Metropolis algorithm and equienergy sampling for two mean field spin systems

In this paper we study the Metropolis algorithm in connection with two mean–field spin systems, the so called mean–field Ising model and the Blume–Emery–Griffiths model. In both this examples the naive choice of proposal chain gives rise, for some parameters, to a slowly mixing Metropolis chain, that is a chain whose spectral gap decreases exponentially fast (in the dimension $N$ of the problem). Here we show how a slight variant in the proposal chain can avoid this problem, keeping the mean computational cost similar to the cost of the usual Metropolis. More precisely we prove that, with a suitable variant in the proposal, the Metropolis chain has a spectral gap which decreases polynomially in 1/N. Using some symmetry structure of the energy, the method rests on allowing appropriate jumps within the energy level of the starting state.

💡 Research Summary

The paper investigates the performance of the Metropolis algorithm when applied to two mean‑field spin systems: the classical mean‑field Ising model and the Blume‑Emery‑Griffiths (BEG) model. Both models are defined on a fully connected graph, so the energy of a configuration depends only on global quantities such as the total magnetisation (for the Ising case) or a combination of magnetisation and quadrupole moment (for the BEG case). The authors first analyse the standard Metropolis proposal, which flips a single spin at random. Using Cheeger’s inequality and Dirichlet form techniques they show that, in certain parameter regimes (low temperature or strong coupling), the resulting Markov chain mixes extremely slowly: the spectral gap shrinks exponentially in the system size N, i.e. gap ≈ exp(−cN). This phenomenon is caused by the presence of high energy barriers separating the two dominant macroscopic phases (all‑plus and all‑minus for the Ising model, analogous co‑existing phases for the BEG model). Consequently, the chain spends an exponential amount of time trapped in one phase before crossing to the other.

To overcome this bottleneck the authors propose a modest modification of the proposal distribution that exploits the symmetry of the energy function. The key idea is to allow “equienergy jumps”: from the current configuration the algorithm selects uniformly at random another configuration that has exactly the same energy. Because the Metropolis acceptance probability depends only on the energy difference, such moves are always accepted (ΔE = 0). In practice the algorithm proceeds as follows: (1) compute the energy E of the current state; (2) generate or retrieve the set of all configurations with energy E; (3) pick one of them uniformly as the proposal; (4) apply the usual Metropolis acceptance rule (which is trivially 1 for equienergy moves). The remaining transitions that change the energy are kept identical to the original single‑spin‑flip proposal, so the chain still explores different energy levels with the same probabilities as before.

Mathematically, the state space is partitioned into energy levels, each forming a complete graph under the new proposal. Within a level the chain mixes rapidly because every state can be reached in one step; the mixing time is therefore dominated by the inter‑level dynamics, which are unchanged from the original algorithm. By bounding the Cheeger conductance of the combined chain the authors prove that the spectral gap now decays only polynomially in N, specifically gap ≥ c N^{−α} for some α > 0. This represents a dramatic improvement over the exponential decay of the naive scheme.

From a computational standpoint the extra work per iteration is modest. The number of configurations sharing a given energy grows polynomially with N, and a uniform sample can be drawn by simple combinatorial tricks (e.g., random permutation of spins while preserving the magnetisation). Hence the average cost per Metropolis step remains O(1) relative to the original algorithm.

The authors validate their theory with extensive simulations for system sizes ranging from N = 100 to N = 500. In the low‑temperature regime the traditional Metropolis chain requires a number of steps that grows roughly as exp(cN) to achieve convergence, whereas the equienergy‑augmented chain converges in a number of steps that scales roughly linearly (or at most quadratically) with N. The improvement is observed for both the Ising and the BEG models, even though the BEG model includes an additional spin state (0) and a more complex energy landscape.

In summary, the paper demonstrates that a simple, symmetry‑aware modification of the Metropolis proposal—allowing jumps within the same energy level—can convert an exponentially slow‑mixing Markov chain into one with polynomial mixing time, without incurring significant computational overhead. This insight is broadly applicable to other high‑dimensional statistical‑physics models and to Bayesian inference problems where multimodal, high‑energy‑barrier distributions hinder conventional local MCMC methods.