Effect of correlated training samples on other neural-network-based Monte Carlo methods

Determine whether incorporating autocorrelated samples from Monte Carlo chains into the training procedures of contour deformation, neural network quantum states, and neural-network-assisted Monte Carlo algorithms improves the performance of these methods.

Background

The paper investigates training neural control variates using autocorrelated configurations produced by Markov Chain Monte Carlo and shows empirically that such correlated samples can improve variance reduction relative to training on only decorrelated data. This challenges the common practice of discarding correlated samples for training due to their redundancy in error estimation.

Building on these findings, the authors suggest that similar benefits might extend to other neural-network-based techniques that rely on Monte Carlo sampling, specifically naming contour deformation, neural network quantum states, and neural-network-assisted Monte Carlo algorithms. They explicitly state that it remains an open question whether training these methods with correlated samples leads to improvements.