Accelerating quantum imaginary-time evolution with random measurements

TL;DR

This paper introduces a novel method to accelerate quantum imaginary-time evolution by efficiently estimating the quantum Fisher information matrix using random measurements, enabling faster ground state preparation in quantum systems.

Contribution

The paper proposes a new approach to estimate the quantum Fisher information matrix efficiently via random measurements, significantly reducing the computational cost of quantum imaginary-time evolution.

Findings

QFIM can be inferred from partial derivative cross correlations of measurement outcomes.

The proposed estimators enable rapid descent in quantum state optimization.

RMITE algorithm effectively prepares ground states in molecular systems.

Abstract

Quantum imaginary-time evolution (QITE) is a promising tool to prepare thermal or ground states of Hamiltonians, as convergence is guaranteed when the evolved state overlaps with the ground state. However, its implementation using a a hybrid quantum/classical approach, where the dynamics of the parameters of the quantum circuit are derived by McLachlan's variational principle is impractical as the number of parameters $m$ increases, since each step in the evolution takes $\Theta(m^2)$ state preparations to calculate the quantum Fisher information matrix (QFIM). In this work, we accelerate QITE by rapid estimation of the QFIM, while conserving the convergence guarantees to the extent possible. To this end, we prove that if a parameterized state is rotated by a 2-design and measured in the computational basis, then the QFIM can be inferred from partial derivative cross correlations of the probability outcomes. One sample estimate costs only $\Theta(m)$ state preparations, leading to rapid QFIM estimation when a few samples suffice. The second family of estimators take greater liberties and replace QFIMs with averaged classical Fisher information matrices (CFIMs). In an extreme special case optimized for rapid (over accurate) descent, just one CFIM sample is drawn. We justify the second estimator family by proving rapid descent. Guided by these results, we propose the random-measurement imaginary-time evolution (RMITE) algorithm, which we showcase and test in several molecular systems, with the goal of preparing ground states.