The Amount of Noise Inherent in Bandwidth Selection for a Kernel Density Estimator.
Abstract
Any practical method of constructing a bandwidth must depend only on a statistical sample, and should produce some sort of estimate of this bandwidth. The purpose of this paper is to show that there a well-defined limits to the accuracy of all data-driven bandwidth estimates. Put another way, there is an unbridgeable gap between the minimum integrated square error attained using a optimal bandwidth and the minimum achievable integrated square error using a data-driven bandwidth estimate. Additional keywords: stochastic processes; cross validation; and random variables.
Document Details
- Document Type
- Technical Report
- Publication Date
- May 01, 1985
- Accession Number
- ADA160235
Entities
People
- J. S. Marron
- P. Hall
Organizations
- University of North Carolina at Chapel Hill