Sign rank, VC dimension and spectral gaps

We study the maximum possible sign rank of $N \times N$ sign matrices with a given VC dimension $d$. For $d=1$, this maximum is $3$. For $d=2$, this maximum is $\tilde{\Theta}(N^{1/2})$. Similar (slightly less accurate) statements hold for $d>2$ as well. We discuss the tightness of our methods, and describe connections to combinatorics, communication complexity and learning theory.

We also provide explicit examples of matrices with low VC dimension and high sign rank. Let $A$ be the $N \times N$ point-hyperplane incidence matrix of a finite projective geometry with order $n \geq 3$ and dimension $d \geq 2$. The VC dimension of $A$ is $d$, and we prove that its sign rank is larger than $N^{\frac{1}{2}-\frac{1}{2d}}$. The large sign rank of $A$ demonstrates yet another difference between finite and real geometries.

To analyze the sign rank of $A$, we introduce a connection between sign rank and spectral gaps, which may be of independent interest. Consider the $N \times N$ adjacency matrix of a $\Delta$-regular graph with a second eigenvalue (in absolute value) $\lambda$ and $\Delta \leq N/2$. We show that the sign rank of the signed version of this matrix is at least $\Delta/\lambda$. A similar statement holds for all regular (not necessarily symmetric) sign matrices. We also describe limitations of this approach, in the spirit of the Alon-Boppana theorem.

Joint work with Noga Alon and Amir Yehudayoff.