Evaluation of mode mixing in the efficiency of Markov chain Monte Carlo simulations

The efficiency of Monte Carlo (MC) simulations depends on the set of trial moves. We propose a simple self-learning algorithm to generate a set of nonlocal trial moves, where the base vectors (modes) of the trial moves are gradually mixed during metropolis sampling. The efficiency of the trial moves obtained was evaluated using the MC simulations of a toy model of a short and moderately rigid chain molecule. It was shown that mode mixing can significantly improve the convergence of several quantities, such as the end-to-end distance and angle between the tangent vectors of both ends.