Previous: The idea behind block factorizations
Up: Block factorization methods
Next: The special case of block tridiagonality
Previous Page: The idea behind block factorizations
Next Page: The special case of block tridiagonality

Approximate inverses

In block factorizations a pivot block is generally forced to be sparse, typically of banded form, and that we need an approximation to its inverse that has a similar structure. Furthermore, this approximation should be easily computable, so we rule out the option of calculating the full inverse and taking a banded part of it.

The simplest approximation to is the diagonal matrix of the reciprocals of the diagonal of : .

Other possibilities were considered by Axelsson and Eijkhout [15], Axelsson and Polman [20], and Concus, Golub and Meurant [55].

Banded approximations to the inverse of banded matrices have a theoretical justification. In the context of partial differential equations the diagonal blocks of the coefficient matrix are usually strongly diagonally dominant. For such matrices, the elements of the inverse have a size that is exponentially decreasing in their distance from the main diagonal. See Demko, Moss and Smith [62] for a general proof, and Eijkhout and Polman [86] for a more detailed analysis in the -matrix case.