Avoiding order reduction of fractional step Runge-Kutta discretizations for linear time dependent coefficient parabolic problems. (Q1426335)

Order reduction is a phenomenon of loss of accuracy that appears in many classical methods for example Runge-Kutta (RK) or Rosenbrock methods when they are used to discretize parabolic problems with time dependent boundary conditions. It can also take place in most of the splitting alternating direction or fractional step methods for solving evolutionary problems with time dependent boundary conditions. The aim of the present paper is to investigate how to avoid this problem. The main idea is based on modifying suitably the boundary conditions for the internal stages in order to attenuate the loss of order. This idea was succesfully developed for RK methods of any order and also for fractional RK methods as well. Such time integrators combined with a standard discretization of the spatial variables involve a class of numerical algorithms which generalize the classical alternating direction, fractional step or splitting methods. A generalization of this technique for parabolic problems with time dependent coefficients is investigated in this article. The problem is of the type: \[ {du(t)\over dt} = A(t) u(t) + f(t), \quad t_0 \leq t \leq T \] \[ u(t_0)=u_0 \in H, \quad \partial u(t) =g(t) \in H_b, \quad t_0 \leq t \leq T, \] where \(H\) and \(H_b\) are two Hilbert spaces. A general convergence theorem for the numerical algorithms that combines these semidiscrete problems and a standard spatial discretization method which admits the abstract formulation and preserves suitable consistency and stability properties is proved. Two numerical experiments are included. They confirm the increase of accuracy that the introduced technique provides.