ResearchTrend.AI
  • Papers
  • Communities
  • Events
  • Blog
  • Pricing
Papers
Communities
Social Events
Terms and Conditions
Pricing
Parameter LabParameter LabTwitterGitHubLinkedInBlueskyYoutube

© 2025 ResearchTrend.AI, All rights reserved.

  1. Home
  2. Papers
  3. 1904.08544
79
35
v1v2 (latest)

Memory-Sample Tradeoffs for Linear Regression with Small Error

18 April 2019
Vatsal Sharan
Aaron Sidford
Gregory Valiant
ArXiv (abs)PDFHTML
Abstract

We consider the problem of performing linear regression over a stream of ddd-dimensional examples, and show that any algorithm that uses a subquadratic amount of memory exhibits a slower rate of convergence than can be achieved without memory constraints. Specifically, consider a sequence of labeled examples (a1,b1),(a2,b2)…,(a_1,b_1), (a_2,b_2)\ldots,(a1​,b1​),(a2​,b2​)…, with aia_iai​ drawn independently from a ddd-dimensional isotropic Gaussian, and where bi=⟨ai,x⟩+ηi,b_i = \langle a_i, x\rangle + \eta_i,bi​=⟨ai​,x⟩+ηi​, for a fixed x∈Rdx \in \mathbb{R}^dx∈Rd with ∥x∥2=1\|x\|_2 = 1∥x∥2​=1 and with independent noise ηi\eta_iηi​ drawn uniformly from the interval [−2−d/5,2−d/5].[-2^{-d/5},2^{-d/5}].[−2−d/5,2−d/5]. We show that any algorithm with at most d2/4d^2/4d2/4 bits of memory requires at least Ω(dlog⁡log⁡1ϵ)\Omega(d \log \log \frac{1}{\epsilon})Ω(dloglogϵ1​) samples to approximate xxx to ℓ2\ell_2ℓ2​ error ϵ\epsilonϵ with probability of success at least 2/32/32/3, for ϵ\epsilonϵ sufficiently small as a function of ddd. In contrast, for such ϵ\epsilonϵ, xxx can be recovered to error ϵ\epsilonϵ with probability 1−o(1)1-o(1)1−o(1) with memory O(d2log⁡(1/ϵ))O\left(d^2 \log(1/\epsilon)\right)O(d2log(1/ϵ)) using ddd examples. This represents the first nontrivial lower bounds for regression with super-linear memory, and may open the door for strong memory/sample tradeoffs for continuous optimization.

View on arXiv
Comments on this paper