On the solvability of 3-source 3-terminal sum-networks

TL;DR

This paper characterizes when a 3-source 3-terminal network can reliably compute the sum of source symbols over any field, establishing conditions for solvability, sufficiency of linear codes, and the capacity with fractional coding.

Contribution

It provides a necessary and sufficient condition for solvability of 3-source 3-terminal sum-networks over any field, and shows linear codes are sufficient in this case.

Findings

A complete characterization of solvability conditions.

Linear codes are sufficient for 3-source 3-terminal sum-networks.

The capacity with fractional coding is 0, 2/3, or at least 1.

Abstract

We consider a directed acyclic network with three sources and three terminals such that each source independently generates one symbol from a given field $F$ and each terminal wants to receive the sum (over $F$) of the source symbols. Each link in the network is considered to be error-free and delay-free and can carry one symbol from the field in each use. We call such a network a 3-source 3-terminal {\it $(3s/3t)$ sum-network}. In this paper, we give a necessary and sufficient condition for a $3s/3t$ sum-network to allow all the terminals to receive the sum of the source symbols over \textit{any} field. Some lemmas provide interesting simpler sufficient conditions for the same. We show that linear codes are sufficient for this problem for $3s/3t$ though they are known to be insufficient for arbitrary number of sources and terminals. We further show that in most cases, such networks are solvable by simple XOR coding. We also prove a recent conjecture that if fractional coding is allowed, then the coding capacity of a $3s/3t$ sum-network is either $0,2/3$ or $\geq 1$.