On the most reliable graphs with fixed redundancy

TL;DR

This paper characterizes the most reliable graphs with fixed redundancy, extending previous results to larger ranges and identifying specific optimal structures near zero failure probability.

Contribution

It provides a structural characterization of most reliable graphs near zero failure probability and extends the analysis to fixed redundancy with large number of vertices.

Findings

Regular graphs with maximal girth are optimal under certain conditions.

Most reliable graphs with fixed redundancy are obtained by subdividing cubic graphs.

Non-existence of uniformly most reliable graphs is proven for some infinite families.

Abstract

The all-terminal reliability of a graph $G$ is the probability that $G$ remains connected when each edge fails independently with probability $p$. For fixed $n$ and $m$, the uniformly most reliable problem asks which graph with $n$ vertices and $m$ edges maximizes reliability for all $p \in [0,1]$. Although such graphs do not always exist, optimal graphs in the regime $p \to 0$ always do and are determined by the structure of their minimal cut sets. We establish a structural characterization of graphs that are most reliable near $p=0$. Our results partially resolve a conjecture of Bourel et al., showing that, under suitable conditions, regular graphs with maximal girth are optimal. Extending this analysis to graphs with fixed redundancy $r=m-(n-1)$ and sufficiently large $n$, we show that the most reliable graphs are obtained by subdividing the most reliable cubic graphs with $2(r-1)$ vertices. The general conjecture remains open. Unlike previous results, which resolved only small redundancy cases or very dense regimes, our approach yields a substantial extension of the known range. We determine the unique cubic candidates for uniformly most reliable graphs for all redundancy levels $m-n \le 19$, and prove the non-existence of uniformly most reliable graphs for several infinite families with fixed redundancy and asymptotically large $n$. These results significantly enlarge both the candidate class and the range of provable non-existence.