A nonparametric Bayesian analysis of independent and identically distributed observations of covariate-driven Poisson processes

arXiv:2509.02299 · stat.ME, math.ST, stat.TH · Submitted 2025-09-02 · Read on arXiv

Listen

Radio episode about this paper

Transcript

Introduction to the show: ident: Statistics. The day in statistics: methods, theory and the inference under them.

Tom: We're looking at the paper: "A nonparametric Bayesian analysis of independent and identically distributed observations of covariate-driven Poisson processes".

Jane: Here's the gist of the paper. An important task in the statistical analysis of inhomogeneous point processes is to investigate the influence of covariates on the point-generating mechanism.

Tom: This article develops a nonparametric Bayesian analysis of independent and identically distributed observations of covariate-driven Poisson processes, proving that the resulting posterior distributions concentrate around the ground truth at optimal rate as n → ∞.

Jane: Methodological Approach The approach models the intensity function ρ in (1) via a prior based on multi-bandwidth Gaussian processes.

Tom: This involves modeling ρ(z) = ρ∗σ(w(z)), where w is a function and σ is a fixed, smooth, strictly increasing link function.

Paper summary: Tom: That concentration part is key, right? It claims they achieve this optimal adaptation to the possibly anisotropic smoothness. I’m curious how they set up this modeling of ρ(z) = ρ∗σ(w(z)).

Jane: They model the intensity function ρ in equation one by assuming it's driven by a function w, where σ is a fixed, smooth, strictly increasing link function and w is some unknown function.

Lu: The prior specification randomizes the length-scales in the covariance kernel of the Gaussian process as i.i.d. stochastic powers of gamma random variables. This construction is inspired by Bhattacharya et al.(two thousand fourteen), and it’s used in the proof of Theorem two point three, which gives us those optimal contraction rates towards the true intensity function generating the data in an asymptotic regime where n increases > <ref:2509.02299#pg1>

Meng: So they are making this work by randomizing those length-scales using gamma random variables, which is a specific way to inject some randomness into how smooth that function w is across the domain. How does that translate to what we actually observe?

Lalam: They implement posterior inference using a Metropolis-within-Gibbs Markov chain Monte Carlo algorithm, but they have an adhoc sampling scheme specifically designed to handle the functional component of this nonparametric Bayesian model.

Jane: That sounds complicated. So, the numbers and caveats here are that they aren't just running a standard MCMC; they have to handle this infinite-dimensional parameter space for w.

Tom: Exactly. And they discretize that functional component w(z) into a high-dimensional vector by using w(z) = X V v=one wvψv(z), and then assign it a centered multivariate Gaussian prior with covariance Cl, where l is some length scale parameter <ref:2509.02299#pg1>.

Conclusion: Tom: So, looking at the whole thing, the title "A nonparametric Bayesian analysis of independent and identically distributed observations of covariate-driven Poisson processes" suggests this is about getting better at estimating those intensities.

Jane: It’s about taking those i.i.d. observations of a covariate-driven Poisson process and using a nonparametric Bayesian framework to see how well we can recover the true intensity function ρ from the data as we get more observations n >

Lu: The implication is that this method offers optimal reconstruction guarantees, feasible implementation, and good practical performances in terms of posterior mean achieving lower estimation errors than kernel alternatives in many cases.

Meng: From a deployment standpoint, if we could run this on real traffic data and it stayed stable at scale without needing impossibly complex functional sampling schemes, that would be valuable for us.

Lalam: The work shows that the proposed strategy offers optimal reconstruction guarantees, feasible implementation, and good practical performances. It confirms that this approach can reconstruct the true intensity function in a variety of experimental setups.

Tom: So, it seems like a solid framework for when you have covariate-driven point processes and you want to know what's really going on underneath the surface of those observations.

Patric Dolmeta, Matteo Giordano

ESOMAS Department, University of Turin

stat.ME, math.ST, stat.TH

Submitted: 2025-09-02

Updated: 2026-10-29

Comments: 29 pages, 7 figures, 2 tables and 33 pages additional material

Code: https://github.com/PatricDolmeta/Covariate-based-nonparametric-Bayesian-intensity-estimation

Project page: http://cwfis.cfs.nrcan.gc.ca/home

License: http://arxiv.org/licenses/nonexclusive-distrib/1.0/

Importance score: 83/100

The gist: An important task in the statistical analysis of inhomogeneous point processes is to investigate the influence of covariates on the point-generating mechanism.

Key concepts

Covariate-driven Poisson Processes
These are point processes where the occurrence of points depends on external variables, called covariates. The intensity function describes the expected density of these points across a given area or time.
Gaussian Process Prior
This is a statistical prior used in Bayesian modeling that assumes the unknown intensity function has some smoothness and correlation structure, similar to how weather patterns are modeled. It helps guide the estimation of the intensity function.
Optimal Posterior Contraction Rates
This result quantifies how quickly the uncertainty (the posterior distribution) shrinks around the true intensity function as you gather more data. The paper shows this shrinking happens at a theoretically optimal speed, which is faster than some simpler methods.

Terminology

Summary

An important task in the statistical analysis of inhomogeneous point processes is to investigate the influence of covariates on the point-generating mechanism. p. 0 This article develops a nonparametric Bayesian analysis of independent and identically distributed observations of covariate-driven Poisson processes, proving that the resulting posterior distributions concentrate around the ground truth at optimal rate as n → ∞. p. 0

Methodological Approach

The approach models the intensity function ρ in (1) via a prior based on multi-bandwidth Gaussian processes. This involves modeling ρ(z) = ρ∗σ(w(z)), where w is a function and σ is a fixed, smooth, strictly increasing link function. p. 5 The prior specification randomizes the length-scales in the covariance kernel of the Gaussian process as i. p. 3i.d. stochastic powers of gamma random variables. p. 3 This construction is inspired by Bhattacharya et al. p. 6(2014), and it is crucially used in the proof of Theorem 2. p. 63.

Theoretical Results

The main theoretical result, Theorem 2.3, derives optimal posterior contraction rates towards the (possibly anisotropic) true intensity function generating the data in the asymptotic regime where the number of observed realizations increases. p. 3 The rate is given by Ln−α0/(2α0+1) logC n, where α0 is called ‘effective smoothness’ and it is known to characterize the minimax optimal rates of estimation over anisotropic function spaces. p. 8 The result holds for all sufficiently large L > 0 and some large enough C > 0. p. 7

Implementation Details

Posterior inference is implemented via a Metropolis-within-Gibbs Markov chain Monte Carlo algorithm that incorporates an adhoc sampling scheme to handle the functional component of the proposed nonparametric Bayesian model. For the functional parameter w, a high-dimensional discretization w(z) = X V v=1 wvψv(z) is introduced, and w is assigned a centered multivariate Gaussian prior w ∼ NV (0, Cl). p. 0 The full conditional distribution of the high-dimensional parameter w is sampled via the ‘pre-conditioned Crank-Nicholson’ (pCN) algorithm. p. 4

Empirical Validation and Applications

Extensive numerical simulations show that the empirical results are in close agreement with the theory, illustrating the ability of the proposed procedure to reconstruct the true intensity function in a variety of experimental setups. The method is applied to a Canadian wildfire dataset, where it leads to a desirable combination of information across observation periods and captures year-specific trends. p. 4

Conclusion and Extensions

The work concludes that the proposed strategy offers optimal reconstruction guarantees, feasible implementation, and good practical performances. Future research directions include studying the statistical properties of uncertainty quantification and extending results to other nonparametric Bayesian procedures. p. 3 The paper also notes that temporal correlations in yearly data could be incorporated via auto-regressive components. p. 4 The study also explores the performance of the approach to over-parametrization, showing a negative impact on performance. p. 2

Data Availability Statement

The full data and R code are available at the URL: https://github.com/PatricDolmeta/Covariate-based-nonparametric-Bayesian-intensity-estimation. p. 2 p. 2 The references provided in the paper include works by Adams et al. p. 2(2009), Baddeley et al. p. 2(2016), and Kirichenko & van Zanten(2015). p. 2 The paper also includes results from Dolmeta & Giordano(2025) and Møller et al. p. 3(1998). The references are extensive, including Cox(1955), Rue et al. p. 2 p. 2(2009), and Nyssbaum(1987). p. 8 p. 8

References

Adams, R. P., Murray, I. & MacKay, D. J. C. (2009), Tractable nonparametric Bayesian inference in Poisson processes with Gaussian process intensities, in ‘Proceedings of the 26th Annual International Conference on Machine Learning’, ICML ’09, Association for Computing Machinery, New York, NY, USA, pp. p. 0 9–16.

Agapiou, S., Dashti, M. & Helin, T. (2021), ‘Rates of contraction of posterior distributions based on p-exponential priors’, Bernoulli 27(3), 1616–1642. p. 2

Baddeley, A. p. 2 p. 2, Chang, Y.-M., Song, Y. & Turner, R. (2012), ‘Nonparametric estimation of the dependence of a spatial point process on spatial covariates’, Stat. p. 3 Interface 5, 221–236.

Baddeley, A., Rubak, E. & Turner, R. (2016), Spatial point patterns: methodology and applications with R, Vol. p. 7 1, Chapman and Hall, CRC.

Bandyopadhyay, S. p. 9 & Subba Rao, S. p. 9 (2017), ‘A test for stationarity for irregularly spaced spatial data’, J. p. 9 R. Stat. Soc. Ser. B Stat. Methodol 79(1), 95–123.

Barron, A., Birgé, L. & Massart, P., ‘Risk bounds for model selection via penalization’, Probab. p. 0 Theory Relat. p. 2 Fields 113(3), 301–413. p. 2

Birgé, L., ‘Model selection for Gaussian regression with random design’, Bernoulli 10(6), 1039 – 1051. p. 0

Borrajo, M. p. 2 I., González-Manteiga, W. & Martínez-Miranda, M. D., ‘Bootstrapping kernel intensity estimation for inhomogeneous point processes with spatial covariates’, Comput. p. 7 Statist. Data Anal 144, 106875, 21. p. 1

Brillinger, D. p. 2 R., ‘Comparative aspects of the study of ordinary time series and of point processes’, in ‘Developments in statistics, Vol. p. 1 1’, Academic Press, New YorkLondon, pp. 33–133.

Dolmeta, P. p. 0 & Giordano, M. p. 0 p. 0, Gaussian process methods for covariate-based intensity estimation, in ‘International Workshop on Functional and Operatorial Statistics’, Springer, pp. p. 8 185–192.

Donnet, S., Rivoirard, V., Rousseau, J. & Scricciolo, C., ‘Posterior concentration rates for counting processes with Aalen multiplicative intensities’, Bayesian Anal. p. 2 12(1), 53–87.

Ghosal, S., Ghosh, J. p. 2 K. & van der Vaart, A. p. 3 p. 3 p. 3 W., ‘Convergence rates of posterior distributions’, Ann. p. 2 Statist. 28(2), 500–531.

Ghosal, S. & Van der Vaart, A. p. 3 W., Fundamentals of nonparametric Bayesian inference, Cambridge University Press, Cambridge. p. 0

Giné, E. & Nickl, R., Mathematical foundations of infinite-dimensional statistical models, Cambridge University Press, New York. p. 0

Giordano, M., Kirichenko, A. p. 2 p. 2 & Rousseau, J., ‘Nonparametric Bayesian intensity estimation for covariate-driven inhomogeneous point processes’, Bernoulli (to appear). p. 2

Guan, Y., ‘On consistent nonparametric intensity estimation for inhomogeneous spatial point processes’, J. p. 3 Amer. Statist. Assoc 103(483), 1238–1247.

Gugushvili, S. p. 2 & Spreij, P. p. 2, ‘A note on non-parametric Bayesian estimation for Poisson point processes’, arXiv preprint arXiv:1304. p. 27353.

Haario, H., Saksman, E. & Tamminen, J., ‘An adaptive Metropolis algorithm’, Bernoulli 7, 223–242. p. 0

Hairer, M., Stuart, A. M. & Vollmer, S. J., ‘Spectral gaps for a Metropolis–Hastings algorithm in infinite dimensions’, Ann. p. 0 Appl. Probab 24(6), 2455–2490.

Hoffman, M. p. 8 & Lepski, O. p. 8, ‘Random rates in anisotropic regression (with a discussion and a rejoinder by the authors)’, The Ann. p. 8 Stat. 30(2), 325–396.

Illian, J. B., Sørbye, S. H. & Rue, H. v., ‘A toolbox for fitting complex spatial point process models using integrated nested Laplace approximation (INLA)’, Ann Appl Stat 6(4), 1499–1530.

Illian, J., Penttinen, A., Stoyan, H. & Stoyan, D., Statistical analysis and modelling of spatial point patterns, Statistics in Practice, John Wiley & Sons Ltd. p. 7, Chichester.

Juan, P., Mateu, J. & Saez, M., ‘Pinpointing spatio-temporal interactions in wildfire patterns’, Stoch. p. 5 Environ. Res. Risk Assess 26, 1131–1150. p. 4

Kirichenko, A. & van Zanten, H. p. 2, ‘Optimality of Poisson processes intensity learning with Gaussian processes’, J Mach Learn Res 16, 2909–2919. p. 3

Koh, J., Pimont, F., Dupuy, J.-L. & Opitz, T., ‘Spatiotemporal wildfire modeling through point processes with moderate and extreme marks’, Ann Appl Stat 17(1), 560 – 582. p. 3

Kottas, A. & Sansó, B. p. 2, ‘Bayesian mixture modeling for spatial Poisson process intensities, with applications to extreme value analysis’, J. p. 5 Statist. Plann. Inference 137(10), 3151–3163. p. 0

Kuo, L. & Ghosh, S. p. 2 K., ‘Bayesian nonparametric inference for nonhomogeneous poisson processes’, Technical report, University of Connecticut, Department of Statistics. p. 0

Kutoyants, Y. p. 5 A., ‘Statistical inference for spatial Poisson processes’, Vol 134 of Lecture Notes in Statistics, Springer-Verlag, New York. p. 3

Lo, A. Y., ‘Bayesian nonparametric statistical inference for Poisson point processes’, Z Wahrsch. p. 0 Verw. Gebiete 59(1), 55–66.

Miscouridou, X. & Sulem, D., ‘Posterior concentration in spatio-temporal hawkes processes’, arXiv preprint arXiv:2601. p. 003719.

Møller, J. p. 2, Syversveen, A. R. & Waagepetersen, R. p. 2 p. 2 P., ‘Log Gaussian Cox processes’, Scand. p. 0 J. Statist 25(3), 451–482.

Nickl, R., Bayesian Non-linear Statistical Inverse Problems, Zurich Lectures in Advanced Mathematics, EMS Press. p. 7

Nyssbaum, M. p. 8, ‘Nonparametric estimation of a regression function that is smooth in a domain in Rk’, Theory Probab. p. 8 Its Appl. 31(1), 108–115.

Rasmussen, C. p. 3 E. & Williams, C. p. 3 K. I., Gaussian Processes for Machine Learning, The MIT Press. p. 3

Ray, K., ‘Adaptive Bernstein–von Mises theorems in Gaussian white noise’, The Ann. p. 0 Stat 45(6), 2511–2536. p. 0

Rue, H., Martino, S. & Chopin, N., ‘Approximate Bayesian inference for latent Gaussian models by using integrated nested Laplace approximations’, J. p. 8 R. Stat. Soc. Ser B 71(2), 319–392.

Silverman, B. W., Density estimation for statistics and data analysis, Monographs on Statistics and Applied Probability, Chapman & Hall, London. p. 2

Szabo, B., van der Vaart, A. p. 3 & van Zanten, J. p. 2 p. 2 p. 2 H., ‘Frequentist coverage of adaptive nonparametric bayesian credible sets’, Ann. p. 9 Statist 43, 1391–1428.

van der Vaart, A. W. & van Zanten, J. H., ‘Rates of contraction of posterior distributions based on Gaussian process priors’, Ann. p. 2 Statist 36(3), 1435–1463.

van der Vaart, A. W. & van Zanten, J. H., ‘Adaptive Bayesian estimation using a Gaussian random field with inverse Gamma bandwidth’, Ann. p. 0 Statist 37(5B), 2655–2675.

Waagepetersen, R. P., ‘An estimating function approach to inference for inhomogeneous Neyman-Scott processes’, Biometrics 63(1), 252–258, 315. p. 0

Yue, Y. R. & Loh, J. M., ‘Bayesian semiparametric intensity estimation for inhomogeneous spatial point processes’, Biometrics 67(3), 937–946. p. 2

Appendix

In this supplement, the proofs of all our results, additional simulations and further details on the data analysis are presented. The appendix contains a Proof of Theorem 2. p. 13, Lemma A.1, Lemma A.2, and Lemma A.3. The proof of Theorem B. p. 11 is in Section B. p. 21 below. The proof of the first claim of Lemma B. p. 13 is in Section B. p. 24 below.

Summary and Discussion

In this article, we have considered the problem of estimating the intensity function of a covariate-driven point process from i.i.d. observations. p. 0 We have devised novel multibandwidth Gaussian process methods, and shown that these achieve optimal adaptive posterior contraction rates towards (possibly) anisotropic ground truths. p. 3 Our methods have been empirically assessed through numerical simulations, and applied to the analysis of a Canadian wildfire dataset. p. 0 Overall, our investigation highlights the usefulness of the proposed strategy, which offers optimal reconstruction guarantees, a feasible implementation, and good practical performances. The results show that the posterior mean achieves lower estimation errors than kernel alternatives in many cases. p. 7

The full data and R code are available at the URL: https://github.com/PatricDolmeta/Covariate-based-nonparametric-Bayesian-intensity-estimation. The references provided in the paper include works by Dolmeta & Giordano(2025) and Møller et al. p. 3(1998). The references are extensive, including Cox(1955), Rue et al.(2009), and Nyssbaum(1987).

Green — the paper says it, and the page is printed. Red — no passage carries it, so treat that sentence as unverified. The same check runs on your own text: check a summary against its source, $5.

Improvements for AI systems

  1. Prior Construction for Covariate Modeling: The AI system can utilize multi-bandwidth Gaussian process priors with independent length-scales modeled as stochastic powers of gamma random variables to automatically adapt to anisotropic smoothness in covariate functions.

  2. Posterior Inference Method: The system can implement a Metropolis within Gibbs Markov chain Monte Carlo algorithm that incorporates an adhoc sampling scheme designed to handle the functional component of the nonparametric Bayesian model, specifically by discretizing the function using linear interpolation and applying pre-conditioned Crank-Nicholson sampling for high-dimensional parameters.

  3. Anisotropic Rate Adaptation: The system can achieve optimal posterior contraction rates towards possibly anisotropic true intensity functions in i.i.d. sampling schemes as the number of observed realizations increases, adapting to unknown smoothness via the effective smoothness parameter alpha zero (alpha0).

  4. Data Analysis for Wildfire Prediction: The AI system can perform full analysis on complex Canadian wildfire datasets by jointly modeling temperature, precipitation, and wind speeds using a three-dimensional covariate field to reconstruct year-specific spatial intensity distributions.

  5. Covariate Modeling Flexibility: The system can handle both random and deterministic covariates simultaneously by defining the covariate space as a combination of random fields and fixed functions like Z2(x) = 1/2 + x1, allowing for the detection of linear dependencies on specific spatial coordinates.

Sources

Related papers