!- Converted with LaTeX2HTML 0.6.4 (Tues Aug 30 1994) by Nikos Drakos (firstname.lastname@example.org), CBLU, University of Leeds ->
One reason that block methods are of interest is that they are potentially more suitable for vector computers and parallel architectures. Consider the block factorization
where is the block diagonal matrix of pivot blocks, and , are the block lower and upper triangle of the factorization; they coincide with , in the case of a block tridiagonal matrix.
We can turn this into an incomplete factorization by replacing the block diagonal matrix of pivots by the block diagonal matrix of incomplete factorization pivots , giving
For factorizations of this type (which covers all methods in Concus, Golub and Meurant  and Kolotilina and Yeremin ) solving a linear system means solving smaller systems with the matrices.
Alternatively, we can replace by a the inverse of the block diagonal matrix of the approximations to the inverses of the pivots, , giving
For this second type (which was discussed by Meurant , Axelsson and Polman  and Axelsson and Eijkhout ) solving a system with entails multiplying by the blocks. Therefore, the second type has a much higher potential for vectorizability. Unfortunately, such a factorization is theoretically more troublesome; see the above references or Eijkhout and Vassilevski .