Common estimation mistakes and how to diagnose them¶
| Symptom | Likely mistake | Discriminating check |
|---|---|---|
| Confident drift | Bias omitted or correlated evidence counted twice | Compare known bias injection with zero-bias run; inspect residual autocorrelation |
| Good pixel fit, wrong distance | Wrong target size or transform direction | Recover a known metric fixture; invert pose explicitly |
| Every delayed image rejected | Arrival time substituted for capture time | Replay the same values with correct capture timestamps |
| Quality changes when input frequency changes | Wrong Q discretization or fixed dt | Compare continuous-noise predictions over subdivided intervals |
| Position right, velocity nonsense | State ordering/units wrong | Track a noiseless constant-velocity sequence with hand-computed prediction |
| Covariance becomes negative | Numerical cancellation, bad matrices, or invalid linearization | Joseph update; inspect eigenvalues and units; compare an analytic scalar case |
| Impossible 358° turn | Unwrapped angle residual | Test 179° versus −179° |
| Online performance looks too good | Future samples/smoothed labels used as current observations | Audit arrival-time availability at every output |
| Filter trusts a repeated sample more each time | Duplicate observation not rejected | Preserve source/sequence identity; test replay idempotence in deployment |
Worked check: two identical copies of one measurement with variance 1 do not provide two independent pieces of evidence. Treating them as independent yields an apparent measurement variance 0.5 and unjustified confidence.
Use lessons 12–21 for derivations and the CPU lab for executable counterexamples. Fix one cause at a time; a lower fitted-sequence error is not proof of consistency on new data.