Convergence of general linear methods on differential-algebraic systems of index 3 (Q1359976)

The author considers systems of differential algebraic equations (DAEs) of the form \[ y'=f(y,z), \qquad z'=v(y,z,u), \qquad 0=g(y) \tag{1} \] with sufficiently smooth functions \(f:\mathbb{R}^m\times \mathbb{R}^t\to\mathbb{R}^m\), \(v:\mathbb{R}^m\times \mathbb{R}^t\times \mathbb{R}^r\to\mathbb{R}^t\), \(g:\mathbb{R}^m\to\mathbb{R}^r\). With the assumption \[ |(g_y(y) f_z(y,z) v_u(y,z,u))^{-1}|\leq M\tag{2} \] in a neighbourhood of the exact solution problem (1) is a so called index 3 problem. Initial values \(y_0\), \(z_0\), \(u_0\) are given in a ``consistent'' way. It is well known that many numerical methods used to solve ordinary differential equations or DAEs can be written as general linear methods. The purpose of the paper is to extend known convergence results for Runge-Kutta and linear multistep methods to a large class of new promising numerical schemes. Especially a multistep three-stage Radau method with excellent stability properties implemented in the Doctoral thesis of the author gives promising results. The author extends his own generalizations of convergence and stability results for Runge-Kutta methods and applies these results to the mentioned multistep Radau and to backward differentiation formula (BDF) methods. Finally he compares on some test problems the performance of three codes using respectively BDF, one-step Radau IIa and multistep Radau methods. One observes that for low accuracy the three codes have more or less the same performance, but for high precision the last one achieves a significant computing time and work gain compared to the other codes.