Primal-dual optimization algorithms over Riemannian manifolds: an iteration complexity analysis (Q2205985)

In this paper, a block optimization problem is studied, in which some constraint sets are \(C^{\infty}\) Riemannian submanifolds embedded in Euclidean spaces. After obtaining a necessary optimality condition for the specific problem under investigation, the notion of an \(\varepsilon\)-stationary solution of the problem is defined. Then, a proximal gradient-based ADMM (Alternating Direction Method of Multipliers) algorithm is proposed. It is shown that the algorithm provides an \(\varepsilon\)-stationary solution. An upper bound of the number of necessary iterates is established. For the case where computing the proximal mapping is difficult, a linearized proximal gradient-based ADMM algorithm is also provided, with similar results, another algorithm is suitable for a stochastic version of the problem. Also, a feasible curvilinear line-search variant of ADMM is provided. The paper is complemented by several applications and numerical results.