An alternating structure-adapted Bregman proximal gradient descent algorithm for constrained nonconvex nonsmooth optimization problems and its inertial variant

We consider the nonconvex nonsmooth minimization problem over abstract sets, whose objective function is the sum of a proper lower semicontinuous biconvex function of the entire variables and two smooth nonconvex functions of their private variables. Fully exploiting the problem structure, we propose an alternating structure-adapted Bregman proximal (ASABP for short) gradient descent algorithm, where the geometry of the abstract set and the function is captured by employing generalized Bregman function. Under the assumption that the underlying function satisfies the Kurdyka-Lojasiewicz property, we prove that each bounded sequence generated by ASABP globally converges to a critical point. We then adopt an inertial strategy to accelerate the ASABP algorithm (IASABP), and utilize a backtracking line search scheme to find "suitable" step sizes, making the algorithm efficient and robust. The global O(1/K) sublinear convergence rate measured by Bregman distance is also established. Furthermore, to illustrate the potential of ASABP and its inertial version (IASABP), we apply them to solving the Poisson linear inverse problem, and the results are promising.