math.OCApr 13, 2025

Mirror Descent Linearized Augmented Lagrangian Methods for Nonconvex Constrained Stochastic Zeroth-Order Optimization

Authors: Qiankun Shi, Han Yuan, Xiao Wang, Hao Wang

Organizations: Sun Yat-sen University, Guangzhou, 510006, China · Pengcheng Laboratory, Shenzhen, 518066, China · Sun Yat-sen University, Guangzhou, 518066, China · ShanghaiTech University, Shanghai, 201210, China.

Abstract

In this paper, we study nonconvex constrained stochastic zeroth-order optimization problems with exact constraints and stochastic objective evaluations. To solve this class of problems, we propose a framework of mirror descent linearized augmented Lagrangian methods that employs two-point stochastic zeroth-order gradient estimators and exploits non-Euclidean mirror descent geometry. Under mild assumptions, we establish oracle complexity guarantees for finding an εε-KKT point parameterized by p≥2p \geq 2. Under Rademacher smoothing, our analysis reveals a trade-off between the variance of the zeroth-order gradient estimators and the smoothness of the mirror map. In the high-accuracy regime, the resulting effective oracle complexity is O(pd2/pε−3)\mathcal{O}(p d^{2/p}ε^{-3}) for p∈[2,2ln⁡d]p \in [2,2\ln d] and O(ln⁡d ε−3)\mathcal{O}(\ln d\,ε^{-3}) for p>2ln⁡dp > 2\ln d. These bounds reduce the dimension dependence in the leading term. When p=2p=2, our method recovers the Euclidean setting with an oracle complexity of O(dε−3)\mathcal{O}(dε^{-3}), improving the εε-dependence over existing methods. Furthermore, to eliminate initial near-feasibility requirements, we introduce a multi-stage scheme that finds an εε-KKT point within O(1+log⁡log⁡(e/ε))\mathcal{O}(1+\log\log(e/ε)) stages while maintaining the leading-order complexity. Numerical tests on QCQPs, black-box adversarial attacks, and fairness-constrained classification demonstrate the effectiveness of our proposed method.

Explore similar work

CardsList