Robustness of optimized numerical estimation schemes for noisy variational quantum algorithms

TL;DR

This paper investigates the robustness of optimized numerical estimation schemes like SPS and FD for variational quantum algorithms under noise, demonstrating their effectiveness and proposing noise-aware optimization methods.

Contribution

It shows that optimized estimators remain accurate under noise without noise knowledge and introduces a noise-model-agnostic error mitigation technique for SPS estimators.

Findings

Optimized estimators outperform unoptimized ones over wide sampling ranges.

Exponential growth of effective sampling range with qubit number.

Noise-aware optimization improves estimation accuracy significantly.

Abstract

With a finite amount of measurement data acquired in variational quantum algorithms, the statistical benefits of several optimized numerical estimation schemes, including the scaled parameter-shift (SPS) rule and finite-difference (FD) method, for estimating gradient and Hessian functions over analytical schemes~[unscaled parameter-shift (PS) rule] were reported by the present author in [Y. S. Teo, Phys. Rev. A 107, 042421 (2023)]. We continue the saga by exploring the extent to which these numerical schemes remain statistically more accurate for a given number of sampling copies in the presence of noise. For noise-channel error terms that are independent of the circuit parameters, we demonstrate that \emph{without any knowledge} about the noise channel, using the SPS and FD estimators optimized specifically for noiseless circuits can still give lower mean-squared errors than PS estimators for substantially wide sampling-copy number ranges -- specifically for SPS, closed-form mean-squared error expressions reveal that these ranges grow exponentially in the qubit number and reciprocally with a decreasing error rate. Simulations also demonstrate similar characteristics for the FD scheme. Lastly, if the error rate is known, we propose a noise-model-agnostic error-mitigation procedure to optimize the SPS estimators under the assumptions of two-design circuits and circuit-parameter-independent noise-channel error terms. We show that these heuristically-optimized SPS estimators can significantly reduce mean-squared-error biases that naive SPS estimators possess even with realistic circuits and noise channels, thereby improving their estimation qualities even further. The heuristically-optimized FD estimators possess as much mean-squared-error biases as the naively-optimized counterparts, and are thus not beneficial with noisy circuits.