Correlated errors from accumulated IMU noise
This software qualification adds a reusable state-space covariance operator. It does not fit a flight, estimate archive noise, change a completed result or open a protected recording. The existing receiver-error operator stays unchanged.
Accept explicitly supplied linear state transitions, independent process-noise loadings, observation matrices, receiver-noise loadings and an initial-error loading. The matrices and Gaussian error law are declared in advance; they are not inferred from the residual being whitened. Loadings produce positive semidefinite process covariance without clipping eigenvalues. Receiver innovations must factor without jitter or repair.
Compute sequential innovations and their log determinant with storage linear in the number of epochs. Apply exactly the same operator to residual vectors and every Jacobian column. Support inversion and explicit continuous block checkpoints; reject changed configurations, skipped epochs or changed column counts. Check the declared storage cap before consuming the step iterator.
Provide a separate constant-level, small-angle example with continuous white gyro and accelerometer errors. Its three-axis covariance must agree with the analytic impulse-response integrals, including cross-time terms. This example has a perfect initial state and does not describe the turning/climbing aircraft or supply calibrated uncertainty for a flight.
Qualification compares small problems with independently assembled dense covariance and Cholesky whitening, including nonuniform epochs, changing transitions, cross-axis receiver covariance and all 27 columns. Test zero and duplicated columns, inversion, block continuity, input preservation, rejection of invalid/singular matrices and failure before allocation at a storage cap.
One scalability check uses 64,080 generated epochs and eight latent states, with a 128 MiB stored-operator cap. No dense reference is constructed for this long case. Short dense controls use at most 32 epochs. All tests use one numerical thread, reserve the three active workers and perform no native trajectory calls, optimizer starts, original reads or protected reads. No stochastic draws or fitted outcomes are used to select these tests.
Source-bound local qualification must pass before publishing the component as software checked. A time-varying flight error model, estimated-parameter uncertainty, empirical noise assessment and coverage calibration remain separate gates. Existing physical accuracy targets and the running noise study are fixed.