A number-of-modes reference rule for density estimation under multimodality.

Abstract

We consider kernel density estimation for univariate distributions. The question of interest is as follows: given that the data analyst has some background knowledge on the modality of the data (for instance, ‘data of this type are usually bimodal’), what is the adequate bandwidth to choose? We answer this question by extending Silverman's idea of ‘normal-reference’ to that of ‘reference to a Gaussian mixture’. The concept is illustrated in the light of real data examples.