Averaged shifted histograms: Effective nonparametric density estimators in several dimensions (Q1074267)

Let \(X_ 1,X_ 2,..\). be i.i.d. from some unknown density f in \({\mathbb{R}}\). For a given bin-width \(h>0\) and an integer \(m\geq 1\) let, for \(0\leq i\leq m-1\), \(\hat a_ i\) be the histogram for the grid \(rh+ih/m\), \(r\in {\mathbb{Z}}.\) The author proposes \(\hat f_ n\), the average of the \(\hat a_ i's\), as an estimator for f, and derives an expansion for the IMSE both as a function of h and m. It turns out that as \(m\to \infty\), \(\hat f_ n\) behaves like a kernel estimate with triangular kernel. Also the multivariate case is discussed.