Approximating Partition Functions of Two-State Spin Systems

Two-state spin systems is a classical topic in statistical physics. We consider the problem of computing the partition function of the systems on a bounded degree graph. Based on the self-avoiding tree, we prove the systems exhibits strong correlation decay under the condition that the absolute value of “inverse temperature” is small. Due to strong correlation decay property, an FPTAS for the partition function is presented under the same condition. This condition is sharp for Ising model.

💡 Research Summary

The paper addresses the computational problem of estimating the partition function of two‑state spin systems (e.g., Ising, hard‑core) defined on graphs of bounded degree. The authors’ main technical contribution is a rigorous proof that such systems exhibit strong correlation decay when the absolute value of the inverse temperature β is sufficiently small relative to the maximum degree Δ. The proof hinges on the construction of a self‑avoiding tree (SAT) for each vertex: a tree that enumerates all self‑avoiding walks starting from that vertex and thereby “unrolls’’ the original graph into an acyclic structure while preserving exact marginal probabilities at the root.

On the SAT, the authors write a recursive message‑passing equation for the ratio R_v of the probability that vertex v takes spin +1 versus –1. By analyzing this recursion they derive an explicit bound
|R_v^{σ} – R_v^{σ′}| ≤ C·(tanh|β|·Δ)^d,
where d is the graph distance between the boundary conditions σ and σ′, and C is a constant independent of the graph size. Consequently, if tanh|β|·Δ < 1/(Δ‑1) (equivalently |β|·Δ < atanh(1/(Δ‑1))) the influence of distant spins decays exponentially fast. This condition is shown to be tight for the ferromagnetic Ising model, matching the known phase‑transition threshold.

The exponential decay enables a deterministic fully polynomial‑time approximation scheme (FPTAS). The algorithm truncates each SAT at depth L = O(log n + log 1/ε). Because the residual influence beyond depth L is bounded by (tanh|β|·Δ)^L ≤ ε/n, the product of the locally computed marginal contributions yields an estimate (\hat Z) that satisfies (1‑ε)Z ≤ (\hat Z) ≤ (1+ε)Z. The runtime is O(n·poly(1/ε)), i.e., polynomial in both the input size and the desired accuracy, provided the graph degree is bounded and the temperature condition holds.

The work extends the seminal self‑avoiding walk tree technique of Weitz (2006) from the hard‑core model to the full class of two‑state spin systems with arbitrary symmetric interaction matrix and external field. It also clarifies the precise relationship between statistical‑physics phase transitions and computational hardness: when the temperature is above the critical value the problem admits an FPTAS, whereas crossing the threshold renders the problem #P‑hard, as known from complexity results for the Ising model.

Beyond the algorithmic result, the paper discusses limitations and future directions. The bounded‑degree assumption is essential; for graphs with unbounded degree the decay factor may exceed one, invalidating the analysis. Extending the correlation‑decay proof to multi‑state Potts models, to graphs with heterogeneous external fields, or to quantum spin systems remains open. Moreover, the authors suggest investigating whether similar deterministic schemes can be designed for the “high‑temperature” regime of other combinatorial counting problems (e.g., matchings, colorings) by constructing appropriate tree‑like expansions.

In summary, the authors provide a clean, mathematically rigorous bridge between statistical‑physics concepts (correlation decay, phase transitions) and algorithmic approximation theory. By leveraging the self‑avoiding tree representation, they obtain a sharp condition under which the partition function of any two‑state spin system on bounded‑degree graphs can be approximated arbitrarily well in polynomial time, and they demonstrate that this condition cannot be improved for the Ising model. The results deepen our understanding of when counting problems are tractable and open avenues for extending these techniques to broader classes of graphical models.