The study of random processes that count events over time, like the number of customers arriving at a store or neurons firing in the brain, is central to probability theory and its applications. A key mathematical question is whether any reasonable quantity derived from such a process can be expressed as a running sum of small updates, each depending only on information available at that moment. This kind of representation is fundamental in financial mathematics for pricing derivatives and in statistics for building efficient estimators. The authors tackle this question for a broad class of counting processes by using a clever construction called Poisson embedding, where a complicated counting process is built by selectively accepting or rejecting events from a simpler, well-understood Poisson process.
The main technical contribution is a rigorous way to derive the martingale representation theorem for counting processes directly from a corresponding result for the underlying Poisson process. The authors use tools from Malliavin calculus, which is essentially a way of differentiating random variables with respect to the randomness in the process itself. They show that any square-integrable martingale in the filtration of the counting process can be written as a stochastic integral against the compensated counting process, and they give an explicit formula for the integrand. This integrand has a natural interpretation: it measures the expected change in the final outcome if one extra event were to occur and be accepted at a given moment, conditional on everything observed so far.
As an extension, the authors introduce a companion process that counts the rejected events in the Poisson embedding construction. They show that any martingale in the combined filtration of both accepted and rejected events decomposes cleanly into two orthogonal parts, one driven by the accepted counting process and one by the rejected one. They then apply all of this machinery to Hawkes processes, which are self-exciting counting processes where past events make future events more likely. These are widely used in modeling earthquakes, financial order flow, and social media cascades. The explicit integrand formula derived here is new for Hawkes processes and opens the door to more concrete calculations in these applied settings.