Minkowski Polytopes of Spherical Designs: High-Order Isotropy and Quantitative Sphericity

What the math gives to ML

The paper turns spherical design exactness into a deterministic, highly isotropic directional discretization: equal-weight nodes reproduce every spherical polynomial up to degree t, while their Minkowski polytope has surface-area measure matching the sphere in the same moment orders. This suggests replacing random directional features, angular bins, or subsets of attention heads with a fixed spherical-design codebook, giving exact low-order cancellation and isotropy rather than merely approximate Monte Carlo isotropy. The quantitative bounds provide a falsifiable scaling law: directional discretization error should decrease like O(t^{-1}) for linear observables, while nonlinear geometric reconstruction degrades to O(t^{-1/2}) in Hausdorff distance.