Optimal kernel estimation of densities (Q2640276)

Let \(X_ 1,X_ 2,...,X_ n\) be an i.i.d. sequence from some unknown density f. For a given kernel \(K_ n\), set \[ \hat f_ n(x)=n^{- 1}\sum^{n}_{j=1}K_ n(x-X_ j). \] As a measure of fit denote by \(MISE_{K_ n}\) the mean integrated squared error between f and \(\hat f_ n\) when choosing the kernel \(K_ n\). Set \(J_ n=\inf_{K_ n} MISE_{K_ n}\). Finally, let \(M_{n,K}\) denote the infimum of \(MISE_{K_ n}\) with \(K_ n(x)=h_ n^{-1}K(x/h_ n)\) generated by some fixed K and a positive bandwidth \(h_ n.\) The objective of this paper is to provide the limit of \(J_ n/M_{n,K}\) as \(n\to \infty\) under certain regularity conditions on the Fourier transform of f.