Extrapolation-based implicit-explicit general linear methods arxiv: v1 [cs.na] 8 Apr 2013

Transcription

1 Extrapolation-based implicit-explicit general linear methods arxiv: v cs.na] 8 Apr 23 A. Cardone, Z. Jackiewicz, A. Sandu, and H. Zhang September 5, 28 Abstract For many systems of differential equations modeling problems in science and engineering, there are natural splittings of the right hand side into two parts, one non-stiff or mildly stiff, and the other one stiff. For such systems implicit-explicit (IMEX integration combines an explicit scheme for the non-stiff part with an implicit scheme for the stiff part. In a recent series of papers two of the authors (Sandu and Zhang have developed IMEX GLMs, a family of implicit-explicit schemes based on general linear methods. It has been shown that, due to their high stage order, IMEX GLMs require no additional coupling order conditions, and are not marred by order reduction. This work develops a new extrapolation-based approach to construct practical IMEX GLM pairs of high order. We look for methods with large absolute stability region, assuming that the implicit part of the method is A- or L-stable. We provide examples of IMEX GLMs with optimal stability properties. Their application to a two dimensional test problem confirms the theoretical findings. Key words. IMEX methods, general linear methods, error analysis, order conditions, stability analysis Dipartimento di Matematica, Università degli studi di Salerno, Fisciano (Sa, 8484 Italy, ancardone@unisa.it. The work of this author was supported by travel fellowship from the Department of Mathematics, University of Salerno. Department of Mathematics, Arizona State University, Tempe, Arizona 85287, jackiewicz@asu.edu, and AGH University of Science and Technology, Kraków, Poland. Department of Computer Science, Virginia Polytechnic Institute & State University, Blacksburg, Virginia 246, sandu@cs.vt.edu. Department of Computer Science, Virginia Polytechnic Institute & State University, Blacksburg, Virginia 246, zhang@vt.edu.

2 Introduction Many practical problems in science and engineering are modeled by large systems of ordinary differential equations (ODEs which arise from discretization in space of partial differential equations (PDEs by finite difference methods, finite elements or finite volume methods, or pseudospectral methods. For such systems there are often natural splittings of the right hand sides of the differential systems into two parts, one of which is non-stiff or mildly stiff, and suitable for explicit time integration, and the other part is stiff, and suitable for implicit time integration. Such systems can be written in the form { y (t = f ( y(t + g ( y(t, t t, T ], y(t = y, (. where f(y represents the non-stiff processes, for example advection, and g(y represents stiff processes, for example diffusion or chemical reaction, in semi-discretization of advection-diffusion-reaction equations 8]. Implicit-explicit (IMEX integration approach discretizes the non-stiff part f(y is with an explicit method, and the stiff part g(y with an implicit, stable method. This strategy seeks to ensure the numerical stability of the solution of (. while reducing the amount of implicitness, and therefore the overall computational effort. IMEX multistep methods were introduced by Crouzeix 3] and Varah 25] and further analyzed in 2, 4]. IMEX Runge- Kutta methods have been investigated in,, 2, 22, 23, 3]. In a recent series of papers the last two authors and their collaborators have proposed the new IMEX GLM family of implicit-explicit schemes based on general linear methods. A general formalism for partitioned GLMs and their order conditions was developed by Zhang and Sandu 28]. The partitioned method formalism was then used to construct IMEX GLMs. The starting and ending procedures, linear stability, and stiff convergence properties of the new family have been analyzed. Zhang and Sandu examined practical methods of second order in 27] and of third order in 28]. A class of IMEX two step Runge-Kutta (TSRK methods was proposed by Zharovski and Sandu 29]. The results in 27, 28, 29] prove that the general linear framework is well suited for the construction of multi-methods. Specifically, owing to the high stage orders, no coupling conditions are needed to ensure the order of accuracy of the partitioned GLM 28]. In addition, it has been shown that 2

3 IMEX GLMs are particularly attractive for solving stiff problems, where other multistage methods may suffer from order reduction 28]. This paper extends our previous work 27, 28, 29] and develops a new extrapolation-based approach for the construction of practical IMEX GLM schemes of high order and high stage order. The organization of this paper is as follows. General linear methods and the implicit-explicit variants are reviewed in Section 2. The new extrapolationbased IMEX GLMs are derived in Section 3, and their order conditions are presented. The stability analysis is performed in Section 4 and specific methods are constructed in Section 5. Numerical experiments are presented in Section 6, and Section 7 gives some concluding remarks and plans for future work. 2 Implicit-explicit general linear methods In this section we briefly review GLMs and the IMEX GLM family. The GLMs for ODEs were introduced by Burrage and Butcher 4] and further investigated in 3, 5, 7,, 2, 5, 6, 7]. We also refer the reader to the review article 8] and the recent monograph 9] and references therein. A diagonally implicit GLM for (. is defined by Y n+] i = h y n+] i = h i ( a ij f ( Y n+] ( n+] j + g Y j + s ( b ij f ( Y n+] ( n+] j + g Y j + r r u ij y n] j, i =, 2,..., s v ij y n] j, i =, 2,..., r, (2. n =,,..., N. Here, N is a positive integer, h = (T t /N, t n = t +nh, n =,,..., N, Y n+] i are approximations of stage order q to y(t n + c i h, i.e., Y n+] i = y(t n + c i h + O(h q+, i =, 2,..., s, (2.2 y n] i are approximations of order p to the linear combinations of the derivatives of the solution y at the point t n, i.e., y n] i = p q ik h k y (k (t n + O(h p+, i =, 2,..., r, (2.3 k= 3

4 and y is the solution to (.. These methods can be characterized by the abscissa vector c = c,..., c s ] T, the coefficient matrices A = a ij ] R s s, U = a ij ] R s r, B = a ij ] R r s, V = a ij ] R r r, the vectors q,..., q s R r defined by q i = q j,i ] j r, and four integers: the order p, the stage order q, the number of external approximations r, and the number of stages or internal approximations s. The method (2. can be written in a compact form ( Y n+] = h(a I f ( Y n+] + g ( Y n+] + (U Iy n], ( y n+] = h(b I f ( Y n+] + g ( Y n+] + (V Iy n], n =,,..., N, and the relation (2.3 takes the form (2.4 y n] = p q k h k y (k (t n + O(h p+. (2.5 k= Applying (2. to the basic test equation y (t = λy(t, t, λ C, leads to the recurrence equation y n+] = S(zy n], n =,,..., z = hλ, with the stability matrix given by S(z = V + zb(i za U. (2.6 We also define the stability polynomial η(w, z by η(w, z = det ( wi S(z. (2.7 The region of absolute stability of the method (2. is the subset of the complex plane A = { z C : all roots w i (z of η(w, z are in the unit circle }. (2.8 The traditional concepts of A(α-stability, A-stability, and L-stability apply directly to GLMs via (2.8. In this paper we will examine only methods of high stage order, i.e., methods where q = p or q = p. It has been shown in 9, 6, 9] that the 4

5 GLM (2. has order p and stage order q = p or q = p if and only if k l= c k k A c k k! U q k =, k =,,..., q, and (2.9 k! l! q k l k B c k k! V q k =, k =,,..., p. (2. An IMEX-GLM 29, Definition 4] has the form Y n+] = h(a exp I f ( Y n+] + h(a imp I g ( Y n+] + (U Iy n], y n+] = h(b exp I f ( Y n+] + h(b imp I g ( Y n+] + (V Iy n], (2. where A exp, B exp correspond to the explicit part and A imp, B imp to the implicit part. The methods share the same abscissa c exp = c imp, which makes (2. internally consistent 29, Definition 2]. The methods also share the same coefficient matrices U exp = U imp = U and V exp = V imp = V. The coefficients q exp k, q imp k in (2.5 can be different, which means that the implicit and explicit components use different initialization and termination procedures. An IMEX-GLM (2. is a special case of a partitioned GLM 29, Definition ]; while in (2. the right hand side is split in two components, stiff and nonstiff, a partitioned GLM allows for splitting in an arbitrary number of components. It has been shown in 29, Theorem 2] that an internally consistent partitioned GLM (and, in particular, the IMEX GLM (2. has order p and stage order q {p, p} if and only if each component method (A exp, B exp, U, V and ( A imp, B imp, U, V has order p and stage order q. We note that no additional coupling conditions are needed for the IMEX GLM (i.e., no order conditions that contain coefficients of both the implicit and the explicit schemes. 3 Extrapolation-based IMEX GLMs 3. Method formulation In this section we derive the new extrapolation-based IMEX GLMs. Consider the following extrapolation formula depending on stage values Y n] k and Y n+] k 5

6 at two consecutive steps f n+] j = s α jk f ( j Y n] k + β jk f ( Y n+] k, j =, 2,..., s. (3. k= Substituting f n+] j k= in (3. for f ( Y n+] j in (2. leads to the proposed class of extrapolation-based IMEX GLMs. The simple example of IMEX method consisting of the explicit Euler method combined with the A-stable implicit θ-method corresponding to θ /2 is presented in 8]. Substituting (3. into (2. leads to Y n+] i = h + h y n+] i = h + h i s k= i s k= i a ij α jk f ( Y n] k a ij g ( Y n+] j s + r b ij α jk f ( Y n] k b ij g ( Y n+] j j i + h + r k= a ij β jk f ( Y n+] k u ij y n] j, i =, 2,..., s, j s + h k= b ij β jk f ( Y n+] k v ij y n] j, i =, 2,..., r, n =,,..., N. Changing the order of summation in the double sums above and then interchanging the indices j and k we obtain Y n+] i = h + h y n+] i = h + h s i k= i s k= s a ik α kj f ( Y n] i j + h a ij g ( Y n+] j s + r b ik α kj f ( Y n] j b ij g ( Y n+] j + r i k=j+ a ik β kj f ( Y n+] j u ij y n] j, i =, 2,..., s, s + h s k=j+ b ik β kj f ( Y n+] j v ij y n] j, i =, 2,..., r. 6

7 These relations lead to IMEX GLMs of the form s Y n+] i = h a ij f ( Y n] i j + h a ijf ( Y n+] j + h y n+] i = h + h i s s a ij g ( Y n+] j b ij f ( Y n] j b ij g ( Y n+] j + r s + h + r u ij y n] j, i =, 2,..., s, b ijf ( Y n+] j v ij y n] j, i =, 2,..., r, (3.2 n =,,..., N, where the coefficients a ij, a ij, b ij, and b ij are defined by a ij = i i a ik α kj, a ij = a ik β kj, b ij = k= k=j+ s s b ik α kj, b ij = b ik β kj. k= k=j+ In matrix notation A = a ij ] R s s, A = a ij] R s s, B = b ij ] R r s, B = b ij] R r s. with A = Aα, A = Aβ, B = Bα, B = Bβ, where α = α ij ] R s s, β = β ij ] R s s. Observe that the matrix A is strictly lower triangular and that the last column of the matrix B is zero. In matrix notation the extrapolation-based IMEX-GLM is defined by: n =,,..., N. Y n+] = h(a If ( Y n] + h(a If ( Y n+] +h(a Ig ( Y n+] + (U Iy n], (3.3 y n+] = h(b If ( Y n] + h(b If ( Y n+] +h(b If ( Y n+] + (V Iy n], 7

8 The explicit part of (3.3, obtained for g(y =, can be represented as a single GLM extended over two steps from t n to t n and t n to t n+, as follows Y n] Y n+] Y n+] y n+] I f ( Y n] = A A U f ( Y n+] A A U Y n]. (3.4 B B V The abscissa vector is c exp = (c e T, c T ] T. Similarly, the implicit part of the IMEX scheme (3.2 corresponding to f(y = assumes the form Y n] Y n+] Y n+] y n+] B V y n] I g ( Y n] = A U g ( Y n+] A U Y n]. (3.5 This method has the order and stage order of the underlying GLM (2., since it is the same method. The abscissa vector is c imp = (c e T, c T ] T, and therefore the method (3.3 is internally consistent. 3.2 Construction of the interpolant We define the local discretization errors η(t n + c j h of the extrapolation formula (3. by the relation y n] f ( y(t n + c j h s = α jk f ( y(t n + c k h + + k= j β jk f ( y(t n + c k h + η(t n + c j h, k= (3.6 j =, 2,..., s. Letting ϕ(t = f(y(t the relation (3.6 can be written in the form η(t n + c j h = ϕ(t n + c j h s α jk ϕ ( t n + (c k h j β jk ϕ(t n + c k h, k= 8 k=

9 j =, 2,..., s. Expanding ϕ(t n + c j h, ϕ(t n + (c k h, and ϕ(t n + c k h into Taylor series around t n we obtain η(t n + c j h = p ( c l j s l! (c k l α jk l! l= k= j k= c l k β jk h l ϕ (l (t n + O(h p+. l! Assuming that the extrapolation procedure given by (3. has order p, i.e., η(t n + c j h = O(h p, leads to the following system of equations for the interpolation coefficients: j s α jk (c k l = c l j β jk c l k, l =,,..., p, j =, 2,..., s. (3.7 k= k= In matrix notation we have α (c e l + β c l = c l, l =,,..., p. ( Stage and order conditions Lemma 3. Assume that the underlying GLM (2.4 has order p and stage order q = p or q = p, and that the interpolation formula (3. has order p (3.8. Then the explicit method (3.4 has order p and stage order q. Proof: The method (3.4 has the coefficients ] ] ] A A c e A exp =, B exp =, c exp =. A A B B c where e =,..., ] T R s, I U exp = U ] U, V exp = V ], and the vectors q exp = e q ] ; q exp i = (c e i i! q i ], i =,..., p. 9

10 We verify directly that the extrapolation-based explicit method (3.4 has stage order q, i.e., it satisfies equations (2.9 (c exp k k A exp (c exp k k! U exp q exp k ] = c k k A ( α (c e k + β c k k! Uq k ] = {from (3.8} c k k Ac k k! Uq k ] =, k =,,..., q {from (2.9}. From (2. we verify that the method (3.4 has order p k l= = k! l! qexp k l k Bexp (c exp k k! V exp q exp k k l= k! l! k k! l! = l= k l= (c e k l (k l! k q k l (c e k l (k l! ( A α (c e k + β c k ] ] U B ( qk α (c e k + β c k k! V q k k Ac k k! U q k k! l! q k l k Bc k k! V q k {from (3.8} ] =, k =,,..., p {from (2.9, (2., and (3.9}. For the first component we have used the fact that c l = ( (c e + e l = k l= l! k! (l k! (c ek l. (3.9 To analyze the order and stage order of IMEX GLMs (3.2 we will impose some conditions on the local discretization errors of the internal and external

11 stages of the underlying GLM (2. and on the accuracy of the extrapolation procedure (3.. We have the following result. Theorem 3. Assume that the underlying GLM (2. has order p and stage order q = p or q = p, and that the extrapolation procedure (3. has order p. Then the IMEX GLM (3.2 has order p and stage order q = p or q = p. Proof: The method (3.3 is an IMEX GLM of the form (2. 28, Definition ]. The explicit (3.4 and implicit (3.5 components have the same order p and stage order q, and they share the same abscissa vector and the same coefficients c exp = c imp, U imp = U exp, V imp = V exp. The result follows directly from 28, Theorem 2]. 3.4 Prothero-Robinson convergence of IMEX GLMs The extrapolation IMEX-GLM schemes (3.2 do not suffer from order reduction phenomenon when applied to stiff systems of differential equations. Following 5, 28, 29] we consider the Prothero-Robinson (PR 24] test problem of the form { y (t = µ ( y(t φ(t + φ (t, t, (3. y( = φ(, where µ C has a large and negative real part and φ(t is a slowly varying function. The solution to (3. is y(t = φ(t. The IMEX scheme (3.2 is said to be PR-convergent if the application of (3.2 to the equation (3. leads to the numerical solution y n] whose global error satisfies p yn] q k h k y (k (t n = O(hp as h and hµ. k= We have the following result.

12 Theorem 3.2 Assume that the implicit GLM (2. has order p and stage order q = p or q = p, and that the extrapolation formula (3. has order p. Then the IMEX scheme (3.2 is PR-convergent with order min(p, q as h, hµ, and hµ S I. Here, S I is the stability region of the implicit GLM (3.5. Proof: The result follows directly from 28, Theorem 3] on PR-convergence of IMEX-GLMs. 4 Stability analysis of IMEX GLMs To analyze stability properties of IMEX GLMs (3.2 we use the test equation y (t = λ y(t + λ y(t, t, (4. where λ and λ are complex parameters. Applying (3.3 to (4. and letting z i = hλ i, i =,, we obtain ( I (z A + z A Y n+] = z AY n] + Uy n], (z B + z BY n+] + y n+] = z BY n] + Vy n], n =,,..., N. This is equivalent to the matrix recurrence relation ] ] Y n+] Y n] = M(z, z, (4.2 y n+] y n] where the stability matrix M(z, z is defined by m (z, z m 2 (z, z M(z, z = m 2 (z, z m 22 (z, z ] with m (z, z = z ( I (z A + z A A, m 2 (z, z = ( I (z A + z A U, 2

13 m 2 (z, z = z ( B + (z B + z B ( I (z A + z A A, m 22 (z, z = V + (z B + z B ( I (z A + z A U. We define also the stability function of the IMEX GLM (3.2 as a characteristic polynomial of the stability matrix M(z, z, i.e., p(w, z, z = det ( wi M(z, z. For z = the stability matrix M(, z and polynomial p(w,, z = w s η(w, z are those of the underlying GLM (2.. For z = we obtain M(z,, and it can be verified that this corresponds to the stability matrix of the explicit method (3.4. We say that the IMEX GLM (3.2 is stable for given z, z C if all the roots w i (z, z, i =, 2,..., s + r, of the stability function p(w, z, z are inside of the unit circle. As observed in 8] in the context of IMEX θ-methods of order one, a large region of absolute stability for the explicit method (3.4 and good stability properties (for example A- or L-stability for the implicit method are not sufficient to guarantee desirable stability properties of the overall IMEX GLM (3.2. We have to investigate the stability properties of the combined as IMEX GLM 29, 28]. In this paper we will be mainly interested in IMEX schemes which are A(α- or A-stable with respect to the implicit part z C. To investigate such methods we consider, similarly as in 8, 28], the sets { } z C : the IMEX GLM is stable for any S α = z C : R(z < and Im(z tan(α R(z. (4.3 For fixed values of y R we define also the sets { } z C : the IMEX GLM is stable for fixed S α,y =. (4.4 z = y / tan(α + iy It follows from the maximum principle that S α = y R S α,y. (4.5 Observe also that the region S α, is independent of α, and corresponds to the region of absolute stability of the explicit method (3.4. This region will be 3

14 denoted by S E. We have S α S E, (4.6 and we will look for IMEX GLMs for which the stability region S α contains a large part of the stability region S E of the explicit method (3.4, for some α (, π/2], preferably for α = π/2. The boundary S α,y of the region S α,y can be determined by the boundary locus method which computes the locus of the curve { S α,y = z C : p ( e iθ, z, y / tan(α + iy } =, θ, 2kπ], where k is a positive integer y =m x Im(z.5.5 max(x Re(z Figure 4.: Points on the intersection of the ray y = mx and S π/2,y for y =.5,.,...,.5 (circles and on the intersection of y = mx and S π/2 (square. This figure corresponds to IMEX GLM scheme with p = q = r = s = 2 for m =, λ =.3, and β 2 = 4.3 We have also developed an algorithm to determine the boundary S α of the stability region S α. For fixed direction m corresponding to the ray y = mx, and for fixed z = y / tan(α + iy (or any z C we can compute the point z = x + iy of intersection of the boundary S α,y of S α,y with the ray y = mx taking into account that such a point satisfies the condition ( max w i z, y / tan(α + iy =. i=,2,...,s+r 4

15 This can be done by the bisection method which terminates when the condition ( max w i z, y / tan(α + iy tol, (4.7 i=,2,...,s+r with accuracy tolerance tol is satisfied. We apply this method to the interval x, ] with x large enough so that the condition (4.7 is not satisfied for the first iteration of bisection method. This process leads to the definition of the function x = f(m, α, y, where x corresponds to the points z = x +iy S α,y. Then the boundary S α of the region S α can be determined by minimizing the negative value of this function for m R and plotting the resulting points z = x + iy. This algorithm is illustrated in Fig. 4., where we have plotted points on the intersection of the ray y = mx and S π/2,y for y =.5,.,...,.5 (circles and on the intersection of y = mx and S π/2 (square. This figure corresponds to IMEX GLM scheme with p = q = r = s = 2 for m =, λ =.3, and β 2 = 4.3. This minimization can be accomplished using the subroutine fminsearch.m in Matlab applied to f(m, α, y for fixed values of m R and α, π/2] starting with appropriately chosen initial guesses for y. This process will be next applied to specific IMEX GLMs. 5 Construction of IMEX GLMs with desirable stability properties In this section we describe the construction of IMEX GLMs (3.2 up to the order p = 4 with large regions of absolute stability S with respect to the explicit part assuming that the implicit part is A- or L-stable. We will always start with implicit DIMSIM with p = q = r = s and the abscissa vector c given in advance. After computing the coefficient matrices A and V so that the resulting method has Runge-Kutta stability with the underlying Runge-Kutta formula which is A- or L-stable, the coefficient matrix B is computed from the relation B = B AB VB 2 + VA. (5. 5

16 Here, B, B, and B 2 are s s matrices with the (i, j elements given by +ci φ j (xdx, φ j (c j φ j ( + c i, φ j (c j ci φ j (xdx φ j (c j, φ i (x = s (x c j,,j i i =, 2,..., s, compare Th. 5. in 6] or Th in 9]. 5. IMEX GLMs with p = q = r = s = Consider the implicit θ-method defined by Y n+] = hθ ( f(y n+] + g(y n+] + y n], y n+] = h ( f(y n+] + g(y n+] + y n], (5.2 n =,,..., N. This method is A-stable for θ /2, ] and L-stable for θ (/2, ]. Consider also the extrapolation procedure f(y n+] = f(y n], (5.3 n =, 2,..., N. Substituting (5.3 into (5.2 we obtain IMEX θ-method of the form Y n+] = hθ ( f(y n] + g(y n+] + y n], y n+] = h ( f(y n] + g(y n+] (5.4 + y n], n =, 2,..., N. We would like to point out that this variant of IMEX scheme is different from IMEX θ-method considered in 8]. Observe that the method (5.4 requires a starting procedure to compute Y ] y(t + θh and y ] y(t. The method (5.2 can be represented by the abscissa c = θ, the partitioned matrix ] ] A U θ =, B V and q =, q =, and the explicit formula corresponding to g(y = in 6

17 (5.4 is GLM with c = θ, θ] T, ] A U = B V and q =, ] T, q = θ, ] T. θ θ, (5.5 Im(z θ=4/5 θ=.2 θ=3/4 θ=2/3 θ=/ Re(z Figure 5.: Stability regions S E = S E (θ of explicit methods for θ = /2, 2/3, 3/4, 4/5, and.5 Im(z Re(z Figure 5.2: Stability regions S π/2,y (θ, y =.,.9,...,. (thin lines, S π/2 (θ (shaded region, and S E (θ (thick line for θ = 3/4 It can be verified that the stability matrix of IMEX scheme (5.4 takes the form ] θz M(z, z =. θz z + ( θz 7

18 .5 Im(z Re(z Figure 5.3: Stability regions S π/2,y (θ, y =.,.9,...,. (thin lines, S π/2 (θ (shaded region, and S E (θ (thick line for θ = 2/3 To investigate stability properties of (5.4 it is more convenient to work with the polynomial obtained by multiplying the characteristic function p(w, z, z of M(z, z by a factor ( θz 2. The resulting quadratic polynomial, which will be denoted by the same symbol p(w, z, z, takes the form p(w, z, z = ( θz 2 w 2 ( θz ( +θz +( θz w ( θz ( θz. The stability polynomial of the explicit methods (5.5 obtained by letting g(y = in (5.4 corresponds to z = and is given by p(w, z, = w 2 ( + θz w ( θz. The stability regions S E = S E (θ corresponding to this polynomial are plotted in Fig 5. for θ = /2, 2/3, 3/4, 4/5 and. Observe that the stability region of the method (5.5 corresponding to θ = is the unit disk S E = S E ( = { z C : z + < }. We will investigate next the regions S π/2 = S π/2 (θ defined by (4.3. We analyze first the case θ =. It follows from Schur criterion applied to the polynomial p(w, z, iy that z S π/2 ( if and only if for any y R. This is equivalent to y 2 2x x 2 y 2 > (x y 2 < or z + < 8

19 and it follows that in this case S π/2 ( = S E (. For θ (/2, S π/2 (θ is no longer equal to S E (θ, but S π/2 (θ contains some part of S E (θ. This is illustrated in Fig. 5.2 for θ = 3/4 and in Fig. 5.3 for θ = 2/3. In these figures we have plotted the regions S π/2,y = S π/2,y (θ defined by (4.4 for y =.,.9,...,. (thin lines, the regions S π/2 = S π/2 (θ defined by (4.3 (shaded regions, and the regions S E = S E (θ corresponding to explicit methods (5.5 for θ = 3/4 and θ = 2/3. We can see that for these values of θ the sets S π/2 (θ contain quite large parts of S E (θ. These figures illustrate also the relation (4.5 and the inclusion (4.6. For θ = /2 it follows from Schur criterion that z S π/2 (/2 if and only if x 2 + y 2 < 4 and x y 2 4yy x (x y 2 (4 + x > for any y R. This implies that x > and (2 + x 2 (x 2 + y 2 < and it follows that S π/2 (/2 is empty. 5.2 IMEX GLMs with p = q = r = s = 2 Consider the implicit DIMSIM with c =, ] T, the coefficient matrices given by A U B V ] = λ 2 λ +2λ 4λ 2 + λ λ 4 2 2, 8λ 3 2λ 2 +λ + λ λ 4(2λ λ 3 +2λ 2 2λ+5 4(2λ+ 8λ 3 +2λ 2 2λ+3 4(2λ+ and the vectors q, q, and q 2 equal to ] λ q =, q = 2λ 2 +λ 2λ+ ], q 2 = 2λ 2 ]. It was demonstrated in 9] that this method has order p = 2 and stage order q = 2. Moreover, this method is A-stable if λ /4 and L-stable for λ = (2 ± 2/2. 9

20 The coefficients α jk of the extrapolation formula (3. of order p = 2 computed from the system (3.7 corresponding to p = s = 2 take the form ] ] α α 2 α = =, α 2 α 22 2 β 2 and the matrices A, A, B, and B appearing in the representation of IMEX GLM (3.2, and the corresponding explicit scheme (3.3 or (3.4, are given by ] ] λ B = A = λ 2+(2 β 2λ+2(2 β 2 λ 2 +2λ 4λ 2 4 λ+2λ 2 +8λ 3 4(+2λ B =, A = β 2 λ 7 β 2 +2( β 2 λ+4(+β 2 λ 2 8( β 2 λ 3 4(+2λ +β 2 +2(9 5β 2 λ 4( 3β 2 λ 2 8( β 2 λ 3 4(+2λ β 2 ( 4λ 2 4 β 2 ( λ+2λ 2 +8λ 3 4(+2λ ]., ], 8 6 area of S E area of S α areas β 2 Figure 5.4: Areas of stability regions S E = S E (β 2 and S α = S α (β 2 for α = π/2 and β 2, 8] To investigate the stability properties of the resulting IMEX scheme we will work with the stability polynomial p(w, z, z obtained by multiplying stability function of the method by a factor ( λz 2. It can be verified that this polynomial takes the form p(w, z, z = w ( ( λz 2 w 3 p 3 (z, z w 2 + p 2 (z, z w p (z, z, 2

21 Im(z Re(z Figure 5.5: Stability regions S π/2,y, y = 2.,.8,..., 2. (thin lines, S π/2 (shaded region, and S E (thick line for λ = (2 2/2 and β λ β 2 Figure 5.6: Contour plots of the area of stability region S π/2 of IMEX GLMs for p = q = r = s = 2 with the coefficients p (z, z, p 2 (z, z, and p 3 (z, z which depend also on λ and β 2. These coefficients are given by p (z, z = 4 β 2 2(4 β 2 λ + 4β 2 λ 2 8β 2 λ 3 z + 4( + 2λ 2 4λ + 2λ2 z 2 2,

22 2.5 Im(z Re(z Figure 5.7: Stability regions S π/2,y, y = 2.,.8,..., 2. (thin lines, S π/2 (shaded region, and S E (thick line for λ.29 and β p 2 (z, z = 2( λ + (2 β 2λ 2 2β 2 λ 3 + 2λ 2 β 2 4(2 β 2 λ + 2(2 β 2 λ 2 z z, 2 z + β 2 4β 2 λ + 2( + β 2 λ 2 z 2 2 p 3 (z, z = β 2 + 2(4 β 2 λ + 4(4 3β 2 λ 2 8β 2 λ 3 z 4( + 2λ + β 2 λ 2 z 2 (2 β 2 z z + ( 2λz + 4λ + 2λ2 z 2 2. The underlying implicit GLM is A- and L-stable for λ = (2 ± 2/2, compare 9], and we choose λ = (2 2/2 since this value leads to explicit methods and IMEX schemes with larger regions of stability S E and S π/2 than those corresponding to λ = (2 + 2/2. We have plotted in Fig 5.4 the area of the stability region S E = S E (β 2 of the explicit method (corresponding to z = and the area of the stability region S π/2 = S π/2 (β 2 of the IMEX scheme for β 2, 8]. It can be verified that the explicit formula attains the maximal area of S E, approximately equal to 7.5 for β , and the IMEX scheme attains the maximal area of S π/2, approximately equal to 5.75 for β On Fig 5.5 we have plotted stability regions S π/2,y for y = 2.,.8,..., 2. (thin lines, stability region S π/2 (shaded region, and stability region S E (thick line. We can see that S π/2 contains a significant part of S E. We have also displayed on Fig. 5.6 contour plots of the area of stability region S π/2 of IMEX methods for β 2 3, 6] and λ.25,.35]. This 22

23 area attains its maximum value approximately equal to 5.83 for β and λ.29. This point is marked by the symbol on Fig On Fig 5.7 we have plotted stability regions S π/2,y for y = 2.,.8,..., 2. (thin lines, stability region S π/2 (shaded region, and stability region S E (thick line corresponding to these values of β 2 and λ. We can see again that S π/2 contains a significant part of S E. 5.3 IMEX GLMs with p = q = r = s = 3 Let λ be a root of the cubic polynomial ϕ(λ = λ 3 3λ λ 6, and consider the implicit DIMSIM with c =, /2, ] T, the coefficient matrix A given by A = , the rank one coefficient matrix V = ev T, where e =,, ] T and ] T v = , and the vectors q, q, q 2, and q 3 equal to q = e, q = ] T, q 2 = q 3 = ] T, ] T. Computing the coefficient matrix B from the relation (5. leads to the method of order p = 3 and stage order q = 3. It was demonstrated in 9] that the resulting method is A- and L-stable. 23

24 The coefficients α jk of the extrapolation formula (3. of order p = 3 computed from the system (3.7 corresponding to p = s = 3 take the form α α 2 α 3 α = α 2 α 22 α 23 = 3 3 β 2. α 3 α 32 α 33 3 β 32 3β β 3 3β 32 To investigate the stability properties of the resulting IMEX scheme we will work with the stability polynomial p(w, z, z obtained by multiplying stability function of the method by a factor ( λz 3, where λ is the diagonal element of the matrix A. It can be verified that this polynomial takes the form p(w, z, z = ( λz 3 w 6 p 5 (z, z w 5 + p 4 (z, z w 4 p 3 (z, z w 3 + p 2 (z, z w 2 p (z, z w + p (z, z, with the coefficients p (z, z, p (z, z, p 2 (z, z, p 3 (z, z, p 4 (z, z, and p 5 (z, z which are polynomials of degree less than or equal to 3 with respect to z and z. These coefficients depend also on β 2, β 3, and β Im(z Re(z Figure 5.8: Stability regions S π/2,y, y = 2.,.8,..., 2. (thin lines, S π/2 (shaded region, and S E (thick line for β 2.3, β 3.45, and β We have performed a computer search in the parameter space β 2, β 3, and β 32 looking first for methods for which the stability region S E of the explicit method is maximal. This corresponds to the parameter values β 2.3, β 3.45, β 32.58, for which the area of S E is approximately 24

25 2.5 Im(z Re(z Figure 5.9: Stability regions S π/4,y, y = 2.,.8,..., 2. (thin lines, S π/4 (shaded region, and S E (thick line for β 2.3, β 3.45, and β Im(z Re(z Figure 5.: Stability regions S π/2,y, y = 2.,.8,..., 2. (thin lines, S π/2 (shaded region, and S E (thick line for β 2.39, β 3.46, and β equal to The stability region S E of the resulting method is plotted on Fig. 5.8 by a thick line. We have also plotted stability regions S π/2,y for y = 2.,.8,..., 2. (thin lines and the stability region S π/2 (shaded region of the corresponding IMEX scheme. We can see that this region S π/2 is substantially smaller than the region S E, the area of S π/2 is approximately equal to.39. However, we can obtain larger regions S α for values of α smaller than π/2, i.e., if we relax the requirement that the implicit part of IMEX scheme is A-stable and require instead A(α-stability for α < π/2. This is illustrated on Fig. 5.9, where we have plotted again the stability region S E of the explicit method (thick line, stability regions of S α,y for y = 2.,.8,..., 2. (thin lines, and stability region S α (shaded region 25

26 2.5 Im(z Re(z Figure 5.: Stability regions S π/4,y, y = 2.,.8,..., 2. (thin lines, S π/4 (shaded region, and S E (thick line for β 2.25, β 3.62, and β for α = π/4. The area of this region is approximately equal to.9. We have also performed a computer search looking for methods for which stability regions S α are maximal for some fixed values of α. For α = π/2 this corresponds to the parameter values β 2.39, β 3.46, and β for which the area of S π/2 is approximately equal to.5. For α = π/4 this corresponds to the parameter values β 2.25, β 3.62, and β for which the area of S π/4 is approximately equal to 2.8. We have plotted on Fig. 5. and Fig. 5. the stability regions S E of the resulting explicit methods (thick lines, the regions S α,y for y = 2.,.8,..., 2. (thin lines and stability regions S α of IMEX schemes (shaded regions for α = π/2 and α = π/ IMEX GLMs with p = q = r = s = 4 Let λ be a root of the polynomial ϕ(λ = λ 4 4λ 3 + 3λ λ + 24, 26

27 and consider the implicit DIMSIM with c =, /3, 2/3, ] T, the coefficient matrix A given by A = , the rank one coefficient matrix V = ev T, where e =,,, ] T and ] T v = , and the vectors q, q, q 2, q 3, and q 4 equal to q = e, ] T q = , ] T q 2 = , ] T q 3 = , ] T q 4 = Computing the coefficient matrix B from the relation (5. leads to the method of order p = 4 and stage order q = 4. It was demonstrated in 26] that the resulting method is A- and L-stable. The coefficients α jk of the extrapolation formula (3. of order p = 4 computed from the system (3.7 corresponding to p = s = 4 take the form α = β 2 β β 32 2(3β 32 β 3 4β 32, α 4 α 42 α 43 α 44 with α 4 = β β 43, α 42 = 36 4β 42 5β 43, α 43 = 6β β 43 45, α 44 = 2 β 4 4β 42 β

28 .5 Im(z Re(z Figure 5.2: Stability regions S π/4,y, y = 2.,.8,..., 2. (thin lines, S π/4 (shaded region, and S E (thick line for β 2.645, β 3.35, and β , β , β , β Im(z π/2 π/6 π/4 π/3 5π/2 π/ Re(z Figure 5.3: Stability regions S α for α =, π/2, π/6, π/4, π/3, 5π/2, and π/2 for β 2.645, β 3.35, and β , β , β , β 43.5 To investigate the stability properties of the resulting IMEX scheme we will work with the stability polynomial p(w, z, z obtained by multiplying stability function of the method by a factor ( λz 4, where λ is the diagonal element of the matrix A. It can be verified that this polynomial takes the form p(w, z, z = ( λz 4 w 8 p 7 (z, z w 7 + p 6 (z, z w 6 p 5 (z, z w 5 + p 4 (z, z w 4 p 3 (z, z w 3 + p 2 (z, z w 2 p (z, z w + p (z, z, with coefficients p (z, z, p (z, z, p 2 (z, z, p 3 (z, z, p 4 (z, z, p 5 (z, z, p 6 (z, z, and p 7 (z, z which are polynomials of degree less than or equal 28

29 .8 Im(z Re(z Figure 5.4: Stability regions S π/2,y, y = 2.,.8,..., 2. (thin lines, S π/2 (shaded region, and S E (thick line for β 2.56, β 3.939, β 32.8, β 4.7, β , and β Im(z Re(z Figure 5.5: Stability regions S π/4,y, y = 2.,.8,..., 2. (thin lines, S π/4 (shaded region, and S E (thick line for β 2.964, β 3.278, β , β 4.63, β , and β to 4 with respect to z and z. These coefficients depend also on β 2, β 3, β 32, β 4, β 42, and β 43. We have performed a computer search in the parameter space β 2, β 3, β 32, β 4, β 42, and β 43 looking first for methods for which the stability region S E of the explicit method is maximal. This corresponds to the parameter values β 2.625, β 3.355, β , β , β , β 43.6, for which the area of S E is approximately equal to The stability region S E of the resulting method is plotted on Fig. 5.2 by a thick line. We have also plotted stability regions S π/4,y for y = 2.,.8,..., 2. (thin lines and the stability region S π/4 (shaded region of the corresponding 29

30 IMEX scheme. The area of S π/4 is approximately equal to.32. We have also plotted on Fig. 5.3 stability regions S α for α =, π/2, π/6, π/4, π/3, 5π/2, and π/2 corresponding to the same values of β ij. We can see in particular that the region S π/2 is quite small, its area is approximately equal to.69. As in Section 5.3 we have also performed a computer search looking directly for methods for which stability regions S α are maximal for some fixed values of α. For α = π/2 this corresponds to the parameter values β 2.56, β 3.939, β 32.8, β 4.7, β , and β 43.32, for which the area of S π/2 is approximately equal to.6. For α = π/4 this corresponds to the parameter values β 2.964, β 3.278, β , β 4.63, β , and β , for which the area of S π/4 is approximately equal to.65. We have plotted on Fig. 5.4 and Fig. 5.5 the stability regions S E of the resulting explicit methods (thick lines, the regions S α,y for y = 2.,.8,..., 2. (thin lines and stability regions S α of IMEX schemes (shaded regions for α = π/2 and α = π/4. 6 Numerical experiments The extrapolation-based IMEX GLMs constructed in Section 5 have been implemented in Matlab. The required starting values y ] and Y ] were computed by finite difference approximations from solutions obtained with the Matlab routine ode5s. The test problem is the two dimensional shallow-water equations system 2], which approximates a thin layer of fluid inside a shallow basin: t h + x (uh + y (vh = t (uh + (u 2 h + 2 x gh2 + (uvh = (6. y t (vh + x (uvh + (v 2 h + 2 y gh2 =. Here h(t, x, y is the fluid layer thickness, u(t, x, y and v(t, x, y are the components of the velocity field, and g denotes the gravitational acceleration. The spatial domain is Ω = 3, 3] 2 (spatial units, and the integration window is t = t t f = (time units. We use reflective boundary 3

31 conditions and the initial conditions at t = u(t, x, y = v(t, x, y =, h(t, x, y = + e (x,y (c,c 2 2 2, (6.2 with the Gaussian height profile c = /3 and c 2 = 2/3. A second order Lax-Wendroff finite difference scheme is used for space discretization, resulting in a semi-discrete ODE system of the form d dt U(t = F (U = F ( U U U(t + ( ( F (U F U U U(t, (6.3 }{{}}{{} g(u f(u where U(t is a combined column vector of the discretized state variables (ĥ, ûh, vh, and F U = F/ U be the Jacobian of right hand side. We consider a splitting of equation (6.3 into the linear part g(u, considered stiff, and the nonlinear part f(u, considered non-stiff. The linear stiff part is treated implicitly, and the non-stiff part is treated explicitly. We compare the numerical results for the solution at the final time with a reference solution computed by the Matlab function ode5s with very tight tolerances atol = rtol = 4. The errors are measured in L 2 norms. The error diagram against the time step size is presented in Fig. 6.. The observed orders for all the methods tested match the theoretical predictions. 7 Concluding remarks General linear methods offer an excellent framework for the construction of implicit-explicit schemes. In this paper we develop a new extrapolation-based approach for the construction of practical IMEX GLM schemes of high order and high stage order. The accuracy, linear stability, and Prothero-Robinson convergence are analyzed. These schemes are particularly attractive for solving stiff problems, where other multistage methods may suffer from order reduction. The extrapolation-based mechanism offers a systematic approach for constructing IMEX GLM schemes. The construction starts with the selection of an implicit component method with suitable stability and order properties. The explicit component is then obtained though an optimization procedure that maximized the combined stability region of the pair. We apply this methodology to construct IMEX pairs of orders one to four. 3

32 Error in L2 norm p=2,imex GLM O(h 2 p=3,imex GLM O(h 3 p=4,imex GLM O(h 4 2 Time step h Figure 6.: Error vs. time step size for several extrapolation-based IMEX- GLMs applied to the shallow water equations test problem. Future work is planned to extend the extrapolation idea to construct other types of partitioned GLMs, including parallel time integrators, and asynchronous pairs of methods that do not share the same abscissae. Acknowledgements. The results reported in this paper were obtained during the visit of the first author to the Arizona State University in January March of 23. This author wish to express her gratitude to the School of Mathematical and Statistical Sciences for hospitality during this visit. The work of Sandu and Zhang has been supported in part by the awards NSF OCI , NSF CCF-96493, NSF DMS-9547, NSF CMMI- 3667, NSF CCF-28454, AFOSR FA DEF, AFOSR , and by the Computational Science Laboratory at Virginia Tech. References ] Ascher, U.M., Ruuth, S.J., Spiteri, R.J.: Implicit-explicit Runge-Kutta methods for time-dependent partial differential equations. Appl. Numer. 32

33 Math 25, 5 67 (997 2] Ascher, U.M., Ruuth, S.J., Wetton, B.T.R.: Implicit-explicit methods for time-dependent partial differential equations. SIAM J. Numer. Anal. 32(3, (995 3] Burrage, K.: Parallel and sequential methods for ordinary differential equations. Clarendon Press, New York, NY, USA (995 4] Burrage, K., Butcher, J.: Non-linear stability of a general class of differential equation methods. BIT Numerical Mathematics 2, (98 5] Butcher, J.C.: The numerical analysis of ordinary differential equations: Runge-Kutta and general linear methods. Wiley-Interscience, New York, NY, USA (987 6] Butcher, J.C.: Diagonally-implicit multi-stage integration methods. Appl. Numer. Math. (5, (993 7] Butcher, J.C.: Numerical methods for ordinary differential equations. John Wiley & Sons Ltd., Chichester (23 8] Butcher, J.C.: General linear methods. Acta Numer. 5, (26 9] Butcher, J.C., Jackiewicz, Z.: Diagonally implicit general linear methods for ordinary differential equations. BIT 33(3, (993 ] Calvo, M.P., de Frutos, J., Novo, J.: Linearly implicit Runge-Kutta methods for advection-reaction-diffusion equations. Appl. Numer. Math. 37(4, (2 ] Cardone, A., Jackiewicz, Z.: Explicit Nordsieck methods with quadratic stability. Numerical Algorithms 6, 25 (22 2] Cardone, A., Jackiewicz, Z., Mittelmann, H.: Optimization-based search for Nordsieck methods of high order with quadratic stability. Math. Model. Anal. 7(3, (22 3] Crouzeix, M.: Une méthode multipas implicite-explicite pour l approximation des équations d évolution paraboliques. Numer. Math. 35(3, (98 33

34 4] Frank, J., Hundsdorfer, W., Verwer, J.G.: On the stability of implicitexplicit linear multistep methods. Appl. Numer. Math. 25(2-3, (997. Special issue on time integration (Amsterdam, 996 5] Hairer, E., Lubich, C., Wanner, G.: Geometric numerical integration : structure-preserving algorithms for ordinary differential equations. Springer-Verlag, New York (22 6] Hairer, E., Nørsett, S.P., Wanner, G.: Solving ordinary differential equations. I, Springer Series in Computational Mathematics, vol. 8, second edn. Springer-Verlag, Berlin (993. Nonstiff problems 7] Hairer, E., Wanner, G.: Solving ordinary differential equations. II, Springer Series in Computational Mathematics, vol. 4. Springer-Verlag, Berlin (2. Stiff and differential-algebraic problems, Second revised edition, paperback 8] Hundsdorfer, W., Verwer, J.: Numerical solution of time-dependent advection-diffusion-reaction equations, Springer Series in Computational Mathematics, vol. 33. Springer-Verlag, Berlin (23 9] Jackiewicz, Z.: General linear methods for ordinary differential equations. John Wiley & Sons Inc., Hoboken, NJ (29 2] Kennedy, C.A., Carpenter, M.H.: Additive Runge-Kutta schemes for convection-diffusion-reaction equations. Appl. Numer. Math. 44(-2, 39 8 (23 2] Liska, R., Wendroff, B.: Composite schemes for conservation laws. SIAM Journal on Numerical Analysis 35(6, pp (998 22] Pareschi, L., Russo, G.: Implicit-explicit Runge-Kutta schemes for stiff systems of differential equations. In: Recent trends in numerical analysis, Adv. Theory Comput. Math., vol. 3, pp Nova Sci. Publ., Huntington, NY (2 23] Pareschi, L., Russo, G.: Implicit-Explicit Runge-Kutta schemes and applications to hyperbolic systems with relaxation. J. Sci. Comput. 25(-2, (25 34

35 24] Prothero, A., Robinson, A.: On the stability and accuracy of onestep methods for solving stiff systems of ordinary differential equations. Math. Comput 28(25, (974 25] Varah, J.M.: Stability restrictions on second order, three level finite difference schemes for parabolic equations. SIAM J. Numer. Anal. 7(2, 3 39 (98 26] Wright, W.M.: The construction of order 4 DIMSIMs for ordinary differential equations. Numer. Algorithms 26(2, 23 3 (2 27] Zhang, H., Sandu, A.: A second-order diagonally-implicit-explicit multistage integration method. Procedia CS 9, (22 28] Zhang, H., Sandu, A.: Partitioned and implicit-explicit general linear methods for ordinary differential equations (23 29] Zharovski, E., Sandu, A.: A class of implicit-explicit two-step Runge- Kutta methods. Tech. Rep. TR-2-8, Computer Science, Virginia Tech., (22 3] Zhong, X.: Additive semi-implicit Runge-Kutta methods for computing high-speed nonequilibrium reactive flows. J. Comput. Phys. 28(, 9 3 (996 35