Exponential multistep methods of Adams-type

Similar documents
Exponential integrators

Exponential integrators for semilinear parabolic problems

Exponential Runge-Kutta methods for parabolic problems

EXPLICIT EXPONENTIAL RUNGE-KUTTA METHODS FOR SEMILINEAR PARABOLIC PROBLEMS

EXPONENTIAL ROSENBROCK-TYPE METHODS

Linear algebra for exponential integrators

Exponential integrators

Exponential integrators and functions of the matrix exponential

Implementation of exponential Rosenbrock-type integrators

Exponential integration of large systems of ODEs

Splitting methods with boundary corrections

Exponential integrators for oscillatory second-order differential equations

Meshfree Exponential Integrators

Approximating the matrix exponential of an advection-diffusion operator using the incomplete orthogonalization method

DIMENSION SPLITTING FOR TIME DEPENDENT OPERATORS. Eskil Hansen and Alexander Ostermann

The collocation method for ODEs: an introduction

McGill University Department of Mathematics and Statistics. Ph.D. preliminary examination, PART A. PURE AND APPLIED MATHEMATICS Paper BETA

Equivalence Theorems and Their Applications

Error analysis of implicit Euler methods for quasilinear hyperbolic evolution equations

THE METHOD OF LINES FOR PARABOLIC PARTIAL INTEGRO-DIFFERENTIAL EQUATIONS

Takens embedding theorem for infinite-dimensional dynamical systems

Exponential Integrators

Solving Ordinary Differential Equations

Finite Differences for Differential Equations 28 PART II. Finite Difference Methods for Differential Equations

Krylov Implicit Integration Factor Methods for Semilinear Fourth-Order Equations

Interval Methods and Taylor Model Methods for ODEs

1 Ordinary Differential Equations

FDM for wave equations

Co-sponsored School/Workshop on Integrable Systems and Scientific Computing June Exponential integrators for stiff systems

Krylov methods for the computation of matrix functions

Lecture Notes to Accompany. Scientific Computing An Introductory Survey. by Michael T. Heath. Chapter 9

RUNGE-KUTTA TIME DISCRETIZATIONS OF NONLINEAR DISSIPATIVE EVOLUTION EQUATIONS

Exponential Integrators

Efficient Wavefield Simulators Based on Krylov Model-Order Reduction Techniques

Computing the Action of the Matrix Exponential

Test you method and verify the results with an appropriate function, e.g. f(x) = cos 1 + sin

Preliminary Examination, Numerical Analysis, August 2016

Ordinary Differential Equations

Convexity of the Reachable Set of Nonlinear Systems under L 2 Bounded Controls

Lecture 9 Approximations of Laplace s Equation, Finite Element Method. Mathématiques appliquées (MATH0504-1) B. Dewals, C.

Lecture 4: Numerical solution of ordinary differential equations

Integration of Vlasov-type equations

Implicit-explicit exponential integrators

A STUDY OF MULTIGRID SMOOTHERS USED IN COMPRESSIBLE CFD BASED ON THE CONVECTION DIFFUSION EQUATION

Scientific Computing: An Introductory Survey

Dynamical Low-Rank Approximation to the Solution of Wave Equations

Marlis Hochbruck 1, Michael Hönig 1 and Alexander Ostermann 2

Chapter 5 Exercises. (a) Determine the best possible Lipschitz constant for this function over 2 u <. u (t) = log(u(t)), u(0) = 2.

Semi-implicit Krylov Deferred Correction Methods for Ordinary Differential Equations

Numerical solution of ODEs

On the efficiency of the Peaceman-Rachford ADI-dG method for wave-type methods

Numerical Methods for Differential Equations Mathematical and Computational Tools

CS 450 Numerical Analysis. Chapter 9: Initial Value Problems for Ordinary Differential Equations

A brief introduction to ordinary differential equations

Lecture 10: Finite Differences for ODEs & Nonlinear Equations

arxiv: v1 [math.na] 6 Nov 2017

Southern Methodist University.

Uniformly accurate averaging numerical schemes for oscillatory evolution equations

Preconditioning for Nonsymmetry and Time-dependence

1. Introduction. In the present paper, we consider a nonautonomous differential equation involving a time-dependent linear operator A

Non-smooth data error estimates for linearly implicit Runge Kutta methods

10 The Finite Element Method for a Parabolic Problem

An Efficient Algorithm Based on Quadratic Spline Collocation and Finite Difference Methods for Parabolic Partial Differential Equations.

Partial regularity for fully nonlinear PDE

Index. C 2 ( ), 447 C k [a,b], 37 C0 ( ), 618 ( ), 447 CD 2 CN 2

Exponential Time Differencing for Stiff Systems

Defect-based a-posteriori error estimation for implicit ODEs and DAEs

Lecture 4. Chapter 4: Lyapunov Stability. Eugenio Schuster. Mechanical Engineering and Mechanics Lehigh University.

u = (A + F )u, u(0) = η, (1.1)

Initial value problems for ordinary differential equations

Dispersive numerical schemes for Schrödinger equations

A Sobolev trust-region method for numerical solution of the Ginz

Comparing Leja and Krylov approximations of large scale matrix exponentials

ITERATIVE METHODS FOR NONLINEAR ELLIPTIC EQUATIONS

Numerical Methods for Differential Equations

Numerical Methods for Differential Equations

Exam in TMA4215 December 7th 2012

Stability of Krylov Subspace Spectral Methods

Ordinary differential equations - Initial value problems

Anton ARNOLD. with N. Ben Abdallah (Toulouse), J. Geier (Vienna), C. Negulescu (Marseille) TU Vienna Institute for Analysis and Scientific Computing

Index. higher order methods, 52 nonlinear, 36 with variable coefficients, 34 Burgers equation, 234 BVP, see boundary value problems

NOTES ON LINEAR ODES

Design of optimal Runge-Kutta methods

2 Numerical Methods for Initial Value Problems

Finite Difference and Finite Element Methods

IEEE TRANSACTIONS ON ANTENNAS AND PROPAGATION, VOL. 57, NO. 2, FEBRUARY X/$ IEEE

Favourable time integration methods for. Non-autonomous evolution equations

ECE257 Numerical Methods and Scientific Computing. Ordinary Differential Equations

A space-time Trefftz method for the second order wave equation

Projection Methods. Michal Kejak CERGE CERGE-EI ( ) 1 / 29

Shifted Laplace and related preconditioning for the Helmholtz equation

Iterative methods for positive definite linear systems with a complex shift

The LEM exponential integrator for advection-diffusion-reaction equations

Partial Differential Equations and the Finite Element Method

CS520: numerical ODEs (Ch.2)

The ReLPM Exponential Integrator for FE Discretizations of Advection-Diffusion Equations

Chapter 1: The Finite Element Method

Discrete Projection Methods for Integral Equations

AMS subject classifications. Primary, 65N15, 65N30, 76D07; Secondary, 35B45, 35J50

Solving PDEs with Multigrid Methods p.1

Transcription:

Exponential multistep methods of Adams-type Marlis Hochbruck and Alexander Ostermann KARLSRUHE INSTITUTE OF TECHNOLOGY (KIT) 0 KIT University of the State of Baden-Wuerttemberg and National Laboratory of the Helmholtz Association www.kit.edu

Outline Motivation Exponential Adams methods Linearized exponential multistep methods Implementation Application: Simulation of optical resonators 1

Motivation general initial value problem u (t) = F ( t, u(t) ), u(t 0 ) = u 0. (in this talk: autonomous problems only) convergence analysis for semilinear initial value problems variation of constants formula u (t) = Au(t) + g ( t, u(t) ), u(0) = u 0 h u(t n+1 ) = e ha u(t n ) + e (h τ)a g ( u(t n + τ) ) dτ 0 t n = t 0 + nh, n = 0, 1,... idea of multistep methods: replace g by interpolation polynomial 2

Exponential Adams methods Given approximations u j u(t j ), consider interpolation polynomial p n through ( tn k+1, g(u n k+1 ) ),..., ( t n, g(u n ) ), given by k 1 p n (t n + θh) = G n + j=1 ( 1) j ( θ j where j G n denotes jth backward difference ) j G n, G j = g(u j ) 0 G n = G n, j G n = j 1 G n j 1 G n 1, j = 1, 2,.... replace nonlinearity g by p n in v.o.c formula 3

Exponential Adams methods, cont d numerical scheme h u n+1 = e ha u n + e (h τ)a p n (t n + τ)dτ 0 k 1 = u n + hϕ 1 ( ha)f(u n ) + h γ j ( ha) j G n j=1 with weights γ j (z) = ( 1) j 1 0 ( ) θ e (1 θ)z dθ, j 0. j Certaine, 1960, Nørsett 1969, Cox and Matthews 2002 rational variants: Lambert and Sigurdsson 1972, Verwer 1976 4

Exponential Adams methods, cont d in terms of ϕ-functions the weights γ j are ϕ j (z) = γ 1 = ϕ 2 1 e (1 θ)z 0 γ 2 = ϕ 3 + 1 2 ϕ 2 θ j 1 γ 3 = ϕ 4 + ϕ 3 + 1 3 ϕ 2 dθ, j 1, (j 1)! γ 4 = ϕ 5 + 3 2 ϕ 4 + 11 12 ϕ 3 + 1 4 ϕ 2 γ 5 = ϕ 6 + 2ϕ 5 + 7 4 ϕ 4 + 5 6 ϕ 3 + 1 5 ϕ 2 5

Examples of exponential Adams methods k = 1: exponential Euler method k = 2: u n+1 = u n + hϕ 1 ( ha)f(u n ) u n+1 = u n + hϕ 1 ( ha)f(u n ) + hϕ 2 ( ha) ( g(u n ) g(u n 1 ) ) interpretation as corrected exponential Euler step implementation can take advantage of j G n =O(h j ) H., Lubich, Selhofer 1998: exp4 methods are not invariant under linearization 6

Starting values use interpolation polynomial p in ( t0, g(u 0 ) ),..., ( t k 1, g(u k 1 ) ) within v.o.c. formula over interval of length mh for m = 1,..., k 1 approximate k 1 u m = u 0 + mhϕ 1 ( mha)f(u 0 ) + h σ m,l ( ha) l G 0, Calvo and Palencia, 2006 solve nonlinear system for u 1,..., u k 1 using fixed point iteration l=1 7

Linearized exponential multistep methods construction involves two steps linearize u = F(u) in each step at u n u(t n ) to get u (t) = J n u(t) + g n ( u(t) ) with J n = F u (u n), g n (u) = F(u) J n u apply explicit exponential integrator to linearized problem generalization to non-autonomous problems (not in this talk) H., Ostermann, Schweitzer (2006, 2009), Tokman (2006) 8

Examples k = 1: linearized exponential Euler method or exponential Rosenbrock Euler method u n+1 = u n + hϕ 1 (hj n )F(u n ) second-order convergent (H., Ostermann, Schweitzer, 2009) k = 2: Tokman 2006 u n+1 = u n + h ϕ 1 (hj n )F(u n ) 2h 3 ϕ 2(hJ n ) ( g n (u n ) g n (u n 1 ) ) third-order convergent (H., Ostermann, 2010) k = 2: u n+1 = u n + hϕ 1 (hj n )F(u n ) 2hϕ 3 (hj n ) ( g n (u n ) g n (u n 1 ) ) third-order convergent (H., Ostermann, 2010) 9

Linearized exponential multistep methods general construction of higher order methods v.o.c. formula applied to linearized ode yields exploit u (t) = J n u(t) + g n ( u(t)) h u(t n+1 ) = e hj n u(t n ) + e (h τ)j ( n g n u(tn + τ) ) dτ 0 g n u (u n) = 0, by approximating g n by Hermite interpolation polynomial p n of degree k satisfying p n (t n ) = 0 10

Linearized exponential Adams method, cont d Hermite interpolation polynomial p n of degree k interpolating in ( tn k+1, g n (u n k+1 ) ),..., ( t n, g n (u n ) ) and satisfying p n(t n ) = 0 k 1 p n (t n + θh) = G n,n + j=1 ( θ ( 1) j+1 θ j ) j 1 l l G n,n, l=1 where G n,m = g n (u m ) and j G n,m denotes the jth backward difference (w.r.t. m) 11

Linearized exponential Adams method, cont d inserting Hermite interpolation polynomial into v.o.c. formula with weights 1 u n+1 = e hj n u n + h e h(1 θ)j n p n (t n + θh)dθ 0 k 1 = u n + hϕ 1 (hj n )F(u n ) + h j=1 γ 2 = 2ϕ 3 γ 3 = 3ϕ 4 ϕ 3 γ 4 = 4ϕ 5 3ϕ 4 2 3 ϕ 3 γ j+1 (hj n ) γ 5 = 5ϕ 6 6ϕ 5 11 4 ϕ 4 1 2 ϕ 3 j 1 l l G n,n l=1 12

Convergence assumptions on semilinear initial value problem u (t) = F ( u(t) ) = Au(t) + g ( u(t) ), u(0) = u 0 X Banach space, A : X X e ta X X + t γ A γ e ta X X C γ, γ, t 0. g locally Lipschitz-continuous in a strip along the exact solution f (t) = g ( u(t) ) sufficiently smooth 13

Convergence theorem u (t) = Au(t) + g(u(t)) exponential (k + 1)-step Adams method f (t) = g ( u(t) ) satisfies f C k+1 ([0, T ], X ) Then, if the error bound u j u(t j ) V c 0 h k+1, j = 1,..., k, u n u(t n ) V C h k+1 sup f (k+1) (t) 0 t t n holds uniformly in 0 nh T. C = C(T ), independent of n and h H., Ostermann, 2010 14

Convergence theorem u (t) = Au(t) + g(u(t)) linearized exponential k-step Adams method f (t) = g ( u(t) ) satisfies f C k+1 ([0, T ], X ) Then, if the error bound u j u(t j ) V c 0 h k+1, j = 1,..., k 1, u n u(t n ) V C h k+1 sup 0 t t n ( f (k+1) (t) + u (k+1) (t) V ) holds uniformly in 0 nh T. C = C(T ), independent of n and h H., Ostermann, 2010 14

Outline of proof using interpolation in exact data, exact solution satisfies with defect 1 u(t n+1 ) = e hj n u(t n ) + h e h(1 θ)j n p n (t n + θh) + δ n+1 0 1 δ n+1 = h e h(1 θ)j ( n f n (t n + θh) p n (t n + θh) ) dθ 0 estimating interpolation error yields the bounds f n (t n + θh) p n (t n + θh) δ n+1 Ch k+2, δ n+1 V Ch k+2 α 15

Outline of proof, cont. error recursion 1 e n+1 = e hj n e n + h e h(1 θ)j ( n p n (t n + θh) p n (t n + θh) ) dθ 0 δ n+1 stability is not trivial (H., Ostermann, Schweitzer 2009) employ Lipschitz condition g(u j ) g ( u(t j ) ) L e j V e n V C max e j V + Ch j=1,...,k 1 n 1 j=0 1 t α n j ( ej V + h k+1) stated error bound follows from a discrete Gronwall lemma 16

Numerical example U t 2 U x 2 = 1 + Φ(x, t), x, t [0, 1] 1 + U2 hom. Dirichlet b.c., Φ s.t. U(x, t) = x(1 x) e t, N = 200 grid points error 10 2 10 4 10 6 10 8 10 10 10 12 10 14 h h 2 h 3 h 4 h 5 h 6 10 2 step size 10 1 Adams k = 1,..., 6 linearized Adams k = 1,..., 5 17

Implementation issues error analysis shows that j G n = O(h j ) for sufficiently smooth solutions Krylov subspace methods become cheaper with increasing j (Tokman, 2006) higher order does not cost very much can be exploited by reformulation of linearized exponential Adams methods k 1 u n+1 = u n + hϕ 1 (hj n )F(u n ) + h j=1 γ j+1 (hj n ) j 1 l l G n,n l=1 k 1 = u n + hϕ 1 (hj n )F(u n ) + h β k,l (hj n ) l G n,n l=1 18

Multiple time stepping construction of exponential Adams methods was based on replacing nonlinearity g by local interpolation polynomial p n in v.o.c. formula h u(t n+1 ) = e ha u(t n ) + 0 e (h τ)a g ( u(t n + τ) ) dτ h u n+1 = e ha u n + e (h τ)a p n (t n + τ)dτ 0 different interpretation: u n+1 = y n (h) exact solution of y n(τ) = Ay n (τ) + p n (t n + τ), y n (0) = u n, multiple time stepping: solve this ode by smaller time steps Grote, Mitkova, 2009, 2010 analogously for linearized exponential Adams methods 19

Numerical example, 2d U t U = 1 1 + U 2, x [0, 1]2, t [0, 0.2] hom. Dirichlet b.c., N = 75 grid points in each direction 10 4 j = 1 j = 2 10 5 j = 3 j = 4 10 6 10 7 10 8 0 10 20 30 40 norms of jth backward differences 45 30 20 10 5 3 2 1 0 10 20 30 40 Krylov dimensions j = 0 20

Partitioned problems partitioned semilinear problem [ ] [ ] [ v L Z v = w Y B w ] [ a(t, v, w) + b(t, v, w) ] [ v = A w ] + g(t, v, w) with L corresponding to stiff components, B to nonstiff components [ ] [ ] [ ] [ ] L 0 0 Z v a(t, v, w) A =, g(t, v, w) = + Y 0 0 B w b(t, v, w) matrix functions [ φ( ha) = φ(hl) 0 hy φ [1] (hl) φ(0)i ], φ [1] φ(z) φ(0) (z) = z attractive, if number of stiff components small 21

Application: Simulation of optical resonators optical discs and ring resonators have become essential building blocks of integrated all-optical circuits filtering multiplexing dispersion compensation advanced sensing challenges: long time simulation necessary to extract full spectral behavior of resonator complex geometries joint work with Kurt Busch, Jens Niegemann, KIT Photonics group and Abdullah Demirel 22

1d test problem Maxwell s equation discontinuous Galerkin discretization, order 4 for E x and B z locally refined grid 10 3 p=2 10 3 p=4 10 3 p=6 10 4 10 4 10 4 10 5 0.2 0.3 0.4 0.6 0.8 1 10 5 0.2 0.3 0.4 0.6 0.8 1 10 5 0.2 0.3 0.4 0.6 0.8 1 error versus coarse grid step size, x fine = x coarse /p 23

Mathematical model of ring resonator modeling and spatial discretization by Kurt Busch and Jens Niegemann Maxwell s equation in 2d PML for transparent boundary conditions locally refined grid discontinuous Galerkin discretization 10 d.o.f. per triangle for each of E x, E y, B z total of 38.760 d.o.f. 300 components are considered as stiff small matrix functions computed via diagonalization work in progress... only preliminary results 24

Histogram of radii of triangles incircles 150 100 Count 50 0 0 0.05 0.1 0.15 0.2 0.25 Radius of incircle 25

Eigenvalues 200 100 0 100 200 400 300 200 100 0 26

Spatial grid gain about 30% of computational time against explicit RK method 27

28 Simulation snapshots

Concluding remarks Summary new class of exponential multistep methods convergence analysis of two classes of exponential multistep methods link to multiple time stepping or explicit local time stepping by Mitkova and Grote application to optical resonators Outlook optimize methods for ring resonator simulation different implementations more comparisons with existing methods multiple time stepping rational Krylov methods preconditioning 29