Univariate density estimation by orthogonal series

Abstract

SUMMARY Orthogonal series estimators of univariate densities are proposed that are derived from and motivated by kernel estimators optimal in Whittle's (1958) sense. A preliminary fit of the data from within a one or two parameter class of densities plays the role of a prior mean density. The ratio of the true density and the prior mean density is assumed to have a series expansion in terms of functions orthogonal with respect to the prior mean density. Coefficients of terms of the series are given a joint prior distribution according to which they are independent, with zero means.