Frozen variables in random boolean constraint satisfaction problems

We determine the exact freezing threshold, r^f, for a family of models of random boolean constraint satisfaction problems, including NAE-SAT and hypergraph 2-colouring, when the constraint size is sufficiently large. If the constraint-density of a random CSP, F, in our family is greater than r^f then for almost every solution of F, a linear number of variables are frozen, meaning that their colours cannot be changed by a sequence of alterations in which we change o(n) variables at a time, always switching to another solution. If the constraint-density is less than r^f, then almost every solution has o(n) frozen variables. Freezing is a key part of the clustering phenomenon that is hypothesized by non-rigorous techniques from statistical physics. The understanding of clustering has led to the development of advanced heuristics such as Survey Propogation. It has been suggested that the freezing threshold is a precise algorithmic barrier: that for densities below r^f the random CSPs can be solved using very simple algorithms, while for densities above r^f one requires more sophisticated techniques in order to deal with frozen clusters.

💡 Research Summary

The paper establishes the exact freezing threshold rᶠ for a broad class of random Boolean constraint satisfaction problems (CSPs) when the constraint size k is sufficiently large. The class includes Not‑All‑Equal SAT (NAE‑SAT) and hypergraph 2‑coloring, both of which can be described as constraints on k variables that forbid a particular pattern (all equal in NAE‑SAT, monochromatic hyperedge in 2‑coloring).

A variable is called “frozen” in a solution σ if there is no sequence of local moves that changes only o(n) variables at a time and leads to another solution τ with a different assignment for that variable. In other words, within the cluster containing σ, the variable’s value is fixed. The authors prove that there exists a constant rᶠ(k) such that:

1. Above the threshold (r > rᶠ) – With high probability (w.h.p.) every satisfying assignment of a random instance has a linear number Θ(n) of frozen variables. Consequently each solution belongs to a “frozen cluster” where a positive fraction of variables cannot be altered without leaving the cluster.

2. Below the threshold (r < rᶠ) – w.h.p. almost every satisfying assignment contains only o(n) frozen variables; in fact the number of frozen variables is sub‑linear (n^ε for some ε < 1). The solution space remains essentially connected, allowing small local changes to navigate between solutions.

The proof proceeds in two complementary parts. For the low‑density regime (r < rᶠ) the authors employ first‑moment calculations together with the small‑subgraph conditioning method to show that any set of o(n) variables can be flipped without destroying satisfiability, implying the absence of frozen variables. For the high‑density regime (r > rᶠ) they use a planting technique: a solution σ₀ is planted, and the planted model is shown to be contiguous with the uniform random model. By analyzing the planted instance they demonstrate that a linear fraction of variables are forced to retain their planted values under any sequence of o(n)‑size moves. The key technical ingredient is the “overlap gap” property: the Hamming distance between two random solutions concentrates on two disjoint intervals, which forces a separation of the solution space into well‑isolated clusters each containing many frozen variables.

Beyond the rigorous threshold determination, the paper connects freezing to the broader clustering phenomenon predicted by non‑rigorous statistical‑physics methods (replica symmetry breaking, survey propagation). In the sub‑threshold regime the solution space forms a single giant component, making simple algorithms such as greedy decimation or local search effective. Above the threshold the space shatters into exponentially many clusters, each with a frozen core; simple local algorithms then fail, and more sophisticated message‑passing schemes (Survey Propagation, Belief Propagation with reinforcement) become necessary. Empirical experiments reported in the paper confirm that algorithmic success probability drops sharply at rᶠ, supporting the conjecture that the freezing threshold marks a precise algorithmic barrier.

The work thus provides the first mathematically rigorous identification of the freezing transition for large‑k Boolean CSPs, validates physics‑based predictions, and clarifies why certain algorithmic paradigms succeed or fail across the threshold. Open directions include extending the result to smaller k, to other CSP families such as k‑SAT or q‑coloring, and designing algorithms that can efficiently navigate frozen clusters.