Beyond Standard Krylov Subspace Methods – Desirable Features for Scientific Software – Andreas Frommer Bergische Universität Wuppertal June 25, 2010 AI Introduction Shifted linear systems Flexible Methods Indefinite inner products Conclusions Basics Computing the basis The methods AI The story of Krylov subspaces History: I 1952: CG by Hestenes and Stiefel I 1975: CG rediscoverd as iterative method by Fletcher, BiCG I 80’s-90’s: General Krylov subspace framework: many authors I 1984: The Faber-Manteuffel theorem I today: established methods I Bye bye: Chebyshev, splitting based methods, SOR, . . . Beyond Standard Krylov Subspace Methods – Desirable Features for Scientific Software –, Andreas Frommer 2/43 Introduction Shifted linear systems Flexible Methods Indefinite inner products Conclusions Basics Computing the basis The methods AI Basics about Krylov subspaces A ∈ Cn×n , (A, r 0 ) Definition: Kk deg(p) ≤ k − 1}. = Ax = b, A non-singular span{r 0 , Ar 0 , . . . , Ak−1 r 0 } = {p(A)r 0 : Known fact: ∃ p, deg(p) ≤ n, p(0) = 1 s.t. p(A)r 0 = 0 (“individualized” Cayley Hamilton) Idea: I r 0 = b − Ax 0 initial residual I p(t) = 1 − tq(t) I x ∗ = x 0 + q(A)r 0 satisfies b − Ax = p(A)r 0 = 0. Krylov subspace method: x k ∈ x 0 + Kk (A, r 0 ) Beyond Standard Krylov Subspace Methods – Desirable Features for Scientific Software –, Andreas Frommer 3/43 Introduction Shifted linear systems Flexible Methods Indefinite inner products Conclusions Basics Computing the basis The methods AI State of the art: I Krylov subspace methods as iterative solvers I common feature: x k from x 0 + Kk (A, r 0 ) I differences: variational characterization of x k available as standard software: I I I I I I Matlab NAG Petsc ITPACK ... Beyond Standard Krylov Subspace Methods – Desirable Features for Scientific Software –, Andreas Frommer 4/43 Introduction Shifted linear systems Flexible Methods Indefinite inner products Conclusions Basics Computing the basis The methods AI Arnoldi process Wanted: (Nested) orthogonal basis v 1 , . . . , v k for Kk (A, r 0 ). Arnoldi: Orthogonalize Av k against v 1 , . . . , v k , normalize → v k+1 Arnoldi summary: Vk = [v 1 | . . . |v k ] AVk = Vk+1 Hk+1,k , Hk+1,k = | ∗ ∗ 0 0 .. . ∗ ∗ ∗ 0 .. . ∗ ∗ ∗ ∗ .. . ··· ··· ··· ··· .. . ∗ ∗ ∗ ∗ .. . 0 0 ··· {z 0 ∗ k " = 0 # Hk ... ∗ } Note: Hk = Vk∗ AVk represents matrix projected on Kk (A, r 0 ) Beyond Standard Krylov Subspace Methods – Desirable Features for Scientific Software –, Andreas Frommer 5/43 Introduction Shifted linear systems Flexible Methods Indefinite inner products Conclusions Basics Computing the basis The methods AI Lanczos process If A is hermitian it is sufficient to orthogonalize against v k and v k−1 Consequence: Short recurrences (length 2 instead of length k) Lanczos summary: Vk = [v 1 | . . . |v k ] AVk = Vk+1 Tk+1,k , Tk+1,k = | ∗ ∗ 0 0 .. . ∗ ∗ ∗ 0 .. . 0 ∗ ∗ ∗ .. . ··· ··· ··· ··· .. . 0 0 0 0 .. . 0 0 ··· {z 0 ∗ k " = 0 # Tk ... . ∗ } Note: Tk = Vk∗ AVk is tridiagonal, Hermitian (and real) Beyond Standard Krylov Subspace Methods – Desirable Features for Scientific Software –, Andreas Frommer 6/43 Introduction Shifted linear systems Flexible Methods Indefinite inner products Conclusions Basics Computing the basis The methods AI Nonsymmetric Lanczos process Idea: Maintain short recurrences for A non-hermitian. Price: Basis Vk is not orthonormal but bi-orthogonal: v k+1 ⊥ Kk (A∗ , r̃ ), w k+1 ⊥ Kk (A, r 0 ). Nonsymmetric Lanczos summary: AVk = Vk+1 Tk+1,k , Tk+1,k = | ∗ ∗ 0 0 .. . ∗ ∗ ∗ 0 .. . 0 ∗ ∗ ∗ .. . ··· ··· ··· ··· .. . 0 0 0 0 .. . 0 0 ··· {z 0 ∗ k " = 0 # Tk ... ∗ } Note: Tk = Vk∗ AVk is tridiagonal, non-Hermitian Beyond Standard Krylov Subspace Methods – Desirable Features for Scientific Software –, Andreas Frommer 7/43 Introduction Shifted linear systems Flexible Methods Indefinite inner products Conclusions Basics Computing the basis The methods AI Variational characterizations Arnoldi or Lanczos based: CG, FOM: r k = b − Ax k ⊥ Kk (A, r 0 ) MINRES,GMRES: r k ⊥ AKk (A, r 0 ) ⇔ x k = argminx∈x 0 +Kk kb − Ax k k2 Nonsymmetric Lanczos based: BiCG: r k ⊥ Kk (A∗ , r̃ 0 ) QMR: quasi minimal resdiual property BiCGStab: nonsymmetric Lanczos, but don’t build Wk . . . and many more: TFQMR, LSQR, SYMMLQ, CGS, IDR, . . . Beyond Standard Krylov Subspace Methods – Desirable Features for Scientific Software –, Andreas Frommer 8/43 Introduction Shifted linear systems Flexible Methods Indefinite inner products Conclusions Basics Computing the basis The methods AI Contents In this talk we want to emphasize very useful enhancements for standard Krylov subspace methods: Beyond Standard Krylov Subspace Methods – Desirable Features for Scientific Software –, Andreas Frommer 9/43 Introduction Shifted linear systems Flexible Methods Indefinite inner products Conclusions Basics Computing the basis The methods AI Contents In this talk we want to emphasize very useful enhancements for standard Krylov subspace methods: I Shifted linear systems I Flexible methods I Indefinite inner products I Deflation I Conclusions Beyond Standard Krylov Subspace Methods – Desirable Features for Scientific Software –, Andreas Frommer 10/43 Introduction Shifted linear systems Flexible Methods Indefinite inner products Conclusions Algorithms Issues for Software Applications AI Definition A family of shifted linear systems is given as (A − σi I )xi = b, i = 1, . . . , p where σi ∈ C, i = 1, . . . , p. Crucial fact: Kk (A, b) = Kk (A − σi I , b) Moreover: In Arnoldi, Lanczos, nonsymmetric Lanczos: I vectors v k , w k independent of σi I Hk → Hk − σi I Consequence: There is a potential for doing the iteration on the shifted systems without any further matrix operations Beyond Standard Krylov Subspace Methods – Desirable Features for Scientific Software –, Andreas Frommer 11/43 Introduction Shifted linear systems Flexible Methods Indefinite inner products Conclusions Algorithms Issues for Software Applications AI Example: Multishift CG choose x 0 , compute r 0 = b − Ax 0 initialize other quantities for j = 1, . . . , until convergence {begin of j-th Lanczos step} δ j = hAv j , v j i, δ̂ j = δ j − σ w j+1 = Av j − δ j v j − β j v j−1 β j+1 = kw j+1 k v j+1 = w j+1 /β j+1 {end of Lanczos step, compute cg iterates} if j > 0 then j j j j ρj = 1/(1 + τ τj−1 βδj ), ρ̂j = 1/(1 + τ̂ τ̂j−1 βδ̂j ) τ j+1 = −ρj τ j β j+1 /δj τ̂ j+1 = −ρ̂j τ̂ j β j+1 /δ̂j x j+1 = ρj (x j + τ j /δ j )v j + (1 − ρj )x j−1 x̂ j+1 = ρ̂j (x̂ j + τ̂ j /δ̂ j )v j + (1 − ρ̂j )x̂ j−1 Beyond Standard Krylov Subspace Methods – Desirable Features for Scientific Software –, Andreas Frommer 12/43 Introduction Shifted linear systems Flexible Methods Indefinite inner products Conclusions Algorithms Issues for Software Applications AI Other methods I multishift BiCG I multishift QMR (Freund 1994) I multishift FOM I multishft GMRES I multishift restarted FOM (Simoncini 06) but not for I multishift restarted GMRES Beyond Standard Krylov Subspace Methods – Desirable Features for Scientific Software –, Andreas Frommer 13/43 Introduction Shifted linear systems Flexible Methods Indefinite inner products Conclusions Algorithms Issues for Software Applications AI Restarted Krylov subspace methods Situation: A non-hermitian, use FOM or GMRES. When long recurrences become too long . . . Beyond Standard Krylov Subspace Methods – Desirable Features for Scientific Software –, Andreas Frommer 14/43 Introduction Shifted linear systems Flexible Methods Indefinite inner products Conclusions Algorithms Issues for Software Applications AI Restarted Krylov subspace methods Situation: A non-hermitian, use FOM or GMRES. When long recurrences become too long . . . . . . we stop, and restart with the current iterate x m and residual r m Beyond Standard Krylov Subspace Methods – Desirable Features for Scientific Software –, Andreas Frommer 15/43 Introduction Shifted linear systems Flexible Methods Indefinite inner products Conclusions Algorithms Issues for Software Applications AI Restarted Krylov subspace methods Situation: A non-hermitian, use FOM or GMRES. When long recurrences become too long . . . . . . we stop, and restart with the current iterate x m and residual r m Fix cycle length m: x m ∈ x 0 +Km (A, r 0 ), x 2m ∈ x m +Km (A, r m ), x 3m ∈ x 2m +Km (A, r 2m ), . . . Restarted methods: FOM(m), GMRES(m) Beyond Standard Krylov Subspace Methods – Desirable Features for Scientific Software –, Andreas Frommer 16/43 Introduction Shifted linear systems Flexible Methods Indefinite inner products Conclusions Algorithms Issues for Software Applications AI Enforcing collinear residuals x m ∈ x 0 +Km (A, r 0 ), x 2m ∈ xm +Km (A, r m ), x 3m ∈ x 2m +Km (A, r 2m ), . . . Problem with shifts in GMRES(m): b − (A − σi I )xim is not collinear to b − Ax m . Approach for multishift GMRES(m): I Enforce collinearity of residual b − (A − σi I )xim to b − Ax m I We do no longer perform GMRES(m) on the shifted system Theorem [F. 98]: If A is positive real and σi < 0, this approach works and kb − (A − σi I )xim k2 ≤ kb − Ax m k. A similar approach yields shifted BiCGStab [F. 03]. Beyond Standard Krylov Subspace Methods – Desirable Features for Scientific Software –, Andreas Frommer 17/43 Introduction Shifted linear systems Flexible Methods Indefinite inner products Conclusions Algorithms Issues for Software Applications AI Issues for Software Which multshift methods to implement? I CG, QMR I restarted FOM I optional: restarted GMRES, BiCGStab A warning: I There are different algorithmic variants . . . I . . . which have different numerical stability I Example: van den Eshof and Sleijpen 04 develop more stable multishift CG Beyond Standard Krylov Subspace Methods – Desirable Features for Scientific Software –, Andreas Frommer 18/43 Introduction Shifted linear systems Flexible Methods Indefinite inner products Conclusions Algorithms Issues for Software Applications AI Families of shifted linear systems arise for I simple parameter dependencies I Tichonov-Philips regularization I Computation of the action of a matrix function Beyond Standard Krylov Subspace Methods – Desirable Features for Scientific Software –, Andreas Frommer 19/43 Introduction Shifted linear systems Flexible Methods Indefinite inner products Conclusions Algorithms Issues for Software Applications AI Simple parameter dependencies Example: Wilson fermion matrix M = κ1 I − D in QCD. I κ is a mass parameter Beyond Standard Krylov Subspace Methods – Desirable Features for Scientific Software –, Andreas Frommer 20/43 Introduction Shifted linear systems Flexible Methods Indefinite inner products Conclusions Algorithms Issues for Software Applications AI Tikhonov-Philips regularization Ah : discretization of a compact operator A between Hilbert spaces. A is not invertible. As h gets smaller, solution xh in Ah x = b is very bad if b is noisy. Tikhonov-Philips regularization: Constrain kxh k ≤ α. Resulting system: (A∗h Ah + αI )xh (α) = A∗h b. Note: Right choice for α found a posteriori. Beyond Standard Krylov Subspace Methods – Desirable Features for Scientific Software –, Andreas Frommer 21/43 Introduction Shifted linear systems Flexible Methods Indefinite inner products Conclusions Algorithms Issues for Software Applications AI An example 0.3 0.2 0.1 0 0 1 Beyond Standard Krylov Subspace Methods – Desirable Features for Scientific Software –, Andreas Frommer 22/43 Introduction Shifted linear systems Flexible Methods Indefinite inner products Conclusions Algorithms Issues for Software Applications AI Matrix functions Wanted: f (A)b, where f : C → C, A large and sparse Examples: I Exponential integrators in ODEs, PDEs: exp(tA)b I QCD: sign(M)b Note: f (A)b = p(A)b, where polynomial p interpolates f on spec(A). Back to Krylov subspaces! Beyond Standard Krylov Subspace Methods – Desirable Features for Scientific Software –, Andreas Frommer 23/43 Introduction Shifted linear systems Flexible Methods Indefinite inner products Conclusions Algorithms Issues for Software Applications AI Short recurrences for f (A)b Stable short recurrence methods use rational approximations f (t) ≈ p X i=1 p X ωi , ⇒ f (A)b ≈ ωi (A − σi I )−1 b. t − σi i=1 20 1 0.8 15 0.6 10 0.4 5 0.2 0 0 −0.2 −5 −0.4 −10 −0.6 −15 −20 −10 −0.8 −8 −6 −4 −2 0 2 4 6 −1 8 −10 6 −10 4 −10 2 −10 0 −10 Beyond Standard Krylov Subspace Methods – Desirable Features for Scientific Software –, Andreas Frommer −2 −10 24/43 Introduction Shifted linear systems Flexible Methods Indefinite inner products Conclusions Algorithms Software issues Applications AI Convergence speeds “Theorem”:(Most) Krylov subspace methods are direct, i.e. ∗ ∃ n∗ = n∗ (A, b) ≤ n s.t. x n = b. Related polynomial approximation problem: k := kp (A)bk is (P) given k, find pk s.t. pk (0) = 1 and cA,b ∗ k smallest. k ≤ c k kbk with c k = inf Note: cA,b ∗ deg p≤k kp(A)k∗ . A A Beyond Standard Krylov Subspace Methods – Desirable Features for Scientific Software –, Andreas Frommer 25/43 Introduction Shifted linear systems Flexible Methods Indefinite inner products Conclusions Algorithms Software issues Applications AI Theorem: a) Let A be Hermitian and positive definite, spec(A) ⊂ [α, β], κ = β α . Then for the CG iterates kx k − A−1 bkA ≤ c k kx 0 − A−1 bkA , √ k κ−1 ck = √ κ+1 α F(A) b) Let F (A) = {hx, Axi : kxk2 = 1} and assume α ≤ F (A) ≤ β. Then for the GMRES iterates kb − Ax k k2 ≤ (1 − β α )kb − Ax k−1 |. β Beyond Standard Krylov Subspace Methods – Desirable Features for Scientific Software –, Andreas Frommer 26/43 Introduction Shifted linear systems Flexible Methods Indefinite inner products Conclusions Algorithms Software issues Applications AI Preconditioning Idea: Ax = b → M1−1 AM2−1 x̃ = M1−1 b. I Systems with M1 , M2 easy to solve I one solve required in each iteration I spec(M −1 AM −∗ ) ⊆ [1 − , 1 + ] or 1 1 F (M1−1 AM2 −1) ⊆ D(1, ). Standard: I ILU and modified ILU preconditioners I Gauss-Seidel or Jacobi preconditioners Less standard: I algebraic multilevel preconditioners I domain decomposition preconditioners I recursive preconditioning Beyond Standard Krylov Subspace Methods – Desirable Features for Scientific Software –, Andreas Frommer 27/43 Introduction Shifted linear systems Flexible Methods Indefinite inner products Conclusions Algorithms Software issues Applications AI Variable preconditioners For simlpicity: Right preconditioing, i.e. M2 = I . Recursive preconditioning: Use a Krylov subspace method to (quite inaccurately) solve Ay = u k Example: GMRESR Domain decomposition preconditioners: Approximately solve smaller subsystems Pi∗ APi and combine individual solutions. Common feature: Preconditioner M depends on iteration number k or on current vector u k Problem: We do not build the Krylov subspace Kk (M −1 A, b). Solution: Flexible methods Beyond Standard Krylov Subspace Methods – Desirable Features for Scientific Software –, Andreas Frommer 28/43 Introduction Shifted linear systems Flexible Methods Indefinite inner products Conclusions Algorithms Software issues Applications AI Examples Recursive preconditioning for (I + Γ5 sign(Γ5 M))x = b in QCD 0 10 SUMR relSUMR relGMRESR(SUMR) −1 10 −2 norm of residual 10 −3 10 −4 10 −5 10 −6 10 −7 10 0 2 4 6 multiplications with Q 8 10 4 x 10 Beyond Standard Krylov Subspace Methods – Desirable Features for Scientific Software –, Andreas Frommer 29/43 Introduction Shifted linear systems Flexible Methods Indefinite inner products Conclusions Algorithms Software issues Applications AI Available methods: I Inexact CG: Golub and Ye 99 I Flexible CG: Notay 00 I Flexible GMRES: Saad 93 I recursive GMRES: Van der Vorst and Vuik 94 I FQMR: Szyld and Vogel, 02 To do: I systematic comparison of competing algorithms I investigate numerical stability Example: Recent results by Roloznik, Gutknecht et al., 08, 09, 10 show that GCR is intrinsically less stable than GMRES. Beyond Standard Krylov Subspace Methods – Desirable Features for Scientific Software –, Andreas Frommer 30/43 Introduction Shifted linear systems Flexible Methods Indefinite inner products Conclusions Algorithms Software issues Applications AI Flexible methods allow for I non-stationary iterative methods as preconditioners including: multigrid, domain decomposition I inexact multiplications with matrix A Example: A = sign(Q) I relaxation strategies for accuracy of multiplication with A (Simoncini, Szyld 03, Sleijpen, van den Eshof 03) Beyond Standard Krylov Subspace Methods – Desirable Features for Scientific Software –, Andreas Frommer 31/43 Introduction Shifted linear systems Flexible Methods Indefinite inner products Conclusions Algorithms Software issues Applications AI Wilson inverter on QPACE Features: I Nearest neighbour couplings on 4d grid I QPACE: no 1 Green TOP 500 parallel computer absed on the CELL We use: I FGMRES with restarts I Multiplicative Schwarz as a domain decomposition method with red black ordering of domains I On domain O(10) steps of GMRES Collaborative Research Center “Hadron Physics from Lattice QCD” (Regensburg, Wuppertal) funded by DFG. Beyond Standard Krylov Subspace Methods – Desirable Features for Scientific Software –, Andreas Frommer 32/43 Introduction Shifted linear systems Flexible Methods Indefinite inner products Conclusions Software issues Deflation Algorithms Software issues Applications AI Non-trivial symmetries In quite some important applications, A is non-Hermitian, but AJ = J ∗ A∗ where J is I simple I hermitian I indefinite Of course AJ is Hermitian, but usually then very indefinite. Beyond Standard Krylov Subspace Methods – Desirable Features for Scientific Software –, Andreas Frommer 33/43 Introduction Shifted linear systems Flexible Methods Indefinite inner products Conclusions Software issues Deflation Algorithms Software issues Applications AI Examples: I I 0 ··· 0 0 ··· 1 Symplectic systems, J = . .. .. .. . . 1 0 ··· I 0 Lorentz symmetry: J = 0 −I 1 0 .. . 0 Note: J induces an indefinite inner product hx, y iJ := hx, Jy i in which A is self-adjoint. CG with this inner product ≡ BiCG with special choice for r̃ 0 Beyond Standard Krylov Subspace Methods – Desirable Features for Scientific Software –, Andreas Frommer 34/43 Introduction Shifted linear systems Flexible Methods Indefinite inner products Conclusions Software issues Deflation Algorithms Software issues Applications I Nonsymmetric Lanczos can exhibit premature breakdowns I can mostly be cured using look-ahead (Nachtigal, Freund 92) I BiCG can have additional sources of breakdown AI Suggestion: I Implement QMR with look-ahead Lanczos for J-symmetric systems I Incorporate J-symmetric preconditioning Beyond Standard Krylov Subspace Methods – Desirable Features for Scientific Software –, Andreas Frommer 35/43 Introduction Shifted linear systems Flexible Methods Indefinite inner products Conclusions Software issues Deflation Algorithms Software issues Applications AI Projections Let A be hermitian and positive definite. Assume we know ` smallest eigenpairs (λi , qi ) of A. Definitions: I Q = [q1 | . . . |q` ] matrix of orthonormal eigenvectors I Λ = diag(λ1 , . . . , λ` ) I P = QQ ∗ orthogonal projector on rangeQ We have Ax = b ⇔ x = x1 + x2 where Ax1 = Pb, Ax2 = (I − P)b. Beyond Standard Krylov Subspace Methods – Desirable Features for Scientific Software –, Andreas Frommer 36/43 Introduction Shifted linear systems Flexible Methods Indefinite inner products Conclusions Software issues Deflation Algorithms Software issues Applications AI Ax = b ⇔ x = x1 + x2 where Ax1 = Pb, Ax2 = (I − P)b. Crucial: I x1 = QΛ−1 Pb I K (A, (I − P)b) = K ((I − P)A(I − P)∗ , b) Consequence: Get x2 with CG for A and (I − P)b, effective n condition number is reduced from λλn1 to λλ`+1 A non-hermitian: Idea works similarly, but with oblique projector built form left and right eigenvectors. Beyond Standard Krylov Subspace Methods – Desirable Features for Scientific Software –, Andreas Frommer 37/43 Introduction Shifted linear systems Flexible Methods Indefinite inner products Conclusions Software issues Deflation Algorithms Software issues Applications AI Deflation on the fly? Facts: I The Krylov subspace Kk (A, r 0 ) can be used to obtain approximation to eigenpairs (Ritz vectors, harmonic Ritz vectors) I Full length methods (like CG, GMRES) can be viewed as gathering eigenvalue information implicitly and using it to deflate ⇒ superlinear convergence I If we restart (example: GMRES(m)), all this information is forgotten and lost ⇒ risk of stagnation Idea for restarted GMRES: Extract eigenpair information from one cycle and used it in the next. Beyond Standard Krylov Subspace Methods – Desirable Features for Scientific Software –, Andreas Frommer 38/43 Introduction Shifted linear systems Flexible Methods Indefinite inner products Conclusions Software issues Deflation Algorithms Software issues Applications AI Many approaches suggested in literature. I Erhel, Burrage, Pohl 06 I Wu, Simon 00 I Morgan 95-09 I Sorenson 92 I Eiermann, Ernst, Schneider 00 I . . . and many more Today, a standard seems to emerge: I deflation via augmentation of the standard Krylov subspace I augmentation s.t. subspace is still a Krylov subspace containing subspaces for each of the approximate eigenvectors I harmonic Ritz values for GMRES, Ritz values for FOM Beyond Standard Krylov Subspace Methods – Desirable Features for Scientific Software –, Andreas Frommer 39/43 Introduction Shifted linear systems Flexible Methods Indefinite inner products Conclusions Software issues Deflation Algorithms Software issues Applications AI Software issues I Implement GMRES-DR à la Morgan I optional: develop and implement FOM-DR I combination with flexible methods (Pinel 10) I combination with multishift methods (based on Schäfer 08) Beyond Standard Krylov Subspace Methods – Desirable Features for Scientific Software –, Andreas Frommer 40/43 Introduction Shifted linear systems Flexible Methods Indefinite inner products Conclusions Software issues Deflation Algorithms Software issues Applications AI Wilson solver on QPACE, again 800 10 ’res_12_8_24_5’ ’res_nodr’ 600 0.01 500 ITERATIONS 0.1 0.001 0.0001 400 300 1e-05 200 1e-06 100 1e-07 0 0.135 1e-08 0 200 400 600 800 1000 1200 1400 sap-fgmres-dr-mx(32,10) 700 1 0.1355 0.136 0.1365 0.137 kappa 0.1375 0.138 1600 Beyond Standard Krylov Subspace Methods – Desirable Features for Scientific Software –, Andreas Frommer 41/43 Introduction Shifted linear systems Flexible Methods Indefinite inner products Conclusions Software issues Deflation Algorithms Software issues Applications AI Seismic exploration application by X.Pinel, CERFACS and TOTAL Beyond Standard Krylov Subspace Methods – Desirable Features for Scientific Software –, Andreas Frommer 42/43 Introduction Shifted linear systems Flexible Methods Indefinite inner products Conclusions AI Conclusions I I Parallelization is in the MVM and the preconditioning 4 enhancements: I I I I shifted flexible J-symmetry deflation I Priority for CG, GMRES(m) . . . I . . . then QMR, BiCGStab I before you implement professionally, study stability in theory and practice I enhancements would be helpful in many areas of scientific computing Beyond Standard Krylov Subspace Methods – Desirable Features for Scientific Software –, Andreas Frommer 43/43