Positive definiteness aspects of vectorizable preconditioners (Q583846)

Two main classes of incomplete block factorization methods for solving discrete elliptic problems are considered. One of them is more suitable for vectorization, but loss of positive definiteness of the preconditioner, and failure of the conjugate gradient iteration might occur even for M-matrices. In order to improve the chances that this will not occur, certain row sum criteria are often used. It is shown in this paper, both theoretically and experimentally, that this does not solve the underlying problem. In fact, for modestly difficult test problems, the methods considered do not perform very satisfactorily at all.