Two limit theorems on ARIMA models (Q1118905)

The object of this paper is to study and discuss the limit a.s. and in distribution of \(\sum^{n}_{t=1}x(t)^ 2\), when \(x=(x(t))_{t\in N}\) is a time series following a standard ARIMA (p,d,q) model constructed from a white noise e(t). The main principal results are extensions of previous papers of \textit{T. L. Lai} and \textit{C. Z. Wei [see e.g. Ann. Stat. 10, No.1, 154-166 (1982; Zbl 0649.62060)] and \textit{T. L. Lai}} and \textit{H. Robbins} [see e.g. ibid. 7, 1196-1221 (1979; Zbl 0426.62059)]. Under the hypotheses: i) X is stable, ii) \(E(e(t)^{4+\delta})<+\infty\) for some \(\delta\geq 0\), and setting: \(H(n)=\sum^{n}_{t=1}x(t)^ 2/t^{2d}\) log n, there holds: a) if \(\delta =0:\) \(\lim_{n\to \infty}H(n)=k_ df(0)\) a.s., b) for some \(\delta >0:\) (log n)\({}^{1/2}[H(n)-k_ df(0)]\to^{D}N(0,R_ df(0)),\) where f(d) is the spectral density of \(y(t)=(1-B)^ dx(t)\) and B is the standard backward operator.