scientific article; zbMATH DE number 4020220

From MaRDI portal
Publication:3763408





[For the entire collection see Zbl 0625.00017.] All truly nonparametric density estimators suffer from embarrassingly poor error properties in higher dimensions. Even with optimal choices, the behavior of an error measure like mean-integrated-squared-error is \(O(n^{-4/(d+4)})\) where d is the dimension. Enormous data sets and fantastic computational power would be required to obtain reasonable density estimates even for modest values of d. This situation stands in marked contrast to such standard statistical techniques as regression, which has good error properties essentially independent of the dimension and which is applied routinely to high-dimensional data. The key to the success of regression is its skillful exploitation of presumed linear relationships in the data. In density estimation, the statistician's presumed knowledge of something of the structure of the density may prove similarly exploitable. In fact, it is shown that if the isopleths (or contours) of the density can be specified or even modelled approximately, then the high-dimensional density estimation problem can be collapsed into a one-dimensional density estimation problem, with the error properties of the latter.











This page was built for publication:

Report a bug (only for logged in users!)Click here to report a bug for this page (MaRDI item Q3763408)