Correlated Sources In Distributed Networks - Data Transmission, Common Information Characterization and Inferencing

Correlation is often present among observations in a distributed system. This thesis deals with various design issues when correlated data are observed at distributed terminals, including: communicating correlated sources over interference channels, characterizing the common information among dependent random variables, and testing the presence of dependence among observations. It is well known that separated source and channel coding is optimal for point-to-point communication. However, this is not the case for multi-terminal communications. In this thesis, we study the problem of communicating correlated sources over interference channels (IC), for both the lossless and the lossy case. For lossless case, a sufficient condition is found using the technique of random source partition and correlation preserving codeword generation. The sufficient condition reduces to the Han-Kobayashi achievable rate region for IC with independent observations. Moreover, the proposed coding scheme is optimal for transmitting a special correlated sources over a class of deterministic interference channels. We then study the general case of lossy transmission of two correlated sources over a two-user discrete memoryless interference channel (DMIC). An achievable distortion region is obtained and Gaussian examples are studied. The second topic is the generalization of Wyner\u27s definition of common information of a pair of random variables to that of N random variables. Coding theorems are obtained to show that the same operational meanings for the common information of two random variables apply to that of N random variables. We establish a monotone property of Wyner\u27s common information which is in contrast to other notions of the common information, specifically Shannon\u27s mutual information and G\u27{a}cs and K {o}rner\u27s common randomness. Later, we extend Wyner\u27s common information to that of continuous random variables and provide an operational meaning using the Gray-Wyner network with lossy source coding. We show that Wyner\u27s common information equals the smallest common message rate when the total rate is arbitrarily close to the rate-distortion function with joint decoding. Finally, we consider the problem of distributed test of statistical independence under communication constraints. Focusing on the Gaussian case because of its tractability, we study in this thesis the characteristics of optimal scalar quantizers for distributed test of independence where the optimality is both in the finite sample regime and in the asymptotic regime