A characterization of the behavior of the Anderson acceleration on linear problems (Q1931717): Difference between revisions

The authors investigate the Anderson iteration with general nonzero mixing parameters and, continuing the work of \textit{H. F. Walker} and \textit{P. Ni} [SIAM J. Numer. Anal. 49, No. 4, 1715--1735 (2011; Zbl 1254.65067)], show that its convergence behaviour can be characterized by the index \(\mu(A,x_0)\) of this acceleration, introduced here. They prove a formula connecting the results of the first \(\mu\) iterations and GMRES iterations and obtain a theorem on convergence or not of the Anderson acceleration. Whereas for nonlinear problems, the method is known to depend sensitively on the choice of the mixing parameters, this is not the case here, but optimizing the parameters at least assures the descent property of the norm of the residual.