Pith. sign in

REVIEW 1 major objections 32 references

Residual-Guided Dictionary Learning for Spectrally Accurate Koopman Approximation

T0 review · 1 major / 0 minor · reviewed 2026-06-30 · grok-4.3

Challenge this review Re-run · record.json Download PDF Read on arXiv ↗ Review JSON

Pith's one-line read Neural dictionaries trained to minimize Residual DMD errors produce Koopman approximations with less spectral pollution and more reliable eigenvalues.

desk verdict The paper folds Residual DMD residuals plus a conditioning penalty into neural dictionary training for Koopman spectra, with decent benchmark gains, but the claim that residuals certify genuine infinite-dimensional spectral objects is not backed by a bound or argument. read the letter →

arxiv 2606.29083 v1 pith:IFGSCCZJ submitted 2026-06-27 math.NA cs.LGcs.NAmath.DS

George Coote , Matthew J. Colbrook This is my paper · ORCID
classification math.NAcs.LGcs.NAmath.DS
keywords KoopmanoperatordictionarylearningresidualDMDspectralapproximationneuralnetworksdynamicmodedecompositionnumericalstabilityforecasting
verification ladder T0 review T1 audit T2 compute T3 formal

The pith

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

The reading

The paper trains neural-network dictionaries for Koopman operator approximation by minimizing Residual Dynamic Mode Decomposition residuals instead of prediction error alone. These residuals serve as operator-level a-posteriori checks that test whether computed eigenvalues and modes are genuine spectral objects of the infinite-dimensional operator. A penalty on the condition number of the lifted data matrix is added to prevent collapse into unstable coordinates. The resulting dictionaries are shown to reduce spectral pollution, tighten residual pseudospectral inclusion, and improve forecast accuracy on both benchmark systems and real sea-surface temperature data.

What carries the argument

Residual DMD residuals, used as a-posteriori operator errors that certify whether eigenvalues and modes are genuine Koopman spectral objects, together with a condition-number penalty on the lifted data matrix.

What would settle it

On a benchmark system whose Koopman spectrum is known analytically, observe whether eigenvalues produced by the residual-trained dictionary still deviate from the true spectrum even when the reported residuals are small.

Watch

Extended reading notes

Core claim

By making minimization of Residual DMD residuals the training objective and coupling it with a condition-number penalty, the learned dictionary becomes expressive, numerically stable, and spectrally disciplined, so that its finite-dimensional eigenvalues and modes more closely reflect the spectrum of the underlying infinite-dimensional Koopman operator.

Load-bearing premise

That small Residual DMD residuals reliably indicate that the computed eigenvalues and modes are true spectral features of the infinite-dimensional Koopman operator rather than artifacts of the finite dictionary.

Editorial extensions

If this is right

  • Spectral pollution is sharply reduced on both conservative and dissipative benchmark systems.
  • Residual pseudospectral inclusion improves, tightening the set of candidate eigenvalues.
  • One-step forecast error decreases relative to standard fixed dictionaries.
  • Koopman diagnostics become cleaner and one-step forecasts improve on noisy sea-surface temperature observations without known governing equations.

Reading between the lines

Editorial extensions of the paper, not claims the author makes directly.

  • The same residual objective could be applied to other linearization methods beyond EDMD to certify spectral objects.
  • If residuals remain small on new data streams, the dictionary may transfer across related dynamical regimes without retraining.
  • The conditioning penalty suggests a general template for dictionary learning that balances expressivity against numerical stability in other operator-learning settings.
Share X Bluesky LinkedIn Reddit HN

Editorial analysis

A structured set of objections, weighed in public.

Desk editor's note, referee report, simulated authors' rebuttal, and a circularity audit.

Referee Report

1 major / 0 minor

Summary. The manuscript proposes Residual-Guided Dictionary Learning, in which neural-network dictionaries for EDMD are trained by minimizing Residual DMD residuals (operator-level a-posteriori errors) together with a penalty on the condition number of the lifted data matrix. The central claim is that this produces dictionaries that are expressive, stable, and spectrally disciplined, sharply reducing spectral pollution, improving residual pseudospectral inclusion, and lowering forecast error on conservative/dissipative benchmarks as well as on sea-surface temperature data.

Significance. If the residual-based objective can be shown to furnish a reliable a-posteriori certificate that computed eigenvalues and modes are close to those of the infinite-dimensional Koopman operator (rather than merely consistent inside the learned finite span), the approach would supply a principled alternative to prediction-error-only dictionary learning and strengthen the trustworthiness of numerical Koopman spectra.

major comments (1)
  1. [Abstract] Abstract: the assertion that Residual DMD residuals 'test whether computed eigenvalues and modes are genuine Koopman spectral objects' is load-bearing for the central claim, yet the provided description supplies no a-posteriori error bound relating the finite-dictionary residual to the distance from the true infinite-dimensional spectrum; the residual is formed from the same lifted data matrix used to build the EDMD operator, so small residuals certify invariance under the empirical operator but do not automatically imply proximity to true eigenfunctions.

Simulated Author's Rebuttal

1 responses · 0 unresolved

We thank the referee for the careful reading and for highlighting this important distinction regarding the interpretation of Residual DMD residuals. We address the major comment below.

read point-by-point responses

  1. Referee: [Abstract] Abstract: the assertion that Residual DMD residuals 'test whether computed eigenvalues and modes are genuine Koopman spectral objects' is load-bearing for the central claim, yet the provided description supplies no a-posteriori error bound relating the finite-dictionary residual to the distance from the true infinite-dimensional spectrum; the residual is formed from the same lifted data matrix used to build the EDMD operator, so small residuals certify invariance under the empirical operator but do not automatically imply proximity to true eigenfunctions.

    Authors: We agree with the referee that the residual, being formed from the same lifted data matrix, certifies approximate invariance under the empirical EDMD operator and does not supply a rigorous a-posteriori bound on the distance to the spectrum of the infinite-dimensional Koopman operator. The manuscript does not derive or claim such a bound. Nevertheless, minimizing the residual during dictionary learning selects observables for which the finite-dimensional operator is more consistent with the observed data in a spectral sense; this is what produces the observed reduction in spectral pollution and improved pseudospectral inclusion in the experiments. We will revise the abstract (and the corresponding claim in the introduction) to state that the residuals provide a certificate of consistency with the empirical operator, which in practice yields dictionaries whose spectra are more reliable, without asserting that they directly test genuineness with respect to the infinite-dimensional operator. revision: yes

Circularity Check

0 steps flagged · score 0.0 of 10

No significant circularity in derivation chain

full rationale

The paper defines a training objective that minimizes Residual DMD residuals (computed from the finite lifted data matrix) while penalizing the condition number, then validates improved spectral properties and forecast accuracy via direct comparison to fixed dictionaries on benchmark systems and real data. The claim that residuals serve as a posteriori certificates for genuine Koopman spectral objects is an interpretive premise supported by empirical outcomes rather than a reduction by construction to already-fitted quantities. No self-citations, uniqueness theorems, or ansatzes from prior author work appear in the provided text, and the central method does not rename known results or force predictions from inputs. The derivation remains self-contained against external benchmarks.

Assumptions & free parameters 0 free parameters · 0 assumptions · 0 invented entities

Abstract-only review supplies no explicit free parameters, axioms, or invented entities; the method implicitly assumes that Residual DMD residuals certify spectral membership and that the neural network can be trained to a stable minimum without additional regularization details being stated.

how reviews work

0 comments
Cite this review

Pith. "Pith review of Residual-Guided Dictionary Learning for Spectrally Accurate Koopman Approximation." pith.science (2026). https://pith.science/paper/IFGSCCZJ

@misc{pith2026260629083,
  author       = {Pith},
  title        = {Pith review of: Residual-Guided Dictionary Learning for Spectrally Accurate Koopman Approximation},
  year         = {2026},
  howpublished = {\url{https://pith.science/paper/IFGSCCZJ}},
  note         = {Machine review of arXiv:2606.29083}
}
read the original abstract

Koopman theory promises linear structure in nonlinear dynamics, but numerical Koopman spectra are easy to compute and hard to trust. A finite EDMD matrix always has eigenvalues; the problem is that many of them may have nothing to do with the infinite-dimensional operator. In this paper we make spectral reliability the objective of dictionary learning. We train neural-network dictionaries not merely to predict the next snapshot, but to minimize Residual Dynamic Mode Decomposition residuals: operator-level a posteriori errors that test whether computed eigenvalues and modes are genuine Koopman spectral objects. To keep the learned observables from collapsing into an unstable coordinate system, the loss also penalizes the condition number of the lifted data matrix. Thus the method couples two requirements that should not be separated: small Koopman residuals and a well-conditioned representation. The result is a learned dictionary that is expressive, numerically stable, and spectrally disciplined. Across conservative and dissipative benchmark systems, the method sharply reduces spectral pollution, improves residual pseudospectral inclusion, and lowers forecast error relative to standard fixed dictionaries. On sea-surface temperature data, it gives cleaner Koopman diagnostics and substantially better one-step forecasts from noisy observations with no governing equations. The message is simple: neural Koopman learning should be judged not by prediction alone, but by whether its spectral claims can be certified. Residuals provide the certificate; conditioning makes it computable.

Discussion (0). Continue with ORCID to comment.

Reference graph

Works this paper leans on

32 extracted references · 32 canonical work pages

  1. [1]

    Benjamin Erichson, Vanessa Lin, and Michael Mahoney

    Omri Azencot, N. Benjamin Erichson, Vanessa Lin, and Michael Mahoney. Forecasting sequential data using consistent Koopman autoencoders. InProceedings of the 37th International Conference on Machine Learning, volume 119 ofProceedings of Machine Learning Research, pages 475–485. PMLR, 2020

    work page 2020

  2. [2]

    Bollt, Felix Dietrich, Ioannis G

    Erik M. Bollt, Felix Dietrich, Ioannis G. Kevrekidis, and Qianxiao Li. Extended dynamic mode decomposition with dictionary learning: a data-driven adaptive spectral decomposition of the Koopman operator.Chaos, 27, 2017

    work page 2017

  3. [3]

    Brunton, Bingni W

    Steven L. Brunton, Bingni W. Brunton, Joshua L. Proctor, and J. Nathan Kutz. Koopman invariant subspaces and finite linear representations of nonlinear dynamical systems for control.PLOS ONE, 11(2):e0150171, 2016

    work page 2016

  4. [4]

    Brunton, Marko Budi ˇsi´c, Eurika Kaiser, and J

    Steven L. Brunton, Marko Budi ˇsi´c, Eurika Kaiser, and J. Nathan Kutz. Modern Koopman theory for dynamical systems.SIAM Review, 64(2):229–340, 2022

    work page 2022

  5. [5]

    Brunton, Bernd R

    Steven L. Brunton, Bernd R. Noack, and Petros Koumoutsakos. Machine learning for fluid mechanics.Annual Review of Fluid Mechanics, 52:477–508, 2020

    work page 2020

  6. [6]

    Colbrook

    Matthew J. Colbrook. Another look at residual dynamic mode decomposition in the regime of fewer snapshots than dictionary size.Physica D: Nonlinear Phenomena, 469:134341, 2024

    work page 2024

  7. [7]

    Colbrook.Infinite-Dimensional Spectral Computations: Foundations, Algorithms, and Modern Applications

    Matthew J. Colbrook.Infinite-Dimensional Spectral Computations: Foundations, Algorithms, and Modern Applications. Cambridge University Press, Cambridge, 2026

    work page 2026

  8. [8]

    Colbrook, Lorna J

    Matthew J. Colbrook, Lorna J. Ayton, and M ´at´e Sz˝oke. Residual dynamic mode decomposition: Robust and verified Koopmanism.Journal of Fluid Mechanics, 955:A21, 2023

    work page 2023

Show all 32 references
  1. [9]

    An introductory guide to Koopman learning

    Matthew J Colbrook, Zlatko Drma ˇc, and Andrew Horning. An introductory guide to Koopman learning. In Operator Theory, pages 1–49. Springer Basel, 2026

    2026

  2. [10]

    Colbrook, Qin Li, Ryan V

    Matthew J. Colbrook, Qin Li, Ryan V . Raut, and Alex Townsend. Beyond expectations: Residual dynamic mode decomposition and variance for stochastic dynamical systems.Nonlinear Dynamics, 112(3):2037–2061, 2024

  3. [11]

    Colbrook, Igor Mezi ´c, and Alexei Stepanenko

    Matthew J. Colbrook, Igor Mezi ´c, and Alexei Stepanenko. Adversarial dynamical systems characterize when data-driven learning succeeds or fails.Nature Communications, to appear

  4. [12]

    Colbrook and Alex Townsend

    Matthew J. Colbrook and Alex Townsend. Rigorous data-driven computation of spectral properties of Koopman operators for dynamical systems.Communications on Pure and Applied Mathematics, 77(1):221– 283, 2024

    2024

  5. [13]

    Froyland, D

    G. Froyland, D. Giannakis, B. Lintner, M. Pike, and J. Slawinska. Spectral analysis of climate dynamics with operator-theoretic approaches.Nature Communications, 12(1):6570, 2021

    2021

  6. [14]

    Deep sparse rectifier neural networks

    Xavier Glorot, Antoine Bordes, and Yoshua Bengio. Deep sparse rectifier neural networks. InProceedings of the 14th International Conference on Artificial Intelligence and Statistics (AISTATS), 2011. 20GEORGE COOTE AND MATTHEW COLBROOK

    2011

  7. [15]

    Haggerty, Michael J

    David A. Haggerty, Michael J. Banks, Ervin Kamenar, Alan B. Cao, Patrick C. Curtis, Igor Mezi ´c, and Elliot W. Hawkes. Control of soft robots with inertial dynamics.Science Robotics, 8(81), 2023

    2023

  8. [16]

    Haykin.Neural Networks: A Comprehensive Foundation

    Simon S. Haykin.Neural Networks: A Comprehensive Foundation. Prentice Hall, Upper Saddle River, N.J., 1999

    1999

  9. [17]

    B. Koopman. Hamiltonian systems and transformation in hilbert space.Proc. Natl. Acad. Sci. USA, 10:315–318, 1931

    1931

  10. [18]

    Koopman and J

    B. Koopman and J. Neumann. Dynamical systems of continuous spectra.Proc. Natl. Acad. Sci. USA, 18:255–263, 1932

    1932

  11. [19]

    Reskoopnet: learning Koopman representations for complex dynamics with spectral residuals

    Nikos Logothetis, Kaidi Shao, Zhongwei Shen, and Yuanchao Xu. Reskoopnet: learning Koopman representations for complex dynamics with spectral residuals. InICML’25: Proceedings of the 42nd International Conference on Machine Learning, pages 69647–69674, 2025

    2025

  12. [20]

    Residual pseudospectra reveal a physics-informed Koopman backbone for tropical pacific variability and ENSO prediction.arXiv preprint arXiv:2606.09369, 2026

    Paula Lorenzo-Sanchez, Matthew J Colbrook, and Antonio Navarra. Residual pseudospectra reveal a physics-informed Koopman backbone for tropical pacific variability and ENSO prediction.arXiv preprint arXiv:2606.09369, 2026

    2026 arXiv

  13. [21]

    Nathan Kutz, and Steven L

    Bethany Lusch, J. Nathan Kutz, and Steven L. Brunton. Deep learning for universal linear embeddings of nonlinear dynamics.Nature Communications, 9(1):4950, 2018

    2018

  14. [22]

    Spectral properties of dynamical systems, model reduction and decompositions.Nonlinear Dynamics, 41(1):309–325, 2005

    Igor Mezi ´c. Spectral properties of dynamical systems, model reduction and decompositions.Nonlinear Dynamics, 41(1):309–325, 2005

    2005

  15. [23]

    Analysis of fluid flows via spectral properties of the Koopman operator.Annual Review of Fluid Mechanics, 45(1):357–378, 2013

    Igor Mezi ´c. Analysis of fluid flows via spectral properties of the Koopman operator.Annual Review of Fluid Mechanics, 45(1):357–378, 2013

    2013

  16. [24]

    Orvieto et al

    A. Orvieto et al. Resurrecting recurrent neural networks for long sequences.Proceedings of Machine Learning Research, 202:26670–26698, 2023. Proceedings of the 40th International Conference on Machine Learning

    2023

  17. [25]

    Otto and Clarence W

    Samuel E. Otto and Clarence W. Rowley. Linearly recurrent autoencoder networks for learning dynamics. SIAM Journal on Applied Dynamical Systems, 18(1):558–593, 2019

    2019

  18. [26]

    Physics-informed probabilistic learning of linear embeddings of nonlinear dynamics with guaranteed stability.SIAM Journal on Applied Dynamical Systems, 19(1):480–509, 2020

    Shaowu Pan and Karthik Duraisamy. Physics-informed probabilistic learning of linear embeddings of nonlinear dynamics with guaranteed stability.SIAM Journal on Applied Dynamical Systems, 19(1):480–509, 2020

    2020

  19. [27]

    Proctor and P

    J. Proctor and P. Eckhoff. Discovering dynamic patterns from infectious disease data using dynamic mode decomposition.International Health, 7(2):139–145, 2015

    2015

  20. [28]

    Rowley, Igor Mezi ´c, Shervin Bagheri, Philipp Schlatter, and Dan S

    Clarence W. Rowley, Igor Mezi ´c, Shervin Bagheri, Philipp Schlatter, and Dan S. Henningson. Spectral analysis of nonlinear flows.Journal of Fluid Mechanics, 641:115–127, 2009

    2009

  21. [29]

    On data-driven unbiased predictors using the Koopman operator.arXiv preprint arXiv:2602.04720, 2026

    Roland Schurig, Pieter van Goor, Karl Worthmann, and Rolf Findeisen. On data-driven unbiased predictors using the Koopman operator.arXiv preprint arXiv:2602.04720, 2026

    2026

  22. [30]

    Learning Koopman invariant subspaces for dynamic mode decomposition

    Naoya Takeishi, Yoshinobu Kawahara, and Takehisa Yairi. Learning Koopman invariant subspaces for dynamic mode decomposition. InAdvances in Neural Information Processing Systems, volume 30, pages 1130–1140. Curran Associates, Inc., 2017

    2017

  23. [31]

    Williams, Ioannis G

    Matthew O. Williams, Ioannis G. Kevrekidis, and Clarence W. Rowley. A data-driven approximation of the Koopman operator: Extending dynamic mode decomposition.Journal of Nonlinear Science, 25(6):1307– 1346, 2015

    2015

  24. [32]

    Learning deep neural network representations for Koopman operators of nonlinear dynamical systems

    Enoch Yeung, Soumya Kundu, and Nathan Hodas. Learning deep neural network representations for Koopman operators of nonlinear dynamical systems. In2019 American Control Conference (ACC), pages 4832–4839. IEEE, 2019

    2019

Pith tools

Reviewed June 30, 2026 · model on record in the stance chip above.