Exact Observer Linearization Framework

• Exact observer linearization is a framework that transforms nonlinear system dynamics into precisely linear observer error dynamics using smooth state transformations.
• It employs nonlinear coordinate maps and Lie derivative matching to design observers with user-prescribed convergence rates in both continuous and discrete-time settings.
• Recent advancements integrate neural network methods and algebraic-geometric approaches to extend the framework to complex, high-dimensional, and non-square systems.

The exact observer linearization framework refers to a suite of methodologies for constructing observer systems whose error dynamics are rendered precisely linear—typically with prescribed convergence rates—either in a reduced coordinate space (functional observers) or in full-order nonlinear settings, including continuous-time, discrete-time, and both linear and nonlinear system classes. Linearization is achieved by identifying a nonlinear (generally smooth or analytic) transformation on the original state or functional variables, under which the observer error evolves according to a linear dynamics chosen by the designer. The framework generalizes classical Luenberger observer theory for linear time-invariant (LTI) systems to much broader contexts and is now realized in both analytic and data-driven (e.g., neural network–based) paradigms, as well as within purely algebraic structures designed for large-scale or non-square systems.

1. Core Mathematical Principles

At the foundation of exact observer linearization is the construction of an observer whose estimation error evolves according to a user-specified linear differential or difference equation. For a smooth, unforced nonlinear system

$$\dot{x}=F(x),\quad y=H(x)\in\mathbb{R}^p, \quad z=q(x)\in\mathbb{R}$$

the problem is to estimate a scalar functional $q(x(t))$ using a lower-order observer system: $\dot{\xi} = A\xi + By, \quad \hat{z} = C\xi + Dy$ together with a smooth immersion $T:\mathbb{R}^n\to\mathbb{R}^v$ satisfying the invariance condition

$\frac{\partial T}{\partial x} F(x) = AT(x) + BH(x)$

and the output reconstruction

$q(x) = CT(x) + DH(x)$

This ensures that the observer error $\tilde{e}:=T(x)-\xi$ evolves according to $\dot{\tilde{e}}=A\tilde{e}$ and is thus globally linear in the observer coordinates, with the decay rate specified by the spectrum of $A$ (Kravaris et al., 2021).

In the discrete-time setting,

$x_{t+1} = \Phi(x_t),\quad y_t = h(x_t)$

an exact linearizing state transformation $T$ is sought, solving the functional equation

$T(\Phi(x)) = AT(x) + b(h(x)),\quad T(0) = 0$

with $A$ required to be Schur (all eigenvalues within the unit circle), ensuring error convergence (Alvarez et al., 2024).

2. Existence Conditions and Algebraic Criteria

A central question is under what conditions an exact observer linearization is possible.

• Functional Observer Lie-Derivative Test: For the continuous-time nonlinear case, the paper (Kravaris et al., 2021) provides a necessary and sufficient criterion. Given a desired characteristic polynomial

$N(s)=s^v+a_1s^{v-1}+\dots+a_v$

an exact $v$-dimensional observer exists if and only if

$$L_F^{\nu}q(x) + a_1 L_F^{\nu-1} q(x) + \cdots + a_\nu q(x) = \sum_{k=0}^{\nu-1} B_k L_F^k H(x)$$

for some row vectors $B_k$, where $L_F^k$ denotes the $k$-th Lie derivative along $F$.

• Linear System Output Invariance: For linear systems with $y=Hx$, an exact observer-based realization exists if and only if the output row space is invariant under $A$:

$$\operatorname{Row}(H)A \subseteq \operatorname{Row}(H)$$

ensuring the existence of $\Xi$ such that $HA=\Xi H$ (Cheng et al., 2024).

• Discrete-Time Nonlinear Analyticity and Non-Resonance: Under analytic system maps, local observability, and non-resonance between system and observer eigenvalues (no integer vector $m$ and index $j$ with $\prod_i k_i^{m_i} = \lambda_j$), a unique, locally invertible analytic $T$ solving the observer linearization equation exists (Alvarez et al., 2024).

3. Observer Construction Methodologies

Construction in the exact observer linearization framework proceeds according to these canonical steps:

1. Polynomial Specification: The designer specifies the target eigenstructure or characteristic polynomial for desired convergence rates.
2. Solving Algebraic Matching or Functional Equations:
   • For continuous-time functional observers, one solves the Lie-derivative equation for $B_k$ (Kravaris et al., 2021).
   • In the discrete-time analytic case, the inhomogeneous functional equation for $T$ is solved, either analytically or (intractable in high-dimensions) via approximation (Alvarez et al., 2024).
3. Coordinate Map Construction: A coordinate transformation $T$ is explicitly constructed based on recursive application of Lie derivatives (continuous-time) or via mesh-based/PINN methods (discrete-time).
4. Observer Design: Observer matrices $(A,B,C,D)$ are formed from the polynomial coefficients and solution vectors, or, in the linear case, from the solution to $HA=\Xi H$ (Cheng et al., 2024).

An explicit formula for $T(x)$ is provided in (Kravaris et al., 2021), detailed step-wise in their construction section.

A concise summary table from (Kravaris et al., 2021):

Step	Continuous-Time (Functional Observer)	Discrete-Time (Full-State PINN)
Problem Variables	$F(x), H(x), q(x)$	$\Phi(x)$, $h(x)$
Transformation $T$	Coordinates in Lie-derived basis, explicit formula	Neural network (PINN) approximation
Linearization Equation	Algebraic Lie-derivative matching $(\star)$	$T(\Phi(x)) = AT(x) + b(h(x))$ (inhomogeneous)
Existence Condition	Lie-derivative span test	Analyticity, local observability, non-resonance
Error Dynamics	$\dot{\tilde{e}}=A \tilde{e}$	$e_z(t+1) = A e_z(t)$

4. Algorithmic and Neural Approaches

Recent advances have enabled exact observer linearization to be achieved using neural function approximators. For nonlinear discrete-time systems, Physics-Informed Neural Networks (PINNs) can be used to learn the requisite state transformation $T_\theta(x)$ that solves the functional equation

$$T_\theta(\Phi(x)) = A T_\theta(x) + b(h(x)),\quad T_\theta(0)=0$$

The network is trained via minimization of a loss comprising a mesh-based enforcement of the transformation equation, the initial value pinning, and a Jacobian constraint at the origin obtained from a linearized "pinning" equation. Training proceeds on nested domains via greedy continuation to handle singularities and improve convergence (Alvarez et al., 2024).

This methodology is demonstrated to outperform traditional local Taylor expansion (power-series) solutions, particularly near domain boundaries and in the presence of steep gradients or singular points, where analytical expansions fail to deliver accurate approximations. Uncertainty quantification is performed through multiple independent network trainings and evaluating error bands on a Chebyshev–Lobatto grid.

5. Algebraic, Geometric, and Operator-Theoretic Extensions

The observer-based realization (OR) framework generalizes the exact observer linearization paradigm to arbitrary linear and affine nonlinear systems using the dimension-keeping semi-tensor product (DK-STP) and Lebesgue-type output-based system descriptions (Cheng et al., 2024).

• DK-STP Product: The DK-STP allows defining observer dynamics directly on output trajectories irrespective of state-space dimension, providing algorithms both for approximate and exact output-side realizations.
• Exactness and Invariance: Existence of an exact OR-system is equivalent to output-rowspace invariance under system dynamics, with extended OR-systems constructed via $A$-invariant closure or geometric algorithms when invariance does not hold.
• Nonlinear Extension via Lie-Derivative Invariance: For affine nonlinear systems,

$\dot{x}=f(x)+\sum_j g_j(x) u_j,\quad y_j = h_j(x)$

if there exist matrices $\Xi_0, \Xi_i$ such that

$L_f(dH) = \Xi_0 dH,\quad L_{g_i}(dH) = \Xi_i dH$

then the observer-based realization

$\dot{y} = \Xi_0 y + \sum_i \Xi_i u_i$

is exact; otherwise, one extends outputs to the involutive closure of the codistribution (Cheng et al., 2024).

6. Comparative Analysis, Applications, and Practical Considerations

Advantages

• Arbitrary Linear Dynamics: The designer can impose observer error spectra using suitable $A$ (Hurwitz/Schur).
• Reduction in Observer Order: It is possible (especially for functional estimation) to construct observers of dimension $v < n$, or even significantly less than $n-p$, surpassing standard reduced-order observer theory (Kravaris et al., 2021).
• Direct Nonlinear Generalization: The above frameworks constitute nonlinear generalizations of Luenberger theory.
• Unified Realization Theory: The OR-system approach offers a purely algebraic and geometric perspective applicable across linear, affine, and some nonlinear classes (Cheng et al., 2024).

Limitations

• Restrictive Existence Conditions: Lie-derivative span (functional observers) and output invariance (OR-systems) tests are restrictive and often rule out linearization for typical functionals or outputs.
• Requirement of High-Order Derivatives/Functional Equations: Explicit construction involves high-order Lie derivatives or solving (possibly intractable) functional equations.
• Nonlinear Output Maps: Extensions to nonlinearly parameterized outputs or observer maps demand more elaborate conditions (see "Extensions" in (Kravaris et al., 2021)).

Example Application

A detailed example in (Kravaris et al., 2021) considers a CSTR (Continuous Stirred-Tank Reactor) with four states and two outputs, constructing a scalar observer for $z=c_A+c_B$ with exact linear error dynamics, exploiting system structure to realize $v=1$.

Neural Implementation Benchmarks

(Alvarez et al., 2024) reports on two discrete-time benchmarks where neural observers outperform power-series solutions, delivering uniformly accurate coordinate transformations even near singularities.

7. Extensions and Generalizations

Extensions of exact observer linearization include

• Observers involving nonlinear functions of outputs [(Kravaris et al., 2021), Section 7].
• Incorporation of invertible nonlinear output correction maps to relax the strict observer output linearity requirement.
• Feedback-extended OR-systems (minimal realization under feedback), with algorithms leveraging geometric control invariants (Cheng et al., 2024).
• Data-driven construction of global observer transformations (e.g., via PINNs), circumventing the limitations of analytic or Taylor-based approaches (Alvarez et al., 2024).
• Algebraic machinery via DK-STP and Lebesgue-type output system formulations enables systematic exact realization and minimality results for both linear and certain affine nonlinear systems.

A plausible implication is that the integration of neural approaches with algebraic/geometric observer theory may broaden the applicability of exact observer linearization, especially in large-scale and high-dimensional control and estimation problems.

---

Principal References

• Functional observer linearization framework, existence, and construction: (Kravaris et al., 2021)
• Neural network (PINN)–based observer linearization for discrete-time nonlinear systems: (Alvarez et al., 2024)
• Algebraic/geometric observer-based realizations (OR-systems), DK-STP, and nonlinear system extension: (Cheng et al., 2024)