Data-Driven Energy-Based Learning via Gibbs Measures on Hierarchical Structures
Extracted source text · arXiv 2606.30064
\documentclass[11pt]{article}
\usepackage[utf8]{inputenc}
\usepackage[T1]{fontenc}
\usepackage{geometry}
\usepackage{setspace}
\geometry{margin=1in}
%\setstretch{1.2}
\usepackage{mathtools}
\usepackage{graphicx, epstopdf}
\usepackage[arrow, matrix]{xy}
\usepackage{xcolor}
\usepackage{cite}
\usepackage{float}
\usepackage{algorithmic}
\usepackage{algorithm}
\usepackage{amssymb, amsmath, amsthm}
\usepackage{wrapfig}
% Theorem environments
\newtheorem{thm}{Theorem}
\newtheorem{defn}{Definition}
\newtheorem{lemma}{Lemma}
\newtheorem{pro}{Proposition}
\newtheorem{rk}{Remark}
\newtheorem{rt}{Result}
\newtheorem{cor}{Corollary}
\newtheorem{question}{Question}
\newtheorem{ex}{Example}
% Math macros
\def\s{\sigma}
\def\f{\varphi}
\def\r{\rho}
\def\O{\Omega}
\def\w{\omega}
\def\e{\varepsilon}
\def\l{\lambda}
\def\z{\eta}
\def\t{\theta}
\def\d{\delta}
\def\a{\alpha}
\def\b{\beta}
\def\m{\mu}
\def\n{\nu}
\def\c{\gamma}
\def\M{\mathcal{M}}
\def\K{\mathcal{G}}
\def\R{\mathbb{R}}
\def\C{\mathbb{C}}
\def\Z{\mathbb{Z}}
\def\L{\Lambda}
\def\G{\Gamma}
\def\N{\mathcal{N}}
\DeclareMathOperator{\Der}{Der}
\def\ann{\operatorname{ann}}
\def\codim{\operatorname{codim}}
\def\Img{\operatorname{Im}}
\def\spa{\operatorname{span}}
% Roman-style enumerate
\newenvironment{romenum}
{\renewcommand{\theenumi}{\roman{enumi}}%
\renewcommand{\labelenumi}{(\theenumi)}%
\begin{enumerate}}{\end{enumerate}}
% Title
\title{Data-Driven Energy-Based Learning via Gibbs Measures on Hierarchical Structures}
\date{}
\vspace{-1cm}
\begin{document}
\maketitle
% Author block with affiliations directly underneath names
\begin{center}
\textbf{L.U. Abdullaev}\\
National University of Singapore, Singapore \\
Email: \texttt{laziz.abdullaev@u.nus.edu}
% \vspace{1em}
\textbf{F. Herrera} \\
Department of Computer Science and Artificial Intelligence, University of Granada, E-18071 Granada, Spain \\
Email: \texttt{herrera@decsai.ugr.es}
% \vspace{1em}
\textbf{U. A. Rozikov} \\
V.I. Romanovskiy Institute of Mathematics, Uzbekistan Academy of Sciences, 9, Universitet str., 100174, Tashkent, Uzbekistan; \\
National University of Uzbekistan, 4, Universitet str., 100174, Tashkent, Uzbekistan; \\
Institute for Advanced Study in Mathematics, Harbin Institute of Technology, \\Harbin 150001, China. \\
Email: \texttt{rozikovu@yandex.ru}\\
\textbf{(Corresponding author)}
% \vspace{1em}
\textbf{M. V. Velasco} \\
Departamento de Análisis Matemático, Facultad de Ciencias, Universidad de Granada, 18071 Granada, Spain \\
Email: \texttt{vvelasco@ugr.es}
\end{center}
%\vspace{1em}
\begin{abstract}
We introduce a data-driven probabilistic framework for learning systems based on Gibbs measures on hierarchical structures. Unlike standard empirical risk minimization, where a dataset is used to identify a single optimal parameter, our approach transforms the empirical loss function into an interaction potential defining an energy-based model. The resulting Gibbs distribution describes a family of equilibrium learning states generated by the data.
We formulate the consistency conditions of the associated finite-volume distributions and derive nonlinear integral fixed-point equations whose solutions characterize the admissible learning states. These equations provide a rigorous connection between empirical loss landscapes and probabilistic inference on trees. For translation-invariant solutions, the problem reduces to the analysis of positive compact operators induced by data-dependent kernels, allowing us to establish existence and uniqueness conditions in the one-dimensional setting.
Furthermore, we show that hierarchical learning systems may exhibit phase-transition phenomena: for certain empirical kernels on Cayley trees, multiple Gibbs measures emerge beyond a critical inverse temperature, corresponding to distinct equilibrium prediction regimes. Numerical experiments with non-separable kernels illustrate the appearance of multiple solution branches and demonstrate the coexistence of several data-induced learning states.
Our results provide a new perspective on energy-based learning, where data do not merely determine an optimal model through minimization but define an entire probabilistic landscape of possible inference states.
\end{abstract}
%\begin{abstract}
%This paper introduces a data-driven construction of Gibbs measures for learning
%systems on Cayley trees, where the empirical loss function is used as the
%Hamiltonian interaction potential. The resulting Gibbs measures, parameterized
%by the inverse temperature $\beta>0$, describe equilibrium states of the
%data-dependent system.
%
%We derive compatibility conditions for finite-volume Gibbs distributions and
%obtain a system of nonlinear integral equations whose solutions (boundary laws)
%characterize the corresponding Gibbs measures. For translation-invariant
%solutions, the problem is reduced to the spectral analysis of compact positive
%integral operators generated by the empirical loss kernel. This allows us to
%establish existence and uniqueness results for the one-dimensional case.
%
%For a kernel on a Cayley tree of order two, we show the existence of
%a critical value $\beta_c$ at which the number of translation-invariant Gibbs
%measures changes, leading to multiple equilibrium states and prediction rules
%in the corresponding learning system. Numerical experiments for a non-separable
%kernel further demonstrate the emergence of multiple branches of solutions
%and several Gibbs measures.
%
%The proposed framework provides a statistical mechanical interpretation of
%learning systems, where data-induced interactions generate an equilibrium
%model rather than only a single optimal parameter.
%\end{abstract}
\noindent \textbf{Mathematics Subject Classifications (2010):} 82B20, 62C10, 68T07, 60J10.\\
\noindent \textbf{Keywords:} Configuration, Ising model, empirical loss function, Gibbs measures, Cayley tree, machine learning.
\section{Introduction}
Machine learning has traditionally been formulated as an optimization problem,
where a training dataset is used to construct a model by minimizing an empirical
loss function.
This approach has been highly successful in modern learning systems. However,
many contemporary models involve complex non-convex landscapes, uncertainty,
and multiple competing solutions, motivating probabilistic approaches where
learning is represented not only by a single optimal parameter but also by a
distribution over possible states.
A natural framework for such probabilistic representations is provided by
energy-based models, where a learning system is described through an energy
function assigning different levels of compatibility to possible configurations.
This viewpoint connects machine learning with statistical mechanics, where
probability distributions are determined by energy landscapes. Energy-based
formulations have played an important role in machine learning through models
such as Boltzmann machines, neural networks, and probabilistic graphical models
\cite{AR,Ari,Bah, Beh, Krza, HRV, Le}. In these approaches, the equilibrium
distribution associated with an energy function provides a probabilistic
description of the possible states of the system.
A fundamental object in statistical mechanics is the Gibbs measure, which
describes equilibrium states of interacting systems through their Hamiltonian.
Gibbs measures provide a rigorous framework for studying phase transitions,
long-range dependence, and probabilistic structures generated by interacting
components \cite{FV,Ge}. In recent years, ideas from statistical mechanics have
become increasingly relevant for machine learning, particularly for understanding
energy landscapes, stochastic inference, and the collective behavior of
high-dimensional systems.
In this paper, we introduce a data-driven construction of Gibbs measures for
learning systems. Unlike classical statistical-mechanical models, where the
interaction potential is specified in advance, we construct the interaction
directly from the training data. More precisely, the empirical loss function of
a learning problem is interpreted as the interaction potential of the
Hamiltonian. Therefore, the dataset does not only determine a minimizer of the
loss function; instead, it generates a complete probabilistic model describing
the equilibrium states of the learning system.
The proposed framework is based on Gibbs measures on Cayley trees. Hierarchical
structures, and in particular Cayley trees, provide an important mathematical
setting for studying recursive probabilistic models and phase transitions. Due
to their tree structure, finite-volume Gibbs distributions satisfy exact
compatibility relations, and these relations lead to nonlinear recursive
equations characterizing the infinite-volume Gibbs measures. Such methods have
been extensively developed in the study of spin systems, phase transitions, and
Gibbs states on trees \cite{EE, Ro,Rbp}.
From the machine learning perspective, Cayley trees provide a natural framework
for hierarchical inference. Their recursive structure is closely related to
message-passing and belief-propagation methods used in probabilistic graphical
models. In our setting, the empirical loss function defines a data-dependent
interaction kernel, this kernel generates a Gibbs measure, and the resulting
equilibrium states represent possible global learning regimes induced by the
dataset.
The central mathematical problem of the paper is the characterization of these
data-induced Gibbs measures. We derive the compatibility conditions for
finite-volume Gibbs distributions and obtain nonlinear integral equations whose
solutions determine the corresponding equilibrium states. For
translation-invariant solutions, these equations reduce to the analysis of
positive compact integral operators generated by the empirical loss kernel.
Consequently, questions about uniqueness, multiplicity, and phase transitions
of learning states become questions about the spectral properties of
data-dependent operators.
The main contributions of this work are summarized as follows:
\begin{itemize}
\item We introduce a statistical-mechanical formulation of learning in which
empirical loss functions define the interaction potential of an energy-based
model.
\item We construct Gibbs measures associated with data-driven interaction
kernels and derive the nonlinear integral equations characterizing their
equilibrium states.
\item We establish existence and uniqueness results for translation-invariant
learning states using properties of positive compact operators.
\item We show that data-induced learning systems may exhibit phase-transition
phenomena, where several Gibbs measures coexist and correspond to different
equilibrium prediction regimes.
\item We provide numerical experiments for non-separable kernels demonstrating
the emergence of multiple solution branches and several data-generated learning
states.
\end{itemize}
The proposed approach provides a new probabilistic perspective on learning
systems. Instead of viewing learning only as the search for an optimal parameter,
we interpret the dataset as generating an equilibrium statistical structure
whose states describe possible inference regimes. This creates a rigorous
connection between empirical learning, energy-based models, and Gibbs theory on
hierarchical structures.
The remainder of the paper is organized as follows. Section 2 introduces the
mathematical framework and the construction of the data-driven Gibbs model.
Section 3 derives the compatibility equations defining the Gibbs measures.
Section 4 studies translation-invariant solutions and establishes analytical
results. Section 5 discusses prediction rules induced by the constructed Gibbs
measures. Section 6 presents numerical experiments illustrating multiple
equilibrium learning states.
%Statistical mechanics provides a powerful mathematical framework for describing
%complex systems composed of many interacting components. A fundamental object in
%this theory is the Gibbs measure, which describes the equilibrium distribution
%of a system through its Hamiltonian (energy function). Gibbs measures have played
%a central role in the study of phase transitions, long-range order, and
%probabilistic structures arising from interacting particle systems
%\cite{FV,Ge}.
%
%In recent years, concepts and methods from statistical mechanics have found
%increasing applications in machine learning and artificial intelligence. In
%particular, energy-based models represent learning systems as probabilistic
%models governed by an energy function, where lower-energy configurations
%correspond to more compatible states of the system. This viewpoint has led to important developments in
%Boltzmann machines, neural networks, and other probabilistic learning models
%\cite{AR,Ari,Bah, Beh, Krza, HRV, Le}.
%
%A natural mathematical framework for studying Gibbs measures is provided by
%hierarchical structures, especially Cayley trees. Due to their recursive
%geometry, Cayley trees allow the analysis of Gibbs states through nonlinear
%functional equations and have become an important setting for investigating
%phase transitions, boundary effects, and multiplicity of Gibbs measures in
%statistical mechanics (see \cite{EE, Ro,Rbp} and references therein). For spin
%systems on trees, the compatibility conditions of finite-volume Gibbs
%distributions lead to boundary law equations, which typically appear as
%nonlinear recursive relations or integral equations. The solutions of these
%equations determine the corresponding Gibbs measures and characterize the
%possible phases of the system.
%
%The aim of this paper is to introduce a data-driven construction of Gibbs
%measures for learning systems on Cayley trees. Unlike classical statistical
%mechanical models, where the interaction potential of the Hamiltonian is fixed
%in advance, we construct the interaction directly from a given training
%dataset. More precisely, the empirical loss function of a learning problem is
%interpreted as the interaction potential of the Hamiltonian. Consequently, the
%resulting Gibbs measure is not prescribed a priori but is generated by the
%statistical structure of the data.
%
%In this framework, a learning problem is transformed into a statistical
%mechanical system: the training dataset determines an interaction kernel, this
%kernel defines a Gibbs measure, and the corresponding equilibrium states
%represent possible configurations of the learning system. Thus, instead of
%selecting only a single optimal parameter through an optimization procedure, we
%associate the dataset with an entire family of equilibrium states described by
%a Gibbs distribution.
%
%The main contribution of this work is the establishment of a direct connection
%between Gibbs theory of interacting systems and data-driven learning models.
%
%The structure of the paper is as follows. In Section 2, we present preliminary
%results and discuss the motivation behind the proposed framework. In Section 3,
%we derive a system of nonlinear integral equations whose solutions generate
%compatible families of finite-dimensional Gibbs measures and, consequently,
%their limiting Gibbs measures. Section 4 is devoted to the analysis of these
%integral equations for Cayley trees of orders $1$ and $2$. In particular, we
%consider specific empirical loss functions and investigate the corresponding
%solutions. In Section 5, we give prediction rules for learning
%systems associated with the Gibbs measures constructed in Section 4.
%Finally, in Section 6, we provide numerical results for a non-separable kernel
%in the equation and show that the solution set has
%multiple branches, resulting in the emergence of several Gibbs measures.
\section{Preliminaries}
In this section, we recall the main definitions and introduce the necessary notation that will be used throughout the paper.
\subsection{Definitions} {\bf Cayley tree.} A Cayley tree \(\Gamma^k=(V,L)\), where \(V\) denotes the set of vertices and
\(L\) the set of edges, with branching factor \(k\geq 1\), is a connected
infinite graph in which every vertex has exactly \(k+1\) nearest neighbors.
The graph \(\Gamma^k\) is acyclic, i.e., it contains no closed paths or
cycles.
Although the tree itself has no boundary due to its infinite
structure, finite subtrees (truncations) of the Cayley tree are frequently
considered in physical and computational applications.
Fix a vertex \(x^0\in V\), which will be referred to as the \emph{root} of the tree.
A vertex \(y\in V\) is called a \emph{direct successor} of a vertex \(x\in V\) if \(x\) is
the predecessor of \(y\) on the unique path connecting the root \(x^0\) to \(y\).
Equivalently, this means that
\[
d(x^0,y)=d(x^0,x)+1,\qquad d(x,y)=1,
\]
where \(d(x,y)\) denotes the graph distance, i.e., the number of edges in the shortest
path connecting \(x\) and \(y\). The set of all direct successors of a vertex \(x\in V\)
is denoted by \(S(x)\).
For a fixed root \(x^0\in V\), the \emph{\(n\)-th level of the tree} is defined by
\[
W_n:=W_n(x^0)=\{x\in V:\ d(x,x^0)=n\},
\]
and the ball of radius \(n\) centered at \(x^0\) is given by
\begin{equation}\label{lp*}
V_n=\{x\in V:\ d(x,x^0)\leq n\},\qquad
L_n=\{l=\langle x,y\rangle\in L:\ x,y\in V_n\}.
\end{equation}
Since the root \(x^0\) is fixed throughout the paper, we omit it from the notation
whenever there is no risk of ambiguity.
For any vertex \(x\in W_n\), the set of its \emph{direct successors} is given by
\begin{equation}\label{sx}
S(x)=\{y\in W_{n+1}: \langle x,y\rangle\in L\}.
\end{equation}
The regular structure of the Cayley tree, characterized by the fact that
each vertex has exactly \(k+1\) neighbors, makes it particularly convenient
for analytical investigations. In particular, Cayley trees play an important
role in statistical mechanics as hierarchical lattices for studying phase
transitions and Gibbs measures. They also appear in various areas such as
network theory, ecology, and computer science, where they provide natural
models for branching processes and hierarchical structures.
{\bf The model with random parameters.} We consider models where the \emph{ spin} $\varphi$ takes
values from $\{-1, 1\}$ and is assigned to the vertices of tree (for motivations, see the next section).
% Hamiltonian and parameters are from the set $[0,1]$,
For $A\subset V$, a parameter configuration, $\sigma_A$, and a spin configuration, $\varphi$, on $A$
are given by an arbitrary functions
$$
\sigma_A : A \to [0,1], \qquad \varphi_A : A \to \{-1,1\}.
$$
Denote by $\Omega_A=[0,1]^A$ and $\Sigma_A=\{-1, 1\}^A$ the sets of all
parameter configurations and spin configurations on $A$ respectively. A configuration $\sigma$ on $V$ is then defined as a function
$$
x\in V\mapsto\sigma (x)\in [0,1].
$$
The \emph{set of all parameter configurations} is $\Omega=[0,1]^V$ and the set of all spin-configurations is $\Sigma=\{-1,1\}^V$.
The (formal) \emph{Hamiltonian} of the model is:
\begin{equation}\label{uy1}
H(\sigma, \varphi)=-\sum_{\langle x,y\rangle\in L}
\xi_{\sigma(x)\sigma(y)}\varphi(x)\varphi(y),
\end{equation}
where $\xi: (u,v)\in [0,1]^2\to \xi_{uv}\in \mathbb R$ is a given bounded,
measurable function\footnote{Given a parameter configuration $\sigma$, the quantity
$\xi_{\sigma(x)\sigma(y)}$ represents the realization of the random variable
$\xi_{\langle x,y\rangle}$ associated with the edge $\langle x,y\rangle$,
where this random variable is defined by
$
\xi_{\langle x,y\rangle}(\sigma)
=
\xi_{\sigma(x)\sigma(y)}.$}, representing interaction parameter between spin-values $\varphi(x),\varphi(y)$ on the $\langle x,y\rangle$
nearest neighbor vertices. We note that if $\xi_{\sigma(x)\sigma(y)}=J$, i.e., is constant, then Hamiltonian (\ref{uy1}) coincides with usual \emph{Ising model}.
Let $\l$ be the Lebesgue measure on $[0,1]$. Below we consider pair $(\sigma, \varphi)$ of configurations $\sigma \in \Omega$, $\varphi\in \Sigma$. On the set of all
configurations on $A$ the a priori measure $\l_A$ is introduced as
the $|A|$-fold product of the measure $\l$. Here and further on
$|A|$ denotes the cardinality of $A$. Therefore, $\lambda_A=\prod_{x\in A}\lambda$.
We consider a standard
sigma-algebra ${\mathcal B}$ of subsets of $X:=\Omega\times \Sigma$ generated
by the measurable cylinder subsets.
A probability measure $\mu$ on $(X,{\mathcal B})$
is called a \emph{Gibbs measure} (with Hamiltonian $H$) if it satisfies the DLR equation, namely for any
$n=1,2,\ldots$, and $\sigma_n\in\Omega_{V_n}$, $\varphi_n\in \Sigma_{V_n}$ :
$$\mu\left(\left\{x=(\sigma, \varphi)\in X :\;
\sigma\big|_{V_n}=\sigma_n, \varphi\big|_{V_n}=\varphi_n\right\}\right)= \int_{X}\mu ({\rm
d}\omega)\nu^{V_n}_{\omega|_{W_{n+1}}} (\sigma_n, \varphi_n),$$ where $\omega\in X$ and $\nu^{V_n}_{\omega|_{W_{n+1}}}$ is the
conditional Gibbs density
$$ \nu^{V_n}_{\omega|_{W_{n+1}}}(\sigma_n,\varphi_n)=\frac{1}{Z_n\left(
\omega\big|_{W_{n+1}}\right)}\exp\;\left(-\beta H
\left(\sigma_n,\,\varphi_n||\,\omega\big|_{W_{n+1}}\right)\right),
$$
and $\b={1\over T}$, where $T>0 $ is the temperature.
Here and below, $W_l$ stands for a `sphere' and $V_l$ for a
`ball' on the tree, of radius $l=1,2,\ldots$,
centered at a fixed vertex $x^0$ (an origin):
$$W_l=\{x\in V: d(x,x^0)=l\},\;\;V_l=\{x\in V: d(x,x^0)\leq l\};$$
and
$$L_n=\{\langle x,y\rangle\in L: x,y\in V_n\};$$
distance $d(x,y)$, $x,y\in V$, is the length of (i.e., the number of edges in) the shortest path
connecting $x$ with $y$. Furthermore, $\sigma\big|_{V_n}$, $\varphi\big|_{V_n}$ and $\omega\big|_{W_{n+1}}$ denote, respectively, the
restrictions of configurations $(\sigma,\varphi), w=
(o,p)\in X$ to $V_n$ and $W_{n+1}$. Next,
$\sigma_n:\;x\in V_n\mapsto \sigma_n(x)$ is a configuration in $V_n$ and
$$H\left(\sigma_n\, \varphi_n||\,\omega\big|_{W_{n+1}}\right)
=-\sum_{\langle x,y\rangle\in L_n}\xi_{\sigma_n(x)\sigma_n(y)}\varphi_n(x)\varphi_n(y)-\sum_{\langle x,y\rangle:\;x\in V_n, y\in W_{n+1}}
\xi_{\sigma_n(x)o_n(y)}\varphi_n(x)p_n(y).$$
Finally, $Z_n\left(\omega\big|_{W_{n+1}}\right)$
stands for the partition function in $V_n$, with
the boundary condition $\omega\big|_{W_{n+1}}$
%
% $$Z_n\left(\omega\big|_{W_{n+1}}\right)=
% \int_{\Omega_{V_n}} \exp\;\left(-\beta H \left({\widetilde\sigma}_n\,||\,\omega
% \big|_{W_{n+1}}\right)\right)\lambda_{V_n}(d{\widetilde\sigma}_n).$$
We use a standard definition of a translation-invariant measure.
The main object of study in this paper
is translation-invariant Gibbs measure for the model (\ref{uy1})
on Cayley tree.
\subsection{Motivation from Machine Learning and Novelty of our Approach}\label{mot}
Machine learning is traditionally formulated as an optimization problem.
Given a training dataset
\[
\mathcal D=\{(x_i,y_i)\}_{i=1}^{N},
\]
a prediction function $\mathcal F:\mathcal X\times\Theta\to\mathbb R$,
and a loss function $\ell:\mathbb R\times\mathbb R\to[0,\infty)$,
one constructs the empirical loss
\[
\mathcal L_N(\theta)
=
\frac1N\sum_{i=1}^{N}
\ell\!\bigl(y_i,\mathcal F(x_i;\theta)\bigr)
\]
and seeks an optimal parameter
\[
\theta^\ast=
\arg\min_{\theta}\mathcal L_N(\theta).
\]
The resulting parameter is subsequently used for prediction,
classification, regression, or inference. Thus, within the standard
learning paradigm, the dataset determines a distinguished optimal model
through empirical risk minimization.
The present work adopts a fundamentally different viewpoint.
Rather than using the dataset to select a single optimal parameter,
we use it to construct a statistical-mechanical model whose interaction
structure is generated directly by the empirical loss.
To each vertex \(x\) of a Cayley tree we assign a local parameter $\sigma(x)\in[0,1].$
For neighboring parameters \(t,u\in[0,1]\) we define the empirical
interaction loss
\[
\mathcal L_N(t,u)
=
\frac1N\sum_{i=1}^{N}
\ell\!\bigl(y_i,\mathcal F(x_i;t,u)\bigr),
\]
which quantifies the average prediction error associated with the pair
\((t,u)\) on the dataset.
Instead of minimizing this quantity, we regard it as an interaction
kernel
\[
\xi_{tu}=\mathcal L_N(t,u),
\]
and use it to define the Hamiltonian (\ref{uy1}).
Consequently, the dataset does not determine a single parameter value.
Rather, it determines an interaction kernel and therefore a Gibbs
probability distribution on the space of parameter configurations.
The primary objects of interest are not minimizers of a loss function
but Gibbs measures and their equilibrium states.
In this way, the learning problem is transformed into the study of the
probabilistic structure generated by the data.
This viewpoint naturally connects machine learning with statistical
mechanics.
Translation-invariant Gibbs measures may be interpreted as macroscopic
learning states induced by the dataset.
Uniqueness of a Gibbs measure corresponds to the existence of a unique
global learning regime, whereas the coexistence of several Gibbs
measures indicates the presence of multiple competing regimes.
From this perspective, phase transitions represent qualitative changes
in the global behavior generated by the data.
A further connection with machine learning arises from the consistency
equations for Gibbs measures.
These equations reduce to nonlinear integral fixed-point equations whose
solutions determine the equilibrium states of the model.
Such equations are closely related to message-passing and
belief-propagation procedures appearing in probabilistic graphical
models and Bayesian inference.
Therefore, the present framework establishes a rigorous bridge between
empirical-loss-based learning models, Gibbs measures on trees, and
modern inference methodologies.
\begin{rk}
The distinction between the present framework and classical machine
learning should be emphasized.
In standard empirical risk minimization, the dataset is used to select
an optimal parameter
\[
\theta^\ast=
\arg\min_{\theta}\mathcal L_N(\theta),
\]
and all subsequent predictions are based on this single parameter.
In contrast, the dataset in our model does not select a minimizer.
Instead, it generates the interaction kernel \(\xi(t,u)\) and therefore
determines a Gibbs probability distribution on the entire space of
configurations.
Consequently, the principal mathematical objects are equilibrium Gibbs
states rather than optimal parameters.
From this perspective, learning is described not by optimization alone
but by the statistical-mechanical structure induced by the empirical
loss.
\end{rk}
\begin{rk}
The dataset $\mathcal D=\{(x_i,y_i)\}_{i=1}^{N}$
may also be interpreted as a collection of observations on a finite
subset $\Lambda_N=\{x_1,\ldots,x_N\}$ of vertices of the Cayley tree.
The observed values \(y_i\) may represent the field values
\(\varphi(x_i)\), the local parameters \(\sigma(x_i)\), labels attached
to the vertices, or more general quantities derived from the local
configuration.
Thus the dataset contains only partial information about an unknown
global configuration on the tree.
The empirical loss generated by these observations determines the
interaction kernel \(\xi(t,u)\), which in turn defines a Gibbs measure
on the space of all admissible configurations.
Consequently, the Gibbs measure provides a probabilistic extension of
the finite dataset from the observed set \(\Lambda_N\) to the entire
tree.
For any unobserved vertex \(x\in V\setminus\Lambda_N\), the information
associated with that vertex can be inferred through the conditional
distributions induced by the Gibbs measure.
In particular, if \(y_i=\varphi(x_i)\), then a natural prediction is
given by
\[
\widehat{\varphi}(x)
=
\mathbb E_\mu
\!\left[
\varphi(x)
\,\Big|\,
\varphi(x_1)=y_1,\ldots,\varphi(x_N)=y_N
\right],
\]
where \(\mu\) denotes the Gibbs measure generated by the dataset.
Therefore, the Gibbs measure plays a role analogous to a probabilistic
predictor.
The finite dataset serves as observed training information, while the
Gibbs measure generated by the empirical loss provides a global
probabilistic description of all compatible configurations.
Prediction is performed through conditional distributions and
expectations rather than through a single optimal parameter.
In this sense, extending information from finitely many observed
vertices to the entire tree becomes an inference problem governed by
equilibrium states of a data-dependent statistical-mechanical system.
\end{rk}
\begin{rk}
The choice of a Cayley tree is motivated not by the geometry of the
dataset but by the role of the tree as a hierarchical inference
structure. A finite dataset provides local information at finitely many
vertices, and the central task is to propagate this information to
unobserved parts of the system. The Cayley tree offers the simplest
infinite graph with a recursive hierarchical organization and without
cycles. Consequently, the corresponding Gibbs measures satisfy exact
recursive consistency equations, which are closely related to
message-passing and belief-propagation algorithms used in machine
learning and probabilistic inference.
Moreover, the exponential growth of the tree reflects the rapid
expansion of possible information states generated from local
observations. For these reasons the Cayley tree provides a natural
mathematical framework for studying how empirical data induce global
probabilistic structures and learning regimes.
\end{rk}
The main results of this paper show that fundamental properties of the
learning system are governed by the spectral characteristics of an
integral operator generated by the empirical loss.
As a consequence, questions concerning the existence, uniqueness, and
multiplicity of Gibbs measures become questions about the global
structure of the learning problem induced by the dataset.
\section{An integral equation}\label{sec:integraleq}
Recall that we write $x<y$ if the path from $x^0$ to $y$ goes through $x$. Call vertex $y$ a \emph{direct successor} of
$x$ if $y>x$ and $x,y$ are nearest neighbors. Denote by $S(x)$ the set of direct successors of $x$.
Observe that any vertex $x\ne x^0$ has $k$ direct successors and $x^0$ has $k+1$.
Let $h:\;x\in V\mapsto h_x=(h_{t,u,x}, (t,u)\in [0,1]\times \{-1,1\}) \in \mathbb R^{[0,1]\times \{-1,1\}}$ be mapping of $x\in V\setminus
\{x^0\}$ with $|h_{t,u,x}|<C$ where $C$ is a constant which does not depend on $t,u$. Given
$n=1,2,\ldots$, consider the probability distribution $\mu^{(n)}$ on $\Omega_{V_n}\times \Sigma_{V_{n}}$ defined by
\begin{equation}\label{uy2}
\mu^{(n)}(\sigma_n,\varphi_n)=Z_n^{-1}\exp\left(-\beta H(\sigma_n, \varphi_n)
+\sum_{x\in W_n}h_{\sigma(x),\varphi(x), x}\right).
\end{equation}
Here $Z_n$ is the corresponding partition function.
% \begin{equation}\label{uy3}
% Z_n=\int_{\Omega_{V_n}}
% \exp\left(-\beta H({\widetilde\sigma}_n) +\sum_{x\in W_n}h_{{\widetilde\sigma}(x),x}\right)
% \lambda_{V_n}({d\widetilde\s_n}).
% \end{equation}
%
% \begin{rk} Note that $Z_n$ is finite, since $\l$ is a probability measure and $\exp(-\b
% H({\widetilde\sigma}_n)+$ $\sum_{x\in W_n} h_{{\widetilde\sigma}(x),x})$ is bounded on $\Omega
% _{V_n}.$
% \end{rk}
The probability distributions $\mu^{(n)}$ are compatible if for any $n\geq 1$ and configurations
$\sigma_{n-1}\in\Omega_{V_{n-1}}$, $\varphi_{n-1}\in \Sigma_{V_{n-1}}$ (see \cite{FV} and \cite{Ro} for several kind of compatibility):
\begin{equation}\label{uy4}
\sum_{p_n\in \Sigma_{W_n}}\int_{\Omega_{W_n}}\mu^{(n)}(\sigma_{n-1}\vee o_n, \varphi_{n-1}\vee p_n)\lambda_{W_n}(d(o_n))=
\mu^{(n-1)}(\sigma_{n-1}, \varphi_{n-1}).
\end{equation}
Here
$\sigma_{n-1}\vee o_n\in\Omega_{V_n}$ is the concatenation of
$\sigma_{n-1}$ and $o_n$.
In this case there exists a unique
measure $\mu$ such that
\begin{equation}\label{muu}
\mu\left(\left\{x=(\sigma, \varphi)\in X :\;
\sigma\big|_{V_n}=\sigma_n, \varphi\big|_{V_n}=\varphi_n\right\}\right)=\mu^{(n)}(\sigma_n, \varphi_n).
\end{equation}
\begin{defn}\label{uyd1}
The measure \(\mu\) defined by \eqref{muu} is called a
\emph{Gibbs measure} corresponding to the Hamiltonian (\ref{uy1}) with the
boundary fields \(h=\{h_x:x\in V\setminus\{x^0\}\}\) if the family of finite-volume
distributions \(\{\mu^{(n)}\}_{n\geq1}\) given by \eqref{uy2} is compatible in the
sense of \eqref{uy4}. In other words, \(\mu\) is the probability measure on
\(X\) whose finite-volume marginals coincide with the Gibbs distributions
\(\mu^{(n)}\) determined by the Hamiltonian and the boundary functions \(h_x\).
\end{defn}
\begin{pro}\label{uyp1} The probability distributions
$ \mu^{(n)}(\sigma_n,\varphi_n)$, $n=1,2,\ldots$, in (\ref{uy2}) are compatible iff for
any $x\in V\setminus\{x^0\}$
the following equation holds:
$$
f(t,x)=\prod_{y\in S(x)}{\int_0^1\eta_{tu}f(u,y)du +\int_0^1\eta^{-1}_{tu}g(u,y)du\over \int_0^1\eta^{-1}_{0u}f(u,y)du+\int_0^1\eta_{0u}g(u,y)du}.
$$
\begin{equation}\label{uy5}
g(t,x)=\prod_{y\in S(x)}{\int_0^1\eta^{-1}_{tu}f(u,y)du +\int_0^1\eta_{tu}g(u,y)du\over \int_0^1\eta^{-1}_{0u}f(u,y)du+\int_0^1\eta_{0u}g(u,y)du}.
\end{equation}
Here
$$f(t, x)=\exp(h_{t,+,x}-h_{0, -, x}), \ \ g(t, x)=\exp(h_{t,-,x}-h_{0, -, x}), \ \ \eta_{tu}=\exp(\beta \xi_{tu}).$$ and
$du=\l(du)$ is the Lebesgue measure.
\end{pro}
\begin{proof} {\sl Necessity.} Suppose that (\ref{uy4}) holds; we shall prove (\ref{uy5}).
Substituting (\ref{uy2}) into (\ref{uy4}), obtain that for any
configurations $\sigma_{n-1}$: $x\in V_{n-1}\mapsto\sigma_{n-1}(x)\in [0,1]$, $\varphi_{n-1}$: $x\in V_{n-1}\mapsto\varphi_{n-1}(x)\in \Sigma$:
\begin{equation}\label{uy6}\frac{Z_{n-1}}{Z_n} \sum_{p_n\in \Sigma_{W_n}}\int_{\Omega_{W_n}}
\exp\left(\sum_{x\in W_{n-1}}\sum_{y\in S(x)} (\beta \xi_{\sigma_{n-1}(x)\omega_n(y)}\varphi_{n-1}(x)p_n(y)+
h_{\omega_n(y),p_n(y),y})\right)\l_{W_n}(d\w_n)
\end{equation}
$$ =\exp\left(\sum_{x\in
W_{n-1}}h_{\sigma_{n-1}(x), \varphi_{n-1}(x),x}\right),$$
% where $\omega_n$: $x\in W_n\mapsto\omega_n(x)$.
From (\ref{uy6}) we get:
$${Z_{n-1}\over Z_n}\sum_{p_n\in \Sigma_{W_n}}\int_{\Omega_{W_n}}
\prod_{x\in W_{n-1}}\prod_{y\in S(x)} \exp\,(\beta \xi_{\sigma_{n-1}(x)\omega_n(y)}\varphi_{n-1}(x)p_n(y)+
h_{\omega_n(y),p_n(y),y})d(\w_n(y))$$
\begin{equation}\label{fh} = \prod_{x\in W_{n-1}} \exp\,(h_{\sigma_{n-1}(x), \varphi_{n-1}(x), x}). \end{equation}
Fix $x\in W_{n-1}$ and consider two configurations $\varphi_{n-1}=\overline{\varphi}_{n-1}$ and
$\varphi_{n-1}=\tilde{\varphi}_{n-1}$ on $W_{n-1}$ which coincide on $W_{n-1}\setminus \{x\}$,
and rewrite now the equality (\ref{fh}) for $\overline{\varphi}_{n-1}(x)=1$ and $\tilde{\varphi}_{n-1}(x)=-1$. Then dividing both of them to the equality corresponding to $\tilde{\varphi}_{n-1}(x)=-1$ and $t=0$
we get (\ref{uy5}).
{\sl Sufficiency.} Suppose that (\ref{uy5}) holds. It is equivalent to
the representations
\begin{equation}\label{uy7}
\prod_{y\in S(x)}\sum_{\epsilon\in \{-1,1\}}\int_0^1\exp\,(\beta \xi_{tu}s\epsilon +
h_{u,\epsilon, y})du= a(x)\exp\,(h_{t,s,x}), t\in [0,1], s=-1,+1
\end{equation}
for some function $a(x)>0, x\in V.$
We have
$${\rm LHS \ \ of \ \ (\ref{uy4})}=\frac{1}{Z_n}\exp(-\b H(\s_{n-1}, \varphi_{n-1}))
\l_{V_{n-1}}(d(\s_n))\times
$$
\begin{equation}\label{uy8}
\prod_{x\in W_{n-1}} \prod_{y\in S(x)}\sum_{\varepsilon\in \{-1,1\}}\int_0^1\exp\,(\beta \xi_{\s_{n-1}(x)u}\varphi_{n-1}(x)\varepsilon+
h_{u,\varepsilon,y})du,
\end{equation}
(where LHS means left hand side).
Substituting (\ref{uy7}) into (\ref{uy8}) and denoting $A_n(x)=\prod_{x \in W_{n-1}} a(x)$,
we get
\begin{equation}\label{uy9}
{\rm RHS\ \ of\ \ (\ref{uy8}) }=
\frac{A_{n-1}}{Z_n}\exp(-\b H(\s_{n-1},\varphi_{n-1})) \lambda_{V_{n-1}}(d\s)\prod_{x\in W_{n-1}}
h_{\s_{n-1}(x),\varphi_{n-1}(x),x}.
\end{equation}
Since $\mu^{(n)}$, $n \geq 1$ is a probability, we should have
$$ \sum_{\varphi_{n-1}\in \Sigma_{V_{n-1}}}\int_{\Omega_{V_{n-1}}}\lambda_{V_{n-1}}(d\s_{n-1})
\sum_{p_n\in \Sigma_{W_n}}\int_{\Omega_{W_n}} \lambda_{W_{n}}(d\omega_{n})
\mu^{(n)}(\sigma_{n-1}\vee \omega_n, \varphi_{n-1}\vee p_n) = 1. $$
Hence from (\ref{uy9}) we get $Z_{n-1}A_{n-1}=Z_n$, and (\ref{uy4}) holds.
\end{proof}
From Proposition \ref{uyp1} it follows that for any $h=\{h_x\in \mathbb R^{[0,1]\times \{-1,1\}},\
\ x\in V\}$ satisfying (\ref{uy5}) there exists a unique Gibbs measure
$\mu$ and vice versa. However, the analysis of solutions to (\ref{uy5}) is
not easy. This difficulty depends on the given function $\xi$. In
the next sections we will find several conditions of such
functions and give some solutions of the corresponding integral
equations.
\section{The case of Empirical loss Hamiltonian}
%\subsection*{1. Empirical loss as Hamiltonian}
Let \(\mathcal X\) be a prescribed input space and let
$
\{(x_i,y_i)\}_{i=1}^{N}\subset \mathcal X\times\mathbb R
$
be a given training dataset, where \(x_i\in\mathcal X\) denotes an input sample and \(y_i\in\mathbb R\) is the corresponding target value.
Assume that a prediction mechanism is specified by a given function
$
\mathcal F:\mathcal X\times [0,1]^2\to\mathbb R.
$
For every input \(x\in\mathcal X\) and every pair of local parameters
\((t,u)\in[0,1]^2\), the quantity
$
\mathcal F(x;t,u)
$
represents the prediction generated by two neighboring vertices carrying parameters \(t\) and \(u\).
Let
$
\ell:\mathbb R\times\mathbb R\to[0,\infty)
$
be a prescribed loss function that measures the discrepancy between a target value \(y\) and a prediction \(\hat y\). Typical examples include the squared loss
$
\ell(y,\hat y)=(y-\hat y)^2
$
and the absolute loss
$
\ell(y,\hat y)=|y-\hat y|.
$
For a fixed pair \((t,u)\), the empirical interaction loss associated with the dataset is defined by
\[
\mathcal L_N(t,u)
=
\frac1N
\sum_{i=1}^{N}
\ell\!\left( y_i,\mathcal{F}(x_i;t,u)\right)
\]
Thus, \(\mathcal L_N(t,u)\) represents the average prediction error produced by the neighboring parameters \(t\) and \(u\) over the training sample.
Set
\begin{equation}\label{eta}
\eta_{tu}=\exp\big(\beta \mathcal L_N(t,u)\big).
\end{equation}
%
% We interpret this as a Hamiltonian:
% \[
% H_N(\theta):=\mathcal{L}_N(\theta).
% \]
%
% The corresponding Gibbs measure on the parameter space is given by
% \[
% \mu_N(d\theta)\propto \exp\big(-\beta H_N(\theta)\big)\,d\theta.
% \]
To adapt this framework to the Cayley tree setting, we proceed as follows:
\begin{itemize}
\item Associate to each vertex \(x\in V\) a local parameter \(\theta_x \in [0,1]\);
\item Identify \(\sigma(x)\equiv \theta_x\);
\item Retain spin variables \(\varphi(x)\in\{-1,1\}\).
\end{itemize}
We define a data-dependent interaction along edges by
\[
\xi_{\sigma(x)\sigma(y)} \;\longrightarrow\;
\xi(\theta_x,\theta_y)
:=
\frac{1}{N}\sum_{i=1}^N
\ell\big(y_i,\mathcal F(x_i;\theta_x,\theta_y)\big).
\]
Then consider the Hamiltonian (\ref{uy1}) with this interaction parameters.
\subsection{Analysis of the solutions of the system (\ref{uy5}) in the case (\ref{eta}).}
{\bf The case \(k=1\).}
The Cayley tree of order \(k=1\) coincides with the one-dimensional integer lattice \(\mathbb Z\).
The root \(0\) is fixed; hence it suffices to solve the system independently on each half of \(\mathbb Z\).
Identifying vertices with \(n=0,1,2,\ldots\), we have
$
S(n)=\{n+1\}.
$
In this case the compatibility equations \eqref{uy5} become
\begin{equation}\label{k1-f}
f_n(t)=
\frac{\displaystyle \int_0^1 \eta_{tu}\,f_{n+1}(u)\,du
+\int_0^1 \eta_{tu}^{-1}\,g_{n+1}(u)\,du}
{\displaystyle \int_0^1 \eta_{0u}^{-1}\,f_{n+1}(u)\,du
+\int_0^1 \eta_{0u}\,g_{n+1}(u)\,du},
\end{equation}
\begin{equation}\label{k1-g}
g_n(t)=
\frac{\displaystyle \int_0^1 \eta_{tu}^{-1}\,f_{n+1}(u)\,du
+\int_0^1 \eta_{tu}\,g_{n+1}(u)\,du}
{\displaystyle \int_0^1 \eta_{0u}^{-1}\,f_{n+1}(u)\,du
+\int_0^1 \eta_{0u}\,g_{n+1}(u)\,du}.
\end{equation}
where $\eta_{tu}$ is given by (\ref{eta}).
We first study translation-invariant solutions
\[
f_n(t)\equiv f(t),
\qquad
g_n(t)\equiv g(t).
\]
Then the system is reduced to
\begin{equation}\label{f1}
f(t)=
\frac{\displaystyle \int_0^1 \eta_{tu}f(u)\,du
+\int_0^1 \eta_{tu}^{-1}g(u)\,du}
{D},
\end{equation}
\begin{equation}\label{g1}
g(t)=
\frac{\displaystyle \int_0^1 \eta_{tu}^{-1}f(u)\,du
+\int_0^1 \eta_{tu}g(u)\,du}
{D},
\end{equation}
where
\[
D=
\int_0^1 \eta_{0u}^{-1}f(u)\,du
+
\int_0^1 \eta_{0u}g(u)\,du.
\]
Introduce
\[
s(t)=f(t)+g(t),
\qquad
r(t)=f(t)-g(t).
\]
Adding and subtracting \eqref{f1} and \eqref{g1}, we obtain
$$ s(t) = \frac1D
\int_0^1
\bigl(\eta_{tu}+\eta_{tu}^{-1}\bigr)s(u)\,du, \ \
r(t)
=
\frac1D
\int_0^1
\bigl(\eta_{tu}-\eta_{tu}^{-1}\bigr)r(u)\,du.$$
Since \(\eta_{tu}=e^{\beta \mathcal L_N(t,u)}\), we have
\[
\eta_{tu}+\eta_{tu}^{-1}=2\cosh(\beta \mathcal L_N(t,u)),
\qquad
\eta_{tu}-\eta_{tu}^{-1}=2\sinh(\beta \mathcal L_N(t,u)).
\]
Thus,
$$
s(t)=\frac{2}{D}
\int_0^1 \cosh(\beta \mathcal L_N(t,u))\,s(u)\,du, \ \
r(t)=\frac{2}{D}
\int_0^1 \sinh(\beta \mathcal L_N(t,u))\,r(u)\,du.
$$
Define the operators
$$
(K_c\phi)(t)=\int_0^1 \cosh(\beta \mathcal L_N(t,u))\phi(u)\,du,
\ \
(K_s\phi)(t)=\int_0^1 \sinh(\beta \mathcal L_N(t,u))\phi(u)\,du.
$$
Then
\begin{equation}\label{eigen}
K_c s=\lambda s,
\qquad
K_s d=\lambda d,
\end{equation}
where \(\lambda=D/2\).
% Assume \(\mathcal L_N(t,u)\) is continuous on \([0,1]^2\). Then \(K_c\) and \(K_s\) are compact operators. Since \(K_c\) is strictly positive, the Krein–Rutman theorem \cite{Du} implies existence of a unique (up to scaling) positive eigenfunction \(s_*\).
%
% If \(\mathcal L_N(t,u)>0\), then \(K_s\) is also strictly positive and
% \[
% r(K_s)<r(K_c),
% \]
% which implies \(d\equiv 0\), hence \(f(t)=g(t)\). Uniqueness follows.
Assume that the empirical loss $\mathcal L_N(t,u)$ is continuous on $[0,1]^2$.
Then the kernels
\[
A(t,u)=\cosh\!\bigl(\beta \mathcal L_N(t,u)\bigr),
\qquad
B(t,u)=\sinh\!\bigl(\beta \mathcal L_N(t,u)\bigr)
\]
are continuous on $[0,1]^2$ as well.
Hence the associated integral operators
$
(K_c\phi)(t)$ and $(K_s\phi)(t)$
are compact operators on $C([0,1])$ and on $L^2([0,1])$.
Indeed, continuity of the kernels implies that $K_c$ and $K_s$ are Hilbert–Schmidt operators on $L^2([0,1])$, and therefore compact; compactness on $C([0,1])$ follows from the Arzelà–Ascoli theorem applied to the image of the unit ball.
%Moreover, both operators are positive in the sense that
%\[
%\phi \ge 0 \quad \Longrightarrow \quad K_c\phi \ge 0,\; K_s\phi \ge 0.
%\]
%
%\medskip
Since
\[
A(t,u)=\cosh\!\bigl(\beta \mathcal L_N(t,u)\bigr) > 0
\quad \text{for all } (t,u)\in[0,1]^2,
\]
it follows that $K_c$ is a strongly positive operator, i.e.,
\[
\phi \ge 0,\ \phi \not\equiv 0 \quad \Longrightarrow \quad (K_c\phi)(t) > 0 \ \text{for all } t.
\]
Therefore, by the Krein–Rutman theorem \cite{Du}, the spectral radius
$r(K_c)$ is a simple eigenvalue of $K_c$, and there exists a unique (up to multiplication by a positive scalar) eigenfunction
$s_* \in C([0,1])$, strictly positive on $[0,1]$, such that
\[
K_c s_* = r(K_c)\, s_*.
\]
Moreover, every nonnegative eigenfunction associated with $r(K_c)$ is proportional to $s_*$.
\medskip
To do positivity comparison between $K_c$ and $K_s$ we assume now that
\[
\mathcal L_N(t,u) > 0 \quad \text{for all } (t,u)\in[0,1]^2.
\]
Then
\[
0 < B(t,u) = \sinh\!\bigl(\beta \mathcal L_N(t,u)\bigr)
< \cosh\!\bigl(\beta \mathcal L_N(t,u)\bigr) = A(t,u),
\]
so that
\[
0 < K_s \le K_c \quad \text{(pointwise ordering of kernels)}.
\]
Both operators are therefore positive, and $K_s$ is also strongly positive.
By the monotonicity property of the spectral radius for positive compact operators,
\[
0 < K_s \le K_c \ \text{and } K_s \ne K_c
\quad \Longrightarrow \quad
r(K_s) < r(K_c).
\]
\medskip
\noindent
\textbf{Consequence for the antisymmetric component.}
Since the normalization constant $\lambda$ is determined by the system \eqref{eigen}, consistency of both equations would require a common eigenvalue of $K_c$ and $K_s$.
However, because
\[
r(K_s) < r(K_c),
\]
the only possible way for both equations to be compatible at the dominant spectral level is that the antisymmetric component vanishes. More precisely, $K_s$ cannot support a nontrivial positive or dominant eigenmode at the same spectral level as $K_c$.
Hence the only admissible solution in the translation-invariant class is
$
r(t) \equiv 0,
$
which implies $
f(t) = g(t).$
Thus the Gibbs-compatible boundary law is symmetric, and the corresponding Gibbs measure is unique in this regime.
%
% In the high-temperature regime \(\beta\ll1\),
% \[
% K_c=J+\frac{\beta^2}{2}M_2+O(\beta^4),
% \qquad
% K_s=\beta M+O(\beta^3),
% \]
% so \(r(K_s)<r(K_c)\) and uniqueness holds.
%
% In the low-temperature regime \(\beta\to\infty\), both kernels become asymptotically equivalent, allowing possible symmetry breaking, analogous to phase transitions.
\begin{thm}
Assume $\mathcal L_N(t,u)>0, (t,u)\in[0,1]^2.$
Then the compatibility equations \eqref{uy5} on a Cayley tree of order \(1\) admit a unique positive translation-invariant solution, determined by the dataset through \(\mathcal F\).
\end{thm}
\begin{proof}
Any translation-invariant solution satisfies \(f(t)=g(t)=h(t)\), leading to the eigenvalue problem
\[
K_N h = r_N h,
\quad
K_N\phi=\int_0^1 \cosh(\beta \mathcal L_N(t,u))\phi(u)\,du.
\]
By Krein–Rutman theorem, a unique positive eigenfunction exists, giving the desired solution.
\end{proof}
% \subsection{Explicit solutions for particular datasets}
%
% (remaining cases unchanged except for $\mathcal F$ replacement where needed)
%
%
% Let
% \[
% \mathcal F:\mathcal X\times [0,1]^2\to \mathbb R
% \]
% be a prediction function. For an input $x\in\mathcal X$ and local
% parameters $t,u\in[0,1]$, the quantity
% \[
% \mathcal F(x;t,u)
% \]
% represents the prediction produced by the pair of neighboring vertices
% carrying parameters $t$ and $u$.
%
% The empirical interaction loss associated with the dataset
% $\{(x_i,y_i)\}_{i=1}^N$ is defined by
% \[
% \mathcal L_N(t,u)
% =
% \frac1N
% \sum_{i=1}^{N}
% \ell\!\bigl(y_i,\mathcal F(x_i;t,u)\bigr).
% \]
% We interpret this as a Hamiltonian:
% \[
% H_N(\theta):=\mathcal{L}_N(\theta).
% \]
%
% The corresponding Gibbs measure on the parameter space is given by
% \[
% \mu_N(d\theta)\propto \exp\big(-\beta H_N(\theta)\big)\,d\theta.
% \]
%
% %\subsection*{2. Tree-structured parameter decomposition}
%
% To adapt this framework to the Cayley tree setting, we proceed as follows:
%
% \begin{itemize}
% \item Associate to each vertex $x\in V$ a local parameter $\theta_x \in [0,1]$;
% \item Identify $\sigma(x)\equiv \theta_x$;
% \item Retain spin variables $\varphi(x)\in\{-1,1\}$.
% \end{itemize}
%
% We define a data-dependent interaction along edges by
% \[
% \xi_{\sigma(x)\sigma(y)} \;\longrightarrow\;
% \xi(\theta_x,\theta_y)
% :=\frac{1}{N}\sum_{i=1}^N
% \ell\big(y_i,f(x_i;\theta_x,\theta_y)\big).
% \]
%
% Then the Hamiltonian takes the form
% \[
% H_N(\sigma,\varphi)
% =-\sum_{\langle x,y\rangle}
% \xi(\theta_x,\theta_y)\,\varphi(x)\varphi(y).
% \]
%
%
%
% Let $(x_i,y_i)_{i=1}^N$ be a dataset and let $\ell(y,f(x;\theta))$ be a loss function.
% Define the empirical loss
% \[
% \mathcal{L}_N(t,u)=\frac{1}{N}\sum_{i=1}^N \ell\big(y_i,f(x_i;t,u)\big),
% \]
% and set
% \[
% \eta_{tu}=\exp\big(\beta \mathcal{L}_N(t,u)\big).
% \]
%
%% Then by the above proved proposition we know that the probability distributions $\mu^{(n)}$, $n=1,2,\ldots$, are compatible
%% if and only if for any $x\in V\setminus\{x^0\}$ the following system of
%% integral equations holds:
%% \begin{equation}\label{emp1}
%% f(t,x)=\prod_{y\in S(x)}
%% \frac{\int_0^1 \eta_{tu}\, f(u,y)\,du
%% +\int_0^1 \eta_{tu}^{-1}\, g(u,y)\,du}
%% {\int_0^1 \eta_{0u}^{-1}\, f(u,y)\,du
%% +\int_0^1 \eta_{0u}\, g(u,y)\,du},
%% \end{equation}
%% \begin{equation}\label{emp2}
%% g(t,x)=\prod_{y\in S(x)}
%% \frac{\int_0^1 \eta_{tu}^{-1}\, f(u,y)\,du
%% +\int_0^1 \eta_{tu}\, g(u,y)\,du}
%% {\int_0^1 \eta_{0u}^{-1}\, f(u,y)\,du
%% +\int_0^1 \eta_{0u}\, g(u,y)\,du}.
%% \end{equation}
%%
% \subsubsection{Solution of the system (\ref{uy5}).}
%
% {\bf The case $k=1$.} The Cayley tree of order $k=1$, coincides with one-dimensional integer lattice $\mathbb Z$. The root 0 is fixed, therefore it suffices to solve the system independently on the each half of $\mathbb Z$. Identifying the vertices with $n=0,1,2,\ldots$, we have
% \[
% S(n)=\{n+1\}.
% \]
% In this case the compatibility equations \eqref{uy5} become
% \begin{equation}\label{k1-f}
% f_n(t)=
% \frac{\displaystyle \int_0^1 \eta_{tu}\,f_{n+1}(u)\,du
% +\int_0^1 \eta_{tu}^{-1}\,g_{n+1}(u)\,du}
% {\displaystyle \int_0^1 \eta_{0u}^{-1}\,f_{n+1}(u)\,du
% +\int_0^1 \eta_{0u}\,g_{n+1}(u)\,du},
% \end{equation}
% \begin{equation}\label{k1-g}
% g_n(t)=
% \frac{\displaystyle \int_0^1 \eta_{tu}^{-1}\,f_{n+1}(u)\,du
% +\int_0^1 \eta_{tu}\,g_{n+1}(u)\,du}
% {\displaystyle \int_0^1 \eta_{0u}^{-1}\,f_{n+1}(u)\,du
% +\int_0^1 \eta_{0u}\,g_{n+1}(u)\,du},
% \end{equation}
% where
% \[
% \eta_{tu}
% =
% \exp\!\bigl(\beta \mathcal L_N(t,u)\bigr).
% \]
%
% We first study translation-invariant solutions
% \[
% f_n(t)\equiv f(t),
% \qquad
% g_n(t)\equiv g(t).
% \]
% Then \eqref{k1-f}--\eqref{k1-g} reduce to
% \begin{equation}\label{TI-f}
% f(t)=
% \frac{\displaystyle \int_0^1 \eta_{tu}f(u)\,du
% +\int_0^1 \eta_{tu}^{-1}g(u)\,du}
% {D},
% \end{equation}
% \begin{equation}\label{TI-g}
% g(t)=
% \frac{\displaystyle \int_0^1 \eta_{tu}^{-1}f(u)\,du
% +\int_0^1 \eta_{tu}g(u)\,du}
% {D},
% \end{equation}
% where
% \[
% D=
% \int_0^1 \eta_{0u}^{-1}f(u)\,du
% +
% \int_0^1 \eta_{0u}g(u)\,du.
% \]
%
% Introduce
% \[
% s(t)=f(t)+g(t),
% \qquad
% r(t)=f(t)-g(t).
% \]
% Adding and subtracting \eqref{TI-f} and \eqref{TI-g}, we obtain
% \begin{equation}\label{s-equation}
% s(t)
% =
% \frac1D
% \int_0^1
% \bigl(\eta_{tu}+\eta_{tu}^{-1}\bigr)
% s(u)\,du,
% \end{equation}
% and
% \begin{equation}\label{d-equation}
% r(t)
% =
% \frac1D
% \int_0^1
% \bigl(\eta_{tu}-\eta_{tu}^{-1}\bigr)
% r(u)\,du.
% \end{equation}
%
% Since
% \[
% \eta_{tu}=e^{\beta\mathcal L_N(t,u)},
% \]
% we have
% \[
% \eta_{tu}+\eta_{tu}^{-1}
% =
% 2\cosh\!\bigl(\beta\mathcal L_N(t,u)\bigr),
% \]
% and
% \[
% \eta_{tu}-\eta_{tu}^{-1}
% =
% 2\sinh\!\bigl(\beta\mathcal L_N(t,u)\bigr).
% \]
% Therefore,
% \begin{equation}\label{spec1}
% s(t)=\frac{2}{D}
% \int_0^1
% \cosh\!\bigl(\beta\mathcal L_N(t,u)\bigr)
% s(u)\,du,
% \end{equation}
% \begin{equation}\label{spec2}
% r(t)=\frac{2}{D}
% \int_0^1
% \sinh\!\bigl(\beta\mathcal L_N(t,u)\bigr)
% r(u)\,du.
% \end{equation}
%
% Define the integral operators
% \[
% (K_c\phi)(t)
% =
% \int_0^1
% \cosh\!\bigl(\beta\mathcal L_N(t,u)\bigr)\phi(u)\,du,
% \]
% and
% \[
% (K_s\phi)(t)
% =
% \int_0^1
% \sinh\!\bigl(\beta\mathcal L_N(t,u)\bigr)\phi(u)\,du.
% \]
% Then \eqref{spec1}--\eqref{spec2} become
% \begin{equation}\label{eigen}
% K_c s=\lambda s,
% \qquad
% K_s d=\lambda d,
% \end{equation}
% where
% \[
% \lambda=\frac{D}{2}.
% \]
%
% Assume that the empirical loss
% \[
% \mathcal L_N(t,u)
% =
% \frac1N\sum_{i=1}^N
% \ell(y_i,f(x_i;t,u))
% \]
% is continuous on $[0,1]^2$.
% Then the kernels
% \[
% A(t,u)=\cosh(\beta\mathcal L_N(t,u)),
% \qquad
% B(t,u)=\sinh(\beta\mathcal L_N(t,u))
% \]
% are continuous. Consequently, the operators $K_c$ and $K_s$ are compact on
% $C[0,1]$ and on $L^2([0,1])$. Therefore their spectra consist of isolated eigenvalues of finite multiplicity, with $0$ as the only possible accumulation point.
%
% Since $ A(t,u)>0$, the operator $K_c$ is strongly positive. By the Krein--Rutman theorem\footnote{https://en.wikipedia.org/wiki/Krein-Rutman$_-$theorem}, its spectral radius $r(K_c)$ is a simple eigenvalue and there exists a unique (up to a multiplicative constant) strictly positive eigenfunction $s_*(t)$ satisfying
% \[
% K_c s_*=r(K_c)s_*.
% \]
% Hence every positive solution of \eqref{spec1} is proportional to $s_*$.
%
% Suppose now that
% \[
% \mathcal L_N(t,u)>0
% \]
% for all $(t,u)\in[0,1]^2$.
% Then $B(t,u)>0$ and $K_s$ is also strongly positive. Again by the Krein--Rutman theorem, the spectral radius $r(K_s)$ is a simple positive eigenvalue associated with a positive eigenfunction $d_*(t)$.
%
% Observe that
% \[
% 0\le \sinh z < \cosh z,
% \qquad z>0.
% \]
% Therefore
% \[
% 0\le B(t,u)<A(t,u).
% \]
% The comparison theorem for positive compact operators yields
% \[
% r(K_s)<r(K_c).
% \]
% Consequently, the equation
% \[
% K_s d=r(K_c)d
% \]
% admits only the trivial solution
% \[
% d\equiv0.
% \]
% Hence
% \[
% f(t)=g(t),
% \]
% and the compatible Gibbs measure is unique.
%
% Next, consider the high-temperature regime $\beta\ll1$. Using the expansions
% \[
% \cosh(\beta L)
% =
% 1+\frac{\beta^2L^2}{2}+O(\beta^4),
% \]
% and
% \[
% \sinh(\beta L)
% =
% \beta L+O(\beta^3),
% \]
% we obtain
% \[
% K_c
% =
% J+\frac{\beta^2}{2}M_2+O(\beta^4),
% \]
% where
% \[
% (J\phi)(t)=\int_0^1\phi(u)\,du,
% \]
% and
% \[
% (M_2\phi)(t)
% =
% \int_0^1L(t,u)^2\phi(u)\,du.
% \]
% Since $J$ has spectral radius $1$,
% \[
% r(K_c)=1+O(\beta^2).
% \]
%
% Similarly,
% \[
% K_s=\beta M+O(\beta^3),
% \]
% where
% \[
% (M\phi)(t)
% =
% \int_0^1L(t,u)\phi(u)\,du.
% \]
% Hence
% \[
% r(K_s)=\beta r(M)+O(\beta^3).
% \]
% Therefore, for sufficiently small $\beta$,
% \[
% r(K_s)<r(K_c),
% \]
% which guarantees uniqueness of the Gibbs measure.
%
% In the low-temperature regime $\beta\to\infty$, we have
% \[
% \cosh(\beta L)
% =
% \frac12 e^{\beta L}(1+o(1)),
% \qquad
% \sinh(\beta L)
% =
% \frac12 e^{\beta L}(1+o(1)).
% \]
% Consequently,
% \[
% K_c-K_s
% =
% \int_0^1 e^{-\beta L(t,u)}(\cdot)\,du,
% \]
% and thus
% \[
% r(K_c)-r(K_s)\to0.
% \]
% Therefore the low-temperature regime is the only possible regime in which non-symmetric solutions may emerge, analogous to the appearance of phase transitions in classical Gibbs measures.
%
% Summarizing, for a Cayley tree of order $1$, the compatibility equations reduce to the spectral problem \eqref{eigen}. The symmetric solution corresponds to the principal eigenfunction of $K_c$, while a non-symmetric solution can exist only if $K_c$ and $K_s$ possess a common eigenvalue. In particular, if $\mathcal L_N(t,u)>0$, then
% \[
% r(K_s)<r(K_c),
% \]
% and the corresponding Gibbs measure is unique.
%
% \begin{thm}
% Assume that
% \[
% \mathcal L_N(t,u)>0,
% \qquad (t,u)\in [0,1]^2.
% \]
% Then the compatibility equations \eqref{emp1}--\eqref{emp2} on a Cayley tree of order $1$ admit a unique positive translation-invariant solution. Moreover, this solution is completely determined by the dataset $(x_i,y_i)_{i=1}^N$ through the empirical loss function
% \[
% \mathcal L_N(t,u)
% =
% \frac1N\sum_{i=1}^{N}
% \ell\!\bigl(y_i,f(x_i;t,u)\bigr).
% \]
% \end{thm}
%
% \begin{proof}
% From the previous analysis, any translation-invariant solution satisfies
% \[
% f(t)=g(t)=h(t).
% \]
% Substituting this relation into \eqref{TI-f}, we obtain
% \[
% h(t)
% =
% \frac{\displaystyle
% \int_0^1
% \bigl(\eta_{tu}+\eta_{tu}^{-1}\bigr)
% h(u)\,du}
% {\displaystyle
% \int_0^1
% \bigl(\eta_{0u}+\eta_{0u}^{-1}\bigr)
% h(u)\,du}.
% \]
% Since
% \[
% \eta_{tu}
% =
% e^{\beta \mathcal L_N(t,u)},
% \]
% it follows that
% \[
% h(t)
% =
% \frac{\displaystyle
% \int_0^1
% \cosh\!\bigl(\beta \mathcal L_N(t,u)\bigr)
% h(u)\,du}
% {\displaystyle
% \int_0^1
% \cosh\!\bigl(\beta \mathcal L_N(0,u)\bigr)
% h(u)\,du}.
% \]
%
% Define the integral operator
% \[
% (K_N\phi)(t)
% =
% \int_0^1
% \cosh\!\bigl(\beta \mathcal L_N(t,u)\bigr)
% \phi(u)\,du.
% \]
% The kernel
% \[
% A_N(t,u)
% =
% \cosh\!\bigl(\beta \mathcal L_N(t,u)\bigr)
% \]
% is continuous and strictly positive. Therefore, by the Krein--Rutman theorem, $K_N$ possesses a unique positive eigenfunction $\psi_N$ corresponding to the maximal eigenvalue $r_N=r(K_N)$:
% \[
% K_N\psi_N=r_N\psi_N.
% \]
%
% Explicitly,
% \[
% \int_0^1
% \cosh\!\bigl(\beta \mathcal L_N(t,u)\bigr)
% \psi_N(u)\,du
% =
% r_N\psi_N(t).
% \]
%
% Since solutions of the compatibility equations are defined up to a positive multiplicative constant, we normalize by
% \[
% \psi_N(0)=1.
% \]
% Then
% \[
% h(t)=\psi_N(t).
% \]
%
% Consequently,
% \[
% \boxed{
% f(t)=g(t)=\psi_N(t),
% }
% \]
% where $\psi_N$ is the unique positive solution of
% \[
% \boxed{
% \int_0^1
% \cosh\!\left(
% \frac{\beta}{N}
% \sum_{i=1}^{N}
% \ell\!\bigl(y_i,f(x_i;t,u)\bigr)
% \right)
% \psi_N(u)\,du
% =
% r_N\,\psi_N(t).
% }
% \]
%
% Therefore the unique compatible Gibbs measure is completely determined by the dataset through the empirical-loss kernel
% \[
% \boxed{
% A_N(t,u)
% =
% \cosh\!\left(
% \frac{\beta}{N}
% \sum_{i=1}^{N}
% \ell\!\bigl(y_i,f(x_i;t,u)\bigr)
% \right).
% }
% \]
% The corresponding boundary law is
% \[
% \boxed{
% f(t)=g(t)=
% \frac{\psi_N(t)}{\psi_N(0)}.
% }
% \]
% This completes the proof.
% \end{proof}
%
\subsection{Explicit solutions for particular datasets}
In general, the positive eigenfunction $\psi_N$, satisfying
\begin{equation}\label{eigen-problem}
\int_0^1
\cosh\!\bigl(\beta \mathcal L_N(t,u)\bigr)
\psi_N(u)\,du
=
r_N\,\psi_N(t),
\end{equation}
cannot be determined in closed form. However, explicit solutions can be obtained for several important classes of empirical loss functions.
\subsubsection{Case 1. Constant empirical loss}
Suppose that
\[
\mathcal L_N(t,u)\equiv c.
\]
Then
\[
K_N\phi
=
\cosh(\beta c)\int_0^1\phi(u)\,du.
\]
Hence $K_N$ is a rank-one operator.
The eigenvalue equation becomes
\[
\cosh(\beta c)\int_0^1\psi_N(u)\,du
=
r_N\psi_N(t).
\]
Therefore $\psi_N(t)$ must be constant. After normalization,
$\psi_N(t)\equiv 1,
$
and
$
r_N=\cosh(\beta c).
$
Consequently, $f(t)=g(t)=1.$
\subsubsection{Case 2. Additive empirical loss}
Assume that the empirical loss has the additive structure
\[
\mathcal L_N(t,u)=a(t)+a(u),
\]
where \(a:[0,1]\to\mathbb R\) is a continuous function. Let
\begin{equation} \label{nuevo}
p(t)=e^{\beta a(t)}, \qquad q(t)=e^{-\beta a(t)}.
\end{equation}
Then the kernel factorizes as
$
\eta_{tu}=p(t)p(u)$, $\eta_{tu}^{-1}=q(t)q(u).
$
The translation-invariant compatibility equations reduce to
\[
f(t)=
\frac{
p(t)\int_0^1 p(u)f(u)\,du
+
q(t)\int_0^1 q(u)g(u)\,du
}{D},
\]
\[
g(t)=
\frac{
q(t)\int_0^1 q(u)f(u)\,du
+
p(t)\int_0^1 p(u)g(u)\,du
}{D},
\]
where
\[
D=\int_0^1 q(u)f(u)\,du+\int_0^1 p(u)g(u)\,du.
\]
Introduce the scalar quantities
\[
A=\int_0^1 p(u)f(u)\,du, \quad
B=\int_0^1 q(u)g(u)\,du, \quad
C=\int_0^1 q(u)f(u)\,du, \quad
E=\int_0^1 p(u)g(u)\,du.
\]
\begin{rk}\label{rkp}
By Proposition 1, we have that \(f(t)\) and \(g(t)\) are strictly positive, while \(p(t)\) and \(q(t)\) are strictly positive by (20). Hence,
\[
A,B,C,D,E>0.
\]
\end{rk}
With this notation, the solutions take the form
\begin{equation}\label{fA}
f(t)=\frac{A p(t)+B q(t)}{D}, \qquad
g(t)=\frac{C q(t)+E p(t)}{D}.
\end{equation}
Thus every solution of the compatibility equations belongs to the
two-dimensional space
\[
\operatorname{span}\{e^{\beta a(t)},\,e^{-\beta a(t)}\}.
\]
\medskip
{\bf Case $f=g$.} We now restrict ourselves to symmetric translation-invariant solutions,
that is,
\[
f=g=h.
\]
In this case,
\[
h(t)=\frac{A e^{\beta a(t)}+B e^{-\beta a(t)}}{A+B}
=\alpha e^{\beta a(t)}+(1-\alpha)e^{-\beta a(t)},
\]
where
\begin{equation}\label{alp}
\alpha=\frac{A}{A+B}.
\end{equation}
By Remark \ref{rkp} we have that $\alpha\in (0,1).$
Substituting this representation into the definitions of \(A\) and \(B\),
we obtain
\[
A=\alpha I_2+(1-\alpha), \qquad
B=\alpha+(1-\alpha)I_{-2},
\]
where
\begin{equation}\label{I12}
I_2=\int_0^1 e^{2\beta a(u)}\,du, \qquad
I_{-2}=\int_0^1 e^{-2\beta a(u)}\,du.
\end{equation}
Using (\ref{alp}), we arrive at the consistency equation
\[
\alpha(1-\alpha)(I_2-I_{-2})+1-2\alpha=0.
\]
Let $\Delta:=I_2-I_{-2}$.
Then the equation can be rewritten as
\[
\Delta\alpha^2-(\Delta-2)\alpha-1=0.
\]
If \(\Delta\neq0\), the solutions are
\[
\alpha_{\pm}
=
\frac{\Delta-2\pm\sqrt{\Delta^2+4}}{2\Delta}.
\]
A direct inspection shows that exactly one root lies in \((0,1)\):
\[
\alpha_+\in(0,1), \quad \alpha_-<0 \quad \text{if } \Delta>0,
\]
\[
\alpha_-\in(0,1), \quad \alpha_+>1 \quad \text{if } \Delta<0.
\]
In the degenerate case \(\Delta=0\), the equation reduces to
$
\alpha=\frac12.
$
%Note that the function
%\[
%h(t)=\alpha e^{\beta a(t)}+(1-\alpha)e^{-\beta a(t)}
%\]
%is strictly positive for all \(\alpha\in (0,1)\).
Define
$$
\alpha^*=\left\{\begin{array}{lll}
\alpha_- \ \ \text{if} \ \ \Delta<0,\\[2mm]
\frac12, \ \ \text{if} \ \ \Delta=0,\\[2mm]
\alpha_+, \ \ \text{if} \ \ \Delta>0.
\end{array}\right.
$$
Then \(\alpha^*\in (0,1)\), and the corresponding function
\begin{equation}\label{hu}
h(t)
=
\alpha^* e^{\beta a(t)}
+
(1-\alpha^*)e^{-\beta a(t)}
\end{equation}
is strictly positive and generates a symmetric translation-invariant Gibbs measure.
Thus the additive empirical loss yields a unique symmetric
translation-invariant Gibbs measure determined by the solution
\(\alpha^*\) of the above quadratic equation.
{\bf Case $f\ne g$.} We now consider the non-symmetric case \(f \neq g\).
Introduce the representation
\begin{equation}\label{fx}
f(t)=x\,p(t)+y\,q(t), \qquad
g(t)=z\,q(t)+w\,p(t),
\end{equation}
for some positive constants \(x,y,z,w>0\). Substituting into the definitions of
\(A,B,C,E\), we obtain the linear relations
\begin{equation}\label{AB}
A = x I_2 + y, \qquad
B = z I_{-2} + w, \qquad
C = x + y I_{-2}, \qquad
E = z + w I_2,
\end{equation}
where $I_2$ and $I_{-2}$ are defined by (\ref{I12}).
On the other hand, comparing coefficients in (\ref{fA}) and (\ref{fx})
yields the algebraic consistency system
\begin{equation}\label{xy}
x = \frac{A}{D}, \qquad y = \frac{B}{D}, \qquad
z = \frac{C}{D}, \qquad w = \frac{E}{D}.
\end{equation}
Substituting (\ref{AB}) into (\ref{xy}) we obtain the system
\begin{equation}\label{sysxy}
\begin{aligned}
xD &= x I_2 + y,\\
yD &= z I_{-2} + w,\\
zD &= x + y I_{-2},\\
wD &= z + w I_2.
\end{aligned}
\end{equation}
Subtracting the fourth equation from the first yields
\begin{equation}\label{eqxw}
(D-I_2)(x-w)=y-z.
\end{equation}
Similarly, subtracting the third equation from the second gives
\begin{equation}\label{eqyz}
(D+I_{-2})(y-z)=-(x-w).
\end{equation}
Substituting (\ref{eqxw}) into (\ref{eqyz}), we obtain
\[
\Bigl[(D+I_{-2})(D-I_2)+1\Bigr](x-w)=0.
\]
We now show that the coefficient is strictly positive. From the first equation of
(\ref{sysxy}),
\[
x(D-I_2)=y.
\]
By Remark \ref{rkp} and (\ref{xy}) we have \(x>0\) and \(y>0\), therefore, a positive solution exists if
$
D-I_2>0.
$
Moreover,
$
D+I_{-2}>0.
$
Therefore,
\[
(D+I_{-2})(D-I_2)+1>1>0.
\]
Consequently, $x-w=0$.
Substituting this into (\ref{eqxw}) yields
$
y-z=0.
$
Hence
\[
f(t)=x\,p(t)+y\,q(t)=
g(t)=z\,q(t)+w\,p(t)
=y\,q(t)+x\,p(t).
\]
We conclude that every translation-invariant solution is necessarily symmetric.
Hence there are no genuinely asymmetric translation-invariant solutions.
\subsubsection{Case 3. Quadratic loss with a linear model}\label{sec:quadratic-nonseparable}
Consider the linear predictor $\mathcal F(x;t,u)=tx+u$,
and the squared-error loss
$\ell(y,z)=(y-z)^2.$
Then the empirical loss takes the form
\[
\mathcal L_N(t,u)
=
\frac1N\sum_{i=1}^{N}
(y_i-tx_i-u)^2.
\]
Expanding the square, we obtain
\[
(y_i-tx_i-u)^2
=
y_i^2+t^2x_i^2+u^2
-2tx_i y_i
-2u y_i
+2tux_i.
\]
Therefore
\[
\mathcal L_N(t,u)
=
At^2+Bu^2+Ctu+Dt+Eu+F,
\]
where
\[
A=\frac1N\sum_{i=1}^{N}x_i^2,
\qquad
B=1,
\qquad
C=\frac2N\sum_{i=1}^{N}x_i,
\]
\[
D=-\frac2N\sum_{i=1}^{N}x_i y_i,
\qquad
E=-\frac2N\sum_{i=1}^{N}y_i,
\qquad
F=\frac1N\sum_{i=1}^{N}y_i^2.
\]
Notice that the mixed coefficient $C$
is proportional to the empirical mean of the input data.
\noindent
{\bf Case \(C=0\).}
In this case the mixed term \(tu\) disappears and the empirical loss
separates into a sum of two one-variable functions:
\[
\mathcal L_N(t,u)
=
p(t)+q(u),
\]
where
\[
p(t)=At^2+Dt,
\qquad
q(u)=Bu^2+Eu+F.
\]
This is the Case 2 discussed above, and there is unique solution.
{\bf Case \(C\neq0\).} In this case the loss contains the interaction term \(Ctu\), and therefore does not
separate into a sum of a function of \(t\) and a function of \(u\).
The kernel of the transfer operator is
\[
K(t,u)
=
\exp\!\Bigl(
-\beta
\bigl(
At^2+Bu^2+Ctu+Dt+Eu+F
\bigr)
\Bigr).
\]
Since the exponent is continuous on the compact square
\([0,1]^2\), the kernel is strictly positive and continuous:
\[
K(t,u)>0,
\qquad
(t,u)\in[0,1]^2.
\]
Hence,
\[
(K_Nf)(t)
=
\int_0^1 K(t,u)f(u)\,du
\]
defines a compact positive integral operator on
\(C([0,1])\) and on \(L^2([0,1])\).
By the Krein--Rutman theorem, the spectral radius
$r(K_N)=\lambda_N$ is a simple positive eigenvalue, and there exists a unique
(up to multiplication by a positive constant)
strictly positive eigenfunction $\psi_N(t)>0$.
%Consequently,
%\[
%\psi_N=\log \lambda_N.
%\]
Unlike the case \(C=0\), the kernel is not separable. Indeed,
\[
K(t,u)
=
e^{-\beta(At^2+Dt+F)}
e^{-\beta(Bu^2+Eu)}
e^{-\beta Ctu},
\]
and the coupling factor \(e^{-\beta Ctu}\) prevents a factorization
into a single product of a function of \(t\) and a function of \(u\). See Section \ref{sec:exp} for complementary numerical experiments for the case $C \ne 0$.
% Using the power-series expansion
% \[
% e^{-\beta Ctu}
% =
% \sum_{n=0}^{\infty}
% \frac{(-\beta C)^n}{n!}\,
% t^n u^n,
% \]
% we obtain
% \[
% K(t,u)
% =
% e^{-\beta(At^2+Dt+F)}
% e^{-\beta(Bu^2+Eu)}
% \sum_{n=0}^{\infty}
% \frac{(-\beta C)^n}{n!}\,
% t^n u^n.
% \]
% Therefore
% \[
% K(t,u)
% =
% \sum_{n=0}^{\infty}
% \frac{(-\beta C)^n}{n!}\,
% \phi_n(t)\psi_n(u),
% \]
% where
% \[
% \phi_n(t)
% =
% t^n e^{-\beta(At^2+Dt+F)},
% \qquad
% \psi_n(u)
% =
% u^n e^{-\beta(Bu^2+Eu)}.
% \]
% Hence \(K_N\) is the norm-limit of finite-rank operators obtained by
% truncating the above series:
% \[
% K_N^{(m)}(t,u)
% =
% \sum_{n=0}^{m}
% \frac{(-\beta C)^n}{n!}\,
% \phi_n(t)\psi_n(u).
% \]
% This provides a natural approximation scheme for computing the dominant
% eigenvalue \(\lambda_N\). In practice, one may either truncate the above
% series or discretize the integral operator using quadrature methods,
% thereby reducing the problem to the computation of the largest eigenvalue
% of a finite matrix.
% {\color{red} Experiment.......}
\subsubsection{Case $k=2$: An explicit example \(\eta_{tu}=e^{\beta(t+u)}\).}\label{k2a}
In this case the translation-invariant equations for the Cayley tree of order \(2\) are
\begin{equation}\label{fexp}
f(t)=
\left(
\frac{
\displaystyle
\int_0^1 e^{\beta(t+u)}f(u)\,du
+
\int_0^1 e^{-\beta(t+u)}g(u)\,du
}
{D}
\right)^2,
\end{equation}
\begin{equation}\label{gexp}
g(t)=
\left(
\frac{
\displaystyle
\int_0^1 e^{-\beta(t+u)}f(u)\,du
+
\int_0^1 e^{\beta(t+u)}g(u)\,du
}
{D}
\right)^2,
\end{equation}
where
\[
D=
\int_0^1 e^{-\beta u}f(u)\,du
+
\int_0^1 e^{\beta u}g(u)\,du .
\]
Introduce
\begin{equation}\label{hk}
h(t)=\sqrt{f(t)},
\qquad
k(t)=\sqrt{g(t)}.
\end{equation}
Then (\ref{fexp})--(\ref{gexp}) are equivalent to
\begin{equation}\label{hexp}
h(t)=
\frac{
e^{\beta t}\displaystyle\int_0^1 e^{\beta u}h^2(u)\,du
+
e^{-\beta t}\displaystyle\int_0^1 e^{-\beta u}k^2(u)\,du
}
{D},
\end{equation}
\begin{equation}\label{kexp}
k(t)=
\frac{
e^{-\beta t}\displaystyle\int_0^1 e^{-\beta u}h^2(u)\,du
+
e^{\beta t}\displaystyle\int_0^1 e^{\beta u}k^2(u)\,du
}
{D}.
\end{equation}
Define
\[
A=\int_0^1 e^{\beta u}h^2(u)\,du,
\qquad
B=\int_0^1 e^{-\beta u}k^2(u)\,du,
\]
\[
C=\int_0^1 e^{-\beta u}h^2(u)\,du,
\qquad
E=\int_0^1 e^{\beta u}k^2(u)\,du.
\]
Then (\ref{hexp})--(\ref{kexp}) become
\begin{equation}\label{hks}
h(t)=\frac{A e^{\beta t}+B e^{-\beta t}}{D},
\qquad
k(t)=\frac{C e^{-\beta t}+E e^{\beta t}}{D}.
\end{equation}
Thus every solution belongs to the finite-dimensional space
\[
\operatorname{span}\{e^{\beta t},e^{-\beta t}\}.
\]
Substituting these expressions into the definitions of
\(A,B,C,E\) yields a closed algebraic system.
For convenience, introduce
\begin{equation}\label{Im}
I_m=\int_0^1 e^{m\beta u}\,du
=
\frac{e^{m\beta}-1}{m\beta},
\qquad m\neq0, \beta>0.
\end{equation}
Since
\[
h^2(u)
=
\frac{
A^2 e^{2\beta u}
+
2AB
+
B^2 e^{-2\beta u}
}{D^2}, \ \
k^2(u)
=
\frac{
C^2 e^{-2\beta u}
+
2CE
+
E^2 e^{2\beta u}
}{D^2},
\]
the original nonlinear integral system reduces to the finite-dimensional algebraic system
\begin{equation}\label{ABCE}\begin{array}{llll}
A=
\frac{
A^2 I_3
+
2AB I_1
+
B^2 I_{-1}
}{(C+E)^2},
\\[2mm]
B=
\frac{
C^2 I_{-3}
+
2CE I_{-1}
+
E^2 I_1
}{(C+E)^2},
\\[2mm]
C=
\frac{
A^2 I_1
+
2AB I_{-1}
+
B^2 I_{-3}
}{(C+E)^2},
\\[2mm]
E=
\frac{
C^2 I_{-1}
+
2CE I_1
+
E^2 I_3
}{(C+E)^2}.
\end{array}
\end{equation}
Therefore the problem of finding translation-invariant Gibbs measures for the kernel given by
\(\eta_{tu}=e^{\beta(t+u)}\) is reduced to solving a system of four algebraic equations in the four unknowns \(A,B,C,E\).
{\bf Case: $A=E, \, B=C $.} Assume that
\[
A=E, \, B=C \ \ {\rm then} \ \
D=C+E=A+B.
\]
Therefore the system (\ref{ABCE}) reduces to
\[
A=
\frac{
A^2 I_3
+2AB I_1
+B^2 I_{-1}
}{(A+B)^2},
\qquad
B=
\frac{
A^2 I_1
+2AB I_{-1}
+B^2 I_{-3}
}{(A+B)^2}.
\]
Let $A=uB,$ for some $u>0$. Substituting into the first equation gives
\[
B=
\frac{
u^2I_3+2uI_1+I_{-1}
}
{u(u+1)^2}.
\]
The second equation gives
\[
B=
\frac{
u^2I_1+2uI_{-1}+I_{-3}
}
{(u+1)^2}.
\]
Equating these two values of $B$ we get
\[
I_1u^3
+
(2I_{-1}-I_3)u^2
+
(I_{-3}-2I_1)u
-
I_{-1}
=0 .
\]
Now use $
I_m=\frac{t^m-1}{m\ln t}$, where $t=e^\beta
$ and \(t-1\ne 0\) to obtain the cubic equation
\begin{equation}\label{cub}
P_t(u):=3t^3u^3
-
t^2(t^3+t^2+t-6)u^2
-
(6t^3-t^2-t-1)u
-
3t^2=0
\end{equation}
\begin{lemma}
For every $t>1$, the equation (\ref{cub}) has exactly one positive real solution $u>0$.
\end{lemma}
\begin{proof}
We study the sign structure and apply Descartes' rule\footnote{Descartes rule states (see \cite{Pra}) that if the nonzero terms of a single-variable polynomial with real coefficients are ordered by descending variable exponent, then the number of positive roots of the polynomial is either equal to the number of sign changes between consecutive (nonzero) coefficients, or is less than it by an even number. A root of multiplicity $n$ is counted as $n$ roots. In particular, if the number of sign changes is zero or one, the number of positive roots equals the number of sign changes.} of signs together with basic continuity arguments.
We evaluate
\[
P_t(0)=-3t^2<0,
\qquad
\lim_{u\to+\infty}P_t(u)=+\infty
\quad (\text{since } 3t^3>0).
\]
By continuity, there exists at least one $u>0$ such that $P_t(u)=0$.
Hence the number of positive roots is at least one in for every $t>1$.
Now study sign structure of coefficients. Write
\begin{equation}\label{Pt}
P_t(u)=au^3+bu^2+cu+d,
\end{equation}
where
$$a=3t^3>0, \ \ b=-t^2(t^3+t^2+t-6), \ \ c=-(6t^3-t^2-t-1), \ \ d=-3t^2<0.$$
For $t>1$, we have
\[
6t^3-t^2-t-1>0,
\]
hence $c<0$ for all $t>1$.
Define
\[
\phi(t)=t^3+t^2+t-6.
\]
Then $\phi(t)$ is strictly increasing for $t>0$ and has a unique root
\[
t_0\approx 1.34872.
\]
Thus:
\[
b>0 \text{ for } 1<t<t_0,\qquad b<0 \text{ for } t>t_0.
\]
We examine sign variations of $(a,b,c,d)$.
Case 1: $1<t<t_0$. Then
\[
(a,b,c,d)=(+,+,-,-),
\]
which has exactly one sign change.
Case 2: $t>t_0$. Then
\[
(a,b,c,d)=(+,-,-,-),
\]
again giving exactly one sign change. Hence the number of positive roots is at most one. Since at least one exists, there is exactly one positive root for each $t>1$.
\end{proof}
Denote by $u_*$ the unique solution of (\ref{cub}). Then, the corresponding solution
of (\ref{ABCE}) is given by $S_*:=(u_*, 1, 1, u_*)B_*$ with
\[
B_*=
\frac{
u_*^2I_1+2u_*I_{-1}+I_{-3}
}
{(u_*+1)^2}.
\]
The corresponding symmetric solution of integral equation is
\begin{equation}\label{fb} f(t)=g(t)=h^2(t)=\left(\frac{u_* e^{\beta t}+ e^{-\beta t}}{1+u_*}\right)^2,
\end{equation}
We summarize the result here
\begin{rt}\label{res1} Independently on values of $\beta>0$ (i.e., of $t>1$) there exists a (symmetric) translation-invariant Gibbs measure $\mu_*$ corresponding to solution $S_*$.
\end{rt}
{\bf Case:
$A\ne E$ and $B\ne C$}. Above we only solved the system (\ref{ABCE}), for \(A=E\) and \(B=C\). To see the case when these conditions are not satisfied we set
\[
D=C+E>0,
\qquad
X=A-E,
\qquad
Y=B-C.
\]
Subtracting the fourth equation of (\ref{ABCE}) from the first yields
\[
\begin{aligned}
XD^2
&=
A^2I_3-E^2I_3
+2I_1(AB-CE)
+I_{-1}(B^2-C^2)
\\
&=
I_3(A-E)(A+E)
+2I_1\bigl[B(A-E)+E(B-C)\bigr]
\\
&\qquad
+I_{-1}(B-C)(B+C).
\end{aligned}
\]
Equivalently,
\begin{equation}\label{eq1}
\Bigl(D^2-\alpha\Bigr)X-\beta Y=0,
\end{equation}
where
\[
\alpha=I_3(A+E)+2I_1B,
\qquad
\beta=2I_1E+I_{-1}(B+C).
\]
Similarly, subtracting the third equation from the second gives
\begin{equation}\label{eq2}
\gamma X+\Bigl(D^2+\delta\Bigr)Y=0,
\end{equation}
where
\[
\gamma=I_1(A+E)+2I_{-1}B,
\qquad
\delta=I_{-3}(B+C)+2I_{-1}E.
\]
Equations \eqref{eq1}--\eqref{eq2} form a homogeneous linear system
\begin{equation}\label{Ds}
\begin{pmatrix}
D^2-\alpha & -\beta\\
\gamma & D^2+\delta
\end{pmatrix}
\binom{X}{Y}
=
\binom00 .
\end{equation}
\begin{rk} By definition, we have \(A,B,C,D,E>0\) and \(I_m>0\) for \(m=\pm1,\pm3\). Therefore, \(\gamma>0\) and \(D^2+\delta>0\). Consequently, it follows from from \eqref{eq2} that \(X=0\) if and only if \(Y=0\).
It remains to consider the case \(X\neq 0\) and \(Y\neq 0\), which is possible only if the determinant of the system \eqref{Ds},
\[
\Delta = (D^2-\alpha)(D^2+\delta)+\beta\gamma,
\]
vanishes. Below, we show that, under suitable conditions on $\beta$, the case \(\Delta=0\) is possible and solutions with $A\ne E$ and $B\ne C$ exists.
\end{rk}
Denote ${A\over B}=u>0$ and ${C\over E}=v>0$.
We assume that $uv\ne 1$, because, the case $uv=1$ is equivalent to $A=E$, $B=C$.
Substituting
$
A=uB, \, C=vE
$ into the original system yields equations for \(B\) and \(E\):
\begin{equation}\label{BE}
E=\frac{v^2I_{-1}+2vI_1+I_3}{(v+1)^2}, \ \
B=\frac{v^2I_{-3}+2vI_{-1}+I_1}{(v+1)^2}.
\end{equation}
In the first equation of (\ref{ABCE})
substituting \(A=uB\) and \(C=vE\), we obtain
\[
uB=\frac{B^2\bigl(u^2I_3+2uI_1+I_{-1}\bigr)}{E^2(v+1)^2}.
\]
Cancelling \(B>0\), we get
\begin{equation}\label{BB}
B=
u\,E^2(v+1)^2\Big/\bigl(u^2I_3+2uI_1+I_{-1}\bigr).
\end{equation}
From the third equation of (\ref{ABCE}),
substituting \(A=uB\) and \(C=vE\), we obtain
\begin{equation}\label{EE}
C=vE=\frac{B^2\bigl(u^2I_1+2uI_{-1}+I_{-3}\bigr)}{E^2(v+1)^2}.
\end{equation}
Now substituting the expression for \(E\), and $B$ given in (\ref{BE}) to (\ref{BB}) and (\ref{EE}),
we obtain system of equations for $u$ and $v$:
\begin{equation}\label{uuvv} \begin{aligned}
u &= \frac{(v^2 I_{-3} + 2v I_{-1} + I_1)(u^2 I_3 + 2u I_1 + I_{-1})}{(v^2 I_{-1} + 2v I_1 + I_3)^2}, \\
v &= \frac{(v^2 I_{-3} + 2v I_{-1} + I_1)^2 (u^2 I_1 + 2u I_{-1} + I_{-3})}{ (v^2 I_{-1} + 2v I_1 + I_3)^3}.
\end{aligned}
\end{equation}
where $I_m$, is given in (\ref{Im}) and $t>1$. Setting $S=t^{2}+t+1$, we obtain
\begin{equation*}
I_{1}=\frac{t-1}{\beta },\quad I_{-1}=\frac{t-1}{\beta t},\quad I_{3}=\frac{%
(t-1)S}{3\beta },\quad I_{-3}=\frac{(t-1)S}{3\beta t^{3}}.
\end{equation*}%
Then, by replacing $I_{\pm 1}$ and \ $I_{\pm 3}$ in (\ref{uuvv}) we get
% \begin{eqnarray*}
% u &=&\frac{(v^{2}\frac{(t-1)S}{3\beta t^{3}}+2v\frac{t-1}{\beta t}+\frac{t-1%
% }{\beta })(u^{2}\frac{(t-1)S}{3\beta }+2u\frac{t-1}{\beta }+\frac{t-1}{\beta
% t})}{(v^{2}\frac{t-1}{\beta t}+2v\frac{t-1}{\beta }+\frac{(t-1)S}{3\beta }%
% )^{2}}= \\
% &=&\frac{\left( \frac{t-1}{\beta }\right) ^{2}(v^{2}\frac{S}{3t^{3}}+2v\frac{%
% 1}{t}+1)(u^{2}\frac{S}{3}+2u+\frac{1}{t})}{\left( \frac{t-1}{\beta }\right)
% ^{2}(v^{2}\frac{1}{t}+2v+\frac{S}{3})^{2}}=\frac{(v^{2}\frac{S}{3t^{3}}+2v%
% \frac{1}{t}+1)(u^{2}\frac{S}{3}+2u+\frac{1}{t})}{(v^{2}\frac{1}{t}+2v+\frac{S%
% }{3})^{2}},
% \end{eqnarray*}
% and similarly,
% \begin{equation*}
% v=\frac{(1+\frac{2v}{t}+\frac{S}{3t^{3}}v^{2})^{2}(\frac{S}{3t^{3}}+\frac{2u%
% }{t}+u^{2})}{(\frac{S}{3}+2v+\frac{v^{2}}{t})^{3}}.
% \end{equation*}%
% That is, e
\begin{align}
u& =\frac{A(v,t)C(u,t)}{B(v,t)^{2}}, \label{eq:reducida1} \\
v& =\frac{A(v,t)^{2}D(u,t)}{B(v,t)^{3}}, \label{eq:reducida2}
\end{align}%
with
\begin{align*}
A& :=A(v,t)=1+\frac{2v}{t}+\frac{S}{3t^{3}}v^{2}, \\
B& :=B(v,t)=\frac{S}{3}+2v+\frac{v^{2}}{t}, \\
C& :=C(u,t)=\frac{1}{t}+2u+\frac{S}{3}u^{2}, \\
D& :=D(u,t)=\frac{S}{3t^{3}}+\frac{2u}{t}+u^{2}.
\end{align*}
Therefore,$\allowbreak $
\begin{equation*}
uB^{2}-AC=0,\qquad vB^{3}-A^{2}D=0.
\end{equation*}%
Since $C=\frac{1}{t}+2u+\frac{Su^{2}}{3}$ and $uB^{2}-AC=0,$ we have $%
uB^{2}-A\left( \frac{1}{t}+2u+\frac{Su^{2}}{3}\right) =0,$ so that
\begin{equation*}
\frac{SA}{3}u^{2}+\left( 2A-B^{2}\right) u+\frac{A}{t}=0.
\end{equation*}%
Let $a_{1}=\frac{SA}{3};$ $b_{1}=2A-B^{2}$ and $c_{1}=\frac{A}{t}.$ Then,%
\begin{equation*}
a_{1}u^{2}+b_{1}u+c_{1}=0.
\end{equation*}%
Similarly, since $vB^{3}-A^{2}D=0$ and $D=\frac{S}{3t^{3}}+\frac{2u}{t}+u^{2}
$ we have $vB^{3}-A^{2}\left( \frac{S}{3t^{3}}+\frac{2u}{t}+u^{2}\right) =0,$
so that
\begin{equation*}
vB^{3}-\frac{A^{2}S}{3t^{3}}-\frac{2A^{2}}{t}u-A^{2}u^{2}=0.
\end{equation*}%
Thus, $A^{2}u^{2}+\frac{2A^{2}}{t}u+\left( \frac{A^{2}S}{3t^{3}}%
-vB^{3}\right) =0$ and setting%
\begin{equation*}
a_{2}=A^{2},\qquad b_{2}=\frac{2A^{2}}{t},\qquad c_{2}=\frac{A^{2}S}{3t^{3}}%
-vB^{3},
\end{equation*}%
we have%
\begin{equation*}
a_{2}u^{2}+b_{2}u+c_{2}=0.
\end{equation*}%
We conclude that the solutions are derived from the equations
\begin{equation}
a_{1}u^{2}+b_{1}u+c_{1}=0\text{ and \ }a_{2}u^{2}+b_{2}u+c_{2}=0\text{ }
\label{sistemas}
\end{equation}%
This system can be written as
\begin{equation*}
\left(
\begin{array}{cc}
a_{1} & a_{2} \\
b_{1} & b_{2}%
\end{array}%
\right) \left(
\begin{array}{c}
u^{2} \\
u%
\end{array}%
\right) =\left(
\begin{array}{c}
-c_{1} \\
-c_{2}%
\end{array}%
\right) ,
\end{equation*}
with
\begin{align*}
a_{1}& =\frac{SA}{3}, & b_{1}& =2A-B^{2}, & c_{1}& =\frac{A}{t}, \\
a_{2}& =A^{2}, & b_{2}& =\frac{2A^{2}}{t}, & c_{2}& =\frac{A^{2}S}{3t^{3}}%
-vB^{3}.
\end{align*}
We claim that $a_{1}b_{2}-b_{1}a_{2}>0.$ Indeed,%
\begin{equation*}
a_{1}b_{2}-a_{2}b_{1}=\frac{SA}{3}\frac{2A^{2}}{t}-A^{2}(2A-B^{2})=A^{2}%
\left[ B^{2}+2A\left( \frac{S}{3t}-1\right) \right] .
\end{equation*}
Since $S=t^{2}+t+1,$ we have that $\frac{S}{3t}=\frac{t^{2}+t+1}{3t}=\frac{%
t+1+\frac{1}{t}}{3}.$ Moreover $t>1$, and hence $t+1+\frac{1}{t}>3,$ so that
$\frac{S}{3t}-1>0,$ and the claim follows.
Consequently (\ref{sistemas}) has a unique solution that can be written as
has a unique solution
\begin{equation*}
\left(
\begin{array}{c}
u^{2} \\
u%
\end{array}%
\right) =\left(
\begin{array}{cc}
\frac{b_{2}}{a_{1}b_{2}-a_{2}b_{1}} & -\frac{a_{2}}{a_{1}b_{2}-a_{2}b_{1}}
\\
-\frac{b_{1}}{a_{1}b_{2}-a_{2}b_{1}} & \frac{a_{1}}{a_{1}b_{2}-a_{2}b_{1}}%
\end{array}%
\right) \left(
\begin{array}{c}
-c_{1} \\
-c_{2}%
\end{array}%
\right) =\allowbreak \left(
\begin{array}{c}
\frac{a_{2}c_{2}-b_{2}c_{1}}{a_{1}b_{2}-a_{2}b_{1}} \\
\frac{b_{1}c_{1}-a_{1}c_{2}}{a_{1}b_{2}-a_{2}b_{1}}%
\end{array}%
\right) .
\end{equation*}%
if and only if the following consistence condition is fulfilled%
\begin{equation*}
\left( \frac{b_{1}c_{1}-a_{1}c_{2}}{a_{1}b_{2}-a_{2}b_{1}}\right) ^{2}=\frac{%
a_{2}c_{2}-b_{2}c_{1}}{a_{1}b_{2}-a_{2}b_{1}},
\end{equation*}%
or equivalently%
\begin{equation*}
(b_{1}c_{1}-a_{1}c_{2})^{2}=(a_{1}b_{2}-a_{2}b_{1})(a_{2}c_{2}-b_{2}c_{1}).
\end{equation*}
$\allowbreak $Now define
\begin{equation}
F(v,t)=(a_{2}c_{1}-a_{1}c_{2})^{2}-(a_{1}b_{2}-a_{2}b_{1})(b_{1}c_{2}-b_{2}c_{1}).
\end{equation}%
Then the compatibility condition is
\begin{equation}
F(v,t)=0.
\end{equation}
By a computer one gets the factorization of $F(v,t)$
\begin{equation}\label{factor}
F(v,t)={1\over 81 t^8}Y(v,t)Q_3(v,t)Q_8(v,t),
\end{equation}
where
\begin{eqnarray}
Y(v,t)&=&(3v^2+6tv+t^3+t^2+t)^2>0.\\
Q_3(v,t)&=&-v^3P_t(1/v)=3t^{2}v^{3}+(6t^{3}-t^{2}-t-1)v^{2}+(t^{5}+t^{4}+t^{3}-6t^{2})v-3t^{3}
\end{eqnarray}
\[
\begin{aligned}
Q_8(v,t)=\;&-243t^{5}v^{8}
+\left(27t^{10}+54t^{9}+81t^{8}+54t^{7}-1917t^{6}\right)v^{7}\\
&+\left(162t^{11}+333t^{10}+513t^{9}+54t^{8}
-5985t^{7}-54t^{6}+189t^{5}+225t^{4}-54t^{3}\right)v^{6}\\
&+\left(27t^{13}+405t^{12}+846t^{11}+1323t^{10}
-810t^{9}-9117t^{8}-1053t^{7}
+2430t^{6}+576t^{5}-216t^{4}\right)v^{5}\\
&+\left(108t^{14}+546t^{13}+1140t^{12}+1653t^{11}
-2372t^{10}-6745t^{9}-3660t^{8}
+8577t^{7}\right.\\
&\qquad\left.
-300t^{6}-150t^{5}-18t^{4}-6t^{3}
+24t^{2}-10t-2\right)v^{4}\\
&+\left(9t^{16}+144t^{15}+426t^{14}+876t^{13}
+831t^{12}-2112t^{11}-2694t^{10}\right.\\
&\qquad\left.
-3432t^{9}+11481t^{8}-948t^{7}
+204t^{6}-456t^{5}+408t^{4}
-96t^{3}-24t^{2}\right)v^{3}\\
&+\left(18t^{17}+73t^{16}+185t^{15}+357t^{14}
-12t^{13}-411t^{12}-1515t^{11}\right.\\
&\qquad\left.
+873t^{10}+4656t^{9}+3021t^{8}
-2839t^{7}+1333t^{6}
-18t^{5}-120t^{4}-12t^{3}\right)v^{2}\\
&+\left(t^{19}+5t^{18}+15t^{17}+36t^{16}
+33t^{15}+3t^{14}-111t^{13}\right.\\
&\qquad\left.
-216t^{12}+903t^{11}+785t^{10}
+2581t^{9}-2550t^{8}
+1476t^{7}-216t^{6}-72t^{5}\right)v\\
&-18t^{14}+54t^{13}+90t^{12}+144t^{11}
+486t^{10}-558t^{9}+360t^{8}
-54t^{7}-18t^{6}.
\end{aligned}
\]
Since $Y(v,t)>0$, we should have $Q_3=0$ or $Q_8=0$. The case $Q_3=0$ gives symmetric solutions which were discussed by roots of $P_t(1/v)$, $u=1/v$ defined by (\ref{Pt}). So it remains to solve equation $Q_8(v,t)=0$.
\begin{lemma} For \(t>1\), let \(N_+(t)\) denotes the number of positive roots of the
equation \(Q_8(v,t)=0\). Then, $N_+(t)\in\{1, 2\}.$
\end{lemma}
\begin{proof} Write $Q_8(v,t)$ as $Q_8(v,t)=\sum_{i=0}^8k_i(t)v^i$.
We study the signs of the coefficients \(k_i(t)\) and then apply Descartes' rule of signs. Since each \(k_i(t)\), \(i=0,\ldots,8\), is an explicit polynomial in \(t\), its zeros can be computed. In particular, the relevant critical values are $t_{i,c}$, $i=0,4,5,6,7$ where
$$t_{4,c}\approx 1.351, \ \ t_{5,c}\approx 1.613, \ \ t_{6,c}\approx 1.857, \ \ t_{7,c}\approx 2.281, \ \ t_{0,c}\approx 4.6545$$ and the following table is obtained.
\begin{table}[h!]
\centering
\begin{tabular}{|c|c|c|}
\hline
Interval in $t$ & Sign of $(k_8(t), \dots, k_0(t))$ & \(N_+(t)\) \\
\hline
$(1, t_{4,c})$ & $(-,-,-,-,-,+,+,+,+)$ & 1\\
\hline
$(t_{4,c}, t_{5,c})$ & $(-,-,-,-,+,+,+,+,+)$& 1 \\
\hline
$(t_{5,c}, t_{6,c})$ & $(-,-,-,+,+,+,+,+,+)$& 1 \\
\hline
$(t_{6,c}, t_{7,c})$ & $(-,-,+,+,+,+,+,+,+)$ & 1\\
\hline
$(t_{7,c}, t_{0,c})$ & $(-,+,+,+,+,+,+,+,+)$& 1 \\
\hline
$(t_{0,c}, +\infty)$ & $(-,+,+,+,+,+,+,+,-)$& 2 \\
\hline
\end{tabular}
\end{table}
For the case $t\in(1,t_{0,c})$, we have $
k_8(t)<0,\, k_0(t)>0.$
Therefore, $ Q_8(0,t)=:\lim_{v\to 0} Q_8(v,t)>0$ and $\lim_{v\to +\infty} Q_8(v,t)<0$.
Hence, by the intermediate value theorem, there exists at least one
positive solution $v>0$ of the equation $Q_8(v,t)=0$. Moreover, applying Descartes' rule of signs, we conclude that
$Q_8(v,t)$ has exactly one positive root (in case $t\in(1,t_{0,c})$).
For the case $t>t_{0,c}$, we have $Q_8(0,t)<0$. Moreover, $Q_8(1,t)>0$. Indeed,
the factorization of \(Q_8(1,t)\) is $Q_8(1,t)=(t-1)R_{18}(t)$,
where
\[
\begin{aligned}
R_{18}(t)=\;&t^{18}+6t^{17}+39t^{16}+157t^{15}
+519t^{14}+1395t^{13}\\
&+2775t^{12}+4614t^{11}+4695t^{10}
+3456t^9\\
&-285t^8-615t^7-1387t^6
+75t^5\\
&-99t^4+180t^3+12t^2+12t+2 .
\end{aligned}
\]
It is easy to see that $R_{18}(t)>0$, for $t>t_{0,c}$ because the
high-degree positive terms dominate the few negative coefficients.
Therefore, by the intermediate value theorem, there exists at least one positive root of
$Q_8(v,t)=0$ in the interval $(0,1)$. Moreover, by Descartes' rule of signs, the
polynomial has at most two positive roots (counting multiplicities). Since we have
already established the existence of one positive root, the possibility of having
zero positive roots is excluded. Hence, the only possible number of positive roots is
two. See Figure \ref{fig:toy-octic} for the visualization of bifurcation.
\end{proof}
Here some numerical analysis for positive roots of $Q_8(v,t)$:
\begin{itemize}
\item for $t=1.63$ one solution, namely $v\approx 2.623$,
\item for $t=2$ one solution, namely $v\approx 7.315$,
\item for critical $t=t_{0.c}$ we have {\bf two} solutions $v_1\approx 0.00000025$ and $v_2\approx 376.98$.
\item for $t=5$ there are two solutions $v\approx0.0002, v\approx524.21$
\end{itemize}
\begin{figure}[t]\centering
\includegraphics[width=.6\linewidth]{fig2_toy_octic.png}
\caption{Positive roots $v$ of the octic $Q_8(v,t)$ versus $\beta$. A second
branch appears at $\beta_c$, the mechanism behind the $2\to3$ jump.}
\label{fig:toy-octic}
\end{figure}
Consequently, we obtain the following result.
\begin{rt}\label{res2}
Let \(\beta=\ln t>0\) and $\beta_c=\ln(t_{0,c})$. The number \(n_G(\beta)\) of non-symmetric
translation-invariant Gibbs measures corresponding to the Hamiltonian
(\ref{uy1}) is given by
\[
n_G(\beta)=
\begin{cases}
1, & \beta<\beta_c,\\[2mm]
2, & \beta\geq \beta_c.
\end{cases}
\]
\end{rt}
Thus, by Results \ref{res1} and \ref{res2}, we conclude that the number of translation-invariant Gibbs measures is
equal to \(2\) for \(\beta<\beta_c\) and equal to \(3\) for \(\beta\geq \beta_c\).
For given solution $(u,v)$ with $uv\ne 1$, the non-symmetric solution of the integral equation
found by formulas (\ref{hk}) and (\ref{hks}):
\begin{equation}\label{ft}
f(t)=\left(
\frac{
v^2I_{-3}+2vI_{-1}+I_1
}
{
v^2I_{-1}+2vI_1+I_3
}
\right)^2\cdot
\left({ue^{\beta t}+e^{-\beta t}\over v+1}
\right)^2, \ \
g(t) =
\left(
{v e^{-\beta t}+e^{\beta t}\over
v+1}\right)^2.
\end{equation}
% From the last system we obtain
% \[
% \frac{u}{v}
% =
% \frac{u^2I_3+2uI_1+I_{-1}}
% {u^2I_1+2uI_{-1}+I_{-3}}
% \cdot
% \frac{v^2I_{-1}+2vI_1+I_3}
% {v^2I_{-3}+2vI_{-1}+I_1}.
% \]
% Hence the system can be written in the separated form
% \[
% \frac{u}{v}=F(u)G(v),
% \]
% where
% \[
% F(u)
% =
% \frac{u^2I_3+2uI_1+I_{-1}}
% {u^2I_1+2uI_{-1}+I_{-3}},
% \]
% and
% \[
% G(v)
% =
% \frac{v^2I_{-1}+2vI_1+I_3}
% {v^2I_{-3}+2vI_{-1}+I_1}.
% \]
% Using notation $q=e^\beta$ and
% \[
% I_3=\frac{q^3-1}{3\beta},\qquad
% I_1=\frac{q-1}{\beta},\qquad
% I_{-1}=\frac{q-1}{\beta q},\qquad
% I_{-3}=\frac{q^3-1}{3\beta q^3},
% \]
% the common factor \((q-1)/\beta\) cancels and
% \[
% F(u):=F(u,q)
% =
% \frac{\frac{q^2+q+1}{3}u^2+2u+\frac1q}
% {u^2+\frac{2}{q}u+\frac{q^2+q+1}{3q^3}},
% \]
% while
% \[
% G(v):=G(v,q)
% =
% \frac{\frac1q\,v^2+2v+\frac{q^2+q+1}{3}}
% {\frac{q^2+q+1}{3q^3}v^2+\frac{2}{q}v+1}.
% \]
% Notice that
% \[
% G(v)=F\!\left(\frac1v\right),
% \]
% so the reduced equation can also be written as
% \[
% \frac{u}{v}
% =
% F(u)\,F\!\left(\frac1v\right).
% \]
% Defining
% \[
% r(x):=r(x,q)=\frac{x}{F(x)},
% \]
% the reduced equation becomes
% \begin{equation}\label{rc}
% 1/r(u)=r\!\left(\frac1v\right).
% \end{equation}
% Using the function $F$ we get the following form of $r$
% \[
% r(x,q)
% =
% \frac{
% 3q^3 x^3 + 6q^2 x^2 + (q^2+q+1)x
% }{
% q^3(q^2+q+1)x^2 + 6q^3 x + 3q^2
% }.
% \]
% \begin{lemma} Then, $r(x,q)$ is strictly increasing in $x\in (0,+\infty )$ if, and only
% if $q\in (q_1, q_2)$, where $q_1=7-4\sqrt{3}, \ \ q_2=7+4\sqrt{3}.$
% \end{lemma}
% \begin{proof} Since $q>0,$ let $y=qx.$ Then, $r(x,q)$ is strictly
% increasing in $x$ if and only if $r(y,q)=r(qx,q)$ is strictly increasing \
% in $y\in (0,+\infty )$. This means, $\widetilde{f}(y)$ is strictly
% increasing where $s=q+\frac{1}{q}$ and
% \begin{equation*}
% \widetilde{f}(y):=\frac{3y^{3}+6y^{2}+(s+1)y}{q^{2}((s+1)y^{2}+6y+3)},\qquad
% (y>0),
% \end{equation*}%
% Since $1/q^{2}>0,\,\ $we have that $\widetilde{f}(y)$ is strictly increasing
% in $y\in (0,+\infty )$ if and only if $\xi(y)$ is it, where
% \begin{equation*}
% \xi(y)=\frac{3y^{3}+6y^{2}+(s+1)y}{(s+1)y^{2}+6y+3}.
% \end{equation*}%
% For $a=s+1,$ we have%
% \begin{equation*}
% \xi(y)=\frac{3y^{3}+6y^{2}+ay}{ay^{2}+6y+3}.
% \end{equation*}%
% Therefore,
% \begin{eqnarray*}
% \xi^{\prime }(y) &=&\frac{(9y^{2}+12y+a)(ay^{2}+6y+3)-(3y^{3}+6y^{2}+ay)(2ay+6)%
% }{(ay^{2}+6y+3)^{2}}= \\
% &=&\frac{3ay^{4}+36y^{3}+(63-a^{2})y^{2}+36y+3a}{(ay^{2}+6y+3)^{2}}.
% \end{eqnarray*}%
% Thus, $\xi^{\prime }(y)>0$ if, and only if, $P(y)>0,$ where $%
% P(y):=3ay^{4}+36y^{3}+(63-a^{2})y^{2}+36y+3a>0.$ Since $y>0,$ we have that $%
% P(y)>0$ if and only if,
% \begin{equation*}
% \frac{P(y)}{y^{2}}=3a\left( y^{2}+\frac{1}{y^{2}}\right) +36\left( y+\frac{1%
% }{y}\right) +63-a^{2}>0.
% \end{equation*}%
% Let $z=y+\frac{1}{y}.$ Since $y\in (0,+\infty ),$ it follows that $z\geq 2$
% \ Moreover, $y^{2}+\frac{1}{y^{2}}=\left( y+\frac{1}{y}\right)
% ^{2}-2=z^{2}-2,$ so that, $\frac{P(y)}{y^{2}}=h(z),$ with
% \begin{align*}
% h(z)& =3a(z^{2}-2)+36z+63-a^{2} \\
% & =3az^{2}+36z+63-a^{2}-6a.
% \end{align*}%
% Since $a=s+1$, it follows that
% \begin{equation*}
% 63-a^{2}-6a=63-(s+1)^{2}-6(s+1)=56-8s-s^{2}.
% \end{equation*}%
% and consequently,
% \begin{equation*}
% h(z)=3(s+1)z^{2}+36z+56-8s-s^{2},\qquad z\geq 2.
% \end{equation*}%
% Since $z\geq 2$ and $s+1>0$ we have
% \begin{equation*}
% h^{\prime }(z)=6(s+1)z+36>0,
% \end{equation*}%
% and hence $h$ is strictly increasing in $z\geq 2.$ Moreover, $h$ has a
% minimum in $z=2$.
% \begin{align*}
% h(2)& =3(s+1)2^{2}+36\cdot 2+56-8s-s^{2} \\
% & =12(s+1)+72+56-8s-s^{2} \\
% & =12s+12+72+56-8s-s^{2} \\
% & =-s^{2}+4s+140=-(s-14)(s+10).
% \end{align*}%
% Since $s=q+1/q>0$, we deduce that $s+10>0,$ so that $h(2)>0$ if and only if,
% $s<14$ which means that $q+\frac{1}{q}<14.$ That is,
% \begin{equation*}
% q^{2}-14q+1=(q-7-4\sqrt{3})(q-7+4\sqrt{3})<0.
% \end{equation*}
% Consequently, $h(2)>0$ if and only if $7-4\sqrt{3}<q<7+4\sqrt{3},$ and the
% result follows.
% \end{proof}
% Assume that \(q\in [q_1, q_2]\), then as above lemma says $r(u)$
% is strictly monotone on \((0,\infty)\). In this case ${\rm Im}(r)\in [0,+\infty)$. Therefore, for every constant \(c>0\), the equation
% \[
% r(u)=c
% \]
% admits a unique solution, denoted by \(u=u(c)\).
% Similarly, the equation
% \[
% r\!\left(\frac{1}{v}\right)=\frac{1}{c}
% \]
% admits a unique solution, denoted by \(v=v(c)\). Hence (\ref{rc}) is satisfied with
% \[
% u(c)=r^{-1}(c), \qquad v(c)=\frac{1}{r^{-1}(1/c)}.
% \]
% \begin{rk} Reduction to $A=E$, $B=C$ only is possible when $c=1$, because $1=u(c)v(c)=r^{-1}(c)\frac{1}{r^{-1}(1/c)}$ and therefor $r^{-1}(c)=r^{-1}(1/c)$, $c>0$ hence $c^2=1$, i.e., $c=1$.
% \end{rk}
% {\bf The complete solution} of the system is then given by the one-parameter family
% \begin{equation}\label{uvE}
% \begin{aligned}
% u(c) &= r^{-1}(c), \\
% v(c) &= \frac{1}{r^{-1}(1/c)}, {\rm with} \ \ (u(c), v(c)) \ \ {\rm satisfying} \ \ (\ref{uuvv}), \\
% E(c) &= \frac{v(c)^2I_{-1}+2v(c)I_1+I_3}{(v(c)+1)^2}, \\
% B(c) &= \frac{v(c)^2I_{-3}+2v(c)I_{-1}+I_1}{(v(c)+1)^2}., \\
% A(c) &= u(c)B(c), \\
% C(c) &= v(c)E(c).
% \end{aligned}
% \end{equation}
% This representation shows that, in the monotone regime of \(r\), the original nonlinear system reduces to a single scalar parameter \(c\), and all variables \(A,B,C,E\) are uniquely determined by \(c\).
% The non-symmetric solution corresponding to this family is found by formulas (\ref{hk}) and (\ref{hks}):
% \begin{equation}\label{ft}
% f(t):=f_c(t)
% =\left(
% \frac{
% v(c)^2I_{-3}+2v(c)I_{-1}+I_1
% }
% {
% (v(c)+1)
% \left(
% v(c)^2I_{-1}+2v(c)I_1+I_3
% \right)
% }
% \right)^2
% \left(
% u(c)e^{\beta t}+e^{-\beta t}
% \right)^2
% \end{equation}
% and
% \begin{equation}\label{gt}
% g(t):=g_c(t)
% =
% \frac{
% \bigl(
% v(c) e^{-\beta t}+e^{\beta t}
% \bigr)^2
% }
% {(v(c)+1)^2}, \ \ c>0, c\ne 1 .
% \end{equation}
% Recall that $q=e^\beta$, $\beta>0$ (see also definitions of $q_1$ and $q_2$ above).
% The results obtained in non-symmetric case can be summarized as follows.
% \begin{thm}\label{tas} If $\beta\in (0, \log(q_2)]$, then the model has a one-parametric family of (non-symmetric) translation-invariant Gibbs measure $\mu_c$ corresponding to solutions (\ref{ft}), (\ref{gt}) with conditions (\ref{uvE}).
% \end{thm}
\section{Prediction problems of Machine Learning with respect to Gibbs measures.}
In this section, for some of the translation-invariant measures $\mu$
constructed above, we calculate, for any finite volume $V_n$, with $n\geq 1$,
the probability (with respect to the measure $\mu$) of observing the system
in the state $(\sigma_n,\varphi_n)$, where $\sigma_n$ denotes the parameter
configuration and $\varphi_n$ denotes the spin configuration. Since a Cayley
tree is cofinal, it is sufficient to consider the volumes $V_n$ instead of
arbitrary finite volumes $\Lambda$; see \cite{Ge}.
By our construction this probability, for Gibbs measure $\mu$, is computed by the formula (\ref{muu}) and reduced to compute
$\mu^{(n)}(\sigma_n, \varphi_n)$ given by (\ref{uy2}).
We do this computations for the measures constructed in Subsection \ref{k2a}.
{\bf Case $\mu_*$:} Consider the translation-invariant solution (\ref{fb}) and corresponding symmetric measure $\mu_*$.
By the definition of the boundary functions,
\[
f(t,x)=\exp(h_{t,+,x}-h_{0,-,x}),
\qquad
g(t,x)=\exp(h_{t,-,x}-h_{0,-,x}),
\]
and since \(f=g\), we obtain
\[
h_{t,+,x}-h_{0,-,x}
=
h_{t,-,x}-h_{0,-,x}
=
\ln f(t).
\]
Therefore,
\[
h_{t,\pm,x}-h_{0,-,x}
=
2\ln
\left(
\frac{u_*e^{\beta t}+e^{-\beta t}}
{1+u_*}
\right).
\]
The finite-volume Gibbs distribution is given (\ref{uy2}) and for the kernel
$\eta_{tu}=e^{\beta(t+u)}$,
the interaction function is
$
\xi(t,u)=t+u,
$
and hence
the Hamiltonian
\[
H(\sigma_n,\varphi_n)
=
-\sum_{\langle x,y\rangle\in L_n}
\big(\sigma(x)+\sigma(y)\big)\varphi(x)\varphi(y).
\]
Hence (absorbing constants into \(Z_n\)):
$$
\sum_{x\in W_n} h_{\sigma(x),\varphi(x),x}
=
2\sum_{x\in W_n}
\ln\left(
\frac{u_* e^{\beta \sigma(x)}+e^{-\beta \sigma(x)}}{1+u_*}
\right).$$
Therefore the Gibbs measure becomes
$$
\mu^{(n)}_*(\sigma_n,\varphi_n) =
\frac{1}{Z_n}
\exp\Bigg[
\beta \sum_{\langle x,y\rangle\in L_n}
\big(\sigma(x)+\sigma(y)\big)\varphi(x)\varphi(y)
+
2\sum_{x\in W_n}
\ln\left(
\frac{u_* e^{\beta \sigma(x)}+e^{-\beta \sigma(x)}}{1+u_*}
\right)
\Bigg]=$$
$$
\frac{1}{Z_n}
\prod_{\langle x,y\rangle\in L_n}
\exp\Big(\beta(\sigma(x)+\sigma(y))\varphi(x)\varphi(y)\Big)
\prod_{x\in W_n}
\left(
\frac{u_* e^{\beta \sigma(x)}+e^{-\beta \sigma(x)}}{1+u_*}
\right)^2=
$$
\begin{equation}\label{m*}
\frac{1}{Z_n}
\prod_{\langle x,y\rangle\in L_n}
q^{\Big((\sigma(x)+\sigma(y))\varphi(x)\varphi(y)\Big)}
\prod_{x\in W_n}
\left(
\frac{u_* q^{\sigma(x)}+q^{-\sigma(x)}}{1+u_*}
\right)^2,
\end{equation}
where $\sigma(z)\in[0,1]$, $\varphi(z)\in\{-1,+1\}$, $q=e^\beta$.\\
{\bf Prediction rule induced by the symmetric Gibbs measure}. Now we give the rule mentioned in Section \ref{mot}, with respect to measure $\mu_*$.
For any unobserved vertex $x\in V\setminus V_N$, where
$$V_N=\{x\in V:d(x,x^0)\leq N\}$$
is the finite volume containing the observed vertices, the information associated
with the vertex \(x\) can be inferred through the conditional distribution
induced by the Gibbs measure.
Let the observed dataset be
\[
\mathcal D=
\{\varphi(x_1)=y_1,\ldots,\varphi(x_N)=y_N\},
\qquad x_i\in V_N .
\]
For an unobserved vertex
$x\in V\setminus V_N$, the prediction is defined by the conditional
expectation
$$\widehat{\varphi}(x)=
\mathbb E_{\mu_*}
\left[
\varphi(x)\mid\mathcal D
\right].
$$
Since $\varphi(x)\in\{-1,+1\},$ we have
$$
\widehat{\varphi}(x)
=
\mu_*(\varphi(x)=+1|\mathcal D)
-
\mu_*(\varphi(x)=-1|\mathcal D).
$$
Recall that the symmetric Gibbs measure (\ref{m*}).
For \(s\in\{-1,+1\}\), define the conditional partition function
\[
Z_x(s|\mathcal D)
=
\int_{\Omega}
\prod_{\langle z,w\rangle\in L_N}
q^{(\sigma(z)+\sigma(w))\varphi(z)\varphi(w)}
\prod_{z\in W_N}
\left(
\frac{u_*q^{\sigma(z)}+q^{-\sigma(z)}}{1+u_*}
\right)^2
\,d\sigma ,
\]
where the integral is taken over all configurations
\(\sigma\in[0,1]^{V_N}\) and the constraints
\[
\varphi(x)=s,
\qquad
\varphi(x_i)=y_i,\quad i=1,\ldots,N
\]
are imposed.
Then
\[
\mu_*(\varphi(x)=s|\mathcal D)
=
\frac{Z_x(s|\mathcal D)}
{Z_x(+1|\mathcal D)+Z_x(-1|\mathcal D)} .
\]
Therefore the Gibbs prediction is
$$
\widehat{\varphi}(x)
=
\frac{
Z_x(+1|\mathcal D)-Z_x(-1|\mathcal D)
}{
Z_x(+1|\mathcal D)+Z_x(-1|\mathcal D)
}=
\frac{
\displaystyle
\int_{\Omega}
\left(W_{+}(\sigma,\varphi)-W_{-}(\sigma,\varphi)\right)\,d\sigma
}{
\displaystyle
\int_{\Omega}
\left(W_{+}(\sigma,\varphi)+
W_{-}(\sigma,\varphi)\right)\,d\sigma
},$$
where
\[
W_{\pm}(\sigma,\varphi)
=
\prod_{\langle z,w\rangle\in L_N}
q^{(\sigma(z)+\sigma(w))\varphi(z)\varphi(w)}
\prod_{z\in W_N}
\left(
\frac{u_*q^{\sigma(z)}+q^{-\sigma(z)}}{1+u_*}
\right)^2.
\]
Thus, the prediction at an unobserved vertex is obtained by integrating over
all possible continuous latent configurations
\(\sigma\in[0,1]^{V_N}\) and averaging the two possible states of the discrete
variable \(\varphi(x)\) according to the Gibbs posterior distribution.
In this sense, the symmetric Gibbs measure defines a Bayesian tree-based
energy model, where the parameter \(u_*\) determines the learned equilibrium
structure and the conditional expectation above provides the probabilistic
prediction rule.\\
{\bf Non-symmetric case.}
Consider a Gibbs measure constructed by a non-symmetric solution of (\ref{uuvv}),
that is \(uv\neq 1\).
Define
\[
R(v,q)=
\frac{
\frac{v^2}{3}(1-q^{-3}) + 2v(1-q^{-1}) + (q-1)
}{
v^2(1-q^{-1}) + 2v(q-1) + \frac{1}{3}(q^3-1)
}.
\]
Then the non-symmetric boundary laws (\ref{ft}) become
\begin{equation}\label{ft-full-new}
f(t)=
R(v,q)^2
\left(
\frac{u q^{t}+q^{-t}}{v+1}
\right)^2,
\end{equation}
\begin{equation}\label{gt-full-new}
g(t)=
\left(
\frac{v q^{-t}+q^{t}}{v+1}
\right)^2.
\end{equation}
The boundary fields satisfy
\[
h_{t,+,x}=\ln f(t)+h_{0,-,x},
\qquad
h_{t,-,x}=\ln g(t)+h_{0,-,x}.
\]
Therefore,
\begin{equation}\label{boundary-sum-new}
\sum_{x\in W_n}h_{\sigma(x),\varphi(x),x}
=
\sum_{x\in W_n}
\Big(
\mathbf{1}_{\{\varphi(x)=+\}}\ln f(\sigma(x))
+
\mathbf{1}_{\{\varphi(x)=-\}}\ln g(\sigma(x))
\Big)
+ C_n,
\end{equation}
where \(C_n\) is absorbed into \(Z_n\).
We compute
\[
\ln f(t)
=
2\ln R(v,q)
+
2\ln\left(\frac{u q^{t}+q^{-t}}{v+1}\right),
\]
\[
\ln g(t)
=
2\ln\left(\frac{v q^{-t}+q^{t}}{v+1}\right).
\]
Substituting into measure gives:
$$
\mu^{(n)}(\sigma_n,\varphi_n)
=$$
$$\frac{1}{Z_n}
\prod_{\langle x,y\rangle\in L_n}
q^{(\sigma(x)+\sigma(y))\varphi(x)\varphi(y)}
\prod_{x\in W_n}\left[
\left(R(v,q)\frac{u q^{\sigma(x)}+q^{-\sigma(x)}}{v+1}\right)^{2\mathbf{1}_{\{\varphi(x)=+\}}}
\left(\frac{v q^{-\sigma(x)}+q^{\sigma(x)}}{v+1}\right)^{2\mathbf{1}_{\{\varphi(x)=-\}}}\right],
$$
where $u,v$ is a solution of (\ref{uuvv}), with $uv\ne 1$.\\
{\bf Prediction rule induced by the non-symmetric Gibbs measure.}
In this case, given data, for an unobserved vertex
$x\in V\setminus V_N$, the prediction is defined by the conditional expectation with respect to the
non-symmetric Gibbs measure \(\mu_{u,v}\):
$$\widehat{\varphi}(x)
=
\mathbb E_{\mu_{u,v}}
\left[
\varphi(x)\mid\mathcal D
\right] =
\mu_{u,v}(\varphi(x)=+1|\mathcal D)
-
\mu_{u,v}(\varphi(x)=-1|\mathcal D).
$$
For \(s\in\{-1,+1\}\), define the conditional partition function $
Z_x^{u,v}(s|\mathcal D)$ by fixing the value of the hidden variable
\(\varphi(x)=s\) and the observed values
\(\varphi(x_i)=y_i\), \(i=1,\ldots,N\):
\[
\begin{aligned}
Z_x^{u,v}(s|\mathcal D)
&=
\int_{[0,1]^{V_N}}
\prod_{\langle z,w\rangle\in L_N}
q^{(\sigma(z)+\sigma(w))\varphi(z)\varphi(w)}
\\
&\times
\prod_{z\in W_N}
\left[
\left(
R(v,q)
\frac{u q^{\sigma(z)}+q^{-\sigma(z)}}{v+1}
\right)^{2\mathbf 1_{\{\varphi(z)=+\}}}
\right.
\\
&\qquad\qquad\qquad\left.
\times
\left(
\frac{v q^{-\sigma(z)}+q^{\sigma(z)}}{v+1}
\right)^{2\mathbf 1_{\{\varphi(z)=-\}}}
\right]
\,d\sigma.
\end{aligned}
\]
Hence the conditional probability of the hidden state is
\[
\mu_{u,v}(\varphi(x)=s|\mathcal D)
=
\frac{
Z_x^{u,v}(s|\mathcal D)
}{
Z_x^{u,v}(+1|\mathcal D)
+
Z_x^{u,v}(-1|\mathcal D)
}.
\]
Therefore the prediction rule generated by the non-symmetric Gibbs measure is
\[
\widehat{\varphi}(x)
=
\frac{
Z_x^{u,v}(+1|\mathcal D)
-
Z_x^{u,v}(-1|\mathcal D)
}{
Z_x^{u,v}(+1|\mathcal D)
+
Z_x^{u,v}(-1|\mathcal D)
}.
\]
Unlike the symmetric case, the prediction depends on the pair of parameters
$(u,v)$, $uv\neq1,$ through two different boundary contributions:
\[
\left(
R(v,q)
\frac{u q^{\sigma(x)}+q^{-\sigma(x)}}{v+1}
\right)^2
\]
for the state \(\varphi(x)=+1\), and
\[
\left(
\frac{v q^{-\sigma(x)}+q^{\sigma(x)}}{v+1}
\right)^2
\]
for the state \(\varphi(x)=-1\).
Thus the learned prediction depends not only on the observed data but also on
the selected non-symmetric Gibbs state. Different solutions $(u_1,v_1)\neq(u_2,v_2)$
of the consistency equations produce different conditional distributions and
therefore different prediction rules.
\begin{rk}[Gibbs states, phase transition, and machine learning interpretation]
The obtained Gibbs measures are completely determined by the solutions of the
corresponding nonlinear consistency equations. In particular, the parameters
appearing in the boundary laws define different probability distributions on the
configuration space and therefore represent different thermodynamic states of the
system.
In the symmetric case, the solution satisfies
\[
f(t)=g(t),
\]
and the corresponding Gibbs measure $\mu_*$ is determined by a single parameter
$u_*$. This state describes a homogeneous phase in which the two possible
boundary configurations have identical statistical behavior. The resulting
probability distribution represents a balanced equilibrium state where no
distinction is made between the two local states.
In the non-symmetric case, the boundary functions become different:
\[
f(t)\neq g(t).
\]
Consequently, the corresponding Gibbs measure $\mu_{u,v}$ describes a
symmetry-broken phase. The parameters $(u,v)$ act as order parameters of the
system, and every solution of the nonlinear system (\ref{uuvv}) generates a
different Gibbs state:
\[
(u_1,v_1)\neq (u_2,v_2)
\quad\Longrightarrow\quad
\mu_{u_1,v_1}\neq \mu_{u_2,v_2}.
\]
Therefore, the existence of several symmetric and non-symmetric solutions
implies the non-uniqueness of Gibbs measures. In physical terminology, this
corresponds to coexistence of several phases. At the critical value
\[
\beta_c=\ln(t_{0,c}),
\]
the structure of the solution set changes and additional non-symmetric
solutions appear. This emergence of new Gibbs measures represents a phase
transition: below $\beta_c$ the system has a smaller set of stable states,
whereas above $\beta_c$ several competing equilibrium states coexist.
From the viewpoint of machine learning, the Gibbs measure can be interpreted
as a hierarchical energy-based probabilistic model on a tree structure. The
interaction kernel
\[
\xi_{tu}=t+u
\]
represents a similarity rule between two neighboring latent variables. The
Hamiltonian
\[
H(\sigma_n,\varphi_n)
=
-\sum_{\langle x,y\rangle\in L_n}
(\sigma(x)+\sigma(y))\varphi(x)\varphi(y)
\]
describes the propagation of information through the hierarchy, where the
field $\varphi$ acts as a modulation or gating variable controlling the
interaction between neighboring representations.
In this framework, the Gibbs measure induces a Bayesian prediction rule for
unobserved vertices. For a vertex $x\in V\setminus V_N$, the prediction of the
hidden state is given by the conditional expectation
\[
\widehat{\varphi}(x)
=
\mathbb{E}_{\mu}
\big[\varphi(x)\mid \mathcal{D}\big],
\]
where $\mathcal{D}=\{\varphi(x_i)=y_i\}_{i=1}^N$. Since
$\varphi(x)\in\{-1,+1\}$, this can be written as
\[
\widehat{\varphi}(x)
=
\mu(\varphi(x)=+1\mid\mathcal{D})
-
\mu(\varphi(x)=-1\mid\mathcal{D}).
\]
Equivalently, the prediction has the Gibbs form
\[
\widehat{\varphi}(x)
=
\frac{
Z_x(+1\mid\mathcal{D})-Z_x(-1\mid\mathcal{D})
}{
Z_x(+1\mid\mathcal{D})+Z_x(-1\mid\mathcal{D})
},
\]
where $Z_x(\pm1\mid\mathcal{D})$ are the conditional partition functions
obtained by fixing the state $\varphi(x)=\pm1$ and integrating over all
remaining latent variables.
In the symmetric regime, this prediction rule is governed by a single
equilibrium parameter $u_*$, and therefore produces a uniform and stable
posterior structure. In contrast, in the non-symmetric regime the prediction
depends on the full pair $(u,v)$ and hence on the selected Gibbs phase. The
posterior becomes multi-modal, and different solutions $(u,v)$ lead to
different prediction rules:
\[
\widehat{\varphi}_{u_1,v_1}(x)\neq \widehat{\varphi}_{u_2,v_2}(x).
\]
Thus, in machine learning terminology, the symmetric measure corresponds to a
single equilibrium representation, whereas the non-symmetric Gibbs measures
describe several competing latent representations. The choice of the solution
$(u,v)$ determines the learned phase of the model, and the prediction
$\widehat{\varphi}(x)$ is the posterior output of the corresponding
hierarchical energy-based model.
The phase transition phenomenon has a natural machine learning interpretation:
when the inverse temperature $\beta$ increases, the model becomes more
sensitive to interactions and may move from a single stable representation to
several competing representations. This mechanism provides a mathematical
framework for describing multi-modal learning, clustering phenomena, latent
state separation, and symmetry breaking in hierarchical energy-based models.
Hence, the family of Gibbs measures generated by the symmetric and
non-symmetric solutions provides a rigorous probabilistic description of how
complex learning systems can move between homogeneous representation regimes
and regimes with multiple specialized latent structures.
\end{rk}
\section{Numerical experiments for the non-separable case: $C \ne 0$}\label{sec:exp}
% The purpose of this subsection is to illustrate, on a
% concrete dataset, that the interaction kernel produced by an empirical loss exhibits the same structural phenomenon established analytically for the model kernel $\eta_{tu}=e^{\beta(t+u)}$: namely, the existence of a critical inverse temperature at which the number of translation-invariant Gibbs measures passes from two to three (and possibly even more).
% We recall the analytic situation. For the kernel $\eta_{tu}=e^{\beta(t+u)}$ on the Cayley tree of order $k=2$, the symmetric translation-invariant solutions are governed by the cubic $P_t(u)$ of \eqref{cub}, which by Lemma~3 has exactly one positive root for every $t>1$, while the non-symmetric solutions are governed by the octic $Q_8(v,t)$ of \eqref{factor}, which by Lemma~4 has one
% positive root for $1<t<t_{0,c}$ and two positive roots for $t>t_{0,c}$. By
% Results~\ref{res1} and \ref{res2}, the total number of translation-invariant Gibbs measures equals two for $\beta<\beta_c$ and three for $\beta\geq\beta_c$, where $\beta_c=\ln t_{0,c}$ and $t_{0,c}\approx4.6545$. See Figure \ref{fig:toy-octic} for the visualization of the roots of $Q_8(v, t)$.
We now replace the model kernel by a kernel generated, in the manner of
Section \ref{sec:quadratic-nonseparable}, from a dataset.
{\bf The dataset and its induced kernel.}
Let $\{(x_i,y_i)\}_{i=1}^{N}$ with $N=400$ be points $x_i\in\mathbb R^{5}$
and labels $y_i\in\{-1,+1\}$, the two classes being generated as samples of
$\mathcal N(\mu,I_5)$ and $\mathcal N(-\mu,I_5)$ in equal proportion, with
$\mu=(\tfrac{13}{10},0,0,0,0)$ (Figure \ref{fig:data}). To each point we associate a local parameter
$\sigma(x_i)\in[0,1]$, obtained by projecting $x_i$ onto the direction
$w=(\bar x_{+}-\bar x_{-})/\lVert \bar x_{+}-\bar x_{-}\rVert$, where
$\bar x_{\pm}$ are the class means, and mapping the projection affinely into
$[0,1]$. The parameter $\sigma(x)$ thus records the position of $x$ relative to
the separating hyperplane.
With the affine pair-predictor $\mathcal F(x_i;t,u)=\sigma(x_i)\,t+u$ and the
squared loss $\ell(y,\hat y)=(y-\hat y)^2$, the empirical interaction loss is,
as in Case~3,
\[
\mathcal L_N(t,u)
=\frac1N\sum_{i=1}^{N}\bigl(y_i-\sigma(x_i)\,t-u\bigr)^2
=At^2+Bu^2+Ctu+Dt+Eu+F,
\]
with coefficients
\[
A=\frac1N\sum_i\sigma_i^2,\quad
B=1,\quad
C=\frac2N\sum_i\sigma_i,\quad
D=-\frac2N\sum_i\sigma_i y_i,\quad
E=-\frac2N\sum_i y_i,\quad
F=\frac1N\sum_i y_i^2 .
\]
For the present dataset these take the values
$A\approx0.31$, $B=1$, $C\approx1.01$, $D\approx-0.39$, $E\approx0$,
$F=1$. Since $C\neq0$, the kernel $\eta_{tu}=e^{\beta\mathcal L_N(t,u)}$ does
not factorize as in Case~2, and one is in the non-separable situation of
Case~3. Moreover $\mathcal L_N(t,u)\geq\min_{[0,1]^2}\mathcal L_N\approx0.88>0$,
so that the positivity hypothesis $\mathcal L_N(t,u)>0$ of Theorem~1 is
satisfied, and the uniqueness assertion of that theorem applies on the Cayley
tree of order $k=1$. See Figures \ref{fig:data-loss} and \ref{fig:data-kernel} for the visualizations of the loss function and the corresponding data kernel.
{\bf The translation-invariant equations for $k=2$.}
For the tree of order $k=2$ we consider the translation-invariant system
\eqref{fexp}--\eqref{gexp} with the data-induced kernel
$\eta_{tu}=e^{\beta\mathcal L_N(t,u)}$, namely
\[
f(t)=\left(\frac{\int_0^1\eta_{tu}f(u)\,du+\int_0^1\eta_{tu}^{-1}g(u)\,du}{D}\right)^2,
\quad
g(t)=\left(\frac{\int_0^1\eta_{tu}^{-1}f(u)\,du+\int_0^1\eta_{tu}g(u)\,du}{D}\right)^2,
\]
with $D=\int_0^1\eta_{0u}^{-1}f(u)\,du+\int_0^1\eta_{0u}g(u)\,du$. Each of these
equations contains integrals of the unknown functions over $[0,1]$, and so
cannot be treated as a finite system as it stands. We therefore replace every
integral by a quadrature rule, which is the only point at which an
approximation enters.
\begin{figure}[t]\centering
\includegraphics[width=.9\linewidth]{fig6_data_embedding.png}
\caption{Synthetic two-Gaussian data (left, first two principal components) and
the one-dimensional embedding $\sigma(x)$ with class-conditional histograms
(right).}
\label{fig:data}
\end{figure}
\begin{figure}[t]\centering
\includegraphics[width=.6\linewidth]{fig3_data_loss.png}
\caption{Data-induced empirical loss $\mathcal L_N(t,u)$. The non-zero cross term
($C=1.01$) makes the kernel non-separable (Case~3); the surface is strictly
positive ($\min=0.88$).}
\label{fig:data-loss}
\end{figure}
\begin{figure}[t]\centering
\includegraphics[width=.95\linewidth]{fig4_data_kernel.png}
\caption{Data-induced kernel $\eta_{tu}=e^{\beta\mathcal L_N(t,u)}$ at increasing
$\beta$. As $\beta$ grows the kernel stiffens and concentrates near a corner of
$[0,1]^2$, which is the source of the high-$\beta$ numerical difficulty.}
\label{fig:data-kernel}
\end{figure}
{\bf Discretization of the integrals.}
A quadrature rule replaces an integral by a finite weighted sum,
$\int_0^1\psi(u)\,du\approx\sum_{j=1}^{m}w_j\,\psi(u_j)$, sampled at prescribed
nodes $u_1,\dots,u_m\in[0,1]$ with weights $w_1,\dots,w_m>0$. We use the
Gauss--Legendre rule, in which the nodes are the (rescaled) roots of the
Legendre polynomial of degree $m$ and the weights are chosen so that the
approximation is exact for every polynomial of degree at most $2m-1$. A function
that is smooth and well approximated by polynomials on $[0,1]$ is thus
integrated to high accuracy already for moderate $m$; the nodes cluster towards
the endpoints $0$ and $1$, where they are most needed. We say that the rule
\emph{resolves} the integrand when the $m$ sampled values $\psi(u_1),\dots,
\psi(u_m)$ capture its variation faithfully, so that the weighted sum is close
to the true integral. This is the case precisely when $\psi$ does not vary
appreciably on the scale of the gaps between adjacent nodes.
Writing $f_i=f(u_i)$, $g_i=g(u_i)$ and
$\eta_{ij}=e^{\beta\mathcal L_N(u_i,u_j)}$, the system above becomes the finite
nonlinear system
\begin{equation}\label{exp:disc}
f_i=\left(\frac{\sum_{j}w_j\eta_{ij}f_j+\sum_{j}w_j\eta_{ij}^{-1}g_j}{D}\right)^{2},
\quad
g_i=\left(\frac{\sum_{j}w_j\eta_{ij}^{-1}f_j+\sum_{j}w_j\eta_{ij}g_j}{D}\right)^{2},
\qquad i=1,\dots,m,
\end{equation}
with $D=\sum_{j}w_j\eta_{0j}^{-1}f_j+\sum_{j}w_j\eta_{0j}g_j$ and
$\eta_{0j}=e^{\beta\mathcal L_N(0,u_j)}$. This is a system of $2m$ equations for
the $2m$ unknowns $(f_i,g_i)$, equivalent to the integral equations in the limit
$m\to\infty$.
{\bf Iterative solution.}
The right-hand side of \eqref{exp:disc} defines a map
$(f,g)\mapsto\Phi_\beta(f,g)$ on positive vectors, whose fixed points are
exactly the solutions sought; the boundary law $(f,g)$ being determined only up
to a common positive factor, we normalize after each application so that the
mean of the entries is fixed. Starting from an initial pair $(f^{(0)},g^{(0)})$,
we form the damped iteration
\[
(f^{(\ell+1)},g^{(\ell+1)})
=(1-\theta)\,(f^{(\ell)},g^{(\ell)})
+\theta\,\Phi_\beta(f^{(\ell)},g^{(\ell)}),
\qquad \theta=\tfrac12,
\]
and continue until successive iterates differ by less than a fixed tolerance. To locate all solutions rather than a single one, the
iteration is run from several initial pairs: the symmetric choice
$f^{(0)}=g^{(0)}\equiv1$; the two polarized choices
$f^{(0)}(t)=e^{\beta t},\,g^{(0)}(t)=e^{-\beta t}$ and its interchange, which
favour the two symmetry-broken branches; and a number of random positive
initializations. Each run that converges yields a fixed point of
\eqref{exp:disc}, hence an approximate boundary law $(f,g)$.
A solution is counted once, two solutions being regarded as identical when they
coincide after multiplication by a positive constant and after the interchange
$(f,g)\mapsto(g,f)$ induced by the symmetry $\varphi\mapsto-\varphi$ of the
Hamiltonian \eqref{uy1}. To exclude artefacts of the discretization, only those
solutions are retained whose count is unchanged when the number $m$ of
quadrature nodes is increased.
\begin{figure}[t]\centering
\includegraphics[width=.9\linewidth]{data_transition.png}
\caption{Data-induced kernel: the grid-stable number of boundary-law solutions
reproduces the $2\to3$ transition at $\beta_c^{\mathrm{data}}\approx2.74$. There is a small region at the end of the green-shaded interval where the number of solutions transition from 3 to 4 after which (the pink-shaded interval) the numerical results for different grids disagree, rendering them inconclusive.}
\label{fig:data-transition}
\end{figure}
{\bf The observed transition.}
Let $n_G(\beta)$ denote the number of translation-invariant solutions obtained
in this way. One finds a critical value $\beta_c^{\mathrm{data}}$, with
$\beta_c^{\mathrm{data}}\approx2.74$ for the dataset above, such that
\[
n_G(\beta)=
\begin{cases}
2, & \beta<\beta_c^{\mathrm{data}},\\[1mm]
3, & \beta\geq\beta_c^{\mathrm{data}}.
\end{cases}
\]
At the transition the unique symmetric solution $f=g$ persists, and a pair of
non-symmetric solutions $f\neq g$, interchanged by $\varphi\mapsto-\varphi$,
appears. This is precisely the structure described analytically in
Subsection~\ref{k2a}: the dataset breaks the symmetry between the two values of
$\varphi$. The numerical value of $\beta_c^{\mathrm{data}}$ differs from the
analytic $\beta_c=\ln t_{0,c}\approx1.538$ because the kernel is different; what
is common to the two situations, and is the point of the illustration, is the
existence of a finite critical inverse temperature together with the passage of
$n_G$ from two to three. See Figure \ref{fig:data-transition} for the visualization of a more sophisticated phase transitions which could not be captured by the analytical theory.
% Two further observations are consistent with the analytic theory. On the tree
% of order $k=1$, the difference $r(K_c)-r(K_s)$ of spectral radii, which by the
% argument preceding Theorem~1 governs the uniqueness of the Gibbs measure,
% remains strictly positive and decreases as $\beta$ increases; this reflects the
% mechanism by which the antisymmetric component is suppressed and uniqueness is
% enforced, the suppression weakening as $\beta$ grows. For large $\beta$ the
% kernel $\eta_{tu}=e^{\beta\mathcal L_N(t,u)}$ is dominated by the region where
% $\mathcal L_N$ is largest and decays steeply away from it, so that its mass
% concentrates in a small neighbourhood of a corner of $[0,1]^2$. The boundary
% laws $f$ and $g$ inherit this steep variation, becoming sharply peaked functions
% that change substantially over distances smaller than the spacing between
% adjacent quadrature nodes. The quadrature rule then no longer resolves them: the
% sampled values $f(u_1),\dots,f(u_m)$ miss the peak, the weighted sums in
% \eqref{exp:disc} cease to approximate the true integrals, and the computed count
% $n_G(\beta)$ changes with the number of nodes $m$. Once $n_G(\beta)$ fails to be
% stable under refinement of the grid, no conclusion is drawn in that range; it is
% shaded and excluded in Figure~\ref{fig:data-transition}.
{\bf Interpretation.} In agreement with the discussion of Sections~2 and 5, the number $n_G(\beta)$ of Gibbs measures plays the role of an order parameter of the learning system. For $\beta<\beta_c^{\mathrm{data}}$ the dataset determines a single equilibrium learning state and, through the conditional expectation of Section~5, a single prediction rule. For $\beta\geq\beta_c^{\mathrm{data}}$ several equilibrium states coexist, the symmetry-broken ones corresponding to the two label polarities, and the prediction rule becomes multi-valued. In keeping with Remark~3, the Cayley tree is used here as a hierarchical inference structure on which the consistency equations \eqref{uy5} are exact, and not as a description of the geometry of the data.
% ---------------------------------------------------------------------
% (C) The algorithm, phrased as a mathematical procedure rather than
% as software pseudocode. Replaces the second red placeholder.
% ---------------------------------------------------------------------
We conclude by recording, as a procedure, the prediction scheme determined by
the construction of this paper.
\begin{itemize}
\item[(1)] Given a dataset $\{(x_i,y_i)\}_{i=1}^{N}$, a predictor
$\mathcal F$, a loss $\ell$, and an inverse temperature $\beta>0$, assign to
each $x_i$ a local parameter $\sigma(x_i)\in[0,1]$ and form the empirical
interaction loss
$\mathcal L_N(t,u)=\frac1N\sum_{i}\ell\bigl(y_i,\mathcal F(x_i;t,u)\bigr)$
together with the kernel $\eta_{tu}=e^{\beta\mathcal L_N(t,u)}$.
\item[(2)] On the Cayley tree of order $k$, determine the translation-invariant
boundary laws $(f,g)$ solving the system \eqref{uy5} for this kernel. By
Proposition~1 each such pair defines a translation-invariant Gibbs measure
$\mu$; for $k=1$ this measure is unique by Theorem~1 whenever
$\mathcal L_N>0$, while for $k=2$ several measures may occur, as in
Subsection~\ref{k2a}.
\item[(3)] For each Gibbs measure $\mu$ so obtained and each unobserved vertex
$x\in V\setminus V_N$, compute the conditional partition functions
$Z_x(\pm1\mid\mathcal D)$ of Section~5 and set
\[
\widehat\varphi(x)
=\mathbb E_{\mu}\bigl[\varphi(x)\mid\mathcal D\bigr]
=\frac{Z_x(+1\mid\mathcal D)-Z_x(-1\mid\mathcal D)}
{Z_x(+1\mid\mathcal D)+Z_x(-1\mid\mathcal D)},
\qquad
\mathcal D=\{\varphi(x_i)=y_i\}_{i=1}^{N}.
\]
The predicted label is $\operatorname{sign}\widehat\varphi(x)$.
\item[(4)] When the parameters $(\beta,k)$ are such that the Gibbs measure is
unique, the rule of step~(3) is single-valued. When several Gibbs measures
coexist, each yields its own prediction rule
$\widehat\varphi(x)$; these may be reported together as competing labelings, or
one of them selected. The number of Gibbs measures is itself an order parameter
of the learning system, equal to one effective prediction rule below the
critical inverse temperature and to several above it.
\end{itemize}
\section*{Data availability statements}
The datasets generated during and/or analysed during the
current study are available from the corresponding author (U.A.Rozikov) on reasonable request.
\section*{Conflicts of interest} The authors declare no conflicts of interest.
\section*{ Acknowledgements}
The authors gratefully acknowledge the University of Granada for awarding the Visiting Scholar Grant (PPVS2024.04) to U.A. Rozikov.
He also expresses his sincere gratitude to the DaSCI Institute and IMAG for their kind invitation and support during his academic visit.
This publication is part of the project ``Ethical, Responsible, and General Purpose Artificial Intelligence: Applications
In Risk Scenarios" (IAFER) Exp.:TSI-100927-2023-1 funded through the creation of university-industry research
programs (ENIA Programs), aimed at the research and development of artificial intelligence, for its dissemination and
education within the framework of the Recovery, Transformation and Resilience Plan of the European Union Next
Generation EU through the Ministry of Digital Transformation and the Civil Service. This work was also partially
supported by Knowledge Generation Projects, funded by the Spanish Ministry of Science, Innovation, and Universities
of Spain under the project PID2023-150070NB-I00. \\
\begin{thebibliography}{99}
\bibitem{AR} L.U. Abdullaev, U.A. Rozikov, \textit{Gibbs measures in machine learning}. World Sci. Publ. Singapore. 2026, 380 pp.
\bibitem{Ari} S. Ariosto, \textit{Statistical Physics of Deep Neural Networks: Generalization Capability, Beyond the Infinite Width, and Feature Learning}, 2025, 10.48550/arXiv.2501.19281.
\bibitem{Bah} Y. Bahri, J. Kadmon, J. Pennington, S. Schoenholz, J. Sohl-Dickstein, S. Ganguli, \textit{Statistical Mechanics of Deep Learning}, Annu. Rev. Condens. Matter Phys. \textbf{11}, 2020, 501-28.
\bibitem{Beh} F. Behrens, N. Mainali, C. Marullo, S. Lee, B. Sorscher, H. Sompolinsky,
\textit{Statistical mechanics of deep learning}, J. Stat. Mech. (2024) 104007.
% \bibitem{Bis} C.M. Bishop, \textit{Pattern Recognition and Machine Learning}. Springer. 2006.
% \bibitem{BR} L.V. Bogachev, U.A. Rozikov, \textit{On the uniqueness of Gibbs measure in
% the Potts model on a Cayley tree with external field}. J. Stat. Mech. Theory Exp. \textbf{7} (2019), 073205, 76 pp.
% \bibitem{Bu} A. Buades, B. Coll, and J-M Morel. \textit{A non-local algorithm for image
% denoising}. In Computer Vision and Pattern Recognition, 2005. CVPR 2005. IEEE Computer Society Conference on, volume 2, pages 60-65. IEEE, 2005.
% \bibitem{Ch} V. Chandola, A. Banerjee, V. Kumar, \textit{Anomaly detection: A survey}. ACM Computing Surveys (CSUR), \textbf{41}(3), 2009, 1-58.
\bibitem{Du} Y. Du, Krein–Rutman Theorem and the Principal Eigenvalue. Order structure and topological methods in nonlinear partial differential equations. Vol. 1. Maximum principles and applications. Series in Partial Differential Equations and Applications. Hackensack, NJ: World Scientific, 2006.
\bibitem{EE} A.C.D. van Enter, V.N. Ermolaev, G. Iacobelli, C. K\"ulske, \textit{ Gibbs-non-Gibbs properties for evolving Ising models on trees}. Annales de l'I.H.P. Probabilit\'es et statistiques, \textbf{48}(3), 2012, 774-791.
\bibitem{FV} S. Friedli, Y. Velenik, \textit{Statistical mechanics of lattice systems. A concrete mathematical introduction}. Cambridge University Press, Cambridge, 2018.
% \bibitem{Gem} S.Geman, D. Geman, \textit{Stochastic relaxation, Gibbs distributions, and the Bayesian restoration of images}. IEEE Transactions on Pattern Analysis and Machine Intelligence, \textbf{6}(6), (1984), 721-741.
\bibitem{Ge} H.O. Georgii, \textit{Gibbs Measures and Phase Transitions}, Second edition. de Gruyter Studies in Mathematics, 9. Walter de Gruyter, Berlin, 2011.
% \bibitem{Gh} Z.Ghahramani, M.I. Jordan, \textit{Factorial Hidden Markov Models}. Machine Learning. \textbf{29}, 1997, 245-273.
% \bibitem{HOR} F.H. Haydarov, B.A. Omirov, U.A. Rozikov, \textit{Coupled Ising-Potts Model: Rich Sets of Critical Temperatures and Translation-Invariant Gibbs Measures}, arXiv:2502.12014 [math.FA]
\bibitem{HRV} F. Herrera, U.A. Rozikov, M.V. Velasco, \textit{Ising Models with Hidden Markov Structure: Applications to Probabilistic Inference in Machine Learning}, Jour. Stat. Mech.: Theory and Exper., 2025. 073201, 21 pages.
% \bibitem{Ki} R. Kindermann, J.L. Snell, \textit{Markov Random Fields and Their Applications}. American Mathematical Society. (1980).
% \bibitem{Krz} F. Krzakala, M. M\'ezard, F. Sausset, Y. Sun, L. Zdeborov\'a, \textit{Statistical-physics-based reconstruction in compressed sensing}. Phys. Rev. X \textbf{2}(2), 2012, 021005.
\bibitem{Krza} F. Krzakala, L. Zdeborov\'a, \textit{Statistical physics methods in optimization and machine learning}, https://sphinxteam.github.io/EPFLDoctoralLecture2021/Notes.pdf
% \bibitem{Ni} H. Nishimori, \textit{Statistical physics of spin glasses and information processing: an introduction}, Oxford University Press, 2001.
\bibitem{Le} Y. LeCun, S. Chopra, R. Hadsell, M. Ranzato, F. Huang, \textit{A tutorial on energy-based learning}. Predicting structured data MIT Press. 2006.
% \bibitem{Pe} J. Pearl, \textit{Probabilistic reasoning in intelligent systems: Networks of plausible inference}. Morgan Kaufmann. 1988.
\bibitem{Pra} V.V. Prasolov, {\it Polynomials} (Springer-Verlag Berlin
Heidelberg, 2004)
% \bibitem{Qa} A. Qawasmeh, F. Mukhamedov, H. Akın, \textit{Analysis of Gibbs Measures and Stability of
% Dynamical System Linked to $(1,1/2)$-Mixed Ising Model on
% $(m,k)$-Ary Trees}. Math. Phys. Anal. Geom. \textbf{28}, 10, 2025.
% \bibitem{QaM} A. Qawasmeh, F. Mukhamedov, O. Khakimov, \textit{On Gibbs measures of the Ising model on
% $(k,m)$-ary trees}, Reviews in Mathematical Physics,
% \textbf{37}(05), 2450042, 2025.
\bibitem{Ro} U.A. Rozikov, \textit{Gibbs measures on Cayley trees}. World Sci. Publ. Singapore. 2013.
\bibitem{Rbp} U.A. Rozikov, \textit{Gibbs measures in biology and physics: The Potts model}. World Sci. Publ. Singapore. 2023.
% \bibitem{Wa} M.J. Wainwright, M. I. Jordan, \textit{ Graphical models, exponential families, and variational inference.} Foundations and Trends in Machine Learning, \textbf{1}(1-2), 2008, 1-305.
% \bibitem{We} Y. Weiss, W T. Freeman, \textit{Correctness of Belief Propagation in Gaussian Graphical Models of Arbitrary Topology}, Advances in Neural Information Processing Systems 12 (NIPS 1999), 673-679.
% \bibitem{Ye} J.S. Yedidia, W. T. Freeman, Y. Weiss, \textit{Understanding belief propagation and its generalizations}. Exploring Artificial Intelligence in the New Millennium, 2005, 239-236.
% \bibitem{Zh} Z.-H. Zhou, \textit{A brief introduction to weakly supervised learning}, National Science Review, Volume 5, Issue 1, January 2018, Pages 44-53.
\end{thebibliography}
% =====================================================================
% Publication-ready experimental section.
% Drop-in replacement for the two red placeholders:
% {\color{red} Experiment.......} (Case 3)
% {\color{red} ADD an algorithm ...} (end of Section 5)
% Requires: \usepackage{amsmath, algorithm, algorithmic, graphicx}
% =====================================================================
% \subsection{Numerical experiments}\label{sec:exp}
% We complement the analytic results of Section~4 with two numerical studies.
% The first \emph{verifies} the closed-form theory for the explicit kernel
% $\eta_{tu}=e^{\beta(t+u)}$. The second \emph{instantiates} the framework on a
% genuine machine-learning task, in which the interaction kernel is generated by a
% dataset rather than prescribed, and shows that the same qualitative phenomenon
% --- a critical inverse temperature at which the number of translation-invariant
% Gibbs measures jumps from two to three --- persists.
% \subsubsection{Verification of the analytic theory}
% For the kernel $\eta_{tu}=e^{\beta(t+u)}$ on the Cayley tree of order $k=2$, the
% translation-invariant boundary laws were reduced in Section~4 to the symmetric
% cubic $P_t(u)$ (eq.~\eqref{cub}) and the non-symmetric octic $Q_8(v,t)$
% (eq.~\eqref{factor}). We computed the positive roots of these polynomials
% numerically and, independently, solved the raw algebraic system~\eqref{ABCE} by a
% multi-start nonlinear solver that uses no knowledge of the factorization. Both
% routes confirm:
% \begin{itemize}
% \item the critical value $t_{0,c}=4.65420$ (hence $\beta_c=\ln t_{0,c}=1.53777$),
% in agreement with the value $t_{0,c}\approx 4.6545$ obtained analytically;
% \item the symmetric cubic $P_t(u)$ has exactly one positive root for every
% $t>1$;
% \item the octic $Q_8(v,t)$ has one positive root for $1<t<t_{0,c}$ and two for
% $t>t_{0,c}$, with the explicitly reported roots reproduced
% ($t=1.63\mapsto v=2.625$, $t=2\mapsto v=7.314$,
% $t=5\mapsto v\in\{1.99\times10^{-4},\,524.2\}$);
% \item consequently the total number of translation-invariant Gibbs measures
% equals two for $\beta<\beta_c$ and three for $\beta\ge\beta_c$ (Results~1--2).
% \end{itemize}
% Figure~\ref{fig:toy-count} shows the measure count, and
% Figure~\ref{fig:toy-octic} the bifurcation of the non-symmetric branch at
% $\beta_c$.
% \subsubsection{A data-induced kernel: semi-supervised classification}
% \paragraph{Task and data.}
% We consider binary classification. The dataset consists of $N=400$ points in
% $\mathbb R^{5}$ drawn from two isotropic Gaussian clusters,
% $x_i\sim\mathcal N(\pm\mu,\,I_5)$ with $\mu=(\mathrm{sep},0,\dots,0)$ and
% $\mathrm{sep}=1.3$, carrying labels $y_i\in\{-1,+1\}$ (balanced). The clusters
% overlap substantially, so the problem is non-trivial (Figure~\ref{fig:data}).
% \paragraph{Embedding.}
% Following the model of Section~2, each data point is identified with a vertex of
% the tree and assigned a local parameter $\sigma(x_i)\in[0,1]$. We obtain
% $\sigma$ by projecting onto the one-dimensional Fisher (LDA) direction
% $w=(\bar x_{+}-\bar x_{-})/\lVert\bar x_{+}-\bar x_{-}\rVert$ and rescaling the
% projection $\langle x_i-\bar x,\,w\rangle$ to $[0,1]$ by a robust (1st--99th
% percentile) min--max map. Thus $\sigma(x)$ is a scalar summary of the point's
% position relative to the class boundary.
% \paragraph{Data-induced interaction.}
% We take the affine pair-predictor
% $\mathcal F(x_i;t,u)=\sigma(x_i)\,t+u$ and the squared loss
% $\ell(y,\hat y)=(y-\hat y)^2$. The empirical interaction loss of Section~2 then
% becomes the quadratic of Case~3,
% \begin{equation}\label{exp:loss}
% \mathcal L_N(t,u)
% =\frac1N\sum_{i=1}^N\bigl(y_i-\sigma(x_i)\,t-u\bigr)^2
% =At^2+Bu^2+Ctu+Dt+Eu+F,
% \end{equation}
% with data-determined coefficients
% \[
% A=\tfrac1N\textstyle\sum_i\sigma_i^2,\quad B=1,\quad
% C=\tfrac2N\textstyle\sum_i\sigma_i,\quad
% D=-\tfrac2N\textstyle\sum_i\sigma_iy_i,\quad
% E=-\tfrac2N\textstyle\sum_iy_i,\quad
% F=\tfrac1N\textstyle\sum_iy_i^2 .
% \]
% For our data $A=0.31$, $B=1$, $C=1.01$, $D=-0.39$, $E\approx0$, $F=1$. The cross
% coefficient $C=\tfrac2N\sum_i\sigma_i>0$ is strictly positive, so the kernel
% $\eta_{tu}=e^{\beta\mathcal L_N(t,u)}$ does \emph{not} factor as in Case~2: this
% is the genuinely non-separable Case~3 (Figure~\ref{fig:data-loss}). Moreover
% $\min_{[0,1]^2}\mathcal L_N=0.88>0$, so the positivity hypothesis of Theorem~1
% holds and the corresponding $k=1$ uniqueness theory applies.
% \paragraph{Solving the tree equations.}
% For $k=2$ we solve the translation-invariant boundary-law
% system~\eqref{fexp}--\eqref{gexp} with this data-induced kernel. The integrals
% are discretized by Gauss--Legendre quadrature on $[0,1]$, and fixed points are
% obtained by damped iteration from several initializations (symmetric, polarized,
% and random). Two distinct boundary laws $(f,g)$ are identified as the same Gibbs
% measure when they agree up to positive rescaling and the label flip
% $(f,g)\sim(g,f)$; to make the count independent of the discretization we
% represent each solution by the first three moments of $f$ and $g$ and retain only
% counts that are \emph{invariant under grid refinement}
% ($n_{\mathrm{grid}}\in\{32,48,72,96\}$).
% \paragraph{Result.}
% The grid-stable number of translation-invariant Gibbs measures equals two up to a
% data-determined critical value $\beta_c^{\mathrm{data}}\approx2.8$ and three above
% it (Figure~\ref{fig:data-transition}). At the transition one symmetric measure
% ($f=g$) persists and a pair of mutually mirror non-symmetric measures
% ($f\ne g$, related by $\varphi\mapsto-\varphi$) appears, exactly as in the
% analytic Case of Section~4: the dataset spontaneously breaks the
% label symmetry. The location of the transition differs from the analytic
% $\beta_c=1.54$ because the kernel is different; what transfers is the structural
% conclusion, namely the existence of a finite critical temperature with a
% $2\to3$ jump in the measure count.
% Two consistency checks accompany this result. First, the $k=1$ spectral gap
% $r(K_c)-r(K_s)$ governing uniqueness in Theorem~1 is positive throughout and
% decreases monotonically with $\beta$ (from $0.46$ at $\beta=0.5$ to
% $\sim10^{-4}$ at $\beta=4$), in agreement with the analytic mechanism whereby
% symmetry breaking is suppressed at high temperature and becomes possible only as
% $\beta$ grows. Second, for $\beta\gtrsim2.9$ the squared fixed-point map becomes
% numerically stiff --- the kernel concentrates almost all its mass near a corner
% of $[0,1]^2$ (Figure~\ref{fig:data-kernel}) --- and the solution count ceases to
% be grid-stable; this regime is shaded and excluded in
% Figure~\ref{fig:data-transition}, and no conclusions are drawn from it.
% \paragraph{Interpretation.}
% The number of Gibbs measures itself plays the role of an order parameter for the
% learning system. Below $\beta_c^{\mathrm{data}}$ the dataset determines a unique
% equilibrium learning state; above it, several coexist. The prediction rule of
% Section~5 is correspondingly single-valued in the unique-measure phase and
% multi-valued in the coexistence phase, where the symmetry-broken states encode
% the two label polarities. We emphasize that the Cayley tree here serves as a
% hierarchical \emph{inference scaffold} --- the recursive consistency equations
% are exact belief-propagation relations --- rather than as a model of the
% geometry of the data.
% --------------------------------------------------------------------
% \begin{algorithm}[t]
% \caption{Gibbs--tree classifier}
% \label{alg:gibbs-tree}
% \begin{algorithmic}[1]
% \STATE \textbf{Input:} dataset $\{(x_i,y_i)\}_{i=1}^N$, labeled subset
% $\mathcal D$, inverse temperature $\beta$, predictor $\mathcal F$, loss $\ell$,
% tree order $k$.
% \STATE Compute embeddings $\sigma(x_i)\in[0,1]$ (e.g.\ rescaled LDA/PCA score).
% \STATE Form the data-induced loss
% $\mathcal L_N(t,u)=\frac1N\sum_i\ell\bigl(y_i,\mathcal F(x_i;t,u)\bigr)$ and the
% kernel $\eta_{tu}=e^{\beta\mathcal L_N(t,u)}$ on a quadrature grid.
% \STATE Solve the boundary-law system~\eqref{uy5} (here $k=2$) for \emph{all}
% admissible $(f,g)$ from several initializations; keep grid-stable solutions
% (symmetric and non-symmetric).
% \STATE \textbf{for} each Gibbs state $(f,g)$ and each unlabeled vertex $x$
% \textbf{do}
% \STATE \quad $\displaystyle
% \widehat\varphi(x)=
% \frac{Z_x(+1\mid\mathcal D)-Z_x(-1\mid\mathcal D)}
% {Z_x(+1\mid\mathcal D)+Z_x(-1\mid\mathcal D)}$
% \quad(conditional partition functions of Section~5).
% \STATE \textbf{Output:} $\operatorname{sign}\widehat\varphi(x)$. If several Gibbs
% states exist ($\beta\ge\beta_c$), return the competing labelings, or select one
% by held-out likelihood. The number of states is an order parameter: a single
% consensus rule below $\beta_c$, several competing rules above it.
% \end{algorithmic}
% \end{algorithm}
% --------------------------------------------------------------------
% Figure stubs (point \includegraphics at the PNGs produced by the code).
% \begin{figure}[t]\centering
% \includegraphics[width=.6\linewidth]{fig1_toy_count.png}
% \caption{Toy kernel $\eta_{tu}=e^{\beta(t+u)}$: the number of
% translation-invariant Gibbs measures jumps from two to three at
% $\beta_c=\ln t_{0,c}=1.538$ (Results~1--2).}
% \label{fig:toy-count}
% \end{figure}
\end{document}