Michael B Woodroofe
Statistics, University of Michigan
,
where 
is the residual sum of squares, D is a
surrogate for the degrees of freedom, and 
. The
derivation uses Stein's unbiased estimator of the risk to suggest the
form of the estimator and asymptotic analysis to determine the
constant c. As a corollary, it is shown that the maximum
likelihood estimator attains the optimal rate of convergence, without
imposing any smoothness conditions on the regression function.
Moderate sample size performance of the estimator is assessed through
simulation.