Splitting methods with complex times for parabolic equations (Q1035797)

Splitting methods constitute useful procedures in the numerical integration of differential equations whose associated vector field \(f\) can be decomposed into several pieces, \(f = \sum_i f_i\), and each of them is explicitly integrable. The idea is to construct an approximation of the exact solution of the differential equation for a time step \(h\) as a composition of the flows corresponding to the vector fields \(f_i\) evaluated at different time steps \(\alpha_j h\), where \(\alpha_j\) are conveniently chosen parameters to achieve a given order of approximation. In close connection with splitting schemes are the so-called composition methods. In this class of integrators, the approximation is constructed as a composition of a basic method successively used with different time steps \(\gamma_k h\). Again, the parameters \(\gamma_k\) are selected to verify certain order conditions. A well known result in splitting and composition methods states that schemes of order greater than two necessarily involve negative coefficients if \(\alpha_j\) (or \(\gamma_j\)) are taken as real numbers. This feature constitutes a fundamental difficulty when the differential equation is defined in a semigroup, as occurs in particular, with the (linear or nonlinear) heat equation. In general, if the differential equation involves operators with large negative spectrum the use of splitting methods is problematic. One way to circumvent this order barrier is to try to develop methods with complex (instead of real) coefficients \(\alpha_j\) (or \(\gamma_j\)) with positive real part. In the paper under review, standard composition techniques are used to build high order splitting methods with these features. In particular, methods with \(\text{Re}(\alpha_j) > 0\) up to order 14 are constructed which are valid for differential equations defined in finite but also in infinite dimensional spaces. To do that, a convergence analysis for unbounded operators in this last setting is provided. Numerical examples illustrate the theoretical analysis on the order of approximation of the schemes. Results along this line have also been derived by \textit{E. Hansen} and \textit{A. Ostermann} [BIT 49, No. 3, 527--542 (2009; Zbl 1176.65066)].