A Stein-Papangelou Goodness-of-Fit Test for Point Processes

Jiasen Yang, Vinayak A. Rao, Jennifer NevilleAISTATS 2019 · OpenReview

Abstract

Point processes provide a powerful framework for modeling the distribution and interactions of events in time or space. Their flexibility has given rise to a variety of sophisticated models in statistics, yet assessing their goodness-of-fit remains a significant challenge. Many realistic models, such as Gibbs or Hawkes processes, involve intractable normalization constants that preclude the use of standard likelihood-ratio tests. We present a novel statistical test that leverages the Papangelou conditional intensity and Stein’s method to circumvent this issue. By deriving a new Stein-Papangelou operator and applying it within Reproducing Kernel Hilbert Spaces (RKHS), we develop a Kernelized Stein Discrepancy (KSD) test. This test provides a closed-form statistic that measures model fit without requiring expensive simulations, outperforming traditional two-sample tests in statistical power.

The Challenge of Intractable Models

Point processes are mathematical frameworks used to model events occurring in time or space—such as earthquake locations, neural firing patterns, or the distribution of stars. While sophisticated models exist to describe these events, checking if a specific model fits observed data is difficult. The primary hurdle is the normalization constant. Many models accounting for point interactions (e.g., repulsion or attraction) can only be written "up to a constant." Calculating this constant requires solving infinite-dimensional integrals, which is computationally impossible, ruling out classical tools like likelihood-ratio tests. Researchers have often relied on heuristic summary statistics, such as Ripley’s K-function, which frequently lack the power to detect subtle mismatches.

Methodology: The Stein-Papangelou Framework

Our approach bridges point process theory and modern nonparametric statistics using two key components:

  1. The Papangelou Conditional Intensity: This function describes the relative probability of an event occurring at a specific location given all other event locations. In this ratio, the intractable normalization constant cancels out, leaving an evaluable formula.
  2. The Stein-Papangelou Operator: Stein’s method identifies a probability distribution by finding an operator $\mathcal{A}$ such that the expectation $\mathbb{E}[\mathcal{A}f]$ is zero if and only if the data follows that distribution. We derive a new operator specifically for point processes that captures the "inclusion-exclusion" dynamics of adding or removing points.

By combining these with RKHS, we derive the Kernelized Stein Discrepancy (KSD). This provides a test statistic that measures the "distance" between the model and the data. A significantly high KSD indicates a poor model fit.

Experimental Results

We evaluated the KSD test against the Maximum Mean Discrepancy (MMD) two-sample test using three fundamental models: the Poisson process (independent events), the Hawkes process (self-exciting temporal events), and the Strauss process (repulsive spatial events).

Comparison of error rates (FPR and FNR) for KSD and MMD across Poisson, Hawkes, and Strauss processes. KSD consistently exhibits lower FNR (higher power) while maintaining controlled FPR.

As shown in the figure above, the KSD test consistently outperforms the MMD test across various scenarios. Specifically:

  • Higher Statistical Power: KSD is better at detecting subtle perturbations in model parameters, reflected in a lower False Negative Rate (FNR).
  • Sample Efficiency: Because KSD utilizes the model's analytical form directly rather than relying on generated samples, it is more efficient than two-sample tests.
  • Robustness: The test maintains a controlled False Positive Rate (FPR) near the chosen significance level.

Discussion and Future Work

A major advantage of the KSD test is that it requires no sampling. Unlike MMD, which requires generating new data from the model to compare against observations, KSD only requires the model's formula. This is a massive benefit for models that are slow or difficult to simulate.

However, limitations remain. Evaluating KSD involves numerical integration over the "ground space," which can become computationally expensive in high-dimensional settings. Additionally, while the Stein identity is a necessary condition for a match, it may not always be sufficient for the most complex non-Poisson processes. Future research will focus on developing "physics-informed" kernels to further enhance the sensitivity of the test for specific process types.

Summary updated

Abstract

Point processes provide a powerful framework for modeling the distribution and interactions of events in time or space. Their flexibility has given rise to a variety of sophisticated models in stat...
BibTeX
@inproceedings{yang2019stein,
  author = {Jiasen Yang and Vinayak A. Rao and Jennifer Neville},
  title = {{A Stein-Papangelou Goodness-of-Fit Test for Point Processes}},
  booktitle = {International Conference on Artificial Intelligence and Statistics},
  year = {2019},
  url = {https://openreview.net/forum?id=ZFXkuyaTInu},
}
Figure