Factorization norms and Zarankiewicz problems
István Tomon
TL;DR
This work studies the γ2-norm (max-norm) of Boolean matrices, linking spectral decompositions to combinatorial structure. It establishes that four-cycle-free Boolean matrices with degeneracy d satisfy γ2(M) = Θ(√d), enabling degree-boundedness results: matrices with bounded γ2-norm and no large all-ones submatrices have few ones, and various decompositions (into thin blocky matrices) are equivalent to bounded degeneracy. These insights yield Zarankiewicz-type incidence bounds for point–polytope configurations, showing that K_{t,t}-free incidence graphs between n points and polytopes from POL(H) have edges bounded by O_{s,t}(n (log n)^{O(d)}) in semilinear settings, with the log-exponent depending only on dimension. The paper also connects to discrepancy theory via the Matoušek–Nikolov–Talwar inequality, demonstrating that γ2 controls hereditary discrepancy up to logarithmic factors for four-cycle-free matrices and recovers Szemerédi–Trotter-type discrepancy in geometric set systems. Collectively, these results advance the understanding of how bounded γ2-norms constrain combinatorial density and have concrete implications for Zarankiewicz problems and incidence geometry in semilinear and geometric contexts.
Abstract
The $γ_2$-norm of Boolean matrices plays an important role in communication complexity and discrepancy theory. In this paper, we study combinatorial properties of this norm, and provide new applications, involving Zarankiewicz type problems. We show that if $M$ is an $m\times n$ Boolean matrix such that $γ_2(M)<γ$ and $M$ contains no $t\times t$ all-ones submatrix, then $M$ contains $O_{γ,t}(m+n)$ one entries. In other words, graphs of bounded $γ_2$-norm are degree bounded. This addresses a conjecture of Hambardzumyan, Hatami, and Hatami for locally sparse matrices. We prove that if $G$ is a $K_{t,t}$-free incidence graph of $n$ points and $n$ homothets of a polytope $P$ in $\mathbb{R}^d$, then the average degree of $G$ is $O_{d,P}(t(\log n)^{O(d)})$. This is sharp up the $O(.)$ notations. In particular, we prove a more general result on semilinear graphs, which greatly strengthens the work of Basit, Chernikov, Starchenko, Tao, and Tran.
