[Paper Review] Reconstruction of colourings without freezing

Allan Sly, Yumeng Zhang|arXiv (Cornell University)|Oct 10, 2016

Digital Image Processing Techniques5 references3 citations

TL;DR

This paper proves that reconstruction in the k-colouring model on trees occurs strictly below the freezing threshold for large k, demonstrating that information about the root spin can be recovered even when the root is not uniquely determined by the leaves. Using a refined analysis of message-passing dynamics and total variation distance, the authors close the gap between reconstruction and freezing thresholds, showing a non-vanishing region of reconstruction below freezing where clusters of colourings remain unfrozen.

ABSTRACT

We prove that reconstruction in the $k$-colouring model occurs strictly below the threshold for freezing for large $k$.

Motivation & Objective

• To resolve the long-standing open question of whether reconstruction is possible strictly below the freezing threshold in the k-colouring model for large k.
• To clarify the relationship between clustering and freezing in random constraint satisfaction problems, particularly in random graphs and tree models.
• To close the gap between the known reconstruction and freezing thresholds in the k-colouring model, which previously differed by a term of order k log 2.
• To demonstrate that clusters of solutions in random k-colourings can remain unfrozen (i.e., all vertices can take all colours within a cluster) in a non-vanishing region below freezing.

Proposed method

• The authors analyze the k-colouring model on Galton-Watson trees with Poisson or d-regular offspring distributions, using the Gibbs measure and broadcast process on trees.
• They define reconstruction via the total variation distance between root spin distributions conditioned on different root values at distant levels.
• A key technical tool is a recursive message-passing analysis using belief propagation and the study of invariant measures on the space of configurations.
• They introduce a refined coupling argument and use concentration inequalities to bound the decay of influence from the leaves to the root.
• The proof involves constructing a coupling between the original and perturbed measures, showing that the influence of the root on distant leaves persists below the freezing threshold.
• They establish a non-trivial lower bound on the reconstruction threshold by analyzing the decay of the total variation distance and using asymptotic expansions in k.

Experimental results

Research questions

• RQ1 Is reconstruction possible in the k-colouring model when the root is not frozen, i.e., below the freezing threshold?
• RQ2 What is the exact location of the reconstruction threshold relative to the freezing threshold for large k?
• RQ3 Does the presence of clustering in random k-colourings imply computational hardness, or is it the onset of freezing that causes the barrier?
• RQ4 Can the reconstruction threshold be strictly below the freezing threshold, and if so, by how much?
• RQ5 What is the structural nature of solution clusters in random graphs when reconstruction is possible but freezing is not?

Key findings

• Reconstruction is possible in the k-colouring model strictly below the freezing threshold for large k, resolving a key open question in spin systems and constraint satisfaction problems.
• The reconstruction threshold lies below the freezing threshold by a region of width Ω(k), specifically by a term of order k log 2, closing the previously known gap.
• The authors prove that for d-regular trees, reconstruction holds when d < k(log k + log log k + 1 - log 2 + o_k(1)), which is strictly below the freezing threshold.
• In the context of random graphs, this implies a range of parameters where solution clusters exist but are unfrozen—meaning all vertices can take all colours within a cluster.
• The result separates the dynamical phase transition (clustering) from the freezing transition, suggesting that freezing, not clustering, may be the true source of computational hardness.
• The proof establishes a non-vanishing influence of the root on distant leaves even when the root is not uniquely determined, via a refined analysis of total variation distance decay.