Pseudo-Rank Leakage: How Inhomogeneous Physical Conservation Laws and Row-Wise Nonlinear Scaling Break Standard SVD Least Squares in Chemical Machine Learning
Mathematical proof and reproducible empirical benchmark identifying Pseudo-Rank Leakage: a numerical failure mode in chemical and physics-informed machine learning where an inhomogeneous linear conservation law (such as the Euler valency-cycle identity) interacts with smooth row-wise nonlinear scaling to bend an exact zero singular value into an unpruned intermediate mode, triggering catastrophic cross-validation variance explosion under standard SVD solvers across public EPA OPERA datasets.
Authors
- Shoji Fujita
Publication Details
- Journal
- Figshare
- Published
- 2026-11-19
- DOI
- https://doi.org/10.6084/m9.figshare.35066902
- Primary Topic
- Machine Learning in Materials Science
- Type
- preprint