Solution of eigenvalue problems. Subspace iteration, The symmetric Lanczos algorithm. Harmonic Ritz values, Jacobi-Davidson s method

Transcription

1 Solution of eigenvalue problems Introduction motivation Projection methods for eigenvalue problems Subspace iteration, The symmetric Lanczos algorithm Nonsymmetric Lanczos procedure; Implicit restarts Harmonic Ritz values, Jacobi-Davidson s method

2 Background. Origins of Eigenvalue Problems Structural Engineering [Ku = λm u] (Goal: frequency response) Electronic structure calculations [Schrödinger equation..] Stability analysis [e.g., electrical networks, mechanical system,..] Bifurcation analysis [e.g., in fluid flow] Large eigenvalue problems in quantum chemistry use up biggest portion of the time in supercomputer centers 16-2 eigbackg

3 Background. New applications in data analytics Machine learning problems often require a (partial) Singular Value Decomposition - Somewhat different issues in this case: Very large matrices, update the SVD Compute dominant singular values/vectors Many problems of approximating a matrix (or a tensor) by one of lower rank (Dimension reduction,...) But: Methods for computing SVD often based on those for standard eigenvalue problems 16-3 eigbackg

4 Background. The Problem (s) Standard eigenvalue problem: Ax = λx Often: A is symmetric real (or Hermitian complex) Generalized problem Ax = λbx Often: B is symmetric positive definite, A is symmetric or nonsymmetric Quadratic problems: (A + λb + λ 2 C)u = 0 Nonlinear eigenvalue problems (NEVP) [ A 0 + λb 0 + n i=1 f i (λ)a i ] u = eigbackg

5 General form of NEVP A(λ)x = 0 Nonlinear eigenvector problems: [A + λb + F (u 1, u 2,, u k )]u = 0 What to compute: A few λ i s with smallest or largest real parts; All λ i s in a certain region of C; A few of the dominant eigenvalues; All λ i s (rare) eigbackg

6 Large eigenvalue problems in applications Some applications require the computation of a large number of eigenvalues and vectors of very large matrices. Density Functional Theory in electronic structure calculations: ground states Excited states involve transitions and invariably lead to much more complex computations. Large matrices, *many* eigen-pairs to compute 16-6 eigbackg

7 Background: The main tools Projection process: (a) Build a good subspace K = span(v ); (b) get approximate eigenpairs by a Rayleigh-Ritz process: λ, ũ K satisfy: (A λi)ũ K V H (A λi)v y = 0 λ = Ritz value, ũ = V y = Ritz vector Two common choices for K: 1) Power subspace K = span{a k X 0 }; or span{p k (A)X 0 }; 2) Krylov subspace K = span{v, Av,, A k 1 v} 16-7 eigbackg

8 Background. The main tools (cont) Shift-and-invert: If we want eigenvalues near σ, replace A by (A σi) 1. Example: power method: v j = Av j 1 /scaling replaced by v j = (A σi) 1 v j 1 scaling Works well for computing a few eigenvalues near σ/ Used in commercial package NASTRAN (for decades!) Requires factoring (A σi) (or (A σb) in generalized case.) But convergence will be much faster. A solve each time - Factorization done once (ideally) eigbackg

9 Background. The main tools (cont) Deflation: Once eigenvectors converge remove them from the picture Restarting Strategies : Restart projection process by using information gathered in previous steps ALL available methods use some combination of these ingredients. [e.g. ARPACK: Arnoldi/Lanczos + implicit restarts + shift-andinvert (option).] 16-9 eigbackg

10 Current state-of-the art in eigensolvers Eigenvalues at one end of the spectrum: Subspace iteration + filtering [e.g. FEAST, Cheb,...] Lanczos+variants (no restart, thick restart, implicit restart, Davidson,..), e.g., ARPACK code, PRIMME. Block Algorithms [Block Lanczos, TraceMin, LOBPCG, SlepSc,...] + Many others - more or less related to above Interior eigenvalue problems (middle of spectrum): Combine shift-and-invert + Lanczos/block Lanczos. e.g., NASTRAN Rational filtering [FEAST, Sakurai et al.,.. ] Used in, eigbackg

11 Projection Methods for Eigenvalue Problems General formulation: Projection method onto K orthogonal to L Given: Two subspaces K and L of same dimension. Find: λ, ũ such that λ C, ũ K; ( λi A)ũ L Two types of methods: Orthogonal projection methods: situation when L = K. Oblique projection methods: When L K eig1

12 Rayleigh-Ritz projection Given: a subspace X known to contain good approximations to eigenvectors of A. Question: How to extract good approximations to eigenvalues/ eigenvectors from this subspace? Answer: Rayleigh Ritz process. Let Q = [q 1,..., q m ] an orthonormal basis of X. Then write an approximation in the form ũ = Qy and obtain y by writing Q H (A λi)ũ = 0 Q H AQy = λy eig1

13 Procedure: 1. Obtain an orthonormal basis of X 2. Compute C = Q H AQ (an m m matrix) 3. Obtain Schur factorization of C, C = Y RY H 4. Compute Ũ = QY Property: if X is (exactly) invariant, then procedure will yield exact eigenvalues and eigenvectors. Proof: Since X is invariant, (A λi)u = Qz for a certain z. Q H Qz = 0 implies z = 0 and therefore (A λi)u = 0. Can use this procedure in conjunction with the subspace obtained from subspace iteration algorithm eig1

14 Subspace Iteration Original idea: projection technique onto a subspace if the form Y = A k X In practice: Replace A k by suitable polynomial [Chebyshev] Advantages: Disadvantage: Slow. Easy to implement (in symmetric case); Easy to analyze; Often used with polynomial acceleration: A k X replaced by C k (A)X. Typically C k = Chebyshev polynomial eig1

15 Algorithm: Subspace Iteration with Projection 1. Start: Choose an initial system of vectors X = [x 0,..., x m ] and an initial polynomial C k. 2. Iterate: Until convergence do: (a) Compute Ẑ = C k (A)X old. (b) Orthonormalize Ẑ into Z. (c) Compute B = Z H AZ and use the QR algorithm to compute the Schur vectors Y = [y 1,..., y m ] of B. (d) Compute X new = ZY. (e) Test for convergence. If satisfied stop. Else select a new polynomial C k and continue.

16 THEOREM: Let S 0 = span{x 1, x 2,..., x m } and assume that S 0 is such that the vectors {P x i } i=1,...,m are linearly independent where P is the spectral projector associated with λ 1,..., λ m. Let P k the orthogonal projector onto the subspace S k = span{x k }. Then for each eigenvector u i of A, i = 1,..., m, there exists a unique vector s i in the subspace S 0 such that P s i = u i. Moreover, the following inequality is satisfied ( λ m+1 (I P k )u i 2 u i s i 2 λ i where ɛ k tends to zero as k tends to infinity. k k) + ɛ, (1) eig1

17 Krylov subspace methods Principle: Projection methods on Krylov subspaces, i.e., on K m (A, v 1 ) = span{v 1, Av 1,, A m 1 v 1 } probably the most important class of projection methods [for linear systems and for eigenvalue problems] many variants exist depending on the subspace L. Properties of K m. Let µ = deg. of minimal polynom. of v. Then, K m = {p(a)v p = polynomial of degree m 1} K m = K µ for all m µ. Moreover, K µ is invariant under A. dim(k m ) = m iff µ m eig1

18 Arnoldi s Algorithm Goal: to compute an orthogonal basis of K m. Input: Initial vector v 1, with v 1 2 = 1 and m. ALGORITHM : 1 Arnoldi s procedure For j = 1,..., m do Compute w := Av j For i = 1,..., j, do { hi,j := (w, v i ) w := w h i,j v i h j+1,j = w 2 ; v j+1 = w/h j+1,j End eig1

19 Result of Arnoldi s algorithm Let H m = x x x x x x x x x x x x x x x x x x x x ; H m = H m (1 : m, 1 : m) 1. V m = [v 1, v 2,..., v m ] orthonormal basis of K m. 2. AV m = V m+1 H m = V m H m + h m+1,m v m+1 e T m 3. V T m AV m = H m H m last row eig1

20 Appliaction to eigenvalue problems Write approximate eigenvector as ũ = V m y + Galerkin condition (A λi)v m y K m V H m (A λi)v m y = 0 Approximate eigenvalues are eigenvalues of H m H m y j = λ j y j Associated approximate eigenvectors are ũ j = V m y j Typically a few of the outermost eigenvalues will converge first eig1

21 Restarted Arnoldi In practice: Memory requirement of algorithm implies restarting is necessary Restarted Arnoldi for computing rightmost eigenpair: ALGORITHM : 2 Restarted Arnoldi 1. Start: Choose an initial vector v 1 and a dimension m. 2. Iterate: Perform m steps of Arnoldi s algorithm. 3. Restart: Compute the approximate eigenvector u (m) 1 4. associated with the rightmost eigenvalue λ (m) If satisfied stop, else set v 1 u (m) 1 and goto eig1

22 Example: Small Markov Chain matrix [ Mark(10), dimension = 55]. Restarted Arnoldi procedure for computing the eigenvector associated with the eigenvalue with algebraically largest real part. We use m = 10. m R(λ) I(λ) Res. Norm D D D D D D D D D D eig1

23 Restarted Arnoldi (cont.) Can be generalized to more than *one* eigenvector : p v (new) 1 = ρ i u (m) i i=1 However: often does not work well (hard to find good coefficients ρ i s) Alternative : compute eigenvectors (actually Schur vectors) one at a time. Implicit deflation eig1

24 Deflation Very useful in practice. Different forms: locking (subspace iteration), selective orthogonalization (Lanczos), Schur deflation,... A little background Consider Schur canonical form A = URU H where U is a (complex) upper triangular matrix. Vector columns u 1,..., u n called Schur vectors. Note: Schur vectors depend on each other, and on the order of the eigenvalues eig1

25 Wiedlandt Deflation: Assume we have computed a right eigenpair λ 1, u 1. Wielandt deflation considers eigenvalues of Note: A 1 = A σu 1 v H Λ(A 1 ) = {λ 1 σ, λ 2,..., λ n } Wielandt deflation preserves u 1 as an eigenvector as well all the left eigenvectors not associated with λ 1. An interesting choice for v is to take simply v = u 1. In this case Wielandt deflation preserves Schur vectors as well. Can apply above procedure successively eig1

26 ALGORITHM : 3 Explicit Deflation 1. A 0 = A 2. For j = 0... µ 1 Do: 3. Compute a dominant eigenvector of A j 4. Define A j+1 = A j σ j u j u H j 5. End Computed u 1, u 2.,.. form a set of Schur vectors for A. Alternative: implicit deflation (within a procedure such as Arnoldi) eig1

27 Deflated Arnoldi When first eigenvector converges, put it in 1st column of V m = [v 1, v 2,..., v m ]. Arnoldi will now start at column 2, orthogonaling still against v 1,..., v j at step j. Accumulate each new converged eigenvector in columns 2, 3,... [ locked set of eigenvectors.] [ ] active {}}{ Thus, for k = 2: V m = v } 1 {{, v } 2, v 3,..., v m Locked H m =

28 Similar techniques in Subspace iteration [G. Stewart s SRRIT] Example: Matrix Mark(10) small Markov chain matrix (N = 55). First eigenpair by iterative Arnoldi with m = 10. m Re(λ) Im(λ) Res. Norm D D D D D D D D D D eig1

29 Computing the next 2 eigenvalues of Mark(10). Eig. Mat-Vec s Re(λ) Im(λ) Res. Norm D D D D D D D D eig1

30 Hermitian case: The Lanczos Algorithm The Hessenberg matrix becomes tridiagonal : A = A H and V H m AV m = H m H m = H H m We can write H m = α 1 β 2 β 2 α 2 β 3 β 3 α 3 β β m α m (2) Consequence: three term recurrence β j+1 v j+1 = Av j α j v j β j v j eig1

31 ALGORITHM : 4 Lanczos 1. Choose v 1 of norm unity. Set β 1 0, v For j = 1, 2,..., m Do: 3. w j := Av j β j v j 1 4. α j := (w j, v j ) 5. w j := w j α j v j 6. β j+1 := w j 2. If β j+1 = 0 then Stop 7. v j+1 := w j /β j+1 8. EndDo Hermitian matrix + Arnoldi Hermitian Lanczos In theory v i s defined by 3-term recurrence are orthogonal. However: in practice severe loss of orthogonality; eig1

32 Lanczos with reorthogonalization Observation [Paige, 1981]: Loss of orthogonality starts suddenly, when the first eigenpair converges. It indicates loss of linear indedependence of the v i s. When orthogonality is lost, then several copies of the same eigenvalue start appearing. Full reorthogonalization reorthogonalize v j+1 against all previous v i s every time. Partial reorthogonalization reorthogonalize v j+1 against all previous v i s only when needed [Parlett & Simon] Selective reorthogonalization reorthogonalize v j+1 against computed eigenvectors [Parlett & Scott] No reorthogonalization Do not reorthogonalize - but take measures to deal with spurious eigenvalues. [Cullum & Willoughby] eig1

33 Partial reorthogonalization Partial reorthogonalization: reorthogonalize only when deemed necessary. Main question is when? Uses an inexpensive recurrence relation Work done in the 80 s [Parlett, Simon, and co-workers] + more recent work [Larsen, 98] Package: PROPACK [Larsen] V 1: 2001, most recent: V 2.1 (Apr. 05) Often, need for reorthogonalization not too strong eig1

34 The Lanczos Algorithm in the Hermitian Case Assume eigenvalues sorted increasingly λ 1 λ 2 λ n Orthogonal projection method onto K m ; To derive error bounds, use the Courant characterization λ 1 = λ j = min u K, u 0 { min u K, u 0 u ũ 1,...,ũ j 1 (Au, u) (u, u) = (Aũ 1, ũ 1 ) (ũ 1, ũ 1 ) (Au, u) (u, u) = (Aũ j, ũ j ) (ũ j, ũ j ) eig1

35 Bounds for λ 1 easy to find similar to linear systems. Ritz values approximate eigenvalues of A inside out: λ 1 λ 2 λ n 1 λ n λ 1 λ2 λ n 1 λn eig1

36 A-priori error bounds Theorem [Kaniel, 1966]: 0 λ (m) 1 λ 1 (λ N λ 1 ) [ tan (v1, u 1 ) T m 1 (1 + 2γ 1 ) where γ 1 = λ 2 λ 1 λ N λ 2 ; and (v 1, u 1 ) = angle between v 1 and u 1. + results for other eigenvalues. [Kaniel, Paige, YS] Theorem 0 λ (m) i λ i (λ N λ 1 ) where γ i = λ i+1 λ i λ N λ i+1, κ (m) i [ = j<i κ (m) i tan (v i, u i ) ] 2 T m i (1 + 2γ i ) λ (m) j λ N λ (m) j λ i eig1 ] 2

37 The Lanczos biorthogonalization (A H A) ALGORITHM : 5 Lanczos bi-orthogonalization 1. Choose two vectors v 1, w 1 such that (v 1, w 1 ) = Set β 1 = δ 1 0, w 0 = v For j = 1, 2,..., m Do: 4. α j = (Av j, w j ) 5. ˆv j+1 = Av j α j v j β j v j 1 6. ŵ j+1 = A T w j α j w j δ j w j 1 7. δ j+1 = (ˆv j+1, ŵ j+1 ) 1/2. If δ j+1 = 0 Stop 8. β j+1 = (ˆv j+1, ŵ j+1 )/δ j+1 9. w j+1 = ŵ j+1 /β j v j+1 = ˆv j+1 /δ j+1 11.EndDo eig1

38 Builds a pair of biorthogonal bases for the two subspaces K m (A, v 1 ) and K m (A H, w 1 ) Many choices for δ j+1, β j+1 in lines 7 and 8. Only constraint: δ j+1 β j+1 = (ˆv j+1, ŵ j+1 ) Let T m = α 1 β 2 δ 2 α 2 β 3... δ m 1 α m 1 β m δ m α m. v i K m (A, v 1 ) and w j K m (A T, w 1 ) eig1

39 If the algorithm does not break down before step m, then the vectors v i, i = 1,..., m, and w j, j = 1,..., m, are biorthogonal, i.e., (v j, w i ) = δ ij 1 i, j m. Moreover, {v i } i=1,2,...,m is a basis of K m (A, v 1 ) and {w i } i=1,2,...,m is a basis of K m (A H, w 1 ) and AV m = V m T m + δ m+1 v m+1 e H m, A H W m = W m T H m + β m+1 w m+1 e H m, W H m AV m = T m eig1

40 If θ j, y j, z j are, respectively an eigenvalue of T m, with associated right and left eigenvectors y j and z j respectively, then corresponding approximations for A are Ritz value Right Ritz vector Left Ritz vector θ j V m y j W m z j [Note: terminology is abused slightly - Ritz values and vectors normally refer to Hermitian cases.] eig1

41 Advantages and disadvantages Advantages: Nice three-term recurrence requires little storage in theory. Computes left and a right eigenvectors at the same time Disadvantages: Algorithm can break down or nearly break down. Convergence not too well understood. Erratic behavior Not easy to take advantage of the tridiagonal form of T m eig1

42 Look-ahead Lanczos Algorithm breaks down when: (ˆv j+1, ŵ j+1 ) = 0 Three distinct situations. lucky breakdown when either ˆv j+1 or ŵ j+1 is zero. In this case, eigenvalues of T m are eigenvalues of A. (ˆv j+1, ŵ j+1 ) = 0 but of ˆv j+1 0, ŵ j+1 0 serious breakdown. Often possible to bypass the step (+ a few more) and continue the algorithm. If this is not possible then we get an Incurable break-down. [very rare] eig1

43 Look-ahead Lanczos algorithms deal with the second case. See Parlett 80, Freund and Nachtigal Main idea: when break-down occurs, skip the computation of v j+1, w j+1 and define v j+2, w j+2 from v j, w j. For example by orthogonalizing A 2 v j... Can define v j+1 somewhat arbitrarily as v j+1 = Av j. Similarly for w j+1. Drawbacks: (1) projected problem no longer tridiagonal (2) difficult to know what constitutes near-breakdown eig1