Observability Matrix in System Analysis

• Observability Matrix is a core construct in systems theory that defines a system's ability to reconstruct its internal state from measured outputs.
• It supports rank-based, determinant-based, and numerical tests for state observability, sensor assignment, and system identification in both linear and nonlinear systems.
• Applications span control design, sensor placement, and parameter identifiability, with metrics like singular values and condition numbers guiding practical implementations.

An observability matrix is a central analytic object in systems and control theory, capturing the ability to reconstruct the internal state trajectory of a dynamical system from its measured outputs and known inputs. For both linear and nonlinear systems, the observability matrix facilitates precise rank-based (and in some settings, determinant-based or numerical) tests for local or global state observability, sensor assignment, system identification, and network-theoretic analysis. Its construction, properties, and quantitative metrics play essential roles across engineering, applied mathematics, and networked sciences.

1. Fundamental Definition Across System Classes

For continuous-time linear time-invariant (LTI) systems with state equation

$$\dot{x}(t) = A x(t) + B u(t), \qquad y(t) = C x(t) + D u(t)$$

the classical observability matrix is defined as

$$\mathcal{O} = \begin{bmatrix} C \ C A \ C A^2 \ \vdots \ C A^{n-1} \end{bmatrix} \in \mathbb{R}^{np \times n}$$

where $n$ is the state dimension, $p$ the output dimension, and the block rows encode multi-step output sensitivity to the initial state. The (structural) observability criterion is that $\operatorname{rank} \mathcal{O} = n$, implying the initial state can be uniquely determined from the output sequence when the input is known or absent (Moeurn, 2024, Villaverde, 2024).

In nonlinear systems, observability is generalized via the Hermann–Krener rank condition: given

$\dot{x} = f(x, u), \qquad y = h(x, u)$

one constructs a (possibly state- and input-dependent) matrix with block rows as the gradients (with respect to $x$) of the outputs and their iterative Lie derivatives,

$$\mathcal{O}(x, u) = \begin{bmatrix} \frac{\partial}{\partial x} h(x, u) \ \frac{\partial}{\partial x} L_f h(x, u) \ \vdots \ \frac{\partial}{\partial x} L_f^{n-1} h(x, u) \end{bmatrix}$$

where $L_f$ denotes the Lie derivative operator along $f$ (Sendiña-Nadal et al., 2019, Villaverde, 2024). Analogous constructions exist for discrete-time, second-order, and high-order systems (Mahmudov, 2019).

2. Observability Tests and Theoretical Criteria

The core test is a rank condition: a system is locally (or structurally) observable at a point $$\mathcal{O} = \begin{bmatrix} C \ C A \ C A^2 \ \vdots \ C A^{n-1} \end{bmatrix} \in \mathbb{R}^{np \times n}$$0 if $$\mathcal{O} = \begin{bmatrix} C \ C A \ C A^2 \ \vdots \ C A^{n-1} \end{bmatrix} \in \mathbb{R}^{np \times n}$$1 (Moeurn, 2024, Villaverde, 2024). This ensures that, in a neighborhood of $$\mathcal{O} = \begin{bmatrix} C \ C A \ C A^2 \ \vdots \ C A^{n-1} \end{bmatrix} \in \mathbb{R}^{np \times n}$$2, state trajectories are distinguishable given output and known input histories.

For linear systems, this criterion is equivalent to the existence of a left inverse for $$\mathcal{O} = \begin{bmatrix} C \ C A \ C A^2 \ \vdots \ C A^{n-1} \end{bmatrix} \in \mathbb{R}^{np \times n}$$3, so the mapping from initial state $$\mathcal{O} = \begin{bmatrix} C \ C A \ C A^2 \ \vdots \ C A^{n-1} \end{bmatrix} \in \mathbb{R}^{np \times n}$$4 to a finite output sequence is injective (Moeurn, 2024). In discrete-time and second-order systems, specialized block-matrix forms serve the same role (Mahmudov, 2019). For nonlinear systems, the Observability Rank Condition extends the test locally: the mapping from state to successive output derivatives is invertible if the corresponding Jacobian (i.e., the observability matrix) is full column rank (Sendiña-Nadal et al., 2019, Villaverde, 2024, Andrews et al., 2024).

Quantitative measures based on the singular values and condition numbers of the observability matrix (or its submatrices) further refine this analysis, enabling continuous metrics of observability or unobservability in empirical and data-driven settings (Cellini et al., 2023, Zhou et al., 2017).

3. Extensions to Networks, Nonlinearities, and Hypergraphs

For high-dimensional or networked systems, direct construction and symbolic analysis of the observability matrix quickly become infeasible. Structural and symbolic techniques have been developed:

• Graph/Symbolic Approach: The observability matrix in a network of coupled dynamical nodes (possibly nonlinear) is assembled using Kronecker structures; symbolic replacments (mapping Jacobian entries to classes such as constant, polynomial, or rational) yield a symbolic observability matrix. Observability coefficients quantify algebraic complexity. Root strongly connected components (rSCCs) in associated pruned fluence graphs dictate minimal sensor sets; measuring one node in each rSCC ensures generic observability (Sendiña-Nadal et al., 2019, Letellier et al., 2018).
• Hypergraphs: In systems whose states interact through higher-order (hyperedge-based) couplings, observability matrices are built using Kronecker-unfolded tensors that encode all polynomial and input-affine terms. Local observability is characterized by the rank of a block-Jacobian of all Lie derivatives; global observability can be checked via polynomial-ideal tests associated with the variety of output polynomial differences (Zhang et al., 2024).
• Randomized Sensing and Sparse Initial State Recovery: In high-dimensional, potentially compressive-sensing regimes, the observability matrix formed by stacking random block-diagonal measurement operators at different time steps admits concentration-of-measure and Restricted Isometry Property (RIP) guarantees for sparse recovery, with sampling requirements scaling as $$\mathcal{O} = \begin{bmatrix} C \ C A \ C A^2 \ \vdots \ C A^{n-1} \end{bmatrix} \in \mathbb{R}^{np \times n}$$5 in the sparsity $$\mathcal{O} = \begin{bmatrix} C \ C A \ C A^2 \ \vdots \ C A^{n-1} \end{bmatrix} \in \mathbb{R}^{np \times n}$$6 and state dimension $$\mathcal{O} = \begin{bmatrix} C \ C A \ C A^2 \ \vdots \ C A^{n-1} \end{bmatrix} \in \mathbb{R}^{np \times n}$$7 (Sanandaji et al., 2012).

4. Empirical and Practical Observability Quantification

Empirical Individual State Observability (E-ISO) offers a simulation-driven methodology for quantifying the practical observability of each state variable individually in nonlinear or black-box systems (Cellini et al., 2023). The approach constructs an empirical observability matrix via finite differences around an operating point and nominal input trajectory, then applies convex optimization to select row subsets (trajectories or sensor/time windows) minimally sufficient to reconstruct individual states. Singular value analysis of these minimal submatrices yields per-state observability and unobservability indices. This framework directly informs sensor placement, trajectory planning, and control law synthesis for engineered and biological systems.

In practical sensor assignment for multi-target tracking, measures of the symmetric (Gram) observability matrix—such as trace, rank, or log-determinant—display monotonicity and submodularity, admitting efficient approximation algorithms for sensor allocation. However, the inverse condition number does not inherit these properties, limiting greedy assignment guarantees. Lower bounds derived from partial Gram matrices facilitate robust assignment under target motion uncertainty (Zhou et al., 2017).

5. Applications, Algorithms, and Sensor Design

The observability matrix framework underpins a variety of algorithmic and applied developments:

• Sensor Selection: Minimal sensor sets are chosen by analyzing the (symbolic or structural) observability matrices, identifying essential outputs via root-SCC decompositions or convex optimization (Sendiña-Nadal et al., 2019, Letellier et al., 2018, Cellini et al., 2023).
• Experiment and Input Design: The structure of the observability matrix depends on input trajectories. Maximizing observability (e.g., optimizing the smallest singular value) with respect to sensor placement and input profiles enables active sensing and improved estimation performance (Cellini et al., 2023, Villaverde, 2024).
• Parameter Identifiability: Observability-parameter identifiability joint analysis is accomplished by treating parameters as augmented states in the observability matrix, ensuring that both states and parameters are locally distinguishable if the extended matrix is full rank (Villaverde, 2024).
• Analysis of Networked and Multiagent Systems: The Kronecker sum structure of observability matrices in networked systems enables tractable block-wise analysis, revealing precise dependencies on network topology, coupling, and measurement allocation (Sendiña-Nadal et al., 2019, Letellier et al., 2018).

6. Quantitative Metrics and Structural Limitations

Quantitative assessment is provided by the minimum singular value, condition number, or symbolic observability coefficients (based on the algebraic complexity of the determinant):

Metric	Mathematical Form	Interpretation
Minimum singular value	$$\mathcal{O} = \begin{bmatrix} C \ C A \ C A^2 \ \vdots \ C A^{n-1} \end{bmatrix} \in \mathbb{R}^{np \times n}$$8	Sensitivity to small state changes; local invertibility
Condition number	$$\mathcal{O} = \begin{bmatrix} C \ C A \ C A^2 \ \vdots \ C A^{n-1} \end{bmatrix} \in \mathbb{R}^{np \times n}$$9	Ill-conditioning; presence of near-unobservable modes
Symbolic coefficient	$n$0	Algebraic observability complexity (networked systems)

A large condition number indicates ill-conditioning and sensitivity to noise; a zero minimum singular value signals local unobservability. In networks, symbolic observability coefficients ascribe a quantitative score to measurement configurations, with $n$1 indicating algebraic full observability (Sendiña-Nadal et al., 2019).

Structural methods certify generic observability, but may fail for specific parameter or bifurcation values; direct symbolic determinants rapidly become intractable in high-dimensional or highly nonlinear models. In large-scale systems, graph-based methods offer scalable, polynomial-time algorithms at the expense of potential conservatism (Letellier et al., 2018).

7. Connections to Identifiability, Duality, and Compressive Sensing

Observability matrices are directly linked to (structural and practical) parameter identifiability in nonlinear parameterized models via the extended observability approach (Villaverde, 2024). Duality with controllability is established via matrix transposition; the observability Gramian and matrix rank connect to the non-singularity and pole-zero cancellation conditions of transfer functions (Moeurn, 2024, Mahmudov, 2019). In high-dimensional systems with randomly assigned observation operators, observability matrices align with compressive sensing theoretical foundations; concentration of measure and RIP properties guarantee stable and robust state recovery in sparse settings (Sanandaji et al., 2012).

---

The observability matrix, its analytic and empirical quantification, and the network-theoretic and randomized extensions remain foundational in system identification, sensor network design, complex systems analysis, and data-driven control science. Continued advances in symbolic, empirical, and scalable numerical techniques for constructing and analyzing observability matrices drive both theoretical and practical innovations across disciplines.