Analysis of a defect correction method for geometric integrators (Q2492795)

This paper is concerned with the application of the defect correction technique to estimate the global error in the numerical solution of initial value problems for ordinary differential equations when the basic numerical method is a geometric integrator. In the standard defect correction the same basic method is applied to the main problem: (1) \( y' = f(t,y), y(t_0)=y_0,\) \(t \in [t_0,t_0+T],\) and to the neighbouring problem: (2) \( z' = f(t,z)+d^{(0)}(t)\), \(z(t_0)=y_0\), \(t \in [t_0,t_0+T]\), where \(d^{(0)}(t)\) is the defect of an interpolation polynomial of the discrete solution of (1). In the case that the basic method is a geometric integrator, since the neighbouring problem (2) does not share, in general, the same structure than (1) we may encounter some difficulties in using the same numerical method to integrate (2). In view of this fact, the authors propose a splitting of the vector field of (2) that uses the same basic method for \(f\) together with a quadrature to the defect term \(d^{(0)}(t)\). In this context, different splittings and quadratures are studied, and after a complicated analysis, the resulting orders attained by successive iterations of this defect correction are determined. Finally to test the orders predicted by the theory, the authors present the results of some numerical experiments with the plane Kepler problem with eccentricity \(0.6\) along one orbit. This problem is integrated with the Störmer-Verlet method and the iterated defect correction uses Yoshida's splitting with interpolating polynomials based either on Gaussian or else on arbitrary nodes.