Leverage scores: Sensitivity and an App; Ilse Ipsen

Leverages scores were introduced in the 1970s for outlier detection in regression problems. About two decades later, Drineas, Mahoney et al. pioneered the use of leverage scores for importance sampling in randomized matrix algorithms. We present bounds for the sensitivity of leverage scores to perturbations, and introduce a Matlab App, kappa SQ, designed to facilitate analytical and empirical evaluation of leverage scores. With the help of a user-friendly interface, kappa SQ makes it easy to compare a variety of probabilistic bounds (for condition numbers of sampled matrices) and to experiment with different sampling strategies.

This is joint work with Thomas Wentworth.