A new preconditioning technique for solving large sparse linear systems (Q808620)

A new approximate LDU factorization technique is described. The triangular matrices L and U are given as a product of elementary triangular matrices with units on the diagonal and nonzero off-diagonal entries in only one column or row. These elementary matrices perform a successive transformation of the factorized matrix. The diagonal entries of D appear during the transformation steps. An obvious advantage of the new technique is the fact that, for symmetric positive definite matrices, it gives the factorization preserving that property. To reduce and control the sparsity of the described approximate factorization, several procedures are suggested: the choice of fill-in by value, deleting some off-diagonal entries accompanied by increasing of the corresponding diagonal entries and reordering by permutations of rows and columns. Numerical experiments help to assess the real efficiency of the conjugate gradient method with preconditioning given by the described factorization and provide some favourable comparison with the ICCG(0) algorithm.