On moving averages and asymptotic equipartition of information (Q343262): Difference between revisions

It is well known that in general a moving average form of the pointwise ergodic theorem fails for certain sequences and holds for others. A sequence of pairs \((n_{\ell},k_{\ell})_{\ell\geq1}\) is said to be Stolz if there is a collection of points \(Z\) in \(\mathbb{Z}\times\mathbb{N}\) and a function \(h\) with \(h(t)\to\infty\) as \(t\to\infty\) such that \((n_{\ell},k_{\ell})_{\ell\geq t}\in \{(n,k)\mid (n,k)\in Z, k\geq h(t)\}\) and there exist \(h_0,\alpha_0\) and \(A>0\) such that \( |\{n\in\mathbb{Z}\mid | n-y|<\alpha|\lambda-r|\text{ for some }(y,r)\in Z\text{ with }r\geq h_0\}|\leq A\lambda \) for all integers \(\lambda>0\). \textit{A. Bellow} et al. [Ergodic Theory Dyn. Syst. 10, No. 1, 43--62 (1990; Zbl 0674.60035)] showed that this technical condition is sufficient for the pointwise convergence of moving averages in the following sense. If \((X,\mathcal{B},\mu,T)\) is a probability measure-preserving system and the sequence \((n_{\ell},k_{\ell})_{\ell\geq1}\) is Stolz, then for any \(f\in L^1(X,\mathcal{B},\mu)\) the moving averages \((1/k_{\ell}) \sum_{i=1}^{k_{\ell}}f(T^{n_{\ell}+i}x)\) converge as \(\ell\to\infty\) for \(\mu\)-almost every \(x\in X\). This result subsumes many of the existing positive results on the question. Here a similar question is taken up for the Shannon-McMillan-Breiman theorem as follows. If \((x_{\ell})_{\ell\in\mathbb{Z}}\) is a two-sided stationary stochastic process taking values in the finite set \(K=\{a_1,\dots,a_s\}\) with joint distribution function of the variables \(x_0,\dots,x_n\) given by \(p(x_0,\dots,x_n)\), then it is shown here that, if \((n_{\ell},k_{\ell})_{\ell\geq1}\) is Stolz, then there is a constant \(H\) for which \((1/k_{\ell})\log p(x_{n_{\ell}},x_{n_{\ell}+1},\dots,x_{n_{\ell}+k_{\ell}}) \to -H\) as \(\ell\to\infty\) almost surely. If \(n_{\ell}=1\) for all \(\ell\geq1\), then this is the Shannon-McMillan-Breiman (SMB) theorem. The fundamental role of the SMB theorem in coding and compression theory is briefly recalled, and a scenario is presented here where a similar role may be played by the moving average version developed here.