Martingale approximation and optimality of some conditions for the central limit theorem (Q1960236)

Let \((\xi_i)_{i\in\mathbb N}\) be a homogeneous and ergodic Markov chain with transition kernel \(Q\) on a state space \((S, \mathcal B, \nu)\) with stationary probability measure \(\nu\), and let \(f\) be a square-integrable function on \(S\). Several authors have proved central limit theorems (CLT's) for the stationary process \((f(\xi_i))_{i\in\mathbb N}\) under various conditions on \(Q\) and \(f\). The paper under review studies the optimality of these conditions by constructing examples of Markov chains which do not satisfy the CLT. For example, if \(Q\) is a normal operator and \(\sum_{n=1}^{\infty} n^{-3/2} \sum_{k=0}^{n-1} Q^k f\) converges in \(L^2\), then the CLT holds; see \textit{M. I. Gordin} and \textit{B. A. Lifšic} [in: Third Vilnius conference on probability and statistics, Vol.\ 1, 147--148 (1981)], \textit{Y. Derriennic} and \textit{M. Lin} [C.\ R.\ Acad.\ Sci., Paris, Sér.\ I 323, No.~9, 1053--1057 (1996; Zbl 0869.60019)]. The paper under review proves that the condition of normality cannot be dropped. Also, it has been shown by \textit{M. Maxwell} and \textit{M. Woodroofe} [Ann.\ Probab.\ 28, No.~2, 713--724 (2000; Zbl 1044.60014)] and \textit{W. B. Wu} and \textit{M. Woodroofe} [Ann.\ Probab.\ 32, No.~2, 1674--1690 (2004; Zbl 1057.60022)] that without assuming the normality of \(Q\), the CLT holds provided that the series \(\sum_{n=1}^{\infty} n^{-3/2} \|\sum_{k=0}^{n-1} Q^k f\|_2\) converges, in particular, if \(\|\sum_{k=0}^{n-1} Q^k f\|_2=o(\sqrt n/\log ^q n)\), \(q>1\). By constructing appropriate counterexamples, the author shows that the conditions \(\sum_{n=1}^{\infty} c_nn^{-3/2} \|\sum_{k=0}^{n-1} Q^k f\|_2<\infty\) for \(c_n\to 0\) or \(\|\sum_{k=0}^{n-1} Q^k f\|_2=o(\sqrt n/\log n)\) are not sufficient for the validity of the CLT.