Anderson Acceleration for a Class of Nonsmooth Fixed-Point Problems
Wei Bian,Xiaojun Chen,C. T. Kelley
DOI: https://doi.org/10.1137/20m132938x
IF: 2.968
2021-01-01
SIAM Journal on Scientific Computing
Abstract:\bfA \bfb \bfs \bft \bfr \bfa \bfc \bft .We prove convergence of Anderson acceleration for a class of nonsmooth fixed-point problems for which the nonlinearities can be split into a smooth contractive part and a nonsmooth part which has a small Lipschitz constant.These problems arise from compositions of completely continuous integral operators and pointwise nonsmooth functions.We illustrate the results with two examples.\bfK \bfe \bfy \bfw \bfo \bfr \bfd \bfs .nonsmooth equatioins, Anderson acceleration, integral equations, nonlinear equations, fixed-point problems \bfA \bfM \bfS \bfs \bfu \bfb \bfj \bfe \bfc \bft \bfc \bfl \bfa \bfs \bfs \bfi fi\bfc \bfa \bft \bfi \bfo \bfn \bfs .65H10, 45G10 \bfD \bfO \bfI .10.1137/20M132938X 1. Introduction.In this paper we prove convergence of Anderson acceleration [1] for a class of nonsmooth fixed-point problems.Anderson acceleration was originally designed for integral equations and is now very common in electronic structure computations (see [6] and many references since then).Anderson acceleration is essentially the same as direct inversion on the iterative subspace (DIIS) [18,19,26,27], nonlinear GMRES [2,21,23,32], and interface quasi-Newton [7,13,20].It is also closely related to Pulay mixing [25], also known as Commutator DIIS [10,15,16,26].Convergence analysis has been reported in the literature only recently, and most of that work assumes at least continuous differentiability of the fixed-point map.There are convergence results for the linear case [30,31], the continuously differentiable case [3], the Lipschitz continuously differentiable case [29,30], and even smoother cases [8,24].In this paper we assume that nonlinearities can be split into a smooth part and a nonsmooth part with a small Lipschitz constant.The splittings we use in this paper are similar to ones used in nonsmooth nonlinear equations [5,14,17].In those papers the norm of the nonsmooth part was small enough so that using the derivative of the smooth part led to a rapidly convergent Newton-like iteration.In this paper the splitting is only used in the analysis, and the algorithm does not change.However, the classes of problems to which the methods apply are very similar.