A New Inexact Manifold Proximal Linear Algorithm with Adaptive Stopping Criteria (2508.19234v1)

Abstract: This paper studies the nonsmooth and nonconvex composite optimization on Riemannian manifolds. We propose a new inexact manifold proximal linear algorithm (IManPL) that incorporates both high- and low-accuracy conditions. At each iteration, IManPL solves convex subproblems on tangent spaces inexactly, guided by two adaptive stopping criteria. We establish convergence guarantees and show that IManPL achieves the best iteration complexity for solving the nonsmooth manifold composite optimization. Numerical experiments on the sparse spectral clustering (SSC) and the sparse principal component analysis (SPCA) demonstrate that our methods outperform existing approaches, including the manifold proximal linear algorithm and manifold proximal gradient algorithm.