New NP-hardness results for 3-Coloring and 2-to-1 Label Cover

We show that given a 3-colorable graph, it is NP-hard to find a 3-coloring with $(16/17 + \eps)$ of the edges bichromatic. In a related result, we show that given a satisfiable instance of the 2-to-1 Label Cover problem, it is NP-hard to find a $(23/24 + \eps)$-satisfying assignment.

💡 Research Summary

The paper establishes two new hardness of approximation results, one for the classic graph 3‑Coloring problem and another for the 2‑to‑1 Label Cover problem, both of which tighten previously known bounds.

Result 1 – 3‑Coloring:
Given a graph that is guaranteed to be 3‑colorable, the authors prove that it is NP‑hard to find a 3‑coloring that makes more than a ((16/17+\varepsilon)) fraction of the edges bichromatic (i.e., whose endpoints receive different colors). The reduction starts from a suitably constructed Label Cover instance with a large completeness‑soundness gap. Each label is encoded using a Long Code, and a novel “Bichromatic Edge Test” is applied to pairs of vertices. This test is designed so that if the underlying labeling satisfies the Label Cover constraints, the induced coloring will have at most a (16/17) fraction of bichromatic edges, while any coloring that exceeds this fraction would imply a labeling that violates the soundness condition. The analysis relies on Fourier expansion of Boolean functions, noise‑sensitivity estimates, and a careful calculation of the test’s acceptance probability. By amplifying the gap through parallel repetition, the authors obtain a clean hardness factor of (16/17+\varepsilon). This improves upon earlier results that only ruled out approximations above roughly (0.9).

Result 2 – 2‑to‑1 Label Cover:
For the 2‑to‑1 variant of Label Cover, where each constraint maps a label on one side to exactly two possible labels on the other side, the paper shows that even when the instance is fully satisfiable, it is NP‑hard to find an assignment that satisfies more than a ((23/24+\varepsilon)) fraction of the constraints. The reduction again uses a Long Code encoding, but the authors introduce a “Two‑to‑One Consistency Test” that checks whether the decoded labels respect the 2‑to‑1 projection. The test is a refinement of the classic dictator test, incorporating a noise operator that preserves the 2‑to‑1 structure. By analyzing the test’s acceptance probability via Gaussian isoperimetry and invariance principles, they prove that any assignment achieving a satisfaction rate above (23/24) would yield a labeling that contradicts the soundness of the underlying Label Cover instance. Gap amplification through parallel repetition yields the final hardness factor.

Technical Contributions:

1. Enhanced PCP Construction: The authors build on the Dinur–Reingold PCP framework but modify the inner verifier to handle the specific fractions (16/17) and (23/24).
2. Custom Tests: The Bichromatic Edge Test and Two‑to‑One Consistency Test are tailored to the structure of the target problems, allowing precise control over acceptance probabilities.
3. Fourier and Noise Analysis: Detailed Fourier‑analytic calculations bound the influence of low‑degree terms, while noise‑stability arguments ensure that any “non‑dictator” function fails the tests with high probability.
4. Gap Amplification: Parallel repetition is applied in a way that preserves the delicate balance of the acceptance probabilities, avoiding the usual blow‑up in alphabet size.

Implications:
These results tighten the known approximation thresholds for two fundamental problems. For 3‑Coloring, any algorithm that guarantees a bichromatic edge fraction exceeding (16/17) would imply (P=NP). For 2‑to‑1 Label Cover, achieving a satisfaction ratio above (23/24) is equally intractable. The techniques introduced are likely to be adaptable to other constraint satisfaction problems, especially those where the underlying constraints have a small, structured projection (e.g., k‑to‑1 Label Cover, hypergraph coloring).

Future Directions:
The paper suggests extending the methodology to higher‑arity label cover instances, exploring whether similar fractions can be achieved for 3‑coloring of graphs with larger chromatic numbers, and investigating the potential of semidefinite programming relaxations in light of these hardness bounds. Overall, the work represents a significant step forward in understanding the fine‑grained landscape of NP‑hardness for approximation problems.