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. 1909.12441
28
8

Total Least Squares Regression in Input Sparsity Time

27 September 2019
H. Diao
Zhao Song
David P. Woodruff
Xin Yang
ArXiv (abs)PDFHTML
Abstract

In the total least squares problem, one is given an m×nm \times nm×n matrix AAA, and an m×dm \times dm×d matrix BBB, and one seeks to "correct" both AAA and BBB, obtaining matrices A^\hat{A}A^ and B^\hat{B}B^, so that there exists an XXX satisfying the equation A^X=B^\hat{A}X = \hat{B}A^X=B^. Typically the problem is overconstrained, meaning that m≫max⁡(n,d)m \gg \max(n,d)m≫max(n,d). The cost of the solution A^,B^\hat{A}, \hat{B}A^,B^ is given by ∥A−A^∥F2+∥B−B^∥F2\|A-\hat{A}\|_F^2 + \|B - \hat{B}\|_F^2∥A−A^∥F2​+∥B−B^∥F2​. We give an algorithm for finding a solution XXX to the linear system A^X=B^\hat{A}X=\hat{B}A^X=B^ for which the cost ∥A−A^∥F2+∥B−B^∥F2\|A-\hat{A}\|_F^2 + \|B-\hat{B}\|_F^2∥A−A^∥F2​+∥B−B^∥F2​ is at most a multiplicative (1+ϵ)(1+\epsilon)(1+ϵ) factor times the optimal cost, up to an additive error η\etaη that may be an arbitrarily small function of nnn. Importantly, our running time is O~(nnz(A)+nnz(B))+poly(n/ϵ)⋅d\tilde{O}( \mathrm{nnz}(A) + \mathrm{nnz}(B) ) + \mathrm{poly}(n/\epsilon) \cdot dO~(nnz(A)+nnz(B))+poly(n/ϵ)⋅d, where for a matrix CCC, nnz(C)\mathrm{nnz}(C)nnz(C) denotes its number of non-zero entries. Importantly, our running time does not directly depend on the large parameter mmm. As total least squares regression is known to be solvable via low rank approximation, a natural approach is to invoke fast algorithms for approximate low rank approximation, obtaining matrices A^\hat{A}A^ and B^\hat{B}B^ from this low rank approximation, and then solving for XXX so that A^X=B^\hat{A}X = \hat{B}A^X=B^. However, existing algorithms do not apply since in total least squares the rank of the low rank approximation needs to be nnn, and so the running time of known methods would be at least mn2mn^2mn2. In contrast, we are able to achieve a much faster running time for finding XXX by never explicitly forming the equation A^X=B^\hat{A} X = \hat{B}A^X=B^, but instead solving for an XXX which is a solution to an implicit such equation. Finally, we generalize our algorithm to the total least squares problem with regularization.

View on arXiv
Comments on this paper