Spectral theory of large data with power-law latent features
cond-mat.dis-nn, cond-mat.stat-mech, physics.data-an
Submitted: 2026-10-28
Updated: 2026-10-28
Comments: 19 pages, 7 figures
License: http://creativecommons.org/licenses/by/4.0/
The gist: Power-law eigenvalue spectra of covariance matrices estimated from data appear in diverse domains ranging from neural populations to proteins.
Abstract
Power-law eigenvalue spectra of covariance matrices estimated from data appear in diverse domains ranging from neural populations to proteins. We introduce a generative latent feature model for such data, optionally corrupted by measurement noise, and develop its spectral theory. We derive closed leading-order self-consistent equations for the Stieltjes transform of the covariance matrix from which the bulk eigenvalue density is obtained and provide a variational principle for the self-consistent equations. We also compute the positions and fluctuations of isolated top eigenvalues. The theory is validated in known algebraic limits and tested against simulations.
Related papers
- Topological metal-insulator transitions in one-dimensional non-Hermitian quasicrystals: beyond PT-symmetry
- Microscopic and collective signatures of feature learning in neural networks
- Synchronous and asynchronous Monte Carlo dynamics of mean-field spin glasses with rotationally invariant couplings: what they share and where they differ
- Reentrant Localization Length in a Quasiperiodic Thue-Morse Chain
- Beyond Ground States: Learning Many-Body Hamiltonians for p-Bit Invertible Arithmetic
- Linear equivalence of nonlinear recurrent neural networks