On the failure of additivity of inversion numbers for dijoins

Abstract

Let $D$ be an oriented graph. For $X \subseteq V(D)$, the inversion of $X$ in $D$ is the oriented graph that is obtained via changing the direction of all edges which have both endpoints in $X$. The inversion number of $D$, $\operatorname{inv}(D)$, is the minimum number of inversions necessary in order to remove all cycles from $D$. A dijoin $D_1 \to D_2$ is the oriented graph obtained by taking disjoint copies of $D_1,D_2$ and adding edges from all $v \in V(D_1)$ to each $u \in V(D_2)$. It was conjectured by Bang-Jensen, da Silva and Havet that $\operatorname{inv}(D_1 \to D_2) = \operatorname{inv}(D_1) + \operatorname{inv}(D_2)$ for all $D_1, D_2$. This was disproved by simultaneous constructions of various $D_1$, $D_2$ with $\operatorname{inv}(D_1 \to D_2) = \operatorname{inv}(D_1)+\operatorname{inv}(D_2)-1$. However, no known constructions achieved more than this gain of 1. Moreover, all of the constructions exploit a gap of 1 between $\operatorname{inv}(D)$ and the minimum rank of a family of related matrices, a gap that can never be more than 1, suggesting that perhaps something close to the original conjecture might be true. We prove that in fact additivity fails badly. In particular, for all $\varepsilon>0$ we show that there exist tournaments $D_1$, $D_2$ with arbitrarily large inversion numbers such that $\operatorname{inv}(D_1 \to D_2) \leq \left(\frac{1}{2} +\varepsilon\right)(\operatorname{inv}(D_1)+\operatorname{inv}(D_2))$.