Error Stability Properties of Generalized Gradient-Type Algorithms

摘要: We present a unified framework for convergence analysis of generalized subgradient-type algorithms in the presence perturbations. A principal novel feature our is that perturbations need not tend to zero limit. It established iterates are attracted, certain sense, an ɛ-stationary set problem, where ɛ depends on magnitude Characterization attraction sets given general (nonsmooth and nonconvex) case. The results further strengthened convex, weakly sharp, strongly convex problems. Our extends unifies previously known stability properties gradient subgradient methods, including their incremental, parallel, heavy ball modifications.