Graph coloring asks: how many colors do you need so that no two adjacent vertices share one? Fractional coloring relaxes the question — vertices can use probability distributions over colors — and the fractional chromatic number is always at most the integral one, often strictly less.
Dhawan's approach is to bound the fractional chromatic number using Shannon entropy. The idea: construct a probability distribution on independent sets, then analyze the entropy of the resulting random coloring. The entropy of the distribution constrains how many “effective colors” are in use, and the structure of the graph constrains which distributions are achievable. The bound emerges from the interplay: graphs that are locally sparse admit high-entropy distributions on large independent sets, and high entropy means few fractional colors suffice.
This recovers and extends classical results. For triangle-free, d-degenerate graphs: fractional chromatic number O(d / log d). For r-uniform hypergraphs with girth at least 4: the bound involves (d / log d)^{1/(r-1)}, generalizing Ajtai, Komlós, and Szemerédi's 1982 independence number result. Both are tight up to constants.
The structural point: coloring is a combinatorial problem, but the tightest bounds come from information theory. The entropy of a random independent set is a proxy for how efficiently a graph can be covered, and the proxy is tight because entropy exactly measures the redundancy in the covering. The technique works as a template — any setting where independent structures can be randomly sampled admits the same entropic analysis. The tool that measures surprise turns out to measure coloring.