Replica exchange and expanded ensemble simulations as Gibbs sampling: Simple improvements for enhanced mixing

The widespread popularity of replica exchange and expanded ensemble algorithms for simulating complex molecular systems in chemistry and biophysics has generated much interest in enhancing phase space mixing of these protocols, thus improving their efficiency. Here, we demonstrate how both of these classes of algorithms can be considered a form of Gibbs sampling within a Markov chain Monte Carlo (MCMC) framework. While the update of the conformational degrees of freedom by Metropolis Monte Carlo or molecular dynamics unavoidably generates correlated samples, we show how judicious updating of the thermodynamic state indices—corresponding to thermodynamic parameters such as temperature or alchemical coupling variables—associated with these configurations can substantially increase mixing while still sampling from the desired distributions. We show how state update methods in common use lead to suboptimal mixing, and present some simple, inexpensive alternatives that can increase mixing of the overall Markov chain, reducing simulation times necessary to obtain estimates of the desired precision. These improved schemes are demonstrated for several common applications, including an alchemical expanded ensemble simulation, parallel tempering, and multidimensional replica exchange umbrella sampling.

💡 Research Summary

The paper reinterprets replica‑exchange (RE) and expanded‑ensemble (EE) simulations as special cases of Gibbs sampling, a well‑known technique in statistical inference where variables are alternately drawn from their conditional distributions. In this view the two variables are the molecular configuration x and the thermodynamic state index k (or the permutation S of states among replicas). While the configuration updates must be performed with molecular dynamics or Metropolis‑Hastings moves and inevitably produce correlated samples, the conditional distribution of the state index is analytically tractable, opening the possibility of redesigning the state‑update step without compromising the target joint distribution π(x,k).

The authors first point out that the standard practice of attempting exchanges only between neighboring thermodynamic states limits diffusion in state space, especially for large numbers of states or multidimensional state grids. This “neighbor‑only” scheme can cause very long autocorrelation times for the state variable, which in turn inflates the overall correlation time of the Markov chain.

To address this, the paper proposes several inexpensive modifications that retain detailed balance (or at least balance) while allowing much broader moves in state space:

1. Global proposal – choose any target state k′ uniformly from the full set of K states and accept with the Metropolis–Hastings probability α = min