Criticality of $\text{AC}^0$ formulae

TL;DR

This paper proves a tight bound on the criticality of any AC^0 formula, resolving a conjecture and leading to improved bounds on correlation, Fourier concentration, and SAT algorithms for AC^0 formulas.

Contribution

It establishes the criticality bound for all AC^0 formulas, not just regular ones, confirming Rossman's conjecture and strengthening previous results.

Findings

Criticality of any AC^0 formula is at most O((log S)/d)^d.

Derived tight correlation bounds against parity for AC^0 formulas.

Improved algorithms for SAT problems on AC^0 formulas.

Abstract

Rossman [In $\textit{Proc. $34$th Comput. Complexity Conf.}$, 2019] introduced the notion of $\textit{criticality}$. The criticality of a Boolean function $f : \{0,1\}^n \to \{0,1\}$ is the minimum $\lambda \geq 1$ such that for all positive integers $t$, \[ \Pr_{\rho \sim \mathcal{R}_p}\left[\text{DT}_{\text{depth}}(f|_{\rho}) \geq t\right] \leq (p\lambda)^t. \] H\"astad's celebrated switching lemma shows that the criticality of any $k$-DNF is at most $O(k)$. Subsequent improvements to correlation bounds of $\text{AC}^0$-circuits against parity showed that the criticality of any $\text{AC}^0$-$\textit{circuit}$ of size $S$ and depth $d+1$ is at most $O(\log S)^d$ and any $\textit{regular}$ $\text{AC}^0$-$\textit{formula}$ of size $S$ and depth $d+1$ is at most $O\left(\frac1d \cdot \log S\right)^d$. We strengthen these results by showing that the criticality of $\textit{any}$ $\text{AC}^0$-formula (not necessarily regular) of size $S$ and depth $d+1$ is at most $O\left(\frac1d\cdot {\log S}\right)^d$, resolving a conjecture due to Rossman. This result also implies Rossman's optimal lower bound on the size of any depth-$d$ $\text{AC}^0$-formula computing parity [$\textit{Comput. Complexity, 27(2):209--223, 2018.}$]. Our result implies tight correlation bounds against parity, tight Fourier concentration results and improved $\#$SAT algorithm for $\text{AC}^0$-formulae.