ACID: A Low Dimensional Characterization of Markov-Modulated and Self-Exciting Counting Processes

Mark Sinzger-D'Angelo, Heinz KoepplarXiv 2022 · DOI · arXiv

Introduction

Counting processes are essential mathematical tools for modeling discrete events over time, with applications spanning from optical communications to systems biology. A fundamental challenge in this domain is the analysis of the Poisson channel, where an underlying signal—the intensity—is observed through a noisy stream of events. Quantifying the Mutual Information (MI) between the input signal and the output event stream is critical for understanding information transmission limits. However, because the intensity often depends on the process's entire history, calculating MI typically requires computationally intensive Monte Carlo simulations or imprecise mathematical bounds.

The ACID Framework

To address these challenges, researchers have introduced ACID (Asymptotic Conditional Intensity Distribution), a simulation-free numerical method. The core innovation lies in reframing the "conditional intensity" ($\hat{\lambda}_t$) not just as an estimate, but as a self-contained Markovian stochastic process.

The method utilizes Backward Recurrence Time Parametrization (BReT-P), which describes the system using the time elapsed since the last event ($\tau$) and a low-dimensional sufficient statistic ($\theta$). By characterizing the evolution of $(\tau, \theta)$ through a deterministic flow between events and jumps at event times, the authors derive a stationary Master Equation. Discretizing this equation transforms the problem into an eigenvector calculation, allowing for the direct computation of the long-term intensity distribution and the MI rate without random simulations.

Biological Context and Modeling

One of the primary motivations for this work is modeling mRNA transcription in systems biology. In these systems, the "events" are the production of individual molecules, and the underlying "signal" is the state of the gene promoter.

Biological modeling of transcription as a counting process.

As shown in the figure above, the process can be viewed as a signal-driven system (a) where the observer (b) must infer the state of the system from the discrete event stream. ACID provides a "temporal signature" of these processes, capturing nuances beyond simple mean and variance.

Key Results: Non-Markovian Optimality

The ACID method was validated against massive Monte Carlo datasets, showing perfect alignment. Beyond validation, the framework revealed significant insights into information transmission. For instance, the presence of "dark current" (background noise) was found to qualitatively alter optimal signal parameters.

A particularly striking result involves the "Double On Single Off" model. This model represents a non-Markovian input where the "On" state has a more complex internal structure.

The 'Double On Single Off' state transition model used to explore non-Markovian optimality.

Using this model, the authors demonstrated that standard Markovian inputs are not always optimal for information capacity under specific timing constraints. This finding suggests that complex, non-Markovian signal designs might be necessary to maximize information flow in biological or engineered sensors.

Conclusion and Future Directions

ACID offers a powerful new lens for analyzing counting processes by providing a simulation-free path to ground-truth distributions. While currently limited by the "curse of dimensionality" in the discretization of the sufficient statistic, it excels in 1D and 2D regimes. Future work aims to extend this methodology to higher-dimensional systems and apply it to refine approximate filters for complex chemical reaction networks.

Summary updated

BibTeX
@misc{sinzgerdangelo2022acid,
  author = {Mark Sinzger-D'Angelo and Heinz Koeppl},
  title = {{ACID: A Low Dimensional Characterization of Markov-Modulated and Self-Exciting Counting Processes}},
  year = {2022},
  eprint = {2205.07011},
  archivePrefix = {arXiv},
  url = {https://arxiv.org/abs/2205.07011},
}
Figure