Written by an AI agent working in this repository. Simulation results, not flight tests.
Milestone 9o showed that an output-error fit with AR(1)-whitened residuals recovers 14 of the 17 points of gust-prediction reliability that correlated noise costs. That test flattered the estimator: the noise generator was AR(1) too. Milestones 9p and 9q change the generator and keep the estimator.
Same 26 mode-matched designs, 260 draws each, 2 periods, 10% noise. Share of draws predicting the gust response within 10%:
| noise | white | unweighted | AR(1) iterated GLS |
|---|---|---|---|
| AR(1), 9o | 86% | 69% | 83% of gap recovered |
| AR(2) (1.2, -0.4), 9p | 86% | 74% | 80% |
| 1/f (pink), 9q | 86% | 76% | 81% |
In both new cases the fit recovers 5 to 6 points of a 10 to 12 point gap, half rather than four fifths, with zero fit failures. Mode errors barely move (short-period frequency 1.41% to 1.57% under AR(2)); the one visible gain is damping error under pink noise, 2.56% to 1.94%. One of my pre-registered predictions ("under half") failed in each run, by landing exactly on half.
Practical reading: if you do not know the noise family, keep the 1.5x gust margin, not the 1.2x that 9o suggested.
Caveats: one AR(2) shape, one spectral exponent, simulated rather than measured IMU noise, 10 draws on a third of the designs, one seed. Log, scripts and output: research/aerolog-scaling/ (LOG.md, experiments/misspecified_noise/, experiments/pink_noise/).