To map a neural circuit, you need to stimulate it and record the responses. The stimulation reveals connectivity — which neurons drive which others. But the stimulation also disrupts the circuit's intrinsic dynamics. The very act of probing changes what you're trying to measure.
This is not merely an engineering limitation. It's a fundamental tradeoff between control and estimation. More stimulation improves identifiability — you get cleaner data about which connections exist. But more stimulation also pushes the circuit further from its natural operating regime, making the inferred connectivity less representative of what the circuit actually does.
Note (2026-11-10): I checked the abstract of arXiv 2603.18497 (I did not search its full text). The paper is a covariance-based method that reconstructs the connectivity matrix of a recurrent network from sparse partial recordings across sessions, refined by Granger causality, and its stimulation-estimation tradeoff is characterized in systematic experiments on synthetic networks modeling small brain circuits. The essay presents the tradeoff as formalizing what every optogenetic, electrical stimulation and calcium imaging experiment faces; the paper's tradeoff is in synthetic networks. The essay does not mention the paper's main surprises: that the incorrect linear approximation acts as implicit regularization and outperforms the oracle estimator with known nonlinearity, and the exact characterization via the Stein-Price identity.
The optimal stimulation level depends on the measurement density. If you can record from many neurons simultaneously, you need less stimulation because the partial measurements constrain the inference. If you can only record from a few, you need more stimulation, which distorts the dynamics further. The tradeoff has a sweet spot that shifts with experimental resources.
This formalizes a problem that neuroscience has known intuitively but not quantified. Every optogenetic experiment, every electrical stimulation study, every calcium imaging paradigm faces this tradeoff. The circuit you measure is not the circuit that was running before you measured it. The question is how far off you are, and the answer depends on how hard you push.
The structural insight: accurate measurement of any active system requires perturbation, and perturbation is measurement's enemy. The optimal observer is not the one who pushes hardest or lightest, but the one who pushes exactly as hard as the recording density can support.