Stochastic optimization over expectation-formulated generalized Stiefel manifold

TL;DR

This paper introduces a new penalty function for stochastic optimization on the generalized Stiefel manifold, enabling effective algorithms with proven convergence and demonstrated efficiency through experiments.

Contribution

It proposes a novel constraint dissolving penalty function with theoretical equivalence to the original problem, facilitating new stochastic optimization algorithms.

Findings

The proposed algorithms have an $ ext{O}( extstyle{rac{1}{ ext{ε}^4})}$ sample complexity.

The penalty function maintains differentiability and equivalence to the original problem.

Numerical experiments confirm the efficiency and robustness of the methods.

Abstract

In this paper, we consider a class of stochastic optimization problems over the expectation-formulated generalized Stiefel manifold (SOEGS), where the objective function $f$ is continuously differentiable. We propose a novel constraint dissolving penalty function with a customized penalty term (CDFDP), which maintains the same order of differentiability as $f$. Our theoretical analysis establishes the global equivalence between CDFCP and SOEGS in the sense that they share the same first-order and second-order stationary points under mild conditions. These results on equivalence enable the direct implementation of various stochastic optimization approaches to solve SOEGS. In particular, we develop a stochastic gradient algorithm and its accelerated variant by incorporating an adaptive step size strategy. Furthermore, we prove their $\mathcal{O}(\varepsilon^{-4})$ sample complexity for finding an $\varepsilon$-stationary point of CDFCP. Comprehensive numerical experiments show the efficiency and robustness of our proposed algorithms.