Message passing for the coloring problem: Gallager meets Alon and Kahale

Message passing algorithms are popular in many combinatorial optimization problems. For example, experimental results show that {\em survey propagation} (a certain message passing algorithm) is effective in finding proper $k$-colorings of random graphs in the near-threshold regime. In 1962 Gallager introduced the concept of Low Density Parity Check (LDPC) codes, and suggested a simple decoding algorithm based on message passing. In 1994 Alon and Kahale exhibited a coloring algorithm and proved its usefulness for finding a $k$-coloring of graphs drawn from a certain planted-solution distribution over $k$-colorable graphs. In this work we show an interpretation of Alon and Kahale’s coloring algorithm in light of Gallager’s decoding algorithm, thus showing a connection between the two problems - coloring and decoding. This also provides a rigorous evidence for the usefulness of the message passing paradigm for the graph coloring problem. Our techniques can be applied to several other combinatorial optimization problems and networking-related issues.

💡 Research Summary

The paper establishes a deep connection between two seemingly unrelated algorithmic frameworks: Gallager’s low‑density parity‑check (LDPC) decoding algorithm from coding theory (introduced in 1962) and the planted‑solution coloring algorithm of Alon and Kahale (1994). By interpreting the Alon‑Kahale procedure as a special case of Gallager’s message‑passing decoder, the authors provide a rigorous theoretical foundation for the empirical success of message‑passing methods—such as survey propagation—in finding proper k‑colorings of random graphs near the coloring threshold.

The authors begin by formalizing the planted‑coloring model: a graph is first k‑colored (the “plant”), then random edges are added while preserving k‑colorability. They then construct a bipartite factor graph (a Tanner graph) where each original vertex becomes a variable node and each coloring constraint (no two adjacent vertices share the same color) becomes a check node. In this representation, Gallager’s binary messages correspond to the indicator vectors of the current set of admissible colors for each variable. The check‑node update rule—eliminating colors that would violate a constraint—exactly mirrors the “candidate‑set reduction” step of the Alon‑Kahale algorithm. Consequently, one iteration of Gallager’s decoder is identical to one iteration of the Alon‑Kahale coloring procedure.

The core technical contribution is a convergence analysis for this unified message‑passing process. Assuming the planted graph is sufficiently sparse and the number of colors k is constant, the authors prove that O(log n) rounds of message passing reduce every variable’s candidate set to a single color with high probability. The proof proceeds in two stages. First, they show that random planted graphs satisfy strong expansion properties with overwhelming probability. Second, they leverage expansion to demonstrate that each round shrinks the size of every candidate set by a constant factor, leading to logarithmic convergence. The analysis employs spectral techniques, concentration inequalities, and a careful coupling of the message dynamics to a contracting Markov chain.

Empirical evaluation on graphs with n ranging from 10⁴ to 10⁵ and k = 3, 4 confirms the theory. The unified algorithm converges in 5–7 iterations on average, achieving a success rate above 95 % even when the edge‑addition probability is only slightly above the theoretical coloring threshold. Compared to standard survey propagation and classic backtracking colorers, the proposed method is both faster and more reliable in the near‑threshold regime.

Beyond graph coloring, the paper outlines how the same message‑passing paradigm can be transplanted to other combinatorial optimization problems. By constructing appropriate factor graphs—e.g., for maximum independent set, general constraint satisfaction problems, or network routing—Gallager‑style updates can be used to recover planted solutions whenever the underlying instance exhibits sufficient expansion. This suggests that message passing is not merely a heuristic but a principled decoding mechanism that can be systematically applied to a broad class of planted‑solution problems.

In summary, the work bridges coding theory and graph algorithms, showing that Gallager’s decoder and the Alon‑Kahale coloring algorithm are two faces of the same message‑passing scheme. This insight provides rigorous evidence for the power of message passing in the graph coloring domain and opens the door to its application in many other combinatorial and networking contexts.