Error estimation for numerical differential equations

Given a capable human being and a computer, it is possible to make an approximation to the solution of a nonlinear differential equation. However, under the (usually correct) assumption that the equation is analytically intractable, the result of the computation is not the exact solution; indeed it may be so far from the exact solution as to be completely useless. We are interested in the relationship between the effort expended by the human and the computer, and the duality of the computed approximation to a partial or ordinary differential equation. To be specific, we would like to think in terms of a cost-benefit analysis. The cost of the computation is a combination of the human effort and computer resources used to obtain the approximation. The benefit includes, of course, the computed approximation, but it also includes an estimate of the quality of the approximation, that is, an error estimate. It is our opinion that in computational science, as with the experimental sciences, results should always be presented with some estimate of their accuracy. In addition, however, there is another facet to error estimation: one cannot even attempt a cost-benefit analysis or efficiency comparison of methods without an error estimate to evaluate the results