ROWMAP -- a ROW-code with Krylov techniques for large stiff ODEs (Q1372711)

This is a highly specialized paper devoted to relatively new methods of numerical approximation of solutions of initial value problems for ordinary differential equations (ODEs): \[ y'(t)=f(t,y(t)),\qquad y(t_0)=y_0. \] Methods of this kind are efficient mainly in the case of very large stiff systems. Such systems occur when, for example, the so-called method of lines is applied to approximation of certain problems of partial differential equations. The authors propose to incorporate a Krylov subspace technique into the so-called Rosenbrock-Wanner (ROW) method. This technique consists in replacing, in consecutive stages of the algorithm, the true Jacobian \(A\) of the function \(f\), by its lower rank approximation of the form \(Q_iQ_i^TA\). Here \(Q_i\) is, in general, a rectangular orthogonal matrix. Its columns are orthogonal vectors of the basis of the Krylov subspace. The matrix \(Q_i\) has to be computed by means of a certain modified Arnoldi orthogonalization process described in the paper. This technique results in some simplification of the algorithm, namely, in reducing of dimension of the linear systems to be solved. Since the systems are, generally, of very large dimension, this simplification may be important. The authors claim that the application of the proposed technique makes it possible to avoid storing of the full Jacobian, exploits the sparsity of the matrix and does not require any special structure of the matrix. In few numerical examples different methods are compared with ROWMAP.