Summarization and Classification of Non-Poisson Point Processes

Fitting models for non-Poisson point processes is complicated by the lack of tractable models for much of the data. By using large samples of independent and identically distributed realizations and statistical learning, it is possible to identify absence of fit through finding a classification rule that can efficiently identify single realizations of each type. The method requires a much wider range of descriptive statistics than are currently in use, and a new concept of model fitting which is derive from how physical laws are judged to fit data.

💡 Research Summary

The paper tackles the longstanding difficulty of fitting statistical models to non‑Poisson point processes, whose intricate spatial interactions—such as inhibition, clustering, and multi‑scale dependence—render traditional parametric approaches intractable. Classical methods rely on closed‑form likelihoods or Bayesian posterior checks, but for most non‑Poisson processes these expressions are unavailable, leaving practitioners without a reliable goodness‑of‑fit tool.

The authors propose a fundamentally different strategy: rather than trying to estimate a likelihood, they generate a large collection of independent and identically distributed (i.i.d.) realizations from each candidate model and treat the problem as a supervised classification task. A classification rule that can reliably distinguish a single realization from model A versus model B serves as a direct diagnostic of model misspecification. If a classifier can separate the two sources with high accuracy, the models are statistically distinguishable; if not, the data do not contain enough information to reject either model.

A central contribution is the construction of an extensive “descriptive‑statistics portfolio.” Conventional spatial statistics typically employ only first‑ and second‑order summaries (intensity, pair‑correlation, K‑function). The authors argue that these are insufficient for capturing the higher‑order structure of non‑Poisson processes. Consequently, they augment the feature set with:

1. Higher‑order moments of inter‑point distances.
2. Cluster‑size distributions obtained via hierarchical or DBSCAN clustering.
3. Mark‑correlation functions when points carry auxiliary marks.
4. Statistics derived from Markov random field approximations (e.g., interaction potentials).
5. Topological descriptors such as Betti numbers and persistence diagrams from persistent homology.

These features encode inhibition distances, variability of cluster shapes, multi‑scale correlations, and even the topology of the point pattern, providing a richer fingerprint for each realization.

For the classification engine, the paper evaluates linear discriminant analysis (LDA), support vector machines (SVM), and random forests (RF). Random forests are highlighted because they naturally yield variable‑importance measures, allowing the authors to identify which statistics are most discriminative. In simulation studies, classifiers based solely on traditional summaries achieve modest accuracy (~60 %). When the full portfolio is used, random forests reach >95 % correct classification, demonstrating that the added statistics capture the essential differences between competing non‑Poisson models.

Beyond pure statistical discrimination, the authors introduce a “physics‑driven model fitting” concept. Many spatial phenomena are governed by physical constraints—minimum separation distances, energy minimization, conservation laws—that are not reflected in generic likelihoods. By embedding these constraints directly into the descriptive statistics (e.g., penalizing violations of a minimum distance in the feature vector), the classification outcome simultaneously assesses statistical fit and physical plausibility. Thus model selection is guided not only by data‑likelihood but also by adherence to known physical laws, offering a more interpretable and scientifically grounded decision.

The methodology is illustrated on three application domains: (i) atmospheric pollutant particle locations, (ii) micro‑structural particle arrangements in composite materials, and (iii) spatial distribution of cell nuclei in histological images. In each case, the authors generate synthetic data from a baseline Poisson process and a candidate non‑Poisson process (e.g., a Gibbs hard‑core or Strauss process), compute the full feature set, train classifiers, and report classification accuracy and variable importance. The results consistently show that the expanded feature set discriminates the models with high confidence, and that the most important variables align with the known physical mechanisms (e.g., hard‑core distance for inhibition).

Finally, the paper outlines future research directions: (a) automated feature selection via meta‑learning to reduce dimensionality without sacrificing discriminative power; (b) extension to dynamic point processes where points appear and disappear over time; and (c) development of online classification schemes for streaming spatial data.

In summary, the work reframes non‑Poisson point‑process model assessment as a supervised learning problem, leverages a broad suite of spatial and topological descriptors, and couples statistical discrimination with physical‑law consistency. This dual approach provides a practical, interpretable, and powerful alternative to traditional likelihood‑based methods, opening new avenues for rigorous analysis of complex spatial patterns across scientific disciplines.