# 32 comparisons of results obtained using sor cg and

This preview shows page 1. Sign up to view the full content.

This is the end of the preview. Sign up to access the rest of the document.

Unformatted text preview: rior, but the SSOR-PCG method overtakes these methods for larger systems. One of the most successful preconditioning techniques utilizes incomplete factorization by Gaussian elimination. Thus, let M = L0D0 LT 0 (9.3.28) 9.3. Conjugate Gradient Methods 47 1=h SOR CG PCG 10 33 26 12 20 60 52 16 40 115 103 22 Table 9.3.2: Number of SOR, conjugate gradient, and SSOR-PCG iterations for Example 9.3.4. J J A= L 0= Figure 9.3.6: Nonzero structure of a discrete Poisson operator A (left) and an approximate lower triangular factor L0 (right). where L0 is determined to have a particular sparsity structure. Example 9.3.5. Consider a Dirichlet problem for Poisson's equation, which has the nonzero structure shown in Figure 9.3.6. Ordering the equations by rows, we have (Ax)n = ;xn + x(xn;1 + xn+1) + y (xn;J + xn+J ): (9.3.29a) If equation n corresponds to mesh point (j k), we may also write (9.3.29a) in the (doublesubscripted) mesh notation as (Ax)jk = ;xjk + x(xj;1 k + xj+1 k ) + y (xj k;1 + xj k+1): (9.3.29b) We insist that (L0 x)n = xn + bnxn;1 + cnxn;J : (9.3.30a) (LT x)n = xn + bn+1xn+1 + cn+J xn+J : 0 (9.3.30b) Taking a transpose 48 Solution Techniques for Elliptic Problems Likewise (D0x)n = dnxn : (9.3.30c) Multiplying (9.3.30b) and (9.3.30c) (D0LT x)n = dn(xn + bn+1 xn+1 + cn+J xn+J ): 0 Now, using (9.3.30a) (L0 D0LT x)n = (D0LT x)n + bn(D0 LT x)n;1 + cn(D0 LT x)n;J : 0 0 0 0 Combining the above two equations (L0D0 LT x)n = dn(xn + bn+1 xn+1 + cn+J xn+J ) + bndn;1(xn;1 + bnxn + cn+J ;1xn+J ;1 )+ 0 cndn;J (xn;J + bn+1;J xn+1;J + cnxn ): Regrouping terms (L0 D0LT x)n = (dn + dn;1b2 + dn;J c2 )xn + dn;1bn xn;1 + dnbn+1 xn+1+ 0 n n dncn+J xn+J + dn;J cnxn;J + bn+1;J dn;J cnxn+1;J + bn dn;1cn+J ;1xn+J ;1 : (9.3.31) Thus, L0 D0LT has seven non-trivial bands as shown in Figure 9.3.7. 0 Using incomplete factorization, we equate the coe cients of xn, xn+1 , and xn+J in (9.3.31) to those of (Ax)n in (9.3.29a). This yields dn = ;1 ; dn;1b2 ; dn;J c2 n n n = 1 2 : : : N: (9.3.32a) bn+1 = x=dn n = 1 2 : : : N ; 1: (9.3.32b) cn+J = y =dn n = 1 2 : : : N ; J: (9.3.32c) Example 9.3.6. ( 5], Section 14.5). Consider the solution of Laplace's equation on the unit square using a squ...
View Full Document

• Spring '14
• JosephE.Flaherty
• Articles with example pseudocode, Gauss–Seidel method, Jacobi method, Iterative method, elliptic problems

{[ snackBarMessage ]}

Ask a homework question - tutors are online