Abstract:We study the statistical resilience of the anti-concentration properties of Rademacher polynomials in face of adversarial deterministic noise taking the form of sign-flips. Given a multilinear polynomial $f:\mathbb{R}^n \to \mathbb{R}$ and a Rademacher vector $\boldsymbol{\xi} \in \{\pm 1\}^n$ (with independent entries), our results provide probabilistic lower bound estimations on the number of sign-flips that $\boldsymbol{\xi}$ can sustain without ``inflating" the atom probability $\sup_{x \in \mathbb{R} } \mathbb{P}\{f(\boldsymbol{\xi}) = x\}$ otherwise resulting in an adversarially biased distribution. Special emphasis is put on bilinear and quadratic forms, for which strengthened estimates are attained. From a computational perspective, our results in this venue are instance-bound in such a way that allows for an efficient computation of the statistical resilience guarantees from the quadratic polynomial itself directly. All of our probabilistic lower bound resilience guarantees are asymptotically tight. On route, we provide a short proof for a new small-ball probability estimate fitting Rademacher multilinear polynomials $f: \mathbb{R}^n \to \mathbb{R}$ removeing a polylog-factor from the classical Meka-Nguyen-Vu bound provided the coefficients are independent of $n$ (dimension-free, hereafter). This removal was conjectured to be possible by Meka-Nguyen-Vu regardless of our assumption. Bilinear Rademacher forms with dimension-free coefficients arise naturally in Combinatorics and specifically in the dense case of the edge-statistics conjecture posed by Alon, Hefetz, Krivelevich, and Tyomkyn. This case of the conjecture was resolved by Kwan and Sauermann. Replacing the appeal to the Meka-Nguyen-Vu classical bound in the work of Kwan, Sudakov, and Tran with our shortly proved result attains an additional proof of the dense case of the edge-statistics conjecture.

Super Non-singular Decompositions of Polynomials and their Application to Robustly Learning Low-degree PTFs

Attribute-Efficient PAC Learning of Low-Degree Polynomial Threshold Functions with Nasty Noise

Active Learning Polynomial Threshold Functions

Testably Learning Polynomial Threshold Functions

Learning Constant-Depth Circuits in Malicious Noise Models

Adversarially Robust PAC Learnability of Real-Valued Functions

On Robustness to Adversarial Examples and Polynomial Optimization

Polynomial Threshold Functions: Structure, Approximation and Pseudorandomness

Robust learning under clean-label attack

Polynomial Bounds for Learning Noisy Optical Physical Unclonable Functions and Connections to Learning With Errors

Efficient PAC Learning of Halfspaces with Constant Malicious Noise Rate

Low-degree learning and the metric entropy of polynomials

Optimal Robust Estimation under Local and Global Corruptions: Stronger Adversary and Smaller Error

Iterative thresholding for non-linear learning in the strong $\varepsilon$-contamination model

On the Computability of Robust PAC Learning

The Curse of Concentration in Robust Learning: Evasion and Poisoning Attacks from Concentration of Measure

A Combinatorial Approach to Robust PCA

Enhancing PAC Learning of Half spaces Through Robust Optimization Techniques

Robust Estimators in High Dimensions without the Computational Intractability

The Full Landscape of Robust Mean Testing: Sharp Separations between Oblivious and Adaptive Contamination

On the resilience of the quadratic Littlewood-Offord problem