Improved Certificates for Independence Number in Semirandom Hypergraphs

TL;DR

This paper develops sharper, nearly optimal certificates for bounding the independence number in semirandom hypergraphs, advancing spectral and Sum-of-Squares methods and removing logarithmic factors in bounds.

Contribution

It introduces improved bounds that eliminate logarithmic factors and nearly reach the conjectured optimal thresholds, along with robust Sum-of-Squares certificates for semirandom hypergraphs.

Findings

Sharper bounds without logarithmic factors

Nearly optimal computational thresholds achieved

Robust SoS certificates for semirandom models

Abstract

We study the problem of efficiently certifying upper bounds on the independence number of $\ell$-uniform hypergraphs. This is a notoriously hard problem, with efficient algorithms failing to approximate the independence number within $n^{1-\epsilon}$ factor in the worst case [Has99, Zuc07]. We study the problem in random and semirandom hypergraphs. There is a folklore reduction to the graph case, achieving a certifiable bound of $O(\sqrt{n/p})$. More recently, the work [GKM22] improved this by constructing spectral certificates that yield a bound of $O(\sqrt{n}.\mathrm{polylog}(n)/p^{1/(\ell/2)})$. We make two key improvements: firstly, we prove sharper bounds that get rid of pesky logarithmic factors in $n$, and nearly attain the conjectured optimal (in both $n$ and $p$) computational threshold of $O(\sqrt{n}/p^{1/\ell})$, and secondly, we design robust Sum-of-Squares (SoS) certificates, proving our bounds in the more challenging semirandom hypergraph model. Our analysis employs the proofs-to-algorithms paradigm [BS16, FKP19] in showing an upper bound for pseudo-expectation of degree-$2\ell$ SoS relaxation of the natural polynomial system for maximum independent set. The challenging case is odd-arity hypergraphs, where we employ a tensor-based analysis that reduces the problem to proving bounds on a natural class of random chaos matrices associated with $\ell$-uniform hypergraphs. Previous bounds [AMP21, RT23] have a logarithmic dependence, which we remove by leveraging recent progress on matrix concentration inequalities [BBvH23, BLNvH25]; we believe these may be useful in other hypergraph problems. As an application, we show our improved certificates can be combined with an SoS relaxation of a natural $r$-coloring polynomial system to recover an arbitrary planted $r$-colorable subhypergraph in a semirandom model along the lines of [LPR25], which allows for strong adversaries.