报告人：张树中 教授 （美国明尼苏达大学）
报告摘要：In this talk we shall present some new results on non-convex block-optimization models over Riemannian manifolds, with binding linear constraints. We introduce some ADMM (Alternating Direction Method of Multipliers) style algorithms for a block optimization model where the objective is non-convex and each block variables are elements of some given manifolds. Moreover, there are also linear constraints linking all the variables. Such models arise naturally in tensor optimization with constraints, including approximative Tucker decomposition with constraints. Iteration complexity bounds for the iterates converging to a stationary solution are presented, together with preliminary numerical results.