Avoiding the exactness of the Jacobian matrix in Rosenbrock formulae (Q919758)

Generalizations of Rosenbrock methods have been introduced by Steihaug and Wolfbrandt, by Kaps and Wanner and by others. Methods of this more general type are usually known as ROW methods and are motivated by one or other of two aims. The first is that of achieving high orders with fewer stages than is possible with standard Rosenbrock method. The second aim is to allow the use of approximate Jacobians without sacrificing order. The present paper develops the latter idea with the assumption that the error in the Jacobian is the stepsize multiplied by some matrix. Additional order conditions are introduced to allow for the effect of this perturbation. Conditions for A-stability are presented for 1 and 2 stage versions of the method. Methods satisfying the new conditions with orders 3 and 4, together in each case with an error estimator one order higher, are derived. From various numerical experiments the author has performed, two are reported in detail. Each of these involves a model stiff problem and each is solved using a third order code.