Kinetic samplers for neural quantum states

TL;DR

This paper introduces kinetic sampling protocols for neural quantum states that significantly reduce autocorrelation times and improve ergodic coverage compared to traditional methods, enhancing the efficiency of quantum state optimization.

Contribution

It proposes kinetic sampling methods for neural quantum states, demonstrating their advantages over Metropolis-Hastings in terms of autocorrelation and symmetry implementation.

Findings

Kinetic sampling reduces autocorrelation times in NQS.

Kinetic sampling achieves more homogeneous Hilbert space coverage.

Kinetic sampling allows easy implementation of lattice symmetries.

Abstract

Neural quantum states (NQS) are a novel class of variational many-body wave functions that are very flexible in approximating diverse quantum states. Optimization of an NQS ansatz requires sampling from the corresponding probability distribution defined by squared wave function amplitude. For this purpose we propose to use kinetic sampling protocols and demonstrate that in many important cases such methods lead to much smaller autocorrelation times than Metropolis-Hastings sampling algorithm while still allowing to easily implement lattice symmetries (unlike autoregressive models). We also use Uniform Manifold Approximation and Projection algorithm to construct two-dimensional isometric embedding of Markov chains and show that kinetic sampling helps attain a more homogeneous and ergodic coverage of the Hilbert space basis.