On geometric ergodicity of nonlinear autoregressive models (Q1347199)

The authors prove elegantly the geometric ergodicity of a general class of nonlinear \(k\)-th order autoregressive processes \(X_n\), by considering \(k\)-tuples \((X_{n + 1 - k}, \ldots, X_n)\) as Markov processes on the state-space \({\mathcal R}^k\). The autoregressive structure means that \[ \bigl\{ X_{n + 1} - h(X_{n + 1 - k}, \ldots, X_n) \bigr\}_{n \geq k - 1} = \{\eta_{n + 1}\}_{n \geq k - 1} \] is an independent and identically distributed real sequence, where \(h\) is a Borel measurable function on \({\mathcal R}^k\). The authors impose the further conditions: (a) \(h\) is bounded on compact sets, and for some constants \(c \geq 0\), \(R > 0\), \(a_i > 0\) with \(\sum^k_{i = 1} a_i < 1\), \(|h(y) |\leq \sum^k_{i = 1} a_i |y_i |+ c\), for \(|y |\equiv \sum^k_{i = 1} |y_i |\geq R\), (b) \(\eta_n\) has absolutely continuous d.f. with a.e. positive density, and \(E |\eta_n |< \infty\). Previous authors, in deriving similar conclusions of geometric ergodicity, have imposed conditions requiring knowledge of \(h(y)\) everywhere (not just for large \(|y |)\), as well as continuity of \(h\). The tools employed here are the sufficient conditions of \textit{D. L. Tweedie} [in: Probability, statistics and analysis. Lond. Math. Soc. Lect. Note Ser. 79, 260-276 (1983; Zbl 0501.60072)] in terms of stochastic Lyapunov functions for geometric ergodicity of \(\varphi\)- irreducible Markov processes.