A new theorem on the existence of invariant distributions with applications to ARCH processes (Q4408540)

In the last time an extensive literature in econometrics has deal with the autoregressive conditional heteroscedastic (ARCH) processes, which can be defined as stationary solutions \((r_k)\) to the system of equations: \[ r_k=\sigma_k \varepsilon_k, \quad\sigma_k^2 =a_0+\sum^\infty_{i=1} a_ir^2_{k-i},\quad k\in\mathbb{Z}.\tag{1} \] In this case the \(\varepsilon_k\) are independent and identically distributed (i.i.d.) random variables with common distribution \(\lambda\), \(a_0>0\) and \(a_i\geq 0\) for \(i\geq 1\). Although conditions for the existence of such processes in particular cases are established (see Section 3), the problem in general is not completely solved. In particular, it is not known whether the following condition guarantees the existence of a stationary solution to (1): \(E\varepsilon_k \sum^\infty_{i=1} a_i=1\). If such a solution exists, then an integrated ARCH process (or IARCH process) is considered. Throughout the paper ``the IARCH problem'' means ``the problem of existence of IARCH processes''. Define \(\varepsilon_k= \varepsilon_k^2\), \[ A(\omega)= \left( \begin{matrix} a_1\omega & a_2\omega\\ 1 & 0\\ 0 & 1\\ \vdots &\vdots\end{matrix} \right),\quad a(\omega)= \left(\begin{matrix} a_0\omega \\ 0 \\ 0\\ \vdots\end{matrix} \right),\quad \xi_k= \left(\begin{matrix} r_k^2\\ r^2_{k-1}\\ r^2_{k-2}\\ \vdots\end{matrix} \right). \] Let \(X=[0,\infty)\) and \(W=[0,\infty)\). If \((r_k)\) is a solution to (1), then \(( \xi_k)\) satisfies the system of equations \(\xi_k=f(\xi_{k-1},\varepsilon_k)\), \(k \in\mathbb{Z}\), where \(f:X\times[0,\infty)\to X\) is defined by \(f(X,\omega)= A( \omega) x+a(\omega)\). The sequence \((\xi_k)\) is a Markov chain evolving on the space \(X\) according to the transition probability \(P(x,B)=\lambda(\omega\mid A( \omega)x+ a(\omega)\in B)\) and the problem is to determine whether \(P\) admits an invariant initial distribution. The authors use an idea of \textit{H. Kesten} and \textit{F. Spitzer} [Z. Wahrscheinlichkeitstheorie Verw. Geb. 67, 363-386 (1984; Zbl 0535.60016)] and reduce the problem to another task, finding a stationary solution to the system of equations \(A(\xi_k)\xi_{k-1} =\rho_k\xi_k\), \(k\in \mathbb{Z}\), where \(\rho_k\) are positive random vectors. The authors present a new theorem on the existence of invariant distribution for Markov chains evolving on a Polish space (Theorem 2.1) and apply it to the Markov chain defined by the iterations of the functions \(f(x,\omega)= A(\omega)x/b^T A(\omega) x\), under the supplementary assumption that the coefficients decay exponentially. The proof of Theoren 2.1 is based on the Schauder fixed point theorem but, unlike \textit{J. Liu} and \textit{E. Susko} [J. Appl. Probab. 29, 363-373 (1992; Zbl 0753.62059)] who considered the topology on the space of probability measures generated by the norm. In Section 2 the authors formulate and prove the main theorem and in Section 4 they apply it to the study of stationarity in ARCH models and prove the existence of IARCH processes.