The dichotomy of stiffness: Pragmatism versus theory (Q1119343)

This is an excellent survey of the recent work on the numerical solution of initial-value problems for stiff ordinary differential equations. The author discusses and compares the performance of various codes for such problems based on linear multistep methods (backward differentiation formulas), cyclic multistep methods, and many classes of Runge-Kutta methods such as diagonally implicit and singly implicit methods. Numerical testing reported in the literature reveals that codes based on the backward differentiation formulas perform well on most stiff problems except for those with Jacobians whose eigenvalues have large imaginary parts. On the other hand, various codes based on Runge-Kutta methods perform well on such problems only when low precision is required. The author suggests that truly efficient and highly stable methods which are appropriate for the numerical integration of stiff systems may be found in the class of general linear methods introduced by Butcher, and he encourages research in this field.