Tight Bounds on $\ell_1$ Approximation and Learning of Self-Bounding Functions

We study the complexity of learning and approximation of self-bounding functions over the uniform distribution on the Boolean hypercube ${0,1}^n$. Informally, a function $f:{0,1}ⁿ \rightarrow \mathbb{R}$ is self-bounding if for every $x \in {0,1}^n$, $f(x)$ upper bounds the sum of all the $n$ marginal decreases in the value of the function at $x$. Self-bounding functions include such well-known classes of functions as submodular and fractionally-subadditive (XOS) functions. They were introduced by Boucheron et al. (2000) in the context of concentration of measure inequalities. Our main result is a nearly tight $\ell1$-approximation of self-bounding functions by low-degree juntas. Specifically, all self-bounding functions can be $\epsilon$-approximated in $\ell1$ by a polynomial of degree $\tilde{O}(1/\epsilon)$ over $2^{{\tilde{O}(1/\epsilon)}$} variables. We show that both the degree and junta-size are optimal up to logarithmic terms. Previous techniques considered stronger $\ell2$ approximation and proved nearly tight bounds of $\Theta(1/\epsilon^{2})$ on the degree and $2^{{\Theta(1/\epsilon2)}$} on the number of variables. Our bounds rely on the analysis of noise stability of self-bounding functions together with a stronger connection between noise stability and $\ell1$ approximation by low-degree polynomials. This technique can also be used to get tighter bounds on $\ell_1$ approximation by low-degree polynomials and faster learning algorithm for halfspaces. These results lead to improved and in several cases almost tight bounds for PAC and agnostic learning of self-bounding functions relative to the uniform distribution. In particular, assuming hardness of learning juntas, we show that PAC and agnostic learning of self-bounding functions have complexity of $n^{{\tilde{\Theta}(1/\epsilon)}$.}