Parallel tempering

Summary

Parallel tempering, also called replica exchange Monte Carlo, is a Markov chain Monte Carlo method that runs several copies of a system at different temperatures. Each copy is updated by its own sampler, and from time to time two copies at neighbouring temperatures propose to exchange their configurations. The exchange is accepted with a Metropolis probability that keeps the product of the Boltzmann distributions invariant. Hot copies cross energy barriers easily, and the exchanges pass these moves down to the cold copies, which would otherwise be trapped in local minima.