Impact of Measurement Noise on Escaping Saddles in Variational Quantum Algorithms

TL;DR

This paper investigates how measurement shot noise affects the ability of stochastic gradient descent to escape saddle points in variational quantum algorithms, revealing a power-law relationship and a key ratio influencing escape times.

Contribution

It provides a theoretical analysis linking measurement noise to optimization dynamics in VQE, introducing a continuous-time SDE model to understand escape behavior.

Findings

Escape time decreases with increased measurement noise.

Escape time depends on the ratio of learning rate to measurement shots.

The SDE model explains the scaling behavior of escape times.

Abstract

Stochastic gradient descent (SGD) is a frequently used optimization technique in classical machine learning and Variational Quantum Eigensolver (VQE). For the implementation of VQE on quantum hardware, the results are always affected by measurement shot noise. However, there are many unknowns about the structure and properties of the measurement noise in VQE and how it contributes to the optimization. In this work, we analyze the effect of measurement noise to the optimization dynamics. Especially, we focus on escaping from saddle points in the loss landscape, which is crucial in the minimization of the non-convex loss function. We find that the escape time (1) decreases as the measurement noise increases in a power-law fashion and (2) is expressed as a function of $\eta/N_s$ where $\eta$ is the learning rate and $N_s$ is the number of measurements. The latter means that the escape time is approximately constant when we vary $\eta$ and $N_s$ with the ratio $\eta/N_s$ held fixed. This scaling behavior is well explained by the stochastic differential equation (SDE) that is obtained by the continuous-time approximation of the discrete-time SGD. According to the SDE, $\eta/N_s$ is interpreted as the variance of measurement shot noise. This result tells us that we can learn about the optimization dynamics in VQE from the analysis based on the continuous-time SDE, which is theoretically simpler than the original discrete-time SGD.