Multiplicative ergodicity and large deviations for an irreducible Markov chain. (Q1879486): Difference between revisions

Let \(\{\Phi _ {n}\}_ {n\geq 0}\) be an irreducible, aperiodic, and recurrent Markov chain on a countable state space \(X\) and \(F: X\to \mathbb R_ {+}\) a nonnegative function such that each sublevel set \(\{x\in X;\; F(x)\leq n\}\) is finite. Assume that \(\mathbf E_ {x}\exp (V_ 0( \Phi _ 1))\leq \exp (V_ 0(x) - \alpha _ 0 F(x) + B_ 0)\) for some function \(V_ 0: X\to \mathbb R_ {+}\) and constants \(\alpha _ 0>0\), \(B_ 0<\infty \). Denote by \(S_ {n}\), \(n\geq 1\), the partial sums \(S_ {n} = \sum ^ {n-1}_ {i=0} F(\Phi _ {i})\). It is shown that there exists a convex function \(\Lambda : \mathbb R\to \left ]-\infty ,+\infty \right ]\), finite on a domain with the interior \(\left ]-\infty ,\alpha _ 1\right [\) for some \(\alpha _ 1\geq \alpha _ 0\), and, for every \(\alpha <\alpha _ 1\), there exists a function \(f_ \alpha : X\to \mathbb R_ {+}\) such that \(\mathbf E_ {x}\exp (\alpha S_ {n}- n\Lambda (\alpha ))\) converges to \(f_ \alpha (x)\) geometrically fast as \(n\to \infty \). The multiplicative ergodic theorem \(\lim _ {n\to \infty }n^ {-1} \log \mathbf E_ {x}\exp (\alpha S_ {n}) = \Lambda (\alpha )\) holds for all \(\alpha \). If \(\alpha <\alpha _ 1\), then \(\Lambda (\alpha )\) is the unique solution to \(\mathbf E_ \theta \exp (\sum ^ {\tau _ \theta -1}_ {k=0} [\alpha F(\Phi _ {k})-\Lambda (\alpha )]) =1\), where \(\theta \in X\) is an arbitrary (but fixed) state, \(\tau _ \theta = \inf \{k\geq 1;\; \Phi _ {k} =\theta \}\), and the function \(f_ \alpha \) solves the multiplicative Poisson equation \(Pf_ \alpha (x) = f_ \alpha (x)\exp (-\alpha F(x) + \Lambda (\alpha ))\), \(P\) denoting the transition probability of the chain \(\{\Phi _ {k}\}\). These results are then used to establish large deviation bounds for the empirical measures.