Moments for stationary Markov chains with asymptotically zero drift (Q642081): Difference between revisions

A stationary Markov chain \(X\) with an invariant distribution \(\pi\) is studied. For a time-homogeneous Markov chain \(X=\{X_n, n\geq 0\}\) taking values in \({\mathbb R}^+\), the random variable \(\xi (x)\) corresponding to the jump of the chain at point \(x\) satisfies \(\operatorname{P}\{\xi (x) \in B\} = \operatorname{P}\{X_{n+1}- X_n\in B| X_n = x\}\) with mean drift \(m(x)\) and second moment \(b(x) = \operatorname{E}\xi^2(x).\) It is well-known that the Lindley recursion \(X_{n+1} = (X_n + \xi_{n+1})^+\), where the \(\xi_n\), \(n\geq 1\), are i.i.d random variables, has invariant distribution if and only if \(\operatorname{E}\xi_1 < 0.\) In queueing theory, it is well-known that the invariant distribution has finite \(\gamma\)-th moment if and only if \(\operatorname{E}(\xi^+)^{\gamma + 1}< \infty\), and the invariant distribution \(\pi\) is light-tailed if and only if the distribution of \(\xi\) is. Theorem 1 shows that almost every stationary Markov chain with asymptotically zero drift (\(m(x) \to 0\) if \(x\to \infty\)) generates a heavy-tailed invariant distribution (\(\operatorname{E}e^{\lambda \xi} < \infty\) for some \(\lambda > 0\)). \textit{J. Lamperti} [J. Math. Anal. Appl. 1, 314--330 (1960; Zbl 0099.12901)] first studied the existence of an invariant distribution in the critical case when \(m(x) \sim - c/x\) at large \(x\). Theorem 1 is somehow sharp since there is no invariant distribution for the Markov chain with \(m(x) = o(1/x)\) as \(x\to \infty\). The approach in this paper to prove that existence or nonexistence of the moment is based on the calculations of the mean drift of the appropriate test function of a Markov chain. Section 3 lists technical lemmas. \textit{R. L. Tweedie} [J. Appl. Probab. 20, 191--196 (1983; Zbl 0513.60067), Theorem 1] first proved an existence result on moments for Harrison-recurrent stationary Markov chains. Under some technical assumptions, the author proves in Theorem 2 the existence of the moments for a class of Markov chains with jump random variables. Theorem 3 in Section 5 gives the nonexistence result of moments for a class of Markov chains with jump random variables under conditions (19)--(22) in the paper. If \(m(x) \sim -\mu x^{-1}\log x\), \(m(x) \sim -\mu/x^{\alpha}\), \(\alpha \in (0, 1)\), then \(xm(x)\to - \infty\), and the moments of the invariant distribution are all finite if the moments of \(\xi^+(x)\) are finite for \(x\geq 0\). Theorem 3 implies the nonexistence of Weibull-type moments (see Corollary 5). For countable Markov chains with asymptotically zero drift and bounded jumps, some results are listed in the last two sections.