Markov chains with stochastically stationary transition probabilities (Q1176357)

From MaRDI portal
scientific article
Language Label Description Also known as
English
Markov chains with stochastically stationary transition probabilities
scientific article

    Statements

    Markov chains with stochastically stationary transition probabilities (English)
    0 references
    0 references
    25 June 1992
    0 references
    In this special invited paper, the author combines a partial summary of results of Nawrotzki and Cogburn on Markov chains whose transition probabilities form a stationary process with new contributions. Section 1 serves to introduce the model: \((\Theta,{\mathcal B}_ 0)\) is a measurable space, \(\Omega=\Theta^{\mathbb{Z}}\), \(\Theta_ n\) the \(n\)-th coordinate map, \(T\) the shift, and \(\pi\) a \(T\)-invariant probability measure. For each \(\omega_ 0\in \Theta,P(\omega_ 0)\) is a stochastic matrix for a finite or denumerable state space \(X\). For fixed \(\omega\in \Omega\), \(P(\Theta_ n(\omega))\) is the matrix of transition probabilities for the \((n+1)\)st transition in \(X\). Some work by Cogburn, who noticed that this model can be viewed as a special skew product in the setting of Hopf's theory of Markov operators in \(L_ 1\), is reviewed. For this, consider the space \(E=X\times\Omega\), and the transition probabilities \(P((x,\omega),\{y\}\times B)=P(\omega_ 0,x,y)\mathbf{1}_ B(T\omega)\). The main result in Section 2 is that if the shift is ergodic, then every ''closed'' (i.e. absorbing) set for the skew product operator \(P\) contains a minimal closed subset. This yields an ergodic decomposition of the skew product into at most countably many components. (The relation to the ergodic decomposition of general conservative contractions in \(L_ 1\), due to Jacobs, is not discussed.) Section 3 deals with the similar model for the one-sided shift, and with its relation to the above model. For example it is shown that there is a bijection between the sets of invariant probability measures in the two models. Let \(m_ N^{(y,\omega)}\) denote the probability measure on the space \({\mathcal M}(X)\) of probabilities on \(X\), which assigns mass \(1/N\) to \(P(\omega_ 0,\omega_ 1,\ldots,\omega_{n-1},\cdot y,\cdot)\), where \(P(\omega_ m,\ldots,\omega_ n)=P(\omega_ m)P(\omega_{m+1})\ldots P(\omega_ n)\). It is shown that when \(\pi\) is ergodic and \(\phi\) invariant for the skew product, then \(m_ N^{(y,\omega)}\) converges weakly for \(\phi\)-a.e. \((y,\omega)\). A number of open problems are stated.
    0 references
    0 references
    random environment
    0 references
    random transition probabilities
    0 references
    stationary process
    0 references
    stochastic matrix
    0 references
    ergodic decomposition
    0 references
    invariant probability measures
    0 references
    0 references