pith. sign in

arxiv: 2604.05633 · v1 · submitted 2026-04-07 · 📡 eess.SY · cs.SY· math.OC

Optimality Robustness in Koopman-Based Control

Yicheng Lin , Bingxian Wu , Nan Bai , Yunxiao Ren , Zhongkui Li , Zhisheng Duan This is my paper

Pith reviewed 2026-05-10 19:18 UTC · model grok-4.3

classification 📡 eess.SY cs.SYmath.OC
keywords Koopman operatoroptimal controlrobustnessuncertainty boundspolicy iterationnonlinear systemsdata-driven controlvalue function
0
0 comments X

The pith

Uncertainties in Koopman-based nonlinear control induce bounded deviations in optimal controllers and value functions, which a robustness-aware design can reduce.

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

The paper aims to quantify how uncertainties in data-driven Koopman models affect the optimal controller and value function in nonlinear control problems. It derives explicit upper bounds on these deviations by representing all uncertainty sources in a single norm-bounded way. A robustness-aware control method is developed that reduces these deviations, showing a trade-off with nominal performance. A policy iteration algorithm is proposed and proven convergent using regularization and PDE techniques. If correct, this would allow more reliable design of data-driven controllers for uncertain systems such as vehicles or chemical processes.

Core claim

We derive explicit upper bounds on the deviations of both the value function and the optimal controller, where uncertainties from multiple sources are systematically integrated into a unified norm-bounded representation. We develop a robustness-aware optimal control methodology that provably reduces such optimality deviations, thereby enhancing robustness while explicitly revealing a quantitative trade-off between nominal optimality and robustness. We further propose a tractable policy iteration algorithm, whose well-posedness and convergence are established via vanishing viscosity regularization and elliptic partial differential equation techniques.

What carries the argument

The unified norm-bounded representation of uncertainties from multiple sources, which carries the derivation of explicit deviation bounds and enables the robustness-aware design.

Load-bearing premise

Uncertainties from multiple sources can be systematically integrated into a unified norm-bounded representation, and vanishing viscosity regularization combined with elliptic PDE techniques guarantees well-posedness and convergence of the policy iteration algorithm.

What would settle it

A numerical simulation with combined uncertainty sources in which the actual deviation of the computed optimal controller exceeds the derived explicit upper bound.

Figures

Figures reproduced from arXiv: 2604.05633 by Bingxian Wu, Nan Bai, Yicheng Lin, Yunxiao Ren, Zhisheng Duan, Zhongkui Li.

Figure 1
Figure 1. Figure 1: The relative changes of value function coefficients an [PITH_FULL_IMAGE:figures/full_fig_p013_1.png] view at source ↗
Figure 3
Figure 3. Figure 3: Quantification of performance-robustness trade-off [PITH_FULL_IMAGE:figures/full_fig_p014_3.png] view at source ↗
Figure 4
Figure 4. Figure 4: Comparison of the optimality deviations of the robus [PITH_FULL_IMAGE:figures/full_fig_p014_4.png] view at source ↗
read the original abstract

The Koopman operator enables simplified representations for nonlinear systems in data-driven optimal control, but the accompanying uncertainties inevitably induce deviations in the optimal controller and associated value function. This raises a distinct and fundamental question on optimality robustness, specifically, how uncertainties affect the optimal solution itself. To address this problem, we adopt a unified analysis-to-design perspective for systematically quantifying and improving optimality robustness. At the analysis level, we derive explicit upper bounds on the deviations of both the value function and the optimal controller, where uncertainties from multiple sources are systematically integrated into a unified norm-bounded representation. At the design level, we develop a robustness-aware optimal control methodology that provably reduces such optimality deviations, thereby enhancing robustness while explicitly revealing a quantitative trade-off between nominal optimality and robustness. As for practical implementation aspect, we further propose a tractable policy iteration algorithm, whose well-posedness and convergence are established via vanishing viscosity regularization and elliptic partial differential equation (PDE) techniques. Numerical examples validate the theoretical findings and demonstrate the effectiveness of proposed methodology.

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

0 major / 3 minor

Summary. The manuscript develops a framework for optimality robustness in Koopman-based optimal control of nonlinear systems. At the analysis stage it derives explicit upper bounds on deviations of the value function and optimal controller by folding multiple uncertainty sources into a single norm-bounded representation. At the design stage it introduces a robustness-aware optimal-control formulation that provably reduces these deviations while exposing an explicit trade-off between nominal optimality and robustness. For implementation it proposes a policy-iteration algorithm whose well-posedness and convergence are established via vanishing-viscosity regularization together with elliptic PDE theory. Numerical examples are presented to illustrate the theoretical claims.

Significance. If the derivations hold, the work supplies a quantitative, unified treatment of optimality deviations under Koopman uncertainty that is currently missing from the data-driven control literature. The explicit bounds and the optimality-robustness trade-off give designers a concrete handle on performance degradation, while the PDE-based convergence argument for policy iteration strengthens the theoretical grounding of the algorithm. These contributions would be of direct interest to researchers working on robust Koopman control for uncertain nonlinear plants.

minor comments (3)
  1. [Analysis section] The transition from the unified norm-bounded uncertainty model to the explicit deviation bounds (presumably in the analysis section) would benefit from an additional sentence clarifying how the individual uncertainty norms are combined; the current wording leaves the aggregation step implicit.
  2. [Notation and preliminaries] Notation for the robustness-aware cost functional and the associated HJB equation is introduced without a consolidated symbol table; adding one would improve readability for readers who are not already immersed in robust HJB literature.
  3. [Numerical examples] The numerical examples section would be strengthened by reporting the precise values of the uncertainty bounds used in each example and by including a direct comparison of the nominal versus robustness-aware controllers on the same plot.

Simulated Author's Rebuttal

0 responses · 0 unresolved

We thank the referee for the careful reading and the positive overall assessment, which accurately captures the analysis-to-design framework, explicit bounds, optimality-robustness trade-off, and PDE-based convergence argument for the policy-iteration algorithm. The recommendation for minor revision is noted.

Circularity Check

0 steps flagged

No significant circularity detected

full rationale

The paper's core claims consist of deriving explicit deviation bounds under a unified norm-bounded uncertainty model, constructing a robustness-aware optimal control design that trades nominal optimality for reduced deviation, and proposing a policy-iteration algorithm whose well-posedness and convergence are established using vanishing-viscosity regularization together with standard elliptic PDE theory. These elements rely on external mathematical techniques (norm-bounded uncertainty representations and PDE regularization) rather than reducing by construction to fitted parameters, self-definitions, or self-citation chains within the paper. No load-bearing step equates a prediction or result to its own inputs via the paper's equations.

Axiom & Free-Parameter Ledger

0 free parameters · 2 axioms · 0 invented entities

Based solely on the abstract, the central claim rests on standard domain assumptions about uncertainty representation and PDE regularization; no free parameters or invented entities are explicitly introduced.

axioms (2)
  • domain assumption Uncertainties from multiple sources can be systematically integrated into a unified norm-bounded representation
    Stated at the analysis level for deriving upper bounds on deviations.
  • domain assumption Vanishing viscosity regularization and elliptic PDE techniques establish well-posedness and convergence of the policy iteration algorithm
    Invoked for the practical implementation aspect to guarantee algorithm properties.

pith-pipeline@v0.9.0 · 5493 in / 1624 out tokens · 148677 ms · 2026-05-10T19:18:31.761085+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

44 extracted references · 44 canonical work pages

  1. [1]

    Linear optimal control of bilinear systems with applications to singular perturbati ons and weak coupling

    Zijad Aganovi and Zoran Gaji. Linear optimal control of bilinear systems with applications to singular perturbati ons and weak coupling . Springer, 1995

    work page 1995

  2. [2]

    Nonlinear H-infinity-control, Hamiltonian systems and Hamilton-Jacobi equations

    MDS Aliyu. Nonlinear H-infinity-control, Hamiltonian systems and Hamilton-Jacobi equations . CRC, 2011

    work page 2011

  3. [3]

    Optimal Control and Viscosity Solutions of Hamilton-Jacobi-Bellm an Equations

    Martino Bardi and Italo Capuzzo-Dolcetta. Optimal Control and Viscosity Solutions of Hamilton-Jacobi-Bellm an Equations. Systems & Control: Foundations & Applications. Birkh¨ auser Boston, Boston, MA, 1997

    work page 1997

  4. [4]

    An Introduction to the Theory of Viscosity Solutions for First-Order Hamilton–Jacobi Equations and Applications, pages 49–109

    Guy Barles. An Introduction to the Theory of Viscosity Solutions for First-Order Hamilton–Jacobi Equations and Applications, pages 49–109. Springer Berlin Heidelberg, Berlin, Heidelberg, 2013

    work page 2013

  5. [5]

    Bertsekas

    Dimitri P. Bertsekas. Dynamic Programming and Optimal Control. Athena Scientific, Belmont, MA, 1995. 2 vols

    work page 1995

  6. [6]

    Reinforcement learning and adaptive optimal control for continuous-time nonlinea r systems: A value iteration approach

    Tao Bian and Zhong-Ping Jiang. Reinforcement learning and adaptive optimal control for continuous-time nonlinea r systems: A value iteration approach. IEEE Transactions on Neural Networks and Learning Systems , 33(7):2781–2790, 2022

    work page 2022

  7. [7]

    Data- driven control via Petersen’s lemma

    Andrea Bisoffi, Claudio De Persis, and Pietro Tesi. Data- driven control via Petersen’s lemma. Automatica, 145:110537, 2022

    work page 2022

  8. [8]

    Linear quadratic control of nonlinear systems with Koopman operator learning and the Nystr¨ om method

    Edoardo Caldarelli, Antoine Chatalic, Adri` a Colom´ e, Cesare Molinari, Carlos Ocampo-Martinez, Carme Torras, and Lorenzo Rosasco. Linear quadratic control of nonlinear systems with Koopman operator learning and the Nystr¨ om method. Automatica, 177:112302, 2025

    work page 2025

  9. [9]

    An iterative method for optimal feedback control and generalized hjb equation

    Xuesong Chen and Xin Chen. An iterative method for optimal feedback control and generalized hjb equation. IEEE/CAA Journal of Automatica Sinica , 5(5):999–1006, 2018

    work page 2018

  10. [10]

    Crandall, Hitoshi Ishii, and Pierre-Louis L ions

    Michael G. Crandall, Hitoshi Ishii, and Pierre-Louis L ions. User’s guide to viscosity solutions of second order partial differential equations. Bulletin of the American Mathematical Society, 27(1):1–67, 1992

    work page 1992

  11. [11]

    Recent developments in numerical methods for fully nonline ar second order partial differential equations

    Xiaobing Feng, Roland Glowinski, and Michael Neilan. Recent developments in numerical methods for fully nonline ar second order partial differential equations. SIAM Review , 55(2):205–267, 2013

    work page 2013

  12. [12]

    Direct data-driven stabiliza tion of nonlinear affine systems via the Koopman operator

    Xingyun Fu and Keyou You. Direct data-driven stabiliza tion of nonlinear affine systems via the Koopman operator. In 2022 IEEE 61st Conference on Decision and Control (CDC) , pages 2668–2673, 2022

    work page 2022

  13. [13]

    Trudinger

    David Gilbarg and Neil S. Trudinger. Elliptic Partial Differential Equations of Second Order . Classics in Mathematics. Springer-Verlag, Berlin, second edition, 20 01

    work page

  14. [14]

    Debdipta Goswami and Derek A. Paley. Bilinearization, reachability, and optimal control of control-affine nonline ar systems: A Koopman spectral approach. IEEE Transactions on Automatic Control , 67(6):2715–2728, 2022

    work page 2022

  15. [15]

    Solution of the continuous time bilinear quadratic regulator problem by Krotov’s method

    Ido Halperin. Solution of the continuous time bilinear quadratic regulator problem by Krotov’s method. IEEE Transactions on Automatic Control , 68(4):2415–2421, 2023

    work page 2023

  16. [16]

    Two Roads to

    Masih Haseli, Igor Mezi´ c, and Jorge Cort´ es. Two roads to Koopman operator theory for control: Infinite input sequences and operator families, 2025. arXiv:2510.15166 [cs.RO]

    work page arXiv 2025

  17. [17]

    A convex approach to data-driven optimal control via perron-frobenius and koopman operators

    Bowen Huang and Umesh Vaidya. A convex approach to data-driven optimal control via perron-frobenius and koopman operators. IEEE Transactions on Automatic Control, 67(9):4778–4785, 2022

    work page 2022

  18. [18]

    Iacob, Roland T´ oth, and Maarten Schoukens

    Lucian C. Iacob, Roland T´ oth, and Maarten Schoukens. Optimal synthesis of LTI Koopman models for nonlinear systems with inputs. In 5th IF AC Workshop on Linear Parameter Varying Systems (LPVS 2022) , IF AC- PapersOnLine, pages 49–54, 2022

    work page 2022

  19. [19]

    B. O. Koopman. Hamiltonian systems and transformation in Hilbert space. Proceedings of the National Academy of Sciences, 17(5):315–318, 1931

    work page 1931

  20. [20]

    On convergence of extended dynamic mode decomposition to the Koopman operator

    Milan Korda and Igor Mezi´ c. On convergence of extended dynamic mode decomposition to the Koopman operator. Journal of Nonlinear Science , 28(2):687–710, 2017

    work page 2017

  21. [21]

    Linear predictors for nonlinear dynamical systems: Koopman operator meets model predictive control

    Milan Korda and Igor Mezi´ c. Linear predictors for nonlinear dynamical systems: Koopman operator meets model predictive control. Automatica, 93:149–160, 2018

    work page 2018

  22. [22]

    Optimal construction of Koopman eigenfunctions for prediction and control

    Milan Korda and Igor Mezi´ c. Optimal construction of Koopman eigenfunctions for prediction and control. IEEE Transactions on Automatic Control, 65(12):5114–5129, 2020

    work page 2020

  23. [23]

    Koopman-based policy iteration for robust optimal control

    Alexander Krolicki, Sarang Sutavani, and Umesh Vaidya . Koopman-based policy iteration for robust optimal control . In 2022 American Control Conference (ACC) , pages 1317– 1322, 2022

    work page 2022

  24. [24]

    Kvalheim and Shai Revzen

    Matthew D. Kvalheim and Shai Revzen. Existence and uniqueness of global Koopman eigenfunctions for stable fixe d points and periodic orbits. Physica D: Nonlinear Phenomena, 425:132959, 2021

    work page 2021

  25. [25]

    Integral reinforcement learning for continuous-time input-affine nonlinear systems with simultaneous invariant exploratio ns

    Jae Young Lee, Jin Bae Park, and Yoon Ho Choi. Integral reinforcement learning for continuous-time input-affine nonlinear systems with simultaneous invariant exploratio ns. IEEE Transactions on Neural Networks and Learning Systems, 26(5):916–932, 2015

    work page 2015

  26. [26]

    Jaeyoung Lee and Richard S. Sutton. Policy iterations f or reinforcement learning problems in continuous time and spa ce - fundamental theory and methods. Automatica, 126:109421, 2021

    work page 2021

  27. [27]

    Lewis, Draguna Vrabie, and Kyriakos G

    Frank L. Lewis, Draguna Vrabie, and Kyriakos G. Vamvoudakis. Reinforcement learning and feedback control : Using natural decision methods to design optimal adaptive controllers. IEEE Control Systems Magazine , 32(6):76–105, 2012

    work page 2012

  28. [28]

    Optimality deviation using the Koopman operator, 2025

    Yicheng Lin, Bingxian W u, Nan Bai, Yunxiao Ren, and Zhisheng Duan. Optimality deviation using the Koopman operator, 2025. arXiv:2512.10270 [math.OC]

    work page arXiv 2025

  29. [29]

    Integrating uncertainties for Koopman-based stabilizati on,

    Yicheng Lin, Bingxian W u, Nan Bai, Zhiyong Sun, Yunxiao Ren, Chuanze Chen, and Zhisheng Duan. Integrating uncertainties for Koopman-based stabilizati on,

    work page

  30. [30]

    arXiv:2508.11533 [eess.SY]

    work page arXiv

  31. [31]

    Giorgos Mamakoukas, Stefano Di Cairano, and Abraham P. Vinod. Robust model predictive control with data-driven Koopman operators. In 2022 American Control Conference (ACC), pages 3885–3892, 2022

    work page 2022

  32. [32]

    A spectral operator- theoretic framework for global stability

    Alexandre Mauroy and Igor Mezi´ c. A spectral operator- theoretic framework for global stability. In 52nd IEEE Conference on Decision and Control , pages 5234–5239, 2013

    work page 2013

  33. [33]

    Global stability ana lysis using the eigenfunctions of the Koopman operator

    Alexandre Mauroy and Igor Mezi´ c. Global stability ana lysis using the eigenfunctions of the Koopman operator. IEEE Transactions on Automatic Control, 61(11):3356–3369, 2016. 15

    work page 2016

  34. [34]

    Koopman operator in systems and control

    Alexandre Mauroy, Y Susuki, and Igor Mezic. Koopman operator in systems and control . Springer, 2020

    work page 2020

  35. [35]

    Autogeneration of mission-oriented robot controllers using Bayesian-based Koopman operator

    Jie Pan, Dongyue Li, Jian W ang, Pengfei Zhang, Jinyan Sh ao, and Junzhi Yu. Autogeneration of mission-oriented robot controllers using Bayesian-based Koopman operator. IEEE Transactions on Robotics , 40:903–918, 2024

    work page 2024

  36. [36]

    Deep robust Koopman learning from noisy data, 2026

    Aditya Singh, Rajpal Singh, and Jishnu Keshavan. Deep robust Koopman learning from noisy data, 2026. arXiv:2601.01971 [cs.RO]

    work page arXiv 2026

  37. [37]

    Kernel-based error bounds of bilinear Koopman surrogate models for nonlinear data- driven control

    Robin Str¨ asser, Manuel Schaller, Julian Berberich, K arl W orthmann, and Frank Allg¨ ower. Kernel-based error bounds of bilinear Koopman surrogate models for nonlinear data- driven control. IEEE Control Systems Letters , 9:1892–1897, 2025

    work page 2025

  38. [38]

    Koopman-based feedback design with stability guarantees

    Robin Str¨ asser, Manuel Schaller, Karl W orthmann, Jul ian Berberich, and Frank Allg¨ ower. Koopman-based feedback design with stability guarantees. IEEE Transactions on Automatic Control, pages 1–16, 2024

    work page 2024

  39. [39]

    An overview of Koopman-based control: From error bounds to closed-loop guarantees

    Robin Str¨ asser, Karl W orthmann, Igor Mezi´ c, Julian Berberich, Manuel Schaller, and Frank Allg¨ ower. An overview of Koopman-based control: From error bounds to closed-loop guarantees. Annual Reviews in Control , 61:101035, 2026

    work page 2026

  40. [40]

    When Koopman meets Hamilton and Jacobi

    Umesh Vaidya. When Koopman meets Hamilton and Jacobi. IEEE Transactions on Automatic Control, 70(12):7843–7858, 2025

    work page 2025

  41. [41]

    Jones, and Boris Houska

    Mario Eduardo Villanueva, Colin N. Jones, and Boris Houska. Towards global optimal control via Koopman lifts. Automatica, 132:109610, 2021

    work page 2021

  42. [42]

    Resilient formation control with Koopman operator for networked NMRs under denial-of-service attacks

    W eiwei Zhan, Zhiqiang Miao, Hui Zhang, Yanjie Chen, Zheng-Guang W u, W ei He, and Yaonan W ang. Resilient formation control with Koopman operator for networked NMRs under denial-of-service attacks. IEEE Transactions on Systems, Man, and Cybernetics: Systems , 54(11):7065– 7078, 2024

    work page 2024

  43. [43]

    Robust tube-based model predictive control with Koopman operators

    Xinglong Zhang, W ei Pan, Riccardo Scattolini, Shuyou Y u, and Xin Xu. Robust tube-based model predictive control with Koopman operators. Automatica, 137:110114, 2022. A Proof of Proposition 1 The full row rank ofW0 ensures A =W0W ⊤ 0 ≻ 0, whose pseudo-inverse satisfies W † 0 = W ⊤ 0 (W0W ⊤ 0 )− 1. For any element in C0, ˜Z⊤ = [A B0 B1 ... B m] = ( Z1)W † ...

    work page 2022

  44. [44]

    We can prove that eigenvalues of Qp are 0 and 1

    (A.1) Define Qp =W ⊤ 0 (W0W ⊤ 0 )− 1W0 ∈ RT × T , which is actu- ally a projection matrix. We can prove that eigenvalues of Qp are 0 and 1. Specifically, for λ p = 0, any non- zero vector v ∈ Nul(W0) is an eigenvector of Qp. For λ p = 1, any non-zero vector v ∈ Col(W ⊤ 0 ) is an eigen- vector of Qp, since any 0 ̸= v = W ⊤ 0 w ∈ Col(W ⊤ 0 ) al- lows Qpv = W ...

    work page