A Decomposition Theory for Binary Linear Codes

TL;DR

This paper explores a decomposition theory for binary linear codes based on matroid theory, discusses its implications for maximum-likelihood decoding via linear programming, and highlights limitations for good codes.

Contribution

It introduces a matroid-based decomposition framework for binary codes and connects it to linear programming decoding, revealing limitations for high-performance codes.

Findings

Certain code families allow polynomial-time ML decoding via LP.

Good codes cannot be decoded reliably using LP due to pseudocodewords.

The codeword polytope can match the fundamental polytope for some codes.

Abstract

The decomposition theory of matroids initiated by Paul Seymour in the 1980's has had an enormous impact on research in matroid theory. This theory, when applied to matrices over the binary field, yields a powerful decomposition theory for binary linear codes. In this paper, we give an overview of this code decomposition theory, and discuss some of its implications in the context of the recently discovered formulation of maximum-likelihood (ML) decoding of a binary linear code over a discrete memoryless channel as a linear programming problem. We translate matroid-theoretic results of Gr\"otschel and Truemper from the combinatorial optimization literature to give examples of non-trivial families of codes for which the ML decoding problem can be solved in time polynomial in the length of the code. One such family is that consisting of codes $C$ for which the codeword polytope is identical to the Koetter-Vontobel fundamental polytope derived from the entire dual code $C^\perp$. However, we also show that such families of codes are not good in a coding-theoretic sense -- either their dimension or their minimum distance must grow sub-linearly with codelength. As a consequence, we have that decoding by linear programming, when applied to good codes, cannot avoid failing occasionally due to the presence of pseudocodewords.