Density Estimation on Compact Manifolds under Intrinsic Spectral Block Variation

What the math gives to ML

The paper's transferable asset is spectral group sparsity defined over complete Laplace eigenspaces rather than individual eigenvectors. Grouping an entire repeated eigenspace makes the representation invariant to arbitrary rotations within degenerate eigenspaces, which is useful for spherical, rotational, mesh, and graph-based neural networks. A practical adaptation is a proximal spectral-block shrinkage layer that computes projectors onto Laplacian eigenspaces, applies one shared gate per block, and removes weak spectral levels without choosing a basis inside the block. This can produce compact geometric representations and reduce downstream computation, although the main engineering cost is estimating or storing spectral projectors.