Monte Carlo for the age of Tensor Networks

Abstract

Modern numerical methods have revolutionized the practice of science, creating a third discipline between traditional theory and experiment. Perhaps the most widely known and successful technique has been the Monte Carlo method in general, and the Metropolis algorithm in particular. In this talk, I will present a new way of performing unbiased Monte Carlo simulations based on highly-accurate tensor network contractions. The resulting technique inherits the legendary precision of tensor networks without any of the variational bias. From a Monte Carlo point-of-view, the method can be seen as an aggressive multi-sampling technique where each sample may account for the vast majority of the entire partition function resulting a a drastic reduction in sample-to-sample variance (in contrast to standard
configuration-based Monte Carlo, where only a small subset of possible configurations are sampled). The presented results are all classical, though applications to quantum systems and the sign problem will be discussed.