Spectral theory for population density dynamics of spiking neurons with refractoriness

What the math gives to ML

The paper turns a refractory population-density model into a closed, dissipative, non-self-adjoint generator with an augmented state containing both active voltage density and refractory history. The transferable asset is not the neuron-specific Fokker–Planck equation itself, but the resulting stable infinite-dimensional-to-finite-dimensional reduction: boundary-driven inputs couple into spectral modes differently from bulk inputs, and defective eigenvalues produce controlled oscillatory transients through Jordan blocks. This suggests a recurrent or state-space neural layer with explicit stable spectral dynamics, separate boundary-like input injection, and optionally learnable near-exceptional-point mode pairs for long memory and oscillation. The first experiment should compare this constrained layer against a standard diagonal state-space model or GRU at equal state size and FLOPs.