Estimation for a class of positive nonlinear time series models (Q1272160): Difference between revisions

Consider an infinitely divisible family of absolutely continuous distribution functions \(\{F_{\alpha}, \alpha \geq 0\}\) which satisfy \(F_{\alpha}*F_{\beta}=F_{\alpha +\beta}\). Let \(F^{\leftarrow}\) be the quantile function corresponding to \(F\). For \(0\leq p\leq 1\) define \[ X_t=F_{\alpha p}^{\leftarrow}(F_{\alpha}(X_{t-1}))+Z_t, \qquad t=1,2,\dots , \] where \(Z_t\sim F_{\alpha (1-p)}\) is a strict white noise and \(X_0\sim F_{\alpha}\). Then \(\{X_t\}\) is a nonlinear, non-Gaussian positive stationary process. The authors investigate an estimate of \(p\) from a sample \(X_0,\dots , X_n\) when \(\alpha \) is known. The asymptotic theory uses a regular variation assumption on the left tail of the innovation distribution.