19
8

Total Least Squares Regression in Input Sparsity Time

Abstract

In the total least squares problem, one is given an m×nm \times n matrix AA, and an m×dm \times d matrix BB, and one seeks to "correct" both AA and BB, obtaining matrices A^\hat{A} and B^\hat{B}, so that there exists an XX satisfying the equation A^X=B^\hat{A}X = \hat{B}. Typically the problem is overconstrained, meaning that mmax(n,d)m \gg \max(n,d). The cost of the solution A^,B^\hat{A}, \hat{B} is given by AA^F2+BB^F2\|A-\hat{A}\|_F^2 + \|B - \hat{B}\|_F^2. We give an algorithm for finding a solution XX to the linear system A^X=B^\hat{A}X=\hat{B} for which the cost AA^F2+BB^F2\|A-\hat{A}\|_F^2 + \|B-\hat{B}\|_F^2 is at most a multiplicative (1+ϵ)(1+\epsilon) factor times the optimal cost, up to an additive error η\eta that may be an arbitrarily small function of nn. 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 d, where for a matrix CC, nnz(C)\mathrm{nnz}(C) denotes its number of non-zero entries. Importantly, our running time does not directly depend on the large parameter mm. 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} and B^\hat{B} from this low rank approximation, and then solving for XX so that A^X=B^\hat{A}X = \hat{B}. However, existing algorithms do not apply since in total least squares the rank of the low rank approximation needs to be nn, and so the running time of known methods would be at least mn2mn^2. In contrast, we are able to achieve a much faster running time for finding XX by never explicitly forming the equation A^X=B^\hat{A} X = \hat{B}, but instead solving for an XX 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