Curve estimation for \(m_ n\)-decomposable time series including bilinear processes (Q1176296)

A sequence of identically distributed random elements is \(m_ n\)- decomposable if it can be written as a sum of two sequences, the first one \(m_ n\)-dependent and the second one converging to zero in probability. The authors consider the following situation: Let \(\{Z_ i\}\) be a sequence of random vectors and for some functions \(\psi\), real valued, and \(\Psi\), \(R^ d\)-valued, define \((X_ i,Y_{\psi,i})=(\Psi(Z_ i)\), \(\psi(Z_ i))\) and suppose that \((X_ i,Y_{\psi,i})\) is \(m_ n\)-decomposable. Define \[ F_ \psi=E\psi(Z_ i)\prod^ d_{j=1}1_{(-\infty,x_ j]}(\Psi_ j(Z_ j)), \] a distribution function of a signed measure and \({\hat F}_{\psi,n}\) its empirical estimator. The authors study the \(\psi\)- empirical processes \(U_{\psi,n}=n^{1/2}(\hat F_{\psi,n}-F_ \psi)\) and they obtain an exponential inequality for its local fluctuations. This is then used to show the uniform consistency of the nonparametric curve estimators for the autoregression functions in linear and bilinear models.