Nonlinear Eigenvalue Problems: An Introduction

Transcription

1 Nonlinear Eigenvalue Problems: An Introduction Cedric Effenberger Seminar for Applied Mathematics ETH Zurich Pro*Doc Workshop Disentis, August 18 21, 2010 Cedric Effenberger (SAM, ETHZ) NLEVPs: An Introduction Pro*Doc Disentis 1 / 37

2 Cedric Effenberger (SAM, ETHZ) NLEVPs: An Introduction Pro*Doc Disentis 2 / 37 Overview 1 Definition 2 Applications 3 Solution 4 Solution of polynomial / rational EVPs

3 Cedric Effenberger (SAM, ETHZ) NLEVPs: An Introduction Pro*Doc Disentis 3 / 37 Linear Eigenvalue Problems Standard eigenvalue problem Given A C n n, we seek eigenvalues λ C, and eigenvectors x C n \ {0} such that Ax = λx. Generalized eigenvalue problem Given A, B C n n, we seek eigenvalues λ C, and eigenvectors x C n \ {0} such that Ax = λbx.

4 Cedric Effenberger (SAM, ETHZ) NLEVPs: An Introduction Pro*Doc Disentis 4 / 37 Generalization to nonlinear EVPs Both problems can be reformulated as T (λ)x = 0, where T (λ) = { A λi, A λb, for the standard EVP, for the generalized EVP. T (λ) = λ d A d + + λa 1 + A 0 polynomial EVP T (λ) = f 1 (λ)a f r (λ)a r nonlinear EVP A j C n n constant matrices, f j : D C analytic, D C open, connected

5 Cedric Effenberger (SAM, ETHZ) NLEVPs: An Introduction Pro*Doc Disentis 5 / 37 Free vibrations of mechanical systems with damping Equations of motion Mü + C u + Ku = 0 M: mass matrix C: damping matrix K: stiffness matrix System of second-order, linear, homogeneous ODEs Eigenfrequencies ω of the mechanical system are eigenvalues of Quadratic eigenvalue problem (ω 2 M + ωc + K)u = 0.

6 Cedric Effenberger (SAM, ETHZ) NLEVPs: An Introduction Pro*Doc Disentis 6 / 37 Free vibrations of string with elastically attached mass vibrating string u = λu in Ω = (0, 1) left end clamped u(0) = 0 mass elastically attached to right end u (1) + αλ λ α u(1) = 0 λ-dependent boundary condition! Discretization with n piecewise linear FE leads to [ K + αλ λ α e nen T ] λm x = 0.

7 Cedric Effenberger (SAM, ETHZ) NLEVPs: An Introduction Pro*Doc Disentis 7 / 37 Delay eigenvalue problems Linear delay differential equation with r discrete delays u(t) = A 0 u(t) + A 1 u(t τ 1 ) + + A r u(t τ r ) models influences which take effect only after some time famous example: Hot shower problem stability analysis involves nonlinear EVPs Delay eigenvalue problem ( λi + A 0 + e τ 1λ A e τr λ A r )x = 0

8 Cedric Effenberger (SAM, ETHZ) NLEVPs: An Introduction Pro*Doc Disentis 8 / 37 Electronic bandstructure computations photonic crystal: lattice of mixed dielectric media control electromagnetic waves by designing the crystal such that it inhibits their propagation complete photonic band gap: frequency range with no propagation of electromagnetic waves of any polarization travelling in any direction

9 Cedric Effenberger (SAM, ETHZ) NLEVPs: An Introduction Pro*Doc Disentis 9 / 37 Electronic bandstructure computations II 2D crystal: periodic in x- and y-direction; homogeneous in z-direction consider only electromagnetic waves propagating in the xy-plane

10 Cedric Effenberger (SAM, ETHZ) NLEVPs: An Introduction Pro*Doc Disentis 10 / 37 Electronic bandstructure computations III Time-harmonic modes of electromagnetic wave (E, H) can be decomposed into transverse electric (TE) polarized modes (E x, E y, 0, 0, 0, H z ), transverse magnetic (TM) polarized modes (0, 0, E z, H x, H y, 0). For TM polarized modes, the macroscopic Maxwell Equations reduce to a scalar equation for E z, r: spatial variable ω: frequency ε: relative permittivity E z = ω 2 ε(r, ω)e z. ω-dependent material parameter!

11 Cedric Effenberger (SAM, ETHZ) NLEVPs: An Introduction Pro*Doc Disentis 11 / 37 Electronic bandstructure computations IV Bloch ansatz: E z = e ik r u(r) ( + ik) ( + ik)u(r) = ω 2 ε(r, ω)u(r) k: wave vector u: periodic function on lattice assumption: lattice consists of 2 materials, one of which is air { ε 1 = 1, r Ω 1 Ω = Ω 1 Ω 2, ε(r, ω) = ε 2 (ω), r Ω 2 discretization using discontinuous Galerkin with p-enhancement (Engström & Wang, 2010) [ G ω 2 M 1 ω 2 ε 2 (ω)m 2 ] u = 0

12 Cedric Effenberger (SAM, ETHZ) NLEVPs: An Introduction Pro*Doc Disentis 12 / 37 Electronic bandstructure computations V nonlinearity caused by ω-dependency of ε 2 popular model for permittivity: Lorentz model K ξ k ε 2 (ω) = α + η k ω 2 iγ k ω k=1 rational EVP ω 2π

13 Cedric Effenberger (SAM, ETHZ) NLEVPs: An Introduction Pro*Doc Disentis 13 / 37 Spectrum of nonlinear EVPs Definition Let T : D C n n be holomorphic. We call ρ(t ) := {λ D : T (λ) is invertible} the resolvent set of T, σ(t ) := D \ ρ(t ) the spectrum of T. Theorem Either the resolvent set of T is empty or the spectrum of T consists of isolated eigenvalues. The number of eigenvalues in σ(t ) may be infinite!

14 Cedric Effenberger (SAM, ETHZ) NLEVPs: An Introduction Pro*Doc Disentis 14 / 37 Newton-based methods Two possibilities to apply Newton s method: 1 det T (λ) = 0 (scalar equation in C) 2 T (λ)x = 0 (vector equation in C n ) We focus on the second case. x, λ n + 1 unknowns T (λ)x = 0 n constraints We have to add one additional constraint. v H x = 1

15 Cedric Effenberger (SAM, ETHZ) NLEVPs: An Introduction Pro*Doc Disentis 15 / 37 Nonlinear inverse iteration F (x, λ) := [ ] T (λ)x v H x 1 Application of Newton s method ] [ ] [ˆx 0 =! x T (λ)ˆx + (ˆλ λ)t F (x, λ) + DF (x, λ) = (λ)x ˆλ λ v H ˆx 1 yields ˆx = (ˆλ λ)t (λ) 1 T (λ)x 1 ˆλ = λ v H T (λ) 1 T (λ)x.

16 Cedric Effenberger (SAM, ETHZ) NLEVPs: An Introduction Pro*Doc Disentis 16 / 37 Nonlinear inverse iteration II Algorithm Input: approximate eigenpair (x 0, λ 0 ) with v H x 0 = 1 for j = 0, 1,... until convergence do solve T (λ j ) x j+1 = T (λ j )x j for x j+1 update λ j+1 := λ j v H x j v H x j+1 normalize x j+1 := 1 v H u j+1 x j+1 end for local quadratic convergence to simple eigenvalues main computational work lies in the solution of the linear system

17 Cedric Effenberger (SAM, ETHZ) NLEVPs: An Introduction Pro*Doc Disentis 17 / 37 Reducing the computational work Replacing T (λ j ) 1 by T (σ) 1 for a fixed σ leads to misconvergence! But... ˆx = (ˆλ λ)t (λ) 1 T (λ)x = x T (λ) 1 T (λ)x (ˆλ λ)t (λ) 1 T (λ)x = x T (λ) 1[ T (λ) + (ˆλ λ)t (λ) ] x = x T (λ) 1 T (ˆλ)x + O ( ˆλ λ 2) Now T (σ) 1 can be used in place of T (λ) 1 safely. Even inexact solution of the linear system is possible.

18 Cedric Effenberger (SAM, ETHZ) NLEVPs: An Introduction Pro*Doc Disentis 18 / 37 Residual inverse iteration Algorithm Input: approximate eigenpair (x 0, σ) for j = 0, 1,... until convergence do solve v H T (σ) 1 T (λ j+1 )x j = 0 for λ j+1 solve T (σ) x = T (λ j+1 )x j for x update x j+1 := x j + x normalize x j+1 := 1 v H x j+1 x j+1 end for local linear convergence to simple eigenvalues convergence speed dependent on distance from σ to eigenvalue

19 Cedric Effenberger (SAM, ETHZ) NLEVPs: An Introduction Pro*Doc Disentis 19 / 37 Convergence string with elastically attached mass smallest magnitude eigenvalue using residual inverse iteration initial guess: σ = 0, x discrete version of function u(z) = z rel. error iterations it. λ

20 Cedric Effenberger (SAM, ETHZ) NLEVPs: An Introduction Pro*Doc Disentis 20 / 37 Difficulties no global convergence can only handle one eigenvalue at a time cannot handle multiple eigenvalues consecutive runs may converge to the same eigenvalue Linear eigensolvers (ARPACK, Jacobi-Davidson) exploit linear independence of eigenvectors to prevent reconvergence. Example (loss of linear independence for NLEVPs) The eigenvalues 1 and 2 of ([ ] [ ] [ ]) λ + λ x = share the same eigenvector [ 1 1 ].

21 Cedric Effenberger (SAM, ETHZ) NLEVPs: An Introduction Pro*Doc Disentis 21 / 37 Several eigenvalues in the linear case Ax = λx Let (x 1, λ 1 ), (x 2, λ 2 ) be two eigenpairs: Ax 1 = λ 1 x 1, Ax 2 = λ 2 x 2. The above equations can be merged: A [ [ ] ] [ ] λ1 x 1 x 2 = x1 x 2 }{{} λ 2 =:X }{{} =:Λ Hence, we have AX = X Λ. X spans an invariant subspace of A.

22 Cedric Effenberger (SAM, ETHZ) NLEVPs: An Introduction Pro*Doc Disentis 22 / 37 Generalization to the nonlinear case f 1 (λ)a 1 x + + f r (λ)a r x = 0 Assume r = 2 and let (x 1, λ 1 ), (x 2, λ 2 ) be two eigenpairs. A 1 x 1 f 1 (λ 1 ) + A 2 x 1 f 2 (λ 1 ) = 0, A 1 x 2 f 1 (λ 2 ) + A 2 x 2 f 2 (λ 2 ) = 0. These equations can again be merged [ ] [ ] f A 1 x1 x 1 (λ 1 ) [ ] [ f 2 + A f 1 (λ 2 ) 2 x1 x 2 (λ 1 ) 2 ] = 0 f 2 (λ 2 ) With X = [ x 1 x 2 ] and Λ = diag (λ1, λ 2 ) as before A 1 Xf 1 (Λ) + A 2 Xf 2 (Λ) = 0.

23 Cedric Effenberger (SAM, ETHZ) NLEVPs: An Introduction Pro*Doc Disentis 23 / 37 Invariant pairs Definition (X, Λ) C n m C m m is called an invariant pair if A 1 Xf 1 (Λ) + + A r Xf r (Λ) = 0. need to exclude degenerate situations, such as X = 0 linear case: require X to have full column rank not suitable for the nonlinear case: [ ] [ ] X =, Λ = is a perfectly reasonable invariant pair of ([ ] [ ] [ ]) λ + λ x =

24 Cedric Effenberger (SAM, ETHZ) NLEVPs: An Introduction Pro*Doc Disentis 24 / 37 Minimality Definition X = An invariant pair (X, Λ) is called minimal if X X Λ. X Λ l 1 has full column rank for some integer l. The smallest such l is called the minimality index of (X, Λ). [ ] [ 1 1 1, Λ = ] [ ] 1 1 X = 1 1 X Λ

25 Cedric Effenberger (SAM, ETHZ) NLEVPs: An Introduction Pro*Doc Disentis 25 / 37 Minimal invariant pairs (X, Λ) is a minimal invariant pair with minimality index 2. Theorem Let (X, Λ) be a minimal invariant pair of a nonlinear EVP T (λ)x = 0. Then every eigenvalue of Λ is an eigenvalue of T.

26 Cedric Effenberger (SAM, ETHZ) NLEVPs: An Introduction Pro*Doc Disentis 26 / 37 Computation of invariant pairs Apply Newton s method to T(X, Λ) := A 1 Xf 1 (Λ) + + A r Xf r (Λ) = 0. X, Λ n m + m 2 unknowns T(X, Λ) = 0 n m constraints We have to impose m 2 additional constraints. N(X, Λ) := V H Theorem X X Λ. X Λ l 1 I = 0, V suitably chosen Let (X, Λ) be a minimal invariant pair of T. The Fréchet derivative DF of F := [ T N ] at (X, Λ) is invertible iff the multiplicities of Λ s eigenvalues match those of T.

27 Cedric Effenberger (SAM, ETHZ) NLEVPs: An Introduction Pro*Doc Disentis 27 / 37 Ongoing work: Continuation of invariant pairs Joint project with Daniel Kressner (SAM, ETHZ) and Wolf-Jürgen Beyn (University of Bielefeld) Parameter-dependent nonlinear EVP T (λ, s)x = 0 Goal: track several eigenvalues as the parameter s varies Idea: use invariant pair ( X (s), Λ(s) ) at parameter value s as an initial guess for invariant pair at s + s apply Newton to obtain invariant pair ( X (s + s), Λ(s + s) ) at parameter value s + s repeat

28 Part II: Solution of polynomial / rational EVPs Cedric Effenberger (SAM, ETHZ) NLEVPs: An Introduction Pro*Doc Disentis 28 / 37

29 Cedric Effenberger (SAM, ETHZ) NLEVPs: An Introduction Pro*Doc Disentis 29 / 37 Linearization Polynomial EVPs can be solved by transforming them into linear ones. Example: (λ 2 M + λc + K)x = 0 introduce auxiliary variable y = λx rewrite as λmy + Cy + Kx = 0 This can be written as a generalized linear EVP with the same eigenvalues: [ ] [ ] [ ] [ ] C K y M 0 y = λ. I 0 x 0 I x linearizations not unique could also rewrite as λmy + λcx + Kx = 0 [ ] [ ] [ ] [ ] K 0 x C M x = λ 0 I y I 0 y

30 Cedric Effenberger (SAM, ETHZ) NLEVPs: An Introduction Pro*Doc Disentis 30 / 37 Linearization II There are entire vector spaces of linearizations (Mackey et al., 2006)! choice of linearization depends on underlying application structure preservation of interest Example The eigenvalues of an even polynomial eigenvalue problem (λ 2 M + λc + K)x = 0, M = M T, C = C T, K = K T occur in pairs (λ, λ). If M is invertible, the symmetric / skew-symmetric linearization [ ] [ ] [ ] [ ] K 0 x C M x = λ 0 M y M 0 y preserves this property.

31 Cedric Effenberger (SAM, ETHZ) NLEVPs: An Introduction Pro*Doc Disentis 31 / 37 Linearization III linearization of polynomial EVPs of degree d > 2 analogously by introducing the auxiliary vector x λx. λ d 1 x Conclusion 1: polynomial EVP of size n linearized EVP of size dn polynomial EVP has exactly dn eigenvalues (counting multiplicities) Conclusion 2: The extended vectors above will always be linearly independent, even though the eigenvectors x themselves may not.

32 Cedric Effenberger (SAM, ETHZ) NLEVPs: An Introduction Pro*Doc Disentis 32 / 37 Solution of rational EVPs most straightforward way: multiplication by the common denominator of all rational terms polynomial EVP degree of resulting polynomial EVP may be very large sometimes smarter ways to solve rational EVPs: Ongoing joint project with Daniel Kressner and Christian Engström (both SAM, ETHZ) [ G ω 2 M 1 ω 2( ξ ] α + )M η ω 2 2 u = 0 iγω M 1 + M 2 is a splitting of the total mass matrix of the problem.

33 Cedric Effenberger (SAM, ETHZ) NLEVPs: An Introduction Pro*Doc Disentis 33 / 37 Rational EVPs in electronic bandstructure computations by suitable ordering of nodes: [ 0 M 2 = R ˇM2] n n with a positive definite ˇM 2 R m m, m n by Cholesky decomposition: M 2 = F T F, F R m n Idea (Bai & Su, 2008): write rational term(s) in transfer function form ω 2 ξ η ω 2 iγω = ξ + bt (A ωe) 1 b

34 Cedric Effenberger (SAM, ETHZ) NLEVPs: An Introduction Pro*Doc Disentis 34 / 37 Conversion to transfer function form by partial fraction expansion p(ω) q(ω) = ζ + σ σ k ω ρ 1 ω ρ k ( ω = ζ + ρ ) 1 1 ( ω + + ρ ) k 1 σ 1 σ 1 σ k σ k = ζ + b T (A ωe) 1 b, where 1 b =., A = 1 ρ 1 σ 1... ρ k σ k, E = 1 σ σ k

35 Cedric Effenberger (SAM, ETHZ) NLEVPs: An Introduction Pro*Doc Disentis 35 / 37 Linearization of rational EVPs EVP with rational term in transfer function form: Ĝu b T (A ωe) 1 bf T F u = ω 2 ˆMu, where Ĝ = G + ξm 2, ˆM = M 1 + αm 2 rearrange: b T (A ωe)bf T F = b T (A ωe) 1 b F T F = B T (A ωe) 1 B where B = b F, A = A I, E = E I

36 Cedric Effenberger (SAM, ETHZ) NLEVPs: An Introduction Pro*Doc Disentis 36 / 37 Linearization of rational EVPs II Ĝu B T (A ωe) 1 Bu = ω 2 ˆMu Introducing the auxiliary variables v := (A ωe) 1 Bu, we obtain the linearized EVP Ĝ B T u ˆM u B A v = ω E v I w I w w := ωu, problem size standard lin. rational lin. p #DoF size comp. time size comp. time s s s s s s s

37 Cedric Effenberger (SAM, ETHZ) NLEVPs: An Introduction Pro*Doc Disentis 37 / 37 Summary introduced nonlinear EVPs showed several sample applications discussed Newton-based technique for determining one or several eigenpairs demonstrated linearization techniques for polynomial and rational EVPs outlined two ongoing projects