English

Sparsifying generalized linear models

Data Structures and Algorithms 2023-12-01 v1 Functional Analysis

Abstract

We consider the sparsification of sums F:RnRF : \mathbb{R}^n \to \mathbb{R} where F(x)=f1(a1,x)++fm(am,x)F(x) = f_1(\langle a_1,x\rangle) + \cdots + f_m(\langle a_m,x\rangle) for vectors a1,,amRna_1,\ldots,a_m \in \mathbb{R}^n and functions f1,,fm:RR+f_1,\ldots,f_m : \mathbb{R} \to \mathbb{R}_+. We show that (1+ε)(1+\varepsilon)-approximate sparsifiers of FF with support size nε2(lognε)O(1)\frac{n}{\varepsilon^2} (\log \frac{n}{\varepsilon})^{O(1)} exist whenever the functions f1,,fmf_1,\ldots,f_m are symmetric, monotone, and satisfy natural growth bounds. Additionally, we give efficient algorithms to compute such a sparsifier assuming each fif_i can be evaluated efficiently. Our results generalize the classic case of p\ell_p sparsification, where fi(z)=zpf_i(z) = |z|^p, for p(0,2]p \in (0, 2], and give the first near-linear size sparsifiers in the well-studied setting of the Huber loss function and its generalizations, e.g., fi(z)=min{zp,z2}f_i(z) = \min\{|z|^p, |z|^2\} for 0<p20 < p \leq 2. Our sparsification algorithm can be applied to give near-optimal reductions for optimizing a variety of generalized linear models including p\ell_p regression for p(1,2]p \in (1, 2] to high accuracy, via solving (logn)O(1)(\log n)^{O(1)} sparse regression instances with mn(logn)O(1)m \le n(\log n)^{O(1)}, plus runtime proportional to the number of nonzero entries in the vectors a1,,ama_1, \dots, a_m.

Keywords

Cite

@article{arxiv.2311.18145,
  title  = {Sparsifying generalized linear models},
  author = {Arun Jambulapati and James R. Lee and Yang P. Liu and Aaron Sidford},
  journal= {arXiv preprint arXiv:2311.18145},
  year   = {2023}
}
R2 v1 2026-06-28T13:36:13.770Z