Partitioned adaptive Runge-Kutta methods for the solution of nonstiff and stiff systems (Q1109532)

The paper deals with adaptive Runge-Kutta (ARK) methods introduced by the second and third author [Computing 29, 153-165 (1982; Zbl 0483.65043)] for solving stiff initial value problems for ordinary differential equations. Roughly speaking these schemes are based on a local linearization \(y'=Ty+r(t,y)\) of \(y'=f(t,y)\). The needed matrix exponentials are approximated by rational functions. For a detailed description see loc. cit. The theory of Butcher series is modified in order to derive the order conditions. A complete characterization of translation invariant ARK-methods is given. An ARK-method is called translation invariant if the stage increments to compute \(y_{n+1}\) do not depend explicitly on the old value \(y_ n.\) A theorem of the first author [Numerische Lösung steifer und nicht steifer Differentialgleichungssysteme mit partitionierten adaptiven Runge-Kutta-Methoden (Dissertation) (1985; Zbl 0603.65048)] is quoted. It basically says that to any explicit s-stage Runge-Kutta method one can construct a translation invariant ARK-method of the same order. For order 3 an example is given. In the present paper partitioned ARK-methods do not decompose a system of ordinary differential equations in stiff and nonstiff components. What is done is that in each time step one uses either an ARK-method (for stiff systems) or the corresponding explicit RK-method (for nonstiff systems). Hence stiffness detection has to be done. A program PAI4 with stiffness detection and stepsize selection is described and extensive comparisons with other codes on test sets of stiff and nonstiff differential equations are presented.