On the asymptotic normality of estimates in the nearly non-stationary AR(1) models (Q1381645): Difference between revisions

Given a family of AR(1) processes \( \{X_{n,k}, 0 \leq k \leq n\}_{n= 1}^{\infty} \) such that \( X_{n,0} = 0 \) and \( X_{n,k} = \beta_n X_{n,k-1} + \varepsilon_k, k = 1,\ldots, n \), set \( V_n = \left(\sum_{k = 1}^{n} X_{n,k-1}^2\right)^{-1} \sum_{k=1}^{n} \varepsilon_k X_{n,k-1} \) if \( \sum_{k=1}^{n} X_{n,k-1}^2 > 0 \). For various classes of noises \( \{\varepsilon_k, k \geq 1\} \), namely, for i.i.d. r.v.'s and for martingale-differences, under the specified moment conditions imposed on \(\varepsilon_k \) and assuming a certain rate of \(\beta_n\to 1\) \((n\to\infty)\) for the autoregressive parameter \(\beta_n\), the authors estimate \[ \Delta_n=\sup_{x}\left| {\mathbf P}\left(\sqrt{n/(1-\beta_n^2)}V_n<x, \sum_{k=1}^{n} X_{n,k-1}^2 > 0 \right)-\Phi(x) \right| \] where \(\Phi(x)\) is a distribution function of a standard normal law. In particular, if \(\varepsilon_k, k\geq 1\), are i.i.d. \({\mathcal N}(0,1)\) variables, then there exists an absolute constant \(C>0\) such that \(\Delta_n\leq C \delta_n^{-1/2} \log \delta_n\) for \(\delta_n=n(1-\beta_n^2)>1\).