How sampling reveals a process (Q921710): Difference between revisions

Let \(X=\{x_ 0,x_ 1,x_ 2,...\}\) be an ergodic, stationary process, then almost all infinite realizations \((x_ 0(\omega),x_ 1(\omega),...)=(\xi_ 0,\xi_ 1,...)\) determine the correct joint distribution of the random variables, and therefore the distribution of the process. However, the infinite realizations cannot be observed. Therefore the authors study the natural problem if one can guess the distribution from the finite sequences \((\xi_ 0,...,\xi_ n)\) in such a way that, as \(n\to \infty\), the guesses converge to the true process. They ask for convergence in the \(\bar d\)-metric, which is a strong requirement. If the state spaces are finite, \(\bar d(X,Y)\leq \epsilon\) means that there exists a stationary process \((\bar X,\bar Y)\) of pairs \((\bar x_ i,\bar y_ i)\) such that \(\bar X\) has the same joint distribution as X, \(\bar Y\) the same as Y, and \(P(\bar x_ 0\neq \bar y_ 0)\leq \epsilon.\) It is shown that such a guessing scheme exists if X is a Bernoulli process (not necessarily independent), i.e. if X is the limit in the \(\bar d\)-metric of totally ergodic k-steps Markov-processes. (They permit quite general state spaces.) It is also shown that, in a certain sense, the class of Bernoulli processes is the largest one for which the desired guessing schemes exist. Outside the class of Bernoulli processes even the question whether two processes are different cannot be answered by an effective scheme.