Constraint Complexity of Realizations of Linear Codes on Arbitrary Graphs

TL;DR

This paper investigates the minimal local constraint complexity of linear code realizations on arbitrary graphs, introducing new bounds and graph parameters to understand the trade-offs between code complexity and graph structure.

Contribution

It introduces the Vertex-Cut Bound and vc-treewidth, providing tight lower bounds on realization complexity and linking code complexity to graph-theoretic properties.

Findings

Lower bounds on realization complexity using new graph parameters.

Good codes require graphs with large vc-treewidth for low complexity.

Logarithmic growth of complexity ratio with code length is achievable.

Abstract

A graphical realization of a linear code C consists of an assignment of the coordinates of C to the vertices of a graph, along with a specification of linear state spaces and linear ``local constraint'' codes to be associated with the edges and vertices, respectively, of the graph. The $\k$-complexity of a graphical realization is defined to be the largest dimension of any of its local constraint codes. $\k$-complexity is a reasonable measure of the computational complexity of a sum-product decoding algorithm specified by a graphical realization. The main focus of this paper is on the following problem: given a linear code C and a graph G, how small can the $\k$-complexity of a realization of C on G be? As useful tools for attacking this problem, we introduce the Vertex-Cut Bound, and the notion of ``vc-treewidth'' for a graph, which is closely related to the well-known graph-theoretic notion of treewidth. Using these tools, we derive tight lower bounds on the $\k$-complexity of any realization of C on G. Our bounds enable us to conclude that good error-correcting codes can have low-complexity realizations only on graphs with large vc-treewidth. Along the way, we also prove the interesting result that the ratio of the $\k$-complexity of the best conventional trellis realization of a length-n code C to the $\k$-complexity of the best cycle-free realization of C grows at most logarithmically with codelength n. Such a logarithmic growth rate is, in fact, achievable.