How Path Integral Diffusion Decides a Finite Spin Glass

Michael ChertkovarXiv 2026 · arXiv

When an AI diffusion model generates an image or a molecule, it starts with pure noise and gradually introduces structure. But at what point does the model "decide" that it’s making a cat instead of a dog, or a specific protein fold instead of another?

In a new paper, Michael Chertkov explores this "speciation" process by applying Path Integral Diffusion (PID) to one of the most challenging targets in physics: the frustrated spin glass. By using a system where the ground truth is mathematically rigorous, the author provides a spectral theory for how and when diffusion models commit to macroscopic choices.

TL;DR

This paper proves that isotropic diffusion is equivalent to "stochastic localization" and that the timing of a model's decisions is governed by the leading eigenvalues of the data's correlation matrix. In spin glasses, this manifests as a two-stage commitment process: first breaking symmetry (the "mirror choice"), followed by a compound decision among glassy energy valleys.

What Problem Does This Paper Address?

While diffusion models are state-of-the-art for generative AI, our understanding of their internal decision-making dynamics remains limited. We know they eventually "commit" to a class, but the sequence and timing of these commitments are often inferred rather than calculated.

The author tackles this by using Ising spin glasses as a laboratory. Spin glasses are notoriously difficult to sample because their energy landscapes are "frustrated"—filled with many competing local minima (valleys). Because the Gibbs measure of these systems is known (or can be exactly enumerated for small sizes), they serve as a perfect "white box" to study how a generator navigates a complex, hierarchical distribution.

Methods

The paper bridges the gap between generative AI and statistical mechanics using several key approaches:

  • PID as Stochastic Localization: The author shows that Path Integral Diffusion from a point source is simply a time-changed version of Gaussian stochastic localization. This means the generator’s "knowledge" at any point is equivalent to the original glass system subjected to a random external field (the "tilt").
  • Partial Commitments: To track the generator's state, the paper defines "partial commitments"—the probability a trajectory assigns to specific nested classes (subsets of configurations) at any given time.
  • The Internal Clock ($\gamma$): The paper moves away from physical time ($t$) and uses the signal-to-noise ratio ($\gamma$) as the universal "internal clock." This reveals that the specific diffusion protocol (how you schedule noise) changes when a decision happens in physical time, but not what is decided.
  • Spectral Prediction: The core theoretical result is that a split between two classes is decided at an internal time $\gamma_{1/2} \simeq 1.044/\lambda_P$, where $\lambda_P$ is the leading eigenvalue of the parent class's correlation matrix.
Conceptual overview of classes, the internal clock $\gamma$, and how different protocols $\nu(t)$ map that clock onto physical time.

Datasets

The study evaluates these theories on:

  • Bethe-lattice glasses: Sparse random graphs where each spin has three neighbors.
  • Sherrington–Kirkpatrick (SK) glasses: Dense systems where every spin interacts with every other spin.
  • Controls: Ferromagnets (unfrustrated), the Mattis model (hidden ferromagnet), and independent spins to ensure the observed glassy structure isn't an artifact of the method.
  • Scale: Exact enumeration for $N \le 24$ spins; sampled posteriors for $N=32$ and $N=64$.

Key Insights and Ideas

  • The "Mirror Choice" Comes First: In zero-field Ising models, the system is symmetric (flipping all spins doesn't change the energy). Every trajectory must first "commit" to one of these two mirror halves.
  • Spectral Ordering: Decisions are not necessarily ordered by the depth of a hierarchy, but by the magnitude of eigenvalues. If two nested levels have similar eigenvalues, the generator decides them simultaneously in a "compound decision."
  • Protocol Independence: For a continuous-time generator, the protocol is just a re-parameterization of time. However, for a discrete generator (taking finite steps), the protocol matters immensely. The author finds that steps are best spent in the "local stage" (where individual spins are fixed) rather than the macroscopic decision windows.
  • Crossovers, Not Transitions: At finite $N$, these commitments are not sharp phase transitions but "crossovers" occurring over a window of SNR (roughly a factor of 25).

Key Results

  • Universal Decision Curves: Information about a level rises along a universal binary-channel curve. The spectral prediction ($\gamma_{1/2} \approx 1.044/\lambda_P$) was accurate within 30% across hundreds of test instances.
  • Trajectory Dynamics: As shown in Figure 3, a single trajectory doesn't descend a tree level-by-level in a clean sequence. Instead, it typically makes two distinct jumps: the mirror choice and then a compound glassy decision.
A trajectory-level view showing the evolution of class probabilities (a, b) and the median commitment times across an ensemble (c).
  • Robustness of Glassy Classes: These glassy classes appear near the glass transition temperature and persist as the system is cooled, whereas they are entirely absent in ferromagnetic controls.

Limitations and Future Work

  • Finite Size: The study is limited to $N \le 64$. While this allows for exact or highly accurate calculations, it doesn't fully address the $N \to \infty$ thermodynamic limit where true phase transitions might occur.
  • Isotropic Assumption: The theory currently assumes isotropic diffusion (noise is added equally to all dimensions). The author suggests that "hierarchy-aware" schedules—using anisotropic noise to target specific eigenvalues—could be a promising direction for more efficient sampling.
  • Compositional Targets: While spin glasses are hierarchical, real-world data like images are often "compositional" (made of independent parts). Future work is needed to see how these spectral rules apply to parallel, independent decisions.

For more details, including the code and data used to generate these results, visit the official repository.

Summary updated

Abstract

When an AI diffusion-type generator samples a frustrated spin glass, which macroscopic alternatives does it decide first, and when? We study this for bridge-diffusion models of the Path Integral Diffusion (PID) type started from a point source. We show that every isotropic PID is a time-changed Gaussian stochastic localization, and conversely every signal-to-noise schedule is realized by a PID protocol. What a continuous-time generator decides therefore does not depend on the protocol, which only sets when and at what cost. We describe the generator's state by its partial commitments: the probabilities it assigns, at each moment, to nested classes of configurations fixed by the target. The partial commitments are martingales that move only along class contrasts. Their ensemble average gives the information about each level at an exact rate set by squared magnetization contrasts. Each split is decided, approximately, like a binary Gaussian channel at signal-to-noise ratio 1.04/λ_P, with λ_P the leading eigenvalue of the parent's correlation matrix. The theory is applied to Bethe-lattice and Sherrington--Kirkpatrick (SK) glasses with up to 64 spins. It is observed generically that a trajectory first shifts its weight to one of the two mirror halves of the zero-field measure and then to the glassy classes inside it. These classes appear near the glass transition and keep their identity on cooling. Nested glassy levels have similar λ_P and are decided together, so a trajectory commits in one or two distinct events---the mirror choice and a compound glassy decision---which separate further with coupling and size, as the spectra predict. For a generator that takes finitely many steps, the protocol matters through the late, local stage rather than the macroscopic decisions.
BibTeX
@misc{chertkov2026how,
  author = {Michael Chertkov},
  title = {{How Path Integral Diffusion Decides a Finite Spin Glass}},
  year = {2026},
  eprint = {2611.11826},
  archivePrefix = {arXiv},
  url = {https://arxiv.org/abs/2611.11826},
}
Figure