Mean-field limit of Age and Leaky Memory dependent Hawkes processes

Valentin Schmutz

Abstract

We propose a mean-field model of interacting point processes where each process has a memory of the time elapsed since its last event (age) and its recent past (leaky memory), generalizing Age-dependent Hawkes processes. The model is motivated by interacting nonlinear Hawkes processes with Markovian self-interaction and networks of spiking neurons with adaptation and short-term synaptic plasticity. By proving propagation of chaos and using a path integral representation for the law of the limit process, we show that, in the mean-field limit, the empirical measure of the system follows a multidimensional nonlocal transport equation.

Introduction

In complex systems such as financial markets or neural networks, events are often modeled as point processes—sequences of discrete occurrences like stock trades or neuronal spikes. A fundamental challenge in these fields is bridging the gap between "microscopic" stochastic models, which describe individual agents, and "macroscopic" deterministic equations, which describe the collective behavior of the population.

While Hawkes processes have long been used to capture how events trigger subsequent activity, standard models often rely on the "renewal" assumption: the probability of a future event depends only on the time elapsed since the very last event (the "age"). However, biological systems like neurons exhibit more complex "non-renewal" behaviors, where a "leaky memory" of the distant past causes synapses to temporarily strengthen or weaken. This paper provides a rigorous mathematical foundation for these complex systems by deriving the mean-field limit for processes dependent on both age and leaky memory.

The Model: Age and Leaky Memory

The proposed framework tracks three primary variables for each process in a system of size $N$:

  1. Age ($A_t$): The time elapsed since the last event.
  2. Leaky Memory ($M_t$): A $d$-dimensional vector representing the history of past activity. This variable drifts toward a baseline but "jumps" whenever an event occurs, capturing phenomena like synaptic facilitation or depression.
  3. Global Interaction ($X_t$): A variable representing the aggregate influence of all other processes in the network.

By incorporating these variables, the model generalizes previous age-dependent Hawkes processes, allowing for a more nuanced representation of systems where the history between the last two events still significantly influences current behavior.

Mathematical Methodology

To derive the behavior of the system as $N$ approaches infinity, the author employs several sophisticated probabilistic techniques:

  • Propagation of Chaos: Using coupling techniques (à la Sznitman), the paper proves that as the system size grows, individual processes become asymptotically independent. This allows the collective empirical measure to be described by the law of a single "limit process."
  • Path Integral Representation: A significant novelty of this work is the use of a path integral to represent the law of the limit process. Unlike abstract measure-theoretic approaches, this provides an intuitive derivation of the resulting partial differential equation (PDE) and ensures that the solution exists within standard function spaces ($L^1$).
  • Quantified Convergence: The study does not merely prove convergence; it establishes a rate of $1/\sqrt{N}$. This provides a "speed limit" for how quickly a finite network begins to approximate the idealized infinite model.

Key Results

The Limit PDE

The central result is the derivation of a multidimensional nonlocal transport equation that governs the system in the mean-field limit. This equation accounts for three distinct dynamics:

  • The linear aging process.
  • The deterministic drift of the leaky memory.
  • The nonlocal "reset" terms that occur when a process triggers an event (e.g., a neuron spikes).

Mapping to Neuroscience

The paper demonstrates the practical utility of this framework by mapping it to the Tsodyks-Markram model, a well-known model of short-term synaptic plasticity (STP). By showing that the Tsodyks-Markram dynamics fit within the Age and Leaky Memory framework, the author provides the "missing link" for the population density formalism in neuroscience, proving that heuristic equations used to study large-scale brain activity are mathematically exact in the mean-field limit.

Limitations and Future Directions

While the paper provides a rigorous foundation, it operates under specific mathematical constraints. The intensity functions and interaction kernels are assumed to be bounded and satisfy Lipschitz conditions. Future research could expand this to include:

  • Singular Kernels: Addressing cases where interactions might be more abrupt or non-smooth.
  • Quenched Randomness: Exploring systems where interaction weights between specific pairs are themselves random variables, rather than a fixed common function.
  • Numerical Implementations: Developing mesh-based methods to simulate these PDEs efficiently as an alternative to computationally expensive Monte Carlo simulations of individual point processes.

Conclusion

This work establishes a rigorous mathematical bridge between stochastic point processes and deterministic population dynamics. By formalizing the role of leaky memory, it provides a robust framework for studying complex, history-dependent systems in neuroscience and beyond, ensuring that macroscopic models of large-scale networks rest on a solid theoretical footing.

Summary updated

Abstract

We propose a mean-field model of interacting point processes where each process has a memory of the time elapsed since its last event (age) and its recent past (leaky memory), generalizing Age-dependent Hawkes processes. The model is motivated by interacting nonlinear Hawkes processes with Markovian self-interaction and networks of spiking neurons with adaptation and short-term synaptic plasticity. By proving propagation of chaos and using a path integral representation for the law of the limit process, we show that, in the mean-field limit, the empirical measure of the system follows a multidimensional nonlocal transport equation.
BibTeX
@inproceedings{schmutz2022mean,
  author = {Valentin Schmutz},
  title = {{Mean-field limit of Age and Leaky Memory dependent Hawkes processes}},
  booktitle = {Stochastic},
  year = {2022},
  doi = {10.1016/j.spa.2022.03.006},
}
Figure