Estimating the parameters of an interacting particle system typically requires observing all the particles. In a financial network, that means watching every institution; in a neural population, every neuron; in a flock, every bird. But in practice, you observe a subset. Some particles are visible; most are not. The hidden particles affect the visible ones through the interaction law, creating latent dependencies that classical estimation methods either ignore or approximate crudely.
Note (2026-11-14; read against the paper's full text): the essay credits the paper to Della Maestra and Oncu; the authors of arXiv:2602.20875 are Louis Sharrock, Nikolas Kantas and Grigorios A. Pavliotis. On substance, the paper proves convergence of the recursive estimator to stationary points of the asymptotic log-likelihood as time goes to infinity for a fixed finite number of particles and in the joint limit of many particles and long time, and, under additional assumptions such as strong convexity of the asymptotic negative log-likelihood, an L2 rate and a central limit theorem. The numerical examples (systemic risk, interacting FitzHugh-Nagumo neurons, Cucker-Smale flocking) are described as corroborating the theory and suggesting the estimator also works when the theoretical assumptions do not hold, so 'works beyond the formal assumptions' is the paper's observation about those examples, not a proved result. I did not find the essay's phrases about hidden particles or expectation-maximization; the paper's comparison is with an estimator that needs all particles to be observed.
Della Maestra and Öncü (arXiv:2602.20875) work directly with the partial observation. They observe a subset of particles continuously over time, compute a stochastic approximation of the gradient of the asymptotic log-likelihood, and update the parameter estimate recursively as data arrives. No reconstruction of the hidden particles. No expectation-maximization over latent states. The visible particles carry enough information about the interaction law because the interaction law is shared — the same parameters govern all particles, visible and hidden.
Convergence is proved both as observation time grows and as the particle count increases. A central limit theorem provides uncertainty quantification. Three applications — systemic risk modeling, FitzHugh-Nagumo neural dynamics, and Cucker-Smale flocking — demonstrate that the method works beyond the formal assumptions.
The general observation: when all components share the same interaction rule, observing a subset over a long time can substitute for observing the whole system briefly. The statistical information about the shared parameter accumulates through time, not through coverage. You don't need to see every particle. You need to watch the ones you can see for long enough.