We consider the solution of a linear system of equations using the GMRES iterative method. In some applications, performing inexact matrix-vector products in this method may be interesting, provided that a reasonable convergence of GMRES is achieved. A GMRES algorithm where the matrix vector product is performed inexactly is termed ”inexact GMRES algorithm”. An application of this idea occurs in computational electromagnetics, where the fast multipole method provides approximations of the matrix-vector product within a user-defined precision, and where these inaccurate matrix-vector products are all the more cheaper (in terms of CPU time) as the user-defined precision is low. The key point is then to design a control strategy of the accuracy of the matrix-vector product so that the GMRES converges better in the sense that 1) the inexact method achieves a satisfactory limiting accuracy, 2) within a reasonable number of steps.
\\
In [1], a relaxation strategy is proposed for general systems and validated on a large set of numerical experiments. This work is based on heuristic considerations and proposes a strategy that enables a convergence of the GMRES iterates $x_{k}$ within a relative normwise backward error $\frac{\|b−Ax_{k}\|}{\|A\| \|x_{k}\| + \|b\|}$ less than a prescribed quantity $\eta$ > 0, on a significant number of numerical experiments. Similar strategies have been applied to the solution of device simulation problems using domain decomposition [2] and to the preconditioning of a radiation diffusion problem in [5].
\\
A step toward a theoretical explanation of the observed behaviour of the inexact GMRES is proposed in [3, 4]. In this talk, we show that in spite of this considerable theoretical study, the experimental work of [1] is not fully understood yet. We give an overview of the questions that still remains open both in exact arithmetic and in floating-point arithmetic, and we provide some insights into the solution of some of them. Possible applications of this work for the preconditioned GMRES method, when the matrix-vector product is accurate but the preconditioning operation is approximate, are also investigated, based on [3].
\\
\\
References
\\
[1] A. Bouras and V. Frayss´e. Inexact matrix-vector products in Krylov methods for solving linear systems: a relaxation strategy. SIAM Journal on Matrix Analysis and Applications, 2004. To appear.
\\
[2] A. Bouras, V. Frayss´e, and L. Giraud. A relaxation strategy for inner-outer linear solvers in domain decomposition methods. Technical Report TR/PA/00/17, CERFACS, Toulouse, France, 2000.
\\
[3] V. Simoncini and D. B. Szyld. Theory of inexact Krylov subspace methods and applications to scientific computing. SIAM Journal Scientific Computing, 25:454–477, 2003.
\\
[4] J. van den Eshof and G. L. G. Sleijpen. Inexact Krylov subspace methods for linear systems. SIAM Journal on Matrix Analysis and Applications, February 2004. To appear.
\\
[5] J. S. Warsa, M. Benzi, T. A. Warein, and J. E. Morel. Preconditioning a mixed discontinuous finite element method for radiation diffusion. Numerical Linear Algebra with Applications, 2004. To appear.