Extracting Wyner's Common Information Using Polar Codes and Polar Lattices

TL;DR

This paper develops explicit polar code and lattice constructions to extract Wyner's common information for both discrete and Gaussian sources, achieving the best-known Gray-Wyner regions without time-sharing and extending to multiple sources.

Contribution

It introduces novel polar code and lattice methods for extracting Wyner's common information, covering lossless and lossy Gray-Wyner problems for various source types.

Findings

Achieves the entire best-known region of the lossless Gray-Wyner problem for discrete sources.

Extends polar lattices to extract common information of Gaussian sources.

Shows that extracting Wyner's common information for Gaussian sources is equivalent to lossy compression of a single Gaussian source.

Abstract

Explicit constructions of polar codes and polar lattices for both lossless and lossy Gray-Wyner problems are studied. Polar codes are employed to extract Wyner's common information of doubly symmetric binary source; polar lattices are then extended to extract that of a pair of Gaussian sources or multiple Gaussian sources. With regard to the discrete sources, the entire best-known region of the lossless Gray-Wyner problem are achieved by specifying the test channels to construct polar codes without time-sharing. As a result, we are able to give an interpretation that the Wyner's common information remains the same to the lossy case when the distortion is small [1]. Finally, the entire best-known lossy Gray-Wyner region for discrete sources can also be achieved using polar codes. With regard to the Gaussian sources, the best-known lossy Gray-Wyner region for bivariate Gaussian sources with a specific covariance matrix [1] can be achieved by using polar lattices. Moreover, we prove that extracting Wyner's common information of a pair of Gaussian sources is equivalent to implementing the lossy compression for a single Gaussian source, which implies that the common information can be extracted by a polar lattice for quantization. Furthermore, we extend this result to the case of multiple Gaussian sources.