Pairwise independent correlation gap conjecture

For every integer n≥1n\ge 1, every monotone submodular function f:2[n]→R≥0f:2^{[n]}\to\mathbb{R}_{\ge 0}, and every marginal vector x∈[0,1]nx\in[0,1]^n with f++(x)>0f^{++}(x)>0, the correlation gap under pairwise independence is at most 4/34/3: f+(x)f++(x)≤43\frac{f^{+}(x)}{f^{++}(x)}\le\frac{4}{3}. Here f+(x)=max⁡{E[f(S)]:S={i:Xi=1}, X∈{0,1}n, E[Xi]=xi}f^{+}(x)=\max\{\mathbb{E}[f(S)]:S=\{i:X_i=1\},\ X\in\{0,1\}^n,\ \mathbb{E}[X_i]=x_i\}, while f++(x)f^{++}(x) is the same maximum restricted to jointly distributed random variables X1,…,XnX_1,\ldots,X_n that are pairwise independent.

References

Progress summary

Refreshed
Claimed solved

A new preprint claims to settle the four-variable case and the limiting worst case, while the earlier five-variable counterexample remains valid.

Ramachandra and Natarajan (2025) conjectured that the correlation-gap ratio is at most 4/34/3 for every dimension. A June 2026 preprint disproved this for n≥5n\geq 5, leaving n=4n=4 open until the latest claimed development.

Known results

  • n=2n=2 and n=3n=3: the 4/34/3 bound was known to hold.
  • The bound was also known for certain marginal probabilities and specific submodular functions.
  • For n≥5n\geq 5, a coverage-function example gives f+(x)/f++(x)≥640/479>4/3f^{+}(x)/f^{++}(x)\geq 640/479>4/3; the example was found with assistance from GPT5.5 Pro.

September 2, 2026 claimed resolution

A new preprint claims the 4/34/3 bound for n=4n=4, proves that bound is tight, and constructs asymptotic examples attaining e/(e−1)e/(e-1), with an extension to tt-wise independence. This is a claimed resolution, not yet independently verified.

Current status (as of September 2026): The n≥5n\geq 5 counterexample and the cases n=2,3n=2,3 are reported, while the n=4n=4 proof, its tightness, and the asymptotic constant e/(e−1)e/(e-1) are claimed by a new preprint but remain unverified.

Sources

Solutions 0

No solutions have been posted yet.