Macroscopic behavior of populations of quadratic integrate-and-fire neurons subject to non-Gaussian white noise

TL;DR

This paper investigates the collective dynamics of quadratic integrate-and-fire neuron populations under non-Gaussian white noise, deriving governing equations and analyzing the applicability of pseudocumulant methods for different noise types.

Contribution

It introduces a theoretical framework for non-Gaussian noise in neuron populations and examines the limitations of pseudocumulant approaches for fractional alpha-stable noise.

Findings

Derivation of the characteristic function dynamics for non-Gaussian noise.

Identification that pseudocumulant methods are mainly applicable for alpha=1 and 2.

Three-pseudocumulant models provide more accurate reductions than two-pseudocumulant models.

Abstract

We study macroscopic behavior of populations of quadratic integrate-and-fire neurons subject to non-Gaussian noises; we argue that these noises must be alpha-stable whenever they are delta-correlated (white). For the case of additive-in-voltage noise, we derive the governing equation of the dynamics of the characteristic function of the membrane voltage distribution and construct a linear-in-noise perturbation theory. Specifically for the recurrent network with global synaptic coupling, we theoretically calculate the observables: population-mean membrane voltage and firing rate. The theoretical results are underpinned by the results of numerical simulation for homogeneous and heterogeneous populations. The possibility of the generalization of the pseudocumulant approach to the case of a fractional $\alpha$ is examined for both irrational and fractional rational $\alpha$. This examination seemingly suggests the pseudocumulant approach or its modifications to be employable only for the integer values of $\alpha=1$ (Cauchy noise) and $2$ (Gaussian noise) within the physically meaningful range $(0;2]$. Remarkably, the analysis for fractional $\alpha$ indirectly revealed that, for the Gaussian noise, the minimal asymptotically rigorous model reduction must involve three pseudocumulants and the two-pseudocumulant model reduction is an artificial approximation. This explains a surprising gain of accuracy for the three-pseudocumulant models as compared to the the two-pseudocumulant ones reported in the literature.