The free state for the Potts model on Cayley trees is either extremal or glassy

Technical Summary: The Free State for the Potts Model on Cayley Trees is Either Extremal or Glassy

Problem Statement
The paper investigates the structure of the free state (the infinite-volume Gibbs measure obtained with free boundary conditions) for the $q$-state Potts model on a Cayley tree $T_d$ with branching factor $d \ge 2$. A central question in statistical mechanics on trees is determining whether a given Gibbs measure is extremal (a pure state that cannot be decomposed into a convex combination of other Gibbs measures) or non-extremal.

While the threshold for the uniqueness of the Gibbs measure (and thus the extremality of the free state) is well-established for the Ising model ($q=2$) as $\beta \le \text{arctanh}(1/d)$, the structure of the free state in the non-uniqueness regime ($\beta > \text{arctanh}(1/d)$) is more complex. Previous work by Gandolfo, Maes, Ruiz, and Shlosman (2020) demonstrated that for sufficiently low temperatures (large $\beta$), the free state decomposes into a continuum of extremal states. However, this result relied on Peierls contour arguments, which are only valid in the low-temperature regime. The open problem addressed here is whether this "glassy" behavior (decomposition into uncountably many extremal components) persists throughout the entire non-extremality regime, specifically for $\beta > \text{arctanh}(1/\sqrt{d})$, and whether a similar dichotomy holds for general $q$ and models with external fields.

Methodology
The authors develop a proof strategy centered on a zero-one law for an overlap order parameter, avoiding the low-temperature restrictions of Peierls arguments.

1. Overlap Order Parameter ($Q$): Motivated by the Edwards-Anderson order parameter, the authors define an overlap $Q$ measuring the probability that two independent samples from the free state, conditioned on their boundary configurations, yield the same extremal Gibbs measure. Specifically, $Q = \mu^\emptyset_\beta \otimes \mu^\emptyset_\beta (\{(\eta^{(1)}, \eta^{(2)}) : \mu^{\eta^{(1)}} = \mu^{\eta^{(2)}}\})$, where $\mu^\eta$ represents the extremal measure obtained as the limit of conditional measures given boundary $\eta$.
2. Characterization of Extremal Measures: Using Proposition 2, the authors establish a necessary and sufficient condition for two extremal Gibbs measures to be equal: they must have identical likelihood ratios (or "fields") at the roots of subtrees deep within the tree. This reduces the problem of measure equality to the convergence of specific recursive sequences.
3. Recursive Equations on the Tree: By leveraging the self-similar structure of the Cayley tree and the Edwards-Sokal coupling (which relates the Potts model to a random-cluster model), the authors derive a system of non-linear equations for the conditional overlap probabilities $c_{ij}$. These equations relate the overlap at the root to the overlaps at the children of the root, raised to the power of the branching factor $d$.
4. Zero-One Law: The core of the proof involves analyzing the fixed points of these recursive equations. By defining $M$ as the maximum value of the overlap probabilities $c_{ij}$, the authors show that $M$ must satisfy $M = M^d$. Since $d \ge 2$, the only solutions are $M=0$ or $M=1$. This implies that the overlap $Q$ is either 0 or 1 almost surely.

Key Contributions and Results

• Theorem 1 (Main Dichotomy): The paper proves that for the Potts model on $T_d$ ($d \ge 2$) at any inverse temperature $\beta \ge 0$, the free state $\mu^\emptyset_\beta$ satisfies a strict dichotomy:

  1. It is extremal (i.e., $\mu^\emptyset_\beta \in \text{ex}\mathcal{G}_\beta$).
  2. It is glassy, meaning its extremal decomposition measure $\lambda_{\mu^\emptyset_\beta}$ has no atoms. Consequently, the support of this decomposition contains uncountably many extremal Gibbs measures.
     There is no intermediate case where the decomposition contains a countable number of atoms.

• Corollary 1 (Ising Model Threshold): Applying Theorem 1 to the Ising model ($q=2$), the authors combine their result with known extremality thresholds (Bleher, Ruiz, Zagrebnov) to establish that the free state is glassy if and only if $\beta > \text{arctanh}(1/\sqrt{d})$. This extends the previous "continuum decomposition" result from a very low-temperature regime to the entire spin-glass regime.

• Generalization to External Fields (Theorem 2): The methodology is extended to the Potts model with a uniform external field. The paper proves that any completely homogeneous splitting Gibbs measure (CHSGM) also satisfies the same extremal/glassy dichotomy.

• Spin Glass Connection: The authors note that the existence of uncountably many extremal states is a hallmark of the spin glass phase. They suggest that the free state of the ferromagnetic Ising model on $T_d$ exhibits a phase transition at $\beta = \text{arctanh}(1/\sqrt{d})$ that mirrors the spin glass transition found in models with random couplings, a phenomenon not observed in the free state of the Ising model on $\mathbb{Z}^d$ (where the free state is typically a simple average of plus and minus states).

Significance
The paper claims to overcome the barrier of Peierls contour arguments, which previously limited the proof of continuum decomposition to low temperatures. By introducing a zero-one law for the overlap order parameter based on the tree structure and Edwards-Sokal coupling, the authors provide a rigorous characterization of the free state across the entire parameter space.

The significance lies in:

1. Completing the Phase Diagram: Providing the exact threshold ($\beta > \text{arctanh}(1/\sqrt{d})$) for the glassy behavior of the free state in the Ising model on Cayley trees.
2. Structural Insight: Establishing that the free state cannot be a mixture of a finite or countable number of pure states in the non-extremal regime; it must be a mixture of an uncountable continuum.
3. Broad Applicability: Demonstrating that this dichotomy holds for general $q$-state Potts models and models with uniform external fields, suggesting a universal structural property of splitting Gibbs measures on trees.

The authors remain modest regarding the Ising model with an external field ($q=2, H \neq 0$), noting that while the dichotomy holds, determining the exact region where the "middle" state (among three possible CHSGMs) is non-extremal remains an open question.