The standard techniques for stepsize control for Runge-Kutta methods, based
on embedding, typically use an error estimate which not only fails to be a
symptotically correct but is even of the wrong order. We will refer to this
sort of estimate as a "one-step zero approximation". Our aim in this paper
is to consider also m-step error approximations for m greater than or equa
l to 2. (C) 2000 IMACS. Published by Elsevier Science B.V. All rights reser
ved.