Block Coordinate Proximal Gradient Method for Nonconvex Optimization Problems: Convergence Analysis

We propose a block coordinate proximal gradient method for a composite minimization problem with two nonconvex function components in the objective while only one of them is assumed to be differentiable. Under some per-block Lipschitz-like conditions based on Bregman distance, but without the global Lipschitz continuity of the gradient of the differentiable function, we prove that any accumulation point of the sequence is a stationary point of the model. We further show that the stationarity is the ``best" one if the global Lipschitz continuity is additionally assumed, and even the local minimizer for some special cases. Convergence analysis without the global Lipschitz continuity and the enhanced stationarity analysis make our results different from existing results in both the convex and nonconvex contexts.