A Dense Neighborhood Lemma: Applications of Partial Concept Classes to Domination and Chromatic Number

In its Euclidean form, the Dense Neighborhood Lemma (DNL) asserts that if $V$ is a finite set of points of $\mathbb{R}^N$ such that for each $v \in V$ the ball $B(v,1)$ intersects $V$ on at least $\delta |V|$ points, then for every $\varepsilon >0$, the points of $V$ can be covered with $f(\delta,\varepsilon)$ balls $B(v,1+\varepsilon)$ with $v \in V$. DNL also applies to other metric spaces and to abstract set systems, where elements are compared pairwise with respect to (near) disjointness. In its strongest form, DNL provides an $\varepsilon$-clustering with size exponential in $\varepsilon^{-1}$, which amounts to a Regularity Lemma with 0/1 densities of some trigraph. Trigraphs are graphs with additional red edges. They are natural instances of partial concept classes, introduced by Alon, Hanneke, Holzman and Moran [FOCS 2021]. This paper is mainly a combinatorial study of the generalization of Vapnik-Cervonenkis dimension to partial concept classes. The main point is to show how trigraphs can sometimes explain the success of random sampling even though the VC-dimension of the underlying graph is unbounded. All the results presented here are effective in the sense of computation: they primarily rely on uniform sampling with the same success rate as in classical VC-dimension theory. Among some applications of DNL, we show that $\left(\frac{3t-8}{3t-5}+\varepsilon\right)\cdot n$-regular $K_t$-free graphs have bounded chromatic number. Similarly, triangle-free graphs with minimum degree $n/3-n^{1-\varepsilon}$ have bounded chromatic number (this does not hold with $n/3-n^{1-o(1)}$). For tournaments, DNL implies that the domination number is bounded in terms of the fractional chromatic number. Also, $(1/2-\varepsilon)$-majority digraphs have bounded domination, independently of the number of voters.