Sie sind auf Seite 1von 7

Department of Mechanical Engineering

TITLE:

INCONPLETE LU FACTORIZATION
SUBMITTED TO:

Dr. Mazhar Iqbal


GROUP MEMBERS:

Muhammad Uzair Gill Muhammad Enam ul Haq Muhammad Rizwan Muhammad Khurram Azad Asad-ur-Rehman

ILU DECOMPOSITION
BASIC IDEAS OF ILU DECOMPOSITION The incomplete LU factorization known as ILU overcomes the problem of complete LU factorization. It computes and stores the main values of L and U matrices from the factorization of large matrix size. Since the ILU gives an approximation to the matrix, it yields in an iterative method. The incomplete factorizations are generalized in the concept of ILU and are stored in the band matrix format or the set of diagonals. Furthermore, a threshold can also be introduced to achieve a significant reduction of coefficients in cases with a wide band width. These versions lead to different degrees of sparsity, approximation and finally in more or less efficient iterative algorithms. The ILU factorization can be further divided into two streams i.e. for symmetric matrices and non-symmetric matrices. There are different methods to solve the different streams of matrices. Cholesky factorization is used to solve symmetric matrices. MATHEMATICAL DERIVATION: Consider a linear system Ax b) Where A = Sparse Banded NxN non singular. As A is not compactly banded, at least O(N2) arithmetic operation would be required by elimination procedure. In Guass elimination procedure O(N3) operation would be needed. Replace A with M nearly as sparse as A but will admit a sparse LU Decomposition. We set

M A E(2) E Errormatrix (M E ) x b Mx Ex b The fixed point form is Mx ( n 1) Ex ( n ) b(3)

M has a LU decomposition, we can express the above as LUx ( n 1) Ex ( n ) b x ( n 1) U 1 L1 Ex ( n ) U 1 L1b(4) This is in our standard linear xed point form x ( n 1) Gx ( n ) k (5) where G U 1 L1 E , k U 1 L1b

The preconditioner is then M = XY. Note that a linear solve involving M reduces to a triangular solve involving Y followed by one with X. If A is sparse, then so are X and Y and the solution of the associated systems can be performed very quickly.

EXAMPLE: INCOMPLETE LU FACTORIZATION


A= 2 0 4 2 X= 1 0 2 1 Y= 2 0 0 0 0 1 1 0 0 -1 0 0 0 3 0 5 0 1 0 0 0 0 1 0 0 0 0 1 0 1 0 0 1 0 1 0 0 3 0 5

EXAMPLE: CHOLESKY FACTORIZATION


A= 4 2 14 2 17 -5 14 -5 83 L= 2 0 1 4 7 -3 0 0 5

ALGORITHM:
close all clear all clc fprintf('\n\tIncomplete LU Factorization\n\n\n') A = [2 0 4 2 0 1 0 0 1 0 1 0 0; 3; 0; 5]

n=4; for i = 2:n for k=1:i-1 if A(i,k) ~= 0 A(i,k)= A(i,k)/A(k,k); for j=k+1:n if A(i,j) ~= 0 A(i,j) = A(i,j) - A(i,k)*A(k,j); end end end end end X = zeros(size(A)); Y = X; for i=1:n for j=1:i-1 X(i,j)= A(i,j); end X(i,i) = 1; for j=i:n Y(i,j) = A(i,j); end end X Y fprintf('\n\n\n\n') fprintf('\tCholesky factorization\n\n\n') A = [ 4 2 14; 2 17 -5; 14 -5 83]

n = 3; L = zeros(size(A)); L(1,1) = sqrt(A(1,1)); for j=2:n L(j,1) = A(j,1)/L(1,1); end for j=2:n sum = 0; for s=1:j-1 sum = sum + L(j,s)^2; end L(j,j) = sqrt(A(j,j)- sum); for p=j+1:n sum = 0; for s=1:j-1 sum = sum + L(j,s)*L(p,s); end L(p,j) = 1/L(j,j)*(A(p,j) - sum); end end L

CONCLUSION: Iterative methods are really the only available options to solve large scale problems due to their computational and memory demands. They are even more attractive if a good preconditioner is used. The strategies are quite promising and effective in helping reduce the number of iterations, and a noticeable improvement in enhancing the accuracy of the incomplete LU factorizations. Furthermore, the total computation time for solving sparse linear system is reduced even though extra processing time is used for the proposed strategies. Sparse matrices, symmetric or non-symmetric, are stored in band matrix formats; A, L, U. It is better to reorder the band width and definitely to save in storage requirements and computations. A well known cholesky method is used for symmetric matrices where only the half part of the banded matrix is needed for the factorization and storage. An ILU factorization is obtained from an approximate Gaussian elimination procedure. When Gaussian elimination is applied to a sparse matrix A, a large number of nonzero elements may appear in locations originally occupied by zero elements. These fill-ins are often small elements and may be dropped to obtain Incomplete LU factorizations. Thus, ILU is in essence a Gaussian elimination procedure in which fill-ins are dropped. ILU is fairly robust and cost effective general purpose preconditioner. Other preconditioners using the same principle for dropping can also be effective. One of the most

unappealing features of ILU-type preconditioners is that they are not easy to update. If some entries in A change slightly, then the whole LU factorization should, in principle, be recomputed.

REFERENCES: 1) Numerical Solution of Partial Differential Equations in Science and Engineering by Leon Lapidus, George F. Pinder 2) Numerical Analysis of Partial Differential Equations by S.H.Lui 3) Numerical Methods for Elliptic and Parabolic PDEs - P. Knabner & L. Angermann 4) Computational Numerical Analysis of Partial Differential Equations by J. M.McDonough 5) Incomplete LU Preconditioning and Error Compensation Strategies for Sparse Matrices by Eun-Joo Lee 6) A class of incomplete orthogonal factorization methods. I: methods and theories by Zhong-Zhi Bai

Das könnte Ihnen auch gefallen