Time correlation functions for quantum systems: validating Bayesian approaches for harmonic oscillators and beyond

TL;DR

This paper improves the calculation of quantum transport properties by enhancing PIMC data accuracy and developing a robust Bayesian inversion method, validated on harmonic oscillators and extended to complex systems.

Contribution

It introduces improved estimators for current correlations and a novel Bayesian inversion workflow combining stochastic maximum entropy and statistical learning.

Findings

Reduced variance in PIMC current correlation data

Effective Bayesian inversion for real-time response functions

Successful validation on harmonic oscillators and complex potentials

Abstract

The quantum harmonic oscillator is the fundamental building block to compute thermal properties of virtually any dielectric crystal at low temperatures in terms of phonons, extended further to cases with anharmonic couplings, or even disordered solids. In general, Path Integral Monte Carlo (PIMC) or Molecular Dynamics (PIMD) methods are powerful tools to determine stochastically thermodynamic quantities without systematic bias, not relying on perturbative schemes. Addressing transport properties, for instance calculating thermal conductivity from PIMC, however, is substantially more difficult. Although correlation functions of current operators can be determined by PIMC from analytic continuation on the imaginary-time axis, Bayesian methods are usually employed for the numerical inversion back to real-time response functions. This task not only strongly relies on the accuracy of the PIMC data, but also introduces noticeable dependence on the model used for the inversion. Here, we address both difficulties with care. In particular, we first devise improved estimators for current correlations which substantially reduce the variance of the PIMC data. Next, we provide a neat statistical approach to the inversion problem, blending into a fresh workflow the classical stochastic maximum entropy method together with recent notions borrowed from statistical learning theory. We test our ideas on a single harmonic oscillator and a collection of oscillators with a continuous distribution of frequencies, and provide indications of the performance of our method in the case of a particle in a double well potential. This work establishes solid grounds for an unbiased, fully quantum mechanical calculation of transport properties in solids.