Core Concepts
This paper presents a simple method to compute two-sided ε-approximations of the ℓp-Lewis weights of an n × d matrix using only poly(d, 1/ε) many poly(1/d, ε)-approximate leverage score computations.
Abstract
The paper focuses on efficiently computing approximate ℓp-Lewis weights, a generalization of leverage scores that measure the importance of rows with respect to the ℓp norm.
Key highlights:
Lewis weights lack an explicit description and are more challenging to compute than leverage scores, especially for p ≥ 4.
Prior work has established algorithms for computing ε-estimates of Lewis weights, but they require precision inversely polynomial in n, which can be prohibitive.
The authors show that by running a simple "fixed point iteration" on a one-sided ε-approximate Lewis weight, one can obtain a two-sided O(εpd)-approximation using only poly(d, 1/ε) many poly(1/d, ε)-approximate leverage score computations.
This effectively also yields a O(p^3 d^3/2 ε)-estimate of the true Lewis weights.
The authors provide an algorithm that combines their result with a previous one-sided approximation algorithm, allowing for the computation of two-sided Lewis weight approximations from low-precision leverage scores.