Pattern Expansion of Spin Glasses

Abstract: We introduce a systematic method for expanding general spin-glass Hamiltonians in terms of Mattis interactions, providing a novel perspective for understanding the fundamental differences between short-range Edwards-Anderson (EA) and mean-field Sherrington-Kirkpatrick (SK) spin glasses. By iteratively extracting patterns from the coupling matrix, we expand the original spin-glass system into a Hopfield-like model (a series of Mattis interactions) plus a residual system. Our analysis reveals profound distinctions between EA and SK models: while EA models in two and three dimensions break into isolated subconnected sections after expansion, the SK model exhibits remarkable self-similar behavior, with the residual system preserving the mean-field structure and Gaussian statistics throughout the expansion process. This self-similarity manifests in exponential decay of residual matrix norms and expansion coefficients, reflecting the inherent mean-field nature of the SK model. Furthermore, we demonstrate that pattern expansion can identify ultra-low energy excitations in EA models, revealing excitations with energies that decrease rapidly with expansion step. Through connected component analysis, we quantify the size-energy relationship of these independent excitation clusters, opening new avenues for understanding the low-energy landscape of spin glasses and providing insights into the nature of metastable states.

• The paper presents a pattern expansion method that decomposes spin-glass Hamiltonians into a series of Mattis interaction terms.
• It reveals that mean-field SK models retain distributed complexity with many needed patterns, while EA models quickly condense into a few dominant terms.
• The framework offers insights into low-energy excitations and paves the way for applications in optimization and neural coding.

Pattern Expansion of Spin Glasses: A Structural Decomposition Framework

Introduction

The paper "Pattern Expansion of Spin Glasses" (2603.29984) introduces a formal framework for systematically decomposing spin-glass Hamiltonians into Mattis interaction terms, providing a principled route toward understanding the microstructure of energy landscapes in both mean-field Sherrington-Kirkpatrick (SK) and finite-dimensional Edwards-Anderson (EA) spin-glass models. By iteratively extracting patterns from the coupling matrix, this work reframes the original Hamiltonian as a truncated Hopfield-like model (composed of Mattis interactions) plus a residual coupling matrix. Through this perspective, the paper uncovers deep distinctions in the organization and excitations of EA and SK spin glasses, elucidating the role of local and global interactions in the stability and complexity of their metastable states.

Pattern Expansion Formalism

The central contribution is the explicit construction of a pattern-based expansion for a general Ising spin-glass Hamiltonian:

$$\mathcal{H} = - \sum_{\langle i j \rangle} J_{ij} \sigma_i\sigma_j.$$

Here, the coupling matrix $J_{ij}$ is iteratively decomposed as follows:

$$J_{ij} \approx \sum_{\kappa=1}^{M}\alpha_\kappa \Xi_{ij}^{(\kappa)},$$

where each $$\Xi_{ij}^{(\kappa)} = \xi_i^{(\kappa)} \xi_j^{(\kappa)}$$ is a rank-one Mattis interaction term associated with a pattern $\boldsymbol{\xi}^{(\kappa)}\in\{\pm 1\}^N$ and $\alpha_\kappa$ is the associated weight. At each expansion step, the optimal pattern and weight are selected to minimize the Frobenius norm between the coupling matrix and its truncated expansion, analogous to successive principal components but with Ising constraint vectors.

This process produces two complementary objects: the cumulative Hopfield-like representation (the sum of extracted Mattis terms) and a residual coupling matrix, which captures the remainder beyond the extracted backbone.

Structural Distinction: EA versus SK Models

The expansion mechanism reveals fundamentally different behaviors in SK and EA models:

Mean-Field SK Model

The SK model exhibits a profound self-similarity under expansion. At every step, the statistics of the residual coupling matrix remain Gaussian and the mean-field structure is preserved. Empirically and analytically, the residual's Frobenius norm and the expansion coefficients $\alpha_\kappa$ exhibit exponential decay with the number of patterns extracted. In the $N\to\infty$ limit, a large number of Mattis terms are required to significantly alter the SK matrix, reflecting the delocalized and collective nature of its interactions. Notably, the energy landscape complexity in SK cannot be captured by a small set of dominant patterns; instead, its intricacy is distributed across a high-rank hierarchy, consistent with the ultrametric structure of replica symmetry breaking.

Finite-Dimensional EA Model

In contrast, for short-range EA models in two and three dimensions, the residual matrix rapidly fragments after a modest number of expansion steps. The expanded part efficiently approximates the dominant features, and the residual couplings are confined to a sparse set of localized clusters. These clusters become isolated and small, and bonds with substantial strength vanish quickly from the system. The majority of the system's complexity is thus captured by a few expansion terms, indicating a more compressible and local energy landscape than in mean-field cases.

Excitations and Low-Energy Landscape Structure

Pattern expansion provides an efficient mechanism to probe ultra-low energy excitations beyond domain-wall or droplet-based approaches. The ground state of the cumulative Hopfield truncation at order $\kappa$, denoted $\boldsymbol{\sigma}^{(\kappa)}$, serves as a systematic low-energy configuration of the original Hamiltonian. The energy difference between this configuration and the true ground state, $J_{ij}$0, falls rapidly with $J_{ij}$1, as does the configurational (spin) distance.

Connected component analysis of the residual in EA models reveals that these excitations are highly localized: the clusters are small, their energy costs are low, and the correlation between excitation size and energy is weak. This challenges strong versions of droplet theory for EA models, which predict size-energy scaling for excitations, and supports the view that low-energy barriers in EA systems are associated with local clusters of spins.

Thermodynamic Implications and Theoretical Insights

Pattern expansion enables a precise control over the roughness and critical properties of the constructed models. By examining observables such as the density of frustrated plaquettes and Binder cumulant under expansion, it is demonstrated that a small number of expansion patterns are sufficient to recover the thermodynamic behavior (transition temperatures and order parameter fluctuations) of EA spin glasses. In the SK model, pattern expansion reveals that its mean-field criticality and complexity arise from deep structural redundancy, resilient to finite-rank perturbations.

From a theoretical perspective, this work situates pattern expansion as a bridge between the Hopfield model (and its neural network interpretations) and general spin-glass physics. The SK model's self-similarity under expansion aligns with the essence of RSB: complexity encoded in a vast, non-compressible ensemble of configurations, rather than dominance by a tractable basis. For EA models, the compressibility of their Hamiltonians in this Mattis basis may underpin the efficacy of local search and approximate optimization heuristics.

Future Directions

This methodology opens new avenues for research:

• Extension to long-range, hierarchical, or frustrated spin-glass models.
• Quantification of the relationship between extracted patterns and overlap observables, such as replica-averaged correlation functions.
• Application to neural coding, where pattern extraction of coupling matrices could reveal functional modes and constraints in associative memory networks.
• Development of new algorithms for optimization and inference by leveraging the compressibility properties highlighted in short-range glasses.

Conclusion

This paper establishes pattern expansion as a method for dissecting spin-glass Hamiltonians, yielding key insights into the localization, compressibility, and self-similarity of mean-field versus short-range systems. The self-similarity observed in SK models reinforces the foundational results of RSB theory, while the rapid loss of complexity in EA models emphasizes their fundamentally local nature. These distinctions have broad implications for the analysis of disordered systems, neural network theory, and the algorithmic design of heuristic solvers for complex optimization problems.