Event № 814
Abstract: We introduce a novel approach addressing the global analysis of a difficult class of nonlinearly composite nonconvex optimization problems. This genuine nonlinear class captures many problems in modern disparate fields of applications. We develop an original general Lagrangian methodology relying on the idea of turning an arbitrary descent method into a multiplier method. We derive a generic Adaptive Lagrangian Based mUltiplier Method (ALBUM) for tackling the general nonconvex nonlinear composite model which encompasses fundamental Lagrangian methods. This paves the way for proving global convergence results to a critical point of the problem in the broad semialgebraic setting. The potential of our results is demonstrated through the study of two major Lagrangian schemes whose convergence was never analyzed in the proposed general setting: the proximal multiplier method and the proximal alternating direction of multipliers scheme. This is joint work with Jerome Bolte (Toulouse 1 Capitole University) and Marc Teboulle (Tel Aviv University).