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. 2311.10680
97
14
v1v2 (latest)

Optimal Embedding Dimension for Sparse Subspace Embeddings

17 November 2023
Shabarish Chenakkod
Michal Dereziñski
Xiaoyu Dong
M. Rudelson
ArXiv (abs)PDFHTML
Abstract

A random m×nm\times nm×n matrix SSS is an oblivious subspace embedding (OSE) with parameters ϵ>0\epsilon>0ϵ>0, δ∈(0,1/3)\delta\in(0,1/3)δ∈(0,1/3) and d≤m≤nd\leq m\leq nd≤m≤n, if for any ddd-dimensional subspace W⊆RnW\subseteq R^nW⊆Rn, P( ∀x∈W (1+ϵ)−1∥x∥≤∥Sx∥≤(1+ϵ)∥x∥ )≥1−δ.P\big(\,\forall_{x\in W}\ (1+\epsilon)^{-1}\|x\|\leq\|Sx\|\leq (1+\epsilon)\|x\|\,\big)\geq 1-\delta.P(∀x∈W​ (1+ϵ)−1∥x∥≤∥Sx∥≤(1+ϵ)∥x∥)≥1−δ. It is known that the embedding dimension of an OSE must satisfy m≥dm\geq dm≥d, and for any θ>0\theta > 0θ>0, a Gaussian embedding matrix with m≥(1+θ)dm\geq (1+\theta) dm≥(1+θ)d is an OSE with ϵ=Oθ(1)\epsilon = O_\theta(1)ϵ=Oθ​(1). However, such optimal embedding dimension is not known for other embeddings. Of particular interest are sparse OSEs, having s≪ms\ll ms≪m non-zeros per column, with applications to problems such as least squares regression and low-rank approximation. We show that, given any θ>0\theta > 0θ>0, an m×nm\times nm×n random matrix SSS with m≥(1+θ)dm\geq (1+\theta)dm≥(1+θ)d consisting of randomly sparsified ±1/s\pm1/\sqrt s±1/s​ entries and having s=O(log⁡4(d))s= O(\log^4(d))s=O(log4(d)) non-zeros per column, is an oblivious subspace embedding with ϵ=Oθ(1)\epsilon = O_{\theta}(1)ϵ=Oθ​(1). Our result addresses the main open question posed by Nelson and Nguyen (FOCS 2013), who conjectured that sparse OSEs can achieve m=O(d)m=O(d)m=O(d) embedding dimension, and it improves on m=O(dlog⁡(d))m=O(d\log(d))m=O(dlog(d)) shown by Cohen (SODA 2016). We use this to construct the first oblivious subspace embedding with O(d)O(d)O(d) embedding dimension that can be applied faster than current matrix multiplication time, and to obtain an optimal single-pass algorithm for least squares regression. We further extend our results to Leverage Score Sparsification (LESS), which is a recently introduced non-oblivious embedding technique. We use LESS to construct the first subspace embedding with low distortion ϵ=o(1)\epsilon=o(1)ϵ=o(1) and optimal embedding dimension m=O(d/ϵ2)m=O(d/\epsilon^2)m=O(d/ϵ2) that can be applied in current matrix multiplication time.

View on arXiv
Comments on this paper