The paper studies a type of self-exciting process called a Hawkes process, which models events that trigger further events, like earthquakes causing aftershocks or trades on a financial market sparking more trades. The focus is on a regime where the process is "nearly unstable," meaning it is close to a tipping point where the cascade of self-excitation would never die out. The paper asks a fundamental question: if you observe such a process for a long time, what can you actually learn about its key parameters? One parameter is the exogenous rate, essentially how often events arrive from outside the system independently of past events. The other is how close the process is to that critical tipping point.
The central finding is a striking asymmetry in what can and cannot be learned. The distance to criticality turns out to be fundamentally unlearnable in a precise statistical sense: no matter how long you watch the process, the statistical information about this parameter remains bounded, even as the number of observed events grows enormously. This is because the relevant information depends on the product of the observation length, the exogenous rate, and the distance to criticality in a way that stays finite. By contrast, the exogenous rate is learnable, because the heavy-tailed nature of the triggering behavior causes the process to spend time near zero in a way that creates a distinctive statistical signature. The mathematical structure of the problem turns out to be singular in one direction and well-behaved in another.
The authors back this up by analyzing the maximum likelihood estimator for both parameters jointly, showing its likelihood function is nicely shaped (concave), so there is a unique best guess. The estimate of the exogenous rate converges to the truth at a rate determined by the observed data itself, while the estimate of the distance to criticality stays bounded but does not converge. They also address a subtlety about whether results depend on where in the process's history you start observing, showing that for windows observed long after the process began, learnability of the exogenous rate holds with high probability due to the long-run statistical regularity of the process.