Algorithmic Differentiation of QR decompositions
Published in AlgoDiff 2024, 2024
Matrix Derivatives to propagate gradients through all QR matrices. This paper proves two formula mathematically equivalent for real matrices with real valued entries and non-equivalence for complex valued matrices. For complex field entries, we derivate an adjustment that enables library developers and maintainers to easily add support for complex matrices. NOTE: Linked paper is a preprint, full paper is in press at Springer.
Recommended citation: Roberts D, Roberts L. QR and LQ Decomposition Matrix Backpropagation Algorithms for Square, Wide, and Deep - Real or Complex - Matrices and Their Software Implementation. arXiv preprint arXiv:2009.10071. 2024
Download Paper