Evaluating the noise resilience of variational quantum algorithms

TL;DR

This paper investigates how different types of noise affect variational quantum algorithms, revealing that overparameterized circuits can be more noise-resilient and that noise impacts the optimization and final state fidelity.

Contribution

The study provides a detailed analysis of noise effects on variational algorithms, introducing the concept of parameter degeneracy and its lifting under noise, and identifying optimal circuit depths for noise mitigation.

Findings

Overparameterized circuits exhibit increased noise resilience.

Noise level linearly affects the deviation from target states.

Optimal circuit depth varies with noise type and level.

Abstract

We simulate the effects of different types of noise in state preparation circuits of variational quantum algorithms. We first use a variational quantum eigensolver to find the ground state of a Hamiltonian in presence of noise, and adopt two quality measures in addition to the energy, namely fidelity and concurrence. We then extend the task to the one of constructing, with a layered quantum circuit ansatz, a set of general random target states. We determine the optimal circuit depth for different types and levels of noise, and observe that the variational algorithms mitigate the effects of noise by adapting the optimised parameters. We find that the inclusion of redundant parameterised gates makes the quantum circuits more resilient to noise. For such overparameterised circuits different sets of parameters can result in the same final state in the noiseless case, which we denote as parameter degeneracy. Numerically, we show that this degeneracy can be lifted in the presence of noise, with some states being significantly more resilient to noise than others. We also show that the average deviation from the target state is linear in the noise level, as long as this is small compared to a circuit-dependent threshold. In this region the deviation is well described by a stochastic model. Above the threshold, the optimisation can converge to states with largely different physical properties from the true target state, so that for practical applications it is critical to ensure that noise levels are below this threshold.