Must the hidden state be measured?

No. Recovering a useful latent state representation is a central purpose.

Does the method identify a unique state coordinate system?

Generally no; invertible state transformations can preserve input-output behavior.

Is every subspace method suitable for feedback data?

No. Closed-loop identification requires methods and assumptions that address the dependence induced by feedback.

AI Glossary / Probabilistic machine learning / Advanced

Subspace identification

Definition

Subspace identification is a family of methods for estimating dynamical state-space models from observed time-series data using matrix projections and decompositions. In its classical form, it identifies linear time-invariant systems by recovering informative row or column spaces associated with hidden states and observability, then estimating the system matrices.

What it means

A system's outputs can depend on a compact internal state that is not directly measured. Subspace identification uses sequences of past and future inputs and outputs to discover a representation of that state. It is particularly relevant when several inputs and outputs interact over time.

The recovered state coordinates need not correspond to named physical quantities such as temperature or pressure. Different coordinate systems can describe identical input-output behavior. The goal is an effective dynamical model with an appropriate order, not necessarily a unique physical decomposition of every internal mechanism.

How it works

Organize lagged observations and, where present, inputs into block Hankel matrices. These matrices arrange overlapping time windows so that relationships between past data, future inputs, and future outputs can be studied algebraically.

Appropriate projections isolate information related to the state while accounting for inputs and the chosen noise assumptions. A singular value decomposition helps estimate a relevant subspace and select a model order. Algorithms then recover a state sequence or extended observability matrix and estimate transition, input, and output matrices, often through least-squares calculations.

N4SID, MOESP, and canonical-variate approaches are related methods with different weighting and projection choices. Noise, finite samples, feedback, and insufficient input excitation complicate identification. Ordinary principal component analysis of an unstructured table is not equivalent to this procedure. Validation should include prediction and simulation on data not used for fitting, along with residual and stability checks appropriate to the application.

Three practical examples

1. Process modeling: infer a compact dynamical relationship between manipulated settings and sensor outputs. Collect inputs that sufficiently excite relevant behavior rather than relying on nearly constant operating conditions.

2. Equipment monitoring: estimate a reference state-space model and inspect changes in predictive residuals. Distinguish model mismatch, sensor changes, and actual equipment changes before assigning a cause.

3. Signal forecasting: model a multivariate sequence through latent dynamical states. Compare alternative model orders and assess performance beyond the one-step fitting horizon.

Think of it this way

Think of observing shadows cast by a moving mechanism. Several time windows reveal patterns that individual snapshots hide. Subspace identification uses those patterns to construct a compact internal description. The coordinates of that description may differ from the mechanism's physical parts even when its predicted shadows are accurate.

Worked example

Consider an original noise-free autonomous example, with no external input. Observe outputs 2, 1, 0.5, and 0.25. Form a two-row Hankel matrix with rows (2, 1, 0.5) and (1, 0.5, 0.25).

The second row is exactly one half of the first, so the matrix has rank one. A one-state explanation is x at the next time equals 0.5 times the current x, with output y = x and initial state x0 = 2. It generates exactly the four observations.

An observability direction for two successive outputs is (1, 0.5). Its shift relation gives transition coefficient 0.5. The matrix can be factored as the column (1, 0.5) multiplied by the row (2, 1, 0.5), illustrating the observability-times-state structure.

The internal coordinates are not unique. Define a new state z = 3x. Then z0 = 6, the transition remains z at the next time = 0.5z, and output becomes y = z divided by 3. The observed outputs are unchanged. With noisy measurements, the second singular value would generally not be exactly zero, so selecting order one would require evidence rather than literal rank counting. This toy illustrates a subspace relationship, not a full noisy input-output identification algorithm.

Illustrative case study

Fictional case: Linden Cooling wants a model relating fan and valve settings to several temperature sensors. The team collects a designed operating record within approved conditions and fits multiple state orders using a documented subspace method. It reserves separate runs for validation.

Engineers compare multi-step simulations, inspect residual correlation, and check whether the learned dynamics remain credible across the intended operating range. Because normal operation uses feedback control, they verify that the identification method and data assumptions suit that setting. This fictional case makes no claim that an automatically selected order discovers the plant's exact physical state dimension.

Common misconceptions and mistakes

Subspace identification does not simply mean reducing the number of columns in a dataset. Small singular values can reflect noise, insufficient excitation, or weakly observable modes, so an apparent rank gap is not definitive proof of system order. State coordinates are generally identifiable only up to suitable transformations. A model that predicts well near one operating condition may fail under nonlinear changes or unsupported interventions.

Business use

The practical benefit is obtaining multivariable dynamical models through well-developed linear algebra rather than beginning with a large nonlinear parameter search. The costs include careful experiment design, preprocessing, order selection, and validation. Teams should record the sampling interval, inputs, outputs, lag-window choices, identification variant, and feedback conditions. Deployment decisions should reflect the model's validated domain and the behavior of forecasts over the actual operational horizon.

Check your understanding

Answer: 1.

Why is the toy Hankel matrix rank one?

Answer: Its second row is one half of the first.

Answer: 2.

What transition coefficient explains the observed decay?

Answer: It is 0.5.

Answer: 3.

Does replacing x by z = 3x change the predicted outputs when the output equation is adjusted?

Answer: No; it changes the internal representation only.

Related terms

Observability (software)
Singular value decomposition
System identification
state-space model
Hankel matrix
observability
N4SID
MOESP
persistent excitation

Learn next

Construct Hankel matrices from short sequences and examine their ranks. Then study state-coordinate transformations, observability, and how projections remove known input effects before applying a noisy-data subspace identification algorithm.

Sources

Van Overschee and De Moor, Subspace Identification for Linear Systems: Theory, Implementation, Applications, explains state-space recovery through projections and matrix decompositions: https://people.duke.edu/~hpgavin/SystemID/References/VanOverschee-1996.pdf

The geometric-decay calculation and Linden Cooling scenario are original teaching examples; the scenario is fictional.

Independent review consulted the cited textbook’s deterministic subspace theorem, projection factorization and similarity-transformation discussion.

Previous: Subspace Gaussian mixture model Next: Substitutability