NeurIPS2020

Acceleration with a Ball Optimization Oracle

Yair Carmon, Arun Jambulapati, Qijia Jiang, Yujia Jin, Yin Tat Lee, Aaron Sidford, Kevin Tian

51 citations

Abstract

Consider an oracle which takes a point xx and returns the minimizer of a convex function ff in an 2\ell_2 ball of radius rr around xx. It is straightforward to show that roughly r1log1ϵr^{-1}\log\frac{1}{\epsilon} calls to the oracle suffice to find an ϵ\epsilon-approximate minimizer of ff in an 2\ell_2 unit ball. Perhaps surprisingly, this is not optimal: we design an accelerated algorithm which attains an ϵ\epsilon-approximate minimizer with roughly r2/3log1ϵr^{-2/3} \log \frac{1}{\epsilon} oracle queries, and give a matching lower bound. Further, we implement ball optimization oracles for functions with locally stable Hessians using a variant of Newton's method. The resulting algorithm applies to a number of problems of practical and theoretical import, improving upon previous results for logistic and \ell_\infty regression and achieving guarantees comparable to the state-of-the-art for p\ell_p regression.