Sparse Regression Faster than d^ω

09/23/2021

∙

by   Mehrdad Ghadiri, et al.

∙

0

∙

share

The current complexity of regression is nearly linear in the complexity of matrix multiplication/inversion. Here we show that algorithms for 2-norm regression, i.e., standard linear regression, as well as p-norm regression (for 1 < p < ∞) can be improved to go below the matrix multiplication threshold for sufficiently sparse matrices. We also show that for some values of p, the dependence on dimension in input-sparsity time algorithms can be improved beyond d^ω for tall-and-thin row-sparse matrices.