A MULTIGRID ALGORITHM FOR. Richard E. Ewing and Jian Shen. Institute for Scientic Computation. Texas A&M University. College Station, Texas SUMMARY

Transcription

1 A MULTIGRID ALGORITHM FOR THE CELL-CENTERED FINITE DIFFERENCE SCHEME Richard E. Ewing and Jian Shen Institute for Scientic Computation Texas A&M University College Station, Texas SUMMARY In this article, we discuss a non-variational V -cycle multigrid algorithm based on the cell-centered nite dierence scheme for solving a second-order elliptic problem with discontinuous coecients. Due to the poor approximation property of piecewise constant spaces and the non-variational nature of our scheme, one step of symmetric linear smoothing in our V -cycle multigrid scheme may fail to be a contraction. Again, because of the simple structure of the piecewise constant spaces, prolongation and restriction are trivial; we save signicant computation time with very promising computational results. INTRODUCTION In the simulation of incompressible uid ow in porous media, we have to solve at least one second-order elliptic equation per each time step. A very important quantity is the Darcy velocity, dened by u =?Krp (1) where p is the pressure of the uid and K is the conductivity. K can be written by K = k, where k is a tensor representing the permeability of the medium which can be discontinuous in general, and represents the viscosity of the uid. is a continuous function of both time and space variables, but may have a very sharp frontal change of values. In other words, can change rapidly inside the interesting domain and the region of rapid change may move as time changes. According to the conservation law of mass balance, the Darcy velocity u must be continuous along the normal direction at an element or domain boundary, no matter whether K is discontinuous or not. Now, we consider the following simple second-order elliptic equation in mixed form. Find a pair (p; u) such that u =?Krp; in = (0; 1) 2 IR 2 ; r u = f; in ; p = 0; This work was supported in part by the Department of Energy under Contract No. DE{FG05{92ER The authors would also like to thank Joe Pasciak for valuable discussions about this work. (2)

2 where the conductivity K(x; y) = diag(a; b) is positive and uniformly bounded above and below. Because of the discontinuity of K, the classical solution of p in (2) may not exist. Let (; ) denote L 2 () or (L 2 ()) 2 inner product and H(div; ) n u 2 (L 2 ()) 2 j r u 2 L 2 () o. We seek the solution pair (p; u) 2 H 1 () H(div; ), such that (K u; v) = (p; rv); 8 v 2 H(div; ); (r u; w) = (f; w); w 2 L 2 (): (3) In [5], error estimates for solving (3) by the cell-centered nite dierence scheme are studied, with the following results: kp? Ppk L 2 + ku? uk L 2 ch s kpk 1+s;n? ; s = 1; 2; (4) where P is the Raviart-Thomas projection,? are the lines of discontinuity which coincide with the grid lines, and (P; U) is the numerical solution of the cell-centered nite dierence to approximate (3)[5]. Actually, we view the cell-centered nite dierence method as a special numerical integration of the Raviart-Thomas mixed nite element method [4{6]. For s = 2, (4) is the superconvergence error estimate. >From the point of view of mass balance and accuracy, the cell-centered nite dierence scheme is one of the best numerical schemes to fulll our goal. In this article, we investigate the eciency of the multigrid algorithm based on the cell-centered nite dierence scheme introduced in [5]. NUMERICAL SCHEME IN MULTIGRID SETTING Let us use the Laplacian operator,?, to explain the cell-centered nite dierence scheme stencil. For an interior node, the stencil for? is (a) in Figure 1. For a corner node, the stencil for? is (b) in Figure 1. For other boundary nodes, the stencil for? is (c) in Figure 1. For discontinuous conductivity, see [5] for details. Now, we consider the uniform grid only. Let M k denote the piecewise constant Raviart-Thomas rectangular pressure space dened on with mesh size h k = 2?(k+1), k = 0; 1; 2; 3;... ; J. It is clear that M 0 M 1 M 2... M J M J L 2 (): (5) With an abuse of notation, for u 2 M k, u is either a piecewise function or a vector with its nodal values as its entries. On M k, the cell-centered nite dierence approximation is to nd P 2 M k, such that ~A k P = F k P k f; k = 0; 1; 2;... ; J: (6) Here P k : L 2! M k is the L 2 -projection into M k dened by (f; w) = (P k f; w); 8 w 2 M k ; (7) and f is the load function of (2). The corresponding stencil of ~ Ak is shown in Figure 1. Our goal is to nd P 2 M J, such that ~A J P = F J = P J f: (8)

3 4 5 6 (a) (b) (c) Figure 1. Stencils for the Laplacian operator. The discrete L 2 -inner product and associated norm on M k are denoted by (u; v) k = h 2 k vt u and kuk 2 k = (u; u) k; u; v 2 M k ; (9) where v T u is the usual algebraic inner product. Let A J = ~ AJ dene an associated bilinear form A J on M J by (A J w; ) J = A J (w; ); 8 w; 2 M J : (10) Before we dene A k for 0 k < J, we rst dene the prolongation operator I k and the restriction operator P 0 k. Let I k : M k! M k, k = 1; 2;... ; J be the natural imbedding from M k to M k. Thus P 0 k : M k! M k, the adjoint of I k in (; ) k, is dened by (P 0 k w; ) k = (w; I k ) k ; w 2 M k ; 2 M k : (11) >From (9) and h k = 2h k, it is clear that P 0 k = 1 4 I T k form A k (; ) and the matrix A k on M k for k = J; J? 1;... ; 2; 1, by and the corresponding matrix relation is in matrix form. Now, we dene the bilinear 2A k (u; v) = A k (I k u; I k v); 8 u; v 2 M k ; (12) A k = 1 8 I T k A ki k = 1 2 P 0 k A ki k : (12 0 ) Remark 1. It is shown in [5] that for piecewise smooth conductivity tensor K, as long as the discontinuities coincide with the coarser grid lines A k = 1 + O(h 2 k ) ~ Ak : (13) In (13) O(h 2 k) = Ch 2 k. C depends on the local smoothness of K but is independent of the jumps. Since I k is a simple operator, it is much easier to generate A k by (12 0 ) than by (6) directly. Of course, A k, k = 0; 1; 2;... ; J? 1, are all positive denite since A J is, and the spaces are nested. Because of (12), our multigrid algorithm can be considered as a black box solver once I k has been dened. We mention that (12) holds for three-dimensional problems of?r (Kru), with (12 0 ) being changed to A k = 1 16 I T k A ki k = 1 2 P 0 k A ki k :

4 We also dene the adjoint of I k in A k (; ), P k : M k! M k by A k (P k u; v) = A k (u; I k v); u 2 M k ; v 2 M k : (14) To dene the smoothing process, we require linear operators R k : M k! M k for k = 1; 2;... ; J. These operators may or may not be symmetric with respect to the inner product (; ) k. Let A k = D k + L k + L T k, D k be the diagonal part of A k, and L k be the lower triangular part of A k. The linear smoothers we have tried are the following relaxation schemes. For 0 <! < 2, (a) Gauss-Seidel: R k = Dk! + L k (b) Jacobi: R k =!D k ; (c) Richardson: R k =! k I; and R T k ; (15) where I is the identity operator on M k and k is the spectral radius of A k. We allow the relaxation parameter! to be dierent for pre-smoothing and post-smoothing processes in the following denition. Following [1] the multigrid operator B k : M k! M k is dened by induction and is given as follows. The pre-smoother is denoted by R k and the post-smoother by Rk. V -Cycle Multigrid Algorithm: Set B 0 = A 0. Assume that B k has been dened and dene B k g for g 2 M k as follows: 1. Set x 0 = Dene x` for ` = 1; 2;... ; m(k) by x` = x` + R k (g? A k x` ). 3. Set y 0 = x m(k) + I k B k P 0 k g? A k x m(k). 4. Dene y` for ` = 1; 2;... ; m(k) by y` = y` + R k (g? A k y` ). 5. Set B k g = y m(k). Remark 2. Since equation (12) holds for all levels, this multigrid algorithm is non-variational according to [1], but the approximation property (4) is valid for each level as long as the non-variational relation (12) is satised. In this algorithm m(k) is a positive integer which may vary from level to level. In general this multigrid algorithm is not symmetric in (; ) k except for R k = R T k. Setting K k = I? R k A k and Kk = I? Rk A k, it is straightforward to check that I? B k A k = K m 2(k) k [I? I k B k P 0 A k k]k m 1(k) k = K m 2 (k) k [I? I k B k A k P k ]K m 1(k) k : (16)

5 Equation (16) gives a fundamental recurrence relation for the multigrid operator B k. COMPUTATIONAL EXPERIMENTS We have tested the multigrid algorithm described in Section 2. We use a power method to compute the largest and the smallest eigenvalues of B J A J. The linear smoothers we have tried are the following. Let m be a positive integer. m(k) = m for all k, S 1 (m) : R k = 1 k I; Rk = R k ; S 2 (m) : R k = 1 k I; Rk = 2R k ; where k is the largest eigenvalue of A k, S 3 (m) : R k = (D k + L k ) ; Rk = R T k, S 4 (m) : R k = 1:35D k ; Rk = 1 2 R k, S 5 (m) : R k = Dk 1:35 + L k ; Rk Dk = 0:675 + LT k, S 6 (m) : R k = 2 3 D k + L k ; Rk = (2D k + L T k ). Note that only S 1 (m) and S 3 (m) make B J A j A J (; ) symmetric. The rest are neither symmetric nor A J (; ) symmetric. We also have tried nonlinear smoothers, conjugate gradient, and diagonally preconditioned conjugate gradient algorithms. We shall use N(m) to represent our nonlinear multigrid by diagonally preconditioned conjugate gradient smoothers. The reason we choose dierent relaxation numbers comes from the suggestion [3] for an algebraic multigrid algorithm, and from our computational experiments. We list our test results in Tables 1{9 at the end of this paper for the following problems: Ex. 1. Poisson problem: K 1 in (2). Ex. 2. Isotropic problem with nearly singular piecewise smooth conductivity: K = h 0: :1 1 + cos(3:561x) sin(3:001y) i q; q = ( 10?4 ; if x 1 2 and y 1 2 ; 1; otherwise.

6 Ex. 3. Same kind of problems as Ex. 2: K = h 0: :1 1 + cos(9:431x) sin(3:001y) i q; q = ( 10 4 ; if x 1 2 and y 1 2 ; 1; otherwise. Ex. 4. Anisotropic problem with smooth conductivity: K = diag(a; b); a = 0: :1 1 + cos(9:431x) sin(9:431y) ; b = 0: :1 (1 + sin(9:431x) cos(9:431y)) : Note that all the solutions of our examples have the superconvergence results proved in [5], i.e., satisfying (4) with s = 2. In Tables 1 and 6, for example, the second row of Table 1 means J + 1 = 3 level multigrid with h J = 1, 8 m; S 1 (1) means m = min (B J A J ) by S 1 (1) smoothers, and M ; S 1 (1) means M = max (B J A J ) by S 1 (1) smoothers. From Table 1, we can see that even when I? B J A J fails to be a reducer, B J may still be a good preconditioner. In Tables 5{7, it is interesting to see the relations of the number of V -cycles (#V ), average contraction numbers (avc) and the time spent on the machine (cpu in seconds) when solving a xed problem on a xed grid by using dierent multilevels. In Tables 3{5, and 7{9, avc is dened by avc = 1 n nx j=1 kr j k 2 J ; kr j k 2 y where n = #V is the total number of V -cycles and kr j k J is the discrete L 2 -norm of the residual after the jth V -cycle. The stop tolerance for all the iterative algorithms is kr n k 2 J = Our coarsest grid solver is a diagonal preconditioned conjugate gradient solver with tolerance 0 = In Tables 9{11, \cg" means the standard conjugate gradient algorithm, its corresponding \#V " means the total iteration steps, when kr n k 2 J = 10 4, and \bpcg" means the incomplete factorization preconditioned conjugate gradient algorithm [2]. REFERENCES 1. Bramble, J. H, Pasciak, J. E., and Xu, J.: The Analysis of Multigrid Algorithm with Nonnested Spaces or Noninherited Quadratic Forms. Math. Comp., vol. 56, 1991, pp. 1{ Concus, P., Golub, G. H., and Meurant, G.: Block Preconditioning for the Conjugate Gradient Method. SIAM J. Sci. Stat. Comput., vol. 6, 1985, pp. 220{ McCormick, S. F., ed.: Multigrid Methods. SIAM, Philadelphia, Pennsylvania, 1987.

7 4. Russell, T.F. and Wheeler, M.F.: Finite Element and Finite Dierence Methods for Continuous Flows in Porous Media. The Mathematics of Reservoir Simulation (R. E. Ewing, ed.). SIAM, Philadelphia, Pennsylvania, Shen, J.: Mixed Finite Element Methods: Analysis and Computational Aspects. Ph.D. Thesis, University of Wyoming, Weiser, A. and Wheeler, M.F.: On the Convergence of Block-Centered Finite Dierences for Elliptic Problems. SIAM J. Numer. Anal., vol. 25, 1988, pp. 351{375. Table 1. For Ex. 1 Grid J m ; S 1 (1) M ; S 1 (1) m ; S 1 (2) M ; S 1 (2) Table 2. For Ex. 1 Grid J m ; S 3 (1) M ; S 3 (1) m ; S 3 (2) M ; S 3 (2)

8 Table 3. For Ex. 1 by B J (S 2 (1)) Grid J = 1 J = 2 J = 3 J = 4 J = 5 J = 6 J = 7 J = 8 #V avc cpu #V avc cpu #V avc cpu Table 4. For Ex. 1 by B J (S 3 (1)) Grid J = 1 J = 2 J = 3 J = 4 J = 5 J = 6 J = 7 J = 8 #V avc cpu #V avc cpu #V avc cpu

9 Table 5. For Ex. 1 by B J (S 3 (2)) Grid J = 1 J = 2 J = 3 J = 4 J = 5 J = 6 J = 7 J = 8 #V avc cpu #V avc cpu #V avc cpu Table 6. For Ex. 2 Grid J m ; S 3 (1) M ; S 3 (1) Table 7. For Ex. 3 Grid J N (1) S 3 (1) S 5 (1) S 4 (1) cg kr 0 k 2 J #V , avc : cpu #V , avc cpu ,835.0 #V , avc : cpu ,003.0

10 Table 8. For Ex. 3 Grid J N (2) S 3 (2) S 5 (2) S 6 (2) S 4 (2) bpcg #V avc cpu #V avc cpu #V avc cpu Table 9. For Ex. 4 Grid J N (3) S 6 (2) S 4 (3) bpcg cg kr 0 k 2 J #V avc : cpu #V avc : cpu #V , avc : cpu