🔢 mathematics

Dynamical Gibbs-non-Gibbs transitions for finite-alphabet models on trees

This paper investigates dynamical Gibbs-non-Gibbs transitions for finite-alphabet spin models on trees under symmetric spin-flip dynamics, demonstrating that while $f$ -stable initial states lead to a recovery of the Gibbs property at large times, $f$ -saddle initial states result in the persistent non-quasilocality of the time-evolved measures.

Original authors: Sebastian Bergmann, Christof Külske, Niklas Schubert

Published 2026-08-18

📖 5 min read🧠 Deep dive

CC BY 4.0

Original authors: Sebastian Bergmann, Christof Külske, Niklas Schubert

Original paper licensed under CC BY 4.0 (http://creativecommons.org/licenses/by/4.0/). ✨ This is an AI-generated explanation of the paper below. It is not written or endorsed by the authors. For technical accuracy, refer to the original paper. Read full disclaimer

Imagine a vast, branching network where tiny units, like spins on a compass needle, can point in one of several directions. In the world of physics, these networks model how materials behave, such as how a magnet holds its shape or how a liquid freezes. For decades, scientists have studied what happens when these systems are left alone to settle into a calm, predictable state known as equilibrium. But a more intriguing question arises when we disturb that calm: if we start with a system in a stable, ordered state and then let it evolve randomly over time, does it remain predictable? Specifically, does the state of one part of the system still depend only on its immediate neighbors, or does it suddenly become influenced by distant, far-away parts of the network? This shift from local predictability to long-range confusion is known as a transition from a "Gibbs" state to a "non-Gibbs" state, and understanding when and why it happens is crucial for grasping the limits of predictability in complex systems.

Researchers Sebastian Bergmann, Christof Külske, and Niklas Schubert have investigated this phenomenon on a specific type of network called a tree, where every point connects to a fixed number of neighbors, branching out endlessly without forming loops. They focused on models where the units can take on a finite number of values, such as the spins in a magnetic material or the colors in a Potts model. The team simulated a process where these units randomly flip their values over time, mimicking the effect of heat or external noise. Their goal was to track whether the system, after being shaken up, would eventually regain its ability to be described by simple, local rules, or if it would fall into a state of permanent, chaotic dependence on distant parts of the tree.

The study reveals a surprising duality in how these systems behave, depending entirely on how they were started. In one scenario, the researchers began with a system in a very stable, ordered configuration. They found that while the random flipping initially caused the system to lose its local predictability, making it impossible to guess the state of a single unit based only on its neighbors, this disorder was temporary. After a certain amount of time, the system spontaneously recovered its order. The influence of distant parts of the network faded away, and the system returned to a state where local rules once again applied. This recovery happens for a broad class of models, provided the initial state was sufficiently stable to begin with. It suggests that some systems have a natural resilience, capable of healing their own structural defects over time.

However, the story takes a sharp turn when the system starts from a different kind of initial state. The researchers identified a specific, precarious starting point that acts like a saddle on a mountain ridge: it is stable in some directions but unstable in others. When the evolution begins from this unstable position, the system does not recover. Instead, as time passes, the disorder deepens until every single possible configuration of the system becomes "bad." In this context, a "bad" configuration means that no matter how you look at the system, the state of one unit is inextricably linked to the state of units far away, making local prediction impossible. The researchers proved that for these specific starting points, the system never regains its local predictability; the loss of order is permanent and total.

To reach these conclusions, the team developed a new mathematical method to track how information flows through the tree. They treated the system as having two layers: the original state at the beginning and the evolving state at a later time. By analyzing how the initial state influences the later state through a series of recursive steps, they could measure how much the distant boundary conditions affected the center of the tree. They showed that for stable starting points, the influence of the distant boundary shrinks exponentially as time goes on, allowing the system to forget the far-away details. Conversely, for the unstable starting points, they demonstrated that the system amplifies tiny differences in the distant boundary, ensuring that the center of the tree remains forever sensitive to the edges.

The findings offer a clear picture of the fate of these evolving systems. They show that the long-term behavior is not a single, uniform outcome but depends critically on the initial conditions. A system can either heal itself and return to a state of local order, or it can spiral into a state of permanent, global dependence. This work extends previous knowledge, which had only shown these behaviors in specific, simpler models, to a much wider family of systems. By proving that recovery is possible for all stable starting points and that total disorder is inevitable for certain unstable ones, the researchers have mapped out the boundaries of predictability in these complex, branching worlds. Their results confirm that while some systems can bounce back from chaos, others are doomed to remain in a state of deep, unresolvable connection, where the past and the distant future are inextricably linked.

Technical Summary: Dynamical Gibbs-non-Gibbs transitions for finite-alphabet models on trees

Problem Statement
This paper investigates the regularity of time-evolved Gibbs measures for finite-alphabet spin models defined on regular trees (Cayley trees) under independent symmetric spin-flip dynamics. Specifically, it addresses the phenomenon of "Dynamical Gibbs-non-Gibbs transitions," where a measure initially Gibbsian (quasilocal) may lose this property (become non-quasilocal) and potentially recover it at later times. The study contrasts the behavior on trees with that on lattices, noting that while lattice Ising models in zero field typically exhibit a permanent loss of Gibbsianness at large times, tree models may display more complex trajectories, including recovery. The central question is to characterize the long-time behavior of these time-evolved measures $\mu_t$ based on the stability properties of the initial Gibbs measure's boundary law.

Methodology
The authors employ a two-layer representation of the time-evolved model, distinguishing between the initial configuration at time $t=0$ (first layer) and the configuration at time $t$ (second layer). The analysis relies on the recursive structure of tree-indexed Markov chains (splitting Gibbs measures).

Recursion Framework: The finite-volume conditional probabilities are expressed via "boundary messages" propagated from the leaves to the root. The time-independent homogeneous recursion $f$ describes the evolution of boundary laws in the static model. The time-dependent, inhomogeneous recursion $G_t^{[k]}$ describes the propagation of messages in the dynamical setting, conditioned on the second-layer spins.
Tree Contraction Analysis: To prove quasilocal Gibbsianness (regularity), the authors analyze the contraction properties of the time-dependent recursions. They establish that for sufficiently large times, these recursions act as contractions on small neighborhoods of the initial boundary law, provided the law corresponds to an $f$ -stable fixed point.
Perturbation of Basins of Attraction: To prove the loss of Gibbsianness (irregularity), the authors study the deformation of basins of attraction for $f$ -stable fixed points under the influence of the time-dependent dynamics. They demonstrate that for unstable fixed points (specifically $f$ -saddles), the dynamics can drive the system into distinct, uniformly separated trapping regions depending on boundary conditions, leading to discontinuities in conditional probabilities.
Specific Models: The general theory is applied to finite-alphabet models, with explicit detailed proofs provided for the three-state ( $q=3$ ) ferromagnetic Potts model on a tree with degree $d=3$ .

Key Contributions and Results

1. Long-Time Recovery for $f$ -Stable States (Theorem 1)
The paper proves that if the initial Gibbs measure corresponds to a boundary law $l$ that is an $f$ -stable fixed point (i.e., the spectral radius of the differential $Df(l)$ is strictly less than 1), then the time-evolved measure $\mu_t^l$ recovers the quasilocal Gibbs property for all sufficiently large times $t \ge t_R$ .

Mechanism: The proof shows that the influence of boundary conditions on the root conditional probability decays exponentially with distance. This is achieved by demonstrating that the time-dependent inhomogeneous recursions become uniform tree contractions on a neighborhood of $l$ as $t \to \infty$ .
Generality: This result holds for all finite-alphabet models with symmetric spin-flip dynamics, extending previous findings specific to the Ising model.

2. Total Badness for Unstable States (Theorem 2)
Conversely, the paper establishes that if the initial measure corresponds to an unstable fixed point (specifically an $f$ -saddle or a repelling fixed point), the time-evolved measure exhibits "total badness" at large times.

Definition: A measure is "totally bad" if every configuration is a point of discontinuity for the conditional probabilities (i.e., no configuration is "good").
Mechanism: For $f$ -saddles, the time-dependent dynamics deform the basins of attraction such that the system is driven into distinct, uniformly separated regions of the simplex depending on the boundary conditions. This results in a persistent long-range dependence where the root probability depends discontinuously on the boundary configuration, regardless of the time $t$ (for $t$ large enough).
Illustration: The authors provide a rigorous proof for the three-state Potts model, showing that for inverse temperatures $\theta > 4$ , the free state (repelling) and $f$ -saddle states lead to total badness.

Significance and Claims
The paper claims to unify and extend the understanding of dynamical Gibbs-non-Gibbs transitions on trees:

Unifying Framework: It generalizes the known "loss and recovery" phenomenon for the Ising model to all finite-alphabet models, identifying $f$ -stability as the precise condition for long-time recovery.
Opposite Phenomenon: It introduces a general mechanism for proving "total badness" (where all configurations are bad) arising from unstable fixed points, contrasting with the lattice behavior where such recovery is often impossible or different.
Optimality: The authors suggest that the $f$ -stability condition for recovery is optimal, excluding only threshold cases with neutral eigenvalues.
Methodological Advance: The work provides a robust method for analyzing the influence of spatially inhomogeneous perturbations on the solutions of time-dependent fixed-point problems, specifically utilizing the geometry of the simplex and tree contraction properties.

The paper does not claim to resolve the intermediate-time behavior (where non-quasilocality may occur even for stable states) nor does it propose experimental applications. Its scope is strictly theoretical, focusing on the characterization of the asymptotic regularity of dynamical Gibbs measures on trees.

More like this