Optimal numerical differentiation using n function evaluations (Q1061450)

An approximation to \(f^{(d)}(0)\) of the form \(D_ d(\vec w,\vec u,h)=h^{-d}\sum^{n}_{i=1}w_ if(u_ ih)\) is known as an n-point numerical differentiation rule of polynomial degree r when \(D_ d- f^{(d)}(0)=C_{r+1}h^{r+1}+O(h^{r+2})\) as \(h\to 0\) for all functions f analytic at the origin. The familiar finite difference approximation \(h^{-d}\Delta^ df\) is a d-point rule of degree d-1. To construct n-point rules of degree n-1 is a matter of solving linear equations. For higher degree, using only n points, nonlinear equations are involved. It is known that no n-point rules having real abscissas of degree \(r\geq n+1\) exist. The authors examine n-point rules of degree n. They consider only those having \(| u_ i-u_ j| \geq 1\) and define as optimal the rule of this class for which \(| C_{n+1}|\) is the least possible. For \(d=1\) and \(d=2\) they show that the optimum rules have \(u_{i+1}-u_ i=1\) for all relevant i; when n-d is odd this rule is symmetrically located. When n-d is even, the equispaced symmetrically located rule is not optimal but there is a pair of offset rules, each of which is optimal. The paper is well arranged. The results are stated clearly and discussed in the first 5 pages. The remaining 12 pages are devoted to long and tedious proofs, which do little to illuminate.