Brandon Augustino, Dylan Herman, Enrico Fontana, Junhyung Lyle Kim, Jacob Watkins, Shouvanik Chakrabarti, Marco Pistoia (Mar 24 2025).

Abstract: We study quantum algorithms based on quantum (sub)gradient estimation using noisy evaluation oracles, and demonstrate the first dimension independent query complexities (up to poly-logarithmic factors) for zeroth-order convex optimization in both smooth and nonsmooth settings. We match the first-order query complexities of classical gradient descent, using only noisy evaluation oracles, thereby exhibiting exponential separation between quantum and classical zeroth-order optimization. Specifically, for the smooth case, zeroth-order quantum gradient descent achieves $\widetilde{\mathcal{O}}(LR^2/\varepsilon)$ and $\widetilde{\mathcal{O}} \left( \kappa \log(1/\varepsilon \right))$ query complexities, for the convex and strongly convex case respectively; for the nonsmooth case, the zeroth-order quantum subgradient method attains a query complexity of $\widetilde{\mathcal{O}}((GR/\varepsilon)^2 )$. We then generalize these algorithms to work in non-Euclidean settings by using quantum (sub)gradient estimation to instantiate mirror descent, dual averaging and mirror prox. We demonstrate how our algorithm for nonsmooth optimization can be applied to solve an SDP involving $m$ constraints and $n \times n$ matrices to additive error $\varepsilon$ using $\widetilde{\mathcal{O}} ((mn^2+n^{\omega})(Rr/\varepsilon)^2)$ gates, where $\omega \in [2,2.38)$ is the exponent of matrix-multiplication time and $R$ and $r$ are bounds on the primal and dual optimal solutions, respectively. Specializing to linear programs, we obtain a quantum LP solver with complexity $\widetilde{\mathcal{O}}((m+\sqrt{n}) (Rr/\varepsilon)^2).$ For zero-sum games we achieve the best quantum runtimes for any $\varepsilon > 0$ when $m = \mathcal{O}(\sqrt{n})$. We obtain the best algorithm overall (quantum or classical) whenever we further impose $\varepsilon=\Omega((m+n)^{-1/4})$.

Arxiv: https://arxiv.org/abs/2503.17356