From Geometry to Generalization: Why Row Normalization Can Beat Adam and Muon

Row-wise normalization achieves higher population accuracy than full-batch Adam or exact-SVD Muon in high-dimensional multiclass classification. Under isotropic Gaussian-cloud data models, row normalization preserves population decision-boundary directions while Adam's coordinate-wise and Muon's spectral geometries introduce nonvanishing distortions. This advantage persists for power-law spectra with class-mean exponents below one.