OU Innovations¶
What it is¶
PQC models short-timescale correlated residual noise with an Ornstein-Uhlenbeck (OU) process. The OU process is a mean-reverting Gaussian stochastic process:
Why PQC uses it¶
TOA residuals are often not independent at short lags. OU innovations provide a local whitening transform so pointwise surprises are measured against a correlated-noise baseline rather than white-noise assumptions.
Discrete innovation form used in PQC¶
For irregular samples at times \(t_i\):
where q is an additional variance term estimated robustly.
Interpretation¶
\(|z_i| \approx 0\): consistent with the OU-noise prediction.
large \(|z_i|\): locally inconsistent, candidate bad measurement.
Assumptions and caveats¶
OU is only an approximation of residual correlation.
Innovation tails are treated approximately Gaussian.
Non-monotonic timestamps or severe model mismatch can distort scores.
Small worked example¶
If \(\tau=10\) d and \(\Delta t=1\) d, then \(\phi=\exp(-0.1)=0.905\). With \(y_{i-1}=1.0\), \(y_i=0.2\): \(e_i = 0.2 - 0.905 = -0.705\). If \(v_i=0.25\), then \(z_i=-1.41\).
References¶
Uhlenbeck, G. E., & Ornstein, L. S. (1930). “On the theory of the Brownian motion.” Physical Review, 36, 823-841.
Gardiner, C. (2009). Stochastic Methods (4th ed.). Springer.