A FAST POISSON SOLVER BY CHEBYSHEV PSEUDOSPECTRAL METHOD USING REFLEXIVE DECOMPOSITION. Teng-Yao Kuo, Hsin-Chu Chen and Tzyy-Leng Horng*

Transcription

1 TAWANESE JOURNAL OF MATHEMATCS Vol. 17, No. 4, pp , August 213 DO: /tjm This paper is available online at A FAST POSSON SOLVER BY CHEBYSHEV PSEUDOSPECTRAL METHOD USNG REFLEXVE DECOMPOSTON Teng-Yao Kuo, Hsin-Chu Chen and Tzyy-Leng Horng* Abstract. Poisson equation is frequently encountered in mathematical modeling for scientific and engineering applications. Fast Poisson numerical solvers for 2D and 3D problems are, thus, highly requested. n this paper, we consider solving the Poisson equation 2 u = f(x, y) in the Cartesian domain Ω=[ 1, 1] [ 1, 1], subject to all types of boundary conditions, discretized with the Chebyshev pseudospectral method. The main purpose of this paper is to propose a reflexive decomposition scheme for orthogonally decoupling the linear system obtained from the discretization into independent subsystems via the exploration of a special reflexive property inherent in the second-order Chebyshev collocation derivative matrix. The decomposition will introduce coarse-grain parallelism suitable for parallel computations. This approach can be applied to more general linear elliptic problems discretized with the Chebyshev pseudospectral method, so long as the discretized problems possess reflexive property. Numerical examples with error analysis are presented to demonstrate the validity and advantage of the proposed approach. 1. NTRODUCTON There have been considerable studies in the development of robust and efficient solvers for Poisson equation. A fast and accurate Poisson solver has many scientific and engineering applications. These include computer simulations of plasma physics [3], industrial plasma engineering [17], and planetary dynamics [4]. The popular projection method for solving Navier-Stokes equations also involves solving a Poisson equation for the pressure field [7]. Received October 3, 212, accepted January 3, 213. Communicated by Wen-Wei Lin. 21 Mathematics Subject Classification: 35J5, 65F5, 65N35. Key words and phrases: Poisson equation, Chebyshev pseudospectral method, Chebyshev collocation derivative matrix, Coarse-grain parallelism, Reflexive property. *Corresponding author. 1167

2 1168 Teng-Yao Kuo, Hsin-Chu Chen and Tzyy-Leng Horng n this paper, we consider solving the Poisson equation 2 u = f(x, y) in a Cartesian domain Ω=[ 1, 1] [ 1, 1] with all types of boundary conditions: (1.1) 2 u = f(x, y), (x, y) ( 1, 1) ( 1, 1), u α 1 u + β 1 x = g 1(y), at x =1, u α 2 u + β 2 x = g 2(y), at x = 1, u α 3 u + β 3 y = g 3(x), at y =1, u α 4 u + β 4 y = g 4(x), at y = 1, where α i and β i, i, are constants and can not be zeros at the same time. Also, at least one of the α is is not zero to ensure the uniqueness of solution. The current paper proposes a reflexive decomposition scheme to solve (1.1) more efficiently when (1.1) is discretized by the Chebyshev pseudospectral method. Most Poisson solvers are based on finite difference or finite element methods. While the geometry of computational domain is rectangular, the Chebyshev pseudospectral method has advantage over traditional finite difference and finite element methods in numerical accuracy. ts convergence rate is of exponential order O ( c N), <c<1, for smooth solutions where N is the number of grid points [18]. There have been several literatures about solving Poisson equation over a rectangular domain by Chebyshev spectral/pseudospectral method, to name a few as follows. Haidvogel and Zang [14] applied both AD and matrix diagonalization techniques to solve Poisson equation over a square domain, and compared their efficiencies. Their method can be only applied to Dirichlet boundary conditions. Dang-Vu and Delcarte [11] presented the solution of Poisson equation using the resolution of the mixed collocation τ equations into two quasi-tridiagonal systems to simplify the differential operators. Shizgal et al. [2, 21, 12] used eigenvalue/eigenvector technique to diagonalize the discretized system, which is a very efficient method but only subject to Dirichlet boundary conditions. So far, the eigenvalue/eigenvector matrix diagonalization method for solving Poisson equation over a rectangular domain is most efficient [9, 13]. However, it can be only applied with ease for Dirichlet boundary conditions. For non-dirichlet boundary conditions like Neumann or Robin ones, this matrix diagonalization method requests messy row operations or costly iterations [22]. To apply Chebyshev pseudospectral method coping with Neumann or Robin boundary conditions, even more general second-order arbitrary-coefficient elliptic partial differential equations, we need to use Kronecker product during discretization for Poisson equation in 2D/3D rectangular domains. The resultant linear system is very costly to solve, especially for 3D problems, by direct Gaussian elimination or iteration methods.

3 Fast Poisson Solver by Chebyshev Pseudospectral Method 1169 To solve this problem, here we explore the reflexive property [1] of Chebyshev collocation derivative matrix for possible coarse-grain decomposition of the resultant huge matrix. To be more elaborate, we apply reflexive decomposition to orthogonally decompose the original linear system into two, or more, smaller decoupled subsystems so that we can save enormous computation time by solving several smaller linear systems instead. This coarse-grain system reduction can further foster coarse-grain parallelism, which can save more computation time by parallel computation. The organization of this paper is as follows. n section 2, we begin with an observation that the second-order Chebyshev collocation derivative matrix is centrosymmetric and then show that the matrix associated with the 2D discretized Poisson equation using Chebyshev pseudospectral method with a tensor product is block-centrosymmetric and therefore reflexive. n section 3, we further decompose the resultant matrix into submatrices via orthogonal reflexive decomposition. Operations count and numerical experiments with error analysis showing exponential convergence are presented in section 4, and conclusion is given in section REFLEXVE PROPERTY OF CHEBYSHEV COLLOCATON DERVATVE MATRX 2.1. Chebyshev pseudospectral method Given Chebyshev-Gauss-Lobatto ( ) points, x j =cos(jπ/n), j=, 1,..., N, satisfying T N (x j) 1 x 2 j =, where T N (x) is the Chebyshev polynomial of degree N, the Lagrange interpolating polynomials based on Chebyshev-Gauss-Lobatto points can be obtained as follows ( L N,j (x) = ( 1)j+1 1 x 2) T N (x) c j N 2,j=, 1,..., N, (x x j ) { 2, for j =,N, where c j = and L 1, otherwise, N,j (x l )=δ jl with δ jl being Kronecker delta. A function u (x) can be approximated by interpolating polynomials above, u (x) N L N,j (x) u (x j ), j= and its derivative values at Chebyshev-Gauss-Lobatto points can be therefore approximated by u (x i ) N L N,j (x i ) u (x j )= j= N (D N ) ij u (x j ), i =, 1,..., N, j=

4 117 Teng-Yao Kuo, Hsin-Chu Chen and Tzyy-Leng Horng where D N is Chebyshev collocation derivative matrix with (D N ) ij = L N,j (x i). The entries of D N have long been available in literatures [15, 16, 18] and are given as (D N ) = 2N 2 +1, (D N ) 6 NN = 2N 2 +1, 6 x j (D N ) jj = ( ), j =1, 2,..., N 1, 2 1 x 2 j (D N ) ij = c i ( 1) i+j, i j, and i, j =, 1,..., N, c j (x i x j ) { 2, for i =,N, where c i = and note that D 1, otherwise, N is a (N +1) (N +1)matrix. Likewise, the k th derivative values of a function u (x) at Chebyshev-Gauss-Lobatto points can be approximated by ( k ) u x k x=x i N j= ( ) DN k u (x j),i=, 1,..., N, ij where D k N represents the kth power of D N. Note that D N is an anti-centrosymmetric matrix, that satisfies or (D N ) ij = (D N ) N i,n j,i,j=, 1,..., N, D N = J N+1 D N J N+1, where J N+1 = 1 1. Accordingly, the second-order derivative matrix DN 2, denoted by S, will be a centrosymmetric matrix [1, 2, 6, 19] satisfying (S) ij =(S) N i,n j,i,j=, 1,..., N, or S = J N+1 SJ N+1. Deriving this will use the property JN+1 2 = N+1, where N+1 is the identity matrix of dimension N +1. Extending this idea, DN k will be anti-centrosymmetric for any odd k and centrosymmetric for any even k, i.e., ( DN k ) ij = ( ( 1)k DN k ) N i,n j, i, j =, 1,..., N. To solve 2D Poisson problem like (1.1), we apply tensor product to discretize the equation as (2.1) L N u = f, with L N R (N+1)2 (N+1) 2, u,f R (N+1)2,

5 Fast Poisson Solver by Chebyshev Pseudospectral Method 1171 where L N is the discretized Laplacian operator that can be expressed as (2.2) L N = N+1 S + S N+1, with representing the Kronecker product [18]. The Kronecker product of two matrices A and B, denoted by A B, with dimensions m n and p q respectively, can be expressed as an m n block matrix with the i, j block being a ij B. For example, a 11 a 12 a 21 a 22 a 31 a 32 [ ] = 1a 11 2a 11 1a 12 2a 12 3a 11 4a 11 3a 12 4a 12 1a 21 2a 21 1a 22 2a 22 3a 21 4a 21 3a 22 4a 22 1a 31 2a 31 1a 32 2a 32 3a 31 4a 31 3a 32 4a mplementation of boundary conditions To cope with coarse-grain decomposition later, here we allow all kinds of boundary conditions at y = ±1, but only consider Dirichlet boundary conditions at x = ±1. t should be noted that at least on one of the coordinate direction the boundary conditions at both end boundaries are of Dirichlet type. Otherwise, the reflexive property is lost and the method may fail. Hence, S should be modified to include boundary conditions mentioned above. Let S and S denote the boundary-operator-included S matrices along x and y directions respectively. We then have 1 (2.3) S = (S) ij and S = 1 Therefore, the resultant matrix L N for this case is α 3 ( N+1 ) 1,j+1 + β 3 (D N ),j (S) ij α 4 ( N+1 ) N+1,j+1 + β 4 (D N ) N,j. (2.4) L N = N+1 S + S N+1 = K 1 + K 2 with explicit expression of K 1 and K 2 being K 1 = S... S and K 2 = s 1 s 11 s 1,N... s N 1, s N 1,1 s N 1,N,

6 1172 Teng-Yao Kuo, Hsin-Chu Chen and Tzyy-Leng Horng where s ij =(S) ij,= N+1, and = N 1. (N+1) (N+1) Actually, we can further express (2.5) K 1 = S and K 2 = S +( ) ( ). Note that L N above is already further modified to include the boundary condition at y = ±1. To take care of the compatibility of different kinds of boundary conditions at the corner boundary points, extra modification of the boundary condition there is further needed Reflexive property of L N Let R be the reflection matrix with R = J N+1 N+1. Herewewanttoshow that L N in (2.4) is reflexive with respect to R. t suffices to say so if both K 1 and K 2 are reflexive with respect to R. n following, we drop the subscripts of and J for notational brevity. First, we can easily observe that K 1 is reflexive with respect to R by S RK 1 R =... S S =... = K 1. S Similarly, K 2 also satisfies the reflexive property by RK 2 R = s 1 s 11 s 1,N... s N 1, s N 1,1 s N 1,N

7 Fast Poisson Solver by Chebyshev Pseudospectral Method 1173 s N 1,N s N 1,1 s N 1, =... s 1,N s 11 s 1 s 1 s 11 s 1,N =... = K 2. s N 1, s N 1,1 s N 1,N Here we used the centrosymmetric property s ij = ( DN 2 ) ij = ( DN 2 ) N i,n j = s N i,n j for the derivation above. 3. REFLEXVE DECOMPOSTON OF L N As seen in the previous section, the matrix L N is reflexive with respect to R. This reflexive property enables us to decompose L N into smaller submatrices using orthogonal transformation. The decomposition is referred to as the reflexive decomposition because of the reflection matrix R General forms of the decomposition The general forms of the decomposition of L N are categorized to N +1 being even or odd. For notational brevity, let (3.1) L N = K. Decomposition 1: N+1 even. Let N +1 = 2k. We evenly partition R and K into 2 2 sub-blocks as [ ] [ ] R1 K11 K (3.2) R = and K = 12, R 1 K 21 K 22 where R 1 = J k N+1 ; K ij R k(n+1) k(n+1), i, j =1, 2. From the reflexive property K = RKR, we can derive the following relations easily, (3.3) K 11 = R 1 K 22 R 1,K 12 = R 1 K 21 R 1, K 21 = R 1 K 12 R 1,K 22 = R 1 K 11 R 1. Let Q be the following orthogonal matrix, (3.4) Q = 1 [ 2 R 1 ] R1.

8 1174 Teng-Yao Kuo, Hsin-Chu Chen and Tzyy-Leng Horng Using (3.3) and the facts R 2 1 = and RT 1 = R 1, it can be easily shown that the transformation Q T KQ is block-diagonal and therefore decouples K into two independent submatrices as follows [5, 8, 1], (3.5) Q T KQ = 1 [ ][ ][ ] [ ] R1 K11 K 12 R1 D1 =, 2 R 1 K 21 K 22 R 1 D 2 where D 1 = K 11 + K 12 R 1 and D 2 = K 22 K 21 R 1. Decomposition 2: N+1 odd. Let N +1=2k +1.LetRbe partitioned as (3.6) R = R 1 N+1 R 1,R 1 = J k N+1. The matrix K is then partitioned in accordance with R as K 11 K 12 K 13 K = K 21 K 22 K 23, K 11,K 13,K 31,K 33 R k(n+1) k(n+1), K 31 K 32 K 33 K 12,K 32 R k(n+1) (N+1), K 21,K 23 R (N+1) k(n+1),k 22 R (N+1) (N+1). From K = RKR, we can derive (3.7) K 11 = R 1 K 33 R 1, K 12 = R 1 K 32, K 13 = R 1 K 31 R 1,K 21 = K 23 R 1. Let Q be the following orthogonal matrix, (3.8) Q = 1 R 1 2N+1. 2 R 1 By (3.7), the orthogonal transformation Q T KQ yields [1] [ ] (3.9) Q T D1 KQ = D 2 with D 1 = [ ] K11 + K 13 R 1 2K12 2K21,D 2 = K 33 K 31 R 1. K 22 Then, the process of solving linear system (2.1) by reflexive decomposition starts with Ku = f,

9 Fast Poisson Solver by Chebyshev Pseudospectral Method 1175 and then or expressed as Q T KQQ T u = Q T f, (3.1) ˆKû = ˆf, with ˆK = Q T KQ, û = Q T u, and ˆf = Q T f. By D 1 and D 2 obtained above, (3.1) can be further expressed as (3.11) [ ] [ ] D1 ][û1 ˆf1 =, D 2 û 2 ˆf 2 or { D1 û 1 = ˆf 1, D 2 û 2 = ˆf 2. After solving û 1 and û 2, we can recover û, and finally obtain u by u = Qû Explicit forms of the decomposition n this section, we want to derive the explicit forms of D 1 and D 2 mentioned above for the convenience of coding. Again, we categorize cases according to N +1being even or odd. Explicit form 1: N +1=2k. Let k and O k denote the identity matrix and null matrix of dimension k. Let also [ ] [ ] [ ] [ ] k 1 1 Ok 1 (3.12) 1 =, 1 =, O 1 =,O 1 =, 1 k 1 and we can easily see [ ] 1 (3.13) =. 1 O k 1 Following the strategy in previous section, we can evenly partition S into 2 2 subblocks as [ ] S11 S12 (3.14) S =, S ij R S k k. 21 S22 By (2.5), (3.13), and (3.14), K 1 and K 2 can then be expressed in their partitioned forms as [ ] (3.15) K 1 = S 1 S = 1 S,

10 1176 Teng-Yao Kuo, Hsin-Chu Chen and Tzyy-Leng Horng and (3.16) K 2 = S +( ) ( ) [ ] [ ] S = 11 S 12 O1 ( + ) O k S 21 S22 O k O 1 ( ) [ ] S = 11 + O 1 ( ) S 12. S 21 S22 + O 1 ( ) Now, from (3.15), (3.16), and (3.2), we have K 11 = 1 S + S 11 + O 1 ( ), K 12 = S 12, K 21 = S 21, K 22 = 1 S + S 22 + O 1 ( ). Accordingly, we then obtain the explicit forms of D 1 and D 2, ( ) ( ) D 1 = 1 S + S 11 + O 1 ( ) + S 12 R 1, (3.17) ( D 2 = 1 S + S ) ( ) 22 + O 1 ( ) S21 R 1. Explicit form 2: N +1=2k +1. n this case, we partition K 1 and K 2 by their tensor-product forms as 1 K 1 = S = 1 S, 1 and K 2 = S +( ) ( ) S 11 S 12 S 13 1 = S 21 S22 S23 + O S 31 S32 S33 O 1 ( ), where S 11, S 13, S 31, S 33 R k k, S12, S 32 R k 1, S21, S 23 R 1 k and S 22 R. Accordingly, we have K = K 1 + K 2 1 S + S 11 + O 1 ( ) S 12 S 13 = S 21 S+ S 22 S 23. S 31 S 32 1 S+ S 33 +O 1 ( )

11 Fast Poisson Solver by Chebyshev Pseudospectral Method 1177 Therefore, the decomposed submatrices D 1 and D 2 are obtained explicitly as follows, (3.18) D 1 = [ 1 S+ S 11 +O 1 ( )+( S 13 )R 1 2( S 12 ) 2( S21 ) S + S22 ], D 2 = 1 S + S 33 + O 1 ( ) ( S 31 )R ERROR ANALYSS AND OPERATONS COUNT The spectral convergence is still preserved during decomposition, and this is demonstrated by the following numerical experiments. Experiment 1. Considering the following 2D Poisson equation 2 u = f(x, y), (x, y) [ 1, 1] [ 1, 1], f(x, y) = [ 2 π 2 ( y 2 +2y +1 )] sin πx, subject to the following boundary conditions u (±1,y)=,u y (x, 1) =, and u (x, 1) + u y (x, 1) = 8 sin πx, with the exact solution being u (x, y)= ( y 2 +2y +1 ) sin πx. Though the aforementioned decomposition is illustrated by 2D cases, it actually can be extended to 3D as well. A 3D Poisson problem below is also computed here for demonstration. Experiment 2. Considering the following 3D Poisson equation 2 u = f(x, y, z), (x, y, z) [ 1, 1] [ 1, 1] [ 1, 1], f(x, y, z)=2 [( y 2 1 ) (z 1)+ ( x 2 1 ) (z 1)+ ( x 2 1 )( y 2 1 ) (z+3)/8 ] e z/2, subject to the following boundary conditions with the exact solution being u (±1,y,z)=u (x, ±1,z)=,u z (x, y, 1) =, 2u (x, y, 1) + 1u z (x, y, 1) = 1 e ( x 2 1 )( y 2 1 ), u (x, y, z)= ( x 2 1 )( y 2 1 ) (z 1) e z/2. The errors of both experiments above are shown in Table 1. Basically, we can observe the exponential convergence, the feature of pseudospectral methods, in Table 1asN increasing until contaminated by round-off errors (in 3D case).

12 1178 Teng-Yao Kuo, Hsin-Chu Chen and Tzyy-Leng Horng Table 1. L error vs. N for experiment 1 and 2 N L error 2D 3D Taking 2D Poisson problem as example, the structure of L N = K will look like Figure 1(a). The operations count of solving (2.1) without reflexive decomposition by Gaussian elimination would be (4.1) C M = 1 3 M 3 + M M, with M =(N +1) 2. With reflexive decomposition, K is transformed to ˆK, and the structure of ˆK is shown in Figure 1(b). Fig. 1. The structure of K and ˆK shown in (a) and (b) respectively. The operations count of solving (2.1) with reflexive decomposition by Gaussian elimination, including overheads, would then be (4.2) C M = { 1 12 M M M, M is even, M M M, M is odd. 6

13 Fast Poisson Solver by Chebyshev Pseudospectral Method 1179 n addition to the advantage of computing time reduction by the obvious comparison between (4.1) and (4.2), the decomposition also yields coarse-grain parallelism, which can be conducted by parallel computing. f higher computing efficiency is requested, especially when N is large, it should be noted that D 1 and D 2 turn out to be reflexive too, and can be decomposed by the same procedure stated above. 5. CONCLUSON n this paper, we first explore the inherent reflexive property of second-order Chebyshev collocation derivative matrix subject to all kinds of boundary conditions. We then apply reflexive decomposition to decompose the matrix resultant from the discretization of Poisson equation by Chebyshev pseudospectral method into two smaller submatrices. Computing these two smaller linear systems definitely saves time from computing the original large system, not to mention its availability for coarse-grain parallel computing. This time saving effect is particularly notable for 3D cases. These two decomposed submatrices can be further decomposed to four even-smaller submatrices by repeating reflexive decomposition. Numerical examples are also presented to demonstrate the spectral accuracy. n addition, the current decomposition can be extended to solve au xx + bu yy + cu y + du = f (x, y) with a, b, c and d being the functions of x and y and symmetric about y axis subject to the constraint of reflexive property. Certainly, this includes the famous Helmholtz equation. ACKNOWLEDGMENTS The second author Dr. Chen wishes to thank his home institution, Clark Atlanta University, U.S.A. for supporting this research area through a sabbatical leave and Feng Chia University and Tatung University, both in Taiwan for giving him enjoyable environments during his visit to the universities in 29. The corresponding author Dr. Horng wants to thank the partial support of National Science Council of Taiwan under grant no. NSC M-35-1 and NSC E-35-1-MY3. REFERENCES 1. A. L. Andrew, Solution of equations involving centrosymmetric matrices, Technometrics, 15(2) (1973), A. L. Andrew, Eigenvectors of certain matrices, Linear Algebra Appl., 7 (1973), C. K. Birdsall and A. B. Langdon, Plasma Physics via Computer Simulation, McGraw- Hill, New York, J. Birney and S. Tremaine, Galactic Dynamics, Princeton University Press, Princeton, 1987.

14 118 Teng-Yao Kuo, Hsin-Chu Chen and Tzyy-Leng Horng 5. H. C. Chen, ncreasing parallelism in the finite strip formulation: static analysis, Neural Parallel Sci. Comput., 2 (1994), A. Cantoni and P. Butler, Eigenvalues and eigenvectors of symmetric centrosymmetric matrices, Linear Algebra Appl., 13 (1976), C. Canuto, M. Y. Hussaini, A. Quarteroni and T. A. Zang, Spectral Methods in Fluid Dynamics, Springer-Verlag, Berlin, H. C. Chen, T. L. Horng and Y. H. Yang, Reflexive decompositions for solving Poisson equation by Chebyshev pseudospectral method, Proceedings of Neural, Parallel, and Scientific Computations, 4 (21), H. Chen, Y. Su and B. D. Shizgal, A direct spectral collocation Poisson solver in polar and cylindrical coordinates, J. Comput. Phys., 16 (2), H. C. Chen and A. Sameh, A matrix decomposition method for orthotropic elasticity problems, SAM J. Matrix Anal. Appl., 1(1) (1989), H. Dang-Vu and C. Delcarte, An accurate solution of the Poisson equation by the Chebyshev collocation method, J. Comput. Phys., 14 (1993), U. Ehrenstein and R. Peyret, A Chebyshev collocation method for Navier-Stokes equations with applications to double-diffusive convection, nternat. J. Numer. Methods Fluids, 9 (1989), T. L. Horng and C. H. Teng, An error minimized pseudospectral penalty direct Poisson solver, J. Comput. Phys., 231(6) (212), D. B. Haidvogel and T. Zang, The accurate solution of Poisson s equation by expansion in Chebyshev polynomials, J. Comput. Phys., 3 (1979), J. d. J. Martinez and P. d. T. T. Esperanca, A Chebyshev collocation spectral method for numerical simulation of incompressible flow problems, J. Braz. Soc. Mech. Sci. & Eng., XXX(3) (27), R. Peyret, Spectral methods for incompressible viscous flow, Applied Mathematical Sciences, 148, Springer-Verlag, J. R. Roth, ndustrial Plasma Engineering, nst. of Phys., London, L. N. Trefethen, Spectral Methods in MATLAB, SAM, Philadelphia, J. R. Weaver, Centrosymmetric (cross-symmetric) matrices, their basic properties, eigenvalues, and eigenvectors, The American Mathematical Monthly, 92(1) (1985), H. H. Yang, B. R. Seymour and B. D. Shizgal, A Chebyshev pseudospectral multidomain method for steady flow past a cylinder up to Re = 15, Comput. & Fluids, 23 (1994), H. H. Yang and B. Shizgal, Chebyshev pseudospectral multi-domain technique for viscous flow calculation, Comput. Methods Appl. Mech. Engrg., 118 (1994), S. Zhao and M. J. Yedlin, A new iterative Chebyshev spectral method for solving the elliptic equation (σ u) =f, J. Comput. Phys., 113 (1994),

15 Fast Poisson Solver by Chebyshev Pseudospectral Method 1181 Teng-Yao Kuo Ph.D. Program in Mechanical and Aeronautical Engineering Feng Chia University Taichung 4724, Taiwan Hsin-Chu Chen Department of Computer and nformation Science Clark Atlanta University Atlanta, GA 3314 U.S.A. Tzyy-Leng Horng Department of Applied Mathematics Feng Chia University Taichung 4724, Taiwan