Optimal concentration inequalities for dynamical systems (Q694976): Difference between revisions

Let \(X\) be a metric space and \(K:X^n\to {\mathbb R}\) a function that is Lipschitz in each variable separately with constants Lip\(_i(K)\). A stationary process \((Z_0, Z_1, \ldots )\) taking values in \(X\) is said to satisfy an exponential concentration inequality if there is a constant \(C\) so that for any separately Lipschitz function \(K(x_0, \ldots , x_{n-1})\), \[ {\mathbb E}\left(e^{K(Z_0, \ldots , Z_{n-1}) -{\mathbb E}(K(Z_0, \ldots , Z_{n-1}))}\right) \leq e^{C\sum_{j=0}^{n-1} \text{Lip}_j(K)^2} . \] This gives rise to a control on the deviation probabilities: for all \(t>0\) \[ {\mathbb P}(\left| K(Z_0, \ldots , Z_{n-1}) -{\mathbb E}(K(Z_0, \ldots , Z_{n-1}))\right| >t) \leq 2e^{-\frac{t^2}{4C\sum_{j=0}^{n-1} \text{Lip}_j(K)^2}} . \] Let \((X, T, \mu)\) be a dynamical system (\(T\) is defined on \(X\) and \(\mu\) is a \(T\)-invariant measure). The main aim of this paper is to study processes arising from dynamical systems. It is shown that certain processes \((x, Tx, T^2 x, \ldots )\) (where \(x\) is distributed following \(\mu\)), satisfy certain concentration inequalities. This type of work has been done for certain piecewise uniformly expanding maps on an interval [\textit{P. Collet} et al., Probab. Theory Relat. Fields 123, No. 3, 301--322 (2002; Zbl 1087.37028)]. Situations where the maps are less expanding (e.g., exponential towers, including Henon maps) have also been studied; see e.g., [\textit{J.-R. Chazottes} et al., Nonlinearity 18, No. 5, 2323--2340 (2005; Zbl 1084.37023)]. In this paper, it is shown that Young towers with exponential tails satisfy an exponential concentration inequality and the case of polynomial tails is studied, the inequalities being optimal. The paper starts by explaining the simplest situation -- subshifts of finite type endowed with a Gibbs measure, and finishes with applications to specific systems.