Biased halfspaces, noise sensitivity, and relative Chernoff inequalities (extended version)
share
In analysis of Boolean functions, a halfspace is a function f:{-1,1}^n →{0,1} of the form f(x)=1_{a· x>t}, where ∑_i a_i^2=1. We show that if f is a halfspace with E[f]=ϵ, then the degree-1 Fourier weight of f is W^1(f)=Θ(ϵ^2 (1/ϵ)), and the maximal influence of f is I_(f)=Θ(ϵ(1,a' √((1/ϵ)))), where a'=_i |a_i|. These results, which determine the exact asymptotic order of W^1(f) and I_(f), provide sharp generalizations of theorems proved by Matulef, O'Donnell, Rubinfeld, and Servedio, and settle a conjecture posed by Kalai, Keller and Mossel. In addition, we present a refinement of the definition of noise sensitivity which takes into consideration the bias of the function, and show that (like in the unbiased case) halfspaces are noise resistant, and on the other hand, any noise resistant function is well-correlated to a halfspace. Our main tools are `relative' forms of the classical Chernoff inequality, like the following one proved by Devroye and Lugosi (2008): Let {x_i} be independent random variables uniformly distributed in {-1,1}, and let a_i ∈R_≥ 0 be such that ∑_i a_i^2=1. If for some t≥ 0 we have Pr[∑_i a_i x_i > t]=ϵ, then Pr[∑_i a_i x_i>t+δ]≤ϵ/2 holds for δ≤ c/√((1/ϵ)), where c is a universal constant.
READ FULL TEXT
