Convergence of Runge-Kutta methods for differential-algebraic systems of index 3 (Q1893542)

Index 3 differential-algebraic equations (DAEs) are notoriously difficult to solve directly and it is a common practice to resort to index reduction. However, in the modeling of many mechanical systems, this is not an appropriate approach, and it is necessary to retain the index 3 formulation. This paper considers the rate of convergence of solutions obtained by the use of implicit Runge-Kutta methods applied to semi- explicit index-3 DAEs in Hessenberg form, \(y' = f(y,z)\), \(z' = k(y, z, u)\), \(0 = g(y)\), where \(g_y f_z k_u\) is invertible in the vicinity of the exact solution. In particular the paper examines the order enhancement that can be obtained by the application of a projection to the \(z\) component, with a corresponding adjustment to the \(u\) component. Denote by \(B(p)\) (\(p\) is the quadrature order), \(C(q)\) (\(q\) is the stage order) and \(D(r)\) the usual simplifying assumptions, so that these hold with \(p =2s - 2\), \(q = r = s - 1\) in the case of Lobatto IIIC methods and with \(p =2s -1\), \(q = s\), \(r = s - 1\) in the case of Radau IIA methods. After a number of preliminary results preparing the ground, the principal theorem and its corollaries show that under certain assumptions on the initial value problem and on the coefficients of the Runge-Kutta method, the global error in the \(y\) component has order \(\min(p,2q - 2, q + r)\), with a slightly improved estimate if \(k\) is linear in \(u\). Moreover, although the global errors in \(z\) and \(u\) are limited to \(q\) and \(q - 1\) respectively, the use of a projection raises these to the same order as for the \(y\) component. The results are supported by numerical experiments in a variable stepsize setting for both a linear (\(k\) linear in \(u\)) problem and a nonlinear problem, using Lobatto IIIC and Radau IIA methods of various classical orders.