Prediction Error Estimation Methods (Ljung, 2001)

Lennart Ljung's 13-page technical report is a dense primary explanation of prediction-error methods (PEM). Its central contribution is not a software recipe but a precise answer to what the estimate means: choose a parameterized predictor, minimize its prediction errors, and understand the limiting model through the data-generating experiment, the candidate model set, the input spectrum, and the assumed noise model.

Actual coverage #

The load-bearing mechanism #

The general linear model is

$$y(t)=G(q,\theta)u(t)+H(q,\theta)e(t),$$

where $G$ describes the measured input/output dynamics and $H$ shapes an unpredictable innovation $e$. The corresponding one-step predictor is parameterized by $\theta$, and PEM selects

$$\hat\theta_N=\arg\min_\theta\frac{1}{N}\sum_{t=1}^{N}\ell\!\left(y(t)-\hat y(t\mid\theta)\right).$$

For a negative log-likelihood loss matched to the innovation distribution, PEM coincides with maximum likelihood. That equivalence is conditional on the probability model; least squares is not automatically maximum likelihood for arbitrary disturbances.

The report's most important engineering result is that the limiting plant model is experiment-dependent when the true system is not exactly in the candidate class. In open loop, input power emphasizes some frequencies more than others, while $H^{-1}$ reweights the error. An estimated transfer function is therefore not an experiment-free object: weakly excited frequencies and an inappropriate noise model can move the best approximation away from the dynamics relevant to control.

Why the model structures differ #

Ljung derives the common polynomial families from choices for $G$ and $H$ rather than presenting them as unrelated acronyms:

• Box–Jenkins gives plant and disturbance separate rational dynamics;
• output-error sets $H=1$ and estimates only plant dynamics;
• ARMAX makes plant and noise share the denominator $A$ while adding a moving-average noise numerator $C$;
• ARX is the $C=1$ special case and becomes a linear regression, but its shared $A$ polynomial couples plant and noise dynamics.

This structural reading is retained in GTK. It explains both computation and failure modes: ARX is easy to estimate but can need excessive order when the disturbance dynamics do not share the plant poles; a flexible noise model can reduce bias but adds parameters and nonconvex optimization.

Closed-loop data #

Feedback correlates the plant input with output disturbances. The report shows that PEM can still be consistent when the joint plant/noise structure is rich enough and external excitation supplies information not generated solely by the feedback disturbance. This is not permission to ignore experiment design. The reference signal, controller output, measured output, controller dynamics, sample timing and mode changes must be recorded, and residual independence must be interpreted using the actual closed-loop architecture.

Limits and use in GTK #

The report is deliberately a theory survey, not an end-to-end field protocol. It does not cover sensor synchronization, anti-aliasing, saturation, drift removal, missing data, multisine/PRBS design, independent validation logistics, modern regularization, or controller-specific uncertainty requirements in operational detail. Those aspects come from current official documentation and modern control practice.

GTK uses this source for the meaning of PEM, the derivation and assumptions of common model structures, closed-loop consistency conditions, and asymptotic uncertainty. It does not treat asymptotic normality as a finite-sample guarantee, nor a small covariance inside the chosen structure as evidence that the structure itself is adequate.

Visual audit #

All 13 pages were rendered and inspected. The paper contains no information-bearing plots or mechanism diagrams; its content is carried by equations and prose. The small institutional mark on the title page is decorative. No source visual is retained. GTK supplies independent diagrams for the experiment/data chain and the separation of plant and noise paths because those relationships benefit from spatial explanation.