Koopman Observability of Nonlinear Systems and Data-Driven Koopman--Luenberger Observer

arXiv:2609.36334 · math.OC · Submitted 2026-09-28 · Read on arXiv

Listen

Radio episode about this paper

Transcript

Introduction to the show: ident: Mathematics. The day in mathematics: new theorems, new proofs, and what they open.

Ada: We're looking at the paper: "Koopman Observability of Nonlinear Systems and Data-Driven Koopman--Luenberger Observer".

Paper summary: Ada: So we're looking at this paper, "Koopman Observability of Nonlinear Systems and Data-Driven Koopman--Luenberger Observer." It tackles the idea that nonlinear systems can actually be treated like infinite-dimensional linear systems through something called Koopman semigroups. This paper is saying that while you can describe these things as infinite-dimensional linear systems, creating state observers using learned Koopman models in a way similar to finite-dimensional linear systems is still a tough problem.

Emmy: It claims they provide concepts for both Koopman observability and detectability and also a data-driven method for synthesizing this specific kind of observer, which is the Koopman–Luenberger observer for continuous-time autonomous nonlinear systems. It sets up the framework by defining the Koopman semigroup on a reproducing kernel Hilbert space, or RKHS, which are these function spaces where they talk about strong continuity of the semigroup.

Kolya: So what does that mean in practice? They establish conditions for Koopman exponential detectability by setting up an operator Lyapunov inequality. The core idea is that this inequality only implicitly involves the infinitesimal generator of the Koopman semigroup, which is the stuff that describes how things evolve in infinite dimensions.

Ada: And then they take a finite-horizon Yosida approximation of that Koopman generator and use an empirical sampling operator to solve for the observer gain. This data-driven solution is reduced to a finite-dimensional subspace defined by the sample, which means it can be handled by convex semidefinite programming.

Emmy: It’s interesting because they show that even with this infinite-dimensional setup, when you use a finite order observer based on the solution of that empirical Lyapunov inequality, the state observation error actually converges to zero asymptotically.

Kolya: That convergence is important for us because it suggests a way to make these observers work reliably in real scenarios where we don't know the exact underlying dynamics beforehand. The question for me is what kind of data would be needed to make that empirical sampling operator estimation accurate enough.

Conclusion: Ada: Looking at this paper, "Koopman Observability of Nonlinear Systems and Data-Driven Koopman--Luenberger Observer," the authors are providing concrete concepts for observability and detectability within the Koopman framework for nonlinear systems. The main thing they claim is that by defining things on an RKHS, they can establish conditions for exponential detectability by solving an operator Lyapunov inequality.

Emmy: And then they use a finite-horizon Yosida approximation to solve this inequality data-driven, reducing the problem to convex semidefinite programming based on sample data. They also contrast their Koopman–Luenberger observer with other types, noting that while both use linear dynamics, the retrieval of state from the infinite-dimensional space in the Koopman version is completely linear.

Kolya: For me, what this means is that they're offering a pathway to synthesize these observers without having to estimate every single Koopman eigenfunction explicitly. The method relies on estimating just enough of the generator through data trajectories to get a finite-dimensional subspace for the observer gain.

Ada: It implies that instead of getting stuck with cumbersome nonlinear pseudoinverse mappings, which some other methods require, this approach boils down to a convex optimization problem when synthesizing the observer gain. That's a significant structural difference in how they handle the synthesis.

Emmy: The authors show performance against linear model-based, nonlinear model-based, and neural network observers across three different system types: asymptotically stable, limit cycle, and chaotic systems. Their results suggest this Koopman–Luenberger observer achieves better performance in these examples than those other methods tested.

Kolya: It’s good to see that empirical data can drive a solution that actually performs better than purely model-based approaches when dealing with complex dynamics like limit cycles or chaos. But we have to remember the paper states their limitation: the synthesis relies on approximating the Koopman generator using a finite-horizon Yosida form, and they need an empirical sampling operator based on data trajectories to estimate that generator.

Ada: So, while it offers a practical, data-driven way to synthesize these observers for continuous-time autonomous nonlinear systems, the success of their method hinges on how well that approximation captures the true dynamics of the system. That's what we need to keep an eye on for future work.

WENTAO TANG

North Carolina State University

math.OC

Submitted: 2026-09-28

Updated: 2026-11-18

Comments: 16 pages, 4 figures, 1 table. Submitted to IEEE Transactions on Automatic Control (2026-11-18)

Code: https://github.com/WentaoTang-Pack/KoopmanLuenbergerObserver

License: http://creativecommons.org/licenses/by-sa/4.0/

Importance score: 74/100

The gist: While nonlinear systems can be generally described as infinite-dimensional linear systems via their Koopman semigroups…

Key concepts

Koopman Semigroup
This concept describes how the dynamics of a nonlinear system can be represented as an infinite-dimensional linear system. It involves mapping the original nonlinear state space onto a function space where the evolution is governed by a linear operator, allowing for analysis using standard linear system techniques.
Koopman Observability and Detectability
These are criteria used to determine if it is possible to reconstruct the state of a nonlinear system from its observed outputs. The paper defines these concepts within a specific function space (RKHS) and provides conditions for synthesizing an infinite-dimensional observer gain.
Operator Lyapunov Inequality
This is the core mathematical tool used to synthesize the observer gain. By solving this inequality, which is approximated using data trajectories, one can derive a finite-order observer that ensures the state estimation error converges to zero asymptotically.
Koopman–Luenberger Observer vs. KKL
The Koopman–Luenberger observer uses LTI dynamics for retrieval in the infinite-dimensional space, making the state recovery linear. This contrasts with KKL, which requires a nonlinear mapping for state retrieval and involves solving a nonlinear pseudoinverse, whereas the Koopman approach is framed as a convex optimization problem.

Terminology

Summary

While nonlinear systems can be generally described as infinite-dimensional linear systems via their Koopman semigroups, the direct synthesis of state observers on the premise of learned Koopman operator models, in an analogous manner to finite-dimensional linear systems, remains an open problem. p. 0

Where the paper says it

KOOPMAN OBSERVABILITY OF NONLINEAR SYSTEMS AND DATA-DRIVEN KOOPMAN–LUENBERGER OBSERVER ∗ WENTAO TANG† Abstract. While nonlinear systems can be generally described as infinite-dimensional linear systems via their Koopman semigroups, the direct synthesis of state observers on the premise of learned Koopman operator models, in an analogous manner to finite-dimensional linear systems, remains an open problem.

This paper provides concepts for Koopman observability and detectability and a data-driven method for synthesizing a Koopman–Luenberger observer for continuous-time autonomous nonlinear systems. p. 1
Where the paper says it

This paper focuses on constructing a Koopman theory and method of state observation for continuous-time autonomous nonlinear systems. 2 Specifically, Koopman observability and detectability are defined in a similar way to those in infinite-dimensional linear systems theory [13], but accounting for the fact that the observation target is only the finite-dimensional state instead of its “lifted” representation in the infinitedimensional RKHS.

Koopman Theory Foundations

The paper introduces the Koopman semigroup of system (2.1) as defined on a space G of state-dependent functions, and discusses the choice of Hilbert spaces like Hs(X) or H˚s(X). p. 3

Where the paper says it

The choice of function class G obviously depends on the regularity of f. We are in particular interested in Hilbert spaces, as infinite-dimensional linear systems theory typically adopts a Hilbert space formulation [13].

It establishes that under suitable conditions, the Koopman semigroup is strongly continuous on these function spaces. p. 1
Where the paper says it

2 Specifically, Koopman observability and detectability are defined in a similar way to those in infinite-dimensional linear systems theory [13], but accounting for the fact that the observation target is only the finite-dimensional state instead of its “lifted” representation in the infinitedimensional RKHS. Under suitable conditions, a Koopman–Luenberger observer, specified by its observer gain as a bounded linear operator on the RKHS, is expected to be obtained by solving an operator Lyapunov inequality, whose numerical solution is to be approximated from a data-based estimate of the infinitesimal generator of the Koopman semigroup.

The paper also details the representation of the system (2. p. 1
Where the paper says it

This paper focuses on constructing a Koopman theory and method of state observation for continuous-time autonomous nonlinear systems. 2 Specifically, Koopman observability and detectability are defined in a similar way to those in infinite-dimensional linear systems theory [13], but accounting for the fact that the observation target is only the finite-dimensional state instead of its “lifted” representation in the infinitedimensional RKHS.

1) in terms of an infinite-dimensional linear system (2. p. 0
Where the paper says it

The operator-theoretic approaches for nonlinear systems have gained attention of researchers in the recent decade [30, 34, 11]. Using the Koopman operator (in discrete time) or Koopman semigroup (in continuous time), a nonlinear system can be viewed as an infinite-dimensional linear systems on a function space.

5) involving a state evaluation operator E. p. 6
Where the paper says it

5) ˙z = Az, y = Cz, x = Ez. The unbounded operator A generates a strongly continuous semigroup {T t} on Z, C ∈ B(Z, ℝm) is the measurement operator, and E ∈ B(Z, ℝd ) is the state evaluation operator.

Observability and Detectability Criteria

The paper defines Koopman observability and detectability by placing the Koopman semigroup in a properly chosen RKHS. It provides explicit conditions for the existence and synthesis of an infinite-dimensional observer gain (operator) boiling down to an operator Lyapunov inequality on the RKHS. p. 2

Where the paper says it

The explicit conditions in the case where the Koopman semigroup has a discrete spectrum are given (Subsection 3. (ii) Explicit conditions for the existence and synthesis of an infinite-dimensional observer gain (operator) are provided, boiling down to an operator Lyapunov inequality on the RKHS that involves the infinitesimal generator of the Koopman semigroup (Subsection 3.

For practical implementation, it proves that a finite-order observer using the solution of the empirical Lyapunov inequality leads to a state observation error that converges to zero asymptotically. p. 2
Where the paper says it

(iv) For practical implementation, the infinite-order observer must be approximated by a finite order one. It is proved that the finite-order observer using the solution of the empirical Lyapunov inequality leads to a state observation error that converges to zero asymptotically (Subsection 4.

Data-Driven Synthesis Methodology

The synthesis of the observer gain is achieved by solving an operator Lyapunov inequality, which is approximated in a Yosida form and further projected to a finite-order form. This approximation relies on data trajectories to estimate the infinitesimal generator of the Koopman semigroup. p. 1

Where the paper says it

2 Specifically, Koopman observability and detectability are defined in a similar way to those in infinite-dimensional linear systems theory [13], but accounting for the fact that the observation target is only the finite-dimensional state instead of its “lifted” representation in the infinitedimensional RKHS. Under suitable conditions, a Koopman–Luenberger observer, specified by its observer gain as a bounded linear operator on the RKHS, is expected to be obtained by solving an operator Lyapunov inequality, whose numerical solution is to be approximated from a data-based estimate of the infinitesimal generator of the Koopman semigroup.

The empirical version of the Lyapunov inequality is formulated as a compactified inequality (4. p. 2
Where the paper says it

(iii) The Koopman generator, as an unbounded operator, is related to two integral operators, and via Yosida approximation, the operator Lyapunov inequality is converted into a version with only compact operators. With finite data trajectories, the empirical version of the Lyapunov inequality is provided (Subsection 4.

6) which can be solved by reducing it to a finite-dimensional matrix inequality (4. p. 0
Where the paper says it

By defining the Koopman semigroup on a reproducing kernel Hilbert space (RKHS), the conditions for Koopman exponential detectability are established, which involves solving an operator Lyapunov inequality that only implicitly involves the the infinitesimal generator of the Koopman semigroup. Through the use of a finite-horizon Yosida approximation of the Koopman generator and an empirical sampling operator, the data-driven solution of the observer gain can be reduced to a finite-dimensional subspace specified by the sample and hence fulfilled by convex semidefinite programming.

14).

Performance and Comparison

The paper demonstrates the performance of the proposed Koopman–Luenberger observer against linear model-based, nonlinear model-based, and neural network observers across three different system types (asymptotically stable, limit cycle, chaotic). The results show that the Koopman–Luenberger observer achieves a better performance in these examples than the other methods. p. 2

Where the paper says it

The elementary expressions in data-based finite-dimensional space for the operator computations are given (Subsection 4. (v) The Koopman–Luenberger observer synthesis is applied to different cases in Section 5, confirming its performance despite complicated state-space geometries.

Conceptual Distinctions

The paper contrasts the Kazantzis–Kravaris/Luenberger observer with the Koopman–Luenberger observer by noting that while both use LTI dynamics, the retrieval of state from the infinite-dimensional state in Koopman-Luenberger is completely linear, whereas in KKL it must be a nonlinear mapping. The synthesis in Koopman–Luenberger boils down to a convex optimization problem, unlike KKL which requires solving a nonlinear pseudoinverse mapping. p. 1

Where the paper says it

Under suitable conditions, a Koopman–Luenberger observer, specified by its observer gain as a bounded linear operator on the RKHS, is expected to be obtained by solving an operator Lyapunov inequality, whose numerical solution is to be approximated from a data-based estimate of the infinitesimal generator of the Koopman semigroup. As such, a convex optimization-based approach is provided for the synthesis of a fully linear Luenberger-type observer, which is ideally infinite-order and realistically approximated by data as a finite-order one.

Green — the paper says it, and the page is printed. Red — no passage carries it, so treat that sentence as unverified. The same check runs on your own text: check a summary against its source, $5.

Improvements for AI systems

  1. Data driven observer synthesis can be achieved by reducing it to a finite dimensional subspace specified by the sample, which is fulfilled by convex semidefinite programming. This allows for a novel type of observer for nonlinear systems that is ideally infinite-order and realistically approximated as a finite-order one.

  2. The system can be modeled as an infinite dimensional linear system governed by the infinitesimal generator of the Koopman semigroup on a reproducing kernel Hilbert space, allowing for the definition of Koopman observability and detectability conditions.

  3. An observer gain can be synthesized by solving an operator Lyapunov inequality on the RKHS that involves the infinitesimal generator of the Koopman semigroup, which is then approximated from a data based estimate of this generator.

  4. The finite order observer uses the solution of the empirical Lyapunov inequality to guarantee a state observation error that converges to zero asymptotically, providing a practical implementation for nonlinear systems.

  5. The proposed Koopman Luenberger observer achieves performance that is better than linearized model based Luenberger observers and neural network KKL observers within more than one order of magnitude lower training time budget.

Abstract

While nonlinear systems can be generally described as infinite-dimensional linear systems via their Koopman semigroups, the direct synthesis of state observers on the premise of learned Koopman operator models, in an analogous manner to finite-dimensional linear systems, remains an open problem. In this paper, the concepts of Koopman observability, Koopman detectability, and Koopman--Luenberger observer of nonlinear systems are provided. By defining the Koopman semigroup on a reproducing kernel Hilbert space (RKHS), the conditions for Koopman exponential detectability are established, which involves solving an operator Lyapunov inequality that only implicitly involves the infinitesimal generator of the Koopman semigroup. Through the use of a finite-horizon Yosida approximation of the Koopman generator and an empirical sampling operator, the data-driven solution of the observer gain can be reduced to a finite-dimensional subspace specified by the sample and hence fulfilled by convex semidefinite programming (SDP). The state observation error is guaranteed to be bounded in a mean-squares sense, under closed-loop dissipation conditions on the sampling operator and its empirical operator.

Sources

Related papers