A new framework for identifying most reliable graphs and a correction to the $K_{3,3}$-theorem

Abstract

Given a multigraph $G$, the all-terminal reliability $R(G,p)$ is the probability that $G$ remains connected under percolation with parameter $p$. Fixing the number of vertices $n$ and edges $m$, we investigate which graphs maximize $R(G,p)$ -- such graphs are called optimal -- paying particular attention to uniqueness and to whether the answer depends upon $p$. We generalize the concept of a distillation and build a framework with which we identify all optimal graphs for which $m-n\in \{1,2,3\}$. These graphs are uniformly optimal in $p$. Most have been previously identified, but with serious problems, especially when $m-n=3$. We obtain partial results for $m-n\in \{4,5\}$. For $m-n=3$, the optimal graphs were incorrectly identified by Wang in 1994, in the infinite number of cases where $m\equiv 5(\mathrm{mod}9)$ and $m\ge 14$. This erroneous result concerns subdivisions of $K_{3,3}$ and has been cited extensively, without any mistake being detected. While the optimal graphs were correctly described for other $m$, the proof is fundamentally flawed. Our proof of the rectified statement is self-contained. For $m-n=4$, the optimal graphs were recently shown to depend upon $p$ for infinitely many $m$. We find a new such set of $m$-values, which gives a different perspective on why this phenomenon occurs and leads us to conjecture that uniformly optimal graphs exist only for finitely many $m$. However, for $m-n=5$, we conjecture that there are again infinitely many uniformly optimal graphs.

44 pages, 26 figures. Some shortened arguments, clarifications, copyediting, two new figures. Comments welcome