Equitable Colorings of Vertex-Weighted Graphs

TL;DR

This paper extends the Hajnal-Szemerédi theorem to vertex-weighted graphs, establishing new bounds and polynomial-time algorithms for equitable colorings with fairness guarantees.

Contribution

It introduces a generalized equitable coloring framework for weighted graphs, providing bounds, algorithms, and conditions for fairness in coloring and division problems.

Findings

Existence of $(1 + ext{epsilon})$-EQ1 colorings for large enough $k$

Existence of 2-EQ1 colorings for $k \,\geq\, \Delta + 1$

Polynomial-time algorithms for constructing such colorings

Abstract

We study a generalization of the classical Hajnal-Szemer\'edi theorem to vertex-weighted graphs. Given a graph with nonnegative vertex weights, a coloring is called $\alpha$-approximately equitable up to one vertex ($\alpha$-EQ1) if, for each color class, the total weight remaining after removing its maximum-weight vertex is at most $\alpha \geq 1$ times the weight of any other color class. For vertex-weighted graphs with maximum degree $\Delta$, we show that there exist instances for which no $k$-coloring is $\alpha$-EQ1 for any $k < \frac{3\Delta}{2}$ and $\alpha < \sqrt{2}$. In light of this impossibility, we relax these parameters and establish the following results for any vertex-weighted graph $G$ with maximum degree $\Delta$: (1) for any $\varepsilon \in (0,1)$ and all $k \geq (\frac{c}{\varepsilon^2}\ln{\frac{1}{\varepsilon}}) \Delta$, there exists a $(1 + \varepsilon)$-EQ1 $k$-coloring of $G$, where $c$ is a fixed constant; and (2) for all $k \ge \Delta + 1$, there exists a $2$-EQ1 $k$-coloring of $G$. Furthermore, such equitable colorings can be computed in polynomial time. En route to our results on equitability under vertex weights, we establish sufficient conditions for the existence of $k$-colorings that are equitable with respect to any given partition of the vertex set. Our coloring results correspond to fairness guarantees in a constrained fair division setting and lead to concentration inequalities for partly dependent random variables.