CORAL: Constrained Oblique Rotation with Anchored Loadings for Fidelity-Constrained Decorrelation

What the math gives to ML

The paper identifies the orthogonal freedom hidden inside whitening: every exact decorrelator can be written as an inverse square-root whitening transform followed by an orthogonal rotation. The transferable asset is that decorrelation and preservation of coordinate identity can be optimized separately, rather than accepting the arbitrary axis mixing introduced by PCA or a standard whitening convention. This suggests an anchored whitening layer for neural activations, with a tunable lower bound on correlation between each output channel and its corresponding pre-whitening channel. The same construction can serve as a stability or conditioning module while retaining semantic channel identity, which is useful for residual paths, interpretability, pruning, and feature reuse.