Dynamical systems and adaptive timestepping in ODE solvers (Q1577721): Difference between revisions

The author studies the convergence behaviour of adaptive time step one-step solvers for initial value problems in ordinary differential equations (ODEs). It is widely recognized that practical ODE solvers must vary the step size in order to take at each step the largest step size compatible with a prescribed tolerance. In practice this is accomplished by giving a user defined tolerance \(\tau \), and a local error estimate so that the local error at each step is close but below the tolerance. Clearly a basic requirement on the discretization method is that, in some sense, the numerical solution tend to the exact solution when \( \tau \to 0\). This is the main subject of the paper, and to study this problem the author formulates adaptive algorithms as discrete dynamical systems of higher dimension to include the variable step size technique. Then under sufficiently strong conditions he proves that when the integration is carried out in a neighborhood of an attractor of the differential system some convergence and upper semicontinuity results hold as the tolerance \( \tau \) tends to 0. In the second part of the paper, for adaptive time step algorithms that use a pair of explicit Runge-Kutta methods of different orders as local error estimator, the convergence properties are studied. He proves that convergence in finite intervals fails around the equilibrium points of the original differential system. However, for embedded Runge-Kutta pairs of consecutive orders in which the local error estimate has a prescribed form and for equilibrium points where the Jacobian is invertible the author proves that the desirable convergence of the adaptive algorithms also holds.