Observability

What Is Observability?

Observability is a property of a dynamical system that determines whether its internal state can be inferred from knowledge of its outputs and inputs over a finite time interval. A system is observable if every distinct initial state produces a distinguishable output trajectory, meaning no information about the state is permanently hidden from an external observer. The concept was introduced by Rudolf Kalman in 1960 as a dual property to controllability, and together these two ideas form the theoretical foundation of modern control and state estimation theory.

Observability is central to the design of state estimators, which reconstruct the full system state from a limited set of sensor measurements. Without it, no observer can recover the complete state, and any control law that depends on the full state will be impractical to implement.

Observability in Linear Systems

For linear time-invariant systems described by state-space equations, observability is determined by the rank of the observability matrix, assembled from powers of the system matrix multiplied by the output matrix. A system is fully observable if and only if this matrix has full rank, meaning its rows span the entire state space. This rank condition is exact and computable from the system parameters without simulation. The observability Gramian provides a quantitative measure: its eigenvalues indicate how easily different state components can be estimated from measurements, with small eigenvalues corresponding to nearly unobservable modes. An analysis of observability-constrained Kalman filtering for vision-aided navigation illustrates how violations of the observability condition corrupt filter performance in practical navigation systems.

Observability and State Estimation

State observers are algorithms that use the system model and output measurements to reconstruct unmeasured states in real time. The Luenberger observer adds a correction term proportional to the output estimation error, with gains chosen to place the observer error eigenvalues at desired locations in the complex plane. The Kalman filter is the optimal linear observer when process and measurement noise are Gaussian: it computes state estimates that minimize the mean squared estimation error while accounting for model uncertainty. Observability, eigenvalues, and Kalman filtering demonstrates that analyzing the eigenstructure of the error covariance matrix reveals which state components are well-estimated and which are poorly constrained by available measurements. Extended Kalman filters and unscented Kalman filters generalize these ideas to nonlinear systems, though observability analysis in nonlinear settings requires differential-geometric tools rather than simple matrix rank conditions.

Observability in Nonlinear and Complex Systems

Nonlinear observability analysis relies on computing the observability codistribution from successive Lie derivatives of the output with respect to the vector field defining the dynamics. A system is locally observable at a point if this codistribution spans the full state space in a neighborhood of that point. This analysis is more demanding than the linear case and may reveal that observability depends on the current operating trajectory, a phenomenon absent in linear systems. In power systems, battery management, and biochemical networks, determining which measurements are needed to make the system observable informs sensor placement decisions. Recent work on network observability examines when the state of a large-scale interconnected system can be recovered from a sparse set of sensing nodes, with graph-theoretic conditions replacing the classical matrix criteria. Studies of observability in nonlinear orbit estimation show how trajectory geometry determines whether range and bearing measurements are sufficient to recover relative state in space applications.

Applications

Observability has applications in a wide range of fields, including: