Recursive Maximum Likelihood Estimation for Interacting Particle Systems using Virtual Particles

Authors	Louis Sharrock et al.
Year	2026
Field	AI / ML
arXiv	2605.00786
PDF	Download
Categories	stat.ME, stat.ML

Abstract

We study recursive maximum likelihood estimation for stochastic interacting particle systems based on continuous observation of a single particle. In this regime, consistent estimation of the finite-particle log-likelihood is not possible, even in the limit as the number of particles $N\rightarrow\infty$ and the time horizon $t\rightarrow\infty$. We thus seek to optimise the stationary log-likelihood of the limiting mean-field system. We achieve this via a form of stochastic gradient estimate in continuous time, with stochastic gradient estimates computed using the continuous trajectory of the single observed particle, alongside a virtual interacting particle system and a virtual tangent interacting particle system, which are integrated with the online parameter estimate. For fixed numbers of real and virtual particles, we show that the resulting algorithms drive the gradient of a finite-particle surrogate objective to zero as $t\to\infty$. We then prove that, in the iterated limit $t\to\infty$ followed by $N,M\to\infty$, these surrogate gradients converge uniformly to the gradient of the stationary log-likelihood of the limiting mean-field system, yielding convergence to its stationary points. We illustrate the method on several numerical examples, including a model with quadratic confinement and interaction potentials, a model of interacting FitzHugh--Nagumo neurons, and a stochastic Kuramoto model.

---

Engineering Breakdown

Plain English

This paper tackles a fundamental problem in learning parameters of stochastic particle systems where you only observe a single particle in continuous time. The authors prove that you cannot consistently estimate the likelihood when the system has many interacting particles, even as both the particle count and observation time go to infinity. To solve this, they propose optimizing the mean-field limit instead—the behavior of infinitely many particles—using a clever stochastic gradient descent approach that combines the observed particle's trajectory with two virtual particle systems that run alongside it to compute gradient estimates. This enables practical recursive maximum likelihood estimation in regimes where standard approaches completely fail.

Core Technical Contribution

The key novelty is proving that finite-particle likelihood estimation is fundamentally inconsistent in the single-observation regime, then providing a constructive solution through mean-field optimization with online stochastic gradient estimates. The authors introduce a three-system architecture: the observed particle provides real data, while a virtual interacting particle system (approximating the mean-field) and a virtual tangent particle system (for gradient computation) run in tandem with an online parameter estimator. This is fundamentally different from prior work because it abandons the impossible goal of finite-particle likelihood consistency and instead leverages mean-field theory to extract learnable signal from single-particle observations. The continuous-time stochastic gradient formulation allows integration directly with observed particle trajectories, making the approach naturally suited to real-time streaming observation scenarios.

How It Works

The method takes continuous observations from a single particle in an interacting particle system as input, along with an initial parameter estimate. For each infinitesimal time step, the algorithm maintains three coupled systems: (1) the true observed particle whose trajectory generates observations, (2) a virtual N-particle system initialized with the current parameter estimate that evolves according to the mean-field dynamics, and (3) a tangent particle system that tracks how the virtual particle system's gradient changes with respect to parameters. The gradient estimate is computed from the difference between the observed particle's observed dynamics and what the virtual system predicts, scaled by the tangent system's sensitivity. These gradient estimates feed into a continuous-time stochastic gradient descent update rule that recursively refines parameters online. The algorithm outputs improved parameter estimates in real-time, with no batch reprocessing needed.

Production Impact

For engineers building systems that learn from streaming particle-based simulations (molecular dynamics, agent-based models, biological systems), this removes a hard barrier: you can now extract parameter estimates from single-particle observations where classical methods would fail entirely. In practice, this reduces data requirements dramatically—you only need to instrument one particle instead of reconstructing full system state, cutting sensor/measurement costs and communication bandwidth. However, the approach requires running two additional virtual particle systems in parallel with your observed system, roughly tripling the computational cost per timestep for likelihood estimation. Integration complexity is moderate: you need a continuous-time dynamics solver, automatic differentiation for the tangent system, and careful numerical stability management when coupling the three systems, making this suitable for research prototypes and simulation environments rather than ultra-latency-sensitive inference. The main advantage is enabling parameter learning from single-node observations in high-dimensional systems where full-state observation is impractical.

Limitations and When Not to Use This

The paper assumes access to continuous (or very high-frequency) observations of the single particle's full state, which may not hold if your sensor data is low-frequency, noisy, or partially observed—the method likely degrades significantly under these realistic conditions. The approach requires knowledge of the correct mean-field dynamics model; if your assumed dynamics are misspecified, the virtual particle systems will mislead gradient estimates, and there is no analysis of robustness to model mismatch. The computational cost of running three coupled systems makes this impractical for very high-dimensional particle systems or real-time inference at scale, and the paper does not provide guidance on choosing the virtual particle system size N or convergence rates as a function of N. Finally, the theoretical results focus on consistency of mean-field optimization but do not characterize finite-time convergence rates, sample complexity, or how performance degrades with observation noise—critical for practitioners to predict when this will work.

Research Context

This paper extends classical maximum likelihood estimation theory to the challenging non-i.i.d., high-dimensional regime of interacting particle systems, building on decades of work in mean-field theory and particle filtering. It complements recent progress in neural network-based learning of particle dynamics (e.g., via neural ODEs and graph neural networks) by providing a principled statistical foundation for when and how to learn from partial observations. The work opens a new research direction: exploring what happens when observations are discrete-time, noisy, or partially observed—moving from the idealized continuous-observation setting toward realistic sensor data. This also connects to broader themes in simulation-based inference and likelihood-free inference, where learning from simulator outputs rather than true data is central.

---

:::tip Subscribe Get weekly breakdowns of papers like this in AI Letters - the newsletter for engineers building production AI systems. :::

---

Back to Research Lab → · Subscribe to AI Letters →