The intersection of two halfspaces has high threshold degree

TL;DR

This paper constructs two halfspaces whose intersection has a high threshold degree, resolving an open problem and impacting learning theory by showing certain techniques cannot be used for intersection problems.

Contribution

It provides the first exponential lower bound on the threshold degree of the intersection of two halfspaces, solving a longstanding open problem and advancing understanding of polynomial representations.

Findings

Constructed two halfspaces with intersection threshold degree Theta(sqrt n)

Proved the intersection of two majority functions has threshold degree Omega(log n)

Improved lower bounds on approximate degree of AND-OR trees

Abstract

The threshold degree of a Boolean function f:{0,1}^n->{-1,+1} is the least degree of a real polynomial p such that f(x)=sgn p(x). We construct two halfspaces on {0,1}^n whose intersection has threshold degree Theta(sqrt n), an exponential improvement on previous lower bounds. This solves an open problem due to Klivans (2002) and rules out the use of perceptron-based techniques for PAC learning the intersection of two halfspaces, a central unresolved challenge in computational learning. We also prove that the intersection of two majority functions has threshold degree Omega(log n), which is tight and settles a conjecture of O'Donnell and Servedio (2003). Our proof consists of two parts. First, we show that for any nonconstant Boolean functions f and g, the intersection f(x)^g(y) has threshold degree O(d) if and only if ||f-F||_infty + ||g-G||_infty < 1 for some rational functions F, G of degree O(d). Second, we settle the least degree required for approximating a halfspace and a majority function to any given accuracy by rational functions. Our technique further allows us to make progress on Aaronson's challenge (2008) and contribute strong direct product theorems for polynomial representations of composed Boolean functions of the form F(f_1,...,f_n). In particular, we give an improved lower bound on the approximate degree of the AND-OR tree.