Local convergence of random graph colorings

Let $G=G(n,m)$ be a random graph whose average degree $d=2m/n$ is below the $k$-colorability threshold. If we sample a $k$-coloring $\sigma$ of $G$ uniformly at random, what can we say about the correlations between the colors assigned to vertices that are far apart? According to a prediction from statistical physics, for average degrees below the so-called {\em condensation threshold} $d_c(k)$, the colors assigned to far away vertices are asymptotically independent [Krzakala et al.: Proc. National Academy of Sciences 2007]. We prove this conjecture for $k$ exceeding a certain constant $k_0$. More generally, we investigate the joint distribution of the $k$-colorings that $\sigma$ induces locally on the bounded-depth neighborhoods of any fixed number of vertices. In addition, we point out an implication on the reconstruction problem.

💡 Research Summary

The paper investigates the relationship between local and global properties of random graph colorings, focusing on the regime where the average degree d of the Erdős–Rényi graph G(n,m) lies below the k‑colorability threshold and, more precisely, below the so‑called condensation threshold d_c(k). The authors consider a uniformly random proper k‑coloring σ of G and ask whether the colors assigned to vertices that are far apart become asymptotically independent. This question originates from a non‑rigorous prediction in statistical physics that such independence should hold for d < d_c(k). Prior rigorous results only covered a more restrictive range d < 2(k‑1)ln(k‑1), which is roughly an additive ln k below d_c(k).

The main contribution is a proof of this independence for all sufficiently large numbers of colors k ≥ k₀ (where k₀ is a universal constant) and any average degree d < d_c(k). The authors establish that for any fixed number ℓ of vertices v₁,…,v_ℓ, the joint distribution of the colors on the depth‑ω neighborhoods ∂_ω(G,v_i) induced by a random coloring σ is asymptotically the product of the uniform distributions on each neighborhood. In other words, the total variation distance between the induced distribution and the product of the marginal uniform distributions is o(1) with high probability. This result is formalized as Theorem 1.1 and its corollaries.

To achieve this, the paper employs the framework of local weak convergence. The depth‑ω neighborhood of a typical vertex in G(n,m) is, with high probability, a tree that can be coupled with a Poisson(d) Galton–Watson tree T(d). The authors define a random coloring process on T(d): the root receives a uniformly random color, and each child independently chooses a color different from its parent. By considering ℓ independent copies of T(d) and the product measure on their rooted colored trees, they construct a limiting probability measure ϑ_{ℓ,d,k}. They then prove that the empirical distribution of local colorings in G(n,m) (denoted λ_{ℓ,n,m,k}) converges in probability to ϑ_{ℓ,d,k} as n → ∞. This convergence implies the asymptotic independence of distant colors.

A crucial technical ingredient is a recent concentration result for the number Z_k(G) of proper k‑colorings of G(n,m). Using this, the authors develop an enhanced “planting trick” called the planted replica model, which links the joint law of two independent random colorings of G to a much simpler distribution. An analysis of the overlap between two random colorings shows that, under the condensation threshold, the overlap concentrates around its expectation, which in turn yields the desired local product structure.

The paper also addresses the reconstruction problem. For a vertex v and a boundary condition given by the colors outside a radius ω, the bias measures how much the color of v is influenced by that boundary. The authors prove (Corollary 1.4) that reconstruction on G(n,m) occurs if and only if reconstruction occurs on the corresponding Galton–Watson tree T(d). This bridges a complex global problem on random graphs to a tractable tree problem. Existing results on tree reconstruction give that non‑reconstruction holds for d < (1−δ_k)k ln k, while reconstruction holds for d > (1+δ_k)k ln k, with δ_k → 0. Hence the paper pinpoints the exact regime (up to the unknown constant k₀) where the random graph exhibits non‑reconstruction.

Overall, the work provides a rigorous confirmation of the physics prediction about asymptotic independence of distant colors below the condensation threshold, extends the known range of parameters for which this holds, and introduces methodological tools—particularly the planted replica model and overlap analysis—that are likely to be useful for other random constraint satisfaction problems. The results have implications for algorithmic sampling, understanding of phase transitions in random graphs, and the broader theory of probabilistic combinatorics.