Is every state measured directly?

No; observers are especially useful when only part of the state is measured.

Is an observer always a Kalman filter?

No. The observer family includes deterministic and stochastic designs.

Does a stable observer guarantee a safe controlled system?

No. Controller design, combined dynamics, constraints, and operational validation also matter.

AI Glossary / Statistics and probability / Intermediate

State observer

Definition

A state observer is a model-based system that estimates the internal state of a dynamic process from its measured outputs and known inputs. Internal states may include quantities that cannot be measured directly, such as velocity when only position is sensed. Observers combine predictions from a process model with corrections based on observed discrepancies. The term comes primarily from control theory; it is retained in the supplied statistics category because state estimation connects control, inference, and uncertainty modeling.

What it means

A sensor usually reveals only part of what is happening inside a system. An observer uses the relationships between states, inputs, and outputs to infer the missing information over time. It is more than a one-off conversion from a reading to an estimate: the estimate evolves with the system.

A useful observer needs enough information in the measurements and a suitable model. If different internal states produce indistinguishable output histories under the available inputs, those distinctions cannot generally be recovered from those observations alone. This information issue is captured by observability and related concepts such as detectability.

How it works

For a discrete-time linear model, write x_(k+1) = A x_k + B u_k and y_k = C x_k, initially ignoring noise and direct feedthrough. A standard observer uses xhat_(k+1) = A xhat_k + B u_k + L(y_k - C xhat_k). The term in parentheses is the output discrepancy, and L determines how it corrects the estimated state.

With an exact model and known inputs, the estimation error e_k = x_k - xhat_k evolves as e_(k+1) = (A - LC)e_k. In discrete time, eigenvalues of this error matrix strictly inside the unit circle give asymptotic error decay. Continuous-time observers use a related equation with a different stability condition: negative real parts of error-system eigenvalues.

Observability allows placement of observer poles in the standard linear setting; detectability is sufficient for a convergent observer when unobservable modes are already stable. Real systems add noise, disturbances, delays, and model errors. Gain selection must address these effects, since faster ideal correction can amplify noisy measurements. Kalman filtering is one important stochastic state-estimation approach under additional modeling assumptions.

Three practical examples

1. A motion-control system estimates velocity from successive position measurements and known motor commands.

2. A building controller estimates internal thermal conditions from a limited set of temperature sensors and heating inputs.

3. A process-monitoring system compares measured outputs with observer predictions to investigate possible sensor or equipment faults.

Think of it this way

Think of following a moving vehicle partly hidden behind buildings. Its previous location and likely motion provide a prediction. Each new glimpse corrects that prediction. If the glimpses cannot distinguish two possible routes, confidence in one route cannot come from prediction alone.

Worked example

Consider an original noise-free discrete-time model with state x = [position, velocity], time step one, A = [[1, 1], [0, 1]], and measured output C = [1, 0]. There is no input in this teaching construction. Choose observer gain L = [1.1, 0.3] as a column vector.

The error matrix is A - LC = [[-0.1, 1], [-0.3, 1]]. Its characteristic polynomial is lambda squared - 0.9 lambda + 0.2, with roots 0.4 and 0.5. Both lie inside the unit circle, so every initial error decays asymptotically in this ideal model.

For initial error e_0 = [2, 1], the next error is [-0.1 times 2 + 1, -0.3 times 2 + 1] = [0.8, 0.4]. Subsequent errors are [0.32, 0.16] and [0.128, 0.064]. These independently calculated values illustrate convergence for one starting error.

The observability matrix stacks C and CA, giving [[1, 0], [1, 1]], which has rank two. Thus position histories contain information about both states. The example demonstrates ideal error dynamics, not a noise-robust gain choice or a validated physical controller.

Illustrative case study

In a fictional warehouse robotics project, engineers use wheel-position readings to estimate motion states needed by a controller. An observer performs well in an ideal simulation but behaves poorly when sensor readings are delayed.

The team incorporates timing behavior into testing and compares estimated states with an independent reference during controlled trials. It evaluates performance during acceleration, sensor noise, and temporary measurement loss. The project treats mathematical convergence under an ideal model as a design foundation, while requiring additional evidence before relying on the observer in operational motion control.

Common misconceptions and mistakes

An observer does not make uninformative sensors informative by itself. Do not confuse observability with controllability: the first concerns recovering state information, while the second concerns influencing the state through inputs.

Convergence of ideal estimation error does not guarantee accuracy under model mismatch or noise. Increasing gains without considering measurement quality can make estimates worse. An observer is also not automatically a fault detector; unexplained residuals can have several causes. Finally, a numerical state estimate is not a calibrated uncertainty interval unless the method provides and validates such uncertainty information.

Business use

Observers can reduce sensing costs and support monitoring, automation, and predictive maintenance. Their value depends on whether estimated states improve decisions and remain dependable across operating conditions. Document the state definitions, sensor timing, model assumptions, gain design, and validation coverage. Provide appropriate responses to missing or inconsistent measurements instead of treating every estimate as equally reliable.

Check your understanding

A discrete-time observer error matrix has an eigenvalue of 1.2. Does that satisfy the usual asymptotic convergence condition?

Answer: No. Its magnitude exceeds one, so the standard linear error system is not asymptotically stable.

Related terms

Observability (software)
Kalman filter
State estimation
observability
detectability
state-space model
Luenberger observer
feedback control
innovation

Learn next

Study state-space models and linear-system stability, then compare observability with controllability. Explore Kalman filtering after understanding deterministic observer error dynamics.

State space

Sources

Murray, State Estimation course overview: https://murray.cds.caltech.edu/CDS_110b:_State_Estimation

Introduction to Modern Controls, Observers and Observer-State Feedback: https://mcimp-book.github.io/chapters/12-observers/

The original discrete-time construction and fictional case are teaching examples. The control-theory scope is explicit despite the supplied category.

Previous: State machine agent Next: State space