The mode-dispersion approach for constructing continuous associated kernels

Abstract

We introduce the mode-dispersion approach for constructing the (asymmetric) continuous associated kernels fromsuitable parametric probability density functions (p.d.f.) that we shall call the type of kernel. This leads us to value the choice of the associated kernel, since it takes into account the support of the unknown density f , to be estimated. All associated kernel density estimators must be without edge effect. For illustrating this, we introduce the extended beta kernel, which is a typical model of kernels with bounded supports. However, in the presence of a large bias of the density estimator, we propose a general but light modification in the same type of the first associated kernel; it leads to improve the mean integrated square error of the new estimator. Some properties of two estimators are investigated and compared, in particular pointwise and global (asymptotical) properties. Several forms of types of kernels and their associated kernel estimators are subsequently examined in detail. Simulation studies are made on three lognormal kernel density estimators for pointing out some behaviors at the boundaries.

Keywords: Cross-validation, dispersion parameter; free of boundary effect; unimodal kernel