Group Invariant Spectral Embedding

What the math gives to ML

The paper provides a principled way to collapse symmetry orbits before constructing similarity graphs: replace pointwise affinities by kernels minimized, averaged, or computed through invariant features over a compact Lie-group action. Its key transferable asset is that the resulting graph operator approximates differential operators on the quotient manifold \(\mathcal{M}/G\), eliminating redundant group directions rather than merely adding augmented samples. This suggests symmetry-aware neural positional encodings and attention kernels that compare inputs modulo known rotations or transformations. The most practical first test is a Nyström/Laplacian positional encoding or attention-logit kernel using Haar averaging over a small sampled group, compared against ordinary embeddings and explicit augmentation at equal compute.