13: A preconditioned min-res iteration for the mixed problem. Math 639 (updated: January 2, 2012) The iterative algorithms considered in this section for the mixed problem: given b ∈ Rn+m , find x ∈ Rn+m solving (13.1) Ax = b, are based on the corollary of the previous lecture, namely: Corollary 1. There are constants c0 and c1 depending only on α, β, kAk, and kBk satisfying c0 kwk ≤ kAwk∗ ≤ c1 kwk for all w ∈ Rn+m . (13.2) We shall develop preconditioned iterative algorithms for (13.1). To this end, we assume that M is a symmetric and positive definite (n+m)×(n+m) matrix satisfying (13.3) c2 kwk2 ≤ (M w, w) ≤ c3 kwk2 for all w ∈ Rn+m . Setting kwkM = (M w, w) (for w ∈ Rn+m ) and combining (13.2) and (13.3) gives c5 kwk2M ≤ kAwk2∗ ≤ c6 kwk2 for all w ∈ Rn+m . n+m , 2 −1 2 Here c5 = c−1 3 c0 and c6 = c1 c2 . Now if we define for w ∈ R kwkM,∗ = sup y∈Rn+m we find that (w, y) , kykM −1 2 2 2 c−1 3 kwk∗ ≤ kwkM,∗ ≤ c2 kwk∗ . Combining the above inequalities gives (13.4) c7 kwk2M ≤ kAwk2M,∗ ≤ c8 kwk2 for all w ∈ Rn+m . 2 2 −2 Here c7 = c−2 3 c0 and c8 = c1 c2 . The whole point of introducing the second dual norm is that it has a particularly simple form. In fact, kwk2M,∗ = = sup y∈Rn+m sup z∈Rn+m (w, y)2 (w, y)2 = sup kyk2M y∈Rn+m (M y, y) (w, M −1/2 z)2 = (M −1 w, w) for all w ∈ Rn+m . (z, z) Thus, (13.4) can be rewritten (13.5) c7 kwk2M ≤ kM −1 Awk2M ≤ c8 kwk2 for all w ∈ Rn+m . 1 2 This inequality is the basis for our iterative algorithms. We first consider a simple preconditioned linear linear iteration for (13.1) based on the preconditioned normal equations: (13.6) M −1 AM −1 Ax = M −1 AM −1 b. We note that the matrix M −1 AM −1 A is symmetric in the inner product (M w, w) and is positive definite by (13.5). In fact, (13.5) implies that the condition number K of M −1 AM −1 A is bounded by K ≤ c8 /c7 . Thus, we get a rapidly convergent iteration by applying the preconditioned conjugate gradient method to (13.6) in the (M ·, ·) inner product. The resulting error satisfies the standard conjugate gradient estimate, √ K −1 i kei kM ≤ 2ρ , ρi = √ . K +1 Note that this algorithm can be carried out with two evaluations each of M −1 and A per iteration and the evaluation of M can be completely avoided. We note that M −1 A is also a symmetric matrix in the (M ·, ·) whose square equals M −1 AM −1 A. As (13.5) provides eigenvalue bounds for the square, we find that the eigenvalues of M −1 A are in the intervals: √ √ { c7 ≤ |λ| ≤ c8}. It follows that the min-res method (see below) can be applied to the system M −1 Ax = M −1 b. It would appear that this iteration may be more efficient as it only requires one evaluation of M −1 and A per iteration. This is somewhat misleading it can be shown that for some initial errors, the error after 2k − 1 of min-res coincides with the cg-normal error with k steps. The min-res algorithm produce xj = x0 +θ where θ ∈ Kn is the minimizer kx − xj kM = argminζ∈Kn kx − x0 − ζkM . Here Kn is the Krylov space n−1 Kn = spani=0 {(M −1 A)i r0 }. Here r0 = M −1 (b − Ax0 ). This implies the following theorem. Theorem 1. Proof. The above conjugate gradient algorithm produce x̃j = x0 + θ where e n is the minimizer θ∈K kx − x̃j kM = argminζ∈Ke n kx − x0 − ζkM . 3 e n is the Krylov space Here K e n = spann−1 {(M −1 AM −1 A)i r̃0 }. K i=0 Here r̃0 = M −1 AM −1 (b − Ax0 ). The estimate follows from the conjugate gradient estimate and the observation that e j ⊂ K2j+1 . K