M-estimation of Boolean models for particle flow experiments

Probability models are proposed for passage time data collected in experiments with a device designed to measure particle flow during aerial application of fertilizer. Maximum likelihood estimation of flow intensity is reviewed for the simple linear Boolean model, which arises with the assumption that each particle requires the same known passage time. M-estimation is developed for a generalization of the model in which passage times behave as a random sample from a distribution with a known mean. The generalized model improves fit in these experiments. An estimator of total particle flow is constructed by conditioning on lengths of multi-particle clumps.

💡 Research Summary

The paper addresses the statistical modeling of passage‑time data collected from a device that records the flow of fertilizer particles during aerial application. The authors begin by reviewing the classic linear Boolean model, which assumes that every particle occupies a fixed, known passage time τ. Under this assumption the arrival of particles follows a Poisson process with intensity λ (particles per unit time). Observed data consist of “clumps” – consecutive groups of particles – whose lengths L are simply N·τ, where N is the number of particles in the clump. Maximum‑likelihood estimation (MLE) for λ is derived by constructing the likelihood of the observed clump lengths and solving for the parameter that maximizes it. While mathematically straightforward, this model fails to capture the variability observed in real experiments, where particle size, shape, aerodynamic conditions, and sensor noise cause passage times to differ from one particle to another.

To overcome this limitation the authors propose a generalized Boolean model. In the new framework each particle’s passage time Ti is treated as an independent draw from a known distribution FT with finite mean μ (the mean is assumed to be known from calibration or prior experiments). Consequently a clump length is expressed as the sum L = Σi=1N Ti, where N again follows a Poisson distribution with intensity λ. The model now contains two key parameters: the Poisson intensity λ governing the arrival process and the known mean μ of the passage‑time distribution.

Because the likelihood of a sum of random passage times does not admit a simple closed‑form expression for arbitrary FT, the authors turn to M‑estimation, a robust generalization of MLE. An M‑estimator minimizes a chosen loss function ρ over the residuals between observed clump lengths and their expected values under the model. The paper discusses several candidate loss functions, including the squared error, Huber’s hybrid loss, and absolute‑value loss, emphasizing the trade‑off between efficiency and robustness to outliers (e.g., unusually long clumps caused by turbulence). The estimation proceeds iteratively: an initial guess for λ is supplied, the expected clump length for each observed L is computed using the convolution of FT, residuals are formed, weights are updated according to the derivative ψ = ρ′, and a weighted least‑squares step refines λ. This iterative re‑weighted least squares (IRLS) algorithm is shown to converge reliably in simulated and real data sets.

A further contribution is the construction of an estimator for the total particle flow over a measurement interval, denoted Λ. Direct counting of particles is impossible because only clump lengths are observed. The authors therefore derive the conditional distribution P(N|L) via Bayes’ theorem: P(N|L) ∝ P(L|N)·P(N), where P(L|N) is the distribution of the sum of N independent Ti’s (easily obtained from the known FT) and P(N) is Poisson with intensity λ. The conditional expectation E