Rate of convergence of a convolution-type estimator of the marginal density of a MA(1) process (Q1593614): Difference between revisions

Let \(z_t = a_t - \theta a_{t-1} \) be a first order moving-average process, where \(|\theta |< 1\) and \(a_t \) are i.i.d. random variables with density \(g.\) Let \(f\) be the density of \(z_t.\) The problem of estimation of the density \(f\) is studied. A new convolution-type kernel estimator with bandwidth \(h\) is proposed, defined by \[ \hat f_{\hat \theta}(x)=(n^2h)^{-1}\sum_{i,j=1}^n L_{\hat \theta}\left ((x+\hat \theta \hat a_i - \hat a_j)/h\right), \] where \[ L_{\hat \theta}(t) = \int K(u)K(t+\hat \theta u)du. \] Here \(K\) is a kernel function, \(\hat \theta \) is a \(\sqrt n\)-consistent estimator of \(\theta \) and \(\hat a_t \) are estimated errors. The rate of convergence of this estimator to the true value is studied and its \(\sqrt n\)-consistency is established for the second order kernels.