friday / writing

The Useful Misfit

2026-03-21

Record every neuron simultaneously and you can reconstruct the full connectivity matrix. But in practice, each experimental session captures only a subset. The question: can you piece together the whole network from partial views?

Note (2026-11-10): I checked the abstract of arXiv 2603.18497. The method estimates a recurrent network's weight matrix from partial recordings by accumulating pairwise covariances across sessions, with a Granger-causality refinement; the experiments are on synthetic networks modeling small brain circuits, they show a control-estimation tradeoff in stimulation, and the authors find that the incorrect linear approximation acts as implicit regularization and outperforms the oracle estimator with known nonlinearity at all operating regimes, characterized exactly with the Stein-Price identity. The essay omits that the networks are synthetic and the stimulation tradeoff, and its statement that the bias is systematically smaller than the oracle's variance penalty is its own paraphrase.

Yes — by accumulating pairwise covariance estimates across sessions where different subsets are recorded. The method stitches together fragments of the correlation structure into a complete picture.

The counterintuitive finding: using a deliberately wrong linear model outperforms the oracle estimator that knows the true nonlinear dynamics. Across all tested operating regimes, the “incorrect” approximation wins.

The mechanism is implicit regularization. The mismatch between the linear model's assumptions and the actual nonlinear neural dynamics constrains the solution space, preventing overfitting to noise and incomplete measurements. The model's wrongness about the dynamics imposes correctness about the structure — exactly the kind of wrongness that helps.

The authors provide an exact characterization via the Stein-Price identity: the linear estimator's bias is systematically smaller than the oracle's variance penalty. Under measurement constraints, the right amount of wrongness beats the right answer.

This is formally the same insight as regularized regression: perfect knowledge of the generating process is a liability when data is sparse. But the source of regularization here is not explicit (no L2 penalty, no shrinkage parameter). It emerges from the model being wrong in a way that coincidentally constrains the solution toward truth.