Some perturbation theorems for nonlinear eigenvalue problemsbindel/present/2013-01-cardiff.pdfNEP...

Post on 07-Aug-2020

4 views 0 download

Transcript of Some perturbation theorems for nonlinear eigenvalue problemsbindel/present/2013-01-cardiff.pdfNEP...

Some perturbation theorems for nonlineareigenvalue problems

David Bindel

Department of Computer ScienceCornell University

8 January 2013

(Cornell University) Dissipative Spectral Theory 1 / 46

Some favorite examples

Why nonlinear eigenvalue problems?

(Cornell University) Dissipative Spectral Theory 2 / 46

The problem

The general setting

Nonlinear eigenvalue problem:

T (λ)v = 0, v 6= 0.

whereT : Ω→ Cn×n analytic on simply connected Ω ⊂ Cdet(T ) 6≡ 0 (i.e. T is regular)

Write the set of nonlinear eigenvalues as Λ(T ).

Source: transform methods on almost anything with damping!For many examples, see:

NLEVP collectionSurvey by Mehrmann and Voss

(Cornell University) Dissipative Spectral Theory 3 / 46

The problem

Quadratic problems

Example: Damped free vibrations of a mechanical system

Mu′′ +Bu′ +Ku = 0.

Laplace transform:(s2M + sB +K)U = 0.

Approach directly or convert to first order:

Bv +Ku = sMv

v = su

(Cornell University) Dissipative Spectral Theory 4 / 46

The problem

Polynomial problems

More general is polynomial eigenvalue problem:

T (λ)v = 0, T (z) ≡ zdI + zd−1Ad + . . .+ zA1 +A0

Common approach: define uj = λjv, and solve−Ad−1 −Ad−2 . . . −A1 −A0

I 0I 0

. . . . . .I 0

ud−1

ud−2...u1

v

= λ

ud−1

ud−2...u1

v

This is one of many possible linearizations.Can do something similar with rational problems.

(Cornell University) Dissipative Spectral Theory 5 / 46

The problem

A special rational problem

Consider the eigenvalue equation[A− λI BC D − λI

] [vv

]= 0.

If λ 6∈ Λ(D), partial Gaussian elimination yields T (λ)v = 0, where

T (z) = A− zI −B(D − zI)−1C.

This is a spectral Schur complement problem.

(c.f. Feschbach, Lifschitz, Grushin).

(Cornell University) Dissipative Spectral Theory 6 / 46

The problem

Solving general NEPs

T (λ)x = 0, x 6= 0, T : Ω→ Cn×n analytic

Computational approaches:Local polynomial / rational approximation of TMethods based on contour integration

Either way, we want:A starting point (expansion point, contour)Error estimates for the results

(Cornell University) Dissipative Spectral Theory 7 / 46

The problem

Perturbation and localization

Many uses for perturbation theory in linear case:Backward error analysis (first-order theory, pseudospectra)Crude bounds for choosing algorithm parameters (Gerschgorin)Crude bounds for stability testing (Gerschgorin)Reasoning about dynamics (pseudospectra)

Want the same theory for nonlinear problems!

(Cornell University) Dissipative Spectral Theory 8 / 46

NEP perturbation theorems

First-order perturbation theory

Small, analytic E, consider

T = T + E

Given a simple eigentriple (λ, u, w∗) of T :

T (λ)u = 0, w∗T (λ) = 0.

First-order perturbation theory gives:

δλ = −w∗E(λ)u

w∗T ′(λ)u

Great! What about large perturbations, multiple eigenvalues, ...?

(Cornell University) Dissipative Spectral Theory 9 / 46

NEP perturbation theorems

Beyond first order

SupposeT,E : Ω→ Cn×n analyticΓ ⊂ Ω a simple contourT (z) + sE(z) nonsingular, all s ∈ [0, 1], z ∈ Γ.

Then T and T + E have the same number of eigenvalues inside Γ.

Proof:The winding number of det(T + sE) stays continuous for 0 ≤ s ≤ 1.

(Cornell University) Dissipative Spectral Theory 10 / 46

NEP perturbation theorems

A general recipe

Analyticity of T and E +Matrix nonsingularity test for T + sE =Inclusion region for Λ(T + E) +Eigenvalue counts for connected components of region

(Cornell University) Dissipative Spectral Theory 11 / 46

NEP perturbation theorems

Matrix Rouché

‖T (z)−1E(z)‖ < 1 on Γ =⇒ same eigenvalue count in Γ

Proof:‖T (z)−1E(z)‖ < 1 =⇒ T (z) + sE(z) invertible for 0 ≤ s ≤ 1.

(Gohberg and Sigal proved a more general version in 1971.)

(Cornell University) Dissipative Spectral Theory 12 / 46

NEP perturbation theorems

Nonlinear pseudospectra

Define the nonlinear ε-pseudospectrum as

Λε(T ) = z ∈ Ω : ‖T (z)−1‖ > ε−1

Let E = E : Ω→ Cn×n s.t. E analytic,maxz∈Ω ‖E(z)‖ < ε. Then

Λε(T ) =⋃E∈E

Λ(T + E).

If E0 = E ∈ Cn×n : ‖E0‖ < ε, we may also write

Λε(T ) =⋃

E0∈E0

Λ(T + E0).

(Cornell University) Dissipative Spectral Theory 13 / 46

NEP perturbation theorems

Nonlinear pseudospectra and backward error

Suppose λ, v an approximate eigenpair with ‖v‖ = 1,

T (λ)v = r, ‖r‖ small.

Then λ ∈ Λ‖r‖(T ), since(T (λ)− rv∗

)v = 0

(Cornell University) Dissipative Spectral Theory 14 / 46

NEP perturbation theorems

Nonlinear pseudospectra and dynamics

Suppose Ψ : [0,∞)→ CN×N , let

R(z) ≡∫ ∞

0e−ztΨ(t) dt.

Ψ bounded =⇒ R(z) defined in RHP and for any ε > 0,

supt>0‖Ψ(t)‖ ≥ αε

ε,

whereαε ≡ sup

‖R(λε)‖>ε−1

Re(λε)

(Similar proof to that for linear pseudospectra.)

(Cornell University) Dissipative Spectral Theory 15 / 46

NEP perturbation theorems

Pseudospectral counting

Let T,E analytic on Ω and define:

Ωε ≡ z ∈ Ω : ‖E(z)‖ < ε.

ThenΛ(T ) ∩ Ωε ⊂ Λε(T + E)

Also, ifU ⊂ Λε(T + E) a connected component.U ⊂ Ωε.

then U contains the same number of eigenvalues of T and T + E,of which there must be at least one.

(Cornell University) Dissipative Spectral Theory 16 / 46

NEP perturbation theorems

Weakly coupled problems

T (z) =

[L1(z) H(z)G(z) L2(z)

]is analytic over Ω, and

‖G(z)‖ ≤ γ, ‖H(z)‖ ≤ η, Λδ1(L1) ∩ Λδ2(L2) = ∅.

Assume γη < δ1δ2, boundary of Λδ1(L1) is strictly inside Ω. Then1 Λ(T ) ⊂ Λδ1(L1) ∪ Λδ2(L2)

2 T and L1 have same eigenvalue counts in Λδ1(L1)

3 For λ ∈ Λδ1(L1), eigenvector v satisfies ‖v2‖/‖v1‖ < γ/δ1.

4 For λ ∈ Λδ2(L2), eigenvector v satisfies ‖v2‖/‖v1‖ > γ/δ2.

(Cornell University) Dissipative Spectral Theory 17 / 46

Perturbing linear problems

Linear problems, nonlinear perturbations

Perturb linear problem with E analytic, “small” on Ω:

T (z) = A− zB + E(z).

Many linear perturbation theorems still hold!

(Cornell University) Dissipative Spectral Theory 18 / 46

Perturbing linear problems

Nonlinear perturbations + pseudospectra

T (z) = A− zI + E(z)

and suppose ‖E‖ < ε on Ω.

If U a connected component of Λε(A), U ⊂ Ω, thenA and T have the same eigenvalue counts in U .The eigenvalue count in U is at least one.

(Cornell University) Dissipative Spectral Theory 19 / 46

Perturbing linear problems

Nonlinear Gerschgorin

For D diagonal, consider

T (z) = D − zI + E(z)

such thatn∑j=1

|eij(z)| ≤ ρi

ThenΛ(T ) ⊂

⋃ni=1Gi where Gi = Bρi(dii)

U =⋃i∈I Gi a connected component, U ⊂ Ω

=⇒ U contains |I| eigenvalues.

(Cornell University) Dissipative Spectral Theory 20 / 46

Perturbing linear problems

Nonlinear Bauer-Fike bound

Suppose |E(z)| ≤ F componentwise on Ω,

T (z) = A− zI + E(z).

and A has eigentriples (λi, vi, w∗i ). Then

Λ(T ) ⊂n⋃i=1

Bφi(λi)

where φi = n‖F‖2 sec(θi) and

sec(θi) =‖wi‖‖vi‖|w∗i vi|

.

Can also count within connected components.

(Cornell University) Dissipative Spectral Theory 21 / 46

Application: DDEs

Application: Delay-differential equation

From NLEVP collection

T (λ) = A0 − λI +A1 exp(−λ)

Corresponding to

u′(t) = A0u(t) +A1u(t− 1)

Double non-semisimple eigenvalue λ = 3πi.

(Cornell University) Dissipative Spectral Theory 22 / 46

Application: DDEs

Pseudospectral plot

−20 −15 −10 −5 0 5 10 15 20−40

−30

−20

−10

0

10

20

30

40

(Cornell University) Dissipative Spectral Theory 23 / 46

Application: DDEs

Pseudospectral plot

−20 −15 −10 −5 0 5 10 15 20−40

−30

−20

−10

0

10

20

30

40

−1.5

−1.5

−1

−1

−1

−1

−1

−1

−1

−1

−1

−1

−0.5

−0.5

−0.5

−0.5

−0.5

−0.5

−0.5

0

0

0

0

0.5

0.5

1

1

(Cornell University) Dissipative Spectral Theory 24 / 46

Application: DDEs

Gerschgorin applied

ConsiderV −1T (λ)V = D − λI + A1 exp(−λ)

Apply Gerschgorin-like bound

Λ(T ) ⊂3⋃i=1

Bρi(dii) ∪ | exp(−λ)| > exp(−σ)

where

ρi = exp(−σ)

∑j

(A1)ij

α∑j

(A1)ji

1−α

(Cornell University) Dissipative Spectral Theory 25 / 46

Application: DDEs

Example: Bounding the spectral abscissa

−20 −15 −10 −5 0 5 10 15 20−40

−30

−20

−10

0

10

20

30

40

−1.5

−1.5

−1

−1

−1

−1

−1

−1

−1

−1

−1

−1

−0.5

−0.5

−0.5

−0.5

−0.5

−0.5

−0.5

0

0

0

0

0.5

0.5

1

1

(Cornell University) Dissipative Spectral Theory 26 / 46

Application: DDEs

Example: Imaginary part of unstable eigenvalues

−20 −15 −10 −5 0 5 10 15 20−40

−30

−20

−10

0

10

20

30

40

−1.5

−1.5

−1

−1

−1

−1

−1

−1

−1

−1

−1

−1

−0.5

−0.5

−0.5

−0.5

−0.5

−0.5

−0.5

0

0

0

0

0.5

0.5

1

1

(Cornell University) Dissipative Spectral Theory 27 / 46

Application: DDEs

Switching terms

ConsiderV −1T (λ)V = D exp(−λ)− λ+ A0

Gerschgorin-like argument now bounds spectrum from left!

(Cornell University) Dissipative Spectral Theory 28 / 46

Application: DDEs

Example: Bounding spectrum from the left

−20 −15 −10 −5 0 5 10 15 20−40

−30

−20

−10

0

10

20

30

40

−1.5

−1.5

−1

−1

−1

−1

−1

−1

−1

−1

−1

−1

−0.5

−0.5

−0.5

−0.5

−0.5

−0.5

−0.5

0

0

0

0

0.5

0.5

1

1

(Cornell University) Dissipative Spectral Theory 29 / 46

Application: Resonances

Schrödinger resonances

(Cornell University) Dissipative Spectral Theory 30 / 46

Application: Resonances

Spectra and scattering

Spectrum for H = −∆ + V , supp(V ) compact.

(Cornell University) Dissipative Spectral Theory 31 / 46

Application: Resonances

Resonances and scattering

1.0 1.5 2.0 2.5 3.0 3.5 4.0k

50

100|φ

(k)|

For supp(V ) ⊂ Ω, consider a scattering experiment:

(H − k2)ψ = f on Ω

(∂n −B(k))ψ = 0 on ∂Ω

See resonance peaks (Breit-Wigner):

φ(k) ≡ w∗ψ ≈ C(k − k∗)−1.

(Cornell University) Dissipative Spectral Theory 32 / 46

Application: Resonances

1D resonances: a quadratic eigenvalue problem

(− d2

dx2+ V (x)− k2

)ψ = 0, x ∈ (a, b)(

d

dx− ik

)ψ = 0, x = b(

d

dx+ ik

)ψ = 0, x = a

Look for nontrivial solutions:Im(k) > 0: Bound statesIm(k) < 0: Resonances

See:

http://www.cs.cornell.edu/~bindel/sw/matscat/

(Cornell University) Dissipative Spectral Theory 33 / 46

Application: Resonances

Is it that easy?

−400 −200 0 200 400−40

−30

−20

−10

0

10

20

All eigenvaluesChecked eigenvalues

(Cornell University) Dissipative Spectral Theory 34 / 46

Application: Resonances

Is it that easy?

−10 −8 −6 −4 −2 0 2 4 6 8 100

0.2

0.4

0.6

0.8

1

1.2Potential

−2 −1.5 −1 −0.5 0 0.5 1 1.5 2−2

−1.5

−1

−0.5

0Pole locations

(Cornell University) Dissipative Spectral Theory 35 / 46

Application: Resonances

Sensitivity for resonances

Resonance solutions are stationary points with respect to ψ of

Φ(ψ, k) =

∫Ωψ[−∇2ψ + (V − k2)ψ

]dΩ−

∫∂Ωψ

(∂ψ

∂n−B(k)ψ

)dΓ

=

∫Ω

[(∇ψ)T (∇ψ) + ψ(V − k2)ψ

]dΩ−

∫∂ΩψB(k)ψ dΓ

If (ψ, k) a resonance pair, then Φ(ψ, k) = 0 and DψΦ(ψ, k) = 0.

(Cornell University) Dissipative Spectral Theory 36 / 46

Application: Resonances

Potential perturbations

If (ψ, k) a resonance pair, then Φ(ψ, k) = 0 and DψΦ(ψ, k) = 0.

Consider perturbed V :

δΦ = DψΦ · δψ +DV Φ · δV +DkΦ · δk = 0

Use DψΦ · δψ = 0:

δk = −DV Φ · δVDkΦ

(Cornell University) Dissipative Spectral Theory 37 / 46

Application: Resonances

Perturbation worked out

So look at how perturbations δV change k:

δk =

∫Ω δV ψ

2

2k∫

Ω ψ2 −

∫Γ ψB

′(k)ψ

Can also write in terms of a residual for ψ as a solution for the potentialV + δV :

δk =

∫Ω ψ(−∆ + (V + δV )− k2)ψ

2k∫

Ω ψ2 −

∫Γ ψB

′(k)ψ.

(Cornell University) Dissipative Spectral Theory 38 / 46

Application: Resonances

Backward error analysis in MatScat

1 Compute approximate solution (ψ, k).2 Map ψ to high-resolution quadrature grid to evaluate

δk =

∫Ω ψ(−∆ + V − k2)ψ

2k∫

Ω ψ2 −

∫Γ ψB

′(k)ψ.

3 If δk large, discard k; otherwise, accept k ≈ k + δk.

(Cornell University) Dissipative Spectral Theory 39 / 46

Application: Resonances

Nonlinear vs linear eigenproblems

Can also compute resonances byAdding a complex absorbing potentialComplex scaling methodsArtificial dampers

Both result in complex-symmetric ordinary eigenproblems:

(Kext − k2Mext)ψext =

([K11 K12

K21 K22

]− k2

[M11 M12

M21 M22

])[ψ1

ψ2

]= 0

where ψ2 correspond to extra variables (outside Ω).

(Cornell University) Dissipative Spectral Theory 40 / 46

Application: Resonances

Spectral Schur complement

0 5 10 15 20 25 30

0

5

10

15

20

25

30

0 5 10 15 20 25 30

0

5

10

15

20

25

30

Eliminate “extra” variables ψ2 to get

T (k)ψ1 =(K11 − k2M11 − C(k)

)ψ1 = 0

where

C(k) = (K12 − k2M12)(K22 − k2M22)−1(K21 − k2M21)

(Cornell University) Dissipative Spectral Theory 41 / 46

Application: Resonances

Apples to oranges?

T (k)ψ = (K − k2M − C(k))ψ = 0 (exact DtN map)

T (k)ψ = (K − k2M − C(k))ψ = 0 (spectral Schur complement)

Two ideas:Perturbation theory for NEP for local refinementComplex analysis to get more global analysis

(Cornell University) Dissipative Spectral Theory 42 / 46

Application: Resonances

Aside on spectral Schur complement

Inverse of a Schur complement is a submatrix of an inverse:

(Kext − z2Mext)−1 =

[T (z)−1 ∗∗ ∗

]So for reasonable norms,

‖T (z)−1‖ ≤ ‖(Kext − z2Mext)−1‖.

Or

Λε(T ) ⊂ Λε(Kext,Mext),

Λε(T ) ≡ z : ‖A(z)−1‖ > ε−1Λε(Kext,Mext) ≡ z : ‖(Kext − z2Mext)

−1‖ > ε−1

(Cornell University) Dissipative Spectral Theory 43 / 46

Application: Resonances

Nonlinear bounds from linear pseudospectra

Recall:

T (k)ψ = (K − k2M − C(k))ψ = 0 (exact DtN map)

A(k)ψ = (K − k2M − C(k))ψ = 0 (spectral Schur complement)

Let Ωε = z ∈ C : ‖C(z)− C(z)‖ < ε. Then:

Λ(T ) ∩ Ωε ⊂ Λε(T ) ⊂ Λε(Kext,Mext)

(Cornell University) Dissipative Spectral Theory 44 / 46

Application: Resonances

Assessing approximate resonances

-5

-4

-3

-2

-1

0

0 2 4 6 8 10

Im(k

)

Re(k)

CorrectSpurious0

0

-2

-2

-4

-4

-6

-8

-8

-8

-10

-10

-10

To get axisymmetric resonances in corral model, compute:Eigenvalues of a complex-scaled problemResiduals in nonlinear eigenproblemlog10 ‖T (k)− T (k)‖

(Cornell University) Dissipative Spectral Theory 45 / 46

Conclusion

Conclusion

Nonlinear eigenvalue problems are as natural as linear problemsLinear perturbation theorems with complex analytic proofs apply“Perturbation Theorems for Nonlinear Eigenvalue Problems”David Bindel and Amanda Hood

(Cornell University) Dissipative Spectral Theory 46 / 46