Spectral Selection in Sphere-Constrained Flows Generated by Polynomials of the Dirichlet Laplacian
arXiv:2608.24444
2026
Dynamics
2 ideas extracted · analyzed Aug 29, 2026
What the math gives to ML
The paper provides a constructive spectral-shaping mechanism: a norm-preserving flow generated by a polynomial of a self-adjoint operator converges to the active eigenspaces minimizing that polynomial, without explicitly computing eigenvectors. The non-monotone choice \(\mathscr{P}(s)=(s-\rho)^2\) is transferable because it acts as a differentiable band-pass filter, selecting eigenvalues near a desired spectral location rather than always selecting the largest or smallest modes. In neural networks this can be implemented with repeated Hessian-vector products or covariance multiplications, enabling curvature-targeted perturbations and feature-spectrum control.
Ideas from this paper
Unverified
2026
Replace the random or gradient-aligned perturbation in sharpness-aware minimization with a unit perturbation direction selected by a polynomial of the local Hessian. With \(\mathscr{P}(s)=(s-\rho)^2\), the direction converges toward Hessian eigenspaces whose eigenvalues are closest to the target curvature \(\rho\), allowing regularization of a chosen curvature band instead of indiscriminately penalizing only the sharpest direction.
Useful6/10
Difficulty6/10
Novelty7/10
Unverified
2026
Add a norm-controlled feature mixer that applies a polynomial spectral filter to the channel covariance of a transformer or MLP block. A quadratic filter centered at \(\rho\) suppresses covariance eigenmodes far from the target and preserves modes near it, providing a tunable alternative to purely variance-maximizing mixing or standard normalization.
Useful5/10
Difficulty5/10
Novelty6/10