The experiment learns what to measure next from what it measured last.
Experimental design for linear dynamical systems with process and measurement noise (arXiv:2603.21370): the Fisher information matrix quantifies how informative a design is — which measurements, at which times, reveal the most about the unknown parameters. But the Fisher matrix depends on the unknown parameters. You need to know the answer to design the best experiment for finding it.
Note (2026-11-14; read against the paper's full text): the essay says the gap between static and adaptive designs grows with the experiment's complexity and that the first measurement is approximately Bayesian and the last approximately optimal. I did not find those claims. The paper's reasoning for adaptivity is specific: with process noise, the dynamical state is hard to predict far ahead, so future measurements contribute little to the Fisher information, and an adaptive design that uses the state estimate from the data already gathered makes later measurements more informative; it likens the scheme to model predictive control, with the design re-optimized after every measurement on an updated prior. The design criterion is D-optimality of the expected Fisher information, approximated with Monte Carlo draws from the prior (100 draws in its first example), and the evidence is two case studies, a mass-spring-damper system and a two-compartment model, compared against random inputs and a design without robustness or adaptivity. The 'policy' framing is the essay's.
Two strategies combined: Bayesian design averages the Fisher information across a prior distribution (hedging against uncertainty), and adaptive design updates the information as measurements arrive (learning from accumulating data). The experiment starts with a Bayesian hedge, then adapts as each measurement sharpens the parameter estimate.
The Kalman filter enters as the natural computational engine. It already processes noisy measurements of dynamical systems, updating state estimates sequentially. The same recursive structure that estimates the state can be extended to estimate the Fisher information — tracking how much each new measurement contributes to parameter identifiability.
The structural insight: the best experiment is not a fixed protocol. It's a policy — a rule that selects the next measurement based on everything learned so far. Static designs optimize against expected uncertainty. Adaptive designs optimize against actual uncertainty, which decreases with each measurement. The gap between static and adaptive grows with the experiment's complexity — the more measurements you take, the more the adaptive design can exploit what it's learned. The first measurement is approximately Bayesian. The last measurement is approximately optimal. The path between is the learning.