On the acceleration of Kaczmarz's method for inconsistent linear systems

Abstract

Let the linear system Ax= b be rectangular but solvable. If A is a large sparse matrix, then one possibility to solve the system is to use the iterative method of Kaczmarz. Even in the case when the system is unsolvable, this method is applicable if the relaxation parameters are kept fixed in a certain interval. To get a meaningful result, however, one has to apply strong underrelaxation. We give an interpretation of the limit in terms of a generalized inverse of A and derive bounds with respect to the least-squares solution. Using a result of Buoni and Varga, we estimate the speed of convergence. It is further shown that the eigenvalues of the iteration matrix are nearly real numbers, and we propose to use Chebyshev acceleration. This results in a significant speedup.