Blanca Bujanda, Juan Carlos Jorge NEW EFFICIENT TIME INTEGRATORS FOR NON-LINEAR PARABOLIC PROBLEMS

Similar documents
Some Applications of Fractional Step Runge-Kutta Methods

Symmetry Labeling of Molecular Energies

Flavius Guiaş. X(t + h) = X(t) + F (X(s)) ds.

The Laplace equation, cylindrically or spherically symmetric case

Stability properties of a family of chock capturing methods for hyperbolic conservation laws

5 Ordinary Differential Equations: Finite Difference Methods for Boundary Problems

Physically Based Modeling: Principles and Practice Implicit Methods for Differential Equations

Chapter 4: Numerical Methods for Common Mathematical Problems

ETNA Kent State University

New Fourth Order Quartic Spline Method for Solving Second Order Boundary Value Problems

An Approximation to the Solution of the Brusselator System by Adomian Decomposition Method and Comparing the Results with Runge-Kutta Method

Implicit-explicit variational integration of highly oscillatory problems

Math 31A Discussion Notes Week 4 October 20 and October 22, 2015

Numerical Differentiation

FEM solution of the ψ-ω equations with explicit viscous diffusion 1

Copyright c 2008 Kevin Long

Dedicated to the 70th birthday of Professor Lin Qun

LECTURE 14 NUMERICAL INTEGRATION. Find

Polynomial Interpolation

1. Introduction. We consider the model problem: seeking an unknown function u satisfying

MANY scientific and engineering problems can be

Numerical Solution of One Dimensional Nonlinear Longitudinal Oscillations in a Class of Generalized Functions

Finite Difference Method

Consider a function f we ll specify which assumptions we need to make about it in a minute. Let us reformulate the integral. 1 f(x) dx.

Click here to see an animation of the derivative

POLYNOMIAL AND SPLINE ESTIMATORS OF THE DISTRIBUTION FUNCTION WITH PRESCRIBED ACCURACY

Research Article Cubic Spline Iterative Method for Poisson s Equation in Cylindrical Polar Coordinates

1 The concept of limits (p.217 p.229, p.242 p.249, p.255 p.256) 1.1 Limits Consider the function determined by the formula 3. x since at this point

Math 102 TEST CHAPTERS 3 & 4 Solutions & Comments Fall 2006

N igerian Journal of M athematics and Applications V olume 23, (2014), 1 13

DELFT UNIVERSITY OF TECHNOLOGY Faculty of Electrical Engineering, Mathematics and Computer Science

Continuity and Differentiability Worksheet

Recent Progress in the Integration of Poisson Systems via the Mid Point Rule and Runge Kutta Algorithm

Fourier Type Super Convergence Study on DDGIC and Symmetric DDG Methods

Differential equations. Differential equations

Continuity and Differentiability of the Trigonometric Functions

Third order Approximation on Icosahedral Great Circle Grids on the Sphere. J. Steppeler, P. Ripodas DWD Langen 2006

Polynomial Interpolation

Lecture 2: Symplectic integrators

ERROR BOUNDS FOR THE METHODS OF GLIMM, GODUNOV AND LEVEQUE BRADLEY J. LUCIER*

Some Review Problems for First Midterm Mathematics 1300, Calculus 1

A method of Lagrange Galerkin of second order in time. Une méthode de Lagrange Galerkin d ordre deux en temps

Parameter Fitted Scheme for Singularly Perturbed Delay Differential Equations

Flapwise bending vibration analysis of double tapered rotating Euler Bernoulli beam by using the differential transform method

On convexity of polynomial paths and generalized majorizations

1. Consider the trigonometric function f(t) whose graph is shown below. Write down a possible formula for f(t).

ch (for some fixed positive number c) reaching c

SECTION 3.2: DERIVATIVE FUNCTIONS and DIFFERENTIABILITY

MATH745 Fall MATH745 Fall

Order of Accuracy. ũ h u Ch p, (1)

CS522 - Partial Di erential Equations

lecture 26: Richardson extrapolation

Parabolic PDEs: time approximation Implicit Euler

Taylor Series and the Mean Value Theorem of Derivatives

Weak second order explicit exponential Runge-Kutta methods for stochastic differential equations

3 Parabolic Differential Equations

Combining functions: algebraic methods

Influence of the Stepsize on Hyers Ulam Stability of First-Order Homogeneous Linear Difference Equations

A CLASS OF EVEN DEGREE SPLINES OBTAINED THROUGH A MINIMUM CONDITION

New Streamfunction Approach for Magnetohydrodynamics

MA455 Manifolds Solutions 1 May 2008

Differentiation in higher dimensions

1. Introduction. Consider a semilinear parabolic equation in the form

Finite Difference Methods Assignments

1 + t5 dt with respect to x. du = 2. dg du = f(u). du dx. dg dx = dg. du du. dg du. dx = 4x3. - page 1 -

Online Appendix for Lerner Symmetry: A Modern Treatment

MATH1131/1141 Calculus Test S1 v8a

Solutions to the Multivariable Calculus and Linear Algebra problems on the Comprehensive Examination of January 31, 2014

Application of numerical integration methods to continuously variable transmission dynamics models a

3.4 Worksheet: Proof of the Chain Rule NAME

A First-Order System Approach for Diffusion Equation. I. Second-Order Residual-Distribution Schemes

MAT244 - Ordinary Di erential Equations - Summer 2016 Assignment 2 Due: July 20, 2016

2.8 The Derivative as a Function

Generic maximum nullity of a graph

4. The slope of the line 2x 7y = 8 is (a) 2/7 (b) 7/2 (c) 2 (d) 2/7 (e) None of these.

HOMEWORK HELP 2 FOR MATH 151

Exam 1 Review Solutions

The Verlet Algorithm for Molecular Dynamics Simulations

Numerical Analysis MTH603. dy dt = = (0) , y n+1. We obtain yn. Therefore. and. Copyright Virtual University of Pakistan 1

OSCILLATION OF SOLUTIONS TO NON-LINEAR DIFFERENCE EQUATIONS WITH SEVERAL ADVANCED ARGUMENTS. Sandra Pinelas and Julio G. Dix

arxiv: v1 [math.na] 9 Sep 2015

AN ANALYSIS OF NEW FINITE ELEMENT SPACES FOR MAXWELL S EQUATIONS

Differential Equations

Quasiperiodic phenomena in the Van der Pol - Mathieu equation

Math 212-Lecture 9. For a single-variable function z = f(x), the derivative is f (x) = lim h 0

Approximation of the Viability Kernel

Exercises for numerical differentiation. Øyvind Ryan

Digital Filter Structures

Lecture XVII. Abstract We introduce the concept of directional derivative of a scalar function and discuss its relation with the gradient operator.

lecture 35: Linear Multistep Mehods: Truncation Error

Two Step Hybrid Block Method with Two Generalized Off-step Points for Solving Second Ordinary Order Differential Equations Directly

Preconditioning in H(div) and Applications

Notes on Multigrid Methods

Math 312 Lecture Notes Modeling

2.3 More Differentiation Patterns

More on generalized inverses of partitioned matrices with Banachiewicz-Schur forms

Energy-preserving variant of collocation methods 1

LIMITATIONS OF EULER S METHOD FOR NUMERICAL INTEGRATION

INTRODUCTION AND MATHEMATICAL CONCEPTS

Efficient algorithms for for clone items detection

Transcription:

Opuscula Matematica Vol. 26 No. 3 26 Blanca Bujanda, Juan Carlos Jorge NEW EFFICIENT TIME INTEGRATORS FOR NON-LINEAR PARABOLIC PROBLEMS Abstract. In tis work a new numerical metod is constructed for time-integrating multidimensional parabolic semilinear problems in a very efficient way. Te metod reaces te fourt order in time and it can be combined wit standard spatial discretizations of any order to obtain unconditinally convergent numerical algoritms. Te main teoretical results wic guarantee tis property are explained ere, as well as te metod caracteristics wic guarantee a very strong reduction of computational cost in comparison wit classical discretization metods. Keywords: fractional step metods, non-linear parabolic problems, convergence. Matematics Subject Classification: 65M6, 65M12. 1. INTRODUCTION In tis paper we deal wit te construction of numerical metods wic efficiently integrate te following non-linear parabolic problems: find y(t) : [t, T ] H, a solution of { y (t) = L(t)y(t) + f(t) + g(t, y(t)), y(t ) = y. Here H is a Hilbert space of functions defined on a certain domain Ω R n, L(t) is a linear second order elliptic differential operator, f(t) is te source term, g(t, y(t)) is a non-linear function wic meets some additional requirements and y y ( x) is te initial condition. Te numerical integration of tis problem can be viewed as a double process of discretization wic adequately combines a time integrator (typically an adapted ODE Solver) wit a discretization of te spatial variables. It is well known tat te use of classical implicit scemes for te numerical integration of tis type of problems poses mainly two problems. On te one and, te convergence of te numerical sceme is usually obtained by imposing excessively restrictive stability requirements wic force (1) 47

48 Blanca Bujanda, Juan Carlos Jorge us to use low order metods (like te implicit Euler rule) or complicated fully implicit metods for reacing iger orders of convergence in time. On te oter and, we must use slow iterative metods to resolve te internal stage equations because tey are defined as te solution of large non-linear systems. Also, in tis type of scemes, te convergence is deduced by imposing strong stability requirements (for example, B-stability). In [5] a semiexplicit metod, wic is used to determinate te contribution of te linear term, is combined in an additive way wit an explicit metod used for te contribution of te non-linear term. Tus, linear systems only appear in te computation of te internal stages. In [7], splitting metods are applied in te numerical integration of some non-linear advention-diffusion-reaction problems; ere te advention terms are integrated wit an explicit metod. In order to avoid tese problems, in [3] a new class of linearly implicit metods is developed; tey are built by combining a standard spatial discretization of Finite Differences or Finite Elements type wit a suitable time discretization, wic is deduced from a Fractional Step Runge Kutta (FSRK) metod to deal wit te linear terms and an explicit Runge Kutta metod to define te contribution of te non-linear term. By applying tis new linearly implicit metods (see [1, 2, 3]) we obtain numerical algoritms wic are convergent under a single requirement of te linear absolute stability type. It is well known (see [8]) tat te Fractional Step metods are very efficient in integrating multidimensional linear parabolic problems if we suitably decompose te linear elliptic operator and te source term; te advantages of tese metods manifest temselves also in tese new linearly implicit metods. Tus in te resolution of te internal stages only we must solve linear systems wit very simple matrices and it is not necessary to apply classical iterative metods to solve tem. For example, if te elliptic linear operator L(t) is n d i(t) 2 y x 2 i + n v i(t) y x i + n k i(t), and we realize a spatial discretization by applying classical Finite Differences, te matrices wic appear at eac stage are tridiagonal; consequently, te computational cost of te final metod is of te same order as tat of an explicit metod. 2. NUMERICAL SCHEME By applying firstly a spatial discretization process of type Finite Difference, Finite Elements,... to problem (1) we obtain a family of Initial Value Problems wic depend on te parameter (, ] used to discretize in space as follows: Find Y (t) : [t, T ] V, a solution of { Y (t) = L (t)y (t) + f (t) + g (t, Y (t)), Y (t ) = y. (2) For eac it is common to consider a finite dimensional space V ; suc space is te space of discrete functions on a mes in Finite Differences, te subspace of H of piecewise polynomial functions in a classical Finite Element discretization. We

New efficient time integrators for non-linear parabolic problems 49 suppose tat suc spatial semidiscretization process is uniformly convergent of order q, i.e., for sufficiently smoot functions y(t) tere is π y(t) Y (t) C q, t [t, T ]. (3) Here, te operator π denotes te restriction to te mes nodes if Finite Differences used and if we use te standard Finite Elements, ten π are suitable projections in te space V. In order to integrate efficiently tis problem by applying te linearly implicit numerical metods wic are first appeared in [3], we must adequately decompose te spatial discretization of te elliptic operator and te source term in n addends as follows: L (t) = n L i(t) and f (t) = n f i(t); te FSRK metod will act on tese terms. Te contribution of te non-linear term g (t, U (t)) is given by te explicit RK metod. From tis process, we obtain te following sceme: s Y m+1 = Y m +τ were i Y m,i = Y m+τ b ki i ( ) L ki(t m,i )Y m,i + f ki(t m,i ) + τ s ( ) i 1 a kj ij L kj(t m,j )Y m,j + f kj(t m,j ) + τ b n+1 i g (t m,i, Y m,i ), (4) a n+1 ij g (t m,j, Y m,j ). In tis algoritm, we denote wit Y m te numerical approximations to te exact solution y(t m ) at te times t m = t + m τ (τ is te time step), U = U (t ), k i, k j {1,, n}, n is te number of te levels of te metod, t m,i = t + (m + c i ) τ and te intermediate approximations U m,i for i = 1,, s are te internal stages of te metod. By setting some coefficients to zero in te latest formula as follows A k = (a k ij) were a k ij = b k = (b k j ) were b k j = a n+1 ij if k = n + 1 y i > j, a kj ij if k = k j and i j, oterwise, b n+1 j if k = n + 1, b kj j if k = k j, oterwise, we can organize tese coefficients in a table wic can be considered as an extension of te tables introduced by Butcer for standard RK metods, in te following way C e A 1 A 2... A n A n+1 (b 1 ) T (b 2 ) T... (b n ) T (b n+1 ) T were C = diag(c 1,, c s ) and e = (1,, 1).

41 Blanca Bujanda, Juan Carlos Jorge We will use (C, A i, b i ) n+1 to refer ro tis metod sortly and (C, Ai, b i ) to denote every standard RK metod involved in (4). To study te convergence of numerical sceme (4) we introduce, as usual, te global error at te time t = t m by E m = π y(t m ) Y m ; tus we say tat (4) is uniformly convergent of order p in time and of order q in space, if E m satisfies E m C( q + τ p ). (5) To get tis bound, we begin by studying te stability of numerical sceme (4); tus we rewrite it in a more compact way by using te following tensorial notation: given M (m ij ) R s s and v (v i ) R s, we denote M (m ij I V ) V s s and v (v i I V ) V s; we group te evaluations of te source terms f i(t) and of te linear operators L i (t) for i = 1,, n and for all m = 1, 2,, Fi m = (f i(t m,1 ),, f i (t m,s )) T V s and ˆL m i = diag (L i(t m,1 ),, L i (t m,s )) ( ) T L(V, V ). We also group te contribution of te stages Ỹ m = Y m,1,, Y m,s ( T V s and te nonlinear term Ĝm (Ỹ m) = g (t m,1, Y m,1 ),, g (t m,s, Y m,s )) V s. In tis way we obtain n n (Ī τ A i ˆL m m i )Ỹ = ē Y m + τ A i Fi m + τ A n+1 Ĝm (Ỹ m ), Y m+1 = Y m + τ n (b i ) T (ˆLm i Ỹ m + F m i ) + τ (b n+1 ) T Ĝ m (Ỹ m ). Wen te operator ( Ī n Aj ˆL j) m is invertible 1), te numerical solution can be written as were Y m+1 = R ( τ ˆL m 1,, τ ˆL m n) Y m + S ( τ ˆL m 1,, τ ˆL m n, τf m 1,, τf m n + T (τ ˆL m 1,, τ ˆL m n, τ Ĝm (Ỹ m )), (6) R(τ ˆL m 1,, τ ˆL m n) = Ī + n (b i ) T τ ˆL n 1ē, m i (Ī A j τ ˆL j) m (7) ) and S ( τ ˆL m 1,, τ ˆL m n, τ F1, m, τ Fn) m = (8) n ( = τ (b i ) T Fi m + ˆL m (Ī n i τ A j ˆL m ) 1 ( n j τ A k Fk m ) ) T (τ ˆL m 1,, τ ˆL m n, τ Ĝm (Ỹ m )) = (9) n = τ (b i ) T (Ī n ˆLm i A j ˆL m 1τA n+1 j) Ĝ m (Ỹ m ) + τ (b n+1 ) T Ĝ m (Ỹ m ). k=1 1) In [4] it is proved tat tis operator is invertible and tat its inverse operator is bounded independently of and τ.

New efficient time integrators for non-linear parabolic problems 411 Note tat first term (7) is te transition operator if we consider linear omogeneous problems; te contribution of te source term appears in second term (8) and, finally, te contribution of te nonlinear term g (t, Y (t)) is defined by tird addend (9). To study te stability of te numerical sceme we must bound tese tree addends (see [1]). Te contribution of te linear terms, given by (7), can be bounded by applying te following result (see [4]): Teorem 2.1. Let an A-stable FSRK metod be given suc tat all teir stages are implicit (i.e., n k=1 ak ii, for i = 1,, s) or its first stage only is explicit (i.e., n k=1 ak ii, for i = 2,, s and aki 11 = for all i = 1,, s) and it also satisfies (,,, 1) T A i = (b i ) T and a k1 ss, and let {L i (t)} n be a linear maximal coercive system of operators suc tat: a) for eac t [, T ] te system of operators {L i (t)} n is commutative, b) tere exist n constants M i suc tat L i (t )Y L i (t)y t t C L i (t)y t, t [t, T ]. Ten tere exists a constant γ, independent of τ, suc tat R( τ ˆLm 1,, τ ˆL m n) e γτ (1) olds. In [4], te following bound for second addend (8) is also proved to old: S( τ ˆLm 1,, τ ˆL m n, τ F m 1,, τ F m n) C τ n F m i To finis te study of te stability, we must bound tird term (9); in [1], te following result is proved, rewritten for our purposes: Teorem 2.2. Let (4) be a linearly implicit metod meeting te conditions of Teorem 2.1. If we use suc metod to integrate nonlinear parabolic problem (2) in time, were te nonlinear part g (t, Y ) satisfies g (t, X ) g (t, Y ) L X Y, t [t, T ], ten:. T (τ ˆL m 1,, τ ˆL m n, τ Ĝm ( X m )) T (τ ˆL m 1,, τ ˆL m n, τ Ĝm (Ỹ m )) C τ X m Y m. (11) To study te consistency of te numerical sceme, we define te local error of te time discretization metod as e m = Y (t m ) Ŷ m, were Ŷ m is obtained wit one step of (4) taking as starting value Y m 1 = Y (t m 1 ). We say tat (4) is uniformly consistent of order p if e m C τ p+1, m. (12) In [2] a teorem wic can be rewritten for tis case as follows is proved:

412 Blanca Bujanda, Juan Carlos Jorge Teorem 2.3. Let us assume tat problem (2) satisfies te smootness requirements Y (p+1) i 1 (t) C, L (ϱ1) i 1 (t) L(ϱ l 1) i l 1 (t)y (ϱ l) i l (t) C, t [t, T ], l {2,, p}, (i 1,, i l ) {1,, n + 1} l, (ϱ 1,, ϱ l 1 ) {,, p 1} l 1 and ϱ l {r + 1,, p} suc tat 2 l + l ϱ k p + 2, were Y i (t) = L i(t)y (t) + f i (t), i = 1,, n + 1 wit L (n+1) (t) = g Y (t, Y (t)) and f (n+1) (t) = g Y (t, Y (t)) Y (t) + g (t, Y (t)). Let (C, A i, b i ) n+1 be a linearly implicit FSRK metod satisfying te reductions (C) k e k A j (C) k 1 e =, j = 1,, n + 1, k = 1,, r wit r = E [ ] p 1 2 togeter wit te order conditions (b i1 ) T (C) ρ1 e = 1 ρ 1 + 1, Ten (12) is true. k=1 (b i1 ) T (C) ρ1 A i2 (C) ρ2 A i l (C) ρ l e = l 1 (l j + 1) + l {2,, p}, (i 1,, i l ) {1,, n + 1} l, (ρ 1,, ρ l 1 ) {,, p 1} l 1 and ρ l {r + 1,, p 1} suc tat 1 l +, l ρ k k=j l ρ k p. To obtain te convergence of te sceme (bound (5)), we decompose te global error as follows: k=1 E m π y(t m ) Y (t m ) + Y (t m ) Ŷ m + Ŷ m Y m were we ave used te intermediate approximations Y (t m ) and Ŷ m. Te space discretization is convergent of order q, tus by applying (3) we obtain E m C q + e m + Ŷ m Y m. Now we use te fact tat te time discretization is consistent of order p (i.e. it satisfies (12)) E m C q + C τ p+1 + Ŷ m Y m. Finally, we bound te tird addend it by using (6) (9) as follows Ŷ m Y m = R ( τ ˆL m 1,, τ ˆL m n) ( π y(t m 1 ) Y (t m 1 ) ) + + T (τ ˆL m 1,, τ ˆL m n, τ Ĝm (π y(t m 1 ))) T (τ ˆL m 1,, τ ˆL m n, τ Ĝm (Y m 1 )).

New efficient time integrators for non-linear parabolic problems 413 Tus, by applying (1) and (11), we obtain E m C q + C τ p+1 + ( e γ τ + C τ ) E m 1. From tis result is easy to conclude tat sceme (4) satisfies (5). 3. A NEW FOURTH ORDER LINEARLY IMPLICIT METHOD Now we present te main ideas of te construction process for a linearly implicit tree-level FSRK metod; suc metod attains fourt order in time wen it is applied to te numerical integration of te semilinear problems of type (1) combined wit spatial discretizations. To te best our knowledge, tis is te only fourt order metod of tis class wic appears in te literature. Te construction process for tis kind of metods is very complicated due to two main drawbacks; te first one is related to te ig number of order conditions tat we must impose, wic makes us cope wit large and complicated non-linear systems. Te second one is related to te additional requirements tat we must impose in order to guarantee te stability of te final metod. Precisely, te order conditions tat we must impose are (see [2]): (r i) A i e = C e for i = 1, 2, 3 (α 1) (b 1 ) T e = 1 (α 2) (b 1 ) T C e = 1 2 (α 3) (b 1 ) T A 1 C e = 1 6 (α 4) (b 1 ) T C C e = 1 3 (α 5) (b 1 ) T C C C e = 1 4 (α 6) (b 1 ) T C A 1 C e = 1 8 (α 7) (b 1 ) T A 1 C C e = 1 12 (α 8) (b 1 ) T A 1 A 1 C e = 1 24 (β 1) (b 2 ) T e = 1 (β 2) (b 2 ) T C e = 1 2 (β 3) (b 2 ) T A 2 C e = 1 6 (β 4) (b 2 ) T C C e = 1 3 (β 5) (b 2 ) T C C C e = 1 4 (β 6) (b 2 ) T C A 2 C e = 1 8 (β 7) (b 2 ) T A 2 C C e = 1 12 (β 8) (b 2 ) T A 2 A 2 C e = 1 24 (γ 1) (b 3 ) T e = 1 (γ 2) (b 3 ) T C e = 1 2 (γ 3) (b 3 ) T A 3 C e = 1 6 (γ 4) (b 3 ) T C C e = 1 3 (γ 5) (b 3 ) T C C C e = 1 4 (γ 6) (b 3 ) T C A 3 C e = 1 8 (γ 7) (b 3 ) T A 3 C C e = 1 12 (γ 8) (b 3 ) T A 3 A 3 C e = 1 24 ( ij) (b i ) T A j C e = 1 6 for {i, j} {1, 2, 3} ( ij) (b i ) T C A j C e = 1 8 for {i, j} {1, 2, 3} ( ij) (b i ) T A j C C e = 1 12 for {i, j} {1, 2, 3} ( ijk ) (b i ) T A 1j A k C e = 1 24 for {i, j, k} {1, 2, 3} One may ceck tat we need at least nine stages to obtain a linearly implicit metod wic fulfils tese order conditions. Restrictions (r 1 ), (r 2 ) and (r 3 ) require te assumption tat te first stage of te metod (C, A 1, b 1 ) is explicit; we must note tat in tis way te computational cost of te final metod will be similar to te cost of an eigt stage metod wit all stages implicit. To obtain a stable metod, it is convenient to impose te restrictions (,,, 1) T A i = (b i ) T for i = 1, 2, 3 (see [3]); in order to get te numerical solution directly from te calculation of te last stage Y m+1 (= Y m,9 ) and, consequently, to reduce ligtly te computational cost of te final metod, we impose te same restriction on te tird level (,,, 1) T A 3 = (b 3 ) T.

414 Blanca Bujanda, Juan Carlos Jorge Te computational cost of te final metod can be reduced (wen te coefficients of te linear operator do not depend on time) by assuming tat a 1 ii = a2 jj = a for i = 3, 5, 7, 9 and for j = 2, 4, 6, 8; besides, using tese restrictions we simplify te study of te stability of te FSRK metod. Tus, te table of te deduced metod will ave te following structure: C e A 1 A 2 (b 1 ) T (b 2 ) T = = c 2 c 3 c 4 c 5 c 6 c 7 c 8 c 9 a 1 21 a a 1 31 a a 2 32 a 1 41 a 1 43 a 2 42 a a 1 51 a 1 53 a a 2 52 a 2 54 a 1 61 a 1 63 a 1 65 a 2 62 a 3 64 a a 1 71 a 1 73 a 1 75 a a 2 72 a 3 74 a 3 74 a 1 81 a 1 83 a 1 85 a 1 87 a 2 82 a 3 84 a 3 84 a b 1 1 b 1 3 b 1 5 b 1 7 a b 2 2 b 2 4 b 2 6 b 2 8 b 1 1 b 1 3 b 1 5 b 1 7 a b 2 2 b 2 4 b 2 6 b 2 8 A 3 (b 3 ) T = a 3 21 a 3 31 a 3 32 a 3 41 a 3 42 a 3 43 a 3 51 a 3 52 a 3 53 a 3 54 a 3 61 a 3 62 a 3 63 a 3 64 a 3 65 a 3 71 a 3 72 a 3 73 a 3 74 a 3 75 a 3 76 a 3 81 a 3 82 a 3 83 a 3 84 a 3 85 a 3 86 a 3 87 b 3 1 b 3 2 b 3 3 b 3 4 b 3 5 b 3 6 b 3 7 b 3 8 b 3 1 b 3 2 b 3 3 b 3 4 b 3 5 b 3 6 b 3 7 b 3 8 We begin te construction process for te metod by resolving te following order conditions: (r 1 ), (r 2 ), (α 1 ),, (α 8 ), (β 1 ),, (β 8 ), ( 12 ), ( 21 ), ( 21 ), ( 12 ), ( 112 ), ( 121 ) ( 211 ), ( 221 ), ( 212 ) and ( 122 ); wic involve te coefficients of te FSRK metods (C, A i, b i ) 2 only. Tus we obtain a four order family of FSRK metods depending on te free parameters a, c 4, c 5, c 6, c 7, a 1 65, a 1 83 and a 1 87. To perform, in an efficient way, te numerical integration of te linear stiff problems, it is necessary to impose additional restrictions of type A-stability. In order to introduce tis concept in te simplest way we apply a two-level FSRK metod to te numerical integration of te following test problem y (t) = (λ 1 + λ 2 ) y(t) wit Re(λ i ), for i = 1, 2;

New efficient time integrators for non-linear parabolic problems 415 obtaining te recurrence: y m+1 = ( 1 + 2 τ λ i (b i ) T ( I 2 τ λ j A j) 1 e )y m. By substituting in tis expression z i for τ λ i for i = 1, 2 we obtain a rational function of two complex variables; wic we call it te Amplification Function associated wit te FSRK metod 2 ( 2 R(z 1, z 2 ) = 1 + z i (b i ) T I z j A j) 1 e. We say tat an FSRK metod is A-stable iff R(z 1, z 2 ) 1, z i C being Re(z i ), i = 1, 2. As te FSRK metod as nine stages, wit te first one of tem explicit, and attains fourt order, te amplification function of te metod can be written as follows: R(z 1, z 2 ) = R 1 (z 1 )R 2 (z 2 ) + Rest were R i (z i ) = 1 + (1 4 a) z i + 1 2 (1 68 a + 12 a2 )z 2 i + ( 1 6 2 a + 6 a2 4 a 3 )z 3 i (1 a z i ) 4 + + ( 1 24 2a 3 + 3 a2 4 a 3 + a 4 )z 4 i (1 a z i ) 4 are te amplification functions associated wit te eac standard RK metod (C, A i, b i ) for i = 1, 2 and were: Rest = e 41 z 4 1 z 2 + e 32 z 3 1 z 2 2 + e 23 z 2 1 z 3 2 + e 14 z 1 z 4 2 + e 42 z 4 1 z 2 2 + e 33 z 3 1 z 3 2 (1 az 1 ) 4 (1 az 2 ) 4 + + e 24 z 2 1 z 4 2 + e 43 z 4 1 z 3 2 + e 34 z 3 1 z 4 2 + e 44 z 4 1 z 4 2 (1 az 1 ) 4 (1 az 2 ) 4, e 41 = E 4,1 1 24, e 14 = E 1,4 1 24, e 23 = E 2,3 1 12, e 32 = E 3,2 1 12, e 42 = E 4,2 4 a e 32 4 a e 41 1 48, e 24 = E 2,4 4 a e 23 4 a e 14 1 48, e 33 = E 3,3 4 a e 23 4 a e 32 1 36, e 43 = E 4,3 4 ae 42 1 a 2 e 41 4 a e 33 16 a 2 e 32 1 a 2 e 23 1 144, e 34 = E 3,4 4 ae 24 1 a 2 e 14 4 a e 33 16 a 2 e 23 1 a 2 e 32 1 144, e 44 = E 4,4 4 a e 43 4 a e 34 1 a 2 e 42 2 a 3 e 41 16 a 2 e 33 4 a 3 e 32 1 a 2 e 24 4 a 3 e 23 2 a 3 e 14 1 576,

416 Blanca Bujanda, Juan Carlos Jorge were E i1,i 2 = (b k1 ) T A k2 A kj e k (k 1,,k j) {1,2} j n l ( k)=i l, l=1,2 wit n l ( k) te number of te times tat te index l appears in te integer vector k. Given tis FSRK metod of Alternating Directions type attains fourt order, and te standard RK metod (C, A 1, b 1 ) may be reduced to a DIRK metod of five stages wit te first stage explicit, te standard RK metod (C, A 2, b 2 ) may be reduced to a four-stage SDIRK metod, we can see in [6] tat suc metods (C, A i, b i ) for i = 1, 2 are L-stable for a =.57281662482134746. Taking into account tat te strong stability and, particularly, te L-stability are very interesting properties wen we carry out te numerical integration of te parabolic problems wit irregular or incompatible data, we ave fixed a =.57281662482134746. Now, if we were able to make null te Rest we would obtain an L-stable FSRK metod. We ave cecked tat it is impossible to cancel te Rest because by substituting te previous values we observe tat e 44 as te constant value.11591. Neverteless, we can determine te remaining free parameters in order to minimize to some extent te contribution of te Rest to te amplification function. Due to te complexity, size and nonlinearity of te last expressions, we ave taken a 1 87 = in order to sorten tem; now, we can ceck tat e 14 = e 41, e 23 = e 32, e 24 = e 42 and e 43 = e 34, simplifying te subsequent study. By cancelling e 43 = e 34 and e 33 we can fix te parameters a 1 65 and a 1 83. By using tis option, te remaining coefficients of te Rest ave te following values: e 14 =.165332, e 23 =.171413 and e 24 =.423441. Wit te imposed restrictions, ave obtained a family of FSRK metods wic depend on te parameters c 4, c 5, c 6 c 7, attain fourt order and are A-stable. In te last part of te construction of tis metod we solve te remaining conditions. In suc order conditions, some coefficients of te last explicit level (C, A 3, b 3 ) appear. Namely we must solve: (C, A 3, b 3 ): (r 3 ), (γ 1 ),, (γ 8 ), ( 13 ), ( 23 ), ( 31 ), ( 32 ), ( 13 ), ( 23 ), ( 31 ), ( 32 ), ( 13 ), ( 23 ), ( 31 ), ( 32 ), ( 113 ), ( 131 ), ( 311 ), ( 331 ), ( 313 ), ( 133 ), ( 223 ), ( 232 ), ( 322 ), ( 332 ), ( 323 ), ( 233 ), ( 123 ), ( 132 ), ( 213 ), ( 231 ), ( 321 ) and ( 312 ). By resolving tese order conditions we obtain a family of metods wic depend on te parameters: a 3 63, a 3 73, a 3 74, a 3 76, a 3 82, a 3 83, a 3 84, a 3 85, a 3 86, a 3 87, c 4, c 5, c 6 and c 7. Suc coefficients are fixed in order to minimize to some extent te main term of te local error and also to simplify coefficients. Tus, te final metod is given by: A 1 (b 1 ) T = 1.572816624821349.572816624821349.572816624821349.233624731962216.1279261371115 =.14324156253764.14324156253764.572816624821349.247525445891478.114613915837628.5337551373812817.16584521264111465.16898133629871276.5743927888629.572816624821349 B C @.519764686433798.51822334268295.4843426395887455 A.657833985428.2245756812245128.45353475817523.6313327555445881.572816624821349.657833985421.2245756812245128.45353475817523.6313327555445881.572816624821349

New efficient time integrators for non-linear parabolic problems 417 A 2 (b 2 ) T = 1.572816624821349 1.1456321249642698.239482729148816.572816624821349.1122415459243.685561679411592.53637655875728.597488254772145.572816624821349.684549593831673.37263822447982.6584879946324 B C @ 1.189389864279955.77389697436742.33636848563513.572816624821349 A 1.23598239745338 1.73598239745343 1.73598239745338 1.23598239745343 1.23598239745338 1.73598239745343 1.73598239745338 1.23598239745343 A 3 (b 3 ) T = B @.572816624821349 1.14563212496414.364583822964946.2596334954886.165341413273 1.59735621929421 2.3353773662664 1.31892539518593.457158517861478.25963349528255.2363466549123.23975256771974143.4271839375178651 1.23598239745865 1.73598239745232 1.23598239745865 1.73598239745232.188919479934955.14276583987761113 1.7359823974529 1.23598239745358 1.7359823974529 1.23598239745358 1 C A Ce=(,.5 728 16 624 821 349,1.1 456 321 249 642 698, 1 3,.5 728 16 624 821 349, 2 3, 1 3,.427 183 937 517 865,1)T. 4. NUMERICAL EXPERIMENT In tis section we sow a numerical test were we ave integrated te following non-linear convection-diffusion problem wit a non-linear reaction term r(y) = k 1 y + k 2 y + y3 (1+y 2 ) 2 y t = y v 1 y x1 v 2 y x2 r(y) + f, (x 1, x 2, t) Ω [, 5], y(x 1,, t) = y(x 1, 1, t) =, x 1 [, 1] and t [, 5], y(, x 2, t) = y(1, x 2, t) =, x 2 [, 1] and t [, 5], y(x 1, x 2, ) = x 3 1(1 x 1 ) 3 x 3 2(1 x 2 ) 3, (x 1, x 2 ) Ω, wit Ω = (, 1) (, 1), v 1 = 1 + x 1 x 2 e t, v 2 = 1 + x 1 t, k 1 = k 2 = 1+(x1+x2)2 e t 2 and f = 1 3 e t x 1 (1 x 1 )x 2 (1 x 2 ). We ave carried out te time discretization using te metod described in te previous section and we ave discretized in space on a rectangular mes by using a central difference sceme. Te combination of metods of finite differences type of order q in te spatial discretization stage and pt order time integrators of te type described in tis paper provides a totally discrete sceme wose global error is [y(t m )] Y m = O(q + τ p ); ere [v] denotes te restriction of te function v to te mes node.

418 Blanca Bujanda, Juan Carlos Jorge In order to apply te integration metod described ere, we ave taken for i = 1, 2 and j = 1,, 9 L i (t m,j )Y m,j = δ xix i Y m,j [v i ] δˆxi Y m,j g (t m,j, Y m,j ) = 3(Y m)2 (Y m)4 (1 + (Y m)2 ) 3 Y m,j (Y m,j ) 3 (1 + (Y m,j ) 2 ) 2 [k i ] Y m,j 3(Y m )2 (Y m )4 2(1 + (Y m )2 ) 3 Y m,j, were δ xix i and δˆxi denote te classical central differences. Note tat, by using tis decomposition in eac stage of (4) we must solve only one linear system wose matrix is tridiagonal; tus te computational complexity of computing te internal stages of tis metod is of te same order as in te cases of an explicit metod. Note tat tese operators L i (t) are not commutative: neverteless, te numerical results obtained are correct. Te numerical maximum global errors ave been estimated as follows E N,τ = max Y N,τ (x 1i, x 2i, t m ) Y x 1i,x 2i,t m were Y N,τ (x 1i, x 2i, t m ) are te numerical solutions obtained at te mes point (x 1i, x 2i ) and te time point t m = m τ, on a rectangular mes wit (N + 1) (N + 1) nodes and wit time step τ, wile Y is te numerical solution calculated at te same mes points and time steps, but by using a spatial mes wit (2N + 1) (2N + 1) nodes, alving te mes size, and wit time step τ 2. We compute te numerical orders of convergence as log 2 E N,τ E 2N,τ/2. Table 1. Numerical Errors (E N,τ ) N = 8 N = 16 N = 32 N = 64 N = 128 N = 256 N = 512 2.2183E-2 5.692E-3 1.4231E-3 3.5569E-4 8.8921E-5 2.2289E-5 5.654E-6 In Table 1, we sow te numerical errors and, in Table 2, teir corresponding numerical orders of convergence. In order to guarantee tat te contribution to te error of te spatial and temporal part are of te same size, we ave assumed te relation Nτ C =.1 between te time step τ and te mes size 1 N. Table 2. Numerical orders of convergence N = 8 N = 16 N = 32 N = 64 N = 128 N = 256 1.9629 1.9995 2.3 2. 1.9962 1.9915

New efficient time integrators for non-linear parabolic problems 419 REFERENCES [1] B. Bujanda, J.C. Jorge, Stability results for linearly implicit Fractional Step discretizations of non-linear time dependent parabolic problems, to appears in Appl. Num. Mat. [2] B. Bujanda, J.C. Jorge, Order conditions for linearly implicit Fractional Step Runge Kutta metods (Preprint), Universidad Pública de Navarra. [3] B. Bujanda, J.C. Jorge, Efficient linearly implicit metods for nonlinear multidimensional parabolic problems, J. Comput. Appl. Mat. 164/165 (24), 159 174. [4] B. Bujanda, J.C. Jorge, Stability results for fractional step discretizations of time dependent coefficient evolutionary problems, Appl. Numer. Mat. 38, (21), 69 86. [5] G.J. Cooper, A. Sayfy, Aditive metods for te numerical solution of ordinary differential equations, Mat. of Comp. 35, (198), 1159 1172. [6] E. Hairer, G. Wanner, Solving ordinary differential equations II, Springer-Verlag, 1987. [7] W. Hundsdorfer, J.G. Verwer, Numerical Solution of Time-Dependent Advection-Diffusion-Reaction Equations, Springer, 23. [8] N.N. Yanenko, Te metod of fractional steps, Springer, 1971. Blanca Bujanda blanca.bujanda@unavarra.es Universidad Pública de Navarra Dpto. Matemática e Informática Juan Carlos Jorge jcjorge@unavarra.es Universidad Pública de Navarra Dpto. Matemática e Informática Received: October 3, 25.