Scale-Invariant Quantum Machine Learning: Zero-Shot Transfer Across Molecular Dimensions via Information-Theoretic Relational Calculus

This paper introduces an information-theoretic relational calculus that projects molecular energies onto a dimensionless phase space using the isolated informational potential of constituent atoms. Trained solely on H₂ and tested zero‑shot on LiH, a lightweight XGBoost model reduces absolute prediction error by 80%, eliminating dimensionality drift without deep neural networks.