pith. sign in

arxiv: 2605.01835 · v1 · submitted 2026-05-03 · 💻 cs.LG

Learning Koopman operators for coupled systems via information on governing equations of subsystems

Tatsuya Naoi , Jun Ohkubo This is my paper

Pith reviewed 2026-05-09 17:22 UTC · model grok-4.3

classification 💻 cs.LG
keywords Koopman operatorcoupled systemsextended dynamic mode decompositionsubsystem equationsnonlinear dynamicsdata-driven modelingoperator learning
0
0 comments X

The pith

Incorporating known differential equations of subsystems yields more stable Koopman operator approximations for coupled nonlinear systems than pure data-driven EDMD.

A machine-rendered reading of the paper's core claim, the machinery that carries it, and where it could break.

The paper develops a method for learning the Koopman operator in nonlinear coupled systems by using the known governing differential equations of each subsystem. Pure data-driven techniques like extended dynamic mode decomposition often become unstable or inaccurate when data is limited for these high-dimensional interacting systems. The new approach integrates the subsystem equations into the learning process to leverage this partial knowledge. A sympathetic reader would care because this could enable reliable modeling of complex real-world systems such as coupled oscillators with fewer measurements. If the method works, it bridges physics-based knowledge with data-driven operator learning.

Core claim

The authors present a procedure to approximate the Koopman operator for coupled systems that incorporates information from the differential equations governing each subsystem. This is achieved by modifying the extended dynamic mode decomposition method to include the subsystem dynamics, resulting in improved stability and accuracy in numerical tests on coupled oscillator systems, especially under limited data conditions.

What carries the argument

A modified extended dynamic mode decomposition (EDMD) framework that incorporates the governing equations of individual subsystems as additional information during the operator approximation.

Load-bearing premise

The governing differential equations of the individual subsystems are known accurately and can be incorporated into the operator learning without introducing inconsistencies or new instabilities in the coupled system.

What would settle it

Numerical experiments on coupled oscillator systems showing that the proposed method produces less accurate or more unstable predictions than standard EDMD under limited data would falsify the claim.

Figures

Figures reproduced from arXiv: 2605.01835 by Jun Ohkubo, Tatsuya Naoi.

Figure 1
Figure 1. Figure 1: FIG. 1. Schematic diagram of the proposed method. The view at source ↗
Figure 2
Figure 2. Figure 2: FIG. 2. Means and standard deviations of one-step-ahead view at source ↗
Figure 3
Figure 3. Figure 3: FIG. 3. Prediction errors of view at source ↗
Figure 5
Figure 5. Figure 5: FIG. 5. Means and standard deviations of the one-step-ahead view at source ↗
Figure 6
Figure 6. Figure 6: FIG. 6. Prediction errors of view at source ↗
read the original abstract

Nonlinear coupled systems are ubiquitous in science and engineering. The analysis and modeling of such systems is challenging due to their high dimensionality and complex interactions among subsystems. In recent years, operator-theoretic methods based on the Koopman operator have attracted attention as a powerful tool for analyzing and modeling nonlinear dynamical systems. Extended dynamic mode decomposition (EDMD) is one of the most popular methods to approximate the Koopman operator. However, EDMD is a purely data-driven method, and it could be unstable and inaccurate for coupled systems under limited data availability. In this paper, we propose a method to learn the Koopman operator for coupled systems using the differential equations governing each subsystem. We also demonstrate its effectiveness through numerical experiments on coupled oscillator systems.

Editorial analysis

A structured set of objections, weighed in public.

Desk editor's note, referee report, simulated authors' rebuttal, and a circularity audit. Tearing a paper down is the easy half of reading it; the pith above is the substance, this is the friction.

Referee Report

2 major / 2 minor

Summary. The paper proposes a hybrid variant of Extended Dynamic Mode Decomposition (EDMD) to approximate the Koopman operator for nonlinear coupled systems. It incorporates the known governing differential equations of individual subsystems to improve stability and accuracy when data is limited, and validates the approach via numerical experiments on coupled oscillator systems.

Significance. If the hybrid injection of subsystem DEs can be shown to preserve consistency with unknown coupling terms, the method would offer a practical bridge between physics-informed and data-driven operator learning for high-dimensional systems. This could improve modeling in applications like multi-body dynamics or networked oscillators where subsystem equations are available but full-system trajectories are scarce. The experiments provide initial support, but the absence of error bounds limits the strength of the claim.

major comments (2)
  1. [Method section (following the abstract)] The method description does not derive or bound how the subsystem DEs are encoded (hard constraints, soft penalties, or dictionary modification) while remaining consistent under the unknown coupling. This is load-bearing for the central claim that the resulting operator remains stable and accurate for the full coupled system.
  2. [Numerical experiments section] No a-priori perturbation analysis or sensitivity bound is given for the effect of coupling terms on the learned operator when the injected DEs are only approximately known. Experiments on oscillators succeed under perfect knowledge but do not test this fragility.
minor comments (2)
  1. [Abstract] The abstract would be strengthened by naming the specific incorporation technique and reporting at least one quantitative metric (e.g., prediction error or eigenvalue stability) from the oscillator experiments.
  2. [Method] Notation for the augmented dictionary or loss function should be introduced with an explicit equation rather than prose description to aid reproducibility.

Simulated Author's Rebuttal

2 responses · 0 unresolved

We thank the referee for the constructive and detailed feedback on our manuscript. We address each major comment below and will incorporate the suggested improvements in the revised version to strengthen the theoretical foundations and experimental validation.

read point-by-point responses

  1. Referee: The method description does not derive or bound how the subsystem DEs are encoded (hard constraints, soft penalties, or dictionary modification) while remaining consistent under the unknown coupling. This is load-bearing for the central claim that the resulting operator remains stable and accurate for the full coupled system.

    Authors: We agree that a rigorous derivation is necessary to support the central claim. In the revised manuscript, we will expand the Method section with a detailed mathematical derivation showing that the subsystem governing equations are incorporated via a dictionary augmentation combined with a soft penalty term in the EDMD least-squares objective. This formulation ensures consistency with the unknown coupling terms because the data-driven regression on full-system trajectories is retained as the primary fitting mechanism, while the penalty enforces the known subsystem dynamics. We will also derive an a-priori bound on the operator approximation error under the assumption of Lipschitz-continuous coupling, using standard results from EDMD convergence theory. revision: yes

  2. Referee: No a-priori perturbation analysis or sensitivity bound is given for the effect of coupling terms on the learned operator when the injected DEs are only approximately known. Experiments on oscillators succeed under perfect knowledge but do not test this fragility.

    Authors: We acknowledge that the current experiments assume exact knowledge of the subsystem equations, which is a limitation for broader applicability. In the revision, we will add a new theoretical subsection providing a perturbation analysis based on operator perturbation theory for the learned Koopman matrix. This will include explicit sensitivity bounds on the eigenvalues and eigenfunctions with respect to perturbations in the injected DEs. We will also augment the Numerical experiments section with additional tests on the coupled oscillator systems where the subsystem equations are corrupted by additive noise or parametric uncertainty, demonstrating that the hybrid method remains more stable than pure EDMD even under approximate knowledge. revision: yes

Circularity Check

0 steps flagged

No significant circularity; method uses external subsystem DEs as independent inputs.

full rationale

The paper presents a hybrid EDMD variant that injects known governing differential equations of individual subsystems as prior information to learn the Koopman operator for the full coupled system. This incorporation is described as using external knowledge rather than deriving or fitting the subsystem equations from the coupled data itself. No self-definitional loops, fitted parameters renamed as predictions, or load-bearing self-citations appear in the abstract or method outline. The central claim remains independent of its own outputs, with subsystem DEs serving as verifiable external constraints.

Axiom & Free-Parameter Ledger

0 free parameters · 0 axioms · 0 invented entities

No details available from the abstract to identify specific free parameters, axioms, or invented entities; the proposal relies on standard Koopman theory and known subsystem equations.

pith-pipeline@v0.9.0 · 5417 in / 965 out tokens · 41016 ms · 2026-05-09T17:22:59.463144+00:00 · methodology

discussion (0)

Sign in with ORCID, Apple, or X to comment. Anyone can read and Pith papers without signing in.

Reference graph

Works this paper leans on

42 extracted references · 42 canonical work pages

  1. [1]

    Parameters of the subsystems are set as follows:    δ1 = 0.23, α 1 =−0.99, β 1 = 0.80, δ2 = 0.15, α 2 =−0.59, β 2 = 0.86, δ3 = 0.11, α 3 =−0.54, β 3 = 0.77

    The neighbor sets are given byN 1 ={2},N 2 ={1,3}, andN 3 ={2}. Parameters of the subsystems are set as follows:    δ1 = 0.23, α 1 =−0.99, β 1 = 0.80, δ2 = 0.15, α 2 =−0.59, β 2 = 0.86, δ3 = 0.11, α 3 =−0.54, β 3 = 0.77. (32) The initial state,x∈R 6, is set tox∼U(−1.5,1.5) 6. Then, we generate a time-series dataset of length 5001 with a time step of ...

    work page 2000

  2. [2]

    (22), and then we obtain the following equation: ∂ ∂t ϕ(x, t) =x 2 ∂ ∂x1 ϕ(x, t) + (−δx2 −αx 1 −βx 3

    ∂ ∂x2 .(A2) We substituteL † into Eq. (22), and then we obtain the following equation: ∂ ∂t ϕ(x, t) =x 2 ∂ ∂x1 ϕ(x, t) + (−δx2 −αx 1 −βx 3

    work page

  3. [3]

    (A3) Next, we substitute Eq

    ∂ ∂x2 ϕ(x, t). (A3) Next, we substitute Eq. (24) into Eq. (A3). Then, we obtain the following equation: ∂ ∂t X n c(n, t)ψn(x) =x 2 ∂ ∂x1 X n c(n, t)ψn(x) + (−δx2 −αx 1 −βx 3

    work page

  4. [4]

    Hence, the multi-indexnis expressed asn= (n 1, n2), i.e

    ∂ ∂x2 X n c(n, t)ψn(x).(A4) Here, the Duffing equation has two variables. Hence, the multi-indexnis expressed asn= (n 1, n2), i.e. ψn(x) =x n1 1 xn2 2 andc(n, t) =c(n 1, n2, t). Then, we rewrite Eq. (A4) as follows: ∂ ∂t X n1,n2 c(n1, n2, t)xn1 1 xn2 2 =x 2 ∂ ∂x1 X n1,n2 c(n1, n2, t)xn1 1 xn2 2 + (−δx2 −αx 1 −βx 3

    work page

  5. [5]

    (A5) Next, we calculate the derivatives on the right-hand side of the above equation, and then we compare the coef- ficients of the monomial functions on both sides

    ∂ ∂x2 X n1,n2 c(n1, n2, t)xn1 1 xn2 2 . (A5) Next, we calculate the derivatives on the right-hand side of the above equation, and then we compare the coef- ficients of the monomial functions on both sides. After some calculations, we obtain the following equation: X n1,n2 ∂ ∂t c(n1, n2, t) = X n1,n2 h (n1 + 1)c(n1 + 1, n2 −1, t) −δn 2c(n1, n2, t) −α(n 2 +...

    work page

  6. [6]

    S. H. Strogatz, Physica D143, 1 (2000)

    work page 2000

  7. [7]

    K. S. Fink, G. Johnson, T. Carroll, D. Mar, and L. Pec- ora, Phys. Rev. E61, 5080 (2000)

    work page 2000

  8. [8]

    J. A. Acebr´ on, L. L. Bonilla, C. J. P´ erez Vicente, F. Ri- tort, and R. Spigler, Rev. Mod. Phys.77, 137 (2005)

    work page 2005

  9. [9]

    Mirzaei, M

    S. Mirzaei, M. S. Anwar, F. Parastesh, S. Jafari, and D. Ghosh, Phys. Rev. E107, 014201 (2023)

    work page 2023

  10. [10]

    Bialonski, G

    S. Bialonski, G. Ansmann, and H. Kantz, Phys. Rev. E 92, 042910 (2015)

    work page 2015

  11. [11]

    R. Xiao, L. W. Kong, Z. K. Sun, and Y. C. Lai, Phys. Rev. E104, 014205 (2021)

    work page 2021

  12. [12]

    Ghadami and B

    A. Ghadami and B. I. Epureanu, Philos. Trans. R. Soc. A380, 20210213 (2022)

    work page 2022

  13. [13]

    S. L. Brunton, J. L. Proctor, and J. N. Kutz, Proc. Natl. Acad. Sci. U.S.A.113, 3932 (2016)

    work page 2016

  14. [14]

    S. H. Rudy, S. L. Brunton, J. L. Proctor, and J. N. Kutz, Sci. Adv.3, e1602614 (2017)

    work page 2017

  15. [15]

    B. O. Koopman, Proc. Natl. Acad. Sci. U.S.A.17, 315 (1931)

    work page 1931

  16. [16]

    S. L. Brunton, M. Budiˇ si´ c, E. Kaiser, and J. N. Kutz, SIAM Rev.64, 229 (2022)

    work page 2022

  17. [17]

    Williams, I

    M. Williams, I. Kevrekidis, and C. Rowley, J. Nonlinear Sci.25(2014)

    work page 2014

  18. [18]

    Mezi´ c, Nonlinear Dyn.41, 309 (2005)

    I. Mezi´ c, Nonlinear Dyn.41, 309 (2005)

    work page 2005

  19. [19]

    Mezi´ c, Annu

    I. Mezi´ c, Annu. Rev. Fluid Mech.45, 357 (2013)

    work page 2013

  20. [20]

    Susuki and I

    Y. Susuki and I. Mezi´ c, IEEE Trans. Power Syst.29, 899 (2014)

    work page 2014

  21. [21]

    Mauroy and I

    A. Mauroy and I. Mezi´ c, IEEE Trans. Autom. Control 61, 3356 (2016)

    work page 2016

  22. [22]

    J. L. Proctor, S. L. Brunton, and J. N. Kutz, SIAM J. Appl. Dyn. Syst.15, 142 (2016)

    work page 2016

  23. [23]

    S. L. Brunton, B. W. Brunton, J. L. Proctor, and J. N. Kutz, PLOS ONE11, 1 (2016)

    work page 2016

  24. [24]

    Schlosser and M

    C. Schlosser and M. Korda, SIAM J. App. Dyn. Syst.21, 2187 (2022)

    work page 2022

  25. [25]

    Tellez-Castro, C

    D. Tellez-Castro, C. Garcia-Tenorio, E. Mojica-Nava, J. Sofrony, and A. V. Wouwer, Actuators11, 151 (2022)

    work page 2022

  26. [26]

    Y. Guo, M. Schaller, K. Worthmann, and S. Streif, Phys- ica D476, 134651 (2025)

    work page 2025

  27. [27]

    Wang and Y

    S. Wang and Y. Lan, Phys. Rev. E104, 034211 (2021)

    work page 2021

  28. [28]

    Hu and Y

    J. Hu and Y. Lan, Phys. Rev. E102, 062216 (2020)

    work page 2020

  29. [29]

    Dylewsky, M

    D. Dylewsky, M. Tao, and J. N. Kutz, Phys. Rev. E99, 063311 (2019)

    work page 2019

  30. [30]

    M. J. Colbrook, SIAM J. Numer. Anal.61, 1585 (2023)

    work page 2023

  31. [31]

    P. J. Baddoo, B. Herrmann, B. J. McKeon, J. Nathan Kutz, and S. L. Brunton, Proc. R. Soc. A479, 20220576 (2023)

    work page 2023

  32. [32]

    Chen and Y

    B. Chen and Y. Wang, Proc. R. Soc. A481, 20240437 (2025)

    work page 2025

  33. [33]

    Ristich, L

    E. Ristich, L. Zhang, Y. Ren, and J. Sun, in2025 IEEE International Conference on Robotics and Automation (ICRA)(2025) pp. 9273–9279

    work page 2025

  34. [34]

    I. Ohta, S. Koyanagi, K. Kinjo, and J. Ohkubo, J. Phys. Soc. Jpn.94, 064001 (2025)

    work page 2025

  35. [35]

    Ohkubo and Y

    J. Ohkubo and Y. Arai, J. Stat. Mech.2019, 063202 (2019)

    work page 2019

  36. [36]

    Ohkubo, J

    J. Ohkubo, J. Phys. A55, 224007 (2022)

    work page 2022

  37. [37]

    Zhang, C

    H. Zhang, C. W. Rowley, E. A. Deem, and L. N. Cattafesta, SIAM J. Appl. Dyn. Syst.18, 1586 (2019)

    work page 2019

  38. [38]

    Takahashi and J

    S. Takahashi and J. Ohkubo, J. Stat. Phys.190, 159 (2023)

    work page 2023

  39. [39]

    I. V. Oseledets, SIAM J. Sci. Comput.33, 2295 (2011)

    work page 2011

  40. [40]

    N¨ uske, P

    F. N¨ uske, P. Gelß, S. Klus, and C. Clementi, Physica D 427, 133018 (2021)

    work page 2021

  41. [41]

    L¨ ucke and F

    M. L¨ ucke and F. N¨ uske, J. Nonlinear Sci.32, 44 (2022)

    work page 2022

  42. [42]

    Q. Li, F. Dietrich, E. M. Bollt, and I. G. Kevrekidis, Chaos27, 103111 (2017)

    work page 2017