Experimental study of ILU preconditioners for indefinite matrices (Q1379014)

Incomplete LU factorization preconditioners have been applied successfully for many cases of general nonsymmetric and indefinite matrices. However, for them to be useful for general matrices, there are still numerical difficulties including very small pivots, unstable triangular solvers, inaccuracy due to dropping, and zero pivots. Furthermore, these four problems usually occur together. So, there is a need to gain a better practical understanding of LU preconditioners and help improve their reliability. This paper shows how these problems evince themselves, how these problems can be detected, and how these problems can sometimes be circumvented through pivoting, scaling, reordering, perturbing diagonal elements, and preserving symmetric structure. Numerical experiments demonstrate that the proposed techniques are efficient.