leq&\!\! \Big[\alpha\big(\lambda \varphi^*(x)+(1-\lambda)\varphi^*(y) \big)^p+\beta\big(\lambda \psi^*(x)+(1-\lambda)\psi^*(y) \big)^p \Big]^{\frac{1}{p}} \\ \!\! &\leq& \!\! \lambda \big[\alpha(\varphi^*)^p+\beta(\psi^*)^p\big]^{\frac{1}{p}}(x) + (1-\lambda) \big[\alpha(\varphi^*)^p+\beta(\psi^*)^p\big]^{\frac{1}{p}}(y)\\ \!\! &=&\!\! \lambda \phi(x) + (1-\lambda) \phi(y).\end{eqnarray*} According to \eqref{prodottoL} and the Fenchel-Moreau theorem \cite{BV10, Roc70}, one sees that
\begin{equation} \label{biconjugate-1} \big(\varphi\cdot_p \alpha\square_{p}\psi\cdot_p \beta\big)^*=\big(\phi^*\big)^*=\phi \ \ \ \mathrm{and}\ \ \ \alpha\cdot_p f\,\oplus_p\,\beta\cdot_p g=e^{-\phi^*}.
\end{equation}
To claim that $\alpha\cdot_p f\,\oplus_p\,\beta\cdot_p g\in\mathscr A_0$ for $p>1$, it is enough to show that $\phi^*\in\mathscr L_0$. To this end, we only need to claim that $\lim_{|x|\rightarrow+\infty} \phi^*(x)=+\infty$ as we already showed that $\phi^*$ is proper and lower semi-continuous. Inequality \eqref{claimab} implies that for all $x\in \mathbb{R}^n$, $\varphi(x)\geq a|x|+b$ and $\psi(x)\geq a'|x|+b'$ for some $a, a'>0$ and $b,b'\in \mathbb{R}$.
Let $c=\min\{a,a'\}>0$ and $d=\min\{b, b'\}$. Hence $\varphi(x)\geq c|x|+d$ and $\psi(x)\geq c|x|+d$ for all $x\in \mathbb{R}^n$. This further yields that $\varphi ^*\leq (c|x|+d)^*$ and $\psi^*\leq (c|x|+d)^*$. Accordingly, due to $p>1$ and $\alpha, \beta>0$, one has
$$\phi=\big[\alpha(\varphi^*)^p+\beta(\psi^*)^p\big]^{\frac{1}{p}} \leq \big[\alpha((c|x|+d)^*)^p+\beta((c|x|+d)^*)^p\big]^{\frac{1}{p}}= (\alpha+\beta)^{\frac{1}{p}} (c|x|+d)^*.$$ It follows from \eqref{def-dual-1} and \eqref{const-mul-dual} that for all $y\in \mathbb{R}^n$,
\begin{equation*} \phi^*(y)\geq \big[(\alpha+\beta)^{\frac{1}{p}} (c|x|+d)^*\big]^*(y) = c|y| +d (\alpha+\beta)^{\frac{1}{p}}.\end{equation*}
In particular, $\lim_{|y|\rightarrow+\infty} \phi^*(y)=+\infty$. This concludes that $\alpha\cdot_p f\,\oplus_p\,\beta\cdot_p g=e^{-\phi^*}\in\mathscr A_0.$
\end{proof}
Let $\chi_E$ be the characteristic function of $E$, i.e., $\chi_E(x)=1$ if $x\in E$ and $\chi_E(x)=0$ if $x\notin E$.
It is well-known that
if $K\in \mathscr K_{(o)}^n$,
then $\chi_K$ is a log-concave function with $\mathbf{I}_K=-\log(\chi_K)\in \mathscr L_0$ and hence $\chi_K\in \mathscr A_0$. By \eqref{support-1} and \eqref{def-dual-1}, one sees that $ (\mathbf{I}_K)^*=h_K$ and $(h_K)^*=\mathbf{I}_K$. Together with \eqref{Lp-addition}, for two convex bodies $K, L\in \mathscr K_{(o)}^n$, one has, for $p\geq 1$, \begin{equation*}
(\mathbf{I}_K)\cdot_p \alpha \square_{p} (\mathbf{I}_L)\cdot_p \beta=\Big(
\big(\alpha((\mathbf{I}_K)^*)^p+\beta((\mathbf{I}_L)^*)^p\big)^{\frac{1}{p}}\Big)^*=\Big(
\big(\alpha h_K^p+\beta h_L^p\big)^{\frac{1}{p}}\Big)^*=\mathbf{I}_{\alpha \cdot_p K+_p\beta\cdot_p L}.
\end{equation*}
Therefore, for $\alpha, \beta>0$ and $p\geq 1$, $$ \alpha\cdot_p\chi_K\oplus_p\beta\cdot_p\chi_L=e^{-(\mathbf{I}_K)\cdot_p \alpha \square_{p} (\mathbf{I}_L)\cdot_p \beta}=\chi_{\alpha \cdot_p K+_p\beta\cdot_p L}.$$ Moreover, for two convex bodies $K, L\in \mathscr K_{(o)}^n$ and for any $\alpha, \beta>0$,
\begin{eqnarray}
J(\alpha\cdot_p \chi_K\oplus_p \beta\cdot_p \chi_L)
=\int_{\mathbb{R}^n} \chi_{\alpha\cdot_p K+_p \beta\cdot_p L} (x)\,dx = V(\alpha \cdot_p K+_p \beta\cdot_p L). \label{volume-p-addition-515} \end{eqnarray} By the $L_p$ Brunn-Minkowski inequality \eqref{p-BM}, for $p\geq 1$, $\lambda\in (0, 1)$, and $K, L\in \mathscr K_{(o)}^n$, one has \begin{eqnarray}\label{p-BM-2-2}
V\big((1-\lambda)\cdot_p K+_p \lambda\cdot_p L\big)\geq \Big((1-\lambda) V(K)^{\frac{p}{n}}+\lambda V( L)^{\frac{p}{n}}\Big)^{\frac{n}{p}}\geq V(K)^{1-\lambda}V(L)^{\lambda},
\end{eqnarray} where the last inequality follows from the fact that the logarithmic function is concave. Combining \eqref{volume-p-addition-515} and \eqref{p-BM-2-2}, one easily sees that, for $p\geq 1$, $\lambda\in (0, 1)$, and $K, L\in \mathscr K_{(o)}^n$, \begin{eqnarray*} J\Big((1-\lambda)\cdot_p \chi_K\oplus_p \lambda\cdot_p \chi_L\Big)&\geq & J(\chi_K)^{1-\lambda}J(\chi_L)^{\lambda}. \end{eqnarray*} This is a Pr\'ekopa-Leindler type inequality and can be extended to all $f, g\in \mathscr A_0$. This result is proved in the following theorem, where $p>1$ is concentrated as the case $p=1$ reduces to the classical {\it Pr\'ekopa-Leindler inequality} \eqref{PLineq}.
\begin{theorem}\label{prekopainequality} Let $f=e^{-\varphi}\in \mathscr A_0$ and $g=e^{-\psi} \in\mathscr A_0$. For $\lambda\in (0,1)$ and $p>1$, it holds that
\begin{eqnarray}\label{prekopainequality1}
J\big((1-\lambda)\cdot_p f\oplus_{p} \lambda\cdot_p g\big)\geq J(f)^{1-\lambda}J(g)^{\lambda}
\end{eqnarray}
with equality if and only if $f=g$ on $\mathbb{R}^n$.
\end{theorem}
\begin{proof} Let $f=e^{-\varphi}\in \mathscr A_0$ and $g=e^{-\psi} \in\mathscr A_0$. For $\lambda\in (0,1)$ and $p>1$, let
$
(1-\lambda )\cdot_p f\oplus_{p} \lambda \cdot_p g=e^{-\phi_{\lambda}^*}.
$ According to Proposition \ref{close-Lp-A-sum}, one has $(\phi_{\lambda}^*)^*=\phi_{\lambda}$ and $\phi_{\lambda }=\big[(1-\lambda)(\varphi^*)^p+\lambda (\psi^*)^p\big]^{\frac{1}{p}}.$ Since the function $t^{\frac{1}{p}}$ is strictly concave on $t\in [0, \infty)$ when $p>1$, one has, \begin{equation}\label{p-conv-equality-515} \phi_{\lambda }=\big[(1-\lambda)(\varphi^*)^p+\lambda (\psi^*)^p\big]^{\frac{1}{p}} \geq(1-\lambda) \varphi^*+\lambda \psi^*. \end{equation} Together with \eqref{inf-conv-1}, this in turn implies that
\begin{eqnarray}\label{relationship of p and 1 infimal convolution}
\phi_{\lambda}^* \leq \big((1-\lambda) \varphi^*+\lambda \psi^* \big)^*
= \varphi (1-\lambda) \Box \psi \lambda.
\end{eqnarray} By the classical Pr\'ekopa-Leindler inequality (\ref{PLineq}), inequality \eqref{prekopainequality1} holds:
\begin{eqnarray}
J\big((1-\lambda)\cdot_p f\oplus_{p}\lambda \cdot_p g\big)= \int _{\mathbb{R}^n} e^{-\phi_{\lambda}^*(x)}\,dx \geq \int _{\mathbb{R}^n} e^{-\big(\varphi (1-\lambda) \Box \psi \lambda\big)(x)}\,dx \geq J(f)^{1-\lambda}J(g)^{\lambda}.\label{p-prekopa-l}
\end{eqnarray}
Now let us characterize the equality condition. It is obvious that equality holds if $f=g$. On the other hand, to have equality in \eqref{prekopainequality1}, equalities must occur in \eqref{p-prekopa-l}, and consequently equalities hold for the classical Pr\'ekopa-Leindler inequality (\ref{PLineq}) and for \eqref{relationship of p and 1 infimal convolution} (equivalently for \eqref{p-conv-equality-515}). The former one implies that there exists $x_0\in\mathbb{R}^n$ such that $f(x)=g(x-x_0)$, which in turn yields $\varphi(x)=\psi(x-x_0)$. Combining with the latter one and the fact that $t^{\frac{1}{p}}$ is strictly concave, one gets $\varphi^*=\psi^*$ and hence $f=g$ as desired. \end{proof}
\begin{corollary} \label{concave-PL-1} Let $f=e^{-\varphi}\in \mathscr A_0$, $g=e^{-\psi} \in\mathscr A_0$ and $p>1$. Then, $J(f\oplus _p t\cdot_p g)$ is log-concave on $t\in (0, \infty)$ and $J\big((1-t)\cdot_p f\oplus_p t\cdot_p g\big)$ is log-concave on $t\in (0, 1)$.
\end{corollary} \begin{proof}
Let $f=e^{-\varphi}\in \mathscr A_0$ and $g=e^{-\psi} \in\mathscr A_0$. For $\lambda\in (0, 1)$ and $t, s\in (0, \infty)$, let $\eta=(1-\lambda)t+\lambda s$. According to Proposition \ref{close-Lp-A-sum}, one gets $f\oplus_p \eta \cdot_p g=e^{-\phi^*_{\eta}}$ with $\phi_{\eta}=\big[(\varphi^*)^p+\eta(\psi^*)^p\big]^{\frac{1}{p}}.$
It follows from Proposition \ref{close-Lp-A-sum} and the fact that the function $t^{\frac{1}{p}}$ is concave on $t\in (0, \infty)$ for $p>1$ that
\begin{align*}
\phi_{\eta}
&= \big[\big((1-\lambda) (\varphi^*)^p+t (\psi^*)^p\big)+\lambda \big( (\varphi^*)^p+ s (\psi^*)^p\big)\big]^{\frac{1}{p}} \\
&\geq (1-\lambda) \big((\varphi^*)^p+t (\psi^*)^p\big)^{\frac{1}{p}} +\lambda \big( (\varphi^*)^p+ s (\psi^*)^p\big)^{\frac{1}{p}}\\
&= (1-\lambda) (\phi^*_{t})^*+\lambda (\phi^*_{s})^*.\end{align*}
By \eqref{A-sum-1} and \eqref{inf-conv-1}, one gets $ \phi^*_{\eta}\leq \big((1-\lambda) (\phi^*_{t})^*+\lambda (\phi^*_{s})^*\big)^*= \phi^*_{t}(1-\lambda) \Box \phi^*_{s} \lambda $
and hence, $$f\oplus_p \eta \cdot_p g=e^{-\phi^*_{\eta}}\geq (1-\lambda)\cdot \big(f\oplus_p t \cdot_p g\big)\oplus \lambda\cdot \big(f\oplus_p s \cdot_p g\big).$$ By the
classical Pr\'ekopa-Leindler inequality \eqref{PLineq}, one gets the desired log-concavity for $J(f\oplus _p t\cdot_p g)$ on $t\in (0, \infty)$, that is,
$$\log J(f\oplus_p \eta \cdot_p g) \geq (1-\lambda) \log J(f\oplus_p t \cdot_p g)+\lambda\log J(f\oplus_p s \cdot_p g).$$
The log-concavity for $J\big((1-t)\cdot_p f\oplus _p t\cdot_p g\big)$ on $t\in (0, 1)$ is a direct result of Theorem \ref{prekopainequality}.
\end{proof}
\section{An $L_p$ Minkowski type inequality} \label{sectoion-4-6-23} \setcounter{equation}{0}
In this section, we propose a definition for $\delta J_{p}(f, g)$, the first variation of the total mass at $f$ along $g$ with respect to the $L_{p}$ Asplund sum. A Minkowski type inequality for $\delta J_{p}(f, g)$ will be established. Our definition for $\delta J_{p}(f, g)$ is given below.
\begin{definition}\label{variation-p-5-15-1}
Let $f, g \in\mathscr A_0$. For $p>1,$ define $\delta J_{p}(f, g)$ by
\[
\delta J_{p}(f,g)=\underset{t\rightarrow0^+}{\lim}\frac{J(f\oplus_{p} t\cdot_p g)-J(f)}{t}
\]
whenever the limit exists.
\end{definition}
Let $J(f)>0$ and $\mathrm{Ent}(f)$ be the entropy of $f$ which may be formulated by \begin{equation}\label{def-entropy}
\mathrm{Ent}(f)=\int_{\mathbb{R}^n}f\log f\,dx-J(f)\log J(f).
\end{equation}
According to \cite[Proposition 3.11]{CF13}, $\mathrm{Ent} (f)$ is finite if $f\in \mathscr A$ such that $J(f)>0$.
Our main result in this section is the following Minkowski type inequality for $\delta J_{p}(f, g)$. We only focus on $p>1$ as inequality \eqref{mink1} and its equality condition for $p=1$
have already been proved in \cite[Theorem 5.1]{CF13}.
\begin{theorem}\label{minkowskiinequality}
Let $f, g \in\mathscr A_0$ such that $J(f)>0$ and $J(g)>0$. For $p> 1,$ one has,
\begin{equation}\label{mink1}
\delta J_{p}(f,g)\geq J(f)\left[ \frac{n}{p} +\frac{1-p}{p}\log J(f)+\log J(g)\right]+\frac{1}{p}{\rm Ent}(f),
\end{equation}
with equality if and only if $f=g$. \end{theorem}
In order to prove Theorem \ref{minkowskiinequality}, we shall need some preparation. The following result shows how to calculate $\delta J_{p}(f, f)$. Again, the case for $p=1$ has been covered in \cite[Proposition 3.11]{CF13} and will not be repeated in the following result. From Lemma \ref{specialcaseoffg}, one sees that $\delta J_{p}(f,f)$ is finite if $f\in \mathscr A$ such that $J(f)>0$.
\begin{lemma}\label{specialcaseoffg} Let $f\in \mathscr A_0$ such that $J(f)>0$. For $p>1,$ one has
\begin{equation}\label{entropu-p>1}
\delta J_{p}(f,f)= \frac{n}{p} J(f) + \frac{1}{p} \int_{\mathbb{R}^n} f \log f \, dx=\frac{n+\log J(f)}{p} J(f)+\frac{1}{p} \mathrm{Ent}(f).
\end{equation}
\end{lemma}
\begin{proof} Let $f=e^{-\varphi}\in \mathscr A_0$ such that $J(f)>0$. It can be checked from \eqref{prodottoL} that, for $p> 1$ and $t>0$, $f\oplus_{p}t\cdot_p f=e^{-\phi_t}$ with $$ \phi_t(x)=[\varphi\cdot_p (1+t)](x)=(1+t)^{\frac{1}{p}}\varphi\bigg(\frac{x}{(1+t)^{\frac{1}{p}}}\bigg).$$ Consequently, by letting $x=(1+t)^{\frac{1}{p}}y$, one gets
\begin{align*}
\delta J_{p}(f,f) &=\lim_{t\rightarrow0^{+}}\frac{J(f\oplus_{p}t\cdot_p f)-J(f)}{t}\\
& =\lim_{t\rightarrow 0^{+}} \frac{1}{t} \bigg((1+t)^{\frac{n}{p}} \int_{\mathbb{R}^n} e^{-(1+t)^{\frac{1}{p}}\varphi(y)}\,dy - \int_{\mathbb{R}^n} e^{-\varphi(y)} \,dy\bigg) \\
& =\lim_{t\rightarrow 0^{+}} \frac{(1+t)^{\frac{n}{p}}-1}{t} \int_{\mathbb{R}^n} e^{-(1+t)^{\frac{1}{p}}\varphi(y)}\,dy +\lim_{t\rightarrow 0^{+}} \int_{\mathbb{R}^n} \frac{e^{-(1+t)^{\frac{1}{p}}\varphi(y)}-e^{-\varphi(y)}}{t} \,dy. \end{align*}
Note that $\varphi\in \mathscr L_0$ is non-negative. It follows from the monotone convergence theorem that
\begin{eqnarray*} \lim_{t\rightarrow 0^{+}} \frac{(1+t)^{\frac{n}{p}}-1}{t} \int_{\mathbb{R}^n} e^{-(1+t)^{\frac{1}{p}}\varphi(y)}\,dy=\frac{n}{p} \lim_{t\rightarrow 0^{+}} \int_{\mathbb{R}^n} e^{-(1+t)^{\frac{1}{p}}\varphi(y)}\,dy =\frac{n}{p}J(f).
\end{eqnarray*}
Similarly, one can also have
\begin{eqnarray*} \lim_{t\rightarrow 0^{+}} \! \int_{\mathbb{R}^n} \! \frac{e^{-(1+t)^{\frac{1}{p}}\varphi(y)}-e^{-\varphi(y)}}{t} \,dy =\!\int_{\mathbb{R}^n} \lim_{t\rightarrow 0^{+}} \frac{e^{-(1+t)^{\frac{1}{p}}\varphi(y)}-e^{-\varphi(y)}}{t} \,dy=-\frac{1}{p}\int_{\mathbb{R}^n}\varphi(x)e^{-\varphi(x)}\,dx.
\end{eqnarray*}
By \eqref{def-entropy}, one gets
\[ \delta J_{p}(f,f)= \frac{n}{p}J(f) +\frac{1}{p}\int_{\mathbb{R}^n} f \log f \, dx= \frac{n}{p}J(f) + \frac{1}{p} \mathrm{Ent}(f)+ \frac{1}{p} J(f)\log J(f).\]
This is exactly the desired equality \eqref{entropu-p>1}.
\end{proof}
Our second lemma is to extend \cite[Lemma 3.9]{CF13} for $p=1$ to all $p>1$.
\begin{lemma}\label{boundness} Let $f=e^{-\varphi}\in \mathscr A_0$ and $g=e^{-\psi} \in\mathscr A_0$. For $p> 1$ and for $t> 0$, set \begin{equation} \label{varphi-t} \varphi_{t}= \varphi\square_{p}( \psi \cdot_p t) \end{equation} and $f_{t}=e^{-\varphi_{t}}$. Then, for any $ x\in\mathbb{R}^n$ and $t, s\in(0,1]$ such that $s<t$, one has,
\[\varphi_{1}(x)\leq \varphi_{t}(x)\leq \varphi_s(x)\leq \varphi (x) \quad \mathrm{and} \quad f(x)\leq f_s(x)\leq f_{t}(x)\leq f_{1}(x). \] \end{lemma}
\begin{proof}
Let $t > 0$, $\delta>0,$ and $p\geq1$. Note that $f=e^{-\varphi}\in \mathscr A_0$ and $g=e^{-\psi} \in\mathscr A_0$ imply $\varphi^*, \psi^*\geq 0$. By Proposition \ref{close-Lp-A-sum}, \eqref{relation-p-sq-1} (or \eqref{biconjugate-1}) and \eqref{varphi-t}, one gets that \begin{equation}\label{varphi-t-star} \varphi_t^*= \big((\varphi^*)^p+t(\psi^*)^p\big)^{\frac{1}{p}}\geq \varphi^*. \end{equation} Clearly, $\varphi_{t+\delta}^*\geq \varphi_t^*$ and $\varphi_{t+\delta}=(\varphi_{t+\delta}^*)^*\leq (\varphi_t^*)^*=\varphi_t.$ Moreover, for any $ x\in\mathbb{R}^n$ and $t\in[0,1]$, \[
\varphi_{1}(x)\leq \varphi_{t}(x)\leq \varphi (x) \quad \mathrm{and} \quad f(x)\leq f_{t}(x)\leq f_{1}(x). \] This completes the proof of this lemma.
\end{proof}
The following lemma proves that $\delta J_p (f, g)$ indeed exists (although may be $+\infty$). Again we only focus on $p>1,$ as $p=1$ has been covered in \cite[Theorem 3.6] {CF13}.
\begin{lemma}\label{diffe} Let $f, g \in\mathscr A_0$ such that $J(f)>0$. For $p> 1,$ one has $
\delta J_{p}(f,g) \in[0, + \infty].$ \end{lemma}
\begin{proof} Let $f=e^{-\varphi}\in \mathscr A_0$ and $g=e^{-\psi} \in\mathscr A_0$ such that $J(f)>0$. Then $\varphi, \psi \in \mathscr L_0$. Let $\varphi_{t}$ be as in \eqref{varphi-t}. It follows from Lemma \ref{boundness} that $\varphi(x) \geq\bar{\varphi}(x):= \lim_{t\rightarrow 0^+} \varphi_t(x)$ for every $x\in \mathbb{R}^n$ and $$
J(e^{-\bar{\varphi}})=\lim_{t\rightarrow 0^+} J(e^{-\varphi_t})\geq J(e^{-\varphi}),
$$ by the monotone convergence theorem. Note that \begin{eqnarray} \delta J_p(f, g)=
\lim_{t\rightarrow0^{+}}\frac{J(e^{-\varphi_t})-J(e^{-\varphi})}{t}. \label{replace-1}
\end{eqnarray} Therefore, $\delta J_p(f, g)=+\infty$ if $J(e^{-\bar{\varphi}})>J(e^{-\varphi})$, and $\delta J_p(f, g)=0$ if $J(e^{-\varphi_{t_0}})=J(e^{-\varphi})$ for some $t_0>0$ (hence $J(e^{-\bar{\varphi}})=J(e^{-\varphi_{t}})=J(e^{-\varphi})$ for every $t\in [0,t_0]$ due to Lemma \ref{boundness}).
Lastly, we consider the case that $J(e^{-\varphi_t})>J(e^{-\varphi})$ for all $t>0$ but $J(e^{-\bar{\varphi}})=J(e^{-\varphi})$. Note that $J(f)=J(e^{-\varphi})>0.$ In this case, one has
\begin{eqnarray}
\frac{J(e^{-\varphi_{t}})-J(e^{-\varphi})}{t}=\frac{\log J(e^{-\varphi_{t}})-\log J(e^{-\varphi})}{t}\cdot\frac{J(e^{-\varphi_{t}})-J(e^{-\varphi})}{\log J(e^{-\varphi_{t}})-\log J(e^{-\varphi})}.\label{split-1}
\end{eqnarray} Corollary \ref{concave-PL-1} and Lemma \ref{boundness} imply that $\log J(e^{-\varphi_{t}})$ is an increasing and concave function on $t\in (0, \infty)$. Thus, \begin{eqnarray}
J(e^{-\varphi}) =\lim_{t\rightarrow 0^{+}}\frac{J(e^{-\varphi_{t}})-J(e^{-\varphi})}{\log J(e^{-\varphi_{t}})-\log J(e^{-\varphi})}, \ \ \mathrm{and} \ \ \lim_{t\rightarrow 0^{+}}\frac{\log J(e^{-\varphi_{t}})-\log J(e^{-\varphi})}{t}\in [0,+\infty]. \label{3.2}\end{eqnarray} Combining \eqref{split-1} and \eqref{3.2}, one gets $\delta J_p(f,g)\in[0,+\infty]$, and this completes the proof.
\end{proof}
We also need the following lemma. The case for $p=1$ has been given in \cite[Lemma 5.4]{CF13}, so we only state the result for $p>1$.
\begin{lemma}\label{lemma35}
Let $f=e^{-\varphi}\in \mathscr A_0$ and $g=e^{-\psi} \in\mathscr A_0$ such that $J(f)>0$. For $p> 1$, one has, \begin{equation}\label{decompose-1}
\lim_{t\rightarrow0^+} \frac{J((1-t)\cdot_p f\oplus_{p} t\cdot_p g)-J(f)}{t}=\delta J_{p}(f,g)-\delta J_{p}(f,f).
\end{equation}
\end{lemma}
\begin{proof} Let $f=e^{-\varphi}\in \mathscr A_0$ and $g=e^{-\psi} \in\mathscr A_0$ such that $J(f)>0$. According to Proposition \ref{close-Lp-A-sum}, for $p> 1$ and $t\in(0,1)$, $(1-t)\cdot_p f\oplus_{p}t\cdot_p g=e^{-\phi_t^*}$ where $$\phi_t= \big((1-t)(\varphi^*)^p+t(\psi^*)^p\big)^{\frac{1}{p}}= (1-t) ^{\frac{1}{p}} \Big((\varphi^*)^p+\frac{t}{1-t} (\psi^*)^p\Big)^{\frac{1}{p}}.$$
It can be checked by \eqref{const-mul-dual} that \begin{eqnarray*} \phi_t ^*(x) =(1-t)^{\frac{1}{p}}\Big[\big((\varphi^*)^p+\frac{t}{1-t}(\psi^*)^p \big)^{\frac{1}{p}}\Big]^*\Big(\frac{x}{(1-t)^{1/p}}\Big).
\end{eqnarray*} By letting $x=(1-t)^{\frac{1}{p}} y$ and $s=\frac{t}{1-t}$, one gets $1-t=\frac{1}{1+s}$ and
\begin{align*}
J((1-t)\cdot_p f\oplus_{p}t\cdot_p g)
& = \int_{\mathbb{R}^n} (1-t)^{\frac{n}{p}} e^{-(1-t)^{\frac{1}{p}}\big(\big((\varphi^*)^p+\frac{t}{1-t}(\psi^*)^p\big)^{\frac{1}{p}}\big)^*(y)}\,dy \\ & = \int_{\mathbb{R}^n} (1+s)^{-\frac{n}{p}} e^{-(1+s)^{-\frac{1}{p}}\big(\big((\varphi^*)^p +s (\psi^*)^p \big)^{\frac{1}{p}}\big)^*(y) }\,dy\\ &= \int_{\mathbb{R}^n} (1+s)^{-\frac{n}{p}} e^{-(1+s)^{-\frac{1}{p}} \varphi_s(y) }\,dy, \end{align*}
where $\varphi_{s}$ is given by \eqref{varphi-t}. Like in the proof of Lemma \ref{diffe}, let $\bar{\varphi}(x)= \lim_{s\rightarrow 0^+} \varphi_s(x)$ for every $x\in \mathbb{R}^n$. By Lemma \ref{boundness} and the monotone convergence theorem, one has $\varphi \geq \bar{\varphi}$ and $J(e^{-\bar{\varphi}})=\lim_{s\rightarrow 0^+} J(e^{-\varphi_s})\geq J(e^{-\varphi}).$ Lemma \ref{boundness} also implies that $(1+s)^{-\frac{1}{p}} \varphi_s(y)$ is decreasing for all $y\in \mathbb{R}^n$ with $\lim_{s\to 0^+} (1+s)^{-\frac{1}{p}} \varphi_s(y)=\bar{\varphi}(y)\leq \varphi (y).$ It follows from the monotone convergence theorem that, $$\lim_{s\rightarrow 0^+} \int_{\mathbb{R}^n} (1+s)^{-\frac{n}{p}} e^{-(1+s)^{-\frac{1}{p}} \varphi_s(y)} \,dy= \lim_{s\rightarrow 0^+} \int_{\mathbb{R}^n} e^{-(1+s)^{-\frac{1}{p}} \varphi_s(y)} \,dy =J(e^{-\bar{\varphi}}). $$
In summary, due to $s=\frac{t}{1-t}$, one has \begin{align}
\lim_{t\rightarrow 0^+} &\frac{J((1-t)\cdot_p f\oplus_{p}t\cdot_p g)-J(f)}{t}\nonumber \\
&=\lim_{s\rightarrow 0^+} \frac{1+s}{s} \int_{\mathbb{R}^n} \Big((1+s)^{-\frac{n}{p}} e^{-(1+s)^{-\frac{1}{p}} \varphi_s(y)}-e^{-\varphi(y)}\Big)\,dy \nonumber \\ &=\lim_{s\rightarrow 0^+} \int_{\mathbb{R}^n} \frac{ (1+s)^{-\frac{n}{p}} e^{-(1+s)^{-\frac{1}{p}} \varphi_s(y)}-e^{-\bar{\varphi}}(y)}{s} \,dy +\lim_{s\rightarrow 0^+} \frac{J(e^{-\bar{\varphi}})-J(e^{-\varphi})}{s}.\label{chang-variable-1}
\end{align}
Clearly, the second limit in \eqref{chang-variable-1} equals to $+\infty$ if $J(e^{-\bar{\varphi}})>J(e^{-\varphi})$. Note that, in this case, $\delta J_p(f, g)=+\infty$ as proved in Lemma \ref{diffe}, and this proves \eqref{decompose-1} if $J(e^{-\bar{\varphi}})>J(e^{-\varphi})$.
Now let us consider the case $J(e^{-\bar{\varphi}})=J(e^{-\varphi})$. As $\varphi \geq \bar{\varphi}$, one has $e^{-\varphi(x)}= e^{-\bar{\varphi}(x)}$ (and hence $\varphi(x)=\bar{\varphi}(x)$) for almost all $x\in \mathbb{R}^n$. According to \eqref{def-entropy} and \eqref{mink1}, one sees \begin{equation} \delta J_p(f, f)=\delta J_p(e^{-\bar{\varphi}}, e^{-\bar{\varphi}}).\label{equality-entropy} \end{equation} Besides, \eqref{replace-1} implies that \begin{eqnarray} \delta J_p(f, g)=
\lim_{t\rightarrow0^{+}}\frac{J(e^{-\varphi_t})-J(e^{-\varphi})}{t}=
\lim_{t\rightarrow0^{+}}\frac{J(e^{-\varphi_t})-J(e^{-\bar{\varphi}})}{t}. \label{replace-5-17}
\end{eqnarray} Replacing $\varphi$ by $\bar{\varphi}$ in \eqref{chang-variable-1}, one has \begin{align}
\lim_{s\rightarrow 0^+} \int_{\mathbb{R}^n} \frac{ (1+s)^{-\frac{n}{p}} e^{-(1+s)^{-\frac{1}{p}} \varphi_s(y)}-e^{-\varphi(y)}}{s}\,dy &= \lim_{s\rightarrow 0^+} \int_{\mathbb{R}^n} \frac{ (1+s)^{-\frac{n}{p}} e^{-(1+s)^{-\frac{1}{p}} \varphi_s(y)}-e^{-\bar{\varphi}(y)}}{s}\,dy \nonumber \\ &= B_1+B_2+B_3.\label{b1-b2-b3}
\end{align} Here $B_1$ is defined and calculated as below:
\begin{align*}B_1&= \lim_{s\rightarrow 0^+} \int_{\mathbb{R}^n} \frac{ (1+s)^{-\frac{n}{p}} e^{-(1+s)^{-\frac{1}{p}} \varphi_s(y)}-e^{-(1+s)^{-\frac{1}{p}} \varphi_s(y)}}{s}\,dy \nonumber \\ &= \bigg(\lim_{s\rightarrow 0^+} \frac{ (1+s)^{-\frac{n}{p}} -1}{s} \bigg) \bigg( \lim_{s\rightarrow 0^+} \int_{\mathbb{R}^n} e^{-(1+s)^{-\frac{1}{p}} \varphi_s(y)} \,dy \bigg)
= -\frac{n}{p} J\big(e^{- \bar{\varphi}}\big), \end{align*} where Lemma \ref{boundness} and the monotone convergence theorem are used. It follows from \eqref{replace-5-17} that
\begin{align*}B_2&= \lim_{s\rightarrow 0^+} \int_{\mathbb{R}^n} \frac{e^{-\varphi_s(y)}-e^{-\bar{\varphi}(y)}}{s}\,dy
=\delta J_p(f, g). \end{align*} The term $B_3$ is defined and calculated as follows: \begin{align} B_3&= \lim_{s\rightarrow 0^+} \int_{\mathbb{R}^n} \frac{e^{-(1+s)^{-\frac{1}{p}} \varphi_s(y)}-e^{-\varphi_s(y)}}{s}\,dy=\frac{1}{p} \int_{\mathbb{R}^n} \bar{\varphi}(y) e^{-\bar{\varphi}(y)}\,dy. \label{estimation-B3} \end{align} Indeed, \eqref{estimation-B3} is a consequence of the dominated convergence theorem and we now provide some details to complete the argument. Clearly, the integral of the second term in \eqref{estimation-B3} is actually over the domain of $\varphi_s$, since $(1+s)^{-\frac{1}{p}} \varphi_s(y)=+\infty$ and $\varphi_s(y)=+\infty$ if $y\notin {\mathrm {dom}} (\varphi_s)$.
Note that $0\leq \varphi_s(y)<\infty$ for any $y\in {\mathrm {dom}} (\varphi_s)$. Moreover, $1-(1+s)^{-\frac{1}{p}}$ is increasing on $s\in (0, 1)$ and $$\lim_{s\rightarrow 0^+} \frac{1-(1+s)^{-\frac{1}{p}}}{s} =\frac{1}{p},$$ which implies the existence of a finite constant $M<\infty$ such that for all $s\in (0, 1)$, $$0< \frac{1-(1+s)^{-\frac{1}{p}}}{s}<M\big(1-2^{-\frac{1}{p}}\big). $$ As the function $\frac{e^x-1}{x}$ is increasing on $x\in (0, \infty)$, one gets, for all $s\in (0, 1)$ and for all $y\in {\mathrm {dom}} (\varphi_s)$, \begin{align*}
\frac{e^{\big(1-(1+s)^{-\frac{1}{p}} \big) \varphi_s(y)}-1}{\big(1-(1+s)^{-\frac{1}{p}} \big) \varphi_s(y)}\leq \frac{e^{\big(1-2^{-\frac{1}{p}} \big) \varphi_s(y)}-1}{\big(1-2^{-\frac{1}{p}} \big) \varphi_s(y)}.
\end{align*} By Lemma \ref{boundness}, for all $s\in (0, 1)$ and for any $y\in {\mathrm {dom}} (\varphi_s)$,
\begin{align*} \frac{e^{-(1+s)^{-\frac{1}{p}} \varphi_s(y)}-e^{-\varphi_s(y)}}{s} &=\Bigg(\frac{e^{-\varphi_s(y)}\big(1-(1+s)^{-\frac{1}{p}} \big) \varphi_s(y)}{s}\Bigg) \Bigg( \frac{e^{\big(1-(1+s)^{-\frac{1}{p}} \big) \varphi_s(y)}-1}{\big(1-(1+s)^{-\frac{1}{p}} \big) \varphi_s(y)}\Bigg) \nonumber \\ & \leq M e^{-\varphi_s(y)} \Big( {e^{\big(1-2^{-\frac{1}{p}} \big) \varphi_s(y)}-1} \Big) \nonumber \\ &\leq M e^{ -2^{-\frac{1}{p}} \varphi_s(y)} \leq M e^{ -2^{-\frac{1}{p}} \varphi_1(y)}.\end{align*} It is easily checked by Proposition \ref{close-Lp-A-sum} that $e^{ -2^{-\frac{1}{p}} \varphi_1}\in \mathscr A_0,$ and hence $\int_{\mathbb{R}^n}e^{ -2^{-\frac{1}{p}} \varphi_1(y)}\,dy<+\infty.$
For convenience, let $\bar{\varphi}(y) e^{-\bar{\varphi}(y)}=0$ if $y\notin{\mathrm {dom}}(\bar{\varphi})$.
By Lemma \ref{boundness}, ${\mathrm {dom}}(\bar{\varphi})\subseteq {\mathrm {dom}}(\varphi_s)\subseteq {\mathrm {dom}}(\varphi_t)$ holds for $t, s\in(0,1]$ such that $s<t$. Thus, if there exists $s_0>0$ such that $y\notin {\mathrm {dom}}(\varphi_{s_0})$, then $y\notin {\mathrm {dom}}(\varphi_s)$ for all $s\in (0, s_0]$ and $y\notin{\mathrm {dom}}(\bar{\varphi})$; this in turn implies that $$\lim_{s\rightarrow 0^+} \frac{e^{-(1+s)^{-\frac{1}{p}} \varphi_s(y)}-e^{-\varphi_s(y)}}{s}=0=\frac{1}{p} \bar{\varphi}(y) e^{-\bar{\varphi}(y)}.$$ On the other hand, if $y\in {\mathrm {dom}}(\varphi_s)$ for any $s\in (0, 1)$, then
\begin{align*} \lim_{s\rightarrow 0^+} \frac{e^{-(1+s)^{-\frac{1}{p}} \varphi_s(y)}-e^{-\varphi_s(y)}}{s}
&=\left(\lim_{s\rightarrow 0^+} e^{-\varphi_s(y)}\right) \left(\lim_{s\rightarrow 0^+} \frac{e^{\big(1-(1+s)^{-\frac{1}{p}}\big) \varphi_s(y)}-1}{s} \right) =\frac{1}{p}\bar{\varphi}(y) e^{-\bar{\varphi}(y)}.\end{align*} Hence, the dominated convergence theorem can be applied to $B_3$ and get \eqref{estimation-B3}.
Summing up $B_1, B_2$ and $B_3$, by \eqref{chang-variable-1}, \eqref{equality-entropy}, \eqref{b1-b2-b3} and Lemma \ref{specialcaseoffg}, one gets \begin{align*}
\lim_{t\rightarrow 0^+} \frac{J((1-t)\cdot_p f\oplus_{p}t\cdot_p g)-J(f)}{t}
&=B_1+B_2+B_3=\delta J_{p}(f,g)-\delta J_{p}(f,f).
\end{align*} Hence, formula \eqref{decompose-1} is obtained. This completes the proof of Lemma \ref{lemma35}. \end{proof}
We are now ready to prove our Theorem \ref{minkowskiinequality}.
\begin{proof}[Proof of Theorem \ref{minkowskiinequality}] Let $f=e^{-\varphi}\in \mathscr A_0$ and $g=e^{-\psi}\in\mathscr A_0$ such that $J(f)>0$ and $J(g)>0$. For $p>1$, let $F(t)=\log J((1-t)\cdot_p f\oplus_{p} t\cdot_p g)$ for $t\in (0, 1)$. Let $$F(0+)=\lim_{t\to 0^+}F(t) \ \ \mathrm{and}\ \ F(1-)=\lim_{t\to 1^-}F(t).$$ Note that $F(0+)\geq \log J(f)$ and $F(1-)\geq \log J(g)$. According to Corollary \ref{concave-PL-1}, $F$ is a concave function on $t\in (0, 1)$ and thus $$F(t) \geq (1-t)F(0+)+tF(1-)\geq \log J(f)+t(\log J(g)-\log J(f))$$ holds for $t\in (0, 1)$. Consequently, if $F(0+)>\log J(f)$, then
$$ \lim_{t\rightarrow0^+} \frac{J((1-t)\cdot_p f\oplus_{p} t\cdot_p g)-J(f)}{t} =+\infty > J(f)\log\Big(\frac{J(g)}{J(f)}\Big).$$ While if $F(0+)=\log J(f)$, then \begin{eqnarray} \lim_{t\rightarrow0^+} \frac{J((1-t)\cdot_p f\oplus_{p} t\cdot_p g)-J(f)}{t} =\frac{\,d e^{F(t)}}{\,dt}\bigg|_{t=0^+}\!\! =e^{F(t)} \frac{\,d F(t)}{\,dt}\bigg|_{t=0^+}\geq J(f)\log\Big(\frac{J(g)}{J(f)}\Big).\label{derivative--1} \end{eqnarray}
Lemmas \ref{specialcaseoffg} and \ref{lemma35} then yield the desired inequality \eqref{mink1} as follows:
\begin{align} \delta J_{p}(f,g)&= \delta J_{p}(f,f) + \lim_{t\rightarrow0^+} \frac{J((1-t)\cdot_p f\oplus_{p} t\cdot_p g)-J(f)}{t} \nonumber \\ & \geq \delta J_{p}(f,f)+J(f)\log\Big(\frac{J(g)}{J(f)}\Big) \label{ine-with-d-ff} \\
&= J(f)\left[ \frac{n}{p} +\frac{1-p}{p}\log J(f)+\log J(g)\right]+\frac{1}{p}{\rm Ent}(f). \nonumber\end{align}
Now let us characterize the equality for \eqref{mink1}. It is obvious that \eqref{mink1} becomes equality if $f=g$ by Lemma \ref{specialcaseoffg}. Conversely, assume that (\ref{mink1}) holds with equality sign which happens only in the case $F(0+)=\log J(f)$; it requires equality in \eqref{derivative--1}. In particular, as $F(0+)= \log J(f)$, one has
$$ F'(0^{+})=\frac{\,d F(t)}{\,dt}\bigg|_{t=0^+} = \log\Big(\frac{J(g)}{J(f)}\Big).$$ Note that $F(t)\leq F(0+)+t F'(0+)$ for all $t\in (0, 1)$ since $F$ is concave on $(0, 1)$. This gives $$F(t) \leq \log J(f)+t(\log J(g)-\log J(f)).$$ Consequently, equality holds in the Pr\'ekopa-Leindler type
inequality (\ref{prekopainequality1}) and then $f=g$.
\end{proof}
The following corollary provides a unique determination of log-concave functions.
\begin{corollary}\label{cormink-uniq} Let $f_1, f_2 \in\mathscr A_0$ such that $J(f_1)=J(f_2)>0$. For $p>1$, if \begin{equation}\label{incroci-uniq}
\delta J_p (f_1, g) = \delta J_p (f_2, g) \end{equation} holds for any $g\in \mathscr A_0$ with $J(g)>0$, then $f_1=f_2$. \end{corollary}
\begin{proof} By letting $g=f_1$ in \eqref{incroci-uniq}, it follows from Theorem \ref{minkowskiinequality}
(or \eqref{ine-with-d-ff}) and $J(f_1)=J(f_2)$ that \begin{eqnarray}\label{incroci-uniq-11}
\delta J_p (f_1, f_1) = \delta J_p (f_2, f_1) \geq \delta J_{p}(f_2, f_2)+J(f_2)\log\Big(\frac{J(f_1)}{J(f_2)}\Big)
=\delta J_{p}(f_2, f_2)
\end{eqnarray} with equality if and only if $f_1=f_2$. Similarly, \begin{equation}\label{incroci-uniq-22}
\delta J_p (f_2, f_2) = \delta J_p (f_1, f_2) \geq \delta J_{p}(f_1, f_1)+J(f_1)\log\Big(\frac{J(f_2)}{J(f_1)}\Big)=\delta J_{p}(f_1, f_1).\end{equation} This means that \eqref{incroci-uniq-11} holds with equality, which in turn gives $f_1=f_2$ as desired.
\end{proof}
\section{An explicit formula for $\delta J_p(f, g)$ } \label{section-varition-516} \setcounter{equation}{0}
Our goal in this section is to obtain an explicit integral formula for $\delta J_p(f, g)$ for $p>1$ under additional conditions on $f$ and $g$. Again the case $p=1$ has been discussed in \cite{CF13} and hence will not be covered in this section. We shall need the subclass $\mathscr A'_0\subset \mathscr A_0$ where $\mathscr A_0'=\{f=e^{-\varphi}: \varphi \in \mathscr L'_0\}$
with $\mathscr L_0'\subset\mathscr L_0$ given by
\[\mathscr L'_0:=\Big\{\varphi\in\mathscr L_0: \varphi \in \mathcal C^1(\mathbb{R}^n)\cap\mathcal C^2_+(\mathbb{R}^n\setminus\{o\}) \ \mathrm{is\ strictly\ convex\ and \ supercoercive\ with}\ {\mathrm {dom}}(\varphi)=\mathbb{R}^n\Big\}.
\] Hereafter, a function $\varphi$ is called supercoercive if $\lim_{|x|\rightarrow\infty} \frac{\varphi(x)}{|x|}=+\infty.$
It is easily checked that $(\mathbb{R}^n, \varphi)$ for $\varphi\in \mathscr L'_0$ is a pair satisfying that $\varphi$ is differentiable and strictly convex on $\mathbb{R}^n$, and \begin{equation} \label{ess-smooth} \lim_{i\to\infty} |\nabla \varphi(x_i)|\rightarrow+\infty\ \ \mathrm{for\ each\ sequence }\ \ \{x_i\}_{i\in\mathbb{N}}\subset \mathbb{R}^n \ \ \mathrm{such\ that}\ \lim_{i\to \infty} |x_i|=+\infty. \end{equation} This pair usually is called {\it a convex function of Legendre type} (see e.g. \cite[Section 26]{Roc70} for more general definitions and properties). In general, \eqref{ess-smooth} holds automatically for $\varphi$ if ${\mathrm {dom}}(\varphi^*)=\mathbb{R}^n$ and $\varphi$ is a differentiable convex function with ${\mathrm {dom}}(\varphi)=\mathbb{R}^n$, due to \cite[Lemma 26.7]{Roc70}.
We say that $(D, \psi)$ is the \emph{Legendre conjugate} of $(C, \varphi)$ if
$$
\psi(y)=\langle x,y\rangle-\varphi(x),\quad \mathrm{for\ any}\ y\in D\ \mathrm{and \ for \ any }\ x \in \nabla \varphi^{-1}(y)=\{z\in C: \ y= \nabla \varphi(z)\},
$$ where $D=\{y\in \mathbb{R}^n: y=\nabla \varphi(x), \ x\in C\}$. Theorem 26.5 in \cite{Roc70} provides a nice result regarding the relation between the Legendre conjugate and Fenchel conjugate. We shall not need the full statement of \cite[Theorem 26.5]{Roc70}, and only the special cases, when both domains are $\mathbb{R}^n$, will be stated in the following lemma.
\begin{lemma}\label{proplegendre} Let $\phi\in \mathcal C^1(\mathbb{R}^n)$ be such that $ {\mathrm {dom}}(\phi^*)=\mathbb{R}^n$. Then $(\mathbb{R}^n, \phi)$ is a convex function of Legendre type if and only if $(\mathbb{R}^n, \phi^*)$ is. When these conditions hold, $(\mathbb{R}^n, \phi ^*)$ is the Legendre conjugate of $(\mathbb{R}^n, \phi)$ (and vice verse). Moreover, both $\nabla \phi: \mathbb{R}^n \to \mathbb{R}^n$ and $\nabla \phi^*:\mathbb{R}^n\to \mathbb{R}^n$ are continuous bijections and satisfy that $\nabla \phi ^*$ is the inverse of $\nabla \phi$ (namely, $\nabla ^{-1} \phi= \nabla \phi ^*$).\end{lemma}
We now prove some lemmas before we state our main result in this section.
\begin{lemma}\label{vainshat0}
If $\varphi\in \mathscr L'_0$, then $\varphi^*\in \mathscr L_0'$. Moreover, $\nabla \varphi(o)=o$, $\nabla\varphi^*(o)=o$, and
$$\{x\in\mathbb{R}^n: \varphi(x)=0\}=\{x\in\mathbb{R}^n: \varphi^*(x)=0\}=\{o\}.$$
\end{lemma}
\begin{proof} It is clear that $\varphi^{**}=\varphi$, as $\varphi$ is convex and $\varphi\in \mathcal C^1(\mathbb{R}^n)$. According to the Moreau-Rockafellar theorem (see e.g. \cite[Proposition 3.5.4]{BV10}), a proper lower semi-continuous convex function $\varphi$ on $\mathbb{R}^n$ is supercoercive if and only if ${\mathrm {dom}}(\varphi^*)=\mathbb{R}^n$. Consequently, ${\mathrm {dom}}(\varphi^*)=\mathbb{R}^n$ and $\varphi^*$ is supercoercive, due to the facts that $\varphi$ is supercoercive, and respectively ${\mathrm {dom}}(\varphi)=\mathbb{R}^n$. It is also trivial to have $\varphi^*(o)=0$ and $\varphi^*(y)\geq 0$ for all $y\in \mathbb{R}^n$, as $\varphi\in\mathscr L_0$. Note that the pair $(\mathbb{R}^n, \varphi)$ is a convex function of Legendre type, so is the pair $(\mathbb{R}^n, \varphi^*)$ by Lemma \ref{proplegendre}. In particular, $\varphi^*$ is strictly convex on $\mathbb{R}^n$. Lemma \ref{proplegendre} also implies that $\nabla \varphi$ and its inverse $\nabla \varphi^*$ are both continuous on $\mathbb{R}^n$ and thus $\varphi^*\in \mathcal C^1(\mathbb{R}^n)$. As $\varphi\in \mathcal C^2_+(\mathbb{R}^n\setminus\{o\})$, the Hessian matrix $\nabla^2\varphi$ is positive definite and continuous on $\mathbb{R}^n\setminus\{o\}$. It follows from the inverse mapping theorem that $\varphi^*\in \mathcal C^2_+(\mathbb{R}^n\setminus \{o\})$, which concludes $\varphi^*\in \mathscr L_0'$. In particular, $\varphi(x)+\varphi^*(y)=\langle x, y\rangle$ holds for all $x, y\in \mathbb{R}^n$ such that $y=\nabla \varphi(x)$ (and $x=\nabla\varphi^*(y)$). The strict convexity of $\varphi$ implies that $\varphi$ has a unique minimizer. As $\varphi\in \mathscr L_0$, $\varphi$ attains its minimum at $o$, hence $\nabla\varphi(o)=o$ and $\{x\in\mathbb{R}^n: \varphi(x)=0\}=\{o\}$. The same arguments clearly work for $\varphi^*$, and this concludes the proof of Lemma \ref{vainshat0}. \end{proof}
\begin{lemma}\label{belongtothesameclass}
Let $\varphi, \psi \in \mathscr L'_0$. For $p>1$ and $t>0$, set $\varphi_{t}:=\varphi\square_{p}( \psi\cdot_pt).$ Then $\varphi_{t}\in \mathscr L_0'$. Moreover, both $(\mathbb{R}^n, \varphi_t)$ and $(\mathbb{R}^n, \varphi_t^*)$ are convex functions of Legendre type.
\end{lemma}
\begin{proof} Let $\varphi, \psi \in \mathscr L'_0\subset \mathscr L_0$ be two convex functions. According to Proposition \ref{close-Lp-A-sum}, $\varphi_t\in \mathscr L_0$ for all $t>0$. In particular, $\varphi_t$ is non-negative on $\mathbb{R}^n$. By Lemma \ref{boundness}, $0\leq \varphi_t\leq \varphi$ for all $t>0$ and $p>1$. This implies ${\mathrm {dom}} (\varphi_t)=\mathbb{R}^n$. This, together with the Moreau-Rockafellar theorem (see e.g. \cite[Proposition 3.5.4]{BV10}), immediately implies that $\varphi_t^*$ is supercoercive. On the other hand, by Proposition \ref{close-Lp-A-sum}, for any $t>0$, ${\mathrm {dom}}(\varphi_t^*)$ is clearly equal to $\mathbb{R}^n$ and thus $\varphi_t$ is supercoercive.
Let us check the differentiability of $\varphi_t$ and $\varphi_t^*$. As explained in the proof of Lemma \ref{vainshat0}, if $\varphi\in \mathscr L'_0$, then
$\nabla \varphi$ and its inverse $\nabla \varphi^*$ are both continuous on $\mathbb{R}^n$ and continuously differentiable on $\mathbb{R}^n\setminus\{o\}$. Moreover, $\nabla^2\varphi$ and $\nabla ^2\varphi^*$ are positive definite and continuous on $\mathbb{R}^n\setminus\{o\}$. Similar properties hold for $\psi$. Let $x\neq o$. In this case, both $\varphi^*(x)$ and $\psi^*(x)$ are strictly positive due to Lemma \ref{vainshat0}. It follows from \eqref{varphi-t-star} that, for all $t>0$ and $p>1$, $\varphi_t^*(x)>0$ and \begin{equation}\label{first-order-deriv}\nabla\varphi_t^*(x)=\frac{ \big(\varphi^*(x)\big)^{p-1} \nabla\varphi^*(x)+t(\psi^*(x))^{p-1} \nabla\psi^*(x)}{ \big(\varphi^*_t(x)\big)^{p-1} }. \end{equation} Clearly $\nabla\varphi_t^*(x)$ is continuous at $o\neq x\in \mathbb{R}^n$.
On the other hand, one can verify $\nabla\varphi_t^*(o)=o$ according to the usual definition of differentiability:
\begin{align*} \lim_{z\to o} \frac{\varphi_t^*(z)-\varphi_t^*(o)-\langle o, z-o\rangle}{|z-o|}&=\lim_{z\to o} \frac{\big((\varphi^*(z))^p+t(\psi^*(z))^p\big)^{\frac{1}{p}}}{|z|}\\ &=\lim_{z\to o} \bigg(\Big(\frac{\varphi^*(z)}{|z|}\Big)^p+t\Big(\frac{\psi^*(z)}{|z|}\Big)^p\bigg)^{\frac{1}{p}}=0,\end{align*} where the last equality follows from $\nabla \varphi^*(o)=\nabla\psi^*(o)=o$ due to
Lemma \ref{vainshat0}.
Moreover,
$$\lim_{x\rightarrow o}\nabla \varphi^*(x)=\nabla \varphi^*(o)=o\ \ \mathrm{and}\ \ \lim_{x\rightarrow o}\nabla \psi^*(x)=\nabla \psi^*(o)=o.$$ These conclude that, for $p>1$, $\varphi_t^*$ is continuously differentiable on $\mathbb{R}^n$, because \begin{align*} \lim_{x\to o} \big|\nabla\varphi_t^*(x)\big|&=\lim_{x\to o} \bigg|\frac{ \big(\varphi^*(x)\big)^{p-1} \nabla\varphi^*(x)+t(\psi^*(x))^{p-1} \nabla\psi^*(x)}{ \big(\varphi^*_t(x)\big)^{p-1} }\bigg|\\ &\leq \lim_{x\to o} \big|\nabla\varphi^*(x)\big| \bigg(\frac{\varphi^*(x)}{\varphi^*_t(x)}\bigg)^{p-1} +t^{\frac{1}{p}} \lim_{x\to o} \big|\nabla\psi^*(x)\big| \bigg(\frac{t (\psi^*)^p(x)}{(\varphi^*_t)^p(x)}\bigg)^{1-\frac{1}{p}} \\ &\leq \lim_{x\to o} \big|\nabla\varphi^*(x)\big| +t^{\frac{1}{p}} \lim_{x\to o} \big|\nabla\psi^*(x)\big| =0.\end{align*}
Now let us check that $(\mathbb{R}^n, \varphi_t^*)$ is a convex function of Legendre type. To this end, as $\varphi_t^*$ is already proved to be differentiable, we only need to verify that $\varphi^*_t$ is strictly convex on $\mathbb{R}^n$ and \eqref{ess-smooth} holds for $\varphi^*_t$. As mentioned before, \eqref{ess-smooth} for $\varphi^*_t$ holds automatically because ${\mathrm {dom}}(\varphi)=\mathbb{R}^n$ and $\varphi_t^*$ is differentiable, due to \cite[Lemma 26.7]{Roc70}. The strictly convexity of $\varphi_t^*$ is easily checked as follows: for $p>1$, $\lambda\in (0, 1)$ and $x, y\in \mathbb{R}^n$ such that $x\neq y$, by the Minkowski inequality for norms, \eqref{varphi-t-star} and Lemma \ref{vainshat0} (which implies the strict convexity of $\varphi^*$ and $\psi^*$), one has \begin{align*}
\varphi_t^*(\lambda x+(1-\lambda)y)& =\Big(\big(\varphi^*(\lambda x+(1-\lambda)y)\big)^p+t\big(\psi^*(\lambda x+(1-\lambda)y)\big)^p\Big)^{\frac{1}{p}}\\ &<\Big(\big(\lambda \varphi^*(x)+(1-\lambda)\varphi^*(y)\big)^p+t\big(\lambda \psi^*(x)+(1-\lambda)\psi^*(y)\big)^p\Big)^{\frac{1}{p}}\\&\leq \lambda \varphi_t^*(x)+(1-\lambda)\varphi_t^*(y).
\end{align*} Therefore, $(\mathbb{R}^n, \varphi_t^*)$ is a convex function of Legendre type, and so is $(\mathbb{R}^n, \varphi_t)$ due to Lemma \ref{proplegendre}. Moreover, both $\nabla \varphi_t^*$ and its inverse $\nabla \varphi_t$ are continuous. Thus $\varphi_t\in \mathcal C^1(\mathbb{R}^n)$ and is strictly convex.
According to Lemma \ref{vainshat0}, $\nabla^2\varphi^*$ and $\nabla^2\psi^*$ are both positive definite and continuous on $\mathbb{R}^n\setminus\{o\}$. Let $x\in \mathbb{R}^n\setminus\{o\}$ and by \eqref{first-order-deriv}, one has \begin{align} \nabla^2\varphi_t^* &=(p-1)\frac{(\varphi^*)^{p-2}\nabla\varphi^*\otimes\nabla\varphi^*+t(\psi^*)^{p-2}\nabla\psi^*\otimes\nabla\psi^*-(\varphi_t^*)^{p-2}\nabla\varphi_t^*\otimes\nabla\varphi_t^*}{(\varphi_t^*)^{p-1}}\nonumber \\ &\quad \quad + \frac{(\varphi^*)^{p-1}\nabla^2\varphi^*+t(\psi^*)^{p-1}\nabla^2\psi^*}{(\varphi_t^*)^{p-1}},\label{second-hession} \end{align} where $y\otimes y$ is the rank $1$ matrix generated by $y\in \mathbb{R}^n$. Clearly $\nabla^2\varphi_t^*$ is continuous on $\mathbb{R}^n\setminus \{o\}$ and positive definite whose determinant is strictly positive (the calculation is standard and hence will be omitted). So $\varphi_t^*\in \mathcal C^2_+(\mathbb{R}^n\setminus \{o\})$ for all $t>0$ and $p>1$. Together with $(\nabla \varphi_t) ^ {-1} = \nabla \varphi_t ^*$, the inverse mapping theorem gives that $\varphi_t\in \mathcal C^2_+(\mathbb{R}^n\setminus \{o\})$ for all $t>0$ and $p>1$. Moreover, for any $o\neq x\in \mathbb{R}^n$, $\nabla^2\varphi_t(x) $ equal the inverse of $\nabla^2 \varphi_t^*(y)$ with $y=\nabla \varphi_t(x)$. This concludes that $\varphi_t\in \mathscr L'_0$. \end{proof}
We will also need the following lemma. The function $\varphi_t$ is as in \eqref{varphi-t}.
\begin{lemma}\label{continuityforgradient}
Let $\varphi, \psi \in \mathscr L'_0$. For $p>1$ and $t>0$, one has,\\
\noindent i) $\lim_{t\rightarrow0^{+}}\varphi_{t}(x)=\varphi(x)$ for all $x\in \mathbb{R}^n$;\\
\noindent ii) for every closed bounded subset $E\subset \mathbb{R}^n$, $\lim_{t\rightarrow0^{+}}\nabla \varphi_{t}(x)=\nabla \varphi(x)$ uniformly on $E$.
\end{lemma}
\begin{proof}
i) Let $\varphi, \psi \in \mathscr L'_0$. By Lemma \ref{boundness}, for any $x\in \mathbb{R}^n$, $\varphi_t(x)$ is decreasing on $t\in (0, 1]$, and thus $
\limsup_{t\rightarrow0^{+}}\varphi_{t}(x)\leq \varphi (x). $ The desired argument in i) follows immediately once the following is checked: for any $x\in \mathbb{R}^n$, \begin{equation}\label{inflimit}
\liminf_{t\rightarrow 0^+} \varphi_{t}(x)\geq \varphi(x).
\end{equation} To this end, let $x\in \mathbb{R}^n$ be fixed and $r>|\nabla \varphi(x)|$. Let $B_r$ be the Euclidean ball with center at the origin and radius $r$. For $p>1$, it can be checked, by $\varphi^*, \psi^*\geq 0$, that for all $t>0$, \begin{equation}\label{compare-5-25}
\varphi_t^*=\big((\varphi^*)^p+t(\psi^*)^p\big)^{\frac{1}{p}}\leq\varphi^*+ t^{\frac{1}{p}} \psi^*.
\end{equation}
It follows from
\eqref{compare-5-25} that, for $t\in (0, 1]$, \begin{align*}
\varphi_{t}(x) =\underset{y\in\mathbb{R}^n}{\sup}\Big\{\langle x,y\rangle -\varphi_{t}^*(y)\Big\} \geq \underset{y\in B_r}{\sup}\Big\{\langle x,y\rangle -\varphi_{t}^*(y)\Big\}\geq\underset{y\in B_r}{\sup}\Big\{\langle x,y\rangle-\varphi^*(y)-t^{\frac{1}{p}}\psi^*(y)\Big\}. \end{align*} Define the finite constant $c$ to be $c=\max\{\psi^*(y): y\in B_{r}\}$. The fact that $r>|\nabla \varphi(x)|$ implies $\nabla \varphi(x)\in B_r$, and hence, for $t\in (0, 1]$,
\begin{align*}
\varphi_{t}(x)
\geq\underset{y\in B_r}{\sup}\Big\{\langle x,y\rangle-\varphi^*(y)-t^{\frac{1}{p}}\psi^*(y) \Big\}
\geq \langle x,\nabla \varphi(x)\rangle-\varphi^*(\nabla \varphi(x))-t^{\frac{1}{p}}c=\varphi(x)-t^{\frac{1}{p}}c.
\end{align*}
The desired inequality (\ref{inflimit}) follows by letting $t\rightarrow 0^+$. This completes the proof for part i).
\vskip 2mm \noindent ii) This is a direct consequence of \cite[Theorem 25.7]{Roc70}; in a slight different form, it reads: if $\{\phi_i\}_{i\in \mathbb{N}\cup \{0\}}$ is a sequence of finite and differentiable convex functions on an open convex set $E$ such that $\phi_i\rightarrow \phi_0$ pointwisely on $E$, then $\nabla \phi_i\rightarrow \nabla \phi$ pointwisely on $E$ and uniformly on every closed bounded subset of $E$. \end{proof}
The following lemma provides the derivative of $\varphi_t$ with respect to $t$.
\begin{lemma}\label{variationorigin} Let $\varphi, \psi \in \mathscr L'_0$. For $p>1$ and for $t>0$, set $\varphi_{t}=\varphi \square_{p}( \psi \cdot_p t)$. Then for all $x\in \mathbb{R}^n$ and $t>0$, one has
\begin{equation}\label{derivative-varphi-5-27--1} \frac{d}{dt}\varphi_{t}(x)=-\frac{1}{p}\Big(\psi^*(\nabla \varphi_{t}(x))\Big)^p \Big(\varphi^*_{t}(\nabla \varphi_{t}(x))\Big)^{1-p}.\end{equation}
In particular, for all $x\in \mathbb{R}^n$, one has
\begin{equation}\label{derivative-varphi-5-27}
\frac{d}{dt}\varphi_{t}(x)\bigg|_{t=0^+}=-\frac{1}{p}\Big(\psi^*(\nabla \varphi(x))\Big)^p \Big(\varphi^*(\nabla \varphi(x))\Big)^{1-p}. \end{equation}
\end{lemma}
\begin{proof} Let $\varphi, \psi \in \mathscr L'_0$ and $p>1$. Formulas \eqref{derivative-varphi-5-27--1} and \eqref{derivative-varphi-5-27} clearly hold for $x=o$ following from Lemmas \ref{vainshat0} and \ref{belongtothesameclass} (the latter one gives $\varphi_t\in \mathscr L_0$ and hence $\nabla \varphi_t(o)=o$).
Let $x\neq o$. By Lemma \ref{belongtothesameclass} (and its proof), one sees that
the mapping $F$ defined by $
F(x, y, t) = \nabla \varphi_t^*(y)-x$
is continuously differentiable on $(\mathbb{R} ^n\setminus\{o\}) \times (\mathbb{R} ^n\setminus\{o\}) \times (0, + \infty)$. Note that $\frac{\partial F}{\partial y}=\nabla^2\varphi_t^*(y)$ is nonsingular for every $y \in \mathbb{R} ^n\setminus\{o\}$ by \eqref{second-hession}. The implicit function theorem yields (locally) the existence of a unique continuously differentiable mapping $y = y (x, t)$ for $(x, t)\in (\mathbb{R}^n\setminus\{o\})\times (0, \infty)$ such that $F(x, y(x, t), t)=o.$ That is, $x=\nabla\varphi_t^*(y(x, t)).$ According to Lemma \ref{belongtothesameclass}, $\nabla \varphi_{t}=\nabla^{-1} \varphi^*_{t}$. Thus, $y(x, t)=\nabla \varphi_t(x)$ and $x= \nabla\varphi_t^*\big(\nabla \varphi_t(x)\big)$ for $x\neq o$. Moreover, for $x\neq o$, one has $\varphi_{t}(x)=\langle x,\nabla \varphi_{t}(x)\rangle-\varphi^*_{t}(\nabla \varphi_{t}(x))$. Taking derivative from both sides, one gets, for any $t\in (0, \infty)$ and (fixed) $x\in \mathbb{R}^n\setminus\{o\}$,
\begin{align*}
\frac{d}{dt}\varphi_{t}(x)&=\langle x, \frac{d}{dt}\nabla \varphi_{t}(x)\rangle-\frac{1}{p}\Big(\psi^*(\nabla \varphi_{t}(x))\Big)^p \Big(\varphi^*_{t}(\nabla \varphi_{t}(x))\Big)^{1-p}-\langle\nabla\varphi^*_{t}(\nabla \varphi_{t}(x)), \frac{d}{dt}\nabla \varphi_{t}(x)\rangle\\
&=-\frac{1}{p}\Big(\psi^*(\nabla \varphi_{t}(x))\Big)^p \Big(\varphi^*_{t}(\nabla \varphi_{t}(x))\Big)^{1-p}.
\end{align*} This concludes the proof of \eqref{derivative-varphi-5-27--1}. Consequently, \eqref{derivative-varphi-5-27} follows from part ii) of Lemma \ref{continuityforgradient} and by letting $t\rightarrow 0^+$ in \eqref{derivative-varphi-5-27--1}.
\end{proof}
In order to obtain an explicit formula for $\delta J_p(f, g)$, we need to define the notion of admissible $p$-perturbation. See \cite{CF13} for the case for $p=1$.
\begin{definition}\label{p-perturbation} Let $f=e^{-\varphi}\in \mathscr A_0$ and $p>1$. The function $g=e^{-\psi}\in \mathscr A_0$ is said to be an admissible $p$-perturbation for $f$, if there exists a constant $c>0$ such that $(\varphi^*)^p-c (\psi^*)^p$ is a convex function. \end{definition}
Our main result in this section is the following integral formula for $\delta J_p(f, g)$. Again, we only focus on $p>1$ and the case $p=1$ has been covered in \cite[Theorem 4.5]{CF13}.
\begin{theorem}\label{variationformula}
Let $f=e^{-\varphi}\in \mathscr A'_0$ and $g=e^{-\psi}\in \mathscr A'_0$. For $p>1$, assume that $g$ is an admissible $p$-perturbation for $f$. In addition, suppose that there exists a constant $k>0$ such that
\begin{equation} \label{compatible-1} \det\Big(\nabla^2 (\varphi^*)^p (y) \Big)\leq k \big(\varphi^*(y)\big)^{n(p-1)} \det\big(\nabla^2 \varphi^* (y) \big)
\end{equation} holds for all $y\in \mathbb{R}^n\setminus\{o\}$. Then \begin{align}
\delta J_{p}(f,g)&=\frac{1}{p}\int_{\mathbb{R}^n}(\psi^*(\nabla \varphi(x)))^{p}(\varphi^*(\nabla\varphi(x)))^{1-p}e^{-\varphi(x)}\,dx \nonumber \\&=\frac{1}{p}\int_{\mathbb{R}^n}(\psi^*(y))^{p}(\varphi^*(y))^{1-p}\,d\mu(f, y).\label{tesi'}\end{align} \end{theorem}
\begin{proof} Let $f=e^{-\varphi}\in \mathscr A'_0$ and $g=e^{-\psi}\in \mathscr A'_0$. Let $t>0$ be fixed. According to \eqref{varphi-t}, Proposition \ref{close-Lp-A-sum}, and Definition \ref{variation-p-5-15-1}, one sees that \begin{equation}\label{difference-5-28}
\delta J_{p}(f,g)=\lim_{t\rightarrow0^+} \frac{J(f\oplus_{p} t\cdot_p g)-J(f)}{t}=\lim_{t\rightarrow0^+} \int_{\mathbb{R}^n} \frac{e^{-\varphi_t(x)}-e^{-\varphi(x)}}{t}\,dx. \end{equation} By Lemma \ref{variationorigin} (in particular, \eqref{derivative-varphi-5-27}), it holds that, for $x\in \mathbb{R}^n$, \begin{equation}\label{difference-5-28-1}
\lim _{t \rightarrow 0^+} \frac{e^{-\varphi_{t}(x)}-e^{-\varphi(x)}}{t}=\frac{1}{p}\Big(\psi^*(\nabla \varphi(x))\Big)^p \Big(\varphi^*(\nabla \varphi(x))\Big)^{1-p} e^{-\varphi(x)}.\end{equation} Consequently, the first formula in \eqref{tesi'} follows immediately from \eqref{difference-5-28} and \eqref{difference-5-28-1} once the dominated convergence theorem is verified. The second formula in \eqref{tesi'} follows directly from Definition \ref{def-moment-measure}.
Now let us verify that the dominated convergence theorem can be applied for \eqref{difference-5-28}. For any fixed $x\in \mathbb{R}^n$, Lemma \ref{continuityforgradient} implies that $\varphi_t(x)$ is continuous at $t=0$ and Lemma \ref{variationorigin} yields the differentiability of $\varphi_t(x)$ on $t\in (0, \infty)$. Together with \eqref{derivative-varphi-5-27--1}, the Lagrange mean value theorem can be applied to the function $t \mapsto e^{-\varphi_t(x)}$ and obtain that there exists an $s\in (0, t)$, such that \begin{equation}\label{Lagrange-5-28}\frac{e^{-\varphi_{t}(x)}-e^{-\varphi(x)}}{t-0}=\frac{d}{dt}e^{-\varphi_{t}(x)} \bigg|_{t=s}=\frac{1}{p}\Big(\psi^*(\nabla \varphi_{s}(x))\Big)^p \Big(\varphi^*_{s}(\nabla \varphi_{s}(x))\Big)^{1-p}e^{-\varphi_s(x)}.\end{equation} The function on the right, for any $s\in [0, t]$, is integrable over $\mathbb{R}^n$, namely,
\begin{align} \Psi(s)=\frac{1}{p}
\int_{\mathbb{R}^n} \Big(\psi^*(\nabla \varphi_{s}(x))\Big)^p \Big(\varphi^*_{s}(\nabla \varphi_{s}(x))\Big)^{1-p}e^{-\varphi_s(x)}\,dx<\infty \label{bounded-5-28-1}. \end{align} Indeed, \eqref{varphi-t-star} yields $s(\psi^*)^p\leq (\varphi_s^*)^p$ for any $s>0$. Together with \eqref{finitezza-1}, one has \begin{align*}
s \Psi(s) &= \frac{1}{p} \int_{\mathbb{R}^n} s\Big(\psi^*(\nabla \varphi_{s}(x))\Big)^p \Big(\varphi^*_{s}(\nabla \varphi_{s}(x))\Big)^{1-p}e^{-\varphi_s(x)}\,dx \\
&\leq \frac{1}{p} \int_{\mathbb{R}^{n}} \Big(\varphi^*_{s}(\nabla \varphi_{s}(x))\Big)^{p} \Big(\varphi^*_{s}(\nabla \varphi_{s}(x))\Big)^{1-p}e^{-\varphi_s(x)}\,dx\\
&= \frac{1}{p} \int_{\mathbb{R}^{n}} \varphi^*_{s}(\nabla \varphi_{s}(x))e^{-\varphi_s(x)}\,dx<+\infty.
\end{align*} For $s=0$, the assumption that $g$ is an admissible $p$-perturbation of $f$ is needed. Recall that $\varphi^*(o)=\psi^*(o)=0$ if $f=e^{-\varphi}\in \mathscr A_0$ and $g=e^{-\psi}\in \mathscr A_0$. According to Definition \ref{p-perturbation}, there exists a constant $c>0$, such that $(\varphi^*)^p-c (\psi^*)^p$ is a convex function. By Lemma \ref{vainshat0}, one has $(\varphi^*(o))^p-c (\psi^*(o))^p=0$ and $ \nabla\big((\varphi^*)^p-c(\psi^*)^p\big)(o)=0$, which in turn yields that, for any $y\in \mathbb{R}^n$,
\begin{eqnarray}\label{p-add-compare-1}
\big((\varphi^*)^p-c(\psi^*)^p\big)(y)\geq \big((\varphi^*)^p-c(\psi^*)^p\big)(o)- \langle y, \nabla\big((\varphi^*)^p-c(\psi^*)^p\big)(o) \rangle =0.
\end{eqnarray} Consequently, by \eqref{finitezza-1}, the following holds:
\begin{align*}
c \Psi(0)=\int_{\mathbb{R}^n} c\big(\psi^*(\nabla \varphi (x))\big)^p \big(\varphi^*(\nabla \varphi(x))\big)^{1-p} e^{-\varphi(x)}\,d x \leq \int_{\mathbb{R}^n} \varphi^*(\nabla \varphi(x)) e^{-\varphi(x)}\,d x <+\infty. \end{align*}
By \eqref{difference-5-28}, \eqref{Lagrange-5-28} and \eqref{bounded-5-28-1}, we obtain that, for all $s\in (0, t)$,
\begin{align}\delta J_{p}(f,g)
= \frac{1}{p}\lim_{s\to 0^+} \int_{\mathbb{R}^n} \big(\psi^*(\nabla \varphi_{s}(x))\big)^p \big(\varphi^*_{s}(\nabla \varphi_{s}(x))\big)^{1-p}e^{-\varphi_s(x)}\,dx. \label{Lagrange-5-28-22}\end{align} Let $y=\nabla \varphi_{s}(x)$. From Lemmas \ref{vainshat0} and \ref{belongtothesameclass} (in particular, its proof), \eqref{Lagrange-5-28-22} can be rewritten as
\begin{align}
\delta J_{p}(f,g) &= \frac{1}{p}\lim_{s\to 0^+} \int_{\mathbb{R}^n\setminus\{o\}} \big(\psi^*(\nabla \varphi_{s}(x))\big)^p \Big(\varphi^*_{s}(\nabla \varphi_{s}(x))\Big)^{1-p}e^{-\varphi_s(x)}\,dx \nonumber \\ &= \frac{1}{p}\lim_{s\to 0^+} \int_{\mathbb{R}^n\setminus\{o\}} \big(\psi^*(y)\big)^p \Big(\varphi^*_{s}(y)\Big)^{1-p}e^{-\varphi_s(\nabla\varphi_s^*(y))} \det \big(\nabla^2\varphi^*_s(y)\big) \,dy. \label{Lagrange-5-31-22}
\end{align} The desired formula \eqref{tesi'} follows once the dominated convergence theorem is verified for \eqref{Lagrange-5-31-22}.
According to Lemma \ref{belongtothesameclass}, one has $\varphi^*_t\in \mathcal C^2_+(\mathbb{R}^n\setminus\{o\})$. For $y\neq o$, \eqref{second-hession} can be rewritten as \begin{align} \nabla^2\big((\varphi_t^*)^p\big)(y)&=\nabla^2\big((\varphi^*)^p\big)(y)+t\nabla^2\big((\psi^*)^p\big)(y)\nonumber \\&= \Big(p(p-1)\big(\varphi_t^*\big)^{p-2} \nabla\varphi^*_t\otimes\nabla\varphi^*_t+p\big(\varphi_t^*\big)^{p-1}\nabla^2\varphi^*_t\Big)(y). \label{second-hession-5-30} \end{align} For a positive definite matrix $A$ and a semi-definite matrix $B$, the following holds: \begin{equation} \label{matrix-ineq} \det (A+B)\geq \det A+\det B,\end{equation} where $\det A$ denotes the determinant of $A$. This inequality can be applied to \eqref{second-hession-5-30} to get \begin{align} p^n \big(\varphi_t^*(y)\big)^{n(p-1)} \det\big(\nabla^2\varphi^*_t(y)\big) \leq \det \big(\nabla^2(\varphi_t^*)^p(y)\big) =\det\big(\nabla^2(\varphi^*)^p(y)+t\nabla^2(\psi^*)^p(y)\big). \label{second-hession-5-30-1} \end{align}
As $g=e^{-\psi}\in \mathscr A'_0$ is an admissible $p$-perturbation for $f$, there exists a constant $c>0$ such that $\phi=(\varphi^*)^p-c (\psi^*)^p$ is convex. Hence $\nabla^2 (\psi^*)^p =\frac{1}{c} \nabla^2 (\varphi^*)^p -\frac{1}{c} \nabla^2 \phi.$ It follows from \eqref{compatible-1}, \eqref{matrix-ineq} and \eqref{second-hession-5-30-1} that, for $y\neq o$ and $p>1$,
\begin{align*}
\det\big(\nabla^2\varphi^*_t(y)\big)
& \leq p^{-n} \big(\varphi_t^*(y)\big)^{n(1-p)} \det\Big(\nabla^2 (\varphi^*)^p (y)+t\nabla^2 (\psi^*)^p (y)\Big)\\
& \leq p^{-n} \big(\varphi^*(y)\big)^{n(1-p)} \det\Big(\nabla^2 (\varphi^*)^p (y)+\frac{t}{c} \nabla^2 (\varphi^*)^p(y) -\frac{t}{c} \nabla^2 \phi(y)\Big)\\
&\leq p^{-n} \big(\varphi^*(y)\big)^{n(1-p)} \det\Big(\nabla^2 (\varphi^*)^p (y)+\frac{t}{c} \nabla^2 (\varphi^*)^p(y) \Big)\\
&\leq \bigg(\frac{t+c}{cp}\bigg)^{n} \big(\varphi^*(y)\big)^{n(1-p)} \det\Big(\nabla^2 (\varphi^*)^p (y) \Big)\\
& \leq k\bigg(\frac{t+c}{cp}\bigg)^{n} \det\big(\nabla^2 \varphi^* (y) \big),
\end{align*} where the constant $k>0$ is given by \eqref{compatible-1}. Let $k_0=k\big(\frac{1+c}{cp}\big)^{n}.$ Hence, if $t\in (0, 1)$, one gets \begin{align}\label{compatible-2} \det \big(\nabla^2\varphi^*_t(y)\big)\leq k_0 \det\big(\nabla^2\varphi^* (y) \big). \end{align}
For any $o\neq y\in \mathbb{R}^n$, by \eqref{def-dual-2}, \eqref{first-order-deriv} and Lemma \ref{vainshat0}, one gets \begin{align*} \frac{\,d }{\,dt}\Big(\big( \varphi^*_t(y)\big)^p-\Big\langle y, \big(\varphi^*_t(y)\big)^{p-1}\nabla \varphi^*_t(y)\Big\rangle\Big) &= \big( \psi^*(y)\big)^p-\Big\langle y, \big(\psi^*(y)\big)^{p-1}\nabla \psi^*(y)\Big\rangle\\ &= \big( \psi^*(y)\big)^{p-1} \Big(\psi^*(y)-\big\langle y, \nabla \psi^*(y)\big\rangle \Big) \\&=-\big( \psi^*(y)\big)^{p-1} \psi(\nabla \psi^*(y) \leq 0. \end{align*}
Together with \eqref{varphi-t-star} and \eqref{p-add-compare-1}, one gets that, for $p>1$, $o\neq y\in\mathbb{R}^n,$ and $t\in (0, 1)$,
\begin{align}
\varphi^*_t(y)-\big\langle y, \nabla \varphi^*_t(y)\big\rangle
& \leq \big(\varphi^*_t(y)\big)^{1-p}\ \Big( \big(\varphi^*(y)\big)^p-\Big\langle y, \big(\varphi^*(y)\big)^{p-1}\nabla \varphi^*(y)\Big\rangle\Big) \nonumber \\ &=\bigg(\frac{\varphi^*(y)} {\varphi^*_t(y)}\bigg)^{p-1} \big(\varphi^*(y)- \langle y, \nabla \varphi^*(y)\rangle\big) \nonumber \\
&\leq \Big(\frac{c}{1+c}\Big)^{\frac{p-1}{p}} \big(\varphi^*(y)- \langle y, \nabla \varphi^*(y)\rangle\big)\leq 0.
\label{s-compare-5}
\end{align}
Formulas \eqref{varphi-t-star} and \eqref{p-add-compare-1} also yield that, for any $t>0$, \begin{eqnarray} (\psi^*)^p\leq c^{-1} \varphi^* (\varphi_t^*)^{p-1}.\label{p-add-com-222} \end{eqnarray}
Now we are ready to check the interchange of orders of limit and integration in \eqref{Lagrange-5-31-22}. Combining \eqref{def-dual-2}, \eqref{compatible-2}
\eqref{s-compare-5},
\eqref{p-add-com-222}, one has, for $s\in (0, 1)$ and $o\neq y\in \mathbb{R}^n$, \begin{align*} h(y)&=k_0 c^{-1} \varphi^*(y) \exp\bigg(\Big(\frac{c}{1+c}\Big)^{\frac{p-1}{p}} \big(\varphi^*(y)- \langle y, \nabla \varphi^*(y)\rangle\big)\bigg) \det \big(\nabla^2\varphi^*(y)\big) \\ &\geq \big(\psi^*(y)\big)^p \Big(\varphi^*_{s}(y)\Big)^{1-p}e^{-\varphi_s(\nabla\varphi_s^*(y))} \det \big(\nabla^2\varphi^*_s(y)\big).\end{align*} The function $h$ is integrable by \eqref{const-mul-dual} and \eqref{finitezza-1}. Indeed, a calculation similar to \eqref{Lagrange-5-31-22} leads that \begin{align*} \int_{\mathbb{R}^n\setminus\{o\}} h(y)\,dy &=
k_0 c^{-1} \int_{\mathbb{R}^n\setminus\{o\}} \varphi^*(y) \exp\bigg(\Big(\frac{c}{1+c}\Big)^{\frac{p-1}{p}} \big(\varphi^*(y)- \langle y, \nabla \varphi^*(y)\rangle\big)\bigg) \det \big(\nabla^2\varphi^*(y)\big)\,dy\\ &=
k_0 c^{-1} \int_{\mathbb{R}^n} \varphi^*(\nabla\varphi(x)) \exp\Big(-\Big(\frac{c}{1+c}\Big)^{\frac{p-1}{p}} \varphi(x) \Big) \,dx\\ &=k_0 c^{-1} \Big(\frac{c}{1+c}\Big)^{\frac{1-p}{p}} \int_{\mathbb{R}^n} \widetilde{\varphi}^*(\nabla\widetilde{\varphi}(x)) e^{- \widetilde{\varphi}(x)} \,dx \in (-\infty, \infty),
\end{align*} where $\widetilde{\varphi}=(\frac{c}{1+c})^{\frac{p-1}{p}} \varphi$. Therefore, the dominated convergence theorem can be applied to \eqref{Lagrange-5-31-22} and the desired formula \eqref{tesi'} holds. \end{proof}
The condition \eqref{compatible-1} is indeed natural and many widely used functions do satisfy this condition, for example, the Gaussian function $e^{-|x|^2/2}$. The following results give some convenient ways to check condition \eqref{compatible-1}.
\begin{corollary} \label{verify-condition-1} Assume that $\varphi\in \mathcal C^2_+(\mathbb{R}^n\setminus\{o\})$. Then \eqref{compatible-1} holds if any one of the following holds:
\vskip 2mm \noindent i) for some $\alpha\in [0, 1)$, the function $\frac{1}{\alpha} (\varphi^*)^{\alpha}$ (understood as $\log \varphi^*$ when $\alpha=0$) is convex; \vskip 2mm \noindent ii) there exists a constant $k_1$, such that, for any $y\in \mathbb{R}^n\setminus\{o\}$, \begin{align} \Big\langle \nabla \varphi^*(y), \big(\nabla^2\varphi^*(y)\big)^{-1} \nabla \varphi^*(y)\Big\rangle \leq k_1 \varphi^*(y).\label{condition-5-31-1} \end{align}
\end{corollary}
\begin{proof} For $p>1$, condition \eqref{compatible-1} is equivalent to, for any $y\neq o$,
\begin{equation} \label{compatible-1-22} H(y)= \det\Big(\nabla^2 \varphi^* (y)+
(p-1)
(\varphi^*(y))^{-1} \nabla\varphi^*(y) \otimes \nabla\varphi^* (y)\Big)\leq k p^{-n} \det\big(\nabla^2 \varphi^* (y) \big).
\end{equation}
\vskip 2mm \noindent i) Let $\alpha\in [0, 1)$ and the function $\frac{1}{\alpha} (\varphi^*)^{\alpha}$ be convex. For any $o\neq y\in \mathbb{R}^n$, $$A(y)=\nabla^2 \varphi^* (y)+(\alpha-1)(\varphi^*(y))^{-1} \nabla\varphi^*(y)\otimes \nabla\varphi^*(y) $$ is a positive semi-definite matrix. Therefore, \eqref{matrix-ineq} yields
\begin{align*}
H(y) &=\det\!\Big(\!
\Big(1+\frac{p-1}{1-\alpha}\Big)
\nabla^2 \varphi^* (y)+
\frac{p-1}{\alpha-1}
A (y)\!\Big) \leq \Big(1+ \frac{p-1}{1-\alpha}\Big)^n \det\Big(\nabla^2 \varphi^* (y)\Big) .
\end{align*}
Hence, \eqref{compatible-1-22} and condition \eqref{compatible-1} hold true.
\vskip 2mm \noindent ii) It can be calculated that $\mathbb{I}_n+z\otimes z$ for any $z\in \mathbb{R}^n$ has its determinant to be $1+|z|^2$. Hence, for any $o\neq y\in \mathbb{R}^n$, \eqref{condition-5-31-1} yields \begin{align*} H(y) &=\det\big(\nabla^2 \varphi^* (y)\big)\Big(1+ (p-1) (\varphi^*(y))^{-1} \Big\langle \nabla \varphi^*(y), \big(\nabla^2\varphi^*(y)\big)^{-1} \nabla \varphi^*(y)\Big\rangle\Big) \\ &\leq \det\big(\nabla^2 \varphi^* (y)\big)\Big(1+k_1 (p-1) \Big). \end{align*} This implies \eqref{compatible-1-22} and hence condition \eqref{compatible-1} holds ture.
\end{proof}
\section{The $L_p$ Minkowski problem for log-concave functions}\label{functional p Minkowski problem} \setcounter{equation}{0}
This section aims to investigate the $L_p$ Minkowski problem for log-concave functions. Actually Theorem \ref{variationformula} suggests a new measure for log-concave functions. Let $\mathscr L^+=\{\varphi\in \mathscr L: \varphi\geq 0\}$, $\Omega_{\varphi^*}=\{y\in \mathbb{R}^n: 0<\varphi^*(y)<+\infty\}$ and $\widetilde{\Omega}_{\varphi^*}=\{y\in \mathbb{R}^n: \varphi^*(y)=0\}.$ The set $\Omega_{\varphi^*} $ is always assumed to be nonempty. The subscript $\varphi^*$ is often omitted if there is no confusion.
The $L_p$ surface area measure of $f$ can be defined as follows.
\begin{definition} \label{p-def-moment-measure} Let $f=e^{-\varphi}$ be a log-concave function with $\varphi \in \mathscr L$ such that $\varphi^*\in \mathscr L^+$ and $\Omega$ is nonempty. For $p\in \mathbb{R}$, the $L_p$ surface area measure of $f$, denoted by $\mu_p(f, \cdot)$, is the Borel measure on $\Omega$ such that \begin{equation}\label{moment-form-p-6-1} \int_{\Omega}g(y)\, d\mu_p (f, y)=\int_{\{x\in {\mathrm {dom}}(\varphi):\ \nabla \varphi(x)\in \Omega \}}g(\nabla \varphi(x))(\varphi^*(\nabla \varphi(x)))^{1-p} e^{-\varphi(x)} \,dx\end{equation} holds for every Borel function $g$ such that $g \in L ^ 1 (\mu_p (f,\cdot))$ or $g$ is non-negative. \end{definition}
In general, $\,d\mu_{p}(f, \cdot)=(\varphi^*)^{1-p} \,d\mu_1(f, \cdot)$ on $\Omega$. The $L_p$ surface area measure of $f$ for $p=1$ in Definition \ref{p-def-moment-measure} is the restriction of the surface area measure $f$ given in \eqref{moment-form-1} on $\Omega$, i.e., $\mu_1(f, \cdot)=\mu(f, \cdot)|_{\Omega}$. If the Lebesgue measure of $\widetilde{\Omega}$ is zero, $\mu_1(f, \cdot)$ can be extended to $\mathbb{R}^n$ (more precisely the interior of ${\mathrm {dom}}(\varphi^*)$), and reduces to $\mu(f, \cdot)$. The measure $\mu_1(f, \cdot)$ is always finite for $f\in \mathscr A$.
As $\varphi^*=0$ on $\widetilde{\Omega}$, one can even extend $\mu_p(f, \cdot)$ for $p<1$ to $\mathrm{int}({\mathrm {dom}}(\varphi^*))$.
When $p=0$, one has the $L_0$ (or the logarithmic) surface area measure of $f$ which is again a finite measure for $f\in \mathscr A$ based on \eqref{finitezza-1}.
A natural problem to characterize the $L_p$ surface area measure of a log-concave function $f$ can be formulated as follows.
\begin{problem}[The $L_p$ Minkowski problem for log-concave functions] \label{problem-lp-6-2} Let $\nu$ be a finite nonzero Borel measure on $\mathbb{R}^n$ and $ p\in \mathbb{R}$. Find the necessary and/or sufficient conditions on $\nu$, so that, $$\nu=\tau \mu_{p}(f,\cdot)\ \ \ \mathrm{or}\ \ \ (\varphi^*)^{p-1}\nu =\tau\mu_1(f, \cdot)$$ hold for some log-concave function $f=e^{-\varphi}$ and $\tau\in\mathbb{R}$.\end{problem}
When $p=1$ and the Lebesgue measure of $\widetilde{\Omega}$ is zero, Problem \ref{problem-lp-6-2} for $p=1$ reduces to the Minkowski problem for moment measures investigated in \cite{CK15} by Cordero-Erausquin and Klartag. See \cite[Section 7]{CF13} for a full formulation to this problem. In this case, a solution has been provided in \cite{CK15}, including \cite[Proposition 1]{CK15} for necessity and \cite[Theorem 2]{CK15} for sufficiency and uniqueness.
When $f=e^{-\varphi}$ is smooth enough so that $\nabla \varphi: \mathrm{int}({\mathrm {dom}}(\varphi)) \rightarrow \mathrm{int}({\mathrm {dom}}(\varphi^*))$ is smooth and bijective, then formula \eqref{def-dual-2} and Definition \ref{p-def-moment-measure} deduce that, by letting $y=\nabla \varphi(x)$,
\begin{align*}
\int_{\Omega}g(y)\, d\mu_p (f, y)&=\int_{\{x\in \mathbb{R}^n:\ \nabla \varphi(x)\in \Omega \}}g(\nabla \varphi(x))(\varphi^*(\nabla \varphi(x)))^{1-p} e^{-\varphi(x)} \,dx \\
&=\int_{\Omega}g(y) \varphi^*(y)^{1-p}\det(\nabla^2\varphi^*(y))e^{\varphi^*(y)-\langle y, \nabla\varphi^*{y}\rangle}\,dy,
\end{align*} holds for every Borel function $g$ such that $g \in L ^ 1 (\mu_p (f,\cdot))$ or $g$ is non-negative. That is, $\mu_{p}(f,\cdot)$ for $p\in\mathbb{R}$ is absolutely continuous with respect to the Lebesgue measure and satisfies that
\begin{align*}
\frac{d\mu_{p}(f, y)}{dy}=\varphi^*(y)^{1-p} e^{\varphi^*(y)-\langle y, \nabla\varphi^*{y}\rangle} \det(\nabla^2\varphi^*(y)) \ \ \mathrm{for} \ \ y\in \Omega. \end{align*} Consequently, if $\nu$ admits a density function with respect to the Lebesgue measure, say $\,d\nu(y)=h(y)\,dy,$ then finding a solution to the $L_p$ Minkowski problem for log-concave functions requires to obtain a (smooth enough) convex function $\varphi$ satisfying the following Monge-Amp\`{e}re equation:
\begin{align*}
h(y) &= \tau \varphi^*(y)^{1-p} e^{\varphi^*(y)-\langle y, \nabla\varphi^*{y}\rangle} \det(\nabla^2\varphi^*(y)), \ \ \ \mathrm{for} \ y\in \Omega,
\end{align*} where $\tau\in\mathbb{R}$ is a constant.
Our main goal in this section is to provide a solution to Problem \ref{problem-lp-6-2} for $p>1$.
Let $M_{\nu}$ be the interior of the convex hull of the support of the Borel measure $\nu$. For convenience, denote by $\mathscr{M}$ the set of all {\em even finite nonzero Borel measures} on $\mathbb{R}^n$, such that, if $\nu\in \mathscr{M}$, then $\nu$ is not supported in a lower-dimensional subspace of $\mathbb{R}^n$, $\nu(M_{\nu}\setminus L)>0$ holds for any bounded convex set $L\subset\mathbb{R}^n$, and
\begin{equation} \label{p-th-moment} \int_{\mathbb{R}^n} |x|^p\, d\nu(x)<\infty.\end{equation} Note that the conditions for $\nu\in \mathscr{M}$ are all natural. Indeed, the assumption that $\nu$ is not supported in a lower-dimensional subspace of $\mathbb{R}^n$ is essential in the solution to the Minkowski problem for moment measures in \cite{CK15}. The requirement for \eqref{p-th-moment} is to guarantee that the optimization problem \eqref{mini problem} is not taken over an empty set. Finally, the condition that $\nu(M_{\nu}\setminus L)>0$ holds for any bounded convex set $L\subset\mathbb{R}^n$ is to guarantee that $\nu(M_{\nu}\setminus \widetilde{\Omega}_{\varphi})>0$ for any $\varphi$ and thus avoid that $\Omega_{\varphi}$ being either empty or a null set.
\begin{theorem} \label{solution-lp-mink-6-2} Let $\nu\in \mathscr{M}$. For $p>1$, there exists an even log-concave function $f=e^{-\varphi}$, such that, $\varphi\in \mathcal C$ is even and lower semi-continuous, $\varphi^*\in \mathscr L^+$, and
\begin{eqnarray*}
\nu =\tau (\varphi^*)^{1-p} \mu_{1}(f,\cdot) =\tau \mu_{p}(f, \cdot)\ \ \ \mathrm{on} \ \ \Omega,
\end{eqnarray*} where the constant $\tau$ takes the following formula:
$$ \tau= \frac{\int_{\Omega} (\varphi^*(y))^{p-1}\,d\nu(y)} {\int_{\Omega} \,d \mu_{1}(f,y) } = \frac{\int_{\Omega} \,d\nu(y)} {\int_{\Omega} \,d \mu_{p}(f,y) } .$$ \end{theorem}
Before we prove Theorem \ref{solution-lp-mink-6-2}, we need some preparation. Let $p>1$ and $\nu\in \mathscr{M}$. Denote by $L_{p, e} (\nu)$ the set of even non-negative functions on $\mathbb{R}^n$ which have finite $L^p$ norm with respect to the measure $\nu$ and whose function values are $0$ at $o$. For $\phi\in L_{p, e} (\nu)$, let
\begin{align*}
\mathbf{\Phi}_{p,\nu}(\phi)=\frac{1}{p}\int_{\mathbb{R}^n} \!\! \phi(x)^pd\nu(x)-\log J(e^{-\phi^*}).
\end{align*} To find a solution to Problem \ref{problem-lp-6-2},
one needs to search for a solution to
\begin{eqnarray}\label{mini problem}
\Theta=\inf\Big\{\mathbf{\Phi}_{p,\nu}(\phi): \phi \in L_{p, e}(\nu) \ \ \mathrm{and} \ \ 0< J(e^{-\phi^*})<\infty \Big\}.
\end{eqnarray}
The following lemma shall be needed to solve \eqref{mini problem}. Similar results for $p=1$ can be found in \cite[Lemmas 14 and 15]{CK15}.
\begin{lemma}\label{lemma1.3} Let $p>1$ and $\nu\in \mathscr{M}$. Then there exists a constant $c_{\nu}>0$, such that, for any $\phi \in L_{p, e}(\nu)\cap \mathcal C$ satisfying $0<J(e^{-\phi^*})<\infty$, the following holds:
\begin{eqnarray} \int_{\mathbb{R}^n}\phi(x)^p\,d\nu(x)\geq c_{\nu}\big(J(e^{-\phi^*})\big)^{\frac{1}{n}}-\int_{\mathbb{R}^n} \,d\nu(x).\label{equation-6-10}\end{eqnarray}
\end{lemma}
\begin{proof} Let $p>1$ and $\nu\in \mathscr{M}$. Assume that $0<J(e^{-\phi^*})<\infty$. For any $\phi \in L_{p, e}(\nu)\cap \mathcal C$, $\phi^*$ must be an even convex function. It is well-known that there exist constants $c_1, c_2>0$ such that for any $n\in \mathbb{N}$ and any even log-concave function $f=e^{-\phi}$ with $J(e^{-\phi})\in (0, \infty)$, the following holds: \begin{align}\label{santalo-func-6-10} c_1^n\leq J(e^{-\phi}) J(e^{-\phi^*})\leq c_2^n. \end{align} We refer the readers to \cite[Theorem 1.1]{KM05} for more details on inequality \eqref{santalo-func-6-10}. The sharp constant for $c_2$ is $2\pi$ and the upper bound is indeed the Blaschke-Santal\'{o} inequalities for (even) log-concave functions, see e.g., \cite{AKM04, Ball88-2, FM07,FM08}. Applying inequality \eqref{santalo-func-6-10} to the log-concave function $e^{-\phi^*}$, one gets that $0<J(e^{-\phi})<\infty$.
For $\phi \in L_{p, e}(\nu)\cap \mathcal C$ satisfying $0<J(e^{-\phi^*})<\infty$, let
$ K_{\phi}=\{x\in \mathbb{R}^n: \phi(x)\leq 1\}.$ Clearly, $K_{\phi}$ is an origin-symmetric bounded convex set containing the origin $o$ in its interior (due to $\phi(o)=0$). As $0< J(e^{-\phi})<\infty$, one has $V(K_{\phi})<\infty$. Denote by $\mathrm{vrad}(\phi)$ the volume radius of $K_{\phi}$, that is, $$\mathrm{vrad}(\phi)=\bigg(\frac{V(K_{\phi})}{V(B_1)}\bigg)^{1/n}\in (0, \infty).$$ Note that $K_{\phi}$ cannot contain an Euclidean ball whose radius is greater than $\mathrm{vrad}(\phi)$. Consequently, one may find a vector $\theta_0\in S^{n-1}$ such that \begin{eqnarray}\label{an upper bound}
\sup_{x\in K_{\phi}}|\langle x,\theta_0\rangle|=\sup_{x\in K_{\phi}}\langle x,\theta_0\rangle \leq \mathrm{vrad}(\phi).
\end{eqnarray}
The fact that $\phi$ is convex yields that, for $p>1$ and for any $x\in \mathbb{R}^n$ with $|\langle x,\theta_0\rangle|\geq \mathrm{vrad}(\phi)$, \begin{eqnarray}\label{an upper bound2}
\phi^p\left(\frac{ \mathrm{vrad}(\phi)}{|\langle x,\theta_0\rangle|}x\right)&\leq& \frac{ \mathrm{vrad}(\phi)}{|\langle x,\theta_0\rangle|}\phi^p(x).\end{eqnarray} According to \eqref{an upper bound} and \eqref{an upper bound2}, if $x\in \mathbb{R}^n$ such that $|\langle x,\theta_0\rangle|\geq \mathrm{vrad}(\phi)$, then $\phi(x)\geq 1$ and
\begin{align}\label{an upper bound3}
\phi^p(x)\geq \frac{|\langle x,\theta_0\rangle|}{\mathrm{vrad}(\phi)} \phi^p\left(\frac{\mathrm{vrad}(\phi)}{|\langle x,\theta_0\rangle|}x\right)
\geq \frac{|\langle x,\theta_0\rangle|}{\mathrm{vrad}(\phi)} \geq \frac{|\langle x,\theta_0\rangle|}{\mathrm{vrad}(\phi)} -1.
\end{align} Indeed, (\ref{an upper bound3}) holds for all $x\in \mathbb{R}^n$ as it is trivial to have (\ref{an upper bound3}) for those $x\in \mathbb{R}^n$ such that $|\langle x,\theta_0\rangle|< \mathrm{vrad}(\phi)$, due to $\phi(x)\geq 0$. Consequently,
\begin{eqnarray*}
\int_{\mathbb{R}^n}\phi^p(x)d\nu(x)\geq \frac{1}{ \mathrm{vrad}(\phi) }\int_{\mathbb{R}^n}|\langle x,\theta_0\rangle|d\nu(x)-\int_{\mathbb{R}^n} \,d\nu(x)\geq \frac{m_{\nu}}{ \mathrm{vrad}(\phi) }-\int_{\mathbb{R}^n} \,d\nu(x),
\end{eqnarray*}
where $m_{\nu}=\inf_{\theta\in S^{n-1}}\int_{\mathbb{R}^n}|\langle x,\theta\rangle|\, d\nu(x)$. Note that $m_{\nu}\in (0, \infty)$ is a direct consequence of the conditions on $\nu\in \mathscr{M}$, in particular, the function $\theta \mapsto \int_{\mathbb{R}^n}|\langle x,\theta\rangle|\, d\nu(x)$ is positive and continuous on $\theta$ due to the dominated convergence theorem. On the other hand,
\begin{eqnarray*} \int_{\mathbb{R}^n}e^{-\phi(x)}\,dx \geq\int_{K_{\phi}}e^{-\phi(x)}\,dx \geq\frac{V(K_{\phi})}{e}=\frac{V(B_1)}{e} \mathrm{vrad}(\phi)^n. \end{eqnarray*} Therefore, an application of inequality \eqref{santalo-func-6-10} with $c_2=2\pi$ immediately yields
\begin{align*} \int_{\mathbb{R}^n}\phi^p(x)d\nu(x) &\geq \frac{ (V(B_1))^{1/n} m_{\nu}}{e^{1/n}} \bigg(\int_{\mathbb{R}^n} e^{-\phi(x)}\,dx\bigg)^{-\frac{1}{n}}-\int_{\mathbb{R}^n} \,d\nu(x) \nonumber \\ &
\geq c_{\nu} \bigg(\int_{\mathbb{R}^n} e^{-\phi^*(x)}\,dx\bigg)^{\frac{1}{n}}-\int_{\mathbb{R}^n} \,d\nu(x), \end{align*}
by letting $c_{\nu}=\frac{ (V(B_1))^{1/n} m_{\nu}}{2\pi e^{1/n}}$. This concludes the desired formula \eqref{equation-6-10}.
\end{proof}
It has been proved in \cite[Lemma 16]{CK15} that, if $\nu$ is a nonzero finite Borel measure on $\mathbb{R}^n$ that is not supported in a lower-dimensional subspace of $\mathbb{R}^n$, for $x_0\in M_{\nu}$, then there exists a constant $C_{\nu, x_0}>0$ with the following property: $$
\phi(x_0)\leq C_{\nu,x_0}\int_{\mathbb{R}^n}\phi(x) \,d\nu(x)
$$ holds for any $\nu$-integrable convex function $\phi: \mathbb{R}^n\rightarrow [0,\infty]$. This can be applied to $\nu\in \mathscr{M}$ and $p>1$ to get that, for $x_0\in M_{\nu}$, then there exists a constant $C_{\nu, x_0}>0$, such that, \begin{align}\label{estimate-p-6-12}
\phi^p(x_0)\leq C_{\nu,x_0}\int_{\mathbb{R}^n}\phi^p(x) \,d\nu(x)
\end{align} holds for any $\phi\in \mathcal C\cap L_{p, e}(\nu)$.
We shall need the following lemma given by \cite[Theorem 10.9]{Roc70}.
\begin{lemma}\label{Rockafellar's result} Let $C$ be a relatively open convex set, and let $\phi_1, \phi_2,\cdots,$ be a sequence of finite convex functions on $C$. Suppose that the real number $\phi_1(x), \phi_2(x),\cdots,$
is bounded for each $x\in C$. It is then possible to select a subsequence of $\phi_1, \phi_2,\cdots,$ which converges uniformly on closed bounded subsets of $C$ to some finite convex function $\phi$.
\end{lemma}
We are now ready to prove the following lemma. The case $p=1$ has been discussed in \cite[Lemma 17]{CK15}.
\begin{lemma}\label{existence}Let $p>1$ and $\nu\in \mathscr{M}.$ Assume that $\phi_l\in \mathcal C\cap L_{p, e}(\nu)$ for any $l\in \mathbb{N}$ satisfy
\begin{eqnarray}\label{condition1}
\sup_{l\in \mathbb{N}} \int_{\mathbb{R}^n} \phi_l(x)^p\, d\nu(x)<+\infty.
\end{eqnarray}
Then there exists a subsequence $\{\phi_{l_{j}}\}_{j\in \mathbb{N}}$ of $\{\phi_l\}_{l\in \mathbb{N}}$ and a non-negative convex function $\phi \in \mathcal C\cap L_{p, e}(\nu)$, such that,
\begin{eqnarray}\label{liminf}
\int_{\mathbb{R}^n} \phi(x)^p \,d\nu(x) \leq \liminf_{j\rightarrow\infty} \int_{\mathbb{R}^n} \phi_{l_j}(x)^p \,d\nu(x) \ \ \mathrm{and}\ \
\int_{\mathbb{R}^n}e^{-\phi^*(x)}\,dx\geq \limsup_{j\rightarrow\infty}\int_{\mathbb{R}^n}e^{-\phi_{l_j}^*(x)}\,dx.
\end{eqnarray}
\end{lemma}
\begin{proof} We prove this lemma following the ideas of the proof of \cite[Lemma 17]{CK15}, with emphasis on the difference and modification.
As $\nu\in \mathscr{M}$ is an even measure which is not supported in a lower-dimensional subspace of $\mathbb{R}^n$, the open set $M_{\nu}$ is nonempty and origin-symmetric with $o\in M_{\nu}$. By \eqref{estimate-p-6-12} and \eqref{condition1}, for any $x\in M_{\nu}$, one has $
\sup_{l\in \mathbb{N}} \phi_l(x)^p<+\infty. $ We would like to mention that, if $\phi\in L_{p, e} (\nu)$, then $\phi$ is finite near the origin and ${\mathrm {dom}}(\phi)\supseteq M_{\nu}$. Lemma \ref{Rockafellar's result} can be applied to $\phi_l^p$ and $C=M_{\nu}$ to obtain the existence of a subsequence $\{\phi_{l_j}\}_{j\in \mathbb{N}}$ of $\{\phi_l\}_{l\in \mathbb{N}}$, which converges to an even convex function $\phi: M_{\nu} \rightarrow \mathbb{R}$ pointwisely on $M_{\nu}$ and also uniformly on any closed bounded subset of $M_{\nu}$. The finiteness of $\phi$ on $M_{\nu}$ implies the continuity of $\phi$ on $M_{\nu}$. Moreover, $\phi$ is non-negative in $M_{\nu}$ and achieves its minimum at the origin with $\phi(o)=0$ (as $\phi_l(o)=0$ for each $l\in \mathbb{N}$). The function $\phi: M_{\nu}\rightarrow \mathbb{R}$ can be extended on $\mathbb{R}^n$, and the new function will still be denoted by $\phi$. That is, let $\phi(x)=+\infty$ for $x\notin \overline{M_{\nu}}$; while $\phi(x)=\lim_{\lambda\rightarrow 1^-} \phi(\lambda x)$ if $x\in \partial M_{\nu}$. The limit in the latter case always exists, although it may be $+\infty$, due to the fact that the function $\lambda\mapsto \phi(\lambda x)$ is increasing on $\lambda\in (0, 1)$ following from the convexity of $\phi$ and $\phi(o)=0$. Moreover, for any $x\in \overline{M_{\nu}}$, $\phi(\lambda x)$ is increasing to $\phi(x)$ as $\lambda$ is increasing to $1$. This shows that $\phi: \mathbb{R}^n\to \mathbb{R}$ is an even, non-negative, and lower semi-continuous convex function with $\phi(o)=0$.
Note that the support of $\nu$ is a subset of $\overline{M_{\nu}}$. It follows from Fatou's lemma and $\phi_{l_j}\rightarrow \phi$ pointwisely in $M_{\nu}$ that, for any given $\lambda\in (0, 1)$, one has \begin{align*}
\int_{\mathbb{R}^n} \phi(\lambda x)^p\,d\nu(x) = \int_{M_{\nu}} \phi(\lambda x)^p\,d\nu(x)& \leq \liminf_{j\rightarrow\infty} \int_{M_{\nu}}\phi_{l_{j}}(\lambda x)^pd\nu(x) \nonumber \\ & \leq \liminf_{j\rightarrow\infty} \int_{\mathbb{R}^n}\phi_{l_{j}}(x)^pd\nu(x)<+\infty,
\end{align*} where the second inequality again follows from the monotonicity of the function $\lambda\mapsto \phi_{l_j}(\lambda x).$ Similarly, by the monotone convergence theorem, one can also obtain that \begin{eqnarray*}
\int_{\mathbb{R}^n}\phi(x)^p\,d\nu(x)= \lim_{\lambda\rightarrow 1^{-}}\int_{\mathbb{R}^n}\phi(\lambda x)^p\,d\nu(x)\leq \liminf_{j\rightarrow\infty} \int_{\mathbb{R}^n}\phi_{l_{j}}(x)^p\,d\nu(x)<+\infty.
\end{eqnarray*}
This completes the proof of the first argument in (\ref{liminf}). In particular, $\phi\in L_{p, e}(\nu)\cap \mathcal C$, $o\in M_{\nu}\subseteq {\mathrm {dom}}(\phi)$ and hence $J(e^{-\phi^*})<\infty$.
The second argument in \eqref{liminf} indeed follows immediately from the proof of \cite[Lemma 17]{CK15}. For completeness, a brief explanation extracted from \cite[p.~3861-3862]{CK15} is provided here. First of all, for any $y \in \mathbb{R}^n$, one has $\phi^*(y)=\sup_{i\in \mathbb{N}}\{\langle x_i,y\rangle-\phi(x_i)\} $ where $\{x_i\}_{i\in \mathbb{N}}$ is a dense sequence in $M_{\nu}$. When $j$ is large enough, the origin lies in the interior of the convex hull of $\{x_1,\cdots, x_j\}$, and this in turn implies that $\exp(-h_j)$ with $h_j(y)=\sup_{1\leq i\leq j}\{\langle x_i,y\rangle-\phi(x_i)\})$ is integrable. Clearly, $h_j$ is increasing to $\phi^*$ as $j$ is increasing to $\infty$. By the monotone convergence theorem, one gets
\begin{eqnarray*}\int_{\mathbb{R}^n}e^{-\phi^*(y)}\,dy=\lim_{j\rightarrow\infty}\int_{\mathbb{R}^n}e^{-h_j(y)}\,dy.\end{eqnarray*} Let $\varepsilon > 0$. An integer $j_0$ (depending only on $\varepsilon$) can be found to have \begin{eqnarray}\label{condition4}
0\leq \int_{\mathbb{R}^n}e^{-h_{j_0}(y)}\,dy-\int_{\mathbb{R}^n}e^{-\phi^*(y)}\,dy <\varepsilon.
\end{eqnarray} The pointwise convergence of $\phi_{l_{j}}\rightarrow \phi$ (on $\{x_1,\cdots, x_{j_0}\}$) yields that, for sufficiently large $j$, $\phi^*_{l_j}(x)\geq h_{j_0}(x)-\varepsilon$ holds for all $x\in \mathbb{R}^n$. Together with (\ref{condition4}), the following holds: \begin{eqnarray*}
\int_{\mathbb{R}^n}e^{-\phi^*(y)}\,dy >\int_{\mathbb{R}^n} e^{-h_{j_0}(y)}\,dy -\varepsilon\geq e^{-\varepsilon}\limsup_{j\rightarrow\infty}\int_{\mathbb{R}^n}e^{-\phi_{l_{j}}^*(y)}\,dy-\varepsilon.
\end{eqnarray*} The second argument in \eqref{liminf} then follows by letting $\varepsilon\rightarrow 0$.
\end{proof}
Now we are ready to prove our main result, i.e., Theorem \ref{solution-lp-mink-6-2}. The $L_p$ surface area measure for log-concave functions in general does not have homogeneity, hence solving the related Minkowski problem (i.e., Problem \ref{problem-lp-6-2}) usually requires more delicate analysis. Most of the time, such problems shall require to solve constrained optimization problems, and the method of Lagrange multipliers as in \cite{GHWXY19, XY17, ZSY17} should be used; this method should work here for a proof of Theorem \ref{solution-lp-mink-6-2}. However, we find that the ideas in the proof of \cite[Theorem 2]{CK15} work well in our case for $p>1$. Therefore, we decide to adopt the ideas in \cite[Theorem 2]{CK15} in the proof of Theorem \ref{solution-lp-mink-6-2}. Unfortunately, due to the lack of homogeneity for the $L_p$ surface area measure for log-concave functions for $p>1$, it is unlikely to have the uniqueness of solutions to Problem \ref{problem-lp-6-2}.
\begin{proof}[Proof of Theorem \ref{solution-lp-mink-6-2}] We will search a solution for the following optimization problem \eqref{mini problem}:
\begin{eqnarray*}
\Theta=\inf\Big\{\mathbf{\Phi}_{p,\nu}(\phi): \phi \in L_{p, e}(\nu) \ \ \mathrm{and} \ \ 0< J(e^{-\phi^*})<\infty \Big\}. \end{eqnarray*}
According to $(\phi^*)^*\leq \phi$ and $((\phi^*)^*)^*=\phi^*$ for any $\phi \in L_{p, e}(\nu)$, it can be checked that
\begin{align*}
\mathbf{\Phi}_{p,\nu}(\phi)=\frac{1}{p}\int_{\mathbb{R}^n} \!\! \phi(x)^pd\nu(x)-\log J(e^{-\phi^*})\geq \frac{1}{p}\int_{\mathbb{R}^n} \!\! ((\phi^*)^*)^p(x)d\nu(x)-\log J(e^{-\phi^*})=\mathbf{\Phi}_{p,\nu}((\phi^*)^*).
\end{align*}
Consequently, to find a solution to the optimization problem \eqref{mini problem}, it is enough to focus on the class of functions $\phi \in L_{p, e}(\nu)\cap \mathcal C$ which are also lower semi-continuous.
We now claim that the optimization problem \eqref{mini problem} is well-defined. First of all, \eqref{p-th-moment} implies that $|x| \in L_{p, e}(\nu)\cap \mathcal C$. Note that $|x|$ is also lower semi-continuous. It is well-known that $|x|^*=\mathbf{I}_{B_1}$. Therefore, $J(e^{-(|x|)^*})=V(B_1)$ is finite. This in turn yields that the infimum of \eqref{mini problem} is not taken over an empty set and hence $\Theta<\infty$. On the other hand, $\Theta$ is bounded from below. To see this, by \eqref{equation-6-10}, for any lower semi-continuous function $\phi \in L_{p, e}(\nu)\cap \mathcal C$ with $J(e^{-\phi^*})\in (0, \infty)$, one has
\begin{align} \mathbf{\Phi}_{p,\nu}(\phi) = \int_{\mathbb{R}^n}\phi(x)^p\,d\nu(x)-\log J(e^{-\phi^*}) \geq H\big(J(e^{-\phi^*})\big)
-\int_{\mathbb{R}^n} \,d\nu(x), \label{est-6-15-1} \end{align}
where $H(t)= c_{\nu} t^{\frac{1}{n}} -\log t$ for $t\in (0, \infty)$. It can be easily checked that $H(\cdot)$ achieves its minimum at $t_0=(n c_{\nu}^{-1})^n$ and $H(t)\geq H(t_0)>-\infty$. Thus, $\Theta>-\infty$ and the optimization problem \eqref{mini problem} is well-defined. We also would like to mention that \begin{align}\label{limit-6-15-1}\lim_{t\rightarrow 0^+} H(t)=\lim_{t\rightarrow +\infty} H(t)=+\infty. \end{align}
Let $\{\phi_l\}_{l\in \mathbb{N}}\subset L_{p,e}(\nu) \cap \mathcal C$ be a minimizing sequence of lower semi-continuous functions such that $J(e^{-\phi_l^*})\in (0, \infty)$ for each $l\in \mathbb{N}$ and \begin{align*} \Theta=\lim_{l\to \infty}\mathbf{\Phi}_{p,\nu}(\phi_l)\leq \mathbf{\Phi}_{p,\nu}(|x|)<+\infty. \end{align*}
Without loss of generality, we can always assume that \begin{eqnarray}\label{formula-6-15}
\sup_{l\in \mathbb{N}}\mathbf{\Phi}_{p,\nu}(\phi_l)\leq \mathbf{\Phi}_{p,\nu}(|x|)+1<+\infty.\end{eqnarray}
Together with \eqref{est-6-15-1} and \eqref{limit-6-15-1}, one sees that
\begin{align}
0<\inf _{l\in \mathbb{N}}J(e^{-\phi_l^*}) \leq \sup_{l\in \mathbb{N}}J(e^{-\phi_l^*}) <+\infty.
\end{align}
Combining with \eqref{formula-6-15}, the following holds: \begin{eqnarray*}
\sup_{l\in \mathbb{N}} \int_{\mathbb{R}^n} \phi_l(x)^p\, d\nu(x)<+\infty.
\end{eqnarray*} This is exactly the condition \eqref{condition1}. Therefore, Lemma \ref{existence} can be applied to get a subsequence $\{\phi_{l_{j}}\}_{j\in \mathbb{N}}$ of $\{\phi_l\}_{l\in \mathbb{N}}$ and a non-negative convex function $\phi_0 \in \mathcal C\cap L_{p, e}(\nu)$, such that \eqref{liminf} holds, namely,
\begin{eqnarray*}
\int_{\mathbb{R}^n} \phi_0(x)^p \,d\nu(x) \leq \liminf_{j\rightarrow\infty} \int_{\mathbb{R}^n} \phi_{l_j}(x)^p \,d\nu(x) \ \ \mathrm{and}\ \
\int_{\mathbb{R}^n}e^{-\phi_0^*(x)}\,dx\geq \limsup_{j\rightarrow\infty}\int_{\mathbb{R}^n}e^{-\phi_{l_j}^*(x)}\,dx.
\end{eqnarray*}
According to \eqref{liminf}, one immediately has \begin{align} \mathbf{\Phi}_{p,\nu}(\phi_0)\leq \liminf_{l\to \infty}\mathbf{\Phi}_{p,\nu}(\phi_l)= \lim_{l\to \infty}\mathbf{\Phi}_{p,\nu}(\phi_l)= \Theta \leq \mathbf{\Phi}_{p,\nu}(\phi_0).\label{max-6-15-1}\end{align}
Hence, $\phi_0$ solves the optimization problem \eqref{mini problem}. Moreover, $0<J(e^{-\phi_0^*})<\infty$ following from \eqref{limit-6-15-1}. Inequality \eqref{santalo-func-6-10} yields that $0<J(e^{-\phi_0})<\infty$ as well. In particular, \begin{align} \lim_{|x|\rightarrow +\infty} \phi_0(x)=+\infty.\label{inf-6-15-1}\end{align}
Let $\widetilde{\Omega}_{\phi_0}=\{y\in \mathbb{R}^n: \phi_0(y)=0\}$ and $\Omega_{\phi_0}=\{y\in \mathbb{R}^n: 0<\phi_0(y)<\infty\}$. By the facts that ${\mathrm {dom}}(\phi_0)\supseteq M_{\nu}$ and $\phi_0$ is even, one gets that $\phi_0$ is continuous on $M_{\nu}$ and both $\widetilde{\Omega}_{\phi_0}$ and $\Omega_{\phi_0}$ are origin-symmetric. Moreover, \eqref{inf-6-15-1} yields that $\widetilde{\Omega}_{\phi_0}$ is a bounded closed (due to the lower semi-continuity of $\phi_0$) convex set. Recall that if $\nu\in \mathscr{M}$, then $\nu(M_{\nu}\setminus L)>0$ holds for any bounded convex set $L\subset\mathbb{R}^n$. Consequently, $\nu(M_{\nu}\setminus \widetilde{\Omega}_{\phi_0})=\nu(M_{\nu}\cap \Omega_{\phi_0})>0$ and $$\int_{\mathbb{R}^n} \phi_0^p(y)\,d\nu(y)=\int_{\Omega_{\phi_0}} \phi_0^p(y)\,d\nu(y)\geq \int_{M_{\nu}\setminus \widetilde{\Omega}_{\phi_0}} \phi_0^p(y)\,d\nu(y) >0.$$ In particular, $M_{\nu}\setminus \widetilde{\Omega} _{\phi_0} \subset \Omega _{\phi_0}$ as $M_{\nu} \subset {\mathrm {dom}}(\phi_0).$ Thus, $ \Omega _{\phi_0}$ is an open set whose Lebesgue measure is strictly positive.
Let $g: \mathbb{R}^n \to \mathbb{R}$ to be any even compactly support continuous function such that the support of $g$, denoted by $\mathrm{supp}(g),$ is a proper subset of $\Omega_{\phi_0}$. Moreover, the compact set $\overline{\mathrm{supp}(g)}$ is contained in $\Omega_{\phi_0}.$ Let $\phi_t=\phi_0 +tg$ and $\phi_t$ is a continuous function on $\Omega_{\phi_0}$. Note that $\phi_0>0$ on $\overline{\mathrm{supp}(g)}$ and hence $\min_{x\in \overline{\mathrm{supp}(g)}}\phi_0(x)>0$. This further yields the existence of $t_0>0$ such that $\phi_t$ is non-negative on $\Omega_{\phi_0}$ for all $t\in [-t_0, t_0]$. It is easily checked that for all $t\in [-t_0, t_0]$, $x\in \mathbb{R}^n$ and $p>1$, one has $\phi_t^p(x)\leq 2^p (\phi_0^p(x)+|t g(x)|^p)$ and hence $$\int_{\mathbb{R}^n} \phi_t^p(x)\,d\nu(x)\leq 2^p \int_{\mathbb{R}^n}\phi_0^p(x)\,d\nu(x) + 2^p |t|^p \int_{\mathbb{R}^n} | g(x)|^p\,d\nu(x)<\infty. $$
This means that $\phi_t\in L_{p, e}(\nu)$ for all $t\in [-t_0, t_0]$. Also note that $$\phi_0(x)-t_0 \max_{x\in \mathbb{R}^n} |g(x)| \leq \phi_t(x)\leq \phi_0(x)+t_0 \max_{x\in \mathbb{R}^n} |g(x)|.$$
By \eqref{def-dual-1}, one obtains that, for all $t\in [-t_0, t_0]$, \begin{equation*}
\phi_0^*(y)-t_0 \max_{x\in \mathbb{R}^n} |g(x)|\leq \phi_t^*(y)\leq \phi_0^*(y)+t_0 \max_{x\in \mathbb{R}^n} |g(x)|.
\end{equation*} This concludes that $J(e^{-\phi_t^*}) \in (0, \infty)$ for all $t\in [-t_0, t_0]$ as
$$e^{-t_0 \max_{x\in \mathbb{R}^n} |g(x)|} J(e^{-\phi_0^*}) \leq J(e^{-\phi_t^*})\leq e^{t_0 \max_{x\in \mathbb{R}^n} |g(x)|} J(e^{-\phi_0^*}).$$
Consequently, for any even and compactly support continuous function $g: \mathbb{R}^n \to \mathbb{R}$ with $\overline{\mathrm{supp}(g)}\subsetneq \Omega_{\phi_0}$, there exists $t_0>0$ such that $ \mathbf{\Phi}_{p,\nu}(\phi_0)\leq \mathbf{\Phi}_{p,\nu}(\phi_t) $ holds for all $t\in [-t_0, t_0]$. Thus, $\phi_0$ satisfies that \begin{align}\label{variation-6-15} \frac{\,d}{\,dt} \mathbf{\Phi}_{p,\nu}(\phi_t)\bigg|_{t=0}=\frac{\,d}{\,dt} \bigg(\frac{1}{p}\int_{\mathbb{R}^n} \!\! \phi_t(x)^pd\nu(x)\bigg)\bigg|_{t=0}-\frac{\,d}{\,dt} \big(\log J(e^{-\phi_t^*})\big)\bigg|_{t=0} =0. \end{align} By the dominated convergence theorem, one can easily get that \begin{align} \frac{\,d}{\,dt} \bigg(\frac{1}{p}\int_{\mathbb{R}^n} \!\! \phi_t(x)^pd\nu(x)\bigg)\bigg|_{t=0} \nonumber &=\frac{\,d}{\,dt} \bigg(\frac{1}{p}\int_{\mathbb{R}^n} \!\! \big(\phi_0(x)+tg(x)\big)^pd\nu(x)\bigg)\bigg|_{t=0}\\ &= \int_{\mathbb{R}^n} \!\! g(x) \phi_0(x)^{p-1}d\nu(x). \label{variation-6-15-1} \end{align} On the other hand, for all $x\in \mathbb{R}^n$ when $\phi_0^*$ is differentiable, it holds that \begin{align}\label{variation-6-15-22}
\frac{\,d}{\,dt}\phi_{t}^*(x)\bigg |_{t=0}=-g(\nabla \phi_0^*(x)).\end{align} We refer the readers to Berman and Berndtsson \cite[Lemma 2.7]{BB13} for a short proof. In other words, formula \eqref{variation-6-15-22} holds for almost all $x$ in the interior of ${\mathrm {dom}}(\phi_0^*)$. The dominated convergence theorem then gives \begin{align} \frac{\,d}{\,dt} \big(\log J(e^{-\phi_t^*})\big)\bigg|_{t=0} = \frac{1}{J(e^{-\phi_0^*})} \int_{\mathbb{R}^n} g(\nabla \phi_0^*(x)) e^{-\phi_0^*(x)}\,dx. \label{variation-6-15-321}
\end{align} Together with \eqref{moment-form-1} for $\varphi=\phi_0^*$, \eqref{variation-6-15}, \eqref{variation-6-15-1}, and \eqref{variation-6-15-321}, one has
\begin{align} \int_{\Omega_{\phi_0}} \!\! g(y) \phi_0(y)^{p-1}d\nu(y)&= \int_{\mathbb{R}^n} \!\! g(y) \phi_0(y)^{p-1}d\nu(y) \nonumber \\ &=\frac{1}{J(e^{-\phi_0^*})} \int_{\mathbb{R}^n} g(\nabla \phi_0^*(x)) e^{-\phi_0^*(x)}\,dx \nonumber \\&=\frac{1}{J(e^{-\phi_0^*})} \int_{\mathbb{R}^n}g(y)\,d\mu(e^{-\phi_0^*}, y) \nonumber \\&=\frac{1}{J(e^{-\phi_0^*})} \int_{\Omega_{\phi_0}} g(y)\,d\mu(e^{-\phi_0^*}, y) \nonumber \end{align} for any even and compactly support continuous function $g: \mathbb{R}^n \to \mathbb{R}$ with $\overline{\mathrm{supp}(g)} \subsetneq \Omega_{\phi_0}$. This concludes that, on $\Omega_{\phi_0}$, \begin{align} \phi_0^{p-1} \nu =\frac{1}{J(e^{-\phi_0^*})} \mu(e^{-\phi_0^*}, \cdot). \label{final-form-6-15-1} \end{align} Let $\varphi=\phi_0^*$. Then
$\varphi\in \mathcal C$ is an even lower semi-continuous convex function, $\varphi^*\in \mathscr L^+$, and
\begin{eqnarray}
\nu =\frac{1}{J(e^{-\varphi})} (\varphi^*)^{1-p} \mu_{1}(e^{-\varphi},\cdot) = \frac{1}{J(e^{-\varphi})} \mu_{p}(e^{-\varphi}, \cdot)\ \ \ \mathrm{on} \ \ \Omega_{\varphi^*}. \label{equation-final-6-15}
\end{eqnarray} By taking the integration from both sides of \eqref{final-form-6-15-1} or \eqref{equation-final-6-15} on $\Omega_{\varphi^*}$, one sees that
$$ \frac{1}{J(e^{-\varphi})}= \frac{\int_{\Omega} (\varphi^*(y))^{p-1}\,d\nu(y)} {\int_{\Omega} \,d \mu_{1}(e^{-\varphi} ,y) } = \frac{\int_{\Omega} \,d\nu(y)} {\int_{\Omega} \,d \mu_{p}(e^{-\varphi}, y)}.$$
This completes the proof of Theorem \ref{solution-lp-mink-6-2}. \end{proof}
\vskip 2mm \noindent {\bf Acknowledgement.} The
research of DY was supported by a NSERC grant, Canada.
\vskip 0.5 cm
|
\section{Data Availability}
The data that support the findings of this study are available from the corresponding author, B. Perez, upon reasonable request.
\end{document}
\endinput
\section{Introduction}
\label{sec:introduction}
Crowdfunding has become a standard means to financially support individuals' needs or ideas, typically through an online campaign appealing to contributions from the community.
What started off as a grassroots movement is now a flourishing industry.
In fact, from \$597M raised worldwide in 2014, to \$17.2B, in North America alone, in 2017, this industry will continue to grow globally~\cite{crowdfundingStatista}.
Over the last decade, the emergence and consolidation of crowdfunding platforms (CFPs) have narrowed the field to a handful of platforms.
Top contenders, such as Kickstarter, FundingCircle, and GoFundMe, have specialized into one of three categories:
investment-based platforms where donors become angel investors in a new enterprise;
reward-based platforms where the backers provide loans with the condition of interest upon repayment;
and donation-based platforms where campaigns are an appeal to charity~\cite{belleflamme2015economics}.
CFPs and their increasing popularity and fundraising ability for many causes (even recently for coronavirus-related costs~\cite{nyt2020}) inevitably attract malicious actors who take advantage of unsuspecting users (e.g.,~\cite{goFraudMe, nyt2019, theSun2020, ctm2020}).
Immediate access to trusting investors and their funds make these platforms particularly attractive to malicious activity.
Some platforms allow fund disbursements to happen immediately following a donation, while others require scheduled intervals (\textit{e.g.},\xspace weekly), or reaching donation goals.
Furthermore, there is a striking lack of regulation in this space~\cite{fbi2020}.
This void leaves a grey area where crimes are hard to define and difficult to prosecute.
The emergence of few highly publicized cases of fraud in crowdfunding campaigns further undermines general public confidence in this trust-based enterprise.
Nevertheless, evidence of campaign and fund misuse is scant.
According to GoFundMe, one of the most prominent CFPs, fraudulent campaigns make up less than 0.1\% of all campaigns posted on the site~\cite{gfmfraud}.
But even at this ``low'' rate of fraud, which has not be substantiated with transparent reports by GoFundMe or other CFPs, in a billion dollar industry, it can amount to tens of millions in defrauded funds every year.
Given CFPs' major source of revenue are these campaigns, through commissions on new campaigns and on every donation, CFPs are not properly incentivized to detect and stop fraud.
Therefore, the lack of tools quantifying this problem is not surprising, effectively preventing the protection of unsuspected contributors.
In this study, we aim to provide such tools to help combat fraud in donation-based CFPs.
We analyze campaigns in North America created to cover medical expenses, a primary reason for these types of appeals (one in three crowdfunding campaigns~\cite{forbesMedical}).
The urgency and strong emotional content of health-related financial constraints easily attract donor attention and donations.
We are interested in quantifying the prevalence of fraudulent behavior in these campaigns, requiring us to classify campaigns as fraud or not.
Our goal is to create a machine learning (ML) classifier to distinguish between campaigns that are fraudulent or not, at the moment of their creation, i.e., using only features extracted from campaigns newly published.
To accomplish our task, we collect and annotate over 700 campaigns from all major CFPs (GoFundMe, MightyCause, Fundly, Fundrazr, and Indiegogo) and derive deception cues from both the text and the images provided in each campaign.
Overall, we find that fraud is a small percentage of the crowdfunding ecosystem, but an insidious problem.
It corrodes the trust ecosystem on which these platforms operate on, endangering the support that thousands of people receive year on year.
Our results show that using an ensemble ML classifier that combines both textual and visual cues, we can achieve a Precision of 91.14\%, Recall 90.77\% and AUC 96.01\%, i.e., approximately 41\% improvement over the deception detection abilities of people within the same culture~\cite{Bond1990}.
This work is also the first to incorporate text and images in the analysis of fraud, and we rely on features available immediately after a campaign goes live.
This is a significant step in building a system that is preemptive (e.g., a browser plugin) as opposed to reactive.
We believe our method could help build trust in this ecosystem, by allowing potential donors to vet campaigns before contributing.
Similarly, CFPs could use it to prompt vetting and request additional information from a potentially fraudulent campaign creators before campaigns are made public.
Our contributions with the present study are as follows:
\begin{itemize}
\item We collect a dataset of over 700 crowdfunding campaigns on different health-related topics, including medical, emergency appeals, and memorials, and annotate them for being fraudulent or not.
\item Through the use of NLP techniques, we extract language cues, including emotions and complexity of language, and study their association with fraudulent campaigns.
\item Using convolutional neural networks, we extract characteristics of the images posted with each campaign, including emotions and content displayed, and associate them with fraudulent campaigns.
\item Using the above features (text and image-based), we train supervised classification techniques to perform automatic detection of fraudulent campaigns, and discuss our results.
\item We make the collected and annotated dataset available for other researchers to further investigate this problem on CFPs.
\end{itemize}
\section{Related Work}
\label{Sec:related}
Previous work on financial fraud highlights this task's complexity.
Financial information has typically been used to predict the likelihood of a given transaction being fraudulent.
Primarily, past works build behavioral profiles for each user to compute the likelihood of a new transaction being legitimate~\cite{abbasi2012metafraud,bhattacharyya2011data,cecchini2010detecting,dechow2011predicting,panigrahi2009credit,sanchez2009association,srivastava2008credit}.
Novel models in finance technology have opened new areas of research in this space.
Some works focus on detecting deception online.
\citet{luca2016fake} take a set of reviews identified as fraud or not fraud by the platform Yelp and explore the determinants of fraudulent behavior.
The authors explore how restaurants' engagement in positive and/or negative review fraud (i.e., fake reviews) interacts with reputation and competition, over time. We use insights from these fraudulent reviews to shape our understanding of fraudulent deceptive behavior in CFP campaigns.
In their work on Peer-to-Peer lending,~\citet{xu2015p2p} explore trust relationships between borrowers and lenders.
Funds raised can be used for private affairs (\textit{i.e.},\xspace there are no business plans and no milestones to rely on for validation), just like crowdfunding campaigns.
However, in their model, the money is meant to be restituted and the network members build up their reputation over time.
They find that soft descriptions of the borrower, \textit{e.g.},\xspace age, gender, race, and appearance, are good predictors of whether a loan will be repaid in time.
Conversely, the physical appearance (gender, race, and attractiveness) of the campaign's creator as understood from their profile picture can be used to predict the trustworthiness and, by extension, the success of a campaign~\cite{lin2013judging,pope2011s,duarte2012trust}.
Again, such features help us better understand trust relationships between funders and requesters.
Similarly to Peer-to-Peer lending, crowdfunding is an online financial tool in the hands of millions.
Both are paradigms that depend on participant trust, and fraud against them ``causes emotional and financial harm to lenders (donors) and great damage to sites (platforms) destroying their reputation"~\cite{xu2015p2p}.
The study of fraud in crowdfunding has been tied to the success of the campaign~\cite{wessel2016emergence}, or to entrepreneurial endeavors.
In fact,~\citet{wessel2016emergence} looked at social capital as a means to influence consumer decision making.
They take 591 campaigns that have been flagged for having fake Facebook likes and find that fake social capital has an overall negative impact on the number of backers of a campaign.
~\citet{siering2016detecting} looked at the problem of deception in crowdfunding campaigns, focusing on investment-based donations, where there is an expectation of a reward and a defined business plan. Importantly, these are business interactions, fundamentally different from the altruistic behavior implied in our campaigns.
Based on linguistic text cues, their model presented in~\cite{siering2016detecting} achieves 75\% accuracy.
We build on textual features and combine them with image features, leading our model to achieve 86\% accuracy.
Finally, and most similar to our work,~\citet{cumming2016disentangling} look at entrepreneurial campaigns (\textit{i.e.},\xspace commercial campaigns with pledges and rewards) and try to understand the difference between campaigns labeled as detected fraud, suspected fraud, and not fraud.
Using theories from economics and behavioral sciences they identify four possible markers: characteristics and background of the campaign creator (\textit{i.e.},\xspace use of names and further participation in the community), a campaigns' affinity to social media, funding and reward structure in the campaign (\textit{i.e.},\xspace the duration of the campaign), and finally, details in the campaign description (\textit{i.e.},\xspace clarity of language and veracity).
They test their markers in a dataset of 207 fraud cases from two major crowdfunding portals (Kickstarter and Indiegogo) and find that fraudulent campaigns can be described as having longer periods of collection, no Facebook page associated to it, and campaign creators with comparatively less time in the crowdfunding community.
Even though relevant to our work, there are several primary differences with~\citet{cumming2016disentangling}: 1) the type and incentive behind the campaigns (entrepreneurial vs. charitable donations for health problems), 2) when the funds are available (the full amount must be raised vs. immediately available), 3) the tone and the content of the text in the campaigns.
Indeed, while we do borrow their insights on the relationship between fraud and the simplicity of text, the problem we are addressing is different.
\section{Crowdfunding Campaign Data Collection and Annotation}\label{sec:dataset}
\begin{table*}
\centering
\caption{Data fields available across crowdfunding sites. These fields were available during the annotation process.
The automated ML analysis focuses on the campaign description, including text and image content.
Note that many fields are variable (e.g., money raised, donors, etc.) and are not available to the ML process for use when the campaign is first published.}
\label{table:campaign_feats}
\begin{tabularx}{\textwidth}{lX}
\toprule
\textbf{Feature} & \textbf{Description}\\ \midrule
Creation Date & Timestamp of the public release date.\\
Campaign Duration & For time-limited campaigns, the difference between release and close date. Otherwise, it reflects the time elapsed between release and the time of scrapping.\\
Campaign Status & Reflects whether the campaign is open for donations or not (boolean).\\
Title & Title of the campaign.\\
Created by & Campaigns may be a direct appeal or an appeal made on behalf of another. This field contains a reference to the owner of the campaign who might be different from the beneficiary.\\
Description & A narrative of the cause for which the campaign is launched (and updates to the story).\\
Category & The general classification of the campaign (\textit{e.g.},\xspace Memorial, Health, Emergencies, etc.).\\
Fundraising Goal & The total amount of money the creator hopes to raise.\\
Money Raised & The money that has been raised by the campaign to date.\\
Number of Donors & The number of individual contributors to the cause.\\
Donation Amount & The individual amounts from each of the contributions.\\
Social Media & The number of likes or shares that the campaign has received over social media.\\
Geo-Tag & The location from which the campaign was launched.\\
\bottomrule
\end{tabularx}
\end{table*}
\subsection{Background on CFPs}
Crowdfunding sites are designed to help people connect funding requests with benefactors.
While each CFP is different, they all provide a search engine to find specific campaigns and a classification system that allows visitors to find campaigns that may be relevant to their interests.
In this paper, we looked at campaigns from the top five crowdfunding platforms online: Indiegogo, GoFundMe, MightyCause, Fundrazr, and Fundly.
The information displayed per campaign varies depending on the platform and on the creator of the campaign.
Table~\ref{table:campaign_feats} summarizes the fields and descriptions that are common across all sites.
Depending on the type of campaign and hosting platform, funds raised are delivered either to the campaign creator or its beneficiary.
Typically, entrepreneurial campaigns require a goal to be met before funds are released (not reaching a funding goal results in contributions being returned to investors), whereas in charitable projects, there is no limit as to how soon or often any funds are withdrawn from the campaign.
In terms of revenue, \textit{gofundme.com}, the most prominent CFP, states that they charge a percentage of the transaction as fees, plus a fixed amount per donation~\cite{gfm_pricing}.
Finally, CFPs are aware of the risk of fraud and some offer a guarantee: any member that made a donation to a fraudulent campaign is entitled to a refund by the CFP.
The reimbursement, however, must be requested by the donor after an internal investigation reveals the campaign to be fraudulent.
When a campaign is reported as suspicious, the CFP will send a request for information to the creator of the campaign.
Following the initial report, continued suspicious behavior might result in a campaign being deactivated, or altogether removed.
A deactivated campaign will show the title, primary image and total funds raised.
A removed campaign will result in a redirect to the CFPs main website.
A missing or deactivated campaign, however, is not always an indication of fraud.
For example, a campaign created to raise funds outside a CFP's ``donation cover area'' results in the campaign being removed from the platform.
Alternatively, campaign creators might close a campaign if the fundraising goal has been met, or an event has passed.
\subsection{Defining Fraud}
Fraud is generally defined as a misrepresentation of an existing fact, made from one person to another, with knowledge of its falsity and for the purpose of inducing the other to act~\cite{fraudDef}.
One requirement of fraud is therefore deception, as it requires the perpetrator to convince the victim that their (false) statement is true.
It also results in damages to the victim and is, most importantly, a criminal offense\footnote{To date, crowdfunding fraud cases have been prosecuted as theft by swindle, larceny, wire fraud, child endangerement, failure to make required distribution of funds, mail fraud, and felony grand theft among others}.
First, we must recognize that fraud is an umbrella term used to define a range of behaviors including embezzlement where (legitimatelly acquired) funds are missappropriated; opportunist fraud where a real story draws criminals to fabricate association to the people/event; or complete fiction in both events and associations, among others. In this work we are attempting to automate the process of recognizing cues available at the time of publication of a crowdfunding campaign where the creator of the campaign is aware of the falsehood of the claims in the campaign. We refer to these campaigns as \textit{fake}.
As an example, opportunist campaigns are \textit{fake}: the creator of the campaign has limited information which can be reflected in his writing style and choice of picture\footnote{In contrast, fraud by emblezzement is \textit{not-fake}}.
In the results we present, our priority is to minimize the number of false positives (\textit{i.e.},\xspace real campaigns mislabeled as fraud). However, it is not possible to completely eliminate type II errors (\textit{i.e.},\xspace fraud campaigns that were misclassified as real).
Cases like embezzlement where the people, events, and description are real and with the appropriate level of detail but, where the funds were never delivered to the rightful recipient cannot be identified before the decision to commit a crime has been carried out.
Therefore, the results we present should be understood as a lower bound of the number of cases to be expected in the wild.
This work, however, is an improvement upon the current state of the art in detection of these types of campaigns where the determination of \textit{fraud} is delegated to the CFPs, and the individual contributors are left on their own devices and judgement to decide if a campaign is fraudulent or not.
\subsection{Datasets}
\label{sec:data}
We have two main sources of data: a set of campaigns that have been confirmed\footnote{Confirmed means either a conviction following a criminal indictment or a condemnation from the victims (more in Section~\ref{gfraudm}).} as fraud and collected from GoFraudMe~\cite{goFraudMe} (we will refer to these as $set\ A$), and two sets of manually annotated campaigns collected from different CFPs ($sets\ B$ and $C$).
\subsubsection{Labeled data from GoFraudMe.com (set A)} \label{gfraudm}
The goal of this website, maintained by an investigative journalist, is to expose fraudulent cases in the GoFundMe platform.
The site serves the dual purpose of holding the CFP accountable for fraudulent campaigns and presenting, preserving, and publicizing the evidence that led to the characterization of fraud.
The site holds 192 confirmed cases of fraud that were shared with us by the website curator.
The process that leads to the inclusion of a campaign in the website varies greatly, but each is accompanied by a narrative that presents the inconsistencies that led to the declaration of fraud.
Some campaigns have the guarantee of a guilty verdict following legal criminal proceedings, whereas others have been denounced by beneficiaries and supported by their community.
Some of the cases presented, typically those that follow from events reported in various news platforms, give rise to several fraudulent campaigns.
For some cases, there is an archived version of the campaign that was used to collect money with a link to the rightful beneficiary.
For others, there is a timeline following the investigation, that led to criminal charges (and subsequent conviction if it is available).
\subsubsection{Annotated Data (sets B \& C)}
In addition to the labeled campaigns collected from $set\ A$, we created two manually annotated datasets.
$Set\ B$: 191 campaigns from the Medical category in GoFundMe.com.
$Set\ C$: 350 campaigns from different CFPs that were directly related to organ transplants.
$Sets\ B$ and $C$ were manually annotated following the methodology described in the Section~\ref{sec:annotation}.
$Set\ C$ is a random sample of 350 campaigns related to organ transplants collected in January 2019 from the top 5 CFPs: Indiegogo, GoFundMe, MightyCause, Fundrazr, and Fundly.
Both $B$ and $C$ sets were collected through automated crawlers written in python.
From each campaign, we collected the features in Table~\ref{table:campaign_feats} as well as all comments, pictures, and individual donations.
Each campaign was visited in the order presented by the CFP's search engine.
While some CFPs provided APIs to connect with their database, the data fields were collected, for the most part, through the corresponding elements in HTML.
\subsection{Campaign Annotation}
\label{sec:annotation}
\begin{table}
\centering
\caption{Breakdown of the labels for the 733 annotated campaigns. All campaigns had a textual description and many had several images as part of their appeal.}
\label{table:annotation_total}
\begin{tabularx}{\columnwidth}{clXX}
\toprule
\textbf{Score} & \textbf{Label} & \textbf{Text} & \textbf{Images}\\ \midrule
0 & invalid & 93 & 117 \\
1 & fraud & 141 & 138 \\
2 & probably fraud & 26 & 71 \\
3 & unknown & 105 & 123 \\
4 & probably not-fraud & 78 & 141 \\
5 & not-fraud & 290 & 517 \\
\bottomrule
\end{tabularx}
\end{table}
Combined, $sets\ A,\ B$ and $C$ make up the ground truth in the study. $Sets\ A$ and $B$ are inversely balanced in the sense that one provides mostly examples of fraud and the other mostly not-fraud\footnote{From our findings, the prevalence of fraud in a random sample of medical campaigns is approximately 10\%, in contrast to the 0.1\% claimed by CFPs. Therefore most of the campaigns we looked at in $Set \ B$ were not-fraud.}.
$Set\ C$ was created as a means to augment the number of campaigns in the study.
All campaigns were manually annotated using the scale proposed in Table~\ref{table:annotation_total}, where (1) indicates certainty of fraud and (5) certainty of not-fraud.
During manual annotation, two expert annotators developed guidelines to determine the label of each campaign.
The considerations in the guidelines included:
\begin{itemize}
\item A personal (offline) knowledge of the circumstances that led to the appeal as evidenced in the support messages posted to the campaign. Knowledge is reflected (but not limited to) in having met the beneficiary (or having first hand knowledge of the circumstances), participating in offline fundraising activities, or familial relationships between donors.
\item A sense of closure to each campaign, particularly those that have been open for donations for several years.
\item Coherency between the description, support documents, pictures, fundraising goal, donors, and level of detail.
\item Participation of the creator in other campaigns.
\item Reverse search of pictures and text diplayed in the campaign leading to unrelated results in the web.
\item Evidence of contradictory information.
\item Overwhelming lack of engagement of campaign donors.
\end{itemize}
\begin{figure*}[t]
\centering
\subfloat[Text]{\includegraphics[width=.45\linewidth]{figures/sentiment_text_isfraud}}
\quad
\subfloat[Images]{\includegraphics[width=.45\linewidth]{figures/sentiment_image_isfraud}}
\caption{Comparing the emotions displayed in the text and image of the campaign.
}
\label{fig:emotion}
\end{figure*}
The label of fraud assigned by the annotators was independent of the features engineered for automated detection. The annotators relied on semantic interpretation.
The features used in the models are stylistic markers of the textual description and the images in the campaign.
To this extent, we are reasonably certain that while the label of fraud might be incomplete (\textit{i.e.},\xspace it will not capture all categories of fraud), it is correct.
Ultimately, we considered 704 campaigns in the study, with some campaigns removed from the dataset because the content was no longer accessible, the text was in multiple languages, or there were too few characters to compute any of the text-based features.
Inter-annotator reliability (or consistency) refers to the validity of the variable being measured~\cite{mchugh2012interrater}.
In this project, we had a structured subjective task, where we iteratively refined and applied a set of guidelines that define fraud to each of the campaigns reviewed.
At the end of the first iteration, annotators reconciled the labels and revised the guidelines.
At the end of the second round of annotations, and for campaigns with contradictory labels, the final decision was agreed by discussion and consensus between annotators.
We measured the consistency across annotators for each iteration using Cohen's Kappa ($\kappa$)~\cite{cohen1960coefficient}.
We applied the interpretation scale proposed by Landis and Koch~\cite{landis1977application}, where values between 0.6 and 0.8 are considered to reflect a substantial agreement between annotators.
At the end of the first iteration, $\kappa$ was found to be $0.451$, reflecting only moderate agreement between annotators.
After revising the guidelines, at the end of the second round of annotations, $\kappa = 0.675$.
Ultimately, the values used as labels for classification were considered of binary form, \textit{i.e.},\xspace 1 for fraud and 0 for not-fraud, that reflect consensus across annotators (\textit{i.e.},\xspace a dataset with $\kappa = 1$).
\section{Campaign Fraud: Textual Cues}
\label{sec:text}
The description of the campaign is the best line of communication between the campaign creator or beneficiary and any potential donors.
Therefore, it is the first place where a malicious actor might leave traces of deception.
In this section, we present five different areas where automated analysis might find quantitative evidence of deception.
We also present some preliminary analysis of the variable categories with respect to our classification variable: \textit{fraud}.
\subsection{Feature Extraction}
\subsubsection{Sentiment Analysis.}
We extract the sentiment and tone expressed in the text for further analysis using IBM services~\cite{IBM}.
The sentiment is computed as a probability across five basic emotions: sadness, joy, fear, disgust, and anger.
Complementary to emotions, the text's tone can also express a campaign's intent.
We analyze confidence scores for seven possible tones: frustration, satisfaction, excitement, politeness, impoliteness, sadness, and sympathy.
\subsubsection{Complexity and Language Choice.}
The need for appeal to a more general population can lead fake campaign creators to adapt (or carefully select) the language used.
Simpler language and shorter sentences can appeal to the emotions of the reader and, therefore, be more successful.
To check the language complexity of the document and word choice, we look at a series of readability scores (\textit{e.g.},\xspace automated readability index, Dale-Chall Formula, etc.) and language features (\textit{e.g.},\xspace function words, personal pronouns, average syllables per word, total number of characters, etc.)~\cite{cumming2016disentangling,wessel2016emergence}.
\subsubsection{Named-Entity Recognition}
Named-Entity recognition is the process of identifying named entities (e.g. proper nouns, numeric entities, currencies) in unstructured text and assigning them to a finite set of categories. In this project, we relied on spaCy~\cite{honnibal2017spacy} a tool released for Python which identifies 18 types of entities in text.
SpaCy models are based on convolutional neural networks built with pre-trained vectors which give an accuracy of 86.42\%.
\subsubsection{Form of the text}
The next group of features we considered was the visual structure of the text. For the entire textual dataset we captured the form of each word: whether the letters were all lower-case, all upper-case, the number of emojis on the text, the number of words with exclamation mark, the words with apostrophes, and many others. We generated a vector with 255 descriptors and evaluated the text in each campaign against the features.
\subsubsection{Word Importance}
Lastly, we considered the numerical vectorial representation of the text given by tf-idf.
This method, similar to a bag-of-words approach, highlights content similarity between different documents. As with the other textual features, we compute word importance on the text included in the campaign description. Ultimately, this description is the primary method of communication between campaign creator and potential donors. While the success of a campaign is mostly determined by the strength of a community and their participation in the system, a good story may persuade chance visitors to donate to the cause.
\subsection{Exploring the Data: Text-Based Features}
\begin{figure} [th]
\centering
\includegraphics[width=\linewidth]{figures/Word_Difference}
\caption{Word importance in campaign description across fraudulent and not-fraudulent campaigns. From left to right, the words are arranged in order of decreasing difference between the two classes.}
\label{fig:text_wordCloud}
\end{figure}
Fig~\ref{fig:emotion}(a) presents the sentiment analysis for the text of the campaigns.
Each bar is the aggregation of emotions for the label indicated.
From the figure, we see that as emotions, joy and disgust contain valuable information in the separation of the binary variable.
Interesting is also the balance between the positive and negative emotions in each campaign.
This figure shows that, as emotions in the text, campaigns that are not-fraud display more joy and less disgust than campaigns that are fraud.
Almost as if the narrator is making an effort to present their friend or relative (\textit{i.e.},\xspace the beneficiary) as they were; then, present the (presumably negative) reason for creating the campaign.
One of the most interesting results we found is evidenced in Figure~\ref{fig:text_wordCloud}.
Word importance for each category shows that while, generally, both sets of campaigns have similar characteristics, fraudulent campaigns are perhaps more desperate in their appeal.
Starting from the left side of x-axis, Figure~\ref{fig:text_wordCloud} shows that the words \textit{money, help, please, cancer}, and \textit{get} are more prevalent in fraudulent campaigns, whereas not-fraud descriptions will emphasize words like \textit{kidney, transplant, heart, medic(al)}, and \textit{work} (right side of x-axis).
In general, legitimate campaigns are more descriptive, being open about the circumstances in making their appeal.
\subsection{Significance: Reducing Dimensionality}
Combined, the five types of text-based analysis result in 8,341 features extracted from the description provided with the campaign, but several of them can be sparse (e.g., TFIDF) and others may not prove to be so helpful in detecting fraud.
Therefore, the final step in our pre-processing of the data is to analyze each feature with respect to the variable we are interested in, and filter out features not useful.
We start by making no assumptions about the distribution of our random variables and choose the non-parametric, two-sample KS test to check whether the difference between the distributions of the fraud and not-fraud data for each feature are significant at level $a=0.05$.
This testing removed features that were not different, and reduced the space to 71 variables from all five textual analysis categories.
Ultimately, in this paper, any result computed with text-based features includes only the 71 KS significant features\footnote{We ran a similar analysis using the t-test which assumes that the random variable is normally distributed and achieved similar performance with the classifier.}.
\section{Campaign Fraud: Visual Cues}
\label{sec:images}
Though the text is the primary means of information, pictures provide the often essential supporting details of the claim.
As with Section~\ref{sec:text}, in this section we present the rational for the features derived from images and the preliminary results of the analysis of the data collected.
\begin{figure}
\centering
\includegraphics[width=\linewidth]{figures/Object_Difference}
\caption{Object prevalence in images across fraudulent and not-fraudulent campaigns. From left to right, the objects are arranged in order of decreasing difference between the two classes.}
\label{fig:objectAnalysis}
\end{figure}
\subsection{Feature Extraction}
\subsubsection{Emotion Representation.}
Psychological studies show that images, as a form of visual stimuli, can be used to induce human emotion~\cite{joshi2011aesthetics}.
Visual emotion prediction has therefore attracted much interest from the computer vision community---framed as a multiclass classification problem using image-emotion pairs as input-output tuples for learning.
Motivated by the foregoing successes for visual emotion prediction in transfer learning, we repurposed a ResNet-152~\cite{he2016deep} a convolutional network pre-trained on the ImageNet dataset~\cite{krizhevsky2012imagenet} containing 1.2 million images of 1000 diverse object categories.
The fine-tuning was performed by replacing the original 1000-way fully connected classification layer with a newly initialized layer consisting of 8 neurons that correspond to the emotion categories of interest.
As defined in~\cite{zhao2014exploring,you2016building}, the eight categories were as follows: amusement, anger, awe, contentment, disgust, excitement, fear, and sadness.
To fine-tune the model, we utilized the Flickr and Instagram (FI) dataset~\cite{you2016building} of 23k images, where each image is labeled as evoking one of the eight emotions based on a majority vote between five Amazon Mechanical Turk workers.
We used 90\% of the images for training and the remainder for validation.
During pre-processing, each image was resized to $256\times256\times3$ and standardized (per channel) based on the original ImageNet training data statistics.
We used 100 epochs, to minimize a negative log-likelihood loss, with stochastic gradient descent, using an initial learning rate of 0.1, momentum 0.9, and a batch size of 128.
The learning rate was multiplied by a factor 0.1 at epochs 30, 60 and 90.
We performed data augmentation by randomly cropping $224\times224\times3$ image patches, which is the resolution accepted by ResNet-152.
During fine-tuning, all layers except the classification layer were frozen.
The final model accuracy on the validation data was 73.9\%, where predictions are based on central $224\times224\times3$ crops.
The semantic evidence over the eight emotions, in the form of logits (unnormalized log probabilities), can then be extracted for the crowdfunding images.
Each image was resized such that its shortest side was 256 pixels and then a central crop was extracted of size $224\times224\times3$.
Semantic emotion category representations were then extracted from the classification layer.
\subsubsection{Appearance and Semantic Representations.}
Again, with the help of a pre-trained ResNet-152 model, trained on the ImageNet dataset, we extracted appearance representations and semantic representations of each of the images present in the campaigns. For pre-processing, each crowdfunding image was resized such that its shortest side was 256 pixels and then a central crop extracted of size $224$$\times$$224$$\times3$. We standardized each image (per channel) based on the original ImageNet training data statistics.
The appearance representations is meant to quantize the picture itself by generating a vector of descriptors from the penultimate layer of the network. These features ( $\in\mathbb{R}^{2048}$) provide a description of each image where the fields, automatically learnt by the network can be, \textit{e.g.},\xspace the dominant color, the texture of the edges of a segment, a (lower level) object --- \textit{e.g.},\xspace an eye, among others.
In contrast, the semantic representation expresses the logit presence of pre-determined objects in each image.
The vector ($\in\mathbb{R}^{1000}$) is extracted from the classification layer over the 1000 ImageNet classes.
Each representation is useful since convolutional neural networks are known to implicitly learn a level of correspondence between particular objects~\cite{zeiler2014visualizing}.
Moreover, the representations invariably outperform their hand-engineered counter-parts~\cite{chatfield2014return}.
Finally, we consider the number of faces present in the image as a possible distinguishing factor between \textit{fraud} and \textit{not-fraud} campaigns. We extract this feature using the dlib~\cite{king2009dlib} HOG-based face detector and estimate the number of faces present per image.
\subsection{Exploring the Data: Image-Based Features}
\begin{figure}
\centering
\includegraphics[width=0.85\linewidth]{figures/CDF_faces_count.pdf}
\caption{Number of faces detected in images across fraudulent and not-fraudulent campaigns.}
\label{fig:faces-detected}
\end{figure}
In our analysis of emotion in images we found that, as compared to the text (in Figure~\ref{fig:emotion}(a)), there is a greater imbalance between positive emotions and sadness. Figure~\ref{fig:emotion}(b) shows the positive emotions as shades of blue and other emotions as indicated in the legend.
Similar to the text, not-fraud campaigns display more positive emotions and proportionally less anger and fear through their images.
In our analysis of objects present in each image (Figure~\ref{fig:objectAnalysis}), we find that not-fraud campaigns have a stronger presence of objects that are associated with hospital stays (as evidenced by the presence of objects like \textit{lab coats, pajamas, stretchers}, and \textit{neck braces}) though the same categories are found to a lesser degree in the fraudulent campaigns.
On the other hand, fraudulent campaigns appear to include images with objects or concepts that are more casual in nature, such as \textit{barbershop, suit, tie, uniform}, which may not fit the context of CFP campaigns launched for medical-related problems.
Compared to the results in Figure~\ref{fig:text_wordCloud}, the signal revealed by the images is not as strong as the one contained in the text, as the separation is not so clear.
The difference between text and images can be explained by considering that CFPs provide, at times, specific instructions regarding the types of images to include.
For example, one such instruction is to include a picture of the fundraising organizer and the person in need looking happy.
Not only does this homogenize the type of images used in the fundraisers, it also provides a clear guidebook for potentially fraudulent campaigns, hence diminishing the predictive power of images, in general, and the objects identified in those images, in particular.
We also analyze the number of faces detected in the images of the two types of CFP campaigns, in Figure~\ref{fig:faces-detected}.
Even though the number of faces in the extreme cases (e.g., above 10 faces detected) follows similar distribution in the two classes, we notice that for the majority of non-fraudulent campaigns, they tend to include images with more faces than fraudulent campaigns.
Interestingly, the median for both classes is 1, but the mean for non-fraudulent campaigns is 1.488 and for fraudulent is 0.8341, which means it is more common to include images with at least one face in the non-fraudulent campaigns.
\subsection{Significance: Reducing Dimensionality}
Combined, the visual cues amount to 3,057 features. As was the case with textual features, we expect the semantic representation of each image to be sparse and some features to be more discriminative than others with regards to our target variable.
As was the case with the text-based features, we used the KS-test to determine the significance of each descriptor.
The result was a vector of 501 features with representatives from all categories: emotion, appearance, semantics, and number of faces.
The classification models contained only these 501 features in all types of image analysis.
\section{Automated Detection of Crowdfunding Fraud}
\label{sec:results}
Next, we present our effort to train a machine learning (ML) classifier to automatically detect fraudulent campaigns using various features discussed in the previous sections.
\begin{table}
\centering
\caption{Average precision and accuracy for different classification algorithms, for the \textit{Label II} setup and with text-based features considered (st. deviation shown in parenthesis).}
\label{table:text_classifiers_all}
\begin{tabular}{lccc}
\toprule
\textbf{Classifier} & \textbf{Accuracy} & \textbf{F1-Score}& \textbf{AUC}\\ \midrule
SVM & 0.6223 (0.078) & 0.6204 (0.079) & 0.6223 (0.076)\\
\textit{k}-NN & 0.6252 (0.077) & 0.6116 (0.083) & 0.6252 (0.070)\\
Naive-Bayes & 0.7980 (0.062) & 0.7967 (0.063) & 0.7980 (0.062)\\
AdaBoost & 0.8061 (0.063) & 0.8060 (0.063) & 0.8061 (0.063)\\
Decision Tree & 0.8130 (0.062) & 0.8129 (0.063) & 0.8130 (0.062)\\
Random Forest & 0.8368 (0.059) & 0.8367 (0.059) & 0.8368 (0.059)\\
MLP & 0.8553 (0.050) & 0.8544 (0.050) & 0.9252 (0.040)\\
\bottomrule
\end{tabular}
\end{table}
\subsection{Experimental Setup}
\subsubsection{Fraud Scale Grouping.}
The fraud scale presented in Table~\ref{table:annotation_total} can be combined in different ways to generate the overall label of \textit{fraud}.
In our first experimental setup, we use the union of campaigns with scores \{1,2\} as \textit{fraud}, scores \{4,5\} as \textit{not-fraud}, omitting the campaigns with score 3, and denote this setup as \textit{Label I}.
In the second experimental setup, we define as \textit{fraud} exclusively the campaigns with scores of \{1\}, and \textit{not-fraud} the campaigns with scores of \{5\}, omitting the other campaigns, and denote this setup as \textit{Label II}.
Practically, in using \textit{Label I}, we prioritize the need to get more observations for the training of the classifier, whereas using \textit{Label II}, we give more importance to the strength of the signal being captured, but in reduced instances.
In our experiments, we observed better performance when minimizing the noise in the signal.
Ultimately, we chose \textit{Label II} for the final results.
\subsubsection{ML Classifiers.}
In choosing a classifier, we need a method that is fast, robust to noise and not prone to overfit the data, thus, allowing the model to be generalizable.
We tested different classical ML methods whose implementation is available in sklearn~\cite{scikit}: Random Forests (RF), AdaBoost, Decision Tree, \textit{k}-NN, Naive-Bayes and Support Vector Machine (SVM), and compare their performance across different metrics. In addition to the classical methods, we also built a multilayer perceptron (MLP) with one hidden layer (followed by a ReLU) of dimensionality equal to its input. Each MLP was trained for 50 epochs using SGD with momentum $0.9$, weight decay \num{5e-4}, a batch size of 1 and initial learning rate of $0.001$.
During training, inputs were corrupted on-the-fly with additive white Gaussian noise ${\sim N(0,\sqrt{0.1})}$.
\subsubsection{Performance Metrics.}
For each classifier, we compute five metrics: accuracy, precision, recall, F1-score, and the area under the ROC curve (AUC), which plots the relationship between true positives and false positives at different operating thresholds of the classifier.
For these metrics, a perfect classifier would score 1 in all.
\subsubsection{Experiment Iterations.}
Initial attempts at classification showed that the classifiers' results for different metrics were dispersed.
To obtain accurate measures of each model's performance, and following the law of large numbers, we increased the number of iterations and looked at the distribution of results for each model.
For each iteration, we perform a random split of train and test data.
As expected, multiple iterations over the different splits of data yielded different results.
Overall, the mean of normally-distributed classification results can approximate the true value of each metric.
Also, the classes are not balanced and, therefore, we forced the same number of observations for each class by under-sampling the bigger class (\textit{i.e.},\xspace a random selection of observations available) while creating a split per iteration.
We perform two experiments: a preliminary one, to test the performance of each feature modality (text vs. images), and then the final one with an ensemble classifier that uses an average of the two preliminary ones.
Results for the classical ML algorithms were computed by executing 2,000 iterations of the classifiers on the available text or image data. For the neural network, we used 1,000 models to obtain the final classification.
\subsection{Predicting from Different Modalities:\\Text vs. Images}
Tables~\ref{table:text_classifiers_all} and~\ref{table:image_classifiers_all} show the performance of the considered classifiers, using \textit{Label II}, with textual and visual features, respectively.
These results were obtained by running 2,000 iterations and computing all metrics for each model.
As shown in the tables, all classifiers outperform the 50\% random baseline of binary classification implying that the signal separating \textit{fraud} from \textit{not-fraud} is present in the data.
Interestingly, tree-based models such as Decision Tree and Random Forest perform fairly well with AUC up to 0.84, just under the 0.93 AUC exhibited by neural networks on textual features.
We note that textual data alone provide better classification power than images alone (AUC=0.93 vs 0.67). However, the classification performance is improved by combining modalities.
\begin{table}
\centering
\caption{Average precision and accuracy for different classification algorithms, for the \textit{Label II} setup and with visual features considered (st. deviation shown in parenthesis).}
\label{table:image_classifiers_all}
\begin{tabular}{llll}
\toprule
\textbf{Classifier} & \textbf{Accuracy} & \textbf{F1-Score}& \textbf{AUC}\\ \midrule
SVM & 0.6590 (0.061) & 0.6586 (0.061) & 0.6622 (0.061)\\
\textit{k}-NN & 0.6350 (0.061) & 0.6319 (0.062) & 0.6620 (0.062)\\
Naive-Bayes & 0.6584 (0.062) & 0.6576 (0.062) & 0.6605 (0.061)\\
AdaBoost & 0.6423 (0.062) & 0.6419 (0.062) & 0.6609 (0.061)\\
Decision Tree & 0.5701 (0.064) & 0.5693 (0.065) & 0.6597 (0.060)\\
Random Forest & 0.6746 (0.061) & 0.6741 (0.062) & 0.6787 (0.061)\\
MLP & 0.6230 (0.056) & 0.6167 (0.061) & 0.6737 (0.063)\\
\bottomrule
\end{tabular}
\end{table}
\subsection{Automatically Detecting Campaign Fraud}
\begin{table}
\centering
\caption{Evaluation metrics for the ensemble classifiers using the \textit{Label II} setup (st. deviation shown in parenthesis).}
\label{table:ensemble_ABCII_compare}
\begin{tabular}{lcc
\toprule
Metric & \textbf{RF Ensemble} & \textbf{MLP Ensemble} \\ \midrule
Accuracy & 0.8517 (0.068) & 0.9014(0.034) \\
F1-Score & 0.8517 (0.068) & 0.9013(0.034) \\
AUC & 0.8539 (0.068) & 0.9601(0.022) \\
\bottomrule
\end{tabular}
\end{table}
The models based on text and images show definite separation between the class (\textit{fraud} or \textit{not fraud}).
The next step is to determine whether combining all features of a campaign into the same model provides improvement over treating them separately. Tables~\ref{table:text_classifiers_all} and~\ref{table:image_classifiers_all} show that RF is the best from the classical algorithms, but MLP outperforms RF in textual features.
Thus, we use both RF and MLP to evaluate the ensemble classifier performance.
Furthermore, Table~\ref{table:annotation_total} showed that information on each campaign varies: some have no images while others have multiple.
We first run the classification task separately for text and images, and then combine results into a single score for each campaign.
As before, we train and test RF with 2K runs, and the MLP on 1K models, on the \textit{Label II} setup. We then run an ablation study to determine whether any of the feature groups (\textit{i.e.},\xspace TFIDF, text Sentiment Analysis, Named-Entity Recognition, the Shape of the word, the Readability Index, the Descriptive elements of an image, the Objects present in an image, Emotions triggered by each image, and the number of faces recognized) have a negative interaction and should therefore be removed.
The results, shown in Table~\ref{table:ensemble_ABCII_compare} indicate that, while the neural network approach was comparable to the classical algorithms in terms of the separate modalities (\textit{i.e.},\xspace images and text), there is a clear improvement in all metrics when we combine all features in the same model, with AUC=$0.96$.
For completeness, Figure~\ref{fig:ABC_ensemble_distribution} presents the distribution of the individual evaluations of the neural network for the \textit{Label I} and \textit{Label II} setups.
As expected, \textit{Label II} performs better than \textit{Label I}.
Also, the models are not dispersed, and the results are consistently over 80\% of median performance.
\begin{figure}
\centering
\includegraphics[width=\linewidth]{figures/nn_ensemble_1Kruns}
\caption{Box diagram of the classification results for 1k models of the neural network (MLP) classifier.}
\label{fig:ABC_ensemble_distribution}
\end{figure}
\subsection{Clarity: Classifying imperfect data.}
In Section \textit{6.1.1}, we discussed the impact the labels have on the classification output.
Here, we investigate another configuration presented as \textit{Label III}.
This corresponds to the scenario where we train on campaigns with scores of \{1\} for \textit{fraud}, and campaigns with scores of \{5\} for \textit{not-fraud} (i.e., \textit{Label II} setup), and then test this model on campaigns with label scores \{2,4\}, corresponding to \textit{fraud} and \textit{not-fraud}, respectively.
These campaigns were dropped in \textit{Label II} setup, and thus were unseen by the classifier.
In Table~\ref{table:ABC_ensemble_distribution}, we compare the performance of modeling fraud with \textit{Label I}, \textit{Label II} and \textit{Label III} setups.
Overall, we observe that classifying on a stronger fraud signal (\textit{Label II}) translates into better performance.
Also, these results seem to indicate that once a model is trained with a sufficiently strong signal, it is able to correctly label noisy data (AUC = 0.936) on \textit{Label III}.
This shows great promise in terms of the extensibility and applicability of our work.
\begin{table}
\centering
\caption{Average results for the neural network ensemble classifier comparing fraud scale labels. The standard deviation is shown in parenthesis.}
\label{table:ABC_ensemble_distribution}
\begin{tabular}{lccc
\toprule
\textbf{Scores} & \textbf{Accuracy} & \textbf{F1-Score}& \textbf{AUC}\\ \midrule
\textit{Label I} & 0.8445(0.358) & 0.8438(0.036) & 0.9227(0.025)\\
\textit{Label II} & 0.9014(0.034) & 0.9012(0.034) & 0.9602(0.022) \\
\textit{Label III} & 0.9077(0.052) & 0.8996(0.067) & 0.9358(0.051) \\
\bottomrule
\end{tabular}
\end{table}
\section{Discussion and Future Work}
\label{sec:conclusion}
In recent years, crowdfunding has emerged as a means of making personal appeals for financial support to members of the public.
These may be simple tasks such as a DIY project at home, or more complex ventures such as starting a new company or medical procedures.
The community trusts that the individual who requests support, whatever the task, is doing so without malicious intent.
However, time and again, fraudulent cases come to light, ranging from fake objectives to embezzlement.
Fraudsters often fly under the radar and defraud people of what adds up to tens of millions, under the guise of crowdfunding support, enabled by small individual donations.
Detecting and preventing fraud is thus an adversarial problem.
Inevitably, perpetrators adapt and attempt to bypass whatever system is deployed to prevent their malicious schemes.
In this work, we take the first step in studying the problem of fraudulent crowdfunding campaigns and detecting them at the time of publication.
We collect appropriate data from thousands of campaigns from different platforms and study fraud cases to better understand their characteristics.
Armed with this knowledge, we perform an annotation study to label hundreds of campaigns as fraud or not, with substantial overall annotation agreement.
We proceed to extract characteristics (features) from the text and image content included in each campaign, and compare these features with the associated label of the campaign.
The dataset we built is useful in training machine learning ensemble classifiers, which can take visual and textual cues from any crowdfunding campaign, and predict if the campaign is fraudulent or not when created, with satisfactory performance (up to AUC=0.96).
Indeed, there is room for improvement, especially regarding feature engineering and classifier complexity and tuning.
However, our results demonstrate that it is possible to detect fraudulent campaigns with high certainty, and allow crowdfunding platforms to remove them semi-automatically, i.e., can be marked for a more detailed inspection by an administrator.
In practice, we are proposing an automatic method that can help donors to have an indication of which of the campaigns they are viewing may be fraudulent.
With this method, we attempt to make the job of fraudsters harder, by proposing a better system than currently available.
In fact, in order to mitigate the risk of fraudsters catching up with the online model and what features it monitors for predicting fraud, we can explore different methods and timings of when to deliver the warning flag to a donor.
In terms of limitations while building this methodology, we attempted to reduce any bias that may have been introduced by the annotators.
During this process, we created checklists and standards into what would be defined as fraud, to minimize subjective bias.
A further unbiased way to conduct this study would be to rely exclusively on convicted cases of fraud, instead of relying on manual annotations of suspected cases.
However, this option would not provide enough examples to develop good enough machine learning models, and it would be again up to annotators to identify \textit{not-fraud} examples.
One solution that would further reduce the risk of bias is to increase the number of annotators that label each campaign.
But this is highly depended on resource availability.
Finally, algorithmic bias could be reduced.
For example, poorly written campaigns by legitimate requestors who are uneducated or non-native English speakers can be mislabeled as fraud -- a clear source of bias.
Also, there may be limited examples of such campaigns, since these users may not be willing or comfortable to post a campaign in the first place.
These aspects point to the problem of fair and balanced representation of characteristics in our training data and labels.
In the future, we plan to improve our classifier to take into account such sources of bias.
We also plan to test our classifier on unlabeled data of medically-related campaigns to investigate its capability detecting such fraud cases, which in the health domain can have a severe monetary and emotional impact on the defrauded.
\subsection{Emotion Repesentations}
Human emotions can be induced by visual stimuli, such as images, as evidenced by psychological studies. As such, visual emotion prediction has attracted much interest from the computer vision community---framed as a multi-class classification problem using image-emotion input-output tuples for learning. Motivated by the foregoing successes in transfer learning, for visual emotion prediction, we repurposed a ResNet-152~\cite{he2016deep} ConvNet pre-trained on the ImageNet dataset~\cite{krizhevsky2012imagenet} of 1.2 million images of 1000 diverse object categories. Fine-tuning was performed by replacing the original 1000-way fully-connected classification layer with a newly initialised layer consisting of 8 neurons, which correspond to the emotion categories of interest. As defined in~\cite{zhao2014exploring,you2016building}, the categories were as follows: amusement, anger, awe, contentment, disgust, excitement, fear, and sadness.
For fine-tuning, we utilised the Flickr and Instagram (FI) dataset~\cite{you2016building} of 23K images, where each image is labelled as evoking one of the eight emotions based on a majority vote between five Amazon Mechanical Turk workers. We used 90\% of the images for training and the remainder for validation. For pre-processing, each image was resized to $256\times256\times3$ and standardised (per channel) based on the original ImageNet training data statistics. The model was fine-tuned for 100 epoochs, to minimise a negative log-likelihood loss, with stochastic gradient descent, using an initial learning rate of 0.1, momentum 0.9, and a batch size of 128. The learning rate was multiplied by a factor 0.1 at epochs 30, 60 and 90. We performed data augmentation by randomly cropping $224\times224\times3$ image patches, which is the resolution accepted by ResNet-152. During fine-tuning all layers except the classification layer were frozen. The final model accuracy on the validation data was 73.9\%, where predictions are based on central $224\times224\times3$ crops.
The semantic evidence over the eight emotions, in the form of logits (unnormalised log probabilities), can then be extracted for the crowdfunding images. For the crowdfunding images, each image was resized such that its shortest side was 256 pixels and then a central crop was extracted of size $224\times224\times3$. Semantic emotion category representations are then extracted from the classification layer.
\subsection{Apperance and Semantic Representations}
Again utilising a pre-trained ResNet-152 model, we extracted image appearance representations (conveying the existence of certain categorical appearance features) and semantic representations (expressing the logit evidence for the semantic categories). The appearance representations $\in\mathbb{R}^{2048}$ stem from the penultimate layer, whereas the semantic representations $\in\mathbb{R}^{1000}$ are extracted from the classification layer over the 1000 ImageNet classes. Such representations are useful since ConvNets are known to implicitly learn a level of correspondence between particular object parts~\cite{zeiler2014visualizing}. Moreoever, the representations invariably outperform their hand-engineerered counter-parts~\cite{chatfield2014return}.
For pre-processing, each crowdfunding image was resized such that its shortest side was 256 pixels and then a central crop was extracted of size $224\times224\times3$. Again, standardisation is performed (per channel) based on the original ImageNet training data statistics. Feature representations are then extracted from the two aforementioned layers for each image.
In addition to the ImageNet induced appearance representations, we also estimate the number of faces present in an image. This effectively binarises the set of images based on whether the image is believed to contain a face or not. For all images, faces were detected using the dlib~\cite{king2009dlib} HOG-based face detector. The number of faces detected, per image, is then used an additional feature.
|
Subsets and Splits
No community queries yet
The top public SQL queries from the community will appear here once available.