# 45c proving that x also minimizes 945a and conversely

This preview shows page 1. Sign up to view the full content.

This is the end of the preview. Sign up to access the rest of the document.

Unformatted text preview: Then, letting v = Vc we write (9.4.3) as w = Wd (9.4.7) (r(0) ; AVc Wd) = 0: The requirement that this result hold for all w 2 W is equivalent to it holding for all possible choices of d. For this to occur, we must have (r(0) ; AVc W) = 0: 58 Solution Techniques for Elliptic Problems When the inner product corresponds to (9.4.2c), we have WT (r(0) ; AVc) = 0: Thus, c satis es WT AVc = WT r(0) : Assuming that the M M matrix WT AV is invertible, we nd c = (WT AV);1 WT r(0) (9.4.8a) ~ x = x(0) + V(WT AV);1WT r(0) : (9.4.8b) and, using (9.4.2a) and (9.3.5c), Let's look at two methods where the subspaces V and W are one dimensional. Example 9.4.1. Let V = W = spanfv1 g where v1 is an N -vector and choose v1 = r(0) . Then, using (9.4.8) (0) )T (0) c = (r T r (0) (r(0) ) Ar and ~ x = x(0) + cr(0) : We recognize this as being one step of the steepest descent algorithm (Figure 9.3.1). Example 9.4.2. In minimum residual (MR) iteration, we select V = spanfv1g = r(0) and W = spanfw1g = Ar(0) . Using (9.4.8) (0) )T T (0) c = (r T A r (0) (r(0) ) AT Ar and ~ x = x(0) + cr(0) : The matrix A need not be symmetric but only be positive de nite for MR to converge ( 4], Section 5.3). Many iterative techniques for symmetric and non-symmetric matrices utilize Krylov subspaces for the trial space V . A Krylov subspace of dimension M is KM (A r(0) ) = spanfr(0) Ar(0) A2r(0) : : : AM ;1r(0) g: (9.4.9) 9.4. Krylov Subspace Methods 59 It seems clear that vectors v 2 KM have the form v = p(A)r(0) where p(A) is a polynomial of degree not exceeding M ; 1 in A. Methods using Krylov subspaces as trial spaces are called Krylov subspace methods or simply Krylov methods. Di erent approaches are distinguished by their choice of the test space W , with the two most popular choices being KM and AKM . The generalized minimal residual (GMRES) method, in particular, uses the latter choice. Thus, if we de ne the Krylov matrix KM = r(0) Ar(0) A2r(0) : : : AM r(0) ] we could select VM = KM WM = AKM choose the M th solution iterate according to (9.4.2a) and (9.3.5c) as x(M ) = x(0) + VM c(M ) (9.4.10) and determine the M th-G...
View Full Document

## This document was uploaded on 03/16/2014 for the course CSCI 6840 at Rensselaer Polytechnic Institute.

Ask a homework question - tutors are online