Dense networks of integrate-and-fire neurons: Spatially-extended mean-field limit of the empirical measure

Pierre‐Emmanuel Jabin, Valentin Schmutz, Datong P. ZhouarXiv 2024 · DOI · arXiv

TL;DR

This paper proves that any dense network of stochastic integrate-and-fire neurons with $O(1/N)$ synaptic weight scaling converges to a deterministic, spatially-extended mean-field partial differential equation (PDE). By leveraging graphon theory, the authors show that spatial structure in the limit is an emergent property of the network's connectivity, even when the original network has no prescribed physical geometry.

What Problem Does This Paper Address?

In classical neuroscience, "mean-field" models describe the collective behavior of massive populations of neurons. Most derivations of these models assume exchangeability—the idea that neurons are essentially identical and interchangeable. While this works for homogeneous "soups" of neurons, real brains are highly structured.

Existing spatially-extended models (like the famous Wilson-Cowan equations) usually assume a pre-defined physical layout, such as neurons arranged on a 2D sheet. The authors ask a deeper question: Is spatial structure necessary to get a spatially-extended PDE? Specifically, if we have a dense network with arbitrary, non-homogeneous synaptic weights, does the $O(1/N)$ scaling of those weights alone guarantee that the network will behave like a spatially-extended population in the limit?

The challenge is that standard "propagation of chaos" techniques—the bread and butter of mean-field limits—break down when neurons are not interchangeable and the connectivity is complex.

Methods

The authors bridge the gap between stochastic dynamics and network topology using three key pillars:

  1. Graphon Theory: They treat the limit of the synaptic weight matrix as a graphon (or digraphon), a kernel $w(\xi, \zeta)$ that represents the connection strength between two abstract "locations" $\xi$ and $\zeta$ in the interval $[0, 1]$. This allows them to handle networks that are dense but lack a simple geometric structure.
  2. Extended Empirical Measure: To track the system, they define an "extended" empirical measure. They assign each of the $N$ neurons a unique coordinate in $[0, 1]$. As $N \to \infty$, these coordinates fill the space, and the discrete network dynamics transform into a density function evolving over both membrane potential and this new "spatial" dimension.
  3. A Kernel-Dependent Weak Metric ($\Phi_w^{-1}$): Standard metrics like the Wasserstein distance or Sobolev norms are often too "stiff" for the discontinuities inherent in integrate-and-fire models (where potentials reset to zero upon spiking). The authors introduce a novel weak metric specifically adapted to the limit kernel $w$. They prove stability by showing that regularity can be "propagated" along the dual-backward equation (the adjoint) of the mean-field PDE.

Datasets

Not discussed in the paper. This is a theoretical mathematical proof; no empirical datasets or simulations were used for evaluation.

Key Insights and Ideas

  • Emergent Space: The "spatial" dimension in the resulting PDE does not necessarily represent physical distance in the brain. Instead, it represents the organization of functional redundancies. Two neurons are "close" in this space if they interact with the rest of the network in similar ways.
  • Universality of Spatially-Extended PDEs: The result suggests that spatially-extended population equations are the universal limit for any dense network with $O(1/N)$ scaling. You don't need to "build in" a grid; the math finds one for you.
  • Convergence via Subsequences: Because the network can be arbitrarily ordered, the convergence happens "up to the extraction of a subsequence" and a re-ordering of neurons. This is a standard but powerful result from dense graph limit theory.
  • Fine Tracking: Unlike some previous methods that only look at "macroscopic observables," this approach allows for the tracking of individual neuron trajectories. The authors show that a single neuron's path in the finite network remains close to its theoretical trajectory in the mean-field limit.

Key Results

  • Theorem 1 (Main Result): Establishes that the extended empirical measure of the membrane potentials converges (weak-*) to the solution of a spatially-extended PDE.
  • The Limit PDE: The resulting equation accounts for the drift of membrane potentials, the "escape noise" (stochastic firing), and the resets. The interaction term is a convolution of the firing rate with the graphon kernel $w(\xi, \zeta)$.
  • Stability Framework: The authors provide a rigorous proof that the mean-field limit is stable, meaning small differences in initial conditions or connectivity don't lead to wild divergences in the population dynamics.

Limitations and Future Work

  • Dense Networks Only: The proof relies on the theory of dense graph limits (where the number of connections per neuron scales linearly with $N$). It does not cover sparse networks or networks with "intermediate" sparsity, which are common in biological systems.
  • Escape Noise vs. Diffusive Noise: The model uses "escape noise" (a soft threshold for firing). While common, many models use diffusive noise (Brownian motion). The authors note that their proof method relies on the absence of diffusive noise, which would require a different analytical toolkit.
  • Identifying the Graphon: While the paper proves a limit exists, it doesn't provide a recipe for finding the specific limit kernel $w$ for a given real-world biological connectivity matrix. That remains a significant challenge for applied researchers.

Summary updated

Abstract

The dynamics of spatially-structured networks of $N$ interacting stochastic neurons can be described by deterministic population equations in the mean-field limit. While this is known, a general question has remained unanswered: does synaptic weight scaling suffice, by itself, to guarantee the convergence of network dynamics to a deterministic population equation, even when networks are not assumed to be homogeneous or spatially structured? In this work, we consider networks of stochastic integrate-and-fire neurons with arbitrary synaptic weights satisfying a $O(1/N)$ scaling condition. Borrowing results from the theory of dense graph limits, or graphons, we prove that, as $N\to\infty$, and up to the extraction of a subsequence, the empirical measure of the neurons' membrane potentials converges to the solution of a spatially-extended mean-field partial differential equation (PDE). Our proof requires analytical techniques that go beyond standard propagation of chaos methods. In particular, we introduce a weak metric that depends on the dense graph limit kernel and we show how the weak convergence of the initial data can be obtained by propagating the regularity of the limit kernel along the dual-backward equation associated with the spatially-extended mean-field PDE. Overall, this result invites us to reinterpret spatially-extended population equations as universal mean-field limits of networks of neurons with $O(1/N)$ synaptic weight scaling.
BibTeX
@misc{jabin2024dense,
  author = {Pierre‐Emmanuel Jabin and Valentin Schmutz and Datong P. Zhou},
  title = {{Dense networks of integrate-and-fire neurons: Spatially-extended mean-field limit of the empirical measure}},
  year = {2024},
  eprint = {2409.06325},
  archivePrefix = {arXiv},
  url = {https://arxiv.org/abs/2409.06325},
}
Figure