Causal adaptive intensity estimation for non-stationary point processes: a Kalman-kernel duality framework

Abstract

Estimating the instantaneous conditional intensity of a non-stationary point process from a single realization of discrete event timestamps is a fundamental problem in computational neuroscience, quantitative finance, and network monitoring. Traditional adaptive kernel methods achieve variable bandwidths through Abramson's square-root law, yet existing implementations require non-causal access to the complete event record and incur $\mathcal{O}(N^2)$ or at best $\mathcal{O}(N \log N)$ computational cost. Causal alternatives based on reciprocal inter-arrival times, meanwhile, suffer from systematic divergence under Poisson and bursty arrival dynamics due to Jensen's inequality. This paper addresses both limitations by establishing a formal duality between steady-state Kalman filtering and adaptive exponential kernel smoothing. An endogenous state-space observer is constructed in which Poisson counting statistics dictate the measurement noise ($R \propto λ$) and geometric diffusion governs the process volatility ($Q \propto λ^2$). Under these physically motivated axioms, the algebraic Riccati equation autonomously yields an effective smoothing bandwidth scaling as $τ\propto λ^{-1/2}$, giving a causal, recursive, $\mathcal{O}(1)$ per-step realization of square-root adaptive bandwidth scaling. A second-order constant-velocity extension produces a $λ^{-1/4}$ scaling that reduces approximation bias, and a probabilistic dormant-state gating mechanism suppresses phantom intensity leakage during structural zeros. Synthetic benchmarks together with neural and financial data show that the observer adapts across orders of magnitude in intensity with an empirical per-step cost of approximately 30\,$μ$s.