A Cascade of Volterra-Operator BBP Transitions in a Correlated Wigner Matrix

What the math gives to ML

The paper provides a nonstandard spectral mechanism: a correlated Wigner matrix can generate a countable sequence of BBP transitions, with one eigenvalue detaching from the semicircle bulk whenever the coupling exceeds a threshold set by a singular value of a compact Volterra operator. This transfers naturally to recurrent or state-space neural networks as a principled way to initialize and control multiple memory modes rather than choosing an undifferentiated random recurrent matrix. The measurable signature is a staircase in the empirical spectrum: as the correlation strength crosses successive thresholds, additional outlier modes appear and acquire distinct decay times. A practical implementation should combine the structured initialization with spectral-radius rescaling so the modes provide long memory without causing exploding dynamics.