Preprint proposes matrix compression method that matches or improves standard approximation

Researchers report a matrix compression method that matches or improves truncated singular value decomposition, which represents data with a smaller low-rank approximation, at equal parameter cost. Their preprint tests the method on synthetic data and real datasets including language-model embedding tables, network traffic and hyperspectral images. The method, called D-SLR, stores selected rows exactly and approximates the remaining rows. Computing its error-versus-parameter tradeoff and selecting a solution costs three singular value decompositions, without regularization or tuning. It also provides a computable bound on potential improvement from other rank and stored-row choices. The claim that separating exact rows from approximated rows sacrifices no optimum applies under squared error.