Receiver and IMU errors over a complete record
The new calculation combines time-correlated receiver error with accumulated IMU (inertial measurement unit) noise. It keeps their correlation across the training/prediction boundary and never resets the noise-integration origin when selecting a later interval.
15 focused software checks pass; 59 pass with the existing inertial-noise and shared-scale checks. The comparisons include irregular times through 438 seconds, all 27 input columns, and agreement with a separately constructed dense covariance reference. Turning off the IMU-noise contribution reproduces the existing receiver-only calculation.
The implementation stores a small evolving error state instead of a square matrix covering every observation. A 1,000-epoch test uses less than 1 MB of operator storage. A larger archive-scale runtime trial has not yet run.
What is assumed
The caller supplies the receiver covariance, accelerometer and gyroscope noise spectra, and a constant matrix connecting small orientation error to acceleration error. There are no hardware-performance defaults. The existing linear inertial-error equations are unchanged. Receiver and inertial driving errors are assumed independent; their accumulated observations remain correlated in time.
The operator predicts the error mean and coordinate variances from training values only. It does not estimate calibration uncertainty, select a noise law, read future positions or decide which physical model wins. Calibration must remain in the separate shared-parameter prediction method.
What this adds—and what remains
This connects the receiver and inertial error components in a calculation suited to long histories. It does not show that the supplied Gaussian/linear law describes the aircraft recordings. Variable flight dynamics, possible processing effects, error-scale assessment and supported numerical allowances still need explicit treatment before aircraft prediction coverage can be claimed.