Asymptotic convergence analysis and influence of initial guesses on composite Anderson acceleration (Q6185810)

In this paper, polynomial residual update formulas for nonstationary \(AA(m, AA(1))\) are investigated and the authors find that \(AA(m, AA(1))\) with general initial guesses is a multi-Krylov method and it possesses a memory effect. For the linear problem, \(AA(m, AA(1))\) potentially has higher order (at most) polynomial degree and a stronger memory effect than that of \(AA(m)\) at the \(k\)-th iteration. The influence of an initial guess on the asymptotic convergence factor of \(AA(1, AA(1))\) is explored. A scaling invariance property of the initial guess \(x_0\) for \(AA(1, AA(1))\) methods in the linear case are presented. The influence of the initial guess on the asymptotic convergence factor of \(AA(n)\) and \(AA(m, AA(n))\) for both linear and nonlinear problems is numerically tested.