On the accuracy of saddle point solvers

Transcription

1 On the accuracy of saddle point solvers Miro Rozložník joint results with Valeria Simoncini and Pavel Jiránek Institute of Computer Science, Czech Academy of Sciences, Prague, Czech Republic Seminar at the Chinese Academy of Sciences, Beijing, July 11-16, 2010

2 Saddle point problems We consider a saddle point problem with the symmetric 2 2 block form ( ) ( ) ( A B x f B T =. 0 y 0) A is a square n n nonsingular (symmetric positive definite) matrix, B is a rectangular n m matrix of (full column) rank m. Applications: mixed finite element approximations, weighted least squares, constrained optimization etc. [Benzi,Golub, Liesen, 2005]. Numerous schemes: block diagonal preconditioners, block triangular preconditioners, constraint preconditioning, Hermitian/skew-Hermitian preconditioning and other splittings, combination preconditioning References: [Bramble and Pasciak, 1988], [Silvester and Wathen, 1993, 1994], [Elman, Silvester and Wathen, 2002, 2005], [Kay, Loghin and Wathen, 2002], [Keller, Gould and Wathen 2000], [Perugia, Simoncini, Arioli, 1999], [Gould, Hribar and Nocedal, 2001], [Stoll, Wathen, 2008],...

3 Symmetric indefinite system, symmetric positive definite preconditioner A = ( ) A B B T P = R 0 T R A symmetric indefinite, P positive definite (R nonsingular) ( R T AR 1) ( ) ( ) x f R = R T y 0 R T AR 1 is symmetric indefinite!

4 Iterative solution of preconditioned (symmetric indefinite) system Preconditioned MINRES is the MINRES on R T AR 1, minimizes the P 1 = R 1 R T -norm of the residual on K n(p 1 A, P 1 r 0 ) H-MINRES on P 1 A with H = P 1 CG applied to indefinite system with R T AR 1 : CG iterate exists at least at every second step (tridiagonal form T n is nonsingular at least at every second step) [Paige, Saunders, 1975] peak/plateau behavior: CG converges fast MINRES is not much better than CG CG norm increases (peak) MINRES stagnates (plateau) [Greenbaum, Cullum, 1996]

5 Symmetric indefinite system, indefinite or nonsymmetric preconditioner P symmetric indefinite or nonsymmetric P 1 A ( ) ( ) x f = P 1 y 0 ( ) ( ) AP 1 x P = y ( ) f 0 P 1 A and AP 1 are nonsymmetric!

6 Iterative solution of preconditioned nonsymmetric system, positive definite inner product The existence of a short-term recurrence solution methods to solve the system with P 1 A or AP 1 for arbitrary right-hand side vector [Faber, Manteuffel 1984, Liesen, Strakoš, 2006] Matrices P 1 A or AP 1 can be symmetric (self-adjoint) in a given inner product induced by the symmetric positive definite H. Then three term-recurrence method can be applied H(P 1 A) = (P 1 A) T H (P T H) T A = A(P T H) H(AP 1 ) = (AP 1 ) T H HAP 1 = P T AH H(P 1 A) symmetric indefinite: MINRES applied to H(P 1 A) and preconditioned with H H-MINRES on P 1 A H(P 1 A) positive definite: CG applied to H(P 1 A) and preconditioned with H; works on K n(p 1 A, P 1 r 0 ) and can be seen as the CG scheme applied to P 1 A with a nonstandard inner product H H-CG on P 1 A

7 Iterative solution of preconditioned nonsymmetric system, symmetric bilinear form if there exists a symmetric indefinite H such that H(P 1 A) = (P 1 A) T H = [H(P 1 A)] T [(AP 1 ) T H] T = H(AP 1 ) = (AP 1 ) T H is symmetric indefinite MINRES method applied to H(P 1 A) or H(AP 1 ) symmetric indefinite preconditioner H = P 1 = (P 1 ) T so that (P 1 ) T (P 1 )A = A(P 1 ) T (P 1 ) (P 1 ) T AP 1 = P 1 AP 1 right vs left preconditioning for symmetric P P 1 K n (AP 1, r 0 ) = K n (P 1 A, P 1 r 0 ) (AP 1 ) T = (P 1 ) T A = P 1 A

8 Iterative solution of preconditioned nonsymmetric system, symmetric bilinear form H-symmetric variant of the nonsymmetric Lanczos process: AP 1 V n = V n+1 T n+1,n, (AP 1 ) T W n = W n+1 Tn+1,n W T n V n = I = W n = HV n [Freund, Nachtigal, 1995] H-symmetric variant of Bi-CG H-symmetric variant of QMR ITFQMR [Freund, Nachtigal, 1995] QMR-from-BiCG: H-symmetric Bi-CG + QMR-smoothing = H-symmetric QMR [Freund, Nachtigal, 1995, Walker, Zhou 1994] peak/plateau behavior: QMR does not improve the convergence of Bi-CG (Bi-CG converges fast QMR is not much better, Bi-CG norm increases quasi-residual of QMR stagnates) [Greenbaum, Cullum, 1996]

9 Simplified Bi-CG algorithm is a preconditioned CG algorithm H = P 1 -symmetric variant of two-term Bi-CG on AP 1 is the Hestenes-Stiefel CG algorithm on A preconditioned with P P ( 1 ) -symmetric Bi-CG(AP ( ) 1 ) x0 x0, r y 0 = b A 0 y 0 PCG(A) with P 1 P 1 p 0 = P 1 r 0, p 0 = r 0 = P 1 p 0 z 0 = P 1 r 0 k = 0, 1,... ( α k = (r ) k, r k ( )/(AP ) 1 p k, p k ) α k = (r k, z k )/(AP 1 p k, P 1 p k ) xk+1 xk = + α y k+1 y k P 1 p k k r k+1 = r k α k AP 1 p k r k+1 = P 1 r k+1 z k+1 = P 1 r k+1 β k = (r k+1, r k+1 )/(r k, r k ) β k = (r k+1, z k+1 )/(r k, z k ) P 1 p k+1 = P 1 r k+1 + β k P 1 p k P 1 p k+1 = z k+1 + β k P 1 p k p k+1 = P 1 p k+1

10 Saddle point problem and indefinite constraint preconditioner ( ) ( ) A B x B T = 0 y ( I B B T 0 P = ( ) f g ), H = P 1 PCG applied to indefinite system with indefinite preconditioner; will not work for arbitrary ( ) right-hand ( ) side, particular right-hand side or initial guess: x0 s0, r y 0 =, here g = 0 and x = y 0 = 0 [Lukšan, Vlček, 1998], [Gould, Keller, Wathen 2000] [Perugia, Simoncini, Arioli, 1999], [R, Simoncini, 2002]

11 Saddle point problem and indefinite constraint preconditioner - preconditioned system ( A B B T 0 ) ( ) x = y ( ) f, P = 0 ( I ) B B T 0 ( ) A(I Π) + Π (A I)B(B AP 1 = T B) 1 0 I Π = B(B T B) 1 B T - orth. projector onto span(b)

12 Indefinite constraint preconditioner: spectral properties of preconditioned system AP 1 nonsymmetric and non-diagonalizable! but it has a nice spectrum: σ(ap 1 ) {1} σ(a(i Π) + Π) {1} σ((i Π)A(I Π)) {0} and only 2 by 2 Jordan blocks! [Lukšan, Vlček 1998], [Gould, Wathen, Keller, 1999], [Perugia, Simoncini 1999]

13 Basic properties of any Krylov method with the constraint preconditioner e k+1 = ( ) x xk+1 y y k+1 ( f r k+1 = 0 ) ( A B B T 0 ) ( ) xk+1 y k+1 r 0 = ( s0 ) r 0 k+1 = 0 B T (x x k+1 ) = 0 x k+1 Null(B T )! ( sk+1 )

14 The energy-norm of the error in the preconditioned CG method r T k+1 P 1 r j = 0, j = 0,..., k x k+1 is an iterate from CG applied to (I Π)A(I Π)x = (I Π)f! satisfying x x k+1 A = min u x0 +span{(i Π)s x j } u A [Lukšan, Vlček 1998], [Gould, Wathen, Keller, 1999]

15 The residual norm in the preconditioned CG method r k+1 = x k+1 x 0 but in general y k+1 y which is reflected in ( sk+1 0 ) 0! but under appropriate scaling yes!

16 The residual norm in the preconditioned CG method x k+1 x x x k+1 = φ k+1 ((I Π)A(I Π))(x x 0 ) r k+1 = φ k+1 (A(I Π) + Π)s 0 σ((i Π)A(I Π)) σ(a(i Π) + Π) {1} σ((i Π)αA(I ( ) Π)) {0} r k+1 = sk+1 0! 0

17 How to avoid the misconvergence of the scheme Scaling by a constant α > 0 such that {1} conv(σ((i Π)αA(I Π)) {0}) ( ) ( ) ( A B x f B T = 0 y 0) ( ) ( ) αa B x B T = 0 αy ( ) αf 0 v : (I Π)v = 0, α = 1 ((I Π)v, A(I Π)v)! Scaling by a diagonal A (diag(a)) 1/2 A(diag(A)) 1/2 often gives what we want! Different direction vector so that r k+1 = s k+1 is locally minimized! y k+1 = y k + (B T B) 1 B T s k [Braess, Deuflhard,Lipikov 1999], [Hribar, Gould, Nocedal, 1999] [Jiránek, R, 2008]

18 Numerical example A = tridiag(1, 4, 1) R 25,25, B = rand(25, 5) R 25,5 f = rand(25, 1) R 25 σ(a) [2.0146, ] ( ) ( ) 1 αa B I B α = 1/τ σ( B T 0 B T ) 0 1/100 [0.0207, ] {1} 1/10 [0.2067, ] {1} 1/4 [0.5170, ] 1 {1} [2.0678, ] 4 {1} [8.2712, ]

19 τ =100 residual norms r k τ =1 τ = iteration number

20 A norm of errors x x k A τ =100 τ = τ = iteration number

21 Inexact saddle point solvers 1. exact method: exact constraint preconditioning, exact arithmetic : outer iteration for solving the preconditioned system; 2. inexact method with approximate or incomplete factorization scheme to solve inner problems with (B T B) 1 : structure-based or with appropriate dropping criterion; inner iteration method 3. the rounding errors: finite precision arithmetic. References: [Gould, Hribar and Nocedal, 2001], [R, Simoncini, 2002] with the use of [Greenbaum 1994,1997], [Sleijpen, et al. 1994]

22 Delay of convergence and limit on the final accuracy

23 Preconditioned CG in finite precision arithmetic ( xk+1 ȳ k+1 ), r k+1 = ( s (1) k+1 s (2) k+1 x x k+1 A γ 1 Π(x x k+1 ) + γ 2 (I Π)A(I Π)(x x k+1 ) ) Exact arithmetic: Π(x x k+1 ) = 0 (I Π)A(I Π)(x x k+1 ) 0

24 Forward error of computed approximate solution: departure from the null-space of B T + projection of the residual onto it x x k+1 A γ 3 B T (x x k+1 ) + γ 2 (I Π)(f A x k+1 Bȳ k+1 ) can be monitored by easily computable quantities: B T (x x k+1 ) s (2) k+1 (I Π)(f A x k+1 Bȳ k+1 ) (I Π) s (1) k+1

25 Maximum attainable accuracy of the scheme ( f 0) (f A x k+1 Bȳ k+1 ) s (1) k+1, B T (x x k+1 ) s (2) k+1 ( A B B T 0 ) ) ( xk+1 ȳ k+1 ) (1) ( s k+1 s (2) k+1 c 1 εκ(a) max j=0,...,k+1 r j [Greenbaum 1994,1997], [Sleijpen, et al. 1994] good scaling: r j 0 nearly monotonically r 0 max j=0,...,k+1 r j

26 r k 10 4 convergence characteristics Π x k (I Π) ( f A x k B y k ) x x k A iteration number

27 r k x x k A convergence characteristics (I Π) ( f A x k B y k ) Π x k iteration number

28 convergence characteristics r k x x k A (I Π) ( f A x k B y k ) Π x k iteration number

29 Conclusions Short-term recurrence methods are applicable for saddle point problems with indefinite preconditioning at a cost comparable to that of symmetric solvers. There is a tight connection between the simplified Bi-CG algorithm and the classical CG. The convergence of CG applied to saddle point problem with indefinite preconditioner for all right-hand side vectors is not guaranteed. For a particular set of right-hand sides the convergence can be achieved by the appropriate scaling of the saddle point problem or by a different back-substitution formula for dual unknowns. Since the numerical behavior of CG in finite precision arithmetic depends heavily on the size of computed residuals, a good scaling of the problems leads to approximate solutions satisfying both two block equations to the working accuracy.

30 Thank you for your attention. miro M. Rozložník and V. Simoncini, Krylov subspace methods for saddle point problems with indefinite preconditioning,siam J. Matrix Anal. Appl., 24 (2002), pp P. Jiránek and M. Rozložník. Maximum attainable accuracy of inexact saddle point solvers. SIAM J. Matrix Anal. Appl., 29(4): , P. Jiránek and M. Rozložník. Limiting accuracy of segregated solution methods for nonsymmetric saddle point problems. J. Comput. Appl. Math. 215 (2008), pp

31 References I M. Arioli. The use of QR factorization in sparse quadratic programming and backward error issues. SIAM J. Matrix Anal. Appl., 21(3): , Z. Bai, G. H. Golub, and M. Ng. Hermitian and skew-hermitian splitting methods for non-hermitian positive definite linear systems. SIAM J. Matrix Anal. Appl., 24: , A. Greenbaum. Estimating the attainable accuracy of recursively computed residual methods. SIAM J. Matrix Anal. Appl., 18(3): , C. Vuik and A. Saghir. The krylov accelerated simple(r) method for incompressible flow. Technical Report 02-01, Delft University of Technology, 2002.

32 Null-space projection method compute x N(B T ) as a solution of the projected system (I Π)A(I Π)x = (I Π)f, compute y as a solution of the least squares problem By f Ax, Π = B(B T B) 1 B T is the orthogonal projector onto R(B). Results for schemes, where the least squares with B are solved inexactly. Every computed approximate solution v of a least squares problem Bv c is interpreted as an exact solution of a perturbed least squares (B + B) v c + c, B τ B, c τ c, τκ(b) 1.

33 Null-space projection method choose x 0, solve By 0 f Ax 0 compute α k and p (x) k N(B T ) x k+1 = x k + α k p (x) k solve Bp (y) k r (x) k α k Ap (x) k back-substitution: A: y k+1 = y k + p (y) k, inner iteration B: solve By k+1 f Ax k+1, C: solve Bv k f Ax k+1 By k, y k+1 = y k + v k. r (x) k+1 = r(x) k α k Ap (x) k Bp (y) k outer iteration

34 Accuracy in the saddle point system f Ax k By k r (x) k O(α 3)κ(B) 1 τκ(b) ( f + A X k), B T x k O(τ)κ(B) 1 τκ(b) B X k, X k max{ x i i = 0, 1,..., k}. Back-substitution scheme α 3 A: Generic update y k+1 = y k + p (y) u k B: Direct substitution y k+1 = B τ (f Ax k+1 ) C: Corrected dir. subst. y k+1 = y k + B u (f Ax k+1 By k ) } additional least square with B

35 Maximum attainable accuracy of inexact null-space projection schemes The limiting (maximum attainable) accuracy is measured by the ultimate (asymptotic) values of: 1. the true projected residual: (I Π)f (I Π)A(I Π)x k ; 2. the residuals in the saddle point system: f Ax k By k and B T x k ; 3. the forward errors: x x k and y y k. Numerical experiments: a small model example A = tridiag(1, 4, 1) R , B = rand(100, 20), f = rand(100, 1), κ(a) = A A 1 = , κ(b) = B B =

36 Generic update: y k+1 = y k + p (y) k τ = 10 2 relative residual norms f Ax k By k / f Ax 0 By 0, r (x) k / r(x) τ = O(u), τ = 10 10, τ=10 6, τ=10 2 residual norm B T x k τ = 10 6 τ = τ = O(u) iteration number iteration number

37 Direct substitution: y k+1 = B (f Ax k+1 ) 10 0 τ = τ = 10 2 relative residual norms f Ax k By k / f Ax 0 By 0, r (x) k / r(x) τ = 10 6 τ = τ = O(u) residual norm B T x k τ = 10 6 τ = τ = O(u) iteration number iteration number

38 Corrected direct substitution: y k+1 = y k + B (f Ax k+1 By k ) τ = 10 2 relative residual norms f Ax k By k / f Ax 0 By 0, r (x) k / r(x) τ = O(u), τ = 10 10, τ=10 6, τ=10 2 residual norm B T x k τ = 10 6 τ = τ = O(u) iteration number iteration number

39 Schur complement reduction method Compute y as a solution of the Schur complement system compute x as a solution of B T A 1 By = B T A 1 f, Ax = f By. inexact solution of systems with A: every computed solution û of Au = b is interpreted an exact solution of a perturbed system (A + A)û = b + b, A τ A, b τ b, τκ(a) 1.

40 Iterative solution of the Schur complement system choose y 0, solve Ax 0 = f By 0 compute α k and p (y) k y k+1 = y k + α k p (y) k solve Ap (x) k = Bp (y) k back-substitution: A: x k+1 = x k + α k p (x) k, inner iteration B: solve Ax k+1 = f By k+1, C: solve Au k = f Ax k By k+1, x k+1 = x k + u k. r (y) k+1 = r(y) k α k B T p (x) k outer iteration

41 Maximum attainable accuracy of inexact Schur complement schemes The limiting (maximum attainable) accuracy is measured by the ultimate (asymptotic) values of: 1. the Schur complement residual: B T A 1 f B T A 1 By k ; 2. the residuals in the saddle point system: f Ax k By k and B T x k ; 3. the forward errors: x x k and y y k. Numerical experiments: a small model example A = tridiag(1, 4, 1) R , B = rand(100, 20), f = rand(100, 1), κ(a) = A A 1 = , κ(b) = B B =

42 Accuracy in the outer iteration process B T A 1 f + B T A 1 By k r (y) k O(τ)κ(A) 1 τκ(a) A 1 B ( f + B Y k ). Y k max{ y i i = 0, 1,..., k}. relative residual norms B T A 1 f B T A 1 By k / B T A 1 f B T A 1 By 0, r (y) k / r(y) τ = 10 2 τ = 10 6 τ = τ = O(u) iteration number k B T (A + A) 1 Bŷ = B T (A + A) 1 f, B T A 1 f B T A 1 Bŷ τκ(a) 1 τκ(a) A 1 B 2 ŷ.

43 Accuracy in the saddle point system f Ax k By k O(α 1)κ(A) 1 τκ(a) ( f + B Y k), B T x k r (y) k O(α 2)κ(A) 1 τκ(a) A 1 B ( f + B Y k ), Y k max{ y i i = 0, 1,..., k}. Back-substitution scheme α 1 α 2 A: Generic update x k+1 = x k + α k p (x) τ u k B: Direct substitution x k+1 = A 1 τ τ (f By k+1 ) C: Corrected dir. subst. x k+1 = x k + A 1 u τ (f Ax k By k+1 ) } additional system with A B T A 1 f + B T A 1 By k = B T x k B T A 1 (f Ax k By k )

44 Generic update: x k+1 = x k + α k p (x) k 10 0 τ = residual norm f Ax k By k τ = 10 6 τ = τ = O(u) relative residual norms B T x k / B T x 0, r k (y) / r0 (y) τ = O(u), τ = 10 10, τ=10 6, τ= iteration number k iteration number k

45 Direct substitution: x k+1 = A 1 (f By k+1 ) 10 0 τ = τ = 10 2 residual norm f Ax k By k τ = 10 6 τ = τ = O(u) relative residual norms B T x k / B T x 0, r k (y) / r0 (y) τ = 10 6 τ = τ = O(u) iteration number k iteration number k

46 Corrected direct substitution: x k+1 = x k + A 1 (f Ax k By k+1 ) τ = 10 2 residual norm f Ax k By k τ = O(u), τ = 10 10, τ=10 6, τ=10 2 relative residual norms B T x k / B T x 0, r k (y) / r0 (y) τ = 10 6 τ = τ = O(u) iteration number k iteration number k

47 Related results in the context of saddle-point problems and Krylov subspace methods General framework of inexact Krylov subspace methods: in exact arithmetic the effects of relaxation in matrix-vector multiplication on the ultimate accuracy of several solvers [?], [?]. The effects of rounding errors in the Schur complement reduction (block LU decomposition) method and the null-space method [?], [Arioli, 2000], the maximum attainable accuracy studied in terms of the user tolerance specified in the outer iteration [?], [?]. Error analysis in computing the projections into the null-space and constraint preconditioning, limiting accuracy of the preconditioned CG, residual update strategy when solving constrained quadratic programming problems [?], or in cascadic multigrid method for elliptic problems [?]. Theory for a general class of iterative methods based on coupled two-term recursions, all bounds of the limiting accuracy depend on the maximum norm of computed iterates, fixed matrix-vector multiplication, cf. [Greenbaum, 1997].

48 General comments and considerations, future work new value = old value + small correction Fixed-precision iterative refinement for improving the computed solution x old to a system Ax = b: solving update equations Az corr = r that have residual r = b Ay old as a right-hand side to obtain x new = x old + z corr, see [?], [?]. Stationary iterative methods for Ax = b and their maximum attainable accuracy [?]: assuming splitting A = M N and inexact solution of systems with M, use x new = x old + M 1 (b Ax old ) rather than x new = M 1 (Nx old + b), [?]. Two-step splitting iteration framework: A = M 1 N 1 = M 2 N 2 assuming inexact solution of systems with M 1 and M 2, reformulation of M 1 x 1/2 = N 1 x old + b, M 2 x new = N 2 x 1/2 + b, Hermitian/skew-Hermitian splitting (HSS) iteration [Bai, Golub, and Ng, 2003]. Inexact preconditioners for saddle point problems: SIMPLE and SIMPLE(R) type algorithms [Vuik and Saghir, 2002] and constraint preconditioners [?].