[Paper Review] Inapproximability After Uniqueness Phase Transition in Two-Spin Systems

Jin‐Yi Cai, Xi Chen+2 more|arXiv (Cornell University)|May 14, 2012

Markov Chains and Monte Carlo Methods20 references4 citations

TL;DR

This paper establishes computational inapproximability for the partition function of two-spin systems beyond the uniqueness phase transition threshold. Using reductions from the E2LIN2 problem and probabilistic analysis of graph ensembles, it proves that no randomized polynomial-time algorithm can approximate the partition function within relative error ϵ = 10⁻⁴ for d-regular graphs when d = Ω(Δ(β,γ)), unless NP ≠ RP, confirming a long-standing conjecture linking statistical physics phase transitions to computational complexity.

ABSTRACT

A two-state spin system is specified by a 2 x 2 matrix A = {A_{0,0} A_{0,1}, A_{1,0} A_{1,1}} = {β1, 1 γ} where β, γ\ge 0. Given an input graph G=(V,E), the partition function Z_A(G) of a system is defined as Z_A(G) = \sum_{σ: V -> {0,1}} \prod_{(u,v) \in E} A_{σ(u), σ(v)} We prove inapproximability results for the partition function in the region specified by the non-uniqueness condition from phase transition for the Gibbs measure. More specifically, assuming NP e RP, for any fixed β, γin the unit square, there is no randomized polynomial-time algorithm that approximates Z_A(G) for d-regular graphs G with relative error ε= 10^{-4}, if d = Ω(Δ(β,γ)), where Δ(β,γ) > 1/(1-βγ) is the uniqueness threshold. Up to a constant factor, this hardness result confirms the conjecture that the uniqueness phase transition coincides with the transition from computational tractability to intractability for Z_A(G). We also show a matching inapproximability result for a region of parameters β, γoutside the unit square, and all our results generalize to partition functions with an external field.

Motivation & Objective

• To resolve the conjecture that the uniqueness phase transition in two-spin systems coincides with the threshold of computational intractability for partition function approximation.
• To establish strong inapproximability results for anti-ferromagnetic two-spin systems in the non-uniqueness regime, particularly for bounded-degree graphs.
• To extend inapproximability results beyond the unit square (β, γ ∈ [0,1]) to parameters outside this region.
• To generalize the hardness results to systems with external fields, broadening applicability to weighted graph homomorphisms and Ising-type models.
• To provide a tight connection between statistical physics phase transitions and computational complexity by proving that hardness begins precisely at the uniqueness threshold.

Proposed method

• Reduces the E2LIN2 problem to a two-spin system via a randomized graph construction, embedding the hardness of E2LIN2 into the partition function of a d-regular graph.
• Constructs a random graph ensemble H(N, ∆) with controlled degrees and local structures (Ui,k, Vi,k) to simulate the Gibbs measure and analyze expected weights.
• Uses the method of moments and exponential moment bounds to estimate the expectation of the partition function ZA(G), focusing on the leading term in the exponent via entropy functions H(x) = −x ln x − (1−x) ln(1−x).
• Applies Chernoff bounds and concentration inequalities to show that with high probability, the number of 0-assigned vertices in local neighborhoods is tightly concentrated around its mean.
• Employs Markov's inequality and union bounds to control the tail behavior of partition function contributions from rare configurations.
• Derives a tight upper bound on the exponent of the expected partition function, showing it is less than 1.22·dim, which implies exponential separation between typical and rare configurations.

Experimental results

Research questions

• RQ1 Does the uniqueness phase transition in two-spin systems mark the exact boundary between tractable and intractable approximation of the partition function?
• RQ2 Can inapproximability be proven for d-regular graphs when the degree d exceeds the uniqueness threshold Δ(β,γ), even for anti-ferromagnetic systems?
• RQ3 Is the hardness result robust beyond the unit square (β, γ ∈ [0,1]), particularly for ferromagnetic or mixed regimes?
• RQ4 Can the inapproximability result be extended to systems with external fields, and how does this affect the computational threshold?
• RQ5 To what extent does the structure of the graph (e.g., regularity, local tree-likeness) influence the hardness of approximating the partition function?

Key findings

• For any β, γ ∈ [0,1] with (β, γ) ≠ (0,0), (1,1), there is no randomized polynomial-time algorithm that approximates ZA(G) within relative error ϵ = 10⁻⁴ for d-regular graphs with d = Ω(1/(1−βγ)), unless NP = RP.
• The hardness threshold d = Ω(Δ(β,γ)) matches the uniqueness threshold Δ(β,γ) up to a constant factor, confirming the conjecture that the phase transition coincides with the complexity transition.
• The inapproximability result holds even when restricted to d-regular graphs, demonstrating that the hardness arises from the degree exceeding the uniqueness threshold, not from unbounded degrees.
• The paper establishes a matching inapproximability result for parameters outside the unit square, showing that the phase transition-complexity coincidence extends beyond the anti-ferromagnetic regime.
• The expectation of the partition function is shown to be bounded by exp(1.22·dim) for typical configurations, while rare configurations contribute exponentially less, enabling a strong separation in the approximation framework.
• The analysis confirms that the algorithmic FPTAS for anti-ferromagnetic two-spin systems with uniqueness condition is optimal, as soon as the uniqueness condition fails (i.e., d ≥ Δ(β,γ)), the problem becomes inapproximable.