Alternative threshold function for Bayesian Optimization of Variational Quantum Circuits

TL;DR

This paper introduces a modified threshold for the EMICoRe Variational Quantum Eigensolver that better accounts for energy fluctuations, leading to improved or comparable results in quantum system ground state approximations.

Contribution

It proposes an alternative, more lenient threshold for EMICoRe's Confident Region that enhances its performance in variational quantum algorithms.

Findings

Improved accuracy in ground state energy approximation for Ising Hamiltonian

Comparable performance on complex optimization regimes

Analysis of threshold evolution during optimization

Abstract

In this paper, we propose an expansion of the Expected Maximum Improvement over Confident Regions (EMICoRe) Variational Quantum Eigensolver (VQE) -- a technique advanced by Nicoli et al., which utilizes both quantum and classical components to approximate the ground state of a quantum system -- by introducing an alternative threshold for EMICoRe's Confident Region that depends on both the Gaussian process (GP) prior variance and the model's change in predicted energy over a set number of iterations. This modification is a more lenient threshold for the Confident Region and accounts for natural fluctuations in the predicted energy that EMICoRe punishes by eliminating the exploratory benefits presented by the Confident Region. We test both algorithms with the original EMICoRe model as a baseline and our results suggest improvement over EMICoRe's state-of-the-art results for a common benchmark for VQEs, the Ising Hamiltonian, and similar performance for more complex optimization regimes. We analyze the accuracy in approximated ground state energy and how the threshold evolves during optimization to compare the EMICoRe model with the proposed alternative. After comparison, we discuss the potential optimization of the degrees of freedom present in the new threshold for better performance and a more varied choice of system to be approximated.