Key Concepts of Observability

Why This Matters

Observability is one of the two fundamental pillars of modern control theory, alongside controllability. It determines whether you can actually know what's happening inside a system based only on what you can measure. Recognizing when a system's internal states can be reconstructed from output data is essential for designing observers, state estimators, and feedback controllers. Without observability, you're flying blind: no Kalman filter, no state feedback based on estimates, no reliable monitoring of critical system behavior.

These concepts connect directly to state-space analysis, matrix rank conditions, canonical forms, and duality principles that appear throughout control systems coursework. Don't just memorize the observability matrix formula. Understand why full rank guarantees state reconstruction, how the Gramian provides energy-based insight, and when nonlinear techniques become necessary. Each concept below approaches the same core question from a different angle: can we figure out where the system is from what we can see?

---

Foundational Framework

Before testing observability, you need the mathematical structure that makes analysis possible. These concepts establish the language and representation used throughout observability theory.

Definition of Observability

A system is observable if you can reconstruct the full state vector from output measurements over a finite time interval. This isn't about convenient or special initial conditions. Complete observability requires that any arbitrary starting state be determinable from the output history alone.

Why does this matter for design? Without observability, state feedback control based on estimated states becomes impossible. Every observer and estimator you'll encounter in this course assumes observability holds, so verifying it is always the first step.

State-Space Representation

State-space models use first-order differential (or difference) equations to describe system dynamics:

$\dot{x} = Ax + Bu$ $y = Cx + Du$

The $C$ matrix is the critical link for observability. It maps internal states to outputs, determining which state combinations actually appear in your measurable signals. All observability tests operate directly on the $A$ and $C$ matrices from this formulation.

---

Rank-Based Testing Methods

The most common exam questions involve determining observability through matrix rank conditions. These algebraic tests give definitive yes/no answers for linear systems.

Observability Matrix

The observability matrix is constructed by stacking $C$ and its successive products with $A$:

$O = \begin{bmatrix} C \\ CA \\ CA^2 \\ \vdots \\ CA^{n-1} \end{bmatrix}$

where $n$ is the state dimension (the number of state variables).

If $\text{rank}(O) = n$, the system is observable. Intuitively, full rank means every state influences the output through some combination of the output and its derivatives. To compute the rank in practice, use row reduction or, for small systems, check whether the determinant of $O^T O$ is nonzero.

Kalman's Observability Rank Condition

This is the definitive algebraic test for LTI observability: a linear time-invariant system is observable if and only if the observability matrix has rank equal to $n$.

When the rank is deficient, the gap tells you exactly how many states are hidden. If $\text{rank}(O) < n$, then $n - \text{rank}(O)$ state directions cannot be distinguished from the outputs. These correspond to the system's unobservable modes.

This test is directly analogous to the controllability rank test, a connection worth internalizing since duality questions appear frequently.

---

Energy and Gramian Approaches

Beyond rank conditions, the observability Gramian provides a continuous measure of how observable a system is, not just whether it's observable at all.

Observability Gramian

The observability Gramian is defined as an integral over a time horizon $T$:

$W_o = \int_0^T e^{A^T t} C^T C \, e^{At} \, dt$

A positive definite $W_o$ (all eigenvalues strictly positive) guarantees observability. But the Gramian goes further than a binary test. Its eigenvalues tell you how strongly each state direction contributes to the output energy. Large eigenvalues correspond to easily observed state directions; small eigenvalues indicate states that are weakly observable. This matters for practical estimator design because weakly observable modes cause numerical conditioning problems in filters.

Observability in Linear Time-Invariant (LTI) Systems

For LTI systems, the constant $A$ and $C$ matrices simplify all tests. The observability matrix and Gramian have fixed structure, so you compute them once and the answer holds for all time.

A key simplification: as $T \to \infty$ (assuming the system is stable), the Gramian converges to the solution of the Lyapunov equation:

$A^T W_o + W_o A + C^T C = 0$

This algebraic equation is often easier to solve than evaluating the integral directly. LTI observability also forms the foundation for Kalman filter design, where it guarantees that optimal state estimation error converges to zero.

---

Canonical Forms and Duality

These structural concepts reveal deep connections in control theory and simplify certain design problems.

Observable Canonical Form

The observable canonical form is a specific state-space arrangement where observability is immediately apparent from the structure of $A$ and $C$. For a single-output system with characteristic polynomial $s^n + a_1 s^{n-1} + \cdots + a_n$, the matrices take a companion-like form where $C$ picks off a single state and $A$ places the characteristic coefficients in a predictable pattern.

This form is useful because pole placement for observer design follows a straightforward coefficient-matching procedure. Any observable system can be converted to observable canonical form through a similarity transformation $\bar{x} = Tx$, which preserves eigenvalues and the transfer function.

Relationship Between Observability and Controllability

The duality principle is one of the most elegant results in linear control theory: a system $(A, B, C)$ is observable if and only if the dual system $(A^T, C^T, B^T)$ is controllable.

This means the controllability matrix of the dual system, $[C^T \; A^T C^T \; \cdots \; (A^T)^{n-1} C^T]$, is just the transpose of the observability matrix. Any theorem or design technique you know for controllability has a direct observability counterpart, and vice versa.

These are independent properties in general. A system can be controllable but not observable, observable but not controllable, both, or neither. A system that is both controllable and observable is called minimal, meaning its state-space representation has no redundant states.

---

Beyond Full Observability

Real-world systems often don't satisfy clean observability conditions. These concepts address what happens when observability is incomplete or when linearity assumptions break down.

Partial Observability

When $\text{rank}(O) < n$, only a subset of state directions can be reconstructed from outputs. This is common when sensors are limited or certain state variables are inherently hidden from measurement.

The Kalman observability decomposition provides a systematic way to handle this. It applies a similarity transformation that separates the state space into observable and unobservable subspaces:

$\bar{A} = \begin{bmatrix} A_o & 0 \\ A_{21} & A_{\bar{o}} \end{bmatrix}, \quad \bar{C} = \begin{bmatrix} C_o & 0 \end{bmatrix}$

The observable subsystem $(A_o, C_o)$ can be estimated normally. The unobservable states require additional assumptions, prior knowledge, or additional sensors to resolve.

Observability in Nonlinear Systems

For nonlinear systems of the form $\dot{x} = f(x, u)$, $y = h(x)$, the linear rank tests don't apply directly. Observability can depend on the specific trajectory the system follows and may vary across the state space.

The standard tool is the Lie derivative approach. You compute successive Lie derivatives of the output function $h(x)$ along the vector field $f$:

• $L_f^0 h = h(x)$
• $L_f^1 h = \nabla h \cdot f(x)$
• $L_f^k h = \nabla(L_f^{k-1} h) \cdot f(x)$

The system is locally observable at a point if the observability codistribution, formed by the gradients of these Lie derivatives, has rank $n$. This is the nonlinear generalization of the Kalman rank condition.

A critical distinction: local vs. global observability. A nonlinear system may be locally observable near an equilibrium but lose observability in other regions of state space. Analysis must specify where observability holds, not just whether it holds.

---

Quick Reference Table

---

Self-Check Questions

1. Given a system with $A = \begin{bmatrix} 1 & 1 \\ 0 & 2 \end{bmatrix}$ and $C = \begin{bmatrix} 1 & 0 \end{bmatrix}$, construct the observability matrix and determine if the system is observable. What does the rank tell you?

2. Compare the observability matrix approach and the observability Gramian approach. When would you choose one over the other in practice?

3. If a system is controllable but not observable, what are the implications for closed-loop control design using state feedback with an observer?

4. Explain the duality between observability and controllability. If you're given a controllability matrix for system $(A, B)$, how would you construct the corresponding observability matrix for the dual system?

5. A nonlinear system is locally observable near its equilibrium point but loses observability in certain regions of state space. What mathematical tools would you use to analyze this, and how does this differ from LTI observability analysis?