Ergodic decomposition of excess entropy and conditional mutual information

The article discusses excess entropy defined as mutual information between the past and future of a stationary process. The central result is an ergodic decomposition: Excess entropy is the sum of self-information of shift-invariant σ-field and the average of excess entropies for the ergodic components of the process. The result is derived using generalized conditional mutual information for fields of events, developed in the paper anew. Some corollary of the ergodic decomposition is that excess entropy is infinite for the class of processes with uncountably many ergodic components, called here uncountable description processes (UDP’s). UDP’s can be defined without the use of measure theory and the article argues for their potential utility in linguistics. Moreover, it is shown that finite-order excess entropies (some approximations of excess entropy) are dominated by the expected excess lengths of any universal code. Hence, universal codes may be used for rough estimation of excess entropy. Nevertheless, the excess code lengths diverge to infinity for almost every process with zero excess entropy, which is another corollary of the ergodic decomposition.