Tuning selection for two-scale kernel density estimators

Abstract

Reducing the bias of kernel density estimators has been a classical topic in nonparametric statistics. Schucany and Sommers (1977) proposed a two-scale estimator which cancelled the lower order bias by subtracting an additional kernel density estimator with a different scale of bandwidth. Different from existing literatures that treat the scale parameter in the two-scale estimator as a static global parameter, in this paper we consider an adaptive scale (i.e., dependent on the data point) so that the theoretical mean squared error can be further reduced. Practically, both the bandwidth and the scale parameter would require tuning, using for example, cross validation. By minimizing the point-wise mean squared error, we derive an approximate equation for the optimal scale parameter, and correspondingly propose to determine the scale parameter by solving an estimated equation. As a result, the only parameter that requires tuning using cross validation is the bandwidth. Point-wise consistency of the proposed estimator for the optimal scale is established with further discussions. The promising performance of the two-scale estimator based on the adaptive variable scale is illustrated via numerical studies on density functions with different shapes.