cs.LGJul 28, 2026

Quantum Speedups for Stochastic Optimization with Heavy-Tailed Noise

Authors: Bin LuoChengchang LiuJonathan AllcockShengyu ZhangJohn C. S. Lui

Abstract

We study stochastic optimization with heavy-tailed gradient noise. We first propose a novel quantum mean estimator for multivariate heavy-tailed random variables that achieves lower query complexity than optimal classical estimators in the low-dimensional regime. We further develop an unbiased quantum mean estimator by applying a generalized multi-level Monte Carlo technique. We prove quantum lower bounds showing that, when the dimension dd of the random vector is small and can be viewed as a constant, our quantum estimators are optimal up to logarithmic factors. We further derive stronger dimension-dependent lower bounds for tail index p>4/3p>4/3, showing that a nontrivial dependence on the dimension is unavoidable in the low-dimensional regime. Based on these estimators, we propose a quantum normalized stochastic gradient descent method (QNSGD\texttt{QNSGD}), which finds an εε-stationary point using O~(dε5p42p2)\tilde{\mathcal{O}}\big(\sqrt d\,ε^{-\frac{5p-4}{2p-2}}\big) queries to the quantum stochastic gradient oracle. For a convex objective function, we propose a quantum projected stochastic gradient descent method (QPSGD\texttt{QPSGD}), which computes a solution with εε-optimal solution using O~(dε3p22p2+ε2)\tilde{\mathcal{O}}\big(\sqrt d\,ε^{-\frac{3p-2}{2p-2}}+ε^{-2}\big) queries in expectation. These sharper bounds improve upon the classical lower bounds Ω(ε3p2p1)Ω\big(ε^{-\frac{3p-2}{p-1}}\big) for nonconvex problems and Ω(εpp1)Ω\big(ε^{-\frac{p}{p-1}}\big) for convex problems in the low-dimensional regimes dεpp1d\lesssimε^{-\frac{p}{p-1}} and dε2pp1d\lesssimε^{-\frac{2-p}{p-1}}, respectively.

Explore similar work

CardsList