A log log law for unstable ARMA models with applications to time series analysis (Q1185829)

Let \(\{X(t)\}\) be an ARMA process defined by \(\phi(B)X(t)=C(B)\varepsilon(t)\) where \(B\) is the backshift operator and \(\varepsilon(t)\) is a white noise. Let the polynomial \(\phi(z)=1-\beta_ 1z-\dots-\beta_ pz^ p\) have \(r\) different roots \(\lambda_ 1,\dots,\lambda_ r\) with multiplicities \(d_ 1,\dots,d_ r\) such that \(| \lambda_ j|=1\) for \(j=1,\dots,r\), and all the remaining roots lie outside the unit circle. Define \(d=\max_{1\leq j\leq r}d_ j\), \(a=2^ d-1\), \(\bar X_ k=k^{-1}\sum^ k_{j=1}X(j)\). Assume \(d>0\). It is proved that for an arbitrary real constant \(b\) \[ \sum^ n_{k=1} [X(k)-b\bar X_{k-1}]^ 2=O(n^{2d}\log\log n) \] and \[ \left(\sum^ n_{k=1} [X(k)-b\bar X_{k-1}]^ 2\right)^{-1}=O(n^{-2d}(\log\log n)^ a). \]