English

Fast Algorithms for $\ell_p$-Regression

Data Structures and Algorithms 2023-10-10 v2 Optimization and Control

Abstract

The p\ell_p-norm regression problem is a classic problem in optimization with wide ranging applications in machine learning and theoretical computer science. The goal is to compute x=argminAx=bxppx^{\star} =\arg\min_{Ax=b}\|x\|_p^p, where xRn,ARd×n,bRdx^{\star}\in \mathbb{R}^n, A\in \mathbb{R}^{d\times n},b \in \mathbb{R}^d and dnd\leq n. Efficient high-accuracy algorithms for the problem have been challenging both in theory and practice and the state of the art algorithms require poly(p)n121ppoly(p)\cdot n^{\frac{1}{2}-\frac{1}{p}} linear system solves for p2p\geq 2. In this paper, we provide new algorithms for p\ell_p-regression (and a more general formulation of the problem) that obtain a high-accuracy solution in O(pn(p2)(3p2))O(p n^{\frac{(p-2)}{(3p-2)}}) linear system solves. We further propose a new inverse maintenance procedure that speeds-up our algorithm to O~(nω)\widetilde{O}(n^{\omega}) total runtime, where O(nω)O(n^{\omega}) denotes the running time for multiplying n×nn \times n matrices. Additionally, we give the first Iteratively Reweighted Least Squares (IRLS) algorithm that is guaranteed to converge to an optimum in a few iterations. Our IRLS algorithm has shown exceptional practical performance, beating the currently available implementations in MATLAB/CVX by 10-50x.

Keywords

Cite

@article{arxiv.2211.03963,
  title  = {Fast Algorithms for $\ell_p$-Regression},
  author = {Deeksha Adil and Rasmus Kyng and Richard Peng and Sushant Sachdeva},
  journal= {arXiv preprint arXiv:2211.03963},
  year   = {2023}
}

Comments

This paper is a coherent algorithmic framework that combines and simplifies our previous works: 1. arXiv:1901.06764 2. arXiv:1907.07167 3. arXiv:1910.10571

R2 v1 2026-06-28T05:23:19.448Z