Linear equations, iterative solutions

From HandWiki


For certain types of systems of linear equations Ax = b methods like Gaussian elimination can become inefficient, e.g. if A is sparse and large. In such cases, iterative methods are preferable. They converge if certain conditions are fulfilled, e.g. if A is diagonally dominant (see Golub89):

Hepa img602.gif

In this case, Ax = b can be rewritten in the form

Hepa img603.gif

where each line solves separately for the x appearing with the diagonal element of A. Any iterative scheme needs an initial guess x(0), whose quality determines the possibility or the speed of convergence. We obtain the (k+1)st iteration xk+1 if we substitute the kth iteration xk into the right hand side. If we compute all the new values on the left side with all the old values on the right side we obtain the Jacobi iteration :

Hepa img604.gif

If we successively use new values of xi as soon as they are computed, we get the Gauss-Seidel iteration :

Hepa img605.gif

A variant of this algorithm is the method of Successive Over-Relaxation:

Hepa img606.gif

where the over-relaxation parameter Hepa img534.gif satisfies Hepa img607.gif . For how to determine Hepa img534.gif , see Golub89 or Young71.