Irreversible Monte Carlo algorithms for efficient sampling (Q629029)

The letter describes how to upgrade a reversible Monte Carlo (MC) algorithm into an irreversible one covering to the same distribution faster. First, Markov Chain Monte Carlo (MCMC) algorithms are discussed and it is shown that the stationary solution of the master equation satisfies the Balance Condition (BC) which is nothing but the incompressibility condition of the stationary probability flow. Then, breaking reversibility with cycles of the Markov chain is considered. The cycle representation suggests how an irreversible MCMC algorithm can be constructed. It is stated that knowing the state space and carefully planting irreversible cycles, one can achieve a significant acceleration of mixing. Then, by aiming to achieve practical and flexible implementation the authors focus on alternative-building irreversible MCMC algorithms based on controlled deformation of an existent reversible MCMC. They adopt and develop a replication/lifting trick with the next main idea. Instead of planting into the system an irreversible probability flux, corresponding to an ``incompressible'' BC, they add a mixing desirable ``compressible'' flux, and compensate for its compressibility by building an additional replica with reversed flux and allowing some inter-replica transitions. To enforce BC one tunes the replica switching probabilities computed ``on the fly'' (and locally). The letter explains the relatively simple implementation of the idea. The authors design a spin-problem specific irreversible MCMC algorithm and test it on the mean-field spin cluster model (spin system with the Metropolis-Hastings-Glauber algorithm). It is shown by choosing N-spins ferromagnetic cluster (equal strength interaction between all the spins) that the irreversible modification can lead to the dramatic acceleration of MC mixing. The presented results state that the irreversible MC algorithms are especially beneficial for the acceleration of mixing in systems containing multiple soft and zero modes, however inaccessible for standard (reversible) schemes. This situation is typical in systems experiencing a critical showdown in the vicinity of a phase transition, and it is an inherent property of systems possessing internal symmetries of high degree.