Transition phenomena for real-valued Markov chains (Q1920090): Difference between revisions

Let \(\{X_n^{(\varepsilon)}; n\in\mathbb{Z}^+\}\) be a family (indexed by \(\varepsilon \geq 0\)) of homogeneous Markov chains taking values in \(\mathbb{R}^+\) with transition functions \(P^{(\varepsilon)}(x,\cdot)\), \(x\geq 0\). Assume that \(\{X_n^{(\varepsilon)}\}\) has a unique invariant distribution \(\pi^{(\varepsilon)}\) for each \(\varepsilon>0\). The main object of the author's investigation is the limiting behaviour of \(\pi^{(\varepsilon)}\) as \(\varepsilon\to 0\). Let \(\xi^{(\varepsilon)}(x)\) be the random variable describing the jump of the Markov chain \(\{X_n^{(\varepsilon)}\}\) starting at \(x\): \(P\{x+\xi^{(\varepsilon)}(x)\in B\}=P^{(\varepsilon)}(x,B)\). Define \(m^{(\varepsilon)}(x)=E\xi^{(\varepsilon)}(x)\), \(b^{(\varepsilon)}(x)=E(\xi^{(\varepsilon)}(x))^2\). Assume that the following regularity conditions hold for \(\{X_n^{(0)}\}\): \[ \lim_{x\rightarrow\infty} m^{(0)}(x)x=\mu,\quad-\infty \leq \mu \leq \infty,\quad\lim_{x\rightarrow \infty} b^{(0)}(x)=b,\quad 0<b<\infty,\quad\sup_x b^{(0)}(x)<\infty. \] The parameters \(\mu\) and \(b\) play the crucial role in classifying the asymptotic behaviour of the distribution \({\pi}^{(\varepsilon)}\). Here are two theorems that illustrate different behaviour of \({\pi}^{(\varepsilon)}\) as \(\varepsilon \rightarrow 0\). Theorem 1. Let \(2\mu<-b\). Suppose that the following asymptotic representation holds: \[ m^{(\varepsilon)}(x)=-{\varepsilon}+{\mu/x}+o(\varepsilon+{1/ x}),\quad x\rightarrow\infty,\varepsilon\rightarrow 0,\quad\sup_{x,\varepsilon} b^{(\varepsilon)}(x) < \infty,\quad\lim_{x\rightarrow \infty, \varepsilon \rightarrow 0}b^{(\varepsilon)}(x)=b. \] Assume that the random variables \(\max(0,({\xi}^{(\varepsilon)}(x))^2)\) are uniformly integrable in \(x\geq 0\) and \(\varepsilon \geq 0\), assume that the Markov chain \(\{X_n^{(0)}\}\) has a unique invariant distribution \(\pi^{(0)}\), and assume that \(P^{(\varepsilon)}(x, \cdot) \Rightarrow P^{(0)}(x,\cdot)\) as \(x\rightarrow y\), \(\varepsilon \rightarrow 0\) (weak convergence). Then \(\pi^{(\varepsilon)} \Rightarrow\pi^{(0)}\) as \(\varepsilon \rightarrow 0\). Theorem 2. Let \(\infty>2\mu > -b\). Suppose that the asymptotic representations from Theorem 1 hold, and suppose that for each compact set \(K\subset\mathbb{R}^+\) there exists \(\varepsilon_*=\varepsilon_*(K)>0\) such that \(\inf_{x\in K, \varepsilon \leq\varepsilon_*} P\{X_n^{(\varepsilon)} \notin K \text{ \;for \;some } n |X_0^{(\varepsilon)}=x \} >0.\) Assume that the random variables \((\xi^{(\varepsilon)}(x))^2\) are uniformly integrable in \(x\geq 0\) and \(\varepsilon \geq 0\). Then, as \(\varepsilon \rightarrow 0\), \[ {\mathcal L}(2\varepsilon X^{(\varepsilon)})\Rightarrow \Gamma_{1/b, 1+2\mu/b}, \] where \(X^{(\varepsilon)}\) is the random variable with law \({\pi}^{(\varepsilon)}\), and \(\Gamma_{\alpha,\beta}\) is the gamma distribution with parameters \(\alpha\) and \(\beta\). The author also considers the critical case \(2\mu = -b\) thus obtaining the complete classification. In the last paragraph the same problem for Markov chains with values in \(\mathbb{R}\) is treated, and similar results are obtained.