Hessian spectrum at the global minimum of high-dimensional random landscapes
Yan V Fyodorov, Pierre Le Doussal

TL;DR
This paper calculates the Hessian spectral density at the global minimum of high-dimensional Gaussian landscapes, revealing how the spectral gap behaves near phase transitions and differs across landscape complexities.
Contribution
It provides an analytical calculation of the Hessian spectrum in various landscape regimes, highlighting the gap behavior and the effects of replica symmetry breaking.
Findings
Hessian spectrum is gapped in simple landscapes with a positive spectral edge.
Spectral gap vanishes as $(-_c)^2$ near the transition point.
In highly complex landscapes, the Hessian remains gapless, indicating marginal stability.
Abstract
Using the replica method we calculate the mean spectral density of the Hessian matrix at the global minimum of a random dimensional isotropic, translationally invariant Gaussian random landscape confined by a parabolic potential with fixed curvature . Simple landscapes with generically a single minimum are typical for , and we show that the Hessian at the global minimum is always {\it gapped}, with the low spectral edge being strictly positive. When approaching from above the transitional point separating simple landscapes from 'glassy' ones, with exponentially abundant minima, the spectral gap vanishes as . For the Hessian spectrum is qualitatively different for 'moderately complex' and 'genuinely complex' landscapes. The former are typical for short-range correlated random potentials and correspond to 1-step…
Peer Reviews
No public reviews on file for this paper yet. If you reviewed it on a platform where reviews are public (OpenReview, ICLR, NeurIPS, ICML), you can paste yours below so the community can read it here.
Videos
No videos yet. Explain this paper in a talk, walkthrough, or lecture? Add one.
