Table of contents

The paper deals with so-called -partially space-homogeneous time-homogeneous Markov chains , , , in the positive quadrant . These Markov chains are characterized by the following property of the transition probabilities : for some the measure depends only on , , and in the domain , , and only on , , and in the domain , . For such chains the asymptotic behaviour of

is found for a fixed set as , , and . Some other conditions on the growth of parameters are also considered, for example, , . A study is made of the structure of the most probable trajectories, which give the main contribution to this asymptotics, and a number of other results pertaining to the topic are established.

Similar results are obtained for the narrower class of 0-partially homogeneous ergodic chains under less restrictive moment conditions on the transition probabilities . Moreover, exact asymptotic expressions for the probabilities are found for 0-partially homogeneous ergodic chains under some additional conditions.

The interest in partially homogeneous Markov chains in positive octants is due to the mathematical aspects (new and interesting problems arise in the framework of general large deviation theory) as well as applied issues, for such chains prove to be quite accurate mathematical models for numerous basic types of queueing and communication networks such as the widely known Jackson networks, polling systems, or communication networks associated with the ALOHA algorithm. There is a vast literature dealing with the analysis of these objects.

The present paper is an attempt to find the extent to which an asymptotic analysis is possible for Markov chains of this type in their general form without using any special properties of the specific applications mentioned above. It turns out that such an analysis is quite possible, though difficult, in the two-dimensional case. However, even in the three-dimensional case one encounters new fundamental difficulties which, at the present state of the art, render the problem insoluble or at least extremely hard to solve.

The Gibbs field is one of the central objects of modern probability theory, mathematical statistical physics, and Euclidean field theory. In this paper we introduce and study a natural generalization of this field to the case in which the background space (a lattice, a graph) on which the random field is defined is itself a random object. Moreover, this randomness is given neither a priori nor independent of the configuration; on the contrary, the space and the configuration on it depend on each other, and both objects are given by a Gibbs construction. We refer to the resulting distribution as a Gibbs family because it parametrizes Gibbs fields on different graphs belonging to the support of the distribution. We also study the quantum analogue of Gibbs families and discuss relationships with modern string theory and quantum gravity.