Coupling for \(\tau\)-dependent sequences and applications (Q1770896)

Let \(X\) be a real-valued random variable and \(M\) a \(\sigma\)-algebra. The authors consider the mixing coefficient \(\beta(M, \sigma(X))\) between \(M\) and \(\sigma(X)\), and its coupling property: if the probability space is rich enough, there exists a random variable \(X^*\) independent of \(M\) and distributed as \(X\) such that \(\mathbf P(X \neq X^*) = \beta(M, \sigma(X))\). The main purpose of this paper is to introduce a new dependence coefficient, denoted \(\tau(M, X)\), and defined as the minimum \(L^1\)-distance between \(X\) and a random variable that is distributed as \(X\) and is independent of \(M\). This newly defined dependence coefficient \(\tau(M, X)\) is compared to other well-known measures of dependence, and its basic and computational properties are investigated. Section 2 presents the corresponding coupling lemma for \(\tau(M, X)\), showing that it does not depend on the choice of the conditional property, and examining the case when \(X\) takes its values in a Banach space. Section 3 compares \(\tau(M, X)\) to the strong mixing coefficient \(\theta(M, X)\) of \textit{M. Rosenblatt} [Proc. Nat. Acad. Sci. USA 42, 43--47 (1956; Zbl 0070.13804)] and to the \(\alpha(M, X)\) \(s\)-dependence coefficient of \textit{C. Coulon-Prieur} and \textit{P. Doukhan} [Stat. Probab. Lett. 47, No. 1, 61--68 (2000; Zbl 0956.60006)]. Three large classes of examples (causal Bernoulli shifts, iterative random functions, and other Markov chains) are studied, which shows that \(\tau(M, X)\) can be easily computed in certain various situations. Section 4 establishes some deviation inequalities for the maximum of partial sums of \(\tau\)-dependent sequences. Specifically, Bennet's inequality and Rosenthal's inequality for independent sequences are extended to the case of \(\tau\)-dependent sequences. Using these results, Section 5 proves a strong invariance principle for partial sums of \(\tau\)-dependent sequences.