On a guaranteed estimation of autoregressive parameters for an unknown variance of noise (Q1914113)

The authors provide a framework for parameter estimation of autoregressive processes. They propose a class of sequential procedures which give prescribed mean square accuracy of the estimators. The idea is to estimate the variance from a fixed number of observations and then use a stopping rule based on it, to estimate the parameters. The stopping time of these procedures is finite almost surely. The paper extends previous work of \textit{V. V. Konev} [Sequential estimates for parameters of stochastic dynamical systems (in Russian) (Izd. Tomskogo Un-ta, Tomsk) (1985)] to the case of unknown innovations variance. The first order autoregression model is studied in more detail. In that case the asymptotic behaviour of the estimation time and its mean are determined. Furthermore, it is shown that the estimator of the autoregressive parameter is unbiased (without any restriction on its range) and is uniformly asymptotically normal in \([-1,1]\). The last property is a sound manifestation of one of the advantages of sequential plans over fixed sample ones, where the asymptotic normality is not uniform, and at the ends of the interval \([-1,1]\), the asymptotic distribution is not normal at all.