On Reduced Rank Regression
We propose a decoupled solution to the Reduced Rank Regression (RRR) problem, i.e. a pair of matrices (A*, B*) whose product A*B* minimizes the least squares error under a rank constraint. Using our formalism, we rewrite two other solutions from the literature and compare them with ours. In particular, we prove that all these formulations are equivalent. We then analyze their time complexity, and compare their running time and numerical precision on synthetic data, from which we derive practical recommendations. An open-source implementation of the algorithms and experiments is available here: https://gitlab.inria.fr/trudkiew/reduced\_rank\_regression.
math.NA↗