Long-term stability of multi-value methods for ordinary differential equations (Q474963)

The authors consider initial value problems for systems of differential equations \[ \dot y=f(y),\quad y(0)=y_0. \] The long-time behavior of multi-value methods, i.e., general linear methods consisting of a forward step procedure \[ Y_{n+1}=V\,Y_n + h \Phi(h, Y_n), \] a starting procedure \[ Y_0=S_h(y_0), \] and a finishing procedure \[ y_n=F_h(Y_n), \] is studied. First, a backward error analysis of multi-value methods is performed by deriving modified equations for the components associated with the expansions of approximations to \(Y_n\). In this context, the role of growth parameters and the effect of symmetry on the modified equations are also discussed. It is known that for irreducible linear multistep methods the growth parameters cannot be zero. However, general linear methods can be constructed for which the growth parameters corresponding to parasitic roots are all zero. This interesting property greatly affects the long-term energy conservation in the numerical solution of Hamiltonian differential systems. Next, as the main results, sharp estimates for the parasitic solution components are obtained and the long-time behavior of multi-value methods applied to Hamiltonian systems is analysed. Finally, numerical experiments for the Hamiltonian system of the mathematical pendulum are given which confirm the sharpness of the theoretically obtained estimates.