This article describes a multiple-bandwidth version of the kernel estimator for nonparametric probability density estimation, in which the bandwidths are chosen using a set of functions, called filter functions, which determine the support of the density appropriate to the different bandwidths. These filter functions are usually defined using a normal mixture fit to the data. Thus the estimator uses different bandwidths in different regions of the support of the distribution, as controlled by the filter functions. Copyright © 2009 John Wiley & Sons, Inc.
For further resources related to this article, please visit the WIREs website.