A Quotient Property for Matrices with Heavy-Tailed Entries and its Application to Noise-Blind Compressed Sensing

Abstract

For a large class of random matrices $A$ with i.i.d. entries we show that the ${\ell }_1$-quotient property holds with probability exponentially close to 1. In contrast to previous results, our analysis does not require concentration of the entrywise distributions. We provide a unified proof that recovers corresponding previous results for (sub-)Gaussian and Weibull distributions. Our findings generalize known results on the geometry of random polytopes, providing lower bounds on the size of the largest Euclidean ball contained in the centrally symmetric polytope spanned by the columns of $A$. At the same time, our results establish robustness of noise-blind ${\ell }_1$-decoders for recovering sparse vectors $x$ from underdetermined, noisy linear measurements $y=Ax+w$ under the weakest possible assumptions on the entrywise distributions that allow for recovery with optimal sample complexity even in the noiseless case. Our analysis predicts superior robustness behavior for measurement matrices with super-Gaussian entries, which we confirm by numerical experiments.

arXiv comment22 pages, 2 figures

Cited by in corpus (5)

• A Farewell to the Bias-Variance Tradeoff? An Overview of the Theory of Overparameterized Machine Learning
• Estimating the probability that a given vector is in the convex hull of a random sample
• Tight bounds for minimum l1-norm interpolation of noisy data
• Configurations related to combinatorial Veronesians representing a skew perspective
• Random polytopes obtained by matrices with heavy tailed entries