Jointly Poisson processes

What constitutes jointly Poisson processes remains an unresolved issue. This report reviews the current state of the theory and indicates how the accepted but unproven model equals that resulting from the small time-interval limit of jointly Bernoulli processes. One intriguing consequence of these models is that jointly Poisson processes can only be positively correlated as measured by the correlation coefficient defined by cumulants of the probability generating functional.

💡 Research Summary

The paper tackles a long‑standing ambiguity in the definition of jointly Poisson processes, a topic that has received considerable attention in stochastic modeling but has never been fully resolved. The author begins by reviewing the conventional “jointly Poisson” construction that appears in much of the literature: each component process (X_i(t)) is expressed as the sum of an independent Poisson process with intensity (\lambda_i) and a common Poisson process with intensity (\gamma). Symbolically, (X_i(t)=Y_i(t)+Z(t)) where (Y_i) and (Z) are independent Poisson processes. While this representation intuitively captures the idea of a shared source of events, the paper points out that no rigorous derivation has ever been offered to justify why this should be the canonical model for joint Poisson behavior.

To fill this gap, the author turns to jointly Bernoulli processes, which are defined on a discrete time grid with a very small time step (\Delta t). In each interval, each component either fires (value 1) or does not fire (value 0). The success probabilities are taken to be linear in the step size: (p_i(\Delta t)=\lambda_i\Delta t+o(\Delta t)) for the idiosyncratic part and (p_c(\Delta t)=\gamma\Delta t+o(\Delta t)) for the common part. By letting (\Delta t\to0) and invoking the standard convergence of Bernoulli sums to Poisson counts, the paper demonstrates that the limiting continuous‑time process is exactly the conventional jointly Poisson model described above. This result is proved by comparing the probability generating functional (PGF) of the Bernoulli vector with the cumulant functional of the limiting process, showing term‑by‑term convergence of all cumulants.

Having established the equivalence, the paper proceeds to examine the correlation structure that inevitably follows from this construction. Using the second‑order cumulant (\kappa_{ij}) of the joint process, the correlation coefficient is defined as (\rho_{ij}= \kappa_{ij}/\sqrt{\kappa_{ii}\kappa_{jj}}). In the Bernoulli‑to‑Poisson limit, the only source of joint variability is the common intensity (\gamma), yielding (\kappa_{ij}= \gamma\Delta t) (up to higher‑order terms). Consequently (\rho_{ij}\ge 0) for every pair of components; negative correlations are mathematically impossible under this model. This is a striking and perhaps under‑appreciated fact: the accepted jointly Poisson framework can represent only positively correlated point processes when correlation is measured via cumulant‑based coefficients.

The author then discusses the practical implications of this restriction. In neuroscience, for example, simultaneous recordings of multiple neurons often exhibit both excitatory (positive) and inhibitory (negative) interactions. Modeling such data with a jointly Poisson process would inevitably miss the inhibitory component, leading to biased inference about network connectivity. In contrast, domains where events tend to co‑occur—such as simultaneous market trades, correlated failure events in reliability engineering, or synchronized arrivals in telecommunications—are well suited to the positively‑correlated structure. The paper therefore argues that the choice of a jointly Poisson model should be guided by a careful assessment of whether the underlying phenomenon can plausibly be restricted to non‑negative correlations.

Finally, the paper calls for the development of more flexible point‑process frameworks that can accommodate both positive and negative dependencies while preserving desirable analytical properties (e.g., infinite divisibility, tractable likelihoods). Possible directions include conditional Poisson processes where the intensity of each component is a stochastic function of the others, mixed point processes that combine Poisson and renewal components, or hierarchical constructions that embed a latent Gaussian field governing the intensities. Such extensions would retain the elegance of the Poisson paradigm but overcome the fundamental limitation identified in the current jointly Poisson definition.

In summary, the paper provides a rigorous derivation showing that the widely‑used jointly Poisson model is precisely the small‑time limit of jointly Bernoulli processes, and it highlights the consequent impossibility of negative correlation under the standard cumulant‑based correlation metric. This insight clarifies the theoretical foundations of joint Poisson modeling and offers a clear guideline for practitioners: if negative dependence is expected, one must look beyond the traditional jointly Poisson construction to more general point‑process models.