Strong variational sufficiency of nonsmooth optimization problems on Riemannian manifolds

Abstract: The Riemannian Augmented Lagrangian Method (RALM), a recently proposed algorithm for nonsmooth optimization problems on Riemannian manifolds, has consistently exhibited high efficiency as evidenced in prior studies \cite{ZBDZ21,ZBD22}. It often demonstrates a rapid local linear convergence rate. However, a comprehensive local convergence analysis of the RALM under more realistic assumptions, notably without the imposition of the uniqueness assumption on the multiplier, remains an uncharted territory. In this paper, we introduce the manifold variational sufficient condition and demonstrate that its strong version is equivalent to the manifold strong second-order sufficient condition (M-SSOSC) in certain circumstances. Critically, we construct a local dual problem based on this condition and implement the Euclidean proximal point algorithm, which leads to the establishment of the linear convergence rate of the RALM. Moreover, we illustrate that under suitable assumptions, the M-SSOSC is equivalent to the nonsingularity of the generalized Hessian of the augmented Lagrangian function, which is an essential attribute for the semismooth Newton-type methods.