Kernel density estimation for linear processes (Q5905553)

Let \(X_ 1,\dots,X_ n\) be \(n\) consecutive observations of a linear process, \[ X_ t=\mu+\sum_{r=0}^ \infty a_ r Z_{t-r}, \] where \(\mu\) is a constant and \(\{Z_ t\}\) is an innovation process consisting of independent and identically distributed random variables with mean zero and finite variance. It is assumed that \(X_ 1\) has a probability density \(f\). It is proven that the uniform (over a compact) distance between \(f\) and its kernel estimator \(f_ n\) can achieve the rate \((n^{-1}\log n)^{1/3}\). The assumptions involve the characteristic function of \(Z_ 1\) and the sequence \(\{a_ r\}\). In particular, for first order autoregressive processes the rate of convergence is \(((nb_ n)^{- 1}\log n)^{1/2}\) under some weak conditions if the sequence of bandwidths \(\{b_ n\}\) tends to zero slowly enough that \(nb_ n(\log n)^{-3}\to\infty\).