Iterative schemes for three-stage implicit Runge-Kutta methods (Q1901533)

The evaluation of the stages in a fully-implicit Runge-Kutta method, applied to a large stiff problem, is notoriously difficult and computationally expensive. If a full modified Newton scheme is used, then linear systems involving the matrix \(I_{sm}- h A\otimes J\) need to be solved in each iteration, where \(A\) is the coefficient matrix for the method, assumed to have \(s\) stages, and \(J\) is an approximation to the Jacobian matrix of the problem, assumed to be \(m\)-dimensional. To reduce the costs, without losing rapid convergence, alternative schemes have been considered which require the solution of linear systems involving only a matrix of the form \(I_m- h \gamma J\). The present paper is based on the scheme of \textit{G. J. Cooper} and \textit{J. C. Butcher} [IMA J. Numer. Anal. 3, 127-140 (1983; Zbl 0525.65052)], which computes iterate number \(k\) by \(Y^k= Y^{k-1}+ (S\otimes I_m)E^k\), where \((I_{sm}- h \gamma I_s \otimes J)E^k= (BS^{-1} \otimes I_m)D^{k-1} +(L\otimes I_m)E^k\) and \(D^k= e\otimes y_n- Y^k+h(A\otimes I_m)F(Y^k)\). Note that \(F\) denotes the function defining the differential equation but applied stage by stage, and \(y_n\) is the inital value for the step of integration being evaluated. In a recent paper of \textit{S. González-Pinto, C. González-Concepción} and \textit{J. I. Montijano} [Comput. Math. Appl. 27, No. 7, 67-81 (1994; Zbl 0792.65055)] the assumption is made \(S\) is regular and that \(L=I-B\) is strictly lower triangular. This has the effect of ensuring that the order of successive iterates increases until the order of the method is reached and that the linear convergence matrix \(M(z)\) satisfies \(M(0)=0\). The present paper considers schemes for three-stage methods which are convergent for a nonlinear model problem and which satisfy \(\rho(M(\infty))=0\). Subject to these requirements optimal schemes are found for the Gauss and Radau IIA methods. The schemes perform well compared with existing iteration methods for a variety of test problems.