computing spectra of uw applied mathematics using finite...

32
UW Applied Mathematics Computing Spectra of Linear Operators Using Finite Differences J. Nathan Kutz Department of Applied Mathematics University of Washington Seattle, WA 98195-2420 Email: ([email protected]) Stability Workshop, Seattle, September 6-8, 2006

Upload: others

Post on 07-Oct-2020

4 views

Category:

Documents


0 download

TRANSCRIPT

Page 1: Computing Spectra of UW Applied Mathematics Using Finite ...depts.washington.edu/bdecon/workshop2012/c1_stability.pdf · UW Applied Mathematics Computing Spectra of Linear Operators

UW

App

lied

Mat

hem

atic

s

Computing Spectra of Linear Operators

Using Finite Differences

J. Nathan Kutz

Department of Applied Mathematics University of WashingtonSeattle, WA 98195-2420 Email: ([email protected])

Stability Workshop, Seattle, September 6-8, 2006

Page 2: Computing Spectra of UW Applied Mathematics Using Finite ...depts.washington.edu/bdecon/workshop2012/c1_stability.pdf · UW Applied Mathematics Computing Spectra of Linear Operators

Introduction: Spectral Stability

Consider the nonlinear PDE evolution

Equilibrium solution

Linear stability

UW

App

lied

Mat

hem

atic

s

perturbation

B. Deconinck and J. N. Kutz, J. Comp. Phys. (2006)

Page 3: Computing Spectra of UW Applied Mathematics Using Finite ...depts.washington.edu/bdecon/workshop2012/c1_stability.pdf · UW Applied Mathematics Computing Spectra of Linear Operators

Associated Eigenvalue Problem

Separation of variables

Eigenvalue problem

Linear stability

Re(λ) ≤ 0Re(λ) > 0

spectrally stable

spectrally unstable

UW

App

lied

Mat

hem

atic

s

Page 4: Computing Spectra of UW Applied Mathematics Using Finite ...depts.washington.edu/bdecon/workshop2012/c1_stability.pdf · UW Applied Mathematics Computing Spectra of Linear Operators

Finite Differences and Taylor Series

Taylor expand

slope formula with error

UW

App

lied

Mat

hem

atic

s

Adderrorapproximation

Page 5: Computing Spectra of UW Applied Mathematics Using Finite ...depts.washington.edu/bdecon/workshop2012/c1_stability.pdf · UW Applied Mathematics Computing Spectra of Linear Operators

Higher-Order Accuracy

Taylor expand again

slope formula with improved error

UW

App

lied

Mat

hem

atic

s

8 x (first) and subtracterrorapproximation

Page 6: Computing Spectra of UW Applied Mathematics Using Finite ...depts.washington.edu/bdecon/workshop2012/c1_stability.pdf · UW Applied Mathematics Computing Spectra of Linear Operators

Finite Difference Tables

neighboring points determine accuracy

UW

App

lied

Mat

hem

atic

s

Page 7: Computing Spectra of UW Applied Mathematics Using Finite ...depts.washington.edu/bdecon/workshop2012/c1_stability.pdf · UW Applied Mathematics Computing Spectra of Linear Operators

Forward and Backward Differences

asymmetric neighboring points

UW

App

lied

Mat

hem

atic

s

Required for incorporating boundary conditions

Page 8: Computing Spectra of UW Applied Mathematics Using Finite ...depts.washington.edu/bdecon/workshop2012/c1_stability.pdf · UW Applied Mathematics Computing Spectra of Linear Operators

Numerical Round-Off

round-off error dominates below

UW

App

lied

Mat

hem

atic

s

Consider the error in approximating the first derivative

The error includes round-off and truncation

Assume round-off and

minimum at

truncationround-off

Page 9: Computing Spectra of UW Applied Mathematics Using Finite ...depts.washington.edu/bdecon/workshop2012/c1_stability.pdf · UW Applied Mathematics Computing Spectra of Linear Operators

Boundary Conditions: Pinned

tri-diagonal matrix structure

UW

App

lied

Mat

hem

atic

s

pinned boundaries

Page 10: Computing Spectra of UW Applied Mathematics Using Finite ...depts.washington.edu/bdecon/workshop2012/c1_stability.pdf · UW Applied Mathematics Computing Spectra of Linear Operators

Boundary Conditions: Periodic

tri-diagonal matrix structure with corners

UW

App

lied

Mat

hem

atic

s

periodic boundaries

Page 11: Computing Spectra of UW Applied Mathematics Using Finite ...depts.washington.edu/bdecon/workshop2012/c1_stability.pdf · UW Applied Mathematics Computing Spectra of Linear Operators

Boundary Conditions: No Flux

no longer symmetry matrix

UW

App

lied

Mat

hem

atic

s

no flux condition

Page 12: Computing Spectra of UW Applied Mathematics Using Finite ...depts.washington.edu/bdecon/workshop2012/c1_stability.pdf · UW Applied Mathematics Computing Spectra of Linear Operators

General Boundaries

no longer symmetry matrix

UW

App

lied

Mat

hem

atic

s

General (Sturm-Liouville) boundary conditions

Difficult to incorporate into matrix structure

• shooting methods

• relaxation methods

Page 13: Computing Spectra of UW Applied Mathematics Using Finite ...depts.washington.edu/bdecon/workshop2012/c1_stability.pdf · UW Applied Mathematics Computing Spectra of Linear Operators

Algorithm

Easily extends to vectors and higher dimensions

• choose domain length and discretization size

• construct linear operator

• implement boundary conditions

• use eigenvalue/eigenvector solver: O(N3) (or shooting/relaxation methods)

• construct eigenfunctionsUW

App

lied

Mat

hem

atic

s

Page 14: Computing Spectra of UW Applied Mathematics Using Finite ...depts.washington.edu/bdecon/workshop2012/c1_stability.pdf · UW Applied Mathematics Computing Spectra of Linear Operators

Example: Mathieu Equation

compute with matlab, maple, mathematica, or homemade code

UW

App

lied

Mat

hem

atic

s

Classic Example

Operator is self-adjoint (real spectrum)

Page 15: Computing Spectra of UW Applied Mathematics Using Finite ...depts.washington.edu/bdecon/workshop2012/c1_stability.pdf · UW Applied Mathematics Computing Spectra of Linear Operators

Spectrum for Mathieu Equation

a is eigenvalue

q

a

UW

App

lied

Mat

hem

atic

s

Page 16: Computing Spectra of UW Applied Mathematics Using Finite ...depts.washington.edu/bdecon/workshop2012/c1_stability.pdf · UW Applied Mathematics Computing Spectra of Linear Operators

Computing the Ground StateU

W A

pplie

d M

athe

mat

ics

Convergence study and CPU time (q=2)

0.43 sec0.01 sec

1 hour0.77 sec

- 2.5 min

• increase domain length

• Floquet theory

beyond Matlab7’s ability

What about band-gap structure

Page 17: Computing Spectra of UW Applied Mathematics Using Finite ...depts.washington.edu/bdecon/workshop2012/c1_stability.pdf · UW Applied Mathematics Computing Spectra of Linear Operators

Calculating the Bands: Domain LengthU

W A

pplie

d M

athe

mat

ics

Increase the domain length (q=2)

traditional way: very costly for recovering bands

doubling gives 8x computational increase

Page 18: Computing Spectra of UW Applied Mathematics Using Finite ...depts.washington.edu/bdecon/workshop2012/c1_stability.pdf · UW Applied Mathematics Computing Spectra of Linear Operators

Calculating the Bands: Floquet TheoryU

W A

pplie

d M

athe

mat

ics

Make use of Floquet (Bloch) theory

with Floquet (characteristic) exponents

• keep fixed domain

• discretize

• solve D O(N3) equations

larger period solutions

Page 19: Computing Spectra of UW Applied Mathematics Using Finite ...depts.washington.edu/bdecon/workshop2012/c1_stability.pdf · UW Applied Mathematics Computing Spectra of Linear Operators

Implementing Floquet TheoryU

W A

pplie

d M

athe

mat

ics

Floquet theory modifies matrix corners

with Floquet slices

Page 20: Computing Spectra of UW Applied Mathematics Using Finite ...depts.washington.edu/bdecon/workshop2012/c1_stability.pdf · UW Applied Mathematics Computing Spectra of Linear Operators

Floquet Theory vs. Domain LengthU

W A

pplie

d M

athe

mat

ics

Compare methods for computing band (q=2)

band density

beyond Matlab7’s ability

9 min8 sec

-1 min

-10 min

Use Floquet Theory!

Page 21: Computing Spectra of UW Applied Mathematics Using Finite ...depts.washington.edu/bdecon/workshop2012/c1_stability.pdf · UW Applied Mathematics Computing Spectra of Linear Operators

Example: Periodic NLS

Consider the system

with

UW

App

lied

Mat

hem

atic

s

and Jacobi sine function

Page 22: Computing Spectra of UW Applied Mathematics Using Finite ...depts.washington.edu/bdecon/workshop2012/c1_stability.pdf · UW Applied Mathematics Computing Spectra of Linear Operators

Spectrum of Periodic NLS

k=0.7k=0 k=0.9 k=1

UW

App

lied

Mat

hem

atic

s

Page 23: Computing Spectra of UW Applied Mathematics Using Finite ...depts.washington.edu/bdecon/workshop2012/c1_stability.pdf · UW Applied Mathematics Computing Spectra of Linear Operators

Importance of Floquet Slicing

µ ≠ 0

µ = 0

UW

App

lied

Mat

hem

atic

s

Page 24: Computing Spectra of UW Applied Mathematics Using Finite ...depts.washington.edu/bdecon/workshop2012/c1_stability.pdf · UW Applied Mathematics Computing Spectra of Linear Operators

Accuracy and Convergence U

W A

pplie

d M

athe

mat

ics

Page 25: Computing Spectra of UW Applied Mathematics Using Finite ...depts.washington.edu/bdecon/workshop2012/c1_stability.pdf · UW Applied Mathematics Computing Spectra of Linear Operators

Example: 2D Mathieu Equation

Consider

with

UW

App

lied

Mat

hem

atic

s

Operation Count: O((N2)3)=O(N6)

Page 26: Computing Spectra of UW Applied Mathematics Using Finite ...depts.washington.edu/bdecon/workshop2012/c1_stability.pdf · UW Applied Mathematics Computing Spectra of Linear Operators

Laplacian in 2DU

W A

pplie

d M

athe

mat

ics

must stack 2D data: periodic boundaries add structure

Consider

Discretize:

Let

Page 27: Computing Spectra of UW Applied Mathematics Using Finite ...depts.washington.edu/bdecon/workshop2012/c1_stability.pdf · UW Applied Mathematics Computing Spectra of Linear Operators

Laplacian in 2DU

W A

pplie

d M

athe

mat

ics

nx=ny=4

Page 28: Computing Spectra of UW Applied Mathematics Using Finite ...depts.washington.edu/bdecon/workshop2012/c1_stability.pdf · UW Applied Mathematics Computing Spectra of Linear Operators

Laplacian in 2DU

W A

pplie

d M

athe

mat

ics

Matlab easily builds 2D Laplacian

nx=ny=4

Page 29: Computing Spectra of UW Applied Mathematics Using Finite ...depts.washington.edu/bdecon/workshop2012/c1_stability.pdf · UW Applied Mathematics Computing Spectra of Linear Operators

Band Gap Structure

Quasi-momentum representation

First three band-gap structures

UW

App

lied

Mat

hem

atic

s

A=1A=-0.3 A=0

Page 30: Computing Spectra of UW Applied Mathematics Using Finite ...depts.washington.edu/bdecon/workshop2012/c1_stability.pdf · UW Applied Mathematics Computing Spectra of Linear Operators

2D Dominant Eigenfunctions

First three eigenfunctions for µx = µy=0UW

App

lied

Mat

hem

atic

s

Page 31: Computing Spectra of UW Applied Mathematics Using Finite ...depts.washington.edu/bdecon/workshop2012/c1_stability.pdf · UW Applied Mathematics Computing Spectra of Linear Operators

2D Eigenfunctions

First three eigenfunctions for µx = µy=1/4

UW

App

lied

Mat

hem

atic

s

Page 32: Computing Spectra of UW Applied Mathematics Using Finite ...depts.washington.edu/bdecon/workshop2012/c1_stability.pdf · UW Applied Mathematics Computing Spectra of Linear Operators

Summary and ConclusionsU

W A

pplie

d M

athe

mat

ics

• simple, simple, simple

• boundary conditions at edge of matrices

• eigenvalue solvers make use of sparse structure

• Floquet theory for resolution of bands

• costly/impractical for 2D-3D problems

B. Deconinck and J. N. Kutz, J. Comp. Phys. (2006)