download-g.meurant.PDFVIP

(will be inserted by the editor) Received: date / Accepted: date In this paper we give explicit expressions for the norms of the residual vectors generated by the GMRES algorithm applied to a non-normal matrix. They involve the right-hand side of the linear system, the eigenvalues, the eigenvectors and, in the non-diagonalizable case, the principal vectors. They give a complete description of how eigenvalues contribute in forming residual norms and offer insight in what quantities can prevent GMRES from being governed by eigenvalues. GMRES convergence · non-normal matrix · eigenvalues · residual norms We consider the convergence of GMRES (the Generalized Minimal RESidual method) for solving linear systems with complex nonsingular matrices A of size n and n-dimensional right-hand sides b; see e.g. [38] or [37] for a description of the algorithm. The kth GMRES iterate xk minimizes, with x = 0, the norm of the kth residual vector rk = b − Axk over all vectors in the kth Krylov subspace Kk (A, b) ≡ span{b, Ab, . . . , Ak−b}. Therefore, residual norms are non-increasing and satisfy ∥rk ∥ = min ∥p (A)b ∥, p ∈πk G. Meurant 30 rue du sergent Bauchat, 75012 Paris, France. E-mail: gerard.meurant@ J. Duintjer Tebbens Institute of Computer Science, Academy of Sciences of the Czech Republic. Pod Vod´arenskou vˇeˇz´ı 2, 18 207 Praha 8 - Libeˇn and Charles University in Prague, Faculty of Pharmacy in Hradec Kr´alov´e, Heyrovsk´eho 1203, 500 05 Hradec Kr´alov´e. 2 G´erard Meurant, Jurjen Duintjer Tebbens where πk is the set of polynomials of degree k with the value one at the origin and ∥ · ∥ denotes the 2-norm. If the Jordan canonical form of A is denoted by A = XJX − , then