Diagonalizable extended backward differentiation formulas (Q1587311)

This paper is concerned with the numerical solution of stiff initial value problems for differential systems by means of some modification of the so called extended backward differentiation formulas (EBDF) introduced by \textit{G.-Y. Psihoyios} and \textit{J. R. Cash} [BIT 38, No. 3, 612-617 (1998; Zbl 0915.65095)]. The EBDF formulas allow to overcome the so called Dahlquist barrier on the \(A\)-stability of linear multistep methods and in particular of standard backward differentiation formulas (BDF), by introducing some auxiliary implicit stages that allow to get \(A\)-stable methods of order higher than two that are also stiffly stable. Its main drawback is that the implicit equations to be solved at each step have a dimension that is several times the dimension of the original differential system and taking into account that, due to the stiffness, modified Newton processes must be used in the iterative solution of these equations the computational cost of new methods is higher than standard BDF's. New EBDF methods up to order 6, which are \(L\)-stable, are proposed. Such a new methods have been constructed taking into account its suitability for a parallel implementation. Furthermore, some numerical experiments with the method of order 6 are presented to show that, by using four processors, a speedup factor of between 2 and 4 is attained for several practical stiff problems.