Thanks to Dragan Stoikovitch, an improved version of the paper
at http://www.cs.berkeley.edu/~fateman/papers/horner.pdf
is now posted. The principal point of the paper remains: that reformulating
some polynomials to evaluate using Horner's rule can result in
less accurate results than other methods.
Specifically this paper investigates Legendre polynomials of the first kind,
and uses Maxima. The typical defining recurrence is far more accurate
especially
at roots of the polynomial or near 1.
RJF