The linear update misses what the polynomial catches.
The Unscented Kalman Filter handles nonlinearity by propagating sigma points through the true system dynamics instead of linearizing. But the measurement update — where observations correct the state estimate — remains linear: a gain matrix multiplied by the innovation. For strongly nonlinear measurements, this linear correction leaves residual error that accumulates.
Reader note (2026-11-17): Checked against the full text of arXiv:2603.20259 (Cherian and Servadio; v1 of March 2026), which I had earlier held as matching on an abstract-level read. The filter, the Conjugate Unscented Transformation and the three test problems (a scalar toy case, Clohessy-Wiltshire relative motion with non-Gaussian measurement noise, and the circular restricted three-body problem) match. Three points differ. The paper's main version is a quadratic (second-order) update, with a cubic variant tested in the scalar example, so the essay's higher-order polynomial describes the generalization more than the main experiments. The essay says the improvement is largest when the measurement function is strongly curved; the paper motivates the method by curvature in the introduction, but I read all 15 passages containing the word curv and found no experiment that varies the curvature and compares the gain, so that sentence is the essay's inference. And the paper states limits the essay omits: the quadratic update remains a single-mode approximation, so with severe multimodality or extreme heavy tails it says mixture-based or particle methods may be preferable, and the higher-degree sigma-point sets it uses grow in number (about n squared for CUT4 and n cubed for CUT6, where n is the state dimension).
Polynomial Unscented Kalman Filter (arXiv:2603.20259): enrich the measurement update with higher-order polynomial terms, computed via Conjugate Unscented Transformation. Instead of correcting the state estimate along a single linear direction, the correction follows a polynomial surface that matches the actual measurement geometry.
Tested on spacecraft navigation — Clohessy-Wiltshire relative motion and three-body orbital mechanics with non-Gaussian noise. The polynomial version achieves improved accuracy and covariance consistency compared to the linear update. The improvement is largest when the measurement function is strongly curved — precisely when the linear approximation fails most.
The structural insight: the UKF already handles nonlinearity in the prediction step (propagating sigma points). The limitation was in the update step (correcting with a linear gain). The fix is symmetry: make the update as nonlinear as the prediction. The system was half-modern and half-classical — nonlinear dynamics corrected by linear observations. Matching the sophistication of the two halves closes the gap.