next up previous
Next: About this document ...

Alexander, R Toth
Local convergence analysis for Anderson acceleration

4546 Mistiflower Dr
Raleigh
NC 27606
artoth@ncsu.edu
Carl, T Kelley

Anderson acceleration is a method used to improve convergence rates of fixed-point iterations. In this method, a new iterate is computed as a linear combination of up to $ m+1$ ($ m$ is a user-defined parameter) previous evaluations of the fixed-point map, where the coefficients are obtained by solving a least-squares problem over linear combinations of the corresponding fixed-point residuals. Past analysis of this method has primarily dealt with showing the relation between this method and GMRES and quasi-Newton methods. We consider the case where the fixed-point map is a contraction and directly prove that the method is locally r-linearly convergent under an assumption of boundedness of the sum of the absolute values of the linear combination coefficients. We show that in a particular case, this boundedness assumption is satisfied and the convergence is q-linear. We also show that the optimization problem for the coefficients may be solved with respect to any norm. Using the $ l_1$ or $ l_\infty$ norm in particular, the optimization problem can be solved as a linear program, which allows the boundedness assumption to be imposed directly as a bound constraint. Several numerical examples are presented illustrating these ideas.




next up previous
Next: About this document ...
Copper Mountain 2014-02-23